Ž
.
Economic Modelling 16 1999 293
]306
Modelling consumption: permanent-income
or rule-of-thumb behaviour?
Dimitris Hatzinikolaou
1
The Flinders Uni
¤
ersity of South Australia, School of Economics, GPO Box 2100, Adelaide,
SA 5001, Australia
Accepted 9 December 1998
Abstract
A number of authors have recently estimated the percentage of aggregate consumption
accounted for by consumers who simply consume their current income in each period
instead of behaving according to the permanent-income-rational-expectations hypothesis.
The magnitude of this percentage has important policy implications. The existing models,
however, incorporate approximations that may destroy the consistency of its estimates. The
paper considers a model that does not share this problem. Using Greek annual aggregate
data, 1960
]1993, the model produces estimates of this percentage that are lower than those
produced by the existing models. Utility-loss calculations suggest that the differences may be
important.
Q 1999 Elsevier Science B.V. All rights reserved.
JEL classifications: E21; E62; H31
Keywords: Life cycle permanent income hypothesis; Rational expectations; Rule-of-thumb consumption;
Greece
1. Introduction
A major policy implication of the life-cycle-permanent-income hypothesis under
Ž
.
Ž
.
rational expectations LC-PIH-RE , as formulated by Hall 1978 , is that short-run
Ž
.
fiscal policy designed to influence private consumption e.g. a temporary tax cut is
1
Current address: University of Ioannina, Department of Economics, PO Box 1186, 45110 Ioannina,
Greece. Tel.:
q30 651 97336; fax: q30 651 97340. Permanent address. Tel.: q61 8 8201 2987; fax: q61
8 8201 5071; e-mail: dimitris.hatzinikolaou@flinders.edu.au
0264-9993
r99r$ - see front matter Q 1999 Elsevier Science B.V. All rights reserved.PII: S 0 2 6 4 -
Ž
.
9 9 9 3 9 8 0 0 0 4 4 - 3
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
294
doomed to failure. The importance of this implication has stimulated a large
number of authors to test empirically the LC-PIH-RE. Most tests reject Hall’s
formulation, however, and a new literature has emerged in an attempt to find a
modification of the LC-PIH-RE that is compatible with the data.
A branch of this literature abandons the representative-consumer assumption
and argues that the population actually consists of two groups of consumers: the
life-cyclers, who maximize their expected lifetime utility subject to their expected
life-time resources, and the ‘rule-of-thumbers’, who simply consume their current
income. Rule-of-thumb behaviour can be explained partly by the presence of
liquidity constraints and partly by ‘myopia’, i.e. the failure of some consumers to
form rational expectations about their future income and make optimal choices. A
Ž
number of studies Campbell and Mankiw, 1989, 1990, 1991; Jappelli and Pagano,
.
1989; Vaidyanathan, 1993 have recently focused on the estimation of the ‘excess-
sensitivity parameter’, which might be interpreted as the percentage of aggregate
consumption accounted for by the rule-of-thumbers and is an indicator of the
degree to which the LC-PIH-RE is incompatible with the data. Thus, the closer
this percentage is to unity the greater is the empirical failure of the LC-PIH-RE
and the greater is the scope for short-run countercyclical fiscal policy designed to
influence private consumption, the largest component of aggregate demand. The
empirical models used in these studies are modified Euler equations for consump-
tion that attempt to capture the behaviour of both groups of consumers. These
models, however, incorporate approximations that might destroy the consistency of
Ž
.
the estimates for the sake of simplicity. But, as Deaton 1992, p. 177 notes, ‘no
amount of mathematical convenience can compensate for the loss of substance’.
After reviewing the existing models, this paper considers an alternative model
that is highly non-linear, but avoids using unnecessary approximations. Using
Greek annual aggregate data, 1960
᎐1993, it obtains estimates of the excess
sensitivity parameter from all models and finds that the estimates produced by the
alternative model are lower than those produced by the existing models. Utility-loss
calculations suggest that the differences may be important.
A motivation for the paper is the tightening of the non-interest terms of
Ž
.
consumer credit by the Bank of Greece in 1996 OECD, 1996, pp. 47, 112 .
Consumer credit in Greece had been restricted until early 1994, when it was
virtually liberalised, a policy that led to a rapid growth of consumer loans, e.g.
Ž
.
64.6% in 1994 Bank of Greece, 1995, p. 115 . In addition to curbing inflation, the
recent tightening aims at preventing consumers from becoming ‘overindebted’
Ž
.
Ethnos, 1996 , which implies that the proportion of ‘myopic’ consumers may be
significant.
2. Basic definitions
To fix ideas, begin by defining the basic variables. First, let N , N , and N be
1 t
2 t
t
the numbers of life-cycle, of rule-of-thumb, and of all workers in the economy,
Ž
.
X
X
X
respectively N
s N q N . Second, let C , C , and C be life-cycle, rule-of-
t
1 t
2 t
1 t
2 t
t
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
295
Ž
X
X
X
.
X
thumb, and aggregate consumption C
s C q C . Third, let C s C rN ,
t
1 t
2 t
1 t
1 t
1 t
C
s C
X
rN , and C s C
X
rN be life-cycle, rule-of-thumb, and aggregate con-
2 t
2 t
2 t
t
t
t
sumption per worker. Thus, the identity C
X
s C
X
q C
X
can be written in per-worker
t
1 t
2 t
Ž
.
Ž
terms as C
s 1 y
C q C , where s N rN assumed to be constant over
t
1 t
2 t
2 t
t
.
X
X
time . Finally, let Y
and Y
be rule-of-thumb and aggregate disposable income.
