Академический Документы
Профессиональный Документы
Культура Документы
Econometric Decision
Models
New Methods of Modeling and Applications
Lectu re Notes
in Economics and
Mathematical Systems
Managing Editors: M. Beckmann and W. Krelle
366
J. Gruber (Ed.)
Managing Editors
Prof. Dr. M. Beckmann
Brown University
Providence, RI 02912, USA
Editor
Prof. Dr. Josef Gruber, Ph. D.
University of Hagen
Chair of Statistics and Econometrics
Feithstr. 140, W-5800 Hagen 1, FRG
This work IS subject to copyright. All rights are reserved, whether the whole or part of the material
IS concerned, specifically the rights of translation, reprinting, re-use of Illustrations, recitation,
broadcasting, reproduction on microfilms or in other ways. and storage in data banks. Duplication
of this publication or parts thereof is only permitted under the provisions of the German Copyright
Law of September 9. 1965, in its current version, and a copyright fee must always be paid.
Violations fall under the prosecution act of the German Copyright Law.
'f') Springer·Verlag Berlin Heidelberg 1991
Originally published by Springer-Verlag Berlin Heidelberg New York in 1991.
Typesetting: Camera ready by author
42/3140-543210 - Printed on aCid-free paper
Preface
This volume contains a refereed selection of revised papers which were originally presented at the
Second International Conference on Econometric Decision Models, University of Hagen (FernUni-
versitat). The conference was held in Haus Nordhelle, a meeting place in the mountainous area
" Sauerland" , some 50 kilometers south of Hagen, on August 29 - September 1, 1989. Some details
about this conference are given in the first paper, they need not be repeated here.
The 40 papers included in this volume are organized in 10 "parts", shown in the table of contents.
Included are such "fashionable" topics like "optimal control", "cointegration" and "rational expec-
tations models". In each part, the papers have been arranged alphabetically by author, unless there
were good reasons for a different arrangement. To facilitate the decision making of the readers, all
papers (except a few short ones) contain an abstract, a list of keywords and a table of contents. At
the end of the proceedings volume, there is a list of authors.
More than ten years ago, I began to organize meetings of econometricians, mainly called "seminar"
or " colloquium". One major purpose of these meetings has always been to improve international
cooperation of econometric model builders (and model users) from "the East" and "the West".
Unprecedented changes to the better have taken place recently ("perestroika"). For a large fraction
of participants from the Soviet Union, the 1989 conference was the first conference in a Western
country.
This improvement in international scientific cooperation regrettably shows up in publications only
with a rather long lag. This proceedings volume is no exception to this rule.
I wish to thank very cordially all persons and institutions which have directly or indirectly con-
tributed to this volume. To mention but a few: the authors of the papers, the anonymous referees,
the editors of the Lecture Notes in Economics and 1Iathematical Systems and the Springer- Verlag
in Heidelberg.
I gratefully acknowledge financial support from the Deutsche Forschungsgemeinschaft (DFG), the
University of Hagen, the Gesellschaft der Freunde der FernUniversitat, Siemens AG, IBM, Com-
merzbank AG (Hohenlimburg and Elsey), Sparkasse Hagen and Sparkasse Witten. PCs for the
demonstration of software and decision support systems have been made available during the con-
ference by IBM and Siemens AG.
Last but not least I want to thank the staff members of the Chair of Statistics and Econometrics,
University of Hagen. Special thanks go to Mrs. Elke Greber, my secretary. She has shared with
me the burden of organizing the conference and of preparing this proceedings volume. Above all,
her admirable skills in text processing have shaped this proceedings volume in a very favorable way.
To alleviate poverty and misery in the world, to contribute to the sustainable existence of mankind,
decision errors in all fields of human activity have to be minimized. In a small but nonetheless
extremely important area, econometric decision models can contribute significantly to reduce de-
cision errors. Considerable gains can accrue to society from the more widespread use of improved
econometric decision models, built on the basis of better data, especially microdata. May this
proceedings volume strongly stimulate appropriate developments. We bear great responsibility!
8. On some properties of the solution of the optimal control problem for the original
long-term macroeconomic model 119
Vladimir Vasiliev
10. Optimal economic policy making with an econometric model using interactive methods
of vector optimization 149
Bernhard Bohm and Peter Brandner
VI
11. Public preferences and their role in the design of economic policy 167
Andrew Hughes Hallett
12. Least squares estimation of quadratic preference functions for econometric decision
models based on survey data 185
Hartmut Hiisges and Josef Gruber
15. Optimal dictatorial and multi-dictatorial choice in Arrow's model with applications to
multicriteria decision making 248
Andranick S. Tanguiane
17. The Finnish rational expectations QMED model: estimation, dynamic properties and
policy results 285
Ari Lahti and Matti Viren
18. Qualms about the linearized expectations hypothesis and variance-bounds studies of
the interest rate term structure 305
Gary S. Shea
19. The power function of the likelihood ratio test for cointegration 323
S{Jren Johansen
20. Long-run relations in a well-defined statistical model for the data generating process.
Cointegration analysis of the PPP and the UIP relations for Denmark and Germany 336
/(atarina Ju.seiiu.s
21. Interest rate linkages between EC countries participating in the European monetary system:
an application of cointegration 358
Wilhelm Fritz
VII
23. Tighter bounds for the effects of ARMA disturbances on tests for regression coefficients 404
Jan F. [(iviet
24. True vs. nominal size of the F-test in the linear regression model with autocorrelated
disturbances 419
Walter [(ramer, Jan [(iviet and Jorg Breitung
30. Forecast and multiplier sensitivity analysis with respect to changes of structural
equations and submodels in dynamic systems 502
Johann Elker
34. Econometric methods as an instrument for estimating external costs per unit of
emission: the case of a groundwater model 551
A/onika Vorreyer
36. Income and price policy making with an econometric model of financial incomes and
expenditures of Poland's population 570
Wieslaw Debski
39. Evaluating econometric models: the 1988 forecast of the RWI-business cycle model
in retrospect 610
Ullrich fIeilemann
40. The macroeconometric model of the Deutsche Bundesbank - a brief review 626
Max Christoph Wewel
Josef Gruber
Fern Uni versitat Hagen
Lehrgebiet Statistik und Okonometrie
Feithstr. 140, D-5800 Hagen 1, FRG
Abstract
This paper contains (with only very minor changes) the introductory remarks I made at the opening
of the Second International Conference on Econometric Decision Models, University of Hagen, held
in Haus Nordhelle, August 29 - September 1, 1989. The major objectives of the conference are
stated in section 2. In this connection some imbalances in our research efforts and in university
education are briefly discussed. Section 3 is about types of econometric decision models; some
historical facts and (in a rather selective manner) some recent developments are briefly indicated,
thus also informing about my motivation for organizing this conference. In section 4 it is stressed
that successfully constructing and applying econometric decision models requires interdisciplinary
and international cooperation, and some concrete measures are suggested.
Keywords: Types of econometric decision models; specifying scalarvalued objective functions;
multiple criteria decision making; vector optimization; E.E. Rosinger's interactive procedure; ref-
erence point approach; generalized small sample research; improved data bases.
Contents
A cordial welcome to the international audience and a plea for more cooperation
3 Types of econometric decision models: some history and selected recent developments
References
from the Federal Republic of Germany, somewhat less than one third from neighbouring countries
in the West (including the USA) and about one third from eastern countries (including the USSR).
From West Germany, there are, for example, colleagues from universities, from the "large" institutes
for economic research and, what is very important for me, from Federal Ministries in Bonn, from
the Bundesbank in Frankfurt and from the Bavarian Ministry of Agriculture in Munich.
This mixture of econometricians with different backgrounds, with different tasks in their daily life
(from purely theoretical work to purely applied econometric work) is an important condition for
a fruitful conference on econometric decision models: Econometric decision models can bear full
fruit to society only if they are seen as a multidisciplinary (an interdisciplinary) effort.
I again very cordially welcome all of you.
I am especially pleased that so many distinguished colleagues from the Soviet Union have been able
to attend this conference: Almost one fourth of all participants come from the USSR, mainly from
the large research institutes of the Academy of Sciences in Moscow, Kiev, Novosibirsk and Tashkent.
This large group of colleagues from the USSR is something new. In other international conferences
(like those of the Econometric Society) in recent years at most one percent of all participants came
from the USSR.
I hope that this large delegation from the USSR signals some important new development: the
improved possibility of international cooperation, the long-run cooperation beyond the borders
of political, social and ideological blocks. I encourage, I even urge all of you to establish strong
personal contacts during this conference, which, as a rule, are a condition for cooperation to the
mutual benefit of all partners involved.
On the last day of the conference it is exactly 50 years that Hitler-Germany began World War II.
Now the so-called cold war, which has lasted some 40 years, seems to come to an end. If this is
true large parts of the gigantic defense budgets in the west and the east can be used to fight such
everlasting enemies of mankind like illiteracy, illness, poverty, overpopulation and the destruction
of nature.
Never before in recent decades the chances have been so good for the development of long-run and
large-scale international cooperation between east and west. If this conference brought nothing but
a strong stimulus to such a cooperation in econometric modeling, it alone would already be worth
the effort and the funds invested in it.
a) \Vhat are recent developments in theory, methods, algorithms and software systems, which
are promising for one or the other phase of constructing and applying econometric decision
models?
b) What has been learned, what can be gained from solid applications of recently developed
theories and methods?
A very large portion of the papers to be presented at this conference will be relevant for
answering these two questions.
c) Are there important developments in the field of data quality and of data availability?
Since even very powerful new methods often bring only a very small return to society if
5
applied to data of poor quality, I am convinced that it is important to be concerned also with
ways and means for improving the data base for econometric modeling. The improvement
may refer to the quality of existing data sets. It may also refer to new data sets which are
required for modeling new problems (e.g. environmental problems).
I would like to mention here the following three areas of promising activities in West Germany
concerning data quality and data availability.
d) \Vhy are econometric decision models not used more frequently as an instrument for preparing
policy decisions?
Behind this question stands my conviction that such models are a powerful tool, that they
should be applied much more often.
e) \Vhat can be done to give the required training in the field of econometric decision models to
a sufficiently large number of model builders and especially to a sufficiently large number of
potential model users?
This question is based on my assertion that, above all, the lack of well-trained people hampers
the more widespread use of econometric decision models.
I guess, for example, that in the Federal and State Ministries in West Germany in which
econometric decision models should playa major role, at most one percent of all employees are
engaged in constructing and applying various types of econometric decision models. Perhaps
only additional two or three percent of the employees are sufficiently well trained so that
modeling results obtained by others can be interpreted correctly and be integrated into the
decision finding process.
What can be done? In view of the problems in public finance (huge public debts etc.), there
is only little chance that a sufficient number of new positions will be created in the ministries.
Maybe that existing positions become available because of retirement. In view of the age
6
structure of the employees in our ministries, I doubt, however, that this could really solve the
problem in the next 10 or 15 years.
What can nonetheless be done'!
I think, it is extremely important to give additional training in econometric decision models
to people already now working in places like ministries where econometric decision models
should be used more frequently.
The econometrics courses of the University of Hagen, written mainly for students working
towards an academic degree in economics, are also available for students in continuing edu-
cation. Without further effort, these econometrics courses alone will not really solve the
problem. They should probably be combined with seminars, summer schools etc. which are
designed in particular to serve the needs of people in continuing education in econometric
decision models. The only summer school on econometric models in West Germany I am
aware of was organized in 1988 for the first time by RWI, the Rhinian-Westfalian Institute
for Economic Research in Essen. I am very glad that Dr. U. Heilemann, the originator of
this summer school, is in the audience.
Let me now say a few sentences on the situation concerning the basic training in econometrics,
especially in West Germany.
1. Students of economics can complete their program of study at most German universities
without any formal training in econometrics. Since very many students of economics
avoid subjects in which more than elementary mathematics is applied, most of our
graduates in economics today still do not study econometrics. In contrast to this, in
the United States most students in economics do take econometrics courses before they
graduate.
2. In West Germany, there are no programs of study in econometrics for which an academic
degree like Bachelor of Science or Master of Science (in German called "Diplom- ... ") is
awarded. Econometrics can be chosen only as one out of four or five fields in programs
of study in economics (leading to the degree" Diplomvolkswirt", "Diplom6konom" or
"Diplomkaufmann"). As a rule, only some 10 to 20 % of our economics graduates take
econometrics as a field of study. At some universities, econometrics can also be selected
by students in informatics and mathematics. In contrast to this, in the Netherlands
there is a program of study in econometrics: Undergraduates specialize in econometrics
already in their first year and later graduate in econometrics.
3. Doctor's degrees based on a dissertation in econometrics can be earned, as a rule, only
in faculties of economics (there are no faculties of econometrics). Graduates in fields
other than economics at some places have to take additional examinations in economics,
before they can be awarded the doctor's degree.
4. Some colleagues at West German universities tend to teach mainly pure mathematical
statistics in courses labeled econometrics. This does not make econometrics attrac-
tive for students of economics. I know several leading applied econometricians who, as
economists, studied a nonmathematical introduction to econometrics. I know hardly
any leading applied econometrician who learned econometrics from a book or course on
mathematical statistics in econometrics.
For these reasons there are even today not enough young people trained well in econometrics
in Germany. This will seriously hamper the more widespread use of econometric decision
models for decades, unless special efforts are made in continuing education.
The best way to overcome this problem is to state preferences explicitly. In case of econometric
decision models this leads to optimization models with a scalarvalued objective function. For
different sets of preferences, the corresponding sets of values of instrument and target variables
can be calculated. Several sessions of this conference deal with this problem. Of great help
may also be model comparison projects like those of the Econometric Modeling Bureau at
the University of Warwick at Coventry, UK. Prof. Kenneth Wallis, the founder and director
of this Bureau and the speaker in the following session, will have to say something also on
this topic.
g) What are areas of research which are most promising, which therefore deserve more research
input?
Standard economic theory tells us that funds (e.g. research funds) are optimally allocated if
the marginal return is the same in each alternative.
I simply assert here the following:
1. The marginal return to society from research for constructing scalarvalued objective
functions for econometric decision models is large. It is probably considerably larger than
the marginal return from developing estimation and testing procedures for traditional
econometric models, especially if these new methods have to be applied to data of poor
quality.
2. The marginal return to society of research and organizational activities for improving
the data base is also large. Microdata on individual households and firms (cross section
data, panel data) can be obtained by sampling and survey methods. In many developing
countries and in all countries with unprecedented changes in society and in the economy
(like in the former East block), econometric models based on time series suffer from
structural breaks. Especially in such cases microdata collection and analyses may be of
great importance.
3. The marginal return from generalized finite sample research is considerable. The applied
econometrician who has to select an estimation method would have to rely almost entirely
on the hope that asymptotic properties of estimators in the correct model are also valid
for small samples under the correct and especially under the incorrect specification, if
no results were available from statistical experiments on the computer (Monte Carlo
studies, MC studies, for short). Such studies are also called "experimental finite sample
research" in contrast to "analytical finite sample research" (see e.g. Mittag 1987, 1989).
It seems advisable to base the choice of an estimation method not on such hopes but
rather on knowledge. Varying a theme of Leontief (1971), it can be stated: Observe
the small sample properties (SSP) of estimators under empirically relevant conditions
(like specification errors) instead of working with irrelevant theoretical assumptions and
unwarranted hopes!
MC studies playa very important role in providing this information on SSP in the near
and probably also the intermediate future. A good case in point is the study of Kramer
(1980) on the rehabilitation of the ordinary least squares estimator (OLS) in economet-
rics. See also Gruber/Fusek/Fuskova (1987). But MC studies produce a voluminous
output that needs to be analyzed statistically in order to "extract" the information
contained in the output. What is even more important: Results of MC studies are
presently often "less general", "more specific" than analytically derived results. But
the "specificity" of MC results can be decreased, their "generality" can be increased
by several measures. They essentially use knowledge about the object modeled (e.g.
about a commodity market or ilbout an economy as a whole), about properties of an
appropriate model (e.g. dynamic stability) and they apply principles and methods of
statistics (e.g. methods of sampling and of regression analysis). Some suggestions for
8
generalizing finite sample research along these lines have recently become available in
Gruber (1991). See also Hendry (1984), Mariano (1982), Phillips (1983, 1984, 1985).
This concludes my statements on major objectives of this conference. Further objectives will
be referred to below. I would be grateful if further objectives could be added and especially
if we could cooperate during the conference and afterwards to reach the objectives as closely
as possible. We bear great responsibility.
a) an econometric equation system (an econometric model in the narrow sense of this word, as
used in section 3.1) and/or
c) a system of definitional equations, identities (i.e. equations in which there are no statistically
estimated parameters) and, frequently,
d) inequality constraints (e.g. some resources can be left unused or a positive deviation of a
target variable (like employment) from a desired value is to be treated differently than a
negative deviation).
The objective function of this type of econometric decision model has, among others, the following
properties:
a) It is scalarvalued, i.e. it "aggregates" so to speak all individual target variables (like growth
of GNP, unemployment rate, inflation rate, external balance) into a single measure.
b) It models the preferences of the decision maker (or of a group of decision makers) concerning
the decision problem with the same thoroughness as the constraints model other aspects of
the decision problem.
All information required for obtaining the optimal solution (i.e. the set of optimal values of target
and instrument variables) is put into the setup of the model. The optimal solution is calculated on
the computer without interaction with the decision maker.
Econometric decision models with a scalarvalued objective function are in the tradition of the
founders of the theory of quantitative economic policy. To mention but a few names and books: Jan
Tinbergen (1952, 1956), Ragnar Frisch (1981), Henry Theil (1964), Karl A. Fox, Jati K. Sengupta
and Eric Thorbecke (1966, 1973). A landmark in the combination of this type of decision model
with control theory is the book of Gregory C. Chow (1975). Further important steps of development
are well documented, e.g. in the volumes of Benjamin M. Friedman (1975), of Gregory C. Chow
(1981), of David Kendrick (1981,1988). See also Szegii (1982), Gruber (1983), Carraro and Sartore
(1987), Rao (1987) and Sengupta and Kadekodi (1988).
1 apologize for being unable to cite corresponding names and books from eastern countries.
These and other prominent econometricians have propagated during the last four decades the use
of econometric decision models with a scalarvalued objective function. Reports on the present state
of development are given in this conference in sessions on optimal control of econometric models.
By and large, the construction of such optimization models (mainly with a quadratic objective
function and constraints in the form of linear equations) has, at least in Germany, been confined to
universities. Even here, the objective function is not based on any estimation; its parameter values
are usually obtained by assumption.
This points to a weakness of this type of econometric decision model: The modeling of preferences,
i.e. constructing scalarvalued objective functions, is much less developed or at least much less po-
pular than the modeling of the constraints (1 have already referred to this imbalance in section 2).
10
An account of the present state of affairs is given in this conference in several sessions on construc-
ting objective functions. The econometric decision models which are optimization models with a
scalarvalued objective function are mathematically closely related to mathematical programming
models (activity analysis models, models of computable general equilibrium). See e.g. Stephan
(1989) and literature stated there.
In modeling oflarge-scale policy problems, one often finds a combination of such models, of econo-
metric models and of input-output models (on the latter, see e.g. Toyomane 1988).
Gruber (1986), Olbrisch (1988) and Barabas (1991, this proceedings volume).
• a sound knowledge of theory and subject matter (including data bases) in the field of appli·
cation (e.g. economics, ecology, engineering)
That is:
"Econometric Society econometrics" which is emphasizing almost exclusively mathematical sta-
tistics and mathematical economic theory is not enough. Interdisciplinary cooperation is required
for modeling problems of economic (or other) policy. Several developments in hardware and soft-
ware for handling large models facilitate such a cooperation. The political developments have never
before been so favorable for international cooperation; they have never before required so much in-
ternational cooperation. For example, new economic cooperation among regions and states requires
also new cooperation among model builders and model users.
This conference should help not to ignore in the future so much of the work of colleagues abroad.
How many econometricians from, say, West Germany have ever taken into account in their research
and teaching the publications of colleagues from, say, the Soviet Union? I fear that an investigation
of this question would show a disastrous situation. We almost entirely ignore what is not available
in one's mother tongue or in English.
That we ignore so many research results obtained in the Soviet Union may have many reasons which
need not all be stated here (see e.g. Gruber 1986, 1988). One major reason certainly is that in the
USSR the division of sciences into branches often differs from the division we have in "western"
countries. For example, during my ten weeks' stay in the USSR as a guest of the Academy of
Sciences I have hardly ever met the term "econometrics", neither in oral communication nor in
papers, preprints and reprints received. Another reason is that often only summaries or only rather
short papers are available in English, and are hardly" digestible". What can be done to facilitate
international cooperation of econometric model builders and model users? Only two groups of
concrete measures are mentioned here:
1. Translate full-length papers into English, always using the internationally accepted termi·
nology. In full-length papers the reasoning of the author(s) is easier to follow; they can also
contain the data input used for constructing econometric models, the structural equations in
detail (including criteria usually applied for assessing the quality), and numerical results from
forecasting and simulation. The use of the internationally accepted terminology in English can
best be guaranteed if an authoritative dictionary of statistical/econometric terms can be used
(I am not sure that such a dictionary is available e.g. in Russian/English). The translation
into English by professional interpreters who are not also specialists in econometric modeling
will hardly lead to the required legibili ty in English.
2. It deserves at least some thought how to organize an econometric modeling bureau which
works on a scale beyond individual countries. In section 2 of this paper I referred to the
12
Econometric Modeling Bureau at the University of Warwick, UK. One of its main objectives
is to compare econometric models for Great Britain. Its activities have to be transferred to
larger economic and political areas, e.g. to the European Common ~Iarket and, hopefully, to
the United States of Europe. Modeling efforts like in the HER~1ES project and in EFOM
(energy flow optimization model) may serve as a starting point for model comparison projects
beyond individual countries in Europe. On the global level the project LINK could serve a
similar purpose. Will it be desirable, necessary and/or feasible to establish e.g. a separate
modeling bureau in Europe, in the world? I don't know. Maybe it suffices to enlarge the tasks
of already existing institutions, instead of creating entirely new institutions. To continue the
examples just mentioned: In Europe something related to the European Communities may
provide the institutional frame; at the globalleyel one or more organizations of the United
Nations may serve this purpose.
The activities of international modeling bureaus would include meetings of international work-
ing groups, international seminars etc., thus also strengthening the international cooperation
of econometric model builders and model users, like this conference.
References
BARABAS, G. 1991
Implementing the reference-point approach to multi-criteria decision making in econometric models
by the computer program LOS (in this volume).
CARRARO, C., and SARTORE, D. (eds.) 1987
Developments of control theory for economic analysis. Kluwer Academic Publishers, Dordrecht,
Boston, Lancaster.
CHOW, G.C. 1975
Analysis and control of dynamic economic systems. John vViley and Sons, New York, a.e.
CHOW, G.C. 1981
Econometric analysis by control methods. John Wiley and Sons, New York a.e.
FOX, K.A., SENGUPTA, J.K., and THORBECKE, E. 1973
The theory of quantitative economic policy with applications to economic growth, stabilization and
planning. North Holland, Amsterdam, London (1. edition 1966).
FRIEDMAN, B.M. 1975
Economic stabilization policy (Methods in optimization). North Holland, Amsterdam.
FRISCH, R. 1981
From utopian theory to practical applications: the case of econometrics. American Economic
Review 71, Special Issue, Dec. 1981,1-16 (Reprint of Nobel Memorial Lecture. June 1970).
GALLER, H.P. 1991
Microeconometric research at the Sonderforschungsbereich 3 (in this volume).
GRUBER, J. (ed.) 1983
Econometric decision models. Proceedings of a Conference at the University of Hagen, June 19-20,
1981. Lecture Notes in Economics and Mathematical Systems, Vol. 208, Springer-Verlag, Berlin,
Heidelberg, New York.
GRUBER, J. 1983a
Das Referenzpunkt- Verfahren zur interakti yen Vektoroptimierung (Einige einfiihrende Erlauterungen
anhand eines einfachen Beispiels). Diskussionspapier NT. 74, Fachbereich Wirtschaftswissenschaft
der FernUniversitat Hagen.
GRUBER, J. 1986
llericht tiber einen ftinfwiichigen Forschungsaufenthalt in der Sowjetunion. Diskussionspapier Nr.
103, Fachbereich vVirtschaftswissenschaft der FernUniyersitiit Hagen. Gekiirzte Version veriiffentlicht
in: Allgemeines Statistisches Archiy 70, 388-409.
13
GRUBER, J. 1987
Interactive vector optimization as a complement to optimal control in econometric models. In:
Carraro, C. and Sartore, D. (eds.) Developments of control theory for economic analysis. Kluwer
Academic Publishers, Dordrecht, Boston, Lancaster, 63-82.
GRUBER, J. 1988
Bericht iiber einen dreiwochigen Forschungsaufenthalt in Moskau, Kiew und Taschkent. Diskussi-
onspapier Nr. 131, Fachbereich Wirtschaftswissenschaft der FernUniversitat Hagen, 35 Seiten.
GRUBER, J. 1991
Improving the generality of results from Monte Carlo studies of small sample properties of esti-
mators. In: Fandel, G., and Gehring, H. (eds.), Operations Research Beitrage zur quantitativen
Wirtschaftsforschung, Springer-Verlag, Berlin, Heidelberg.
GRUBER, J., FUSEK, I., and FUSKOV A L. 1987
Evaluating the relative performance of estimators by means of combined response surfaces: the
case of EV estimators and OLS. Jahrbiicher fiir Nationalokonomie und Statistik 203,507-516.
HENDRY, D.F. 1984
Monte Carlo experimentation in econometrics. In: Griliches, Z., and Intriligator, M.D. (eds.),
Handbook of Econometrics Volume II, North Holland, Amsterdam, Oxford, 937-976.
JOHANSEN, 1. 1979
The report of the committee on policy optimization - UK. Journal of Economic Dynamics and
Control 1, 101 - 109.
KENDRICK, D.A. 1981
Stochastic control for economic models. McGraw-Hili, New York, a.e.
KENDRICK, D.A. 1988
Feedback - a new framework for macroeconomic policy. Kluwer Academic Publishers, Dordrecht,
Boston, Lancaster.
KRAMER, W. 1980
Eine Rehabilitation der gewohnlichen Kleinst-Quadrate-Methode als Schitzverfahren in der Okono-
metrie. Haag und Herchen Verlag, Frankfurt.
LEONTIEF, W. 1971
Theoretical assumptions and nonobserved facts. The American Economic Review 61,1-7.
MARIANO, R.S. 1982
Analytical small-sample distribution theory in econometrics: the simultaneous equations case. In-
ternational Economic Review 23, 503-533.
MITTAG, H.-J. 1987
Modifizierte Kleinst-Quadrat-Schatzung im Modell mit fehlerbehafteten Daten. Mathematical
Systems in Economics 109, Athenaum- Verlag, Frankfurt.
MITTAG, H.-J. 1989
Estimating parameters in a simple errors-in-variables model: a new approach based on finite sample
distribution theory. Statistical Papers (Springer-Verlag) 30,133-140.
OLBRISCH, M. 1988
Das interaktive Referenzpunktverfahren als Losungsmoglichkeit okonometrischer Entscheidungs-
modelle - Darstellung und Einordnung der Theorie und praktische Anwendungen. (Bochumer
Wirtschaftswissenschaftliche Studien Nr. 121), Studienverlag Dr. N. Brockmeyer, Bochum.
OLBRISCII, M. und GRUBER, J. 1986
Das Referenzpunktverfahren zur interaktiven Vektoroptimierung in okonometrischen Entscheidungs-
modellen. Diskussionspapier Nr. 100, Fachbereich Wirtschaftswissenschaft der FernUniversitat
Hagen.
PHILLIPS, P.C.B. 1983
Exact small sample theory in the simultaneous equations model. In: Griliches, Z., and Intriligator,
M.D. (eds.), Handbook of Econometrics Volume I, North Holland, Amsterdam, 449-516.
14
Kenneth F. Wallis
ESRC Macroeconomic Modelling Bureau
University of Warwick
Coventry CV 4 7AL
England
Abstract
In many areas of economics competing models coexist, and the decision maker is
typically given little guidance on how to choose between them. Many countries have
competing macroeconometric models of their national economies. This paper describes
some recent developments in policy making with macroeconometric models, drawing on
the research programme of the ESRC Macroeconomic Modelling Bureau. This unique
project was established to improve the accessibility of models of the UK economy and to
undertake comparative research which improves understanding of the reasons for
differences between them and, where possible, resolves those differences. The paper first
reviews the techniques employed in model-based policy analysis. Next, the issues that
arise in designing and implementing model comparison exercises based on their use are
considered. Finally, a recent development in the structure of the models, namely the
incorporation of expectations variables treated in a rational or model--{;onsistent
manner, is described. This is an important challenge to conventional policy analysis,
because the analogy between economic and engineering systems on which it rests no
longer holds. A pervasive theme is that model-based policy analysis should be fully
documented and entirely open. When differences between models cannot be resolved,
the reasons for their existence should be made completely transparent.
Contents
1 Introduction
2 Policy analysis in conventional macroeconometric models
3 Model comparisons
4 Models with forward expectations variables
References
16
1 INTRODUCTION
In many areas of economics competing models coexist, and the decision maker is
typically given little guidance on how to choose between them. Many countries have
competing macroeconometric models of their national economies. The properties of such
models are often summarized through dynamic multipliers or ready-reckoners, which
describe the response of key endogenous variables to changes in policy instruments,
treated as exogenous. Different models, constructed for different purposes and from
different theoretical perspectives, perhaps also emphasizing different features of the
economy, may give different estimates of these responses. Likewise, different models
may give different answers to more elaborate questions, such as those that arise in policy
optimization exercises or in the design of feedback policy rules. Such disagreement
places the policy analyst or adviser in a difficult position, moreover little information on
the relative reliability of the different estimates is typically available. The policy maker
ultimately has to make a choice, however, hence it is important to seek to reduce such
disagreement, and where this is not possible, to make clear the reasons for its existence,
so that an informed choice can be made.
This paper describes some recent developments in policy making with macroeconometric
models, drawing on the research programme of the ESRC Macroeconomic Modelling
Bureau. This is a unique research project which brings together the different
econometric models of a national economy, in this case the UK economy, makes these
models accessible to other researchers, and conducts its own independent comparative
and methodological research (see, for example, Wallis et at., 1984-87). Section 2 first
surveys the techniques employed in model-based policy analysis, and Section 3 then
considers the issues that arise in designing and implementing model--{:omparison
exercises based upon their use. A recent development in the structure of
macroeconometric models is the incorporation of explicit future expectations variables,
treated in a model--{:onsistent or "rational" manner, and this is the subject of Section 4.
This development presents some challenges to conventional policy analysis, in particular
acknowledging that "the analogy between engineering and economic systems which
inspired much of the early work on control theory in economics is, quite simply,
misleading. Economic systems, like all systems of social interaction, are intelligent in a
way that engineering systems are not. In economics, we are concerned with the control
of intelligent systems that think about what policy-makers do to them and act
accordingly" (Currie, 1985). Some recent responses to these challenges, in the context of
model-based policy analysis and its associated numerical procedures, are described in
the closing part of Section 4.
17
The techniques used in model-based policy analysis are reviewed in this section. We
begin with the standard linear simultaneous equation model, to fix ideas, and then
consider the more realistic nonlinear model. Extensions to models with explicit future
expectations variables are reserved until Section 4.
The model describes the relations between the elements of a vector of current
endogenous variables Yt and a vector of predetermined (exogenous and lagged
endogenous) variables Zt' and is written in structural form as
(1)
contemporaneous feedbacks among the endogenous variables are solved out in the
reduced form
For those elements of the vector Zt that are policy instruments, the associated
reduced-form coefficients give the policy multipliers or ready-reckoners, describing
theresponse of the endogenous variables (targets) to a unit shock in a policy instrument.
In a linear model such ready-reckoners may be combined to describe the effect of a
policy package. For presentational purposes, however, the impact of the policy change
is often described in terms of the difference between two solutions of the model, which,
of necessity, is the procedure followed once the linearity assumption is relaxed. First a
control solution or base run Yt , t = l, ... ,T, satisfying the equations
(2)
is obtained over some historical or simulation period for given values of the exogenous
variables. Then the vector Zt is replaced by Zt + 0t' t = l, ... ,T, where the vector 0t has
most of its elements equal to zero, with non-zero values corresponding to perturbations
in selected policy instruments. The perturbed solution or simulation Yt' t = l, ... ,T,
satisfying the equation
(3)
is obtained, and the differences Y - Yt estimate the effect of the policy change. The
differences obey the relation
(4)
and this is the equation that must be solved for 0t if it is desired to find the policy
change that achieves a given deviation in target variables. Clearly, as stated by
Tinbergen's (1952) counting rule, at least as many instruments as targets are required;
other conditions are discussed below.
solution is then obtained from the given model augmented by residual adjustments,
namely
(5)
and the problem is to assign appropriate values to the relevant elements of at. The
standard case of the preceding paragraph corresponds to this equation if at = - Ct5t ,
but in many cases the model does not immediately provide the information required for
this calculation. A simple example is that of the introduction of a new tax. If the tax
serves to augment a price variable that appears in the model in an appropriate way,
then the relevant adjustment is given as the tax multiplied by the price coefficient, in
every equation in which the price variable appears. However, if the price variable does
not appear in the model, perhaps because it has exhibited insufficient variation in the
historical data to allow its coefficient to be estimated, then evidence from statistical or
accounting sources, or from other econometric or survey research, must be drawn on to
calculate the relevant adjustment; such evidence should always be fully documented.
Assumptions about a policy often relate not to its effect on a particular behavioural
equation but to its impact on the solution (the reduced-form values) for particular
endogenous variables. Indeed in many cases the central questions are how to achieve
target values of endogenous variables, and what are the consequences of doing so. There
are various methods of obtaining a model solution in which certain endogenous variables
are constrained to have given values. First, this may be achieved by the choice of
appropriate values for the variables' own residual adjustments, and procedures for
producing these adjustments automatically are known as type 1 fixes. A second
procedure, known as exogenization, is to include the endogenous variables with
preassigned values among the predetermined variables, delete the equations which are
normalized on those variables, and solve the slimmed model for the remaining
endogenous variables.
To consider these possibilities we partition the vector y t into subvectors y 1t and Y2t'
and assume that the variables whose values are constrained are collected in the first
subvector. Partitioning the structural form conformably, equation (1) may be rewritten
as
(6)
20
Denoting the preassigned (target) values of the endogenous variables in the first block
by Yit, the associated solution for the unconstrained variables is then obtained (with u 2t
= 0) as
(7)
In this calculation the variables Ylt are treated in the same way as Zt' hence the term
exogenizationj one might alternatively say that the first block of equations has been
switched out. The type 1 fix that ensures that Yit and Y2t also represent a solution to
the first block of equations is
with this residual adjustment, the subvector a2t being zero, is clearly identical. Thus
using alt to control y 1t is numerically equivalent to switching out the first block of
equations, that is, treating y 1t as predetermined in the second block.
These methods assume that yi t can be achieved, and concentrate on the consequences
for the remaining variables as given in (7). The exogenization approach is completely
silent on the question of how Yit is achievedj with a type 1 fix some explanation may be
offered if a specific residual adjustment is given a policy intervention interpretation, as
an incomes policy intervention in a wage equation, for example. A further method of
achieving yi t is to adjust some other variable(s) to which y 1 is related, and the more
explicit is this assignment and the more transparent the relationship, the more likely it
is that a sensible discussion about the policy and its implications can be pursued. A
price variable may be adjusted in order to achieve a desired variation in a quantity
variable, for example. The calculation and imposition of the required adjustment is
known as a type 2 fix. It can be accomplished by adjusting the residuals on other
endogenous variables, or by adjusting appropriate exogenous variables (policy
instruments) by an amount Dt , using the vector notation introduced in equation (3). In
the context of the partitioned model (6), the adjustment required to achieve the target
value of the first subvector satisfies the equation
21
(8)
Given Yit, the existence of a solution for 8t rests on the existence of appropriate
channels of influence, which may be direct, as represented by the matrix C I , or indirect,
as represented by C2 and B 12 . In the context of a single equation, the required
calculation amounts to an inversion of the equation, obtaining the value of the
explanatory variable that is needed to produce the desired value of the dependent
variable. For example, the required adjustment to interest rates that achieves a
monetary target might be calculated in a single-equation context by inverting a
demand-for-money equation. Note that the classification of variables as endogenous or
exogenous is not changed by this calculation. In this example it is precisely because
interest rates are treated as exogenous that they can be manipulated to produce the
desired response of endogenous money.
In a type 2 fix, the associated solution for the variables that are not targeted is
This differs from the case of exogenization or a type I fix, given in equation (7), by
including the explicit policy adjustment. If the policy instruments used in targeting y 1
have a direct effect through the second block of equations, that is, the relevant elements
of C2 are non-zero, then the solution for Y2 given by the exogenization procedure is
biased. For example, the exchange rate might be targeted via adjustments to
short-term interest rates calculated and imposed in a type 2 fix. If, on the other hand,
the exchange rate is exogenized then the influence of the required interest rate
movements on the rest of the economy is neglected, and hence biased estimates of the
response of output, inflation, employment, and so forth are obtained.
The foregoing discussion has neglected the nonlinearity of the practical models, and has
paid no attention to their dynamic nature. In practice, however, dynamic aspects of the
relations among variables are of concern, and the time form of the reaction of a target
variable to a policy instrument is often as important as the overall magnitude of the
reaction. To consider dynamic models explicitly we distinguish between exogenous and
lagged endogenous variables, partitioning Zt into a vector of purely exogenous variables
22
xt and the lagged endogenous variables yt-l' Yt-2'.... The nonlinear system with
first-order dynamics may then be written formally, in its structural form, as
where f is a vector of functions having as many elements as the vector yt' and a is a
vector of parameters. This notation is convenient, although it is more general than is
necessary, since the equations of empirical macroeconomic models are usually linear in
simple instantaneous transformations (such as logarithms) or simple functions (such as
products, ratios or proportionate changes) of variables. Nonlinearity in parameters is
relatively rare.
It is assumed that a unique solution for the current endogenous variables exists,
although in general it does not have an explicit analytic form. It can be approximated
numerically to any desired degree of accuracy, and there is a large literature on
numerical solution methods, to which we refer below. Forecasting and policy analysis
exercises usually ignore the impact of the random error term and utilize the dynamic
deterministic solution of the model. This is obtained, for given x-values and given
initial conditions, as the period-by-period numerical solution, to the desired degree of
accuracy, of the equations
Policy multipliers are calculated as the difference between base and perturbed solutions,
and residual adjustments, type 1 and type 2 fixes, and exogenization exercises likewise
proceed through numerical solution methods. The underlying concepts are exactly as
discussed above, although in the nonlinear case we are no longer able to express them in
convenient linear algebra. In particular multipliers no longer correspond to constant
coefficients, but are potentially base dependent, that is, they depend on the values of the
predetermined variables around which they are calculated.
Adjustments and interventions also need to take account of the model's dynamics. For
example, a residual adjustment held at a constant level typically has a cumulative
effect, and fixes designed to achieve a constant deviation in an endogenous variable
typically vary over time. The instrument sequence bt , t=I, ... ,T, determined in a type 2
23
Control techniques
In a simple one target-Dne instrument case a multi-period quadratic loss function may
be written
24
where Yidenotes the target value of the endogenous variable and x t the base-run value
of the instrument. The k2-{;oefficients indicate the relative cost of changing the
instrument from one period to the next, and the k3-{;oefficients the cost of departing
from the base setting. Type 2 fixes are a special case in which such costs are absent,
thus k2t = k3t = 0 in all periods, and a minimum loss of zero is achieved by attaining
the target in all periods. As noted above, the most common difficulty with this
technique is that the resulting instrument paths may be unreasonably erratic or may be
unstable, in turn destabilizing other variables. This instability may reflect hitherto
unsuspected dynamic properties of the model, or it may simply reflect an inadequate
formulation of the policy problem, in that inappropriate target values have been chosen
or instrument instability has not been penalized.
The classical control-theory literature is more concerned with stability and robustness
than with narrow notions of optimality. The emphasis is on the design of linear
feedback rules for policy instruments that yield acceptable behaviour of the controlled
system, which leads to the proportional, integral and derivative (PID) controller
introduced into economics by Phillips (1954).
Some recent research - see, for example, various papers in Britton (1989) and references
therein - attempts to link the classical and more recent strands of the control-theory
literature by seeking explicit feedback rules that approximate the fully optimized
solution in large nonlinear models. This research is motivated on the one hand by the
technical challenge, and one procedure is to linearize the nonlinear model, then apply
linear techniques to the resulting approximation, and finally apply the optimal feedback
rule so obtained to the original nonlinear model. On the other hand a second motivation
is a desire to obtain simple policy rules whose ease of understanding and verification
helps overcome the credibility problem, discussed below. Since the optimal rule involves
feedback on the whole state vector, which moreover is difficult to interpret in the linear
approximation case, these requirements are in conflict, and simplicity is often achieved
by arbitrary selection of a small number of indicator variables; the PID approach
suggests that these should include target variables. Similarly the choice of target values
is often treated in an arbitrary fashion, although in principle these should be obtained as
the fully optimized solution of the nonlinear model, which at least ensures that the
targets are attainable.
25
Given that in nonlinear models multipliers are base dependent in principle, to ensure
comparability experiments should be conducted relative to the same base, as far as
possible. This may be a historical or a forecast base.
dynamic models the only problem that remains from the paragraph above concerns
initial conditions, for if the starting period represents a current or recent past episode
that is difficult to model for one reason or another, then base--dependence problems are
not completely avoided. In principle a forecast data base can be continued into the
indefinite future, or at least sufficiently far to allow an assessment of the model's
stability and of its convergence (or otherwise) in the absence of shocks. In practice
statistical models are local approximations, valid within the range of sample experience,
and some difficulties may be expected when the model is driven far outside that
experience.
In a relatively open model with a relatively large number of exogenous variables, the
construction of exogenous forecast assumptions in an internally consistent manner
requires a certain amount of care. Implicit relations among exogenous variables and
restrictions on their range are not usually included in a model's specification because
they are satisfied automatically, in principle, in historical data. It is necessary to ensure
that they are also satisfied in projected data. Examples are stock-flow identities (with
stocks remaining positive), adding-up requirements, the government budget constraint,
and the foreign reserve position.
To aid model comparisons and to avoid base dependence, the solution over the forecast
period for key endogenous variables might be required to be common to all models,
analogous to a common perfect-tracking solution in the case of a historical base. This
was a stated objective in the Brookings exercise on multicountry model comparisons
(Bryant et at., 1988), where the project organizers specified a common base to be
implemented by each model proprietor. However this caused as many problems as it
solved, as they report (p.30):
Unfortunately, many of the model groups could not or did not follow the
instructions for the baseline path precisely. Establishing a common baseline was
found to be considerably more difficult than running the simulation experiments.
For some models, considerable model manipulation and trial and error were
required to generate the baseline even approximately. Thus, although the
baseline paths for most variables for most model groups conform moderately
closely to the specified common paths, complete conformity was not achieved.
Some model ~roups were concerned that the "add factors" necessary to generate
the baseline l adjustments to the constant terms or residuals of equations or to
the values of exogenous variables) were very large in relation to the specified
baseline values of the variables in question. Although such add factors
incorporated in linear equations do not affect the absolute differences between
baseline and simulation values, they can affect the percentage differences and
thereby give misleading indications of model responses.
27
In our own analysis of the UK models (Wallis et al., 1984-87; Fisher et al., 1988, 1989)
simulations are performed relative to the modellers' own published forecasts and
supplied forecast assumptions. Although the forecasts are not identical, there tends to
be a broad consensus, which also applies to the different modellers' views of external
developments, and there is no evidence that the observed forecast differences distort
multiplier comparisons. (However if a particular endogenous variable is temporarily
exogenized in a forecast, this must be changed to the equivalent residual adjustment
before simulations begin.)
Additional assumptions about the general policy environment are invariably required,
however, and must be implemented in a standard manner across different models. For
any perturbation, some view must be taken of the accompanying stance of monetary
and fiscal policy, and again this must be made explicit. In general, policy changes lead,
directly or indirectly, to changes in the public sector borrowing requirement (PSBR),
and different results may be obtained according to the way in which this is financed.
The standard possibilities are that the change in the PSBR is accommodated by
additional money creation (money finance), financed by the sale of government debt to
the non-bank public sector (bond finance), or offset by a change in other elements of the
PSBR, such as additional taxes (tax finance). Each may be adopted on its own, or some
combination of these methods may be employed, for example, balanced money and bond
finance. Policy in the foreign exchange market, and its relationship with monetary
policy, must also be specified. More generally a statement of an overall stance of policy,
such as that suggested by a nominal income target, might be relevant, which would
require the specification of policy instruments and the way in which they adjust to
28
ensure the achievement of such a target. In many cases, therefore, some modification of
policy instruments other than those of immediate concern is necessary in order to retain
the internal consistency of the overall policy stance. Then the sensitivity of the
simulations to different policy environments may be of interest.
After such broad differences in model type have been taken account of, what of the
remaining differences? Their resolution is a matter for econometric analysis. There is
after all only one data generation process, and different models are simply different
approximations to it, whose relative quality can be assessed by econometric techniques.
Data can often discriminate between rival models; if this is not possible the choice rests
on other criteria, which econometric analysis helps to focus more sharply. Exercises in
which cross-model simulation differences are explained and resolved by detailed
empirical investigation of model structures are presented by Turner et al. (1989) and
Fisher et al. (1990).
in the hands of the model proprietors, hence complete standardization was not achieved.
Giving the whole exercise to an independent third party avoids most of these problems.
Direct comparisons across models can then be made at all stages of the project - design,
execution, and analysis - and if particular side assumptions or adjustments are required
in particular circumstances, these can be made completely explicit, rather than have
model proprietors taking them for granted as part of their normal modus operandi and,
perhaps, forgetting to write them down.
Introduction
The rational expectations literature is usually assumed to start with Muth (1961),
although following much older discussions of the influence of forecasts on outcomes
Grunberg and Modigliani (1954) had already shown that where agents react to forecasts
and thereby alter the course of events, this reaction can be taken into account to yield a
correct, self-fulfilling forecast. This is the same kind of internal consistency achieved by
modern algorithms for calculating model-consistent expectations. Whereas Muth
introduced the term rational expectations into the English language, Keuzenkamp
30
(1989) has recently drawn attention to a remarkable early paper by Tinbergen (1932),
published in German, which anticipates much of Muth's analysis. Muth's paper in turn
had little impact on empirical econometrics for a while, perhaps because its leading
example was one in which rational expectations and adaptive expectations coincide, and
so might have been thought to provide further justification for the then rather new
adaptive expectations hypothesis. In the United States the eventual adoption of the
rational expectations hypothesis was associated in addition with new classical
equilibrium business cycle models and the policy ineffectiveness proposition, which again
hindered its wider dissemination, although Barro and Fischer (1976) drew a clear
distinction between the rational expectations hypothesis as a theory of expectations and
the type of equilibrium model in which the hypothesis was typically embodied at the
time. This distinction was seen more clearly in other countries, including the United
Kingdom, where explicit forward-looking expectations handled in a model-<:onsistent
manner have been incorporated in mainstream "sticky price" and quantity adjustment
models.
give a closed form, analytic expression for the solution and consider its uniqueness and
stability; a solution satisfying the uniqueness and stability conditions is referred to as a
saddlepoint path. For the nonlinear model numerical solution methods are necessary.
Again it is not possible to use period-by-period solution techniques appropriate for
backward-looking dynamic models; at least conceptually it is necessary to iterate over
the complete solution path, as in the method of Fair and Taylor (1983), although further
research has led to improved methods (Fisher et al., 1986; Fisher and Hughes Hallett,
1987). Often there exists a multiplicity of consistent solution paths, of which only one
is stable, and it is necessary to impose additional conditions to ensure that this stable
path is selected. Typically this selection is made via terminal conditions that specify
the forecast values and expectations at the end of the forecast horizon or beyond.
If the nature of the stable path is known a priori, from analysis of the equilibrium
properties of the model, for example, then the terminal conditions may explicitly
incorporate this knowledge. This may require a relatively long solution period, to ensure
that the model has reached an approximate equilibrium. In the absence of such
information, or in a shorter solution period, terminal conditions should be specified that
approximate the stable path after the end of the solution period, typically by requiring
constant growth rates of relevant variables. In order to ascertain whether the time
period is sufficiently long or whether appropriate terminal conditions have been used, so
that the solution over the period of interest is unaffected, numerical sensitivity analysis
may be undertaken. The sensitivity of the solution to variations in the terminal date
indicates how much of the solution can be regarded as a genuine product of the model
rather than a product of the assumed terminal conditions. Changes in the type of
terminal condition, together with evidence from the different solution periods, can also
suggest whether or not a unique stable solution exists. Examples of such sensitivity
analysis are presented by Wallis et al. (1986, Ch.2.5).
Experimental design
Policy optimization
The single-economy policy game typically assumes noncooperative behaviour, with the
two players interacting according to Nash or Stackelberg assumptions. In the former
each player takes the other's actions as given, and an equilibrium position is one that
provides no incentive for either player to move, again taking the other player'S (optimal)
policy as given. In the latter one player acts as a leader, anticipating the reactions of
the follower to announced policies and optimising accordingly. For practical
implementation in large-scale empirical models the Stackelberg assumption presents
fewer difficulties, since the econometric model describes the behaviour of economic
agents who form expectations on the basis of the announced policy, which they treat as
given, and the policy maker proceeds to optimize this. Under Nash assumptions,
however, the policy maker must optimize taking agents' expectations as given, and an
additional level of iterations is necessary to ensure that the expectations are consistent
with the optimized policy. Nevertheless it might be argued in an open government
context that this is more in keeping with the rational expectations hypothesis - that
economic agents understand and hence correctly anticipate the policy maker's actions -
and the resulting policy is time consistent.
REFERENCES
BurTER, W.H. and MARSTON, R.C. (eds.) (1985). International Economic Policy
Coordination. Cambridge: Cambridge University Press.
CHOW, G.C. (1975). Analysis and Control of Dynamic Economic Systems. New York:
Wiley.
CURRIE, D.A. (1985). Macroeconomic policy design and control theory - a failed
partnership? Economic Journal, 95, 285-306.
FAIR, R. C. and T AYLO R, J. B. (1983). Sol u tion and maximum likelihood estimation
of dynamic nonlinear rational expectations models. Econometrica, 51, 1169-1186.
FISHER, P .G. (1990). Simulation and Control Techniques for Nonlinear Rational
Expectations Models. Ph.D. Thesis, University of Warwick.
FISHER, P.G., HOLLY, S. and HUGHES HALLETT, A.J. (1986). Efficient solution
techniques for dynamic non-linear rational expectations models. Journal of
Economic Dynamics and Control, 10, 139-145.
FISHER, P.G. and HUGHES HALLETT, A.J. (1987). The convergence characteristics
of iterative techniques for solving econometric models. Oxford Bulletin of
Economics and Statistics, 49, 231-244.
FISHER, P.G., TANNA, S.K., TURNER, D.S., WALLIS, K.F. and WHITLEY, J.D.
(1988). Comparative properties of models of the UK economy. National Institute
Economic Review, No.125, 69-87.
HALL, S.G. (1986). Time inconsistency and optimal policy formulation in the presence
of rational expectations. Journal of Economic Dynamics and Control, 10, 323-326.
(Revised and abridged version of National Institute of Economic and Social
Research Discussion Paper No.71, 1984.)
KEUZENKAMP, H.A. (1989). The prehistory of rational expectations. Discussion
Paper No. 8931, Center for Economic Research, Tilburg University.
KLEIN, L.R. (1987). The ET interview: Professor L.R. Klein interviewed by Roberto S.
Mariano. Econometric Theory, 3, 409-460.
KYDLAND, F.E. and PRESCOTT, E.C. (1977). Rules rather than discretion: the
inconsistency of optimal plans. Journal of Political Economy, 85, 473-49l.
LUCAS, R.E. (1976). Econometric policy evaluation: a critique. In The Phillips Curve
and Labor Markets, (K. Brunner and A.H. Meltzer, eds.), pp.19-46.
(Carnegie-Rochester Conference Series on Public Policy No.1, supplement to the
Journal of Monetary Economics, January 1976.)
MUTH, J.F. (1961). Rational expectations and the theory of price movements.
Econometrica, 29, 315-335.
36
WALLIS, K.F. (ed.), ANDREWS, M.J., BELL, D.N.F., FISHER, P.G. and WHITLEY,
J.D. (1985). Models of the UK Economy: A Second Review by the ESRC
Macroeconomic ModeUing Bureau. Oxford: Oxford University Press.
WALLIS, K.F. (ed.), ANDREWS, M.J., FISHER, P.G., LONGBOTTOM, J.A. and
WHITLEY, J.D. (1986). Models of the UK Economy: A Third Review by the ESRC
Macroeconomic ModeUing Bureau. Oxford: Oxford University Press.
WALLIS, K.F. (ed.), FISHER, P.G., LONGBOTTOM, J.A., TURNER, D.S. and
WHITLEY, J.D. (1987). Models of the UK Economy: A Fourth Review by the
ESRC Macroeconomic Modelling Bureau. Oxford: Oxford University Press.
PART 2:
Gabor Bagdy
Institute of Economics
Hungarian Academy of Sciences
H-1502 Budapest, Pob 262
Hungary
Abstract
Contents
1 Introduction
2 The decision model
3 Specification of the objective function
4 The numerical results
1 Introduction
A possible application of econometric models is to make control
experiments based on them. An econometric model is usually a
system of simultaneous discrete-time difference equations, that
involve endogenous, exogenous and policy variables. If a crite-
rion or welfare function is added to the model, the question is,
what values the instruments should have in order to maximize
welfare. In this paper we try to answer the question: To what
extent do different scalar-valued objective functions and the
length of the time horizon influence the results of the optimal
control exercise?
Two approaches are used to solve the optimal control problems.
In the first case, the model is assumed to be linear and the ob-
jective function quadratic. This approach is used in most cases.
The linear-quadratic problem can be formulated in the follmling
way. Assume that the following reduced-form linear econometric
model is given:
40
T
w(X,u) Z(Xt-x*t)TQ(xt-x*t) + (Ut-u*t)TR(Ut-U*t),
t=l
where, x t * and u t * are the given target vectors and Q and Rare
given symmetric matrices.
subject to
subject to
and
T
w2 :E DSS (t) IT ~ min.
t=l
3. To achieve smoother growth in the ec?nomy. The optimal
path should avoid fluctuations as much as poss~ble:
w3= + +
At first sight, the optimal control problem with the third ob-
jective function appears to be linear-quadratic. However, we
chose to use the GRG code instead, and added constraints on sev-
44
\ I '?~
11B I 578
I /'
" I
I 568 -' W3
.xcl....nl ..... t •• \ I I Wl
I , ---- ... ""' "'
pi ", . . . xport con.",.pt Ion sse - ,,2
.uhs: Ld Ie .. [ ]
- 002
,aid lor I.I.fllt 188 \\ " , ,/"'\
. , ~,'
, :
,i/ ~"'':'-:::,........ G.. , "" 1
doll.,. export. - 003 548 ' .
(ac.t ual ., !ee)
.' / 538
,~ , ' ./
/ :./'\ \.... '
\ \ .. ~-"......
98 , /" ...~-................ .................. 528
"' SIB
see I I I I I I I I
88 I ~ I 79 sa 81 12 83 84 85 86
79 98 81 82 83 84 8S 8&
."."f'.I
ye .. r ..
~
(J1
118 2eB
•• • • • •• •• ••• •• ~. - •• • - " • • • •• • ••• •• • • • • • ••• •• •••• • • • - - _ _ . ft . . . . . . .
I ... ,
I .." 199
18B I "\
I \
excJotan, ..... t •• I \
.. I nUl. i.port "3 168 , W3
I ~.
"t .. xo. / Y', lorelrn debt ,,
98 / ~ [- ;
w2] [- ;
.. 2]
paU lor unit I ~ (1979 , 18B)
.011 .. ,. IMport. , I ,\ ~~ w1
( .. ctr.l .. l ~ 1Be)
... wi 14e
v ~
'. .... ~~ ...,Jt. /·~<··""·
,, '. . .....~~ ,. . . . .~-.~ .... ....... ....••
~
,,' 128 .,;,;:-::'--- "... ...
- _ ..........•......;.;: ....
,,<> - - - ...... -"
78
188
~
218 228
218
" /.#... ./f-
'~ ~ .
288 1
288 '\ ,
\ "l
198 "l
Iln..... stl'llsn1.:I 198 "., - ,,2 do 1 lI:r exports .. t ~ .
- ,,2
conatlknt pric ••
....· 1.11
" 188 ... wi
188
\ ..',.~ ~
',,~,~,.~, /-"",
B B
" '" ,. 118
,.•..,:.:---- ,---,,/'
178 ....., ....•" .,.., ......., ...
--' 168
168 I I I ·,··,·, .. ' ..····r...... ....... ,. I I
1581 I I I I I I
13 sa 81 92 8J 8i es 86
13 88 81 82 B3 84 BS 86
v· .. r. 'Y."Z".:II
.I>-
m
Ellects 01 dirt.rent objective fL,tnctions Ettec::t. or 411/erent ohJect l\ol'8 I'u. nct Ions
(1~1~1~961 ( 191~198 6 1
2488 2ee
1ge
2388
18e
I
,r-, -,
. ,,;-, /
.•. .....••. 118 ,,
2288 /
"3 I "3
dol1&f' i .. ports .t 168 I
gross ou.lp"" t - .. 2 con.t .. nt pr Ice. [ ]
- .. 2
~!--y"it"~' ,.",;''''/
••• 1011 ... t..I1
2188
B 148
\ ~"~J-~:// .......,.,.,.,...,........ ..
..., ..., .....~ ..~ .........
118
Elt.ct. at ,au ' ..... nt t.l •• hal" lzon. Ett.ch 01 ,. IF' et'o "''t. t. SIilO "'ot'13:0"'.
( w2 ) ( w2 )
59B 13B
/~:~'::::::.:::.::
seB
12B
- 01-86
- 79-84 ex c"ange rates 118t\
plu.s export '.,
.l.Ibsid,l •• - 79-86
- 79-85
consWlptlan paid. fo ... "nit
.., •.. 79-85
_. 79-8£ dol hI" ex.rort.s
s.S (actu.a 1.. lUI )
1eet ·". \, . ,/ .,\~ /" .- 79-8.
.- 81-8£
53S ,......\ , ~~~ '~ "".;
.'\ ... ~'1t.:: : :.:...... \....J'"
,,'" \ . . -....-~-
S2S ge J. "tI"
S19
1&8
--- ....
.... ' ...
118 ..... .... '\
lS8
/>' ";
148 I ,. J
;/.,/
- 79-04
138 // /</ - 79-84 .xc:h .. nf. rat.*
.lnus Illiport
"'~>
98
1 "" .\,\\
- 79-85
- 79-85 t.x ... '\ \
lor.lin d.lI.t
1 / . . ... . ..=:-,.....
~
..... ",,,,' ....... , .... , .•.• .' ,aU (or unit
(1970:.188) 1Ze .•. ,... ,r , , / ' " ,,\~..... " •••"" / ... 79-86 do 1 hr t_,ports
"' \ ··· 79-86
CactuAl ll:188 ) \ \ .- 81-86
..- 81-86 "
';~:~~<:~~? \ \\;
76
:::r'
98
. . ..................,.' \\
;
..
8S I I I I I I 68 I I I I I I I i
79 8B 81 82 83 8i 85 86 79 08 81 02 83 84 OS 86
\018"1'. v·· rs
Chart II •
228 ./~o
228
218 //
//)/' 218 '.,
"~,'\. >, .//
288 - 79-84 .....\ .,l' - 79-84
,~" 288 .
- 19- 85 - 79-85
dol Jar- e .ports at 198 InveJl:t. e"tJl: ",,/
const.nt prleaS" ... 79-86
I _. 79-86
.,.,
\ \. ' '","
198
188 .- 81-86 -- 81-86
188
118
\<~~~~--~:<:~:>,/
178 ......... -.................
16et ..~.:.:>·
128' I I I I I I I
79 88 81 82 83 84 85 86 1988~
1 ----~----+-----~--~-----+----~----~
79 88 81 82 B3 84 85 86
v.·,.. \iI.ars
50
References
Michael Leserer"
Institut fur Agrarokonomie
Universiiit Gottingen
Platz der Gottinger Sieben 5
D-3400 Gottingen
Abstract
Starting from a "behavioristic point of view" causality is discussed in the framework of
"networks of decisions" (NOD's) in which a conditional probability formalism is used in
order to represent causal relations. A worked example confronting the analysis of NOD's
with the usual "mapping analysis" of causality leads to some conclusions about (mecha-
nistic) optimal control models. It can be shown that a constant structure of decision
making in a dynamic NOD implies a "variable parameter model" and a time dependent
incontrollability problem in the traditional control model formulation.
Keywords: Causality and its representation; networks of decisions; control in variable
parameter models.
Contents
1 Orientation
2 Example
3 Discussion
References
Appendix
1 Orientation
Causality is a fundamental concept in human thinking. The ability of man to ask for cau-
ses behind phenomena implies model building in order to visualize causal relations. But
in our mechanical world the "picture" of causality is solely determined by the "heritage of
Newton" [1] in the following sense: If one formalizes causality then one automatically
constructs a mapping from causes to effects. This procedure is totally motivated by expe-
rimentation in the natural sciences: Mathematical mappings are the natural analogies of
"experimental mapping" which is done by setting causes and observing their effects. But
this is not the usual situation in economics! Here the behavior of economic agents origina-
tes causal phenomena far from experimentation facilities. Thus a necessary condition for
I am indebted to Lidwina \\"ohlfahrt for her engagement in managing the numerical part
of this paper.
52
causality in economics is the psychologically or legally based possibility for agents to influ-
ence other agents, and a sufficient condition for causality is that economic agents behave
or must behave in order to make this influence possible. Thus in many cases, the origin
of causality in economics is not a physical "transfer of energy" ("Energieubertrag") [ 4]
but an accepted voluntary interaction for realizing economic intentions. So in my opinion
this "non-experimental causality" implies a causal analysis which is not originally based
on the ideas of mappings. I have investigated this matter including the implications for
control analysis "in foregoing papers [2], [3]. In this paper I confront a dynamic version of
my analysis of "networks of decisions" (N 0 D) with the traditional analysis of mapping.
In this context, the philosophy of NOD's is the following: Like in chemistry, it is assumed
that economic variables are explained in terms of atomistic decision structures consisting
of causally connected sets of alternatives which can be chosen in a probabilistic way, i.e. in
each period economic agents may have an economically well-defined finite set of decision
alternatives. These alternatives are chosen with a certain probability. Such a chance of
realisation may be derived from utility aspects. But typically, one agent's alternative - if
it is chosen - may induce reactions of other agents. This phenomenon defines causality.
Thus, a NOD may be seen as an economically well-defined reaction scheme.
Based on such decision structures, economic accounting systems then define economIC
variables which are random. In the space of these random variables it is possible to
reconstruct the causal structure in a NOD phenomenologically by regressing effect random
variables on cause random variables. In this way the causal structure of a NOD determines
the causal representation of economic variables. The following example will show that
constancy in a dynamic NOD over time (equal number of agents, equal number of decision
alternatives per agent, constant probabilities, constant mode of causality described by
directed arrows) induces temporarily non-constant parameters in affine transformations if
they are used for a causal representation in the space of random variables.
2 Example
Let ..1ij(t) (i = 1, . . , L , j = 1, ... , Af" • t = 1,2. L. Ali' eN) be the decision alternatives
of L agents (or groups of agents) over time where there are Mit alternatives for the i-th
agent in period t. Let further Bij(t) denote the set composed of decision alternatives
which influences ..1ij(t) in the sense that ..1<j(t) occurs if and only if Bij(t) occurs. Finally
let PI3,,(t) (..1ij(t») be the conditional probability of .4,)(1) under the restriction of Bi)(I)
Then the pair ( ..1ij(t), PI3.,(t) may be called a network of decisions.
In order to illustrate this definition let us introduce the following example which also may
be the basis of further statements.
53
t =1
t =2
Figure 1: Dynamic NOD with two agents and with time invariant causal structure
For all periods two agents can choose between two decision alternatives each, for simplicity
both equally probable, but agent 2 in period t = 1 only chooses if agent 1 chooses
alternative A 11 (l) , in period t = 2 he decides only if agent 1 has chosen A 11 (2) and he
himself has chosen A 21 (1) in period t = 1 and so on. Thus, for example, the probability of
choosing A21(2) is conditional on 8 21 (2) = .'1 11 (2) n A 21 (1), whereas .4 21 (1) is conditional
on 8 21 (1) = A11{l). In this way alternatives constitute causality.
For further analysis note that these alternatives induce a probability space over which
random variables are defined. So one may define the following:
C "'. = (Al1(t) , . )
~t (.;) t = 1,2, ...
otherwise
C
I)t("')
w =( .421(t) )
1=1.2, ..
othenvise
for t = 1,2,. which are derived from the (time dependent) common probability mass
functions of these random variables (see Appendix) by :
[:: ]
where bt = [bit b2tl', (= ['1t-l ~.]' L( (
(assumed to be regular) ,
= cov( ()
O"(ry = cov('1t) , p.ry = E'lt and p.( = E( With these settings one optimizes the
representation of 'It via at + b1l 'lt-l + b2t~t according to min E ('It - at - b;()' ('It - at - b;() .
at ,b j
The results of this procedure are listed in Table 1:
Table 1: Parameters of the representation '1* = at + blt 'lt-l + b2t~t for t = 1,. .. ,10
t at b1l b2t
1 0 0 0.5
2 -0.062 o .y
.~o 0.125
3 -0.01526 0.25 0.031252
4 -0.003906 0.25 0.007812
3 Discussion
First note that this paper does not intend to establish generallmvs 01' even a final theory.
Rather its only purpose is to demonstrate an effect which may serve as a starting point
for further \vork and which is best described as follows:
Considering the causal structure described in Figure lone intuitively receives confirmation
that there is a constancy in the influencing characteristics. Precisely this structure offers
a time invariant mode of causal dependence. It is the main implication of the analysis
carried out in section 2 that this constancy in a NOD is obviously not transferred to the
"Newtonian" causal analysis using affine transformations. NOD's generally would lead to
"variable parameter models"! But, if \'alues of the random variables defined for a NOD are
observed, a "data"-based analysis of such models by methods of inferential statistics may
or may not regard this variability of parameters. (Time dependence for all parameters
in the proposed example is obtained if the lagged random variable having a cumulative
memory over the entire history of the system is defind as '1t-l:= II + ... + It-I, where It IS
the indicator function for A2dt).) :\Ioreover one can observe a sequence of ~t-coefficients
converging monotonically towards zero as t grows. This results in serious consequences
for the controllability of systems over many period,,: There lIlay exist a time dependent
kind of incontrollability if "structure wins over mechanism"!
55
References
[1] Ching-Yao Hsieh and Stephen L.Mangum, A Search for Synthesis in Economic Theory.
::Jew York-London 1986, pp.71-75.
[2] Leserer,:VI. Kogniti\'e Inferenz als okonometrische Aufgabe - Einige Bemerkungen
zur okonometrischen Grundsatzdiskussion. "J ahrbucher fur Nationalokonomie
und Statistik", Bd.201/2(1986), S.97-106.
[3] Leserer.:VI. Zur Kontrolle aus Regressionen erster Art. In: H.Isermann et al.
(Eds.). Operations Research Proceedings 1986. DGOR-Vortrage der 15.Jah-
restagung. Berlin-New York 1987. S.557-567.
[4] Vollmer, G. \Vas konnen wir wissen? Bd.2: Die Erkenntnis der Natur. Stuttgart
1986.
Appendix
period t=l
period t=2
period t=3
period t=4
period t=10
Abstract
We describe a new algorithm for the optimal control of nonlinear stochastic models subject
to quadratic objective functions and report on its application to a small macroeconometric
model of the Austrian economy. In addition to the usual additive uncertainty the algorithm
allows for random parameters in the model. The optimal values of the control variables are
computed in an iterative fashion: First, the time-invariant nonlinear system is linearized
around a reference path and approximated by a time-varying linear system. Second, this
new problem is solved by applying Bellman's principle of optimality. The resulting feed-
back equations are used to project expected optimal state and control variables. These
projections then serve as a new reference path, and the two steps are repeated until con~
vergence is reached. The algorithm has been implemented in the statistical programming
system GAUSS. This implementation, called OPTCON, has been used to perform con-
trol experiments with an IS-LM-type model of the Austrian economy. The optimal policy
derived without considering parameter uncertainty is compared with two scenarios where
stochastic parameters are considered in order to assess the influence of model uncertainty
on the optimal trajectories of control and state variables.
Keywords: Optimal control, stochastic control, nonlinear models, control algorithm, ap-
plication to a macroeconometric model for Austria.
Contents
1 Introduction
2 OPTCON - An algorithm for the optimal control of nonlinear stochastic models
3 Derivation of the algorithm
4 An application to a macroeconometric model for Austria
5 Concluding remarks
References
Financial support from the "Fonds zur Forderung der wissenschaftlichen Forschung" (re-
search project no. P .5846) is gratefully acknowledged.
58
1 Introduction
Several algorithms for the optimal control of stochastic econormc systems have
already been published. Typically, they either allow for additive uncertainty
only [Chow & Butters 1977] or rule out nonlinear system equations [Coomes 1985]
[Kendrick & Coomes 1984J [Norman 1979J [MacRae 1975] or have not been implemented
[Kendrick 1981, Chapt. 9J.
In this paper we describe a new algorithm for the optimal control of nonlinear economic
models that allows for additive uncertainty as well as for the presence of a stochastic
parameter vector in the system equations. In Section two, we describe some elements
of the class of problems to which the algorithm can be applied, and some devices used
within the algorithm. Section three gives the derivation of this algorithm, which has been
developed following the lines of [Chow 1981, Chapt. 3J. Section four presents a small
nonlinear macroeconometric model of the Austrian economy consisting of eight equations.
The design of the control experiments and the results obtained are described. Section
five lists some experiences got in the course of the implementation of the algorithm and
indicates directions of further research.
We start the description of the algorithm by specifying the form of the objective and
system function. We introduce the notation used and discuss first and second derivatives
of the system equations. Furthermore, we evaluate certain expectations which are needed
in the main part of the derivation, the application of Bellman's principle of optimality
to our problem. This principle enables us to show by induction the validity of the main
recurrence equations of the algorithm.
In its most general form, the objective function of a control problem is described by
(1)
and quadratic in the deviations of state variables Xt and control variables Ut from their
"ideal" levels Xt and Ut, respectively:
(3)
with W t defined as
t = S, ... ,T. (4)
The quadratic tracking form of the objective function defined in (3) has an obvious inter-
pretation: Deviations from "ideal" values are punished. The so-called "general quadratic"
form, however, simplifies notation and computation. It will therefore be used for the
derivation of the algorithm:
(7)
~ ( ~:)' W (~:)
t •
(9)
The system dynamics (the equations of the econometric model) are assumed to be given
by
xt=f(Xt-l,Xt,Ut,(),Zt)+E t , t=S, ... ,T, (10)
where Xt-l denotes an (n X 1 )-vector of lagged state variables,
Xt denotes an (n X I)-vector of state variables,
Ut denotes an (m X 1 )-vector of control variables,
() denotes a (p x 1 )-vector of unknown parameters,
Zt denotes an (l x I)-vector of non-controlled exogenous variables, and
Et denotes an (n X 1 )-vector of stochastic disturbances.
() and ft, t = S, . .. ,T, are assumed to be independent random variables with known finite
expectations and covariance matrices:
() (O,L;88), (11)
E ~ (O,L;«). (12)
60
f1(.) )
f(.) = ( : (13)
m(. )
We demand that the first and second derivatives of the system function with respect to
:l:t-1, Xt,Ut and () exist and are continuous. The following notation is used:
i = 1, ... ,n
(14)
j = 1, ... ,n.
f". denotes an (n X n )-matrix the elements of which are defined by
(f ) . = 8£i(.) i = 1, ... ,n
(16)
u. ',) BUt,j j = 1, ... ,m.
fo denotes an (n X p )-matrix the elements of which are defined by
(,..'::", ) (~'~:'" )
ofl af l
:'"
OX._I.109. a"._I,nD9p
ofn
OX._l.106. 8 X t_l,n OO p
61
f",. ,9 denotes an (n . p x n )-matrix the structure of which is very similar to the above save
that all occurences of Xt-1 are replaced by Xt.
( au~fa91 )
af'
8u,t,Tn88 p
(19)
The system dynamics given in equation (10) are very comfortable for economic applications
of control theory. In most practical cases the so-called structural form of an economic
model can be easily transformed into the given system dynamics. For the derivation of an
algorithm, however, the state-space representation used in most control-theoretical texts
[Aoki 1967] is more adequate. It differs from (10) mainly in that it does not include Xt
at the left-hand side. We will show how one can eliminate this variable in the course of a
linearization of the system dynamics (cL [Chow 1981, pp. 15ff.]).
For known values of Xt-1, Ut, e, and Et we can compute a value Xt such that
(20)
Equation (25) implies for the expectation and the covariance matrix of ~t, conditional on
the information given at t - 1, that:
E t - 1 (tt) 0, (26)
COV t _ 1 (tt,tt) (I - ['",)-1 COY h, Et) [(1 - f""rT. (27)
62
It has to be kept in mind that all these matrices are functions of the reference path along
which they have been evaluated. If this path changes the matrices will also change. Thus,
we have been able to show how one can approximate the time-invariant nonlinear system
f(.) by time-varying linear system functions.
From the section above it is obvious that the matrices At, B t and the vector Ct are functions -
of the random parameter vector () and are, therefore, random themselves. Of course, both
matrices can be written as collections of their column vectors
Each of these column vectors as well as Ct are functions of (), which we indicate by the
notation
In order to obtain the matrices DB"i, Db"j, and DC, introduced above, one has to calculate
the derivatives of At, Be, Ct with respect to (). We use analogous notations as for the second
derivatives of the system equations.
(~)
8e,
8a,,11
CO.)
8e,
8at-ln
ae;- ao;:-
DA, = (36)
C~) (~)
86, 86,
8 a t,nl Bat,nn.
~ an;-
63
DB, denotes an (n. p x m )-matrix and DC, an (n· p x 1 )-vector of the following structures:
( ~) ( ~)
80, ao;-
8b",m) 8e,
(
8b~'11
88
8b,~,m
ao p
aCt,!
ae;;
p
DC, = (37)
8;;;n)
( ~) ( ~) (
88, 80 ,
The relationship between the matrices defined above and the matrices Da"i, Db"j, and
DC, can be found through inspection of their structure and is expressed by the following
identities:
[vec ((D a", ),) , ... , vec ((Da',i)') , ... , vec ((Da"n)')] , (38)
[vec ( (Db',,)') , ... , vec ( (Db',j)') , ... , vec ( (Db"m )') ] , (39)
Using the matrix differential calculus established by [MacRae 1974] it is possible to prove
that the matrices introduced above can be evaluated as follows:
Thus, we have found a way to compute the linear approximations (33)-(35) and can evalu-
ate the expectations and covariances of the parameters of the linearized system equations:
In the process of deriving a control algorithm for this model it will become necessary to
evaluate expecations such as E t- 1 (~A~KtBt), where K t denotes a deterministic symmetric
64
(50)
where
i = l, ... ,n
(52)
j = l, ... ,m·
i = l, ... ,n
[l:KA] ..
'.J
= tr [KtDa •.iCOV t_1 (0,0) (Dat,.)'],
j = I, . .. ,n,
(53)
i = 1, ... ,m
[l~KA] .. tr [KtDa.,iCOVt_l (0,0) (Db •.• )'] , (54)
t,J j = 1, ... ,n,
i = 1, .. . ,m
[l~KB] .. tr [KtDb"i COVt-d 0,0) (Db •• )'] , (55)
'.J j = 1, ... ,m,
[l;tKt tr [KtDc'COV t _1 (0,0) (Dat .• )'] , i = ,n,
1, ... (56)
The key idea of our new algorithm is to use Bellman's principle of optimality:
(59)
where J;( Xt-l) denotes the loss which is expected at the end of period t-1 for the remaining
periods t, . .. , T if the optimal policy is implemented during these periods. E t - 1 (.) denotes
65
conditional expectation; Xk-l, k = S, ... t, and Uk-I, k = S + 1, ... t, are known at the
time when we have to decide about Ut.
We are going to show by induction that J;( Xt-l) can be expressed as a quadratic function
of Xt-l:
(60)
for all periods t = S, . .. , T + 1.
Obviously, this is true for J T+1 (XT), as we do not count any losses after the end of the
optimization horizon and can thus define
HT+l 0, (61)
hr+l 0, (62)
h~+1 0, (63)
hY+1 0, (64)
hj+1 0. (65)
As the second step of the induction proof, we now have to show that if J;+1 (Xt) can be
expressed as a quadratic function of Xt then J;(Xt-d will also be quadratic in Xt-I' For
this purpose we define some auxiliary variables:
A"'x
t y~KA + E t- 1 (Ad K t E t- 1 (At), (68)
A XU
t (A~"')' , (69)
AU"
t y~KA + E t- 1 (E t )' KtE t- 1 (At) + WtXE t _1 (Ad, (70)
AtUU y~KB + E t- 1 (E t )' K tE t_1 (Et) + 2E t- 1 (Ed Wt" + W t U
", (71)
These definitions render it possible to combine L t( Xt, Ut) and J~+l (Xt) and write J t( Xt-l, Ut),
the function of expected accumulated loss, in a compact manner:
(77)
Minimization of this function with respect to 1lt yields the following feedback rule
(78)
66
where
_ (A,",")-1
t
AtU:C , (79)
_ (A~,")-l A~. (SO)
By substituting the feedback rule for Ut into J t(Xt-1,Ut) we can derive J;(Xt-I), the func-
tion of IlliIlliIllal expected accuIllulated loss, as
(Sl)
where
A"""
t
- A"'u
t
(A uu
t
)-l AU'"
t , (S2)
X"t - A"'u uu
t (A t )-l AU t , (S3)
A~ - ~ (An' (A~u)-1 A~, (S4)
A:, (85)
Af· (86)
Thus, it has been proved by induction that J;(X t _1) is a quadratic function of Xt-1 for all
periods t = S, ... , T.
As can be seen from the derivation, it would be necessary to evaluate several conditional
expectations such as E t - 1 (At), E t - 1 (B t ), and E t - 1 (Ct). To simplify the problem, we
exclude any learning about the parameters and substitute ES-1 (.) for E t - 1 (.) for all t =
S + 1, ... ,T +1. Analogous assumptions are used for the covariance matrices, all occurences
of COV t _ 1 (.,.) are replaced by COY S-1 (., .). These simplifying assumptions make the
algorithm which is summarized in Figure 1 fully operational.
In this section we report about an application of the optimal control algorithm described
in the previous sections to a small macroeconometric model of the Austrian economy. The
primary purpose of this exercise is to test the algorithm and its implementation in the
programming system GAUSS. Moreover, to some extent the results of the optimal control
experiments may also serve as a first basis for an evaluation of Austrian macroeconomic
stabilization policies, although for this purpose the interpretations must be very carefully
carried out due to the simple nature of the model and of the objective function applied.
An optimization experiment is described which considers all the parameters of the model
as deterministic; this is supplemented by two experiments where the stochastic nature of
some key parameters is taken into account.
We consider a small macroeconometric model for Austria, which is a slightly revised and
updated version of the model OPTPOL-1 previously used for optimal control experiments
67
1. Use the Gauss-Seidel algorithm and a tentative policy path (Ut);=s (which has to be
submitted as input) to calculate a trial solution (Xt);=s'
(a) Linearize the system equations around the reference values Xt-I, Xt, Ut, 0, and
€ = On to get At, B t , Ct, and COV t _ 1 (~t, ~t).
(b) Compute influence of stochastic parameters. Obtain TfKA, TfKA, TfKB,
TfKc, TfKc, T~Kc.
3. Use the computed feedback rules and the system equations to project expected op-
timal state (x;);=s and control trajectories (u;);=s'
5. Check convergence. If for all state and control variables the relative deviations of the
reference values (Xt);=s, (Ut);=s used for linearization from the projected optimal val-
ues (x;);=s, (u;);=s are smaller than a predefined tolerance (e.g. 0.001), then STOP.
Otherwise, if the number of iterations already exceeds a predefined maximum, then
STOP without convergence. Otherwise, use the projected optimal state variables
(x~);=s and the projected optimal control variables (u;);=s as the new reference path
(Xt)t=s, (ut);=s and go to step 2 to start a new iteration.
with the Chow program [Neck & Posch 1982] [Neck & Posch 1984]. Its theoretical basis
is a Keynesian one in the sense of textbook-IS-LM-models. The model consists of four
behavioral equations and four identities. Data have been taken from the database of the
Austrian Institute of Economic Research (WIFO). All real data have dimension Billions
of 1976 Austrian Schillings.
The model contains eight endogenous (state) variables: real private consumption (CR),
real fixed investment (IR), real imports of goods and services (MR), the long-term bond
yield as an interest rate variable (R), real gross domestic product at market prices (YR),
real total aggregate demand (VR), the total-demand deflator (1976 = 100) as a measure
for the general price level (PV), and the rate of inflation (PV%). There are three control
variables: the net tax rate (net tax receipts as percentage of GDP, T%) and real public
consumption (G R) as instruments of fiscal policy, and the nominal stock of money supply
M1 (M1) as an instrument of monetary policy. Finally, three exogenous non-controlled
variables are present: the deflator of imports of goods and services as a measure of the
import price level (PM), real autonomous expenditures (real exports of goods and services
plus inventory changes including errors and omissions from the national income accounts
statistics, AR), and the deflator of GDP at market prices as a measure for the domestic
price level (PY).
Behavioral equations for CR, IR, MR and R have been estimated for annual data by OLS,
using the software system lAS. The estimation period has been 1964 to 1988. All behav-
ioral equations are of a form that allows their immediate input into the optimal control
algorithm. They are formulated according to the partial-adjustment hypothesis. Thus
the consumption function conforms to the permanent-income and the habit-persistence
hypothesis; in addition, it takes into account an influence of the real rate of interest on
consumption expenditures. The investment function embodies an accelerator and an influ-
ence of the real rate of interest on investment expenditures. Imports depend on aggregate
total demand and on the relative price of imports. The equation for the rate of inter-
est is a reduced-form version of a demand-for-money function, assuming real demand for
money to depend on real GDP and the nominal rate of interest, and to be equal to real
money supply. Although some of the estimated coefficients are not significant according
to the usual statistical criteria, the present version has been used for the optimization
experiments; those coefficients which are not significant qualify as natural candidates for
stochastic parameters in the experiments where random parameters are taken into account.
Figure 2 shows the results of the estimations and the identities of the model, together with
the statistical characteristics of the four regressions. Values in brackets below regression co-
efficients denote estimated standard deviations, t-values and standardized beta-coefficients
of the parameters, respectively. R2 is the coefficient of determination, R~ the coefficient of
determination adjusted for the degrees of freedom, SE ist the estimated standard error of
the equation, MAPE ist the mean absolute percentage error, DW is the Durbin-vVatson
statistic for serial correlation, and p is the estimated first-order autocorrelation coefficient
of the residuals.
As can be seen from the equations in Figure 2, the model is nonlinear. For simulations, it
has to be solved numerically; in the optimal control algorithm, this is done by the Gauss-
69
T%t
CRt = 0.3061 CR t _ 1 + 0.63312 Y R t (1- - ) - 1.81043 (Rt - PV%t) + 5.27457
(0.14364) (0.13237) 100 (0.74461) (5.904)
(2.13) ( 4.78) (2.43) (0.89)
(29.1) (66.9) ( 4.0)
PMt
MR t = 0.21599 M R t - 1 + 0.28844 V R t - 0.93284 ( - ·100) + 13.43473
(0.14035) (0.05414) (0.55423) PVi (69.10811)
(1.54) (5.33) (1.68) (0.19)
(20.8) (73.1) (6.1)
MIt
Rt = 0.792 R t _ 1 - 0.01857 (-·100)+ 0.00169 Y R t + 2.76811
(0.1604) (0.01755) PVi (0.00228) (1.31912)
( 4.94) (1.06 ) (0.74) (2.10)
(57.5) (24.7) (17.8)
8. Rate of inflation
PV%, = (PVi - PVi-tl/PVi-l ·100
Seidel method. In a dynamic simulation of the model over the estimation period using
historical values for all exogenous variables, it turns out that the dynamic behavior of total
demand and its components is rather close to the historical movements of these variables;
thus the model gives a satisfactory picture of Austrian macroeconomic developments. One
major drawback of the model is its neglect of the supply side of the Austrian economy;
it is essentially a demand-side model with the price sector being in fact exogenous. Most
of the macroeconometric models for Austria that include aggregate supply are fairly com-
plex; so far, we have not been able to integrate any of them into a model to be used
for optimization experiments. Furthermore, it should be mentioned that for the present
purpose we assume that fiscal and monetary policies are conducted by the same authority;
thus, government and central bank are considered to be one homogeneous policy-maker.
In Austria, this might be not too unrealistic since the independence of the central bank
from the government is rather limited.
In addition to the model, we have to specify an intertemporal objective function for the
optimal control experiments. Unfortunately, economic theory or the theory of economic
policy do not provide much support for this task; thus the objective function is inevitably
rather ad-hoc. For the present purpose, we assume the quadratic-tracking function as used
by our algorithm. The time horizon for the optimal control experiments has been chosen
to be 1971 to 1988. From mid-1970 to 1986 the Austrian government was dominated
by the Socialist Party (SPO); thus, it may be suspected that during this period historical
stabilization policies were relatively homogeneous. The last two years are included because
it is well known that in finite-horizon dynamic optimization problems the controls may
exhibit strange behavior during the last years due to the neglect of the future. Thus,
even if we are only interested in optimal policies until 1986, it is advisable to extend the
optimizations a few years further.
For the "ideal" values of the state and control variables, we make the following assumptions:
For 1970, historical values of all variables are assumed to be given. For the following years
(1971 to 1988), we postulate growth rates of 3.5% p. a. for real total demand and its
components, i. e. for real GDP, private and public consumption, fixed investment and
imports of goods and services. A growth rate of 2% p. a. is considered as "ideal" for the
price level (the deflator of total demand) and a constant value of 2 is assumed as "ideal"
rate of inflation. The rate of interest has an "ideal" constant value of 2 for all periods. For
the control variable net tax rate, the historical value of 1970 is used as "ideal" value for
all periods. For nominal money supply, we assume an "ideal" growth rate of 5.5% p. a.
from its 1970 historical value.
This choice of the "ideal" values for state and control variables can be justified as follows:
We assume that the values of the economic variables in 1970 cannot be influenced by a
fictitious government starting to plan for a time horizon which coincides with the period
when the Socialist-dominated government was in office. Thus, the initial values of 1970
are regarded as the heritage from the previous administrations. The aim of the fictitious
government is to stabilize the economy in the sense of achieving smooth growth paths of real
71
and nominal variables without excessive inflation. The value of 3.5% for the growth rate of
real variables is slightly above the average growth rate of historical GDP at market prices
over the planning horizon and can therefore be regarded as a realistic goal. Attaching the
same "ideal" growth rates to all real variables of the model means that the hypothetical
policy-maker has the aim of achieving a "balanced growth" of all components of total
demand. The rate of interest and the money supply can be interpreted as intermediate
targets; a low rate of interest may serve to expand aggregate demand, and a money stock
growing at the same rate as "ideal" nominal GDP (3.5% real plus 2% price level growth)
may be interpreted to be neutral with respect to inflation.
For the weighting matrices W t we assume constancy over time (a = 1); all off-diagonal
elements of this matrix are zero, and the main diagonal elements have the following weights:
variable CR IR MR R YR VR PV PV% T% GR M1
weight 5 5 5 2.5 10 0 5 0 5 5 1
Real GDP at market prices is regarded as the main target variable; its components and
real imports get smaller weights to express the idea of "balanced growth" and to achieve
stabilization of the shares of the components of total aggregate demand which itself is
not penalized directly for deviations from its "ideal" values. The general price level is
also penalized, although this will not result in a visible effect since this variable (though
formally endogenous) is in fact determined by the exogenous price level variables PM and
PY. The constant growth target for real public consumption and the constant level target
for the net tax rate have the same weights; in this way, the aim of a constant allocation
between the private and the public sector is expressed. This device may also be regarded
as a substitute for a government budget constraint absent from our model. The weights
given to the intermediate target variables rate of interest and money supply are intended
to prevent excessive fluctuations of these variables.
Three different control experiments are performed: In the first one, all parameters of the
model are regarded known with certainty. The only stochastic influences considered are
the additive error terms in the behavioral equations, whose variances contribute to the
optimal value of the objective function but do not alter the optimal policies as compared
to a purely deterministic set-up. Since we have only 01S estimates at our disposal, we
have no estimate of the variance-covariance matrix of the additive error terms. Instead,
we assume it to be a diagonal matrix, with the squared estimated standard errors of the
behavioral equations in the main diagonal. Also, the values ofthe exogenous non-controlled
variables are assumed to be known for all time periods in advance.
In the second and third experiments we tentatively introduce some stochastic parameters.
In the second experiment those parameters whose estimates have the lowest t-values are
regarded as stochastic. These are the interest rate coefficient in the investment function
and the money stock and the GDP coefficient in the interest rate equation. The estimated
values of the regression coefficients and their estimated standard deviations are used as
expected values and standard deviations, respectively, of these parameters in the algorithm.
Since these insignificant coefficients are exactly those which are crucial for the monetary
72
policy multiplier, by this experiment we can also investigate what happens to optimal
policies when the effectiveness of monetary policy becomes more uncertain.
In the third experiment the marginal prospensity to consume (the disposable-income coef-
ficient in the consumption equation) is added as fourth stochastic parameter, again taking
the estimated coefficient and its estimated standard deviation as the first and second mo-
ments of that parameter, respectively. This allows us to study the effects of making fiscal
policy multipliers stochastic. Although the effectiveness of fiscal policy depends on several
parameters in our model, the one chosen here is particularly crucial for both the tax rate
and the public consumption multiplier.
Fully taking account of the stochastic nature of all the parameters of the model would
require an estimate of the entire variance-covariance matrix of the coefficients of the model.
It could be obtained from a simultaneous-equations estimation, but the running time of
the current version of the computer program would be very high.
