The Great Crash, the Oil Price Shock, and the Unit Root Hypothesis

Author(s): Pierre Perron

Reviewed work(s):
Source: Econometrica, Vol. 57, No. 6 (Nov., 1989), pp. 1361-1401
Published by: The Econometric Society
Stable URL: http://www.jstor.org/stable/1913712 .
Accessed: 05/06/2012 13:00
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to Econometrica.

http://www.jstor.org

Econometrica, Vol. 57, No. 6 (November, 1989), 1361-1401

THE GREAT CRASH, THE OIL PRICE SHOCK, AND THE UNIT
ROOT HYPOTHESIS
BY PIERRE PERRON1

We consider the null hypothesis that a time series has a unit root with possibly nonzero
drift against the alternative that the process is "trend-stationary." The interest is that we
allow under both the null and alternative hypotheses for the presence of a one-time change
in the level or in the slope of the trend function. We show how standard tests of the unit
root hypothesis against trend stationary alternatives cannot reject the unit root hypothesis
if the true data generating mechanism is that of stationary fluctuations around a trend
function which contains a one-time break. This holds even asymptotically. We derive test
statistics which allow us to distinguish the two hypotheses when a break is present. Their
limiting distribution is established and selected percentage points are tabulated. We apply
these tests to the Nelson-Plosser data set and to the postwar quarterly real GNP series. In
the former, the break is due to the 1929 crash and takes the form of a sudden change in the
level of the series. For 11 out of the 14 series analyzed by Nelson and Plosser we can reject
at a high confidence level the unit root hypothesis. In the case of the postwar quarterly real
GNP series, the break in the trend function occurs at the time of the oil price shock (1973)
and takes the form of a change in the slope. Here again we can reject the null hypothesis of
a unit root. If one is ready to postulate that the 1929 crash and the slowdown in growth
after 1973 are not realizations of an underlying time-invariant stochastic process but can be
modeled as exogenous, then the conclusion is that most macroeconomic time series are not
characterized by the presence of a unit root. Fluctuations are indeed stationary around a
deterministic trend function. The only "shocks" which have had persistent effects are the
1929 crash and the 1973 oil price shock.
KEywoRDs: Hypothesis testing, intervention analysis, structural change, stochastic
trends, deterministic trends, functional weak convergence, Wiener process, macroeconomic
time series.

1. INTRODUCTION

THE UNIT ROOT HYPOTHESIShas recently attracted a considerable amount of

work in both the economicsand statisticsliterature.Indeed, the view that most

economic time series are characterizedby a stochasticratherthan deterministic
nonstationarityhas becomeprevalent.The seminalstudy of Nelson and Plosser
(1982) which found that most macroeconomicvariableshave a univariatetime
series structurewith a unit root has catalyzeda burgeoningresearchprogram
with both empiricaland theoreticaldimensions.
Nelson and Plosser'sstudy was followed by a series of empiricalanalyses
which basically confirmedtheir findings.Some (Stulz and Wasserfallen(1985)
and Wasserfallen(1986), among others)applieda similarDickey-Fuller (1979)
statistical methodologyto other economic series. On the statisticalfront, there
11 wish to thank Brian Campbell, Larry Christiano, Jean-Marie Dufour, Clive Granger, Whitney
Newey, Hashem Pesaran, the referees, and the editor for useful comments. Christian Dea and
Nicholas Marceau provided useful research assistance. This research was supported by the Social
Sciences and Humanities Research Council of Canada, the Natural Sciences and Engineering Council
of Canada, and Quebec's F.C.A.R. grants. The first draft of this paper was written while the author
was Assistant Professor at the Universite de Montreal.

1361

1362

PIERRE PERRON

emerged an interest in developingalternativeapproachesto test the unit root

hypothesis.Examplesinclude:the class of tests proposedby Phillipsand Perron
(1988) and the methodologysuggestedby Campbelland Mankiw(1987, 1988)
and Cochrane(1988) using an estimateof the spectraldensityat frequencyzero.
Empiricalapplicationsof thesemethodologiesgenerallyreaffirmedthe conclusion
that most macroeconomictime serieshave a unit root (e.g., Perron(1988)).
These studies had many effectson economictheorizing.They seem to confirm
previous analyses which had advancedthe unit root hypothesisfor particular
economic series, e.g., consumption(Hall (1978)),velocity of money (Gould and
Nelson (1974)), and stock prices(Samuelson(1973)).They also launcheda series
of theoreticalinvestigationswith implicationsconsistentwith a unit root, e.g.,
Blanchardand Summers(1986) for employment.Furthermore,a considerable
stock of statisticaltools was developedfor more generalmodels with integrated
variables;these includethe cointegrationframework(Engleand Granger(1987))
and multivariatesystems (Stock and Watson (1988) and Phillips and Durlauf
(1986)).
As far as macroeconomictheoriesare concerned,the most importantimplication of the unit root revolution,is that underthis hypothesisrandomshockshave
a permanenteffect on the system.Fluctuationsare not transitory.This implication, as forcefullyarguedby Nelson and Plosser,has profoundconsequencesfor
businesscycle theories.It runscounterto the prevailingview that businesscycles
are transitoryfluctuationsarounda moreor less stabletrendpath. It is therefore
of importanceto assess carefullythe reliabilityof the unit root hypothesisas an
empiricalfact.
The aim of this paper appears startling, given the results in the above
mentionedliterature.Ourconclusionis that most macroeconomictime seriesare
not characterizedby the presenceof a unit root and that fluctuationsare indeed
transitory.Only two events(shocks)have had a permanenteffect on the various
macroeconomicvariables:the Great Crashof 1929 and the oil price shock of
1973.
Of course, to reach such a conclusion,a particularpostulatemust be introduced which differentiatesour approachfromthe previousones. This postulateis
that the Great Crash and the oil price shock were not a realizationof the
underlyingdata-generatingmechanismof the various series. In this sense, we
consider these shocks as exogenous.The exogeneityassumptionis not a statement about a descriptivemodelfor the time seriesrepresentationof the variables.
It is used here as a deviceto removethe influenceof these shocksfrom the noise
function.A more detaileddiscussionof theseissuesand theirimplicationscan be
found in Section 6.
These two shocksare ratherdifferentin nature.On one hand, the GreatCrash
created a dramaticdrop in the mean of most aggregatevariables.On the other
hand, the 1973 oil priceshockwas followedby a changein the slope of the trend
for most aggregates,i.e., a slowdownin growth.In this light, our aim is to show
if one allows a single
that most macroeconomicvariablesare "trend-stationary"

UNIT ROOT HYPOTHESIS

1363

changein the interceptof the trendfunctionafter1929 and a singlechangein the

slope of the trendfunctionafter 1973.
Our approachis in the spiritof the "interventionanalysis"suggestedby Box
and Tiao (1975). Accordingto their methodology,"aberrant"or "outlying"
events can be separatedfrom the noise functionand be modeledas changesor
"interventions"in the deterministicpart of the generaltime seriesmodel. Using
such a strategy makes it "possible to distinguishbetween what can and what
cannot be explainedby the noise"(Box and Tiao (1975,p. 72)). These "interventions" are assumedto occur at a known date. The same strategyis used in the
presentanalysisin that we considerthe time of the changesin the trendfunction
as fixed ratherthan as a randomvariableto be estimated.
To make our point as unambiguousas possible,we use the same data set as
Nelson and Plosser,as well as the real GNP series analyzedby Campbelland
Mankiw.The data set used by Nelson and Plossercontainsfourteenmacroeconomic variablessampledannually.All series end in 1970 and contain only one
break, the 1929 GreatCrash.We shall not analyzethe unemploymentrate series
since there is a generalagreementthat it is stationary.The real GNP series is
postwarquarterlyfrom1947:1to 1986:IIIand so containsonly one breakas well,
the 1973 oil shock. Furthermore,to make our analysisas similaras possible to
previous ones, the statisticalmethodologyapplied here is an extension of the
Dickey-Fuller methodology(as used by Nelson and Plosser) to test for the
presenceof a single unit root in a univariatetime series.
The plan of the paperis as follows. Section 2 motivatesthe ensuinganalysis
and presentsthe alternativemodels considered.Section 3 shows that usual tests
will not be able to rejectthe unit root hypothesisif in fact the deterministictrend
of the serieshas a singlebreak(eitherin the interceptor the slope). In Section4,
we develop formalstatisticaltests of the null hypothesisof a unit root whichcan
distinguish the unit root hypothesisfrom that of a stationaryseries around a
trend which has a single break. The asymptoticdistributionsunder the null
hypothesis are derived and tabulated. Empiricalresults from applying these
proceduresare presentedin Section 5. Section 6 contains a discussionof some
issues raisedby our analysisand suggestionsfor futureresearch.All theoremsare
proved in AppendixA.

2. MOTIVATION

The null hypothesisconsideredis that a given series {Yt

}, (of which a sample
of size T + 1 is available)is a realizationof a time seriesprocesscharacterizedby
the presenceof a unit root and possiblya nonzerodrift. However,the approach
is generalizedto allow a one-time change in the structureoccurringat a time
TB(1 < TB< T). Threedifferentmodelsareconsideredunderthe null hypothesis:
one that permitsan exogeneouschangein the level of the series(a "crash"),one
that permitsan exogenouschangein the rateof growth,and one that allowsboth

1364

PIERRE PERRON

change: These hypothesesare parameterizedas follows:

Null hypotheses:
Model (A)

yt = I + dD(TB) t +yt-

Model(B) yt = i +yt- + (M2-

+ et,

)DUt + et,

Model(C) Yt=
=l+Yt-l+dD(TB)t+(y2-#1)DUt+et,
D(TB)t = 1

if

DUt= 1

if t> TB,

A (L) et = B(L)

t=TB +l,

where

Ootherwise;
0 otherwise; and

Vt,

vt - i.i.d. (0, 2), with A(L) and B(L) pth and qth orderpolynomials,respectively, in the lag operatorL.
The innovation series {et}) is taken to be of the ARMA(p, q) type with the
orders p and q possibly unknown. This postulate allows the series {yt to
representquite generalprocesses.More generalconditionsare possible and will
be used in subsequenttheoreticalderivations.
Instead of consideringthe alternativehypothesisthat Ytis a stationaryseries
around a deterministiclinear trend with time invariantparameters,we shall
analyze the followingthreepossiblealternativemodels:
Alternative hypotheses:
Model(A)

Model(B)
Model (C)

Y1=P1+#t+GL2-ttj)DU,+et,
Yt= L+ 13lt+ (/2 - 1l)1DTt* + et,
Yt= 1 + ,81t+ (A2- Aj)DUt + (/2-

I)DTt + et

where
D1Tt

t TB,

and

DTt = t if

t > TB and O otherwise.

Here, TB refersto the time of break,i.e., the periodat whichthe changein the
parametersof the trendfunctionoccurs.Model(A) describeswhatwe shall refer
to as the crash model. The null hypothesisof a unit root is characterizedby a
dummy variable which takes the value one at the time of break. Under the
alternativehypothesis of a "trend-stationary"system, Model (A) allows for a
one-timechangein the interceptof the trendfunction.For the empiricalcaseswe
have in mind, TB is the year 1929 and JU2< 1.- Model (B) is referredto as the
"changing growth" model. Under the alternativehypothesis,a change in the
slope of the trendfunctionwithoutany suddenchangein the level at the time of
the breakis allowed.Underthe null hypothesis,the model specifiesthat the drift
parameter ,u changes from 1,k to tL2 at time TB. In the empiricalexamples
presentedin Section 5, TB is the first quarterof 1973 and /2 < /1, reflectinga
slowdownin growthfollowingthe oil shock.Model(C) allowsfor both effectsto
take place simultaneously,i.e., a sudden change in the level followed by a
differentgrowthpath.

1365

UNIT ROOT HYPOTHESIS

9.2 -

9
8.8
8.6
8.48.287.8-

7
6.6
6.4

1900

1910

1920

1940

1930

1950

1960

1970

Note: The broken straight line is a fitted trend (by OLS) of the form Y,= ,u+ IDU, + fit where
DUt = O if t < 1929 and DUt = 1 if t > 1929.
FIGURE 1.-Logarithm of "Nominal Wages."

To motivate the use of these threemodels as possible alternativesto the unit

root with drift hypothesis,we presentin this section some descriptiveanalyses
for three series: "nominal wages" (1900-1970), "quarterly real GNP"
(1947:1-1986:III)and "commonstock prices"(1871-1970).
Figure1 shows a plot of the logarithmof the nominalwage series.A featureof
this graph is the marked decreasebetween 1929 and 1930. Apart from this
change, the trend appearsfairlystable (same slope) over the entire period.The
solid line is the estimatedtrendline froma regressionon a constant,a trendand
a dummyvariabletakinga value of 0 priorand at 1929 and value 1 afterwards.
Table I presents the results from estimating (by OLS) a regressionof the
Dickey-Fullertype, i.e.:
k
(1)

Yt

= /I

t +

ECiAYt-i

ot

i=l

The first row presentsthe full sample regression.The coefficienton the lag
dependent variableis 0.910 with a t statistic for the hypothesisthat a = 1 of
-2.09. Using the criticalvaluestabulatedby Dickeyand Fuller,we cannotreject
the null hypothesisof a unit root. When the sampleis split in two (pre-1929and
post-1929), the estimated value of a decreases dramatically:0.304 for the
pre-1929sample and 0.735 for the post-1929sample.However,due to the small
samplesavailable,the t statisticsarenot largeenough(in absolutevalue)to reject
the hypothesisthat a = 1, even at the 10 percentlevel.
Two features are worth emphasizingfrom this example:(a) the full sample
estimate of a is markedlysuperiorto any of the split sample estimates and
relativelyclose to one. It appearsthat the 1929 crashis responsiblefor the near
unit root value of a; and (b) the split sampleregressionsare not powerfulenough

PIERRE PERRON

1366

TABLE I
REGRESSIONANALYSISFORTHEWAGES, QUARTERLYGNP, AND COMMONSTOCKPRICE SERIES
Ay i + it

Regression:yt = &+ ,#t + Fyty5 +

k

Series/Period

(a) Wages
1900-1970a
1900-1929
1930-1970
(b) Common stock prices
1871-1970a
1871-1929
1930-1970
(c) Quarterly real GNP
1947:I-1986:111
1947:I-1973:1
1973:II-1986:III

2
7
8

0.566
4.299
1.632

2
3
4

0.481
0.3468
-0.5312

2
2
1

0.386
0.637
0.883

t#

ta

0.004
0.037
0.012

2.30
2.73
2.64

0.910
0.304
0.735

-2.09
-2.82
-3.19

0.060
0.0803
0.0269

2.02
2.13
-1.64

0.003
0.0063
0.0166

2.37
2.70
1.96

0.913
0.732
0.788

-2.05
-2.29
-1.89

0.158
0.1209
0.1376

2.90
3.04
2.23

0.0004
0.0008
0.0008

2.71
2.99
2.27

0.946
0.910
0.878

-2.85
-3.02
-2.23

0.010
0.0099
0.0102

fi

tA

2.30
2.84
3.60

S(e)

aResults taken from Nelson and Plosser(1982, Table 5).

to rejectthe hypothesisthat a = 1 even thoughthe estimatesare well below one.

