Consumption and Income

Paneleconometric Evidence for West Germany

Christian Dreger*and Reinhold Kosfeld**

Abstract. In this paper the permanent income hypothesis (PIH) of consumption is tested
by means of paneleconometric techniques, which are applied to West German regional
data. Panel unit-root tests which are robust to cross section correlation indicate the sta-
tionarity of the savings rate, revealing weakly support for the PIH. However, consumers
may be subject to liquidity constraints. The relevance of non-optimizing consumers is
examined within the error-correction λ-model. Approximately a fraction of 45 percent
of disposable income is earned by households that do not to behave according to the
PIH, and the share is positively related to regional unemployment rates.

Key words: Permanent Income Hypothesis, Liquidity Constraints, Panel Unit Root
Tests

JEL classification: C23, E21

1 Introduction

Consumption of private households is by far the biggest aggregate in GDP. In Germany

consumption expenditures amount to 55 percent in GDP on average over the last two
decades. Other industrialized countries experience quite similiar shares. Explaining con-
sumption spending is therefore a prequisite for accurate forecasts of GDP.

Despite its relevance only a few actual studies analyse aggregate consumption behav-
iour in Germany. Wolters (1992) e.g. discusses the empirical performance of several
specifications, especially error correction models build on consumption and disposable
income series. Reimers (1997) and Hassler (2001) extend the cointegration analysis by

*
Institute for Economic Research Halle (IWH), Kleine Märkerstraße 8, 06108 Halle, Germany. Tel.:
+49-345-7753854, Fax: +49-345-7753825, E-Mail: [email protected].
**
University of Kassel, Department of Economics, Nora-Platiel-Straße 5, 34127 Kassel, Germany. Tel.:
+49-561-8043084, Fax: +49-561-8043045, E-Mail: [email protected].
 2

introducing a wealth variable, representing accumulated savings. The former author

reports tests for seasonal cointegration, while the latter discusses multicointegration
topics between consumption, income and wealth.

This paper chooses the life-cycle permanent income hypothesis (PIH) for consumption
as the point of departure. The PIH implies cointegration between consumption and dis-
posable income or a stationary saving series. As a novelty, this hypothesis is tested
within a paneleconometric framework by the means of recently developed panel unit
root tests. These tests utilize a broader long run information set and are therefore better
suited to detect a false null hypothesis than the conventional unit root procedures. The
panel is based on data from West German regions. To our knowledge no other study has
analysed the regional data set so far.

However, note that cointegration analysis can only provide weak tests of the validity of
the PIH. Stationarity of the savings series is also consistent with several specifications,
including an ordinary Keynesian type consumption function. Here expenditures depend
not on permanent, but on actual income, which can be justified on the basis of myopic
or credit rationing behaviour, see Campbell and Mankiw (1991). Thus, the relevance of
market imperfections has to be investigated and this is also done within the panel error
correction framework. Essentially, regional consumption functions are estimated using
the SUR technique, and the restriction of an equal feedback parameter across the panel
members is tested by the means of a standard Wald test. The coefficient of actual in-
come growth may be interpreted as the income share earned by liquidity constrained
consumers. This view is supported by a correlation analysis between the individual
shares and measures of regional economic performance such as the unemployment rate.
As a result of the analysis, regional weights of actual and permanent income in explain-
ing consumption expenditures per capita are estimated.

The rest of the paper is organized as follows. The next section briefly reviews the rele-
vant theory of the PIH including the modifications in the presence of liquidity con-
straints. Second, panel unit root tests are discussed. After describing the regional data
set, the empirical analysis is performed. Section 6 concludes.
 3

2 Permanent Income and λ-model

The PIH in a stochastic environment results from optimizing behaviour of infinitely

long living agents, c.f. Hall (1978). Representative consumers have a felicity (instanta-
neous utility) function u(Ct ), relating utility only to the level of current consumption Ct
per capita. In fact, the expected life cycle utility