One final remark should be made concerning the interpretation of the optimization ex-
periments. We do not claim to have a positive model of stabilization policy-making in
Austria. Obviously, for that purpose many additional features would have to be intro-
duced, apart from the question whether Austrian policy-makers really want to optimize
an intertemporal objective function. Also, a normative interpretation should be subject
to several caveats. In particular, the policies calculated are optimal only under the par-
ticular objective function and provided that the structure of the model is in fact valid.
Moreover, using a model which is estimated with data up to 1988 and assuming all values
of exogenous variables up to that year to be known, means using more information than
a policy-maker of 1970 could have obtained even under ideal circumstances. Instead, the
optimization experiments could be regarded as exercises in an "experimental evaluation"
of the model and the objective function; by performing optimal-control experiments, we
want to learn about the response of the model economy to variations in the assumptions
about the objectives of policy-making and about the uncertainties inherent in the model.
Thus we are performing extended simulation experiments guided by an objective function
to be optimized. Needless to say, even this restricted interpretation of the optimal control
framework for macroeconomic policies would be meaningless if the Lucas critique were
fully valid; a Keynesian view of the potential effectiveness of monetary and fiscal policies
is a necessary requirement for these optimization experiments.
The three optimal control experiments described above were carried out on an IBM com-
patible 12 MHz PC-AT with a 80287 mathematical coprocessor. The algorithm was imple-
mented in the programming language GAUSS. It took about 20-30 minutes to arrive at an
approximately optimal solution. 'With a tolerance level of 0.001, on average, 6 iterations
were required for the Gauss-Seidel algorithm to converge, and the optimal policies were ob-
tained after three to five iterations of the stochastic control algorithm. Thus, convergence
appears to be rather rapid for our small econometric model, but the running time on the
PC is unsatisfactory and prevents the use of larger (and hence more realistic) econometric
73
The time paths of the expected optimal control variables are displayed in Table 1. Addi-
tionally, in Figure 3 and in Figure 4 historical and target values are compared with the
expected optimal values for two selected variables, namely the control variable real public
consumption and the main target variable real GDP. All data in Table 1 and Figures 3
and 4 are given in Billions of Austrian Schillings. In Figures 3 and 4 the labels 'Stoch(3)'
and 'Stoch( 4)' refer to the control experiments with three and four stochastic parame-
ters, respectively, and the label 'Determin' denotes the results of the experiment without
stochastic parameters.
The upper portion of Table 1 shows the results when all parameters are assumed to be
deterministic. This experiment shows that fluctuations of the main objective variables can
be stabilized to some extent by optimal policies. The optimal values of the control variables
exhibit counter-cyclical behaviour; particularly fiscal policies (the variables T% and GR)
are set in a more active way than historically to combat the recessions in 1975 and 1981.
During most years the optimal net tax rate is lower and real public consumption is higher
in the optimization experiment than they were historically; thus fiscal policy is exerted in
an expansionary way, which is plausible given the Keynesian features of our econometric
model. On the other hand, optimal money supply has a relatively smooth path over
the optimization horizon, with values higher than historical until 1973 and lower values
afterwards. This results in higher values for the rate of interest, which nevertheless do not
seem to crowd out either real private consumption or fixed investment. The optimal values
of real imports and real GDP are below the historical values during the initial years and
above afterwards. Thus, within our model optimal policies are capable of obtaining smooth
growth paths for the main target variables and can effectively overcome the prolonged
period of low growth of the eighties following the two oil price shocks.
The central part of Table 1 shows the results of the experiment where the three parameters
relating to monetary policy are assumed to be stochastic. The differences to the previously
described run are minor. Money supply is smaller during most years; this, however, has no
visible effect on the endogenous variables apart from the rate of interest which is slightly
higher. This indicates that monetary policy is rather ineffective in our model. Making its
impact more uncertain means that it is even less used for stabilization purposes and kept
closer to its "ideal" path (which is unrealistically low due to the low "ideal" values of the
price level).
Finally, the lower portion of Table 1 shows the results from a run where in addition to the
above the marginal propensity to consume is regarded stochastic. Here we can observe
much more pronounced differences to the first run: The net tax rate is much higher, and
so are real public consumption and the money supply which counteract the restrictive
effects of high taxes on real GDP. As must be expected, this scenario results in lower levels
of consumption and imports and also of real GDP, whereas investment is higher due to
lower rates of interest as compared to the run where the parameters were assumed to be
deterministic. Similar results were obtained from a run where only the marginal propensity
to consume was regarded as stochastic. These outcomes are somewhat surprising because
they indicate that when fiscal policy effects become more uncertain higher levels of fiscal
74
/I CR I IRI MR I R YR I VR PV PV% T% GR M1
no stochastic parameters
1971 324.620 148.225 166.493 7.151 586.280 752.773 70.930 5.856 38.841 108.607 94.278
1972 335.023 155.289 182.030 7.028 604.576 786.606 75.401 6.304 39.469 111.426 98.482
1973 341.049 163.356 199.151 7.020 626.022 825.173 80.755 7.100 41.193 109.380 103.155
1974 349.097 170.733 206.420 7.206 649.048 855.469 90.004 11.453 43.606 108.295 107.555
1975 373.661 174.037 216.951 7.393 663.989 880.940 95.302 5.887 37.140 126.251 113.147
1976 377.100 180.654 233.373 7.593 690.086 923.458 100.000 4.929 39.971 120.291 118.283
1977 390.756 185.480 248.294 7.793 712.147 960.440 104.992 4.992 38.980 125.131 123.931
1978 407.213 189.984 264.636 7.986 734.144 998.780 109.623 4.411 37.866 132.678 129.576
1979 413.458 195.505 279.784 8.194 762.988 1042.772 114.623 4.561 39.943 128.935 135.092
1980 423.592 200.452 292.070 8.453 791.198 1083.268 121.772 6.238 40.955 128.815 140.462
1981 446.513 203.622 301.723 8.773 814.377 1116.100 130.756 7.377 39.080 141.702 145.477
1982 468.262 207.219 318.270 9.097 838.257 1156.527 137.277 4.987 37.113 153.073 150.497
1983 489.637 210.414 337.466 9.378 863.715 1201.181 140.761 2.538 35.372 163.856 155.725
1984 499.887 215.947 357.276 9.684 896.047 1253.322 147.322 4.661 37.630 162.225 160.723
1985 516.852 219.377 375.330 9.967 926.303 1301.633 151.708 2.977 36.702 167.352 166.395
1986 548.483 221.745 400.401 10.176 950.108 1350.509 153.081 0.905 33.207 187.307 172.422
1987 574.419 225.636 425.612 10.337 981.300 1406.912 154.354 0.832 32.645 200.762 178.603
1988 588.865 233.136 451.485 10.491 1027.241 1478.726 156.692 1.514 34.804 201.884 185.649
three stochastic parameters
1971 324.556 148.060 166.412 7.383 586.082 752.494 70.930 5.856 38.725 108.558 85.413
1972 335.010 154.958 181.919 7.417 604.364 786.284 75.401 6.304 39.262 111.448 90.140
1973 341.109 162.835 199.024 7.498 625.791 824.816 80.755 7.101 40.937 109.485 95.758
1974 349.235 170.032 206.283 7.693 648.806 855.089 90.004 11.453 43.341 108.479 102.245
1975 373.857 173.184 216.800 7.832 663.721 880.521 95.302 5.887 36.889 126.491 110.380
1976 377.348 179.691 233.216 7.974 689.812 923.028 100.000 4.929 39.746 120.576 116.488
1977 391.044 184.436 248.134 8.124 711.871 960.005 104.992 4.992 38.777 125.452 122.284
1978 407.530 188.884 264.476 8.273 733.868 998.345 109.623 4.411 37.681 133.024 128.050
1979 413.799 194.374 279.627 8.443 762.721 1'042.347 114.623 4.561 39.778 129.301 133.744
1980 423.949 199.305 291.917 8.666 790.938 1082.855 121.772 6.237 40.808 129.193 139.379
1981 446.874 202.477 301.573 8.946 814.123 1115.697 130.755 7.377 38.951 142.082 145.115
1982 468.618 206.094 318.123 9.226 838.006 1156.129 137.277 4.988 37.004 153.444 151.085
1983 489.984 209.326 337.325 9.471 863.476 1200.801 140.761 2.538 35.279 164.216 156.401
1984 500.228 214.907 357.145 9.752 895.830 1252.975 147.322 4.661 37.552 162.576 161.103
1985 517.178 218.388 375.208 10.017 926.098 1301.306 151.708 2.977 36.635 167.687 166.707
1986 548.795 220.805 400.286 10.221 949.916 1350.202 153.082 0.906 33.144 187.627 171.903
1987 574.726 224.740 425.508 10.394 981.129 1406.637 154.355 0.832 32.583 201.076 176.832
1988 589.181 232.279 451.400 10.570 1027.110 1478.510 156.692 1.514 34.740 202.210 182.778
four stochastic parameters
1971 300.690 145.926 162.899 5.660 577.442 740.341 70.924 5.848 45.185 122.405 150.664
1972 308.530 153.957 177.357 4.528 595.749 773.106 75.399 6.309 44.857 125.751 151.437
1973 314.524 163.555 194.365 3.865 617.699 812.064 80.759 7.108 46.574 122.598 153.662
1974 324.787 172.736 201.987 3.689 641.728 843.715 89.996 11.438 48.508 118.850 156.291
1975 341.660 176.851 211.766 3.603 654.533 866.299 95.298 5.892 43.816 140.798 163.852
1976 346.109 185.179 228.279 3.647 681.403 909.682 100.000 4.934 45.777 132.980 168.353
1977 357.673 191.314 243.079 3.778 703.093 946.172 104.993 4.993 45.129 138.111 173.392
1978 370.866 196.925 259.109 3.962 724.397 983.506 109.629 4.415 44.337 146.809 178.437
1979 377.350 203.811 274.325 4.216 753.651 1027.976 114.627 4.559 45.943 141.941 182.949
1980 387.635 209.960 286.723 4.575 782.102 1068.825 121.770 6.232 46.714 140.823 187.103
1981 407.573 213.853 296.082 5.039 804.488 1100.570 130.747 7.372 45.140 154.881 190.938
1982 425.675 217.982 312.156 5.519 827.402 1139.557 137.275 4.993 43.381 167.927 195.058
1983 442.255 221.502 330.712 5.962 851.621 1182.334 140.767 2.543 41.973 181.303 198.370
1984 454.014 227.879 350.616 6.477 884.663 1235.279 147.329 4.662 43.465 178.123 198.978
1985 469.371 231.667 368.512 6.986 914.457 1282.969 151.715 2.977 42.590 i
183.879 200.744
1986 494.122 233.719 392.707 7.442 936.194 1328.901 153.103 0.915 39.771 1208.086 201.244
1987 514.881 237.365 417.106 7.906 966.029 1383.134 154.384 0.837 39.315 224.793 198.612
1988 516.275 243.940 441.520 8.430 1008.979 1450.499 156.721 1.514 42.719 235.443 194.537
o
Re 1 public consumption
....
C\l
o
C\l
C\l
,.
0 - ~-!
C\l ./
o
co -: /
. I' "
/::-... ,..... - ~---
o
C\l
7 1 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88
Hi s toric Sloch( 4)
Targel Stoch(3)
----. --- Determ in
Figure 3: Real public consumption
.-~
CD " ;~ /; ---:_/.
o
. ...
. ..-
~ .
.
/ "
",.... r-- .
t --:
<0
c:i
.-r. --t ---
~ L-~~__~-L~~~~__L--L~__~-L~__~-L~__~~
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88
Historic Stoch(4)
Targ l Stoch(3)
---- ~ --- Determ in
policy variables become optimal, which is difficult to interpret in economic terms. Further
experiments are required to check whether this conclusion holds more generally.
5 Concluding remarks
In this paper, we have developed an algorithm for the optimal control of nonlinear dynamic
macroeconometric models with stochastic additive error terms and stochastic parameters
under a quadratic intertemporal objective function. This algorithm has been implemented
in the programming language GAUSS and applied to a small econometric model of the
Austrian economy in order to show the feasibility of the algorithm. The optimization
experiments show that optimal policies lead to a considerable stabilization of the time paths
of the main objective variables of the model. Results of two different runs with stochastic
parameters have been presented to demonstrate the influence of parameter uncertainty.
For the currently available PC version of the program, runtime is a serious concern. The
most time-consuming part of the program happens to be the Gauss-Seidel solution of the
system. Thus, the projection of the state variables is a very expensive operation. Also,
computing first and second derivatives of the system equations accounts for an important
part of the runtime as it has to be done for each time period. The use of a symbolic
instead of a numerical derivation algorithm could certainly help reducing the runtime by
a considerable amount.
For the algorithm itself, there exist several possible extensions. By adding updating equa-
tions for the stochastic parameters such as the ones used by [MacRae 1975] it could be
expanded into a passive learning algorithm. A more difficult task would be the exten-
sion of the algorithm to include active learning, i. e., the derivation of approximations
to closed-loop policies in the sense of [Kendrick 1981]. Another interesting extension to
be considered in the future will be the examination of the effects of decentralized policy-
making; here results of decentralized control theory (dynamic team theory) and dynamic
game theory will have to be incorporated.
77
References
[Neck & Posch 1982J Reinhard Neck and Ursula Posch. "On the 'Optimality' of
Macroecononric Policies: An Application to Austria". In:
G. Feichtinger (ed.). Optimal Control Theory and Economic
Analysis. Amsterdam et al.: North Holland, 1982, pp. 209-
230.
[Neck & Posch 1984J Reinhard Neck and Ursula Posch. "On the Sensitivity of Opti-
mal Macroecononric Policies to Econometric Models: An Em-
pirical Analysis for Austria". In: T. Basar, L.F.Pau (eds.).
Dynamic Modelling and Control of National Economies 1983.
Oxford et al.: Pergamon Press, 1984, pp. 91-98.
[N orman 1979 J Alfred 1. Norman. "First Order Dual Control", Annals of Eco-
nomic and Social Measurement, 5/3 (1976), pp. 311-321.
Game Theoretical Approach to Optimal Control of Economic Systems
Krystyna Strzala
Department of Econometrics
University of Gdansk
Armii Czerwonej 101
PI-81-824 Sopot
Abstract
Contents
min E(W) (1 )
x E X
subject to:
Yt = At Yt - 1 + CtX t + b t + v t (2 )
YO given (3 )
where:
Yt p-vector of state variables,
xt m-vector of control variables,
At' C t pxp and pxm nonstochastic parameter matrices,
Et expectation operator conditional on information
available at the beginning of period t,
at p-vector of targets towards which Yt is aimed,
K positive semi-definite matrix (often diagonal),
t
T planning horizon,
bt p-vector which measures the influence of
noncontrollable exogenous variables,
v p-vector of random disturbances with expectation
t
2
(pxp)-covariance matrix E(vtv )=a
s v t
for s=t and E(Vtvs)=O otherwise.
Diagonal elements of the K matrix are penalty weights of squared
deviations of the state and control variables from their
targets. Therefore, a linear-quadratic stochastic optimal control
problem is to minimize the expectation of the quadratic loss function
(1) given a linear stochastic dynamic model in a state-space form
(2) and initial conditions (3).
General techniques used for solving a linear-quadratic optimal
control problem are the Minimum Principle and Dynamic Programming
(in the former case, see: Pontriagin et al. (1962), Intriligator
(1971), Feichtinger and Hartl (1986), in the latter case: Bellman
(1957), Chow (1975, 1981).
The most favored algorithm in economic applications is that
proposed by Chow (197 5) which leads to a set of time-varying
feedback control rules of the form :
t=1,2, ... ,T (4 )
where: Ht ; Kta T = hT ; c T =
-(CtHtCt)-1(CtHtAt)
-(CtHtCt)-lCt(Htbt - h t ) ( 5)
81
t=1,2, ... ,T (7 )
where the functions lt' t=l,2, ... ,T, are known at t=l.
In other words, under a regime of closed-loop information
structure, values of present and future policy instruments are at
t=1 known functions of some information that will be available
when values of these instruments have actually to be assigned.
Strategies or control rules are in this case functions of time and
the state of the system. If uncertainty is present, this means
that values of instruments in periods ">1 can respond to
observations (realizations) of random variables between periods 1
and T.
In both cases each player forecasts the other players' actions and
formulates his best response. An important distinction, however,
is that under a closed-loop information structure each player is
announcing a rule describing how he wi 11 respond to some new
information that may accrue during a course of a game. ThUS, each
player, when computing his response, recognizes that his decision
affects his rivals' response, because each player act ion affects
the state of a system and the state of the system determines
control rules via the closed-loop assumption.
Without loss of generality let's assume there are only two
decision makers. Following the traditional approach, (compare with
(1)-(3», a two-person nonzero-sum noncooperative dynamic game
over the horizon (1, TI can be stated straightforwardly in the
linear-quadratic setup by distinguishing between instruments under
84
i = 1,2
where:
x it - (m i x1)-vector of policy variables (instruments) of
decision maker i at period t, i=1,2, m1 + m2 = m;
Cit - (nxmi)-nonstochastic parameter matrices, i=1,2;
L. - loss of decision maker i after T planning periods;
1
a it - p-target vector of the i-th decision maker;
Kit - (pxp) weighting matrix of the i-th decision maker at
period t;
other symbols are defined as in (1)-(3).
The optimization problem regarding more than one decision maker
cannot be solved without assumptions about the behavior of all
decision makers. Several modes of behavior are possible e.g.
noncooperative behavior in the sense of Nash; cooperative behavior
in the Nash-Harsanyi sense; and leader-follower behavior in the
sense of Stackelberg (see Nash (1951), Harsanyi (1977), Kydland
(1975) ) .
Under the Nash assumption, all players (the government and private
agents in an economic system) have equal status in the game. They
have no perceptible influence on the behavior of the others and
therefore they treat the rivals' strategies as parametric.
The solution concept of Nash has the important prop'~rty that in
equilibrium a non-decision maker has an incentive to deviate from
his Nash-strategy if other decision makers continue to use their
equilibrium strategies (see e.g. Basar and Olsder (1982».
A noncooperative Nash-equilibrium strategy
S~ E S. is the strategy with the property:
l l
(10)
-I< 1<-1<
2 1<- I<
Y A Yo + E C.X. + b (12)
1. 1.
i=l
yO and decisions of other players are given, (13)
= Rt Yt-1 + r t ,
where:
Rt=At+CtG t , (17)
rt=bt+Ctg t ,
all symbols marked by a bar mean deterministic variables, other
symbols have the same meaning as in (8)-(9).
This yields an open-loop Nash solution xt for t=l, .. , T, since
vectors of policy variables are incorporated in the state vector.
As in the former case a solution is obtained for the deterministic
model,"the certainty equivalent" of the original stochastic model.
Two other algorithms for obtaining an open-loop Nash solution in
the linear-quadratic discrete time setup have been invented by
Pindyck (1976,1977) and de Zeeuw (1984).
In both algorithms the method employed for the calculation of
open-loop Nash strategies is the discrete-time minimum principle
(see Canon, Cullum and Polak (1970), Pindyck (1973)).
Formally the optimization problem is stated as follows:
Jo
1
2
+ E
j=1
(u °t-U °t) 'Ro ° (u °t- U °t) I
J J 1J J J
'* min i=1,2 (18)
subject to:
where:
Xt - vector of n state variables;
U j t - vectors of q1 and q2 control variables, j=1,2, ql+q2=q;
Zt - vector of s noncontrollable exogenous variables;
A,B l ,B 2 , and nxn, and nxs matrices;
87
(21 )
88
At + CtG t
b t + Ctg t
~~_~~~~_~~~~~~~_~~~~!_~~~!~g~!~~~_~f_~~~~=g~~~_!~_!~~_~~~~~~g~!
_~g~~~~ig_~~~!Y§1~
References
Ivan Sujan
Institute of Socio-economic Information
and Automation in Management (VUSEI-AR)
Dubravska 3, 842 21 Bratislava, CSFR
and
Osvald Vasicek
Institute for Automation of Management
in Industry (INORGA), branch Brno
n. kpt. Jarose 32, 656 01 Brno, CSFR
Abstract
The solution of the optimal control problem for nonlinear systems may be properly
approximated by an iterative procedure, in which the optimal control algorithm is
subsequently applied to specific linearizations of the original system. The weights for
target and control variables used in the optimization functional may be derived from the
variation coefficients of their deviations from nominal values and from the assigned
relative importance of variables in a given variant.
Contents
Introduction
2 Application of optimal control methods to nonlinear systems
3 Constructing the weights for deviations of variables from their desired trajectories
4 The econometric model used in optimal control simulations
5 Scenarios of simulated development strategies
6 The results of optimal control simulations
7 Conclusion
References
Appendix: Structural equations of the Model CEM-4M
95
Introduction
For the proper evaluation of alternative development strategies, it may be helpful to find
the optimal values of selected policy instruments, which are needed for reaching the
target values of individual strategies. Such problems may be solved by using recent
developments in optimal control methods applied to a properly adjusted econometric
model. The optimal control problem is to find such values of selected control variables,
which minimize the weighted sum of squares of the deviations of target and control
variables from their desired (nominal) values [see e.g. Chow (1981) ].
Applications of optimal control methods are usually limited to systems of linear equations
[King et al. (1982), Vasicek and $ujan (1987)]. In this paper we present the procedure
and a practical application of the optimal control of a nonlinear system. In section 2 we
describe the problem of optimal control of a linear system with parameters varying in
time and a procedure for the iterative approximation of the optimal control of
a nonlinear system. Section 3 contains some rules for constructing the weigths for the
squared deviations of target and control variables from their nominal trajectories. In
further sections we present the nonlinear econometric model of the Czechoslovak
economy used, the scenarios of the baseline projection and three alternative
development strategies, and the results of the optimal control simulations of individual
strategies. Conclusions are summarized in section 7.
where Yt' v!' and et are pX1, qX1, and pX1 vectors of endogenous, exogenous, and
residual (random) variables at time t, respectively, F is a vector of given continuous
functions, h is the maximum lag and N is the number of periods under consideration.
For the application of optimal control methods some transformations of the model (1)
are needed:
1) This means a separation of the dynamic core of the model with lagged variables.
96
(f) transformation of the time t to the model time i according to the relations:
Let us consider first the optimal control problem under the assumption that the system
(1) is linear with parameters varying in time. In this case the system (1) after
transformations according to (a) - (f) may be written as an identified discrete dynamic
system described by the vector difference equation of the first order (state space
equation) for Xi+ 1 ' the output equation for Yj and the known initial condition for xo:
where xi' YI' uj , <; are nX1, mX1, rX1, sX1 vectors of state, output, control and
exogenous variables at time i, respectively; A., Bi , GI , Dj are given nXn, nXr, nXs, mXn
matrices of time-varying parameters and c is a constant vector of the initial state of
XI·
Let 91 and ul be the given desired (nominal) values of output and control vectors for the
optimization periods i = 0, 1, ... , N-1. Ignoring the mathematical expectations operator
E, the optimization criterion is the minimum value of the quadratic functional J given by
In this case the optimal control problem is to find an optimal sequence of the control
vectors(l.f}such that (6) is minimized subject to the constraints (3) - (5). The
procedure of solution is given e.g. in Vasitek (1988).
Optimal control methods may also be applied to systems of nonlinear equations. Let us
assume now that the system (1) is nonlinear and after transformations (a) - (f) may be
written as
(7) x +1 =
j f(xj' u l ' Zj),
(8) Yj = g(x j ),
(9) Xo = c,
where f and g are given continuous vector functions. The optimal control problem may
be formulated like in the preceding case, Le. find the optimal sequence {un which
minimizes (6) subject to the contraints (7) - (9).
The optimal control of the nonlinear system (7) - (9) may be practically solved by an
iterative procedure as a limit of optimal control sequences for linearized systems with
97
time-varying parameters [Vasicek (1989)]. In the starting iteration (k=D) we solve the
optimal control problem for the system (7) - (9) linearized along the trajectories of the
variables obtained as a response of the system to the nominal values of the control
vectors u, .
In further iterations the system is linearized along the trajectories obtained
as a response to the optimal vectors (k-,') u· found in the preceding iteration. For
convergent iterations the optimal control of the original nonlinear system is approximated
with the desired precision when the following relations hold simultaneously
Linearization of the system (7) - (9) along the trajectories (k) u,1={(k-l) un and {(k) x,!,
(k)y,) (obtained as a response to {(k)U,),{Zt) and xo) yields in subsequent iterations
a linearized model for the differences
with the matrices of parameters calculated as the partial derivatives of the functions
f and g along the response trajectories in periods i = 0, 1, ... , N-1 :
The nominal output and control vectors for the linearized model are given in subsequent
iterations and periods by
After substitution of l:,.(k) x" l:,.(k) y" l:,.(k) U,' (k) A, (k) B" (k) 0, for x" y" u" A, Bp 0"
respectively, into (3) - (5)2) and substitution of (k) 9, and (k) 0, for 9, and into (6), the u,
optimal control problem for the linearized system (3) - (6) is solved, yielding the optimal
sequence {l:,.(k) ui1. The optimal control vectors for the original nonlinear system are in
the kill iteration approximated by
When (10a) and (10b) are satisfied, the iterations are finished and the sequence of the
optimal control vectors for the nonlinear system u~ is approximated by the results from
the last iteration, i.e. (k) u~
The convergence of the described algorithm has not yet been generally proved.
However, for nonlinear relations which are typical for econometric models (such as
power, exponential and logarithmic functions, products and ratios of variables, etc.) we
have always obtained satisfactory results. The algorithm is implemented by the program
RIC written in FORTRAN IV and working under the operation system OS/SVS on the
computer EC 1045 [Vasicek (1989)].
3 Constructing the weights for deviations of variables from their desired trajectories
The matrices Q I and RI in (6) are usually taken as diagonal and constant in time. Their
diagonal elements Qu and Ru represent the weights of the squared deviations of output
and control variables from their desired trajectories. These weights serve for
transformation of the deviations of individual variables to equal dimension (baSic
weights), as well as for the discrimination between the relative importance of individual
variables (relative weights).
The basic weights for the output variables Yj may be derived from the condition that the
variation coefficients which express the ratios of their mean squared deviations from the
desired values (Sj) to the average desired values (y) should be equal for all variables:
where
_ 1 N-l •
Yj = -N L Ylj·
1-0
From the comparison of (6) and (18) it follows for the basic weights Q(~) :
(20) n(v)
-U
= [ItYj WVlJl2
Y.I'
j = 1,2, ... , m
99
where 'I'Iy,)j is the relative weight for variable YI in variant v. In selecting the relative
weights, the preferences of a decision maker are brought to bear: The higher (in
absolute value) the relative weight is, the higher is the importance ascribed to a given
variable and the lower will be its relative deviations from the desired trajectory in the
optimal control solution.
If the relative weights for some output variables are set near zero, then the deviations
of these variables do not affect significantly the functional (6). Therefore the desired
trajectories may be given only for variables with non-zero weights ("target" output
variables).
The basic weights for the control variables may be constructed so that the variation
coefficients of deviations from the desired values be equal for both output and control
variables. Then the total weights for the control variables R(~) are given by
(21 ) R(V)
u
= ll1
r
~
y../,v).
u.J
]2 '
j = 1, 2, ... , r,
where
wt~\ = relative weight for the variable uj in the variant v, to be given according to the
preferences of a decision maker.
In the case that only some selected control variables should be used in a given variant,
we set up the weights of the other control variables very high (prohibitive weights).
Then these variables in the optimal control simulation are not allowed to differ
significantly from their desired values, i.e. they behave like pure exogenous variables.
Prohibitive weights may be assigned also to selected target variables, which we like to
exogenize in a given variant. On the other hand, we may in fact endogenize selected
control variables by setting their weights near zero.
The list of variables of the CEM-4M model is given in Table 1. The second columm of
Table 1 indicates the partitioning of the endogenous variables (9 target variables and 23
other state variables) and exogenous variables (14 control variables and 28 purely
exogenous variables) according to the transformation of the model for optimal control
simulations.
3) Regarding the currendy available computers in Czechoslovakia. the application 01 optimal control methods to nonlinear
models is presentiy possible only to models with limited size and complexity.
100
Table 1. (continued)
Symbol Type Unit Definition
The parameters of the regression equations were estimated using annual ~time series for
the period 1967-1987.4 } The estimation methods used were ordinary least squares and
ridge regression (for the production function). The system of structural equations of the
CEM-4M model with estimated parameters is given in the Appendix (estimated standard
errors of parameters are in parentheses, estimated standard errors of residuals se are
given in percent of the average values of the dependent variables; the time index t is
omitted except of t-1 for lagged variables).
We present below a brief description of the structural equations of the model. 5}
') The authors are indebted to Milota Sujanova from the VUSEI-AR Bratislava for processing the database of the CEM-4M
model, estimation and testing of equations and for the experimental dynamic simulations of the model.
') A more detailed discussion of most equations can be found in t/1e papers on the CEM-4 and CEM-4A models [see
Sujan (1988), Sujan and Strauch 1988)].
102
Equation 1 expresses the formation of gross change in fixed capital stock Av from
investment Iv with distributed lags (according to the Koyck model). In the following
identity, this change is transformed into the fixed capital stock Kv. Equation 3
approximates the rate of embodied technical progress Q TR as a modified share of new
fixed assets and imported machinery (from nonsocialist countries) in the fixed capital
stock.
Identity 8 defines the average nominal wage in the material sphere Wv as a product of
the labour productivity and the exogenous ratio Qwv' Further identities define the
formation of labour income Ywand real disposable income YRD'
Identity 17 defines total fixed capital investment I by means of the exogenous investment
ratio QI' A part of total investment Iv is allocated to the material sphere in identity 18.
The volumes and prices of Czechoslovak exports are modelled under the assumption of
equilibrium, regarding some specific conditions. The volume of exports to nonsocialist
countries EK is determined by foreign demand in equation 24 and the corresponding
price index P EK is determined by domestic supply in equation 22. On the other hand,
the volume of exports to socialist countries Es is determined by supply function 25 and
its price index PES is derived from the price index of world imports with distributed lags
in equation 23. Both volumes and prices of exports are influenced (with different
intensity) by technical progress 0TR' In the following identities 26-28, the volumes of
exports are transformed into current prices (EKP ' Esp) and domestic constant prices (Ex).
Identities 29 and 30 define the trade balances SKP and SSP as differences between
exports and imports in current prices.
103
Identity 31 is derived from the basic macroeconomic identity between the formation and
the use of gross national income, the change in inventories J being defined as
a residual component. The last identity 32 defines the synthetic efficiency of production
factors E3 as a ratio of an adjusted sum of end-use components to a specific weighted
aggregate of fixed capital and labour employed in the material sphere.
A dynamic ex-post simulation of the CEM-4A model showed its good fit to actual data:
the mean absolute error averaged across all variables (excluding balances SKP' SSP
and J) reached 0.82 percent only (the systematic part from this error was 0.09 percent
only).
The main assumptions and results of the baseline projection are illustrated in the first
columm of Table 3. The projection reveals certain improvement over the last decade,
particularly in synthetic efficiency, personal consumption and some other indicators.
However, the speed and quality of growth according to this projection neither meet the
official targets for the economic and social development of Czechoslovakia up to the
year 2000, nor are they sufficient to diminish the lagging of the Czechoslovak economy
behind the developed marl<:et economies. The transition to a marl<:et economy itself is
a necessary, but not sufficient condition for a substantial qualitative change in the
Czechoslovak economic development. Another necessary condition for such a change is
an active development strategy adopted by the economic centre and its effective
implementation with proper economic policy instruments.
104
Table 2. Weights for deviations of target and control variables in the optimal control
simulations (1990-2000)
Variables values of A B C
nominal Intensive Extensive Combined
trajetories growth growth strategy
acceleration acceleration
Target variables:
Control variables:
Table 3. Average annual growth rates of target and control variables simulated for the
period 1990-2000 (in percent)
Target variables:
Control variables:
The purpose of the further optimal control simulations is to investigate some alternative
development strategies oriented towards increasing - the rate and the quality of economic
growth in the period 1990-2000. The target (desired) values for these simulations are
given in the second column of Table 3. In comparison with the baseline projection the
target growth rate of gross national income (the main economic indicator -used in
Czechoslovakia) is higher by about 1 percentage pOint; at the same time the growth
rates of synthetic efficiency and personal consumption are higher by 1.35 and 1.74
points, respectively, while the growth rate of investment remains unchanged. Another
target is the substantial improvement in the share of satisfied demand on the
consumption goods market. As it follows from some previous model simulations [see
$ujan and $ujanova, 1987], one of the effective strategies oriented to meet these goals
may be based on a substantial increase in the imports of machinery and consumer
goods from the West even at the costs of -increasing the foreign debt. Therefore we
have added trade balances to the target variables. The target value for the balance with
nonsocialist countries is much more negative than in the baseline projection.
Employing the CEM-4M model and the iterative procedure for optimal control of
nonlinear systems (described in preceding sections), we have simulated three alternative
development strategies according to the following scenarios: 6)
The scenario for this variant supposes a substantial increase in machinery imports from
developed market economies, which is supposed to speed up technical progress,
structural changes and other driving forces of intensive growth. This would also gradually
increase the export performance which would create real possibilities for the repayment
of foreign debt. It is obvious that for a successful implementation of this strategy the
reconstruction of the economic system towards a much greater role of the market
mechanism is necessary. Within this framework the motivation for work efforts should be
increased. Therefore. in the scenario for this simulation variant. we assume some
increase in the ratio of average wages to labour productivity (Qwv) and a considerable
increase in the suplly of industrial consumer goods, i.e. in its imports (McK' Mcs) and in
its share in total domestic output (Q x2 d. while the investment ratio 0 1would decrease.
The nominal values of all control variables were fixed at the trajectories of the baseline
projection. The deviations from the nominal values are penalized by the relative weights
given in Table 2. In the simulation variant A, greater deviations are allowed for imports
components M1K , MCK and Mcs , therefore the relative weights for these variables are
lower. On the other hand, the control variables representing extensive growth factors
(employment in the material sphere, imports of raw materials, etc.) are kept at their
nominal baseline values by prohibitively high weights. The deviations of the target
variables are penalized by equal (unity) weights. as their nominal values are consistent
with this variant.7)
.) The starting values of the state variables in 1990 (an initial condition for the optimal control procedure) were set at the
values of the baseline projection for 1990.
1) Besides the variables included in Tables 2 and 3, also some additional (less important) target and control variables
As the strategy of intensive growth acceleration according to the scenario A calls for
high foreign indebtedness, we have also simulated an alternative strategy of extensive
growth acceleration. o) In the scenario for this simulation variant we have assumed
a higher investment ratio 01' more employment in the material sphere Lv, higher imports
of raw materials (M NS ' MNK ) and higher use of intermediate product (expressed by the
dummy variable Ux). For the last three control variables higher deviations from their
nominal trajectories are allowed by their lower penalization (see Table 2). All the
"intensive" control variables employed in the simulation A (except of 0 1) are kept at
their nominal trajectories by prohibitively high weights. As this variant is not consistent
with the nominal trajectories of all target variables, the desired nominal growth of gross
national income YG is preferred, while for other target variables (especially for trade
balances) greater deviations from their nominal values are allowed by lower relative
weights.
For the practical implementation, the most feasible strategy might be some combination
of intensive and extensive growth acceleration. The scenario for such a strategy
(simulation variant C) is, in fact, a combination of the preceding scenarios A and B. In
this variant all the control variables are employed, and their relative weights (except
of 0 1) are set to twice the unprohibitive weights from the variant A or B. The weights
for the target variables are set similarly (see Table 2).
The main results of our optimal control simulations of the Czechoslovak economy for the
period 1990-2000 according to the scenarios A, B, C are given in Table 3 and Figures
1-10. All three simulated strategies lead to an acceleration of the growth rate of gross
national income Y G (in comparison with the baseline projection) by 1 percentage point,
yielding the growth rate of 3.75 percent p.a. However, they differ in the growth rates of
other target variables and used control variables.
In the simulation variant A, the trajectories of all the target variables are very close to
their nominal values. This means that in this variant a considerable acceleration in the
growth of synthetic efficiency E3 and personal consumption C would be achieved (the
latter growing as fast as national income). At the same time, equilibrium on the
consumer goods market (expressed by aCeD) would be improved. The price for this
would be a considerable increase of foreign debt: the cumulative passive trade balance
with nonsocialist countries (SKP) over the period 1991-2000 would reach more than 10
billion USD. rlowever, the baseline projection also imply a passive cumulative baiance of
2 billion USD. It should be noted that in the variant A a considerable increase in the
imports of advanced machinery and technology would imply an acceleration of technical
0) II should be noted that In this strategy only the acceleration of growth above the baseline projection is purely extensive.
while the baseline growth (2.75% p.a. for Y G' see Table 3) is predominantly intenSive, Le. this strategy is in fact mixed.
108
370
320
270
220L-L-~-L-i~L-L-~~~~--~L-~~-i~L-~~~~
98
96
94
92
185
175
165
155
145
135 ~~~~-L~~--L-~~~-L~~~L-~~~-L~~
Or-------------~~~----~~~-------------~
-5
- 10
4~~~-L~~~~L-~~~~~~~~-L~~~~~
"
4
39
38
37
35 L-~~-L~~--L-~~-L-L~-- -L _~~-L-L~--L-~
28
25
24
"
2 2 ~~~~-L~--L-~~~-L~--L-~ -L-~~~~~~
,/
5.9
6.8
5.7
5.6L-~~-L~--L-~~-L~~L-L-~-L~--L-L-~-L~~
26
24
22
progress and increase in export performance: The growth rate of exports to nonsocialist
countries would reach 5.5 percent (in the baseline projection only 3.5 percent), which
would create a real possibility for the subsequent repayment of foreign debt.
In the variant A, the "extensive" control variables are not employed (they remain at their
nominal trajectories), but the "intensive" control variables would grow faster, especially
the imports of machinery and consumer goods from nonsocialist countries (M 1K , McJ.
Also the share of the light industry in total output (QJ(2C) would be higher and the
investment ratio Q 1 would be lower. The growth rate of investment I would be almost
equal to the baseline projection.
In the variant B, the growth rate of personal consumption would be lower than in the
variant A by about 1 percentage point in favour of the growth rate of investment. As
a consequence, synthetic efficiency would grow more slowly. Due to higher raw
materials imports, this variant would also increase the cumulative passive balance with
nonsocialist countries to 4.8 billion USD. However, faster growth of production would not
be accompanied by an acceleration of technical progress, so the increase in exports to
nonsocialist countries would be very small. Only the exports to socialist countries would
increase significantly, so the cumulative balance with these countries (Ssp) would be less
passive than in the variant A (see Table 3).
In comparison with the baseline projection and variant A, the variant B would require
higher growth of raw materials imports, intermediate product,g) investment ratio and
employment in the material sphere. On the other hand, the trajectories of "intensive"
control variables would not differ from the baseline projection.
In the variant C, the growth rates of target and control variables would be between
those from the variants A and B (see Table 3 and Figures 1-10). The growth rate of
personal consumption (3.2 percent) would be somewhat higher than that of investment
and sufficient for the motivation of working efforts, taking into account growing imports
of consumer goods and improvement of equilibrium on the consumer goods market. 10)
The cumulative passive trade balance with nonsocialist countries would reach 7.8 billion
USD, i.e. considerably less than in the variant A. The growth rate of exports to these
countries would reach 4.65 percent with favourable developments of both exports
volumes and prices.
The size of foreign debt in this variant would be about 500 USD per capita in the year
2000. tt seems to be an acceptable risk as compared with contemporary net debt of
some other socialist countries (GOA about 600 USD, Poland 1 000 USD, Hungary 1 300
USD per capita, etc.). It may be expected that the reconstruction of economic system in
Czechoslovakia will create favourable conditions for effective use of foreign loans and for
their repayment.
"The share of intennedialB product in total product would decrease in this variant only by 0.2 percent p.a. while in the
baseline projection by 0.3 percent and In the variant A by 0.47 percent.
'0) As Shown In Figure 2. the share of satisfied demand aeco In this variant would be even higher than in the variant A.
except 01 the last year 01 the simulation.
114
7 Conclusion
The solution of the optimal control problem for nonlinear systems may be properly
approximated by an iterative procedure, in which the optimal control algorithm is
subsequently applied to specific linearizations of the original system. The weights for
target and control variables used in the optimization functional may be derived from the
variation coefficients of their deviations from nominal values and from the assigned
relative importance of variables in a given variant (the latter expressing preferences of
a decision maker).
We are, however, aware of the limited practical use of modelling procedures. Therefore,
the presented results of the simulations should be considered only as a demonstration
of possible applications of the described methods. Official forecasts and plans of the
Czechoslovak economic centre may be substantially different.
115
References
3.
MIK' 0 MX 1100 ) 0.1217 . (KY 1~.\-1
OTR -- 3.221 (Ay. 0.5) -0.4984 . 00.5
TA.\-l
(0.0030) (0.0121)
se = 0.36%, R2 = 0.9893, OW = 1.80
NO.2318 . (W • l )0.2427 . e Ux
E2 YAA. \-1 y
(0.0347) (0.0409)
7. YG = X - N
12. CCD = -7.165 + 0.5356 (YAD - Y AD .I-1) + 0.2354 YAD. 1 + 0.7547 CCD.l-1 -
t- 0.2717 CC.l-1
(0.0728) (0.0721) (0.1854) (0.1325)
17. 1= Y G • Q I /100
22 • PEK = 2 . 814,..,0.5241
r MK
. y'~.3531 • MO.3984. 0.0554 UpEK
G KP e
•
e
0.1 UTA (OTR -1.4)
(0.0423) (0.0838) (0.0459) (0.0059)
24. EK = 1.0530 M~416 • (PEK/PWKD )-0.7058 e0.0668 UEKL . e 1.0 UTA (OTR-1.4)
(0.024) (0.0544) (0.0093)
25. Es = 0.3595 y~.4103 • M~~601 • p~.455 . eO.0468 UESL • eO.05 UTR (OTR - 1.4)
(0.1412) (0.111) (0.158) (0.0059)
31. J = YG + Mx - (Ex + C + G + I + H)
Vladimir Vasiliev
Insti tute I'or Systems Studies (V1UISI)
USSR Academy of Sciences
9, Prospect 60 Let Octyabrya
117312 Moscow, USSR
Abstract
A new approach to the construction of macroeconomic growth
models is considered. This approach is based on an application
of ordinary differential equations to the description of the
long-term development for large economic systems. Principles of
construction of the original macroeconomic model taking into
account a non-stationary endogenous technical progress and
non-renewable energy resources depletion influence on the
economic growth are described. The analytical properties of the
model based on this approach as well as the existence and the
behaviour of the solutIons of the optimal control problem are
investigated.
Contents
Introduction
The model description
'rhe existence and asymptotic behaviour of model trajectories
4 Statement and solution of the optimal control problem
5 Properties of the optimal solution
6 ConclUSion
References
120
Introduction
Some years ago in the USSR in the field of macroeconomic
modelling a new approach was developed. It is based on applying
ordinary differentIal equations to Gescribe long-term economic
development of different countries and regions of the world
(see [1] - [ 4] ) •
Although at that time this approach was widely used in
different fIelds (see [5]-[8]), it was not applied to study the
analytical properties of models and the existence and behaviour
of optimal control problem solutions.
In this paper the above mentioned problems are investigated
for one verSlon of the long-term macroeconomic model described
in [9].
2 The model description
In paper [9] the followir~ macroeconomic model with non-
stationary endogenous technical progress was suggested for the
description of the development 01 large economic systems:
. ..
~ :::: (1-U 7 )Y ~ + U7Y(~ + U )
, K v L U
. (1 )
K :::: ~Y -ilK
U: : <£(t)~
K
rU
where Y(t) denotes gross output at time t,
K(t) denotes production capital,
L(t) denotes labour,
U(t) is the average level of technology in industry,
<£(t) is a flITlction reflecting the real dynamics of the de-
velopment of tecl1.'101ogies and possiblli ties of their
realization in in i1 ustry,
Uk (t) is the sh'?,re i~api tal investment,
(Y!'
U
- - c U if.:':. (2 )
U - 'n l{ K
. (3 )
K ::. ~y -- ~K
~(t) E [0,1 J.
f :8(t):dt
00
to
and the function Uk(t) and the parameters cn ' VI and )J. are such
that for any t)t o the following inequality is true:
'l:
f
t
a('l:)exP[f b(Od(U Jd'l: < ~o (5 )
y
o
to to
where a('l:) = cn Uk('l:)+(1-~('l:», bl.~) = (1-~(;»)(Vl+)J.) - 8(;),
and, moreover, the function ~~t) is such that
then for any initial values Yo' Ko in the semi-interval [to' (0)
;;here exists only one continuous strictly positive solution of
system (3) with the following properties:
. .
lim X1I1 qoo>O, lim Xl12 = lim ~1!1 vk>O (7 )
t ..... oo K(t) t .....oo Yl.t) t-+oo K (t )
co
Jto ,Y (t) -
1----
K(t)
~Idt < co . (8 )
~-{ - -~
-- cit
[ Xl!)
Kl t )
1 = alt)q~t) ~ O(t)q(t) (9 )
~ ~ (10)
(HC n )
(11 )
. (12 )
K = ~y - IJ.K ,
'-:>
2 Y'-
+ C",u. (13 )
u--K K
Note that both the Hamilton function (13) and the conjugate
system (14) are nonlinear in both the phase variable Y and the
control variable ~(t). Usually this situation does not occur
in the same optimal control problems using production functions.
Since the solution of thIs problem,based on the Pontrjagin
maximum principle (see [17j) is very bulky, only the statement
of the theorem about the existence of optimal control, the ex-
plicit analytical solution and the interpretation of the re-
sults will be_pre~ented in this paper. - I <
'I'heorem 2. 'Tne only control variable ~(t) satisfying the
Pontrjagin maximum principle for problem (11 )-(12) is defined
by the following expression:
r I] if t E CO,ts ]
t
*
'1, (t)=
K 1
( 15 )
1I' t E It,,,Tj
.:>
l+C n
-Oft-t.)
, Ij
'I
. ~n
e
8(t)--.=:. e ------------ ( 16 )
KO T -o('1:-t)
e 0
J---------d'1:
8('1:)
t
If the solution of equation (16) does not exist then for*the
whole interval [to,T] for the optimal control variable ~(t)
there should necessarily be fulfilled the identity ~(t) = O.
t E [t o ,t 1 ],
(17 )
t E (t 1 ,t 2 ],
t E (t2'T],
where I i and t2 are times of optimal control switches, Uo and u1
are equal to 0 or 'I depending on boundary conditions, and ~* E
(0,1). In this case output-capital ratio 'iiI1 remains constant
K (t )
on the semi-interval (t 1 ,t 2 ]. This means that describing an
economic system being in the optimal equilibrium state (if, of
course, t 1 and t 2 exis t) and at this period the ou tpu t is
optimally distributed between consumption and production
investment.
In contrast to this, the solution (15) for the problem
(11) - (12) differ,s markedly from the optimal control solution
in 1,17): initially consumption is equal to zero and the whole
output is spent on capital investment and the development of
technical progress. After that, when the economic system comes
<;0 a rather high level of development, all the means are
,'3pent on (~onsuJ.'1lption. One should underline that the time of
'.:he optimal control switch t s depends, as is seen from
,=qual':ion (16), not only on the initial conditions and the
parame1:ers (JI' the model but also on the value of the discount
rate o. One may show that time ts exists only in the case that
the ini t11al conditions and the parameters of system (12)
satisfy the following inequality:
126
t=t 1
t=r2
t2(t 1
Fig.1
Note that this property depends only on the form of system
(12) which we use for describing the economic development and
will be valid for any optimization functional linear in ~.
6 Conclusion
The results presented in this Raper show that those
trajectories of tIle macroeconomic model described above which
are the solutions of the optimal control problem have highly
127
Gyorgy Barabas
Fern Universi tat Hagen
Lehrgebiet Statistik und Okonometrie
FeithstraBe 140, D-5800 Hagen 1, Germany
Since April 15, 1991:
Rheinisch- Westfilisches Institut fiir Wirtschaftsforschung (RWI)
Hohenzollernstral3e 1/3, D-4300 Essen, Germany
Abstract
Contents
1 Introduction
2 The sample model
2.1 The model
2.2 When only the no-derivative optimization can be used
2.3 The reference-point approach
2.3.1 Defining the objectives and instruments
2.3.2 Pareto-optimality and the achievement scalarizing function
2.3.3 Decision support matrix
2.3.4 Finding the optimum by changing the reference-point
3 Case study with an econometric model of West Germany
3.1 Selecting the objectives and the instruments
3.2 The achievement scalarizing function for the case study
3.3 The base-line forecast
3.4 Decision support matrix
3.5 Finding the optimum
4 Re fe ren ces
A research grant from the Deutsche Forschungsgemeinschaft (DFG) to Prof. J. Gruber (contract No.
Gr 410/6-2) is gratefully acknowledged.
132
1 Introduction
The PC program LOS (=Large Optimizing System) has been developed at the University of Ha-
gen in order to optimize large, non-linear, deterministic econometric models. In this paper the
application of LOS to multi-criteria decision making (MCDM) problems is shown.
Making decisions is not an easy task. The objectives are usually in conflict: If you improve one
objective at least another suffers. Furthermore, there are a lot of uncertainties in the assumptions
which constitute a risk. The use of MCDM methods doesn't make the psychological aspect of the
decision making easier in general because a choice has to made among alternatives. But since the
alternatives rely on more information (the alternatives are more relevant) you can expect that the
decision you find will be better.
The decision you make in MCDM is "subjective". It may be good ("optimal", "a good com-
promise") for one person (or board, institute, community), but may be suboptimal for another
one.
We understand under the reference-point approach an iterative procedure in order to find the sub-
jective optimum in a model with multiple objectives, where the achievement scalarizing function
(named by Wierzbicki) plays an important role. This function leads to the scalarization of the
multiple objectives with very useful properties. It means further that for the reference-point ap-
proach you don't need to construct a utility function, that is a scalar-valued objective function
with economic meaning.
The decision-maker has to decide at the beginning of the reference-point approach whether he
wants to maximize, minimize or stabilize each of the single objectives. During the iterations he
always chooses the better from two alternatives. An alternative is a value combination of the
target variables, instrument variables and endogenous variables which are obtained by the solution
(minimization of the achievement scalarizing function) of the model.
The reference-point is a vector corresponding in its elements to the objectives (target variables)
and therefore may have economic meaning. In the mathematical sense the reference-point is a
parameter of the achievement scalarizing function. If you change the reference-point and minimize
the achievement scalarizing function (i.e. you make an iteration), changing the reference-point
influences the solution like changing the aspiration level or the goal value (desired value) of the
objectives.
Nonetheless the achievement scalarizing function works differently than the commonly known
quadratic objective function (see e.g. Chow, 1975) with the vector of desired values. The lat-
ter works actually like goal programming in order to find the smallest distance to the goals. On the
other side the reference-point approach respects the initial aims, i.e. maximization or minimization
of the objectives, and if possible it yields a better solution in this sense than the corresponding
reference-point elements would suggest as goal values. The task of the quadratic goal function is
equivalent to the stabilization with the reference-point approach.
There are some computer codes wich apply the reference-point approach with a better user-interface
than LOS (Bohm, Brandner, 1989). In LOS you have to manipulate manually the achievement
scalarizing function. However, other codes are specialized on small linear models, and making even
small changes in the model or the data may be difficult. The flexibility of LOS in changing the
133
goals, the restrictions and even the scalarizing function (and everything with non-linearities and
not necessarily derivable) is the main advantage of LOS.
The topic of this paper is a rather technical discussion of the implementation of the reference-point
approach, especially of the achievement scalarizing function. Therefore some remarks should be
made:
(i) The mathematical description of the reference-point approach can be found in Lewandowski,
Wierzbicki (1987) and Olbrisch (1988). However, if you are interested mainly in economic
aspects of decision making it is not important to know it in detail.
(ii) This paper shows a relatively simple procedure how you can determine your optimium by
means of the decision support matrix and by making iterations after changing the reference-
point. The flexibility of LOS allows you to apply also other, more sophisticated procedures
which converge more quickly to the optimum.
(iii) No detailed economic analyses and conclusions are made in this case study.
(iv) It might be helpful to read the User's Guide of LOS before this paper. However, if you
read subsection 2.1 of this paper and if you are familiar with some model solving code, that
subsection could be sufficient to understand the terminology of LOS in this paper.
(v) You shouldn't be confused by the different contexts of the word iteration in this paper. Firstly,
if you change the reference-point and minimize the achievement scalarizing function you
make an iteration of the reference-point approach. Secondly, LOS uses for the optimization
of a function (e.g. the achievement scalarizing function) iterative optimization procedures.
Thirdly, LOS solves an interdependent model (system of equations) by an iterative method
(Gauss-Seidel) in each iteration of its optimization procedures.
(vi) Everything written by typewriter type letters refers to computer input/output or similar
statements in LOS. Quarters are referred to as, for example, <1990q3> which means the 3rd
quarter of 1990. The notation <1987y1> means simply the year 1987 in an annual model.
In this paper the phrase 'guess' appears rather often. This means that economic modelling and
decision making remains a kind of art, where feelings, heuristics and external information play an
important role.
Since this paper deals with econometric models there are some aspects to take in account. The
decision maker is in most cases an expert or research worker and not a real decision maker. A very
important role of modelling is the learning effect of the expert by which he gets more knowledge
not only of the operation of the model but also of the real economy.
A relatively simple model with two objectives helps YOll to understand the reference-point approach
and its implementation in LOS.
134
In this subsection we introduce the sample model and give information about the terminology
and inputs of LOS. The scalar-valued objective function in this subsection is a commonly known
quadratic goal function which has nothing to do with the reference-point approach. It will be
compared to the achievement scalarizing function of the reference-point approach in subsection
2.3.2.
The sample model in this paper is the same as in the User's Guide of LOS (Barabas, 1990a).
This is not a real econometric model, but nonetheless illustrates some essential similarities: The
variables denote time series, the model is dynamic, the problem contains besides linear constraints
also a non-linear constraint. The sample model differs from the usual econometric models, since
the equations have no economic meaning, the equations form no interdependent (simultaneous)
system, the structure of the problem is not the same in all periods.
The diagrammatic layout of the variables used in the sample model is the following, where variables
beginning with the same letter can be considered as the coordinates of a two-dimensional point,
for example (Il,I2) , (G1,G2) , (Yl,Y2):
-----
1st period II-Yl Y2- 12
------ ! !
2nd period Yl Y2
------ ! !
3rd period Yl Y2
G2
Gl ~/
----: z:'----
- - - -- - - - - -
The function Z measures the distance in the third period between the points (Yl,Y2) and (Gl,G2)
by the following formula:
If this problem were only a simple simulation task, the initial values of (11,12) in the first period
and of (Gl,G2) in the third period would determine the value of Z in the third period.
But now let us define a control problem: What (11,12) values are required in order to minimize the
distance Z ?
The variables Il and 12 are the instrument (control) variables. Z as a function is the objective
135
(goal) function, Z as a variable is the objective variable to be minimized. The exogenous point
(Gl,G2) is the goal point, it consists of "desired" values. The functions Yl and Y2 define the
structural dependences of the system. In our sample model, they have the following simple form
(Yl<-I> denotes the one period lag):
Explicit restrictions are applied for the I1 and 12 variables, see also exhibit 1:
Q), [) and [) are constant restrictions: 12 ~ 1 and 12 2: -1.5 and 11 2: -1.9 ;
® and Q) are linear restrictions: 12 ~ -0.5*11+1 and 12 ~ -2*11+4 ;
Q) is a non-linear restriction: 12 2: 1l~2 - 2 (II 2: -1.9 !) .
The closed set in exhibit 1 denoted by A is the restricted set of the 11,12 variables. C is the feasible
set of the Yl,Y2 variables in the third period. It has the same form as A, but has been shifted by
(4,3). This shift comes from the functional dependences between (11,12) and (Yl,Y2), see above.
The coordinate system involves both the (11 ,12) space in the first period and the (Yl,Y2) space in
the third period.
®
-3 1 2 4 5 6 11 in the 1st period
-I A YI in the 3rd period
[)
-2
The problem we defined above is the mode! definition in the terminology of LOS and is written
into a file in the following form:
Y1 if(firstperiod(O» + then(I1) +
e1se(Y1<-1>+2)
Y2 if(firstperiod(O» + then(I2) +
e1se(Y2<-1>+1.5)
Z sqrt(sqq(Y1-G1)+sqq(Y2-G2» +
ubound(I2,1) + 1bound(I2,-1.5) + lbound(Il,-1.9) +
ubound(I2,-O.5*I1+1)+ubound(I2,-2*I1+4)
+ lbound(I2,1/(Il+2)-2) ;
136
The if(firstperiod(O)) ... then... else ... functions act upon the condition whether
the computation is running in the first or in later periods.
In the Z equation sqrt is the square root, sqq the square function. The restrictions, i.e. the upper
and lower bounds, are defined by the ubound and lbound functions.
In the model definition we did not state what we want to optimize. This information comes next
into the control file. We set the three computation periods from 1985 to 1987 by the period
command. The output, model and vari commands define the output, model and variant files.
Computation (simu command) is to be made for all periods, dynamically already from the first
period on. The format of the result-table is set by the ptable command and the tab1 file.
After the minimize command we state the variable whose value we want to minimize. This is Z
in the last period: minimize Z<last>. We choose the I1 and 12 variables as instrument variables
in the first period (1985):. .. instrument I1<first> I2<first> ... The whole control file is
as follows:
The values of the goal point (reference point) and the initial values of the instrument variables are
defined in the variant file. The goal point (G1,G2) takes the value (3,3.5) in 1987. The initial
point of (Il,12) is (3, -2.4), that is an infeasible (outside of the restrictions) value. The variant file
then is as follows:
Gl <1987yl> 3
G2 <1987yl> 3.5
Ii <1985yl> 3
I2 <1985y1> -2.4
LOS uses two types of numerical optimization methods (Barabas, 1990b) in order to compute the
maximum or minimum of a scalar-valued function (in our case this is the quadratic goal function and
the achievement scalarizing function). One method is no-derivative (actually in two modifications:
the Neider-Mead and Turn-the-worst), the other is a gradient method (steepest descent).
In many cases the gradient method is faster than no-derivative methods, but in this subsection
I show an example in our sample model when the gradient method fails to obtain the optimum.
137
This causes no problem for LOS. A no-derivative method is started automatically in LOS to check
whether the optimum is correct.
Exhibit 2 shows the situation when the gradient method fails to obtain the optimum. The quadratic
goal function (1) is now in use with the goal point of G = (8,4). The initial values of the instruments
(Ii, 12) have been set in 1987 to (2, -1.5) in order to determine an initial value of Pl=(6, 1.5) in
the space of the objective variables Yl, Y2. The procedure starts by this way from the PI initial
objective point.
Y2<1987>
4
®
°G
3
2
@
PI
-3 -2 -1 2 3 4 5 6 7 8 9
Yl<1987>
-1
-2
Exhibit 2. The gradient method fails to reach the point P3 starting from P 1
Yl <1987yl> 6.461479
Y2 <1987yl> 2.077041
which is P2 in exhibit 2 after the 54th iteration. Continuing the computation from P2 the Turn-
the-worst method stops after the 361th iteration in the point P3:
Yl <1987yl> 6.001
Y2 <1987yl> 2.999
If you work from the beginning with the NeIder-Mead method instead with the gradient method,
the solution is reached in only 195 iterations. One iteration of the optimization method means one
evaluation of the objective function (model solution).
138
No real decision problem is formulated by the sample model because the feasible solution is known
as you see in the exhibit, i.e. you could determine a good objective combination in C looking at only
the picture (exhibit 2). In real problems you don't know explicitly the feasible set and therefore an
MCDM method should be used. By the reference-point approach you scan the feasible set you are
interested in and choose a point which you accept as your optimum.
The objective and instrument variables should be chosen by economic reasoning. You should also
take into account the time consumption of the computation by LOS and reduce therefore the
problem size (the number of objective and instrument variables) if possible. The optimization of
the sample model takes only some seconds, but the optimization oflarge models often takes several
hours.
At the beginning you should define the social preferences of your objectives: Which of the objectives
to maximize, minimize or stabilize? The stabilization means that you want to achieve a particular
constant value (a desired value).
Let's define two objectives (target variables) and two instruments in the sample model. The I1
and 12 in 1985 are the instrument variables, while Yl and Y2 in 1987 the objective variables. We
want to maximize both objectives.
A point in the objective variable space is (roughly speeking) Pareto-optimal if you cannot find
another point which is better with respect to (at least) one objective without sacrificing with respect
to (at least) another objective. This means that a different point improves (some) objectives but
deteriorates at least one objective. Obviously the decision maker should be able to select the better
point. But the aim of introducing the definition of the Pareto-optimal set is to reduce the size of
the set to be investigated in a general way. Especially the case when the number of Pareto-optimal
points is very large and many of these points differ greatly in the opinion ("the preferences") of
the decision maker is of interest for applying the reference-point approach.
Looking again at exhibit 2 the points on the sections (edges) tID - © - @ are Pareto-optimal if
your social preferences are to maximize both Yl and Y2. If you want to increase for example the
first objective Yl on the Pareto-optimal set you should decrease (deteriorate) Y2 in order to remain
in the feasible set.
The aim of the decision making is to select one point among all the Pareto-optimal points which can
be accepted as a compromise solu tion ("optimal solution"). The difficulty is that you don't know
the Pareto-optimal set in an econometric decision model. Your task is to scan the Pareto-optimal
surface in a small number of steps. The small number is important because the determination and
evaluation of one Pareto-optimal point may cost a lot of computation time.
139
In this paper the main topic is the implementation of the achievement scalarizing function, which is a
central concept in the reference-point approach. This function scalarizes the vector of the objective
variables and yields a point on the Pareto-optimal set (Pareto-optimal point) if the function value
is minimized.
A simple form of the achievement scalarizing function s(w) to be minimized is the following:
(2)
yf is the i-th component of the reference-point, Yi the i-th objective variable, yf is a weighting
factor. p is the number of the objective variables. { is a small constant, e.g. 0.0001.
yf is proposed by 'Wierzbicki to be the absolute value of the mean of the available actual values
(think e.g. of time series) of Yi.
Without the term E* 2::;=1 (Wi) the minimization of s(w) could yield also weakly efficient points, i.e.
points on the section ® - ® excluding ® in exhibit 2.
More sophisticated forms of achievement scalarizing functions can be found in Wierzbicki (1986)
and Olbrisch (1988).
The next two equations should be added to the model file of LOS implementing the achievement
scalarizing function in the sample model, when Y1 and Y2 are the objectives to be maximized in
1987:
S represents the function s(w) in (2). The equation EPSTAG defines the term £ * 2::f=I(Wi). Gl and
G2 are the elements of the reference-point. The value of { is 0.0001 . The weighting factor yf is
set to 1 and therefore not represented in the equations.
How could the main difference between the achievement scalarizing function and the quadratic goal
function be represented? If you start the minimization of S by the reference-point (goal point)
G1<1987> 0
G2<1987> 0
the achievement scalarizing function respects the aim of maximizing the objectives and yields
140
Y1 <1987> 4
Y2 <1987> 4
while the quadratic function yields the smallest distance to the goal values, as:
Y1 <1987> 2.878
Y2 <1987> 2.139
Y2<1987>
S=3 S=2 S=l S=O
4
G
3
2 c
-3 -2 -1 2 3 4 5 6 7 8 9
Yl<1987>
-1
L
-2
Exhibit 3. Contours of the function S = max;=l (w;) if both objectives are to be maximized and the
reference-point G is (8,4)
Yl <1987> 6.333
Y2 <1987> 2.333
which is a slightly different solution than the solution of the quadratic goal function:
Y1 <1987> 6.001
Y2 <1987> 2.999
The contours of the achievement scalarizing function reported above consist of two half-lines like
in exhibit 3. The vertices of the half-lines form the line L through the reference-point. The angle
of L to the coordinate axes is 45 degrees if the weighting factors yf
are equal. The angle between
141
the two half-lines is 90 degrees if only the term max;=l (Wi) is used and a bit obtuse (i.e. slightly
greater than 90 degrees) if the term E * I:f=l(Wi) is added. The contours of the quadratic goal
function (not presented) are concentric circles around the goal point.
The decision support matrix helps you to make a first guess in the exploration of the region of the
objective variables you are interested in.
The decision support matrix is built up by the partial optimization of the objectives.
If you maximize Y2<1987> in the sample model, Y2 takes its maximal value anywhere on the section
® - ® (refer to exhibit 2) while Y1 changes from 2.33 to 4. For the decision support matrix only
the Pareto-optimal value of Y1= 4 is of interest.
Suppose, the maximization of Y2<1987> yields the point of (3,4). Starting a new optimization
yields necessarily a Pareto-optimal point, since now the achievement scalarizing function (2) will
be minimized involving all (now two) of the objectives. The previous result of the objective variables
(3,4) is attained as reference-point. This optimization yields the Pareto-optimal point (4,4) which
is the maximal value of Y2 at the same time. In our example the NeIder-Mead method needed
many thousands of iterations, while the Turn-the-worst method required only 476 iterations for
this calculation.
Y1<1987y1> Y2<1987yl>
maxzmlze Yl<1987yl> 6.750 1.500
maxImIze Y2<1987y1> 4.000 4.000
The vector which contains the worst element from each column of the decision support matrix is
called the nadir point. The vector which contains the best element from each column (Le. the
diagonal elements) is the utopia point. The utopia point is generally not a feasible solution. The
utopia point cannot be improved in any element while the nadir point can be improved in each
element.
The utopia point is (6.75,4) while the nadir point is (4,1.5). We know now that
- the solution cannot be better than 6.75 for Y1, and not better than 4 for Y2;
- the compromise solu tion for the vector optimizing problem will be better than 4 for Yl and
better than 1.5 for Y2. These were our first guesses about the available compromise region of the
objectives, if we didn't know it in advance from the exhibits in our artificial case.
You can choose the utopia point from the decision support matrix or another desirable value as the
first reference-point. Gl ,G2 are the elements of the reference-point, while Yl, Y2 are the objectives:
Suppose that we are not satisfied with Y2 and want to improve it by 0.5 . Change G2 by 0.5 . It
is not obvious that Y2 will also move by 0.5 but we don't have any advance information about the
way how the reference-point controls the objectives:
Suppose that 5.5 for Yi is by our preferences too small. Let Y2 deteriorate by 0.1 . We could also
improve the reference-point element for Yi which is Gl but in changing G2 we have got already
some guesses.
Suppose that we accept the last result with values for Yl,Y2 of (5.567,3.217) as our "subjec-
tive" optimum. Then the hypothetical MCDM problem formulated in our sample model has been
successfully solved by the reference-point approach as implemented in LOS.
First let us summarize briefly the econometric model used in the case study. The model is a
quarterly business cycle model of West Germany, built in the Institute for Economic Research
(RWI) in Essen 2 . It consists of more than 140 partly non-linear equations, 58 of which form an
interdependent block. The number of behavioral equations is 41.
The time horizon for the optimization in our case study is set from the 4th quarter of 1988 to the
4th quarter of 1990. These nine periods make a system of more than 1260 equations.
2 You find a more detailed description of the model in this volume: see the contribution of V. Heilemann.
143
The computations have been made on an 80386-SX PC with coprocessor, clock rate 16 MHz. LOS
issue 1.32 has been used for the computations.
The objective variables should be chosen according to economic aspects. In the case study we use
the state deficit (DEFKK) and the number of unemployed (AL) as objectives.
Among the exogenous variables of the model those variables should be selected as instruments
which can be controlled by the decision maker. In the case study we use the value-added tax
(MWST) and the public investments (IBST).
DEFKK denotes the state deficit in billions of Deutsche Mark, cumulated from 1989q1 through
1990q4. If there is a deficit, DEFKK is negative, if there is a surplus, DEFKK is positive. We want to
maximize its value in 1990q4.
The cumulation is defined in the model file of LOS by the following equation:
It means, DEFKK takes in 1989q1 the value of the deficit of the current quarter (DEF), later the
lagged value is added to the current period value.
AL denotes the number of unemployed people in thousands. We want to minimize this in each
quarter of 1989 and 1990, this makes eight objectives over the eight periods.
MWST is the value-added-tax rate (V.A.T.). We use it as an instrument in the following way:
By this definition MWST takes the first quarter value of MWSTI in all four quarters of a year,
which assumes some stability in the V.A.T. The instruments are actually MWSTI<1989ql> and
MWSTI <1990q1>.
Explicit restrictions in the model have been defined for the instrument variables since economically
unrealistic values would otherwise be obtained for the optimum. IBST can assume values between
Sand 20 billions of DM, MWST between 10 and 20 per cent. The following equation defines these
restrictions added to the model file:
The exogenous base-line data of IBST were assumed here to take values between 7 and 14 billions
of Deutsche Mark, of MWST exactly 14 per cent, like in RWI research.
The achievement scalarizing function (2) is coded and added to the model file in the following way:
where ALMAX is the term maxf~l(wi) applied for the pI = 8 periods of AL;
ALSUH is the term L:f~l(Wi) applied for the pI = 8 periods of AL;
S is the achievement scalarizing function s(w), summarizing all of the objectives;
DEFREF and ALREF are the components of the reference-point. The weighting factor of AL is 2000, the
weighting factor of DEF is 35. These seem to be representative absolute values of the corresponding
time series.
It is important to present the control file parameters of LOS used later in this section in order to
facilitate a repetition of the calculations. You find the meaning of the parameters in the User's
Guide of LOS:
The base-line forecast with the assumed values (RWl) of the exogenous variables MWST and IBST is
the following, reporting only the variables introduced in section 3.1:
You can compare later this table with the results of the reference-point approach.
The decision support matrix is gained from the partial optimization of the single objectives. Maxi-
mizing DEFKK<1990q4> yields the following result:
1989 1990
variable q1 q2 q3 q4 q1 q2 q3 q4
AL 2312.1 2186.5 2089.5 2271.4 2436.8 2325.6 2198.5 2334.9
DEFKK -2.085 19.427 25.979 37.466 27.537 42.875 46.858 56.854
IBST 5.0 5.0 5.0 5.0 5.0 5.0 5.0 5.0
KVST 20.0 20.0 20.0 20.0 20.0 20.0 20.0 20.0
RuntIme 45 mmutes.
1989 1990
variable q1 q2 q3 q4 q1 q2 q3 q4
AL 2232.2 1967.5 1748.4 1770.6 1798.3 1622.4 1444.3 1495.8
DEFKK -23.13 -19.62 -30.40 -33.77 -61.19 -62.72 -75.13 -80.97
IBST 20.0 20.0 20.0 20.0 20.0 20.0 20.0 20.0
KVST 10.0 10.0 10.0 10.0 10.0 10.0 10.0 10.0
RuntIme 15 mlllutes, but IllItIal values of the Illstruments were suggested from previous calculations.
Observe that the instruments (IBST. MWST) take in the last two tables their explicitly restricted
values, however the opposite bounds in the two tables.
Minimizing AL in a period before <1990q4> yields the same value for the periods up to the period
chosen for minimization as minimizing AL<1990q4>. This means, the minimal values of the AL-s
are time invariant. Therefore we present only one row for AL in the decision support matrix:
The utopia point consists of the first eight values minimizing AL<1990q4> and of the last column's
val ue maximizi ng DEFKK <1990q4>.
The decision support matrix shows that there is a big conflict between the optimal scenarios of
DEFKK and AL: For example, targeting a small state defici t we obtain a surplus of 56.854 billions
of DM, while targeting the low level unemployment the state deficit is 80.97 billions of DM. This
means that further calculations by an MCDM method are required to find a compromise solution.
This is the subsection where we actu;;.lly do calculations with the reference-point approach. Our
aim is a solution with a low level of the state deficit and a low number of unemployed, together
with low seasonal fluctuations of unemployment.
146
The tables contain the results of each calculation. AL and DEFKK are the objective (target) variables
while IBST and MWST are the instrument variables. ALREF and DEFREF are the reference-point
elements corresponding to AL and DEFKK. S is the value of the achievement scalarizing function.
Other variables of the RWI model are not presented in the tables.
The runtime reported is sometimes very high. However, the manual preparation of a new run for
LOS, including the analysis of the results and thinking about what to do next, takes often only
some minutes.
The long computation-time raises also the question: Do you need to optimize the model at all?
Optimization is of advantage if the number of economically different alternatives (possible scenarios)
is high. On the other side if you can restrict the number of scenarios to a small number, making only
simulations is more effective. For example, if you want to investigate only the effects of 15, 16 and
17 per cent of MWST and of an increase of IBST by 20 per cent, making simulations is much quicker
than using an MCDM method. But let's continue with our topic, the reference-point approach.
The utopia point from the decision support matrix (rounded) is used as the first-reference point
(ALREF ,DEFREF):
The unemployment is too high in all periods, near to the nadir values in the decision support ma-
trix. Let's set the reference-point component DEFREF<1990q4> to the more realistic value of zero
in the next iteration while ALREF remains unchanged. We hope that a reduced surplus reduces the
unemployment.
The unemployment hasn't decreased in a satisfactory way. In <1989q4> and <1990ql> it increases
to a high level and in the last period it also jumps up. These three reference-point elements are
reduced in the 3rd iteration by the desired differences.
147
Then we tried to reduce AL<1990ql> by 130 in the fourth iteration with the reference-point ap-
proach.
I accept this result as optimal. It seems that the seasonality of the unemployment is not avoidable
with the used instruments. The solution suggests a 15.8 per cent V.A.T. in 1989 and 19.9 per cent
in 1990. The public investments should be on a high level.
Enlarging the time horizon could be profitable in further analyses with the model, since lags in
the system could influence the optimal solution in another way through more periods. Maybe, the
remarkably low values of IBST<1990q3> and IBST<1990q4> would cause a deterioration for the
objectives in the future.
It should be also noted that a real analysis requires watching the changes of all endogenous variables,
not only of the here selected seven ones. Maybe, the deterioration of any other endogenous variable
would suggest a different compromise solution.
This concludes the brief presentation of selected results from our case study, in which we applied
the reference-point approach as implemented in LOS to the RWI business-cycle model of the West
German economy.
4 References
B6hm B., Brandner P. (1989): Interactive Optimization. Manual for the lAO-Program Package.
Institute for Advanced Studies, Vienna. Research Report No. 95.
148
Chow, G.C. (1975): Analysis and Control of Dynamic Economic Systems. John Wiley and Sons,
New York, London.
Gruber, J. (1981): Uber Finf Typen okonometrischer Entscheidungsmodelle und einige Anwen-
dungsprobleme. In: Clever, P., Hefihaus, W., Liicke, M., Mus, G. (Hrsg.): Okonomische Theorie
und wirtschaftliche Praxis. Verlag Neue Wirtschafts-Briefe, Herne, Berlin, 177-197.
Gruber, J. (ed.) (1983): Econometric Decision Models. Proceedings of a Conference at the Uni-
versity of Hagen, June 19-20, 1981. Lecture Notes in Economics and Mathematical Systems, Vol.
208, Springer-Verlag, Berlin, Heidelberg, New York.
Lewandowski, A., Wierzbicki, A. P. (1987): Theory, Software and Testing Examples for Decision
Support Systems. Working Paper WP-87-26 of the International Institute for Applied Systems
Analysis (IIASA), Laxenburg, Austria.
Bernhard B6hm
University of Technology
Argentinierstrasse 8, A-1040 Vienna, Austria
Peter Brandner
Austrian Institute of Economic Research
A-l103 Vienna, P.O.Box 91, Austria
Abstract
This paper presents an evaluation of an economic policy optimization experiment using
interactive methods of vector optimization. A medium sized nonlinear macroeconometric
model for Austria is transformed to yield a set of linear constraints for macroeconomic
policy optimization using a set of mainly fiscal and monetary instruments. To assess the
practical impact of different methods and their implementation to support macroeconomic
decision making, four techniques of vector optimization have been tested with participants
in an experimental economic planning problem for the year 1989. The techniques based on
linear programming are applied interactively and differ from each other in the information
requirements and the conduct of the interaction. We consider the Step-Method (Benayoun
et al. (1971)), Interactive Multiple Goal Programming (Spronk (1981)), a method proposed
by Zionts and Wallen ius (1976) - all contained in the optimization program lAO - together
with the reference point approach, implemented in DIDAS (developed and provided to us
by the International Institute of Applied Systems Analysis, Laxenburg). People involved
in practical economic policy decision making for Austria have been invited to participate
in the experiment. Among them were members of the chamber of commerce, the ministry
of finance, the national bank and members of the national council. A survey of the results
achieved and a critical comparative evaluation is presented.
Keywords: Interactive vector optimization; optimal economic policy making; macroeco-
nometric model.
Contents
1 Introduction
2 Economic policy making with multiple objectives
3 Interactive optimization with multiple objectives
4 The optimization problem and the econometric model constraints
5 The experimental setup for interactive economic policy optimization
6 Conclusions
References
We would like to thank M. Lupuicik and P. MillIer for discussion and J. Gruber for helpful comments to
improve the paper.
150
1 INTRODUCTION
The problem of economic policy making is to search for the joint realization of economic
objectives which in most cases are in conflict with each other. There exist numerous
attempts by economists to provide solutions to such problems by formulating some kind
of decision models (Gruber (1983), Despontin et al. (1984), Hughes Hallett (1984». By
taking a normative approach of decision making the question of how to choose values
for the decision variables (instruments) arises. Most approaches use an a priori specified
preference function to derive an optimal policy. Considering the difficulty of formulating
explicit preference or goal functions we intend to draw attention to the potential benefits
of interactive optimization techniques for decision making with econometric models. \Ve,
therefore, shall not assume knowledge of an objective function but rather show the potential
merits of four interactive procedures between the policy maker and some optimization
algorithm to reach a subjectively optimal set of target values given the economic conditions
in the form of an econometric model. Even for this approach one is confronted with a great
number of potential techniques. We shall use a particular selection and provide evidence
with respect to their qualities when applied to a medium sized econometric model.
Most applications of multiple objective decision making (MODM) methods were in-
tended to demonstrate the use of a particular algorithm (e.g. Despontin (1982». Other
studies that compare the performance of MODM procedures use fictitious problems which
have to be solved by students, staff members of research departments or managers from
industry (Wallenius (1975), Brockhoff (1985), Buchanan and Daellenbach (1987». As far
as we know, a comparative evaluation of MODM methods with a real world problem and
active policy makers has not been reported.
The paper is organised as follows: Section two gives a short review of approaches to
optimize economic policy with multiple objectives. The third section discusses shortly
the interactive optimization techniques used and their implementation. Essentials of the
nonlinear econometric model and its properties are discussed in the fourth section together
with a statement of the optimization problem the participants in the experiment were
supposed to solve. Section five describes the experimental design, presents results and a
comparative evaluation. Section six reports some conclusions.
Facing real world economic policy problems, a well known approach is to simulate an
econometric model for alternative policies which means to select some values for the in-
struments and solve for the objectives ("fixed instruments approach"). Obviously, such
simulation exercises are a poor tool for finding the best policy. By considering only a small
number of a priori specified policies, it is an inefficient method. Simulation experiments
by trial and error do not exploit all information of the econometric decision model, trade-
offs between objectives can not be properly examined. A consequence of the unsystematic
search process is the possible consideration of alternatives which are not Pareto optimal.
Additionally, no operative framework guides the decision maker to select an optimal policy.
Tinbergen (1952, 1956) adopted the opposite view in order to reach an instrument/objec-
tive combination. He suggested to choose some values for the objectives and solve for the
instruments ("fixed targets approach"). Generalizations along this line were discussed in
Russel (1975), Preston (1974, 1977), and Preston and Sieper (1977).
151
Theil (1964) introduced explicitly an objective function supposed to represent the de-
cision maker's preferences. He assumed that decision makers are able to express desired
values for objectives and instruments. Then a quadratic target function can be minimized
under the restrictions derived from a linear econometric model. Friedman (1975) analysed
piecewise quadratic preference functions within this "flexible targets" approach. Further
generalizations consider a dynamic objective function which leads to optimal control prob-
lems (Chow (1975), Pitchford and Turnovsky (1977), Holly et al. (1979), among others).
Preston and Pagan (1982) present a unified treatment of the Tinbergen/Theil approach
to the theory of economic policy. Restricted to the linear/quadratic model, the existence,
uniqueness and design of policy is analysed in a static as well as a dynamic framework.
Hughes Hallett and Rees (1983) give an integrated account of these methods of quantitative
economic planning in which decision makers play an explicit role through descriptions of
their preferences, information, and risk aversion by extending the (dynamic) optimization
framework to problems with qualitative properties.
Frisch (1961, 1976) (see also Johansen (1974)) proposed to construct an explicit pref-
erence function intended to express the preferences of the decision making authority via
interviews. Basically policy makers have to be asked three types of questions in order to
identify and to estimate parameters of the preference function and points of indifference.
The estimation of preference parameters rather than applying interview schemes has also
been considered as a tool for constructing an explicit preference function (Ancot et al.
(1982), Ancot and Hughes Hallett (1983), and others in Gruber (1983)).
'While the explicit knowledge of a preference function is not required for conducting
simulation studies, its specification is essential for the traditional optimizing framework
starting with Theil's approach. Regardless of how scalar valued preference functions (or
cost functions) have been constructed, a policy maker is no longer needed. Optimal policy
is the outcome of a mathematical optimization exercise, a decision making process does not
evolve. Such a solution to the economic policy problem may satisfy optimality conditions,
but aspects concerning the quality of the decision are not taken into account. No decision
process takes place. This is at variance with R. Frisch's (1976, p.92) ideas of economic
policy making:
"Only through such a cooperation with demonstration of alternatives will
it be possible to map out to the authorities the feasible alternatives and to
help them to understand which one - or which ones - amongst the feasible
alternatives are the most desirable from their own viewpoint. To develop a
technique of discussing feasible policy alternatives in such a scientific way is
one of the most burning needs in economic policy-making today."
It turns out that interactive methods for solving multiple objective decision models are
well suited for economic policy making as an additional tool to simulation and control
theoretical approaches (cf. Gruber (1987)).
• help the decision maker to clarify the complex structure of the system and confront him
with trade-off choices
• pay attention to psychological and sociological aspects of decision making
• assist the decision maker to find his optimal solution
• take into account learning effects.
Obviously, methods that first collect all preference information and then solve the problem
are not the relevant ones. Decision makers often decide in a stepwise and iterative manner
because they are not always able or willing to express their preferences once and for all.
By the use of interactive optimization techniques the decision process is formalized: The
decision maker operates from one solution to another, guided by local preferences. Local
preferences are those which are formed with respect to a current solution. The decision
process is an interactive collection of preference information and iterative calculation of
compromise solutions. The nature of including preference information articulated in an
interaction differs between methods. The decision maker may determine or judge trade-
offs among objectives, define aspiration levels (minimum or maximum values) for one or
more objectives, or choose a most preferred solution from a set of (efficient) solutions.
There exists a great number of available techniques to solve multiple objective deci-
sion problems (see e.g. Hwang and Masud (1979), Despontin et al. (1983), Grauer and
Wierzbicki (1984». Here we concentrate on four interactive methods which seem to us to
possess just the right qualities in order to explain the basic ideas to policy makers!. We
shall use the Step-Method (STEM) developed by Benayoun et al. (1971), interactive mul-
tiple goal programming (IMGP) by Spronk (1981), the method of Zionts and Wallenius
(1976)(ZW) and the reference point approach as implemented in DIDAS by Rogowski,
Sobczyk, and Wierzbicki (1988). As detailed descriptions of these methods are readily
available in the relevant literature we shall only present the basic ideas and point out a
few amendments which have been made in our computer application.
lObviously other methods also possess such qualities (e. g. Rosinger's algorithm as applied by Streuff and
Gruber (1983». Our selection was determined by easily available PC-software.
153
until it shrinks to a satisfactory point. Both STEM and IMGP belong to this group of
methods.
In STEM the decision maker is always confronted with a compromise solution which has
been derived from the ideal (and unfeasible) solution by applying the minimax criterion.
By comparing the Payoff-matrix of the initial stage with the compromise solution at each
iteration, the decision maker is asked which and how much of the target values may
deteriorate in order to improve at least one other objective. Using the minimax criterion
as a measure of distance, the efficiency of the compromise solution is not ensured. In the
lAO-program package a modified version of STEM has been implemented, so the decision
maker is always confronted with an efficient solution.
In the IMGP method the decision maker has to provide his preference information on the
basis of a proposed solution and a Potency-matrix which consists of a vector of ideal and
a vector of pessimistic (minimum requirement) objective values. It starts from the worst
possible point and enables the decision maker to indicate his wishes for improvements and
to state aspiration levels. At each iteration the decision maker has to indicate wether a
proposal solution is satisfactory or not and wether shifts in this solution are outweighed by
shifts in a new Potency-matrix which represents a reduced objective space. A number of
features already discussed in Spronk (1981) have been added to the basic lMGP algorithm
and are available in the lAO-program package. Based on Muller (1985), it is possible
to state aspiration levels at any iteration step. If targets are to be improved simultane-
ously the decision maker has the opportunity to state a weighting vector via improvement
qualifications. Furthermore, an efficient solution can be presented at each iteration step.
A combination of the above mentioned two principles of interactive techniques can be
found in the reference point approach as implemented in DIDAS (Rogowski, Sobczyk,
Wierzbicki (1988)). The decision maker is assumed to specify aspiration levels for all
objective outcomes (the reference point) in order to influence the selection of efficient
decisions. A parametric scalarization of the multiobjective problem is then obtained by
maximization of an order approximating achievement function where the reference point
enters as a parameter. As a result the user is confronted with an attainable efficient solution
that is uniformly as close to the aspirations as possible. Scanning several representative
efficient solutions and outcomes controlled by changing aspirations, the user should learn
enough to select either an actual (subjective) decision or an efficient solution proposed by
DIDAS as a basis for actual decisions. Further discussion concerning this approach can be
found in Olbrisch (1988).
2 PC-GIVE (Hendry (1987)) was used to carry out the estimations. A listing of model equations and
variables is available from the authors on request.
154
Thus most behavioural equations exhibit an error correction mechanism. This specification
approach captures both short run dynamics as well as long run information of the time
senes.
The basic long run economic theory underlying the model specification follows the Key-
nesian spirit but contains neoclassical elements as well. Aggregate demand is the driving
force of the economy, production is assumed to adjust. Disequilibrium unemployment may
occur in the labour market. As a main objective in model building was the explicit con-
sideration of government economic activity with a spectrum of possible instruments, em-
phasis is placed on a distinction between private and public sector variables. Components
of public consumption and public investment at current prices are treated as exogenous
instruments. Private sector behaviour is assumed to be governed by profit and utility
maximization. Prices are determined by the mark-up principle. Hourly wage formation is
based on an expectations augmented Phillips-curve with additional working time effects.
The latter are included to provide a wage compensation rule for a working time reduction
policy. Price expectations are derived from the estimated consumer price equation. Im-
port and export equations depend on relative prices and an activity variable. The small
monetary sector explains the money stock (M3) via the multiplier mechanism of money
creation transmitting the indirect effects of current and capital account on base money as
well as required reserve and savings policy of the central bank. Interest rate movements
reflect the policy of pegging the Austrian shilling to the German mark. Indirect taxation
and subsidies influence the gross price level and thus the nominal value of the income to
be distributed. Direct taxation of firms and households as well as the institution of social
insurance affect their disposable income. The public sector budget constraint determines
the changes in the stock of total public debt which in turn influences the budget via the
necessary debt service.
The policy maker has the following instruments at his disposal: government transfers
to private households (TRANSV), public investment (OIFS), the average rates of indirect
and direct wage taxation (FTIV, FLST), the level of subsidies (SUBV) and the employers'
share of social security contributions (FDBSS). These variables are in reality controlled
by the government alone or jointly with the "social partners" (i.e. firms and labour union
representatives). Further instrument variables are the legal working time per week (NAZ),
the rate of interest on savings deposits (RSP) and the (effective) required reserve ratio
(RRR). The exchange rate is derived from the DM/US$ relation and set to some level
expected by the policy maker before the optimization procedure is started. Table 1 presents
forecasts of the instrument values (made in March 1989) which were used for comparisons
with experimental solutions. Reasonable lower and upper bounds are added to the model
constraints in order to keep the solution bounded. These values could be changed in an
experimental session.
From the large number of endogenous variables we have chosen five target variables:
growth of real gross domestic product (GDP), the inflation rate of consumer prices (PC),
the unemployment rate (UR), the foreign balance of goods and services (HDLB) and the
net government deficit as a percentage of the nominal gross domestic product (DEFQ).
Real growth of GDP was to be maximized, all other targets were to be minimized (HDLB
in absolute terms). Table 2 contains forecast values for these objective variables which
served as guidelines for possible outcomes.
The set of linear model constraints is derived from the nonlinear model by using the
155
Yt = DoXt +R t
where Do is the matrix of relevant reduced form multipliers and R t contains all remaining
terms and errors. We use the values of Do given in Table 3.
TABLE 1
INSTRUMENTS LOWER FORECAST UPPER UNIT DESCRIPTION
Numbers in parentheses refer to percentage changes with respect to the previous year.
TABLE 2
OBJECTIVES FORECAST U1'iIT DESCRIPTION
Numbers in parentheses refer to percentage changes with respect to the previous year.
156
TABLE 3
VALUES OF Do (TRANSPOSED)
Since the grO\vth rates of CDP and PC are chosen as targets, whereas the respective
columns of Do in Table 3 refer to levels, the columns are transformed appropriately.
The planning horizon for the experimental problem has be~n deliberately restricted to
one period. It could easily be extended to cover more periods by using the linearized final
form of the model. However, this would have made the experimental problem much too
complicated for the majority of participants because of the increased numbers of objec-
tives. We have, therefore, kept the problem relatively simple and avoided the intertemporal
problem. It turned out that even this setup proved to be rather demanding for practition-
ers.
Together with a brief model description the forecast situation has been presented to the
participants in the experimental evaluation. They were asked to attempt a "best" solution
to the decision problem with the stated five economic objectives. A little assistance was
provided to handle the software but the de<:isions were taken without any guidance by the
operator. Each participant was presented with the same problem and only asked initially
if he agreed to the proposed set of objeccives and instnunents, and to the bounds applied.
Generally, the exchange rate was kept fix at a level of 12.5 A.S./US $. If some participants
did not agree they were free to choose their own set. The majority, however, accepted the
set-up without changes.
The order in which the techniques were applied was always the same: STEM was used
first, then followed IMGP and finally ZW. A part of the participants was further confronted
with an implementation of DIDAS by Rogowski, Sobczyk and Wierzbicki (1988) to have
an opportunity to experience a different software implementation of yet another approach
to multiobjective decision support. After experiments with all methods the participants
were asked to fill in a questionnaire.