2 t
t
In per-worker terms, let Y
s Y
X
rN and Y s Y
X
rN . By the definition of
2 t
2 t
2 t
t
t
t
rule-of-thumb consumption, we have C
s Y . In this paper, we assume that
2 t
2 t
disposable income per rule-of-thumb worker equals aggregate disposable income
per worker, i.e. Y
s Y , so that C s Y . This holds if the fraction of aggregate
2 t
t
2 t
t
disposable income that accrues to the rule-of-thumbers is exactly equal to
, i.e. if
Y
X
s
Y
X
.
2
2 t
t
3. Previous models
Assuming quadratic separable preferences and constant real interest rates,
Ž
.
Jappelli and Pagano 1989 derive the following modified Euler equation, which
allows for rule-of-thumb behaviour:
Ž
.
Ž .
C
s a q a C
q
Y y a Y
q e .
1
t
0
1
t
y1
t
1 t
y1
t
They estimate
for each of the seven countries in their sample by using aggregate
Ž
.
data and by applying non-linear instrumental variables NLIV as well as full-infor-
Ž
.
mation maximum likelihood FIML . As instruments, they use a linear trend and
the first lag of consumption, of disposable income, of government expenditure, and
of exports. However, using the le
¤
els of these variables, which may be non-sta-
tionary; lagging them only once; assuming that preferences are separable in
consumption and other goods; and that the real interest rate is constant may
destroy the consistency and the asymptotic normality of the estimates.
Ž .
Adding and subtracting the quantity C
q
Y
to Eq. 1 and then rearrang-
t
y1
t
y1
Ž
.Ž
.
ing gives
⌬C s a q a y 1 C
y
Y
q
⌬Y q e , where ⌬ is the differ-
t
0
1
t
y1
t
y1
t
t
Ž .
ence operator, e.g.
⌬C s C y C . This equation is equivalent to Eq. 1 . Note,
t
t
t
y1
however, that if C and Y are cointegrated unit-root processes, then it is a
t
t
Ž
.
Ž
.
non-linear error correction model ECM . By letting u be the residual from the
ˆ
t
Ž
.
regression of C on a constant and on Y
the cointegrating regression ,
can be
t
t
Ž
.
estimated by applying linear instrumental variables LIV to the following equa-
tion, using as instruments stationary variables lagged at least twice:
Ž .
⌬C s b q b u
q
⌬Y q e .
2
ˆ
t
0
1
t
y1
t
t
Ž
.
Using aggregate data for 59 countries, Vaidyanathan 1993 estimates
using Eq.
Ž .
Ž .
2 as well as restricted and unrestricted versions of Eq. 1 .
2
Ž
.
Campbell and Mankiw 1991 , among others, also assume that C
s Y , thus assuming Y s Y .
2 t
t
2 t
t
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
296
Ž
Since consumption and income seem to evolve as log-linear rather than as
.
linear stochastic processes and since it is desirable to allow for variable real
Ž
.
interest rates, Campbell and Mankiw 1989, 1990, 1991 prefer the model
Ž
.
Ž
.
Ž
.
Ž .
⌬log C s
q log 1 q r q ⌬log Y q ⑀ ,
3
t
t
t
t
where r is a real interest rate and
is the elasticity of intertemporal substitution.
t
Using aggregate data for a group of industrialised countries, they estimate variants
of this equation using LIV, emphasising that the instruments must be stationary
Ž .
variables lagged at least twice. Eq. 3 is a modification of the log-linearised Euler
equation that is derived by assuming separable preferences represented by a
Ž
.
standard constant-relative-risk-aversion
CRRA utility function. In their 1990
paper, which uses US data only, Campbell and Mankiw allow for non-separable
Ž
.
Ž
.
preferences by replacing log 1
q r with ⌬log X , where X s labour supply, or
t
t
government expenditure, or stock of consumer durables. These experiments con-
Ž .
firm their earlier conclusion, however, that their simple model, i.e. Eq. 3 , ‘will be
hard to beat as a description of the aggregate data on consumption, income, and
Ž
.
interest rates’ Campbell and Mankiw, 1989, p. 212 .
Despite its popularity, the log-linearisation of a Euler equation may not be an
innocuous transformation, however, because it will destroy the consistency of the
Ž
estimates if the untransformed Euler equation is the correct model Gallant, 1987,
.
p. 427 . Also, it requires the restrictive assumption that C and r be jointly
t
t
Ž
.
conditionally lognormal and homoskedastic see Deaton, 1992, p. 64 . Finally, as
Ž
.
Campbell and Mankiw 1991, p. 729 note, the interpretation of
as the fraction of
rule-of-thumb consumers is not exact in their log-linear model.
4. An alternative model
Consider a life-cycler whose one-period utility function is
␥
␣

1
y
␣y
Ž
.
Ž .
u
s C L G
y 1 r
␥ , ␥ / 0,
4
1 t
t
t
where L
s leisure per worker and G s government expenditure per worker,
assumed to be exogenous and to take on the same values for both groups of
Ž
.
consumers so there is no need to distinguish between L
and L , for example .
1 t
2 t
The assumption of exogeneity of L is based on the empirical observation that
t
Ž
hours of work are determined institutionally see Hatzinikolaou and Ahking, 1995,
.
Ž .
Ž
p. 1118 . The utility function 4 has already been used in the literature e.g.
.
Hatzinikolaou and Ahking, 1995 . It has the following characteristics. First, it
implies that the consumer considers a geometric weighted average of C , L, and G,
1
with weights
␣, , and 1 y ␣ y . Second, it reduces to other utility functions by
placing various restrictions on its parameters. For example, if
␥ s 0, it reduces to
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
297
logarithmic utility; and if
␣ s 1 and  s 0, it reduces to the standard CRRA
Ž
utility function, which has been widely used in the empirical literature see, e.g.