∞
(1) ∑ (1 + δ ) − i Et u (Ct + i )
i =0

is maximized subject to the wealth accumulation equation

(2) Wt +1 = (1 + r )Wt + Yt − Ct

where Wt is financial wealth at the beginning of the period t and r the real interest rate.

rWt is capital income, and Yt labor income received in the current period. The time pref-

erence rate δ and the real interest rate are assumed to be constant. Et indicates expecta-
tions conditional on the information set available to the representative consumer at time
t. This amounts to consumers having rational expectations. Recursive substitution in
equation (2) and taking expectations leads to the infinite horizon budget constraint

∞ i ∞ i
 1   1 
(3) ∑  1 + r  t t + i
E C = (1 + r )Wt ∑
+
 1 + r
 Et Yt + i

i =0 i =0

where no Ponzi games have been ruled out.1 According to equation (3), the present
value of lifetime consumption is equal to the sum of financial and human wealth, and
the latter is defined as the present value of expected future labour incomes. Differentiat-
ing the Lagrangean of the constrained optimization problem stated in equations (1) and
(3) gives the familiar stochastic Euler equation

1+ δ
(4) Et u ′(Ct + i ) = u ′(Ct )
1+ r

1
The no Ponzi game assumption sets the present value of financial wealth (debt) to 0 as i goes to infinity.
Otherwise agents can borrow indefinitely to finance an increase in consumption.
 4

which implies that on the optimal path marginal utilities of consumption follow a first
order Markov process. Moreover, marginal utilities and thus the corresponding levels of
consumption per capita are constant, when the time preference rate is equal to the real
interest rate. In the following, this restriction is assumed to be met.

Further analysis requires the specification of the instantaneous utility function. Utility
may be quadratic in consumption,

u (Ct ) = −0.5(Ct − C ∗ ) 2

*
where C is some constant bliss level of consumption. Inserting into equation (4) and
shifting time back one period yields:

(5) Ct = Ct −1 + vt

where v is a white noise forecasting error, stemming from the rational expectations as-
sumption. According to equation (5) consumption follows a random walk (without
drift).2 Moreover EtCt+i=Ct holds for all i, implying that the intertemporal budget con-
straint may be rewritten as

i
r ∞ 1 
(6) Ct = rWt + ∑   Et Yt + i
1 + r i = 0 1 + r 

which is the consumption function under the PIH, c.f. Flavin (1981). Ex ante consump-
tion is equal to permanent income, and the latter is defined as the right hand side of
equation (6) which is the sum of capital income and expected future labor income. The
terms are obtained as present values at the beginning of the period.

Disposable income is equal to the sum of both income components: YDt=rWt+Yt.

Utilizing the definition of saving, St = YDt -Ct and equation (6) the PIH may be rewritten
as

2
See Hall (1978). The random walk hypothesis can be reformulated in the logs of consumption, when a
CES utility function is assumed, see Campbell and Mankiw (1991).
 5

∞ i
 1 
(7) St = − ∑   Et ∆Yt + i ,
i = 0
1+ r 

where ∆ is the first difference operator. This formulation essentially captures the rainy
day feature of the PIH. Saving occurs if a decline in labour income is expected. More-
over, saving serves as an optimal predictor for anticipated changes in labor income, c.f.
Campbell (1987). Finally, if labour income has a unit root, saving is stationary since it
is the sum of stationary terms with exponentially declining weights, provided the real
interest rate is positive. As a result, consumption and disposable income are cointe-
grated, and the vector corresponding to the long run equilibrium relation is (1,-1).

According to the PIH and the assumptions needed for equation (5), changes in con-
sumption correspond to revisions in expected labor income changes. This can be seen
by inserting equation (2) in equation (6). Surprises in the income process are purely
white noise implying that changes in consumption are generally unforecastable.