We understand that an assessment of the methods is not independent of how they are
implemented. The questions should, therefore, reflect aspects which are relevant to the
user and also allow some discrimination among the methods.
The following criteria were chosen:
1. Ease of use of the method
2. Ease of understanding the logic of the method
3. Ease of providing information required by the method
4. Confidence in the final solution
5. Satisfaction with the implementation of the methods
6. Degree of learning-effect
7. Potential for use of interactive MCDM techniques in concrete practical work of the
participants.
Criteria 1 and 2 were assessed by ranking, 4, 5 and 6 on a point scale from 1 to 10 (with
higher numbers preferred to lower numbers). Criterion 3 was measured according to a
scale from 1 (easy) to 6 (difficult). Criteria 1, 2 and 4 are comparable to some of the
criteria used by vVallenius (1975) and Bucha..'lan and Daellenbach (1987).
The compromise solution is given by the vector of objective values (3.50, 4.78, 3.52, 3.18,
2.21). The decision maker is asked whether a deterioration of one of the objectives is
accepted in order to improve at least another one. Suppose the deficit ratio may deteriorate
by 0.8 percentage points. Then the proposed compromise solution is (3.61, 4.85, 3.22, 1.63,
3.01). Now suppose in the following step a deterioration of 0.3 in the inflation rate and of
8.0 bill. A.S. in the foreign balance is permitted, but the deficit ratio remains unchanged.
The result is (3.41, 4.62, 3.01, 3.86, 3.01). If no further deterioration is permitted a
(subjectively) optimal solution is found yielding the values of the instruments:
transv 334.30 :!lst 14.50 subv 49.67
oifs 53.00 exrc 12.50 fdbss 0.52
ftiv 16.50 rsp 2.50 rrr .055
naz 38.00 opas 185.00
Using IMGP in the next part of the session the decision maker is again confronted with
the same pay-off matrix from which the following potency matrix is generated:
POTENCY-MATRIX for PROPOSAL SOLUTION
gdp-r, ur pc-Yo Ihdlbl defq
Ideal sol. 4.1161 3.9126 1.5927 0.0000 -0.9752
Pess. sol. 1. 8158 5.3507 3.9411 23.1849 7.2410
Now the aspiration levels are set (e.g. we choose 3.4 percent for GDP growth and 3.5
percent for the deficit ratio) and wishes to improve target values are indicated. Suppose
we want to improve all objectives but we are willing to accept only a minor improvement for
the foreign balance target. The new potency matrix will reflect these additional constraints
(if a feasible solution exists) and a new iteration is entered. Further interaction, finally,
leads to
POTENCY-MATRIX for PROPOSAL SOLUTION
gdp-r, ur pc-r, Ihdlbl defq
Ideal sol. 3.4010 4.5970 2.7650 3.6820 3.4981
Pess. sol. 3.4000 4.5983 2.7660 13.9594 3.5000
The corresponding efficient solution is accepted and yields the following vectors of target
values: (3.4,4.6,2.8,3.7,3.5). The instruments are:
transv 342.61 :!lst 16.00 subv 50.00
oHs 53.00 exrc 12.50 fdbss 0.52
ftiv 16.19 rsp 2.50 rrr .055
naz 38.00 opas 185.00
We note that in this case we have achieved a lower unemployment and inflation rate than
in the previous run with STEM, though at the cost of a higher deficit ratio. The (optimal)
policy differs only in the values of three instruments (TRANSV, FLST and SUBV).
Proceeding to the third technique we can already assume that the decision maker has a
fairly good idea of the combination of target values which might be achievable. In the ZW-
method we start with the question on weights for the objective function. In case no specific
weights are given equal weights are assumed. The solution based on equal weights is in our
case: (3.76, 5.02, 3.43, 0.00, 2.91). The search for efficient nonbasis variables leads to the
presentation of the following three trade-off vectors (rows): (0.00267, -0.04872, 0.00720,
0.00000, -0.05991), (-0.00429, -0.01040, -0.05465, 0.00000, 0.09758), (0.01909, 0.07858,
0.05203, 0.00000, -0.02586). We accept the first vector and reject the third one, being
indifferent to the second one. The new solution after this choice is: (3.72, 4.92, 2.89,
159
0.00, 3.89). Two further trade-off vectors are presented. Let us assume that both are not
accepted. Then the solution remains unchanged and is considered optimal, yielding the
following instrument values:
transv 345.65 flst 14.50 subv 50.00
oifs 53.00 exrc 12.50 fdbss 0.52
ftiv 16.50 rsp 2.50 rrr .055
naz 38.00 opas 185.00
Ranking the four methods under scrutiny produced Table 4. It contains the average of
the ranks achieved (lower figures preferred). STEM and DIDAS are the easiest to use -
probably the implementation in IAC-DIDAS-L is partially responsible for this good result
- but participants had problems to understand the logic of the reference point method. It
turned out that most of the participants did not see how their result varied with variations
in the reference point. In contrast the guiding principle of STEM was the easiest to
understand and also easy to use. The term "ease of use" was often misunderstood by the
participants as "user friendly implementation".
Providing the required information (criterion 3) was generally found relatively easy,
sometimes with some need to think. Most participants did not find much difference across
methods, but for STEM and IMGP answering deemed a little easier than for the other
two methods.
Considering aspects of the implementation Table 5 gives point averages from a scale of
1 to 10 (higher figures preferred to lower figures):
TABLE 5
CRITERION 5 - AVERAGE POINTS
6 CONCLUSIONS
Utilizing all advantages of an optimization framework for actual decision making but with-
out the requirement of a preference function, interactive methods for solving multiple ob-
jective decision models are well suited for economic policy making. The policy maker is
closely involved in the solution process. Consequently, the decision quality is considerably
161
REFERENCES
Brandner P., B. Bohm, Interactive Optimization. Manual for the lAO-Program Pack-
age, Research Report No. 95, Institute for Econometrics and Operations Research,
University of Technology, Vienna (1989).
Brockhoff K., Experimental test of MODM algorithms in a modular approach, European
Journal of Operational Research 22 (1985), 159-166.
Buchanan J. T., H. G. Daellenbach, A comperative evaluation of interactive solution
methods for multiple objective decision models, European Journal of Operational Re-
search 29 (1987), 353-359.
Chow G. C., "Analysis and Control of Dynamic Economic Systems," 'Wiley, New York,
1975.
Deissenberg C., Interactive solution of multiple objective dynamic, macroeconomic sta-
bilization problems. A comparative study, in "Gruber, J. (ed.), Econometric Decision
Models," Lecture Notes in Economics and Mathematical Systems Vol. 208, Springer-
Verlag, 1983, pp. 305-309.
Despontin M., Regional multiple objective quantitative economic policy: a Belgian model,
European Journal of Operational Research 10 (1982), 82-89.
Despontin M., J. Moscarola, J. Spronk, A User-oriented listing of multiple criteria de-
cision methods, The Belgian Journal of Statistics, Computer Science and Operation
Research 23 (4) (1983).
Despontin M., P. Nijkamp, J. Spronk, "Macro-Economic Planning with Conflicting
Goals," Lecture Notes in Economics and Mathematical Systems Vol. 230, Springer-
Verlag, 1984.
Friedman B. M., "Economic Stabilization Policy: Methods in Optimization," North
Holland, 1975.
Frisch R., Numerical determination of a quadratic preference junction jor use in macro-
economic programming, Giornale Degli Economisti a Annali Di Economia (1961).
Frisch R., "Economic Planning Studies," International Studies in Economics and Eco-
nometrics Vol. 8, edited by F. Long, D. Reidel Publishing Company, Dordrecht. 1976.
Grauer M., A. P. Wierzbicki, "Interactive Decision Analysis," Lecture Notes in Eco-
nomics and Mathematical Systems Vol. 229, Springer-Verlag, 1984.
Gruber J. (ed.), "Econometric Decision Models," Lecture Notes in Economics and Math-
ematical Systems Vol. 208, Springer-Verlag, 1983.
Gruber, J., Interactive vector optimization as a complement to optimal control in econo-
metric models, chapter 5, in "C. Carraro and D. Sartore (eds.), Developments of Con-
trol Theory for Economic Analysis," Martinus Nijhoff Publishers, Dordrecht, 1987,
pp.63-82.
Hendry D. F., "The Econometrics of PC-GIVE," Version 5.0, Oxford, 1987.
Hendry D. F., A. R. Pagan, .J. D. Sargan, Dynamic specification, chapter 18, in "Z.
Griliches and M.D. Intriligator (eds.), Handbook of Econometrics Vol. II," Amster-
dam, 1986, pp. 1025-1100.
Holly S., B. Riistem, M. Zarrop (eds.), "Optimal Control for Econometric Models: An
Approach to Economic Policy Formation," Macmillan, London, 1979.
Hughes Hallett A., "Applied Decision Analysis and Economic Behaviour," Advanced
Studies in Theoretical and Applied Econometrics 3, Martinus Nijhoff Publishers, Dor-
drecht, 1984.
Hughes Hallett A., H. Rees, "Quantitative Economic Policies and Interactive Planning,"
Cambridge University Press, 1983.
163
Hwang C. 1., A. S. M. Masud, "Multiple Objective Decision Making - Methods and Ap-
plications. A State-of-the-Art Survey," Lecture Notes in Economics and Mathematical
Systems Vol. 164, Springer-Verlag, 1979.
Johansen L., Establishing preference functions for macroeconomic decision models, Eu-
ropean Economic Review 5 (1974), 41-66.
Muller P., Interactive Multiple Goal Programming (Pascal-Programm), Master Thesis,
Institute for Econometrics and Operations Research, University of Technology, Vienna
(1985).
Olbrisch, M., "Das interaktive Referenzpunktverfahren als Losungsmoglichkeit okono-
metrischer Entscheidungsmodelle - Darstellung und Einordnung der Theorie und
praktische Anwendungen," Bochumer Wirtschaftswissenschaftliche Studien Nr. 121,
Studienverlag Brockmeyer, Bochum, 1988.
Pitchford J. D., S. J. Turnovsky (eds.), "Application of Control Theory to Economic
Analysis," North Holland, Amsterdam, 1977.
Preston A. J., A dynamic generalization of Tinbergen's theory of policy, Review of Eco-
nomic Studies 41 (1) (1974),65-74.
Preston A. J., Existence, uniq'ueness and stability of linear optimal stabilisation policies,
in "J. D. Pitchford, S. J. Turnovsky (eds.), Application of Control Theory to Economic
Analysis," North Holland, Amsterdam, 1977, pp. 255-292.
Preston A. J., A. R. Pagan, "The Theory of Economic Policy," Cambridge University
Press, 1982.
Preston A. J., E. Sieper, Policy objectives and instrument requirements for a dynamic
theory of policy, in "J. D. Pitchford, S. J. Turnovsky (eds.), Application of Control
Theory to Economic Analysis," North Holland, Amsterdam, 1977, pp. 215-254.
Rogowski T., J. Sobczyk, A. P. Wierzbicki, IAC-DIDAS-L Dynamic Interactive Deci-
sion Analysis and support System Linear Version, Working Paper WP-88-110, IIASA,
Laxenburg (1988).
Russel C. S., V. K. Smith, Targets, instruments and generalized in'verses, European
Economic Review 6 (1975), 143-153.
Spronk J., "Interactive Multiple Goal Programming for Capital Budgeting and Financial
Planning," Martinus Nijhoff Publishing, Boston, 1981.
Streuff, H., J. Gruber, The interactive multiobjective optimization method by Elemer
E. Rosinger: A computer program and Aspects of Application.l, in "Gruber, J. (ed.),
Econometric Decision Models," Lecture Notes in Economics and Mathematical Sys-
tems Vol. 208, Springer-Verlag, 1983, pp. 334-364.
Theil H., "Optimal Decision Rules for Government and Industry," North Holland, 1964.
Tinbergen J., "On the Theory of Economic Policy," North Holland, 1952.
Tinbergen J., "Economic Policy, Principles and Design," North Holland, 1956.
Wallenius H., Optimizing macroeconomic policy: A review of approaches and applica-
tions, European Journal of Operations Research 10 (1982),221-228.
Wallenius J., Comparative evaluation of some interactive approaches to multicriteria
optimization, Management Science 21 (1975), 1387-1396.
Zionts S., J. Wallenius, An interactive programming method for solving the multiple
criteria problem, Management Science 22 (6) (1976), 652-663.
Zionts S., J. vVallenius, An interactive multiple linear programming method for a class of
1Lnderlying nonlinear utility junctions, Management Science 29 (5) (1£)83). 519-529.
PART 4:
Abstract
This paper surveys some recent attempts to estimate revealed preferences econometrically.
It concentrates on the underlying assumptions of these exercises such as the implied axioms
of choice; the conditions for the existence of a collective preference function; the local
nature of the estimates obtained; and the extent to which the results can be interpreted as
preferences built up from a policy search. We distinguish between policy searches with
preferences implicit and with preferences explicit - the former being appropriate for policy
selection, the latter for identifying a preference ordering. An alternative way of structuring
the preferences implicit approach (for policy selection) is to examine efficient policy
trade-Dffs.
Keywords: Optimal policy selections; axioms of choice; revealed preferences; efficient policy
trade-Dffs.
Contents
1 Introduction
2 Economic planning models and economic decisions
3 The policy search: preferences implicit and preferences explicit techniques
4 Can preferences be represented by a quadratic function?
5 Examples: parameters in the preference function
6 Preferences implicit again: policy choice and policy trade-offs
References
168
1 Introduction
The reconstruction of the preferences and priorities which are implicit in certain economic
decisions - and, by implication, the desire to be able to use numerical values for planning
priorities which are based on the observed behaviour of the decision makers rather than
imposed by econometricians - has been an enduring theme of economic research. The
reality is that specifying numerical values for the relative priorities and ideal values
assigned to target variables is an unavoidable necessity if any numerical decision values or
quantitative policy choices are to be made on a rational basis. And that holds whether
decisions are to be optimised or chosen using conventional simulation techniques.
The trouble with estimating policy reaction functions is that the necessary assumptions on
the constancy of priorities, on risk aversion, of nonsatiation, for specifying error correction
dynamics, or on the exogeneity and separability of exogenous variables simply cannot be
made in most cases. That means a numerical speCification of the objective function will be
needed in order to get explicit policy decision rules. Policy preferences may change quite
sharply over time, not least because policy makers and governments change over time and
also because, for most policy makers, policy priorities tend to be rather sensitive to the
current state of the economy and its expected development. We might characterise the
latter problem by saying that policy makers have complicated (nonlinear) preference
functions which they cannot specify fully for all eventualities, and any approximation
which they might use has partial derivatives which are very sensitive to any potential
changes in the target variables. In any event, that fact, electoral calendars, and shifts in
the government itself, will imply priorities which vary over time. As a result estimated
169
reaction functions will almost certainly suffer from parameter instability and poor
performance characteristics. The same may be said of variations in the degree of risk
aversion. Moreover nonsatiation is clearly an inappropriate assumption for many targets of
economic policy and priorities may vary depending on which side of equilibrium the
economy finds itself (as well as how far it may be from the equilibrium point). Finally any
policy feedback mechanism is likely to compromise the exogeneity assumptions
incorporated in the simple reaction function approach.
For all these reasons, it is going to be necessary to specify an explicit objective function
and compute policy decision rules from there. However this necessity for specifying
objective functions directly conflicts with the weakness of the theoretical case for using
such objective functions at all. There are problems of whether one may assume that public
or collective preferences even exist and, if they are estimated, whether they reflect social
welfare in any sense or whether they are just a device enabling policy makers to make
reasonable policy choices more easily. Much of that debate turns on what underlying
axioms of choice should be ascribed to the policy makers and what restrictions may
reasonably be placed on the functional form, structure and "freedom of choice" represented
in the objective function. This conflict explains why there has been a continuing interest in
reconstructing "actual" preference functions while, at the same time, this branch of the
literature has had rather little impact on the theory and practice of economic policy.
Over the years a number of techniques have been proposed for recovering the relative
priorities implicit in certain policy or decision values, given a model which describes the
limits of feasible economic performance. These techniques have been of increasing
generality, both in terms of the number of restriction which must be imposed on the
preference function in order for the technique to work, and in terms of the range of
problems where the technique can be used. These techniques and some typical economic
applications are discussed in van Ejck and Sandee (1959), Friedlaender (1974), Ancot et al
(1982), Streuff and Gruber (1983) and Hordijk et al (1976). These investigations have
generally restricted themselves to linear-quadratic problems; i.e. to the optimisation of
quadratic preference functions subject to linear constraints on economic performance,
which are either deterministic or certainty equivalent and which contain only one decision
maker. However further generalisations which pay greater attention to the risk and
uncertainty aspects, and which allow nonlinearities (especially inequali ty constraints) and
multiple players in a game theory setting, have been undertaken by Brandsma et al (1988),
Boehm et al (1989) and Dolado et al (1989). On the other hand, the numerical properties,
and more importantly the statistical properties, of the estimates of the relative priorities
obtained from these algorithms have not been investigated. This survey will make no
170
attempt to deal with these very recent extensions. Instead it will review the role of revealed
preferences in empirical policy problems, and the information which those revealed
preferences can provide for the investigator.
The main reason for wishing to analyse economic decisions within a formalised model
framework is that to do so ensures that all components involved in the prediction and
decision making process are made quite explicit. This will be necessary if the resulting
decisions are to be open and accountable to all parties concerned. If those decisions are not
understood, or at least capable of being criticised, by all interested parties, then it is most
unlikely that they will be able to command the support and agreement which is needed if
they are to be carried out effectively.
Hence accountability means that the assumptions, priorities, information which have been
incorporated must be open to examination and debate, and to make that possible we need
an explicit planning model. To create such a planning model is of course a more extensive
operation than just constructing an econometric model of the economy's feasible responses
to policy changes. It must in fact contain four components:
(a) A dynamic econometric model, which states the economic responses to given policy
interventions
(b) An objective function specification, which lists the target variables, their ideal values
and their relative priorities in each year (plus any restrictions on the use of the policy
instruments)
(c) The conditioning information set, which states the expected values of the exogenous
variables and possibly the actions expected from other (foreign) policy makers over
the planning period
(d) A measure of risk aversion, which states how policy makers plan to allow for the
unavoidable uncertainties of economic planning.
These components are needed to ensure a systematic and internally consistent set of
policies. Consistency requires that the planned output of different sectors (or the planned
expenditures of different categories) do not preempt all the resources so that the planned
expansion in another sector becomes impossible. It also requires that plans are consistent
over time; resources should not be used today in such a way as to make tomorrow's plans
impossible. Finally internal consistency means that a full examination of the policy options
171
must be undertaken to check which option matches society's needs best and that the
maximum scope for intervention in that direction has been exploited. Naturally that choice
must make some allowance for risk because random events, unpredictable actions of others
and imperfect models all imply substantial uncertainties. That must be taken into account
in formulating plans, at least in the form of a mechanism for revising plans as new or better
information becomes available.
We suppose that rational choices are made by planners when selecting the values for the
variables (instruments) under their control. The assumption of rational choices implies no
more than that a preference ordering exists over values of all the variables of interest to the
planners, and that this preference ordering should satisfy a minimal set of axioms of choice.
Self-interest maximising choices are made on this basis.
where R = [~ll.
. .
1· So Rtj = {OftlfJx j ~f t ~ ~_
0 otherWIse, t,] - L.T
are dynamic multipliers
RTl····RTT
evaluated numerically at some level of economic activity, x and s.
If values are rationally chosen for z then some index of economic performance can be
associated with each z value. The ability to do no more than rank z values implies the
index is an ordinal preference function for z,
w = w(z) (4)
which satisfies the minimal set of axioms consistent with rational choices. These axioms are
reflexivity and completeness to yield a properly defined set of options; and transitivity so
that a consistent preference ordering exists, plus convexity so that an optimal choice for Z
can be represented by minimising (4).
There is a second reason why we need to work with implicit preference functions. The
specification of a w(z) suitable for formulating collective decisions cannot be known a
priori. There may be uncertainty as to the existence of collective preferences, doubt as to
the precise specification of relative priorities in areas of economic performance different
from recent experience, or for larger subs paces of F, or uncertainty over the correct
173
These points are crucial in representing the behaviour of a self-managed system because
decisions are scarcely ever taken by a single dictator without the influence of advisors. At a
national level decisions are influenced by the individual interests of the government,
organised labour, the business community, and the administrators. It applies equally in a
micro-view of organisations. A government is composed of individual cabinet (or polit-
buro) members; trade unions have executive councils, and aggregate to a confederation.
Not only may these different decision makers have different variables and different
priorities in their own preferences, but they may also judge success by altogether different
cri teria. Therefore our preference structures must not restrict the criteria which different
agents might apply.
There may still be a question of how far any revealed preferences could be interpreted as
collective preferences, since there are situations where a social welfare function cannot exist
(e.g. where Arrow's "impossibility theorem" l.pplies) However their non-existence requires
certain stnct conditions to be met. These conditions have been reviewed in detail by
Johansen (1969, 1974) and Bailey (1979), and they require that the decision process must
not be the outcome of a (cooperative) game; that the dcmain of the choice set may not be
restricted; and that the choice set must be rtducible to no more than two options. By the
nature of economic planning these conditions cannot be met. For example, the fact that
prices, nomina.l interest rates, exchange rat'~s, government expenditures etc. cannot be
negative irr.plies that the domain of the choice set must, in practice, be restricted. Since
there are 5,~vcral variable values to be seleted for several time periods (where to do
nothing is cJways a possibility) there is little chance of rt~ducing the process to a series of
dichotomous options. Finally the preceding paragraph makes it clear that policies will
generally emerge from a (cooperative) bargaining game. Thu~ the nonexistence conditions
174
Consider how a single planner would actually select optimal policies as his proposal.
Suppose he takes some sequence of candidate policies z(s)fF; s = 0,1, .... His selection
process is to rank the candidates according to the Euclidean distances
IIz(s+l) - z II < IIz(s) - z II (8)
P 2 P 2
where z is any preferred IIsecond bestll but generally infeasible policy while zd remains the
p
ideal. We can allow z to be updated during these comparisons, where necessary, in order
p
175
to maintain it as preferred to all the candidates considered so far. The policy finally
selected will be the final member of the z(s) sequence, z(p) say, introduced by (8).
Within each individual planner's scale of preferences we assume that some z exists
p
preferred to z(s), and that the intersection of F and the set of z values acceptable to the
planner is nonempty. If each member of the sequence {z(s)} is the constrained optimum of
(5), corresponding to each member of the possibility set {Q(s)} for Q, then zp is any point
such that w(zp) < w(z(i)), i=O, ... ,s.
We next suppose the policy maker picks a sequence of zp vectors in order to sample the
Pareto optimal frontier of F with respect to zd. Each zp is picked as one in a series of
experiments to locate which feasible positions are preferable to each z(s); and therefore
ultimately to locate the preferred feasible position (the policy proposal) while zd remains
the ideal. We define each z to ensure a systematic search of the frontier, but otherwise the
p
switch from one zp to another (when, and by how much) is unrestricted. Our scheme now
has an inner iteration in z(s), and an outer one in z ; but that in z(s) need not be fully
p
exploited. If the inner iteration is completed we obtain the solution, for a given z ,to
P
z(p) = m!n{llz - zpl12 s.t. Hz - b} (9)
i.e.
(10)
Evidently (10) is a special case of (7) with Q = I, z = zd and q = o. Consider any point
p
on the Pareto optimal frontier of F; say z(o) generated by (7) with Q = q(o). Now consider
any other arbitrary point on that frontier, say z* = z(o) + dz*. Since z* is also feasible,
Hdz* = O. The optimisation to give z* would, if appropriate Q were known, be (7), from
which
pz d = z* - Q-1[ H'(HQ -1 H' )-1( s + HQ -1 q) - q1 (11)
Q(o) = I and zd replaced by the zp sequence (10). It is only necessary for the planner to
experiment with different z values in order to sample the entire Par·~to optimal frontier of
p
F.
On the other hand, every feasible z* can also be generated by varying Q over the set of
positive definite, symmetric, normalised matrices while zd remains fixed. One way of doing
that is the algorithm described in detail in Ancot et al (1982) and Hughes Hallett and
Ancot (1982). We therefore have two fully equivalent search proced'lres for policy
selection. One keeps preferences implicit in the z values (Le. Q is fixed, zd iterated); the
p
other makes the preferences explicit (Le. zd is fixed, Q iterated).
Hence policy selection and the "estimation" of preferences are tWG facds of the same
problem, and the same interactive planning mechanism therefore permits the analysis of
two interesting but separate problems. Given any policy selection z(p), where w(z( is
unknown, the preferences implicit in that choice (up to a quadratic approximation at z p))
may be recovered by running pseudo policy iterations from an arbitrary start to z(p),
through the "preferences explicit" algorithm. Alternatively, given any p'Jlicy (zp) preferred
to the current situation or choice (z( s )), where w(z) is imperfectly known, we can generate
an improved policy recommendation by running actual policy iterations either through the
"preferences implicit" algorithm, (10), or through the "preferences explicit", (7), ·)ne if the
preference structure is also of interest.
4.1 Homogeneity
w in (5) displays Euler's theorem for a function homogeneous of degree 2 in z "ince q=O.
The priorities over correcting the failures in z are linearly related to z itself. ':.nd to the
extent that the level of economic activity varies by scalar multiples of b - and this
contains all the uncertainty - the true preferences will select failures for e,,,:h policy
variable in constant ratios to all the other optimal failures. That is to say that, :1.long the
linear expansion path, the marginal rates of substitution of zi for 22 at tht'" optimal
solution are constant. If all the exogenous factors in b, such as world trade, imput prices,
agricultural conditions, etc., conspire to prod1lce boom conditions with, say, 2%' 'lemploy-
177
ment and 2% lllilation, and if we also accept falling short of the unemployment target by
100.000 jobs is tradeable without penalty for 1% less inflation, then the same rate of trade
is desired in a situation of 20% unemployment and 20% inflation. Clearly this is absurd.
Every government would press somewhat harder on the inflation target in the former case
and very much harder on unemployment in the latter.
This example reminds us just how restricted the validity of a particular Q and zd
specification is even where no references are made to (5) being an approximation to
something more general. If we need to consider policies significantly removed from the
economic conditions for which a particular set of preference parameters apply then the new
ones must be supplied.
The difficulty with (5) is the restricted validity of this preference function specification. A
large shock to the economy, a radical change in the political atmosphere, or just dis-
satisfaction with current recommendation may well lead policy makers to identify a
preferred trajectory. Could the original preference specification still be used? The examples
above suggest not. One result discussed by Hughes Hallett and Ancot (1982) is that for any
given zp the associated Q is unique only up to additive factors such that
There is no disadvantage in practical terms to have the estimated preferences only locally
valid: it is a special feature of particular significance in problems of social choice that the
relative penalties on the policy failures z are themselves heavily dependent on the values of
178
Other studies in this field have attempted to obtain globally valid preference orderings by
imposing prior zero restrictions and time invariance on Q. We relax the maintained
hypothesis on Q, and trade global results for local ones since in any event, no more
generality is justified. Indeed our results suggest that these prior restrictions do not
generally hold. A series of tests by Hordijk et al (1976) in which these prior restrictions are
varied when the implicit preferences of a fixed problem are calculated show that varying
the restrictions will typically produce greater changes in the results than are seen within
the preference structure itself.
(i) The diagonal elements of Q, (Qii)kk i =1...T, k=l...n+m, show the relative
penalties assigned to squared deviations of variables from their desired values in
each period. In particular we can look cross-sectionally to see how the implied
preference ranking alters over time and how the relative gaps in that ordering
increase or decrease. We may also look at the same time series behaviour of those
weights in order to rank preferences on anyone variable over the planning
exercise. These results will show which policy variables were treated as most
important at each stage; which were unimportant, whether those rankings
become inverted at (say) elections or mid-term elections compared to nominal
years, and whether policy targets become much more important in the run-up to
an election whereas policy instruments are the focus for the new administration.
(iii) Finally the remaining submatrices, Q .. i/:j, contain penalties which perform a
1J
similar function but for the joint failures at different points of time. Inter-
temporal penalties may be assigned so that accelerating penalties accrue to
persistent failures, to penalise the timing with which a unit failure occurs, or to
encourage substitutability over time. This is in addition to the basic pattern of
preferences in each time period as set up in the block diagonal components of Q.
Examples of this are, respectively, penalising units of a balance of payments
deficit proportionally more the longer they persist; penalising undershooting the
employment target proportionally more the closer to an election it occurs;
penalising consecutive consumption failures of the same sign; or rewarding a
failure in receipts and a consecutive expenditure failure of the same sign when a
balanced budget is desirable in the long run, but a deficit or surplus is
permissible in the short run.
For more detail on these points and empirical examples see Ancot et al (1982) which
examines the economic priorities underlying the New Deal policies and their manipulation
for President Roosevelt's reelection ambitions. Hughes Hallett and Ancot (1982) contrast
President Eisenhower'S concern with mid-term congressional elections with his lack of
concern for Nixon's campaign against Kennedy in 1960 and study the effect of that on US
policies of the 1956-62 period. Ancot and Hughes Hallett (1983) look at the preferences
underlying Soviet economic policies of the 1966-70 and 1971-75 five year plans and find
two strands of policy - there are "economic" variables which are used for steering the
internal economy and there are "strategic" variables which are used to reinforce the
independence of the Soviet economy, and the latter group comes to dominate in importance
by the 1970s. Ancot and Hughes Hallett (1984) estimate the preferences implicit in the
election manifestos of two Dutch political parties who subsequently formed a coalition
government, and compare those preferences with those revealed by the policies of the
subsequent coalition government. The party preferences appear to account for only about
10% of the resulting government's preferences; the remaining 90% are presumably supplied
180
by the "reality of office", the civil service, t,he central planning bureau ',!LC.
Less provocative but typical examples will be found in the analysis r;1 :\llitrian economic
priorities obtained by Boehm et al (lSg9:, and of German priorities ''I'hich are calculated
within a game between different governments (D'Jlado et al (1989)). 01'~U '~xaiilples appear
in Friedlaender (1974) on US macroeconomic policies of the Kennedy-.3 ,iL.JV'l era, and van
Ejck and Sandee (1959) who discuss estlmates of Dutch planning p;:~\Jntil;s in the 1950s.
Few other empirical examples appear to ha.ve been published.
Indeed it is not entirely clear how a policy maker would actually arrive at '/alues for his zp
vector when he does cooperate. If he experiments fully, he must inevita1:ly end up setting
those values in terms of the best trade-offs between target variables that he can get. For
example he might search for the minimum increase in the inflation rate which must be
accepted in order to get a given reduction in the unemployment rate; or the minimum
sacrifice in terms of output growth needed in order to achieve a given improvement in the
balance of trade? Starting from some z vector, the values should be changed to alternative
p
trade-off values which the policy maker would find acceptable. (That does not necessarily
mean they expect those trade-offs to be feasible.) Policy makers can make those changes
relatively easily since trade-offs are a currency they understand - the abstract penalties
specified in an artificial objective function like (5) is not a formulation which they can set
up for themselves or use themselves. In that case, why not trace out these efficient policy
frontiers (target trade-offs) explicitly and have the policy makers choose their own
preferred position on those trade-offs? That should, in principle, be an easier way of
getting policy makers to reveal their preferences (albeit implicitly) while choosing policies
for their future plans, It is also the "menu" approach recommended in Hughes Hallett and
181
Rees (1983), in which the role of policy maker and advisor are separated but linked by
two-way information flows. Advisors are expected to provide technical advice to problems
specified by the policy makers, but not to take decisions; policy makers are expected to
choose a policy from the range of projected outcomes and commentary provided by the
advisors, but not to dictate the techniques of analysis. The simplest way of doing that is to
present the policy makers with a few representative options and their projected outcomes,
and allow them to make their choice from that "menu". If those options are presented in
the form of a set of optimised trade-offs between policy targets, then policy makers can
choose optimal policies independently of any particular (arbitrary) objective function
specification that might be introduced on their behalf and it does show the scope and
limitations of policy at best (i.e. independently of any particular relative priorities).
The curious thing is that, although economists are clearly aware of the importance of
examining policy trade-offs and of reducing the arbitrariness of introducing some guessed
at objective function for the policy makers, there are so few examples in a large empirical
literature on policy choice and policy evaluation. Inflation vs. unemployment or output
trade-offs are studied by Chow and Megdal (1978), Henry et al (1982) and Wallis et al
(1987). The Chow approach may be represented by taking a special case of (5):
T[ d2 d2 d2 d2] (16)
w = ~ (Ylt-YIt) +a(Y2t-Y2t) +,8(x It -x lt ) +1(x2CX2t)
t=l
subject to the constraints in (2), where Y1 = inflation and Y2 = unemployment are targets,
and Xl = public expenditures and x2 = money supply (say) are instruments. Optimal
combinations of target values (and corresponding policy instruments) can be generated by
evaluating (7) for a whole range of values of m(O,m), keeping the rest of (16) and (2)
constant. The results can be examined by sketching the average inflation-unemployment
p
o
p
.
B
u (J
Figure la u
Figure lb
Of course we are not obliged to study average trade-{)ffs over several periods. We could
construct Figure 1 from the outcomes at t=l (a short run trade-{)ff) or at t=T (the long
run trade-{)ff, with instruments either free to change or restricted after t=l). Standard
theory would predict a steeper (possibly vertical) trade-{)ff in the long run, although that
result does not always emerge in practice. (See Henry et aI, 1982 and Wallis et aI, 1987 for
results on a variety of British models.) This kind of trade-{)ff can be studied for any set of
target variables - most obviously between inflation and output growth, which will rotate
the frontier in Figure la clockwise to an upward sloping curve.
Finally, as Wallis et al (1987) poin' out, simply varying one of the objective function
parameters in a multivariate objen:,', :unction such as (16) does not generate a genuine
trad€-{)ff in a strict efficiency sen.)! ;: ;1(' ;; confuses a movement along the trad€-{)ff curve
(which is what we get by varying '.' Wh"L :)=,=0) with shifts in the trad€-{)ff curve itself
(which is what we get by varying c xi:.en ({3, 'Y) are nonzero) since the latter increases the
target weights relative to instrurnC:'l. weights and therefore shifts the frontier of what is
feasible towards the ideal values in Figure 1. This however can be corrected by adjusting
the weights which would otherwise be left constant to bear the same relation as a set to
those on the targets which are being traded off (Hughes Hallett and Petit, 1990).
183
Another problem is that these policy trade-offs can become "reversed" from what we
would normally expect; i.e. Figure 1a rotates to an upward sloping curve. In those cases
there is no trade-off (at least in absolute terms, there may be a relative change along the
curve) since one end of the "frontier" is unambiguously better than the other. Investigating
the reasons for such trade-off reversals reveals something of the role that different
preferences or preference rankings play in policy choice; i.e. it tells us where different
priorities matter and where they do not, and why different policy makers perceive policy
problems differently, and why they have different ideas over which components,
disagreements or uncertainties, form serious constraints on successful policy design. These
are important factors in understanding how to construct policy packages which will best fit
the policy makers wishes, and they have been analysed in Hughes Hallett and Petit (1988,
1989 and 1990). Those papers show how the slope of the optimised trade-off depends on
the policy makers' preferences (Q, yd), on their expectations for exogenous events (Et(s)),
and on their perception of the economy's responses (R). One can in fact test whether
alterations in anyone of these factors would materially change that slope (in sign or
steepness) and hence identify which of the relative priorities, which expectation of the
future, or which component of the economic model is crucial to the choice of policies even
when we are not exactly sure what the preference structure should be. That also allows us
to say where disagreements or uncertainties over the correct specification of the
preferences, model or expectations would cause genuine difficulties or confusion in the
process of selecting a policy. Some empirical examples for the Italian economy are given,
and in Hughes Hallett and Petit (1989) they are extended for the case where two policy
makers (the Central Bank and the Treasury) cannot cooperate. The latter reveals that not
only does noncooperation imply inferior policy outcomes, but it may impose a policy
conflict (trade-off) which otherwise would not be present and it can also reduce the range
of policy options open to policy makers. There are therefore more reasons than just
efficiency gains for wanting to be able to specify (collecti ve) preferences correctly. If
revealed preference methods are not possible, or if they are thought to be too restrictive an
approximation for developing new policies, a study of the relevant policy trade-offs will
help in that endeavour.
References
Ancot, J.P. and A.J. Hughes Hallett (1982), "Consensus Decisions and Individual
Preferences in Economics: an Empirical Study", Greek Economic Review, 4, 3-35.
Ancot, J.P. and A.J. Hughes Hallett (1983), "The Determination of Collective Preferences
in Economic Decision Models: with an Application to Soviet Economic Policy" in J.
Gruber (ed), Econometric Decision Models, Springer Verlag, Berlin and New York.
184
Ancot, J.P. and A.J. Hughes Hallett (1984), "Establishing Public Preferences for Coalition
Government: an Empirical Study in Economic Planning Behaviour", De Economist,
132, 153-70.
Ancot, J.P., A.J. Hughes Hallett and J.H.P. Paelinck (1982), "The Determination of
Implicit Preferences: Two Possible Approaches Compared", European Economic
Review, 18, 267-89.
Bailey, M.J. (1979), "The Possibility of Rational Social Choice in an Economy", Journal of
Political Economy, 87, 37-86.
Boehm, B. and P. Brandner (1989), "Optimal Economic Policy Making with Interactive
Methods of Vector Optimisation: Practical Experiences for Austria" in J. Gruber
(ed), Econometric Decision Models: New Methods of Modelling and Applications,
Springer Verlag, Berlin and New York, 1990.
Brandsma, A.S., G. Gelauff, B. Hanzon and A. Schrijver (1988), "Retracing the
Preferences behind Macroeconomic Policy: the Dutch Experience", De Economist,
136, 468-90.
Dolado, J., M. Griffiths and J. Vinals (1989), "Preference Revelation in International
Policy Coordination" paper presented to the European Economic Association,
Augsburg, September 1989.
Chow, G.C. and S.B. Megdal (1978), "Econometric Definition of the Inflation-Unemploy-
ment Trade-off", American Economic Review, 68, 446-53.
van Ejck, C.J. and J. Sandee (1959), "Quantitative Determination of an Optimum
Economic Policy" Econometrica, 27, 1-13.
Friedlaender, A. ~1973), "Macro Policy Goals in the Post-war Period: a Study in Revealed
Preferences' , Quarterly Journal of Economics, 87, 25-43.
Henry, S.G.B., E. Karakitsos and D. Savage (1982), "On the Derivation of the "Efficient"
Phillips Curve", Manchester School, 50, 151-77.
Hordijk, L., A.P. Mastenbrock and J.H.P. Paelinck (1976), "Contributions Recontes a
l'etude Empirique de Fonctions de Preferences Collectives", Revue d' Economie
Politique, 4, 505-34.
Hughes Hallett, A.J. and J.P. Ancot (1982), "Estimatin~ Revealed Preferences in Models
of Planning Behaviour" in J.H.P. Paelinck (ed), Quantitative and Qualitative
Mathematical Economics, Nijhoff-Kluwer, Boston and Dordrecht.
Hughes Hallett, A.J. and H.J.B. Rees (1983), Quantitative Economic Policies and Inter-
active Planning, Cambridge University Press, Cambridge and New York.
Hughes Hallett, A.J. and M.L. Petit (1988), "Trade-off Reversals in Macroeconomic
Policy", Journal of Economic D'!jnamics and Control, 12, 85-9l.
Hughes Hallett, A.J. and M.L. Petit (1989), "A Note on the Costs of Uncoordinated Fiscal
and Monetary Policies" in Operations Research Proceedings 1988, Springer Verlag,
Berlin and New York, pp. 279-85 - an abridged version of "Decentralised Policies
and Efficient Trade-offs" Discussion Paper 251, Centre for Economic Policy
Research, London (July 1988).
Hughes Hallett, A.J. and M.L. Petit (1990), "Stagflation and Phillips Curve Instability in
a Model of Macroeconomic Policy", Manchester School (forthcoming).
Johansen, L. (1969), "An Examination of the Relevance of Kenneth Arrow's Possibility
Theorem for Economic Planning", Economics of Planning, 9, 5-4l.
Streuff, H. and J. Gruber (1983), The Interactive Multiobjective Optimisation Method by
Elemer E. Rosinger: a Computer Program and Aspects of Applications" in J. Gruber
(ed), Econometric Decision Models, Springer Verlag, Berlin and New York.
Wallis, K.F., P.G. Fisher, J. Longbottom, D.S. Turner and J.D. Whitley (1987), Models of
the UK Economy: a Fourth Review by the ESRC Macroeconomic Modelling Bureau,
Oxford University Press, Oxford.
Zionts, S. and J. Wallenius (1976), "An Interactive Programming Method for Solving the
Multiple Criteria Problem", Management Science, 22, 652-63.
Zionts, S. and J. Wallenius (1983), "An Interactive Multiple Linear Programming Method
for a Class of Underlying Nonlinear Utility Functions", Management Science, 29,
519-29.
Least squares estimation of quadratic preference func-
tions for econometric decision models based on survey
datal
Abstract
In an econometric decision model which is an optimization model in the usual sense, a scalar-
valued preference function is needed. In the frequently applied "linear-quadratic" control
theory approach a quadratic objective function is used. We deal mainly with estimating
such a function on the basis of survey data and with constructing the questionnaire for such
a survey.
After briefly discussing some utility theoretic aspects of estimating preference functions
in the framework of a regression model, methods are presented for generating D-optimal
questionnaire designs with hypothetical policy alternatives. Such optimal policy alterna-
tives can be used in a survey to obtain the required information about preferences. We
also report about a pilot study in which students of the University of Hagen were asked to
evaluate 28 "D-optimal alternatives" of economic policy. The method of least squares was
used to estimate five versions of preference functions of the persons who had filled in the
questionnaires.
The results obtained show that it is possible to estimate quadratic and other scalar-valued
preference functions on the basis of survey data. It is no longer justifiable to work with
hypothetical values of e.g. weights in quadratic objective functions. After having discussed
the most important results of the analysis, some lines of future research are briefly outlined.
Keywords: Econometric decision model; preference function; estimation; survey data;
multiple objectives; control theory approach; linear-quadratic model.
Contents
1 Introduction
1.1 Relevance of econometric decision models .
1.2 Contents . . . . . . . . . . . . . . . . . . .
4 Empirical analysis
4.1 Data collection
4.2 Preference functions estimated .
4.3 Results . . . . . . . .
References
1 Introduction
1.1 Relevance of econometric decision models
In many cases economic policy makers are faced with decision problems characterized by
multiple and conflicting objectives. An example for this kind of decision problem is to
stabilize and regulate the economy in a way that the following objectives are achieved
simultaneously: Unemployment is minimized, inflation is under control, economic growth
is maximized and imbalances in the foreign trade of goods and services are minimized.
Such decision problems are well known as decision problems with multiple objectives. There
is a large group of interactive procedures for solving MCDM (Multiple Criteria Decision
Making) problems (see e.g. Changkong, Haimes (1983), Zeleny (1982); some research on
MCDM in relation to econometric models is presented e.g. in Gruber (1983, 1987), Gruber,
Streuff (1983) and Streuff (1983)). One major disadvantage of MCDM procedures is that
the solution obtained is "subjective".
An econometric decision model with a scalar-valued objective function avoids this
disadvantage. The different conflicting objectives are aggregated into one evaluation cri-
terion (i.e. in a scalar-valued objective function). The evaluation of alternatives can be done
by this criterion. The criterion chosen should reflect the decision-makers preferences and
can be interpreted as a utility function. An optimal decision can be found by optimizing the
objective function (utility function) subject to restrictions derived (mainly) from an econo-
metric model.
A special case is the so-called linear-quadmtic approach: The scalar-valued objective func-
tion is a quadratic function and the restrictions are linear equations in the instrument and
target variables.
In practice this approach has been used rather early to derive optimal economic policies, es-
pecially optimal stabilization policies (Theil (1964), Fox, Sengupta, Thorbecke (1966), Chow
(1975)).
Ragnar Frisch, who together with Jan Tinbergen pioneered this approach, was convinced of
the relevance of econometric decision models as a means to optimize (macro-) economic de-
cision problems. He suggested a new type of cooperation of poli ticians and econometricians
(Frisch (1976), p. 41):
187
1.2 Contents
The purpose of this paper is to present one possible way of assessing scalar-valued preference
functions: regression analysis of preference functions on the basis of survey data. The pref-
erence information necessary for the analysis can be gained by directly interrogating those
persons (i.e. they fill in a questionnaire) whose preference function has to be estimated.
Having collected preference data there are two possibilities of estimating the regression
model: non-parametric estimation procedures (e.g. kernel estimation of regression func-
tions) and parametric estimation procedures (especially the method of least squares). Non-
parametric estimation procedures don't require a priori assumptions about the mathematical
form of the preference function to be estimated. In order to apply parametric methods as-
sumptions of that type are necessary. In this paper only results of parametric procedures
are presented.
In a pilot study Merkies and Nijman (1983) assumed a quadratic preference function. They
used the method of least squares to estimate the regression model. In this paper we are also
concerned with the method of least squares. Some theoretical difficulties are discussed and
empirical results from a small survey of students at the University of Hagen are presented
in this paper. In more detail, the paper is organized as follows:
First, the regression function showing the dependence of the utility index u on the values
of n target variables Yi (i = 1,2"", n) is presented (section 2.1). Then the random utility
model is taken as an appropriate utility theoretical basis for the analysis to follow (section
2.2). Finally, it is shown how the unknown parameters of a scalar-valued quadratic prefer-
ence function can be estimated in a linear regression model by OLS (section 2.3).
Section 3 of the paper deals with an important question concerning the development of the
questionnaire to be filled in by those persons (e.g. decision makers or others) whose pref-
erence function is to be estimated: How to formulate the hypothetical policy alternatives
(which form the rows of the data matrix) so that the utility index value for each alternative
can be stated with minimal strain on the person filling in the questionnaire and with a low
level of error? Statistical principles usually used in the experimental design literature are
applied here and adapted to answer the question stated above.
In section 4, a selection of empirical results is presented: For each of 13 persons who filled in
the questionnaire which contained 28 policy alternatives, the results of estimating 5 models
(out of a larger number of models) are gi\'en in tables and briefly discussed.
In section 5, after a short summary, some problems are indicated which could or should be
investigated next.
188
Roberts (1979) shows the conditions for the existence of functions of this type.
This utility model is called the Random Utility lv/odel: It is taken here as a theoretical basis
for handling the problem of estimating scalar-valued preference functions based on survey
data in the framework of regression analysis.
The t-th row refers to the t-th policy alternative. The total number of alternatives to be
evaluated by the actual or hypothetical decision maker is T.
Before being able to estimate the parameters of the decision-maker's preference function on
the basis of this data set the mathematical form of the regression function g(.) must be
determined.
In a pilot study Merkies, Nijman (1983) have assumed that the preference function is
quadratic in the target variables. It can be written as follows:
t=l,···,T.
This function consists of a linear and a quadratic part and a term considering interactions 4
between the target variables. After redefinition we obtain the regression model for estimating
4
aij = dy,d'u h d I . ..
dYJ' t e un er ymg questiOn IS: How does the utility index change if target variable Yi is
The linear part consists of n variables (regressors) Zti, with i = 1,2,···, n, and n being
the number of target variables. In the quadratic part there are n regressors Zti, with i =
n + 1,n + 2,··· ,2n. The interaction part contains ~n(n - 1) regressors Zti, with i = 2n +
1, 2n + 2, ... , 2n + tn( n -1). In matrix notation we write the regression model (3) as follows:
(4)
u is a T x I-vector, Z a T x K-regressor matrix, f3 a K x I-vector of regression coefficients,
and EaT x I-vector with error variables. The regression coefficients can be estimated by
the method of least squares.
It is important to point out that a utility function which is quadratic in the undesired
deviations of target variables from the corresponding desired values yti can always be trans-
formed into a quadratic function where undesired deviations no longer explicitly appear.
The following example shows this:
Y1i is the desired value of target variable i in alternative t. This simple example indicates that
the regression model ( 2) is a correct representation of a utility function which is quadratic
in the undesired deviations of the target variables.
To illustrate this approach we consider the classical linear model of normal regression (see
e.g. Schonfeld (1969)) that can be written as follows:
y=Xf3+L (5)
maxdet(ZIZ). (9)
z
1. det(X'X) as a measure of the amount of information within the design matrix X and
According to the literature available to us, until now only term 1 has been used in order
to construct an optimal design. The underlying assumption is that the variance of the
evaluation errors is independent of the design matrix X. This will often be an unrealistic
assumption.
A more realistic assumption is that the design of the questionnaire has a considerable influ-
ence upon the decision maker's evaluation errors. As a consequence we can no longer assume
a classical linear regression model because it has to be expected that the evaluation errors
are auto correlated as well as heteroscedastic.
Therefore the construction of an optimal design requires a generalized linear regression
model. It can be written as follows:
y=XP+f., (10)
where now the error variables f. are assumed to be normally distributed with covarIance
matrix
~(,( = % and with E(f.) = 0,
with 0 a positive definite TxT-matrix.
Best linear and unbiased estimates can under model assumptions be calculated by the
Generalized Least Squares Method (GLS):
(12)
Since 0 IS a positive definite and symmetric matrix the following representation for 0- 1
holds:
0- 1 = p'P.
P is a non-singular matrix. Equation (12) can therefore be written as:
Defining P X = X we obtain:
~(O""
(3,(3)
= (X X) -1 0'(.2
- I -
(14)
This formal representation corresponds to the case of the classical linear regression model.
If the matrix P were known an optimal design matrix X could be generated (section 3.2) by
maximizing the determinant of .X"' X.
However, another problem arises. The starting point of this section is the hypothesis that
the evaluation error is influenced by the design matrix X. As a consequence the matrix P
is influenced by the design matrix X. Referring to (13) and (14), the optimization problem
therefore has the following objective function:
respectively
maxdet(X'X) subject to R(X) = O. (17)
X
R( X) is an equation system of restrictions describing the subset ~k of 3{k, where L:<.< = (J'< 2 I
holds.
Forming the restrictions R(X) = 0 requires a priori information about the relationship
between the decision maker's evaluation errors and the design matrix X. For illustration
purposes an example is presented. About the decision maker's evaluation behaviour we
assume the following:
The variance of the evaluation error is the larger the more different consecutive
alternatives are.
The evaluation of alternatives can be done the more easily the fewer target va-
riables are changed from one alternative to the next.
A design taking into account an evaluation behaviour like this can be written as follows
(compare with section 2.3):
Alternative 1
(Yl.l, Y1.2, Yl.3, YI.4) = ('flg.1 Yg.2 Yg.3 Yg.4)
Alternative 2
(Y2.1, Y2.2, Y2,3, Y2.4) = (Yg.l + 8 Yg.2 - 8 Yg.3 Yg.4)
Alternative 3
(Y3.1, Y3.2, Y3,3, Y3,.1) = (Yg,l + 8 Yg.2 Yg,3 - 8 Yg,4)
Alternative 4
(Y4.h Y4,2, Y4.3, Y4,4) = (Yg.1 + 8 Yg,2 Yg.3 Yg.4 - 8)
Alternative 5
(YS,1, YS.2, YS.3, YS.4) = (Yg,l Yg.2 + 8 Yg.3 - 8 Yg.4)
Alternative 6
(Ys,], YS.2, YS,3, YS.4) = (Yg,] Yg.2 + 8 Yg.3 Yg.4 - 8)
Alternative 7
(Y7.1, Y7.2, Y7.3, Y7.4) = (Y'J.] Yg.2 Yg.3 + 8 Yg,4 - 8)
Yt.i is the value of target variable i in a.lternative t and Yg.i is a specific level 9 of the target
variable i. We start from a fixed level Yo,; of the four target variables. Then the other
alternatives are generated by changing a( the same time two target variables by 8. The
design can be constructed for different levels of the four target variables.
194
The resulting design structure is the basis for the empirical analysis presented in section 4
of this paper. The design was generated for four different levels of the four target variables
(Yg,i' 9 = 1,2,3,4, i = 1,2,3,4) so that with the 7 alternatives (given above) per level there
are 28 alternatives (number of levels x 7). The system of restrictions (denoted by R(X) in
(17)) that should guarantee this design-structure is as follows:
Besides this, lower and upper bounds of the target variables should be taken into account in
order to avoid "utopian" values of the target variables:
!!lax det(X'X)
Yg,i'O
subject to
(18 )
R(Yt,i, Yg,i, 0) = 0, Y! ~ Yt,i ~ Y't,
Yt,i, Yg,i, 0 E ~.
i = 1, 2, 3, 4, 9 = 1, 2, 3, 4 an d t = 1, 2, ... , 28.
R(Yt,i,Yg,i'O) = 0, (i = 1,2,3,4, 9 = 1,2,3,4 and t = 1,2"" ,28) represents the system of
restrictions presented earlier.
The interpretation of this optimization problem is as follows: Working with four target
variables (i = 1,2,3,4) we are looking for those four levels Yg,i (g = 1,2,3,4) and the
"change parameter" 0 which maximize the objective function subject to the restrictions
presented.
We solved this optimization problem by the Simplex method 6.
4 Empirical analysis
4.1 Data collection
The procedure described in section 3.4 has been used to generate a questionnaire containing
28 policy alternatives. Students of the University of Hagen were presented this questionnaire
and asked to evaluate the 28 alternatives of that questionnaire by stating their utility index
for each alternative. Thus, for each student who filled in the questionnaire, a 28 X 5-data
matrix resulted. On the basis of these data, mainly quadratic scalar-valued preference func-
tions were estimated (section 2.3).
We selected the following four target variables:
Yl = yearly percentage increase in the cost of living index.
Y2 = registered unemployment as a percentage of the dependent working population.
Y3 = yearly percentage increase of GNP in real terms.
Y4 = yearly percentage increase of public debt.
In order to avoid implausible or utopian values of the target variables III the policy al-
ternatives to be evaluated, lower and upper bounds were fixed.
Lower bounds: y~ = l.0 Upper bounds: y~ = 6.0
y~ = 6.0 yz = 12.0
y~ = l.0 y~ = 7.0
y~ = 3.0 y~ = 10.0 .
The resulting questionnaire containing 28 policy alternatives to be evaluated IS shown III
Appendix A.
6We used a subroutine of the NAG Fortran pe50 Library (Release 1) which is a collection of Fortran
routines for solving problems that occur in numerical computation.
196
n
Modell 1 u(Ytl,"',Ytn) = ao + LaiYti + LaiiY;i
i=1 i=l
n-l n
+2L L aijYtiYtj + Et
i=1 j=i+1
n
Modell 2 u(Ytl,··· ,Ytn) = aO + LaiYti + Et
i=l
n
Modell 3 u(YtJ, ... ,Ytn) = ao + LaiY;i + Et
i=l
n
Modell 4 log u(Ytl, ... ,Ytn) = ao + Lai log
Yti + Et
i=1
n
Modell 5 log u(Ytl, ... ,Ytn) = ao + Lai[log Yti]2 + Et
i=1
t = 1"" ,T.
Model 1 is a quadratic utility function with a linear, a quadratic and an interaction term.
Model 2 is a linear utility function. The quadratic term and the interaction term of Model
1 were omitted.
Model 3 is also a special case of model 1: the interaction term and the linear part were
dropped.
Models 1 to 3 are very well suited for practical applications in econometric decision models
with a scalar-valued objective function because they are easy to handle from a mathematical
point of view (convexity, differentiation). rvloreover models 1 and 3 obey a fundamental law
in economics: the decreasing marginal rate of substitution. Using models 1 or 3 respectively
as the objective function in a decision model with a linear (or linearized) system of restric-
tions we obtain the so-called linear-quadratic decision model.
Model 4 is a logarithmic utility model. Model 5 is well known in demand theory as the
translog utility function 7 (e.g. Jorgensen, Lau (1975), Christensen, Jorgensen, Lau (1975),
Krelle, P allaschke (1981)). The advantage of this type of function is that it is flexible to
handle and consistent with utility maximization theory. Therefore translog functions are
very well suited to test hypotheses about the decision maker's evaluation behaviour.
4.3 Results
Tables 1 to 5 (see Appendix B) summarize the estimation results for 13 students who had
filled in the questionnaire. B1 to B15 denote the estimated regression coefficients. A star
marks statistically significant coefficients (significance level 5 % ). R2 and R2C denote the
coefficient of determination and the corrected coefficient of determination respectively. MAPE
is the mean absolute percentage error, DW is the value of the Durbin-Watson test-statistic and
RHO the estimated first-order autocorrelation coefficient of the regression residuals. (Strictly
speaking, DW is applicable only to time series.)
Several specification test statistics which are not presented in the tables usually indicate the
behaviour of actual decision makers some improvements of the questionnaire should be made.
Results from psychology could be used.
C. Further empirical studies could employ optimal designs based on other models (i.e. not
on the "full" quadratic model 1 used in this paper). Moreover other hypotheses about the
decision maker's evaluation behaviour could be used.
D. An important further step would be to analyze preference structures of actual decision
makers in order to get more experience with respect to interview techniques.
£. Moreover, alternative ways of gathering information about the decision maker's prefer-
ences should be explored. Interactively interrogating the decision maker (dialogue between
the analyst and the decision maker) could perhaps improve the quality of the data. It seems
fruitful to look for possibilities to estimate a scalar-valued preference function in the frame-
work of interactive vector optimization procedures (or at least to obtain the information
required for estimating the preference function in a separate model).
F. Individual preferences and the parameter estimates derived therefrom show considerable
variability. To reduce the undesirable variability in the parameter estimates, a scalar-valued
preference function could be estimated on the basis of data about preferences of several
persons who together form a rather homogeneous group (like groups of consumers, members
of a labour union, members of a political party).
199
Alternatives Yl I Y2 I Y3 I Y4 Utility
1. 2.00 8.50 2.50 4.70
2. 1.50 9.00 2.50 4.70
3. 1.50 8.50 3.00 4.70
4. 1.50 8.50 2.50 5.20 8 = 0.5
5. 2.00 8.00 3.00 4.70
6. 2.00 8.00 2.50 5.20
7. 2.00 8.50 2.00 5.20
8. 4.00 8.80 5.40 7.70
9. 3.10 9.70 5.40 7.70
10. 3.10 8.80 6.30 7.70
11. 3.10 8.80 5.40 8.60 8 = 0.9
12. 4.00 7.90 6.30 7.70
13. 4.00 7.90 5.40 8.60
14. 4.00 8.80 4.50 8.60
1.5. 5.90 8.70 4.70 5.40
16. 4.90 9.70 4.70 5.40
17. 4.90 8.70 5.70 5.40
18. 4.90 8.70 4.70 6.40 8 = 1.0
19. 5.90 7.70 5.70 5.40
20. 5.90 7.70 4.70 6.40
21. 5.90 8.70 3.70 6.40
22. 2.00 9.00 2.20 7.00
23. l.l0 9.90 2.20 7.00
24. l.l0 9.00 3.10 7.00
25. l.l0 9.00 2.20 7.90 8 = 0.9
26. 2.00 8.10 3.10 7.00
27. 2.00 8.10 2.20 7.90
28. 2.00 9.00 1.30 7.90
B11 -0.16124 2.53081 0.15795 1. 25364 3.JOm 7.03 123* -1.26754 0.10433 -0.14477 O.mOl -1.96249 -2.62887 t 1. 71025
B12 -11.57421 t -9.60134 -7.65545* -0.m45 -4.86271 7.64 555 t 4.27082 -3.42676 6.80714 -9.62625 t 0.42036 -6.67367 t -0.26529
S
P,
B13 8.25585 t -5.06616 3.12261 0.50503 umo 7.25 654 t -1.11656 0.78190 1.42837 1.94754 -1.32963 -1.59482 -0.31046 M-
~.
B14 -1.J8820 2.55001 1.08560 0.86319 -U5065 t 6.79 l59* 1.54086 -3.31791 0.09852 -3.18HJ -0.58639 -3.15994* -4.88637* 0
B15 733.06628 1686.92615 t 15.41185 13.53481 1165. 38477 t 1697.0 ;04 * 342.16904 360.08m 810.10559 173.56508 -762.08459 t -357.45508 195.73587 =:s
~= ------ ~---
- - = = . ===-== =---= =--==--== - - - - - - - - - - - - - ---- --=---== ===---=_.
===--== I\)
R2 0.949 0.919 0.980 0.937 0.945 O. 198 0.911 '1).912 0.919 0.939 0.991 0.900 0.944 ~ 0
CD 0
R2C 0.894 0.831 0.957 0.870 0.885 O. 196 0.815 0.942 o.m 0.874 0.983 0.793 0.88' Ul
~
MAPE 10.85 51.99 6.25 2.37 12.36 1.(8 11.06 42.00 20.03 6.77 12.04 4.78 6.16 M-
SE 6.14289 0.41588 4.179(8 2.9804 6.16385 1.44 397 U469S 7.16798 9.93467 6.52491 3.95374 3.50295 6.51700 Ul
-
- - r---- ----- ----
DW 2,555 2.705 2.656 2.515 3.008 2. 651 2.W 2.m 2.469 2.305 2.357 2.50 2.390
RHO(l) -o.n -0.39 -0.39 -0.26 -0.51 -0 .36 -0.24 -0.07 -0.26 -0.20 -0.19 -0.28 -0.23
'-----'--- - - - ---- - -- --
Table 1: Estimation results, model 1_
~ Person 1 Person 2 Person 3 Person 4 Person 5 Person 6 Person 7 Person 8 Person 9 Person 10 Person 11 Person 12 Person 13
(1) - - 1------1--- --- - - - - - - -
rou 81 -11.48396' -5,27169 -4021465' -3,73847' -8,9(516' -9,9WO' 2.17975 1 9,26970* -U,65885 1 -1.70567* 18,7H71* -1.81007 -0,31263
82 -17,1181(1 -33,07628' -23,37116 1 -10,39745' -13,91488* -15,71933 1 -10.56800* -10.03779 1 -lU96~7' -17.89550 1 18,83929 1 -3,80730 -30.05232*
t;? B3 10,036361 7,51703 1 12,77713 1 2.36902* 9.19358 1 11.m73* 5.01197* 8.22744* 1.22061* 13.01224* -1.33447 1.32663 2.m22
rt
84 -8,37428* -2,17953 -2,61056* -2.24310* -9,63727* -9.66897 1 -4.11381* 3,32786' -8,23763* -0.19955 4.13803* -2,48641* -0,29241
~' ------- ---- --
rt
1-" 85 236.53377' 313.62317* 231,57234* 168,50595* 201.95284 1 250,10538* 127 .06709* 58.24237 227,75444* tn.66168' -200,68018' 93,36122' 319,0286*
g - = -=--::= ===--::= =--==
R2 0.851 o,m 0,918 0.7J6 0,800 0.532 0,859 0,909 0,565 0.854 0.918 0,291 o.s88
R2C 0,825 0,621 0.903 0.737 0.766 0,451 0,834 0,893 0,489 0.829 0.974 0.168 0,869
~
f-' -- f - - - - -
it HAPE 19.65 15.70 1U6 U8 22052 26,59 11. 76 09.05 55.43 10.33 24.09 12.59 9.H
SE 7.88623 5.30466 6,75200 4.24004 8,79823 7,75317 6.00095 9,70878 7.31U0 7,61115 1.90201 7,0204 6.92372
~ DW 20611 1.838 1. 781 1. 730 2,366 0,956 2.062 1.032 1.194 2.815 1.674 1. 200 J.131
f-'
RBO(l) -0.32 0,07 0,08 0,13 -0,19 0.51 -0,05 0,45 0,38 -0,41 0,15 0, (0 -0.08
tv
, -- I\)
o
~
,..-------.....- - - --'--- -------- ------- --------- ------ --------- -------,
~ Person Person 2 Person 3 Person 4 Person 5 erson 6 I Person 7 Person 8 1 Person 9 Person 10 Person 11 Person 12 Person 13
--------- - - - - ---------- -----~--- - ---------- ------------ --------- ----------.
~
w B1 -1.392 -0,67750 -0,43853* -0,49761* -1.11182* -1.33580* 0,32161' 1. 35641' -1.84561* -0,81005' 2.10233* -0.16654 0,03865
82 -0,964 -1.85112* -1.32085* -0,60247* -0.79732* -0,93448' -0,60024* -0,56001 * -0,83153' -0,98983' 1.01724* -0,20404 -1.70868*
t;? 83 1.014 0,91456 1 1.49818' 0,23883* 0,95154' 1.24911* 0.65497' 1.08511* 0,84878* 1.38410 1 0.17129 0,08722 0,24998
rt 84 -D.680 -0,21280 -0,23428* -0,18392* -0.76151* -0,76096* -0,31623* 0,26957' -0,71400* -0,09822 0,36593* -0,19910* -0,02417
~. --------- ----- --------- ---- --------- --------- ------- ---------_.
[1 85 1 139,093 167,22688* 137 ,87245* 116,87622* 123,21904* 62,32034'1 78,88372* 1 48,58874* 1 142,84924 123,78547* -84,75568' 68,37716* 190,69485*
:.::::.===== ==::~=== ======- =========- ::: : =::::::=-==== =========
8 ========= =======:
R2 0,8 0,615 0,911 0,808 0,769 0,540 0,857 0,9081
I 0.665 0.195 0,970 0,276 0,883
R2C 0.7 0.618 0.903 0,775 0,128 0.460 0,832 0,892 0.601 0.759 0,965 0,150 0,863
[ - - - - --------- ------- . --------- ------- ------ --------
rt MAPE 22 11.23 14,16 4.61 22.21 26,981 13,53 11.71 32.68 12,61 9.45
til 11.22 I 48.36
S8 8,861 5,35888 6,16402 3.92282 9,47118 7,59962 6.04620 9,77105 5.18646 9.03016 5.63526 7,10138 7,07842
------ - . ------- ------ ----- --------
D1I 2, 1.838 1.446 1.852 U48 1.080 2.183 1.3061 1.195 2.111 1.554 1.238 2,125
~ RBO(l) -0 0,07 0,26 0,06 -0,23 0,45 -0,12 0.31 0,22 -0,22 0,20 0.37 -0,09
w ----- ------ --------- - ------ --------- - - - - - -------
Person 1 Person 2 Person 3 Person 4 Person 5 Person 6 Person 7 Person 8 Person 9 Person 10 Person 11 Person 12 Person 13
~
(D r----- r---" ----
ol>o B1 -1.07350* -0.58836* -0.26023* -0.15374* -0.7988(t -0.29360 0.18022 0.38315 -0.53186 -0.38577* 1.90704 * -0.18380 -0.05639
B2 -4.63981* -13.56075 t -4.95938* -1.52517 t -4.10690 t -2.29270* -2.81166* -2.40559 -3.20886 -2.82105* 6.82309* -0.87957 -4.40775*
~ B3 1.16957* 1.37059* 0.99791* 0.11337 1.30865* 0.61289 t 0.65347* 1.09573* 0.28612 0.81372* -0.46764 0.17230 0.16178
rt -0.93424t -0.35021* -0.00019
B4 -1.14935* -1.82873 t -0.41094* -0.17903* -1.81013* -0.91868* 1.25412 -0.92605 0.00843 2.64784*
~. 1------1 - - - - - - -1------- ---- -- - - - - - --- - - ----- 1-------- -------
rt 6,270B1* 13.49514*
1-" 85 15.18314* 34.21077 t 14,21085- 7.73m* 14.663B1t 10.22827 t 10.23695 t 4.77979 12.28225 t 9.44245* -17.64310*
g F~-== F=--== - - = - - -=:=:===:=--
R2 0.848 0.854 0.872 0.690 0.867 o.m 0.881 0.573 0.244 0.865 0.802 0.337 0.867
R2C 0.822 0.829 0.850 0.636 0.844 0.382 0.860 0.499 0.113 0.842 0.767 o.m 0.844
[ ~- 1 - - - - --- 0 - - - -1 - - - - - - --- ----- ----- r------ --------
rt MAPE 5.U 13.44 U8 1.49 5.57 6.76 4.02 10.60 15.58 2.56 9.30 3.19 2.37
til
BE 0.23516 0.48376 0.20460 0.08308 0.23887 0.33340 0.18494 0.72609 0.65691 0.12909 0.60691 0.16627 0.12720
1----- - - - - ------ ---- -- ---- -----
~ DW 2.052 1.866 1.969 1.317 1.074 0.659 2.112 1.750 0.734 2.780 1.755 1.2481 2.408
f-' RHO(l) -0.04 0.01 -0.02 0.33 0.45 0.66 -0.09 0.12 0.63 -0.50 0.11 0.37 -0.20
~--- ----- ------- - - - - - --------- ----- - - - - - - - - - - -- - - - - - --------- ---------
.*'" ----- ------ I\)
0
I\)
- --- -------- - - -
Person 1 Person 2 Person 3 Person 4 Person 5 Person 6 Person 7 Person 8 Person 9 Person 10 Person 11 Person 12 Person 13
~ -----1-- 1--- -- ---- -----
V1 B1 -0.55912* -0.34927* -0.16415* -0.09356* -0.45269* -0.23837 0.03801 0.27724 -0.55471 * -0.19873* 0.97318 t -0.05629 -0.01669
B2 -0.99241* -3.13052* -1.15029* -0.34884t -0.91763* -0.55042 *I -0. 69556 t -0.57235 -0.79224* -0.62653* 1. 44599* -0.17479 -1.01822*
~ B3 0.51463* 0.60557* 0.44815* 0.06132* 0.58272* 0.33270 0.31216 t 0.36887 0.38464* 0.3U2Bt -0.19107 0.04010 0.05558
rt
B4 -0.39239* -0.56452* -0.16147* -0.05919* -0.56747* -0.29103 t I -0.28007* 0.32884 -0.33983 t -0.04031 0.77509 t -0.10260* -0.00441
~. - - - - 1------ ------- ------ - - - - ---- - - - - - - - - ------- ------ -----
rt
1-" 85 9,25214* 18.63669* 9.09001* 5.94793* 9.10787* 7.35668 7.06024 t l 4.18916 8.38718*1 6.nl79*1 -7.05972*1 4.93148*1 8.81540*
g --- --- == F==== ====--= f=-=== ===== ---=--==Io====*---=====:t========:t======
@ R2 0.895 0.854 0.889 0.777 0.903 0.561 0.857 1 0.567 0.842 0.271 0.867
o.m j0.859
R2C 0.876 o.m 0.870 0.738 0.886 0.485 0.832 0.492 0.404 0.834 0.814 0.145 0.843
~f-' ------- ------ -------- -------- 1 - - - - - - ----- ------- ------ -------- --------
fir MAPE 4.60 13.85 4.07 1.16 UB 5.86 4.58 1 10.05 11.61 2.37 7.44 3.36 2.46
BE 0.19592 D.48432 0.19019 0.07050 0.20430 0.30443 0.20284 0.73074 0.53841 0.13214 0.54245 0.17434 0.12726
- - - - - - - - 1------ - --- - - - ----- - - - ------
DW 1.887 2.136 2.106 1.820 2.054 1.094 2.235 1 1.870 2.594 2.106 1.206 2.352
~
f-' 0. 972 1
RHO(l 0.02 -0.21 -0.08 0.08 -0.19 0.44 -0.18 0.06 0.50 -0.31 -0.08 0.39 -0.19
. L - - - - - 1______ ~ _____ _
V1 - - - ------- ' - - - - - ----- '------ ----
.
203
References
Ash, A., Hedayat, A. (1978)
An introduction to design optimality with an overview of the literature.
Commun. Statist. A7 14, 1295-1325.
Bariow,R.E.,Bartholomew,D.J.,Bremner,J.M.,Brunk,H.D. (1972)
Statistical inference under order restrictions. John Wiley, New York.
Changkong, V., Haimes, Y. Y. (1983)
Multiobjective decision making, theory and methodology. North-Holland,
New York, Amsterdam, Oxford.
Chow, G. C. (1975)
Analysis and control of dynamic economic systems. John Wiley, New
York.
Christensen, L. R., Jorgensen, D. W., Lau, L. J. (1975)
Transcendental logarithmic utility functions. American Economic Review
65, 367-383.
Fox, K. A., Sengupta, J. K., Thorbecke, E. (1966)
The theory of quantitative economic policy with applications to economic
growth, stabilization and planning. North-Holland, Amsterdam, London.
Friedman, J., Tibshirani, R. (1984)
The monotone smoothing of scatterplots. Technometrics 26, 243-250.
Frisch, R. (1976)
Cooperation between politicians and econometricians on the formali-
zation of political preferences. In: Long, F. (ed.), Economic Planning
Studies. D. Reidel Publishing Company, Dordrecht, 41-86.
Gruber, J. (1979)
Approaches to determining the weights in the objective function of eco-
nometric decision models. Discussion Paper No. 35, Department of Eco-
nomics, University of Hagen.
Gruber, J. (1981)
Uber funf Typen okonometrischer Entscheidungsmodelle und einige An-
wendungsprobleme. In: Clever, P., Hesshaus, W., Lucke, M. und Mus,
C. (Hrsg.), Okonomische Theorie und wirtschaftliche Praxis. Verlag Neue
Wirtschafts- Briefe, Herne, Berlin, 177-197.
Gruber, J., ed. (1983)
Econometric decision models. Lecture Notes in Economics and Mathe-
matical Systems 208, Springer-Verlag, Berlin, Heidelberg.
Gruber, J. (1987)
Interactive vector optimization as a complement to optimal control in
econometric models. In: Carraro. C., Sartore, D. (eds.), Developments
of control theory for economic analysis. Kluwer Academic Publishers,
Dordrecht, Boston, Lancaster.
Gruber, J., Leserer, M. (1982)
Prognose- und Entscheidungsmodelle. Fernstudienmaterialien cler Fern-
universitat - Cesamthochschule - Hagen, Kurs Nr. 0862.
Husges, H. (1986)
Methoden zur Schatzung von Zielfunktionen fiir okonometrische Ent-
scheidungsmodelle. Diskussionsbeitrag Nr. 113, FernUniversitat - Ce-
samthochschule - Hagen.
204
Hiisges, H. (1990)
Regressionsschatzung skalarwertiger PraJerenzfunktionen fiir okonome-
trische Entscheidungsmodelle auf der Grundlage von Befragungsdaten.
Dissertation am Fachbereich Wirtschaftswissenschaft der Fern TJ niversitiit
- Gesamthochschule - Hagen (Publication planned in 1991).
Jorgensen, D. W., Lau, L. J. (1975)
The structure of consumer preferences. Annals of Economic and Social
Measurement 4, 49-101.
Krelle, W., Pallaschke, D. (1981)
A general demand system. Zeitschrift fiir Nationalokonomie 41,223-252.
Merkies, A. H. Q. M., Nijrnan, TH. E. (1983)
The measurement of quadratic preference functions with small samples.
In: Gruber,J.,ed., Econometric Decision Models. Lecture Notes in Eco-
nomics and Mathematical Systems 208, Springer Verlag, Heidelberg a.e.,
242-263.
Merkies, A. H. Q. M. (1989)
The Impact of a Projection of a Multidimensional Preference Function
Into an Operational Macroeconometric Model. Discussion paper presen-
ted to the 2nd international conference on econometric decision models
in Hagen.
Prakasa Rao, B. L. S. (1983)
Nonparametric functional estimation. Academic Press, Orlando.
Roberts, F. S. (1979)
Measurement theory. Addison Wesley, New York.
Rosinger, E. E. (1981)
Interactive algorithm for multiobjective optimization. Journal of Opti-
mization Theory and Applications 3.5, 339 - 365.
Schonfeld, P. (1969)
Methoden der Okonometrie. Band 1. Verlag Franz Vahlen, Berlin, Frank-
furt a.M.
Silvey, S. D. (1980)
Optimal design, an introduction to the theory for parameter estimation.
University Press, Cambridge, London, New York.
Streuff, H. (1983)
Ein interaktives Verfahren der Vektoroptimierung angewandt auf
okonometrische Entscheidungsmodelle: der Rosinger-Algorithmus: zur
Theorie, seine programmtechnische Realisation, Aspekte seiner Anwen-
dung. Haag und Herchen, Frankfurt/~tain.
Streuff, G., Gruber, J. (1983)
The interactive multiobjective optimization method by Elemer E. Rosin-
ger: A computer program and aspects of applications. In: Gruber,J.,ed.,
Econometric Decision Models. Lecture Notes in Economics and Mathe-
matical Systems 208, Springer Verlag, Heidelberg a.e., 242-263.
Theil, H. (1964)
Optimal decision rules for government and industry. North-Holland, Am-
sterdam.
Zeleny, M. (1982)
t-.Iultiple criteria decision making. :'lcGraw-Hill, New York.
OPERATIONALIZING A MACRO-ECONOMIC PREFERENCE FUNCTION
Abstract
Contents
1 Introduction
5 The Implications
6 Summary
206
I INTRODUCTION
2.1 The specification of the preference function and the sources of informa-
tion collected
max W(y,x)
(2.1)
s.t. G(y,x,z) - 0,
where y, x and z are vectors of targets, instruments (controls) and data re-
spectively. The projection of W(y,x) in targets alone is written as U(y) and
can be approximated by
where the constant ao' the row vector a' with n "linear" parameters a i and A,
the positive definite matrix of "quadratic" parameters a i j , may depend upon
the prevailing situation y •.
To estimate the parameters of (2.2) we need observations. We have for each
207
U - Y/3 + £, (2.3)
The introduction of the matrix A without any restriction would imply too many
parameters to be estimated, viz. K-(n2 +3n)/2 «n2+n)/2 "quadratic" parameters
a ij and n "linear" ones at). In view of the reluctance of the enquirees to
answer questionnaires with much more than 25 programs the estimation is not
possible unless restrictions are introduced. Therefore the matrix A is re-
stricted to be diagonal. Hence K is taken as 2n. The vector /3 is estimated as
(2.4)
(2.5)
with Pi some weight. The diagonality of the matrix V reflects the plausible
assumption that each policy program is evaluated independently. The weights
express the idea that if a program is considered "odd" or its probability Pi
to occur low no serious attempt is made to quantify Ui ' leading to a large
variance of £i. This assumption on V will be discussed in more detail in
section 2.2.
(2.7)
a + Ayo + v - 0, (2.8)
R/3+v-O, (2.9)
208
where the first row of R equals R1-[l,1,0, ... ,O,y~ ,0, ... ,0] and so on and fi
is a column vector of 2n+l elements, being a constant, n "linear" parameters
a i and n "quadratic" ones a i i .
Equation (2.9) is added as a restriction to (2.3). Note that R is an
nx(2n+l)-matrix so that (2.9) puts n constraints on the estimation of the
parameters. The estimation problem is solved by combining (2.3) and (2.9)
into
(2.10)
(2.11)
i f A"'O, (2.12)
and by
i f A=O, (2.13)
with a3 the variance of the variable Yj over the past, expressing the idea
that the more volatile a variable Yj has been in the past, the more difficult
it will be to specify its optimal value y~.
In the analysis given above it is tacitly assumed that the global preference
209
(2.15)
o - (~+qr)S , (2.17)
At the end of 1983 we held an inquiry among political parties, more specifi-
cally among their representations in the "Tweede Kamer der Staten-Generaal"
(Parliament). Those parties who provided answers that could be processed are
210
mentioned in table 1 together with the number of seats they held in that
chamber.
Number of seats
CDA Christian Democrats 45
VVD Liberals 36
PvdA Labour Party 47
D'66 Democrats 6
GPV Reformed Political Association 1
The ques tions concerned a set of 25 "policy programs" (N-25). Each program
refers only to the values of five policy variables (n-5). A set provides two
sources of information; firstly the evaluation Ui j of party j on each of the 25
programs and secondly the content of the optimal program of party j. We col-
lected information on two different sets of policy programs. A and B.
The 25 possible scenarios for 1987 of these variables are given in table 2.
The information asked for is indicated by dots. The first source of informa-
tion concerns the evaluations of the programs in the column indicated by U,
the second source is the specification of the optimal program in the first
row. Hence this row (= yO) differed among the parties.
1) The measurement of the "actual" situation is not unambiguous. It depends upon the time of
measurement. It remains always possible that more recent information gives corrections on
earlier estimates. Moreover. measurement errors always remain.
211
(2.16) we specified T=20, n=5 and r=3 (so ignoring that no quadratic term is
included in 3 equations).
i U Yl Yz Y3 Y4 Ys pl)
1 100
2 16 -1 7 13 2 100
3 12 2 8 9 2 78
4 16 2 7 14 2 99
5 16 -5 7 14 2 100
6 16 -2 9 14 2 100
7 16 -2 4 14 2 100
8 16 -2 7 14 -1 95
9 16 -2 7 14 5 94
10 12 -2 7 14 2 62
11 18 -2 7 14 2 98
12 16 -2 7 9 2 66
13 16 -2 7 17 2 82
14 14 -2 7 9 2 81
15 16 -2 8 9 2 64
16 12 1 5 8 2 61
17 18 -3 7 9 3 44
18 17 0 4 9 4 70
19 20 1 3 10 5 58
20 16 2 3 13 4 94
21 16 3 3 10 4 87
22 16 2 2 10 4 84
23 16 2 3 10 0 77
24 10 2 3 10 4 16
25 16 2 3 7 4 42
The first value was found in the previous studies to be a reasonable compro-
mise. It expresses the view that the degree of uncertainty of the future is
ten times as great as suggested by the variance of our regression. The para-
meter T, which we will refer to as the "horizon of the enquiree" , is given a
value of one to indicate that the enquirees considered these scenarios
implicitly as to refer to 1984 (1 year ahead from the time of the inquiry).
In other words, the horizon of the evaluators was taken to be one year and
not 4 years as suggested by the inquiry (see table 2). The probabilities Pi
were calculated, according to these rules, prior to the inquiry. Thus we
could compile a set of scenarios of different degree of attainability. For
instance, program 24 was thought to be so unrealistic that its probability to
appear (or rather the probability that this program or a more odd vector in
the same Cartesian direction would materialize in 1985) was thought to be
212
1. The probabilities of table 2 are calculated with (2.16); they are there-
fore of a mechanical character. They were needed prior to the inquiry to
assure sufficient variety in the set of scenarios.
2. Parties' views on the reality of the scenarios were also asked for. They
were quite different, see table 3. Only the Democrats' views were posi-
tively correlated (r=.20) with the mechanically obtained probabilities.
3. Ex post comparison between the mechanical and the parties' views revealed
that changing the "degree of uncertainty of the future" }1 or the
"horizon" T would not have improved the correlation, see table 3.
4. It is of interest to see how the parties looked upon the reality of their
own preferred plans, see table 4.
5. The real situations for the years 1984-1987 have been added for compari-
son. We have:
Y1 Y2 Y3 Y4 Y5
1984 17 2 3 9 5
1985 16 2 3 7 5
1986 15 3 0 6 3
1987 14 1 0 8 2
T~l T=2
CDA VVD PvdA D'66
}1=10 1'=15 1'=10 1'-15
---
CDA 1.00 0.80 0.68 0.33 -0.46 -0.40 -0.61 -0.57
VVD 0.80 1.00 0.78 0.13 -0.54 -0.52 -0.62 -0.63
PvdA 0.68 0.78 l.00 0.00 -0.36 -0.31 -0.41 -0.38
D'66 0.33 0.13 0.00 l.00 0.20 0.25 -0.07 0.02
213
Programs yO Probabilities
Y1 Yz Ya Y4 Y5 own T= 1 T- 1
1) No information given.
No one was near to the realized view on 1985. Probabilities were computed as
before, a quadratic term was selected only for variable Yl. Again we speci-
fied T-20, n=5 and r=3 (ignoring zero parameters of quadratic terms). Again
we chose J.I-lO and r=l.
T=l T=2
CDA WD PvdA D'66 J.I=lO J.I=15 J.I-lO J.I=15
---
CDA 1.00 0.15 0.21 0.15 -0.46 -0.48 -0.53 -0.59
VVD 0.15 1. 00 0.83 -0.36 -0.39 -0.34 -0.53 -0.52
PvdA 0.21 0.83 1. 00 -0.13 -0.37 -0.35 -0.47 -0.47
D'66 0.15 -0.36 -0.13 l.00 0.51 0.49 0.42 0.43
The optimal vectors of the parties and their probabilities are given in table
6.
programs yO probabilities
1) No information given.
CDA -11.14 -19.93 0.11 -4.55 -8.10 -3.25 -1. 99 -0.01 -2.26 -2.02 -1. 29
VVD 1. 66 -2.43 4.98 -1.94 -10.28 1. 22 -0.10 -1.00 -0.97 -0.79 0.30
PvdA 20.84 -4.90 8.84 0.03 0.03 -7.90 -0.80 -1.47 -1. 29 -5.84 -3.94
D'66 -6.83 -10.68 2.04 0.40 -0.97 -1.85 -0.41 -0.20 0.10 -0.08 -0.31
• 0 • 0 '0 • 0 • 0
Computed optima Yl Yz Y3 Y4 Ys U
CDA 10.05 -3.15 -0.00 -1. 72 -5.71 18.62 -0.39 0.19 -0.41 -0.95 -3.10
VVD -0.04 -1. 52 -0.00 -1. 76 -7.26 2.75 -0.07 0.97 -0.10 -0.45 -1. 38
PvdA 0.0 -112.86 -10.0 -5.0 -0.0 -8.57 -28.21 5.00 -1. 25 -0.0 1.43
D'66 -0.03 -1. 95 0.04 -0.74 0.03 -0.00 -0.08 0.02 -0.36 0.01 0.04
-0 -0 - 0 -0 -0
Computed optima Y1 Y3 Ys Y7 Ya U
The most striking feature of the above result is that the computed optima y
hardly differ from the revealed optima Yo in tables 4 and 6. This is not
surprising because the parties' own weights are in general rather close to
zero. This implies that such observations are more or less ignored in the
computations. With so many weights close to zero the content of the informa-
tion source 1 shrinks away leaving the conclusions to depend only on informa-
tion of source 2. This is also indicated by the U-values of the various
optima which are much closer to 100 than those with mechanical weights. It is
not surprising that the measured preference functions become rather
unreliable as they depend only upon a very mutilated set of observations. The
slight differences between computed and revealed optima must be handled with
prudence. We have verified whether our conclusions are depending upon the
weights. Table 9 shows for the variables of set A for each party and for
each variable whether the computed optimum was above (+) or below (-) the
revealed optimum if the mechanical weights were used. If the same feature was
found with the parties' own probabilities an asterisk (*) was attached to the
indicated sign.
Table 9. Estimated optima higher (+) or lower (-) than revealed optima;
set A; modelweights*
Y1 yz Y3 Y4 Y5
unemployment growth inflation govt. balance of
finance payments
CDA +*
VVD + +
PvdA +* +*
D'66
GPV + +
The hidden wish for higher growth rates and lower unemployment appears to be
independent of the set of weights.
For the second set of variables (set B) sufficient information is lacking to
construct a table.
Whether preferences are stable through time can be verified with the help of
the variables of set B. In Merkies and Nijman (1983) preference functions on
the same set of variables are discussed. They referred to 1977, the same
procedure was followed and optima were computed in the same way. The results
are reported in Merkies and Hofkes (1986). The remarkable point is that much
lower unemployment figures were strived for in 1977. A saying often heard is:
"Politics is the art to attain what is possible". The differences between
computed and revealed optima vanish against differences over time. But this
mainly concerns unemployment. Actual possibilities in the economy seem to be
more important than preferences.
5 THE IMPLICATIONS
These two community preference functions can be handled in the same way as
before. The optimal values for the two sets are:
The indifference curves of the two quadratic preference functions are ellips-
oids. Their projections for U-95 and U-99 (or 95% resp. 99% of the optimal
value) are drawn in figure 1 (for set A) and figure 2 (for set B). The opti-
mal points of table 10 are represented in the figures by a +sign and referred
to by point A.
Comparison of both figures shows that the projection depends upon the varia-
bles in the preference function that are suppressed. In both figures the
symmetry axes of the ellipses are parallel to the main axes of the diagram.
This is not coincidental but follows from the diagonality of the matrix A in
(2.2). This was assumed to save degrees of freedom. Figure 1 and figure 2
suggest that it may not have been the most plausible assumption.
Apart from the indifference curves we have represented the possibilities in
the figures. The point estimate of 1984, made with model (2.15), i.e. 14%
unemployment and 6,5% inflation, is represented in both figures with a black
box and referred to by point B. We computed a 95% probability interval around
this point in the way we also computed our probabilities in section 3. These
intervals are ellipsoids in five dimensions. The projection is represented in
the figures 1 and 2. As the variables suppressed in both preference functions
are not the same the ellipses of the figures 1 and 2 differ. They are
projections from two different five-dimensional spaces. Note that these
projections depend upon the values adopted for the suppressed variables.
If the weighted preference functions (5.1) and (5.2) reflect indeed the pref-
erences of the society and the possibility-ellipses are indeed true represen-
tations of what the economy allows, an optimal policy in the line of Theil
(1964) can be constructed. The solution of the maximization of the preference
function (5.1) and (5.2) subject to the constraints described by our possi-
bility ellipses leaves one degree of freedom: the relation between ends and
means. For given means the possibilities-ellipse is fixed and the optimal
policy is obtained by varying the indifference ellipse such that the two
ellipses are tangent. The same solution can be obtained with a given
indifference curve and a varying possibilities-ellipse. The smaller we take
this given indifference curve, the closer the optimal point will be to the
society's optimum of table 10. It is clear that for different given indif-
ference curves the optimal points will be on a curve that goes through both
points A and B. A little geometry will indicate that this curve is only a
straight line if both ellipses degenerate into circles. But although in gen-
eral a variety of curves can appear, a glance at the figures shows that in
our example the path between A and B is not far from a straight line.
We have selected the two variables inflation and unemployment to throw some
light on the significance of the Phillips-curve. Although the original form
launched by Phillips (1958) connected money-wages with unemployment, the
literature on this subject, see e.g. Desai (1984) often selects other repre-
sentations of the concepts discussed. Whatever the actual choice of the
variables, the underlying significance remains somewhat obscure. Few authors
attempt to describe the true nature of the relation. Does it describe demand
or does it describe supply? Is it the locus of clearing points of the labour
market or is it the result of a disequilibrium solution? Whatever the true
nature, we would like to know what the scatter diagram originally studied by
Phillips reflects. Is it allowed, as for instance done by Reuber (1964), to
218
interpret the graph through the scatter by the society's indifference curve?
To clarify the point we assume that usually the optimal policy, as given by
Theil's method, is more or less attained. This assumption is not as bold as
it seems. Authors who try to deduce community's preferences from revealed
preference theory, see e.g. Friedl~nder (1973) do in fact the same.