It would be useful,in this light, to have a morepowerfulprocedurebased on the
full sample that would allow the 1929 breakto be exogenous.
Figure2 graphsthe postwarquarterlyreal GNP series.Here,the seriesbehave
accordingto Model(B) wherethereis no sharpchangein the level of the seriesat
the 1973:1breakpoint but rathera changein the slope. The solid line is a fitted
trend where a dummyvariableis includedin the regression,taking the value 0
prior and at 1973:1and the value (t - 105) after 1973:1(1973:1being the 105th
observationin the sample).Table I comparesregressionsof the form(1) with full
and split samples.Again,the estimateof a is lowerin both subsamplesthanwith
TABLE II
OF THE "DETRENDED"SERIES
SAMPLEAUTOCORRELATIONS
Period

Series

T Variance r5

r3

Quarterly GNP

B 47:I 86:III 159 0.001 0.94 0.83 0.70

62
62
62
111
81
82
111
71
71
82
102
71
100

0.010
0.023
0.012
0.017
0.005
0.015
0.066
0.016
0.003
0.023
0.036
0.587
0.066

0.45
0.31
0.54
0.44
0.59
0.63
0.89
0.47
0.40
0.69
0.79
0.58
0.53

r4

r5

r6

0.23 0.11 0.05 0.04

0.12 0.08 0.11 0.12
0.33 0.20 0.13 0.09
0.32 0.17 0.08 0.12
0.43 0.30 0.20 0.15
0.45 0.31 0.17 0.06
0.80 0.71 0.63 0.54
0.26 0.12 0.03 -0.03
0.12 -0.12 -0.27 -0.33
0.52 0.38 0.25 0.11
0.70 0.62 0.57 0.52
0.38 0.25 0.15 0.11
0.36 0.20 0.10 0.08

A
A
A
A
A
A
A
A
C
A
A
A
C

1909-1970
1909-1970
1909-1970
1860-1970
1890-1970
1889-1970
1860-1970
1900-1970
1900-1970
1889-1970
1860-1970
1900-1970
1871-1970

0.77
0.68
0.81
0.71
0.82
0.82
0.96
0.76
0.74
0.87
0.90
0.77
0.80

r2

Real GNP
Nominal GNP
Real per capita GNP
Industrial production
Employment
GNP deflator
Consumer prices
Wages
Real wages
Money stock
Velocity
Interest rate
Common stock prices

0.57

0.45

0.35

Note: A, B, and C denote the detrending procedure corresponding to the given model under the alternative
hypothesis.

1367

UNIT ROOT HYPOTHESIS

8.38.28SI
8797877767574737.27.1'
7
.
6.9
1950
1945

1955 1960 1965 1970 1975 1980

985
+
of
the
form:
fitted
trend
=
is
a
line
OLS)
broken
(by
Note: The
straight
fit + yDT* where
Yj ,i
DTt *=O if t < 1973:I and DT,* = t-TB if t > 1973:I = TB.
FIGURE2.-Logarithm of "Postwar Quarterly Real GNP."

the full sample (given the quarterlynature of the series, the difference is
important).The same featuresdiscussedabove appearto hold when there is a
change in the slope of the trendfunction.
As a final example,considerthlecommonstock price seriesgraphedin Figure
3. The break point is again in 1929 but in this case there appearsto be both a
sudden change in the level of the seriesin 1929 and a highergrowthrate after.
The solid line is the estimated trend with two dummy variables added, an
interceptdummy(Opriorand at 1929, 1 after 1929) and a slope dummy(Oprior
5.4.5-

4
3.53

21.51870 1880 1890 t00

1910 1920 1930 1940 1950 1960 1970

?
Note: The broken straight line is a fitted trend (by OLS) of the form Y-,= jI+
where DU, = DT, = 0 if t 61929 and DU, = 1, DTt = t if t > 1929.
FIGURE 3.-Logarithm

of "Common Stock Prices."

DU, + fit + Y2DTt

1368

PIERRE PERRON

and at 1929 and t after1929).The estimatedvaluesof a (in regression(1)) with

the full sample are 0.913 but are only 0.732 using the pre-1929sampleand 0.788
using the post-1929.Hereagain,the t statisticsare not largeenough,however,to
reject the unit root hypothesisat even the 10 percent level using any of the
subsamples.2

Table II presentsthe autocorellationfunctionof the "detrendedseries"for the

full set of variablesanalyzedby Nelson and Plosser, along with the postwar
quarterlyreal GNP series.All seriesare detrendedaccordingto Model (A) (with
a constant, a trend, and an interceptdummy)except for the postwarQuarterly
Real GNP Series (with a slope dummyinstead of the interceptdummy,Model
(B)) and the real wage and common stock price series (with both a slope and
interceptdummy,Model(C)). Unlike the "standard"detrendedseries(see Table
4 of Nelson-Plosser),the autocorrelationsdecay quite rapidly for all variables
except for the consumerpricesand velocityseries.This behaviorof the autocorrelation function is certainlynot the one usually associatedwith either a random
walk or a detrendedrandomwalk.Indeed,the "detrended"seriesappearstationary.
The results of this section motivate the analysis presentedin the following
sections. We first investigatethe effectsof the two types of changesin the trend
function that we consideron the statisticalpropertiesof autoregressiveestimates
of the type found in regression(1) (both in finite samples and asymptotically).
We find that such changescreatea spuriousunit root that may not vanish,even
asymptotically.To overcome the problem of the low power associated with
testing for a unit root using split samples,formaltest statistics,whichpermitthe
presence of either or both an interceptand a slope shift, are developedin Section 4.
3. THE EFFECTOF A SHIFT IN THE TREND FUNCTION ON TESTS
FOR A UNIT ROOT

To assess the effectsof the presenceof a shift in the level of the seriesor a shift
in the slope (at a singlepoint of time) on tests for the presenceof a unit root, we
first presenta small Monte Carloexperiment.Consider firstthe "crashhypothesis" (Model (A)). We generated10,000replicationsof a series {y)} of length 100
definedby
(2)

YtA=i1+ (2-

lic)DUt+ Pt + et

(t = 1,. .., T),

where DUt =1 if t > TB,and 0 otherwise.

2Dickey-Fuller tests for the presence of a unit root using split samples are presented in Appendix
B for all the series considered. The results are presented for values of k ranging from 1 to 12. These
results show that (i) the conclusions drawn are not sensitive to the value of k chosen; (ii) for some
series it is possible to reject the unit root hypothesis, especially when considering the post-1929
subsample. Furthermore, the statistical significance of the lagged first-differences (not reported)
suggest that a large value of k may be needed. For example, the t statistics on the eighth lagged
first-difference is often statistically significant. A similar pattern will occur in the full sample tests
reported in subsequent sections.

UNIT ROOT HYPOTHESIS

1369

0.9
0.8

0.70.6
0.50.40.3-

0 II
0

I
0.8
0.6
0.4
mechaThe
data-generating
(4).
parameterin regression
Note: a is the estimatedautoregressive
nism is givenby equation(2) with t,u= 0, fi = 1.0 and {et} i.i.d. N(0, 1), T = 100 and TB= 50.
-0.4

-0.2

0.2

4.-C.D.F. of &underthe "Crash"Model.

FIGURuE

For simplicity, t,z = 0, / = 1, TB= 50, T = 100 and the innovations et are i.i.d.
N(O, 1). For the "changing growth" hypothesis, a similar setup is considered
except that yt is generated by
(t =l,~...,~T),
Yt=P+Plt+ (#2- #,)DTt* +et
(3)
where DTt* = t -TB if t> TB, and O otherwise.
Again, ,u= 0, f31= 1, TB= 50, T = 100, and et - i.i.d. N(O,1). For each replication, we computed the autoregressive coefficient x in the following regression,
using ordinary least squares:
t(4)
Yt=A+t+?'Yt-l+
Figure 4 graphs the cumulative distribution function of a when the data
generating process (D.G.P.) is given by (2) for various values of IL2. This
experiment reveals that as the magnitude of the crash increases (.U2 decreases),
the c.d.f of & becomes more concentrated at a value ever closer to 1. The
corresponding mean and variance of the sample of a generated are shown in
Table III. Figure 5 graphs the c.d.f. of a when the D.G.P. is given by (3) for
various values Of fP2. As fP2 diverges from /3l, again, the c.d.f. becomes more
concentrated and closer to one. The computed mean and variance of & presented
in Table III confirms this behavior.
andhencecan be used
3Note thatwhen the errorstructureis i.i.d., a is freeof nuisanceparameters
as a formaltest statisticon the samegroundas the t statistic.However,we also performeda similar
experimentwith the t statisticon a (a = 1) in regression(4) as well as in a regressionwith additional
as regressors.The resultsobtainedshow the samebehavior.If anything,the t
lags of first-differences
as regressorsshowsa still greaterbias towardnonrejection
statisticwith extralags of first-differences
of the null hypothesisof a unit root.Theseresultsareavailableuponrequest.We preferto reportour
resultin termsof the behaviorof the estimatora insteadof its t statisticbecauseit makesclearthat
what causes the nonrejectionis not due solely to the behaviorof the varianceestimator.Whatis of
importanceis that &is biasedtowardsunity.

1370

PIERRE PERRON
TABLE III
MEAN AND VARIANCE OF a

(a) Crash Simulations,

2= 0

2 =-2

-0.019
0.00986

Mean
Variance

0.172
0.01090

1 = 0,

fi =

P2 = - 5

F2

0.558
0.00471

0.795
0.00089

10

L2 =-25

0.899
0.00009

(b) Breaking Trend Simulations, /It = 1, it= 0

2 = 1-.0

Mean
Variance

= 0.9

0.334
0.00938

-0.019
0.00986

= 0.7

0.825
0.00094

= 0-4

2 = 0.0

0.949
0.00009

0.981
0.00001

See notes to Figure 4 for case (a) and Figure 5 for case (b).

What emerges from this experiment is that if the magnitude of the shift is
significant, one could hardly reject the unit root hypothesis even if the series is
that of a trend (albeit with a break) with i.i.d. disturbances. In particular, one
would conclude that the shocks have permanent effects. Here, the shocks clearly
have no permanent effects, only the one-time shift in the trend function is
permanent.
To analyze the effect of an increase in the sample size on the distribution of a
with a shift of a given magnitude, we derive the asymptotic limit of a. To this
end, we again consider processes generated by Models (A), (B), or (C) under the
alternative hypotheses, but we enlarge the framework by allowing general conditions on the error structure {et }. Many such sets of conditions are possible and
would allow us to carry out the asymptotic theory. For simplicity, we use the
I

0.9
0.8
0.70.6
0.504
0.3
0.2-02

OI

/-/-

.82

O.9

92O.7

I-

A=4~

0.I
0

-0.4

-0.2 ,.2

, ,,

0,

0.2

0.4

0.6

0.8

'

Note: a is the estimated autoregressive parameter in regression (4). The data-generating mechanism is given by equation (3) with ,u= 0, f1 = 1.0, { et } i.i.d. N(O,1), T = 100, TB = 50.
FIGURE 5.-C.D.F.

of & under the "Breaking Trend" Model.

1371

UNIT ROOT HYPOTHESIS

"mixing-type"conditionsof Phillips(1987)and Phillipsand Perron(1988).These

are stated in Assumption1.
(a) E(e,) = 0 all t; (b) SUpt Elet1?e+ < X for some /3> 2 and
{e1}
T-1E(ST) exists and a2>O, where ST= YTef;(d)
is strong mixing with mixing numbers am that satisfy: OOa1-2/f < 00.

ASSUMPTION 1:
E>O; (C) aJ2 =limTM

These conditionsare generalenoughto permitthe series {et} to be generated

by a finite orderARMA(p, q) processwith Gaussianinnovations.To carryout
the asymptoticanalysis,we shall requirethat both the pre-breakand post-break
samples increase at the same rate as the total number of observations, T,
increases.To this effect, we assume,for simplicity,that TB = XT for all T. We
referto A as the "breakfraction."The asymptoticlimits are takenas T increases
to infinity in a sequencethat ensuresan integervalue of TB for a given A. This
type of increasingsequenceis assumedthroughoutthe paper.The resultsproved
in AppendixA are presentedin the followingtheorem.
THEOREM 1: Let {Yt}o be a sample of size T+ 1 generated under one of the
alternative hypotheses with the innovations { et } satisfying Assumption 1. Let '-'
denote convergencein probability. Furthermore,let TB = AT, for all T and 0 < A <
1; then as T-* oo:
(a) The "crash hypothesis": Under Model (A)
{ [It'l

-2

]2A

+ 71) } {[ul

- P2 ]2A

ae2) }1

where
T

A=[A-4A2+6A3-3A4],

Y1= lim T-1' E(e,e,1),

T- oo

and
T

2 =

lim

T-1' EE(e2)
1

(b) The "breaking trend hypothesis": Under either Model (B) or (C)
(i)

(ii)

cx

1,

T(a- -1) {- 3(-1 + 4A - 52 + 2A3)}

3A2+3A3-A4)}1
-{2(-3+4A-

Part (a) of Theorem1 shows that under the crash hypothesis,the limit of a
depends on the relativemagnitudeof I,1 - Iu212 A and ae2.In particular,this
limit gets closer to one as 1,I - ,U2]2increases.Anotherfeatureis that the limit of
a is always greater than the true first-orderautoregressivecoefficientof the
stationarypart of the series, yJ/a2. However,since &-does not convergeto 1, the

1372

PIERRE PERRON

usual statistics for testing that a = 1, such as T(d - 1) or the t statistic on a,

would eventually reject the null hypothesis of a unit root. Nevertheless, added to
the generally poor power properties of tests for a unit root is the consideration
that the limit of a is inflated above its true value. These conditions are such that
it could be difficult to reject the unit root hypothesis in finite samples.
There is another interpretation to the results under the crash hypothesis. As
stated in model (A), the change in the intercept of the trend function is given by
a fixed value. This implies that in the asymptotic derivations we are
(M2 -,1),
considering a shift which decreases relative to the level of the series as the sample
size increases. It may be more appropriate to specify the change in the intercept
as a magnitude relative to the level of the series at the time of the break. Since at
this period the level of the series is proportional to TB, we can specify ( L2as a proportion of TB, say, (A2 - Al) = YTB. In this case the "crash" is proportional to the level of the series. Since TB-- o as T -x 00, it is clear, from part (a),
that under this interpretation, a
1.
Such ambiguity does not occur under the "breaking trend hypothesis" (Models
(B) or (C)) as is shown by part (b) of Theorem 1. Here, the limit of a is 1
irrespective of the behavior of the intercept and the limit of T(& - 1) is invariant
to the relative magnitude of the shift (f2 versus fl). The expression in part (b, ii)
is a function of X. However, it varies from 0 to 1/2 for values of X in the range
(0, 1). Since the one-sided 5 percent asymptotic critical value of T(5 - 1) is
- 21.8 under the null hypothesis of a unit root (Fuller (1976)), Theorem 2 implies
that the unit root hypothesis could not be rejected, even asymptotically.4
These results could be extended to more general test statistics, such as the t
statistics. Nevertheless, the picture is clear. Tests of the unit root hypothesis are
not consistent against "trend stationary" alternatives when the trend function
contains a shift in the slope. Although they are not inconsistent against a shift in
the intercept of the trend function (if the change is fixed as T increases), their
power is likely to be substantially reduced due to the fact that the limit of the
autoregressive coefficient is inflated above its true value. When interpreting the
"crash" as proportional to the level of the series as T increases, a unambiguously
converges to one and implies a considerable loss in power. There is therefore a
need to develop alternative statistical procedures that could distinguish a process
with a unit root from a process stationary around a breaking trend function.
4. ALTERNATIVE STATISTICAL PROCEDURES

In this section, we extend the Dickey-Fuller testing strategy to ensure a

consistent testing procedure against shifting trend functions. We shall present
several ways to do so, all of which are asymptotically equivalent, and discuss the
main differences between each.
After the first draft of this paper was written, we became aware of a result similar to part (b, i) of
Theorem 1 proved by Rappoport and Reichlin (1987). In fact, in the case of deterministic trends with
multiple shifts in slope, they prove the following more general result: "If the true model contains
K + 1 segments, then any fitted model involving K or less segments will, asymptotically, yield a larger
sum of squared residuals than [a difference stationary] model" (p.9).