However, empirical evidence suggests that predictable changes of income do have an

impact on actual changes in consumption, see Muellbauer (1994) for a survey. Accord-
ing to Campbell and Mankiw (1991), this so-called excessive sensitivity of consumption
may be a reflection of liquidity contraints. Credit-rationed agents do not save and spend
their whole income, which solely stems from labor. For example, young people with
low income, who expect to earn more in the future are faced by such restrictions, when
desired credits are denied. In opposite, optimizing households consume without con-
straints, and their expenditures fluctuate randomly. The shares of the two distinct groups
in the income distribution are λ and 1-λ, respectively. Therefore, the changes in con-
sumption may be reproduced by the so called λ-model

(8) ∆Ct = λ∆YDt + (1 − λ )vt + γ (Ct −1 − YDt −1 ) ,

see Campbell and Mankiw (1991). Short-run movements of consumption are explained
via a linear combination of the ordinary Keynesian function and the PIH. A test for the
 6

importance of liquidity constraints relies on the null hypothesis λ=0.3 If the null is re-
jected, the estimated parameter may be interpreted as the fraction of credit-restricted
consumers in the economy; the complement fraction 1-λ is earned by optimizing con-
sumers.4 Unpredictable fluctuations in permanent income are accounted for by the white
noise error term v. In any case, the lagged saving series must be included as an addi-
tional regressor if consumption and disposable income are cointegrated (γ <0).

3 Panel Unit Root Tests

In the following, testing for cointegration between consumption and disposable income
is implemented as a test for a unit root in the saving series. The analysis is carried out
within a paneleconometric framework. Panel unit root tests have greater power than the
conventional unit root tests. Since the time series dimension is enhanced by the cross
section, the tests rely on a broader long run information set. Thus they are better suited
to detect a true alternative hypothesis even in the presence of nearly unit root alterna-
tives, see Levin and Lin (1993) and Maddala and Wu (1999). However, the increase in
power is not costless. In particular, contemporaneous correlation can arise and the unit
root properties may differ across the panel members. These two pitfalls are discussed in
turn.

First, the contemporaneous correlation may be attributed to common shocks which af-
fect jointly all the panel individuals. As a consequence, independent long run informa-
tion is lost. If the correlation structure is not recognized, test statistics will suffer from
substantial size distortions, see O'Connell (1998). In fact, the true size of the unit root
tests can be far above the nominal level even if the correlation is moderate. Thus, the
results of the tests are highly questionable when the dependencies are not modelled ap-
propriately.

Second, panel unit root tests do not reveal any individual specific information in respect
to the unit root feature. If the joint null hypothesis of a unit root is rejected, the series

3
This also may reflect some kind of myopic behaviour. Furthermore, if future income is more heavily
discounted due to income uncertainty, current income is necessarily of higher relevance in determining
actual consumption.
 7

may be stationary for all cross sections or only for a subgroup. The null might be re-
jected even if there is only one stationary individual in an otherwise unit root environ-
ment, see the Monte Carlo evidence presented by Taylor and Sarno (1998). Therefore
test results must be interpreted with some caution. However, this vagueness seems to be
less important in the present study. The panel consists of West German regions, and
consumer behaviour should not differ substantial across the individuals.5

Several panel unit root tests are available, see Banerjee (1999) and Breuer, McNown
and Wallace (1999) for recent surveys. Most popular are the tests of Levin and Lin
(1993) and Im, Pesaran and Shin (1997), hereafter IPS. Both rely on ADF principles,
and the joint null is always a unit root for all panel members. The tests differ in their
alternatives, respectively. In particular, if the null is rejected, Levin and Lin (1993) con-
clude the stationarity of the series for all regions, while in the IPS (1997) setting, at least
one individual is stationary. A different degree of heterogeneity allowed for the cross
sections is responsible for this outcome. In the Levin and Lin (1993) procedure, hetero-
geneity is restricted to the deterministic components and short run dynamics, but the
unit root properties are restricted to be the same. Instead, IPS (1997) relax the assump-
tion of an equal order of integration across the panel members.