Mankiw, 1981; Hansen and Singleton, 1982; Campbell and Mankiw, 1989, 1990,
.
1991 . Third, if
␣ ) 0,  ) 0, and 1 y ␣ y  ) 0, it is strictly increasing in C , L,
1
Ž
.
and G non-satiation ; and if
␣ ) 0 and 1 y ␣␥ ) 0, it is strictly concave in the
Ž
.
choice variable C . Fourth, if
␥ ) 0 ␥ - 0 , in addition to the non-satiation
1
restrictions, it implies that C , L, and G are pairwise Edgeworth complements
1
Ž
.
3
substitutes . Finally, it leads to an estimating equation whose variables are likely
to be stationary, an empirically attractive property.
Ž
Under time separability, the first-order condition Euler equation for consump-
.
tion is
1
q r
t
Ž
.
␣␥y1
␥
1
y
␣y ␥
Ž
.
Ž
.
Ž
.
Ž .
E
C
rC
L
rL
G
rG
s 1,
5
t
1 t
q1
1 t
t
q1
t
t
q1
t
ž
/
1
q
␦
Ž
.
where
␦ is a subjective discount rate ␦ ) 0 and E is a mathematical expectation
t
conditional on information available at the beginning of time t.
4
The following
steps lead to an estimating equation that allows for rule-of-thumb behaviour. First,
define l
s L rL , g
s G rG , y
s Y rY , and apc s C rY . Sec-
t
q1
t
q1
t
t
q1
t
q1
t
t
q1
t
q1
t
t
t
t
Ž .
ond, remove E from 5 , add a prediction error, e
, to its right-hand side, and
t
t
q1
solve the resulting equation for C
to obtain C
s k
C , where k
is a
1 t
q1
1 t
q1
t
q1 1t
t
q1
non-linear function of l
, g
, e
, and r . Third, multiply C
s k
C
by
t
q1
t
q1
t
q1
t
1 t
q1
t
q1 1t
Ž
.
Ž
.
Ž
.
1
y
; add the quantity k
C y Y y C
y Y
, which is zero, since
t
q1
2 t
t
2 t
q1
t
q1
Ž
.
C
s Y ; use C s 1 y
C q C and the definition of k ; and rearrange, to
2 t
t
t
1 t
2 t
t
q1
obtain
1
q r
t
␣␥y1 ␥
Ž1
y
␣y .␥
w
Ž
.
Ž
.
x
Ž .
y
apc
y
r apc y
l
g
y 1 s e
.
6
t
q1
t
q1
t
t
q1 tq1
t
q1
ž
/
1
q
␦
Ž .
Ž
.
The variants of Eq. 3 considered by Campbell and Mankiw 1990 should, in
Ž .
Ž
.
principle, be special cases of the log-linearised version of 6 that uses log 1
q e
t
q1
f e . However, again, the meaning of
in these variants is different. To see
t
q1
Ž
.
Ž
.
why, consider Eq.
18
of Campbell and Mankiw
1991 , which in fact uses
Ž1
y
.
Ž
.
C
s C
Y
instead of using C
s 1 y
C q Y .
t
1 t
t
t
1 t
t
5. Econometric methodology
Ž . Ž .
Ž .
Eqs. 1
᎐ 3 and 6 are estimated by the generalised method of moments
Ž
.
GMM , which exploits the orthogonality conditions between the error term and
3
Ž
.
Ž
.
Two goods are Edgeworth complements substitutes if the marginal utility of one increases decreases
as the quantity consumed of the other increases.
4
Ž .
Ž
Setting up the problem completely and interpreting Eq. 5 is a standard exercise see, e.g. Hatzinikolaou
.
Ž
.
and Ahking, 1995 . Assuming Ricardian equivalence and applying Hall’s, 1978, p. 986 perturbation
Ž .
Ž .
argument, Eqs. 5 and 6 are derived in an appendix, which is available upon request. A non-Ricardian
model leads to the same Euler equation.
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
298
the variables in the information set. The results are then evaluated by constructing
several specification tests. The methods are briefly described here by referring to
Ž .
Eq. 6 .
Let
be a K = 1 parameter vector and z an M = 1 vector of instruments. For
Ž .
Ž
.
X
Eq. 6 ,
s
␣, , ␥, ␦, . Under the joint hypothesis of rational expectations
and of a correctly specified model, the error term e
has zero mean, is serially
t
q1
uncorrelated, and is uncorrelated with any information known at the beginning of
time t, when expectations are formed. Thus, permissible instruments are relevant
Ž
variables dated t
q 1 y h, where h G 2. See also Hall, 1988, p. 348; and Campbell
.
and Mankiw, 1990, p. 268. The joint hypothesis implies the orthogonality condi-
Ž
.
Ž .
tions E z
e
s 0 whose sample counterpart is denoted here by g s
t
q1yh tq1
Ž
.
T
1
rT ⌺
z
e
, where T is the number of observations that are available for
t
sh tq1yh tq1
estimation after the construction of growth rates and of lagged variables. Assuming
Ž
.
that M
G K a necessary condition for identification , estimation of and testing
˜
Ž .
of the joint hypothesis are based on g
. In particular, GMM chooses to
˜
˜
X
ˆ
y1
˜
ˆ
Ž .
Ž .