1373

UNIT ROOT HYPOTHESIS

Considerfirst detrendingthe raw series { y,} accordingto either model (A),

(B), or (C). Let {Y,}, i = A, B, C be the residualsfrom a regressionof Y, on (1)
i = A: a constant, a time trend, and DU,; (2) i = B: a constant, a time trend, and

DT,*; (3) i= C: a constant,a time trend, DUE,and DT. Furthermore,let a' be

the least squaresestimatorof a in the followingregression:
(5)

yt'= alYt'_1 + et

(i =A, B, C; t = 1,29,.... T).

Up to this point the extensionsfrom the no breakmodel are straightforward

enough. However,mattersare not so simple concerningthe distributionof the
statisticsof interest,namelythe normalizedbias, T(&i- 1), and the t statisticon
&, t'i (i = A, B, C). Needless to say, the only manageable analytical distribution

theory is asymptoticin nature.But two additionalfeaturesare also presentover

the usual procedure:(a) extraregressorsand (b) the split samplenatureof these
extra regressors.To this effect,we derivethe asymptoticdistributionof T(a&- 1)
and t,i under the null hypothesisof a unit root. As in Section3, we requirethat
the breakpoint TB increasesat the same rate as the total samplesize T. Again,
for simplicity,it is assumedthat TB= XT with both T and TB integervalued.
The method of proof is similar to that of Phillips (1987) and Phillips and
Perron (1988). We use weak convergenceresults that hold for normalized
functions of the sum of the innovationswhen the latter are assumedto satisfy
Assumption1. The limitingdistributionsobtainedunder this generalsetting are
then specializedto the i.i.d. case. The asymptoticdistributionsin the i.i.d. case
are evaluatedusing simulations,and criticalvalues are tabulated.We then show
how the results can be extended to innovations {et} that follow the general
ARMA(p, q) process in the same way that the Dickey-Fullerregressionsare
modifiedby addingextralags of first-differences
of the data as regressors.
The main results concerningthe asymptoticdistributionsof the normalized
bias estimatorsand the t statisticsof the autoregressivecoefficientunderthe null
hypothesisof a unit root are presentedin the next theorem.
THEOREM 2: Let{ytj be generated under the null hypothesis of model i
(i = A, B, C) with the innovation sequence { et I satisfying Assumption 1. Let =*
denote weak convergencein distributionand X = TB/T for all T. Then, as T-- oo:

(a)

T( a-1)

(b)

*
i

HI/Ki,

(giKi)12,

where
HA = gAD1-D541-D642;
HB =giD,

+ D53

+ D8s44;

Hc = gcDg + D13&5- D1446;

KA= 9AD2-D402-D301;
KB=gD2

+ D74 + D3043;

Kc = gcD10 - D12&6 + D1145;

1374

PIERRE PERRON

with

P1= 6D4 +

6D3 + (1

#2 =

12D3;

+3 =

(1 + 2X)(1-A)

+4=

(1 + 2X)(1

X) 1kD3-(1-

IP5 = D12

- Dl;

X)1-1)D;

+ 3X)D3;

D7-(1

3
D7;

A6 = A5 + (1-_A)2

D12/A3;

and
( )(w(1)2

D=

-_ e2/02)

D2 =

w(r)2dr

D3 =

rw(r) dr-(
w(1)/2 - jw(r)

D7 =

frw(r)

dr;

-(w(1)

jw(r)

dr;
( w(r) dr)

w(r) drX(-)
e2/2)/2

- w(X))(1

X-lw(X)

X)1

Xw(r) dr
(r) dr;
dr + (2)jXw(r)

w(r) dr;

+ w(X)/2 - jw(r)

Xw(X)/2 - jw(r)

A = 1- 3(1 - X)X;

+ A)|jw(r)

rw(r) dr - (X/2)j

D13= (1-X)w(l)/2

D14=

w(r) dr;

dr- ((1-)2/2)j

Dll f = rw(r) dr-(2(

D12=

w(r) dr-X fw(r)

D4=

D6 = w(X) -Xw(l);

dr-A Xw(r)

Ds=|w(r)d--(
(w(1)2-

dr;

)fw(r)

D8= ((l-X2)/2)w(1)-

Do=

w(r) dr;

w(1)

w(r) dr];

-[

D=

dr;

dr;
gB= 3A3;

gc= 12(1 - X)2;

dr;

dr;

1375

UNIT ROOT HYPOTHESIS

and where w(r) is the unit Wiener process defined on C[O,1], a2 = limToo
=
E[T 1 ST] ST = ' et, and a0
limTOE[T1XTe,I.
Theorem 2 provides a representation for the limiting distribution of the
normalized least squares estimators and their t statistics in terms of functionals
of Wiener processes. These limiting distributions are functions of the parameter
X, the ratio of the pre-break sample size to total sample size. It is easy to verify
that when X is either 0 or 1, the limiting distributions are identical over all
models and are given by:
H/K

T(a-1)=

and

ta=:(a1/qe)H/K1/2

2/02)

+ 12[

where

H=

(,)(w(1)2

*[j1w(r) dr-()w(1)]

K=

w(r)2

dr- 12( jrw(r)

rw(r) dr - (21w(r)

-w(l)j

dr]

w(r) dr,

dr)

+ 12j1w(r) drj rw(r) dr - 4(

w(r) dr)

These latter asymptotic distributions correspond to those derived by Phillips

and Perron (1988) in the case where no dummy variables are included.
The expressions for the limiting distributions in Theorem 2 depend on additional nuisance parameters, apart from X, namely a2 and a,2. As in Phillips
(1987) and Phillips and Perron (1988), a2 is the variance of the innovations and
a2 is, in the case of weakly stationary innovations, equal to 2 Tf(0) where f(0) is
the spectral density of {et} evaluated at frequency zero. When the innovation
sequence {et} is independent and identically distributed, a2 = a2 and, in that
case, the limiting distributions are invariant with respect to nuisance parameters,
except X.
Therefore, when a 2= a2, percentage points of the limiting distributions can be
tabulated for given values of X. Tables IV, V, and VI present selected percentage
points that will allow us to carry hypothesis testing. The critical values are
obtained via simulation methods. We briefly describe the steps involved. First, we
generate a sample of size 1,000 of i.i.d. N(O,1) random deviates, {et }. We then
construct sample moments of the data which converge weakly to the various
functionals of the Wiener process involved in the representation of the asymptotic distributions.For example,as T - oo, T-12
lTet w(1) T l/2STBe =
etc. With a
(_ (w(1)2-1),
w(X), T 3j2St =_eej Jow(r) dr, =L(gkej1)et
sample size of 1,000 and i.i.d. N(0, 1) variates, we can expect the approximation
to be quite accurate. Once the various functionals are evaluated, we construct the
expressions in Theorem 2 and obtain one realization of the limiting distributions

1376

PIERRE PERRON
TABLE IV.A
PERCENTAGEPOINTSOF THEASYMPTOTICDISTRIBuTIONOF T (a -1)
Time of Break Relative to Total Sample Size: X

A=

1%
2.5%
5%
10%
90%
95%
97.5%
99%

0.1

34.17
28.93
25.04
21.45
- 4.57
- 3.40
-2.35
- 1.28

0.2

0.3

- 35.85
- 30.35
-26.00
-22.16
-5.19
- 3.90
-2.92
-1.70

-35.07
- 29.92
-25.90
-21.93
-5.13
- 3.80
-2.85
-1.60

0.4

- 34.44
- 29.26
-25.40
-21.61
-4.28
- 2.83
-1.69
-0.61

0.5

- 34.07
- 29.00
-25.25
-21.55
- 3.85
- 2.38
-1.42
-0.40

0.6

-35.83
- 29.80
- 25.56
-21.79
-4.36
- 2.92
-1.89
-0.78

0.7

-35.59
-29.61
-25.99
- 22.33
-5.15
- 3.86
-2.78
-1.58

IN MODEL A

0.8

- 34.86
-29.40
- 25.82
-22.10
- 5.32
- 3.87
-2.84
-1.78

0.9

- 34.65
- 29.35
-25.40
-21.48
-4.62
- 3.27
-2.13
-1.39

TABLE IV.B
PERCENTAGEPOINTSOF THEASYMPTOTICDISTRIBuTIONOF ta IN MODEL A
Time of Break Relative to Total Sample Size: X
A=

1%
2.5%
5%
10%
90%
95%
97.5%
99%

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

-4.30
- 3.93
- 3.68
- 3.40
-1.38
-1.09
- 0.78
-0.46

-4.39
-4.08
- 3.77
- 3.47
-1.45
-1.14
-0.90
-0.54

-4.39
-4.03
-3.76
-3.46
-1.43
-1.13
-0.83
-0.51

-4.34
-4.01
-3.72
- 3.44
-1.26
-0.88
-0.55
-0.21

-4.32
-4.01
- 3.76
- 3.46
-1.17
-0.79
-0.49
-0.15

-4.45
-4.09
-3.76
-3.47
-1.28
-0.92
-0.60
-0.26

-4.42
-4.07
- 3.80
- 3.51
-1.42
-1.10
-0.82
-0.50

-4.33
- 3.99
- 3.75
- 3.46
-1.46
-1.13
-0.89
-0.57

-4.27
-3.97
- 3.69
- 3.38
-1.37
-1.04
-0.74
-0.47

TABLE V.A
PERCENTAGEPOINTSOF THEASYMPTOTICDISTRIBUTIONOF T(a -1)
Time of Break Relative to Total Sample Size: A
A=

1%
2.5%
5%
10%
90%
95%
97.5%
99%

IN MODEL B

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

- 34.34
- 28.74
-25.00
- 21.26
-4.27
- 3.12
-2.20
-1.11

- 37.16
- 31.97
-27.16
-23.10
- 5.09
-3.85
- 2.69
-1.58

- 38.07
-32.78
-28.61
-24.20
- 5.92
-4.50
-3.30
-2.19

- 39.21
- 33.42
-29.23
-25.04
-6.62
- 5.06
- 3.90
- 2.50

- 39.77
- 33.60
-29.65
-25.40
-6.96
-5.31
-4.14
- 3.01

-40.08
- 33.21
-29.51
-25.15
-6.71
-5.15
-3.94
-2.54

- 38.70
- 32.31
- 28.68
-24.30
-6.08
-4.59
- 3.36
-2.20

-36.18
-31.45
-27.24
-23.01
- 5.26
- 3.82
- 2.72
-1.50

- 34.69
- 29.42
-25.25
-21.24
-4.45
-3.16
-2.21
-1.24

of the statistics T(ai

1), tji (i = A, B, C). We replicate this procedure 5,000

times and obtain the criticalvaluesfrom the sortedvectorof replicatedstatistics.

This procedureis performedfor each statisticwith nine valuesof the parameter
X, the ratio of pre-breaksamplesize to total samplesize.5
Several features are worth mentioningwith respect to these critical values.
First, as expected, for a given size of the test, the criticalvalues are larger(in
5For some evidence on the adequacy of this method to obtain critical values for limiting
distributions involving functions of Wiener processes, see Chan (1988).