In the presence of contemporaneous correlation the distributions of the test statistics are
no longer valid. IPS (1997) consider the special case where cross-correlation between
the panel members is caused by time specific effects. The dependencies can be re-
moved, when testing relies on mean-adjusted data. Time effects are identical for all in-
dividuals and so the mean is computed over the cross sections. Unfortunately, this ap-
proach is useful only for a specific correlation structure and hence, not applicable in
general.

In order to control for any type of contemporaneous correlation the multivariate exten-
sion of the ADF test recently proposed by Taylor and Sarno (1998) is preferred. The
multivariate ADF test (hereafter MADF test) investigates the joint null that the series

4
If λ ≠0, saving is a multiple (1-λ) of (7), see Jin (1995). Thus, saving will be smoother than under pure
PIH conditions.
5
Breuer, McNown and Wallace (1999) suggest complementary ADF tests in a SUR setting in order to
test for individual unit roots. Taylor and Sarno (1998) present a Johansen style likelihood ratio test in
order to determine whether all individual series are stationary. Due to the degree-of-freedom problem,
the latter test is only applicable in panels where the cross section dimension is small.
 8

has a unit root for all regions. Robust testing is done on the basis of a SUR analysis,
which is essentially GLS using an estimate of the contemporaneous covariance matrix
of the errors obtained from individual OLS estimation in the previous step. The ith
equation of the system is given by

pi
(9) ∆yit = α i + δ i yit −1 + ∑ φij ∆yit − j + vit
j =1

where vit denotes the individual White Noise error and i and t indicate the cross section
and the time series dimension of the panel, respectively (i=1,..,N; t=1,..,T). The errors
are allowed to be contemporary correlated. Member specific effects, short run dynamics
and unit root properties are permitted to vary across the individuals, as in the procedure
suggested by IPS (1997). The joint null of a unit root δ1=...=δN=0 is judged by the
means of a Wald test. If the null is rejected, at least one individual is stationary. As far
as similarity of West German states may be justified, stationarity can be concluded for
all panel members.

Note that the distribution of the MADF is nonstandard mainly because of the presence
of nuisance parameters which represent the cross section correlations under the null
hypothesis.6 Thus, the empirical finite sample distribution has to be derived through
simulation and details for the present study are given in the appendix. The critical val-
ues are specific to the estimated covariance matrix, the sample size and the number of
panel members.7 Due to this requirement panel testing might be less attractive, but nev-
ertheless it is a reliable way to proceed.

After testing for cointegration between consumption and disposable income, an error
correction model in the style of equation (8) is implemented to adress the liquidity con-
straint feature. This model is also estimated along the lines of the SUR framework. At
this step, the usual asymptotics apply, since all variables are stationary.

A normal limit theory can be obtained if the errors are cross-sectionally independent, see Im, Pesaran
and Shin (1997).
7
When the tests of Levin and Lin (1993) and Im, Pesaran and Shin (1997) are applied, a non-parametric
bootstrap may also be appropriate in order to deal with the problem of contemporaneous correlation,
see Berkowitz and Kilian (2000).
 9

4 Data

The analysis is based on real per capita data for private consumption expenditure and
disposable income. The price index of consumption (1991=100) is used to deflate the
consumption and income series, and per capita variables are constructed by dividing
through the population figures. Annual data is employed and the sample period ranges
from 1970 through 1997. All series are obtained for 10 West German regions, which
constitute the panel: Baden-Württemberg (BAD), Bavaria (BAV), Bremen (BRE),
Hamburg (HAM), Hesse (HES), Lower Saxony (LSX), North Rhine Westphalia
(NRW), Rhineland Palatinate (RHP), Saarland (SAL) and Schleswig Holstein (SHO).
Berlin is excluded from the sample because data are available only up to 1994. In total
280 observations are included.