In our framework this assumption would imply that, if neither preferences nor
the economies' possibilities change, the scatter of historical points would
lie around the line-segment through A and B. The scatter of actual points in
figure 1 and figure 2 shows that this is not the case. The points may be on a
remote contour around A but they cannot be on a line-segment through A and B,
unless these points are also shifted.
We have no definite answers yet on the questions raised above. There are,
however, various features that are worth to draw attention to:
6 SUMMARY
The comparison of our results with earlier attempts supports the view
mentioned above that preferences shift with actual opportunities of the
economy. If this is true it affects the interpretation of certain relations
in economic theory, e.g. the Phillips-curve. The empirical relation between
price increases and unemployment ratios can no longer be interpreted as an
indifference-curve as e.g. done by Reuber (1964). The relation may indicate
the opportunities of the economy, but our sample period (1964-1983) is
probably too short to verify such a statement.
REFERENCES
Desai, M. (1984), "Wages, Prices and Unemployment a Quarter Century after the
Phillips Curve" in: Hendry, D.F. and K.F. Wallis (eds.), "Econometrics
and Quantitative Economics", Basic Blackwell, Oxford, pp.342.
FriedHl.nder, A.F. (1973), "Macro Policy Goals in the Postwar Period: a Study
in Revealed Preference", Quarterly Journal of Economics, 87, pp.25-43.
Merkies, A.H.Q.M. and T.E. Nijman (1983), "The Measurement of Quadratic Pref-
erence Functions with Small Samples" in: J. Gruber (ed.), "Econometric
Decision Models", (Proceedings, Hagen, West Germany, 1981), Lecture Notes
208, Springer-Verlag, Berlin, pp.364.
220
Merkies, A.H.Q.M. en A.J. Vermaat (1981), "De Onmacht van een Kabinet. Een
Empirisch Onderzoek naar Sociaal-Economisch Preferentiefuncties en hun
Gebruik als Welvaartsindicator" , Maandschrift Economie, 45, pp.101-118.
Phillips, A.W. (1958), "The Relation between Unemployment and the Rate of
change of Money Wage Rates in the United Kingdom, 1861-1957", Economica,
25, pp.283-299.
'75
0
a '7'b
a
'7)
0
'72
0 '7fa
'7'0
a
cD
'RI)
'6'b 0
0 '69 '77
0
.D 0
'66
0
'70 '78
0 '798
'65
~
0
'"6
'68 0
'sF
~
N
a
0i------r----~-----,r-----r_----~----_r----_,--,L_,r_--__,
o 2 6 8 18
UNmpLOYI1ENTRATE
221
a
N
'75
0
a '74
a 0
'73
'72 0
0 '760
, 71
a
(I)
. 0
'64 !80
0 0
, 81
0
z
a a
H
' 69 0 0
U)
E-<
<t:
,....l
~ '66 0
z
H
'70 '78
' 7cB
.,..
0
a '65 0
'67
0
'68 '84
0 0
a
N
a
a~------r------.------'---~--r------.~----,---f---r------+------'
o
On Constructing Objective Functions 1
Abstract
According to our results, the features of this algorithm are the following: It requires minimal
input information which may be obtained by interviewing an expert; the algorithm requires
only ordinal information about preferences, like preference, or indifference; the objective
function can be uniquely constructed from the obtained data; from incomplete data the
objective function can be constructed approximately, and this construction is stable with
respect to inaccuracies in the data; any input data are consistent with the formal assumptions
of the model; consequently, no contradiction will arise in applications.
In this paper we also discuss briefly how the algorithm considered can be adapted to construct
a quadratic objective function f(ul,' .. ,un) = 2::'=1 ai U ; + 2: i #J bijUiUj + 2::'=1 Ci1(i·
Keywords: Separable objective function; quadratic objective function, data from inter-
viewing experts, interactive procedures.
Contents
1 Introduction
7 Appendix: Proofs
References
1 Introduction
Recall that the statement of a choice problem presupposes determining a set of alternatives
which are commonly identified with the vectors of their characteristics, as in linear program-
ming. The preference structure is often formalized by a scalar-valued objective function
(called also utility function, or goal function) which adopts greater values at better alter-
natives. The physical, or economical constraints like budget, resources, technology, etc. are
taken into account as restrictions. The optimal choice is realized as finding the maximum of
the scalar-valued objective function under the given restrictions.
We can say that the desire is compared with reality-the objective function with the restric-
tions, and their interaction determines the optimal decision.
(-1) finding the global maximum of the objective function under the given restrictions.
Although no item from the above four can be considered as quite simple, we think that the
weakest links are the first two. Indeed, solving an optimization problem is a technical task,
and although mathematical methods may differ in efficiency, in principle any optimization
problem is solvable by scanning. The physical and economical constraints are usually known,
and their representation by restrictions is not so difficult. However, the preference function,
or the objective function, is not so easy to reveal. It may depend on subjective intentions
and goals which are not realized clearly, or which are not commensurable. Finally, in most
cases we are not even capable to recognize possible alternatives. Even if they are given, they
may not be characterized by numbers, or be incomparable.
In the present paper we deal with the second item. The main assumption of our model is
therefore the existence of a set of alternatives characterizable by the same specifications, i.e.
representable by vectors of the same n-dimensional space. The main attention will be paid
to the commensuration of goals and the substantiation of the related procedure.
Let X be a set of alternatives. For example, X may consist of different models of cars to
be compared. Suppose that each alternative x E X is identified with the n-dimensional
real-valued vector of its characteristics or specifications
For example, each car x can be characterized by its power Uj(x), gas consumption U2(X),
number of doors U3(X), etc.
We shall consider only the characteristics Ui which may adopt all intermediate values from
certain segments [ai; bi]. Obviously, there are characteristics which don't meet this assump-
tion. For example, for cars the number of doors can be 2, 4, or 5, but never 2.5. However,
dimensions, power, maximal speed, gas consumption, etc. are continuous variables. There
are also specifications which are discrete but we can assume their continuity, as in the case
of price measured in dollars.
Thus we identify the set of possible alternatives (not actual!) with an n-dimensional paral-
lelepiped
v = VI X ... X Vn = {(UlJ ... ,u n ): ai:::; Ui:::; bi, i = 1, ... ,n},
where Vi = [ai; bi], ai < bi for all i, i = 1, ... ,n.
To formalize the idea of preference, we consider a scalar-valued objective function
225
We shall consider only continuous objective functions. The continuity implies that close
alternatives cannot differ dramatically in preference. This assumption is quite reasonable
in applications, since the distinguishability of alternatives is limited, in particular due to
inaccuracy of measurement.
We shall also assume that an increase in any specification Uj, while other specifications
remain constant, results in an increase in preference. For example, the more powerful a
car is, while gas consumption, price, tax, etc. remain the same, the more preferable it is.
To meet such an assumption, some specifications should be bounded (like the dimensions,
since bigger cars are preferable only within reasonable limits), others should be inversed (like
the price, since cheaper cars with other specifications being the same are preferable). Thus
objective functions are supposed to be increasing (monotone) with respect to all of their
arguments.
Note that, varying one specification Uj, we suppose that an increase in preference is under-
stood equally whatever the fixed values of the other specifications are. Such a preference
structure is called independent of the other specifications.
Let us illustrate the difference between "independent" and "dependent" specifications with
a simplified version of the example by Winterfeld & Fisher (1975).
Consider two specifications of a car: Length of the body, and steering type (mechanical, or
powered). If the car body is not very large-up to 4.5 meters long, then the larger car is
always preferable regardless of the steering type. Therefore, we can write down:
In this case the specification "lengt h of the body" is not independent of the "steering type".
In the above example the assumption of independence is violated, because a new factor, the
safety, is taken into account. The violation of independence means a revision of preferences
with regard to specifications, which is natural under great changes in parameters, but is not
226
typical for their small deviations. For close alternatives the independence of specifications
is more likely, since the alternatives are regarded from the same point of view and with
respect to the same criteria. Sometimes it is possible to classify alternatives and assume the
independence of preferences within these classes.
Now we generalize the idea of independent preferences from a single specification (charac-
teristic) to a pair of specifications (two characteristics). The idea of independent preferences
means that we can consider a group of characteristics of the alternatives separately from
the other characteristics. Such a consideration is meaningful, only if our preference on the
given characteristics is not influenced by the particular values of the other characteristics.
In particular, the pairwise independence implies that it is meaningful to speak of a global
improvement obtained by an improvement in a pair of specifications which is understood
equally for arbitrary fixed levels of others. For our purposes it suffices to define the pairwise
independence.
Let a preference order on the space V = VI x ... X Vn of n-dimensional vectors u = (Ul, ... , un)
be represented by the objective function
This preference order is called independent with respect to VI X V 2 (of V3 x ... x Vn) if
for arbitrary U3, •.. , Un. The independence of preferences with respect to other pairs of
specifications Vi X Vj is defined similarly.
The above definition is illustrated by Fig. 1 for the case n = 3. The pairwise independence
implies that the related preference cuts (the preferences in two-dimensional planes VI x V 2
shown by indifference curves) are the same at any level of the other specification U3 (in
the case n > 3 we would have n - 2-dimensional vectors (1l3, ... , lin) instead of 113). The
dependence means deviations in this preference, while shifting the cut plane U I x U2 up and
down the axis V3 .
It was established that under rather general assumptions the pairwise independence of pref-
erences with respect to all pairs of specifications is equivalent to that the associated objective
function is separable (=additiue), i.e. it has the form
The discussion of related questions goes back to Ramsey (1931). Later it has been contin-
ued in the context of econometric modelling (Leontief 1947; Samuelson 1948; Neumann &
Morgenstern 1947). A topological approach to independent preferences was formulated by
Debreu (1960) and Koopmans (1972). An algebraic approach based on Holder's theorem on
ordered groups was developed by Krantz, Luce , Suppes & Tversky (1971). One can find
more references in Fishburn & Keeney (1974).
227
_ - - - - - - . (u'\. It; )
_-
_---- _ - - -.- -.
(II ".
U\. U~)
.,.
-------- ---
... -.-..-..--
-----------------
-.."..~ ~...
.... --_..--_.-
/
-...~-..-....
/
....,,. ..-
....
........ -
--------
. .'" ,.... .------ - - - - - - - - - - -----------
/ ../~.-.------ --------------
...-----_ ...--
'."
The interest in justifying separable objective functions is explained by their common use in
econometric decision models. Besides, for them there are special methods to solve optimiza-
tion problems (Hadley 1964).
Considerable attention has been paid to the practical construction of separable objective
functions. Thus, Fishburn (1967) describes 24 heuristic methods for constructing separable
objective functions. However, all these methods have a common disadvantage, namely, the
lack of uniqueness in the determination of the objective function.
We therefore require that every procedure for constructing an objective function must be
backed by a uniqueness theorem, or, at least, by an estimate of the accuracy of the con-
structed objective function. These questions are discussed in the following sections.
function
where 1;( Ui), i = 1, ... ,n, are continuous and increasing, but neither the order .:!/, nor the
objective function f( UI, ... , un) are known.
Assume that two compensation curves (indifference curves) Ml and M z , which are level
curves of the function fl(ud + fz(uz), are given in the coordinate plane U1 x Uz (see Fig. 2,
top). Let Ml correspond to a lower level of the objective function fl(Ul) + h(uz) than M z ,
implying Ml to be below M z .
Also assume that a compensation curve M i , which is a level curve of the function fl (Ul) +
fi(Ui), is given in each coordinate plane Ul x Ui , i = 3, ... , n (see Fig. 3, top).
Suppose that the curve Mi passes through the vertices (aI, bi ) and (b 1 , ai) of the rectangle
U1 x Ui for every i, i = 2, ... , n, as shown in Fig. 2 top and Fig. 3 top (otherwise, points ai
and bi can be chosen appropriately, diminishing, if necessary, the parallelepiped U).
By means of these n compensation curves we are going to construct an approximation
g( U}, • •• , un) of the separable objective function f( UI, ... , un) which corresponds to the pref-
erence order :5. 9 . This preference order approximates the unknown preference order :5.'. At
first we shall present the related algorithm, and then we shall explain its background.
and
gl(s~)=t:+k
for every k, k = 0, ... , I<, where t: = 0 if s~ = aI, and 0 < t: < 1 if al < s~ (see Fig. 2
bottom), i.e. t: f. 0 if the upper "stair" in Fig. 2 top is incomplete.
3. Transfer the obtained segmentation from U1 to other axes Ui by "reflection" from the
curves A{. In other words, on each axis Ui , i = 2, ... , n, determine the sequence
o K
ai = Sj < Sj1 < ... < Sj ::; bj
such that
k K-k
(Sj,Sj )EMj
for all k, k = 0, ... , I<, as shown in Fig. 3 top.
229
{ ,
(/,
.0
-'I
!i;(U,)
o
[/,
Fig. 3. Construction of gi( lli) which approximates ji( lli) for i = 2,3, ... , n
231
4. For i = 2, ... ,n define an increasing piecewise linear function gi( Ui) on Ui , put ting
By the way note that the number 1( + 1 of points s?, s~, . .. ,sf' E UI is always finite;
otherwise, the continuous function II (uIl would be unbounded on the compactum [al; bl],
which is impossible.
A strict substantiation of the algorithm is given in the next section.
Note that, in fact, the compensation curves MI and M2 in Fig. 2 top and Mi in Fig. 3
top are not needed for our approximate construction. For each axis Ui we need only the
segmentation
sio < siI < ... < s,J\ ,
corresponding to an equal increase in I,( H,) by a certain conventional unit. Such segmenta-
tions can be obtained directly by interviewing an expert in the following way:
Suppose that two pairs of specifications are given (cf. Fig. 2 top): (b l , a2) and (st· - I , a2)
which is worse only in the first specification, since sf' - I < bl. This means that the points
(b l , a2) and (st· -I, a2) belong to different compensation curves (indifference curves) M2 and
:'vII, respectively. The expert has to answer the question: 'Which increase in the second
specification does compensate the difference in the first specification'? In other words, the
expert has to determine the point s~ on U2 such that (s{' -1, sD is equivalent to (b l , a2), i.e.
belongs to lvl2 • Now, with reference to the obtained point on /v12 , the expert has to answer
the question: Which decrease in the first specification does make this imaginary alternative
equivalent to the point (st'- I ,a2) on M 1 "? Thus the value s{'-2 on C1 is obtained. Next the
232
expert has to answer again the question: Which increase in the second specification does
compensate the decrease in the first specification? That is, which value s~ on U2 brings the
interviewed expert back to Mz?
This way the "staircase" in Fig. 2 top, which determines the required segmentation of UI
and U2 , can be constructed. The upper corner points of the "staircase" are located on
indifference curve lv[2, the lower ones on indifference curve MI. The indifference curves may
be interpreted as the result of interpolating the related corner points of the "staircase".
The segmentation of the axes corresponds to an increase in the separable objective function
by the same conventional unit. That is, in a certain round of the interviewing process, one
specification value is always moving back and forth between the same imaginary indifference
curves, i.e. between All and /V[2 in Fig. 2 top. Therefore, in Fig. 2 bottom the segments of
the gl (Ul )-axis are always of equal length. The distances between the points s~, si, ... ,s{'
will, as a rule, vary, thus determining the shape of the function gi (ud in Fig. 2 bottom. Of
course, this function is also the result of interpolation.
The segmentation of every other axis Ui can be obtained by interviewing the expert in a
similar way. For that purpose, referring to the segmentation s~, si, ... ,sf of the axis U1 , the
expert has to determine the segmentation s?, s; , ... ,sf of the axis Ui such that the pairs
of specifications (s~, sf) are equivalent to a certain point (all bi ) (cf. Fig. 3 top). Here, only
one indifference curve Mi has been drawn. Since the objective function fl(ud + fi(ui) is
constant along Mi and it (Uj) increases by one on each segment on Uj , the component fi( Ui)
must also increase by one on each segment on Ui. Fig. 3 bottom illustrates the construction
of the functions 9i( Ui) for i = 3, ... ,n.
Since all of our construction activit.ies relate to certain points of indifference, one can see
that the knowledge of compensation curves in their entirety is not needed. Rather, the
compensation curves are the result of interpolation.
To estimate the closeness of binary relations, we shall refer to the oscillation of a function
(Shilov 1973), with which we characterize the compensation curves M 2 , ..• , !lIn- For example,
consider an order on the rectangle Y x Z = [a; b] x [c; d] represented by a continuous increasing
233
objective function f = f(y,z). Then each level curve of the function f(y, z) is a graph of
some decreasing continuous function
z = m(y)
which has the oscillation
Recall that the oscillation is a continuous non decreasing function, defined for all P 2 0, and
equal to zero at p = 0.
Thus we can estimate the distance between the unknown objective function
where Wi, i = 2, ... , n, are the oscillations of the compensation curves NIi E UI X Ui, i =
2, ... ,n, considered as functions of UI, and p is the horizontal distance between NIl and M 2 ,
l. e.
p = sup I y' - y" I.
y' ,y" EUj :(y,z)EA'fJ,(y" ,z)EM2 ,ZEU2
order ~f. Since the Hausdorff distance is not a pseudometric but a metric, the unknown
preference order ~f can be uniquely reconstructed as the limit of its approximations. Thus
we obtain the following theorem.
where ai < bi, i = 1, ... , n. Suppose that there exists an unknown preference order ~f on U
representable by a separable objective function f( UI, ... ,un) = fl (ud + ... + fn( un), where
the fi(ui) are continuous and increasing for all i, i = 1, ... , n. Suppose that compensation
curves 1vIi - the level curves of the functions h(ud + fi(ui) which pass through the points
(ai, bi ) and (6 1 , ai) - are given in the coordinate planes U, X Ui for every i, i = 2, ... ,n, and
that a sequence of compensation curves {;\1,k : k = 1,2, ... }, converging to the compensation
curve AI2 , is given in the coordinate plane U 1 x U2 . Then the preference o7'der ~j can be
uniquely reconstructed from these data.
234
f( t)
v ]
o
-I
Fig ..j. The indifference curves in two-dimensional preference cuts from example 1
for all p, p = 1, ... , m, are the level curves of the function f(ui) + f(u;), see Fig. 5_
Indeed, since f(t) on [0; 2m] is composed by reiterating the curve f(t) on [0;2]' the compen-
sation curves on [0; 2m] x [0; 2m] are also composed by reiterating the compensation curves
on the square [0; 2] X [0; 2]. Consequently, it suffices to show that the objective function
f(u;) + f(llj) is constant along the line tii + H] = 2. In fact,
f(Ui) + f(2 - 11,) = (Ui - 1)3 + 1 + (2 -lli - 1)3 +1= 2,
as required.
Mi C UI X Ui, i = 2, ... , n,
gi : UI -+ Ui , i = 2, ... , n,
where al < c < bi and gI(C) = g2(b I ), as in Fig. 2 top. Suppose that the curve MI lies below
M2! i.e.
gI(ud < g2(ud
for all UI E [aI; c). Then there exist continuous increasing functions II (UI), ... ,fn (un) on
UI , ... , Un! respectively! such that lvfI and 1'.12 are the level curves of the function fI(ud +
f2 (U2)! and Mi are the level curves of the functions fl (ud + Ii (u;) for all i, i = 3, ... , n.
Note that the algorithm described in section 3 can be modified in different ways. For example,
if it is not convenient to compare the first specification with others, the comparisons can
be carried out for other pairs of specifications. However, for our purposes it is necessary
to transfer the segmentation of UI and U2 obtained by means of the "staircase" to every
other axis. For that purpose it suffices to follow a certain tree of pairs of specifications,
transferring the obtained segmentation to all other axes by the corresponding compensation
curves. Estimating the accuracy of the approximation of the preference order may become
complicated, since to estimate h in the proof of theorem 1 we have to consider compositions
of oscillations.
237
c = s~
1+t
o
.0
- I
Fig. 6. The construction of gl(u.) which approximates II(u.) by means of "saw teeth"
For example, if the compensation curves are given in the planes Ui - 1 x Ui, i = 2, ... , n, and
,,;)i(P) are the oscillations of !vIi as functions from Ui- I to Ui , i = 2, ... , n, then the estimate
of the accuracy of the constructed preference order looks like
Also note another modification of the algorithm for constructing separable objective func-
tions. Instead of constructing a "staircase" in the first step one can construct "saw teeth"
as follows .
lao Determine c = s1 E U2 and S~-I E UI to make (sf"-I,C) and (b"a2) indifferent with
respect to our preference. Then determine sf" -2 E U I to make (sf" -2, c) and (sf" -1, a2)
equivalent with respect to the preference, etc. (Fig. 6). In this way the segmentation
is determined so that the first component II (ut} of the unknown objective function
I( UI, ... ,un) = II (lld + ... + In (Un) increases by the same value along every segment
i. SIi+IJ ,Z. -- 1 , ••. , }'
[ SI' .. - 1.
238
In designations of the proof of theorem 1, the Hausdorff distance between the preference
orders ~f and ~g is less than 3nh, where h is the length of the maximal segment [s~'; Sfi- 1]
and the maximum is taken over all ki and i. Since the separable objective function increases
by 1 along each of these segments, and since it is uniformly continuous on the parallelepiped
U, we obtain that h ~ 0 if s1 - sg ~ 0, i.e. if c ~ a2'
It is important that the convergence of the whole construction is determined by the conver-
gence of a single parameter which may be varied and measured with ease.
• Any input data are consistent with the formal assumptions of the model (theorem
3); consequently, no contradiction will arise in applications. Moreover, our approach
requires only ordinal information about preferences, so no cardinal information about
preferences is needed.
Firstly, note that a quadratic objective function without the terms biJUiUj, where i f= j, is
separable. Consequently, our procedure can be applied without adaptation to constructing
incomplete quadratic objective functions, i.e. quadratic objective functions without the terms
biJUiUJ' where i f= j.
239
In the general case, such terms can be eliminated by a linear transform of the coordinates
which reduce the associated quadratic form to a simpler form involving square terms only,
see e.g. Rao (1973). However, the specifications of alternatives (old coordinates) may be
transformed into "factors" of their linear combinations. The main problem is therefore to
compare such factors with each other in order to construct compensation curves (indifference
curves).
According to our algorithm, an expert should answer questions whether he is indifferent
between certain combinations of specifications. In the case of a separable objective func-
tion these questions relate to pairs of specifications. In the case of a quadratic objective
function, more complex combinations of specifications should be considered. The role of the
computer is to help in determining these combinations for comparisons, and in constructing
an objective function from the obtained data.
The computer system may be provided with options for optimal decision making with respect
to the constructed objective [unction. Since our algorithm constructs a separable piecewise
linear objective function, some special optimization programs can be used (Jacquet-Lagreze,
Meziani & Slowinski 1987; Slowinski 1988). With regard to the obtained optimum, the
parameters of the model can be adjusted by recursive estimation in an interactive mode
to provide better consistency between the expert's intuition (subjective optimization) and
computed results. Finally, the objective function and the optimal solution can be determined.
The work with the comput.er can be organized as a series of questions to an expert about
his preferences (indifference) on certain possible combinations of specifications according to
our algorithm. After the expert has answered the questions, the computer can construct the
objective function and determine the optimal decision with respect to the objective function
constructed. If the expert is not satisfied with the computed optimal decision and can suggest
an alternative intuitive optimum, the computer might suggest necessary minor changes in
the expert's answers in order to make his rational answers consistent with his intuitive
judgement. After the expert has accepted certain changes, the objective function can be
constructed again, and the new optimal decision with respect to the corrected objective
function can be determined.
A decision making model can therefore consist of two interacting levels with a feedback.
At the upper level the data about the preferences of the expert are processed; at the lower
level the corresponding optimal decision is determined. The analysis of the optimal decision
can require preference adjustments at the upper level of the model. In turn the optimal
decision with respect to the adjusted preferences can be determined. To make the preferences
consistent with the decisions, several iterations may be needed.
Since the objective function is nothing but the formalization of goals, the decision model
can give a means to recognize and, if necessary, to revise real intentions of the expert, and
to find and to substantiate the optimal decision.
Certainly, to implement such an idea in a man-machine decision making system on a com-
puter, a lot of mathematical and programming work must be done.
The first step in future research could be an integrative review of the approach to constructing
scalar-valued objective functions presented in this paper and the one discussed in Husges &
Gruber (1991).
240
7 Appendix: Proofs
and
S k -- S,
( 8 k) , ... , Sn
1
kn) E
sm = (s~l, ... ,s;;'n) E S.
PROOF OF LEMMA. Normalize the separable objective function f(u], ... ,u n ) = JI(ud +
... + fn (un) in order to make
fi(ai) = 0
for all k, kl = 0, ... , K, where (= 0 if al = s?, and 0 < ( < 1 if al < s? Hence,
1,(s7') = kj
and
Since by assumption
we obtain
hj = max {I Siki -
ki-I
Sj
I} ,
3~1 ESj
and put
Let U = (Uh ... ' un) E U and v = (VI, ... , v n ) E U be such that U ~f v. We shall show that
there exist sk, sm E S such that
p(U,Sk):::; (n + y'n)h,
p(v, sm) :::; (n + y'n)h,
where p is the distance in n-dimensional Euclidian space En and sk -:5. 9 sm. For this purpose
we take Sk', Sm' E S such that U E [sk'; sk'+eJ and V E [sm'-e; sm'], where [sk'; sk'+eJ denotes
a small parallelepiped with the diagonal vertices indexed by k' = (k~, ... , k~) and k' + e =
(k~ + 1, ... ,k~ + 1), i.e.
k' k'+e]
[s;s ={
u=(UI, ... ,U n ) k' k'+1
:Sj':::;Uj:::;s' ,i=l, ... ,n }
242
u
k'
S
sm
nt'
S
. m'-e
.S
a1 = s~
C
a I = ""I ,<? .5: >
~i .< .~
(Fig. 7). The parallelepiped [sm'-o; sm'] is defined similarly. Since -:5. 1 is strictly monotone,
and
Note that Sk' -:5. 1 sm' (otherwise v -:5. 1 sm' ~1 Sk' -:5. 1 u, hence v ~1 u, against the assumption).
Therefore,
k~ + ... + k~ - n ::; m; + ... + m~
such that
and
k1 + ... + k n = k~ + ... + k~ - n.
If there is no such k, take
1.:=(-1,0, ... ,0) ,
243
i.e. let
such that
m~ :::; mi, i = 1, .. , n,
and
ml + ... + mn = m~ + ... + m~ + n.
If there is no such m, take
m = (J(,J( + 1, ... ,J( + 1),
i.e. let
Sm = b = (b l , ... , bn ) .
If Sk = a, or sm = b, then by increase of 9 we get sk -<.g sm; otherwise, taking into account
the above proof and the definition of k and m, we have
kl + ... + k
n = k; + ... + k~ - n :::; m~ + ... + m~ = nIl + ... + nI n - n,
Con versely, for any u, v E U such that u ~g v there exist sk, sm E S such that
p(U,Sk):::; (n + .;n)h,
p(v,sm):::; (Il + vn)h,
and sk ~f sm. Therefore, the Hausdorff distance between ~f and ~g in E 2n is not greater
t.han
Finally, note that hi :::; p and hi :::; ""'i(hd for every i = 2, ... , n, whence
its required.
244
<k+2
<2
U;+I [
Sk+1
2
~r;
:
[ Sk
Uf [
.2
S~ = a2
Such a construction is possible and unambiguous, since the functions gl(ud and g2(UI) are
continuous and decreasing. The number 1( is finite, because gl (u\) < g2( UI) for all u\ E [at; c]
implies
inf 1y' - y" 1= t > O.
y',y" EU,:(y' ,z)EM, ,(y" ,z)EIIf2 ,zEU2
Hence, the differences s~ - s~-\ > t for every k; consequently, J( cannot be infinite.
and
245
and
2. Suppose that for a non-negative integer k < f{ the continuous increasing functions
fl(UI) on U{-k and !2(U2) on U;, satisfying the conditions
3. Since 91(Ul) E U; for UI E U{'-k-I, by the previous item the continuous increasing
function
fl(ud = -f2(91(UI))
is defined for all UI E U{-k-I, and
Since 92( UI) E U;+I for HI E [:/" -k-I, the continuous increasing function
246
This way the continuous increasing functions fl(ud on UI and f2(u2) on U2 can be
constructed.
as required.
References
Chow G.C. (1981) Econometric Analysis by Control Methods. New York: Wiley.
Debreu G. (1960) Topological ~[cthods in Cardinal Utility Theory. In: K. Arrow (Ed.).
Mathematical ivlclhod., in the Social Sciences, 1959, Stanford. Stanford: Stanford
University Press, 16-26.
Fishburn P.C., Keeney R.L. (1974) Seven Independence Concepts and Continuous Mul-
tiattribute Utility Functions. Journal of Mathematical Psychology, 11,294-327.
Gruber J. (Ed.) (1991) Economdric Decision Models: New Ale/hod., of Modeling and
Applications. Proceedings of the Second Intemalional Conference on Econometric De-
cision Models, University of !fagen, Held in Halls NOl'dhelle, Allgllst 29-September 1,
1989. Springer-Verlag, Berlin. etc.
Hadley G. (1964) Nonlinear and Dynamic Programming. Reading. Mass.: Addison Wes-
ley.
247
Hiisges H., Gruber J. (1991) Least Squares Estimations of Quadratic Preference Func-
tions for Econometric Decision Models Based on Survey Data. In: Gruber J. (Ed.)
Econometric Decision Models: New Methods of kfodeling and Applications. Proceed-
ings of the Second International Conference on Econometric Decision klodels, Univer-
sity of Hagen, Held in Haus Nordhelle, August 29-September I, 1989. Springer-Verlag,
Berlin, etc., 1991.
Jacquet-Lagreze E., Meziani R., Slowinski R. (1987) MOLP with an Interactive As-
sessment of a Piecewise Linear Utility Function. European Journal of Operational
Research, 31, 350-357.
Krantz D.H., Luce R.D., Suppes P., Tversky A. (1971) Foundations of Measurement.
Vol. 1. New York: Academic Press.
Ramsey F.P. (1931) The Foundations of Mathematics and Other Logical Essays. New
York: Harcourt Brace Jovanovich Inc.
Rao C.R. (1973) Linear Statistical Inference and Its Applicahons. 2nd ed. New York:
Wiley.
Shiloy G.E. (1973) Elementary Real and Complex Analysis. Cambridge, Mass.: M.LT.
Press.
Winterfeld D., Fisher G. (1975) Multiattribute Utility Theory: Models and Assessment
Procedures. In: D. Wendt and Ch. Vlek (Eds.). Utility, Probability, and Human Deci-
sion Making. Selected Proceedings of an Interdisciplinary Research Conference. Rome,
3-6 September, 1975. Dordrecht, Holland, 1975,47-85.
Optimal dictatorial and multi-dictatorial choice in
Arrow's model with applications to multicriteria deci-
sion making
Andranick S. Tanguiane
Computing Centre of the USSR Academy of Sciences,
40 Vavilov Street, 117333 Moscow, USSR
Abstract
We apply this result to Arrow's collective choice model and revise Arrow's paradox. It
follows that there always exists a dictator who is a representative of the collective rather
than a dictator in a proper sense. The refinement of the concept of a dictator leads to the
consistency of Arrow's axioms.
Besides single representatives, we consider the cabinet (named by analogy with a cabinet
of ministers) which consists of a few dictators with delimited domains of competence. We
show that the representativeness of optimal cabinets tends to 100%-values with the increase
in their size, not depending on the size of the collective. We suggest a geometric interpre-
tation of optimal dictators and cabinets. It is based on the approximation formulas for the
indicators of representativeness derived for the model with independent individuals. Finally,
for cabinets we establish the consistency of different concepts of optimality, resulting from
the use of different indicators of representativeness.
Our model has applications to multicriteria decision making. The appointment of a cabinet
corresponds to the selection of a few partial criteria. Therefore, our results can be used for
the reduction of the set of criteria to a certain sufficient minimum.
Keywords: Arrow's collective choice model; collective choice; multicriteria decision making;
dictator; representative; cabinet; mathematical theory of democracy.
Contents
1 Introduction
2 Definitions
3 Main Results
249
6 Conclusions
References
1 Introduction
In the present paper a possibility to overcome Arrow's paradox in social choice is discussed.
K.Arrow (1951) proved that under certain assumptions the collective choice is determined by
an individual, called a dictator. To overcome Arrow's paradox, two different approaches have
been suggested. Usually, Arrow's conditions are interpreted quite restrictively. Therefore, to
overcome the paradox, it is possible to omit one of the conditions, or weaken their formulation
(see Kelly (1978) for several variants of Arrow's theorem). Another approach was developed
for the model with an infinite set of individuals interpreted as a large collective. For such
a model Fishburn (1970) proved that the dictator may disappear. However, Kirman &
Sondermann (1972) showed that in the infinite model the dictatorial origin is still present
but in an indirect form, being realized by coalitions grouped around an "invisible" dictator.
In Schmitz (1977), Armstrong (1980, 1985), and Fishburn (1987) it is established that the
dictator makes decisions together with certain deciding coalitions. These coalitions form a
kind of deciding hierarchy (Tanguiane 1980).
The definition of a dictator doesn't specify whether the dictator compels other individuals
to bend to his will, or whether the collective preference is formed in the first turn, and
only afterwards an individual who shares this preference is found out. Nevertheless, the
interpretation of the concept of a dictator in Arrow's model depends on what we put first--
dictatorial or collective preferences. The former possibility corresponds to our understanding
of dictatorship in a common sense and the latter possibility meets the interpretation of a
dictator as a representative of the collecti ve.
The present paper and other publications (Tanguiane 1989a,b; 1991a-c) are devoted to the
analysis of the latter possibility, i.e. to the interpretation of a dictator as a representative.
As follows from the cited works, the dictator may be considered as a representative of the
deciding coalitions. The problem therefore is to estimate their weight and show whether the
dictator can be interpreted as a representative of the whole collective.
For this purpose we define probability measures on the elements of Arrow's model which are
interpreted as the weight of coalitions, and the probability of different cases in decision mak-
ing. For arbitrary measures we show that a dictator can be appointed to express a majority
preference in most cases. Owing to the interpretation of a dictator as a representative, we
can speak about overcoming Arrow's paradox. The paradox can be regarded as the result of
an excess of information in Arrow's model, which comes into contradiction with itself. An
excess of information can be caused either by an excess of rules with which the contradictive
250
However, even for an optimal representative, the part of cases, when his preference contra-
dicts that of a majority, can be quite significant. In this connection the question emerges,
whether it is possible to take a few representatives and to delimit their competence in such
a way that the decisions of the resulting "cabinet" (named by analogy with the cabinet of
ministers) would meet the majority preference in all, or almost all situations.
In section 2, "Definitions", we introduce two quantitative indicators of representativeness of
individuals: The average representativeness (the mathematical expectation of the weight of
the coalition represented by the given individual), and the majority representativeness (the
probability of the event when the given individual represents a majority). Then we introduce
the concept of a cabinet which is a multi-dictatorial representation with delimited domains
of competence. This concept generalises the concept of a single dictator.
In section 3, "Main Results", we estimate the saturation of cabinets (understood as the
closeness of indicators of representativeness to their maximal possible values). In particular,
we show that for every positive integer k there exists a cabinet with k members such that
the part of the cases when it represents a majority is greater than 1 - 2- k . For example,
it is always possible to find a cabinet with 7 members, representing a majority in at least
(1- 2- 7 ) ·100% > 99% of all cases, and with 10 members, representing a majority in at least
99.9% of all cases. We also formulate corollaries for single individuals. It follows that there
always exist individuals who can be representatives of the collective rather than dictators in
a proper sense. Restricting the prohibition of dictators to dictators in a proper sense, we
obtain the consistency of Arrow's axioms and overcome Arrow's paradox.
In section 4, "Independent Individuals and Geometric Interpretation", we consider a model
with a large number of individuals with independent preferences. Owing to the indepen-
dence, we can use the law of large numbers to derive simple approximation formulas for the
indicators of representativeness. Using these formulas, we suggest a geometric interpretation
of our model. The individuals and cabinets are associated with the vectors of their chacter-
istics, and it is shown that the vectors associated with the optimal dictators and cabinets
provide maximal projections on certain characteristic vectors of the collective. We also prove
that cabinets which are optimal with respect to one indicator of representativeness are at
the same time almost optimal with respect to the other indicator. This way the ambiguity
in understanding optimal representation, inherent in single representatives, is overcome for
cabinets.
In section 5, "Discussion of the Results", some general remarks and observations concerning
the interpretation of our model are outlined. In the final section 6, "Conclusions", the
251
2 Definitions
In the present section we define a collective choice model and introduce indicators of represen-
tativeness. We refer to such common basic concepts as individuals, coalitions, alternatives,
preferences, using the designations from (Fishburn 1970; Kirman & Sondermann 1972; Tan-
guiane 1980). In addition, we also consider three probability measures on the elements of
the model and introduce the concept of a cabinet.
By V we shall denote a finite set of individuals, called the collective. In the present paper
any subset of the collective A C V is said to be a coalition. For simplicity, we restrict our
consideration to the discrete finite case, although our results are also true for an infinite set
of individuals with a countable coalition Boolean algebra (Tanguiane 1991a).
To distinguish between majority and minority, we define a probability measure f.1 on V. By
f.1( v) we denote the weight of individual v E V; by p( A) we denote the weight of coalition
A C V. Since we refer to a probability measure, the weight of a coalition can be understood
as the percentage of the total number of individuals. Obviollsly, the collective V has the
weight 100%, and the weight of a non-strict majority is greater than or equal to .50%.
Let X be a finite set of alternatives. In social choice the alternatives may be candidates for
the president, different social programs, particular important decisions as "war", "peace",
"20% income tax", "2.5% income tax", etc. In multicriteria choice they may be competitive
projects, different models of cars to be bought, vectors of quantities of consumer products
delivered to certain regions, school-graduates entering the university, etc. Detailed examples
of alternatives and further references one can find in Gruber (1983) and Hwang & Lin (1987).
Let P = {P} be the class of weak orders P on X. A weak order P on X is defined to be an
asymmetric and negatively transitive binary relation on X, i.e. (.f, y) E P =} (y, x) tt. P and
(x,y) tt. P & (y,z) tt. P =} (x,z) tt. P. In this case the notation (x,y) E P means that x is
preferred to y. For example, the relation "x is cheaper than y" is a preference.
To specify the probability that a given pair of alternatives (x, y) will be considered in decision
making, i.e. it will be judged whether x should be preferred to y, deflne a probability measure
( on X x X. The measure ( can be also interpreted as the significance of decisions on the
252
pairs of alternatives. If ~(x,y) > 0, the pair of alternatives (x,y) is said to be significant.
A situation, or profile, is defined to be a combination of individual preferences, i.e. a mapping
f: V -+ P, for each individual l' E V, putting into correspondence his preference f(v) E P.
By F we denote the set of all situations.
Since F is finite, we can define a probability measure v on F to characterize the forecasts
on individual preferences. We interpret the value v(J) as the probability of situation f.
Consider the sample space
n=XxXxF.
A simple event is therefore "a pair of alternatives and a situation":
w = (x,y,j).
P=~®v
The first coalition consists of the individuals who prefer x to y under the situation f. The
second coalition consists of the individuals who don't prefer x to y under the situation f.
Fix an individual v. Since he belongs to one of these coalitions, we can consider him as a
representative of the coalition to which he belongs and whose preference on the given pair
of alternatives under the given situation he shares. The measure of his representativeness is
the weight of the given coalition. We denote it by
( ) ( f) r(Aw) if v E A w ,
m v,w = m V,X,y, =
{ r(A~) = 1- r(Aw) if v E A~, (2)
and call it the representativeness of individual v for event w. Note that since {L(V) = 1, the
value m(v,w) is the part (percentage) of the individuals represented by v for event w.
Now suppose that simple events (pairs of alternatives and situations) are taken randomly.
In this case the individual v represents a certain coalition for each simple event, and for each
simple event he has a certain magnitude of representativeness, associated with the weight
of the represented coalition. This way we define a random variable on n. We call it the
representativeness of individual v and denote it by
The majority representativeness of individual v IS the probability of the event when the
individual v represents a majority:
Em(v) ~ 1/2,
Mm(v) ~ 1/2,
Bya cabinet with k members we understand a set of cabinet members ~V = {VI, ... , vd c V.
We suppose that for every simple event w = (x, y, 1) the cabinet VV accepts or rejects the
decision "x is preferred to y", entrusting it to the member with the highest representativeness
for the given event. Therefore, the representativeness of cabinet W is defined to be the
random variable
(3)
The majority representativeness of cabinet W is the probability of the event when the cabinet
W represents a majority:
A cabinet ~V* = (v~, ... , vi;) E V k is called optimal with respect to the average representa-
tiveness if it has the greatest average representativeness among the cabinets with the same
number of members, i.e. for all WE V k it holds
A cabinet W" = (v~·, ... , vZ*) E V k is called optimal with respect to the majority repre-
sentativeness if it has the greatest majority representativeness among the cabinets with the
same number of members, i.e. for all WE V k it holds
Similarly, an individual v** E V who has the greatest majority representativeness is called
an optimal dictator with respect to the majority representativeness.
From (3) it follows that the representativeness of cabinets increases while adjoining new
members. To characterize the saturation of cabinets, we shall compare their representa-
tiveness with the representativeness of the maximal possible cabinet which consists of the
whole collective V. Obviously, for every simple event w such a cabinet represents a majority.
Its average representativeness will be denoted by Em(V). One can see that its majority
representativeness Mm(V) is equal to 1.
3 Main Results
In the present section we show that the indicators of representativeness of optimal cabinets
with k members tend to their maximal values at the speed of the exponent 2- k • We shall
also formulate corollaries for single dictators with applications to Arrow's paradox.
Emk(W*) ~ 1/2,
1
Emk(W*) ~ Em( V) - 2- kk- 1 (1 - k + 1 )k+l,
Mmk(W") > 1-2-k.
255
In particular, an optimal cabinet with only k = 4 members has the average representativeness
greater than 99.5% of its maximal possible value. It means that already four members almost
saturate the cabinet with respect to the average representativeness. One can see that the
saturation of optimal cabinets with respect to the majority representativeness is somewhat
slower than with respect to the average representativeness. However, the interpretation of
the saturation estimate is more clear. We can say that an optimal cabinet with k = 7
members represents a majority more often than in 99% of all simple events.
Recall that Arrow's paradox (1951) is a theorem about the inconsistency of the following
five axioms.
A3 (Unanimity) If all individuals prefer one alternative to another, then the first one is
also preferable for the collective, i.e. if under situation f E F it holds (x, y) E f( v) for
all v E V, then (x, y) E aU).
A mapping a : F ~ P which satisfies axioms A 1-A4 is called Arrow's social welfare function.
From the standpoint of our consideration a general prohibition of dictatorship turns out
to be rather rough. Indeed, the common meaning of the word "dictator" implies that his
decisions contradict interests of other individuals. If he expresses opinions of large coalitions,
he should be called a representative and shouldn't be prohibited. Owing to the indicators of
representati veness, we can distinguish between proper dictators and dictator-representatives
whose prohibition we cannot justify. In other words, we refine the concept of a dictator to
a proper dictator, not considering dictator-representatives as dictators. Taking into account
these arguments, we reformulate Arrow's axiom AS.
256
A5a (Prohibition of Proper Dictators) There must be no proper dictator with respect
to the average representativeness, i.e. there must be no individual v E V such that
Em(v) < 1/2 and f(v) C O"(J) for all situations f E F.
Take an optimal dictator v· E V and determine Arrow's social welfare function by the rule
O"(J) = f(v·)
Note that we could overcome Arrow's paradox, referring not to average but to majority
representativeness.
Comparing Arrow's formulation with ours, one can notice that our model assumes the ex-
istence of certain probability measures, whereas Arrow's model doesn't. In particular, one
can ask whether theorem 3 really overcomes Arrow's paradox. \Ve can consider the measures
used as some additional information about the individuals and their preferences, always in-
herent in real situations. Theorem 2 is valid for arbitrary measures, implying the potential
existence of an appropriate representative, even if these measures are not known. This is
analogous to the theory of differential equations where the existence of the solution can be
proved, but its analytical expression may be not found. Similarly, theorem 3 implies the
potential existence of a non-dictatorial social welfare function, even if we cannot determine
it. Certainly, we overcome Arrow's paradox at the level of its interpretation, not revising
the mathematical result.
In the light of our consideration one can ask why our approach to overcoming Arrow's
paradox wasn't proposed earlier. Probably, the cause is the common understanding of de-
ciding coalitions as consisting of individuals sharing the dictatorial preference (Kirman &
Sondermann 1972; Schmitz 1977; Armstrong 1980, 1985; Tanguiane 1980). Identifying the
dictator's representativeness with the weight of such coalitions, we can see that it is too
small to interpret dictators as representatives. Our approach is much more refined, since we
consider the deciding coalitions with respect to each pair of alternatives, but not with respect
to preferences. To explain it in simple words, one can imagine a president who shares the
preference of all peasants with respect to one decision, of all workers with respect to another
decision, etc. Nevertheless, as a whole program, his decisions (described by his preference)
may meet no appreciation of these coalitions. Obviously, then he represents a large coalition
257
with respect to each particular decision, and no coalition with respect to his preference as a
whole. Thus, defining deciding coalitions with respect to pairs of alternatives, we take into
account the individual representativeness in the most complete form.
As in most probability models, the most interesting results can be obtained for independent
events. The assumption of independence implies the validity of the law of large numbers
which is useful for the description of sums of random variables. In our case we assume the
independence of individuals, meaning that they can choose their preferences independently.
For the model with independent individuals we use the following designations.
Fix a pair of alternatives (x,y). The preference probability of individual V on a pair of
alternatives (x,y) is defined as the probability that the individual v prefers x to y. This
probability is denoted by
Denote the probability that all members of cabinet W prefer x to y, and that they don't
prefer x to y, respectively, by
II PXy(v),
vEW
II (1 - PXy(v)).
vEW
It is easy to verify that the characteristic vector of a cabinet W with a single member v
coincides with the characteristic vector of the individual v.
By E denote a diagonal matrix of size J( x [{, with the diagonal elements being ~k
> 0:
~(x,y)
E = ( ;' .~, ~ ~.
. .. : : : .. ) (15)
o 0 0 ... ~K
Finally, denote the scalar product of vectors of the considered vector space by
( , ).
where the accuracy of approximation increases with n regardless of the choice of individual
v in (18) and (19).
The above approximation formulas result from the application of the law of large numbers
to our model, and the larger the collective is, the more accurate is the approximation. The
formula (16) implies that, within the accuracy of approximation, the cabinet W whose vector
c(W) provides the maximal projection onto the vector e under the linear transform 3 is
the optimal cabinet with respect to the average representativeness. Similarly, from formula
(17) it follows that the optimal cabinet with respect to the majority representativeness is
the cabinet W whose vector c(W) provides the maximal projection onto the vector sunder
the linear transform 3. The same relates to dictators.
Note that each coordinate of the vector s has the same sign as the corresponding coordinate
of the vector c. Therefore, sand e belong to the same multidimensional angle, bounded by
the coordinate hyperplanes (quadrant if these vectors are two-dimensional). Since all the
elements of the diagonal matrix 3 are positive, the vectors 3e and 3s are also located in
the same multidimensional angle as e and s. Therefore, the angles between all these vectors
cannot exceed 90°. Consequently, the posi ti ve half-spaces determined by the hyperplanes
passing through a with normals 3c and 3s have a non-empty intersection, containing the
whole multidimensional angle mentioned. Since these hyperplanes, within the accuracy of
approximation, separate the characteristic vectors c( v) of dictator-representatives from that
of proper dictators, we obtain the following proposition.
(3c,c(v)) 2: 0, (20)
(3s,c(v)) 2: O. (21)
(b) If vectors 3e and 3s are collinear, then the optimal dictator with respect to one
indicator of representativeness is also optimal with respect to the other indicatcr.
260
Let us show that the collinearity of the vectors Bc and Bs in the formulation of theorem 6
is essential.
The vector space for our example is shown in fig. 1. Since only two pairs of alternatives
are significant, the dimension of our vector space is f{ = 2. Since these two pairs are
equisignificant, the matrix
~
=,=(10 1 0)=~(1
201 0) '
2
In fig. lone can see the characteristic vectors c( vd and c( V2) of individuals VI E VI and
V2 E V2 , the mean characteristic vector 2c, and the sign vector !s.
Since all the assumptions of theorem 5 are satisfied, by (18)-(19) we have
and, similarly,
Em(V2) ~ 0.5482,
Mm(v2) ~ 0.4900.
Comparing the obtained values Em(vd, Mm(vl), Em(v2)' and Mm(v2), we see that in our
example every individual is an optimal dictator with respect to one indicator of represen-
tativeness, but not a dictator-representative with respect to the other indicator, which is
even weaker than an optimal dictator. To show this graphically, two dotted lines, passing
through 0, with the normals C and s, respectively, are drawn in fig. 1. These lines sepa-
rate the space of characteristic vectors c( v) into the half-planes of that of proper dictators,
and dictator-representatives with respect to each indicator (cf. with (20)-(21)). The dotted
domains correspond to the vectors C(Vl), C(V2) which are inconsistent with the system of
261
inequalities (20)-(21) for the given vector c, binding C(VI), C(V2) by virtue of (6) and (9) as
follows:
which is equivalent to
We can conclude that our two definitions of dictator-representatives (with respect to the
two indicators) are inconsistent in rather specific cases, when the following conditions hold
simultaneously:
(a) The angle between the vectors C and s is close to 45° (since s is the bisectrix of the
multidimentional angle to which belongs the vector c).
(b) The absolute value of the vector C is small , otherwise the dotted domain is outside the
domain of admissible characteristic vectors restricted by the inequalities I cxy ( v) I::; 0.5
(by virtue of (9) and the property of probabilities 0 :::: Pry:::: 1).
(c) The individual preferences are antipodal: Pxy(vtl is small for VI E VI, whereas Pry(V2)
262
is great for V2 E V2, and, conversely, Pxz( vI) is great for VI E VI, whereas Pxz( V2) IS
To satisfy the above conditions, the characteristic vectors c(v) must have a quite special
localization. When the characteristic vectors c( v) are widely scattered, we can expect the
existence of the desired dictator-representative with respect to both indicators. In any case,
an individual v with the characteristic vector c(v) from the same multidimensional angle as
c, is the dictator-representative with respect to both indicators of representativeness.
As seen from example 1, our two indicators of representativeness may result in the ambiguity
of an optimal representative (with respect to average and majority representativeness). In
theorem 6 we have formulated the conditions for the consistency of the two definitions of
optimality, which are not inherent in the general case. As shown below, the problem can be
solved for large cabinets.
It follows from theorem 6 and example 1 that for single dictators the consistency of our two
definitions of optimality is true only for the one-dimensional space of preference probability
vectors p = {PXy(v)} which arises when only one pair of alternatives is significant. The
consideration of cabinets, in other words, the increase in the dimension of the space of
representation allows us to extend this result to the multidimensional space of preference
probabili ty vectors.
(x,y)EXxX
lI:here
C= max
(x,y)EX xx :~(x,y) >0
I cxy I,
Q= min
(x,y)EX x X~(x,y»o
I cxy I,
and t(n) ~ 0 has the same exponent as the accuracy of approximations (18) and (19).
Thus the concepts of optimality of cabinets wj;J.-: 'espect to different indicators of represen-
tativeness tend to become equivalent with the enlargement of thE' cabinets.
263
The probabilities PXy( v) can be interpreted as indicators of the individual interest, expressing
the degree of preference. Such an interpretation is similar to that of lotteries in Neumann-
Morgenstern utility theory. Therefore, the methods of revealed preferences can help in the
numerical determination of the probabilities Pxy( v) and, indirectly, the measure v on the
set of situations. Furthermore, if the probabilities Pxy( v) are interpreted as indicators of
the individual interest, then the mean probabilities Pxy can be interpreted as the average
collective interest. In this case, according to (18)-(19), the optimal dictator is the individual
whose interest is "closest" to the interest of the collective. In the interaction of collective
and individual interests we see parallels with the theory of economic equilibrium, cooperative
games, etc.
In comparison with the results concerning optimal dictators, we can draw conclusions about
the advantages and the disadvantages of the collective representation in the form of cabinets.
The advantages are:
(a) The possibility to attain high representativeness by means of limited groups of indi-
viduals (theorem 1).
The disadvantages of the cabinet in comparison with the dictators are caused by some general
principles. Since Condorcet (1743-1794) it is known that the part of the situations when
a majority rule leads to intransitive orderings increases rapidly with the increase in the
number of alternatives and individuals. The same disadvantage, consequently, is inherent
in preferences of cabinets which represent a majority preference. Therefore, we come to a
principal methodological alternative:
This alternative cannot be avoided, and the choice of one of these possibilities should be
made, taking into account many factors outside the model.
6 Conel usions
1. The common approach to overcoming Arrow's paradox suggests to reject the least
evident of Arrow's axioms-the universality condition, or independence of irrelevant
alternatives. We suggest to refine the concept of dictator, implying the refinement
of the related axiom. This approach is analogous to overcoming paradoxes in the set
theory. (Recall that its paradoxes were overcome by a refinement of the concept of a
set itself.) Such an approach was never before applied to collective choice.
265
4. This way we prove that the collective choice can be entrusted to single representatives,
or several representatives organized in certain structures. Their interaction with the
represented collective has many common properties which allows considering cabinets
as a kind of "agents". The cabinet can provide high representativeness, but its pref-
erences may lack "consistency", being intransitive. A dictator is not as representative
as a cabinet, but has an evident advantage in the "consistency" of his preference and
the simplicity of its presentation.
References
Armstrong T.E. (1980). Arrow's Theorem with Restricted Coalition Algebras. Journal
of l'vlathematical Economics, vol. 7, No. 1: 55-75.
Armstrong T.E. (1985). Precisely Dictatorial Social Welfare Functions. Erratum and
Addendum to 'Arrow's Theorem with Restricted Coalition Algebras'. Journal of Math-
ematical Economics, vol. 14, No.1: 57-59.
Arrow K.J. (1951). Social Choice and Individual Values. Wiley, New York.
266
Fishburn P.C. (1970). Arrow's Impossibility Theorem: Concise Proof and Infinite Voters.
Journal of Economic Theory, vo!' 2, No. 1: lO3-106.
Fishburn P.C. (1987). Interprofile Conditions and Impossibility. Harwood Academic Pub-
lishers, Chur.
Hwang Ch.-L. &. Lin H.J. (1987). Group Decision Making Under Multiple Criteria.
Methods and Applications. Springer-Verlag, Berlin.
Kelly J.S. (1978). Arrow Impossibility Theorems. Academic Press, New York.
Kirman A. &. Sondermann D. (1972). Arrow's Theorem, Many Agents, and Invisible
Dictators. Journal of Economic Theory, vol. 5, No.2: 267-277.
Lariehev 0.1. (1979). The Science and the Art of Decision Making. Nauka, Moscow.
(Russian).
Tanguiane A.S. (1991b). Overcoming Arrow's Paradox and Development of the Mathe-
matical Theory of Democracy. The University of Hagen: Discussion Paper No. 160.
Gregory C. Chow
Department of Economics
Princeton University
Princeton, NJ 08544-1017
Abstract
Using data on stock price and dividends, and on long-term and short-term
interest rates, we test an important implication of present value models, that
current value is a linear function of the conditional expectations of the
next-period value and the current determining variable. This implication,
combined with rational expectations RE, is strongly rejected. Combined with
adaptive expectations AE, it is accepted. The latter model can also explain
the observed negative relation between the rate of return and stock price.
Thus the RE assumption should be used with caution; the AE assumption may be
useful in econometric practice.
Contents
1 Introduction
2 Models and method
3 Statistical evidence
4 Adaptive or rational expectations?
5 Stock price model in logarithmic form
6 Conclusion
References
*I would like to thank Ben Bernanke. John Campbell. Angus Deaton. Mark
Gertler. Dwight Jaffee. Guy Laroque. Jianping Mei. Robert Shiller. Kenneth
West. three referees. and especially Whitney Newey for advice and comments.
Xu Zheng for able research assistance. and the National Science Foundation
for financial support.
Editor's note: Reprinted with the permission of Elsevier Science Publishers
from The Review of Economics and Statistics. Vol. 71. August 1989.
pp. 376-384.
270
1 Introduction
The hypothesis of rational expectations (RE) has stimulated much interesting
research in economics for over a decade, although many economists remain
skeptical of its empirical validity. Lovell (1986) provides a survey of some
evidence bearing on this issue. This paper presents further evidence on the
validity of the RE hypothesis by applying it to present value models. A number
of studies have attempted to estimate and test present value models under the
RE assumption, including the recent studies of Campbell and Shiller (1987),
Fama and French (1988), Poterba and Summers (1987), West (1988), and the lit-
erature cited. Many of these studies have found the restrictions imposed by
the RE hypothesis on present value models inconsistent with the data, but the
authors usually prefer to reject the models tested while maintaining the RE
hypothesis. In this paper, I employ a different method of estimation and
testing from those employed in previous studies, and find one implication of
the present value model under RE to be strongly rejected by the data. However,
rather than searching for more complicated models and maintaining RE, I find
that the simple model under adaptive expectations (AE) can explain the data
very well. Such results might persuade some readers to reconsider the possible
validity of the AE hypothesis as compared with the RE hypothesis in some econo-
metric applications.
Section I describes the models and the method used for estimation under the RE
assumption. In section II empirical results on present value models for stock
price and long-term interest rate will be presented. Since the models are
strongly rejected, I will examine in section III the alternative AE hypo-
thesis and show that it explains the data much better than RE. Section IV
tests the model in logarithmic form and shows that the AE version is not only
preferable, but helps to explain the phenomenon of mean reversion in stock
prices.
[1]
[2a]
[2b]
Assumption [1] implies [2] but not conversely; [2] is much weaker than [1]. In
the stock price example. [2a] asserts only that the price of a stock (or a
group of stocks) at the beginning of period t is equal to the discounted value
of the expected sum of money. from selling the stock and keeping the dividend.
to which the owner of the stock is entitled at the beginning of period t+1.
As it allows for the effects of speculation. this arbitrage relation for any
one period is much weaker than the assumption that the price of a stock is the
discounted value of all future dividends. It is well known [see Sargent.
(1987. p. 95)] that the solution to the difference equation [2a] is [1] plus a
speculative term At(l/o)t. where yt is a martingale. with EtYt+1 = At. This
paper is concerned mainly with hypothesis [2] under rational or adaptive ex-
pectations.
By the method of Chow (1983. Chapter 11) we solve [4] for Yt+1 and reduce the
time subscripts of the resulting equation by one to give
-1 -1 -1
Yt = 0 Yt-1 - 0 eE t _1zt _1 + c(l - 0 ) + ut [5]
To estimate equation [5] for stock price. I assume the following model for Zt
Zt = a 1z t _1 + ••• + apz t _p + YOYt + ••• YpY t - p + b + vt [6]
Since Zt-1 is correlated with the residual Ut + Vt-1. I will estimate [7] by
using as instrumental variable the least squares estimate Zt-1 of Et-1Zt-1.
For long-term interest rate. since Zt has the same timing as Yt. Et-1Zt-1 =
Zt-1 and equation [5] can be estimated by least squares.
3 Statistical evidence
Thanks to Campbell and Shiller (1987). I have been able to use their data for
studying the relations between stock price and dividends, and between long-
term and short-term interest rates. The stock price and dividend series are
deflated beginning-of-year price and dividends paid during the year for the
Standard and Poor's 500 stocks from 1871 to 1986 used by Shi11er (1984) and
updated to 1987. The interest rate series are U.S. Treasury 20-year yields
from Solomon Brothers' Analytical Record of Yields and Yield Spreads and the
1-month Treasury bill rate from the U.S. Treasury Bulletin. These series are
monthly from 1959.2 to 1983,11.
The estimates of both coefficients of [9] strongly contradict the theory. The
coefficient of Yt-1 should be 6- 1 > 1. but is about two standard errors small-
er. If we assume asymptotic normal distribution for this coefficient when 6- 1
> 1, the hypothesis that 6- 1 = 1.02, say. would be rejected at about the 2.1
percent level using a one tail test. The coefficient of Zt-1 should be -1, but
is over two and a half standard errors larger. The F(2, 111) statistic for
testing the joint hypothesis that the coefficient of Yt-1 is 1.02, say, and the
coefficient of Zt-1 is -1 turns out to be 5.453, leading to rejection at the
.82 percent level.
In particular, the t statistic for testing the first coefficient of [9] being
1.02 has theoretical left-tail probability of .025 and Monte Carlo relative
frequency of .041. The F(2,110) statistic for testing the joint hypothesis of
the two coefficients being 1.02 and -1 respectively has theoretical right-tail
probability of .022 and Monte Carlo relative frequency of .032. Thus the above
Monte Carlo experiment has demonstrated that the rejection probabilities re-
ported by the standard t and F tests applied to the coefficients of equation
[7] estimated by IV, with 0- 1 as low as 1.02, are reasonably accurate. In the
rest of this paper, when I apply t and F tests in similar situations, I will
assume the standard distribution theory to hold.
One possible shortcoming of the estimates given by equations [8] and [9] is the
assumption of homoskedasticity. Since both dividends and stock price have in-
creased substantially from 1875 to 1986, the standard deviations of the resi-
duals of equations [8] and [9] might be expected to increase proportionally. I
will assume that the variance of the residual of [8] is proportional to (Zt-1)2
and the variance of the residual of [9] is proportional to (Yt-1)2. Indeed, a
regression of log of the square of the residual of [9] on the log of (Yt-1)2
has a coefficient of 1.075 with a standard error of .166. Using weighted least
squares, I have obtained an equation [8a] very similar to equation [8]. Using
[8a] to form the instrument ~t and applying weighted IV, I have modified equa-
tion [9]:
Yt = .919 Yt-1 + 2.358 Zt-1 [9a]
(.077) (1.534)
The F(2,111) statistic for testing the joint hypothesis for the coefficients to
be 1.02 and -1.0 respectively is 9.930, leading to rejection at the .000108
level, much stronger than the evidence from equation [9]. Since the method of
weighted least squares is equivalent to ordinary least squares applied to re-
scaled data, the Monte Carlo experiment reported above is relevant here also.
Another useful way to look at equation [7] is to regress Yt + Zt-1 on Yt-1 and
other variables in Ht-1. Hypothesis [2a] implies that Et-l(Yt+Zt-1) equals
0-l Yt -1 and that variables other than Yt-1 are not important in explaining
Yt+Zt-1. If we regress Yt+Zt-1 on Yt-1 alone, using data from 1875 to 1987,
275
[17]
or
[18]
When equation [18] is estimated by the stock-dividend data from 1875 to 1987,
again assuming the residual variance to be proportional to (Yt-1)2, the result
is
Yt = .899 Yt-1 + 2.767 Zt-1 [19]
(.069) (1.357)
If an intercept is added to [19], its t statistic is only .059. Solving the
coefficients of [19] for 0 and a, we get
o = .965 a = .761 [20]
(.008) (.104)
which are very reasonable. The standard errors in [20] are obtained by apply-
ing weighted nonlinear least squares to [18]. [19] is consistent with AE, but
not with RE.
To see this, let us assume [19] to be correct and try to force a RE interpre-
tation into it. We advance its time subscript by one and take expectations
given Ht to yield
[21]
[21] contradicts [2a], as equation [9a] did, in giving a discount factor 0 lar-
ger than one and a coefficient for Etzt very different from -1. Note the simi-
larity between [19] and [9a]. The coefficients of both make good sense under
AE, but not under RE. Under RE, the coefficient of [19] would be 0- 1 and -1
respectively. Given the restrictions on 0 and a, the coefficient for Yt-1 in
[18] should be less than one and the coefficient for Zt-1 should be positive
under AE. Obtaining a coefficient for Yt-1 close to 0- 1 > 1 and a coefficient
for Zt-1 close to -1 in [19] would contradict AE and support RE. The fact that
[21J, with Etzt added on both sides, is consistent with the regression [11J
also confirms AE, for under AE we can take conditional expectation Et of [18]
for Yt+1 and obtain [21] by using EtEt=O. Hence [11] supports AE while reject-
ing RE.
Multiplying [22] by a and using the right-hand side of [23] to measure aEtYt+1
and aEt-1Yt in the resulting equation, we obtain
+ ca(1-a)(1-oa)-1 + a(l-aa)-lEt
This equation will be compared with equation [5] under RE.
[27] is similar to [25] and is consistent with the AE model [24]. However.
[27] is inconsistent with the RE model [5] in having a significantly positive
coefficient for Et-1Zt and an estimate of the coefficient of Yt-l very signi-
ficantly below any reasonable value of 0- 1• The F(2. 288) statistic for test-
ing the jOint hypothesis under RE that the coefficient of Et-lZt is zero and
the coefficients of Yt-l and Zt-l sum to one is 5.491. leading to rejection at
the 0.46 percent level.
Campbell and Shiller (1987) suggest using a shorter sample 1959-1978 because of
a possible structural shift after 1978. and estimating Zt by 11 instead of 6
own lags and 11 lagged y's. Equation [24] for the AE hypothesis is so reesti-
mated using monthly data from 1959.12 to 1978.12 to yield
Yt = .0324 + .9729 Yt - l + .1263 Zt - .0959 Zt-1 R2 = .9878 [28]
(.0451) (.0124) (.0264) (.0272) s = .1769
OW = 2.161
An estimate of 1-0 is .0959/.9729. giving 0 = .901. An estimate of 1-e is
.0959/.1263. giving e = .241. Using the third coefficient of [24] these esti-
mates imply a coefficient of .0956 for Zt-1. very close to the estimate .0959
in [28]. Applying nonlinear least squares to estimate [24] yields
a = .9006 e = .2342 c = 1.710 R 2= .9878 [29]
(.0276) (.0949) (.455) s = .1765
OW = 2.154
The data are consistent with the AE hypothesis with a < 1. e = .234 and a po-
sitive risk premium. Again. to examine the RE hypothesis in the framework of
[24] we estimate. for the short sample 1959.12-1978.12. a regression equation
of Yt on Yt-1. Zt-1 and Et-1Zt. using Zt as instrumental variable. The result
is similar to [25] and is consistent with the AE model [24]. The coefficient
of Yt-1 is about two standard errors below any reasonable value of 0-1 > 1.
Hence the evidence suggests rejection of [5]. as before.
280
which corresponds to equation [5], and can therefore be estimated and tested by
equations corresponding to [6] and [7].
To examine how well the AE hypothesis can explain the data. we assume
Etln(Yt+1+Zt) to be formed adaptively as in [16]:
Estimation of [36] by least squares for the period 1875 to 1987 yields
implying p = .947 and S = .768, which are reasonable results. Furthermore, the
two coefficients of [37] sum to one. as the AE theory implies.
[38]
Equation [38]. just like [34]. is inconsistent with RE given by [33]. for the
two coefficients should be p-1 = 1.0684 and _p-1(1-p) = -.0684 respectively.
The fact that [37] estimated by LS is very similar to [34] estimated by IV for
dt-1 supports the assumption of [35] and [36] that Et-1Et = O. Hence the evi-
dence confirms AE while it rejects RE.
282
Hence. the log of one-period return is related negatively to Pt-l and posi-
tively to dt-l. with coefficients of the same absolute value. Assuming [39]
to be the true model and taking expectation given Ht-l. we derive
[40] contradicts hypothesis [2a] or [32] under RE. for the latter hypothesis
implies zero coefficients for both Pt-l and Et-ldt-l. Estimating [39] by least
squares for the period 1875 to 1987. with the dependent variable denoted by
~l.t-l as in Campbell and Shiller (1988). yields
It is of interest to note that the model (40] is consistent with. and provides
an explanation for. the regression equation for ~1.t-l reported in Campbell and
Shiller (1988. Table 3):
30 R2 = .086
~l.t-l = .008 0t-l + .137 6d t _2 + .126 £t-l + constant
(.125) (.155) (.085)
where £tQl = «et-2 + •.. + et-31)/30)-Pt-l. et being real earnings. In this
283
6 Conclusion
Using data on stock price and dividends, and on long and short-term interest
rates, I have tested hypothesis [2], which is an important implication of the
present value model [1]. Combined with RE, hypothesis [2] is strongly reject-
ed. Combined with AE, it is accepted. The AE formulation, if accepted, im-
plies a model which is inconsistent with the RE formulation. The data support
the former and reject the latter. The former model is also capable of explain-
ing the observed negative relation between the rate of return and stock price,
but how well it does, as compared with alternative models which maintain ra-
tional expectations, remains to be studied. The acceptance of hypothesis [2]
under AE does not imply the acceptance of the present-value model [1]. Hence
the finding that stock price is too volatile to be consistent with [1] under RE
does not affect the validity of hypothesis [2] under AE. How the latter hypo-
thesis explains the volatility of stock price is another topic deserving fur-
ther study.
This paper has illustrated a well-known fact, that incorrectly imposing the
assumption of rational expectations on an otherwise correct model can lead to
unreasonable estimates of important parameters. Assuming hypothesis [2] under
adaptive expectations to be correct, we have seen that taking conditional ex-
pectations of equations [19], [25] and [36] would lead to models which, if a
rational expectations assumption is forced on them, will yield unreasonable
parameter estimates. Hence the assumption of rational expectations should be
used with caution. This paper also suggests that the assumption of adaptive
expectations can sometimes be a useful working hypothesis in econometric
practice.
References
Ari Lahti
The Finnish Bankers' Association
P.O. Box 1009
SF-OOlOl Helsinki
FINLAND
and
Matti Viren
Bank of Finland
Research Department
P.O. Box 160
SF-OOlOl Helsinki
FINLAND
Abstract
This paper reports some policy experiments carried out with the Quarterly
Model of the Economics Department of the Bank of Finland (QMED). These
experiments illustrate the dynamic and long-run properties of the model. Thus,
it is investigated how different temporary and permanent, random and nonrandom
shocks affect the cyclical path and the long-run growth rate of total output.
The main issue is, however, the role of expectations. We compare a static
expectations version with two rational expectations versions of the model.
These two versions differ in terms of the time horizon of expectations. When
various policy simulations are carried out with these different versions
- both in terms of anticipated and unanticipated shocks - it turns out that
the whole short-run dynamics is crucially affected by the way in which
expectations are modelled. In particular, we find that advance effects can be
of considerable magnitude in the case of the rational expectations versions.
1 We are indebted to Anneli Majava and Pekka Sulamaa for research assistance
and the Yrjd Jahnsson Foundation for financial support. Thanks are due to Paul
Fisher, Josef Gruber and Kenneth Wallis as well as to the other participants
of the Second International Conference on Econometric Decision Models, in
Hagen, and the European Meeting of the Econometric Society, in Munich, for
useful comments.
286
Contents
1 Introduction
2 A short description of the QMED model
3 Some simulation results of the QMED model
4 Examining the dynamic properties of the model
5 Examining the role of expectations
6 Concluding remarks
Reference~
1 Introduction
This paper reports some policy experiments carried out with the Quarterly
Model of the Economics Department of the Bank of Finland (QMED). In designing
these experiments we have paid particular attention to the dynamic and
long-run properties of the model, and, even more, to the role of expectations.
Thus, in essence we have two model specifications to be compared: a model
version in which expectations - to be more precise, inflation, wage and income
expectations - are modelled using a simple static scheme and a model version
in which these expectations are modelled using the rational expectations
hypothesis. Obviously, the rational expectations version is of special
interest because it allows us to examine the difference between anticipated
and unanticipated policy changes and the importance of the time horizon of
expectations. In addition, it is of some interest to investigate how various
computational aspects affect the simulation results.
QMED model is a small, aggregative quarterly model of the Finnish economy. The
current version consists of 35 equations and of 79 endogenous and exogenous
variables. The number of stochastic equations is 21. However, 6 of these
equations are some sort of auxiliary equations for income accounting, the
structure of private consumption expenditure, and for employment and the
labour fource. The remaining 15 main equations are reported in Table 1. In
order to save space, only the coefficient estimates and the basic equation
statistics are reported here. (The corresponding diagnostic statistics, which
rather well pass muster, are reported in Appendix 1, however.) The estimates
are OLS estimates; Hatanaka's iterative IV estimates are presented in Lahti
and Viren (1989). The model uses seasonally adjusted data which are almost
entirely derived from the Finnish National Accounts. A novel feature of the
current version of the model is the treatment of expectations. It is assumed
that expectations concerning the rate of inflation, the rate of change of real
wages, and real income are formed rationally, given information of the current
period. Inflation expectations, in turn, determine the expected real interest
rate, and all these variables together affect, in the first place, private
287
Table 1: OLS Estimates of the Main Behavioural Equations of the QMED Model
+ .204*(m-z)(-3) - .250*cap(-1)
(1.72) (1.92)
+ .023*d1 + 1.800
(2.57) (4.73)
+ .053*pm(-2)
(2.10)
+ .013*d6
(1. 21)
+ .057*dll
(5.96 )
All variables, except r, are expressed in logs, and all expenditures are
defined in real terms. The number of lags in quarters is shown in parentheses
after each lagged variable (i.e. (-1) refers to period t-1 and (+1) to period
t+l). /:,. denotes the first backwards differencing operator and A4 denotes the
fourth backwards differencing operator. The numbers in parentheses below
parameter estimates are t-ratios. R2 = coefficient of determination, D-W
Durbin-Watson statistic and SE = standard error of estimate. The list of
variables is the following (exogenous variables are underlined):
c private consumption
ca current account
cap capacity utilization rate in manufacturing
d1-dl7 dummy variables
er exchange rate
i foreign import demand
R public consumption
gp rate of change in labour productivity (five-year moving average)
hk stock of housing capital
1.& public consumption and investment
if manufacturing investment
ih housing investment
k stock of capital, manufacturing sector
1 wage and salary earners' employment
m imports (excluding oil)
mr scale parameter for capacity utilization
.!l working-age population
pc private consumption prices
pcih pc - pih
.£i foreign producer prices, manufacturing
pg public consumption prices
pi investment prices
pih housing investment prices
.P!!l import prices
pmo import prices of oil
pq GDP deflator
Q£ prices of raw materials
prz pr - pz
px export prices of goods (excluding bilateral exports)
pxf px - pf
pz domestic demand prices
pzm pz - pm
q manufacturing production
r interest rate (government bond yield)
rd discount rate
~ employers' social security contributions
~ linear trend
w wage rate
wc contract wage rate
wn w*(l+s)
wr w*(l+s) - pq
x exports of goods (excluding bilateral exports)
290
xe bilateral exports
y gross domestic product at constant 1985 market prices (GDP)
~ instrumental variable for output (determined by f and g)
yh households' disposable income
yhr yh - pc
Yi instrumental variable for output (determined by xe, f, (px-pq) and ig)
z domestic demand
We are not able to discuss all features of (the rational expectations version)
of the QMED model here. May it suffice to say that it is basically a Keynesian
macromodel in which effective demand plays a crucial role. There are, however,
some features which abstract from this standard Keynesian framework: First,
prices, wages and interest rates are not (completely) rigid; second, the
3 Various values for the path extension parameter k and different tolerance
levels were used in solving the model. Thus, k was taken to be between 2 and
10, and the tolerance levels were varied between .01 and .000001 for the three
iteration types which were used in solving the model for the estimation period
(see Fair and Taylor (1983) and Fisher et al. (1985) for details of the
procedure). Different values for k did not make any noticeable difference.
Also, the tolerance level turned out to be of little importance. For instance,
as regards GDP, we could detect an average quarterly difference of a magnitude
of .01 % between two successive simulations, which differed by .10 in terms of
the tolerance level (with k, this difference was of the magnitude of .0005 %).
The average total number of passes through the model for the overall solution
was about 13 000. However, when the time horizon of expectations was extended
to two periods (see the discussion in Section 4), about 30 000 passes were
required. The corresponding average amount of CPU time with a Burroughs A12
computer was about 6 minutes (in the case of the static expectations version
of the model, only about 20 seconds were needed). Most of the computation work
was carried out using the lAS-SYSTEM software (see Sonnberger (1985) for
details). In the present study, the tolerance level was set at .005 for all
three types of iteration; k, in turn, was set at 3. The model was estimated
with three iterations of Hatanaka's IV estimation procedure. On another
occasion, we experimented by continuing the iterations until ten, but the
further iterations did not produce any noticeable difference in results.
291
capacity variable is endogenous, allowing for supply side effects; third, the
demand for labour and capital depends on relative prices and some demand shift
variables - not on actual output; and, finally, expectations with respect to
some key variables are modelled using the rational expectations framework.
We cannot describe the functioning of the model in any detail here. Thus, we
limit reporting to the presentation of the mean absolute percent errors (MAPE)
for some key variables (see Table 2). These statistics are derived from a
dynamic simulation which makes use of data for the estimation period
1971.1-1986.4. As pointed out by e.g. Fisher and Wallis (1988) and Pagan
(1989), these dynamic simulation results should not, however, be used as a
proof for model validation. Rather, the diagnostic statistics reported in
Appendix 1 could serve this purpose. Although we accepted this point. the
dynamic simulation results may still be worth presenting because they show the
relative tracking performance of different equations. In addition, we report
some standard policy simulations using changes in interest rates, contract
wages, public consumption and investment as well as taxes, and, finally, oil
prices as 'policy variables,.4
Clearly, the MAPE values in Table 2 indicate that the model generates the
actual data rather well. This is true particularly if one takes into account
the relatively (in a cross-country sense) high volatility of Finnish exports
and price movements during the sample period. Moreover, there do not appear to
be any noticeable problems in terms of the long-run properties of the model.
OL5 H25
We now turn to the simulation results presented in Table 3. The effects of the
following sustained policy changes on gross domestic product (y). consumption
prices (pc). the interest rate (r) and current account balance (ca) are
tabulated for an increase in: a) the discount rate. b) the contract wage rate,
c) public consumption. d) public consumption and taxes, e) public investment
and taxes, and f) oil prices. All these changes take place at the beginning of
the first quarter of 1972, the model being solved, however. from the first
quarter of 1971. Thus, the changes are already known in advance. We do not
discuss here the time paths of these policy effects in detail. Rather. we hope
that it suffices to point out that these time paths are - at least in our
op~n~on - fairly standard (cf. e.g. Taylor (1985». An increase in interest
rates has a persistent - albeit declining - effect on GDP and a persistent -
albeit very small - negative effect on prices. In the oil price simulation,
the GDP effect is similar but the price effect is - as one might expect - much
stronger. An increase in contract wages has only a temporary positive effect
on GDP. which becomes negative after just two years (this is due to the
deterioration in competitiveness; notice that the exchange rate index is kept
unchanged in this simulation). Finally. expansive fiscal policy has a positive
effect on output.
5 The contract wage, public demand and tax rate simulations are based on the
assumptions that the central bank pegs the discount rate. Moreover. the
exchange rate index is treated as exogenous. Thus. the resulting real interest
rate and real exchange rate effects do to some extent affect the long-run
adjustment paths, which. however. are not of the main interest here.
293
Let us first turn to the exercise with temporary policy shocks. The exogenous
(policy) variables which are analyzed here include public consumption (g),
contract wages (wc), discount rate (rd), income tax rate (tax), foreign import
demand (m) and import prices of oil (pmo). A 10 per cent shock is introduced
to each of these variables and the time path of GDP for each particular
variant is derived by means of dynamic simulation. Then the cumulative
percentage differences between the base and the variant solutions are
computed; these differences are presented in Figure 1. The sample period is
1971.1-1997.2. The data for the exogenous variable covering the period
1987.1-1999.4 are based on the extrapolated values of these variables. 7
Clearly, the effects of the temporary shocks on GDP die out rather quickly and
the long-run cumulative effect is zero. The discount rate represents some sort
of exception. The effect of the one-quarter shock on GDP lasts several
quarters and dies out rather slowly (notice, however, that the magnitude of
the whole effect is very small). Towards the end of the sample period, the
time path of GDP clearly deviates from the corresponding control solution
path. (A similar effect can be discerned with other policy shocks as well).
The obvious reason for this is that the model is here solved without any
terminal constraints, which can be of cPlcial importance given the rational
expectations form of the model. 8 Anyway, these simulations suggest that the
stability properties of the model are not far from satisfactory.
After this, we seek to analyze the cyclical properties of the model. This is
done by introducing, in the first place, purely random shocks to both the
endogenous and the exogenous variables of the model and by solving the model,
and then scrutinizing the cyclical pattern of the resulting time paths of GDP.
This exercise is carried out in the following way. First, we generate data for
the exogenous variables for a sample period of 150 quarters. In the case of
the control solution, these variables obtain constant values for all quarters,
the values being equal to the actual val'les of the last quarter of 1985. Then,
in the first experiment we introduce random shocks to ten exogenous variables:
public consumption (g), foreign import demand (f), discount rate (rd), foreign
producer prices (pf), working age population (n), bilateral exports (xe),
import prices excluding oil (pme), import prices of oil (pmo), income tax rate
(tax) and contract wage rate (wc).9
G See Howrey and Klein (1972), Naylor et al. (1969) and Bianchi et al. (1978)
fer earlier examples of this kind of 2.:\ l",is. See also Hm.ney (1980) for a
:pvi€~ of basic methodology.
9 It also turned out that the time pat!, cf the dynamic S01ution for the last
observations in the data sample was rather sensitive in terms of computationa:
u,--:;uracy (cf. fn. 2). In general, this cer1sitiveness ceald be decreased by
ipcreasing accuracy from, say, the conventional .01 to .0001. As far as the
role of terminal conditions is concerne~. see e.g. Fisher (1987) and Minford
et a1. (1979).
9 In the case of the discount rate (rd; .ru~ the tax pa~ameter (tax), the
~ tandard dev ia tion is computed around t.~i2 :nean.
295
I
I ! !
!I I
:
2 0.5
! -+~
i 0 .0
i
i i j i.
~T"'" I
a
I - 0 .5
I II
V
=- I
- 1 1975
=t I I I I I II
ao
I II I
85
III
90
i II
95 2000 - 1.0
-1 1/ 1 111 1
197 5 80 85 90 95 2000
- (' r--
0 .0
0 . 02
-
( I'- ./
il - 0 .5
0.00
~I
-1.0
- 0 . 02
-1.5
- I I -I I II I :1 I II I! /I I 1111 1 1 11 1 111 1 11 1
- 0 . 04 1975 I " /I -2 .0
80 85 90 95 2000 1975 80 85 90 95 2000
1.0 0 . 00
rr- r-... . . . . '1.
0.5 - 0 .01
0.0 - 0 .02
- 0 .5 f - 0 .03
I
- 1.0 - 0 . 04 II I II I I Il l!
1975 80 85 90 95 2000 1975 80 85 90 95 2000
The random shocks are generated by using the normal distribution so that the
standard deviations correspond to those obtained from the computed deviations
between the actual values and the time trend. Thus, random shocks occur in all
of these ten variables in each one of the 150 periods. The model is solved
using these shocked exogenous variables, and spectral densities are computed
for the corresponding log difference of GDP. The exercise is repeated ten
times and the resulting average value of the spectral densities of ~log(GDP)
is presented in Figure 2. 10
In the same way, stochastic shocks are introduced to endogenous variables (cf.
equations (1) - (13) in Table 1). Thus, a nonzero error term (with the
standard deviation corresponding to the standard error of the estimate) is
introduced to all structural equations and the model is solved by means of
dynamic simulation for the same sample of 150 observations but now using
constant values of exogenous variables. Again, spectral densities of ~log(GDP)
are computed for the resulting 149 observations and the exercise is repeated
ten times. The average value of these densities is also presented in Figure 2.
Now, what can be said of the cyclical properties of the QMED model? As far as
shocks in exogenous variables are concerned, it is immediately obvious that
the resulting cyclical behaviour is characterized by short-term movements with
a duration of less than one year, so it is really more a question of some sort
of seasonal cycle. There is a weak cyclical component representing a two-year
cycle but this is really overshadowed by the short-term cyclical behaviour.
Finally, there are no signs of long swings (lasting more than, say, ten
years). Given the earlier results with temporary policy shocks presented in
Figure 1 these long-run properties are not really very surprising.
10 The Tukey-Hanning window is used, the lag length being 40. Notice that
because of the rational expectations specification of the model the random
shocks are not genuinely random but they are in a sense expected. Thus our
results cannot be directly compared with some earlier results from models
which use backward-looking expectations or have no expectations at all.
II See also Salmon and Wallis (1982) for a review of earlier results.
297
0.0
0.0 L L - - - - - - - - - L - - - - - -__~__________
L __ _ _ _ _ _~0.5
0.0
0.0 LL--------~----------~--------~--------~O.5
298
The dynamic simulation experiment with these autocorrelated shocks was carried
out in the same way as the above-presented experiment with the purely random
shocks. The resulting spectral density functions are presented in Figure 2.
Clearly, the GDP series display some cyclical behaviour. When exogenous
variables are shocked, the GDP cycles are still rather short. although a clear
cyclical peak at the frequency of 2.2 years can also be discerned. In the case
of shocks to endogenous variables, the corresponding cycle length is about 3
years. Obviously. these cycle lengths are too short compared with conventional
business cycles, although with Finnish data for the 19703 and 1980s business
cycles are not easily found. Only a rather weak 4-year-cycle can be discerned
but that does not represent an important point of reference. This is not.
however, important. It is obvious that the cycle length of the generated
series can be increased by increasing the persistence of shocks. In this
context. the important thing is to know that reasonably persistent shocks do
generate medium-term cycles while purely random shocks only generate
short-term cycles (or no cycles at all).
Next. we deal with the question of how expectations functi.·:)n in this model. To
do so. we consider four alternative versions of the model. or ways of
performing the policy simulation. The policy simulation on which we
concentrate here is the contract wage simulation. in which the contract wage
rate is increased by one per cent at the beginning of the first quarter of
1972. The alternatives are:
3) We repeat simulation 1 (explained above) but assume now that the policy
change is not known in advance.
4) We change the model by using static expectations for the rate of change
in consumption prices, wages and households' disposable income,
respectively.
The results of these simulations are reported in Figure 3. When the results
are scrutinized, it turns out that there is at least one important difference
in the results. Namely, simulations 1 and 2, on the one hand, and simulations
3 and 4, on the other hand, differ considerably in terms of the advance
effect. Thus, if we use a rational expectations specification and allow
economic agents to know the policy change in advance this change already has
significant (real) effects before the policy change actually takes place.
In this model, the advance effects are mainly due to the income effect:
households' real disposable income will increase and households react to this
increase in advance by increasing consumption and investment in residential
construction. In addition to this "income effect", the model reacts to
inflationary expectations. A change in inflationary expectations operates
through the direct inflation effect (which affects consumption negatively) and
the real interest rate effect (which tends to increase consumption) (see
equations 3 and 4 in Table 1).12
It is not only in terms of the adva~ce effects that these four simulations are
different. The whole short-run dynamics are very different. In the case of
simulations 1 and especially 2, the advance effect dominates the whole
short-run dynamics. This effect is of considerable magnitude. If the time
horizon of expectations is two periods, a one per cent increase in contract
wages increases GDP by .06 per cent during the previous year. If the time
horizon is only one period - as in the basic version of our model - the effect
en an annual basis is a little less than one half of this. By contrast, in the
case of simulation 3 (the "surprise case") the positive short-run effect is
strikingly small. The positive consumption (real wage) effect is nearly offset
by the simultaneous negative net exports effect. Finally, in the case of
simulation 4 (with static inflation and income expectations) the simulated
time paths are very close to those of the "surprise" simulation 3 (the
positive output effect being, however, slightly larger). Thus, in a sense one
can say that old-fashioned static-expectations macromodels assume that all
policy changes are, in fact, unanticipated.
i2 '..men evaluating these results one should keep in mind that prices and wages
,,1:':not completely flexib12 ire Lhe short run. Hence, the results may be more
consistent with models .;£ tee 'Teary cL:.l Stiglitz (1983) type than with
:'t'~;'eotype market-clc:a~ c.:-q, .:u:r rational e1q:ectations models.
~~! ~-~
-0.05 "-_ _ _ _..L_ _ _ _- '_ _ _ _---'
,.0012)FlE;~'WO
lead'
~.~i --.p:.~~2)
E-- FIE:-+-;
IwO I"_ad' -----1d ::: E-----
====~
0.25 E-_-----f7%?1:~~~:>-jo7;lt7.:~:::>l
0.00 t:...._ _ _"'"'~::.L.:::.dLL1L.o""""O>:::LoI:.Go~
-0.05 1 : . - ---
- ' --
- --
-----
'-------'
;.u'P';=..::1
~.~ :=t=:::~:::3lRE:: j3)RE;SU,P.';se
1.00
::: ~-----~~---+~~~~~
~~~
_o.ost= ___--'______ ...1..._ _ _ _- - ' .
0.25 E-=-----~;:y:$i?::F7~~~~ii7;~
0.00 t:....~ _ _ _c.::::..:lG=a:::L.I.~~~O:::\':;.:l
0.00
-0.05 L---'--,9~7:-,-'-.L.J.-:-::'=-'--'-..L.:-:::::-1--'
One could of course discuss the proper ways of specifying the transmission
mechanisms of price and income expectations, for instance in the context of
the present model. It is clear that the present specification is deficient in
many respects. But perhaps the most important thing is, after all, the fact
that rational expectations do matter when evaluating the policy effects. That
is, the short-run effects of various "policy" changes are extremely sensitive
with respect to the way expectations are allowed to have an influence in the
model.
6 Concluding remarks
Equation numbers are the same as in Table 1. The ri's refer to Godfrey's
autocorrelation test statistics for lags of 1, 2, 3 and 4 quarters, CHOW to
the Chow stability test statistic for the period 1977.2 (due to dummy
variables this statistic could not be computed for all equations), J-B to the
Jarque-Bera test statistic for normality, and ARCH(4) to Engle's
autoregressive conditional heteroscedasticity test statistic for four lags.
For other details, see Kramer and Sonnberger (1986). In this context we can
mention that the OLS residuals are typically not contemporaneously correlated.
Thus, only 4 out of 78 coefficients of correlation exceed the 5 per cent
critical value. In the case of squared residuals, however, the corresponding
number is 12. Moreover, if principal components are computed for these (13)
squared residuals, the first two account for about 80 Z of total variation.
Quite clearly the variability of the price and volume series is affected by
some common source(s) of variation.
References
Barro, R.J. and D.B. Gordon (1983): "Rules, Discretion and Reputation in a
Model of Monetary Policy", Journal of Monetary Economics, 12: 101-121.
Fair, R.C. and J.B. Taylor (1983): "Solution and Maximum Likelihood Estimation
of Dynamic Nonlinear Rational Expectations Models", Econometrica,
51: 1169-1185.