Ž .
minimize the quadratic form q
s g S g , where S is an M = M symmet-
ˆ
ric weighting matrix. In order to achieve positive semidefiniteness of the matrix S
ˆ
and to obtain autocorrelation-consistent standard errors, the definition of S is that
Ž
.
of Newey and West 1987, 1994 , which sets the value of the ‘lag truncation
w Ž
.
2
r9
x
w x
parameter’ equal to 4 T
r100
, where ‘ . ’ means ‘integer part of’.
Ž
The instrument vector used here is z
s constant, c
, l
, g
,
t
q1yh
t
q1yh
t
q1yh
.
X
y
, w
,
¤
, where c
s C rC is one plus the rate of growth in
t
q1yh
t
q1yh
t
q1yh
t
q1
t
q1
t
consumption, w
s W rW is one plus the rate of growth in a real wage, and
¤
t
q1
t
q1
t
Ž
.
is an ex-post real interest rate. Tauchen 1986 and Davidson and MacKinnon
Ž
.
1993, pp. 222 and 604 point out that in small samples efficiency gains from using
more instruments are obtained at the cost of a greater bias in the estimates. In the
light of this result, the values of h and M are chosen to be as small as possible so
as to ensure parameter identification while minimizing the bias. Thus, for Eqs.
Ž . Ž .
Ž
.
Ž .
1
᎐ 3 , the choice h s 2 is used so M s 7 and T s 31 ; whereas for Eq. 6 , the
Ž
.
Ž
.
choices h
s 2, 3 so M s 13 and T s 30 and h s 2, 3, 4 so M s 19 and T s 29
are also used, since choosing h
s 2 does not lead to statistically significant
estimates. The value of the ‘lag truncation parameter’ is set equal to 3, since
w Ž
.
2
r9
x
4 30
r100
s 3.
To evaluate the models statistically, the following tests are carried out. First, the
hypothesis that e
is serially uncorrelated is tested against the alternative
t
q1
hypothesis of ith order serial correlation by estimating Gauss-Newton regressions
Ž
and by calculating the values of an F-statistic, denoted by F
see Davidson and
ARi
5
Ž .
As an example, here is how F
for Eq. 6 is constructed. Let R be the series of the GMM residuals;
AR 2
Ž
R1 and R2 be the first two lags of R, where the lost observations are set to zero so the series R1 and
.
Ž
.
R2 also have T observations ; R1H and R2H be the fitted values from the ordinary least squares OLS
regressions of R1 and R2 on the instruments z; and XDH, XAH, XBH, XGH, and XLAH be the fitted
Ž .
values from the OLS regressions of the derivatives of the non-linear part of 6 with respect to
˜
Ž
.
evaluated at
on z. The statistic F
is calculated as the F-test for the hypothesis that the
AR 2
coefficients of R1H and R2H in the OLS regression of R against XDH, XAH, XBH, XGH, XLAH,
Ž
.
R1H, and R2H without a constant are zero.
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
299
Ž
.
5
Ž
.
MacKinnon 1993, Sections 7.7 and 10.10 . Note that when h
s 2 M s 7 is
chosen and the value of F
is significant, then the validity of the instruments is
AR2
questionable, since the error term may be correlated with the instruments. For the
Ž .
Ž
.
same reason, when Eq. 6 is estimated with h
s 2, 3 M s 13 the statistic F
is
AR3
Ž
.
used as an additional test; and when it is estimated with h
s 2, 3, 4 M s 19 the
statistics F
and F
are also used. Second, the validity of the instruments is
AR3
AR4
checked by regressing the GMM residuals against the instruments and by calculat-
2
2
Ž
ing the adjusted R , denoted by R
, and the marginal level of significance or
e
⭈z
.
p-value of the F-statistic for the hypothesis that all the slope coefficients in these
regressions are zero. Third, since M
) K, the joint hypothesis is tested by using the
well-known test of overidentifying restrictions, also an instrument-validity test.
6
˜
Ž .
The relevant statistic is defined as J
s T = q and is asymptotically distributed
2
Ž
.
as
. Newey 1985 shows, however, that the statistic J often fails to reject a
M
y K
misspecified model, because it has little power against some local alternatives, so it
can lead to the acceptance of GMM estimates that are in fact inconsistent. In
Ž
particular, it may easily fail to detect parameter instability Newey, 1985, pp.
.
239
᎐240 .
Ž
.
Thus, as a fourth test, a likelihood-ratio LR type statistic is constructed to test
the hypothesis of structural stability. The test is discussed in Ghysels and Hall
Ž
.
Ž .
1990, pp. 127
᎐128 and can be implemented as follows. Again, refer to Eq. 6 , but
Ž . Ž .
Ž .
note that Eqs. 1
᎐ 3 must be written in the form of 6 , i.e. with the error term as
the only variable on the right-hand side, before applying the test to them. Suppose
that we want to test for a structural break in period t . Let D be a dummy variable
0
t
defined as D
s 0 for t - t and D s 1 for t G t . Also, let T and T denote the
t
0
t
0
1
2
number of observations available for estimation in the first and in the second
Ž
.
subsample T
q T s T . Parameter instability, if any, is assumed to occur in the
1
2
Ž .
second subsample. First, use the instrument vector z to estimate 6 with the first
T observations only and let J be the value of the statistic J obtained from this
1
1
Ž .
estimation. Second, construct two equations, one by multiplying Eq.
6
by
Ž
.Ž
.
Ž
.
T
rT 1 y D and the other by multiplying it by T rT D , and use z to estimate
1
t
2
t
them as a system. Let J
be the value of J from this estimation. The statistic
2
LR
s J y J is asymptotically distributed as
2
and can be used to test for
t 0
2
1
M
structural stability. All estimates and tests are derived using the econometric
computer programme RATS, version 4.2.
