Вы находитесь на странице: 1из 7

OXFORD BULLETIN OF ECONOMICS AND STATISTICS, 48,4(1986)

0305-9049 $3.00

PRACTITIONERS' CORNER
Kinked Exponential Models for Growth Rate
Estimation

James K. Boyce

I. INTRODUCTION
The usual technique for estimating growth rates in the subperiods of
a time series is to fit separate exponential trend lines by ordinary least
squares to each segment of the series.' These trend lines are likely to
be discontinuous, which can result in anomalies such as subperiod
growth rates which all exceed, or are less than, the estimated growth
rate for the period as a whole. Indeed, it is possible for each estimated
subperiod growth rate to be negative while the estimated growth rate
in the period as a whole is positive, or vice versa.
Discontinuities between segments of a piece-wise regression can be
eliminated via the imposition of linear restrictions (see Poirier, 1976).
In the case of log-linear models, such an approach yields kinked
exponential functions. This note suggests that these provide a better
basis than conventional estimates for intertemporal and cross-sectional
growth rate comparisons. Kinked exponential models with one, two and
multiple kink points are derived. These can be easily estimated with
standard OLS regression packages by using composite independent
variables. The underlying rationale for such models is elaborated, and
their advantages are illustrated by means of a comparison of agricultural
growth rates in two adjoining regions in two subperiods.

II. KINKED EXPONENTIAL MODELS

Single-Kink Model
First consider the simplest case, in which a time series for the period
t = 1,..., n is broken at point k. Discontinuous growth rate estimates
for the two resulting subperiods could be derived by estimating them
separately or, equivalently, by fitting the single equation:
lnYt a1D,+a2D2+(1D1+2D2)t+u (1)

See, for example, the widely used estimates of rates of inflation, GDP and sectoral
output growth, public and private consumption, investment, exports and imports for 126
countries, in the subperiods 1965-73 and 1973-83, reported by the World Bank (1985).

385
386 BULLETIN

where D1 is a dummy variable which takes the value 1 in the /th sub-
period and O otherwise.2
Discontinuity between the two trend lines can be eliminated via a
linear restriction such that they intersect at the break point k:
a1 + 1k = + 2k. (2)
Substituting for a2 (and noting that a1D1 + a1D2 = a1), we get the
restricted form:
in Y = a1 + 1(D1t + D2 k) + f32(D2t Dk) + u. (3)
The OLS estimates of 31 and 2 from (3) give the exponential growth
rates for the two subperiods.3 There is a kink between the two trend
lines whenever

Two-Kink Model
If the time series is broken at two points k1 and k2, so as to create three
subperiods, the unrestricted (discontinuous) model becomes:
lnY ==a1D1 +a2D2+a3D3+(jDi+2D2+133D3)t +u,. (4)
The estimated growth rates from (4), 31, 32 and 33, are the same as if
exponential trends were fitted separately to the data for each subperiod.
The two-kink exponential model is derived by imposing linear restric-
tions such that the subperiod trend lines meet at k1 and k2:
a1 + 1k1 = a2 + 32k1 (5a)
a2 + 2k2 = a3 + 3k2. (Sb)
Substituting for a2 and a3, we obtain the two-kink exponential model:
in Y. = a1 + 31(D1t + D2k1 + D3k1) + 32(D2t D2k1 D3k1 + D3k2)
+ 3(D3t D3k2) + ut. (6)
The growth rates in the three subperiods are now given by the OLS esti-
mates of the coefficients of the resulting composite variables.

General Model
The same technique can be used to derive a generalized kinked
exponential model for m subperiods and m - 1 kinks. Let the kink
2
Where k coincides with an observation in the time series - rather than falling between
two observations - it can be assigned to either subperiod (but not to both) in the construction
of the dummy variables, without affecting the results.
Equation (3) can, of course, be simplified (note that D1 = 1 D3). The more complex
form is presented here to facilitate the following exposition of models with more than one
kink.
PRACTITIONERS' CORNER 387
points be denoted as k1, ... , km_i, and the subperiod dummy variables
as D1.....Dm. The unrestricted model for joint estimation of the
subperiod growth rates, with no continuity requirement imposed, is
then:
lnYaiDi+a2D2+...+amDm
+(131D1+2D2+...+mDm)t+Ut. (7)
Applying the appropriate m - I linear restrictions,
a1 + 31k, = + f31+1k1, for all j = 1, ... , m - 1, (8)
we obtain the generalized kinked exponential model:
m m m
lnY1.=a1+1(D1t+ D1k1)+2(D2t D1k1+ D1k2)
=2 j=2 j3
m m
+ ... + ,(Dt - D1k,1 + D1k,) +
=1 1=1+ 1

+ m(Dm t Dmkmi) + ut. (9)


The number of subperiods into which a given time series can be
meaningfully partitioned will vary from case to case, depending upon
such considerations as the amount of instability, the presence of
cyclical fluctuations and the a priori grounds for expecting growth
rates to change. The single-kink and two-kink models can be readily
derived as special cases of (9).