Howrey, E.P. (1980): "The Role of Time Series Analysis in Econometric Model
Evaluation", in J. Kmenta and J. Ramsey (eds.) Evaluation of Econometric
Models, Academic Press, New York, 275-307.
Howrey, E.P. and L.R. Klein (1972): "Dynamic Properties of Nonlinear Economic
Models", International Economic Review, 13: 599-618.
Kramer, W. and H. Sonnberger (1986): 'The Linear Regression Model under Test',
Physica-Verlag, Heidelberg.
Kydland, F.E. and E.F. Prescott (1982): "Time to Build and Aggregate
Fluctuations', Econometrica, 50: 1345-1370.
Lahti, A. and M. Viren (1989): "The Finnish Rational Expectations QMED Model:
Estimation, Dynamic Properties and Policy Results", Bank of Finland
Discussion Papers 23/89, Helsinki.
Lybeck, J.A., Bergstrom, R., Bergstrom, W., Carlsson, E., Johansson, S.,
Jarnhall, B., Kukkonen, P., Tarkka, J., Visholm, T., and A. Willman (1984):
"A Comparison of the Dynamic Properties of Five Nordic Macroeconomic
Models", Scandinavian Journal of Economics, 86: 35-51.
Naylor, T.H., Wertz, K. and T.H. Wonnacott (1969): "Spectral Analysis of Data
Experiments with Econometric Models", Econometrica, 37: 333-352.
Salmon, M. and K.F. Wallis (1982): "Model Validation and Forecast Comparisons:
Theoretical and Practical Considerations", in Chow, G. and P. Corsi (eds.)
Evaluating the Reliability of Macro-Economic Models, John Wiley & Sons, New
York, 219-245.
Gary S. Shea
The Pennsylvania State University
Department of Finance
609 Business Administration Building
University Park, P A 16802, USA
Abstract
Campbell (1986) has shown that the linearized term structure hypothesis can reconcile the
conflicting objections to traditional term structure hypotheses raised by Cox, Ingersoll, and
Ross (1981). We build a cOintegrated model of the nonstationary zero-coupon term
structure which imposes the constraints of the Local Expectations Hypothesis (LEH).
When we simulate coupon-bearing bond term structure data, we find the linearized excess
volatility tests specified by Shiller (1979) and Singleton (1980) are nearly always
unreliable. Current practice is to linearize models around 'duration' points which result in
accurate approximations to yield data, but which also result in unreliable volatility tests.
Contents
1 Introduction
5 Conclusions
Acknowledgemen ts
Appendix
References
306
1 Introduction
In this paper we examine a recently developed method for testing simple present-value
models of the interest rate term structure. The linearized expectations hypothesis and its
associated test statistics were introduced by Shiller (1979), a means of attaching standard
errors to the statistics were presented by Singleton (1980), and since that time the
theoretical underpinnings of the model have been laid down by Campbell, Shiller, and
Schoenholtz (1983), Campbell and Shiller (1984), and Campbell (1986). Linearization is
common practice among econometricians to simplify an intrinsically nonlinear model, to
translate nonlinear constraints into a set of equivalent linear ones in a transformed model,
or even to approximate otherwise unobservable data. Each of these respective applications
of linearization technique has appeared in the papers cited above and in an additional
paper by Campbell and Shiller (1987). The intent of this paper is not to cast doubt on the
linearized term structure hypothesis itself, nor on all its results that have appeared so far in
the literature. We do intend to demonstrate however that the linearized hypothesis cannot
be uniformly successful in solving problems to which it has been applied in the study of the
term structure. In particular, we focus upon the reliability of linearized variance-bounds
tests whose results are widely cited as unfavorable to the simple expectations hypothesis
and bond market rationality and whose results are consistent with anomalous pricing
bubbles in fixed term securities.
The literature on excess volatility implied by inefficient market expectations has focused
primarily upon equilibrium relationships in the stock market (between prices and
dividends) and the term structure (long-term and short-term yields) as cases for study. In
the second round of papers in this literature, wide-ranging criticisms were leveled at the
empirical technique of the first papers. These cri ticisms are all well surveyed by Gilles and
LeRoy (1987), but we will summarize a few of these points here to distinguish them from
the criticisms presented in this paper. The criticisms primarily concern stock market
applications of excess-volatility tests and among these we can discern two broad themes.
First, test statistics based upon ex-post rational (perfect foresight) prices will, by
construction, contain finite sample biases; this point was first raised by Flavin (1983).
Related to this was a second critical argument, advanced by Kleidon (1986), that the
smoothing implications of present-value models cannot be reliably inferred from but one
time-series sample (even of very large size) of actual prices and approximated ex-post
rational prices. Such criticisms are not pertinent to the term structure literature, however,
since the use of ex-post rational prices or yields has been quite limited in that context. In
Shiller's (1979) paper, for example, comparable time series plots oflong-term yields and ex-
post rational long-term yields are subject to this criticism, but are offered only as
preliminary evidence of variance bounds violations.
Test statistics for excess volatility within a linearized framework, however, have been a
primary means used to study the term structure. When introduced, it was intended that
linearized term structure tests for excess volatility could substitute for methods based upon
multivariate stochastic models which are the approach taken by LeRoy and Porter (1981).
To date, there has been no examination of the empirical reliability of these techniques as
there has been for the stock market tests. In this paper we examine whether the linearized
framework is an adequate approach to studying excess volatility in the term structure.
The paper proceeds as follows. In the following section we briefly explain the linearized
term structure model, its motivation, and its variance-bounds implications. In a very
heuristic fashion we explain our qualms with the variance-bounds tests used to date. In the
next two sections, we develop and simulate a term structure model which conforms to a
simple expectations model and see whether in that model our qualms are justified. Then
we conclude with a summary of the results, lessons, and speculations on further research.
307
The linearized term structure equation as it appeared, for instance, in the originating
Shiller (1979) paper was
(1)
where {3 is a constant between 0 and 1 and 1> is a constant premium. The long-term yields
(R) refer to yields of coupon-bearing bonds, of course, and the truncated exponential sum
of expected future short-term yields (r) puts greater weight on near-term expectations than
on far-term expectations. The linearization point commonly chosen is (3 = l/(l+C), where
C is the coupon rate of long-term bonds at issue, although in empirical practice, the sample
average yield on long-term bonds is used instead of C. We mi~ht call this 'duration
linearization' since Equation (1) "may be described as setting the (n)-period rate equal to a
duration-weighted average of expected future interest rates of any maturities that cover the
period from t to (t+n)." [
It is at this point that our first qualms about the linearized term structure hypothesis arise.
It is often asserted that the usefulness of the linearized hypothesis lies in its accuracy, but
this has usually connoted the accuracy with which linearized expressions of yields, for
example holding-period yields, approximate the exact yields on coupon-bearing bonds. As
long as we consider only long-term bonds near issue so that yields cannot be far from the
original coupon rates, these approximations can be very good and much evidence has been
marshaled in this direction (cf. footnote 9, Shiller, 1979, Table 1, Shiller, Campbell, and
Schoenholtz, 1983, and Tables I and II, Campbell, 1986). One such approximation used in
this paper is for holding-period yields,
(2)
. 1
with fJ. = fJ(l-&-) .
J (1-&)
If Equation (2) is a good approximation for exact holding-period yields, does that imply·
that its subsequent use in variance-bounds tests is reliable? The question has not been
addressed, but it seems a doubtful proposition. Yield is, in general, a nonlinear function of
coupon and price, and duration linearization is designed to eliminate the nonlinear coupon-
related terms of the function in circumstances when they are likely to be small, that is,
when long-term bonds are selling near par. But in a variance-bounds test we try to infer
the relative variances of zero-coupon yields from the sample variances of coupon-bearing
yields or their approximations. The two objectives of accurately approximating yield data
in levels and accurately measuring the relative variances of yields seem to be quite
dissimilar. Could it be possible that duration linearization defeats the construction of
reliable variance-bounds tests even though it results in accurate yield approximations?
This is our basic qualm and to give it more substance, we now examine the linearized
variance-bounds tests themselves.
The linearized statistics that test the LEH were derived by Shiller (1979) as follows. If
short-term and long-term yields are stationary, the standard deviation of holding-period
yields in the linearized framework is bounded so that a(H) $ aa(r), for large j (thUS fJj~ fJ)
and a = (1_;)-1/2. If yields require differencing in order to render them stationary, the
bounds are a(H-r) $ ba(6.r) and, further, if the covariance between R-r and 6.r is
-
negative, the bounds are a(H-r) $ ca(6.r), where b
2 1/2 /(l-fJ)
= fJ(1-fJ)- and c = fJ/(l-fJ).
Shiller claims an advantage for these variance-bounds inequalities because in them "we
make no assumption, it should be emphasized, about the nature of the random processes or
the information used in forecasting." 2 But, to the contrary, such assumptions are clearly
being made since the relative variances of holding-period yields and short-term yields are
being directly related to interest rate levels; empirical evidence and modelling practice for
the term structure does not treat such a restriction between relative variances and interest
rate levels as relevant. We acknowledge that a linearized variance bound must be some
function of interest rate levels to properly account for duration effects in coupon-bearing
bonds' yields, but is it likely that, only because of linearization, the relative variances of
holding-period yields and short-term yields in the underlying zero-coupon term structure
can be summarized in one parameter and that this parameter is nothing but the simple
function of interest rate levels used in current practice?
We have other qualms. A linearization point is properly a function of an exoiSenously
supplied model parameter or population datum. In this respect, treating fJ as l/l1+C) is
proper, but when the sample average R (Rsa) is used so that fJ itself becomes a random
variable (l/(l+Rsa) whose sampling variation is not accounted for in linearized test statis-
tics. Finally, if we acknowledge that interest rates might be nonstationary, Rsa itself will
be undefined in the population, so it is doubly problematic what the sampling reliability is
of test statistics that depend upon Rsa. With respect to these two latter qualms we note
there is a presumption in this literature that linearization must be around an interest rate
level and a, b, and c are therefore not sensitive to Rsa since Rsa will generally be small and
"( = l/(1+Rsa) will therefore not vary much as a function of interest rate levels.
Shiller confirms (1979, page 1203) that the elastici ty of a with respect to Rsa for his data
appears to be -.5, but for band c it can similarly be calculated that they respectively are
-1.4 and -1. There is also tacit acknowledgement (cf. footnote 16, page 1204) that with
The Local Expectations Hypothesis for nonstationary term structures implies that Hand r
are unit-root processes, but that the difference, H- r, is a stationary forecast error process.
Cointegrated time series models developed by Granger and Engle (1987) define equilibrium
between series of this type in terms of their low frequency coherence. We shall build a
model for Hand r that is cointegrated of order (1,1) and in this way we can treat H- r as
the forecast error that the Local Expectations Hypothesis requires it to be and also flexibly
model the short-term dynamics of Hand r. We adopt the notation of Granger and Engle
and employ a number of the results of the Granger Representation Theorem. Let
r t
H( 1)
t
x t -- H( 21 , where r t denotes the short-term spot yield at time t and Hi j 1 is the short-
t
term holding-p(Oriod yield realized at time t from the sale of a j-period term-to-maturity
3Gilles and LeRo.v (1986) extend Cox, Ingersoll, and Ross' work in exploring the
circumstances Ir, wrllc~ the LF.H '.vc"11:1 re true in a discrete time framework.
310
H( 9)_ r
t t
in our simulations. We put these elements together in a ditIerence equation,
where B is the backshift operator, <' is a vector of multi variate-gaussian errors, A *(0 )=1,
A*(l) has all elements finite and ,*0. This is the so called error correction representation
of a cointegrated system. The Local Expectations Hypothesis implies that rt equals the
expected short-term holding period lield for any long-term bond and this constraint is
easily imposed on the system. Part (5) of the Granger Representation Theorem justifies a
stationary representation of Zt that can be wri t ten,
The Local Expectations Hypothesis merely requires that 0' {-I and J(O)=J(B). We have
chosen a structural form for our simulation models, described in detail in an Appendix,
which conforms to the LEH and also lets us control the relative variances of .0.r and H- r
so that var[H-r]/var[.0.r] smoothly increases with j which is a hallmark of short-term
holding-period yield data. We can also control for the sign of the correlation coefficient
between .0.r and R- r, a matter of some importance in judging the reliability of linearized
volatili ty statistics studied in the next section.
4 The Simulations
The simulations are designed to answer several questions. First, what can we say about
the influence of the relative variances of the zero-coupon holding-period yields and short-
term spot yields upon the linearized variance-bounds tests? Is it possible that the tests
perform adequately if interest rates in some sense are not too variable as has been
frequently asserted in the literature? We have chosen two models in which the relative
variances are quite different. The ratio of standard deviation of H( 9)_ r to that of.0.r is at
a minimum of 2 in one model and at a maximum of 9 in the other. This wide range
encompasses the relative standard deviations observed in most samples of the zero-coupon
interest rate term structure that we have. 4 Moreover, in Shiller'S (1979) paper, similar
ratios of standard deviations for annual series of yields for coupon-bearing bonds are 7.5,
5.6, and 4.1 and are also well wi thin this range. We emphasize that our model imposes
only the LEH constraint, E t -1(H t ) = rt, and so while our simulations will always produce
smooth .0.r relative to H-r, restrictions on the relative variances of long-term yields and
short-term yields (Cox, Ingersoll, and Ross' Return-to-Maturity restrictions) are not
imposed.
4The difference between one-week holding period yields on la-week Treasury Bills and one-
week yields to maturity has a standard deviation 5 times the size of the standard deviation
of differenced one-week yields in Roll's (1970) long series of Treasury Bill yield curves
Remarkably, the ratio is about the same for the difference between one-year holding period
yields on la-year NTT bonds and one-year yields to maturity and differenced one-year
yields. The data set is described by Shea (1984).
311
The second question we wish to address concerns the reliability of the linearized variance-
bounds tests as a function of the sample level of interest rates. Our simulation models are
zero-mean models for f:.r , 6.H, and H- r. By adding constants and drift to simulated data,
we can generate a term structure at any level we desire, but we should also ideally expect
that the rejection rates for the LEH would not be affected by the levels or the drift
characteristics of the models. The linearized variance-bounds test statistics, however,
have interest rate levels as their only parameter and we might therefore expect that the
rejection rates of the linearized variance-bounds tests would vary with the sample level of
interest rates. This would seem to be an undes~rable character of these statistics, but
perhaps the reliability of the statistics at realistic mean simulated sample levels for interest
rates does not suffer. We perform several sets of simulations normalized at different levels
of the term structure to study this matter.
The third question we address concerns the small sample reliability of the test statistics
themselves. There has been no study for such effects even though the only formula for
standard errors for linearized variance-bounds statistics is admittedly asymptotic
(Singleton (1980)). The simulations are therefore performed with two sample sizes and the
linearized statistics are compared to the performance of benchmark statistics pertaining to
the zero-coupon term structure. Since we drive our cointegrated models with simulated
gaussian errors, we know that the simulated variances of 6.r, f:.H, and H- r will be multi-
normally distributed around their population variances. The asymptotic variances and
covariances of the simulated variances are descri bed by
2 , _ 41f
a O"6.r [y 6.r,f:.r Y f:.r,H-r ]
vr [ 2 ----r- ' (5)
O"H_r y 6.r, H-r y H-r,H-r
where Y wr,wr
A A, for example, is the cumulated squared spectral density of 6.r over all
frequencies and YAH
wr, -r is the cumulated squared cospectrum plus the cumulated squared
quadrature spectrum of f:.r and H- r over all frequencies. (See page 465, T.W. Anderson,
1971.) When sample spectral estimates are used in Equation (5), a test statistic for excess
volatility in the zero-coupon term structure can be formed as
0" 2
6.r
1sample - r 2
0"-0"
r
2
f:.r 1actual
SD [ a 2 - 0" 2 ]
H-r 6.r sampl e
We designate these statistics respectively BSTAT and CSTAT and note that they are
hypothesized (Singleton (1980)) to be asymptotically normally distributed with positive
means and unit standard deviations. 5 Such linearized statistics may perform better or
worse in finite samples than the equivalent statistics (ZSTAT) we could use if we could
directly observe the zero-coupon interest rate term structure.
In Table 1 we report the performance of the statistics for 100 simulations of length 200
each. The levels and the drift for each model have been set so that the average short-term
yields for 100 simulations are the same across rows of the table. In this way we know that
the difference in performance between the models is due solely to different population
SD[H(j}-r]/SD[t.r]. We have run simulations of longer length (500) and the results are
not materially changed. The 'low-level' simulations produce short-term yields of
approximately 4% and the 'high-level' simulations produce an average level of 12%. The
low-variance simulations are for models which set SD[H( 9>-r]/SD[t.r] to 2 and the high-
variance simulations have the same ratio set to 9. Finally, all the simulation models have
a positive population correlation between R( 10)_ rand t.r. The details of the simulation
procedures are explained in the Appendix.
Simulations of length 200 seem to be sufficiently long to make ZSTAT behave as the
asymptotic N(O,l) variable it is and the bad behavior of BSTAT and CSTAT does not
appear to be improved by increasing sample size. The major qualms we have previously
stated do seem to be largely justified by our tabled results. First, the sizes of BSTAT and
CSTAT, as opposed to that of ZSTAT, are affected by the volatility built into the models;
as the volatility is increased, the linearized statistics become smaller and therefore more
likely to reject the LEH. Within the realistic ranges of volatility presented, however, we
see that BST AT is not capable of rejecting the LEH and at some point CST A T cannot help
but reject the hypothesis. 6 Clearly, realistic volatility in the term structure introduces
terms into an ideal statistic (ZSTAT) which are fatally ignored in linearized statistics such
as BST AT and CST AT. Second, our qualm that the linearized statistics will be more or
less reliable as a function of interest rate levels is also justified by the results. The high-
level simulations of the term structure are more likely to result in rejection of the LEH
even though the volatility of the simulations remains the same as it was when performed at
a lower level.
We emphasize that linearized procedures fail in this fashion not because they do not
approximate holding-period yields well. The exact holding-period yields on long-term
coupon-bearing bonds are almost perfectly approximated in our simulations. The test
statistics fail because the points around which the term structure model is linearized bear
little resemblance to the linearization point which is optimal for testing variance bounds.
5Calculation of BSTAT and CSTAT employ Hlj) as specified in Equation (2). The
empirical standard deviations are based on the same spectral approach embodied in
Equation (5). BSTAT and CSTAT are not calculated to have a zero expected value as is
ZST AT because, without knowledge of the model, we cannot expect the investigator to
know what the true positive means of BSTAT and CSTAT to be. Therefore, the fact that
BSTAT rarely rejects the null hypothesis does not cast as much doubt on the reliability of
the linearized test statistics as does the nearly uniform rejection of the null hypothesis of
which CST AT is capable. I thank an anonymous referee for pointing this out.
6Population correlation between R-r and t.r is positive for these models and so the
objection might be raised that the tighter bound tested by CST AT should not be used to
pillory the linearized test methodology. But the typical investigator would not know the
sign of the true correlation and so in a certain portion of cases would mistakenly apply
CSTAT. In any case, neither BSTAT nor CSTAT behaves like ZSTAT to give normally
specified rejection rates. Secondly, in simulation models in which the correlation is
rejection (refer to the Appendix for details), practically the same results are obtained; in
hightvolatility models CSTAT cannot fail to reject the true null hypothe is.
Table 1
The Summary of 100 SiJDulated Excess Volatility Statistics, N per simulat ion =200
Low-Level Simulations
Significance .01 0 0 0 0 0 14
Levels for .05 5 0 0 2 0 31
N(O,l) .10 7 0 0 7 0 44
.25 17 0 0 25 0 67
Excess-Volat ili ty Statistics
Mean 0.2 220.0 41.2 0.2 34.0 -1.2
Std. Dev. 1.1 84.3 11.2 1.2 8.1 1.0
Minimum -2.3 69.0 17.6 -2.1 16.9 -3.5
Maximum 3.8 573.4 82.8 3.4 59.9 1.1
0J
......
High-Level Simulations 0J
Significance .01 0 0 0 1 0 75
Levels for .05 2 0 0 6 0 92
N(O,l) .10 5 0 0 9 0 96
.25 26 0 0 26 0 99
Excess-Volatility Statistics
Mean 0.0 90.9 21.7 0.1 17.5 -3.2
Std. Dev. 1.0 29.3 5.2 1.0 6.9 1.1
Minimum -1.9 39.8 11.4 -2.3 2.1 -5.7
Maximum 2.6 191.5 38.0 3.5 35.5 -0.6
The normalization in the' LOll-Level Simulations' is a level of 4 percent. The models in the' High-
Level Simulat ions' panel were normalized to simulate short-term interest rates at a mean level of
12 percent over 100 simulations.
314
5 Conclusions
The linearized system has been called a 'well-tempered' one by analogy to a tuned musical
instrument in which the playable notes are separated by a fixed frequency. Campbell and
Shiller (1984) state that 'well-tempered' refers to the ability to tune a musical instrument
lito more than one key at a time." Well-tempered as precisely used by Bach referred to a
piano in which the octaves and fifths are mean-tempered (have fixed frequency differences)
but the thirds vary in their frequency span (Sloane, 1988). The linearized system of the
term structure is well-tempered in the sense that all of the traditional expressions of the
expectations hypothesis are consistent, whereas they are dissonant when applied to a
nonlinearized, zero-coupon term structure. The musical analogy is apt except that
composers will at times compose a piece containing certain intended effects that can be
achieved only if the musical instrument is tuned to a particular key. Not all pieces will
sound as the composer intended if played on an instrument tuned to the usual reference
key. By the same token the term structure model may be linearized around different
points (keys), but will perform adequately in one key for one purpose (approximating
holding-period yields) and inadequately for other purposes (testing variance-bounds).
There is evidence in this paper that variance-bounds tests 'play poorly' in a linearized
system tuned to the particular keys used by previous researchers, even though their
holding-period yield approximations play very well indeed. Even if we acknowledge that
the 'duration' keys are alright for variance-bounds tests as long as interest rates do not
vary too much, the results of our simulations suggest that this variation threshold is very
low indeed and too far below empirical volatilities for linearized variance-bounds tests to be
practical.
In summary, the linearized system of studying the interest rate term structure is an
important disciplinary step in that it forces our attention onto a model that is time
consistent and does not contain contradictory testable implications. In doing this,
however, not all practical econometric problems are solved or even made simpler. Our
qualms about the variance-bounds test methodology in a linearized framework first arose
because it seemed almost too good to be true; it promised a means for testing for excess
volatility without having to explicitly model the term structure. But the smoothing
properties of expectations theories of the term structure cannot, unfortunately, be
summarized in one parameter that is a function of nothing more than the sample level of
interest rates. It is the selection of a linearization scheme, not linearization itself, which we
question in this paper. The role played by linearization in other variance-bounds studies
and in the closely related asset price bubble studies may similarly reveal Significant effects.
Our small simulation study suggests that the choice of a linearization scheme is not
secondary to such an empirical study, but is one of the primary empirical problems to be
solved.
ACKNOWLEDGEMENTS
This paper has benefited from comments given by John Y. Campbell, Eric Renault, S0ren
Johansen, Katarina Juselius, an anonymous referee and seminar participants at Penn State
University, The Board of Governors of the Federal Reserve System, the European Meetings
qf the Econometric SOciety, the Ecole Nationale de la Statistique et de l' Administration
Economique, and the Second International Conference on Econometric Decision Models.
Any remaining shortcomings are, of course, solely the author's responsibility.
315
REFERENCES
Anderson, T.W. (1971). The Statistical Analysis of Time Series. Wiley, New York.
Campbell, J. Y. and R. J. Shiller (1987). Cointegration and tests of present value models.
Journal of Political Economy 95, pp. 1063-1088.
Cox, J. C., J. E. Ingersoll, and S. A. Ross (1981). A re-examination of traditional
hypotheses about the term structure of interest rates. Journal of Finance 36, pp. 769-799.
Gilles, C. and S. F. LeRoy (1986). A note on the local expectations hypothesis: a discrete-
time exposition. Journal of Finance 41, pp. 975-979.
Gilles, C. and S. F. LeRoy (1987). The variance-bounds tests: a critical survey, Working
Paper, University of California, Santa Barbara, August.
Kleidon, A. W. (1986). Variance bounds tests and stock price valuation models. Journal of
Political Economy 94, pp. 953-1001.
LeRoy, S. F. and R. D. Porter (1981). The present-value relation: tests based on implied
variance bounds. Econometrica 49, pp. 555-574.
Nelson, C. R. and C. 1. Plosser (1982). Trends and random walks in macroeconomic time
series: some evidence and implications. Journal of lvJonetary Economics 10, pp. 139-162.
Roll, R. (1970). The Behavior of Interest Rates. Basic Books, New York.
Shea, G. S. (1984). Pitfalls in smoothing interest rate term structure data: equilibrium
models and spline approximations. Journal of Financial and Quantitative Analysis 19, pp.
253-269.
Shiller, R. J. (1979). The volatility of long-term interest rates and expectations models of
the term structure. Journal of Political Economy 87, pp. 1190-1219.
Shiller, R. J., J. Y. Campbell, and K. L. Schoenholtz (1983). Forward rates and future
policy: interpreting the term structure of interest rates. Brookings Papers on Economic
Activity, no. 1, pp. 173-217.
Singleton, K. J. (1980). Expectations models of the term structure and implied variance
bounds. Journal of Political Economy 88, pp. 1159-1176.
Sloane, C. (1988). A brief survey of keyboard temperaments. Continuo 12, pp. 11-12.
316
As elsewhere in the paper we use the notation of Engle and Granger (1987). The LEH
strictly requires a' 7=1 and J(O)=J(l) in Equation (4)'s representation of the model. By
the definition of Zt as a' Xt, a' is defined as
-1 10 . . . . 0
-1 0 1 0 . . . 0
-1 0 0 1 0 . . 0
a'
9x 10
. 0
-1 0 0 0 0 .. 1
Under the LEH, 7 will be the pseudo-inverse of a'. The choice of J(B) so that J(O)=J(l) is
important in that it determines the dynamics of the term structure model. From Engle
and Granger, we take
where C(B) is C(l) + (l-B)C*(B). But under the LEH constraints, A(B) equals
1+ (Al* + la' - I)B - A2*B3 and, since C-l(B)(l-B) equals A(B), it is readily seen that
Var(z) is a complex function of the choices made for Al*' A2*, and E = E[ff']. This
approach led to a tedious trial and error process in choosing A*(B) and E to satisfy all the
constraints we wished to impose upon Var(z) and Var(Hl n) - r)/Var(!:::..r).
By making assumption i) instead, many of these problems were reduced and the
replicability of our simulations by other investigators was enhanced. Letting A*(B )=1 not
only guarantees J(B)=a', but (along with the other LEH condition 0'7=1) also makes
Zt = a' ft. Therefore, Var(z) becomes a'Ea. We also note that
Letting O'ij denote an element of E, we have constructed E's such that 0'11<0'22<' •••• <O'nn.
The off-diagonal elements of E were also made nonzero. In constructing E it was possible to
guarantee that SD[H(j)-r]/SD[!:::..r] increased with j, but it was not Simultaneously
possible to guarantee that that ratio be greater than 1 for any j, which is generally
observed in term structure data even for fairly small j. We emphasize, however, that this
feature in no way vitiates the model as an equilibrium model of term structure volatility.
To show how this feature arises we present the following mathematics. " being the pseudo-
inverse of a', makes la' = 1- (l/n)u', where L is an n-length vector of ones. Thus
Var(fl.r) is
where ~j denotes summation over i=l, ... ,n and ~i~jO'ij denotes the sum of all elements in E.
Since the variance of Zt is a'Ea, it is easily seen that, for example, Var(H( 1)_ r) is
0'11+ 0'22- 20'12 and Var(H ( n) - r) is 0'11+ 0'22- 20'1n. It is possible therefore in our
simulations that E is chosen so that SD[H (j) - rll SD [fl.r] is less than 1 for some j.
Another result of making assumption i) is that our term structure model imposes a positive
correlation between !:::..r and R( n) - r. This is in contrast to the negative correlation we
were able to impose in our models used in the previous version of this paper. The negative
correlation was obtained by manipulating the autoregressive structure (under assumption
ii) of A*(B). The results in the two versions are about the same, especially with respect to
the overly large rejection rates displayed by CST AT, which suggests that the reliability, or
the unreliability rather, of the linearized volatility statistics is not a function of this
particular feature of the underlying term structure model. We demonstrate the positive
correlation in the present model as follows. It is by definition that
But since
m
It is not immediately obvious that, for any positive definite };, Cov[R~nL r t _ 1,,0,r t J is
positive. But if we consider the positive quadratic form [I(I")-+L']};[I(.,I)-+LJ,
n n
Now for b O'lj $ 0, it is clear that the covariance is positive and for E O'lj > 0 it is simi-
j=2 j=2
n-l 1 n I l 1
larly clear that ---O'll - - - b O'lj > - n bj O'lj- ---;:;2n Eibj O'ij >- ---;:;2n Eibj O'ij· By add-
n n . 2
J=
ing ~2 EibjO'ij to all sides of these inequalities we obtain Cov[R~nL r t _ 1,,0,r t ] > O.
Two versions of}; were constructed: Version 1 gave rise to a simulation model in which the
population ratio, SD[H( 9) -r]/SD[,0,rl, was 2 and in Version 2 the same ratio was set to 9.
In the simulations, }; was used to de1ine the multi-normal distribution from f'S which were
then used to drive Equation (3). Equation (3), along with the appropriate identities, was
simulated repeatedly to generate time series of zero-coupon term struct ure data. This basic
term structure data was then used to generate term structure data for coupon-bearing
bonds. We assumed that coupon-bearing bonds were issued at par with 10 periods term to
maturity and that they bore a coupon for each period. After forming this data in each
simulation, we calculated the one-period yields from holding a coupon bond from issue until
it was a nine-period bond. Since the bonds were issued at par, we found that the exact
holding-period yield and its linearized version (Equation (2)) were nearly identical so, as
emphasized in the text, the failure of the linearized variance-bounds tests is not related to
any inaccuracies in the linearized yields that might be used in the tests.
The simulations are distinguished not only by the version of}; employed, but also by their
length (lengths of 200 and 500 were used), their average sample level, and the
nonstationary drift imparted to Xt. We found that 200 and 500 were sufficient so that
small sample effects were not evident in our simulations. Likewise, in experiments with
drift, we found no appreciable effects on our simulations. We varied the drift from 0.0 to
0.03 percent per period and found, in no circumstance, are the rejection rates by ZSTAT,
BSTAT) and CST AT affected. The 'low-level' simulations in the tables have a 0.03
percent per period drift and the 'high-level' simulations have none. Finally, the
319
simulations are also di vided into high and low level groups. The low-level simulations have
a mean level of short-term interest rates set at 4 percent. By this we mean that the
average level of rt over all 100 simulations was 4 percent after checking that the mean level
of the term structure in anyone simulation was never negative. (Imparting a positive drift
to these simulations aided us in this task.) The high-level simulations were normalized
around an average level of rt at 12 percent.
PART 6:
S0ren Johansen
Institute of Mathematical Statistics
University of Copenhagen
5 Universitetsparken
DK-2100 Copenhagen 0, Denmark
Abstract
The distribution of the likelihood ratio test for cointegration under local alternatives is
found to be given by a suitable functional of a p-r dimensional Ornstein-Uhlenbeck
process. The results are related to those of Phillips (1988), who considered
near-integrated processes and derived the power function of a test for unit roots based
on a regression estimate. The power function is investigated numerically for a
one-dimensional alternative.
Contents
1 Introduction
2 Asymptotic properties of the process X and the power function under
local alternatives
3 Numerical investigation of the power function under the alternative
of one extra cointegration vector
4 Appendix
5 Acknowledgement
References
1 Introduction
The concept of cointegration was introduced by Granger (1981) in order to define the
notion of a stable economic relation among non-stable economic variables. He
considered non-stationary economic variables, i.e. a non-stationary vector process,
and defined a cointegrating relation as a linear combination of the components with
the property that it determined a stationary stochastic process. This makes precise
one of the many meanings of the notion of stability and as such it can be investigated
and tested by statistical techniques, see Engle and Granger (1987), Park and Phillips
(1986) and Phillips and Ouliaris (1990). Many papers have since then been devoted to
finding properties of various regression and eigenvector estimators for cointegration
vectors.
This paper deals with vector autoregressive processes with Gaussian errors where
maximum likelihood estimators and likelihood ratio tests can be found, see Johansen
(1988b). There are now a number of papers that describe this method in detail and
apply it to economic problems, see Johansen and Juselius (1990), Ju~elius (1991),
Kunst (1988), (1989), Kunst and Neusser (1990), Liitkepohl and ReImers (1989),
Hoffman and Rasche (1989), and Garbers (1989). The method will be described briefly
below.
Consider the p-dimensional vector autoregressive process X t , t= l, ... ,T, defined by the
equations:
324
The max.imum likelihood estimation and likelihood ratio test of this model were
investigated in Johansen (1988b), and can be described as follows. First the model
(1.1) is rewritten as
k-1
~Xt = L:1 ri~Xt_i + Ct(3'X t _ k + f t , t = 1, ... ,T, (1.3)
where r.1 = - I + III + ... + II.,
1
i = 1, ... ,k-1. One can also rewrite the
model with the error correction term as Ct(3'X t _ 1 and a suitable transformation of the
other parameters, but the analysis remains the same. In model (1.3) the parameter
matrices r F .. ,r k_1,Ct,(3, and A are variation independent and one can easily maximize
the likelihood function with respect to the parameters r F··,r k-1 by regressing ~Xt
and X t _ k on the lagged differences, and obtain residuals ROt and R kt . We define
I),Skk - o5
Sko S S0k I = 0, (1.4)
giving the maximized likelihood function
-2/T -_
Lmax ISao I IIIr( 1 -).- ..)
1
The likelihood ratio test statistic for the hypothesis H of (at most) r cointegrating
relations is given by
-2 In(Q) = - T Ei+1In(1 -~) (1.5)
The limiting distribution of this statistic is non-standard and has been tabulated by
simulation in Johansen (1988b).
The theory for the model which also allows a constant term is more involved, and is
given in Johansen (1991).
It is the purpose of this paper to derive the power function of the above likelihood ratio
test, applying the results of Phillips (1988).
The alternative we are interested in is clearly that there are one or more extra
cointegrating relations than assumed under H. If we investigate the power of the test
under the assumption that there is in fact another cointegration relation with some
non-zero loadings, then it is not difficult to see that the power tends to one. More
interesting is it to consider local alternatives of the form
HT : IIT = Ct(3' + Ct 1(3i/T,
325
where a 1 and {J1 are pxs matrices. Under the alternative HT we are thus allowing s
extra cointegration vectors to enter the model with small weight aliT. These extra
linear components of X t are "near-integrated" in the terminology of Phillips (1988),
and the basic mathematical results are directly taken from that paper and adopted to
this slightly different framework of analyzing the multivariate Gaussian distribution.
In the next section we describe the asymptotic properties of near-integrated series, and
apply these to find an asymptotic expression for the power function. In section 3 we
investigate by simulation the power function for the local alternative of one extra
cointegration vector with a small loading. The appendix contains the proofs of the
mathematical results.
2 Asymptotic properties of the process X and the power function under local
alternatives
In order to discuss the asymptotic properties we recall some of the properties of the
process X t as given in Theorem 4.1 (Granger's representation theorem) in Johansen
(1991). We assume that rank(a) = rank({J) = r and let a .I. and (J.I. denote px(p-r)
matrices orthogonal to a and (J and of full rank. The condition for the process X t in
a{J' and the mean lag matrix III = I;fiII i , and is given by
rank(a'llI{J ) = p-r.
.I. .I.
(2.1)
Under this condition we have the moving average representation of the differenced
process D.X t = C(L)c t and the expression
THEOREM 1. Under the local alternative H T .' II T = a,B' + a 1{J ;/T, and under the
condition (2.1) and
rank{(a,al)~IlI({J,{Jl)) = p-r-s (2.4)
the process X~T) converges weakly to the process X t' and the process T-~ z(VtJ
converges weakly to the (p -r) -dimensional process Gt' which satisfies the stochastic
326
differential equation
The proof of this result will be given in the appendix. The main result about the
power function is given in the next Theorem:
THEOREM 2. The asymptotic distribution of the likelihood ratio test statistic - 2lnQ
for the hypothesis H: II = a(3' is, under the local alternative H T : II = a(3' + a lf3 liT,
given by
This result is closely related to the result by Phillips (1988), p.1031, on the power
function for a certain test statistic for the hypothesis II = 0 based on a normalization
of a regression estimate of II. In this situation, where we want to test the hypothesis
that II = 0, the present test which exploits the Gaussian distribution can not be
distinguished by its asymptotic properties for local alternatives from the test given by
Phillips.
It is seen that the asymptotic power function for local alternatives depends on the
parameters only through a and b. Thus it depends on how the extra loadings (a l ) and
cointegrating relations (f3 1 ) are related to the a and f3 assumed under H. A tabulation
of the power function thus involves 2(p-r)xs parameters. It turns out, however, that
fewer parameters will do. In order to see this, we shall exploit the invariance of the
multivariate Brownian motion under rotation and the invariance of the test statistic.
327
Let 0 denote a (p-r)x(p-r) orthonormal matrix. Clearly B t and OB t have the same
distribution and hence after multiplying (2.8) by 0 we see that OK t solves the
equation (2.8) for the Brownian motion OB t with the pxs matrices Oa and Ob. The
test statistic is invariant under this transformation and thus the power function is the
same for any alternatives ab' and a 1bi such that Oab'O' = a 1bi. We can simplify still
further by choosing new coordinates using the orthonormal vectors
e 1 = b(b'b)-t
CO RO LLAR Y. Under the local alternative II = a/3' + a 1/3 liT, where a land /3 1 are
px 1 vectors, the power function for the likelihood ratio test for cointegration depends on
a 1 and /3 1 only through the quantities f and g given by
f= a'b = /3jCa 1 < 0, (2.12)
and
l = a'ab'b - (a'bl
= (aja/a;Aa)-la;a l )(/3jCAC'/31) - (/3jCa l l , (2.13)
Note that the expressions above for C as well as a (a' Aa )-1 a' do not depend on the
.L .L.L .L
particular choice for a and /3. Note also that for p - r = 1 the power function
.L .L
depends only on f, since (2.9) above describes the process K t . The proof of the
Corollary is given in the appendix.
In order to interpret these results we consider the hypothesis HT for the simple
example we get from (1.3) by setting k = I, that is we consider the system
We can express the results as follows: The power for finding a stationary relation (31
with weights T-1 a 1 depends on the position of the vectors a 1 and (31 through the
angle between them (fig) and the area spanned by them (g), as well as the dimension
of the space in which we are looking for them (p-r).
3 Nu.merical investigation of the power function under the alternative of one extra
cointegration vector
For s = 1 the equations (2.9), (2.10) and (2.11) are of the form
t
- f f OKlu du + Kit = Bit, (3.1)
t
- g f OKlu du + K2t = B2t , (3.2)
K3t = B 3t . (3.3)
The distribution of the functional (2.7) is too complicated to find analytically but it
can be found by simulation. In order to simulate the processes we shall apply the
discrete version of these equations
(3.6)
We then easily solve the equations (3.4), (3.5) and (3.6) recursively and form the
Tx(p-r) matrix j\l with elements Mti = Kit. Then we calculate .0.M and M_ 1 , i.e. the
differences and lagged variables respectively and find the test statistic
Test = tr{ .0.M'M_1(M~l M_l)-lM~l.0.M}.
The number of observations T has to be chosen so large that the approximation of the
random walk to the Brownian motion is sufficiently good. We have chosen T = 400.
329
We find the results for stationary alternatives and p-r = 1,2, and 3 in Table I and
Figures 1 and 2 based. It is seen that, not surprisingly, the power decreases as the
dimension increases, i.e. if there are many dimensions to hide in. This means that it is
difficult to find the cointegrating vector, if it has a small loading. It was found that
the non-stationary alternatives (not shown) are readily picked up by the test with
large power. The test appears unbiased as f and g move away from O.
The table gives the limit distribution for the process defined by (3.1)-(3.3). If we
apply the results for the approximation defined by (3.4)-(3.6) one can for a large value
of T interpret the coefficient 1 + f/T = 1 + ai.81/T as the autoregressive parameter,
see (3.4), in the stationary (or near-integrated) relation we are trying to find. Hence
we can apply Table I as follows: Assume we have T = 100 observations and an
autoregressive coefficient of .79 say. Then the relation .79 = 1-21/100 = 1 + f/T
shows that we can use Table I with f = -21. Now the power of finding such a process
depends on the relation between the loadings and the cointegration vector. If a1 and
.81 are proportional, so that g = 0, then, if the number of non-stationary components
is p-r = I, we have a probability of .998 of rejecting the hypothesis of
non-stationarity, and hence of finding a stationary relation. If, however, the system
has p-r = 3 non-stationary components, then the probability of rejecting the
non-stationarity hypothesis is only .346. For a given angle, that is for fixed f/ g, it is
seen from the table that the larger the vectors the easier it is to find them.
Table I
The asymptotic power of the likelihood ratio test at 5% for r co integration vectors
among p variables under the local alternative of one extra cointegration vector. The
quantities f and 9 are defined in {2.12} and {2.13}. The number of simulations is 5000
for p-r = 1 and 2000 for p-r = 2,3, and T = 400.
p -r = 1
f 0 -3 -6 -9 -12 -15 -18 -21
.052 .141 .350 .620 .820 .945 .987 .998
g P -r = 2
g p-r=3
0 .054 .059 .073 .105 .136 .187 .266 .346
6 .510 .211 .143 .142 .188 .233 .305 .383
12 .834 .694 .536 .424 .407 .429 .476 .543
4 Appendix
We have here collected the proofs of Theorem 1, Theorem 2 and the Corollary. We
first define the variance and covariance matrices conditional on the lagged differences
by
330
Power
p-r
1.00
.as ~::::::--.:::::;.--
o -18 -21 f
Fig. 1:
The power function of the LikeLihood ratio test for cointegration, as a
function of f, see (2.12), for g = 0, see (2.13), and different vaLues of
the dimension p - r.
Power
100
0.67
0.33
-7
-11.
Fig. 2:
f e p~wer function for the LikeLihood ratio test for cointegration
Th as a
unctton of f and g, see (2.12) and (2.13) and p-r = 2.
331
~Ok,B = Cov(6.Xt,,B'Xt_kl6.Xt_F··,6.Xt_k+1)·
It follows from the representation (1.3) that, since £t is independent of the past values
of the process X t ' one can find the relations
~OO = a,B'~kk,Ba' + A, (4.1)
and
~Oki1 = ai1'~kk,B, ( 4.2)
which can be solved for a
a = ~Oki1(,B'~kk,B)-l. (4.3)
One can apply these identities to prove
( '
= a.!. a.!. ~ OOa.l.
)-1,
a.!.
-1 -1
= ~ 0 0 - ~ 0 Oa (a 'r; OOa
-1 )-1 -1
a 'r; 0 O·
PROOF. The second equality sign holds by the relation (4.1) and the two others can
be proved by multiplying by (a,Aa ) .
.I.
We now give the proof of Theorem 1. The result follows from the results in Phillips
(1988), and we here only sketch the calculations involved in the proof.
Now apply the decomposition (2.3) and use the result that
a'altJl'tJ
1. 1.
fotG U duo The second term converges to a'1J!tJ
1.1.
Gt .
It is easily checked that the solution of (2.5) is given by (2.6). In order to see that the
matrix a~ altJitJl. has rank s we have to apply the conditions (2.1) and (2.4) which
guarantee that the process X t is at most 1(1) both under Hand HT · Let b l =
tJl. (tJ~tJ)-ltJ~tJl and let b 2 be orthogonal to (tJ,b 1) such that (tJ,b 1,b 2) has full rank.
Then tJl. = (b 1 ,b 2) and (tJ,tJ1)1. = b 2. Similarly we define a l and a 2 from a1 and al.'
The condition (2.1) can now be expressed as
[ a i 1J!b 1 ai1J!b2]
rank a' 1J!b a' 1J!b = p-r,
2 1 2 2
whereas the condition (2.4) can be expressed as
2
rank(a 1J!b 2) = p-r-s.
These conditions clearly imply that
2 2
ai 1J!b 1 - ai 1J!b 2 ( a 1J!b 2)-1 a 1J!b l
is offull rank S. This therefore also holds for ai1J!b l and hence for a l and b l . Now
Next we turn to the proof of Theorem 2. This follows closely the proof given in
Johansen (1988b) for the asymptotic distribution under H, and it will therefore not be
given in full technical detail. The likelihood ratio test is calculated from (1.5) solving
the eigenvalue problem (1.4). The basic idea in the proof is to study what happens as
T -I rn in equation (1.4) under different normalizations, and then apply the continuity of
the ordered eigenvalues as a function of the coefficients.
1
First multiply the matrix in equation (1.4) by (/3,T-'/31. ), and its transposed
Theorem 1, /3'.L Skk/3.L IT converges weakly to /3'.I. /3.I. fGG'd/3'.I. /3.I. = 4>, say, and this shows
that in the limit). has to satisfy the equation
0
I ),/3'E kk/3 - /3'E ko E 6 E ok/3II),4> I = 0,
which has r positive solutions and p-r null solutions. Thus the r largest eigenvalues of
(1.4) converge to those of
Now normalize ).. by T and define p. = T)'.. We multiply the matrix in (1.4) by
1 1 1
(/3,/3 ), and (/3,/3 ), and let T .... ill. We then get that the limiting value of p will satisfy
.I. .I.
the equation
0
I/3'EkOEoOIEok/311 p4> - F'(E 6 - Eo6EOk/3(/3'EkoEOOIEOkf3)-1 /3' Eko E o6)F 1=0.
Since a = E ok /3(/3'E kk /3)-l one can express the matrix between F' and F as
-1 -1 ( -1 )-1 -1
EOO - EOOa a'E OO a a'E OO '
which by Lemma 1 equals
a (a' Aa )-l a , .
.I. .1..1. .I.
/3'.I. /3.I. (jGG'du /3'.I. /31 a'1 a .I. + fG(dW)') = /3'.I. /3.I. fG(dG)'/3'. 1Il!'a
..1.
.
Inserting this result into (4.7) we find that the limiting values of p are the roots of the
334
equation
1(,8',8
.L.L
){pjGG'du-jG(dG)',8'w'a
.L
(a'Aa )-l a.,w,8
.L.L.L L.L
j(dG)G'}(,8',8)1 =0 .
.L.L
Finally we ~ive the proof of the Corollary. This result follows from the equations (2.9),
(2.10) and (2.11), and the expressions for the coefficients f and g are easily deduced by
applying the expressions for a and b. We find
f= b'a = ,8',8 (a'w,8 )-l a 'a = ,8'Ca
1.L .L.L .L 1 1 I'
which is non-zero by the result of Theorem 1. Since we are interested in stationary
alternatives only, it follows from (2.9) that b'a < o. Furthermore
a'a = a'a (a'Aa )-l a 'a
1.L.L.L .L1
and
5 Acknowledgement
The author thanks Neil Ericsson for some interesting discussions on the topic at an
early stage of the work, and a referee for suggestions on how to improve the
presentation.
References
Engle, R.F. and Granger, C.W.J. (1987): "Cointegration and error correction repre-
sentation, estimation and testing", Econometrica, Vo1.55, pp. 251-276.
Garbers, H. (1989): "The relationship between spot- and forward exchange rates and
the relative money supply: the Swiss franc case", Technical Report, University of
Zurich.
Granger, C.W.J. (1981) : "Some properties of times series and their use in econometric
model specification", Journal of Econometrics, Vol. 16, pp. 121-130.
Hoffman, D.L. and Rasche, R. (1989): "Long run income and interest elasticities of
money demand in the United States", Technical Report, Arizona State Univer-
sity.
Johansen, S. (19SSa): "The mathematical structure of error correction models",
Contemporary lvlathematics, Vol. SO, pp. 359-386.
Johansen, S. (1988b): "Statistical Analysis of Cointegration", Journal of Economic
Dynamics and Control, Vol. 12, pp. 231-254.
Johansen, S. (1991): "Estimation and hypothesis testing of cointegration vectors in
Gaussian autoregressive models", to appear in Econometrica.
Johansen, S. and Juselius, K. (1990): "Maximum likelihood estimation and inference
on cointegration - with applications to the demand for money", Oxford Bulletin
of Economics and Statistics, Vol. 52, pp. 169-210.
Juselius, K. (1991): "Long-run relations in a well defined statistical model for the
DGP. Cointegration analysis of the interest parity relations", in this volume.
Kunst, R.l\I. (1988): "Cointegration in a l\IacTO-€conomic system: sensitivity and
335
Katarina Juselius
Institute of Economics
University of Copenhagen
Studiestrade 6
DK 1455 Copenhagen K
Denmark
Abstract
The concept of a well-defined statistical model for the data generating process is given an
empirical formulation in the vector autoregressive model under the assumption of
cointegration in an analysis of prices, interest rates and exchange rates between Denmark
and Germany. The long-run relations are estimated as stationary linear combinations
between the non-stationary variables. Comparative analyses of the long-run relations
based on the single equation error-correction model as well as the Engle-Granger two-step
procedure are performed. The full system versus the partial system analysis approach is
discussed in terms of optimal inference on the long-run parameters. By testing hypotheses
about the weight coefficients it is empirically demonstrated that the full five-dimensional
system cannot be reduced without loosing some efficiency.
Keywords: Non-stationary time-series; cointegration; long-run relations; VAR models;
purchasing power parity; uncovered interest rate parity.
Contents
1 Introduction
2 The dynamic linear regression model
3 The maximum likelihood estimation of the general cointegration model
4 The long-run purchasing power parity and the uncovered interest rate parity
relation between Denmark and Germany
5 Comparison with other methods
6 Concluding remarks
Appendices
References
1 Introduction
It is a well-known fact that economic time series are typically non-stationary. Never-
theless, the effect of non-stationarity on model specification and estimation had not
received much attention among econometricians before the discussion of cointegration in
Granger (1983). Most empirical models based on time-series data are of the following
categories:
i) Models using differences instead of levels. Since economic variables usually are
integrated of order one the differences do not have unit roots and standard distributional
results apply to the model estimates. However, all long-run information is removed by
differencing unless the difference operator is applied to the error process as well. When
337
short-run effects are equal to long-run effects, this procedure gives satisfactory results, but
empirical evidence seems to indicate that these are often fundamentally
different (see for instance Juselius, 1989).
ii) Partial adjustment and rational expectations models. These models assume a
long-run steady-state relation between the levels of the variables and a simple adjustment
mechanism when the process is out of equilibrium. Thus the long-run information in the
data is preserved and some dynamics are allowed for, but it is often too restrictive for the
model to be a satisfactory statistical description of the complicated data generation process
of economic time-series.
iii) Error-correction-mechanism (ecm) models. These models which originate from
Sargan (1964) and were further developed in Hendry (1987), Davidson et al. (1978),
explicitly postulate the possibility of different short-run and long-run dynamics. The
long-run information is accounted for by including so called error-correction mechanisms
(equilibrium errors) defined as Yt-Yt* where Yt* = f(x t ) is an assumed steady-state relation
between the levels of y and the levels of the explanatory variables in the model. The
short-run information is described by differences of the explanatory variables, the dynamic
specification of which is usually data determined. Most empirical ecm-models have been
estimated as single equation models. The concept can easily be extended to the
multivariate case, but the estimation becomes more complicated due to non-linear
cross~quation restrictions. An example of a multivariate ecm-model is given in Davidson
(1983).
The mathematical formulation of the multivariate error-correction model was first given
by Granger (1983). In this paper, he defined the concept of cointegration and demonstrated
the connection with error-correction mechanisms using the moving average representation
of a vector time series model. In Johansen (1988a) the results were generalized to the
autoregressive representation which greatly facilitated estimation. The statistical analysis
then developed along two lines, on one hand the full information maximum likelihood
procedure developed in Johansen (1988b), Johansen (1991a) and Johansen & Juselius
(1990) and on the other hand the two step procedure in Engle & Granger (1987), Stock
(1987), Phillips (1987), Phillips & Durlauf (1986), Park & Phillips (1988) and Park &
Phillips (1989). In t.he latter procedure the cointegration relation is first estimated by a
static regression analysis, the estimates of which have been shown to be super consistent
(Stock, 1987) when the variables are non-stationary. In the second step the residuals from
the first step are tested for stationarity and, if accepted, included as ecm-mechanisms in
the final model analysis. Thus the final model is similar to the traditional ecm-model but
estimation and testing of the steady-state relations are di fferen t.
In this paper we analyze the hypothetical long-run relations in a simultaneous model
approach by applying the Johansen method of multivariate cointegration. The main
difference to the methods discussed above is that the long-run relations are estimated
jointly with the short-run effects thus using all information in the data. The concepts are
illustrated by an empirical analysis of the uncovered interest rate parity versus the
purchasing power parity relation between Denmark and Germany. Graphs of the data are
given in Appendix A.
In section 2 we attempt to demonstrate empirically the close connection between the vector
autoregressive model and the concept of a well-defined statistical model for the data
generating process (DGP) following mainly Hendry & Richard (1983) and Spanos (1986),
and then show that non-stationarity can easily be incorporated in this general framework
such that the presence of cointegration vectors can be statistically estimated and tested. In
section 3 some of the basic concepts and some important results on estimation and
inference on maximum likelihood cointegration are discussed. In section 4 the estimated
cOintegration vectors and their corresponding weights are presented. The adjustment
338
coefficients in the full simultaneous model framework are analyzed to gain insight into how
deviations from equilibrium states interact with the rest of the system. Comparative
analyses of the long-run relations based on the single equation ecm-model as well as the
Engle-Granger two-step procedure are presented in section 5 and the relative importance
of a full system analysis is discussed.
In this section some important principles developed by the "British school" of econometric
methodology (see for instance Hendry & Richard, 1983, Hendry, 1987 and Spanos,1986) are
discussed in the context of the dynamic linear regression model. This school explicitly
stresses the distinction between a theory model and an estimable empirical model on one
hand, and the reality that has generated the data and the statistical generating model on
the other. The statistical properties of the latter are then used to investigate the appro-
priateness of the former.
When the data are realizations of time series neither the sampling model nor the proba-
bility model for the actual data is usually known to the investigator. This means that many
economic hypotheses may not be testable or possibly only indirectly testable based on the
available data set. On the other hand, if a general well-defined statistical model can be
found that describes the data satisfactorily, this can be used as a framework in which the
economically interesting questions can be asked and statistically tested. Moving from "the
general to the specific" will ensure that the restrictions imposed on the final model are
consistent with the information in the data. This is in contrast to the experimental design
case where the sampling and probability model are designed first and then the data
collected accordingly, such that the theoretical questions can be answered in some optimal
way.
In the following, we will analyze a general vector autoregressive (V AR) model and
demonstrate that it can be reparameterized such that questions about short-run and
long-run economic phenomena can be given a statistical answer. This is made possible by
exploiting the information inherent in the stationary and the non-stationary part of the
process using the concept of cointegration.
Consider the p-dimensional vector Zt = [zlt, ... ,Zpt l ' where the elements Zit' t = 1, ... ,T, are
realizations of stochastic variables. Let us further assume that Zt and Zt_i are time
dependent. The joint probability of all the sample data has then to be given conditional on
the initial values, i.e.: D(Zt IZO;8) or in the form of the factorized density function:
T
II D(zt IZt_l' ZO;O),
t=l
Let us further assume that the conditional process {Zt:Zt_l;O} N NIID(ll't'~) and that Il t
has an autoregressive representation of order k, i.e.
where Al'" .. ,A k are pxp matrices of autoregressive parameters. Under these assumptions
339
we can now define a statistical model for the data generating process:
k
Zt = E A.z t _· + Ct· (2.1)
i=l 1 1
In an economic sense equation system (2.1) corresponds to the so-called reduced form and
can be efficiently estimated by the ordinary least squares method applied to each equation
in turn. This means that estimation and testing are straightforward using any computer
package. If the residuals <t pass misspecification tests for normality, independence and
homoscedasticity it is concluded that the statistical model (2.1) is an approximate
description of the data generating process.
Up to this point we have only assumed that the conditional process is stationary but not
considered explicitly that the data are non-stationary. We will now discuss the statistical
implications of Zt being a non-stationary process and begin with a formal definition of the
order of integration and cointegration (see for instance Johansen, 19S5a) .
Thus the notion of cointegration implies that certain linear combinations of the compo-
nents of the vector process are integrated of lower order than the process itself. The impli-
cation of cointegration is that the non-stationarity in one variable corresponds to the
non-stationarity in a companion variable such that a linear combination of them becomes
stationary. Another way of expressing this is that when two or several variables have
common stochastic (or deterministic) trends they will tend to move together in the
long-run, since they are driven by the same persistent shocks. Such a linear combination
!3'Zt is of considerable interest for the decision maker since it can often be interpreted as an
adjustment process for a long-run economic relation. This makes the association with the
error-correcting mechanisms straightforward.
340
To begin with, we will demonstrate the association between the data generating mechanism
as given in (2.1) and the general cointegration model. For notational simplicity we assume
that k = 2 and add an intercept to each equation of the model. The implications of a
constant term in the model will be discussed subsequently in section 3. Using the lag
operator L, where Liz t = Zt_i the model is formulated as:
,
2
(I - AIL - A2L )Zt = /.L + ct ' ft ,.., NIIDp(O,E). (2.2)
Assuming that Zt is 1(1) and noting that L = l-(l-L) = 1-t:.., where t:.. is the difference
operator, (2.2) can be rewritten as
(2.3)
The system of equations (2.3) is a reparameterized version of (2.2) which is now more
appropriate for the analysis at a first order non-stationary process. Note that there is no
loss of long-run information in (2.3), a phenomenon usually associated with models for
differenced data. All long-run information is contained in the levels component IIz t _ 2.
Since Zt is 1(1) and f t is stationary it follows that the matrix II cannot be offull rank, since
this would mean inconsistent interpretation of the model. On the other hand if II = 0 the
model is statistically consistent but there would be no long-run information in the data,
and (2.3) would be reduced to the standard V AR model in first differences. Finally when
the rank is greater than zero, say r, the model implies the existence of r linear cointegration
vectors, defining the number of long-run relations in the data. If these can be given an
economic interpretation they can be considered the equivalents of the error-correcting
mechanisms.
In the next section model (2.3) will be discussed and estimated under the assumption of
reduced rank.
To simplify the derivation of the ML estimates for a and (3, all short-run dynamics are
conditioned out by regressing t:..Zt and Zt-2 respectively on t:..Zt_l and the constant term /.L,
giving rise two (px1) vectors of residuals, the "difference" residuals called ROt and the
II levels II residuals called R
2t . Model (3.1) can now be written as:
(3.2)
341
The non-linear estimation problem is solved by first keeping /3 fixed, which gives the esti-
. -1 .
mate a(/3) = 5 02 /3(/3'5 22 /3) and the residuals sum of squares function ~(/3) =
SOO-a(/3)(/3'S22/3)a(/3)', where 500 , 522 and 5 20 are defined as the product moment
matrices based on the residuals ROt and R 2t . Minimizing the determinant of this function
for /3 gives the maximum likelihood estimator. It has been shown that the estimates are
found by first solving the eigenvalue problem:
(3.3)
which gives the eigenvalues Al > ... > Ap ~nd the corresponding matrices of eigenvectors V
= (v 1 ,... ,v p ) normalized such that V'S22V = L For further details, se.e Johansen (1988b),
Johans,en (1991a) and Johansen & Juselius (1990). The eigenvalues \ correspond to the
squared canonical correlations between the ~'levels" residuals and the "difference" !esiduals
as defined above. The r first eigenvectors v·1 determine the linear combinations v!Zt
1
(i =
1,,,.,r) which are most correlated with the conditional process {tlZtl tlZt_i' i = 1,,,.,m-l}.
Thus the magnitude of A.1 is a measure of how strongly the cointegration relation v!Zt
1
is
correlated with a linear combination of the stationary process tlZt. The last (p-r)
(3.5)
Note that the parameters a and /3 are not identified in the sense that given any choice of a
and /3, then for any non-singular matrix ~(rxr), the choice a~ and /3(0-1 will give the
same matrix IT = a/3', and hence determine the same probability distribution for the varia-
bles (Johansen & Juselius, 1990). It is, however, not unusual that the /3. vectors, deter-
I
mined by the normalization /3'S22/3 = I, can be interpreted directly in terms of interesting
economic relationships. In these cases the estimated /3/Zt are the relations of interest and
not linear combinations of them, but to make interpretation easier it is often
342
The exposition above was based on the assumption of known cointegration rank. In
practical situations the rank r is not known but has to be inferred from the data. The
corresponding statistical problem is to discriminate between non-zero and zero eigenvalues
and the likelihood solution is given by the likelihood ratio test procedure, two versions of
which are presented below:
The first is a test procedure for the hypothesis r ~ q against the general unrestricted model
r = p and is given by:
p
-2In(Q;qlp)=-T ~ In(l-'x.) (3.6)
. =q+ 1
I
I
The second test procedure is for testing the hypothesis r < q against the previous
hypothesis r ~ q+1:
,
These tests will be described in more detail in connection with the empirical application.
For a mathematical derivation, see Johansen (19SSb).
3.3. Two interesting cases in a hypothetical example.
Below we shall investigate in some detail two cases where we first assume one cointegrating
vector and then two cOintegrating vectors in order to illustrate the potential use of this new
concept in econometric modelling. In both cases we will assume that Zt is a four-dimensio-
nal vector described by the AR(2) model:
It would often be natural to normalize by the coefficient of one of the variables, say zl' by
multiplying 13 by 1/131 and a by 131 to facilitate the interpretation of the cointegrating
343
*
vector as the long-run equilibrium error, zlt - zlt:
If we denote :
.811 z1 + .821 z2 + .831 z 3 + .841 z4 ci 1
and
then
In this case we have two cointegrating relations (cil and ci 2) and the interpretation of
these need not always be straightforward, due to the non-uniqueness discussed above. For
instance, assume that {.811 = .82l f 0 and .831 = .841 = O} in the first relation, and {.812 =
.822 = 0 , .832 = .842 f O} in the second. This would imply that zl and z2 are cointegrated,
i.e. that they have a common stochastic trend, and also that z3 and z4 are cointegrated,
possibly with a different common trend. In most applications the structure is not so
clear-cut as in this hypothetical example; instead, all f]. . will typically differ from zero. By
IJ
taking linear combinations of the estimated vectors structurally simpler cointegration
vectors could be found.
Cointegration analysis has frequently been concerned with the analysis of the ,B-vectors
without paying much attention to the corresponding a-coefficients. However, the analysis
of the O!ij-coefficients provides additional insight into the complicated functioning of the
economic system which can be of great importance. Consider for instance the first two rows
of O! in case 2 and assume that ci 1 and ci 2 describe a long-run demand and supply relation,
respectively, such that O! and .8 can be assumed to be well-defined. Assume then that {O!ll
f 0, 0!12 =O} and {0!21 =0, 0!22fO}. This would imply that the DGP of zl has been affected
344
by the equilibrium error ci 1 but not by ci 2, whereas the opposite is true for the DGP of z2.
Clearly, such a result would be stron~ evidence that a demand relation and a supply
relation have actually been identified. (An illustration of this is given in Juselius, 1989).
Thus the a· .--<:oefficients are the weights of each cointegration relation in each of the four
IJ
equations of the system. Note that some of the coefficients might be statistically significant
from zero whereas others will not. From an economic point of view these hypotheses are
often of considerable interest, as will be shown later.
The inclusion of a constant term in the model turns out to have strong implications both
on the probability theory (Johansen (1991b)) and the economic interpretation (Johansen
and Juselius (1990)). This is a consequence of the dual role the constant plays in a model
with stationary and non-stationary components, namely as the intercept in a long-run
relation in levels and as the slope of possible linear trends in the non-stationary part of the
model. This can be stated more generally by expressing fL as the sum of the a and a
.L
projection:
fL = a{30 + a.L '0' (3.9)
where {30 = (a'a)-l a 'fL is a vector of (rx1) intercepts, '0 = (a .L.L
'a )-l a .L 'fL is a «(p-r)x1)
vector of linear slope parameters and a is a px(p-r) matrix of full rank, the columns of
.L
which are orthogonal to the columns of a, Johansen (1991a) and Johansen & Juselius
(1990).
The hypothesis that there are no linear deterministic trends in the data can be tested by
applying the likelihood ratio test procedure described in Johansen & Juselius (1990). The
procedure consists of comparing the restricted model
~Zt = r 1~Zt_1 + a{3' *Zt_2
* + tt' (3.10)
where {3*= *
[{3',{301',
Zt= [zt',11' and {30 is from (3.9), with the unrestricted model (3.1).
Model (3.10) is chosen if the restriction a .L 'fL = 0 is accepted. However, it should be pointed
out that the distributions of the likelihood ratio test, (3.6) and (3.7) respectively, are not
the same under these alternatives. In fact three different distributions are needed, one for
model (3.10) and two for model (3.1). All three are non-standard but their asymptotic
distributions have been estimated by simulation and are reported in Johansen & Juselius
(1990) as Table D.1, Table D.2 and Table D.3. The latter table corresponds to model (3.10)
whereas Tables D.1 and D.2 correspond to model (3.1), i.e. the model where we allow for
the possibility of a linear trend. In this model we are in the unfortunate case of having two
limiting distributions under the null: Table D.1 corresponds to the case where deterministic
trends are correctly included in the model, whereas Table D.2 corresponds to the case when
deterministic trends are allowed for but there are actually no trends in the data. Since one
does not usually know which case is true it is advisable to use Table D.2 since it has fatter
tails, or alternatively simulate a distribution which corresponds as closely as possible to the
actual problem being investigated.
4 The long-run purchasing power parity and the uncovered interest parity relation
between Denmark and Germany
In this section we will investigate two popular theories for the international transmission
effects between two trading countries. The first is based on the purchasing power parity
(PPP) relation, Pd = Pg + ed / g , and the second on the uncovered interest rate parity
345
(UIP) relation, id = ig + 6ee , where Pd is the natural logarithm of the Danish consumer
price index, id is the Danish bond rate, p and i are the corresponding German variables,
g g
ed / g is the logarithm of the exchange rate measured as Dkr/Dmk and the superscript e
denotes the expected exchange rate. For a general review of theories and evidence in the
foreign exchange market see Baillie and McMahon (1989).
The link between the two relations is obtained by assuming that 6e~+1= (ed/g-Pd+Pg)t.
Given that expectations are unbiased this would lead to E(id-i g-€d/g +Pd-Pg)t = 0 where
E(x t ) denotes the mean value of the process x t . The two parity concepts are fundamentally
different. The PPP is a long-run relation and the adjustment towards it can be expected to
be backward looking and probably very slow, if at all present, whereas the UIP is basically
a forward looking market clearing relation. The interaction between them is clearly an
interesting empirical question for which the full information maximum likelihood
cOintegration methods should be ideal to provide an answer.
We will here give an illustration of the full information maximum likelihood method
applied to the vector zi
= [Pd,Pg,ed/g,id,iglt, t =1973.1-1987.3. The sample period
coincides with the post Bretton Woods period, which for Denmark has been characterized
by pegged exchange rates within certain ranges specified by the snake and later on by the
EMS. Graphs of the data are given in Appendix A.
The first step of the empirical analysis is concerned wi th the unrestricted estimation of the
long-run parameters in the full cointegration model. If the data support the existence of
long-run relations and these can be interpreted in terms of the purchasing power parity
relation and/or the uncovered interest rate parity relation, the next step is to analyze the
effect of these disequilibria by testing restrictions on the matrix a of adjustment
coefficients. This is likely to give valuable information about the complicated interactions
of the system. The ML method estimates model (2.3) under the restriction that II=a,B'.
Thus the estimation is non-linear in the parameters and requires an eigenvalue routine
and a computer program that can handle matrices. A program written in RATS is
available and can be obtained from the author on request. The estimated coefficients of
model (2.4) are given in Appendix B.
In the upper part of Appendix B we have reported the estimates of the unrestricted
VAR-model and in the lower part some misspecification test statistics. The independence
assumption is tested by the Box-Pierce Q-statistic which is asymptotically distributed as
X2(12). As a complement to this the individual residual autocorrelations are reported. It
can be noticed that the Box-Pierce test statistic is quite high for the first equation,
whereas for the other equations it is more satisfactory. There also seems to be excess kurto-
sis both in the Danish price equation and the German interest rate equation. This could
indicate either that the second order autoregressive model is not sufficiently general to
describe all the systematic variation in the data or that the information set we have used is
not sufficiently large. The estimated coefficients of the short-run dynamics r l' also
reported in Appendix B, are not of particular interest in the context of this empirical
investigation. They correspond to the reduced form of the system and, due to the general
model formulation, are not effiCiently estimated. The simultaneous structure of this system
is described by the error covariance matrix. To facilitate interpretation, it is given in
correlation form in Appendix B.
In the following we will concentrate on the investigation of the long-run impact matrix II
given in Appendix B. The rank is determined by the two test procedures described in
346
section 3. The first is based on the test statistics '\race (3.6) and the second on Amax
(3.7). The following six hypotheses are being tested by each test:
i.e. each hypothesis H. is tested against the unrestricted general model H5. The A test
1 max
is performed conditional on the previous model, i.e.:
There is no obvious rule for which of the test statistics to choose. The trace statistics will
usu~y have more power to detect the r-th stationary relation when several of the estima-
ted A., i=r+l, ... ,p, are quite close to being significant, whereas A has more power when
1 max
the estimated \ are either large or close to zero. Consequently Amax is usually preferable
when the estimates discriminate effecti::ely between small and large eigenvalues, whereas
Atrace is preferable when the estimated Ai values are rather evenly distributed.
In Table 4.1, the eigenvalues and both of the test statistics are given for models (3.1) which
allows for deterministic trends in the non-stationary part of the model. The choice of
model (3.1) instead of (3.10) is motivated by the obvious linear trend in the Danish price
variable.
Table 4.1 The likelihood ratio tests for the rank of IT in the model with linear trends.
Critical values are from Table D.2 in Johansen & Juselius (1990).
A. -Tln(l-A.) A . 90 A . 95 -~Tln(I-A.) A· 9O A· 95
1 1 max max 1 trace trace
1 .469 36.13 30.81 33.26 78.12 65.96 69.98
2 .229 14.80 24.92 27.34 41.99 45.25 48.41
3 .224 14.44 18.96 21.28 27.19 28.44 31.26
4 .127 7.73 12.78 14.60 12.75 15.58 17.84
5 .084 5.02 6.61 8.08 5.02 6.69 8.08
Based on the test results of Table 4.1. it would be natural to accept r = 2, but the third
eigenvector (see Table 4.2 below) seems to measure exactly what we are looking for,
347
namely the PPP and the interest differential. Therefore it does not seem reasonable to omit
the third cointegration vector from further analysis and we choose r = 3 although the
p-value is above 0.10. In this context it can be pointed out that the power of the tests is
quite poor for small values of \' Johansen (l991b), which in some situations can motivate
the choice of larger significance values than the standard ones. In this sense the decision to
include the r-th cointegration vector in the stationary set is a question of whether it can be
interpreted as a reasonable economic relation or not.
,
The decomposition of II into eigenvectors and weights, based on model (3.1) are given in
Table 4.2. The three stationary vectors are denoted by {J. and the two non-stationary ones
1
by v1.. Correspondingly the first three weights vectors are denoted by a·1 and the last two by
Wi· In Appendix C, the graphs of the ;1i'Zt relations are given. The first three ;1j'zt relations
determine the stationary relations, whereas the last two determine the common trends.
The e~genvectors:,
The w,eights:
a1 a2 a3 w4 w5
We will now try to give an interpretation of the first three /1-vectors. It might seem
surprising that we find strong evidence on the PP P relation in the third vector and not as
one might expect in the first one. However, if the adjustment to the long-run steady-sta!e
is very slow, one should expect the corresponding cointegration relation to have a small A.
I
value. Since the first three coefficients of ;13 seem to determine the PPP relation quite
accurately we have normalized ;13 by the Danish price level. The last two coefficients seem
!o determine the nominal interest rate differential between Denmark and GermanJ- The
;12-vector seems primarily to be a relation for the interest rate differential whereas ;11 does
not seem to have an intuitive interpretation. It will be shown b,elow that the role of the
first one will become clear after the analysis of the corresponding a-vector.
348
We will now demonstrate how the investigation of the ai-vectors facilitates a full
understanding of the complicated functioning of the system, by evaluating the relative
A
importance of each a ij for i=l, ... ,5 and j=l,2,3. This is done by testing the
hypotheses a·1. =0 for i= 1, ... ,5, which gives us the relative importance of each
cointegrating vector rvzin each equation. The test results are presented in Table 4.3
together with the new estimates of /3.1 under the restriction that a·1. =0.
Since the third cointegration vector, /33'z, can be considered to be the most interesting from
an economic point of view we will look at the third column of test values in table 4.3. They
indicate that the third cointegration vector is most important for the exchange rate
equation, has some significance for both interest rate equations, but has no significance for
the German price equation and only a minor impact on the Danish price equation. This
result tells us that the information about the PPP is mainly in the exchange rate equation
and that prices hardly adapt at all, an empirical result which is consistent with the notion
of sticky nominal prices.
A A
-Tln(1-A 1) -Tln(1-A 2)
Full model 36.13 14.80
i
/31 {32 /33 {31 {32 /33 /31 /32 /33 /31 /32 {33 /31 /32 /33
.1 -1.1 .-.3 -.2 -1.0 .0 .0 -.5 2.8 -.0 -1.2 -.8 .0 -1.0 -.0
-.3 1.4 -.1 .4 1.0 -.0 -.0 -.0 -4.9 -.0 1.5 .6 -.1 1.0 -.0
-.1 1.0 .7 -.1 1.0 -.1 -.1 1.0 -.5 -.0 1.0 1.0 -.1 1.0 .1
.1 -.6 1.0 .1 .5 1.0 .4 2.5 1.0 .4 -.9 1.0 .6 1.0 1.0
1.0 .6 -.6 1.0 -.3 -.4 1.0 -2.6 1.2 1.0 .9 -1.6 1.0 -.2 -.9
349
The first cointegration vector is strongly significant in the German price equation as well
as in the Danish price equation. It also has some impact on the German interest rate. This
can actually provide us with an interpretation of the first cointegration vector: namely that
it measures the positive association between the inflation rate t:.Pt and the nominal interest
rate.
It is interesting to notice the effect on /32 and /33' when the information from the DGP of
the German price ~nd the Ge~man interest rate equations, respectively, is suppressed.
Firstly, notice that /33 becomes /32 and vice versa. This can econometrically be explained by
the fact that the estimates of ).2 and ).3 are very close to each other and therefore the
ordering is non-informative. Secondly, and more interestingly, the relation between devia-
tions in the PPP and the interest rate differential seems to be very sensitive to these impo-
sed restrictions.
This illustrates how sensitive the cointegration results can be for small changes in the
information set. One way of circumventing this indeterminacy is to specify the structural
hypotheses of interest and then test whether they are contained in the space spanned by (3.
These test procedures are described in Johansen & Juselius (1990).
As a final step in the cointegration analysis the estimates of IT = a/3' for r = 3 are calcu-
lated and given in Table 4.4. The rows indicate the combined effect of the three cointegra-
ting relations on each of the Zit processes. The empirical results can be summarized for:
1) the Danish price equation: No effect from the PPP, positive effects from the
German and the Danish interest rates.
2) the German price equation: No effect from the PPP, some positive effects from the
interest rates (own country interest rate).
3) the exchange rate equation: Strong PPP effects, no effect from the Danish interest
rate, some effects from the German interest rates.
4) the Danish interest rate equation: Indication of a long-run real interest rate
relation.
5) the German interest rate equation: No effects except from own interest rate.
The most widely used method for estimating long-run relations is the Engle-Granger
two-step procedure (Engle & Granger,l987). The first step is simply to estimate the coeffi-
cients of the suggested long-run relation with a simple linear regression model:
350
(5.1)
where Yt' xlt,··,x kt are one-dimensional time-series and Dit are seasonal dummies. The
second step is then to test the hypothesis that the residuals ut are a stationary process. If
the stationarity hypothesis is accepted the estimated linear relation is assumed to be an
estimate of the long-run steady-state relation. In the second step the full dynamic model
is specified so that the changes in the variable y t are explained by present and lagged
~hanges of the explanatory variables and by the estimated error-1:orrecting mechanism
u t- 1. For instance:
The second method estimates the steady-state relation as the static long-run solution of a
dynamic linear regression model, given for instance by:
where ct is a white noise process. The long-run steady-state relation is then derived as:
*
y = fO + f1 x 1 + f2 x 2 + ... + fkxk' (5.4)
where fi = -1:/c, i = 1, ... ,k. The estimates in (5.4) are shown to be equivalent to the
maximum likelihood estimates of model (3.1) for the special case when there is just one
cointegration vector and just one a i f- 0 , i.e. the one corresponding to the equation for y
(Johansen & Juselius, 1990).
To allow for full comparability, each of the models is estimated with Pd' e d / g and id
respectively as regressands. The estimates are presented in Table 5.1, where a regress and is
indicated by -l.0, and a coefficient with a significant t-value (> 121) is indicated by an *.
For each model the Durbin-Watson and the Box-Pierce test statistic is calculated as a
measure of residual autocorrelation. The D-W test statistic is used to test the null hypo-
thesis that ut in (5.1) has a unit root, whereas the Wald test statistic is used to test the
overall significance of the long-run steady-state relation (5.4). Based on the reported
D-W statistics the estimated long-run relations would be considered stationary though
most of them are close to non-stationarity. The striking exception is the exchange rate
equation of model (5.3).
It is, however, not straightforward to compare the results from a single equation analysis
with the results of a system analysis. We have chosen to compare the results in Table 5.1
with the corresponding estimates of II = af3' given in Table 4.4, since we believe that the
combined estimates are likely to correspond most closely to the single equation estimates.
351
Table 5.1 Estimation of the long-run real interest parity relation by the static regression
model (5.1) and the dynamic linear regression model (5.3)
The static linear regression estimates
Pd Pg ed / g id i DW Q(21)
g
-1.0 1.75* .44* .08 -.18 .25 264.11
.70 -.48 -1.0 -.20 .45* .34 122.70
.10 -.10 -.15 -1.0 .72* .18 180.40
The dynamic linear regression estimates
Pd Pg ed / g id i DW Q(21) Wald(8)
g
-1.0 1.68* .23 .96 .75 2.08 40.92 853
.94 -1.00 -1.0 .54 .04 2.10 6.39 48250
.11 .05 .29 -1.0 .45 2.15 19.95 30
The Danish price equation: Both the static and dynamic regression methods find cointegra-
tion between prices and the exchange rate, though the coefficients are not consistent with
the PPP-hypothesis. In the static regression case the interest rates enter with very small
and insignificant coefficients, whereas in the dynamic regression case the interest rates
seem to have an effect similar to the one found in the ML case.
The exchange rate equation: In the dynamic regression case the PPP seems to have the
expected coefficients, whereas the result is less satisfactory in the static regression case.
The interest rate equation: In the static regression case the relation between nominal
interest rates dominates though the coefficient of the German interest is quite different
form the ML estimate. In addition the PPP-relation comes out very nicely but with small
and insignificant coefficients. In the dynamic regression case the estimates are completely
different from the ML estimates.
6 Concluding remarks
In this paper we have demonstrated empirically the potential use of cointegration analysis
based on a second order VAR model. Based on the assumption that the VAR model is a
satisfactory statistical description of the non-stationary vector process Zt it is possible to
derive procedures for likelihood inference on long-run relations. This is done under the
assumption of reduced rank r<p of the long-run impact matrix II which then can be parti-
tioned into the (pxr) matrices a and {J with IT = a{J'. These matrices summarize all data
information about the long-run relations of interest and a variety of structural hypotheses
can be tested against the general model formulation.
The full system maximum likelihood approach is compared to two single equation proce-
dures. We conclude by noting that when there is more than one cointegrating relation in
the data any single equation method would usually be unsatisfactory since one would not
be able to discriminate between the different relations by such a method. On the other
hand, one would usually expect the economic structure to be quite complicated in those
cases when the data suggest several long-run relations. This clearly motivates the use of all
information in the data as effiCiently as possi ble.
352
LEVELS DIFFERENCES
DANISH PRICE
I.-r------------------------------------__
.~
..
•• 4-----------------~~----------------_1 l.
-. z
I.
-.1
_ I I o.4------4--------------------------~
-.1 ·1
13 1. 15 11 11 IS n 10 II IZ ., s. 15 U 81 73 1~ 15 11 11 IS 1~. ao. S1 52 n a. 85 18 81
GERMAN PRICE
·,
1.-r----------------------------~--~~_,
z '-r--~------------------------------~
, .,
••
• .3
13 1~ 15 11 7T 18 71 eo II IZ 13 S. l5 el 81
13 n 15 71 17 I. n. 10 II 12 13 I_ 15 9S 87
EXCHANGE RATES
I·°lr--------------------______________-, 7· l r----------------------__________
5. ,
s .• s.
I .7
l.
5 .1
I.
5.5
""
-I .
73 T. 15 71 11 18 l' eo II IZ 13 8~ 85 sa 87
353
APPENDIX A. cont.
LEVELS DIFFERENCES
INTEREST RATE IN DENMARK
~~.,--------------------------------------, 1.,----------------------------------,
20. z.
II.
II.
O.-r~_+~+_~~~~~~~~~~~--+_~~
I~ .
·1.
12 .
·a.
10.
.J .
. ~
71 15 TS 77 7S I'. 10 51 !Z Il 55 'I 81
7~ 8~
71 1~ 15 TJ 17 re l' eo 51 ez 11 8~ 85 n 81
Il.-r--------------------~--------------, ,,------------------------------------,
II.
2 .
..
• L
1.
1·~r-_r_r~--~~_r~--r_~_r_r_.--~~~·1
APPENDIX B. The unrestricted V AR model ~Zt = r 1~Zt-1 + IIz t _ 2 + Jl. + ~D t + "t '
"t niid(O,E) where z' = [Pd,Pg,ed/g,id,igJ and If>Dtaccounts for fixed seasonal effects.
OJ
... ., ,----------------------------------------~
... ., 0
,/
·10 ., I
·H
.,
:j
."
· 51 .,
·n
·e 0 .1
v 5 'Zt
J t~ i · 10 :I
:: h
\
·10 Ie
'10
'0
A
'n~1
'10 n
u
\}V '10 /,
II I
I
'10
"\
I~ ·10
111
1 '2~----~--~~--------~~--------------- ·10 10
as II
IS " 75 7S 77 18 7' 10 II 52 I' "
t33Z t
·1. I I
·1. 10
·10 I I
·11 7~
·1. I.
'1 ~
,.
·10 10
·IO . 1l
·10 I"
71 I, 1 S re 77 11
7' lO II n !3 ,- 'I Ie
356
REFERENCES
Baillie, R. & McMahon, P. (1989): "The foreign exchange market. Theory and econometric
evidence. II Cambridge University Press.
Engle, R.F. and Granger, C.W.J. (1987): "Co-integration and error correction: representa-
tion, estimation, and testing."Econometrica, Vol. 55, pp. 251-276.
Granger, C.W.J. (1983): "Cointegrated variables and error correcting models", UCSD
Discussion Paper 83-13a.
Hendry, D.F. and J-F. Richard (1983): "The econometric analysis of economic time series,
(with discussion)." International Statistical Review, Vol. 51, pp. 111-163.
Johansen, S.(1991 b): "The power function of the likelihood ratio test for cointegration." In
this volume.
Nickell, S.J. (1985): "Error Correction, partial adjustment and all that: An expository
note." Oxford Bulletin of Economics and Statistics, Vo1.47, pp. 119-130.
Park, J.Y. & Phillips, P.C. (1988): "Statistical inference in regressions with integrated
processes. Part one." Econometric Theory, Vol.4, pp. 468-497.
357
Park, J.Y. & P.C. Phillips (1989): "Statistical inference in regressions with integrated
processes. Part two." Econometric Theory, Vol. 5, pp. 95-13l.
Phillips, P.C.(1987): "Time series regression with a unit root." Econometrica, Vol. 55,
pp.277-301.
Phillips, P.C. & S.N. Durlauf (1986): "Multiple time series with integrated variables."
Review of Economic Studies, Vol. 53 pp.473-496.
Sargan, J.D. (1964): "Wages and prices in the United Kingdom: a study of econometric
methodology." in Econometric Analysis for National Economic Planning. ed. by P .E. Hart,
G. Mills and J.N. Whittaker, London.
Spanos, A. (1986): "Statistical foundations of econometric modelling." Cambridge Univer-
sity Press. London.
Stock, J.H. (1987): "Asymptotic properties of least squares estimators of cointegrating
vectors." Econometrica, Vol.55, pp. 1035-1056.
INTEREST RATE LINKAGES BETWEEN EC COUNTRIES PARTICIPATING IN THE
EUROPEAN MONETARY SYSTEM: AN APPLICATION OF COINTEGRATION
Wilhelm FritzI
Bank for International Settlements
Monetary and Economic Department
Centralbahnplatz
CH-4002 Basle, Switzerland
Abstract
Contents
I Introduction
2 Interest rate linkages in the EMS
3 The concept of cointegration
4 Econometric specification
5 The order of integration
6 Testing for cointegration
7 The error correction representation
8 Conclusions
References
Appendix
I Introduction
IViews expressed in this paper are those of the author and are
not necessarily shared by the BIS.
359
Germany vis-a-vis the United States. Less attention has been paid
to the interest rate linkages of other countries which also main-
tain close international economic relations. The EC countries are
no exception.
i
d
= if+exp(de). (1)
(3)
2Since June 1989 the Spanish peseta has participated in the EMS
(Footnote Continued)
361
(Footnote Continued)
exchange rate mechanism; the margin for fluctuation is 6%. Since
September 1989 the peseta and the Portugese escudo have been
included in the ECU basket.
362
(4)
xt = c Yt + ct (5 )
dX t a 2t _ 1 + ~t (6 )
with:
2t = x t - C Yt = 6 t
which implies that the change in the dependent variable x from
one period to the next is determined by the lagged deviation be-
tween the two variables x and y - taking a scaling factor into
account - and a random component ~. It is possible to combine the
differenced 1(1) series and the error correction term in the same
regression without creating the spurious regression problem since
all variables are stationary.
(7 )
(8 )
4 Econometric specification
3Monte-Carlo experiments show that the OF and ADF tests have good
power properties and their critical values are quite robust
compared with alternative unit root tests (Engle and Granger
1987); however, it has been shown recently that specific
conditions in the data-generating process (like moving-average
components) can have considerable influence on the distribution
reported by Dickey and Fuller (Schwert 1989).
365
The estimated OF values are actually very high and the non-
stationarity hypothesis can be rejected for all equations without
lagged endogenous variables. Including the lagged endogenous
366
As regards the sign, it may be seen from (3') that when the
endogenous interest rate x is above its long-run average position
vis-a-vis the other country's interest rate y, the error correc-
tion term is positive. A downward movement of the endogenous in-
terest rate towards the long-run equilibrium path is required to
reduce the disequilibrium, thus the coefficient (a) should be
negative. By contrast, when the endogenous interest rate (x) is
below its long-run average, the error correction term is negative
and the change (dx) should be positive, which again requires that
the coefficient be negative. A negative coefficient is therefore
to be expected, should a linkage between the interest rates ex-
ist. (These movements towards the long-run equilibrium should not
be confused with movements of the two interest rates towards each
other. Whenever the endogenous series is located between the
equilibrium path and the other series, the likely movement is
away from the latter.)
Given the I(l) nature of the interest rates, the endogenous vari-
able is stationary, and, therefore, ordinary least-squares re-
gression is appropriate at this second stage to find an estimate
for the coefficient (a), which measures the speed at which the
endogenous interest rates adjust to the long-run equilibrium. The
error correction model has been estimated for both sample periods
separately. For interest rate pairs which failed to pass the unit
root tests the error correction model should not be estimated
(nevertheless such results are presented in the interest of com-
pleteness; see Table 4b).
Nearly all estimated coefficients of the error correction term
are negative and none of those with the "wrong" sign is signifi-
cant (Table 4). The interest rates of one country are markedly
different from all others: viz. those of Germany. Nearly all co-
efficients from the sub-samples - for both nominal and real in-
terest rates - where the German series entered the equation on
the left-hand side are insignificant. Conversely, the coeffi-
cients for the other interest rates vis-a-vis the German series
are practically all significant. Ignoring Ireland in the first
sub-sample, the smallest t-value was computed for the Netherlands
(-1.776; Table 4a).
(Footnote Continued)
371
For the real interest rates, the error correction model is also
presented for both sample periods (Table 4b). "Good" results were
obtained for the first sub-sample, but no attempt has been made
to interpret them owing to the ambiguous results of the
cointegration test. The estimates for the second sub-sample are
very similar to those found for the nominal interest rates. This
is not surprising, given the diminished inflation differentials.
The Durbin-Watson statistics for the Netherlands, Belgium and,
particularly, Ireland are low, suggesting that the dynamic speci-
fication is not appropriate.
All in all, the results clearly identify the German series as the
dominant interest rate in the EMS throughout the whole period.
Given the weight of the Deutsche Mark in the EMS currency basket
and its importance in world financial markets, some influence
would be expected. Nevertheless, the strength of this influence
could not have been predicted. In the early part of the decade
the German interest rate was closely followed by the rates in all
other EMS countries, with the sole exception of the Netherlands.
However, by the end of the observation period, the Netherlands'
series was also closely linked with the German series.
8 Conclusions
(Footnote Continued)
The domestic one-month interest rates show the same one-sided
relationship as the Euro-market rates, though clearly less
pronounced.
372
Acknowledgement
The author would like to thank his colleagues at the BIS and es-
pecially P. Andersen for suggestions and support.