6. The data
The sources of the data are given in Appendix A. The empirical definitions of
the variables are as follows. First, C
s national private consumption expenditure
on non-durables, services, and semidurable goods, divided by the consumer price
6
Note that testing for a zero-mean error term in each equation is part of the test of overidentifying
restrictions, since the instrument set includes a constant.
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
300
Ž
.
Ž
.
index CPI and by total employment N . Second, leisure per worker is measured
as L
s 5840 y 50 = HR , where HR s hours of work per week in manufacturing,
t
t
Ž
.
assuming that the time endowment per worker per year is 5840
s 365 = 16 hours
and that the average worker works 50 weeks per year. Third, G
s central govern-
ment expenditure, excluding payments for amortization and interest on the public
debt, social security,
7
and transfers to the rest of the world, divided by CPI and by
N. Fourth, Y
s income of households, including private non-financial unincor-
porated enterprises, less direct taxes and social-security contributions, divided by
CPI and by N. Fifth, W
s weighted average compensation per hour of male and
female wage earners in all manufacturing firms employing 10 or more persons,
Ž .
deflated by CPI. Sixth, the ex-post real interest rate
¤
is the rate offered by
commercial banks for 3
᎐12 month demand, time, and regular savings deposits in
Ž
.
Ž .
drachmas less the CPI inflation rate
. Seventh, the ex-ante real interest rate r
Ž
e
.
is the nominal interest rate just defined less expected inflation
.
A series for
e
is constructed recursively by regressing
on lagged variables
t
Ž
e.g. the rate of growth in money supply, in labour productivity, in the US
.
dollar
᎐drachma exchange rate, and in import prices in US$ and by taking
one-step-ahead forecasts.
8
To produce each value of
e
, the sample is updated and
also possibly the regressor set, so that the forecasting regression fits the data well
and passes a series of diagnostic tests at the 5% level of significance. The series
e
Ž
e
.
fails one test of rationality at the 5% level, namely the forecast error
y
t
t
fails to be a white-noise process, but passes that test at the 1% level and passes all
Ž
.
other tests of rationality at the 5% level see Lovell, 1986 .
Ž
.
Finally, reliable data on N are available only for 1961 census year and from
1966 onward. The values of N for 1960 and 1962
᎐1965 are obtained by regressing
Ž
.
⌬ N on a constant and on emigration, using data for the period 1966᎐1989. This
t
regression assumes that the massive emigration from Greece in the 1960s was the
reason for the negative trend in N in that decade.
A possible structural break might have occurred in 1981, when the socialists won
a national election. Since the share of central government expenditure in GDP
grew from approximately 23% in 1980 to approximately 46% in 1993 and since
Ž . Ž .
government expenditure is absent from Eqs. 1
᎐ 3 , a structural-stability test is
carried out by choosing t
s 1981.
0
7. Results
Ž . Ž .
Table 1 reports the results from the estimation and testing of Eqs. 1
᎐ 3 and
Ž .
6 . Each column reports parameter estimates, their t-ratios, an estimate of loss of
7
Social security benefits are excluded from the definition of government spending in order to avoid
Ž .
double counting in 4 . If they are included, the empirical results remain virtually unchanged, however.
8
Ž
.
Most studies either assume that the real interest rate is constant, e.g. Hall 1978 , or use an ex-post
Ž
.
Ž
.
real interest rate, e.g. Mankiw 1981 . An exception is Hall 1988 , which uses an ex-ante real rate.
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
301
Table 1
Ž . Ž .
Ž .
GMM estimation of Eqs. 1
᎐ 3 and 6 ᎏ unrestricted and restricted
Ž .
Ž .
Ž .
Ž .
Ž .
Ž .
Ž .
Ž .
Equation
1
2
3
6
6
6
6
6
Restrictions None
None
None
None
None
None
␣ s 1,  s 0
s 0
M
7
7
7
7
13
19
13
13
U
UUU
UU
␣
᎐
᎐
᎐
0.29
0.38
0.26
᎐
0.37
˜
Ž
.
Ž
.
Ž
.
Ž
.
0.59
1.66
2.60
2.20
U
UUU
UU
˜

᎐
᎐
᎐
0.52
0.47
0.57
᎐
0.48
Ž
.
Ž
.
Ž
.
Ž
.
0.62
1.63
5.30
2.46
UUU
UUU
UU
␥
᎐
᎐
᎐
0.94
1.03
1.62
0.52
1.18
˜
Ž
.
Ž
.
Ž
.
Ž
.
Ž
.
0.40
1.47
3.37
4.04
2.18
UU
␦
᎐
᎐
᎐
y0.006 y0.001
0.006
y0.006
0.001
Ž
.
Ž
.
Ž
.
Ž
.
Ž
.
y0.42
y0.15
2.06
y1.08
0.16
UUU
UUU
UUU
UU
UUU
UUU
˜
0.39
0.52
0.71
0.34
0.27
0.19
0.36
᎐
Ž
.
Ž
.
Ž
.
Ž
.
Ž
.
Ž
.
Ž
.