Figure 1: Per capita savings Rates of West German states

BAD BAV
.175 .16
.15 .14
.125 .12
1970 1975 1980 1985 1990 1995 1970 1975 1980 1985 1990 1995
BRE .25 HAM
.175
.2
.125 .15
.075 .1
1970 1975 1980 1985 1990 1995 1970 1975 1980 1985 1990 1995
HES .16 LSX
.16
.13 .135
.1 .11
1970 1975 1980 1985 1990 1995 1970 1975 1980 1985 1990 1995
NRW RHP
.16 .175
.14 .15
.12 .125
1970 1975 1980 1985 1990 1995 1970 1975 1980 1985 1990 1995
SAL SHO
.15 .175
.125
.15
.1
.125
1970 1975 1980 1985 1990 1995 1970 1975 1980 1985 1990 1995
 10

The series are taken from the Arbeitskreis Volkswirtschaftliche Gesamtrechnungen der
Länder that conducts the regional accounting for Germany. The variables are measured
in natural logarithms and a cointegration relation between the logs of consumption and
income requires the stationarity of the consumption income ratio or the savings rate,
respectively.8 Per capita saving series are plotted in Figure 1.

Despite some dissimilarities, per capita saving in percent is roughly comparable across
the regions. Relatively high saving rates can be observed in particular at the beginning
of the sample period. Since the first oil crisis in 1973/74, the rates have declined for
some panel members, namely Baden-Württemberg, Rhineland Palatinate, Schleswig
Holstein and the city states Bremen and Hamburg, wheras for the rest of the regions, the
savings rate tends to be mean reversible. Starting with German unification in 1990, a
fall in the savings rate seems to be more significant. According to the PIH, this may be
attributed to the low growth rates of the economy at the end of the sample period. For
example between 1995 and 1997 the growth rate of German GDP was only 1.3%, which
is significantly lower than the average growth rate in the whole sample period. Cur-
rently the saving rate partially recovers from the historically low levels in the second
half of the 1990s. In the first quarter of 2001, it has risen up to 11% for the whole econ-
omy, compared to 9% in the corresponding quarter of 2000.

5 Results

As a preliminary for setting up the λ-model, the MADF test for a unit root in the sav-
ings rate is conducted. In a first step the standard ADF tests are in order to motivate
carried out the panel data approach. Here the lag length k in the individual ADF-
regressions is determined by the procedure suggested by Campbell and Perron (1991).
Specifically, an upper bound of k=3 is set for all regions. Then k is reduced sequentially
by 1 until the last lag becomes significant. If no lags are significant, the straightforward
choice is k=0. To assess the significance of the lags, the 10% value of the normal distri-
bution is used. This procedure suggests k=0 or k=1 for exactly one half of the panel

8
The behaviour of saving and the change in consumption under the PIH can be restated in terms of the
log consumption-income ratio and the growth rate of consumption, see Campbell and Mankiw (1991).
 11

members. In addition the ADF regression includes a constant, but no time trend. Results
are reported in Table 1.

Table 1: ADF tests for the savings rate

k ADF K ADF

BAD 0 -1.580 LSX 1 -2.658(*)

BAV 0 -2.246 NRW 1 -2.811(*)

BRE 1 -1.686 RHP 0 -0.910

HAM 1 -1.615 SAL 0 -1.715

HES 0 -2.344 SHO 0 -1.846

Notes: * Significance at a level of 5 %; (*) significance at a level of 10 %

The 5% critical value is -2.980. Hence the null of a unit root in the savings rate cannot
be rejected in any case. At least at the 10% level the rate seems to be stationary for
Lower Saxony and North Rhine Westphalia. These results might be substantial i.e. they
may indicate no integration of the logs of per capita consumption and income with the
vector (1,-1) but they could also reflect the low power of standard unit root tests.

In order to overcome the problem the MADF approach is employed and this is done for
two panel sizes. Panel A includes the entire regions, while Panel B excludes Lower
Saxony and North Rhine Westphalia. The subpanel is choosen to ensure that any prob-
lems arising from potentially different unit root features of the individuals are avoided.
The MADF procedure is based on a SUR analysis, where each series is modelled ac-
cording to the specifications found in the conventional ADF settings.9 The finite sample
distributions have to be derived by Monte Carlo simulation, and details of the methods
applied are given in the appendix. Table 2 reports the results of the MADF test for the
different panel sizes.