373
References
A ~ ~e n d l x
J
40 40 '0
I
l
n~t'''Ilnal
__ rea~
30 3C JO 30
20 t 2J 20 20
10 10 10 10
0 0 0 0
- 10 - ,0 - 10 ·10
78 BO 82 8. 86 88 78 80 82 84 86 dS
Nether l an ds Italy
40 40 .0 .0
30 30 30 30
20 20 20 20
10 10 ,0 ,0
0 0 0 0
Belgium Denmark
40 .0 .0 4C
30 30 30 30
20 20 20 2C
'0 ,0 ,0 ,e
0 0 a 0
- 10 - 10 . a . Ie
73 BO 82 8. as a8 78 ao 82 8. 86 88
375
Germany -0.006 -0.473 2.37 0.00 -0.042 -1. 229 2.55 0.02 77
France -0.040 -1.146 2.65 0.02 -0.291 -3.385 2.34 0.14 69
Netherlands -0.012 -0.739 2.35 0.01 -0.058 -1.477 2.25 0.03 77
Italy -0.018 -0.751 2.36 0.01 -0.338 -3.782 2.05 0.17 69
Belgiwn -0.003 -0.221 1.86 0.00 -0.018 -0.779 1.88 0.01 69
Denmark -0.021 -0.943 2.24 0.01 -0.174 -2.568 2.17 0.09 69
(Ireland -0.017 -0.888 2.36 0.04 (-0.152 -1.138 1. 70 0.07 19)
Ireland 2.1325 * IDE 0.97 0.38 (-.216 -1. 542 0.12 2.02 19)
0.9056 * IFR 0.94 1. 79 (-.931 -4.035 0.47 2.08 19)
2.1729 * INL 0.95 0.56 (-.333 -2.071 0.19 2.18 19)
0.7904 * lIT 0.96 1.26 (-.645 -2.983 0.33 2.07 19)
1. 2902 * IBE 0.99 0.92 (-.459 -2.183 0.21 1. 70 19)
1.1553 * IDK 0.95 0.55 (-.270 -1. 548 0.12 1. 30 19)
IDE. IFR. INL. lIT, IBE. IDK and lIE denote the nominal interest rates
for Germany, France, the Netherlands. Italy, Belgium. Denmark and Ireland
respectively.
377
t
de
t
(a-I) e ~
t-1
+ fL t
IDE, IFR, INL, lIT, IBE, IDK and lIE denote the nominal interest rates
for Germany, France, the Netherlands, Italy, Belgium, Denmark and Ireland
respectively.
378
RDE, RFR, RNL, RIT, RBE, RDK and RIE denote the real interest rates for
Germany, France, the Netherlands, Italy, Belgium, Denmark and Ireland
respectively.
379
RDE, RFR, RNL, RIT, RBE, RDK and RIE denote the real interest rates for
Germany, France, the Netherlands, Italy, Belgium, Denmark and Ireland
respectively.
380
Sub-sample 1 Sub-sample 2
Country vs. a t-value R*2 DW obs. a t-value R*2 DW obs.
Germany IFR .013 .351 0.00 2.42 69 .007 .178 0.00 2.04 69
INL -.090 -1. 775 0.04 2.35 77 .079 1. 029 0.02 2.13 69
lIT -.031 -.682 0.01 2.31 69 -.040 -.922 0.01 1. 96 69
IBE -.107 -1. 745 0.04 2.21 69 .013 .226 0.00 2.05 69
IDK -.038 -1.120 0.02 2.33 69 -.009 -.215 0.00 2.01 69
(lIE -.106 -1. 422 0.10 1. 37 19) .012 .334 0.00 2.06 69
France IDE -.370 -4.315 0.21 2.34 69 -.254 -3.318 0.14 1.89 69
INL -.313 -3.697 0.17 2.31 69 -.239 -3.086 0.12 1.88 69
lIT -.567 -4.007 0.19 2.36 69 -.796 -6.323 0.37 2.04 69
IBE -.430 -4.359 0.22 2.22 69 -.324 -2.925 0.11 1.83 69
IDK -.284 -3.110 0.12 2.38 69 -.199 -2.711 0.10 1. 97 69
(lIE -.849 -3.653 0.43 2.09 19) -.131 -1.480 0.03 1. 94 69
Italy IDE -.287 -3.995 0.19 2.13 69 -.161 -2.462 0.08 2.06 69
IFR -.008 -.076 0.00 2.38 69 .132 1.153 0.02 2.42 69
INL -.287 -3.797 0.17 2.09 69 -.154 -2.386 0.08 2.08 69
IBE -.354 -3.904 0.18 2.10 69 -.156 -1.797 0.05 2.05 69
IDK -.245 -3.130 0.13 2.13 69 -.101 -1.642 0.04 2.19 69
(lIE -.473 -2.039 0.19 2.12 19) -.004 -.064 0.00 2.24 69
Denmark IDE -.127 -2.456 0.08 2.12 69 -.194 -2.950 0.11 2.12 69
IFR -.106 -1. 787 0.04 2.10 69 -.106 -1.880 0.05 2.27 69
INL -.174 -2.662 0.09 2.08 69 -.245 -3.319 0.14 2.04 69
lIT -.170 -2.302 0.07 2.04 69 -.145 -2.361 0.08 2.25 69
IBE -.149 -1.930 0.05 2.14 69 -.183 -2.560 0.09 2.09 69
(lIE -.186 -1.076 0.06 1. 37 19) -.199 -2.889 0.11 2.11 69
Ireland (IDE -.148 -1.416 0.10 2.16 19) -.108 -2.392 0.08 1. 33 69
(IFR - .135 -1.910 0.17 2.05 19) -.101 -1. 927 0.05 1. 34 69
(INL -.083 -.917 0.04 2.17 19) -.100 -1.890 0.05 1. 30 69
(lIT -.200 -2.243 0.22 2.39 19) -.169 -3.569 0.16 1. 37 69
(IBE -.275 -1.402 0.10 2.04 19) -.168 -2.969 0.11 1. 28 69
(IDK - .111 -1.239 0.08 2.34 19) - .117 -2.056 0.06 1. 36 69
IDE, IFR, INL, lIT, IBE, IDK and lIE denote the nominal interest rates for
Germany, France, the Netherlands, Italy, Belgium, Denmark and Ireland
respectively.
381
Sub-sample 1 Sub-sample 2
Country vs. a t-value R*2 DW obs. a t-value R*2 DW obs.
France RDE -.403 -4.340 0.22 2.26 69 -.222 -2.670 0.09 1.87 69
RNL -.349 -3.778 0.17 2.25 69 -.221 -2.664 0.09 1.88 69
RIT -.366 -2.755 0.10 2.45 69 -.599 -5.282 0.29 2.15 69
RBE -.356 -3.873 0.18 2.28 69 -.291 -2.590 0.09 1.83 69
RDK -.358 -3.584 0.16 2.28 69 -.248 -2.906 0.11 2.00 69
(RIE -.291 -1.683 0.14 2.70 19) -.148 -1. 621 0.04 1. 93 69
Netherlands RDE -.141 -2.279 0.06 2.20 77 - .133 -2.845 0.11 1.65 69
RFR -.065 -1.474 0.03 2.35 69 -.043 -1. 542 0.03 1. 62 69
RIT -.058 -1.396 0.03 2.34 69 -.048 -1.572 0.04 1. 57 69
RBE -.369 -3.453 0.15 2.02 69 -.097 -2.421 0.08 1. 61 69
RDK -.136 -2.420 0.08 2.28 69 -.006 -.180 0.00 1.56 69
(RIE -.097 -1.149 0.07 2.47 19) -.086 -2.381 0.08 1. 58 69
RDE, RFR, RNL, RIT, RBE, RDK and RIE denote the real interest rates for Germany,
France, the Netherlands, Italy, Belgium, Denmark and Ireland respectively.
PART 7:
Karsten-Dietmar Freimann
Institut fUr Angewandte Wirtschaftsforschung
Ob dem Himmelreich 1
D-7 400 Tiibingen, Germany
Abstract
In this paper the small and large-sample properties of six estimators of the variances
and covariances of random coefficients are compared. The criteria chosen for the
evaluation in small samples were bias, mean-squared error, the fraction of negative
variance estimates, and consistency for large samples. It turns out that five of these
estimators are consistent, but that only two of them always show moderate or small
mean-squared error values, even in the presence of data generated by a linear trend
and a random walk process.
Contents
Introduction
2 The Model
6 Summary
References
386
1 Introduction
A large number of authors have discussed the problem of estimating the expected
values and the variances of random coefficients since the well-known article by
Hildreth and Houck appeared in 1968. This paper is a further contribution to this
topic; it also generalizes the problem to the estimation of covariances.
If the researcher believes that a random coefficient model is suitable for his empirical
problem he must interpret the regression coefficients very carefully: The values of the
variances and covariances of the random coefficients have also to be taken into
account. The elements of the covariance matrix are important because
their size indicates changes in the behaviour of the people or the severity of the
specification error the researcher may have committed,
they can be used for a GLS estimation of the parameter vector, i. e. for im-
proved least squares estimation. (In a maximum likelihood procedure estimates of
the variances and covariances are always computed.)
In this paper the small and large-sample properties of various estimators of the
covariance matrix of random coefficients are compared with the corresponding
maximum likelihood estimator, namely some variants of a generalization of the
Hildreth/Houck procedure and two estimators which are based on a decomposition of
the covariance matrix, an idea by Schwallie (1981).
In section 2 of this paper the random coefficient model and its transformation into a
fixed coefficient model is described. Section 3 deals with some variants of a
generalization of the Hildreth/Houck procedure, and in section 4 the maximum
likelihood estimator as well al) two other estimators are introduced which make use of
the positive definiteness of the true covariance matrix. In order to investigate the
small-sample properties of these estimators the results of a Monte Carlo study are
presented in section 5. Section 6 contains a summary of the results.
2 The Model
The estimation problems discussed in this paper are connected with the following
random coefficient model:
°
(A2) - The elements Xtk of the vector Xt are nonstochastic and satisfy
< IXtkl < 00 for k = 1, 2, ... , K and t = 1, 2, ... , T.
- The TxK matrix X defined by X': = (Xl, Xz, ... , xT) has full column rank K
for T ~ K.
- The matrix
T
Mx:= lim 1 L XtXt'
T~oo T t=1
[ x: ,2 2x, ~ x12 2x'.' x13 ... 2xllxI K X12 2 2X12X13 ... XIK 2
X** -
XTl 2 2XT1 xT2 2XTlxT3 ... 2xTl xTK XT2 2 2xT2 xT3 ... XTK 2
I,
has full column rank for T ~ K(K + 1)/2, where Xtk 2 := (Xtk)2,
t = 1, 2, ... , T and k = 1, 2, ... , K.
- The matrix
(2/3)
By (AI) the vector Vt is distributed as N(O, 1:) and serially independent. The random
variable Ut is distributed as N(O, 0/) where O"t 2 XtTxt. Hence Ut is also serially
independent but heteroscedastic.
The problem we are interested in is to estimate the elements of the covariance matrix
1: = E(vtv t'), i. e. the variances and covariances of the random coefficients.
388
(3.1/1)
with u' := (UI, ... , uT)' X' := (Xl, ... , XT) and with I being the identity matrix. Hence
the relation E(u·u·') = M4»M' holds with E(uu') =: 4», and we find that
In this equation we defined u·" = (UI· 2, U2· 2, ... , UT· 2), M* M*M, where "."
denotes the Hadamard product of two matrices, X· is given by
and cr' : = (crll, cr2Z, ••. , crKK)' Now, adding u·· and subtracting E(u h) on both
sides of (3.1/3) yields
The parameter vector cr in (3.1/4) can now be estimated by means of OLS yielding
an estimator cr· (also denoted by HH for Hildreth/Houck in the sequel) which is un-
biased and consistent under (A 1) and (A2), but not (asymptotically) efficient.
The main problem connected with this estimator is, however, that it very often
produces negative estimates for the variances crkk, k = 1, 2, ... , K In order to
overcome this problem some authors suggest a GLS-like estimation of the parameter
vector cr (see e.g. Hildreth/Houck (1968) or Raj/Ullah (1981»; others propose in-
strumental variable techniques (see Raj/SrivastavajUpadhyaya (1976», but none of
these estimators provides an entirely satisfying solution. Another way to overcome the
problem would be inequality restricted estimation, e.g. by means of maximum
likelihood methods. Such estimators are not considered in this paper. The most
convincing solution to the problem of negative variance estimates was provided by
389
Schwallie (1982) who proposed the decomposition of L into the product of two
nonsingular triangular matrices tJ.:
L = tJ.'tJ. (3.1/5)
The Hildreth/Houck procedure outlined so far only estimates variances. The off-
diagonal elements of L, however, the covariances of the random coefficients, will in
practice not be restricted to zero and may give valuable information about the model
to be specified. Therefore, the HH method was generalized (see for instance
Raj/Ullah (1981» to the estimation of covariances, based on a regression equation
which is very similar to (3.1/4):
That is, X** contains products of the exogenous variables, and the vector cr* consists
of the elements of the upper triangle and the main diagonal of the matrix L Again,
we consider the OLS estimation of the vector cr* proposed in Raj/Ullah (1981):
The OLS estimator cr*· of <r* is unbiased and consistent, but
not (asymptotically) efficient, and
it can produce negative estimates of variances (and very often does).
Since the GHH estimator often shows large mean-squared error values, usually caused
by large variances of the elements of cr*·, it was tried to improve its small-sample
performance by dropping the matrix M* in (3.2/1). This modified estimator (denoted
by MGHH) has the following properties:
MGHH is consistent if (A 1) and (A2) hold (see Freimann (1988) for a more
general case), but
not unbiased and
not (asymptotically) efficient.
MGHH can also yield negative variance estimates.
390
The corrected estimator (denoted by CMGHH) has the following properties under
(A 1) and (A2):
The estimator is consistent (see Freimann (1988) for a more general model), but
not unbiased and
not (asymptotically) efficient.
It always yields a positive definite estimate of L
(4.1/1)
The first step in the derivation of the PDM estimator is to define the Kx1 vector
(4.1/2)
such that
(4.1/4)
holds with rt·' = (rn·, rt2·, ... , rtK·)' Since the vector rt • is not known, its ele-
ments are estimated by
r tk •• 1, 2 .... , K, (4.1/5)
391
where the symbol "I . I" denotes the absolute value and where Ut - is the OLS resi-
dual. Now we have the following system of regression equations:
(4.1/6)
IUt-I/V"K
IUt-I/V"K
(4.1/7)
The coefficients 0ij are the elements of the upper triangular matrix fl. They are
estimated by OLS applied to each equation of the system (4.1/7). The resulting
estimator E - = /l -, /l- of the covariance matrix of the random coefficients has the
following properties if (A 1) and (A2) hold (compare Freimann (1988)):
It is biased and
not consistent, but as we shall see later,
has small variances in finite samples,
small mean-squared errors in finite samples, and
always yields a positive definite matrix E-.
The estimator of the elements of E derived from the OLS estimates of the Oij' PDM,
is in general not consistent because the disturbances I':tk of the equation system
( 4.1/7) need not have zero expectations, not even asymptotically.
The POM estimator has the advantage that it can be calculated without any restnctlOn
on the coefficients to be estimated: One can always order the elements of Xt in such
a way that the element 011 - of the upper triangular matrix /l -, which corresponds to
the standard deviation of the random variable vtl in (2/3), is positive.
The matrix E - may be approximately positive semidefinite since the estimates of the
variances can approach zero very closely.
In order to make use of both the satisfactory small-sample properties of the POM
estimator and of the consistency of the Hildreth/Houck-related estimators, POM is
combined with the CMGHH estimator (compare Freimann (1988». The resulting
mixed estimator (denoted by ME in the sequel) is a weighted average:
(4.2/1)
where cr*PDM and cr*CMGHH denote the POM and CMGHH estimators of the vector
cr* in (3.2/1). and where 0 :0:; cT :0:; 1 is the weighing function, defined by
The symbol "MSE" denotes here the trace of the mean-squared error matrix of the
corresponding estimator (in brackets). As MGHH is consistent and POM is not, the
weighing function cT converges to 0 if T approaches infinity, and hence (1 - CT)
converges to 1.
(4.2/3)
using the consistency of CMGHH. This choice of weights makes the mixed estimator
converge to the true covariance matrix L in probability, i.e. ME is consistent under
(AI) and (A2).
T T
LL - T In(21T) - L I In (Xt'!1' !1Xt) - L I (Yt - Xt'(3)2 (4.3/1)
2 2 t= 1 2 t= 1 Xt'!1' !1Xt
In order to find the unconstrained maximum of this function, several procedures were
tested, among them variants of the Gauss and Newton-Raphson methods, e.g. the
Marquardt algorithm. Most of them showed good convergence results for one
replication of the simulation experiment; the 200 replications this study is based upon
took however too much time on the AT-compatible personal computer used for the
simulation. Therefore a variant of the direct search method by Ludeke, Hummel and
Rudel (1988) was used for the calculations.
393
The original procedure operates on a standardized parameter space and uses a lattice
which is refined in subsequent iterations. The ML version investigated in the present
Monte Carlo study takes the estimates of the mixed estimator (ME) outlined above
as starting values and establishes a lattice in the neighborhood of the starting point
and refined ones in subsequent iterations. As the reader will see below, this procedure
often improves the already good results of the ME and PDM estimators. It usually
finds a local maximum of the likelihood in the neighborhood of the starting point in
at most three iterations, often already during the first iteration. Its advantage for
simulation experiments is that the number of calculations necessary to reach the local
maximum is nearly the same for each replication and that only the calculation of the
likelihood has to be done. First or second derivatives are not needed.
(5.1/1)
where btl - N(/31' 0'11)' bt2 - N(/32' 0'22), Cov(b tl , b t2 ) = 0'12, i. e. the random
coefficients were generated from a bivariate normal distribution. Also, the data for XI>
the exogenous variable, were chosen as realizations of normally distributed random
variables and produced by the random number generator of RATS. Three different
processes were taken into account for the generation of the exogenous variable Xt:
The last two data generating processes are nonstationary (a linear trend and a random
walk process). They proved to be important for the small-sample properties of the
estimators. The numbers of observations available for each variable in the simulation
experiments were T = 20, 40, 80.
In each experiment the expectations (31 and (32 of the random coefficients were held
constant at 1.0, and, unless otherwise stated, the parameter values 0'11 = 0'22 = 0.05
and 0'12 = 0.01 were used. Each experiment was repeated 200 times.
394
The first series of experiments consists of a variation of the variances and covariances
of the random coefficients and was done for untrended data. Table 5.2.1 shows the
results for the mean-squared errors (MSE, now interpreted as the mean-squared
errors of the estimated elements). In order to save place, only typical outcomes are
reported here. 1 It turns out that the mean-squared errors
rise independently of T with increasing variances of the random coefficients,
may rise or fall if the covariances increase,
fall if the number of observations, T, increases,
of POM, ME and ML never exceed 0.5 for 0"11 = 0"22 = 0.05 and never exceed
1.5 for 0"11 = 0"22 = 0.1 while the MSEs of the other estimators sometimes take
very large values, especially in the case of 0"11·.
- The results show the following order:
MSE(ML) < MSE(POM) < MSE(ME) < MSE(CMGHH) as well as
MSE(CMGHH) < MSE(MGHH) < MSE(GHH).
No exception to this ordering occurred among the simulation results.
For a deeper analysis of these results it is useful to look at table 5.2.2: The biases
corresponding to the above MSE values
increase in absolute value, apart from only few exceptions, if the variances of the
random coefficients rise,
rise or fall, depending on the estimator, on the coefficient to be estimated and
on the number T of observations, if the covariances between the random
coefficients increase,
of PDM, ME and ML never exceed 0.5 for 0"11 = 0"22 = 0.05 and never exceed
1.0 for 0"11 = 0"22 = 0.1 in absolute value while the biases of the other esti-
mators often exceed 1.0 in absolute value, expecially the biases calculated for the
estimates of 0"11 at T = 20,
of GHH do not indicate that this estimator is unbiased.
This outcome reveals that the fall of the mean-squared errors caused by an increase
in the number of observations T (see table 5.2.1 above) is due to the performance of
the variances. A clear ordering of the estimators with respect to the absolute value of
the bias is difficult to find.
The fraction of negative variance estimates which occurred in each experiment was
very large for GHH and MGHH. The minimum fraction was 181/400 (calculated for
2 variances in all 200 replications) found for GHH at 0"12 = 0.04 and T = 80, the
maximum fraction was 269/400 computed for MGHH at 0"11 = 0.05 and 0"12 =
0.01 as well as at 0"11 = 0.1 and 0"12 = 0.02, in both cases at T = 20. The fraction
of negative variance estimates was always larger for MGHH than for GHH although
this relation does not hold for the absolute bias.
1The interested reader can get the complete set of simulation results from the author.
395
Table 5.2.1: MSE values dependent on 0"11. 0"12. 0"22 and T for untrended Xl
G H H MGHH C MGH H
0"12 T 0"11 • 0"12 • 0"22 • O"ll • 0"12 • 0"22 • O"ll • 0"12 • 0"22 •
I I
~1
P D M ME ML
0"11 • 0"12 • 0"22 • O"ll • 0"12 • 0"22 • O"ll • 0"12 • 0"22 •
I I
GHH MGHH CMGH H
0"12
-
j 0"11 •
P DM
0"12 • 0"22 • 0"11 •
ME
0"12 • 0"22 • O"ll •
ML
0"12 • 0"22 •
Table 5.2.2: Bias values dependent on 0"11. 0"12. 0"22 and on T for untrended XI
0.05
P DM ME ML
0"12 T 0"11 • 0"12 • 0" 22 • 0"11 • 0"12 • 0" 22 • 0"11 • 0"12 • 0"22 •
0"22 0.1
G H H MGH H CMGH H
0" 12 T 0"11 • 0"12 • 0"22 • a II • a12 • a22 • all • a l2 • a22 •
dJ
0.0
20
40
all •
PDM
a l2 •
Table 5.2.3 shows the mean-squared errors obtained for different versions of a linear
trend in the exogenous variable Xt. The results reveal that the MSEs
rise or fall if the slope of the linear trend increases, they usually rise for er11·,
usually fall for small values of the slope if the number T of observations increa-
ses,
rise or fall for moderate and large values of the slope if T increases,
of ML never exceed 1.0 and those of POM and ME never exceed 1.5 while the
MSE values of the other estimators often become very large, especially for er11·.
An analysis of the bias values corresponding to table 5.2.3 is even more difficult and
is therefore omitted here. The only fact to note is that the absolute values of the
biases of ML, POM and ME never exceed 1.0 while the absolute biases of the
Hildreth/Houck-related estimators often exceed this margin in the case of eru •.
Again, the fraction of negative variance estimates was large for GHH and MGHH,
but not as large as the one found in the last section. As before, GHH produced
fewer negative variance estimates than its modification. All these fractions lie in the
interval between 149/400 and 231/400.
As an alternative to the linear trend, the exogenous variable Xt was generated as the
realization of a random walk process. Table 5.2.4 shows the mean-squared error
values calculated for different values of the variance of Et. It turns out that the MSEs
for er11· first fall and then rise and those for er12· and er22· monotonously fall
in the case of the HH-related estimators if the variance of E t increases,
for er11· monotonously decrease, for er12· first fall and then rise and for er22·
remain (nearly) constant in the case of POM and ME if the variance of Et in-
creases,
monotonously fall in the case of ML if the variance of Et rises,
for er 11· rise and for er 12· and er 22· fall if the number T of observations in-
creases, independently of the estimator,
of POM and ME never exceed 0.01 and those of ML never exceed 0.03 while
the MSE values of the HH-related estimators do.
An ordering of the estimators with respect to the size of their mean-squared error
values yields:
MSE(POM) < > MSE(ME) < MSE(CMGHH) < MSE(ML) as well as
MSE(ML) < MSE(MGHH) < MSE(GHH)
400
Table 5.2.3: MSE values dependent on the slope a 1 of the linear trend of XI and on T
P DM ME M L
at T all • a 12 • a22 • erll • erl2 • er 22 • erll • a12 • er 22 •
GHH M G H H eM GHH
(1 2
f- T (111 • (112 • (122 • erll • (112 • (122 • (111 • (112 • (122 •
PDM ME M L
er 2
f- T (111 • er 12 • (122 • erll • (112 • (122 • (111 • (112 • er 22 •
where "< >" denotes that there is no clear ordering between PDM and ME. It is sur-
prising that the random walk process of XI worsens the small-sample properties of the
maximum-likelihood estimator in such a manner that its MSE values are even worse
than those of CMGHH.
The bias values (which are not tabulated here) reveal that the influence of the
random walk process of XI is different for each estimator and for each coefficient. The
size of the biases of all estimators is however much smaller than in the case of a
linear trend. Surprisingly, GHH now shows the smallest absolute bias values which
increase with T.
The fraction of negative variance estimates found in the experiments of this section is
much smaller than before for both GHH and MGHH. Now, there is no clear
ordering with respect to this criterion, the fraction of negative variance estimates
found for GHH and MGHH always lies in the interval between 46/400 and
231/400.
6 Summary
The results of the investigation described in this paper show that although five of the
six estimators of the variances and covariances of random coefficients are consistent
their small-sample properties need not be satisfactory. Only two estimators always
show moderate or small mean-squared error values, namely PDM and ME, even in
the presence of a linear trend and a random walk process of the exogenous variable.
While PDM is not consistent, the estimator ME, which is consistent, turns out to be
the most robust method. The maximum likelihood estimator ML shows small mean-
squared errors in the case of untrended data and in the presence of a linear trend, its
behaviour worsens, however, when the data are generated by a random walk process.
Four of the estimators, namely ML, PDM, ME and CMGHH, always yield positive
variance estimates while the Hildreth/Houck-related estimators GHH and MGHH
often produce negative ones in combination with large mean-squared errors.
References
Jan F. Kiviet
University of Amsterdam
Department of Actuarial Science & Econometrics
23 Jodenbreestraat, 1011 NH Amsterdam, The Netherlands
Abstract
The usual t- and F-tests for linear constraints on regression coefficients
involve an unambiguous assumption on the distribution of the disturbances,
whereas in most practical applications the correlations of the disturbances
are unknown. However, it is possible to use the same test statistics in a
bounds test under a composite assumption on these correlations. In this paper
such bounds for critical values and also for actual significance levels are
derived which take into account that particular regressors (such as the
constant or a time trend) occur in the regression. These bounds are
calculated for tests on restrictions in multiple regressions with a constant
term and with AR(l) or ARMA(l,l) disturbances. They are presented in such a
way that the bounds test is easily applicable, as is illustrated.
Contents
1 Introduction
2 Bounds on t- and F-Tests
3 Boundary Levels and Critical Values
4 Illustrations and Conclusions
Acknowledgements
References
1 INTRODUCTION
is supposed that the regressor vectors are orthogonal, or that some of the
regressors are similar to particular eigenvectors corresponding to the above
mentioned eigenvalues.
Less restrictive results on the bias of the variance estimates are
presented by Sathe and Vinod (1974), and as far as the variance of the
disturbances is conerned - by Neudecker (1977). In Neudecker (1978) one
constraint is put on the regressor matrix; it is supposed that the regression
contains a constant term, and now quite definite results on the direction and
possible magnitude of the bias in the residual variance are obtained. Vinod
(1976) calculates bounds on the significance points for the F-test of the
complete coefficient vector for arbitrary regressor matrices and he indicates
how the practitioner can use these bounds to infer from an Ordinary
Least-Squares (OLS) statistic in what in fact is a Generalized Least-Squares
(GLS) model. In Kiviet (1980) some critical comments on Vinod (1976) are made
and in addition attainable bounds are derived and calculated for the
significance points of the test statistic for an arbitrary set of linear
constraints on regression coefficients. These bounds, which generalize
Watson's results, also apply to the significance test of a subvector of the
coefficients, hence they include the usual t-test. The above mentioned
contributions are reviewed in Vinod and Ullah (1981, pp.97-l02); a more
sophisticated but computationally much morn complex procedure has
recently been developed by Dufour (1990).
In this paper we try to make the results on the t- and F-test of more
particular interest to the practitioner. This is pursued by presenting the
bounds in convenient diagrams, and also by tightening the bounds through
exploiting the presence of particular regression variables in the regression,
along the same lines as Neudecker (1978) did with respect to bounds on the
bias of the disturbance variance estimate. The assumptions made here with
respect to the presence of particular regressors have practical relevance,
contrary to the assumptions made by Watson concerning the correspondence of
regressors with particular eigenvectors. In the figures presented here, we
assume and exploit that a constant term is present. The diagrams enable to
interpret - without any further calculations the standard statistical
results of a regression analysis embodied in t-ratios and F-statistics, when
the assumption is dropped that the disturbances are independent. Some of the
diagrams can also be used when it is recognized that - in an (unsuccessful)
attempt to make Ordinary Least Squares applicable - an incorrect assessment
of the correlation matrix may have been employed in the transformation of the
model. Doubts about a regression analysis originating from a fallacy in the
adopted correlation matrix can thus be removed, because the bounds permit an
appeal to the robustness of t- and F-tests with respect to departures from
the assumption of independent disturbances.
(1)
(2)
N = X(X'X)-lR' [R(X'X)-lR'j-1R(X'X)-lX'
This is easily seen, since e=Mu, and upon substituting r=Rp it is found
that
e = y - Xb
e + Nu (M+N)u
~
h,n-k
< Q
- h,n-k
where
h
2
2: >.
n-h+i 'Ii
n-k i=l n-k
~
h,n-k 11 n-k and FU
11
h,n-k
2
2: >.. '1 +
~ k i
i~l
, 2 -1/2
with >'l~" ·~>'n the eigenvalues of 0 and ('11"" ,'In) =(a 0) u-N(O,I)
407
PROOF: See Kiviet (1980), but note that here the eigenvalues are arranged
in decreasing order.
From this Theorem bounds on the actual critical values of this test
statistic follow, and also bounds on the actual significance level of the
standard test procedure when the a-level F significance points are
(erroneously) employed. In Kiviet (1980) tables of calculated boundary
(5)
2
THEOREM 2: If y = X~+u ~ Xl~1+X2~2+u wHh u-N(O,o 0) whereas the
constraints are such that R~=R2~2-r is valid (and hence they do not concern
the coefficients ~l)' then the statistic Qh -k is bounded by the statistics
~* ~ ,n
h,n-k and Fh,n_k (which do not depend on X2 and R2 ) so that
where
h h
~* ~ n-k i=l
l;
"
*
n-kl-h+i 'Ii
*2
and F
u* =
n-k
l;
i=l
".
* *2
.1 'Ii
h,n-k 11 n-k h,n-k 11 n-k
l;
i=l
".
*
.1
*2
'I k +i
2
2:
i=l
*
"k2+.1.
*2
'I k +i
2
Here *
"I?:" *
'?:"n-k are the possibly non-zero eigenvalues of 1'11 °, and
* * l 2 -1/2
C'u - N(O,I) , where C is the nx(n-k 1 ) matrix
('11'··· n k )' = (0 C'OC)
,'I _
1
of eigenvectors corresponding to the unit eigenvalues of 1'11 =I-X1 (Xi Xl)-lXi
408
PROOF: For the idempotent matrix Xl we have Ml-CC', whereas C'C-I k'
n- 1
Upon premultiplying model (5) by C' we obtain
y * (6)
If it is assumed that O-In this yields 0 -I and then the appropriate test
n-k l
statistic for the hypothesis R2fi2=r is
* * *
n -k
Qh,n* -k* = - h -
e *' e *
Since R=[O:R 2 ] this statistic is equivalent to Qh,n-k of (3), as is seen as
follows. From a standard result on partitioned regression we obtain Rb=R 2b 2 =
1 * *'*-1*'* -1
R2 (X 2 Ml X2 )- X2M)Y and since b 2=(X 2 X2 ) X2 y =(X 2 CC'X 2 ) X2 CC'y=b 2 we
find Rb=R 2 b 2 Also R(X'X)-lR'-[O:R ](X'X)-l[O:R ]'~R (X'M X )-lR'=
*' * -1 * * * * 2 2 2 2 1 2 2
R2 (X 2 X2 ) R2 And finally e -y -X2b2=C'(y-X2b2)=C'(y-Xlbl-X2b2)=C'e, so
the residual sum of squares is e *' e *=e'CC'e=e'M 1 e-e'e.
Q* * *=Q
Since they have equivalent boundary critical values.
h,n -k h,n-k *
Upon application of Theorem 1 to model (6) - hence replacing n by n =n-k l and
k by k2 etc. these bounds are easily found, and they involve now
expressions in the n-k l eigenvalues of C'OC, which conform to the n-k l
possibly non-zero roots of CC'O=MlO. o
L*
Like the boundary statistics of Theorem 1, the statistics Qh,n-k and
u*
Qh,n-k of Theorem 2 imply on the one hand bounds on the actual significance
level of the test procedure when the dependence of the disturbances is
ignored, i.e. the test statistic Qh,n-k is naively used in combination with
the a-level F significance points. On the other hand they imply bounds on the
exact critical values of Qh,n-k in the presence of serial correlation.
The actual significance level (probability of a type I error) of the
Qh,n-k test statistic when employed at the nominal size a is defined as
a = (7)
h,n-k
(9)
on the actual critical value; these boundary critical values are defined by
L* L*) u* u* )
P ( Fh,n_k ~ Fh,n_k(a) = a = P(Fh,n_k ~ Fh,n_k(a) (10)
The bounds of (8) and (9) are attainable, and in contrast with ah,n_k and
Qh,n-k(a), they do no depend on X2 . Notwithstanding the multitude of their
determining factors, such bounds can be tabulated. The bounds on ah,n-k in
(8) indicate the maximal deviation of the actual from the nominal
significance level which is possible, given 0 and Xl' due to X2 .
The bounds on Qh,n_k(a) permit a bounds test which is immune to serial
dependence of the disturbances. This test procedure is employed as follows.
From Theorem 2 we find
(11)
and
a . (12)
Now let q be the realisation of the test statistic Qh,n-k in some model.
U*
Consequently, if q ~ Fh n_k(a) then the null-hypothesis should be rejected,
L*
and if q ::5 Fh,n_k(a) then the null-hypothesis should not be rejected; in the
L*
event that one obtains Fh,n_k(a) u*
< q < Fh,n_k(a) the test is inconclusive.
The bounds given in (8) and (9) depend on n,k,h,O,X l and a. The boundary
values can be calculated numerically, since the boundary statistics are
ratios of simple quadratic forms in normal variates. In all calculations that
410
(13)
0.3 15 0.15 0.16 0.19 0.17 0.20 0.21 0.18 0.21 0.23 0.23
50 0.15 0.16 0.20 0.16 0.21 0.25 0.16 0.22 0.26 0.29
0.9 15 0.56 0.63 0.73 0.66 0.78 0.80 0.69 0.80 0.83 0.83
50 0.67 0.71 0.87 0.74 0.89 0.94 0.76 0.91 0.96 0.97
411
Notice that when the serial correlation coefficient is only 0.3, the type I
errors may already have tripled, and - irrespective of the size of the sample
- they may be a tenfold of a, or even larger when p-0.9 . Upon comparison of
these results with those for general X matrices, we see that the
incorporation of only a constant term does not lead to bounds which are
considerably tighter.
0.0 0.3 0.11 0.11 0.12 0.12 0.12 0.12 0.13 0.12
0.6 0.15 0.15 0.17 0.16 0.18 0.16 0.20 0.17
0.9 0.17 0.17 0.19 0.17 0.21 0.18 0.24 0.18
0.3 0.6 0.09 0.09 0.10 0.10 0.10 0.10 0.11 0.10
0.9 0.11 0.11 0.12 0.12 0.13 0.12 0.14 0.12
0.6 0.3 0.20 0.23 0.21 0.24 0.22 0.25 0.23 0.26
0.9 0.08 0.08 0.08 0.08 0.08 0.08 0.09 0.08
0.9 0.3 0.47 0.63 0.52 0.66 0.54 0.69 0.56 0.54
0.6 0.29 0.50 0.31 0.52 0.32 0.53 0.32 0.54
when k increases, the critical t-values increase too, because of the decrease
in the number of degrees of freedom n-k. One can derive (and observe from the
diagrams) that
(14)
and a comparable inequality holds for h-l concerning the bounds for the
critical t-values.
In comparison with the unrestricted bounds tabulated in Kiviet (1980) we
find that the incorporation of a constant term of course leads to tighter
bounds, but that this gain is impressive only for high values of p and for
small or moderate sample sizes.
Y.
~
11.60 + 0.789 Xi +
(1.53) (4.41)
Note that these diagrams cannot be used to test the constant term,
because we restricted ourselves in the derivations to tests involving the
coefficients corresponding to X2 only.
Figure 2 presents bounds on F-values, where the dashed lines mark the
Fh,n_k(a) critical values.
IG k - 2 I. lG k - 3 I.. :.
12 12 : p_ 0.9
101i-------~~~~·
----------~--~
IOIf-----~------~-=~~----~----~p- 0 9 -
...... L .... :.. .. ,
8 8 .......
- - - - -~ . . ..... : . . . . _. . i . , .. : ...•• .•• . j.,"
6 ~~~~~~~~'~~~~'~~~.
~ ......... . .................. . ........ :: .. :.:: ... : p-07 ' .. '::::::::::::::.. .. . " . : : :: ::: ::: :: ::: IT" 0.7"
.... ... .. --........ .. ....... ,,; . . . . . . . . : p - O.~ ..
2 ~
...
... ... .. ........ ......... _ ~........ , _. _ .~'
:. 7 . ~ . ~ . ":"' . : . ":'" p. O
1.6 :------:
.... ~::- : ~
..::- .."" ..~~-----.:: p- OJ .
..""
1.6 :------...:
.. .... ~-::
. :""
...,..,
...,.,
..,..,.....:- -...:.....- ---..:: IT" O.J ..
0." .............. . ....................... ,; . . ....... ; p . 0.9 06 .... . . .•• , ...... ~ ~ ~: ',:::'.'. ~:.: .......... : . •. ~ ::'.::',::'.: .~. 09
...... .. ..... .... . ' ., .... ........ .
o,' ~--~~--±-------~~~~~~ 0~-----1~0----~20~----JO
------4~0-----5-0----~
10 20 30 ~o ~o
n ( , umpl . su o) n ( ~amph .t l2: f' )
16 .1 t - 4 IG k - :;
~ P- 09
Il I~
. : . . .
10 IOI+-----7-----~--~----~----~
---_ ... : ' .. " .... ! .... ..... :' .. . -_ ... ........... \
8 ......... .. ....... . .. . ........................ .
.
:- ... ,- .:'.: .- ."':'.7 ; , ~ , ~ ,,,:,, ,,,,, , ,,, , ,,! __ __ __ ._ "~ p. O
1.6 16
1.2
o o
10 20 30 40 ~O 10 ZO 30 40 50
n (lo.mp' " loU . ) n (samph J'Uf')
414
I k - ~ ; h - JL''''
loof::====.:==~~~~--~-~--I
w /~ '-O.' .
........... . . .
--\'······_-·-········ 1·
. "; p- 07 "
20 ·········;· · ·······j·········:·····,,··I·········
. . ...... .:... ..... . p .. 07
.
~o
p - O.J ••
.\ ., : p- o, ~ :-:
,.,', ;::::: ....... :
0.6
O. ~ ... _........ ......... .. : .... : : : ~ f ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ p . 0.7 ..
0~:tlrk==_;5=.h==_=J'I-:·-··-··-·.-..~f.-..-..-..-..-:.-..-..-..-..:~.-..-..-..1.
~IOJ-~~~~~-:::::==~====~==~~p-o~
"':: :: :.:::::. :::::::::::
~
~ b .... . .......... .. ............. .
.::::.:~;~ ~~ :.~::~~:~ ~ :~::~~':~;.: ~ :::::.::. : :::.: : ' : . ; p- o.J :.
,: K- ~ct~,J,~:YH : ~ •
I
O.!>
04
... , : : : : : ~: : : : : : : . . : : . ~ ~ ~ ~ ~ " ~ : '. .'. : '. .".~: ~ . -. .-. .....'"'::~
. .:- . ~ p- 0.9 .:
f~ 1I--..i-----i- -@,:~-
.. ...
20 ......... :'""-...c_.......
' :-;'-:7:-:-'t:;
, '-:-':":
".........
c.,';';''-:-:.::
' '.,.:.;
' .....
':'":';''':''':.;.'
. .';'":.;.;..:.:..:..:.:.:..;
,.. -: p. 07
.. : . . . . . : p-O.J ··
.j k • 3 : r • 0 .3 1- ..
. .
101~----~------~----~----~------~---- 1 10+-. .-.-..-.-.-:
. :~ ..-.-...".."...".=..,;.: p_ 0.9
. .-.-..-.-.~.:-.-..-.-..-.-T.---~
'-.- ~
:·: ~: ~~~~~U<H«TT~/l//U .:
I! • . ::: :: : :;~; P.0.9 "
G : ~ ;;; ; ;;r;;;; : :: c; ;; ;;ri :: ::iLi;;; ;) .
: .' .' .' .' .' ::.' :.' .' .' .' .' .' .' .' .' :.' .' :::.- ...:....:::: .- :.- .' .' .' .' .' .' .' .' ~ ..... ..... .... . , ........... ...... ....... ... ...... .
. : : -: . . .. - . .. . : p . 0 ,(" . ,
3 • • p- 0.0 ~_H:-:':--,'-c'·cT:.,.
· ',.,'.,. ' ._._.'.;.:~
' ", ' .~._._.._. _..~:_
.._._. _••_. .....
..:
; -.. ... -
:::" ..
- _. - - - -
!
. ( .
2 p. 0.0
---.:::::==:
H
) :::.' ::: ::: : :: .':: :: : :~ : : : ,- ::: :::: .'.' .'.' .' .' :!: : .'.'.' .'.'.'.' ~ .'.':'.~? - .- .- :::::::i::: ::~; .- :~ : ::::::::~ :~ :::::::::: ::::::~ ~ :~: :~.!: :-.
; mmn.:.u····:··
0 .6 ... ... .......
. ...... ................ ... ... ..... .............
.
O.I!
..... . : .... ..... : , . . .. . . .. : .. .. .... . : .. ....... .. ... ... "
:
0.3 .. . ...... ; ......... ! .. . .. . ... ~ ......... : .. . ...... ; ...... .
... ......... .. .. ..: .... .... ... .... .. .: ..... .. ....... .
, . .... ... ... ..... .... ...... .. .... ...... .. .
O' 0'
101i-----~----~----;-----~----;----- 1
6 . ... . . . ::; :.':: ::: : :; ::: : : -. :::; :: : ::::::~ :: : : ::::: ; : :: :: : :'.
. ... ............ ... ... .... ... ... ....... .
10~-----~----~.----~----~:------:----- 1
6 ........ . ; .. . .. . . .. , ......... ; .... .. ... , .. . ...... ; .. ... .. .
a /:iii;~'~~'~<; ;:::~~;;;;
3 :::~ ~"';" " ": : :~::.::::.:::!::.::.: .: :: :;
~r~···~· ~~~t-~·~·~· ~· r~·~· ~-l :::9 ..
; . .• . • . . . ., 1 . .. ... ... ;
II- -0.0 · : . . .
o.e · : :~::·:::~?S+:::·::::;: : :::: ::: i : ::::: :::~ : ::::::::
M ::.::::
004
~~ : ~~ L: ~ ~ ~ ~ ~ ~ ~ j~ ~::::: ~~~ E~ ~ ~ ~ ~ ~:: ~ i: ~ ~ ~~: ::J~: ~:::::
·· .. · i ·· .. ..... ,· .... . . . . ; .. .. . . . ··; .. ··· .... ;· · ···· ..
.. .... ...., ... ...... ...... .... ... .. ... ... ..... .
0,3 ~ : : : : :: : .. ~: .-: : : .- : .. : ~ : : : : : .- : :: ~ ~ ~ : :: : : : : ~ .. : : : : .. :: :
. .... ....... .. ... ..... . .. ..... ..... ... .. .
t: ::. :::. .
02:t;======~--:---~--:--1 02t;::=====::;---~--:--~--1
·1k - 2: r - 0.9 I·.. ····; ........ ·: .. .. .. .. ·~. . , Ie - 3 : r • 09 , . .. .... ....... ............ . . .
.. ........... .... ... ...... ...... . ,., . ..... , , ..
104-----~----~----~----~~----~---- .. 10' .
. .... ..... .... ..... ... ...... .... ..... ..... ... ...... .. . .
: ;::;;;;:; i;;;;;;;: :i;;;;: ::; ;t; :;;;;;;; 1;;;;;;:;;1;;;;:; ;;.: : :::::: : : r::::~ ~ ~ ~ r~ ~ ~;~:;;;;:::;:::::;;::::::::r:::::::;
.. .... ... ... ..... ..... , .............. .... ... ... ......... . .
J: •~~~-~LL~I~~j'-.
~
2. 2. ' . ,
:: ~ :~:G~uuu
0.4 : ~'. :, ~ :.~r:::·:
:. ::, ·: ~:: i:::.~:. ~ ~ ': ~ ~ ':':'::. ~:::;:::. ~~'::: ~;: ~ ~ ~ ~~: ~:
OJ
j:.:....:.\ . ...:. :::~ ........::.j:::·:3:::::::·.·.
0.:1 : :. : : .••.....
O'
10 20 30 ~o 0;0 O.~
· ~--~----~--------~~~----~
10 ZO 30 40 50
" (.remp" $''') n (."mph .... )
416
..;-.-..-.-. -. .-."C
10-1--..- .-.-. .-. ...; ..'=.=
" ,"".---:.=.::i ..~.-.-..-."C.."'.-..-.-. -..-. -;'
. ~ ,- 0.9 -
e .. :: . : . ::~ . ':.:: .. :. ;':::::.: ':. ~'.: : .::: ':.:':::.:'.::::
6· : ~::. :.: ::: ~:::::: '. ',:. i:·:· :-:.:::::. ~:.t • :.•.:.:• :.. :..:.:.:.
• , •• • ••• 1: •• •• • , ••• ~ ........ . •• . :
L:·:·:::·::: ~
04 · . .. . ..... ........... . ..... . .... ... . . . . . .. • • • • • • •• , •• • •
, - 00 :~. ; .. . : . :: .. . ; : ... . . .. . ; . . ...... . ~ ,-0.6 '
3
... ... ... .: .. .... .. . ... .. : .... •.• .
.. ::
~ , ;
~ 2
I
~_ 0.0 ~ . ~:;:~:~:~.~.T.~.- -.~.-.-.-.-.~ 06 ..
0.3 ~ ~::. ~ ~.: ~.: {~~ ~.~ ~ ~ ~ ~.: ~.~.:.: .: .:.:.: .: ~~~~ ~ :~.:.:.: ~ j.: .: ~ ~.:.: .:.: .:!.:.~.:.:.:.: ~.:: 0.3 '.:: , :'::::C ::::::L:: ::::::L::':':':: :.'!.'.'.'.'.'.'.'.' .' i:.'.'::::::
0"
I J: - 4: - 0.(; I · .. <. · . . ·. . :........ ·
·· ~ ·
02
... .... ... ... ........... ...... .. ...
~;~::m! ; ;:
G •
....... ~
........ . ' . .. H ...... ; ...... . ...
.. p- 00 : ... : .. ... ; ~::::: .... :::: '.:: ' : ' ;::::::::: ;
4 .. '. . .. . ............ . .. ... .. ... ..
i! . :~: ~'_ /
':::.:::~~ . :::::
'-'"~
Figures 3.1 and 3.2 concern ARMA( 1,1) disturbances for simple tests
(h=l). For models subject to an autoregressive transformation with r-0.3, 0.6
and 0.9 the bounding significance points for p=O, 0.3, 0.6 and 0.9 are
mapped. We see that the bounds for overcorrection (r>p) are interlaced with
those for undercorrection (r<p). Application of the bounds test is now
handicapped because interpolation is required with respect to both p and r,
as is seen in the following example.
the assumption p~O the hypothesis is rejected. Figure 1 shows that for
0~p~0.5 the hypothesis of constant returns can still be rejected. But, as we
are cautious with respect to serial correlation, we of course should also
test the hypothesis HO: p=O The Durbin-Watson statistic turns out to be
0.839, which is significant at the 5 percent level. Although this may as well
be an indication of some other form of misspecification, we will here assume
that an autoregressive transformation of the model is the correct remedy. We
choose r=0.58 (since the Durbin-Watson statistic is used to estimate p)
which leads to the (conditional) GLS estimates
and the test statistic for .8 2 +.8 3=1 is now 2.96 . With some trouble Figure
3.2 indicates that now even for 0~p~0.6 the hypothesis that the returns to
scale are constant must be rej ected. When r-O. 3 is used, the t-value is
3.95 and now the diagram clearly shows that rejection of the null hypothesis
is indisputable for 0~p~0.6 . Apparently as a result of the transformation
of the model the estimates and tests are a bit more efficient and the
robustness has improved.
From the above we may conclude that straightforward tests on the values
of (functions of) regression coefficients may acquire more reliability with
respect to unknown serial correlation by making reference to the boundary
critical values. The ensuing test procedure guarantees Type I errors not
exceeding 5 percent under a composite assumption concerning serial
correlation. The bounds for actual type I errors given in the Tables 1 and 2
show that the significance level of the ordinary t- and F test procedures is
completely out of control when the disturbance terms are dependent. This
problem is extremely serious when the first-order autocorrelation coefficient
is close to unity, as is also shown in Kramer et al. (1991). From our
diagrams it is obvious that the probability of inconclusiveness of the bounds
test procedure approaches unity when very high values of the correlation
coefficient are not excluded. So, in cases where very serious serial
correlation may be relevant, inferences are better not based on the usual t-
and F-statistics; then also the hounds procedure falls short, and the more
418
Acknowledgements
This is a revised version of an unpublished paper presented at the European
Meeting of the Econometric Society in Athens, 1979. Helpful discussions over
the years with many people, notably Jean-Marie Dufour and Walter KrAmer, are
gratefully acknowledged, as well as the comments by a referee and the editor
of this volume.
References
Dufour, J-M. (1990), "Exact Tests and Confidence Sets in Linear Regressions
with Autocorrelated Errors", Econometrica, 58, 475-494.
Huang, D.S. (1970), Regression and Econometric Methods, New York: John Wiley.
Kiviet, J.F. (1980), "Effects of ARMA Errors on Tests for Regression
Coefficients: Comments on Vinod's Article; Improved and Additional
Results", Journal of the American Statistical Association, 75, 353-358.
KrAmer, W., J.F. Kiviet and J. Breitung (1991), "True vs. Nominal Size of the
F-Test in the Linear Regression Model with Autocorrelated Errors", this
volume.
Maddala, G.S. (1977), Econometrics, Tokyo: McGraw-Hill Kogakusha. 2. the
Neudecker, H. (1977), "Bounds for the Bias of the LS Estimator of a l.n
Case of a First-Order Autoregressive Process (Positive utocorrelation)"
Econometrica, 45, 1257-1262. 2
Neudecker, H. (1978), "Bounds for the Bias of the LS Estimator of a in the
Case of a First-Order (Positive) Autoregressive Process when the
Regression contains a Constant Term", Econometrica, 46, 1223-1226.
Sathe, S.T. and H.D. Vinod (1974), "Bounds on the Variance of Regression
Coefficients due to Heteroscedastic or Autoregressive Errors",
Econometrica, 42, 333-340.
Vinod, H.D. (1976), "Effects of ARMA Errors on the Significance Tests for
Regression Coefficients", Journal of the American Statistical
Association, 71, 929-933.
Vinod, H.D. and A. Ullah (1981), Recent Advances in Regression Models, New
York: Marcel Dekker.
Watson, G.S. (1955), "Serial Correlation in Regression Analysis In,
Biometrika, 42, 327-341.
TRUE VS. :;OYlINAL SIZE OF THE F-TEST IN THE LINEAR REGRESSION MODEL
Walter Kramer
Universitat Dortmund, Fachbereich Statistik
Lehrstuhl Wirtschafts- und Sozialstatistik
Vogelpothsweg, D-4600 Dortmund 50, FRG
Jan Kivi et
University of Amsterdam
Faculty of Economics and Econometrics
23 Jodenbreestraat, 1011 NH Amsterdam, Netherlands
Abstract
Contents
1 The problem
2 Non-robustness to autocorrelation
3 Empirical examples
4 Conclusion
1 The prob I em
This note is concerned with the null distribution of the well-known F-test of
a general linear restriction in the standard linear regression model, when the
disturbances are ill-behaved. More formally, we consider the model
y ="
VJ X(1)3(l)
+ u =, I +
X(2)(3(2) + u , (1)
where y and u are Tx1, X is TxK, nonstochastic and of rank K < T and
the design matrix has been partitioned into X(l)(Txq), 1 < q < K, and
X (2) (Tx (K-q) ) .
(2 )
for some fixed (q;.;l)-vector bel) This particular way of putting the null
hyphothesis involves no loss of generality, since any linear hypothesis about
the unknown regression coefficients can always be cast in the form (2) by
appropriate reparameterization.
The familiar F-test of the null hypothesis (2) is based on the statistic
1\ 1\
u'(M(2)_
rv '"
(u'u - u'u)/q M)u/q
F A A
(3)
u'u I(T-K) u'Mu/(T-K)
A A
where u y X~ is the vector of the residuals from the unrestricted
F > Fa (4)
- q,T-K
It is well known (see e.g. Lehmann, 1986, p. 369) that (4) is uniformly
most powerful among all invariant tests, if the disturbances are indeed
n.i.d.(O,(J"2 l , as has been required in the derivation of the test. Much less
u
is known about the test if this crucial condition fails. One can then either
think of completely different tests or check the robustness of the standard
test in nonstandard situations. In particular, statisticians have long been
concerned with the size of the test, i.e. with the true probability of the
event (4), under conditions of (i) nonnormality, (ii) heteroskedasticity, and
(iii) autocorrelation of the disturbances in the regression model.
This issue has been addressed bv various authors from different view-
points. ~euwirth (1982) and Tranquil\i and Baldessari (1988) derive conditions
on X and V such that the true size equals the nominal size. Watson (1955) and
Watson and Hannan (1956), among others, show that these can differ markedly.
421
Vinod (1976) and Kiviet (1980) derive explicit and attainable bounds for the
size of the test, given some specific V, and arbitrary X. Kramer (1989 a,b)
and Kramer et al. (1989) on the other hand fix the design matrix X and
consider the size of the test as a function of the correlation matrix V, with
the main result that the true size can be as large as unity.
2 Non-robustness to autocorrelation
where
T-l
1 p
PT-2
P 1 P
2
P P
V ( 6)
T-l T-2
P P 1
The rationale is: If the null distribution of the F-test is quite erratic in
this simple case, the same applies a fortiori to more general error models
which comprise AR(1}-disturbances as a special case.
Given (5) and (6), the null hypothesis under test can be rephrased as
(7 )
2
where we take explicit notice of the nuisance parameters ~u
true
a P(F (8 )
< 1
which is always greater than the nominal a. The only question is, how much?
Kramer (1989 a,b) and Kramer et al. (1989) show that a true = 1 is
possible. Below we characterize design matrices where this happens. To this
purpose, let j 11,1, . . ,11' and e 11,_1,1,.,(_I)T-1 1, Let <X> and
(2) denote the column spaces of X and X(2), respectively
<X >
422
Theorem: The true size of the F-test equals unity if and only if one
or more of the following conditions hold:
Comment: The condition i ftc <x> implies that there must be no intercept in the
regression (since i itself would then appear as a column of X). For all
practical purposes, case (i) therefore applies to regressions without an
intercept. By the same token, (iii) applies to regressions with an intercept,
where the intercept is excluded under HO'
Conditions (i) and (ii) involve the nominal a-level of the test. They
mayor may not hold in a given application, depending on a. In fact, they can
always be made to hold by choosing a large enough. Conditions (iii) and (iv)
on the other hand do not depend on a, i.e. they induce a true size of unity no
matter how small the nominal significance level is.
Proof of the Theorem: The rejection probability of the test can be expressed
as
u'(M(2)_ Mlu/q
P ( ~ Fa
u'Mu/(T-K) q,T-K
q
Fa u'M u ~ 0 1
q,T-K
T-K
q a
P (u'lM(2) - dM)u ~ 0 ) (where d 1 - - F
T-K q,T-K
(where f. V- 1 / 2u ~N(O,I))
(J
T 2
PC I Tt t ~t ~ 0 ) , (9 )
t=l
423
c 2 are 11
.. d 2 . 1/2 (2) 1/2
where ~t XI 1 I and ~t are the eIgenvalues of V (M - dM)V ,and
therefore also of
(10 )
The probability (9) is strictly less than one for all correlation
matrices V of the form (6), where Ipl< 1. This follows from the fact that
M(2 )_ dM has at least two non-zero eigenvalues of different sign, so by
Sylvester's law of inertia (see result 5.7.3 in Marcus and Minc, 1964) the
same is true for V1/2 (M(2)_ dM)V 1/2 . Since in addition the probability (9) is
a continuous function of p, a supremum of one can only occur as the limit of
(9) as p ~ 1 or as p ~ -1.
wi th
1 1 1
1 1
y+ ii' ( 13)
1 1
and
1 -1 (_1)T-1
-1 1 (-1) T-2
V ee' (14)
Since both y+ and V have rank one, at most one of these limiting eigenvalues
is different from zero. It can be obtained as
(15 )
and
(16 )
424
For the proof of the necessity, assume that all conditions fail. The
expressions (15) and (16) are then either zero or negative. If they are
negative, the limiting rejection probability tends to zero as ipi 4 1. If they
are zero, the situation is more complicated, because this implies that (11)
and (12) are zero, so all the 7r'S from (9) will also tend to zero, and their
limit can no longer be obtained from the limit of the 7r s. t
Consider first the case where p 4 1, and rewrite (9) as
p (17)
where 7r t /(1-p) , t = 1, ... ,T, are the eigenvalues of l~P V(M(2)_ dM). These
eigenvalues have limits equal to the eigenvalues of
0 1 T-1
0 T-2
2 T-3
W+ lim (V - ;;;+) (19)
1-p
p41
T-1 T-2 0
Since W+ has full rank, we infer that W+(M(2 1- dM) ~ 0, with eigenvalues not
all equal to zero. Again by Sylvester's law of inertia, we see that at least
two nonzero eigenvalues must have opposite signs, so the limit as p41 of (17)
must be strictly less than one.
T 7r t 2
(20)
lim P ~ l+p Et ;;: 0 ) < 1
p4-1 t=l
425
where
o 1 (T_l)(_I)T-l
1 o (T_2)(_I)T-2
-2 (T_3)(_I)T-3
w 1
(22)
(T-l) (_I)T-l o
3 Empirical examples
(23)
where K = 2, T = 11, and the x 's are from Huang (1970, p. 21, table 3.3). If
we test HO ~2 = 0, neither of the conditions (i) - (iv) in the theorem
appl ies, so the size of the test is less than unity. Table 1 reports the
exact rejection probabilities, given a nominal significance level of Q = 5%.
for various values of p (The entries under p = -1 and p = 1 are the limiting
values as p ~ -1 and as p ~ +1). The table also gives the exact rejection
probabilities of the test of HO ~1 = O. For p ~ +1, condition (iii) applies,
Table 1 shows that the test is conservative for both null hypotheses for
negative p, even producing a limiting rejection probability of zero as p ~ -1.
As expected from the proof of the theorem, the true size of unity of the test
of HO : ~1 = 0 is attained as the limiting rejection probability as p ~ 1.
426
Table 1 True size of the F-test in a simple regression model with two
regressors (percent)
a) positive rho
b) negative rho
(24)
a) positive rho
b) negative rho
For p --) +1, the size of the test equals one, like in table 1.
Table 3 shows the same as table 2, but now for T 31. The results in
table 3 are essentially the same as those in table 2.
hypothesis
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.99 1
a) positive rho
(33=(34=(35=0 5 4.1 3. 3 2.4 1.7 1.1 0.7 0 .4 0.2 0.2 0.1 0.1
(31 = 0 5 6.8 9. 1 12.0 15.8 20.6 26.8 34. 7 45.1 58.5 76.3 100
b) negative rho
(33=(34=(35=0 5 5.8 6 .8 8.2 10.2 13.3 18.0 25.3 36.5 54.7 86.4 100
(31 = 0 5 3.6 2 .5 1.8 1.2 0.9 0.7 0.7 1.2 4.8 50.0 100
4 Conclusion
References
Kramer, W., Kiviet, J. and Breitung, J. (1989), "The Null DiSlrihution of the
f-Test in Ihe Linear Regression Model wilh Autocorrelated Dislurhances·,
to appear in Statistica.
Lehmann, E., (1986), Tesling Sialislical Hypolheses (2nd ed.) New York: Wiley.
Manfred Losch
Ruhr-Universitat Bochum
Faculty of Economics
P.O.Box 10 21 48
D-4630 Bochum 1, FRG
Abstract
The extension of WOLD's distribution free fixpoint-(FP-)
estimation principle to simultaneous systems, which contain
rational expectations variables, has been shown to be a very
powerful estimation technique especially with respect to
econometric system predictions. The FP-estimator is introduced
and its asymptotic properties (consistency, asymptotic
distribution) are presented. Furthermore, EFRON's bootstrap and
TUKEY's jackknife are available to assess the variability of the
estimates. A brief review of the bootstrap and jackknife idea in
the context of the considered rational expectations model will be
given. An application of the presented techniques is illustrated
in an example.
Keywords: Rational expectations; identification; fixpoint-
estimator; bootstrap; jackknife.
Contents
o Introduction and summary
1 The RE-model
2 Identification
3 The fixpoint-(FP-)estimation principle
4 Asymptotic properties of the FP-estimator
5 Alternatives to the conventional asymptotics: TUKEY's
jackknife and EFRON's bootstrap
6 Application
References
The first main part of the paper deals with the problem of
identification of simultaneous RE-models. As PESARAN (1981)
pointed out, it is important to distinguish between exogenous
variables, say Xl , which cannot be anticipated precisely, and
430
between those, say X2, like dummies, time trends etc. which are
perfectly predictable. With respect to the not perfectly
predictable variables Xl a further distinction can be considered:
(i) The predictions Xl+ are given or (ii) the predictions Xl+ are
generated by a process which is not completely known. For case
(i) identification criteria have been derived by WALLIS (1980),
PESARAN (1981) and WEGGE and FELDMAN (1983a). Furthermore WEGGE
and FELDMAN (1983b) have shown that PESARAN's result is not
correct to some extent. The present paper is restricted to case
(ii) and, following PESARAN (1981), an autoregressive process of
known order is assumed to generate Xl. The actual structural
model is given by a linear simultaneous system containing current
expectations and lagged endogenous and exogenous variables. Under
linear identifying restrictions on structural coefficients within
the same equation an order and rank condition is presented which
contrasts with PESARAN (1981). As a simple counter-example
SARGENT's (1976) two equations model is adduced.
In the third chapter the extension to the above RE-model of
WOLD's (1965, 1980) distribution free fixpoint-(FP-)estimation
principle is introduced. This extension is not new at all. It has
already been treated in detail by LOSCH (1984) even for models
containing future rational expectations variables. It has been
shown to be a very powerful estimation technique especially with
respect to econometric system predictions; see SCHIPS (1988).
Chapter 4 deals with the asymptotic properties of the FP-
estimator. Under the given assumptions the strong consistency of
the estimator can be stated. Furthermore the asymptotic
distribution of the FP-estimator is given, generalizing
LYTTKENS' (1973) result for classical models, i.e. models without
rational expectations. The derivation of the asymptotic
distribution makes use of the presented identifiability
criterion. Since the asymptotic distribution is explicitly given,
it is possible to get estimates for the standard errors of the
FP-estimator by using the asymptotic variance-covariance matrix.
Therefore an objection to the FP-estimation of RE-models is
removed, which arose from the fact that it was up to now
impossible to assess the standard errors on the basis of
conventional asymptotics.
Since the computation of the standard errors has to be done
simultaneously and requires the inversion of large matrices this
computation is almost impossible for large models. Thus it is of
great importance that two further methods, TUKEY's jackknife and
EFRON's bootstrap, are available to assess the variability of
estimates; see TUKEY (1958), MILLER (1974) and EFRON (1979).
Chapter 5 gives a brief review of the jackknife and the bootstrap
idea in the context of the considered RE-model and the FP-
estimates. The result is presented that the bootstrap is
asymptotically as correct as the conventional asymptotics. But in
general there is good empirical evidence that the bootstrap
outperforms the conventional asymptotics; see EFRON (1979, 1982),
or FREEDMAN and PETERS (1984a,b).
The objective of chapter 6 is to illustrate the performance of
the three competing procedures in a real example. The
431
1 The RE-Model
where
zQt:= {z(t),z(t-1),oo. } ,
( 1. 3)
+Qt : = z Qt v {y ( t ) }
and
E(e(t)u(t)')=OMlxN (A5)
and
where
We get
with
where
(1.13b)
TI3:=(IN-B-A)-lDO,2 , (1.13c)
and
V(t):=(IN-B)-lU(t) (1.14)
(1.15)
u(t):= [
e(t)]
u(t)
X2 (t) : =
[ X2 (t)
X2 (t-l)
]
B .· -- [
OHlllHl
Do, 1
~H1I1N ] Bl : = [
1(1
Dl, 1
OHlll N ]
Bl
G ·. -
[ OHlll H2
DO,2 Dl,2 ] .
A· -
[ OHlll Hl
ONII Hl
~Mlll N ]
Furthermore we get
Therefore
Define
(1. 19)
nl := [ Kl OHlxN Jl
(IN-B-A)-l (Dl, l+Do, lKl) ( IN - B-A) - 1 Bl
- [
n2 :=
OHl J[ Hl
(IN-B-A)-lDo,2
1[2
(IN-B-A)-l (Dl, 2+Do, 11[2)
] and
=[ e(t)]
(IN - B) -1 Do, 1 e (t) + v (t)
Furthermore, we assume
11Th II < 1, where II II is the operator norm (AB)
11Th U=sup{lfhyl; IYI~l } .
If ~(C) denotes the spectral radius for any (nxn)-matrix C, we
get max(n:n )'nnl »= ~(i'h )S;IITI111<1 and the final form
(1. 21)
and
( 1. 23)
of (1.8) exist.
If we admit that the first Nb equations of (1.8a) are behavioural
equations and that the last Ni:=N-Nb equations are identities, we
introduce the following partitions:
B = lBll
r
B2l
B12]
B2Z
, A = [All
AZl
A12]
Azz
Nb
Hi
( 1. 24 )
Nb Hi Nb Hi
with a priori known matrices B12, BZ2, AZl and Azz,
(1.25 )
with u i (t)=ONixl
435
INi -B22 , INi -B22-A22 and E(u b (t)u b (t)' )=:2:u, bare (A9)
nonsingular.
e(t) =
i
(t)]
rebe (t) MIb
Mli
( 1. 26)
2 Identification
S= (B , (X2, U, e; Z) ) (2.1)
where
fa H -> e
436
by
fa(5):=8 for every 5 E M (2.6)
and assume 8=fe(M).
Furthermore we write
JT:= {1, ... ,T} for any TEN (2.7)
and
(2.8)
<S>T:= <y,z;JT>
distribution of the process {(y(t),z(t»;t E JT}
which can be generated by using the final forms (1.22) and (1.23)
and which depends on 5 E M.
Then we introduce the following definitions:
Definition 2.1: For 5 E M and TEN 5* is said
T-observationally equivalent to S in H
in M ) iff 5* E Hand <S>T=<S*>T.
Write
KT(5):= {5* E H; 5-T5* in M} (2.9)
Definition 2.2: Call e T-identifiable in M for 5 E H iff
faIKT(5) is constant.
Call 8 T-identifiable in H iff e is T-
identifiable in M for every 5 E M.
Definition 2.3: For 5 E M 5* is said to be observationally
equivalent to 5 in M (5-S* in M) iff S-TS*
in M for every TEN.
Write
K(5):= {S* EM; S*-S in M} = n TEM KT(S). (2.10)
Definition 2.4: e is called identifiable in M for S E H iff
faIK(5) is constant.
e is called identifiable in M iff e is
identifiable in H for every S E M
Thus the matrix x of the AR-Modells (1.8b) and the reduced form
matrices TIl, TI2, TI3 and TI4 are identifiable under the given
assumptions, since they can consistently be estimated by ordinary
least-squares.
Write
where
(a) S~S* in H .
and x=x*.
[Bl]n=[b1,n , Ql,n ]
Nn 1 H.n.l
Mn 1 Mn 1
(2.15)
1
~23B ~3B P4 B £4 B 1 NnB,B
TI3 r
I
Q3*
Q3 n
Q3 A
~3*
~3n
il3 A J
P4
rL
P4 *
P4 n
P4 A
i::j
f4 A
Nn*
1
NnA,A
I Q3+ Q3+ P4+ £4+ Nn+
L
X rL Xll
X2l
21
'J{l
X22J
Mln O
l:U.ll. 0
Mnl Mnl
Define
j
ill * Q2* il3* £4 *-Q2 *Xl2
p( n)
r
I
ill *
I ill A
0
0
Q3*
Q3 A
£4*
£4 A
(2.16)
L
440
P( n) [ ~ln 0 ~3n
and
{
p( n) an, n a priori known
n(n) [ p( n)
P( n)
1 otherwise
(2.18)
Furthermore, we define
(2.19)
Nn B := number of all a priori not known elements of [B]n
= NnB. B + Nn*
y+(tlt-1) = (B+A)y+(tlt-1)+Bly(t-1)+Do.IXI+(tlt-1)+Do.2X2(t)
+DI x( t-l)
and
f( 1) , v 0 0
0 f( 2) , v 0
fv:=
0 0 [( Nb) ,v
I( 1) , v 0 0
0 I( 2) , v 0
£v:=
r
l 0 0 I< Nb) , v
Notice that fv has rank equal to the number of its row vectors
if the same is true for TI(n) for every n=l, ... Nb. Postmultiplying
(3.9) by [INb !Xi z(t)]'fv' and summing up leads to
(3.10)
(l/T)~t=l,T{[yb(t) 00 z(t)]'fv' -~ [v[INb 00 z(t)z(t)']fv'}
or
Syb.z fv' - g [v[INb ® Szz]fv' (3.11)
= a fv[INb ~ Szz]fv' + Svb,z fv'
with
Szz: = (1/T)~t=l, T z(t)z(t)' (3.12)
Syb,z:= (l/T)~t=l,T [yl(t)Z(t)' .... ,yNb(t)Z(t)'] and
Sv b • z : = (1 / T ) ~ t =1 • T [VI (t) z ( t) , , . .. ,VN b (t) z ( t ) , ]
Similarly to (3.7) we find for the m-th equation of AR-model
( 1. 8b)
Xl • m (t) = ~m f ( m) . e Z (t ) + im [( m) . e Z (t ) + em (t ) (3.13)
= ~m f ( m) ,i X ( t -1) + im [( m) , i X( t -1) + em ( t )
where
~m respectively 1m denotes the vector which contains the
unknown respectively known elements of [~]m ,
f(m),e respectively [(m},e denotes the matrix containing the
rows of [OMxN,OMxMl ,OMxM2 ,1M] which correspond to ~m
respectively to 1m,
f(m), i respectively [(m},i denotes the matrix containing the
rows of 1M which correspond to ~m respectively to 1m.
Furthermore define
~: = [~l, ... , ~Ml b ] and i: = [11, ... , IMI b ] (3.14)
and take f(m},e, t(m),e, f(m) , i and t(m),i. respectively, to
construct the blockdiagonal matrices reo [e. fi and [i.
respectively, analogously with fv and [v in (3.8). Now the full
system of the first M1b equations (3.13) can equivalently be
given in the two stacked forms below.
(3.15)
(Xlb(t»' - I te[IMlb !Xi z(t)]= ~ fe[IMlb IX' z(t)] + (eb(t»'
and
(Xlb(t»' - i Ii[IMlb 00 x(t-l)] = ~ fi[IMlb !Xi x(t-l)] (3.16)
+ (e b (t) ) ,
~ [i[1Mlb 00 :l:uJfi'
= :l:x I b, xi [i' - 1 .ri ' [1MI b 00 :l:lIlI J[i '
(ii) (3.22)
a [v[INb 00 :l:zzJfv'
= :l:yb,z fv' - g .rv[1Nb 00 :l:zzJfv'
By (A4), E(Seb,xi)=Olx(Nb.M) and, by (A2), E(Svb,z)=Olx(Nb.K)
Thus (i) and (ii) follow from (3.11) and (3.18).
Let ~2 be the vector which contains the elements of the matrices
:l:zz and :l:yz. Then we can write (3.21) and (3.23) in the following
way:
446
and
where
Fl ~~,a X ~K+N)K --) ~~ and
F2 .Q~, a X R< K+N) K --) ~a
with
~~x ~a: ~ .Q~, a
.Q~,a:= {(~,a) (~,!,a,Q) satisfies (A1) and (AB)}
N~:= length of vector ~ and Na:= length of vector a.
With respect to (3.23) and (3.24) there exists an open subset
C~,a of .Q~,a such that
Fl C~, a X R( K + N ) K - - ) ~~
and
F2 C~, a: X R( N+K) K --) ~a:
Then we get
and
(4.2)
Theorem 4.2: Let M be the model of theorem 2.8 and assume (AFP).
Then
(ii )
If [ ~ ~'<xl
Qa. ~ Qa
is the partitioning of Q
with
Se: =fe (};e, b ® };z z ) fe'
Sv: =fv (};v, b \XI };zz )fv'
Qe : =fe (1M1 b \XI };z z ) fe'
Qv: =fv (Hv \XI };zz )fv' and
Qv, e: =fv (He, v \XI };z z ) f e
,
Remarks:
As has been pointed out by some authors (see EFRON (1979) and
BERAN (1984» there is a close theoretical relationship between
the jackknife and the bootstrap.
The basic bootstrap is described by EFRON (1979, 1982), related
papers are BICKEL and FREEDMAN (1981. 1983), FREEDMAN
(1981, 1984) and SHORACK (1982). Later on, various modifications
of the basic bootstrap approach have been considered by several
authors (BERAN (1987). RUBIN (1981), HALL (1986), EFRON (1987).
SILVERMAN and YOUNG (1988) and BANKS (1988». In essence. the
bootstrap is a procedure for estimating standard errors by
resampling the data in a suitable way, so the model is tested
against its own assumptions. The objective of this chapter is to
show how the bootstrap can be applied to the considered RE-model
(1.8).
(5.4)
(l/T)L:t=l,T e(t) = OHlxl
(l/T)L:t=l, T U(t) = ONxl
( 1 /T ) L:t =1 , T e(t)u(t) , = OMlxN
(l/T)L:t=l,T e (t )X2 (t) , = OHlxH2
(l/T)L:t=l,T U(t)X2 (t)' = ONxH2
Now consider a model like (1.8), but where all the ingredients
are known:
Set the coefficients at the FP-estimates.
Make (X2(t),u(t),e(t», t=l, ... ,T, independent, with
common distribution equal to the empirical distribution,
which puts mass l/T at each of the triplets
(X2(t),u(t),e(t», t=l, ... T.
Using this simulation model, pseudo-data can be generated
for periods t=l, ... ,T. This pseudo-data will be denoted by
stars: (Xl * (t) ,y* (t», t= 1, ... ,T .
For theoretical considerations the pseudo-data will be
successively calculated using the final forms (1.22) and (1.23),
where the (X2*(t),u*(t),e*(t»'s are independent with the common
empirical distribution specified above:
(5.5)
y* ( t) : = y*, z (t It) + ( IN - B ( T) ) - 1 U* (t )
Now pretend the pseudo-data (5.5) come from a model like (1.8)
with unknown coefficient matrices. Using the FP-estimation
principle we get new estimates ~*(T) and a*(T). The distribution
of the pseudo-errors ~*(T)-~(T) and a*(T)-a(T) can be computed
and used to approximate the distribution of the real errors
~(T)-~ and a(T)-a. This approximation is the bootstrap. The
distribution of the pseudo-errors ~*(T)-~(T) and a*(T)-a(T) can
be computed by Monte Carlo.
Theorem 5.1: Let H be the model of theorem 2.8 and assume (All)
and (AFP). Then along almost all sample sequences,
as T-)m, conditionally on the data, the conditional
law of Tl/2[(~*(T),a*(T»-(~(T),a(T»]
has the same limit as the unconditional law of
Tl /2 [ ( ~ ( T) ,a ( T) ) - ( ~ , a ) ] .
6 Application
(1981) and which has been slightly modified by the author. The
model tries to explain the movements in the real gross national
product and in the consumer price level for West Germany. Since
monetary variables have been taken into account the RE-hypothesis
is involved owing to the FISHER-equation which explains the
nominal rate of interest as a function of the real rate of
interest and of the difference between the expected and the
observed consumer price level. The model consists of the
following equations:
+ ~3[LNPV+(tlt-1)-LNPV(t)] + ~4LNAF(t)
+ ~5TETA(t) + U2(t)
References
BANKS, D.L.(1988). Histospline smoothing the Bayesian bootstrap.
Biometrika 75 673-684.
BERAN, R.(1984). Jackknife approximations to bootstrap estimates.
The Annals of Statistics 12 101-118.
Andreas Schepers
Gerhard Arminger
Ulrich L. Kiisters
Bergische Uniyersibit Wuppertal
Fachbereich 6: Wirtschaftswissenschaft
CauJ3-Str. 20 ~U4.2S
D-·'i600 Wnppcrtal-Elberfeld 1
Federal Republic of Germany
Abstract
In this paper we present models for the analysis of microeconomic data that can
handle limited dependent variables (metric, censored metric, dichotomous, ordinal)
simultaneously. In these models the dependent variables are indicators of latent variables
that are the endogenous variables in a simultaneous equation system. Such models may
be considered as generalizations of models with errors in the variables. \Ve also introduce
a new computer program (MECOSA: MEan and COvariance Structure Analysis) for
estimating the model paramet.ers. Unlike the LISCO~IP program (Muthen 1988), ?vlECOSA
can treat the case where the latent yariables are hierarchically structured. and ,'vIECOSA
can also handle more general restrictions on th.' parameters, subject only to the condition
that each parameter can be expressed as a continuously differentiable function of the set
of fundanwnt"l parameters. The program itself is implemented in GAUSS (Version 2.0)
for IB!\I-PC's and compatibles.
Keywords: Factor analysis, hierarchical structures, L1S8.EL, L1SCO!lIP, polychoric
correlcttion coefficient, probit model, tobit model.
Contents
1 Introduction
:2 The gelleral MECOSA model
:3 Estimation and implementatioll in the ~IECOSA program
4 Example
References
1 Introduction
III tltis section tbe building blocks of the general \IECOSA (MEan and COvariance
Structure clnalysis) model - simultanpol1s equation systems, factor analysis and limited
dcpPllfjellt variablr:' models - aT" brieflv rc\·ip\\'f,d. In section :2 the gelleral lVIECOSA
model (including nested Sil1111ltancolls equation models ane! factor analytic models yielding
the hicrarchical model structure) j" described. Section:} introduces the concept of
I For b,'lpftil COlllltlCnts on a previolls dnft of the p3per tbe alltllors are indebted to M. E. Sobel,
Ulli\'ersity uf Ariwll;) at TUCSOll, all 1illkllUII'Il referee ai"! to G. ROllning;, Cniwl'sity of [(onstanz. FRG.
who also ;Slll'i_-djt_~d tile d~lt<-t set .Jl1alyzcd ill this P;\PCl'.
460
fundamental parameters, the marginal maximum likelihood estimation strategy and its
implementation in the MECOSA program. In section 4, an empirical example wi th business
test data is given.
The building blocks of the general model are the following:
BT) = rx + ( (1)
y* = AT) +8 (2)
The matrix A is usually called the matrix of factor loadings, and 8 is the vector of
error terms. The factor loading matrix and the covariance matrix of the vector 8
are the parameters of interest. Note that setting A to the identity matrix I yields
an errors in variables model.
All the parameters of interest in the three building blocks are combined in a parameter
vector 13 which must be estimated from a random sample of observations {Yt, Xt }t=l .....T
where T is the sample size. N" ote that the combination of thesE' building blocks allows one
to estimate latent variable models with endogenous variables that have only been observed
on a censored, dichotomous and/or ordinal level. ]\loclels depicted above have been
considerE'd by Muthen (1979, 1984). However, the MECOSA model is more general because
it can handle hierarchical structures and more general types of parameter restrictions
(Ki.isters 1987).
461
x- variables
where N'" denotes an n-variate normal distribution. Here ~f(li) + I1(v)· Xt is the mean
structure of V; and 2::(19) is the covariance structure. Note that in contrast to the
hierarchical structures below the mean and the covariance structures are unrestricted, i.e.
not yet parameterized in a factor analytical or simultaneous equation framework. Hence,
the mean structure depends on exogenous variables x but the covariance structure does
not. The assumption of multivariate normality of (v7lxd implies that each component
(Y7ilxt) is univariate normal. If the obsenoed variable Yti is censored or ordered categorical,
the threshold models connecting Vti and Y~i discussed in section 2.4 and the assumption
of univariate normality imply that the reduced form coefficients for Y~i can be estimated
by using a tobit model for censored and an or'dinal probit model for ordered categorica.l
dependent variableso The assumption of multi\Oariate normality also allows a meaningful
specification of correlations between equations and of autocorrelated terms that is difficult
462
to formulate for other models such as the logit model (Arminger & Eiisters 1991). The
mean and the covariance structure of y;
are generated by the mixture of factor analytic
submoclels and simultaneous equation models described below.
In this subsection the unrestricted mean and covariance structures are restricted by factor
analytic and simultaneous equation parametrizations. The idea of a hierarchical structure
goes back to McDonald (1978, 1980). It is most simply explained by the concept of
second order factor analysis (Thurstone 1947). Let y* = 1)0 be a vector of observed metric
variables which are reduced to first order factors 1)1 in the model
(4)
where 171 and El are independent, and E( El) = o. Now the first order factors themselves
are reduced to second order factors 1)2 in the equation
where '1)2 and E2 are independent, and E( '1)2) = E( E2) = o. The hierarchical structure for
is given by 112, the two error terms El and
'1)0 E2 and the two level multiplicative structure
1\11\2, i.e.,
(6)
(7)
The expectecl value of rio is o. This idea may be extended from two to H hierarchical
levels, with nonzero mean values, and exogenous variables x (I(iisters 1987). In genera.!,
we assnme II + 1 hierarchical levels of latent endogenous variables '1)h of dimension nh x 1
conditional on exogenous variables Xh:
(8)
for gi\·en Ilh+l ~ nh+1 x 1 and Xh+1 rv 171h+1 x 1. The latent variable 1)0 is equal to the vector
y* = '10 whose components are connected with the obsen·ationo by threshold models. 1)h+l
is the latent endogenous vector on the level of the hierarchy. The \·ector x h+1 contains
elements of Xt, while Eh+1 rv nh X 1 is an error term. Each level is parameterized in the
matrices B"+1 rv nh xn,,, /lh+1 rv nil x 1, Ah + 1 rv TIh X Ilh+1 and f h + 1 rv 71h X mh+l· B h + 1
is a nonsingular matrix that either represents relations between the components of Tlh in
a simultaneolls equation system or serves as a scaling matrix in a variance components
model. The vector Ph+1 is the general mean of the vector 1)h. The matrix f h + 1 consists of
the coefficients of the linear regression of '1)h on Xh+l, while Ah+J is usually either a matrix
of factor loadings or a design matrix in a variance component model. :-lany special cases
of this model arc gi\·en in IGisters (1987).
463
Successive substitution of level h + 1 in level h yields the reduced form of the system for
y~,
H-1 H-1 H
y* = T)o = ( IT (FhC h))FHi1(II) + (IT (FhCh))FHKHx(H) + (IT (FhCh))T)(H+l)' (0)
h=l h=l h=l
In this formulation the substitution is achieved by using the following structured matrices:
Ir
0 0
(I
In! 0
Fh = ( 10)
0 I nh _ 2
0 0 Bh
n
0 0 0
In! 0 0
[1
Cit = (11 )
0 I nh _ 2 0
0 0 Tflh _!
n
0
( f,
K, ~ 1 f2
( 12)
0
P(It) = T
( Ill"" ,Ph
T)T
(13)
T
,7:(h)= ( X 1 "",X"T)T (1·1 )
T T T)T (1,5)
11th) = ( c1"'" c"_l' IIh-l
By using obvious substitutions for ;, IT and C the reduced form may be written as:
y~ = ~( + ITx + E* (16)
H-l
,(v) = (IT (FhGh))FHJ-i(H) (18)
h=l
H-l
IT(19) = (IT (FhGh))FHJ(H (19)
h=l
and
H H
~(19) = (IT (FhG h))f2(V)( IT (FhGh))T (20)
h=1 h=l
In the last equation the covariance matrix f2( 19) is block diagonal
(21 )
(22)
J
,(v) = IT Li(19) (23)
i=l
J
IT ( v) = IT M j (() ) (24)
J=1
J\ J\
~(D) = (IT N k(D))f2(v)(IT Nk(v)f (2.5 )
k=! k=l
Another \\'ay of looking at the last three equations is to consider the multiplicative
structure as a completely general structure for ,(v), IT(v) and L:(v) parameterized in v,
rather than a consequence of the model fOIllulated in equation (8). This is the approach
taken in the ~vIECOSA program. In the third step of the estimation procedure the user
has to specify the number of levels and the matrices L i , il1j , f2 and Nk as functions of !?
465
Depending on the measurement level each component 1J1l is connected with 1Jri by one of the
following measurement relations. For convenience of notation the case index t = 1, ... , T
is omitted.
Vi == yt (26)
• Viis ordered categorical wi th unknown thresholds Ti,l < Ti,2 < ... < T,,I\',
and categories Yi = 1, ... , !{i -'- 1 (ordinal probit relation, McI<:elvey and Za-
voina 197.5).
Note that for reasons of identification the threshold Ti 1 is set to 0 and the variance
of the reduced form error term 0-;( 0) is set to 1. '
• Classified metric variables may be treated analogously to the ordinal probit case with
the difference that the class limits are now used as known thresholds (Stewart 1983).
• Vi is one-sided censored with a threshold value Ti,l known a priori (tobit relation,
To bin 19,58).
if vi > Ti,l }
(28)
if vi :::; Ti,l
• V, is double-sided censored with threshold values Ti,l < T,,2 known a pnorl (two-
limit-probit relation, Rosett and :\elson 197.5).
1. In the first stage the threshold parameters T, the reduced form coefficients I and
II of the regression equation, and the reduced form error variance 0-; of the i-th
equation are estimated using marginal maximum likelihood. Note that this first
stage is the estimation of the mean structure without restrictions as in equation (3).
The parameters to be estimated in the i-th equation are the thresholds denoted by
the vector Ti, the regression constant denoted by Ii, the regression coefficients, i.e.
the i-th row of II denoted by IIi. and the variance denoted by 0-;' The loglikelihood
function that is maximized with respect to {Ti, Ii, IIi., o-?} of the i-th equation is
given by
T
fi(Ti,I;,II;.,oD = LlnP(Yt;!Xt) (30)
t=1
The formulation of P(Vt; IXt) depends on the measurement level of the observed
variable Vi. If Yi is metric then P(YtilXt) is the univariate normal density with
expected value Ii + IIi.Xt and variance 0-;,
i.e.,
(31)
in which <p(vlp, 0- 2 ) denotes the univariate normal density. Another example is the
ordinal probit model that is applied if Vi is an ordinal variable. In this case the
probability that Vti = /.., must be computed as:
(32)
The probability of the last equation is the basis of the likelihood used in the ordinal
probit model (McKeh'ey and Za\'oina 1975). If other models are used such as the
tobit model (Tobin 1%8) for dependent variables censored on one side only or the
two limit probit model (Rosett &; Nelson 1975) for dependent \'ariables censored on
two sides then the probabilities P(YtiIXt) have to modified accordingly.
A solution to the likelihood equations obtained by setting the first derivatives
of C(Ti.I"IIi.,o-1) to 0 is computed in I\IECOSA by applying a Quasi-Newton
algorit hm proposed by Polak (1971) using analytical first deri\'ati yes. The second
order derivatives are approximated by computing the crossproduct of the first
order derivatives. The first estimation stage yields strongly consistent estimates
of Ti, Ii, IIi and o-f for each equation i. Regularity conditions and the proof of
strong consistency are given in Ktisters (1987).
2. In the second stage the problem is to estimate the covariances of the error terms in
the reduced form equations. Note that in this stage the covariances are estimated
wi thout parametric restrictions. Since the eITors are assumed to be normally
distributed (lnd strongly consistent estimators of the reduced form coefficients
467
have already been obtained in the first stage the estimation problem reduces to
maximizing the loglikelihood function
T
£ij(aij) = L InP(Vti,Vtjlxt, 7i,1i, It, 0-;, 7j,';fj, TI j , o-],aij) (33)
t=l
in which P( Vti, Yt; Ixt, 7;, ';/i, TIi.' 0-;' Tj,';(j, TI j., 0-], ai)) is the bivariate probability of Vti
and Vtj given Xt and the reduced form coefficients. A typical example of this bivariate
probability is the case when Vi and Vj are both ordinal. Then the probability that
. Vti = k and Vtj = I is given by:
(31 )
:3. In the third stage the vector K of thresholds, and the reducecl form regression
coefTicients ane! the reduced form covariance matrix are written as a function of
t he structural parameters of interest, collected ill the pa rameter vectorl? The
parameter \·ector () is thell estimated by minimizing the quadratic form
(:3.5)
In the MECOSA program the function QT( 1J) is minimized using the Davidon-
Fletcher-Powell algorithm (Luenberger 1984) with numerical first derivatives. The
MECOSA program offers two features that are available neither in LrSREL 7 nor in
LrSCOMP. First, the general multiplicative structure of equations (23) - (25) can be
used, with up to 10 levels. The individual matrices Li(1J), M;('r'J) , Ni(1J) and n(1J)
are specified by the user. Second, the parameter vector 1J itself may be thought of
as a continuously differentiable function of a vector ~ of fundamental parameters.
Hence, the 1J's may be restricted in a very general way. Typical examples are the
restriction that 1J k must be positive, that the 1J's are ordered or that the 1J's lie
within given intervals. All of these restrictions are easily formulated in terms of the
ts and are easily implemented in MECOSA.
4 Example
The MECOSA approach is demonstrated by analyzing business test data that have
previously been considered by Ronning and Kukuk (1988) using a model of Nerlove
(1988). Nerlove models future changes in the demand for the goods of a firm t with
the adaptive expectations hypothesis:
(36)
D;,s is the expected demand for the time period s + 1 for the goods of firm t, the
expectations being formed at time period s. D; 3-1 is the same variable lagged one
period. Dt,s is the demand observed at time s. ~t,s is an error term. The coefficient
(3 describes how quickly the firms adapt to discrepancies between expected and
observed demand.