1377

UNIT ROOT HYPOTHESIS

TABLE V.B
PERCENTAGEPOINTSOF THEASYMPTOTICDISTRIBUTIONOF t- IN MODEL B
Time of Break Relative to Total Sample Size: X
A

1%
2.5%
5%
10%
90%
95%
97.5%
99%

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

- 4.27
- 3.94
- 3.65
- 3.36
-1.35
-1.04
-0.78
-0.40

-4.41
-4.08
- 3.80
-3.49
-1.48
-1.18
-0.87
-0.52

-4.51
-4.17
- 3.87
- 3.58
-1.59
-1.27
-0.97
-0.69

-4.55
-4.20
-3.94
-3.66
-1.69
-1.37
-1.11
-0.75

-4.56
-4.26
- 3.96
-3.68
-1.74
-1.40
-1.18
-0.82

-4.57
-4.20
- 3.95
-3.66
-1.71
-1.36
-1.11
-0.78

-4.51
-4.13
-3.85
-3.57
-1.61
-1.28
-0.97
-0.67

- 4.38
-4.07
- 3.82
-3.50
-1.49
-1.16
-0.87
-0.54

-4.26
- 3.96
- 3.68
- 3.35
-1.34
-1.04
-0.77
-0.43

TABLE VI.A
PERCENTAGEPOINTSOF THEASYMPTOTICDISTRIBUTIONOF T(& -1) IN MODEL C
Time of Break Relative to Total Sample Size: X
0.1

A=

1%
2.5%
5%
10%
90%
95%
97.5%
99%

36.17
30.65
26.63
22.68
-4.74
- 3.41
- 2.51
- 1.31

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

- 39.97
- 34.92
-29.95
-25.50
- 5.85
-4.34
- 3.19
-2.14

-42.98
- 36.48
- 32.47
-27.90
- 7.35
- 5.50
-4.14
- 2.82

-45.52
- 37.12
-33.22
-29.39
- 8.43
-6.67
- 5.37
- 3.96

-44.07
- 37.56
- 33.79
-29.41
- 8.84
-7.19
- 5.82
-4.39

-44.75
- 37.72
- 33.19
-29.04
-8.55
-6.79
- 5.47
-4.24

-43.02
- 37.50
-33.11
-28.14
-7.41
- 5.66
-4.33
-2.80

-41.48
- 35.16
- 30.70
-25.79
-6.17
-4.52
- 3.35
- 2.02

- 36.58
- 31.82
-27.16
-22.62
-4.89
- 3.52
- 2.49
-1.28

TABLE VI.B
PERCENTAGEPOINTSOF THEASYMPTOTICDISTRIBUTIONOF ta IN MODEL C
Time of Break Relative to Total Sample Size: X
A-

1%
2.5%
5%
10%
90%
95%
97.5%
99%

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

- 4.38
-4.01
- 3.75
- 3.45
-1.44
-1.11
- 0.82
-0.45

- 4.65
-4.32
- 3.99
- 3.66
-1.60
-1.27
-0.98
-0.67

-4.78
-4.46
-4.17
- 3.87
-1.78
-1.46
-1.15
-0.81

-4.81
-4.48
-4.22
-3.95
-1.91
-1.62
-1.35
-1.04

-4.90
-4.53
-4.24
- 3.96
-1.96
-1.69
-1.43
-1.07

-4.88
-4.49
-4.24
-3.95
-1.93
-1.63
-1.37
-1.08

-4.75
-4.44
-4.18
- 3.86
-1.81
-1.47
-1.17
-0.79

-4.70
-4.31
-4.04
- 3.69
-1.63
-1.29
-1.04
-0.64

-4.41
-4.10
- 3.80
- 3.46
-1.44
-1.12
-0.80
-0.50

absolute value) for each model than the standardDickey-Fullercritical values

when considering the left tail. One would thereforeexpect a loss in power.
Secondly,althoughthe criticalvaluesare not significantlyinfluencedby the value
of the parameterX, the maximum(in absolutevalue) occurs aroundthe value
X = 0.5, i.e., for a break at mid-sample.6In the left tail, critical values of the
statistics are smallest(in absolutevalue)when X is close to 0 or 1. This is to be
6The simulated critical values suggest that the limiting distributions

are symmetric around X = 0.5.

This feature seems intuitively plausible. We have not, however, proved that such is the case.

1378

PIERRE PERRON

expected since, as previouslymentioned,the criticalvalues are identicalto those

of Dickey and Fullerwhen X= 0, 1.
Some criticalvaluesare worthnoticing.Considerthe t statistics.Under Model
(A), the "crash hypothesis",the 5 percent criticalvalue has a minimum(over
values of X) of -3.80. Under Models(B) and (C), the correspondingfiguresare
- 3.96 and -4.24 respectively.The criticalvalues under the variousmodels are
therefore noticeably smallerthan the standardDickey-Fullercritical value of
- 3.41 (see Dickey (1976) and Fuller(1976)).
These sets of results can be used to performhypothesistesting. One simply
picks the critical value correspondingto the sample value of X at the chosen
significancelevel. Sincewe only providecriticalvaluesfor a selectedgrid of X's,
the proceduresuggestedis to choose the criticalvaluecorrespondingto the value
of X nearest to its sample value, i.e., TB/T. Given that the differencesin the
critical values over adjacentvalues for X in the tables are not substantially
different,this procedureshouldnot cause misleadinginferences.
4.1. Extensions to More GeneralError Processes: Case (1)

Using regressions(5) (i = A, B, C) and the criticalvaluesin Tables IV, V, and

VI is valid only in the case wherethe innovationsequence{e, } is uncorrelated.
When there is additional correlation,as one would expect, an extension is
necessary.Two approachesare possible.One is to follow the approachsuggested
by Phillips (1987) and Phillipsand Perron(1988). This involvesfindinga set of
transformedstatistics that would convergeweakly to the limiting distributions
expressedin Theorem2 with a2 ae2. The other approachis that suggestedby
Dickey and Fuller(1979) and Said and Dickey (1984).
Consider first the extensionto Phillips'(1987) procedure.It is useful first to
write the limiting distributionsof Theorem2 in a different,more compactform.
To do so, we adopt the frameworksuggestedby Ouliaris,Park, and Phillips
(1988). Define wi(r) (i = A, B, C) to be the stochasticprocesson [0,1] such that
wi(r) is the projection residual of a Wiener process w(r) on the subspace
generated by the following set of functions: (1) i = A:1, r, du(r), where du(r) = 0
if r s X and du(r) = 1 if r > X; (2) i = B: 1, r, dt*(r), where dt*(r) = 0 if r < X
and dt*(r) = r- X if r > X; (3) i = C: 1, r, du(r), dt(r), where dt(r) = 0 if r < X
and dt(r) = r if r > X. Adoptingthis notation, an alternativerepresentationof
the limiting distributionsin Theorem2 is given by:
T(&a- 1) (

wi(r)dw(r)

+ 8)(wi

tai~ (a/ae)(fl

wi(r)dw(r)

(r)dr )

(i =A, B, C)

dr)

(i =A, B, C)

and

where 8 = (a2 - a2)/(2a2).

8)

wi(r2

1379

UNIT ROOT HYPOTHESIS

Now, define 67 and 67ito be, respectively,any consistentestimatorof a2 and

a based on the estimated residuals from regression (5)(i

A, B, C).7Also define

S12 (i-A,

B, C) to be the residualsum of squaresfrom the regressionYt-i on (1)

i=A: 1,t,DU,; (2) i=B: 1,t,DTt*; and (3) i=C: 1,t, DUt,DTt. We then
define the transformedstatisticsas:
(6)

Z(a=i)

(7)

Z(tei)

(8)

Z(a)

1)- T2( i

T(ai-

(i=A,

ife6)/2S7

B, C),

T( -i2-6e)/2i1Si
(i = A, B, C).
Following Ouliaris,Park, and Phillips (1988), it is straightforwardto show
that:
( -ei/Ii) tji

(fwi(r)dw(r))(flw,(r)

2dr)

(i =A, B, C)

)2

(i = A, B, C).

and
(9)

Z(tai)

(
wi (r dw) fr))

w(

dr

The limiting distributionsin (8) and (9) are those whose critical values are
presented in Tables IV, V, and VI derived using the representationgiven by
Theorem2.
The other approachadopts the proceduressuggestedby Dickey and Fuller
(1979, 1981) and Said and Dickey (1984) which add extra lags of the first
differencesof the data as regressorsin equation(5). This extendedframeworkis
characterizedby the followingregression(againestimatedby OLS):
k

(10)

= diyt
&YYt

E
j=l

EijdYt'i- + it

(i

= A, B, C)

where
In the above representation,axis the OLS estimatorof a, the sum of the
autoregressivecoefficientsand the test is again that a = 1. The parameterk
specifiesthe numberof extraregressorsadded.In a simpleAR(p) process,k =p.
In a more generalARMA(p, q) processwith p and q unknown,k must increase
at a controlledrate with the samplesize. Argumentssimilarto those developed
by Said and Dickey can be used to show that the limiting distributionsof the
statistics tai (i = A, B, C) are the same when the innovation sequence is an
ARMA(p, q) processand regression(10) is used, as they are when the errorsare
i.i.d. and regression(5) is used. However,slightly more restrictiveassumptions
are needed with respect to the innovation sequence { et} and the truncation
parameterk for this asymptoticequivalenceto hold. They are detailedin the
followingAssumption(see Said and Dickey (1984)).
7SeePhillips(1987),Phillipsand Perron(1988),and Perron(1988)for details.

1380

PIERRE PERRON

ASSUMPTION 2: (a) A(L)e =B(L)v,; (b) v, is a sequence of i.i.d. (0,a2)

random variables with finite (4 + 8)th moment for some 8 > 0; (c) k -- oo and
T-'k3

-*

0 as T

oo.

4.2. Extensions to More GeneralError Processes: Case (2)

A possible drawbackof the methodssuggestedabove is that they imply that

the changein the trendfunctionoccursinstantaneously.Given,for instance,that
the Great Depressionwas not an instantaneousevent but lasted severalyears,
one may wish to allow for such a transitionperiodduringchangesin the trend
function. One way to model this is to supposethat the economyreactsgradually
to a shock to the trend function.8Consider,for instance, Model (A) where a
crash occurs.A plausiblespecificationof the trendfunction,say A, is given by:
(11)

(L)DUt

As=yi+t+y

where {'(L) is a stationaryand invertiblepolynomialin L with A(0) = 1 and

i=
2 - 1l The long run changein the trend functionis given by y4(l) while
the immediate impact is simply y. A similarframeworkholds for models (B)
and (C).
One way to incorporatesuch a gradualchange in the trend function is to
suppose that the economyrespondsto a shock to the trend function the same
way as it reacts to any other shock, i.e. to impose 4i(L) = B(L)-1A(L) (see
Section2). In the literatureon outliersspecificationthe frameworksuggestedhere
is analogous to the so-called "innovationaloutlier"model whereasthe framework consideredin Section4.1 is analogousto the "additiveoutlier"model (see,
e.g., Tsay (1986)).We can then implementtests for the presenceof a unit root in
a frameworkthat directlyextendsthe Dickey-Fullerstrategyby addingdummy
variablesin regression(1). The followingregressions,correspondingto Models
(A), (B), and (C) are constructedby nestingthe correspondingmodelsunderthe
null and alternativehypotheses:
k

(12)

yf= ,uA +

ADU+

'At + JAD(TB)t

+ay

At

+
i=l
k

(13)

+ 6BDU+

y=,i

yAD J;*+&ByAB

1+
i=l

(14)

CDU+

vYtP +A
k

+ 2 C^At_i

Ct+ yCDT +dCD(TB), + ACy

+ et,

i=l

The null hypothesisof a unit root imposesthe followingrestrictionson the true

parametersof each model: Model (A), the "crashhypothesis":aA =1,13A = 0,
8

Again, this treatmentis analogousto the methodologyproposed by Box and Tiao (1975)
concerninginterventionanalyses.

1381

UNIT ROOT HYPOTHESIS

B = 0;
0; Model (B), the "breaking slope with no crash": a' = 1, yB =0,
and Model (C), where both effects are allowed: aC = i, yC= 0, cC=0. Under

OA =

the alternativehypothesisof a "trend stationary"process, we expect aA,

aeC < 1; #A,

[B3,

# 0;

OA,

C, yB,yC

aB,

# 0. Finally, under the alternative hypoth-

esis, dA, dC, and OB shouldbe close to zero whileunderthe nul hypothesisthey
are expected to be significantlydifferentfromzero.
The asymptoticdistributionof the t statisticstjA and t C in (12) and (14) are
the same, respectively,as the asymptoticdistributionof t,A and taC in (10).
However such a correspondencedoes not hold for the t statistic teB in (13).
Apart from the one-timedummyvariabled(TB)t, regressions(13) and (14) are
equivalent;hence the asymptoticdistributionof taB is identical to the asymptotic distributionof taC. This implies that in the above framework,it is not
possible to test for a unit root under the maintainedhypothesisthat the trend
function has a changein slope with the two segmentsjoined at the time of the
change.Consequently,tests for the presenceof a unit root in Model (B) will have
less power using regression(13) than using (10) where the asymptoticcritical
values of the t statisticon a are smaller(in absolutevalue). However,it is still
possible to test for a unit root with constant drift against a trend stationary
process as in Model (B) and use the criticalvaluesof Table V. One simplyruns
the regression:
k

(15)
(15)

= 1B +fBt+
y~=AB
+:t+y
Xt

A?BDT*
T+ +&5AtA+
a Yt1+Eci

t
y-i+et.

i= 1

The asymptoticdistributionof the t statistic t&Bin (15) is the same as the

asymptoticdistributionof the t statisticteB in (10). Given that the regressorDUt
is absent from (15), this case, however,implies that the change in drift is not
permittedunder the null hypothesis.
Finally, note that it is possible to apply Phillips' nonparametricprocedure
regressors
using regressions(12) through(14) withoutthe laggedfirst-differenced
and applyingthe correctionsgivenby (6) and (7). However,such a procedurehas
the unattractivefeatureof imposingonly a one-periodadjustmentto the change
in the trend function. In the notation of (11), it imposes A(L) = 1 - 41L where

A is the coefficienton the firstlag in the polynomialB(L)-1A(L).

The proceduresoutlinedin this sectionpermittestingfor the presenceof a unit
root in a quite generaltime seriesprocesswhich allows a one-timebreakin the
mean of the series or its rate of growth(or both). In the next section,we apply
these proceduresin the specificcontext of breaksat the time of the 1929 crash
and the 1973 oil shock.
5. EMPIRICALAPPLICATIONS

We apply the test statisticsderivedin the previoussection to the data set used
by Nelson and Plosserand to the postwarquarterlyreal GNP series.The data set
considered by Nelson and Plosser consists of fourteen major macroeconomic
series sampled at an annualfrequency.We omit the analysisof the unemploy-

1382

PIERRE PERRON

ment rate series given that it is generallyperceivedas being stationary.The

sample varies for each series with a starting date between 1869 and 1909.
However,each series ends in 1970. Given that we entertainthe hypothesisthat
only the 1929 GreatCrashand the 1973 oil price shockcauseda majorchangein
trend function,each seriesin this data set containsonly one break.It is therefore
possible to apply the tests described in the previous section. Similarly,the
quarterlypostwar real GNP series contains a single break as the sample goes
from 1947:1 to 1986:11L.FollowingNelson and Plosser,we consider the logarithm of each seriesexceptfor the interestrate for whichwe use the level.
Of the thirteenseriesin the Nelson-Plosserdata set that we analyze,preliminary investigationsshowed that eleven were potentiallywell-characterized
by a
trend function with a constant slope but with a major change in their level
occurringright afterthe year 1929. For these series,the maintainedhypothesisis,
therefore, that of Model (A) and given that the Great Crash did not occur
instantaneouslybut lastedseveralyears,we applyregression(12) to carryout our
testing procedure.The two series that were not modeled as such are the "real
wages" and "commonstock price"series.For these series,it appearedthat not
only a change in the level occurredafter 1929 but therewas also an increasein
the slope of the trendfunctionafter this date. For these reasons,the maintained
hypothesis is that of Model (C), and we use regression(14) to implementour
tests.
The postwarquarterlyreal GNP seriesoffersyet a differentpicture.The 1973
oil price shock did not causea significantdropin the level of the series.However,
after that date, the slope of the trend function has sensibly decreased.This
phenomenonis consistentwith the much discussedslowdownin the growthrate
of real GNP since the mid-seventies;see, for example,the recentsymposiumin
the Journal of Economic Perspectives (1988). For these reasons, the maintained

hypothesisis that of Model(B). Given the inherentdifficultyin testingfor a unit

root allowing lagged effects for the change in the trend function, we apply
regression(10) (i = B) to carrythe testingprocedure.Modelingthe changein the
trend function following the 1973 oil price shock as instantaneous,at least
appearsmore plausiblethan the changethat occurredduringthe Great Depression.