9
The average contemporaneous correlation between the residuals obtained from separate OLS regres-
sions is about 0.5. The absolute value drops to 0.3 if the analysis relies on mean-adjusted data. Hence,
the correlation remains substantial and SUR is recommended.
 12

Table 2: MADF unit root tests for the savings rate

MADF 5% critical value 10% critical value

Panel A 72.16(*) 72.50 63.47

Panel B 53.23(*) 59.88 53.02

Notes: * Significance at a level of 5 %; (*) significance at a level of 10 %

Almost at the 5% level at significance, the null of a joint unit root is rejected in Panel A.
The test statistic declines substantially in the smaller panel but it is still significant at the
10% level. The critical values will also be lower in absolute value as correlation is re-
moved from the system in the smaller Panel B. Although the empirical evidence is not
overwhelming, the results seems to be in favor of a stationary saving rate per capita.
Equivalently the logs of consumption and disposable income are cointegrated, and the
cointegrating vector is (1,-1).10 However a cointegration analysis can not discriminate
between the PIH and a Keynesian type consumption function.

Therefore the λ-model which explains the growth rate of consumption by income
growth and the lagged consumption-income ratio as the error correction term is ana-
lysed. Note that OLS applied to equation (8) will inevitable result in inconsistent esti-
mators. The reason is that changes in actual and permanent income are likely to be posi-
tively correlated, where the latter constitutes the error term. The correlation biases the
parameter upwards, implying that the relevance of market imperfections is likely to be
overestimated.

A solution to this problem is to use the instrumental variable approach. In particular,

lagged growth rates of consumption and disposable income, a constant, a linear time
trend and impulse dummies are treated as instruments, and the concrete set of variables
is allowed to differ across the panel members.11 The impulse dummies capture shocks

10
Wolters (1992) and Hassler (2001) also report a cointegrating relationship between consumption and
disposable income, while in Reimers (1997) the long run relationship can only be established after a
wealth variable is included in the analysis.
11
Details are available from the authors upon request. The lagged saving rate is insignificant in these
regressions and therefore it does not serve as an instrument. This implies that disposable income is
weakly exogeneous with respect to the cointegration structure, and thus a single equation error correc-
tion model can be estimated.
 13

traced to the oil crises and to the German unification, respectively.12 On average, the
2
adjusted R of a regression of the growth rate of income on its instruments is approxi-
mately 0.6, where the lowest figures 0.3 and 0.5 are computed for the two city states
Bremen and Hamburg, respectively. The fitted values replace the original regressor i.e.
the growth rate of income, which renders the estimation consistent.

However, IV estimation is not efficient. Efficiency gains can be realized by exploiting

the contemporaneous correlation structure across the regions. This is the rational of a
SUR analysis. Moreover, in this framework cross equation restrictions can be evaluated
for example by the means of standard Wald tests. In particular, the hypothesis of an
equal feedback parameter, i.e. the coefficient of the error correction term, can be
broadly accepted. The test statistic is distributed as chi-square with N-1=9 degrees of
freedom. It has a value of 4.74 while the 5% critical value is 16.9.13 The results of the
restricted error correction model are presented in Table 3.

Table 3: IV-SUR estimation of the λ-model (8)

∆yt ect-1 ∆yt ect-1

0.439 0.507
BAD (6.956)
LSX (7.872)
0.497 0.368
BAV (9.326)
NRW (6.297)
0.519 -0.075 0.453 -0.075
BRE (4.853) (6.099)
RHP (6.777) (6.099)
0.374 0.531
HAM (4.276)
SAL (5.884)
0.467 0.445
HES (5.636)
SHO (6.716)

Notes: The terms in brackets are the t-values, ect=ct-yt is the consumption-income ratio with ct=log (Ct)
and yt=log(Yt). As explained above, instruments are used instead of the original ∆yt series.