III. ADVANTAGES OF KINKED MODELS

A distinctive feature of kinked exponential growth models, as opposed


to conventional discontinuous models, is that they make use of in-
formation regarding the values of the variable in question throughout
the time series in estimating the growth rate for a given subperiod. The
rationale for preferring fitted trends over simple point-to-point growth
rate calculations is, of course, that OLS estimates are less affected by
instability or cyclical fluctuations. In other words, neighbouring
observations are understood to provide information as to the trend
level of the variable at any given point in time. Yet conventional growth
rate comparisons proceed, in effect, by estimating subperiod growth
rates in a state of primordial ignorance as to those values taken by the
variable outside each given subperiod - a mask of ignorance which is
dropped whenever the resulting growth rates are compared. Kinked
exponential models, by contrast, make use of the full set of available
information from the outset of the estimation exercise. There is no
388 BULLETIN

asymmetry in the treatment of the information contained in observa-


tions falling before and after the beginning and end points of the
subperiod; all observations are used to distinguish the underlying
growth trends from instability or fluctuations around them.
The incorporation of this information eliminates what can be termed
the 'discontinuity bias' of conventional subperiod growth rate
estimates. If, for example, large positive deviations from the growth
trend happened to occur immediately prior to the break point between
two subperiods, and large negative deviations occurred immediately
after it, the discontinuous growth rate estimates for both subperiods
would be biased upwards. In a kinked exponential model the impact of
the deviant observations would be diminished, since they would tend to
offset each other. The sensitivity of growth rate estimates to
instability is thus reduced by the kinked exponential method. The
result, in general, is more accurate estimates of underlying trends. It
is no longer possible, for example, that each of the subperiod growth
rates estimated from a time series could be greater than, or less than,
the exponential growth rate estimated for the period as a whole.
In certain circumstances, however, there may be reasonable a priori
grounds for permitting a discontinuity between subperiods. An
obvious instance would arise if changes in the definition or geographical
coverage of the variable resulted in lack of comparability between
subperiods. When the rationale for discontinuity is less clear-cut, a
comparison of alternative specifications can shed light upon the
robustness of growth rate estimates. If, for example, a war or
natural disaster severely disrupted an economy, causing a sharp decline
in output, it could be argued that before-and-after growth rate
comparisons should incorporate a discontinuity between the trend
lines. Even so, the implications of such a discontinuity deserve explicit
mention. For instance, our appraisal of a finding that the post-disaster
output growth rate is higher will be modified if we also observe that
output levels have yet to surpass those projected by the earlier trend.
In the absence of such special circumstances, the discontinuous ap-
proach to subperiod growth rate estimation has little to recommend it.

IV. AN EXAMPLE

To illustrate the advantages of kinked exponential models, this section


compares alternative estimates of agricultural output growth rates in
Bangladesh and in the neighbouring Indian state of West Bengal, before
and after the advent of the new seed-fertilizer technology, popularly
termed 'Green revolution', in the mid-1 960's. Discontinuous growth
trend estimates for the subperiods 1949-64 and 1965-80 yield the
following results:
PRACTITIONERS' CORNER 389
Bangladesh: 1nQ = 4.583D1 + 4.499D2 + O.0127D1t + O.0218D2t
Total R2 = 0.91 withF(3, 28) = 94.4 and DW = 1.94. (10)
West Bengal: lnQ = 4.617D1 + 4.342D2 + 0.0l67D1t + 0.0273D2t
TotalR2 = 0.84 with F(3, 28) = 47.7 and DW = 1.93. (11)
Here Q = the index numbers of agricultural output derived by Boyce
(1984, 1985); t = 1, ..., 32; and D1 and D2 are defined as in equation
(1) above.
In both subperiods, the discontinuous estimates show West Bengal
to have experienced more rapid agricultural growth than neighbouring
Bangladesh: 1.67 per cent as opposed to 1.27 per cent in the first
subperiod, and 2.73 per cent as opposed to 2.18 per cent in the second.
Yet when exponential growth rates are estimated for the 1949-80
period as a whole, we obtain the opposite result:
Bangladesh: 1nQ = 4.526 + 0.0203t (12)
TotalR2 = 0.89 withF(1, 30) 247.8 and DW = 1.59.
West Bengal: inQ = 4.599 + 0.0174t (13)
Total R2 = 0.80 withF(1, 30) = 119.0 and DW = 1.69.
In the period as a whole, Bangladesh thus experienced more rapid
agricultural growth than West Bengal: 2.03 per cent as opposed to 1.75
per cent. Our previous finding, that West Bengal had higher growth
rates in both subperiods hence seems rather odd. The key to this
anomalous result is discontinuity bias. It so happens that 1964 was a
bumper crop year in West Bengal, while 1965-67 were years of
drought, imparting an upward bias to the discontinuous growth rate
estimates in both subperiods. In Bangladesh, by contrast, discontinuity
at the break point results in a downward bias. The discontinuous
subperiod growth rate estimates are hence inflated in the case of West
Bengal, and deflated in the case of Bangladesh.
The kinked exponential model removes this discrepancy. Fitting
equation (3) with a kink at the midpoint of the time series (i.e.,
k = 16.5), we obtain the following results:
Bangladesh: lnQt = 4.565 + 0.0157(D1t + D2k) + 0.0249(D2t -D2 k)
TotalR2 = 0.90 withF(2, 29) = 135.3 and DW = 1.77. (14)
West Bengal: in Q = 4.644 + 0.0120(D1t + D2k) + 0.0227(D2t - D2k)
Total R2 = 0.82 withF(2, 29) = 64.9 and DW = 1.87. (15)
The kinked exponential estimates indicate that in both subperiods, as
in the 1949-80 period as a whole, Bangladesh experienced more rapid
agricultural growth than West Bengal: 1.57 per cent as opposed to 1.20
390 BULLETIN