Table VII presents the correspondingestimatedregressionsfor each series

along with the t statisticon the parametersfor the followingrespectivehypotheses: IL= 0, B= 0, 0 = 0, y = 0, d= 0, and xx= 1. Recallthat underthe hypothesis
of a unit root process ,L 0 (in general), ,B= 0, 8 = 0 (except in regression
Models (C)), y = 0, d # 0, and a = 1. Under the alternativehypothesis of
stationary fluctuationsaround a deterministicbreakingtrend function: ,t # 0,
9#0, 8 0 , y 0(in general),d=0, and a< 1.
The value of k chosen is determinedby a test on the significanceof the
estimated coefficientsci. We actually used a fairly liberal procedurechoosing
a value of k equal to say k* if the t statistic on c, was greaterthan 1.60 in
absolute value and the t statistic on cl for 1> k* was less than 1.60 (with a
maximum value for k of 8, except for the postwarquarterlyreal GNP series

1383

UNIT ROOT HYPOTHESIS

TB=
NOTE:
a,
1973:I
Quarterly
b,

real

and
c

prices

statistical
0.66

A
6.977
significance
at
t,,

the

1160.51

level

(10).

0.41
0.59
8 1 k

Model
A
B;
0.353
2.115
y,
tAi,
+
?t

4.33
4.09
-

6
1.051
-0.190
yDT7* +

fi

2.5%,

5%

71 100
Regression

0.0087
to

97.73
-

+y;
y

3.71
-4.29

0.0107
0.0070

fi

0.0031 +
I
respectively.
_l

0.86

(14),

Regression

#DU,
+

fit
+

jDT,
+

0.0066
0.0139
t

3.33
3.98

-0.088
0.932
-

0.031
0.128

3.98c
0.78
0.76

td
&

t&

-4.28c
-4.87b

=
4

-0.343
-0.071
-0.098
0.005
-0.046
-0.298
-0.190
-0.102
-0.360
-0.189
-0.004
t6

-0.21
-2.65
-3.16
-4.58
-2.76
-4.28
-2.06
-2.59
-4.32
-0.20
-4.77
fi

0.0121
0.0111
0.0323
0.0105
0.0005
0.0197
0.0359
0.0267
0.0057
0.0002
0.0070
to

4.01
-0.35
1.75
5.05
4.18
2.64
5.37
4.26
5.42
4.00
5.44
0.033
0.085
-0.095
0.136
-0.025
-0.036
0.197
0.026
0.018
-0.070
0.100
td

2.01
0.53
0.68
1.36
-0.77
0.64
-0.79
-0.99
-1.09
1.09
-0.30
&

0.471
0.941
0.531
0.976
0.978
0.776
0.667
0.322
0.812
0.619
0.282
- - -

t&

S(e)

0.1402
0.0330

Model
A;
y,
,u
+
6DU,
+
fit
+

-0.45
-1.66
-1.28
-5.41a
4.51a
5.03a
-4.04b
-5.42a
5.47a
4.29b
4.09b

S(e)

0.0663
0.0445
0.0295
0.0875
0.0555
0.0438
0.0509
0.2787
0.0694
0.0440
0.0532

TESTS
FOR
A

TABLE

UNIT
VII

dD(TB), ROOT
+
6y,tt
+
I
-

d
AYE
+

0.298
0.718

(12),

&y,-

I-

S(e)

0.41
0.41
0.63
0.63
0.33
0.33
0.33
0.59
0.49
0.49
0.49
2 0 6 7 2 5 7 8 7 8 8 k

u
Model
2.38
C;
0.301
3.441
-0.018
0.065
3.325
0.050
5.692
0.669
3.402
0.120
y,
=
A
t'A
+
5.45
4.11
5.07
4.37
4.72
1.12
4.09
4.54
5.44

0.0097

(a)
X

dD(TB),

ta

71 10282 71 11182 81 11162 62 62

t'A

Y
+

GNP
production

Regression

4.43
3.79

t1

12.06

GNP Real Real

Wages
Money
Interest
1929
Velocity
perNominal
Industrial
Consumer
GNP
rate stock
Employment
deflator
GNP
capita
prices

(b)

X
k

10

and

1929
Common
wages

stock

GNP
159

TB=

(c)

denote

1%,

TB
=

Real

Ay,
+

1384

PIERRE PERRON

where we used a maximum of 12). This liberal procedure is justified in the sense
that including too many extra regressors of lagged first-differencesdoes not affect
the size of the test but only decreases its power. Including too few lags may have
a substantial effect on the size of the test.
Consider first the series for which we applied Model (A). To evaluate the
significance of the t statistic on a. we use the critical value presented in Table
IV.B with a value of X closest to the ratio of pre-break sample size to total
sample size. Of the eleven series in that group, the unit root hypothesis cannot be
rejected even at the 10 percent level for three of them: "consumer prices,"
"interest rate" and "velocity." However, we can reject the null hypothesis of a
unit root at that 2.5 percent level or better for all other eight series. We can reject
it at the 1 percent level for the following series: "real GNP," "nominal GNP,"
"industrial production," "employment," and "wages," and at the 2.5 percent
level for the series "real per capita GNP," "GNP deflator," and "money stock."
In some cases the coefficient a, which is an estimate of the sum of the
autoregressive coefficients, is dramatically different from one. For example, it is
0.282 for "real GNP" and 0.322 for "industrial production."
Given that the unit root hypothesis can be rejected for the eight series
mentioned above, we can assess the significance of the other coefficients using the
fact that the asymptotic distribution of their t statistic is standardized normal. In
all cases, the estimated coefficients on the constant (,i), the post-break dummy
(0), and the trend (/B) are significant at least at the 5 percent level. All series
showed a trend function with a positive slope and a significant decrease in level
just after 1929. For these eight series the coefficient on the break dummy (d) is
not significant. These results strongly suggest that, except for the "consumer
price," "velocity," and "interest rate" series, the underlying process is one of
stationary fluctuations around a deterministic trend function.
Consider now panel (b) of Table VII which presents the results for the
"common stock price," and "real wages" series estimated under Model (C). We
can reject the null hypothesis of a unit root at the 2.5 percent level for "common
stock prices" and at the 5 percent level for the "real wages" series. In both cases,
the constant (,i), the post-break constant dummy (0), the trend (/B), and the
post-break slope dummy (y) are highly significant, while the break dummy (d) is
not. The coefficients a for the "real wages" series is very low at 0.298 while for
the "common stock price" it is at 0.718 showing substantial mean reversion
effects. This finding about the "common stock price" series is particularly striking
given the vast amount of theoretical and empirical studies supporting the random
walk hypothesis in this situation.
Finally, panel (c) of Table VII presents the results for the postwar quarterly
real GNP series using regression (10) corresponding to Model (B). In this case,
the null hypothesis that a = 1 can be rejected at the 5 percent level with an
estimated coefficient 5 equal to 0.86. This result is especially significant given the
usual poor power properties of tests for a unit root against stationary alternatives
when using a data set with a small span sampled frequently (see, e.g., Perron
(1987) and Shiller and Perron (1985)). The other estimated coefficients in panel
(c) confirm the relevance of the "trend stationary" model versus the "unit root"
model. The coefficient on the post-break slope dummy coefficient (y) is highly

UNIT ROOT HYPOTHESIS

1385

significant. The estimated regression is therefore consistent with an underlying

process characterized by stationary fluctuations around a deterministic trend
function with a decrease in the slope after 1973.9
Table A3 in Appendix B presents the estimated value for the sum of the
autoregressive coefficient, a, and its t statistic for the null hypothesis a = 1, for
all values of the truncation lag parameter k between 1 and 12. In general, the
results are quite robust to which value of k is selected. 10
The results presented in this section are quite striking. The unit root hypothesis
can be rejected for all but three series. To obtain these results, only a rather weak
postulate needed to be imposed, namely the presence of a one-time change in the
trend function. We claim that this is a weak postulate for the following reasons.
As shown by Nelson and Plosser and Campbell and Mankiw, all the series
analyzed have a unit root if the trend function is not allowed to change. This
view implies that the 1929 crash is simply one big outlier in the innovation
sequence. On the other hand, it also implies that the post-1973 growth slowdown
is a succession of smaller innovations and that the mean of the innovations is
different for the pre-1973 and the post-1973 period. These alternative interpretations are, we think, less appealing than the hypothesis of a break in the trend
function, especially given that we allow such a break under both the null and
alternative hypotheses.
Given that the "consumer price," "velocity," and "interest rate" series appear
to be characterized by the presence of a unit root, it seems worthwhile to see if
this feature is stable in the pre and post 1929 samples. The subsample DickeyFuller type regressions are presented in Appendix B for values of k between 1
and 12. Consider first the "consumer price index." With k = 2, the estimated
value of the sum of the autoregressive coefficients, &-is 0.965 for the pre-1929
sample with a t statistic of - 1.28. We therefore cannot reject the null hypothesis
of a unit root for this subperiod. However, for the post-1929 sample, the picture
is rather different; when k = 7, & has a value of 0.704 with a t statistic of -4.56
significant at the 1 percent level. The nonrejection of the unit root hypothesis
using the full sample is due to the pre-1929 sample. After 1929, the unit root is no
longer present.
The case for the "velocity" series has a special feature. It has been well
documented that the U.S. velocity series declined steadily until 1946, remained at
a fairly constant level until 1970, before increasing (see, for example, Poole (1988)
and Gould and Nelson (1974)). For this reason, we discuss applications of the
standard Dickey-Fuller procedure for the following three samples: 1869-1929,
1930-1945, 1946-1970. We cannot reject the unit root hypothesis with the
pre-1929 sample even though & is only 0.865 (k = 0). However, the picture is

coefficienta and its t statisticwere

9Basically,the sameestimatesof the sumof the autoregressive
obtainedusing regressions(13) and (14), for all valuesof k, showingsome robustnessfor the results
presented.We also appliedPhillips'nonparametric
procedureto the detrendedseries(equations(6)
and (7), i = B). These test statisticsdid not allow the rejectionof the null hypothesisof a unit root.
10Onenotableexceptionis the quarterlyrealGNP serieswherethe t statisticis significantat the 5
percentlevel with k = 2 or k = 10. It is significantat the 2.5 percentlevelwith k = 11, witha valueof
- 4.32. We choose to reportthe resultof k = 10 becausethe 10th laggedfirst-difference
was highly
significant(t statisticof 2.29)whilethe 11th and 12th lags werenot.

1386

PIERRE PERRON

dramaticallydifferentusing post-1929samples.For the period 1930-1945, with

k = 1, a is estimated at

0.011 with a t statistic (for testing a = 1) of

3.44

which is significantat the 5 percentlevel.Thoughthis resultshouldbe takenwith

caution due to the small number of observations,it is suggestiveof a quite
differentbehavior.The period1946-1970 affordsa modest amountof additional
informationand yields still more dramaticresults:with k= 4, a is equal to 0.00
with a t statistic(for a = 1) of - 5.86, significantat the 1 percentlevel using the
Dickey-Fullercriticalvalues.It thereforeappearsthat the "velocity"and "consumer price"seriesyield similarresults:the presenceof a unit root before 1929
but not after.
The results for the interestrate seriesindicatethat, in this case, the unit root
hypothesis cannot be rejectedat usual significancelevels for both subsamples
(even though a is estimatedat 0.540 (k = 3) with the pre-1929sample).Indeed,
given our previousresults,we can concludethat only the "interestrate"seriesis
characterizedby the presenceof a unit root after 1929. All the other series are
better construedas stationaryfluctuationsarounda deterministictrendfunction
for this period.
6. DISCUSSIONAND CONCLUDING COMMENTS

When testing for the presenceof a unit root in a time seriesof data againstthe
hypothesis of stationaryfluctuationsarounda deterministictrend function,the
use of a long span of data has definite advantages.It allows tests with larger
power comparedto using a smallerspan, in most cases even if the latter allows
more observations (see Shiller and Perron (1985) and Perron (1987)). The
drawback,however, is that a data set with a large span has more chance to
include a major event which one would rather consider as an outlier or as
exogenousgiven its relativeimportance.The argumentsin this paperrest on the
postulatethat two sucheventshave occurredin the 20th century:the 1929 Great
Crash and the slowdownin growth after the oil shock of 1973. We therefore
considered, as a relevant alternative,a trend function with a change in the
interceptin 1929 and a changein the slope after 1973.
Let us discuss,in moredetail,whatare the relevantissuesin drawingparticular
conclusions about the nature of economic fluctuationsfrom our results. It is
particularlyimportantto put our resultsinto perspectiveand also highlightwhat
has not been shown.
The first importantissue to point out is that we have not provideda formal
unconditional statistical model of the time series properties of the various
aggregates.A rejectionof the null hypothesisof a unit root conditionalon the
possibility of shifts in the underlyingtrend function at known dates does not
imply that the variousseriescan be modeledas stationaryfluctuationsarounda
completelydeterministicbreakingtrendfunction.As a matterof generalprinciple, a rejectionof the null hypothesisdoes not imply acceptanceof a particular
alternativehypothesis.However, since the tests were designed to have power
against a specificclass of alternativehypotheses,it is useful to look amongclose
membersof that class to proposean interestingstatisticalmodel for the various

1387

UNIT ROOT HYPOTHESIS

aggregates.Only with such a model is it possible to provide forecasts with

appropriatestandarderrors.
We certainlydo not entertainthe view that the trend function includingits
changes are deterministic.This would imply that one would be able to forecast
with certaintyfuturechanges.This is indeedquite unappealing.Whatwe have in
mind in specifyingour class of maintainedhypothesescan be parameterizedas
follows:
(16)

Yt ,= Qt + zt,

qt

At + 9tt,

where A(L)Zt = B(L)et; et,i.i.d.(O,

a2);

tt=pt-,

+ V(L)vt, and

f = fit-, +

W(L)wt. Here, the Zr'sare (not necessarilystationary)deviationsfrom the trend

function

qt.