3.91
4.50
7.33
1.21
1.96
2.89
3.26
Utility loss
0.12
0.28
30.71
0.09
0.07
0.05
0.10
᎐
w
x
w
x
w
x
w
x
w
x
w
x
w
x
w
x
F
0.80
0.82
0.63
0.25
0.50
0.88
0.28
0.43
AR 1
w
x
w
x
w
x
w
x
w
x
w
x
w
x
w
x
F
0.82
0.63
0.86
0.29
0.80
0.98
0.54
0.71
AR 2
w
x
w
x
w
x
w
x
F
᎐
᎐
᎐
᎐
0.84
0.70
0.68
0.82
AR 3
2
R
y0.17
y0.16
y0.22
y0.12
y0.41
y0.94
y0.34
y0.41
e . z
w
x
w
x
w
x
w
x
w
x
w
x
w
x
w
x
0.94
0.92
0.99
0.82
0.98
0.99
0.95
0.98
w
x
w
x
w
x
w
x
w
x
w
x
w
x
w
x
J
0.55
0.31
0.96
0.19
0.84
0.93
0.91
0.83
w
x
w
x
w
x
w
x
w
x
w
x
w
x
w
x
LR
0.42
0.53
0.71
0.64
0.49
0.69
0.39
0.43
1981
Note.
UUU
significant at the 1% level;
UU
significant at the 5% level;
U
significant at the 10% level. The
numbers in parentheses are t-ratios; those in brackets are p-values.
Ž
utility if a consumer behaves as a rule-of-thumber rather than as a life-cycler see
.
the next section , and information on the diagnostic tests. Since the focus of the
paper is the magnitude of the parameter
, the table does not report estimates of
Ž . Ž .
the other parameters of Eqs. 1
᎐ 3 , but reports estimates of all of the parameters
Ž .
of Eq. 6 , since the latter is under evaluation.
Before discussing the results, the following comments are in order. First, given
the small size of the sample, it is difficult to test the stationarity assumption
Ž
.
Ž . Ž .
Ž .
required by GMM . For the variables of Eqs. 2 , 3 and 6 , stationarity seems to
be defensible, however. In particular, the variables used for the estimation of Eq.
Ž .
6 have no obvious trends and the unit-root hypothesis can be rejected at the 5%
level for c, l, g, y, w, r, and
¤
, though only at the 10% level for apc.
9
Second, the
9
The unit-root hypothesis for the variables c, l, g, y, and w is tested with augmented Dickey-Fuller
Ž
.
tests. These tests also suggest that the variables c, g, y, and w may have a mild negative trend. For r
and
¤
, that hypothesis is tested by regressing each variable on its first lag only and by using the critical
Ž
.
values of the top part of Fuller’s Table 8.5.2 Fuller, 1976, p. 373 . A constant and a time trend are not
significant, but in the case of
¤
, unless they are excluded the hypothesis cannot be rejected. Finally, in
the case of apc, a constant is included, so a critical value is obtained from the middle part of Fuller’s
table.
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
302
tests on the parameters assume that each estimate is a random drawing from a
normal distribution, but for non-linear models and for small samples convergence
to the normal distribution may be problematic.
The basic results can be summarized as follows. First, the four models perform
reasonably well, since parameter estimates lie in the expected range
10
and the
diagnostic tests provide no evidence of misspecification for any of the equations.
11
Ž .
Note, in particular, that the estimates of the parameters of Eq. 6 , unrestricted as
well as restricted, satisfy the non-satiation and concavity restrictions and suggest
that consumption, leisure, and government expenditure are Edgeworth comple-
ments. Second, unless we impose restrictions on the parameters, choosing M
s 7
Ž .
for Eq. 6 causes an identification problem, reflected in the failure of all of its
estimates to be statistically significant. Also, choosing M
s 13 leads to estimates of
␣ and  that are significant only at the 10% level.
12
Without imposing any
restrictions, only when M
s 19 are all estimates significant at the 5% level. Third,
whenever statistically significant and unrestricted, the estimates of
obtained from
Ž .
Eq. 6 range from 0.19 to 0.27, so they are lower than those obtained from Eqs.
Ž . Ž .
1
᎐ 3 , which range from 0.39 to 0.71, and those reported by Jappelli and Pagano
Ž
.
Ž
.
Ž
.
13
1989 , 0.54 NLIV and 0.60 FIML .
Based on the information of Table 1, however, it is not possible to decide which
estimate of
is more reliable; and further testing does not lead to conclusive
evidence, either. The following observations might be suggestive, however. First, as
Ž .
the last column of Table 1 shows, when the restriction
s 0 is imposed on Eq. 6 ,
the diagnostic tests do not provide evidence of misspecification, which suggests that
the value of
may be low. Non-rejection of the model might simply reflect the low
power of the tests, however, a consequence of the small sample size. Second, the
˜
Ž
.
Ž .
coefficient of log 1
q r in Eq. 3 ,
s 0.20, has a t-ratio of 2.34. In addition, if r
t
t
Ž .
is treated as constant in Eq. 6 , there is evidence of misspecification. Thus,
Ž .
Ž .
treating r as a constant in Eqs. 1 and 2 might cause the estimates to be
t
inconsistent. Treating it as a variable in these equations has little effect on the
Ž .
results, however. Third, the results from Eq. 6 provide some evidence that the
10
Ž .
In the case of the unrestricted Eq. 6 , the estimate of
␦ is negative when M s 7 and when M s 13,
Ž
.
but is statistically insignificant. When M
s 19, it has the correct sign positive and is significant.
11
Note the following results, which are not reported in Table 1. First, although the data provide some
Ž .
evidence that C and Y are cointegrated, the error-correcting term in Eq. 2 fails to be statistically
significant. Second, the possibility of a structural break in 1974 was also considered for the reasons
Ž
.
explained in Hatzinikolaou and Ahking 1995, footnote 10 ; but, again, no evidence of structural
Ž .
instability was found for any of the equations. Third, when Eq. 6 is estimated with M
s 19, the
p-value of the statistic F
is 0.49.