All regressors are significant at the 1% level. Both the PIH and an ordinary keynesian
approach seem to explain the growth rate of consumption per capita. The estimator of λ

12
Dummies are not needed in the case of the panel unit root tests because of co-breaking, see Hassler
(2001). The shocks affect jointly consumption and income and their net impact, which is included in
the savings rate, is negligible.
13
In addition, the restriction of an equal λ-parameter is not rejected either, but it is not applied in the
following. Without this restriction the interpretation of is easier.
 14

suggests that roughly 45 percent of disposable income is earned by non-optimizing con-

sumers.14 This figure is broadly comparable across the panel regions, where North
Rhine Westphalia (37), Hamburg (37) and Saarland (53) mark the edges of the scale.
The complement is received by optimizing agents and therefore, the PIH is still impor-
tant for explaining consumption behaviour in Germany. Furthermore the results indicate
a similiar consumption behaviour across the regions.

Finally, the correlation between the individual λ parameters and measures of regional
economic performance is considered. This may shed some light upon the sort of the
market imperfections. For example, the correlation between the λ coefficients and the
mean regional unemployment rates is 0.29 and this number is significant positive at the
5% level. If unemployment serves as a proxy for credit rationing, the findings of the λ-
model can be partly traced to the liquidity constraints issue.15

6 Conclusion

In this paper the permanent income hypothesis (PIH) of consumption is examined on

the basis of West German regional data. Standard unit root tests fail to reject the joint
null of a unit root in the saving rate, mainly due to the power of these procedures. In
fact panel unit-root tests which are robust to cross section correlation indicate the sta-
tionarity of the saving rate. This finding supports the PIH, but also other types of the
consumption function.

In particular, the relevance of non-optimizing consumers is examined within the λ-

model which is extended by an error correction term. Here approximately a fraction of
45 percent of disposable income is earned by households that do not to behave accord-
ing to the PIH. The share is positively related to measures of regional economic per-
formance i.e. the unemployment rate and may therefore reflect the relevance of liquidity

14
The only previous result for Germany is reported by Wolters (1992). He estimates the model without
the error correction term and finds a fraction of 0.29 for the whole West German economy.
15
The assumption that an individual that is unemployed is also faced by credit restrictions merely serves
as an approximation. Specifically, a distinction should be drawn between sustained and temporary peri-
ods of unemployment. Individuals who expect to be permanently unemployed might behave according
to the PIH, and this potentially explains the relative low correlation. However, the data are only avail-
able at a yearly frequency and therefore, the argument can not be stressed further.
 15

constraints. All in all the results are broadly comparable for the panel members, thus
indicating similiar consumption behaviour across the regions.

Appendix

Due to the dependence between the cross section units, the finite sample distribution of
the MADF statistic has to be derived by simulation. The critical values will depend on
both the cross section and time series dimension of the panel and on the contemporane-
ous correlation structure as well. The Monte Carlo experiments are based on 10 autore-
gressive models describing the fluctuations in the regional saving rates, and the lag
structure is reported in Table 1. After performing OLS for the cross sections separately,
the contemporaneous covariance matrix of the residuals is computed. This matrix is
employed in executing the Monte Carlo experiments.

In particular, error series of length 100+T are drawn from the distribution. To reduce the
sensitivity of the results on initial conditions the first 100 observations are discarded,
leaving time series of length T=30, 50, 100 and 250 for the analysis. Using the esti-
mated values of the coefficients of the lagged variables, pseudo saving rates are gener-
ated. Then the system corresponding to equation (9) is estimated by the SUR technique
and the Wald test for the joint null of a unit root for all panel members is performed. In

Table 4: Critical values for the MADF test

N=8 N=10

T T
α
30 50 100 250 30 50 100 250

10% 53.02 39.66 30.61 24.32 63.47 47.99 37.42 29.75

5% 59.88 44.78 34.53 27.50 72.50 53.41 41.38 33.26

1% 76.78 55.94 42.10 33.69 94.08 63.04 49.23 40.73

Notes: N = number of cross-section units; T = number of time periods; α = level of significance.