The intercept and the slope of the trend functions, ti and Pt, are

themselvesrandomvariablesmodeledas integratedprocesseswith W(L), V(L)

stationaryand invertiblepolynomials.However,the importantdistinctionis that
the timing of the occurrenceof the shocks vt and wt are rare relative to the
sequence of innovations{et}, for example,poisson processeswith arrivalrates
specified such that their occurrencesare rare relative to the frequencyof the
realizationsin the sequence{ et}. The intuitiveidea behindthis type of modeling
is that the coefficients of the trend function are determinedby long-term
economicfundamentals(e.g., the structureof the economicorganization,population growth, etc.) and that these fundamentalsare rarely changed. In our
examples, vt is nonzeroin 1929 (the greatdepression)and w, is nonzeroin 1973
(the oil price shock).
In this sense, our exogeneity assumptionabout the changes in the trend
functionis a device that allowstakingthese shocksout of the noise functioninto
the trend function without specific modeling of the stochastic nature of the
behaviorof It and 8t. It is in this sense that our approachdoes not providean
unconditionalrepresentationof the time seriespropertiesof the variousvariables.
Estimationof modelsof the form (16) by specifyinga probabilitydistribution
for the error sequences {et,wt,
is clearly an important avenue of future
vtw
research. Interesting recent advances on this topic have been provided by
Hamilton(1987) and Lam(1988)wherethe slope of the trendfunctionis allowed
to take two differentvaluesand the changesare modeledas a binomialprocess.
However,no methodsare currentlyavailableto test whetherZt is integratedor
not in this framework.11
Problemsin estimationof models of the form (16) are
furthercompoundedby the fact that, accordingto our view, only one nonzero
realizationof both v, and wt would be presentin the data set typicallyavailable
for the series of interest.
In the above framework,the purposeof this paperis to test whetherZt is an
integratedprocess or not, i.e. to test whetherthe shocks {et} have persistent
effects that do not vanishover a long horizon.Our approachis to removefrom
the noise function two events that occurredat two dates where we believe
positive occurrencesof the shocks { vt,w)} happenedand to model them as part
"1Bothauthorsstudiedthe behaviorof the postwarquarterlyreal GNP with possibleshiftsin the
slope of the trendfunction.Hamilton(1987)imposesZt to be an integratedprocesswhileLam(1988)
leaves Z, unconstrained.

1388

PIERRE PERRON

of the trendfunction.The fact that we model thesechangesas exogenousimplies

that our results are conditional.That is, conditional upon the presence of a
change in the trend function in 1929 and 1973, the fluctuationsare transitory
(i.e.,

Zt

is stationary).

An importantdirectionfor futureresearchis to make this conditionalresult

into an unconditionalstatement.This could,in principle,be achievedby a direct
test for structuralchangein the trendfunction.In a sense, our procedureallows
such a test, but conditionalon a changeoccurringat a fixed knowndate. Hence,
problemsof pre-testingand "datamining"could be raisedregardingthe role of
looking, ex-post, at the data on the choice of the date.12Accordingly,what is
needed is a test for structuralchangesin the trendfunctionoccurringat unknown
dates. The problem,however,is that care must be appliedto ensurethat the test
has an adequatesize underboth the unit root and trend-stationaryhypotheses.
No such test is currentlyavailablein the literature.We hope to report,in the near
future,developmentsin this areaand applicationsin this context.
However, an importantissue of observationalequivalencecould not even be
settled by such a formal test for structuralchange. Consider,for instance, the
following limiting case in the crash model.13 A trend-stationarymodel with a
break and where the errorshave zero varianceis observationallyequivalentto a
unit root model with drift wherethe errorshave a high probabilityof being zero
but are occasionallynonzero and finite. In general,when the varianceof the
errorsis nonzero,the two modelswill be nearlyobservationallyequivalentwith
the disturbancesin the unit root model having fat tails. We are able to make a
distinction in our empiricalresult throughthe mixing conditions(see Assumptions 1 and 2) which prohibit fat tailed disturbances.Any formal test for
structuralchangewould,presumably,also have to impose some mixingcondition
prohibiting fat-tailed disturbances,thereby not resolving this issue of nearobservationalequivalence.
In fact, any test for the presence of a unit root against trend-stationary
alternativesis subjectto anothertype of observationalequivalence,as recently
argued by Cochrane(1987) and Blough (1988). Indeed, in finite samples, any
trend-stationaryprocess is nearly observationallyequivalent to a unit root
process with a strong mean-reversioncomponent,i.e. where the errorshave a
moving-averagecomponentwith a root nearminusone.'4 The fact that we reject
the unit root hypothesisexcludingthe eventof 1929 suggeststhatif thereis a unit
root at all the correlationstructureof the innovation sequence must exhibit
substantialmean reversion.
12See, for example,Christiano(1988).
1This issue and its followingillustrationwereraisedby a referee.
14PThisobservationalequivalenceproblemonly disappearsasymptotically.
In otherwords,in finite

samples,any test for a unit root with ARMA errorsshouldhave zero power.Formally,the critical
values should be determinedsuch that the test has a given fixedsize over all possiblevalues of the
nuisance parameters(here, the additionalcorrelationin the errors).Given the near observational
equivalence,any suchtest wouldhavezeropowerby definition.The unitroot testscan be rationalized
by arguingthat we are willing to have the wrong size over some of the parameterspace'exactly
becausefor all practicalpurposesit does not matterwhetherwe label a seriesas trendstationaryor
differencestationarywith a strongmean-reversion
component.

1389

UNIT ROOT HYPOTHESIS

To sum up: trendstationaryprocesseswith a breakare nearlyobservationally

equivalent to unit root processes with strong mean-reversionand a fat-tailed
distributionfor the errorsequence.Whicheverview one adoptscannotbe decided
by data alone. Nevertheless,the pictureunderany of these views is basicallythe
same: shocks had little, if any, persistenceeffectover a long horizon.Only those
associatedwith the GreatDepressionand the oil price shock significantlyaltered
the long run behaviorof the series.
While choosing one view over the other is a matter of convenience for
interpretingthe data, it has profoundimplicationsfor a multitudeof statistical
procedures.Indeed,underthe unit root view one must ensurethe validityof the
procedures under fat-tailed disturbances,and at the moment very few are
appropriatein a time seriescontext. Hence for all practicalpurposes,it may be
more advantageousto adopt the trend-stationary
view with breaksand detrend
our series accordinglypriorto analyzingthe remainingnoise.
Department of Economics, Princeton University, Princeton, NJ 08544, U.S.A.;
and Centre de Rechercheet De'veloppementen Economique, Universitede Montreal,
C.P. 6128, Succ. A, MontrJal, Canada, H3C - 3J7.
Manuscriptreceived October,1987; final revisionreceived February,1989.

APPENDIX A
The conditions imposed by Assumption 1 permit us to use a functional weak convergence result
due to Hermdorf (1984). Let St = tle, (So = 0) and define the following variable lying in the space
D[0, 13:
XT(r)=-u1T-'/2S[Trl=r'J7T-1/2
XT (1)-

_l,

(j-l)/Ts

r<j/T

( .j

T)

1T-1/2ST.

Hermdorf's theorem states that under the conditions of Assumption 1, XT(r) =* w(r) where
denotes weak convergence to the associated probability measure and w(r) is the unit Wiener process
defined on C[0,1], the space of all continuous functions on the interval [0,11. Following Phillips
(1987) and Phillips and Perron (1988), it is easy to derive the following lemma related to functions of
St. These results will be used in proving both Theorems 1 and 2.
LEmMAA.1: Let St = EX.le(So =0) and assume that the innovation sequence { e,t} satisfies the
conditions of Assumption 1. Furthermore,let TB= XTfor all T; then as T-3 oo:
IB
B

(a)

T-3/2

(b)

TB
T 2 jS12 =2fw(r)2

(c)

T- 5/2

(d)

T 3/2 >je2e

(e)

Si

af w(r) dr;

dr;

TB

EA Crfrw(r) dr;
TB

a Xw(X) -

w(r) dr);

5?S.J_ej
I

(s /2)(w(1)

1390

PIERRE PERRON

The results in (a) through (d) are simple extensions of those in Phillips (1987) and Phillips and
Perron (1988); part (e) is proved in Phillips (1987). The previous results can be recovered by simply
letting X = 1 in which case TB= T. For example, to prove part (a):
TB

TB

S =a

T3/29

T l/2a -S[Trjdr=a

fi/

xXT(r)dr

=*J

w(r)dr

using Herrndorf's weak convergence result and the continuous mapping theorem (see, e.g., Billingsley
(1968)). The proofs of (b), (c), and (d) are entirely analogous.
PROOF OF THEOREM1: We consider first the most general model (C) in which both ,B (the slope)
and ,u (the intercept) are allowed to change, i.e.:
Yt =IL + 81t + et
Y=2+

(A1)

( 2

#1

(0
)TB

,B2t+

< t A TB)

(TB < t < T).

et

The following lemma provides a convenient representation of the sample moments of {Yt
}y
LEMMAA.2: Assume that {Yt)} is generated according to (A.1) with the innovationsequence {et,
satisfying Assumption 1; then,
T

(a)

Ety,I

= (1/6)[,81(3X - A3) +f,2(2 - 3X + 3)]T3

+( l4L

2+ A2(1 -2)

? [(3/2)jLX

- X)f2

+ (3/2)(I

+ [T-3/2

+ X(1 - X)(,81 -,2)]T2

- 12-

(1/3)#,X

- (1/3)#2(l-

X)] T

+Op()

te,ij]T3/2

(b)

Eyt-=

2
(1)[#(X2-2X)

(I + 2A-3A?)]T2

+ [XAI+ (1 - X)js2 + (3/2)#1X - (')p2(1

+ 3A)]T

(c)

Eyt2l=

(113)[,#2A+(3 2/ + [lAJX2

+ F22(1

X)

l,2+ (1/ _ A)2] T3

_ 3(- +
?)3

22(1

+)-2,2(_

-_A2)_(-)(A
T

+2

X(/pl-#2)T-

TB

et + #T-3/2

1/
TRB+1

+[

?A)2)]Te2

tet - 2T -3/2

E: tet T3/2
TB+ 1

+ (1/6),822 (1 - A)
21+ (1/6)#B2A?+ ,ul,BX + (1 - XA),22

+y2R(1-A)-2y2

T-'
Tlet2

lT+

op (T),

1391

UNIT ROOT HYPOTHESIS

T

(d)

(1/3)[
1~~~~~~~~~

EYtYt,

#2?2j(3 - 2X) + #? (1 -)3

+ P212(1-2

lX2

+ 3,81,82X(1 - X)2] T3

2)]T2

+ 2IT-3/2

TB

tet?1 +2T-3/2

te t_l +X(#I -

2)T-1/2

TB+1

+ [,2+22(1

+ Al) - 2XA2/02] T+ o,(T).

+ X1(L2

It is straightforward but tedious to show that a is given by

A

e, T3/2

TB+1

X) - (1/3)(X132 + (1- X)1322)

+ X( #2-131132)

(A.2)

-= A/D where

(T2/2 + T/2) > tyt-1>Yt

-(T T3/3 + T 2/2
+(T T2/2

+ T16)E yt- I yy Y-T

+ T/2) Ey,-, E tYt+ ( T4/12

tyt_lyEtyt
- T2/12) SY,tyt-

and
(A.3)

D = ( T4/ 1-

l-T(

2)yt-

+ (T + T)-(Ptyt

, t_1i)

+ T2/2 + T16)

,yt--(T3/3

where all summations run from 1 to T. To prove part (b) we derive the limits of T- 7A and T- 7D
using Lemma A.2:

ty2

(1/12)T-

T- 7D

T-6(Ety1tl)

-T

? T-3Etyt-,
= (1/36)[I32X3

-t_l

(1/3)T4(y11)

ot(l) +

- A2)]
+ B22(1- 3X + 2X3) + 3132#,3X(1

- (1/36)[13X(3

- A2) + 12(2 - 3X + A3)]2

+ (1/12)131X(3

-A2)

-(1/12) [ ,8l(2

+ #2(2 - 3X ?+ 3)][ 1X(2

- A) +^1

-X)

+?

2(1-

X)2]

)]+o()

Simple algebra yields:

(A.4)

T-7D

4A
3 A6)(,Bl
+ 4_82 -s
3A 3A
(1/36)(

)2.

Similarly,
=A

+ (1/12)T 3y1y
=

_ty1_T*tT

(1/12)[131X(3 -(1/12)[,81X(2-X)
- (1/36)[1

(1/3)T4(Fyt1)2

T6(tty1)2

p1 + op(l)
2) + 132(2- 3X + 3)] *[,13(2

- X) + 132(1 _ X)21

+ 12(1 _ X)2]2

1X(3 - X2) +

2(2 - 3X +)]

+ (1/36)[ #2X3 + #22(1 - 3 + 2X3) +3#2

1 f- V)] + op(l

1392

PIERRE PERRON

and
T-7A

(1/36)(-3A2

+ 4A3 -

3A4

+ 3A5- A6)(fl,

,2)2

1 proving part (b, i). To prove part (b, ii), note that T(a
Therefore, a = T-7A/T-7D
T 6(A - D)/T-7D. Simple manipulation yields:
(A - D) = (1/12)T2(1y,y,1

- (1) T-4YTE

+ (1/6)T-YTYt1

+ T

2y,/_

-(1/12)T

- 1)=

ty_1

(Y,1)2

(1/12)T-4

+ op(1)
= (1/24)[fij#A2

+ f?(1

A)2 + 2fl,1f2A(1 - A)]

? (1/12)[/?1X(3 -A2) + ,82(2 -3A + 3) [,81A(2 -X) + /2(1 -X)2]

?+R2(1 -

?(1/12)[(f1

- (1/12)[(f1

- 12)X?+2] [P1X(3 - A2) + f2(2-

2)X?+2][t *

(2-XA)

- (1/12) [f1X (2 - A) + R2(1 - A)2]

A)2]
3A + X3)

? op(1)

and

- D) -(1/24)(f,B 4A3
+
5A 2A].
_p2)2[_A?

T-6(A

Therefore,
-- (3/2)[-1+

T(a-1)

4A -5A2

+ 2A3] [-3

+ 4A -3A2

+ 33 -A4]

proving part (b, ii).