AR 4
12
Note that
␣, , and are expected to be positive. Thus, for example, a t-ratio of 1.66 for the
hypothesis H :
␣ s 0, is significant at the 10% level, since the alternative hypothesis is H : ␣ ) 0. Note
0
1
also that it is possible to obtain estimates that are significant at the 5% level when M
s 13 by imposing
restrictions that are supported by the data, e.g.
 s 2␣.
13
Ž
.
Ž
.
Using consumption expenditure which includes spending on durable goods , Vaidyanathan 1993
produces several estimates of
, claims that they are similar, and reports the one that has the smallest
standard error, without making clear which model generates it. For Greece, that estimate is 0.36.
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
303
variables L and G may be important,
14
so excluding them might lead to inconsis-
Ž .
tent estimates. When these two variables are excluded from Eq. 6 , by imposing
the restrictions
␣ s 1 and  s 0, there is still no strong evidence of misspecifica-
tion, but the estimates of
are larger than in the unrestricted regressions. In
˜
Ž
particular, when M
s 13, we get
s 0.36 see the second-to-last column of Table
˜
.
Ž
1 ; and when M
s 7, we get
s 0.49 with a t-ratio of 5.58 this regression is not
.
reported in the table because it produces insignificant estimates of
␥ and ␦ . This
finding casts some doubt on the reliability of high estimates of
. In addition,
Ž .
Ž
.
Ž
.
despite the significance of L and G in Eq. 6 , the variables
⌬log L and ⌬log G
t
t
Ž .
are insignificant in Eq. 3 ; and their presence causes the estimate of
to be
somewhat lower, namely 0.64. This finding suggests that the log-linearisation of the
Euler equation might not be an innocuous transformation.
Finally, note that a low estimate of the percentage of rule-of-thumb consumers is
Ž
consistent with the high net household saving rate in Greece relative to other
.
Ž
OECD countries as well as with the possibility to borrow from relatives thanks to
.
strong family ties , which reduces the severity of liquidity constraints for many
consumers.
8. Utility loss of rule-of-thumb behaviour
Do the differences in the estimates of
matter? If yes, then the choice of model
becomes crucial. An answer can be derived from utility-loss calculations.
15
Fol-
Ž
.
lowing Campbell and Mankiw 1991, pp. 747
᎐749 , it is possible to measure
approximately the utility loss to a rule-of-thumber, who chooses C
s Y in each
2 t
t
period instead of choosing the optimal level of consumption, C
U
, the level chosen
t
Ž
U
.
by a life-cycler i.e. C
s C . The monetary value of the loss in each period,
1 t
t
expressed as a percentage of C
U
, is approximately given by
t
2
U
1
C
y C
2 t
t
Ž .
E
,
7
t
U
ž
/
2
C
t
where
s yu C
U
ru , u s
⭸ ur⭸ C , and u s ⭸
2
u
r
⭸ C
2
, which are evalu-
t
C C
t
C
C
2 t
C C
2 t
U
Ž
.
U
ated at C . As was shown earlier, C
s 1 y
C q C ; and since C s C and
t
t
1 t
2 t
1 t
t
U
Ž
Ž
..Ž
.
Ž
U
.
U
Ž
C
s Y , it follows that C s 1r 1 y
C y Y and C y C rC s 1 y
2 t
t
t
t
t
2 t
t
t
. Ž
.
Ž .
apc
r apc y
. Thus, the sample counterpart of the quantity in 7 is
t
t
2
T
1 1
1
y apc
t
Ž .
.
8
˜
Ý
t
ž
/
˜
2 T
apc
y
t
s1
t
14
The hypothesis H :
 s 0 can be rejected in three out of four regressions; whereas the hypothesis
0
H : 1
y
␣ y  s 0 can be rejected only when M s 19, in which case the p-value of the Wald test is
0
0.0008.
15
I am grateful to an anonymous referee for making this suggestion.
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
304
If the utility function is quadratic then
depends on the level of ‘bliss
˜
t
Ž
.
consumption’, which may be difficult to estimate, see Cochrane 1989, pp. 327
᎐328 .
Ž .
Alternatively, if utility is given by Eq. 4 , then
s 1 y ␣␥. Thus, using the
˜
˜˜
t
estimates
␣ s 0.26 and ␥ s 1.62, we get s 0.58.
˜
˜
˜
t
Ž .
The utility-loss calculations reported in Table 1 are derived from quantity 8
˜
s 0.58 using and the various values of . Note that this measure of loss of utility
˜
t
Ž
.
corresponds to measure L of Campbell and Mankiw 1991 , but avoids using the
3
Ž
.
approximations used in the last line of their Eq. 22 . Also, note that it is an
˜
˜
˜
increasing function of
, since apc ) for all values of t and of considered
t
here. Thus, although the loss of utility is small for the estimates of
generated by
˜
Ž
.
the alternative model e.g. only 0.05% of optimal consumption when
s 0.19 , it is
significant for some of the estimates of
obtained from the other models. In
˜
particular, for
s 0.71 the loss is almost 31%, which is a serious loss of utility.
9. Summary and conclusion
The focus of the paper has been the estimation of the percentage of rule-of-
thumb consumers-workers using Greek aggregate data and four different models,
three of which already exist in the literature. For Greece, the existing estimates of
this percentage range from 0.36 to 0.60. The existing models incorporate approxi-
mations that might destroy the consistency of the estimates, however. The paper
considers an alternative model that avoids using unnecessary approximations.
Here, the estimates of this percentage obtained from the existing models range
from 0.39 to 0.71; whereas those from the unrestricted alternative model, whenever
they are statistically significant, are lower, ranging from 0.19 to 0.27. The paper
does not produce conclusive empirical support for the alternative model, but
provides reasons why its estimates may be more reliable.