For the estimation of (3 we use the data of the German business test conducted by
the IFO Institute for the third quarter of 1980 collected from T = 5389 German
firms. Since only one time point is observed, s is set to 1. The time index is
equal for all cases and hence is eliminated for the sake of simplicity. The observed
demand for the third quarter of 1980 of firm t, t = 1, ... ,T, is denoted by D t . The
expected demand for the fourth quarter of 1980 of firm t is denoted by D;. The
lagged expected demand for the third quarter of 1980 of firm t is denoted by D;(_I)'
Hence, the model for the t-th firm is given by
(37)
If D;, D;(_l) and D t were known, a simple regression analysis would have to be
performed. However, we have only ordinal measurements Vt3, Ytl and Yt2 of these
variables with the categories negative change (1), no change (2) and positive change
(3). Additionally, we have the restriction that in the equivalent regression equation
for firm t, t = 1, ... ,T,
(38)
the coefficient 0:1 is equal to (1 - 0:2)' To put the equations (37) and (38) into the
1IECOSA program we consider the three variable simultaneous equation system for
each firm t
(39)
469
(40)
(41 )
with ordinal threshold models for the observed variables Ytl as an indicator for
D~( -I)' YtZ for D t and Yt3 for Dr The possible combinations of these three variables
and their observed frequencies are given in table 1.
Table 1:
Observed frequencies of lagged expected demand (YI), observed demand (yz)
and expected demand (Y3) for the third quarter 1980 of T = 5389 firms.
YI Y2 Y3 f YI Y2 Y3 f
1 1 1 24 2 3 2 456
1 2 1 52 2 1 3 43
1 3 1 29 2 2 3 437
1 1 2 43 2 3 3 588
1 2 2 97 3 1 1 11
1 3 2 39 3 2 1 20
1 1 3 19 3 3 1 33
1 2 :3 46 3 1 2 51
1 3 3 56 3 2 2 'Y,)
~o~
2 1 1 45 3 3 2 173
2 2 1 71 3 1 3 59
2 3 1 43 3 2 3 393
2 1 2 162 3 3 3 914
2 2 2 1233
In the first step we estimate the parameters of the reduced form equation system
(42)
I
Table ~: Results from the first estimation stage
I Thresholds
!
Tl~ T22 TJ2
loS 1cl 1.5.Jc:3 1.G1~
Regression Constants
[I "'/2 ~i3
I
PolycllOric Correla bon ~Iatrix
VI Y2 !/3
1
.2SS 1
.451 .407 1
Unlike Ronning and Kukuk however, we are also able to compute the weight matrix
TV which is an estimate of the asymptotic cO\'ariance matrix of the reduced form
parameters. This covariance matrix is given in table :3.
0.761.Jc
0.09e15 0.668;)
0.lS62 0.11.56 0.799j
o.,'ieiS 1 0.Oc19~ O.l1Si O.CUi
0.056:) 0.185 cl O.ln28 0.080:' 0 ..5968
0.1~20 0.0(j~)2 0.6193 0.1(;20 0.0960 0.i:30!J
0.07:jcl 0.0807 0.02H 0.0618 0.0713 0.OE1S 0.~elS6
0.10i~ O.O~lli 0.10!)! 0.09U 0.0 I !):) O.mlO:2 0.0:513 0.lS16
0.0442 0.0938 0.1:3131 0.0:381 0.0151 0.1 ~69 (i.0118 0.01-+:3 0.21:33
In the third estimation stage tbis covariance matrix is used to calculate the weighted
least squares estimates of Ctl and Ct2 if the coefficient:' of equation (:38) ,l1'C not
restricted and the estilllillt' of J of (''Illatioll (:37). The results arc giW'll in table -1.
471
The estimates of 001 and 002 agree with the estimates of Ronning and Kukuk (1988).
However our standard errors are larger and the t-values are accordingly smaller since
the MECOSA approach takes explicitly the fact into account that the estimates from
the first estimation stage are stochastic and not fixed. Our final result is the estimate
~ = 0.4797 for f3 which is quite different from the unrestricted value pointing to a
possible misspecification of the model.
To check for misspecification we have included a yVald test for Ho : 001 + 002 = 1
against HI : 001 + 002 i- 1. The Wald statistic equals 429.0187 wich is significant on
the 0.05 test level. Hence, Ho has to be rejected.
The present - very simple - model will be extended to deal with multiple periods
and exogenous variables such as seasonal dummy variables in the future.
References
Arminger, G. (1979). Faktorenanalyse. Stuttgart.
Arminger, G. and Kusters, U. (1991). Toward a General Model for Longitudinal
Change in Categorical and Continuous Variables Observed with Error. In: Dwyer,
J. (ed.), LOTlgitudinal Data Analysis in Epidemiology, Oxford University Press.
Oxford. in press.
Daganzo, C:. (1979). A1ultinomial Pro bit - The The07'y and Its Application to
Demand Forecasting. New York.
Dupacm'<i, .1. and \Vold, H. (1£182). On some identification problems in l\IL modeling
of systems with indirect observation. In: .1oreskog, K.G. and Wold, H. (eds.).
Systems under Indirect Observation - Causality * Structure * Prediction - Part
II. Amsterdam, 293-315.
Joreskog, E.G. and Sorb011l, D. (1988). LISREL 7 -- A Guide to the Program and
Applications. SPSS Inc., Chicago.
Ei.isters. U. (1987). Hierarchische JIittelwert- llnd I\ol)arian::sirllkturmodelle mit
Illchtmeil'ischen endogenen variablen. Heidelberg.
LueniJerger, D.C. (1984). Linear and .Vonfinear Programmil)(f. Reading, l\Ias-
~ilC huset t s.
472
Abstract
Starting with critical remarks on the compatibility of the probability approach adopted
in econometrics with the widely accepted requirements of Popperian epistemology, a
new, non-stochastic approach to macroeconomic modelling is presented. This new
methodology is based on the assumption that coefficients in a structural relationship
vary erratically within numerically fixed interval boundaries. Within this framework,
allowance for observational errors can easily be made by extending the interval concept
to the variables. Thus, the stochastic foundations of the traditional econometric
approach are replaced by elements of interval mathematics. Since conditional
predictions of the endogenous variable derived from such a model are fixed intervals
as well, econometric interval hypotheses can be tested effectively even in situations of
scarce empirical data, as is typically the case in the macroeconomic field. From the
falsificationist point of view, this constitutes a clear advantage of the interval model,
compared to the stochastic model. The issue of "estimation" of the interval coefficients
is also attempted by using epistemological instead of statistical criteria. The suggestion
is to select from all a priori feasible and hitherto corroborated structures of a given
econometric interval model, the one with the highest empirical content. It is shown that
this decision rule results in a linear programming task, if the economic relationship is
linear. Finally, some implications of the interval model are indicated.
Keywords: Econometric methodology, Non-stochastic modelling, Interval mathematics,
Linear programming, Errors in variables, Selection of structures.
Contents
Almost fifty years ago, Haavelmo wrote his famous monograph on "The Probability
Approach in Econometrics", in which he pleaded for the acceptance of a stochastic
framework in econometric models. He argued that, if one wants to apply mathematical
methods of statistical inference such as parameter estimation and hypothesis testing,
economic theories necessarily have to be formulated in a probabilistic way. In the
meantime, the stochastic approach has become the fundamental, mandatory paradigm
of econometric theory formulation. However, there has been more and more criticism
of this paradigm from the epistemological side. According to the Critical Rationalism
of Popper, the probabilistic formulation of hypotheses entails the risk of immunizing
these hypotheses against experience. This applies particularly to those scientific fields
where the empirical basis for hypothesis testing is small.
Unfortunately, the analysis of relationships bet\veen macroeconomic time series
constitutes to a large extent such a field, since the data occur only at yearly, quarterly
or, at best, monthly intervals. Moreover, these data are usually subject to considerable
observational errors. For this reason, Blaug characterized the working philosophy of
science of modern economics as "innocuous falsificationism" [I]. In his words, the
diagnosis of the crisis in modern economics is this: "Unfortunately, we lack both reliable
data and powerful techniques for distinguishing sharply between valid and invalid
propositions in positive economics ... These weaknesses, not so much of theoretical
econometrics as of the actual procedures followed by applied econometricians, go a long
way toward explaining why economists are frequently reluctant to follow their avowed
falsificationist precepts. In many areas of economics, different econometric studies reach
conflicting conclusions and, given the available data, there are frequently no effective
methods for deciding which conclusion is correct. In consequence, contradictory
hypotheses continue to coexist sometimes for decades or more." [2J
What kind of conclusion has to be drawn from this diagnosis? Of course, dismissing
stochastic methods from macroeconomics cannot qualify as a reasonable research
strategy, as long as no substitute with a better epistemological foundation is available,
which allows as refined investigations of macroeconomic relationships as do the existing
macroeconometric models based on probability theory. Also, the obvious discrepancy
between the requirements of Popperian epistemology and traditional econometric
methodology cannot serve as a pretext to reject Critical Rationalism as an inappropriate
epistemological basis of economic theories. Instead, this diagnosis should be a motiva-
tion to search seriously for alternatives to the probability approach in econometrics.
Although there has been little discussion of non-stochastic approaches to empirical
macroeconomic modelling in the past, few authors have addressed this issue. In a
comprehensive article on the statistical methodology in economics, Grohmann presents,
apart from deterministic hypotheses (which obviously cannot give a perfect explanation
of human behaviour) and stochastic hypotheses, a third type of hypotheses described
as a conjectural approach [3]. In line with the requirements of Critical Rationalism,
conjectural statements are characterized by a logical structure which emphasizes the
aspect of testability rather than the aspect of hypothesis formation, the latter playing a
decisiw role in the traditional stochastic concept. To state it briefly, the degree of
testability of an economic hypothesis is a decreasing function of the amount of data
necessary for its potential falsification. Therefore, econometric models, while taking into
account the imperfections of economic explanations by allowing a certain variability in
the mathematical relationships, must set well-defined boundaries to the extent of
variability to be tolerated.
475
+ b- , a
+
(3) tal + [bl [a + b+J
[a , a +] * [lib +, lib - I
-
(6) [ a] / [b 1 (0 rj [b I)
The interval arithmetic operations + and * are both commutative and associative.
However, the distributive law is only valid in the weaker form of subdistributivity:
(8) tal e [bj, [e] e [d] ~ tal w [el e [b] ,,] ld].
476
The neutral elements of interval addition and multiplication are [0,0] and [1,1 J,
respectively, but the inverse elements generally do not exist.
Interval addition and subtraction can easily be generalized for interval vectors
[a] E In(lR) and interval matrices [AJ E Inxm(IR) by application of the respective operation
to each component. Similarly, an interval matrix is multiplied by a scalar interval
[k] E I(lR):
and the matrix product of two interval matrices [A] E Inxm(lR) and [B] E ruxh(lR) is defined
by:
m
(10) [A) * [B] (I [a iJ·] * [b J· k ]) E rnxh (lR)
j =1
The most important interval extensions are the joint range extension [F] defined by
and the natural extension [F] of a rational function f, which results from the substitution
of interval arguments [xi) for real arguments xi in the function term [10]. Both these
interval extensions are continuous functions [II], and, if f is a rational function, the
following relation holds:
However, if_each variable x i occurs only once in the function term of f, the interval
extensions [F] and [F] are identical [12].
As described above, the interval approach is based on the assumption that a coefficient
b i in an econometric relationship generally varies between the boundaries of a specified
477
interval [bi]. As a result, the exact value of the endogenous variable (conditional on the
values of the explanatory variables) remains undetermined, but captured in a well-
defined interval, providing continuity of the underlying function. Mathematically, the
interval model is expressed by the general equation
(IS) [y] =
-
[F](x,[b]) =
-
[F](x" ... ,xj,[b,l, ... ,[b k ]) ,
which is obtained from the according deterministic relationship y = f(x 1, ... ,Xj>b, , ... ,~)
by a joint range extension with regard to the coefficients bi .
If the underlying function f is rational with each coefficient b i occurring only once in
the term, the joint range extension is equivalent to the natural extension, which makes
the formal treatment of the interval model much easier. Specifically, this applies to the
linear interval model:
k
(16) [y] = I y=x'blb db]} = [min x'b, max x'b] = x' [b] = I [b.] x.
b€[b] b£[b] i=l 1 1
The interval model can easily be generalized by introducing a second joint range
extension with regard to the explanatory variables, thereby allowing for (bounded)
observational errors in the empirical data:
(17) [y] =
-
[F] ([x], [b]) resp [y] [x]' [b] .
Figure I:
Predictions with a linear interval model using exact versus interval explanatory data
y y
x x
478
From the epistemological point of view, it is very important to explain clearly what the
econometric hypothesis associated with a certain structure of an econometric model
states about reality. Though this is rarely discussed in traditional stochastic
econometrics [13], an exact definition of econometric hypotheses is absolutely necessary
as a precondition for effective empirical testability, because an empirical statement must
explicitly state what is permitted and what is prohibited [14]. For this reason, we define
an econometric interval hypothesis (EIH) as a statement of the logical form:
(18) for t E S .
Translated into words, this statement declares for all empirical situations t within the
scope S of the model that the numerical value of the endogenous variable y t is
contained in a prediction interval which is mathematically determined by the vector of
explanatory variables Xt and the fixed structure [b*]. The logical form of this statement
ensures a high degree of testability, since the compatibility of an EIH with economic
reality can be judged on the basis of single observations, which is very important in view
of the rare opportunities to obtain suitable data for testing macroeconomic theories.
Now, suppose an EIH is confronted with a set of exactly observed test data
{ (x t,Yt) I t ET } defined by the index set T c S. Then the logical form of statement (18)
implies the test rule that the EIH is ultimately refuted, if
(19)
holds for at least one observation t E T. Otherwise. the EIH is tenable further on as a
tentative (or conjectural) statement. Only by withstanding extensive and rigorous
testing of this kind docs an EIH gradually corroborate.
However, it might be argued that the above rule of falsification is too strict compared
to the inaccuracy of the basic statements in economics, especially in view of highly
aggregated macroeconomic data. To avoid the refutation of an EI H by mistake,
researchers should therefore allow for an appropriate error margin, which depends on
the specific data employed in the test. Consequently. the test decision has to be
modified for the case of interval data {([x ],[y D I tET }. This is done by requiring for
the falsification of an EIH that even t t
has to be true for at least one observation t E T. ThIS means that, taking into account
the error margin of the data, an EIH is not refuted until statement (18) is proven to be
false for any possible combination of values (x t ,y t ) E ([x t l.[Y t D. This conservati\e
479
formulation of the test rule is necessary because of the asymmetry of ultimate rejection
and temporary retention. So, as a consequence of using flawed data, it has to be
accepted that sometimes falsifying facts are not detected.
Obviously, the size of w([y*]) in its turn is determined by the specification of the EI M,
the selected structure [b*], and the numerical values of vector x. Since the latter factor
does not refer to the EIH itself, the relative empirical content of an EIH is defined in
terms of average prediction intervals using some weighting function g(x) [16]. Let HI
and H2 be two EIHs corresponding to the structures [hI] and [bz] of the same EI M [17],
then the empirical content of H I is equal to or higher than that of H2, if the following
inequality holds:
According to this definition, the borderline case of the deterministic hypothesis where
w([y *]) = 0 has the maximum empirical content; however, high empirical content by
itself is not a criterion for the appraisal of a scientific hypothesis unless coupled with
corroboration. Essentially, scientific theories qualify by withstanding rigorous
falsification attemps despite a high risk of failure (reflecting high empirical content).
This consideration provides a good starting point for tackling the issue of structure
selection (or "estimation") in the ElM. For the formal treatment of this problem, assume
an EI M specified by function [F] representing an economic relationship between varia-
bles x and y, certain a priori restrictions defining M, and the scope S. Moreover, the
empirical evidence E C S is defined as the set of all relevant empirical data available at
the present time [18]. Given this basis, the selection of the "best" structure of the EI:-"1
is attempted from the epistemological point of view, leaving aside statistical
optimization criteria used in traditional econometric estimation.
First, all a priori structures which are inconsistent with the empirical evidence are
eliminated in the selection process. Hence, fundamentally, only those structures
[b] EM are eligible, which satisfy, for all tEE, the following condition of consistency (for
both exact and in terval da ta, respectively):
[ 19].
480
!j>([b]) = J w([F](x,[bJ))g(x) dx
JRk
(23)
[b] E C(E)
Since the objective function takes no negative values, this optimization problem is
always solvable, provided C(E) is not empty. However, the optimal solution is not
always easy to obtain. Difficulties of an especially· serious nature arise if [F] is not
equivalent to a natural extension of a rational function, because in this case the above
optimization problem cannot be expressed in terms of the interval bound.aries
b- and b+; i.e., it does not reduce to an optimization problem in the real space JR Zk.
Structure selection becomes comparatively simple in case of a linear ElM based on (16),
with a finite number of linear a priori restrictions on the coefficients,
r ,
k
(25) g(x) II g. (x.) .
1 1
i=l
Then, the resulting optimization problem (23) takes the form of a linear programming
task. In particular, the objective function becomes
- + + -
k\ +-
(26) !j>(b,b) c'(b -b ) = L c.(b.-b.)
i=l 1 1 1
<
o ,
whereas the latter group comprises constraints refering to the available observations
(for both exact and interval data, respectively):
- +
[y] E X[b] for all X E [X] [X ,X ] .
It can be shown that these empirical constraints can also be written in the form of linear
inequalities, making use of
[20 ]
(29) X[b] (Xl+Xz) (b] Xl [b] + Xz[b]
where the matrix X is expressed as the sum of non-negative matrix X 1 and non-positive
matrix X Z defined by
Thus, the coefficients of a linear ElM can be obtained as the solution of the linear
programming task composed of items (26), (27), (24), (31):
(32) b- b+ <
0
<
Rlb + Rzb+ r
- +
Xlb + XZb ~
Y
-x2b- Xlb
+ ~ -y
Allowance for interval data is easily made by substituting the appropriate boundaries
Xl' X;, y- and X~, X;, y+ for the exact data Xl' X Z' Y [21].
482
6 Concluding Remarks
As can be seen from this representation, the shift from well established statistical
methods of inference to empirical methods guided by epistemological criteria is not only
desirable in order to improve the testability of econometric hypotheses: It is also
practicable from the mathematical point of view. However, it must be granted that the
"estimation procedure" loses some elegance. Moreover, the technique developed in the
previous section is not intended to provide an ideal method for determining the
numerical values of the coefficients in the ElM. It is simply a rationale to obtain those
numerical \"alues which produce the most informative hypothesis, i.e. the statement with
the highest risk of falsification among all possible hypotheses not yet defeated.
Therefore, the solution of optimization problem (23) should be interpreted merely as a
useful guideline for the numerical setting of an EIH. After all - irrespective of the
"estimation" method - an EIH, like any other hypothesis, remains a more or less vague
conjecture, which only corroborates through repeated and rigorous empirical testing.
Finally, two implications of the ElM seem worth mentioning [22]. First, an EIH may
be used not only in macroeconomic forecasting, but also in policy decision making. Then
eq uation (15), with a fixed target interval [y], has to be solved for some policy instru-
ment variable xi. Again, the solution is an interval [x i], unless the target interval is
inadequately narrow compared to the forecast precision of the EIH. In the latter case,
no setting of the instrument variable exists which guarantees the realization of the
target. This result ref1ects the almost evident rule that the precision of policy target
formulation has to correspond to the stringency of the supposed inf1uence of the in-
strument variable on the target variable.
References
Maier, T. [1987]: Interval Input-Output Analysis as the Basis of a Price System with
Inexact Data Material (in German), Jahrbucher fur Nationalokonomie und Statistik,
Vol. 203, pp. 274-294
Moore, R.A. [1966]: Interval Analysis, Prentice Hall: Englewood Cliffs (New Jersey)
Popper, K.R. [1959]: The Logic of Scientific Discovery, Basic Books: New York
Rohn, J. [1980]: Input-Output Model with Interval Data, Econometrica, Vol. 48,
pp.767-769
Footnotes
[4] For convenience, the discussion of the interval approach is restricted to the case of
a single economic relationship, in which the endogenous variable is a continuous
function of several explanatory variables.
[5] For this section see textbooks of interval mathematics by Moore [1966], Alcfeld
Herzberger [1974] and Nickel [1978].
[8] There are, e.g., applications of interval mathematics to input-output analysis; see
Rohn [1980], Maier [1985] and [1987], Lorenzen! Maas [1989].
[10] See Moore [1966], p.18, and Nickel [1978], Part I, p. 13. Note that the natural
extension is not unique, because [F] depends on the specific representation of term f.
[16] g(x) has to fulfill the same requirements as a density function in probability theory
(non-negativity, unit integral).
[17] This means that both EIHs must be commensurable in the sense that they refer to
the same endogenous variable; see Wewel [1987], pp. 130-131.
[18] The necessary condition for obtaining unique estimation results in the stochastic
model, namely that the number of available observations, i.e. the cardinality of E, has
to equal or exceed the- number of coefficients, applies to the interval model as well.
[19] Note that, in order to be on the conservative side in case of interval data, all
structures which are possibly refuted by the empirical evidence have to be excluded
from the selection (which is asymmetric to test rule (20)).
[20] Here the distributive law holds, because at least one of any two corresponding
components of matrices Xl and Xz equals zero.
Jiirgen Bartnick
FernUniversitat Hagen
Lehrgebiet Statistik und Okonometrie
Feithstr. 140, D-5800 Hagen 1, FRG
Abstract
A measure of interdependence of a system of equations facilitates e.g. Monte Carlo studies of small
sample properties of estimators and tests.
To find such a measure we reorder the equation system such that the sum of the absolute values
above the the main diagonal of the coefficient matrix is maximal. If in the reordered system there
are only zeroes below the main diagonal, the system is recursive and the degree of interdependence
is zero. We use the 'degree of linearity' proposed by Helmstiidter 1965 for an equivalent problem
(triangulation of an input-output matrix) to measure the degree of interdependence. The measure
is refined later: First the classes of interdependence are constructed (using the transitive closure
operation) and then the measures of interdependence are computed within these classes. This
means block-triangulation of the coefficient matrix and corresponds to reordering the equation
system so as to achieve a blockrecursive system of equations.
Keywords: reordering of equations; optimal ranking; linear ordering problem; measure of inter-
dependence; degree of linearity; triangulation; block-triangulation, blockrecursive system.
Contents
1 Linear econometric equation systems
2 The position matrix of an equation system
3 Construction of the transitive closure of the position matrix
4 Classes of interdependence
.5 The optimization problem
6 Construction of a ranking for the classes of interdependence
7 Measure of interdependence within a class of interdependence
8 Critical remarks on the method
9 Appendix: Some short programs in PASCAL
References
where
C is a real matrix of type m X m, nonsingular, and all its elements on the main diagonal are equal
to -1 (the coefficient matrix of the system),
Yt is a vector with m real components, containing the unlagged endogenous variables of the model.
B is a real matrix of type m X k,
488
Xt is a vector with k real components, containing the predetermined variables of the model: exo-
genous and lagged endogenous variables,
Ut is a vector with m real components, containing the error terms,
t is time.
Two variables Yi and Yj are called interdependent if Yi depends on Yj and vice versa. They are
directly interdependent if the elements eij and eji of the coefficient matrix C are both different
from zero. But it can happen that interdependence is not recognized at once from the coefficient
matrix, because the interdependence is indirect. Therefore, we construct the transitive closure
of the position matrix (section 3). In a first approach we reorder the equation system such that
the sum of all absolute values of the coefficients above the main diagonal is maximal. If one gets
only zeros below the main diagonal, the system is (with respect to the form of C) recursive and
the degree of interdependence should be zero. The "degree of linearity" defined by Helmstadter
as the measure of success for triangulating input-output matrices corresponds to the "degree of
recursivity" for econometric equation systems. Therefore, we can use this degree to formulate our
measure of interdependence.
But this first approach is very hard to compute for large equation systems. Therefore we refine our
method. \Ve first compute classes of interdependence by taking the transitive closure of the position
matrix. \Vithin these interdependence classes we use our first approach. Thus the problem is to
block-triangulate the econometric equation system. Triangulation does not mean transformation
into an exact triangular matrix (that is not possible in general) but it means transformation into a
matrix, where the sum of absolute values of the elements below the main diagonal of the coefficient
matrix C is minimal. This definition was given by Korte and Oberhofer (1969).
Pij 0 if ei) = 0
pi) else.
This matrix is called relation matrix by Krupp (1976) and position matrix by Bodin (1974). There
is another situation, if P is chosen before C is estimated. Then eij = 0 a priori, if Pij = O.
If one considers the position matrix P instead of the coefficient matrix C some information is lost.
We will later return to the coefficient matrix C. But for the moment using the position matrix has
some advantages: Numerical values are ignored, only direct dependences are indicated. Variable
Yi is directly dependent on variable Yj iff Pi; = 1. (This follows from equation number i because
of Pii = 1.) \Ve could even use Boolean values (false and true) as components of P, but for later
computations it is simpler to use the integer values zero and one.
Example:
\Ve consider the example from Gruber (1968), pp.38-41. There are m 5 endogenous variables.
The coefficient matrix is
C = (~ ~ 0 0 *
1 1 -1
~ ~:~~ ~)
0.13 0
* 0
o 0 0 *
(we use * as a neutral symbol for -lor 1 in the main diagonal).
489
id)
The position matrix in this example is
p~(H
Dependence should be a transitive relation, therefore we build the transitive closure. In the example
Y2 depends directly on Y4 and Y4 depends directly on Y3, therefore (by transitivity) Y2 depends on
Y3. General definition: The transitive closure is the smallest transitive extension (of a relation or,
in our case, a matrix with elements 0,1).
4 Classes of interdependence
Two variables Yi and Yj are interdependent if tij == 1 and tji == 1. The relation of interdependence
is an equivalence relation and we are interested in equivalence classes for this relation, see Krupp
(1976).
In the example the classes of interdependent elements are obvious. The first class is formed by 1,
2, 3, 4, i.e. the first four structural equations of the .5-equation-model taken as an example, and
the second class consists only of the element 5, i.e. of the 5th structural equation.
In the appendix we give a PASCAL procedure giving all classes of interdependent elements for a
given transitive closure T of a position matrix P.
For the example the output is:
1 234
5
The elements of the classes of interdependence are ordered by their numerical value in the computer
output (of course, each class of interdependence is a set and there is no order of the elements). In
the computer output the classes are ordered by the numerical value of the first element. But we
are interested in an optimal linear ordering. We will consider the internal order of the classes of
interdependence later. We first have to define an optimality criterion resp. an objective function.
As we now consider classes as elements, it is adequate to take a more abstract point of view. Up
to now we only considered econometric equation systems.
490
a) alternatives
b) production sectors
c) endogenous variables.
In all three cases the problem can be extended to weak relations (search for optimal preference-
indifference-relations, block- triangulation of an input-ou tput matrix, block- triangulation of an equa-
tion system). These problems are reconsidered later (section 8). We used the transitive closure of
the position matrix and the classes of interdependence to reduce the problem to strong relations.
In all three problems a typical quadratic matrix is used:
a) voting matrix
b) input-output matrix
The voting matrix contains in row i and column j the number of votes for" ai is better than ai".
The input-output matrix contains in row i and column j the transactions from sector i to sector j.
The coefficient matrix contains in row i and column j the factor (oy;j aYj) measuring how strong
variable Yi depends on variable YJ. For the following optimization problem only non-negative values
make sense (otherwise positive and negative values could sum up to zero). Therefore in the case of
econometric equation systems we take the absolute value of the coefficients. This is also reasonable
for the transformation of problem c) into problem a) - there is no negative number of votes.
We now have the fourth problem of optimally reordering classes of interdependent elements. The
matrix to be triangulated has as elements only zero or one and the number of rows is equal to
the number of different classes of interdependence. We get the matrix to be triangulated from the
transitive closure: For every class we choose a representative variable (e.g. the first variable of
the class in the numerical ordering). Then we remove all other variables which are not chosen as
representative variables and the corresponding rows and columns of the transitive closure.
In the example we have a matrix with 2 rows. vVe choose YI and Ys as representative variables and
remove rows and columns no. 2,3, 4. Result:
1: o
5:
491
We look for an optimal ranking of these elements (alternatives, production sectors, endogenous
variables, classes of interdependence).
Optimality criterion: A ranking is called optimal, if the sum of the absolute values above the main
diagonal of the matrix is maximal.
Objective function: The sum of the absolute values above the main diagonal of the matrix.
Equivalently we could minimize the sum of the absolute values below the main diagonal. In the
special case when the sum of the values below the main diagonal is zero, we have an upper triangular
matrix. Therefore in the case of input-output matrices the name "triangulation" is usual, though
in general it is not a precise triangulation in the sense of linear algebra (Korte/ Oberhofer, 1968,
1969) .. Methods of linear algebra for block-triangulation cannot be used for our problem. The
optimization problem is not trivial, it is NP-hard (it is equivalent to the minimal feedback arc
set problem, see Garey/Johnson 1979, problem GT8). This optimization problem is solved in
Korte/Oberhofer (1968,1969), BartMlemy/Monjardet (1981), Bartnick (1983,1989), Reinelt (1985).
The problem has different names: triangulation of input-output matrices, construction of a median
relation, aggregation of preference relations, linear ordering problem. Applications are described in
Gruber (1978) and Wessels (1981). Methods of artificial intelligence are used to improve computing-
time (Bartnick 1989). Another question is the number of optimal solutions. For the problem of
econometric equation systems one optimal ordering is sufficient, but there are several solutions in
general, which is not only of theoretical interest. Already Korte and Oberhofer remarked that the
number of optimal solutions is a structural measure for the considered matrix (Korte/Oberhofer,
1969, p.429: "This fact ... shows that the number of essential different optima can be interpreted
as a further structural indicator (Strukturkennzahl)." Bounds for the number of optimal solutions
are given in Bartnick (1990). The number of optimal solutions seems to be closely connected with
the computing time for the A * algorithm (Bartnick 1989).
A= Li<j I Cij I,
Liofj I Cij I
where it is assumed that the matrix C is already ordered optimally.
The degree of interdependence is now defined as
8:=2·(1-A).
As A is the degree of "recursivity", we have to take the complementary value (1 - A). The factor
2 is chosen because the minimal value of A is 0.5.
If we have only zeroes below the main diagonal, then
A = 1 and 8 = O.
If lambda has its minimal value then
A = 0.5 and 8 = 1.
In the example of GRUBER (1968) we get for the interdependence class of the first four variables
(the fifth variable forms a separate interdependence class):
(~o
100
0
~ ~* ~~)
100
13
100 *
We now look for an optimal reordering of this system, which maximizes the sum of values above
the main diagonal.
The program "measure" (written in PASCAL) is available in Bartnick (1991). It computes an
optimal ranking with the algorithm of Bartnick (1989) and uses the formula above. We get as a
measure of interdependence
Ii = 0.56.
Optimal rankings for this problem are not of interest. There may be several rankings, but all must
give the same value for the measure of interdependence.
One could use the method of sections 5, 6 and 7 and get the indifference classes via the elements
remaining below the main diagonal in the optimal solution. But as econometric equation systems
are very large and the optimization problem is NP-hard, this is quite impractical.
And there are theoretical reasons for not doing this: The objective function used in section 5 is not
adequate for preference-indifference relations.
Remember that the general formulation of "optimal fit" is a minimum (least sum of distances,
Barthelemy /Monjardet 1981) and this minimum can be transformed in a maximum in the case of
preference relations without indifferences (Bartnick, 1983, p.131).
In the case of a general preference-indifference relation resp. in the case of block triangulation this
minimum cannot be transformed into a maximum in the same way (consider the special case of
total indifference).
Tiishaus (1983) proposes a general method for all types of relations (order, preorder, equi valence, ... ).
An example for preference-indifference relations (complete preorders) is given on p.95.
This algorithm is too slow for large econometric equation systems. Transitive closure is simpler and
very fast in computation, but there is no clear optimization criterion. Which objective function is
to be maximized or minimized, if we build the transitive closure?
For this question it is interesting to compare results of voting theory with econometric equation
systems.
If we take for example the coefficient matrix
c = (~~ ~~ ~:~)
0.6 0.3 -1
there is nothing remarkable: The equation system is totally interdependent, if we use the transitive
closure of the position matrix. But for the analogous voting matrix
C = (
*
10
20* 50)
40
60 30 *
the famous voting paradox arises. Solving the voting paradox by declaring all three alternatives as
indifferent, is obviously too simple.
Final remark: In this paper only discrete methods were used. There are approaches which use con-
tinuous methods (GauB-Seidel and Newton-type algorithms), e.g. Garbely/Gilli (1984), Don/Gallo
( 1987).
494
procedure interdep(t:imat);
var i, j: integer;
ok: array [l .. m] of boolean;
begin
for i:=l to m do ok[i] :=false;
for i:=l to m do
if not ok[i] then
begin
Ilrite(i:3);
for j:=l to m do
i f (t[i,j]=1) and (t[j ,i]=1) then
begin
Ilrite(j:3);
ok[j] :=true;
end;
Ilriteln;
end;
end.
495
References
BARTHELEMY, J.P. ; MONJARDET,B.: The Median Procedure in Cluster Analysis and Social
Choice Theory, Mathematical Social Sciences 1 (1981), pp. 235 - 267
BARTNICK, J.: Bewertung und Kompromifibildung, Eine Weiterentwicklung der Nutzwertanalyse
mit Beispielen aus der Raumplanung, Baden-Baden 1983
BARTNICK, J.: An A * Algorithm for the Aggregation of Preference Relations, Methods of Ope-
rations Research 59 (1989), pp. 311 -322
BARTNICK, J.: On the Number of Optimal Solutions of the Linear Ordering Problem, Extended
Abstract, Methods of Operations Research 63 (1990), pp. 425-426
BARTNICK, J: Ein Mail fUr die Interdependenz okonometrischer Gleichungssysteme und Ver-
fahren zur Ermittlung blockrekursiver Strukturen. Diskussionsbeitrag Nr. 166 des Fachbereichs
Wirtschaftswissenschaft, FernUniversitat Hagen 1991
BARTNICK, J: Aggregation of individual preference relations to a collective cardinal evaluation
(to appear in Methods of Operations Research)
BODIN, L.: Recursive Fix-Point Estimation. Theory and Applications, University of Uppsala,
Department of Statistics, Selected Publications, vol. 32, 1974
DON, F.J.H.; GALLO, G.M.: Solving Large Sparse Systems of Equations in Econometric Models,
Journal of Forecasting 6 (1987), pp. 167 - 180
GAREY, M.R.; JOHNSON, D.S.: Computers and Intractability - A Guide to the Theory of NP-
Completeness, San Francisco 1979
GARBELY, M.; GILLI, M.: Two approaches in reading model interdependencies, in: ANCOT,
J.P. (ed.), Analysing the Structure of Econometric Models, The Hague, Boston, Lancaster 1984,
chapter 2
GRUBER, J.: Okonometrische Madelle des Cowles-Commission- Typs: Bau und Interpretation,
Hamburg-Berlin 1968
GRUBER, J.: Zur Interdependenz iikonometrischer Gleichungssysteme, in: FROHN, J. (ed.), Ma-
kroi:ikonometrische Modelle fUr die Bundesrepublik Deutschland, Gottingen 1978, pp. 185 - 213
HELMSTADTER, E.: Ein Mail fiir das Ergebnis der Triangulation von Input-Output-Matrizen,
Jahrbiicher fiir Nationaliikonomie und Statistik 177 (1965), pp. 456 - 463
KORTE, B.; OBERIIOFER, W.: Zwei Algorithmen zur Losung eines komplexen Reihenfolgepro-
blems, Unternehmensforschung 12 (1968), pp. 217 - 231
KORTE, B.; OBERHOFER, W.: Zur Triangulation von Input-Output-Matrizen, Jahrbiicher fUr
Nationaliikonomie und Statistik 182 (1969), pp. 398 - 433
KRUPP, R.: Interdependenz und Zerlegbarkeit iikonometrischer Madelle, Jahrbiicher fUr Natio-
nalokonomie und Statistik 190 (1976), pp. 105 - 120
REINELT, G.: The Linear Ordering Problem: Algorithms and Applications, Berlin 1985
REINGOLD, E.M., NIEVERGELT,J. ; DEO, N.: Combinatorial algorithms: Theory and Practice,
Englewood Cliffs, New Jersey 1977
ROD DING, W.: Aggregation of Individual Preferences, Gottingen 1975
SZPILRAJN, E.: Sur l'extension de l'ordre partiel, Fundamenta Mathematicae 16 (1930), pp. 386
·389
TUSHA US, U.: Aggregation binarer Relationen in der qualitativen Datenanalyse, Kiinigstein/Ts.
1983
WESSELS, n.: Triangulation und Blocktriangulation von Inpu t-Output- Tabellen und ihre Bedeu-
tung, Deutsches Institut fiir \Virtschaftsforschung, Beitrage zur Strukturforschung, Heft 63, Berlin
1981
Evaluating the number of zero eigenvalues in a
dynamic model
Abstract
If a sparse matrix can be put into a block triangular form its eigenvalues can simply be
computed from the submatrices on the diagonal and most zero eigenvalues are recognized
without numerical computations. It is shown how to find the block triangular form of a
dynamic multiplier matrix based only upon the causal content of the coefficient matrices.
For the indecomposable matrices on the diagonal additional zero eigenvalues may also be
identified on a qualitative basis. Moreover the approach provides insight on which coefficients
are relevant for the existence of eigenvalues and which coefficients are not.
Keywords: Structural properties of econometric models; graph theory; sparsity.
Contents
1 The problem
2 Case with decomposable matrix D
3 Maximum matching and minimum cover
4 Illustration
5 Comments
References
1 The problem
The investigation of the dynamic properties in a model consists, among others, in evaluating
the eigenvalues of a particular matrix. Let us consider a deterministic linear model in its
structural form
(1 )
where Yt is the (n xl) vector of endogenous variables and Zt is the (m xl) vector of exogenous
variables. D, E and F are coefficient matrices. If D is nonsingular the reduced form of (1)
will be
(2)
and it is well known that the steady-state solution of (1) depends on the eigenvalues of
D- 1 E. From a numerical point of view, the identification of zero eigenvalues is a delicate
matter, e~pecially if the dimension of D- 1 E becomes large. However. for applications where
D and E Me large, but sparse, as it is the case for almost all econometric models, D- 1 E
497
might be put into a block triangular form by appropriate column and row permutations and
the eigenvalues are those of the submatrices on the diagonal.
If E has p nonempty columns, i.e. there are p lagged endogenous variables, it is straightfor-
ward that D- l E can be put into the following block triangular form
Our purpose is to investigate, relying only upon the structure of the matrices D and E,
whether the p X P submatrix Spp of D-lE contains additional zero eigenvalues.
If matrix D is indecomposable, i.e. if all n equations form one interdependent model, the
submatrix Spp in (3) will also be indecomposable and we will show in section 3 how zero
eigenvalues may be identified in such a case.
If D is decomposable it can be put into a block triangular form and D- l will have the same
block triangular pattern. In such a case the submatrix Spp in (3) might also be decomposable,
but of course with a different pattern. We can take advantage of such a decomposition to
identify additional zero eigenvalues.
A nonzero eigenvalue corresponds to a feedback in the model, i.e. a loop which goes over
the same variable from period t - 1 to period t.
If the matrix D is indecomposable every entry in matrix E creates a feedback. To identify,
whether in the case of a decomposable matrix D the lagged variables create feedback, we
drop all the lags, so that all lagged variables become current variables. This leads to a new
model with a matrix D*, corresponding to D + E, for which we seek the block triangular
form 2 •
If we now partition the original matrices D and E according to the q blocks of D* as shown
in (4), they will be block triangular, however, the diagonal blocks may be decomposable. In
such a case we partition the zero and nonzero columns of each matrix Ei as described in (3).
S1
[D": Di :
1
[E'.: E2. . S2 (4)
(5)
which corresponds to the matrix E. The set of vertices H represents the rows of E, which
are the equations, and the set of vertices y(-I) represents the columns of E. The set of edges
AE corresponds to the nonzero elements in matrix E, i.e. for a given element eij which is
nonzero there will be an edge [hi,y)-I)] in G E •
Let us now consider a matching ~VE of G E , i.e. a set of edges none of which being adjacent,
thus defining a one to one correspondence between some vertices belonging to H and some
vertices belonging to y-I. For matrix E a matching corresponds to a set of elements such
that each row and each column contains at most one of these elements. We then have
where P( n) is the set of n! permutations of set I = (1,2, ... , n) and p( i) is the i- th component
of permutation p. The function s takes the value +1 if the permutation is even and -1 if it
is odd. We conclude that a necessary condition for I E 1# 0 is that there exists at least one
permutation p, such that the product in (7) becomes nonzero. The set of elements of matrix
E satisfying
eip(i)#O for i=1,2, ... ,n (8)
corresponds in G E to a set of edges which define a matching. Hence it comes that a matching
with maximum cardinality WE will define the largest submatrix of E for which (7) is nonzero.
If the maximum cardinality of the matchings of matrix E is r < p, matrix Epp in (3) can
be partitioned in a way to show which columns or rows of Epp are linearly dependent. To
do so, we consider a minimum cover C' in G E, i.e. the smallest set of vertices which meets
every edge. According to the theorem of Konig-Egervary4 we have
I.e. the maximum number of edges in a matching equals the minimum number of vertices in
a cover.
Let VI be the number of vertices of the cover belonging to 1I and V2 the number of vertices
of the cover belonging to y(-l), VI + V2 = r. We then partition matrix Epp as follows:
(10)
where the dimension of the zero matrix satisfies p - VI + p - V2 > p, which characterizes
a singular matrix. If V2 > VI matrix E V ,(P- V 2) contains the V2 - VI dependent columns, if
VI > V2 it is matrix E(P-VJ) V 2 which has VI - V2 dependent rows.
4 Illustration
All the computations of the following illustration have been carried out with the program
CAUSOR, see [1], for which mainframe and PC versions can be obtained from the authors.
The matrices given hereafter are incidence matrices S of D and E of a thirteen equation
model, such as defined in (1).
It;
I
2 2
3 . 1 I 3
4 1 1 4 1
5 I 1 1 5
6 I I 6
7 7
D= S 1 1 E= 8 1 1
9 I I 9
10 10 1 I
11 11 1 1
12 I .~
.. I 12
13 I . . I 1 13 I
Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y
2 I S I I 4 3 6 978 I 1 2 1 5 1 1 4 3 6 9 7 8 1 1
0 I 2 3 0 1 2 3
Matrix D is supposed to be nonsingular and matrix E has one zero column corresponding
to variable Ys, which makes matrix Spp in (:3) be of rank less than or equal to twelve.
Matrix D is decomposable and therefore we compute D* which happens to be decomposable,
too. Indeed we verify that D + E can be put into a block triangular pattern and we give
hereafter matrix D and matrix E partitioned according to the block triangular pattern of
matrix D*:
7 1 7
6 1 I 6
4 1 I 4 1
I I 1 5
13 1 I 13
3 3
D= 8 1 I E= 8
10 I I 10
9 1 1 9
2 I 2
12 I
1l
Y Y Y Y Y Y Y Y Y Y Y Y Y
'l4-n
. . I
12
11 I
Y Y Y Y Y Y Y Y Y Y Y Y Y
I :4-h
2 3 4 I I I 6 7 9 1 I 8 2 3 ·1 I I I 5 6 7 9 I I 8
I 2 o 3 I 2 o 3
Finally we have to seek a maximum matching in the block corresponding to variables Yl, Ys,
Y6, Y7, Y9 and YlO. The bipartite graph as defined in (5) is:
equations
@
o
o
@-
The edges drawn with solid lines form a matching with maximum cardinality which is five.
The vertices in the boxes form a minimum cover and we partition the block according to
(10):
8[0"
9
10
3
.1.
. . 1 1 . 1
1 . . . . .
. . 1 . . .
13 1 . . . . .
2 . 1 . . . .
y y y y y y
691157
o
which shows that the equations 2, 13,3 and 10 are linearly dependent in the subspace defined
by the variables Y6, Y9, YlO, Y1, Ys and Y7·
We conclude that the thirteen equation model has at most seven nonzero eigenvalues.
5 Comments
References
[1] Gilli, M., 1984, CAUSOR - A Program for the Analysis of Recursive and Interdepen-
dent Causal Structures, Cahiers du Departement d 'Econometrie, No 84.03, U niversite
de Geneve.
[2] Murata, Y., 1977, Mathematics for Stability and Optimization of Economic Systems,
Academic Press, New York.
[3] Roy, B., 1969, Algebre modeme et tMorie des graphes, tome I, Dunod, Paris.
[4] Schoonbeek, L., 1983, On the Number of Non-Zero Eigenvalues of a Dynamic Linear
Econometric Model, Empirical Economics, Vol. 8, pages 119-123.
[5] Strang, G., 1986, Introduction to Applied Mathematics, Wellesley-Cambridge Press,
Massachusetts.
Forecast and multiplier sensitivity analysis with respect to chan-
ges of structural equations and submodels in dynamic systems
Johann Elker
Fern U niversitiit Hagen
Lehrgebiet Statistik und Okonometrie
Feithstr. 140, D-5800 Hagen 1, FRG
Abstract
In this paper improved methods are presented for computerized automatic analysis of the sensiti-
vity of forecasts and multipliers in interdependent dynamic models. These methods facilitate the
diagnostic checking of the effects of structural coefficients on point forecasts and multipliers. The
effects of changes of all coefficients in one or more structural equations (i.e. in a sub model) on
point forecasts and multipliers are quantified. In addition, measures are constructed for measu-
ring the sensitivity to be investigated. The presented results may not only be helpful instruments
for sensitivity analysis of point forecasts and multipliers in the "finished" model. They may also
be useful tools for comparison of distinct submodels and variants of equations already during the
construction of econometric models.
Keywords: Sensitivity analysis in nonlinear and linear models; computing inverse matrices with·
out repeated inversion; measures of sensitivity; diagnostic checking.
Editor's Note:
Johann Elker, a blind mathematician, was a staff member of the Chair of Statistics and Econome-
trics, University of Hagen, and a doctoral student from June, 1985,till August, 1987. At that time
he suffered a severe stroke from which he never recuperated. He died on ~Iarch 13, 1989, at the
age of 4l.
In honor of Johann Elker, this paper is included in this proceedings volume. It was written as Dis-
cussion Paper No. 112 in 1986. the present revised version was finished on July 17, 1987. It briefly
summarizes several papers listed at the end of this paper (there, also various proofs and derivations
can be found). An earlier version of this paper was presented at the VlIIth International Confe-
rence on Problems of Building and Estimation of Large Econometric Models. Macromodels '86, in
Szczyrk-Bila, Poland. December 3-5, 1986, organized by the Committee of Statistics and Econo-
metrics, Polish Academy of Sciences, and the Institute of Econometrics and Statistics, University
of Lodz.
The methods of sensitivity analysis developed by Johann Elker can also be applied to nonlinear
models: In a Taylor series expansion. only the linear terms are used. For the linearization the
solution of the nonlinear system is used which is generally computed iteratively only once.
The methods for sensitivity analysis developed by Johann Elker are also available in a FORTRAN
computer program. It has been written in 1987 by Paul Chr. Gruber in close cooperation with
Johann Elker. It is available on request from the editor.
This program has been extended by parts for computing sensitivity measures (which J. Elker and
P. Chr. Gruber could not finish because of J. Elker suffering a stroke on August 29, 1987) by
Wiltrud Terlau. She also adapted the program to particular problems arising during the sensitivity
analysis of a 49-equation quarterly model of the "Vest German textile and apparel industry (see
paper by W. Terlau in this section).
503
Contents
1 Introduction
6 Sensitivity measures
7 Outlook
504
1 Introduction
In this paper, instruments for automatically analyzing the sensitivity of point forecasts
and multipliers in large interdependent dynamic models are developed. In many cases
of forecasting by econometric models and calculating multipliers, considerable errors
occur, which may be due to the particular structure of the underlying model. Even small
changes of the structure may cause relatively large increments of forecasts. The forecast
sensitivity analysis serves for localizing strong influences on the forecasts, caused by
one or several structural coefficients. Like incorrect forecasts, erroneously estimated
multipliers can lead to disadvantageous decisions in economic policy. The structure of
the model may be a reason for the errors. It may be localized by sensitivity analysis of
the multipliers. Sensitivity analysis of forecasts and multipliers helps in diagnosing the
dependencies of multipliers and forecasts on the structural coefficients. Thereby, the
increments of multipliers and forecasts caused by a change in one or several structural
coefficients are calculated (under the ceteris paribus condition that all other coefficients
are not changed and interdependencies of coefficients are not taken into account). The
sensitivity analysis of forecasts and multipliers in nonlinear interdependent dynamic
systems with regard to influences of one or all coefficients of an equation is described
in detail by Elker (1986a,b,c,d).
In the present paper these results are developed further for analyzing the effect of
changing the structural coefficients of a submodel of a linear model. The changes of the
structural coefficients in the sensitivity analysis usually being small, the instruments
described can be transferred for use with nonlinear systems without severe loss of accu-
racy (using in a Taylor series expansion only the linear term and using for linearization
the solution of the nonlinear system which is generally computed iteratively only once).
In the linear case the presented algorithms permit the investigation of structural
changes of any size. So they arc suitable for diagnosing in the "finished" model and
for comparing the quality of various equations and su bmodels in the pbase of model
construction. Meanwhile, the author of this paper is transferring the techniques for
sensitivity analysis for use with nonlinear models.
The described methods may be of considerable importance not only in the "finished"
model, they may even be of help in the phase of construction. So they might improve
the accuracy of multipliers and forecasts, because the constructor of a model can easily
investigate the structural changes of several variants of equations or submodels, the
estimated coefficients and explanatory variables of which differ from those of the other
ones. Judging the quality of a variant by statistical and other criteria often does not
lead to a clear decision.
The classical method for sensitivity analysis requires the inversion of a matrix and
(in dynamic models) frequently a considerable number of matrix multiplications. It
always requires also considerable manpower input. Therefore it is rather expensive and
hardly applied. The techniques presented in this paper are, so to.speak, automatized
and, by and large, cheaper. Thus they facilitate simplified computerized sensitivity
505
analysis especially of large systems. The proofs are neglected, since they are given In
Biker (1986a,b,c,d) or can be derived in similar ways.
The paper is organized as follows: In section 2, we present the structural form of
a dynamic linear system of (difference) equations. We also derive the corresponding
reduced form, the final form, impact multipliers and delay multipliers. Then we are
concerned with the sensitivity analysis of the reduced coefficients (section 3), of the
multipliers (section 4) and of the point forecasts (section 5). In the first subsection of
these three sections, we deal with the general case in which changes are made in all
equations. In the second subsection, changes are restricted to a submodel consisting of
one or more (h) equations. In the third subsection, a special case is treated: changes
of structural coefficients are made in only one equation. In section 6, measures of
the sensitivity of point forecasts with respect to changes of structural coefficients are
presented and briefly discussed.
I
where
1. Yt,Yt-I, ... ,Yt-S are the g x 1 - random vectors of the endogenous variables in
the periods t, t - 1, ... , t - S,
2. Xt, Xt-l, ... ,Xt-R are the p x 1 - vectors of exogenous variables III the periods
t, t - 1, ... , t - R,
4. Co, C I , ... ,Cs are the g x g - matrices of the corresponding structural coefficients
of the endogenous variables and
Assuming that Co is nonsingular, the reduced form can be derived by solving (1) for
Yt:
Yt = F1Yt-1 + ... + FSYt-s + DoXt + D1Xt-l + ... + DRXt-R + Vt. (2)
In this equation, F., Dr and Vt are given by
(3)
506
(4)
Vt = - C-1 0 Ut· (5)
The coefficients of the reduced form conLained in Lhe matrices F.• , s = 1, ... , S, and
Dr) r = 1, ... , R, are called reduced coefficients.
Appropriately lagged versions of (2) are substituted aL first for Yt-l in (2), then for
Yt-2 in the resulting equation etc. In this way, one obtains after K substitutions:
Yt + ... + pK
= PK+IYt-K-l
K ](+sYt-K-S + QK
0 Xt + 1 Xt-l + ...
QK
The equation (6) is called the final form with fixed starting conditions.
The matrices p!!and Qt; may be calculated using the recursively defined auxiliary
matrices <Pn:
n n-l
<Po =I and <Pn = L F~CJ>n-~ = L i1>k Fn-K (7)
.• =1 k=O
In all equations of this paper, it holds by definition: F3 = 0 for s < 1 or s > Sand
Dr = 0 for r < 0 or r > R. In (6), the coefficient matrices can be constructed as
follows:
J(
Furthermore we have:
wr
K
= L <PkVt-k (10)
k=O
If the coefficients of the lagged endogenous variables approach zero for K -> 00, (if
and only if the modulus of the equation det(>.s 1+ >.S-1 Fl + ... + Fs) = 0 is less than
one; I denotes the unit matrix of rank g) (6) takes the form
In the following the matrices C .• , s = 0,1, ... , S, and B r , r = 0,1, ... , fl, are replaced by
the matrices C,. and 13r of the estimated structural coefficients. The reduced coefficient
matrices F .• , s = 1,2, ... , S, and Dr, r = 0, 1, ... , R, are replaced by Fs and Dr, which
can be constructed corresponding to (3) and (4). The matrix <Pn and the matrix Qn,
n = 0, 1, ... , of the lagged multipliers are replaced by ell n and Q", which can be derived
according to (7) and (12).
6.F tl
= _C-0 1 XlI -l-
' Cn
y~ (;-1
0
X)-I(y!C ~ + y~C
lI
F)
$
( 13)
(14 )
The change matrices 6.C, and 6.E r can always be decomposed in products of the men-
tioned form. Thus, for example, for X the unit matrix I and for Y!, and Y! the change
c. B,
matrices 6.C.• and 6.E r themselves can be chosen. In this case, It = 9 holds, However,
the methods for sensitivity analysis derived from (13) and (14) provide advantages in
comparison to traditional procedures (i.e, recomputing by (3) and (4)) only if h is
substantially smaller than g, because in the other case the effort for computation may
not decrease sufficiently.
!lC, = Ei!lC3 ,i
!lEr = E;!l Br,;
where i = (i l , ... , i h ) and E, denotes the 9 X h - matrix, the columns of which are the
unit vectors C;I" .. , CiJ, (Cikj = 1 for j = ik, 1 ::; k ::; h, else Cikj = 0). The h X 9 -
matrix !lC."j and the h X P - matrix !lEr ,;, respectively, arc identical with the vectors
of changes as their rows.
With X = Ei 'Yc! = !lC 3 ,i and Y! = !lEr,i, the following relations, which are
',. Br
fundamental for sensitivity analysis of the reduced form coefficients, may be derived
from (13) and (14):
(IS)
and
(16)
where
(17)
Furthermore, (COI)i denotes the matrix consisting of the columns i 1 , ... , i h of the
matrix COl.
When h is substantially smaller than g, the sensitivity of the reduced form coeffi-
cients can be computed without much effort, using (15) and (16). So it is not necessary
to compuLe repeatedly the changed system (1) of sLructural equations in Lhe usual way
(e.g. by inverting the 9 x 9 - matrix Co cLc.). These formulas may be applied for
sensitivity analysis separately to one structural coefficient or to more than one structu-
ral coefficient. Thus, the foundations for the simply feasible, computerized sensitivity
analysis of the reduced form coefficients with respect to all structural coefficients are
available.
The results in (15) and (16) may be transferred to changes of coefficients of several
variables in all structural equations (changing columns) In a similar way. But in the
present paper this case is not dealt with.
(18)
509
and
(19)
where
1
In these equations, cO,.i is the i-th column vector of COl, the 1 X 9 - vector 6.c 3 ,;. is the
change of the i-th row of the matrix C. and the 1 x p - vector 6.b r ,;. denotes the change
of the i-th row of the matrix 13r .
where
r = -(I + y!Gil C-0 1X)-l. (21 )
The matrices en are recursively defined by
n
eo = 1 , en = L nme m - n n = 1,2, ... , (22)
m=l
where
m
nm- r "\~vl
L-,lC,.~m-3
~ c--lx
0 . (23)
The explanation in section 3.1 about simplified sensitivity analysis of reduced coef-
ficients holds also true for the simplified sensitivity analysis of multipliers: The techni-
ques based on (20) are only favourable if h is very much smaller than g.
where rand (Colr are defined like in section 3.2 and 8 n -k-1 is defined 1Il (22), with
the exception that
m
In this special case, the matrices 8 n - k- 1 also reduce to a scalar, which can be computed
by
n
where
(29)
One might presume that the compuLa\,ion of the matrices of increments D.Qn) which
result from changing several equations at a time, may be simplified and accelerated
by stepwise changes using the less complicated formula (26). However) this procedure
requires the inversion of Co after each step before calculating the next one (in reality,
repeatedly calculating the col umns iI, i"l., ... ) i h is sufficient). Especially it can happen
that one s\,ep produces a (nearly) singular matrix of coefficients Co (in the worst case
even independent of the order of the changes), although the matrix is nonsingular after
changing all rows at a time. This problem can only be avoided by special precautions,
which are not discussed in this paper.
(30)
where
K-k 1 5 R
ak = L 'CL 1";.)1t-](+I-"
8K-k-I I + L Yf.Yt-K+I-,o + LYb,Xt-K+l-r) (31)
1=0 ,,,,.(1 s=l+l r=O
where
K-k l S R
ak,i = L 8K-k-lr(L ~C.,dJt-K+l-., + L ~Cs,iYt-K+l-. + L ~Br,iXt-K+l-r)
1=0 .=0 3=l+ I r=O
(33)
In these equations (COl)i is defined like in section 3.2, 8K-k-1 is defined by (22), with
TIm like in (23) and r being defined by (17).
where
K-k 1 S R
G:k = L ,8K - k- l (L 6.c~,i.Yt-K+l-s + L 6.i:~".Yt-K+l-., + L 6.b~,i Xt-K+l-r)'
l=O .• =0 ~=I+ I r=O
(35)
These formulas clearly show the advantages of the simplified techniques for forecast
sensitivity analsis developed in this paper. The 9 x 1- vectors 4>kcO,.i , k = 0, 1, ... , K ,
i = 0,1, ... , (<1>0 = I) and the vectors of forecasts Yt-K +1 , l = 0,1, ... , K, have to be
computed only once, independent of the actual change. Then every change only requires
the computation of the scalars G:k, k = 0,1, ... , K, by (35) and 6.Yt by (34), which is
rather simple.
6 Sensitivity measures
In this section measures for the sensitivity of point forecasts are presented. Here a
denotes the vector which contains all structural coefficients to be varied; is the 6.a.
513
corresponding change vector. The effects of these changes on the point forecasts of the
jointly dependent variables can be measured by t,he following terms:
7 Outlook
in the past sensitivity analysis of forecasts and multipliers was made only in rare cases,
probably because of the large inputs of manpower and computer time. In contrast to this"
the presented simply and fastly executable analysis procedures may find widespread
applications. For especially in large equation systems their advantages are very conspi-
cuous in relation to traditional techniques. In the phase of model construction they may
serve as a diagnostic instrument and may therefore promote the quality of forecasts
The efficiency of the simplified sensitivity analysis procedures presented in this paper
still has to be demonstrated in practice. An application to large equation systems of
the FRG is in preparation.
The generalization of the presented techniques to sensitivity analysis of interval
forecasts (forecast ellipsoids) is as desirable as the sensitivity analysis of forecasts with
respect to changes of data.
514
References
Diick, W. (1970)
Numerische Methoden der WirtschafLsmaLhematik 1. Akadcmie- Verlag, Berlin.
Elker, J. (1986a)
SensiLivitaLsanalyse der Multiplikatoren in dynamischen okonometrischen Modellen.
Diskussionspapier Nr. 105, Fachbereich Wirtschaftswissenschaft, Fern Universitat
Hagen.
Elker, J. (1986b)
Sensitivitatsanalysc der Punktprognosen in dynamischen okonometrischcn Modellen.
Diskussionspapier Nr. 106, Fachbereich Wirtschaftswissenschaft, Fern Universitat
Hagen.
Elker, J. (1986c)
Sensitivity analysis of multipliers in dynamic econometric models. Discussion paper
No. 110, Fachbereich Wirtschaftswissenschaft, Fern Universitat Hagen.
Elker, J. (1986d)
Sensitivity analysis of point forecasts ill dynamic econometric models. Discussion
paper No. Ill, Fachbereich Wirtschaftswissenschaft, Fern Universitat Hagen.
Goldberger, A.S. (1966)
Econometric Theory. John VViley & Sons, New York.
Goldberger, A.S. (1959)
Impact Multipliers and Dynamic Properties of the Klein-Goldberger-Model.
(2nd Edition. 1970) North-Holland Publishing Company. Amsterdam.
Gruber, J .,Rosemeyer, 13. (1982)
Prognosc-Sensitivitiitsanalyse in groBen, nichtlinearen okonometrischen Modellcn -
einige Ergebnisse fiir das RWI-Konjunturmodell. In: I\:rallmann, H. (Ed.) ] 982.
Sozio-okonomische Anwendungen der Kybernetik und Systemtheorie. Betriebliche
Kommunikationssysteme 2. Erich Schmidt Vcrlag, Berlin, S. 173-195.
Gruber, J., Rosemeyer, B. (1983)
Sensitivitiitsanalyse der Prognosen in okonometrischen Mehrgleichungsmodellen. In:
Buhler, W., Fleischmann, B., Schuster, K.P., Streitferdt, L. and Zander, H. (Eds.)
1983. Operation Research Proceedings 1982. DC OR, Vortrage der 11. Jahrestagung.
Springer- Verlag, Berlin, Heidelberg, New York. S. 531-541.
Kenkel, J. (1974)
Dynamic Linear Economic Models. New York, Paris, London.
515
Wiltrud Terlau
University of Munster
Forschungsstelle fur allgemeine und textile Marktwirtschaft (FATM)
Alter Fischmarkt 21, D-4400 Munster, Germany
Since March 1, 1991:
Rheinisch-Westfalisches Institut fur Wirtschaftsforschung (R WI)
Hohenzollernstr. 1-3, D-4300 Essen, Germany
ABSTRACT
CONfENfS
1 Introduction
2 The Structure of the Forecasting Model for the Textile Pipeline (Qualitative Ana-
lysis)
3 Structural Sensitivity Analysis of Forecasts
4 Selected Empirical Results for the Model of the West German Textile and Apparel
Industry
5 Conclusions
Appendix
References
517
1 INTRODUCTION
This paper reports on a methodological study of procedures for examining some basic
behavioral properties of dynamic equation systems, applied to the quarterly model of
the West German textile and apparel industry. Until now, questions of the structure of
econometric models, which illuminate the properties of the observed system, have
received relatively little attention.
The traditional multiplier analysis gives indicators of the reactions of the endogenous
variables caused by variations of the exogenous variables. The computed multipliers
measure the direct and indirect effects of such variations. As parameters of the reduced
form equations, each multiplier consists of a conglomeration of many structural
coefficients. When an understanding of the relations between economic theory and
economic behavior through the model is of main interest, the structural equations have
to be analyzed.
Structural sensitivity analysis studies the structure of a model by investigating the sensi-
tivity of point forecasts to changes in structural coefficients. It detects those structural
parameters to which model "outputs" (endogenous variables) are most sensitive. High
sensitivity to certain structural coefficients indicates their importance in the model. If
statistically poor values are estimated for the endogenous variables, it may signal a
defect of the model [Kuh, Neese (1982a), 120, Gruber, Rosemeyer (1982), 3].
In the context of a forecasting model, the computed point forecasts are of main interest.
Even small changes of the structure may cause relatively large increments of point
forecasts. The influence of an individual parameter in a given equation is revealed by
structural coefficient variations. The results of such an investigation help to illuminate
the interactions of the model elements. Furthermore structural sensitivity analysis can be
used to assess the robustness and the reliability of the model [Kuh, Neese (1986a), 126].
Investigations of the structural sensitivity have so far, to the best of my knowledge, been
carried out for models of the economy as a whole. The transfer to sector-specific models
will be of significant interest.
The object of this study is the disaggregated forecasting model of the economic
development of the West German textile and apparel industry. The applied structural
analysis is divided into two parts: a qualitative and a quantitative one. With regard to a
modified incidence matrix the qualitative analysis gives a first overview of the
interactions of the variables. The intensity of these links is quantified by the structural
sensitiviy analysis. Improved sensitivity analysis methods and several sensitivity measures,
to scale the reactions of point forecasts of the endogenous variables caused by changes
in structural coefficients, are applied to the forecasting model. The software program
used is briefly described. Selected empirical results are presented in the last section of
this paper.
518
21 The Model
For more than 20 years the Forschungsstel1e fUr allgemeine und textile Marktwirtschaft
(FATM) at the University of Munster has been working on and with a quarterly
econometric forecasting model of the West German textile and apparel industry
[Helmstiidter, Krupp, Meyer (1976), Kleine, Thiele (1982)]. Its purpose is to make every
year forecasts of 5 - 9 quarters ahead. The results are presented to interested parties in
associations and business.
The most recent version of the forecasting model considers the various stages of the
textile pipeline [Terlau (1991)]. It is based on the demand-side approach: Basic
assumption of the model is the dependency of the development of production on the
domestic and foreign demand for textile goods. The model yields point forecasts starting
from the final demand for textile products up to production, turnover and foreign trade.
In the textile pipeline there are reaction lags of the economic units. This makes it
necessary to formulate a dynamic equation system. The structural form of the linear
dynamic equation system is as follows:
where
yp Yt-l' ... , Yt-S are the gx1-random vectors of the endogenous variables in the
periods t, t-1, ... , t-S,
are the px1-vectors of exogenous variables in the periods t, t-1, ... , t-
R,
is the gx1-random vector of disturbances in period t,
are the gxg-matrices of the corresponding structural coefficients of
the endogenous variables and
are the gxp-matrices of the corresponding structural coefficients of
the exogenous variables.
In our case, g=49, p=25 (including the intercept), S=4 and R=4.
The model is in real terms. The estimation period is from 1970,1 to 1987,4, using
quarterly, seasonally unadjusted data. The sources are statistics of the Bureau of
Statistics (Statistisches Bundesamt) and from the lio-Institute for Economic Research.
Seasonal components are taken into account by means of dummy variables. The list of
variables is given in the appendix.
The parameters of the 49 behavioral equations have been estimated by ordinary least
squares. The software package SAS has been used for model construction.
519
One major goal of qualitative analysis of causal structures in dynamic forecasting models
is to obtain a first overview of the interactions of the variables.
The incidence matrix is a conventional tool to describe the structure of a model [Krupp
(1971), 203 ff, Helmstadter (1973), 123 ff]. Variable links (structural coefficients
different from zero) are usually characterized by "1" and their absence by a dot. For a
dynamic model containing lagged variables, one incidence matrix for each lag can be
formed. A hierarchical ordering of the matrices makes it possible to reveal the structure
of the system [Simon (1953), Gruber (1978)].
But the disadvantage of this form of structural analysis is that it only considers the direct
interactions between the economic magnitudes [Heile mann (1989), 258 ff]. The indirect
links, which are also important, have also to be taken into account.
The modified incidence matrix of the forecasting model of the textile and apparel
industry is given in Figure 1. Its rows contain the explained variables and its columns
contain explanatory variables. Only a few coments are possible here:
The modified incidence matrix, reproducing the minimal (reduced) graphs, contains
all 49 endogenous variables. This means, there are no interdependent subblocks of
equations. The endogenous variables form so-called trivial blocks, each containing
just one variable. The demand-side approach to the textile forecasting model leads
to a recursive formulation; the structure of the model is triangular [HelmsHidter
(1962)].
48% of the entire 462 links (structural coefficients) are direct links (B, D) and 52%
indirect links (I). The greatest part (71 %) is composed of the structural coefficients
of the 24 exogenous variables (columns AEBA to ZEIT).
IThe program CAUSOR is used for the follOwing computations; it is implemented in the software
program TROLL of M.LT., USA
520
Fig. 1: Triangulated Structure of the Forecasting Model for the Textile Pipeline
- Qualitative Analysis of Causal Links -
""...
UIXIU
. . . 8 II . ..
' • • • el • • • •
. I!I •• • • • • • • I!I
• • .
1
. 1
0". ... ..•• ... .•. . 1
ac;c;vAtJl •••• 11.. . .. .1 ... 1 .. • •. 1
...
mlU • . . . 11 . . . . • • . . .1
£HAU5 • • • • • 11! ••• • 1 • •• • • •• •. . 1
EUW 1. . . • , 1 . . . . .11 . .. .1
DGGSfOit • • • • • I iii • a .. . . 1
DlWW I . .. • • • . . . . . I .. . .• 1
mtXI . . 1 • .•• . • • . . I .. .. 1
ElVIM . . . . 1.1 • . • . . 1
1...... . . . . I I . . . . . •. • •• ••• . .1
UlnA . . . . . I I . .. . 1
UBUI .II •• IIB . .• . I . .1
UTlXA • • 1 . '.1 • • • • · 1
.. ,II".
.
UTili . 1 ' • •••• . 1
VICl.1l1 .•. I . .1
. .
....
Uu.RH1 . 0 . 0011 • • • • , I ..
UIll( I 1. . I I I . . . . . . I . ·a I . . ,
.,
UTOt •. 11111. • • . . · II I .
[OOM ••• • 01001 •• ••• 1 . 1. ·
OWW ••• • • • • 1..1 •• 1.1. · .,
·
. .
EJlfOAS •• DQlOI •• • • • O . I .
tC6E1t •• • • ID11 ••• • • 0 • •• · . . 1
.,
·,
[M • • •• IDDDI •• • • • • • I . · I
[WAUl • • •• 1111 • • • • • 1 . 1 . . ·
....... I I • • DO f • • •• I • • I I • • iii • •
,
IO!AI(I • • 1 I D! I • • I ••• I • . I . I 1 . • I . . 1
ce,t '" 1 • t , 01 0 0 • • . . .. . J I • . , . 1 I . . I . . 1
OGGlTO . . 11 . . . . 1. . . • 1. , • III , I 1. • • • . 1
OHUO • . 1D . . . . . . . II . . I!!I . .1
...
OITIl . . 111.0.1. l- . e. I I . . I- 1
OlIn · .110.1. . . . . . . . .1. . 11 • . 1. · 1
10..,.. •. .. 1011.1. . • 111 . . I . I . .. · . 1
WI. . . • . 1111 . . . . . •. I .• . • B · I ••• 1
IIIGGVU, . • • • 10.1 • . . . • 181 . · I . I.
Imll • • • 1 . . . 1. . . . • 1.1. .... I • I 1 . I- .1
.
CIIe(I( 11. • • • • 1. . • . . 1.1. .11 •. 1 • • •• • • 1. .1
·,
..... 11 • • • 1.1. . . . . 1.1. II .. I . I ..... I. . . 1
CGGOU •• 111111 •• •• ••• 1.1 •• • • 1 • I 1 I • I . • •. 1
·
... .
OCiGII.l( •• 1 I I II II I • ••• • 1 • 1 ••• · I I I . • . . ... 1
1I!CNSTOtt •• I I 0 I II I I • • • • 1 I I I • • %. .1 111
°
. 1. . I . I .
. ,
. .,
IIN[IU • • 11011111 .1 •• 1.11 •. • .• • 1 • I.
11.(011 •• 01101 • • • • 1 .1 • • • • • • . 1
I . . 1 I
I iii •• 1 I . 1. .. .
IIIWWI 11 . . 1101. 1 . . 1.11 . . . . .a. . . I 1 •. 1 I .. I . I . .... ..... 1
JJN,ua 1 1 I 1 1 I 1 I • . • 1 • I I 1 I I 1 1 . I . I' . I . . . . Ii . 1
CXOWI •• IIDa.tl . I ... III . I ·a. . ... . e . ... . ,
Symbols: (B) = Basic (Direct) Link, (D) = Direct Link, (I) = Indirect Link
521
The temporal disaggregation of the direct links shows that 69% of them operate
without lag.
The qualitative structural analysis, which is presented in Figure 1, only describes the
causal links of the variables, but no statements about the important economic question
of the intensity of the links are possible.
In the last years, improved methods have been developed to quantitatively analyze the
structure of large equation systems. Traditional sensitivity analysis may, especially in
large models, be extremely laborious. Every change in the structure requires the
inversion of a matrix and the multiplication of a large number of matrices. But recent
research has made it possible to quantify the structural sensitivity of point forecasts with
respect to structural parameters by using algorithms for calculating inverse matrices
without repeated inversions [Gruber, Rosemeyer (1982), (1983); Elker (1986), (1986a),
(1986b)]. On the basis of a given structural equation system and the derived final form,
a simplified computation of the vector of forecast changes !l.y is possible.
522
At the University of Hagen a corresponding software package has also been developed
[for the most recent version see EIker, Gruber (1987)] . This computerized structural
sensitivity analysis has been applied by Gruber and Rosemeyer (1982, 1983) to the
nonlinear RWI-model. The effects of parameter perturbations are computed for one
period (static structural sensitivity analysis).
These methods have been extended by EIker (1986, 1986a, 1986b), to dynamic equation
systems, considering also the effects over time (dynamic structural sensitivity analysis).
They refer to linear systems, but they can be transferred to nonlinear models by using
only linear terms in a Taylor series expansion and using for linerarization the solution of
the nonlinear system which is generally computed iteratively only once.
Following Elker (1986a, 1986b), we start with the structural form of a linear dynamic
equation system, like in equation 1. Assuming Co to be nonsingular, the corresponding
reduced form is derivable:
(2)
where
The .sensitivity of the changes in the forecasts of the endogenous variables caused by
variations in structural coefficients may be difficult to interpret, because the units of
measurement of the variables involved have to be taken into account, and these units of
measurement usually differ widely.
To facilitate the interpretation of the results of the study, various sensitivity measures
are used. Elasticities, i.e. the relative change of the forecast with respect to a small
percentage change of a structural coefficient (aij)' are widely used and serve
substantive/theoretical goals. From the statistical point of view a measure considering,
for example, the variable's standard deviation may be preferable. Choosing a scale
523
depends mostly on the immediate purposes of the investigation [Kuh, Neese, Hollinger
(1985),30].
Several sensitivity measures are used in this study.1 As a measure for the sensitivity of
the individual forecasts with respect to changes of individual structural coefficients, we
use the following:
1. The unscaled difference D tg in the point forecast of the g-th endogenous variable,
caused by a change in the structural model:
(3)
* 100. (4)
3. The arc elasticity Eltg,ij of the point forecast of the g-th endogenous variable with
respect to the structural coefficient aij
JiYtg aij
Eltg,ij= * (5)
Ytg Jia ij
4. The following elasticity which uses the Euclidian norm II-II is a useful tool to
measure the responsiveness of a single endogenous variable with respect to
simultaneous changes of several structural coefficients [Schintke (1976), 65, 155 ff):
Iiall
*--- (6)
II Mill
where a denotes the vector which contains all structural coefficients to be varied and
.11ft is the corresponding vector of changes.
5. When the sensitivity of the point forecasts of all endogenous variables y of the
model with respect to changes in the vector a of structural coefficients as a whole is
of main interest, then the following generalized elasticity is a suitable measure:
For further scaling possibilities see Kuh, Neese, Hollinger (1985), 24, Elker (1986a), 16 f, Gruber,
Rosemeyer (1983), Terlau (1991). The question, which sensitivity measure is best suited or whether
there are better scaling possibilities, is of interest for further research.
524
/I AYtg /I /I a II
Nt = - - - - * (7)
II Ytg I /I Aajl
The measures (eq.6) and (eq.7) may be seen as generalized elasticity measures [Gruber
(1978),204, Elker (1986a), (1986b)].
3.4 SOFIWARE
(1) The initial program produces four files containing the inverse of Co' the multipliers
Q, the forecast Yand the matrices (~ * inverse of Co (named "auxvec")). This part
needs to be run only once .
(2) The main program calculates the unscaled changes of the multipliers (AQ) and of
the point forecasts Ay, resulting from various changes of the structural coefficients.
It is run as often as needed.
The output consists of point forecasts of the endogenous variable vectors Yt of the size
gx1 for the periods 0 to K, i.e. t to t+ K..
As sensitivity measures only the unscaled differences (eq.3) are used in Elker and
Gruber (1987). Therefore an extended version of the program of Elker and Gruber
(1987) has been developed by Terlau (1989). It calculates also other forms of structural
sensitivity measures. Variations of the point forecasts are scaled as percentage changes
(eq.4), arc elasticities (eq.S) and, when the sensitivity of all point forecasts as a whole is
of main interest, elasticities expressed in Euclidian norm (generalized elasticity, eq.7)
are computed. A sensitivity measure for simultaneous changes (eq.6) is also
considered. 1
1 Caused by the great dimensions of the matrices the program of Elker and Gruber (1987) had to be
converted from the eMS-System to the MVS-System. Furthermore it is possible now to insert the
desired coefficient perturbations in different periods as percentage changes.
525
The structural sensitivity analysis methods presented in the previous section have been
applied to the model of the textile and apparel industry. For variations of all structural
coefficients of the system a tremendous amount of output has been produced.ln
selecting results here, we focus our attention to the interactions among turnover,
production and foreign trade in order to serve the dual purpose of learning more about
a matter of substantive/economic interest and understanding how the structural
sensitivity methods apply to a meaningful problem. 1
4.1 Specification
2.57492 D2 + 8.6904 D3
(1.69785) ( 4.27624) (8)
The domestic turnover of apparel depends mostly on the domestic order intakes (AEBI)
and on the business climate index for retailing of textiles and apparel (G03TEXBEK),
which quantifies the opinions of the firms about the state and the expectations of the
corresponding trade-sector (final demand). The lagged endogenous variable UBEKI(-4)
is responsible for the economic dynamics. Saisonal effects are considered by the dummy-
variables Dl, D2 and D3.
1 The present paper is a condensed version of an extensive study to be published later [Terlau (1991)).
526
Table 1: Basic information about structural coefficient variations for equation UBEKI
The qualitative structure analysis (Fig. 1 in section 2.2) provides a first overview of the
interrelations of the economic variables. The set of variables affected by a variation of a
parameter in equation UBEKI is given by the set of variables with the non-zero entries
in the UBEKI column (qualitative structural sensitivity):
A change in equation UBEKI will affect the point forecasts of production (ONBEK,
KDOBR, KHAKAR, KMWAE), imports (IMBEKR, IMHAKAR, IMWAER), the
entire revenues of apparel (UBEK) and UBEKI itself. Most of the links are of indirect
nature, whereas UBEK and IMBEKR are characterized by direct links. 1
Varied coefficient
Endogenous variable 80.25 BO.2 BO.l0 C4.21
0 EL 0 EL 0 EL 0 EL
UBEK[ ·0.32 ·0.27 0.22 0.19 0.45 0.38 0.77 0.65
UBEK ·0.28 ·0.25 0.19 0.18 0.39 0.36 0.66 0.61
ONBEK ·0.15 ·0.12 0.10 0.08 0.21 0.17 0.35 0.29
KOOBR ·2.48 ·0.13 1.71 0.09 3.46 0.18 5.91 0.31
KHAKAR ·1.07 ·0.11 0.74 0.07 1.49 0.15 2.55 0.25
KM\lAE ·0.1 ·0.03 0.07 0.02 1.44 0.04 0.24 0.07
[MBEKR ·2.29 ·0.46 1.58 0.32 3.20 0.64 5.46 1. 10
[MHAKAR 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
[MIIAER 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Varied coefficient
Varied coefficient
Endogenous variable BO.25 BO.2 BO.l0 C4.21
0 EL 0 EL 0 EL D EL
UBEKI ·0.69 ·0.55 0.52 0.42 0.93 0.74 1.78 1.42
UBEK ·0.59 ·0.51 0.45 0.39 0.80 0.69 1.52 1.31
OHBEK ·0.88 ·0.69 0.69 0.54 1.19 0.93 2.17 1.69
KOOBR -10.50 -0.49 8.20 0.39 14.08 0.66 26.17 1.23
KHAKAR ·8.80 -0.82 6.87 0.64 11.88 1. 11 21.44 2.00
KMIIAE .1. 06 -0.37 0.82 0.29 1.44 0.50 2.55 0.89
IMBEKR ·4.65 -0.69 3.56 0.53 6.21 0.93 12.38 1.85
IMHAKAR 1.76 0.62 -1.31 ·0.46 ·2.43 0.86 -4.16 -1.47
IMIIAER 0.11 0.05 ·0.09 ·0.04 ·0.16 ·0.07 -0.27 -0.12
Norm 0.24 0.19 0.32 0.60
'---.
528
Varied coefficient
Endogenous variable 80.25 80.2 BO.l0 C4.21
0 EL 0 EL 0 EL D EL
UBEKI ·0.95 ·0.86 0.64 0.58 1.14 1.04 2.35 2.15
UBEK ·0.81 ·0.78 0.55 0.52 0.98 0.93 2.02 1.93
ONBEK ·1.55 ·1.47 1.08 1.03 1.85 1.75 3.68 3.49
KDOBR ·16.39 ·0.87 11.46 0.61 19.52 1.04 39.25 2.08
KHAKAR ·16.09 ·1.92 11.28 1.35 19.19 2.30 38.08 4.55
KMIIAE ·3.03 -1.55 2.16 1.11 3.69 1.89 7.24 3.71
IHBEKR -2.80 ·0.31 2.02 0.22 3.44 0.38 8.90 0.97
IHHAKAR 4.52 0.84 ·3.09 -0.58 ·5.43 -1.01 -10.76 ·2.01
IMIIAER 0.74 0.32 ·0.54 -0.23 ·0.92 ·0.39 ·1.80 ·0.76
To summarize the results reported in Tables 2-5, the following major points emerge:
2. Increasing the structural parameter of domestic order intakes (BO.2) by the amount
shown in Table 1, has positive effects on 7 and negative effects on 2 endogenous
variables (only in Table 2 are the corresponding effects zero).
3. Similar results are obtained by varying the structural coefficient of the business
climate index of retailing on a higher level: A one percent increase of BO.1O causes
an increase in point forecasts nearly twice as large as a corresponding change of
BO.2.
4. The autoregressive term (C4.21) has the greatest influence on the 9 endogenous
variables: both on impact (Table 2) and in the longer run (Tables 3-5). An increasing
parameter produces a strong increase in production and imports. The great sensitivity
of endogenous variables to "own-dynamics" parameter perturbations, apparent also
for the other endogenous variables in Tables 3, 4 and 5, indicates its importance in
the context of this model. This is a result appearing also in corresponding
investigations of other models [Kuh, Neese, Hollinger (1985), 210, 229].
The dynamic analysis (Tables 3-5) shows an amplification of structural sensitivity, mostly
at a factor of five from impact to twenty-five quarter computations. In the short-run, the
greatest structural sensitivity with respect to variations of the parameters of the equation
explaining domestic turnover of apparel appears for imports of apparel (IMBEKR). In
the long-run this effect will be weakened and the response of the production of men's
outer garments (KHAKAR) is largest.
The results of the structural sensitivity analysis of the model of the West German textile
and apparel industry give also information about the forecast quality of this model. For
the prediction of the endogenous variables possible structural changes have to be taken
into consideration. The empirical analysis of the responses of the point forecasts of the
endogenous variables with regard to structural parameter variations shows at the
impact and in the long-run relatively small sensitivity effects. The point forecasts of the
model of the West German textile and apparel industry are relatively robust against
structural changes.
A graphical summary is given in Figures 2 - 10. The time paths of several elasticities
with respect to each coefficient are scaled so that their individual largest magnitude will
be + 1. For example, the scale factor of the elasticities series of the endogenous
variable ONBEK with regard to the coefficient C4.21 is 3.59 (see Table 6).
Constructing corresponding overlay graphs for each endogenous variable facilitates the
comparison of the individual time paths [Kuh, Neese (1982), 149 ff, Kuh, Neese (1982a),
128]. The reactions of point forecasts of an endogenous variable to each coefficient
change are nearly identical in course of time. The variables show more or less damped
oscillations. The fluctuations of elasticities are influenced by the time path of the
associated variable of the parameter (see Table 1) and by the fluctuations of the point
forecast itself. The maximum values in Table 6 complete the summary.
BO. 25 ·1.01 -0.94 -1.53 -1.08 -2_04 -1.63 -0_84 1.1 0.38
C4. 21 2.15 1.94 3.59 2.45 4.78 3.89 1.85 -2.62 -0.92
Fig.2: Scaled elasticity response~ tor UBEKI to Fig.3: Scaled elasticity responses tor UBEK
variations 01 the tour coefficients to variations ot the tour coefficients
1,00 , A •. _ _ • 1,00 I A ,· jL .' " 1
0.25 r ¥ 0.25 r V
U1
W
o
Fig. 4: SC<lled elasticity responses tor ONBEK to Fig. 5: Scaled elasticity responses tor KDOBR to
variations ot the tour coefficients variations 01 the tour coerticients
0.25 I --,7
FATM FATM
Fig. 6: Scaled elasticity responses for KHAKAR Fig. 7: Scaled elasticity responses for KMWAE to
to variations of the four coeffic ients variations of the four coefficients
0,00 .fr;=,-
, .,-.,..-,-,--,-.,-.,..-,--r-,-.,-.,..-,--.--,-.,-,.....,.-.--,--,-""""
..73 1176 ,err IIU, 1Ii~13 liU linT
U1
W
~
Fig. 8: Scaled elasticity responses for IMBEKR Fig. 9: Scal ed elaslicity responses lor IMHAKAR
to variations of the four coefficients to varialions 01 the four coefficients
1,00 I .~ j 1,00 I X
0,75 I . Y 1 f , 0,7&
" I I '~ V ,
£f14 I IIV;S' V""
0 ,50 .J----------""dt~f------j
0 ,25 + - - - - - -- -....~-----------i
5 CONCLUSIONS
This paper illustrates an approach to the structural sensitivity analysis of point forecasts
with respect to structural coefficients in linear dynamic equation systems. Improved
methods have been developed by Elker (1986a, 1986b), using e. g. algorithms for
calculating inverse matrices without repeated inversions. The corresponding software
package which was written at the University of Hagen has been further developed and
been applied to the econometric forecasting model for the textile pipeline of West
Germany. Empirical results illustrate the structure of the forecasting model and give
information about its forecast sensitivity. The available space made it necessary to limit
the presentation to a few examples. A much greater volume of results will be published
in Terlau (1991).
APPENDIX
List of Variables of the Model of the West German Textile and Apparel Industry
Endogenous Variables
Exogenous Variables
RAUSB Construction
REFERENCES
Belsley, D.A, Kuh, E., Welsch, R.E. (1980): Regression Diagnostics - Identifying Influential Data and
Sources of Collinearity, Wiley Series in Probability and Mathematical Statistics, New York,
Chichester, Brisbane, Toronto, Singapur.
Elker, J. (1986a): Sensitivity Analysis of Point Forecasts in Dynamic Econometric Models, Discussion
Papers of the Department of Economics, University of Hagen, No. 111.
Elker, J. (1986b): Forecast and Multiplier Sensitivity Analysis with Respect to Changes of Equations and
Submodels in Dynamic Systetns (revised version of July 17, 1987, is published in this volume).
Elker. J .. Gruber, P. (1987): Program for Sensitivity Analysis of Multipliers and of Point Forecasts in
DynamiC Econometric Models, University of Hagen.
Gilli, M. (1984): CAUSOR, A Program for the Analysis of Recursive and Interdependent Causal
Structures, Version 3.10, Cahiers du d~partement d'~conom~trie, Universit~ de Geneve,
Cahier 84.03.
Gruber, J. (1978): Zur Interdependenz i:ikonometrischer Gleichungssysteme, in: J. Frohn (Ed.),
Makroi:ikonomische Modelle fUr die Bundesrepublik Deutschland, Sonderhefte zum
Allgemeinen Statistischen Archiv, No. 12, GOttingen, 185-213.
Heilemann, U. (1989): "Was leisten Prognosemodelle?" Eine empirische Untersuchung am Beispiel des
RWI-Konjunkturmodells, in: B. Gahlen u.a.(Eds.), Wirtschaftswachstum, Strukturwandel und
dynamischer Wettbewerb, Festschrift fUr Ernst Heltnstiidter, Berlin, Heidelberg, New York,
253-272.
Helmstadter. E. (1962): Die geordnete Input-Output-Struktur, in: Jahrbucher fur Nationali:ikonomie und
Statistik 174, 322-361.
He1mstiidter, E. (1973): Was leisten Prognosemodelle?, in: Zeitschrift fur allgemeine und textile Markt-
wirtschaft, Tagungsband Textilwirtschaftlicher Prognosen, No. 2/3, 118-150.
Helmstadter, E., Krupp, R., Meyer, B. (1976): Ein i:ikonometrisches Prognosemodell fur die Textilwirt-
schaft der Bundesrepublik Deutschland, Schriften zur Textilwirtschaft 19, FATM (Ed.),
Miinster.
Kleine, E., Thiele, W. (1982): Ein disaggregiertes Prognosemodell fUr die Textilwirtschaft der Bundesre-
publik Deutschland, Schriften zur Textilwirtschaft 27, E. Helmstadter (Ed.), Frankfurt a.M.,
Bern.
Kuh, E., Neese, L. (1982): Parameter Sensitivity, Dynamic Behavior and Model Reliability: An Initial
Exploration with the MQEM Monetary Sector, in: E.G. Charatsis (Ed.), Proceedings of the
Econometric Society European Meeting 1979, in Memory of Stefan Valavanis, Athens, 121-
168.
536
Kuh, E., Neese, L. (l982a): Econometric Model Diagnostics, in: G.e. Chow, P. Corsi (Eds.), Evaluating
the Reliability of Macro-Econometric Models, Chichester, New York, 119-146.
Kuh, E .. Neese, J.W., Hollinger, P. (1985): Structural Sensitivity in Econometric Models, Wiley Series in
Probability and Mathematical Statistics, New York, Chichester, Brisbane, Toronto, Singapore.
Simon, H.A (1953): Causal Ordering and Identifiability, in: W.C. Hood, T.e. Koopmans (Eds.), Studies
in Econometric Methods, New York, 49-74.
Terlau. W. (1989): Program of Structural Sensitivity Analysis in Dynamic Econometric Models, A Modi-
fied Version of Elker, J., Gruber, P. (1987), Munster.
Abstract
A program for Linear Energy Optimization (LEO) which was used to investigate ther-
modynamical and technical options of reducing the energy consumption of industrialized
countries is extended to handle the cost associated with the realization of the calculated sa-
vings. We do vector optimization a) by parameterizing the cost function and b) by forming
a scalarizing function from the individual objective functions. As a result, we are able to
show the dependence of the possible energy savings on the energy price and determine the
trade-off between energy conservation and cost.