per cent in the first subperiod, and 2.49 per cent as opposed to 2.27
per cent in the second. In the present instance - in which a single kink
occurs at the midpoint of a complete time series - the average of the
kinked exponential growth rate estimates for the two subperiods
precisely equals the exponential growth rate estimate for the period
as a whole.4 By eliminating discontinuity bias, the kinked exponential
model thus provides a superior measure of the two regions' relative
performance in the two subperiods.

V. CONCLUSION

When exponential growth trends for subperiods of a time series are


estimated separately, there is no guarantee that the terminus of one
trend line will coincide with the beginning of the next. Unless it does,
the fitted trend level of the variable at the break point between
subperiods will take two different values, one given by the earlier trend
and one by the later. This entails a fundamental inconsistency when-
ever the resulting growth rate estimates are compared. The attendant
discontinuity bias can produce anomalous results, with subperiod
'growth rates' which bear little relation to the growth rate in the
period as a whole. Since the magnitude and direction of this bias may
vary from one case to another, discontinuous estimates can be
misleading for cross-sectional as well as intertemporal growth rate
comparisons. Yet today this is the standard methodology employed
for growth rate estimation.
Kinked exponential models, which impose a continuity restriction at
the break points between subperiods, eliminate this bias and thereby
provide an improved basis for growth rate comparisons. These estimates
make use of the full set of observed values of the variable from the
outset of the analysis, rather than excluding those values outside each
subperiod as in the discontinuous estimates.
In the absence of special circumstances, such as definitional changes
or natural disasters, kinked exponential models are preferable to
discontinuous ones for growth rate comparisons. An inevitable element
of arbitrariness remains in any growth rate estimation procedure:
the choice of end points of the time series, the demarcation of
subperiods, and the inclusion or exclusion of exceptional observations
may affect the results. Moreover, for forecasting purposes an alternative
functional form (such as the log-quadratic, which assumes constant
Also, it may be noted that in this case the magnitude of the trend break - the difference
between the growth rates in the two subperiods - is the same in the discontinuous and
kinked exponential models: 0.91 per cent in Bangladesh and 1.07 per cent in West Bengal.
This too is a special property of models with a single kink at the midpoint. In other
specifications, the magnitude and even the direction of the trend break can vary between the
two classes of models.
PRACTITIONERS' CORNER 391
acceleration or deceleration) might be more appropriate in some cases.
For the estimation and comparison of past growth rates, however,
kinked exponential models represent a distinct improvement over
conventional estimation procedures, and have potential applications in
a wide range of cases.

Magdalen College, Oxford


and University of Massachusetts, Amherst
Date of Receipt of Final Manuscript: February 1986

REFERENCES

Boyce, J. K. (1984). 'Agricultural Growth in West Bengal, 1949-50 to 1980-81:


A Review of the Evidence', Economic and Political Weekly, Vol. 19, No. 13,
pp. A9-A16.
Boyce, J. K. (1985). 'Agricultural Growth in Bangladesh, 1949-50 to 1980-81:
A Review of the Evidence', Economic and Political Weekly, Vol. 20, No. 13,
pp. A31-A43.
Poirier, D. J. (1976). The Econometrics of Structural Change, Amsterdam,
North.Holland.
World Bank (1985). World Development Report 1985, New York and Oxford,
Oxford University Press.

Вам также может понравиться