PROOF OF PART (a): The proof of part (a) is straightforwardbut tedious. The following arguments
present the main steps. First, the expressions for the moments of { y, } in Lemma A.2 can be written
as follows:
T

ty,l =ajT3+bjT2+ciT3/2+diT+op(T),
Tyt-

= a2T2 + b2T+ c2T1/2

y,_1 =a3T3
T1YtYt-

+d

+ b3T2+c3T32 +d3T+op(T),

==a4T3 + b4T2 +?c4T3/2 +d4T+

op(T).

are given by the corresponding expressions in

The coefficients a,, b,, c, and d, (i= 1,...,4)
Lemma A.2 where the condition ,= = #2= / is imposed. From (A.3), the denominator of a is given
by:
T-5D

a3/12

- al + aja2 -2a/3]T2

+ [b3/12

2a1b1 + a12 + b1a2+ a1a2 -(2/3)

+ [c3/12

2a1c1 + alc2 + cla2 -(2/3)

?[d3/12

a3/12 -b2-

a2b2-

a2/2]T

a2c2] T1/2

2ald + ald2+ blb2

?d1a2 ? alb2 + bla2 -b22/3

(2/3)a2d2 -a2b2-a

2/6]

+ op(l).

1393

UNIT ROOT HYPOTHESIS

Now, tedious algebra shows that the 0(T2), O(T), and 0(T1/2) coefficients all cancel out and the
sum of the 0(1) coefficients yields the following result:
T

+ 6A3- 3A4].

-D S (1/12) a72 + (1/12) [/L1-,]2[A-4A2

2

To analyze the limit of T-5A, the numerator of &,we need the expansions for E2y,and S2ty,.These
are given by:
1tYt= a1T3 + (b, + ,B/2) T2 + c1T3/2 + (1/3)(d1

+ (5/6)p + 2) T + oP(T),

?llyt=a2T
+(b2?+#)T+c2T2

+?d2.

Then we can write,using(A.2):

T-5A

[aia2 -a 2/3 -a2 + a4/12] T2

+ (b2 +

?(a,a2

)(a1/2-a2/3)

+ (b1 + //2)(a2/2

- a,) - a2/2 + b4/12

+ a1b2/2 - alb, - a2b2/3 + b1a2/2] T

+ [a1c2 + c1a2 - (2/3) a2c2 - 2a1c1 + c4/12] T1/2

+ a1/2 - b2/3 - a2/2)

+ [(b2 + 1)(bl/2

? (1/3)(d1 + (5/6)3 + A2)(a2/2

+ (b1 + P/2)(b2/2

a,)

+ a2/2 - bl) + d1a2/2 + b1a2/2 - a2d2/3

a2b2/2

-a 2/6 - d1al + ald2/2 + d4/12 + a4/12].

Again, tediousalgebrashowsthat the 0(T2), O(T),and 0(T1/2) termsall cancelout and that the
sum of the coefficientsremainingyields the followingresult:
T-5A -(+ (1/12)

[,]2[A A2 -

4X2+ 6A3- 3A4].

This provespart (a).

PROOFOF THEOREM
2: We firstnote someinvariancepropertiesof the estimatorsa' (i = A, B, C):
LA is invariantwith respectto ,u and d; aB is invariantto g, and Au2;and ac is invariantto d, ,
and AL2. Hence, withoutloss of generality,we can studythe limitingdistributionof T(L' - 1) and t&

underthe null hypothesisthat the sequence{Yt} is generatedaccordingto:

(A.5)

(t= 1.

yt =yt-, + et

T)

with the innovation sequence satisfying Assumption 1. It is also straightforward to show that under
(A.5), a' (i = A, B, C) areasymptotically
equivalentto the least-squaresestimatorsaL(i = A, B, C) in

the followingregressions:
Yt

?
+

(A.7)

yiB

+ B t + ,BDTt* + yt

(A.8)

Y = Ac +
?cDUt

(A.6)

DUt +

+?ct

At

Ay 1?t,A,

+
?cDTt

+
a?cyt1+t

whereDUt= 1 and DT7= DTt* = if t TBand DU, = O, DT t -TB, DTt= t if t > TB.Since we
are also concernedwith regressionsof the type (A.6)-(A.8) later in the text, we shall derive the
limitingdistributionsconcerningL (i = A, B, C) using the representation
of aL (i = A, B, C). Note,
however,that LA in (A.6) and a in (A.9) arenot invariantto the valueof the parameterd underthe
null hypothesis.To achieveinvariance,one mustintroducea dummyvariableD(TB)t takingvalue 1
at t = TB+ 1 and 0 elsewhere,as is donein regressions(12) and(14). Underthe null hypothesis,LB is
not invariantto a drift takingtwo distinctvalues:i,u and /2. To achieveinvariance,the variableDUt
must be introducedin (A.7). However,this affectsthe limitingdistributionof LB which becomes
equivalentto that of ac. Hence, one cannot analyzedirectlythe case of a joint segmentedtrend
functionin a one-steptyperegression.We now turnto the proofof the theoremderivingthe limiting
distributionof T(a' - 1) and t&i(i = A, B, C) in (A.6) through(A.8).

1394

PIERRE PERRON

The following Lemma provides weak convergence results for the sample moments of the data and
will be used extensively. Its proof is a simple extension of Lemma A.1 and follows the methods of
Phillips (1987) and Phillips and Perron (1988).
LEMMAA.3: Let {Y}Tobe a stochastic process generated according to (A.5) with the innovation
sequence satisfying Assumption1. Furthermore,let TB= AT; then as T oo:
TB

T- 3/2

(a)

yt= ajw(r)

dr,

1o
TB

T- 5/2

(b)

ty,

of

rw(r) dr,

TB

(c)(c)

'y7 = a2jX(r)2

T~2

dr,

(d)

T-

yt- l e, =>(/2)

( w()-e/a).

Parts (a), (b), and (c) are simple generalizations of results in Phillips (1987) and Phillips and Perron
(1988) where TB = T and hence X = 1; part (d) is proved in Phillips (1987).
Using the property that a regression of the form y = X1,f1+ X2,f2 + e^ yields a numerically
identical estimator /2 as obtained in a regression of the form y* = X2*j'2+ e where y* and X2* are
projections of y and X2, respectively, on the space spanned by the vectors in X1, (A.6), (A.7), and
(A.8), can be written as:
(A.9)
(A.9)

~Y"t ~ ,#iXl,t
~ +~ 'i~~-ii
1+
X2,t+ &ii*l+e

(t=1
,...
t

T;; i=A, ,BB,C),)

where
PA=

YBt Yt

YAt

C = yC

= YBY;

B;

Y,

YA*t-1 YBt-1

Yt-1

-1'

XlAt = X1Bt = t -

Xlct = t - cl if
X2A, = DUt - X,
X2Bt = - t*

X2ct = tc- c

if
if

if

Yct=y,-A

t < TB and
t < TB and
TB and

t <TB

Yc*t-i =Yt- -A'

and

=t - TB - i*
=0

otherwise,

otherwise,

otherwise,

=yt-B

t , TB and

if

otherwise,

=t -C2

=Y-

- B' otherwise,

and
T
t

=T-

T-TB

2t = (T +1)/2,

t* =T-1

2:,t =T(l

-) A)/2

+ (1 - A)/2,

1
c2 =t + TX/2,

cl = t-T(1-X)/2,
TB

A =-Y+ A-lTl

TB

B = Y- (1-A)

(y--Yi),

lT-1(yt-Yi),
1

1
TB

A' = Y_l ?X-'T-'

TB

ii),

(Y,-1-

B'= Y1Q~-(1 A)

lT1l

E(Y,1

1
T

T_ 'yt,
1

Y_11=T-lyy-l.
1

-Y_

1395

UNIT ROOT HYPOTHESIS

Now, let Yl1=(yi*.,Yy,T-1),
E' = (el.
eT),
(X21,11. X2i T); then, under the null hypothesis,
(A.10)

&I -1

= (Yi'y,)

[ Z,'(I-Py)Z] Z

i'E- (Y1'1Y) 1'Zi

[Xli, X2J], Xli

(X,11,

'

Xli,T),

X2,

where Py, = Y1(Y1'1)- ?7'. Furthermore, define the following terms:

Zi'Z, = [b,

c ]'

Zil Yi [=ji

['

]L

Straightforward manipulation yields the following representation for a"- 1:

(A.ll)

a-1

[(aic, - b?) Yi'E- ciHiKi + biJiKi - aiJiLi + b,H1Lj]/

ai
c1H72+ 2b1HRJj]
[(aic, - b)

--E, /F.

The quantities involved are defined as follows with their respective limits as T
Lemma A.3:
aA

aB

T3/12 - T/12,

T 3aA = T3aB

CA= (1-X)XT,

1/12,

(1)A/2,
-

T2bA

bA= -(1-X)XT2/2,

oo, obtained using

T1cA-+ (1-X)X,
T

YA'YA = YBYB-,(Y,l-

-1)2

T2YAYA = TY2

jw(r)

[j w(r)2dr

YBy

dr)],

T
-

YA'E= YBE=-?(yl-1

V-l)e,,
w(r) dr,

T-1YA'E= T-1YYBE (a2/2)(w(1)2-ae2/a2)-a2w(1)f

T

HA= HB D=t(Yt-lY

5/2HA =

T-5/2H8

rw(r) dr- (a2)

jw(r)

dr,

TB

JA=

T 3/2JA =*a

,(yt-l1-Y-j),
T

w(r)dr-aX

lw(r)dr,

KA = KB = Etut
1

ut,

T32KA

= T32KB

* (a/2)w(1)

- a wf(r)

TB

LA =

E(et-e

),

T-112LA OcrW(X) -aXw(1),

1
T

T-TB

cB=

T *2 +

+ 3A)/12,

(_-*2T-3CB
-(1-A)3(l

JB= E

(t

TB)(Yt-1

Y-l)9

TB +1

T-/2JsB=*

aj 1rw(r)

dr-

Xcaj1w(r) dr- ((1

X)2/2)

aj

w(r)

dr,

dr,

1396

PIERRE PERRON

E
tT8 (t

LB

TB

ac=

w (r) dr,

((1-A W)/2) caw(1)-rr

T- 3/2LB-(

),

TB) (et

(t- C1)2+
1

T-3ac (1-

(t-c2)2,

)2/12 +X3/12,

TB+1

TB

(t

- c1)2,

T 3bC

T-3cc

X3/12,

TB

CC= F (t

C1)2,

13/12,
T

TB

)2 +

(yt--A

Yc'yc=
1

TB+1

CYYc-rq

dr

w(r)

yt_- -A') et + F

T- 1Y'E

a2(w(1)

w(A))(l

w(r) dr)],

w(r) dr

XA)'f1w(r)

dr,

TB

-A')+ ?

, (t - cl)(y,l
1

(t-C2)(Y,-1-B'),
TB+1

aG[f'o rw(r)

Y.(t-cj)(Y,--

+X)f

dr-(1)(j

w(r)

dr? +()

C =: a [fXrw(r)

T-A'),

dr-(X/2)f

w(r) dr]j

w(r)

dr],

TB

Kc = E (t-Cl)e, + F,
1

(t-C2)et,

TB+1

T-3/2KC
Lc =

1(f

( Yt--B')et,,

W
(a2/2)(w(1)2-a2/a2)-X-1a2w(X)f
-

Jc=

-X)

TB+1

T5/2Hc

-(1

x1(fXw(r)dr)
T

TB

YcE = F,(

Hc =

-B')2,

(y,-

a[21)(1

- X)w(l)

+ (1)w(AX)

cl e,,

T-3/2Lc

*a[(X/2)w(X)j-

, (t-

w(r)dr]
Xw(r)drj

Now, using (A.11), we can write the statistics as follows:

/Th6FA,

(A.12)

T('A-1)

(A.13)

taA = T EA/[SA2 T6FA

(A.14)

T(a& -

(A.1)

T-

1) = T-7EJ/T8,

t& = T7E,/S,2

T-(aAcA

-bA)]1/

i = B, C,

-8F - T-6 (aici -b2)

1/2,

i = B, C,

where 92 = T-12Tet,.
The results of Theorem 2 follow taking the limits of the expressions in (A.12) through (A.15) as
T-+ oo, using the weak convergence results of the relevant moments (given above) and the fact that
?,2(i = A, B, C) converges in probability to a2e as T-. oo.

1397

UNIT ROOT HYPOTHESIS

APPENDIX B
EXTENDEDSET OF RESULTSFOR TESTSOF A UNIT ROOT USING SPLIT
AND FULL SAMPLES

Real
Real
Real
GNP
Money Wages
Interest
Velocity
per
Nominal
Common
Industrial
Quarterly
GNP
wages Consumer
rate
Employment
stock
deflator
real stock
GNP
capita
prices

aThe

sample
for

GNP
production

GNPa
prices
Quarterly
Real
GNP
is

ta

ta a t- & ta &ta & ta a t&a ta a ta

-

ta a

ta a

ta &ta & t- &

-2.33
0.44
0.76
-2.44
-2.16
0.39
-2.14
0.60
0.69
-2.83
0.95
3.14
-2.44
0.81
0.75
2.32
0.76
-1.89
0.73
0.84
-2.51
-1.73
0.93
-1.95
-2.88
0.44
-1.52
0.89

k=1

-1.91
-1.83
0.42 k=2
0.37
0.52
-1.65
-2.23
-2.57
0.72
0.71
0.73
-2.04
-2.79
0.71
-2.22
-1.28
-1.33
0.80
0.91
-1.53
-2.89
0.97
0.80
0.50
0.86
0.90
-3.02
-1.27
-

TESTS

47:1-73:1.