 16

order to find the empirical distribution of the test statistic, the process is repeated 5000
times. The results for the different panel sizes examined in this paper are reported in
Table 4. For example at the 5% level of significance and 30 time periods unit the critical
value is 72.50, when all cross sections are included (N=10). If the MADF test exceeds
this level, the joint null is rejected. Note that the critical values depend on a particular
contemporaneous correlation structure. Hence they are specific to the analysis and can-
not be applied in other studies. The TSP386 program which generates the critical values
is available from the author upon request.

References

Banerjee, A. (1999): Panel Unit Root Tests and Cointegration: An Overview, Oxford
Bulletin of Economics and Statistics, Special Issue, 607-629.
Berkowitz, J. and Kilian, L. (2000): Recent Developments in Bootstrapping Time Se-
ries, Econometric Reviews 19, 1-54.
Breuer, J.B., McNown, R. and Wallace, M. (1999): Series-Specific Tests for a Unit
Root in a Panel Setting with an application to Real Exchange Rates, Discussion Paper
99-9, Department of Economics, University of Colorado, Boulder.
Campbell, J.Y. (1987): Does Saving anticipate declining Labor Income? An alternative
test of the Permanent Income Hypothesis, Econometrica 55, 1249-1273.
Campbell, J.Y. and Mankiw, N.G. (1991): The Response of Consumption to Income,
European Economic Review 35, 723-767.
Campbell, J.Y. and Perron, P. (1991): Pitfalls and Opportunities: What Macroecono-
mists should know about Unit Roots, NBER Macroeconomics Annual, 141-201.
Flavin, M.A. (1981): The Adjustment of Consumption to Changing Expectations about
future Income, Journal of Political Economy 89, 974-1009.
Hall, R.E. (1978): Stochastic Implications of the Life Cycle-Permanent Income Hy-
pothesis: Theory and Evidence, Journal of Political Economy 86, 971-987.
Hassler, U. (2001): Wealth and Consumption: A Multicointegrated Model for the Uni-
fied Germany, Jahrbücher für Nationalökonomie und Statistik 221, 32-44.
Im, K.S., Pesaran, M.H. and Shin, Y. (1997): Testing for Unit Roots in Heterogeneous
Panels, Working Paper, Department of Applied Economics, University of Cambridge,
Cambridge, UK.
Jin, F. (1995): Cointegration of Consumption and Disposable Income: Evidence from
twelve OECD countries, Southern Economic Journal 62, 77-88.
Levin, A. and Lin, C.-F. (1993): Unit Root Tests in Panel Data: New Results, Discus-
sion Paper 93-55, University of California, San Diego.
 17

Maddala, G.S. and Wu, S. (1999): A Comparative Study of Unit Root Tests with Panel
Data and a new simple test Oxford Bulletin of Economics and Statistics, Special Issue,
631-652.
Muellbauer, J. (1994): The Assessment: Consumer Expenditure, Oxford Review of Eco-
nomic Policy 10, 1-41.
O'Connell, P.G.J. (1998): The Overevaluation of Purchasing Power Parity, Journal of
International Economics 44, 1-19.
Reimers, H.-E. (1997): Seasonal Cointegration Analysis of German Consumption Func-
tion, Empirical Economics 22, 205-231.
Taylor, M.P. and Sarno, L. (1998): The Behaviour of Real Exchange Rates during the
post-Bretton Woods Period, Journal of International Economics 46, 281-312.
Wolters, J. (1992): Der Zusammenhang zwischen Konsum und Einkommen: Alternative
ökonometrische Ansätze, RWI-Mitteilungen 43, 115-132.