Different estimates of the percentage of rule-of-thumb consumers lead, of
course, to different policy conclusions. For example, temporary tax cuts will be
much less effective in stimulating consumption if this percentage is only 0.19 than
if it is 0.71. As another example, the recent tightening of consumer credit in
Greece, which is expected to raise this percentage, may cause only a small decline
in utility if the tightening raises it from 0.20 to 0.30, but may cause a serious
decline in utility if it raises it to 0.70.
Acknowledgements
I am grateful to two anonymous referees of this journal and to my colleague
Peter Wagstaff for their comments and suggestions, which improved the paper
significantly. The usual disclaimer applies.
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
305
Appendix A: Data sources
The data are available from the author upon request. Their sources are as
follows: household income, direct taxes, social security contributions, consumption,
Ž
saving, government expenditure, transfers to the rest of the world OECD, Natio-
.
nal Accounts, vols. I and II ; CPI, GDP, interest rate, exchange rate, money supply,
Ž
import prices in US dollars International Financial Statistics, series 64, 99b, 60l,
.
Ž
ahx, 34, 75d ; total employment OECD, 1987, 1991, 1992, Report of the Governor
.
Ž
of the Bank of Greece, Statistical Yearbook of Greece ; emigration United
.
Ž
Nations, Demographic Yearbook ; hours of work, wage rate International Labor
.
Office, Yearbook of Labor Statistics ; interest payments on the public debt, price
Ž
level of imported machinery and transportation equipment in drachmas Bank of
.
Greece, 1992, Economic Bulletin of the Bank of Greece ; productivity of labour
Ž
.
Summers and Heston, 1991 .
References
Ž
.
Bank of Greece, 1992. Long-run Statistical Series of the Greek Economy in Greek . Athens.
Bank of Greece, 1995. Report of the Governor for the Year 1994. Athens.
Campbell, J.Y., Mankiw, N.G., 1989. Consumption, income, and interest rates: reinterpreting the time
Ž
.
series evidence. In: Blanchard, O.J., Fischer, S. Eds. , NBER Macroeconomics Annual 1989. MIT
Press, Cambridge.
Campbell, J.Y., Mankiw, N.G., 1990. Permanent income, current income, and consumption. J. Bus.
Econ. Stat. 8, 265
᎐279.
Campbell, J.Y., Mankiw, N.G., 1991. The response of consumption to current income: a cross-country
investigation. Eur. Econ. Rev. 35, 723
᎐756.
Cochrane, J.H., 1989. The sensitivity of tests of the intertemporal allocation of consumption to
near-rational alternatives. Am. Econ. Rev. 79, 319
᎐337.
Davidson, R., MacKinnon, J., 1993. Estimation and Inference in Econometrics, Oxford University Press,
New York.
Deaton, A., 1992. Understanding Consumption. Oxford University Press, New York.
Ethnos, 5 May 1996.
Fuller, W.A., 1976. Introduction to Statistical Time Series. John Wiley, New York.
Gallant, R.A., 1987. Non-linear Statistical Models. John Wiley, New York.
Ghysels, E., Hall, A., 1990. Are consumption-based intertemporal capital asset pricing models struc-
tural? J. Econom. 45, 121
᎐139.
Hall, R.E., 1978. Stochastic implications of the life cycle-permanent income hypothesis: theory and
evidence. J. Pol. Econ. 86, 971
᎐987.
Hall, R.E., 1988. Intertemporal substitution in consumption. J. Pol. Econ. 96, 339
᎐357.
Hansen, L.P., Singleton, K.J., 1982. Generalized instrumental variables estimation of non-linear rational
expectations models. Econometrica 50, 1269
᎐1286.
Hatzinikolaou, D., Ahking, F.W., 1995. Government spending and consumer attitudes toward risk, time
preference, and intertemporal substitution: an econometric analysis. S. Econ. J. 61, 1117
᎐1126.
Jappelli, T., Pagano, M., 1989. Consumption and capital market imperfections: an international
comparison. Am. Econ. Rev. 79, 1088
᎐1105.
Lovell, M.C., 1986. Tests of the rational expectations hypothesis. Am. Econ. Rev. 76, 110
᎐124.
Mankiw, G.N., 1981. The permanent income hypothesis and the real interest rate. Econ. Lett. 7,
307
᎐311.
Newey, W.K., 1985. Generalized method of moments specification testing. J. Econom. 29, 229
᎐256.
(
)
D. Hatzinikolaou
rEconomic Modelling 16 1999 293᎐306
306
Newey, W.K., West, K.D., 1987. A simple, positive semi-definite, heteroskedasticity and autocorrelation
consistent covariance matrix. Econometrica 55, 703
᎐708.
Newey, W.K., West, K.D., 1994. Automatic lag selection in covariance matrix estimation. Rev. Econ.
Stud. 61, 631
᎐653.
Organization for Economic Cooperation and Development, 1987, 1991, 1992. Economic Outlook 42, 49,
51, OECD, Paris.
Organization for Economic Cooperation and Development, 1996. Economic Surveys 1995
᎐1996. Greece
OECD, Paris.
Ž
.
Summers, R., Heston, A., 1991. The Penn World Table Mark 5 : an expanded set of international
comparisons, 1950
᎐1988. Q. J. Econ. 106, 1᎐41.
Tauchen, G., 1986. Statistical properties of generalized method-of-moments estimators of structural
parameters obtained from financial market data. J. Bus. Econ. Stat. 4, 397
᎐416.
Vaidyanathan, G., 1993. Consumption, liquidity constraints and economic development. J. Macroecon.
15, 591
᎐610.