0.43
-1.43
-1.51
0.45
0.65
0.37
-2.05
-2.35
0.91
-2.66
0.78
0.80
-1.80
3.00
0.55
0.68
0.96
-1.72
-1.37
0.73
0.49
-2.06
0.54
-1.22
-4.20
-2.52
0.92
-2.29

k=
3

-1.91
0.23
0.32 k=4
-2.83
-1.36
0.61
0.78
-1.12
0.44
-1.27
-2.71
0.64
0.43
0.77
0.96
-2.09
-2.23
0.66
-2.54
0.93
-1.83
0.75
-1.17
-1.89
-1.50
0.54
0.90
-1.59
0.93
0.76
-2.18
-1.60

-1.43
0.63
-2.18
0.69
-2.21
0.82
0.93
-2.53
0.96
-2.10
0.55
-1.47
0.79
-1.90
-0.82
0.24
-

-2.41
-1.55
0.75
0.92

0.73
-1.71
-2.05
-1.78
0.91
0.75
2.08
-2.19
-0.93
0.28
-1.29
0.74
0.46
0.97
0.54
-2.54

0.91
-2.09
0.64
-2.60

0.61
-1.93
-2.85
0.68
-2.05
-3.01
0.25
0.96
-1.37
-2.82
0.09
0.74
0.89
-1.84
0.30
-1.12

0.57
-0.89

k=5

FOR

Regression:
A
y,
=
+

P3t
+
5ytY,
+

k=6 -1_1d,

0.71
-2.63
-1.43
0.71
-1.41
0.43
-2.05
0.56
-0.63
0.96
0.68
-1.67
-1.08
-3.37
0.90
-0.04

ROOT

TABLE

ON Al

PRE-1929

AY,-,
k=7 +
.t
k

-2.83
-1.75
0.68
0.90

UNIT

0.61
-2.68
-2.06
0.90

k=9

0.88
-3.04

k=10

0.86
-3.40

k=11

0.89
-2.62

k=12

SAMPLES

1398

PIERRE PERRON

Real

aThe

Interest Money
Common Velocity
Quarterly
wages
rate
stock
real
stock

sample
for

per Nominal
Industrial
GNP

Consumer

Employment
deflator

capitaGNP

prices

GNP'prices
Quarterly

t-

Real
GNP
is

Real

Real

GNP

Wages

GNP
production

ta & ta & ta a ta a ta
-

ta a ta a ta & ta & ta & ta & ta a ta

-

-2.23
1.111.07
0.85-3.14
0.88
3.480.65
0.78-2.84
-3.05
0.78
0.68
3.370.763.48
3.820.613.010.893.330.723.370.78-2.92
0.84
0.72
-2.60

k=1

-2.91
0.71-1.69
0.91
0.98
0.93-3.10
-1.58
-0.28
0.88-1.64
-1.87
0.79
0.75
0.84
0.80-2.75
-2.74
-2.70
0.67-1.85
-1.32
0.69
0.92-2.66
0.87-3.00
0.84
-

k=2

0.91-2.34
2.390.85
-2.55
0.79
-1.99
0.89-1.76
-1.68
0.75
0.79
0.87-2.57
0.92-2.55
0.90
0.802.650.66
0.141.01
2.800.63-2.03
0.70-1.41
-1.60

k=3

TESTS
FOR

73:11-86:111.

Regression:
A

UNIT

0.83-2.52
0.712.550.91-1.84
0.63
-1.35
0.75-1.46
0.88-2.10
0.82
-2.44
0.88
-2.29
0.87-1.80
-1.96
-1.89
0.79
0.89-1.96
0.78
0.84
0.371.03
-1.97

k=4

0.71
-2.13
0.511.04
0.85
-2.25
-1.95
0.85-2.50
0.78
0.86-2.16
-2.48
0.75
0.87-2.02
-1.66
0.69-2.50
0.82
0.68-1.77
0.80-1.76
-2.19
0.89-2.14
0.74
-

k=5

ROOT
ON
=1

-2.63
0.83-2.64
-1.91
0.81-2.72
0.81
0.71
0.722.130.64
0.66 k=6
0.66-2.59
0.80
3.070.64
0.87-2.79
-2.40
0.792.65
0.601.06
0.54-2.12
-2.64
-2.64
-

k=7 +

SAMPLES

2.190.80
-1.99
0.64
-

2.330.52-4.57
0.635.32
0.33 k=8
3.080.765.610.475.750.23
5.360.43
-4.10
0.793.310.363.190.763.450.68
-

2.780.73
-2.00
0.62
-

0.75-2.94
0.554.380.03
-3.91
0.783.210.62
2.160.495.930.67-2.14
-2.57
3.390.563.97
0.40-2.02
0.34
0.22
-

-1.63
3.300.65
0.66

0.813.240.423.61-0.11
-1.43
0.633.610.63-2.34
0.232.480.75-1.98
-1.73
0.72
3.080.32
0.18 k=10
3.200.51-3.19
-

k=9

2.910.41-2.05
0.46
0.67
-1.82
-1.89
0.502.830.59-1.89
0.143.060.66-2.04
2.140.753.300.242.40-0.03
0.30 k=11
-

-1.56
3.170.52
0.52

2.810.602.390.58
-2.13
-0.35
0.36-4.05
2.480.68-2.79
0.39-2.11
0.102.33-0.24
-2.54
0.30
-2.24
0.40-2.34
0.09 k=12

-4.03
0.52
-1.64
0.58

POST-1929

2.110.60-2.57
0.63
2.930.723.03
-3.03
-2.96
0.672.550.53
-0.25
0.582.270.77-4.55
0.72-2.70
0.77
0.58
0.56
-4.47
-2.80
0.97
0.86-2.05
0.70
-

TABLE
A2

1399

UNIT ROOT HYPOTHESIS

aThe

number
in

column
(1)
refers
to

Real
Real
Real
GNP
Wages
Money
Interest
Velocity
per
Common
Quarterly
Industrial Nominal
Consumer
GNP
wages
rate
Employment
stock
deflator
real stock
GNP
capita
prices
GNP
GNP
production
prices

i
=
(12) (12) (12) (12) (12) (12) (12) (12) (1)a
B (10) (14) (12) (12) (12) (14)
ta a t a ta a t.t a te & t45 a t a te a t45 & ta a te & ta a te a t- &

the

0.71
-4.43
1.01
-4.31
0.93
-4.63
-1.87
-4.04
-3.33
0.68
0.76
0.70
-3.79
-3.62
0.87
-3.87
0.77
0.97
0.78
-3.77
0.92
0.72
0.12
-1.82
0.84
-4.87
-3.94
0.67
regression
model
used
to
test
for
a
unit
root
using
the
full

-3.61
0.73
-4.06
0.68
0.73
0.63
-4.15
0.76
0.69
-0.45
-4.17
0.98
-1.53
0.98
0.87
-1.28
-3.97
-3.99
0.65
-3.58
0.94
0.82
-4.39
-3.29
-4.46
0.80
-3.97
0.90
-

k=1
TESTS
k=2
FOR
A

0.81
3.86
0.66 k-3
-1.41
-4.11
-0.53
-3.87
0.69
0.91
-4.43
0.95
0.85
-4.25
-1.68
-3.86
-3.89
0.97
0.57
0.74
0.97
-3.72
-3.37
0.72
0.77
-3.47
0.72
0.59
-4.84
-

UNIT

0.63 k=4
3.73
-3.58
-1.11
0.73
-3.21
3.91
-3.66
-4.06
-4.05
-4.16
0.98
0.96
0.86
0.78
0.70
-2.19
0.97
0.76
0.52
-3.78
0.56
-3.18
-4.69
0.70
-0.34
0.90
0.74
-

ROOT

-4.01
0.71
-4.21
3.48
0.66
0.62
-0.31
-3.83
0.61
-3.53
0.76
0.98
0.97
-2.06
0.78
-3.12
0.78
-3.16
0.49
0.97
-3.82
-3.89
0.70
-0.74
0.84
-4.04
-2.94
0.90
-

k=5

0.55 k=6
0.81
-4.73
0.73
3.87
0.75
0.68
0.97
-3.98
-3.27
-4.70
0.60
0.76
0.99
-0.79
0.97
-4.29
-3.62
0.47
-1.89
0.57
0.64
3.52
-4.20
-3.94
-0.24
-3.39
0.90
-4.81
0.43 k=7
0.53
-2.01
-4.51
-5.41
0.56
3.51
-4.88
-3.51
-4.09
0.75
0.96
0.78
0.38
-4.65
0.48
0.62
0.97
0.67
-4.69
-4.02
0.74
0.07
0.89
-0.94
-4.10
1.004
5.03
-2.01
0.71
0.43
-4.91
1.01
0.28 k=8
0.95
3.41
-5.42
0.47
0.96
0.75
0.76
-4.76
-5.47
-4.08
-4.56
-4.28
0.62
0.32
0.89
0.22
-1.09
0.30
-4.32
0.60
-3.42
-

the

-4.11
1.01
-0.91
-4.15
0.73
-5.18
0.46
-4.89
0.19
0.96
-3.89
0.39
0.96
-4.36
0.29
-4.28
-3.55
0.88
-4.62
0.70
-3.59
0.64
0.40
3.52
0.12
0.74
0.64
-1.90

k=9

discussion
0.41
-1.93
-5.38
-4.63
1.03
0.63
-4.05
0.19 k=10
0.96
0.28
-0.37
0.42
-4.14
0.96
0.69
-3.58
0.62
-3.14
0.46
-0.89
-3.98
-4.29
-3.82
0.86
-4.19
0.67
0.74
-4.34
in
the
text.

TABLE
A3

THE

samples.
See

ON

-3.23
0.23
0.15 k=11
-4.16
-4.45
-7.86
-4.35
0.37
0.59
-0.29
-0.08
0.99
-1.22
-4.32
0.72
0.27
0.97
0.67
-3.59
0.94
-3.64
-4.69
0.60
-1.49
-4.32
-4.50
0.62
0.84
0.13 k=12
-5.98
0.58
-4.08
-0.32
-3.17
-4.20
-4.35
0.98
-3.18
-3.37
0.68
0.34
0.34
0.97
-0.49
0.77
0.29
0.67
0.97
-3.49
-3.28
0.87
-1.44
-4.00
-0.40
0.60
-3.64

FULL

SAMPLES

1400

PIERRE PERRON
REFERENCES

BILLINGSLEY,
P. (1968): Convergenceof ProbabilityMeasures. New York: John Wiley.
BLANCHARD,
(1986): "Hysteresis and the European Unemployment
0. J., AND L. H. SUMMERS
Problem," in NBER MacroeconomicsAnnual 1986, ed. by Stanley Fisher. Cambridge: M.I.T. Press,
15-77.
BLOUGH,S. R. (1988): "On the Impossibility of Testing for Unit Roots and Cointegration in Finite
Samples," Working Paper No. 211, John Hopkins University.
Box, G. E. P., AND G. C. TIAO (1975): "Intervention

Analysis with Applications

to Economic

and

Environmental Problems," Journal of the American Statistical Association, 70, 70-79.

J. Y., ANDN. G. MANuw (1987): "Permanent and Transitory Components in MacroecoCAMPBELL,
nomic Fluctuations," American Economic Review, Papers and Proceedings, 77, 111-117.
(1988): "Are Output Fluctuations Transitory?" QuarterlyJournal of Economics, 102, 857-880.
CHAN,N. H. (1988): "The ParameterInference for Nearly Nonstationary Time Series," Journal of the
American Statistical Association, 83, 857-862.
L. J. (1988): "Searching for Breaks in GNP," N.B.E.R. Working Paper No. 2695.
CHRISTIANO,
J. H. (1987): "A Critique of the Application of Unit Root Tests," mimeo, University of
COCHRANE,
Chicago.
(1988): "How Big is the Random Walk in GNP," Journal of Political Economy, 96, 893-920.
DICKEY,D. A. (1976): "Estimation and Hypothesis Testing for Nonstationary Time Series," Unpublished Ph.D. dissertation, Iowa State University, Ames.
DICKEY,D. A., ANDW. A. FULLER(1979): "Distribution of the Estimators for Autoregressive Time
Series with a Unit Root," Journal of the American Statistical Association, 74, 427-431.
(1981): "Likelihood Ratio Statistics for Autoregressive Time Series with a Unit Root,"
Econometrica, 49, 1057-1072.
ENGLE,R. F., ANDC. W. J. GRANGER
(1987): "Co-Integration and Error Correction: Representation,
Estimation and Testing," Econometrica, 55, 251-276.
FULLER,W. A. (1976): Introductionto Statistical Time Series. New York: John Wiley and Sons.
GOULD, J. P., AND C. R. NELSON(1974): "The Stochastic Structure of the Velocity of Money,"
American Economic Review, 64, 405-418.
HALL, R. E. (1978): "Stochastic Implication of the Life Cycle - Permanent Income Hypothesis:
Theory and Evidence," Journal of Political Economy, 86, 971-987.
J. D. (1987): "A New Approach to the Economic Analysis of Nonstationary Time Series
HAMILTON,
and the Business Cycles," Econometrica,57, 357-384.
HERRNDORF,
N. (1984): "A Functional Central Limit Theorem for Weakly Dependent Sequences of
Random Variables," Annals of Probability, 12, 141-153.
LAM, P. S. (1988): "The Generalized Hamilton Model: Estimation and Comparison with Other
Models of Economic Time Series," mimeo, The Ohio State University.
NELSON,C. R., AND C. I. PLOSSER
(1982): 'Trends and Random Walks in Macroeconomic Time
Series," Journal of Monetary Economics, 10, 139-162
OULIARIS,S., J. Y. PARK,ANDP. C. B. PHILLIPS
(1988): "Testing for a Unit Root in the Presence of a
Maintained Trend," in Advances in Econometrics and Modelling, ed. by Baldev Raj. Dordrecht:
Klewer Academic Publishers, 7-28.
PERRON,P. (1987): "Test Consistency with Varying Sampling Interval," Cahier de Recherche, No.
4187, C.R.D.E., Universite de Montreal.
(1988): "Trends and Random Walks in Macroeconomic Time Series: Further Evidence from a
New Approach," Journal of Economic Dynamics and Control, 12, 297-332.
PHILLIPS,P. C. B. (1987): "Time Series Regression with Unit Roots," Econometrica, 55, 277-302.
PHILLIPS,P. C. B., AND S. N. DURLAUF(1986): "Multiple Time Series Regression with Integrated
Processes," Review of Economic Studies, 53, 473-496.
PHILLIPS,P. C. B., AND P. PERRON(1988): "Testing for a Unit Root in Time Series Regression,"
Biometrika, 75, 335-346.
POOLE,W. (1988): "Monetary Policy Lessons of Recent Inflation and Disinflation," Journal of
Economic Perspectives, 2, 73-100.
RAPPOPORT,P., AND L. REICHLIN(1987): "Segmented Trends and Nonstationary Time Series,"
mimeo.

SAID, S. E., AND D. A. DICKEY(1984): "Testing for Unit Roots in Autoregressive - Moving Average
Models of Unknown Order," Biometrika, 71, 599-608.
P. A. (1973): "Proof that Properly Discounted Present Values of Assets Vibrate
SAMUELSON,
Randomly," The Bell Journal of Economics and Management Science, 4, 369-374.

UNIT ROOT HYPOTHESIS

1401

SHILLER,R. J., AND P. PERRON(1985): "Testing the Random Walk Hypothesis: Power Versus
Frequency of Observation," Economics Letters, 18, 381-386.
STOCK,J. H., AND M. W. WATSON(1988): "Testing for Common Trends," Journal of the American
Statistical Association, 83, 1097-1107.
STULZ,R. M., AND W. WASSERFALLEN
(1985): "Macroeconomic Time Series, Business Cycles and
Macroeconomic Policies," in UnderstandingMonetary Regimes, ed. by Karl Brunner and Allan H.
Metzler. Carnegie-Rochester Conference Series on Public Policy, 22, 9-54.
TSAY.,R. S. (1986): "Time Series Model Specification in the Presence of Outliers," Journal of the
American Statistical Association, 81, 132-141.
WASSERFALLEN,
W. (1986): "Non-Stationarities in Macro-Economic Time Series -Further Evidence
and Implications," Canadian Journal of Economics, 19, 498-510.