Keywords
Contents
1 Introduction
2 Definitions and Assumptions in the Energy Optimization Model
3 The Cost Function
4 Vector Optimization
5 Conclusions
1 Introduction
law of thermodynamics), its quality decreases in any real process of work performance and
information processing (second law of thermodynamics). Excess heat of industrial processes
is often emitted into the environment (e.g., by cooling towers), even if its temperature T
is still higher than the temperature of the environment and its quality is therefore still
greater than zero. Based on energy demand profiles that indicate the process heat demand
of industrialized countries at various quality levels, we present an optimization model that
determines strategies to make the best possible use of this excess heat before its quality
finally reaches zero. The model has been described in detail in Refs. 3 and 4. We repeat
some important definitions and assumptions in Sec.2. It is the purpose of this paper to
develop a vector optimization procedure that may be helpful in determining the conditions
under which market forces will drive the development of industrial systems towards greater
energy efficiency. The guiding principles are those of thermoeconomics [5].
34% in the United States. For this result it was assumed that waste heat from the use of
electricity cannot be reused as secondary energy.
In a more realistic scenario energy losses during heat transportation are included [4]. If
all heat from heat exchangers has to be transported over 50km and heat from cogeneration
plants over 25km and if at the same time the efficiency of heat pumps is reduced to 0.6 of
the Carnot efficiency, the remaining industrial saving potentials are 25% for West Germany,
30% for the United States and about 45% for the Netherlands and Japan.
The definitions for the cost function given in Ref.6 are C = total annual cost for complete
satisfaction of process-heat-demand of the considered industrial system; Co = total cost
before optimization and b(F) = price of one unit of fuel F at the location where it is used,
including all cost like buying price, transportation etc.
All following cost factors are defined to include investment, capital cost and operation
(e.g., maintenance) cost as long as they are not fuel cost: c(q, F) = cost of producing one
enthalpy unit q from fuel F; « q, q') = cost of providing one enthalpy unit q from secondary
energy q' via technology i (i = ex, p); cc( 10) = cost of producing one uni t of electrical energy
in a cogeneration plant. The cost of producing one enthalpy unit q via cogeneration is set
equal to zero because all cost for cogeneration will be charged to electricity production.
With the assumptions of Sec.2 still valid, we can write down the function for the total
cost C per year:
The first row of Eq.(3.1) contains the cost of process-heat-supply on levels q = 1, ... ,9
before optimization. The demanded amount of energy, n( q), is multiplied by the specific
cost, which consists of the cost factor derived from investment, capital cost and operation
cost, c(q, F), and the fuel price b(F). While the cost factors are given with respect to a
produced enthalpy unit q, the fuel price is based on one unit of fuel F. The latter must
therefore be divided by the efficiency 7]. The next three rows contain saved and added cost
by heat exchangers (ex), cogeneration (c) and heat pumps (p). The specific cost factors
542
for conventional energy supply multiplied by the amount of energy supplied by the new
technologies (e.g., .,,~x( q, q')xex( q, q')n( q') in the second row) yield the saved amount of money.
On the other hand there will be new costs for the installation and operation of the new
technology, i.e., <Aq, q')"'~x(q, q')xex(q, q')n(q') for each pair of levels q, q' in the second row.
The cost of cogenerated electricity is indicated in the fifth row. The last row gives the cost of
electricity production in all-electric power plants. It includes the electricity demand of heat
pumps and the respective cost. e is the Heavyside function and Pp is a function that has
positive values if a heat pump works efficiently and negative values otherwise [4]. The two
underlined terms indicate the cost without optimization, Co. Using an adequately defined
specific transportation-cost factor, it is possible to include the (considerably high) cost of
heat transportation into the model.
4 Vector Optimization
The mathematical optimization problem consists of two objective functions, namely the
amount of primary energy, N(F), needed to satisfy the electricity and process-heat demand
of an industrial system and the total cost C of the energy system. While environmental
protection and cost are often considered as examples of usually conflicting objectives of
vector optimization, energy conservation may lead to a reduction of cost and pollution, if
only energy prices are high enough.
Since for conflicting objective functions there is no single optimal solution to the mini-
mization problem, a possible solution is called functional efficient or pareto optimal if none
of the objective functions can be reduced further without increasing the other one.
It is the goal of this investigation to determine under what circumstances the minimiza-
tion of energy consumption and minimization of the cost of energy services may be parallel
objectives, which means that pursuing one objective will at the same time optimize the other
one. In our case there are two ways of tackling the problem: (a) parameterization of one
objective function [6] or (b) the linear combination of the two weighted objective functions
[7]. In the following sections we summarize the results of both methods.
The cost factors used in this paper were obtained by van Gool and Kummel for an earlier
investigation [2]. 'While, in Ref.2, they were given in dollars, we now prefer to use arbitrary
currency units (ACU) in order to avoid problems with exchange rates and inflation. For
orientation, the reader may note that a primary energy price of 4 ACU jGJ corresponds to
an oil price of 24$/ barrel before 1986. The cost factors include an annuity factor of 0.175 / a.
They are c(q, F) = 0.39 ACUjGJ·a, c(10, F) = 3.33 ACUjGJ·a, c:x(q, q') = 1.40 ACU/GJ·a,
c~(q, q') = 3.50 ACUjGJ·a, cc(IO) = 3.85 ACUjGJ·a. The cost factor for transportation-cost
has been chosen to be 1.0 ACUjGJ [8].
The cost function of Eq.(3.1) is used to introduce an upper cost limit, which serves
as an external parameter and which is defined as a function of the cost without saving
technologies, Co. An energy conserving technology is called cost efficient, if its employment
is less expensive than using primary energy directly [at a given primary energy price, b(F)].
If the cost restriction implies that energy savings must be realized without allowing for
additional cost, a saving potential can only be found, if at least one of the considered
543
Co
3.0 32
'-
~
2.8 " 24
2.6 16
N(F)
2.4 8
~ - -<>- - -<>- -
2.2+----.---.----,---.----.----.---.----.---.----+0
2.0 4.0 6.0 8.0 10.0 12.0
b( F) [ACU /GJ]
Fig.I. Optimization results for West Germany; upper cost limit at 1.0 . Co (solid
curves) and 1.1 . Co (dashed curve); displayed are the requi red primary energy
N(F) (for both scenarios) and the cost with (C) and without (Co) energy
saving technologies (for the first scenario) vs. the fuel price b(F). The optimi-
zation value without cost restriction is 2321 P J / a. Nate the suppressed zero
point on the N(F)-scale on the !.h.s. [6].
technologies is cost efficient. Taking into account the additional energy needed for heat
transportation, all technologies considered in this paper need extra exergy input for their
operation.
In Fig.1, optimization results for vVest Germany are shown at fuel prices rising from 2 to
12 ACU/GJ. The first scenario (solid curves) allows for no extra cost, while the second one
(dashed curve) tolerates cost which are 10% higher than Co. At the current very low energy
prices, which lie somewhere on the left edge of Fig.1, none of the discussed technologies is
cost efficient. With rising energy prices. heat pumps would be the first technology to become
cost efficient between 4 and.5 ACU/GJ. This is the point, where the solid N(F)-curve, which
represents the amount of primary energy needed to satisfy all energy demand, starts to
decrease. It keeps decreasing with rising fuel prices until it reaches its minimum value of
2321 PJ/a, which is about 74% of the amount of primary energy without energy savings.
The minimum is identical with the value one gets for optimization without cost restriction.
For fuel prices higher than 10 ACU /GJ, all energy saving technologies are cost efficient.
The cost without energy saving measures, Co, is linearly rising with increasing fuel price.
Due to the restriction of the first scenario, the actual cost, C, has to be smaller than or
at most equal to Co- For fuel prices up to 10 ACU /GJ the allowed cost is fully used, the
curves for Co and C are identical. For prices higher than 10 ACU/GJ it will be increasingly
cheaper to use saving technologies, the curve C is now running below Co. For the energy
544
demand-profiles of the ."l"etherlands, the United States and Japan one finds analogous results
which lead to similar conclusions.
The dashed curve in Fig.1 shows the results of optimization if the cost C of energy supply
may be 10% higher than Co. Now, even for the lowest fuel price of 2 ACU /GJ, savings are
possible. They continue to increase until they reach their maximum at 7 ACU/GJ.
Fig.2. Trade-off between primary energy consumption N( F) and the total cost C for
West Germany at fuel prices b(F)=4 ACU/GJ (0) and 8 ACU/GJ (+) [7].
where (3 is the weighing factor. The index '0' indicates the magnitude of N(F) and C when
no saving technologies are applied. The weighing factor is neither a physical nor an economic
quantity but an external parameter expressing preferences of the decision-maker.
Optimization was performed for values of (3 between 0.2 and 0.7. Each value of (3 leads to
a pareto-optimal solution that can be viewed as a compromise between the two conflicting
objectives. For a given value of (3, the amount of primary energy needed to satisfy the
demand for process heat and electrical energy, N( F), was determined together with the
respective cost, C. In Fig.2 we plot N(F) versus C, which means that every point in the
figure represents a destinct value of (3. One can now directly determine the trade-off, i.e.
evalutate how much money each unit of additionally saved primary energy will cost. For the
considered low energy price scenario [b( F) = 4 ACU /GJ] there is an increase of nearly 30%
545
in total cost between the maximum and minimum primary energy use. For higher fuel prices
[b(F) = 8 ACU/GJ], this increase is reduced to about 6%, but the whole curve is shifted
towards higher total cost.
There is a second advantage of the new objective function in Eq.(4.1): For values of
j3 > 0.7 the program LEO selects from the possibly very many energy optimal solutions
the one that requires the least cost. On the other hand, for values of j3 < 0.2, among the
solutions without cost increase the one which uses the smallest amount of energy is selected.
5 Conclusions
Using a highly aggregated model of the industrial energy demand in West Germany, we
have demonstrated the dependence of energy saving potentials on the energy price. As a
result of vector optimization, we may say that, at current energy prices, the realization of
the maximum savings potential will lead to more than 30% higher cost. If, however, energy
prices are increased, the relative increase in cost is reduced, while, at the same time, the total
cost at minimum energy use is higher. Therefore, if we want to stimulate energy conservation
and thus pollution abatement we have two choices: Either double or triple the energy price,
e.g. by high energy taxes, so that energy saving and cost minimization become parallel
objectives on a high total cost level, or accept a moderate cost increase on a much lower
level. The latter could be achieved by appropriate subsidies, the funds for which might be
raised by a relatively low energy tax.
References
1. W. van Gool, "The Value of Energy Carriers", Energy 12, 509 (1987).
2. W. van Gool and R. Kiimmel, "Limits for Cost and Energy Optimization in Macrosystems",
in: Miyata, Matsui eds., Energy Decisions for the Future, VoU, Tokyo (1986), pp. 90-106.
3. R. Kiimmel, H.-M. Groscurth and W. van Gool, "Energy Optimization in Industrial Models",
in: M. Iri, K. Yajima eds., Proceedings of the 13th IFIP-Conference on System Modelling and
Optimization, Springer, Tokyo (1988), pp. 518-529.
4. H.-M. Groscurth, R. Kiimmel and W. van Gool, "Thermodynamic Limits to Energy Optimi-
zation", Energy 14, 241 (1989).
5. VV. F. Kenney, Energy Conservation in the Process Industries, Academic Press, Orlando, FL
(1984).
6. H.-M. Groscurth, R. Kiimmel and W. van Gool, "Cost Aspects of Energy Optimization", in:
W.W. Koczkodaj, R. Janicki eds., Proceedings of the International Conference on Computing
and Information (ICCI '89), Elsevier, Amsterdam (1989), pp. 481-486.
7. H.-:"1. Groscurth, R. Kiimmel, "Energy and Cost Optimization in Industrial Models", in:
N.N. (eds.), Proceedings of the 14th IFIP-Conference on System Modelling and Optimization,
Springer, Berlin, Heidelberg (1989), in print.
8. A. Tautz et al., "Fernwarmeschiene Niederrhein", BMFT research report T 84-167, University
of Dortmund, FRG (1984).
ENERGY RESOURCES DEPLETION AND ECONOMIC GROWTH
01 eg A. Eismon t
Institute for Systems Studies
USSR Academy of Sciences
9, Prospect 60 Let Octyabrya
117312, Moscow, USSR
: L,,=L
c..
: Lr,=L
c:.
(9 )
549
,.1'
.'
I
.'~
--- , ... ,
1"·'-- -....... I'
.' "- .......
It I "
,.It II,. .........
.... . ....
~ ',~
,.",
t' -:,~
t
Fig.1
REFERENCES
1. A.Lc,vins, L.H.Lovins, F.Krause, W.Bac11, "Energy Strategy
for Low Climatic Rl sks" , Report for the German Federal
Environmental Agency, 1981.
~. SJ'!Tlposium 011, th~ EcoTlOmics of Exhaustible ResoLU'ces,
ReVIew of ~conomic ~tudles, 1974 vol.41.
3. H.Hotelling, "The Economics of Exhaustible Resolu~ces",
Journal of politf{~al Economy, 1931, vo1.39, pp. 137-175.
4. W.Krelle, "Theorie des Wirtscl1at"tIicl1en Wachstul11s",
Springer Verlag, Heidelberg, 1985.
Econometric methods as an instrument for estimating external
costs per unit of emission: the case of a groundwater model
Monika Vorreyer
Fern U niversitat Hagen
Lehrgebiet Statistik und Okonometrie
Feithstr. 140, D-5800 Hagen 1, FRG
Abstract
Econometric methods are used in this study to analyze the cause-effect relationships of the sulfur
dioxide and nitrate emissions on the damage costs of groundwater pollution (purification costs)
and on the nitrate induced avoidance costs. A simultaneous 4-equation groundwater model was
constructed for this purpose. By means of the reduced form coefficients the results provide
estimates of the external costs per unit of emission which are caused by power economical and
agricultural activities and which are imposed on the water-supply enterprises (68.2 DM/t sulfur
dioxide and 327 DM/t nitrate).
Keywords: Ecological modeling; external costs; damage costs of groundwater pollution;
emission-multipliers.
Contents
1 Introduction
2 Emissions, immissions and environmental damages
3 External costs of groundwater pollution
4 Econometric groundwater model
5 Concluding remarks
Appendix
References
1 Introduction
Econometric methods have been applied mainly in the field of modeling macroeconomic theories
and relationships. More recently, they have also been used in health-care economics, social,
natural and technical sciences. A number of studies l have linked different variables (among
other things environmental factors like specific air pollution levels) to health damage (the damage
indicators being mortality and morbidity).
The aim of this paper is to outline some major results of an econometric ~roundwater mo-
del. This model was especially constructed to show that the cause-effect relatIOnships between
emissions, immissions and monetarized environmental damages can be modeled.
4.2 Specification
The model specification stands at the beginning of building an econometric model (see e.g.
GRUBER 1968, 1981). One task is the determination of the explanatory variables of the ground-
water model. By expert considerations and by the availability of data the explanatory variables
are given for three equations, the WGS-, KB3- and KANNEU-equations listed below. For
the WGN2-equation, which explains the effect of variations in agricultural p{)llution on the
groundwater Immission by nitrate, there exists a broader spectrum of potential explanatory
553
variables. The proced ure to select the explanatory variables of this equation involved stepwise
regression.
The specified interdependent groundwater model has the following structural form 2 :
4.3 Estimation
The second phase of the construction of the econometric groundwater model is the parameter
estimation. The four equations of the groundwater model are linear in the parameters and
there are no lagged endogenous variables. The first and second equations are estImated by OLS,
the third and fourth ones, which above all describe the effect of the variations in groundwater
immissions on the additional purification and avoidance costs, are estimated by TSLS. Table 1
reproduces the regresssion results for some interesting relationships, that means for the struc-
tural and reduced form coefficients of the explanatory variables EMS (sulfur dioxide emissions),
EMN (nitrate emissions), WGS (sulfur dioxide induced immissions) and WGN2 (nitrate induced
immissions ).
4.5 Emission-multipliers
The estimated structural coefficients quantify the external costs per unit ofimmission. By means
of the relevant calculated reduced form coefficients, it is possible to attribute the external costs
to the polluters energy and agriculture.
The marginal external costs, which are caused by the sulfur dioxide emissions from the
combustion of fossil energy and imposed on the water-supply enterprises are estimated at 68.2
DM/t S02 (sulfur dioxide emission-multiplier with respect to the additional purification costs).
The marginal external costs which result from fertilizing activities of agriculture and which
the water-supply enterprises have to bear are estimated at 327 DM/t Noi (nitrate emission-
multiplier with respect to the nitrate induced avoidance costs).
5 Concluding remarks
The econometric groundwater model provides the basis for causal-analytically quantifying partial
effects of the emissions on the additional costs of the water-supply enterprises (purification costs
and nitrate induced avoidance costs). The estimated structural regression coefficients and the
reduced form coefficients derived therefrom indicate the order of magnitude for the marginal
external costs per unit of emission to be internalized.
Appendix
List of endogenous and exogenous variables of the econometric groundwater model
Endogenous variables:
WGS - fraction of the output of raw water with a sulphate content 2: 60 mg/I, in %
WGN2 - output of raw water with a nitrate content 2: 40 mg/l, in 1000 m 3
KB3 - by groundwater pollution induced damage/abatement costs, in 1000 DM
KANNEU - nitrate induced avoidance costs, in 1000 DM.
Exogenous variables:
EMS - sulfur dioxide emissions, in tons
DG - meadows, in ha
AL - arable land, in ha
WF - forest, in ha
WSGB - water protection areas, in ha
SOKU - special cultures, in ha
EMN - nitrogen oxide emissions, in tons
LHKW2 - raw water with a content of volatile halogenized hydrocarbons 2: 0.013 mg/I, in 1000 m 3
ABI - raw water-supply plants with chemical-physical purification, number
3 As there are no data available for the groundwater pollution by nitrates arising from agricultural activities,
the external costs per unit of (nitrate) emission are calculated by means of the external costs per ha arable land
and the degree of the losses by washing out and denitrification.
555
References
1. CHRISTAINSEN, G.B., DEGEN, C.G. (1980)
Air Pollution and Mortality Rates: A Note on Lave and Seskin's Pooling of Cross-Section
and Time-Series Data. Journal of Environmental Economics and Management 7, 149-155.
2. GRUBER, J. (1968)
Okonometrische Modelle des Cowles-Commission- Typs: Bau und Interpretation. Verlag
Paul Parey, Hamburg - Berlin.
3. GRUBER, J. (1981)
Okonometrie 2: Formen. Typen und Spezifikation iikonometrischer Modelle. Kurse 0860-
3-01-S1 und 0860-4-02-S I, Fernuniversitat - Gesamthochschule - in Hagen, Fachbereich
vVirtschaftswissenschaft, Hagen.
4. GYSIN, Ch.H. (1985)
Externe Kosten der Energie in der Schweiz. Reihe Okologie Band 5, Verlag Rliegger,
Grusch.
5. KRAMER, W., SONNBERGER, H. (1986)
The Linear Regression ~Iodel Under Test. Physica-Verlag, Heidelberg - \Vien.
6. LAVE, L.B., SESKIN, E.P. (1970)
Air Pollution and Human Health. Science Vo!' 169, 723-733.
7. LEU, R.E., GYSIN, Ch.H., FREY, R.L., SCHMASSMANN, N. (1984)
Externe Kosten der Energie am Beispiel von !"[ortalitiit und Morbiditat. Schweiz. Zeit-
schrift fUr Volkswirtschaft und Statistik Heft 3.
8. OSTRO, B.B. (1983)
The Effects of Air Pollution on Work Loss and Morbidity. Journal of Environmental
Economics and Management 10,371-382.
9. RIDKER, R.G. (1971)
Economic Costs of Air Pollution: Studies in Measurement. Praeger Publishers, New York
- Washington - London.
Heinz P. Galler
F achbereich Wirtschaftswissenschaften
Johann Wolfgang Goethe-UniversiUit
Senckenberganlage 31, D-6000 Frankfurt 11, FRG
Abstract
Contents
In this approach, a quantitative model is used to describe the processes relevant for politics
as well as to assess the effects of different policy measures on these processes. Traditio-
nally, macroeconometric models have been used in economics for this purpose based on
aggregated data derived from the national accounts. However, there are some short-
comings in such an approach related to the level of aggregation.
First of all, policy makers are usually not only interested in broad aggregates like national
income, total consumption or investment, but they also have to consider distributional and
structural aspects of the economic process. This implies that aggregated models cannot
provide all information needed. In addition, models are required which allow to assess the
distributional and structural impact of policy measures. This can only be achieved by more
disaggregated approaches.
Secondly, at least some of the instruments of economic policy have a rather complex design
and depend strongly on structural characteristics of the economy. An example is income
taxation with its set of complex regulations, depending on individual income and other cha-
racteristics of the tax payer. In most cases, such a tax schedule cannot be modelled appro-
priately on an aggregated level since the impact of taxation depends strongly on the struc-
ture of the tax base. Structural aspects must be considered in great detail in order to assess
the effects of changes in the tax law appropriately. This requires disaggregated models
allowing to analyse such structures.
In addition, models based on aggregated data also suffer from some methodological defi-
ciencies 2. One problem is that the reliability of econometric forecasts depends strongly on
the stability of the underlying structural relations. Only if the basic structure remains
unchanged, valid forecasts can be derived from a model. But since the parameters of
macroeconometric models represent weighted means of the corresponding parameters on
the micro level of individual economic agents, structural change on the micro level will
result in parameter shifts on the macro level, even if the individual micro relations remain
unchanged. As a consequence, aggregated models may be unreliable in periods of structu-
ral change. However, since structural effects are considered explicitly, models on a lower
aggregation level are less affected by this problem.
1 A general overview on the Sonderforschungsbereich 3 can be found in Hauser and Hochmuth (1989).
Cf. Orcutt (1986).
561
Another problem concerns the efficiency of parameter estimates. Since usually only a
comparatively small number of observations is available for time series data on the macro
level, estimates based on macro data are of limited statistical efficiency. In micro data sets
on the other hand, a comparatively large number of practically independent observations
with detailed information on structural characteristics is available. This allows not only
more detailed analyses, but also more efficient parameter estimates. Thus, behavioral rela-
tions on the micro level do not only tend to be more stable but the parameters can be esti-
mated more efficiently than in macro relations.
However, in the past the lack of suitable micro data, the predominance of cross sectional
data not allowing dynamic analyses and comparatively large computational requirements
have been an obstacle for a vast application of microeconometric approaches. But since
sample surveys have become a standard source of information, an increasing number of
micro data sets has become available to researchers not only in the USA but also in Euro-
pean countries. Additionally, the limitations of cross sectional data have been overcome by
the availability of panel surveys containing longitudinal information on micro units 3. In
parallel, the costs of computing have been reduced dramatically. Thus, even dynamic
microeconometric analyses have become feasible which could not be conducted a few years
ago.
3 One of the major projects at the Sib 3 was the German Socio-economic Panel Study which is now continued
by the Deutsches Institut fUr Wirtschaftsforschung, Berlin. In this panel, a sample of about 6000 households
in the Federal Republic of Germany has been followed up since 1984 in annual interviews. Besides other
information, the panel contains detailed longitudinal information on the the socio-demographic structure of
the households, on education, labor force participation, employment and unemployment, on earnings and
other income, on taxes and transfers on the level of individual persons. Cf. Autorengemeinschaft Panel 1990.
4 At the Sib 3, a medium sized macroeconometric model of the business cycle has been developed. Cf.
Hujer/Hansen/K1ein 1988.
5 Selected references are given in the bibliography at the end of this contribution.
562
model for the individual propensity to leave unemployment was used with individual cha-
racteristics like nationality, education and age as explanatory variables. In addition, the
impact of unemployment compensation was controlled for as well as seasonal factors.
Duration dependence of the hazard rate is modelled in the Weibull specification by an
exponential time trend.
In table 1 parameter estimates are given for a specification without unobserved heteroge-
neity of the micro units as well as a specification with unobserved heterogeneity, using the
semiparametric approach of Heckman and Singer (1982). Except for the Weibull parame-
ter for duration dependence, the estimates for the two specifications do not differ signifi-
cantly. German nationals show a higher propensity to leave unemployment than foreigners.
Better education also reduces the duration of unemployment. The chances to leave unem-
ployment are reduced in the higher age brackets. There is a bigger chance to become
unemployed during spring and summer 6. The results also indicate that the entitlement to
unemployment benefits, granted in the first months of unemployment (Arbeitslosengeld),
does not affect the exit rate significantly whereas subsidiary benefits (Arbeitslosenhilfe),
granted after unemployment benefits have expired, seem to reduce the exit rate.
The two different specifications for unobserved heterogeneity result in significantly diffe-
rent estimates for duration dependence. While the specification without unobserved hete-
rogeneity shows no significant duration dependence, the estimates based on the semipara-
metric Heckman/Singer specification imply a significant positive duration dependence, i.e.
the exit rate increases with prolonged duration of unemployment. This result is in accor-
dance with the results of search theory. However, analyses by Galler and Potter (1990) cast
some doubt on such estimates. Their research shows that the Heckman/Singer estimator
seems to be rather sensitive with regard to deficiencies of the data. Therefore peculiarities
of the data may result in an estimate of negative duration dependence, even if it is not the
true relationship.
6 The large coefficient for December results at least partially from a methodological artefact. Information on
the unem ployment spells is collected annually by retrospective interviews. Since some respondents forget to
report the continuation of unemployment in the next interview, transitions out of unemployment are
overestimated for December.
563
Table 1: Weibull Model for the Transition of Men from Unemployment to Employment
A! 1.0000 0.0665
A2 (State Probabilities) 0.4219
A3 0.5116
-2.4302 0.2822 -1.2652
'\ 0.0880 0.0000 0.0000
02 (Support Points) 0.1073 -2.2320 0.0000
A certain disadvantage of the micro approach is that no direct conclusions can be drawn
from micro relations with regard to variables and structures on the macro level. But on the
other hand the ultimate interest, at least of policy orientated research, is to infer about the
implications of given processes or of policy measures on a more aggregated level. In order
to derive such aggregated results, the structure of the population must be considered in
addition to the micro relations. For this purpose, microanalytic simulation techniques have
been developed at the Sonderforschungsbereich 3 allowing to derive aggregated estimates
from micro relations 7.
In these simulation models, the micro relations are applied to a representative sample
from the population considered. Based on the results for the individual micro units, the
corresponding aggregates are estimated for the total population or subgroups of interest.
Thus, the micro simulation approach provides an instrument for inferences about the
aggregated consequences of policy measures, using micro relations for individual units.
Detailed structural properties of the economy can be taken into account explicitly in esti-
mates derived from microsimulation models.
An example is an analysis of the distributional effects of the German reform of direct taxa-
tion in 1990, conducted by van Essen et al. (1988,1988a) using a static microsimulation
model. While the effects of changes in the tax schedule on individual income can be asses-
sed easily, the net effect for different groups is much more difficult to estimate. Depending
on the individual characteristics of the tax payer, the different measures imply reductions
and eventual increases in the tax burden which may compensate each other. Therefore the
joint distribution of taxable income and other characteristics of the tax payers must be
taken into account when estimating the net effect of the tax reform. This cannot be done in
a macroeconomic model.
For this simulation study, the micro data of the first wave of the German Socio-economic
Panel study of 1984 have been used. This data set contains rich information on family and
household structures, on taxable income and its components as well as on other variables
required for computing the income tax. However, since some pieces of information like
costs of maintaining houses are missing, some imputations had to be applied to the data.
Monetary variables have been adjusted to the levels of 1988 and 1990 by adjustment fac-
tors. However, the basic socio-demographic structure of the sample was not changed by re-
weighting since no significant changes occur in such a short period of time. Besides, no
provisions have been made to account for behavioral responses to changing tax rates. Thus,
the simulation represents the formal incidence of the tax reform.
The effects of the different tax schedules have been estimated by computing first taxes for
the individual micro units in the sample according to their characteristics and to the regula-
tions of the tax law. Afterwards, these micro results have been aggregated into estimates of
the effects for different groups as for example households in given income brackets or for
the total population. They represent the net effect of the reform, given the distribution of
the characteristics in the sample. Some basic results concerning the distribution of benefits
are presented in figure 2. In terms of the percentage of the taxes paid according to the pre-
vious schedule, low income households benefit most since a larger part of income remains
untaxed or is taxed at lower rates. Tax reductions become relatively smaller for the middle
income groups, up to about an annual income of about DM 75000, but rise again for high
incomes.
Figure 2
Formal Inoldeno. 01 I". 111110 I n r.form
(S Imulation r.eulle)
14"
eo" 12"
10"
40" 8S
es
'''''''1'1.''''''''''''''''''
J ------r. . . .. . .. . ,··· ··,. . . 1
...•..........
2"
2 20 200
I ncome (1000 0 1.4)
This is mainly a consequence of the reduction of the maximum marginal tax rate. However.
if the tax reduction is compared to income, the effects are smallest for low income groups
and increase with rising income. This effect is strongest for households with an annual
income of more than DM 60000.
When the effects are broken down by income brackets as well as the number of dependent
ch ildren, it becomes evident that households with no children benefit less than the average.
In general the absolute tax reduction increases with the number of children and with rising
income. However, in the income bracket between DM 50000 and DM 100000 the reduc-
tion for families with up to two or three children is smaller than for households without
children. This is a consequence of the reduction of tax exemptions depending on the num-
ber of children, especially for contributions to retirement funds and to thrift institutions,
that had benefitted most the middle income groups in the past. This shows that in a com-
plex set of policy measures compensating effects may occur which cannot be detected
easily unless very detailed analyses are performed. As it is possible to take into account all
relevant variables, microsimulation models are a superior instrument for such analyses.
566
4 Conclusions
In the last decade, the availability and the quality of micro data as well as the computatio-
nal facilities have been improved substantially. Besides, new methods have been developed
allowing to analyse especially longitudinal micro data appropriately. As a result, microeco-
nometric analyses have become an important tool for applied economic research. As the
example of research conducted at the Sonderforschungsbereich 3 shows, analyses based on
micro data are not only feasible but moreover can provide additional insights that cannot
be achieved by traditional macroeconomic approaches.
The micro approach has even become more important in the last years due to the structu-
ral changes taking place. Changes in society in general and in particular in politics affect
economic behavior. As a consequence, economic relations have become less stable. The
impact of structural change is probably stronger for macro relations, since the parameters
of such relations depend implicitly on structural aspects of the economy. Behavioral
models on the micro level on the other hand should show more stability since the indivi-
dual preferences governing individual behavior in general show a strong latency. Thus, one
may expect that micro relations describing individual behavior will be more stable than
macro relations in such a situation. If this is true, microeconometric analyses can give use-
ful information on the economic process even in a period of substantial structural change.
5 Bibliography
Autorengemeinschaft Panel 1990: Das Sozio-okonomische Panel nach fUnf Wellen, in: Vierteljahrshefte zur
Wirtschaftsforschung, Heft 1
Blossfeld, H.-P.; Hamerie, A. 1988: Using Cox Models to Study Multiepisode Processes, in: Sociological
Methods and Research, 17, pp. 432-448
Blossfeld, H.-P.; Hamerie, A. 1990: Unobserved Heterogeneity in Hazard Rate Models: A Test and an Illustra-
tion from a Study of Career Mobility, in: Mayer, K.U.; N.B. Tuma (eds.), Event History Analysis in
Life Course Research, Madison, pp. 241-252
Blossfeld, H.-P.; Hannan, M.; Schomann, K 1988: Erwerbsveriauf und die Entwicklung der Arbeitseinkommen
bei Miinnern, in: Zeitschrift fill Soziologie, 17, pp. 407-423
Carroll, G. R.; Mayer, K-U. 1984: Organizational Effects in the Wage Attainment Process, in: The Social
Science Journal, 21, pp. 5-22
Carroll, G. R.; Mayer, K-U. 1986: Job-Shift Patterns in the Federal Republic of Germany: the Effects of Social
Class, Industrial Sector, and Organizational Size, in: American Sociological Review, 51, pp. 323-341
Galler, H. P. 1981: On the Evaluation of Social Security Programs: The Case of the German System of Old Age
Pensions in: lSI-Bulletin, 49, pp. 1021-1051
Galler, H. P. 1986: The microsirnulation model of the Sfb 3 for the analysis of economic and social policies, in:
G.Orcutt et al. (eds.), Microanalytic models to support social and financial policy, Amsterdam,
pp.227-247
567
Galler, H. P. 1986a: Ubergangsratenmodelle bei intervalldatierten Ereignissen, in: Statistische Hefte, 27, pp.1-
22
Galler, H. P. 1987: Microsimulation of Household Formation and Dissolution, in: Keilman, N. u.a. (eds.),
Modelling Household Formation and Dissolution, Oxford, pp. 226-258
Galler, H. P. 1990: Simulation of Tax Transfer-Systems Using Microsimulation Models, in: Brunner, J.K; H.-P.
Petersen (eds.), Simulation Models in Tax and Transfer Policy, Frankfurt, pp. 279-299
Galler, H. P. 1991: Mikrosimulationsmodelle in der Forschungsstrategie des Sib 3, in: Hauser, R. et al. (eds.):
Mikroanalytische Grundlagen der Gesellschaftspolitik - Fortschritte in der Wirtschafts- und Sozial-
berichterstattung, Weinheim
Galler, H. P.; Ott, N. 1990: Zur Bedeutung familienpolitischer MaJ3nahrnen fur die Familienbildung - Eine ver-
handlungstheorctische Analyse familiarer Entscheidungen, in: Felderer, B. (Hrsg.), Bevolkerung und
Wirtschaft, Schriften des Vereins fur Socialpolitik, Berlin, pp. 111-134
Galler, H. P.; Potter, U. 1990: Unobserved Heterogeneity in Models of Unemployment Duration, in: N.Tuma
and K-U. Mayer (eds.): Event History Analysis in Life Course Research, pp. 226-240
Galler, H. P.; Wagner, G. 1983: Arbeitszeitverkiirzung und Arbeitsangebot, in: Wirtschaftsdienst, 1983/VII, pp.
329-336
Galler, H. P.; Wagner, G. 1986: The Microsimulation Model of the Sib 3 for the Analysis of Economic and
Social Policies, in: G.Orcutt et al. (eds.): Microanalytic Models to Support Social and Financial Policy,
Amsterdam, pp. 227-247
Hauser, R.; Hochmuth, U. 1989: Grundlagenforschung zur Unterstutzung einer aktiven GeseUschaftspolitik, in:
Th. Ellwein et a!. (eds.): Jahrbuch zur Staats- und Verwaltungswissenschaft, 3, Baden Baden, pp. 365-
387
Heckman, J.; Singer, B. 1982: The identification problem in econometric models for duration data, in:
W.Hildenbrand (ed.): Advances in Econometrics, Cambridge, pp. 39-67
Hujer, R.; Hansen, H.-J.; Klein, E. 1988: Makrookonomische Wirkungen der Steuerreform 1990. Eine okono-
metrische Analyse mit dem Frankfurter Modell, in: Finanzarchiv NF Bd.46, pp. 38-55
Hujer, R.; Hansen, H.-J.; Klein, E. 1990: The Contribution of Macroeconometric Models to the Evaluation of
Tax Policies - Problems of Application Exemplified by the Frankfurt Model and the West German Tax
Reform of 1990, in: Brunner, 1.K; Petersen, H.-G. (eds.), Prospects and Limits of Simulation Models
in Tax and Transfer Policy, New York
Hujer, R.; Lowenbein, 0.1987: Latente-Variablen-Modelle zur Analyse von Paneldaten, in: Krupp, H.-J.;
Hanefeld, U. (eds.), Lebenslagen im Wandel, Frankfurt/New York
Hujer, R.; Lowenbein, 0.; Schneider, H. 1990: Wages and Unemployment: A Microeconometric Analysis for
the FRG, in: Konig, H. (ed.), Economics of Wage Determination, Berlin/Heidelberg/New York, pp.
299-323
Hujer, R.; Schnabel, R. 1990: Analyse von BeschMtigung und Erwerbslosigkeit anhand des Soziookonomischen
Panels, in: Allgemeines Statistisches Archiv, 74, pp. 98-116
Hujer, R.; Schneider H. 1989: The Analysis of Labour Market Mobility Using Panel Data, in: European
Economic Review, 33, pp. 530-536
568
Hujer, R.; Schneider, H. 1990: Unemployment Duration as a Function of Individual Characteristics and
Economic Trends, in: Tuma, N.B.; Mayer, K.U. (eds.), Event History Analysis in Life Course
Research, Madison, pp. 113-126
Hujer, R.; Schneider, H.; Zapf, W. (eds.) 1991: Herausforderungen an den Wohlfahrtsstaat im strukturellen
Wandel, Frankfurt
Mayer, K.-U.; Huinink, J. 1990: Age, Period and Cohort in the Study of the Life Course: A Comparison of
Classical A-P-C-Analysis with Event History Analysis or Farewell to Lexis?, in: Magnusson, D.;
Bergman, L.R. (eds.), Data Quality in Longitudinal Research, Cambridge u.a., pp. 211-232
Mayer, K.-U.; Papastefanou, G. 1983: Arbeitseinkommen im Lebensverlauf, in: Schmahl, W. (ed.), Lebensein-
kommensanalyse, Tiibingen, pp. 101-125
Merz, J. 1980: Die Ausgaben privater Haushalte - Ein mikrookonomisches Modell fur die Bundesrepublik
Deutschland, Frankfurt/New York
Merz, J. 1983: FELES: The Functionalized Extended Linear Expenditure System - Theory, Estimation Proce-
dures and Application to Individual Household Consumption Expenditures Involving Socioeconomic
and Sociodemographic Characteristics, in: European Economic Review, 23, pp. 359-394
Merz, J. 1989: Zum Erwerbsverhalten von Frauen im Haushaltszusammenhang - Eine Mikroanalyse von
Lohnsatz- und Arbeitszeitdeterminanten mit Daten des Sozio-okonomischen Panels, in: Emmerich, K.;
Hardes, K.-H.; Sadowski, D.; Spitznagel, E. (eds.), Einzel- und gesamtwirtschaftliche Aspekte des
Lohnes, Niirnberg, pp. 133-154
Merz, J. 1991: Zur Dynamik markt- und nichtmarktmaBiger Aktivitaten: Zeitverwendung verheirateter Frauen
in Beruf und Haushalt - eine Panelanalyse, in: Hujer, R.; Schneider, H.; Zapf, W. (eds.), Herausforde-
rungen an den Wohlfahrtsstaat im strukturellen Wandel, Frankfurt (in print)
Orcutt, G. 1986: Views on Microanalytic Simulation Modelling, in: Orcutt, G. et aI. (eds.): Microanalytic
Models to Support Social and Financial Policy, Amsterdam, pp. 9-26
Ott, N. 1986: Ausscheiden erwachsener Kinder aus dem elterlichen Haushalt, in: Zimmermann, K.F. (ed.),
Demographische Probleme der Haushaltsokonomie, Bochum
Ott, N. 1991: Verlaufsanalysen zum Ehescheidungsrisiko, in: Hujer, R.; Schneider, H.; Zapf, W. (eds.), Her-
ausforderungen an den Wohlfahrtsstaat im strukturellen Wandel, Frankfurt
Schneider, H. 1990: Determinanten der Arbeitslosigkeitsdauer - Eine mikrookonometrische Analyse fUr die
Bundesrepublik Deutschland auf der Grundlage von Langsschnittdaten des Sozio-okonomischen
Panels, Frankfurt/New York
Tuma, N. B.; Huinink, J. 1990: Postwar Fertility Patterns in the Federal Republic of Germany, in: Mayer, K.U.;
Tuma, N.B. (eds.), Event History Analysis in Life Course Research, Madison, pp. 146-169
van Essen, U.; Kaiser, H.; Spahn, P. 1988: Verteilungswirkungen der Einkommenssteuerreform 1986-1990,
Sfb 3 Working Paper No. 262, Frankfurt
van Essen, U.; Kaiser, H.; Spahn, P. 1988a: Verteilungswirkungen der Einkommenssteuerreform 1986-1990 -
Eine Simulationsstudie fUr die Bundesrepublik Deutschland auf der Basis des Sozio-okonomischen
Panels, in: Finanzarchiv 46, pp. 56-84
Wagner, G. 1985: Ein Simulationsmodell zur Erzeugung von Lebenslaufen - Die Langsschnittversion 82.w des
Sfb 3-Mikrosimulationssystems, in: Zeitschrift fUr Bevolkerungswissenschaft, 4, pp. 593-623
569
Wagner, M. 1990: Education and Migration, in: Mayer, K.U.; N.B. Tuma (eds.), Event History Analysis in Life
Course Research, Madison, pp. 129-145
INCOME AND PRICE POLICY MAKING WITH AN ECONOMETRIC MODEL OF
FINANCIAL INCOMES AND EXPENDITURES OF POLAND'S POPULATION
Wieslaw Debski
University of Lodz
Institute of Econometrics
and Statistics
41 Rewolucji 1905r.
PL-902l4 Lodz
Abstract
Contents
1 Introduction
2 Structure of the model
3 Income and price policy in the light of the simulation analysis
References
571
1 Introduction
Indicator at
Index of. d ise9vuilibrium
Labour q,ualificClhon !
produchvity or employee5 state l
[ Profit sUboidies
L J
I I
,- -- -. - - f - - - I
1
financial
accumulation
:-~ /r'.e.raqe U
L-1 wage in the ~
'production h IA veraqe wage Prfce deflator Price i nclex
'------~~I of N M P i t'l h I c.~~t or U1
5ector~ I I-----l in socialiud h t- ---J
OJ
I or material I f-.l 1,Iving
socia lized I ~ qoods
I economy economY Index
I
I
., Price index ot
nomonst,..lmer Price index ot 'Price i nctex ot
Averuge wage commodltle5 procurement ~
nonmateria 1 t--
in the or agricu Lturol services
nonproduction Capita 1 u5i n oods
sectors i c05r
I
•
I I I
L _________ J
Price il1c1e~ at Price detlCltor 01"
material qoods consumption of
buyinq on free nondurable goods
marKet
0.3 -
0.50
0.2
0.1 - I 0.25
O.O~~----~--~----~--~----~-J
1909 1'J'Xl 1991 19'J2 UJJ 19')<1 O.CJJ'-~----~--~----~--~----~--'
1909 1'J'Xl 1991 19')2 1993 1994
0.5
0.50 -
0.4 -
0.3 -
0.25 -
0.2 -
0.1
O.CJJL-~----~-- __----~--______~~
lWJ J9'.-Q 19')1 1m lW 19'.)01 0.0
1989 1m 19') 1 1992· 1993 1994
o. ~%r-------------------------------, 0.6 %
O.~
0.:Al 0.4
0.3
0.25 - 0.2
o. J -
1989 1m 1991 1992 1993 1994 O.U
1989 1m 1991 1992 1993 1994
579
1.0
0.50
0.25 0.5
1.0
1.0
0.5
0.5
U.UL~-~-~--~-~-_._--'
1909 1m 19')1 1992 1993 1994 O.uL~---~-~--_-~-I
1989 1m 1991 1992 1993 1994
1.5%;:------------------, 1.00%:':.--------------,
O.r,)
1.0
0.50
0.5
0.25
O.OL_-~-_- _ _ _~-~--'
1989 1990 1991 1992 1993 1994 U.WL------~-~-----i
1~ 1990 1991 1992 1993 1994
582
-0.5 -
-0.51--
-1.0 -
-1.0 -1.5 -
-2.0
-1.5 -
-2.5·
-J.OL..·_ _-_-~-_-_--~.J
-2.0 '-~-~-~--~-~-~-- 1989 1990 19'Jl 1992 1:J93 1994
1S@ 1990 1:J91 1992 199J 19')4
-0.:5 -
-0.5
-1.0
-1.0
-1.5 -
-1.5 -2.0 - -
-2.0 '--~-~----~-~--~ 1m 19'.AJ 1991 1992 199J 1994
1989 1990 1991 I9'JZ 1m 1994
-0.5 '0.5
-1.0 '1.0
1.5 Z.ll -
1.5 -
1.0
1.0
0.5 0.5 -
O.O'--~----~--~--~----~--~-
O.O'-~----~-~-~-_~-
1989 19s() J991 1992 1993 1994 1909 1990 1991 1992 1993 1994
1.5
-
1.0
1.0
0.5 -
0.5 f-
O.B ~ 1.0
--------------------
0.6
0.4 - 0.5 -
0.2
4 References
-0.5 -0.5
-1.0
-1.0
-1.5
-2.0 '--~-~-~-~-~--~-I
1969 1990 19:11 19:12 I9:lJ 1994 1989 1990 19:11 1992 1m 1994
-0.5
-0.5
-1.0
-1.0
-1.5
-1.5
-2.0
-0.5
-0.5
-1.0
-1.0
-1.5
- Ekonomiczne, forthcoming.
2. Debski W.(1990), Model pienieznych dochodow i wydatkow
ludnosci Polski - ekonometryczna analiza, symulacje (A Model of
Financial Incomes and Expenditures of Poland's population - An
Econometric Analysis, Simulations), Lodz.
3. Hajnovic F.(1981), Prijmovo spotrebny model v systeme
modelov (Model of Incomes and Expenditures in the Systems of
Models), Informacne Systemy, No.3.
4. Hajnovic F.(1983),Zobrazenie penaznych priJmov a vydatkov
obyvatelstva v systeme modelov (An Outline of the Financial
Incomes and Expenditures of Population in the System of Models),
Informacne Systemy, No.6.
5. Welfe A. (1988), Prognoza rynku dobr konsumpcyjnych do roku
1992 (pazdziernik 1987) (The Forecast of the Consumer Goods'
Market up till 1992) (October 1987), Prace Instytutu Ekonometrii
i Statystyki UL, Seria F, No. 61, Lodz.
6. Welfe A., Welfe W.(1986), Modele i prognozy rynku w warunkach
nierownowagi (Models and Forecasts of Market under
Diseqilibrium),Ekonomista, No.2.
7. Welfe W.(1989), The Outlook for the Polish Economy 1988 - 1993
and the Plan of Economic Consolidation, Prace Instytutu
Ekonometrii i Statystyki UL, No. 68, Lodz.
Keynesian or Classical Unemployment in West Germany?
- An Empirical Analysis with a Disequilibrium Model -*)
Hermann-Josef Hansen
Sonderforschungsbereich 3
Johann Wolfgang Goethe-Universitat
Senckenberganlage 31
D-6000 Frankfurt am Main, FRG
Since the mid-seventies the labour market of the Federal Republic of Germany is marked
by persistently high unemployment. Macroeconomicaliy, the most prominent candidates
competing for the explanation of unemployment are insufficient demand for domestic
goods (Keynesian unemployment) and too high wage costs (classical unemployment). The
significance of these two approaches can be tested with econometric disequilibrium
models.
A further developed Sneessens/Dn!ze (1986) model, which was originally conceived for
Belgium, serves as the basis for the empirical analysis. The essential assumption in this
model refers to the structure of production: While in the long run the possibility of
substitution between the factors of production, labour and capital, exists. the short-term
production function is limitational. If we designate the ratio of output to labour input as A
and the ratio of output to capital input as B, then A· LS and B· K are the upper production
limits, with LS for labour supply and K for disposable total capital. If producers consider
possible rationings on output markets, employment L results - considering the
heterogeneity of the economy - from the equation
(1)
with
LP = potential employment (LP = A-I . B . K)
LK= Keynesian labour demand (LK = A-I . yd)
LS = available labour supply.
The model is thus composed of submodels for production, the goods market and the labour
market.
*) This paper was prepared within the project 'Macroeconomic Simulation' of the SonderfLlrschungsbereich
3 'Microanalytical Foundations of Social Policy'. I would like to thank the German Science Foundation
(DFG) for financial support.
589
In the production sector submodel the technical coefficients A and B as well as total capital
are explained. Given the assumption that producers are risk neutral, producers select the
technical coefficients and the capital stock so that their expected profits are maximized.
Possible output-market restrictions are considered. The resulting optimum technical
coefficients are a function of relative prices and technical progress. The optimum capital
stock is a function of relative prices, expected goods demand and expected profits at full
capacity [Lambert/Mulkay (1987)]. The process of adjustment of the present production
structure to the optimum structure is represented by an error-correction model.
The model has been estimated for the Federal Republic of Germany for the period 1961-
1988 using annual data. For estimating the parameters we used a sequential block-by-block
limited information procedure. In the first block the planned goods demand yd is
determined. It is needed for the determination of the capital stock as an explanatory
variable, which is jointly estimated with the technical coefficients in the second block.
Finally, the employment equation (1) is estimated by inserting for yd, A and B the
estimates from the first two blocks.
Figure 1 presents the components of employment over time. Figure 2 shows the share of
firms in each regime. The shares can be calculated from the elasticities of employment with
respect to LP, LK, and LS [For a detailed description of the model and the estimation
results see Hansen (1991)].
The results show that the two recessions of 1974/75 and 1981/1982, which were mainly set
off by supply shocks on the world markets, carried the West German economy each time
into a state of Keynesian unemployment. During recent years, the share of firms in the
regime of classical unemployment has increased. This development is to be explained by
insufficient investment activities in the first half of the eighties.
25.0
- - employment
24-.5 potential employment
24.0 Keynesian labour demand
23.5 _.- labour supply
23.0
22.5
22.0
21.5
21.0
1.0
0.9
0.8
0.7
0.6
0.5 Keynesian
0.4 unemployment
0.3 repressed
0.2 inflation
0.1
0.0
61 65 70 75 80 85 88
The results are comparable to those of other empirical analyses concerned with this
complex of questions. One advantage of the approach used here is that it can be integrated
into a larger macroeconometric model. This makes it possible to study alternative policy
scenarios and make a contribution to well-founded policy counseling.
References
Hansen, H.-J., 1991, Keynesianische oder Klassische Arbeitslosigkeit in der Bundesrepublik
Deutschland - Empirische Analysen mit einem okonometrischen Ungleich-
gewichtsmodell (Schulz-Kirchner, Idstein).
Sneessens, H.R. and J.H. Dreze, 1986, A Discussion of Belgian Unemployment, Combining
Traditional Concepts and Disequilibrium Econometrics, Economica 53, S89-
S 119.
Jerzy RomaCIski
University of Toru~
Chair of Econometrics and Statistics
Armii Czerwonej 25, Toru~, Poland
and
Wlaclyslaw Welfe
Uni ver'si ty of l:..6dz
Institute of Econometrics and Statistics
Rewolucji 1905 r, 41, 90-214 t.6dz, Poland
Abstract
The paper presents the structure of a new mini-model W-6 of the
Polish economy. It combines the long-term relationships with the
short-term adaptation mechanisms. It also assumes that the growth
of the economy was in the past unbalanced. Hence, i t defines
demand, supply and excess demand for commodities and labour, being
(partly or fully) unobservables. To deal with this problem excess
demand indicators were developed and used in the estimation
process after suitable respecification of the initial equations.
The paper discusses the results obtained for the simulation
version of the model which is composed of initially specified
equations, including accounting identities.
Contents
1 Intr-oduction
2 The structure of the W-6 model
3 Specification of the equations of the W-6-85 simulation mini-
model
4 The changes in the structure of the Polish economy 1961-1985 and
the adjustments of the model structure
References
Appendix; List of W-6 minimodel variables
1 Introduction
Total til 46
1. Employment 15 6
Ill'..n:!sf.ifler'!':' uutlr3.y'=' ):J
3. Fi ;·~ed assets 20 6
4.
3.
Production
Consump t i on and i rlv',·ntc.... i. t:!S
18
9 .,.
6
6. Foreign trade 20 8
7. Personal income, wages ~
.,,1
8. Personal savings 1
9.. Pr- i C}:!s 6
Estimation
Per-i od: 1961 - 1985
Mel. hod",; OLS <)_nd TSI_S
Hildreth-Lu for serial correlation
Dt for Et I)
0 i-F Et (I
:J t { 1 if E ..' i)
t
594
Dt = d( * ) (3 )
and use the identity (2) to obtain the equation explaining the ob-
servable realization Q from the demand side:
for (5)
NXD=13.332+0.00324*X(-1)-0.0429*X(-1)/NX(-l'- O.6334*U6164+
-0. 3391*U7679
NOD=O.1662+0.000103*KO+40.554*GD/KO+O.2375*U6769
JXD=-120.93+0.5961*JX(-1)+O.tt33*XD+O.0294*XD*U7375+
-O.0348*XD*U8082
II.2. JOD - investment outlays i.l the NMS sector, demand
Inves~ment outlays in the NMS sector are composed of residential
and communal structures. They depend on the e~pected increase of
urban population DLe and social consumption G. To allow for the
lags in response the lagged investment outlays were introduced.
JOD=-25.173+0.7892*JO(-1)+74.523*DLC+O.3563*GD+74.002*U74+
-95.037*U80
JAD=37.796+0.8643*JD-O.1327*JD*U8285-114.87*U76
rVX~-82.17S+0.9U91*(JX79D*U6179+JX80A*U8085)+
III",2 .. IKO - in\lP5r::mE~:li_::. I!U.t intu upt_"'f"6.tion if"' titEI Nt-,S st=\r:.L(lr~
TK"'-II<X+H::O
c~eJ ec ted 10llJ oJ i·., tr' .ll:JUt L '.J!l f ,~,,,. the per· i Del 1980-1985:
JR80S=0.096*JR+O.774*JR(-1)+O.13*JR(-2)
KOX~-3.4792+0.003;9*~X~(-1)+O.1~95*IKX(-1)+121.36*U82
IIL8. f;XK - fi':ed assets in the MF' ;:,ector' (end c:f ye.~r)
600
CXD=44.717+0.8595*(Y+OlEP(-1)/PY)+
-3.6781*[oO%/(PY-PY(-1»*100)]*U80S5
MD=t1ND+MRD
SNEMF'=EN*FEN-MN*F'~1N
ZUXP=-38.155+0.0335*(X-XR)/(NX-NR)+230.47*PY-66.256*PY*U8285
Y=YP/py
where ~O = 0.48077 .
PX=c).001214+1).6039*(YF'+AFOPl/X+!.1059*O.5*(AFP+AFP(-I) l/X(-l)+
+O.04927*PM+O.1767*U82
PM=(MN*PMN+MR*PMRJ/M
Ref<:>ro;>nc85
Ullrich Heilemann1
Rheinisch-Westfalisches Institut fur Wirtschaftsforschung
Hohenzollernstr. 1/3, D-4300 Essen 1, FRG
Abstract
The paper analyzes the causes of the large errors in the 1988 FRG macroeconomic
forecasts. The study is based on a forecast and various simulations with the RWI-business
cycle model, a medium-sized macroeconomic short-term model. It is shown that the most
important causes were errors in the assumptions of government construction outlays and -
more important - of world trade. As to endogenous errors, the forecasts of changes in stocks
and the distributed profits proved to be most damaging. By extending the error analyses on
such endogenous variates, the paper explores possibilities of model analysis only rarely
employed up to now. However, it should be clear that only such an extensive analysis gives a
true picture of a model's strengths and weaknesses.
Keywords: macroeconomic models; forecasting with macroeconomic models; forecast
analysis
Contents
1 Introduction
2 The RWI-business cycle model - an overview
3 The forecast for 1988
4 The accuracy of the forecast: exogenous and endogenous errors
5 Summary and conclusions
References
1 Introduction
The year 1988 might be remembered in the FRG's economic history for two reasons: its
comparatively strong economic growth and the fact that this growth was not foreseen by any
of the major forecasters. Both aspects provide a reason to take another look at the 1988
forecast of the RWI business-cycle model, focusing on three questions. First, do the
forecast errors lie in or outside the range expected from previous experience? Second, are
I am indebted to David Belsley, Josef Gruber and Kenneth Wallis for helpful comments.
611
the errors attributable to certain aggregates? And third, since scientific forecasts rely on
both assumptions and hypotheses (Hempel-Oppenheim forecast deduction scheme) what
are their contributions to the forecast errors?
Of course, it might well be argued that the last question is primarily of interest to the
producer and not to the consumer of a forecast - indeed, most evaluations do not distin-
guish between assumptions and hypotheses. But a failure to do so here, in the context of an
econometric model, means to give up an opportunity to learn and to ignore some of the
models' truly unique features.
In addition, one might also ask for an evaluation of the forecast errors in utility terms - we
are, after all, economists! But, whereas the specification of a forecast utility function may
seem possible for the "truth-seeking" scientist, typically, we do not have the information to
determine utility functions of other forecast consumers (Menges 1975, p. 5f.). Faute de
mieux, one might speculate on how the economic policy of the federal government might
have differed, if there had been no forecast errors. But, since in 1988 the government
tended to be much more optimistic than most forecasters, it is unlikely that economic policy
would have differed much. Things might have looked somewhat different for the policy of
the Deutsche Bundesbank (Deutsche Bundesbank 1989).
The paper tries to answer the three questions in the context of the RWI business-cycle
model. Analyzing the predictive quality of econometric models has a long tradition (Evans,
Haitovsky, Treyz, Su 1972), but it has only been since the beginning of the 80s that this is
made on a regular and systematic base (Wallis (ed.) 1986; McNees 1989, p. 43). As to the
R WI business-cycle model, such analyses form a central part of every forecast production
and have been the subject of various studies before (Heile mann 1985; 1989). The current
study differs methodologically from previous studies primarily by examining not only the
influence of errors in exogenous variables (assumptions) but also the role of "endogenous"
errors.
This paper is organized as follows. The next section gives a synopsis of the model and an
overview of its ex post forecasting performance. Section 3 presents the 1988 forecasts, both
from the R WI-model and from other sources. Section 4 analyzes the forecast accuracy, and
section S is devoted to a summary and conclusions.
The RWI business-cycle model is a medium-sized, short-term (quarterly) model which has
been used for forecasting and simulation for more than 10 years. It consists of roughly 40
stochastic equations and 75 identities, which together form an interdependent, weakly
612
Exhibit 1
MAEI RMSPE 2
Prices
Consumer prices, % change 0.5 71.7 0.8
GNP prices, % change
Government budget
Deficit 2.9 182.9 0.8
nonlinear model (Heile mann, Mi.inch 1984a). It can be segmented into five sectors:
demand (30 equations), distribution (21 equations), production (16 equations), price (12
equations), and government (36 equations). The exogenous variables consist of "politically-
determined" variates, such as the social-security contribution rate, government construction
outlays, short-term and long-term interest rates, on the one hand, and internationally
determined ones, such as world trade and import prices, on the other. The theoretical
613
Exhibit 2
quarters
1 2 3 4 1 2 3 4
Assumptions
Public investm. 9.2 2.2 1.0 -2.2 1.4 3.5 3.9 10.4 2.6 -2.1 -1.8
Social sec. contrib. 17.7 17.7 17.9 17.6 17.6 17.8 17.9 17.9 17.9 17.9 17.9
World exports 4.2 3.2 3.4 2.7 1.8 3.4 4.7 4.5 3.7 2.6 2.7
Price index of imports -10.7 -4.4 1.2 -11.0 -4.8 -0.4 -0.8 0.8 0.8 1.2 2.1
Interest rate, short term 4.6 4.0 4.0 4.1 3.8 3.9 4.2 3.8 3.9 4.0 4.2
Interest rate, long term 6.1 5.9 6.1 5.9 5.6 6.1 6.1 5.7 6.0 6.2 6.5
Negotiated wages 4.2 3.8 3.7 4.0 3.7 3.8 3.8 3.4 3.8 3.8 3.8
Forecast
GDP, origin, volume
GDP, volume 2.6 1.4 1.8 2.2 0.8 1.4 1.1 2.6 2.0 1.3 1.4
Employed (x 1000) 1.0 0.6 0.8 0.9 0.7 0.5 0.5 0.7 0.6 0.8 0.9
Prices
Consumer prices, % change -0.5 0.6 1.5 -0.5 0.8 0.8 1.1 1.8 1.3 1.4 1.6
GNP prices, % change 3.1 2.2 2.0 3.1 2.6 1.4 1.8 2.2 2.0 2.2 1.7
Govemment budget
Deficit -23.5 -38.4 -43.9 -16.2 -2.6 -11.1 -8.6 -19.7 -4.6 -11.0 -8.5
Source: Rheinisch-Westfalischcs Institut fUr Wirtschaftsforschung (Hrsg.), Vierteljahrliche Prognose mit dem
R WI-KonjunkturmodellI987-3 bis 1988-4, Nr.26 v. 20.12.1987 bearbeitet von T. Weil3. Essen 1987,3.
614
Exhibit 3
Model-adjustments
1988 - forecast of the RWI-model
BiII_ DM
quarters
3 4 1 2 3 4
foundations of the equations are, as for most applied econometric models, somewhat
eclectic, including neoclassical, Keynesian, and monetarist elements. The architecture of
the model, however, is Keynesian, in the Keynes-Klein tradition. With respect to the roles
played by demand and money and the stability of the private sectors, the model may rather
generally be labelled as post-Keynesian (Davidson 1980; Chick 1983). Expectations are, in
most cases, adaptive or "weakly rational" (Eckstein 1983, p. 41).
The model is re-estimated twice a year; the sample period is the same for all equations and
covers the last 40 quarters of the data available ("moving window"). For the current study it
is 1977-3 to 1987-2. The parameters are estimated by ordinary least squares (OLS). Exhib-
it 1 presents the values for various forecast-error summaries of a full dynamic simulation of
the model over the sample period. In general, they do not differ from previous results
(Heilemann, Munch 1984a, p. 360f.; Heilemann 1989, p. 256f.), except that the tracking of
private consumption, and, as a consequence, of GNP, has deteriorated somewhat.
615
The forecast for 1988 (and the last two quarters of 1987) was made and published in De-
cember 1987 (Rheinisch-Westfalisches Institut fur Wirtschaftsforschung (Hrsg.) 1987). The
main assumptions and results are given in Exhibit 2. Additional assumptions, introduced by
add-factoring, are shown in Exhibit 3. While most of these seem quite clear, the "Other
adjustments" items deserve some comment.
The reduction in ''wage and salary income" has been made to account for the reductions in
the hours of the working week which became effective in 1988 (reflecting the fact that the
negotiated wage rate used here is a kind of gross rate, representing only the cost effects).
(The model has since been expanded to cope with such problems in a more straightforward
manner (WeiJ3 1989).) The reduction in "distributed profits" is a complementary change
reflecting the fact that this variate is supposed to follow not only gross profits but also wage
and salary income. The reduction in the "price index of government consumption" is meant
to represent the more restrictive wage policy in the public sector. The other changes reflect
the effects of the revaluation of the OM, which is not captured explicitly in the model.
Finally, because the increase in many of the negotiated wages had been fixed in the 1987
wage round, negotiated wages have been exogenized for the present forecast.
It should be noted that the main assumptions are very similar to those of the forecast made
by the Joint Diagnosis in October 1987 (Die Lage der Weltwirtschaft und der
westdeutschen Wirtschaft im Herbst 1987 (1987) (hence: Joint Diagnosis) with their model
so that both the RWI-model's and the Joint Diagnosis' forecast seem in accord with their
major assumptions 2.
Under these assumptions, the model predicts a (real) GNP growth of 1.8%, a O.4%-point
more than 1987. Prices at the consumer level increase 1.5%, while those for GNP remain
near 2.0% - results of the expected end of a phase of OM revaluation and of a reversal of
the price developments on world markets. Given a productivity growth of roughly 1.1 %, the
GNP growth leads to a growth in employment of about 0.8%, or of about 195,000 workers,
an increase of roughly 30,000 workers more than in 1987. The government deficit rises to
43.9 bill. OM, 5.3 bill. OM more than in 1987, mainly a consequence of the 1988 tax
reduction of roughly 12 bill. OM (gross, cf. Exhibit 2), although, in part because of the "self-
financing" nature of this measure, it is likely to accomplish only about 70% of the original
"expenditure."
2 World trade in the RWI model forecast was assumed to be about 0.5 %-point, price index of imports about
1.3 %-point lower than previously assumed by the Joint Diagnosis. The reductions of the interest rate were
about 1.0 %-point (short-term) and 0.6 %-point (long term).
616
Exhibit 4
Selection of 1988-forecasts
% change from previous year
Prices
Consumer prices, % change 1.3 1.5 1.3 2.0 1.5 1.3
GNP prices, % change 1.5 2.0 1.9 2.0 1.5 1.5
Government budget
Deficit, bill. OM -42.2 -43.9 -38.0 -45.0 -49.5
In short: the model led us to expect a continuation of the recovery, although it was mainly
the effect of the tax reduction that produced the increase in the growth rate. The rising
inflation rate mainly reflected a turn in world inflation and in the exchange-rate
developments of the OM. It should be noted that no other consequences of the crash of
October 19, 1987, were expected than those incorporated in the (reduced) assumptions of
world-trade. The RWI model forecast was numerically well in line with that of the Joint
Diagnosis (1987, p. 18f), which shared much of its assumptions, and of the Council of
Economic Experts (Sachverstandigenrat zur Begutachtung der gesarntwirtschaftlichen Ent-
wicklung 1987, p. 120), at least within the published figures (rounded), as seen from
Exhibit 4.
Before examining the forecast numbers, some methodological clarification seems advisable.
Econometric models like the RWI-businesscyc1e model may be written formally in period t
as:
(1) 0,
with
y vector of endogenous variates,
x vector of exogenous variates,
B matrix of parameters,
e vector of disturbance terms, and
order of lags, 1, ... , n.
For various reasons, (2) might be modified (elements in B might be altered) or com-
3 The presence of Yt +' on both sides of the equality here reflects the fact that a closed-form
solution for Yt +' may Jot exist in the general, ronlinear context. The form of (2) is intended to
convey the idea hat the Yt +' are the implicit solutions of this set of equations conditional upon
the values of the other vari!tes. The equations (2) can also be normalized with respect to the
individual endogenous variate so that the current value of the endogenous variate that occurs on
the left-hand-side does not also occur on the right-hand-side. It is assumed that a unique solution
for (2) exists within some appropriate region of the space of endogenous variates.
618
where
xa vector of additions to the constant terms.
The reasons for such additions may be of either an a priori economic or a statistical nature -
or they may just stand for a feeling that the outcome of an equation appears implausible. In
the RWI model, corrections are usually not made for statistical reasons (see, for example,
WeiB 1987 for some experience with Cochrane-Orcutt corrections). Thus, the model's
forecast may be presented as:
where
xe :vector of additions to the constant terms made for a priori economic reasons,
xP :vector of additions to the constant terms made for reasons of plausibility.
This short representation illustrates not only the way some econometric model forecasts are
made, but also the way their predictive accuracy should be ascertained. The simplest way, of
course, is just to compare forecast and actual values:
where
et + j Ix : vector of forecasting errors.
Further, since the forecasts developed here are conditional forecasts, the error e can be
cleansed of the effects of any errors that are subsequently observed to occur in all or part of
the assumptions by repeating the forecast with meanwhile available observations of x:
where
e t +j Ix vector of forecasting errors, corrected for observed errors in the assump-
tions.
The process lying behind (6) can be used to examine the effects of various types of
corrections. For example, the forecasts may be repeated after exogenizing some of the
endogenous variates, or after different assumptions have been made with respect to some
619
Exhibit 5
1987 1988
quarters
1987 1988 3 4 1 2 3 4
Author's computations.
The present analysis of forecast accuracy will follow the lines described in the introduction.
It will concentrate on the forecasts for 1988, ignoring the small errors in the base-year
(1987). The same holds for the effects of the mean while National Accounts-revisions which
the Deutsche Bundesbank (1989, p. 34) has calculated to be about 0.3% of GNP for their
model forecast.
We begin with the errors that have occurred in the main assumptions on the exogenous
variates X. These are seen in Exhibit 5 to be rather small, falling, for the most part, well
within the bounds experienced by the RWI model to date (Heilemann 1989, p. 269). Two
exceptions, however, are noteworthy: public investment and world trade. The error
relevant to public investment is twice its usual size and is due mainly to the very mild winter
(first quarter 1988) whose effects were not compensated during the rest of the year because
of the favorable finance position of the government sector. Of a more serious nature are
the errors that materialized in the assumptions relevant to world-trade development. The
usual 6-quarter forecast error in world exports (made outside the RWI-model) is 0.4 %-
points, but the figures for 1988 are approximately eight times as large. A detailed analysis
of the causes of this error is beyond the scope and possibilities of the present study. It seems
clear, however, that it was a consequence of an underestimation of the general trend of
620
Exhibit 6
Errors of 1988-forecasts
% changes from previous year
Actual Errors
Joint Council
RWI-model1 Diagnosis of
Economic
Experts
II III IV V VI
Prices
Consumer prices, % change 1.3 0.2 0 0 0 -0.5 0.2 0.7 0.2
GNP prices, % change 1.5 0.5 0.4 0.3 0.4 -0.3 0 0.5 0
Government budget
Deficit, bill. DM -42.2 1.7 4.2 -1.3 0.5 16.0 4.0 -2.8 -7.3
Sources: s. Exhibit 4 and author's computations.- 1) I: Original forecast; II: forecast with actual values of
exogenous variables; III: forecast II with employment exogenized; IV: II with "Distributed profits" exogenized;
V: II without parts of "other adjustments" (Exhibit 3); VI: ex post forecast with a 1979-1 to 1988-4 based model
verSIOn.
actual events. Otherwise, it is interesting to note that the interest-rate assumptions did
surprisingly well. With minor exceptions in the first two quarters - the monetary authorities
were likely too busy to prevent a slump in consequence of October 19 -, the interest-rate
assumptions proved nearly correct.
All of the 1988 forecast errors, both those of the RWI model and of the other forecasts
mentioned above, are brought together in Exhibit 6. Neglecting for the moment the role of
the assumptions, it is clear that the RWI model (column RWI-model I) performs poorly.
Ignoring the interdependencies, this is especially true relative to private consumption and
the changes in stocks, whereas the RWI-model proves superior with respect to the dynamics
of investment activity4.
Now, while such a comparison may reveal things about the forecaster's general under-
standing of the economic process, it can say nothing about the validity of the hypotheses
employed and little about the usefulness of the model as an instrument for scientific
forecasting. To gain insight into these issues, the entire 1988 forecast was repeated using
the actual values for the exogenous variates. It is seen from the column labelled RWI-
model II in Exhibit 6, that, in this case, the situation is very much changed. The 1.6 %-point
underestimation for real GNP becomes a 0.4 %-point overestimation, and the figure for
investment growth is reduced to only 0.6 %-points. Unfortunately, the forecast error with
respect to private consumption more than doubles, and the forecast of the changes in stocks
is only slightly improved. In assessing the other figures, given the uncertainty that prevailed
in late 1988, it is less surprising that their errors tend to lie outside the usual limits than that
the differences are rather small (cf. Heilemann 1989, p. 256).
Referring to the model's link matrix (Gilli 1984; Heilemann 1989, p. 260), the error in
private consumption might be linked via wage and salary income to employment which is
central for private consumption. To illustrate the role of employment for the model's
forecast, a further simulation was conducted in which employment (the number of wage and
salary earners) was exogenized. The results are presented in the column labelled RWI-
model III of Exhibit 6.
The forecast errors of nearly all demand categories are reduced, but perhaps not by as
much as expected. This, again, is mainly due to offsetting effects: The smaller prediction
errors for private consumption are balanced by the increased errors for imports. While this
finding could give cause to think of a structural change in the consumption function - a
previous study has cast some doubts upon its inter-temporal stability (Heilemann, Munch
4 Though these kinds of comparisons and statements are often made, they are basically inadequate since they
ignore the links (at least by definition) between the various parts of the forecast.
622
1984b, p. 117f.) - a comparison of the present regression results with those of the 1979-1 to
1988-4 sample showed rather similar explanatory power and reaction-coefficients. A closer
examination of the forecasting accuracy of the explanatory variables proved more
successful. In fact, an analysis of the various income components revealed that the model
much overestimated "distributed profits". A comparison of the present regression results
and of the 1979-1 to 1988-4 sample reveals that the large overestimation in the forecast
period is mainly explained by the reduction of the gross profits-elasticity from 0.44 to 0.34
(the wage-elasticity changed from 1.11 to 1.18). Though the introduction of the Add-factors
corrected the variate in the right direction, the correction was too small. Whether the
unexpected slight increase in the savings rate that occurred during this period is due to the
"fallout" of October 19 or to specific attitudes held by consumers towards the "windfall
income" of the tax reduction ("tax fall"), or to a mixture of both, is difficult to judge.
The RWI-model forecasts are those of an econometric model and not those of eco-
nometricians (Evans, Haitovsky, Treyz, Su 1972, p. 952): Add-factors are set primarily to
account for special events (xe ) which the model is unable to capture explicitly. To illustrate
the role played by the "pure" add-factors (x P) in the 1988 forecasts, the forecast was redone
after eliminating the corrections for the three price indices (Exhibit 3). The effects of this
are seen in Exhibit 6 (RWI-model V). Comparing these results with those using the actual
exogenous variates (RWI-model II) reveals at least three characteristics of the model.
First, such corrections are primarily of local importance, as is illustrated by the effects on
consumer prices as well as those on private consumption and exports. Second, in both these
instances, the real responses are much smaller than the price-adds, especially with respect
to export prices. And third, the effects on nominal GNP and its distribution are more
substantial, although more difficult to detect because of various offsetting effects (cf. gross-
wage income, gross profits, and national income).
To complete the analysis, the current version of the model (Spring 1988, sample period
1979-1 to 1988-4) was used to forecast the period under study (i.e. an ex post forecast within
the sample period) was made. Although some of the equations of the model have since
been respecified (independently of the forecasting experience reported here), a comparison
with the previous results seems to be meaningful. In general, these forecast errors are
smaller (RWI-model VI) than those of the ex ante forecast with actual values for the
exogenous variates (RWI-model II). However, it is interesting to note some important
shifts in the hierarchy of the forecasting quality. While the errors for private consumption
and for the changes in stocks are much reduced, that for investment in machinery has
substantially increased. These are normalizations which can be attributed to the well-known
levelling-off effects of OLS.
623
1. Examining the 1988 forecasts of the RWI model in retrospect has shown that the
1.6 %-point underestimation of the FRG's economic growth has mainly been due to a large
underestimation of the 1988 expansion of world trade. While usually the errors in the va-
lues assumed for this variate have not tended to exceed greatly the 0.4 %-point margin, in
1988 it jumped to 3.3 %-points.
2. A model can, of course, hardly be blamed for having to operate with the wrong as-
sumptions, so it is interesting to note that once corrected for the preceding and all other
errors in the assumptions, the speed-up in the 1988 growth rate is well explained, and the
accuracy of the forecast does not differ from previous experience. This outcome is not
quite as obvious as one might expect, since models operating with wrong assumptions often
produce better forecasts than do those using correct ones (e.g. McNees 1989, p. 54) - a re-
sult caused by a tendency of some forecasters to orient their prediction at some "Star of
Bethlehem" (Waelbroeck, De Roo 1983, p. 400), technically achieved by extensive mas-
saging (exogenous variates as well as add-factors). At the same time, these results suggest
that the 1988 forecast errors do not stem from a shift in the short-term behaviour of the
FRG-economy. If some analysts see the growth of the FRG-economy in 1988 as a long
overdue result of a successful supply-side policy the present analysis hardly backs such a
view.
3. Private consumption and changes of stocks provide an exception to some of the previous
conclusions, as they show errors outside their usual margins. While the poor results for the
changes in stocks simply indicate that the explanation has to be revised, for private
consumption things look different. Judging from the empirical quality of the estimation of
the private consumption function used here (and in most other approaches), it is difficult to
blame this result on it. Rather, it is mainly due to the model's large errors in explaining
distributed profits. First of all it is the much-reduced reaction to gross profits which
accounts for most of the underestimation of this variate. The problems with the explanation
of the comparatively large withholdings of profits, however, do not seem to be restricted to
the RWI model. It also besets the forecasts by the Joint Diagnosis and by the Council of
Economic Experts - although both had additional difficulties in matching the dynamics of
private investment. Indeed, there is much evidence that this misperception has not been
restricted to the FRG. In 1988 the under-estimation of GNP growth has been of the same
magnitude (0.6 %-points) for the world as for the FRG, and there is some evidence that it is
mainly investment and exports that lie behind this drive, while private consumption lags
behind. This result once again raises the questions (1) whether expectations are an
important force for transmitting business cycles among countries (Waelbroeck, De Roo
624
1983, p. 400), and (2) if so, how models should take them into account.
References
V. Chick (1983), Macro Economics after Keynes. Cambridge, Mass.
P. Davidson (1980), Post Keynesian Economics. "Public Interest", Special Issue 'The Crisis
in Economic Theory'. New York, vol.20, pp.151-173
Deutsche Bundesbank (Hrsg.) (1989), Die Verwendung des okonometrischen Modells der
Deutschen Bundesbank zu gesamtwirtschaftlichen Vorausschatzungen. "Monatsberichte der
Deutschen Bundesbank", 4l.Jg., Nr.5, pp.29-36
Die Lage der Weltwirtschaft und der westdeutschen Wirtschaft im Herbst 1987 (1987). Be-
urteilung der Wirtschaftslage durch folgende Mitglieder der Arbeitsgemeinschaft deutscher
wirtschaftswissenschaftlicher Forschungsinstitute e.V., Essen: Deutsches Institut rur Wirt-
schaftsforschung, Berlin; HWWA-Institut rur Wirtschaftsforschung, Hamburg; Ifo-Institut,
Munchen; Institut fur Weltwirtschaft an der Universitat Kiel; Rheinisch-Westfalisches Insti-
tut fur Wirtschaftsforschung, Essen. 0.0.
O. Eckstein (1983), The DRI-Model of the U.S. Economy. New York
M.K. Evans, A. Haitovsky, G.!. Treyz, V. Su (1972), An Analysis of the Forecasting
Properties of U.S. Econometric Models. In: B.G. Hickmann (ed.), Econometric Models of
Cyclical Behavior. (Studies in Income and Wealth, 36) vol.2, New York u.a., pp.949-1139
M. Gilli, (1984), CAUSOR - A Program for the Analysis of Recursive and Interdependent
Causal Structures. (Cahiers du Departement d'Econometrie, 84.03) Universite de Geneve
U. Heilemann (1985), Zur Prognosegenauigkeit okonometrischer Modelle. "Zeitschrift fur
Wirtschafts- und Sozialwissenschaften", 105.Jg., pp.683-708
U. Heilemann (1989), "Was leisten Prognosemodelle?" Eine empirische Untersuchung am
Beispiel des RWI-Konjunkturmodells. In: B. Gahlen, B. Meyer, J. Schumann (Hrsg.), Wirt-
schaftswachstum, Strukturwandel und dynamischer Wettbewerb. Festschrift fur Ernst
Helmstadter zum 65. Geburtstag. Berlin, Heidelberg, New York, pp.253-272
U. Heilemann und H. J. Munch (1984a), Einige Bemerkungen zum RWI-Konjunktur-
model!. In: G. Langer, J. Martiensen und A. Quinke (Hrsg.), Simulationsrechnungen mit
dkonometrischen Makromodellen. Munchen, pp.355-385
U. Heilemann and H. J. Munch (1984b), The Great Recession: A Crisis in Parameters? In:
P. Thoft-Christensen (ed.), Proceedings of the 11th IFIP Conference on System Modeling
and Optimization, Copenhagen, Denmark, July 1983. (Lecture Notes in Control and
Information Sciences, 59.) Berlin, Heidelberg, New York, pp.117-228
S.K. McNees (1989), Why Do Forecasts Differ? "New England Economic Review",
January IFebruary, pp.42-54
G. Menges (1975), The Scientist and the Policy Maker as Forecaster. Paper presented at
the 3rd World Congress of the Econometric Society. Toronto
Rheinisch-Westfalisches Institut fur Wirtschaftsforschung (Hrsg.) (1987), Vierteljahrliche
Prognose mit dem RWI-Konjunkturmodell 1987-3 bis 1988-4. Nr. 26, Essen, 20.12.1987,
bearb. von T. WeiB
625
Abstract
This short review presents the main characteristics of the macroeconometric model of
the Deutsche Bundesbank. This medium-size quarterly model describing the West
German economy is regularly used in macroeconomic forecasting and policy evaluation.
It is divided into seven interdependent blocks dealing with (1) the demand side and
(2) the supply side of the economy, with (3) costs and prices, (4) the public sector,
(5) foreign trade and payments, and (6-7) the financing of the economy. The latter two
blocks were respecified recently, when a portfolio selection model was incorporated to
explain private demand for financial assets. This portfolio block is described in more
detail. Moreover, estimation results obtained by application of Zellner's restricted joint
generalized least squares (RJGLS) method are presented.
Keywords: Macroeconometric modelling, Portfolio theory.
Contents
sources are the national accounts of the Federal Statistical Office (Statistischcs Bun-
dcsamt), and the banking statistics of the Deutsche Bundesbank.
The coefficient estimates in the stochastic equations are regularly updated as soon as
the national accounts data of a new quarter are available. In most equations, the esti-
mation period starts with the first quarter of 197I; thus, current estimations, for the
most, draw on more than 70 observations. The estimation method is 0 LS in linear
equations, NLS in non-linear equations, and restricted least squares (RLS) in equations
involving Almon lag distributions. The newly specified portfolio selection model (see
section 3) is estimated by Zellner's restricted joint generalized least squares (RJG LS)
method ([ II], pp. 295-317).
The Bundesbank Model is used on a regular basis to set up short-term macroeconomic
forecasts for the Central Bank Council; for the most part, these forecasts cover the
current year and the following one. Moreover, the Bundesbank Model is applied to
analyse the impact of exogenous shocks on the German economy by means of
simulation. Of course, the consequences of monetary policy measures are of specific
in terest, but in the past there were also studies dealing with oil price and other external
shocks, with a reduction of working time, or with fiscal policy changes like the recent
tax reforms. Although these studies; like the short-term forecasts, serve mainly internal
purposes, the Deutsche Bundesbank occasionally reported on these simulation results
[6) [8J (9). Recently, there was also an article on the performance of internal forecasts
using the Bundesbank Model [I OJ.
As mentioned above, the new version of the Bundesbank Model consists of seven
interdependent blocks dealing with the following aspects of the West German economy:
(A) the demand side, (B) the supply side, (C) costs and prices, (D) the public sector,
(E) the foreign sector, (F) financial assets and liabilities, and (G) interest and exchange
rates.
The first five blocks (A) to (E) form the real part of the Bundesbank Model, whereas
the last two blocks (F) and (G) refer to the financial sphere. Since the model is
structured from the theoretical point of view, the market behaviour of the different
economic agents is subsumed under separate blocks. Thus, block A represents private
households, block B refers to private firms, and blocks D and E obviously represent the
public and the foreign sectors. Each of these four blocks results in the definition of the
respective sector's financial surplus or deflcit, which then enters the financial part of the
model. However, the sectoral disaggregation is somewhat different in the financial part;
there, one finds the division: central bank, commercial banks, private sector (households
and production firms), government, and foreign sector.
Economic activities of private households explained in the demand block (A) are private
consumption and labour supply. The consumption function is specified in real per
capita terms, using an error correction form. In the long run, private consumption is
determined by real disposable income, net financial wealth of private households, the
functional distribution of income, and the real rate of return of the private portfolio.
The specification of labour supply is based on a utility maximization approach
comparing the benefits of working and leisure time. Thus, the labour force participation
628
rate depends on real per capita wages, the comparative real non-wage income per
capita, and the age structure of the population. Moreover, the demand block contains
the definition of final demand and GNP, both in real and in nominal terms. Thereby,
in a technical sense, this block becomes the central part of the Bundesbank Model, since
there are direct links to all other blocks.
The core of the supply block (B) is the definition of potential output, which is derived
from the estimation of a CES production function relating final demand to total hours
worked. capital stock actually used, and imports. However, potential output is defined
only on the basis of standard working time of the total labour force and fully utilized
capital stock, not taking into account imported inputs. Apart from the determination
of aggregate supply, the second block contains the factor demand equations. According
to the assumption of profit maximizing firms, the demand for capital (i.e. net invest-
ment), labour, and imported inputs positively depends on final demand and the ratio
of product prices (after taxes) and the respective factor costs. The definitions of the rate
of capacity utilization, the unemployment rate, and the capital stock are also included
in the supply block, which is closely connected to the other real blocks, but has no direct
link to the financial part of the model.
In the third block (C), costs and prices are determined. Wages are explained by a
Phillips curve specification using the unemployment rate, consumer prices, and the
standard working time as explanatory variables. Together with the user costs of capital
(determined by real interest rates, depreciation rates, capital taxation and related
allowances) and import prices, they make up the costs of production index. To avoid
problems caused by multicollinearity of the cost components, this index is used to
explain prices. Also, capacity utilization and indirect tax rates enter the price equations.
Price deOators are defined by the ratio of the respective nominal and real aggregates.
Thus, the price block interacts closely with the other real blocks and also with the
financial part via interest rates.
The public sector block (D) deals with direct and indirect taxes, contributions to the
social security system, and transfer payments. The revenues of the various taxes are
determined by the respective bases of \'aluation and the corresponding tax rates.
Contributions of employees and employers to the social security system and tranfer
payments (unemployment benefits and pension payments) are specified in a similar
way. Thus, the variables explained in this block largely depend on those of the demand
block and the supply block. Conversely, there are several repercussions on these two
blocks. As noted above, taxes also affect costs and prices in manifold ways.
The foreign sector block (E) completes the real part of the Bundesbank Model. In this
comparatively small block, real exports and import prices are essentially explained by
external factors, whereas real imports (explained in block B) as well as export prices
(explained in block C) largely depend on domestic economic conditions. So, real exports
of goods and services are specified as a function of foreign demand expressed in terms
of production output and the real exchange rate based on deOators of final demand.
The explanatory variables of import prices are world prices of energy and of other raw
materials as well as foreign deOators of final demand, all these prices being converted
into O-marks. The definitions of this block serve to compute total exports and imports
both in real and in nominal terms, and the surplus or deficit in the current account. The
foreign sector block has mutual direct links to the other real blocks except the public
sector block, and there is of course an influence from the financial sphere through
exchange rates.
629
r
A: Demand Side B: Supply Side
(Priv. Househ.) (Private Firms)
• pro consumption • potent. output
• labour supply • factor demand
• GNP (= demand) • capacity util-
• fin. balance of • fin. balance of
priv. househ. +-- private firms
5 S 26 D 13 S 32 D
E: Foreign Sector
• exports
• import prices
• current account
7 S 19 D
REAL PART
FIN. PART
I
F: Financing I G: Financing II
(Portf. Modell
• fin. assets and • short and
liabilities long-term
• money stock interest rates
• fin. wealth • exchange rates
• fin. accounts
1& S 16 D 12 S 3 D
The financial part of the Bundesbank Model was respecified recently; it is now divided
into two blocks. The first block (F) contains the newly specified portfolio model
explaining private demand for several financial assets as described in the next section.
Moreover, it includes equations for financial assets of the public sector and bank
lending to the private and public sectors, definitions of monetary aggregates and
financial wealth of the sectors as well as accounting identities. In the second financial
block (G), interest and exchange rates are explained. The pivotal variable in interest
rate determination is the three-months money market rate, which hinges on the
Bundesbank rates, the equivalent Euro-dollar rate, the inflation rate, and a pressure
variable reflecting a liquidity surplus or deficit. The other interest rates, especially those
corresponding to the assets in the portfolio model, are appended to the money market
rate using term structure equations. The external value of the O-mark against the U.S.
dollar and other non-EMS currencies is explained by the respective price and interest
rate differentials and a risk premium, which depends on the net foreign position of West
Germany. Consistent with the view that realignments within the EMS are an instru-
ment of discretionary exchange rate policy, the external value of the O-mark against
other EMS currencies is assumed to be exogenous.
Obviously, exchange rate movements have multiple effects on the current account. In
this context, it is interesting to note a result from simulations carried out using the
foreign sector block. An appreciation of the D-mark against the U.S. dollar leads to a
slight improvement of the current account due to the pronounced reduction of import
prices, whereas an appreciation of the O-mark against other currencies produces the
normal reaction: the current account clearly deteriorates as a consequence of the
prevailing negative effect on exports. There arc also direct influences of interest rates
on the real sphere, since real interest rates play an important role in the determination
of private investment (block B) and consumption (block A). On the other hand, the
structure of the private portfolio depends on the volume of transactions.
The core of the new financial part of the Bundesbank Model consists in an integrated
system of asset demand equations describing the portfolio selection of the private sector.
This system, built in the tradition of the Brainard-Tobin model [3], can be derived from
the maximization of the expected utility of the portfolio, subject to the constraint that
portfolio shares must add up to one (see [2] for mathematical details). The utility of the
portfolio is assumed to be a function of both the vector of individ ual rates of return of
the considered assets and their variance-covariance matrix representing the risk
structure of the assets. Additional explanatory variables, e.g., the volume of
transactions, can be included in the utility function. By choosing a suitable
mathematical form of the utility function, the maximization results in a system of linear
asset demand equations:
n m
(1 ) + L b ..
lJ
r.
J
+ L c ..
lJ
X.
J
(i=l, ... ,n).
j=l j=l
Thus, the optimal portfolio share yI of asset i depends on the rates of return
r i (i = I, ... ,n) of all assets of the system, apart from additional explanatory \'ariables
631
x j (j = I, ... ,m). To allow for temporary deviations of the actual from the optimal
portfolio, usually a multivariate model of partial adjustment is used:
n
(2 ) y.(-l) +
1
L d ..
1J
( y.
J
- y'~ (-1)
J
} (i=l, ... ,n).
j=l
Consequently, asset demand equations explaining the actual portfolio structure always
contain the whole set of lagged endogenous variables. However, the coefficients of these
asset demand equations must satisfy various linear restrictions implied by the
adding-up constraint.
n
(3 ) r L y.(-l) r . .
i=l 1 1
In each asset demand equation, this portfolio yield is substituted for all interest rates
of competing assets so that only two interest rates remain in each equation. Moreover,
by imposing dij = 0 for i F j, the adjustment process (2) is assumed to be univariate,
which implies (due to the adding-up constraint) that the remaining coefficients
d ii (i = 1, ... ,n) are of the same size. According to these assumptions the asset demand
eq uations are specified in the form:
(4) Yi
v·1
o
portfolio share of asset i (per cent)
r i interest rate of asset i (per cent p.a.)
r portfolio yield (per cent p.a.)
x GNP divided by total private portfolio
t dummy variable for withholding tax
q I ,q2,q3 : seasonal dummies.
In the portfolio block of the Bundesbank Model, nine financial assets are included, the
selection being made with reference to the available banking statistics. Table I shows
the list of assets as well as their respective shares and their yields a\eraged over the
years 1971 to 1988.
632
It should be noted that the yield on foreign assets is calculated as an average yield on
long-term government bonds in selected countries, corrected by changes in the effective
D-mark exchange rate in the past. Thus, the volatility of the yield on foreign assets
exceeds that of domestic assets by far.
In order to take into account the various constraints on the coefficients, the portfolio
system is estimated simultaneously using Zellner's RJG LS method. The resulting
estimates are presented in table 2.
References