Sunteți pe pagina 1din 19

Review of Economic Studies (1993) 60, 631-649 0034-6527/93/00320631$02.

00
© 1993 The Review of Economic Studies Limited

Consumption Growth, the


Interest Rate and Aggregation
ORAZIO P. ATTANASIO

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
Stanford University and CEPR
and
GUGLIELMO WEBER
Unioersita di Venezia and IFS

First version received May 1989; final version accepted January 1993 (Eds.)

In this paper we present empirical evidence on aggregation problems with Euler equations
for consumption. Our main results are: estimates of the elasticity of intertemporal substitution
for consumption are consistently lower for aggregate data than for average cohort data and the
theoretical model is statistically rejected on aggregate data, not rejected on average cohort data.
In trying to explain these differences we find that a major role is played by the non-linearity of
the estimable equation and by omitted demographic factors (normally unobservable on aggregate
data). However, even when these sources of aggregation bias are corrected for, the estimates of
the elasticity of intertemporal substitution obtained from aggregate data remain lower than those
obtained from average cohort data, and excess sensitivity tests reject the implications of the model.
This can be explained as the result of imposing identical coefficients to cohorts who differ in
preferences and/or opportunity sets.

1. INTRODUCTION
The relationship between consumption growth and the real interest rate has attracted
considerable interest, largely because under suitable assumptions on consumer preferences
and the stochastic properties of the two processes, it can be used to estimate the elasticity
of intertemporal substitution. In particular, if preferences are time and group additive,
utility is isoelastic and consumption growth and the real interest rate are jointly log-
normally distributed, we have:
a log (CHI) = intercept + O'rH I + 6'+1' (1)
where C denotes (real) expenditure on a group of commodities (normally on non-durable
goods and services), r is the real interest rate (computed as the logarithm of 1 plus the
nominal interest rate, net of the inflation rate), and 0' is the elasticity of intertemporal
substitution. The intercept depends on conditional second moments of consumption
growth and real interest rate, and the error term 6,+1 is uncorrelated with information
dated t or earlier!
Relation (1) is usually estimated on aggregate data. However, its derivation relies
on the equality of each individual intertemporal rate of substitution with the (economy-
wide) marginal rate of transformation. The conditions under which relation (1) holds at
the macro level may not be met in practice.'
1. For a derivation in an expected utility framework, see Hansen and Singleton (1983). As Hall (1988)
argues, a similar relation holds in models where risk aversion and intertemporal substitution are allowed to
differ. Attanasio and Weber (1989) substantiate Hall's claim within the class of preferences used by Epstein
and Zin (1989).
2. Grossman and Shiller (1982) derive conditions under which equation (1) aggregates across
heterogeneous consumers.
631
632 REVIEW OF ECONOMIC STUDIES

The availability in the U.K. of a long time series of cross sections offers the unique
opportunity of comparing estimates obtained on aggregate data with those derived from
micro data. The main aim of this paper is to determine to what extent and possibly why
these results differ.
The data used in this paper are obtained from the Family Expenditure Survey and
are described in Section 2. In that section we also discuss some econometric problems.
In Section 3 we compare estimates of (1) obtained on both aggregate and cohort
data to establish if there are systematic differences and how these can be detected. All

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
our estimates are based on quarterly time series over the 17-year period 1970-86. The
sample period is dictated by the availability of the survey data. Our results indicate that
there are substantial differences between estimates of the elasticity of intertemporal
substitution obtained from aggregate and average cohort data. They also reveal that
standard tests of serial correlation of the residuals and the overidentifying restrictions
are sensitive to incorrect aggregation of non-linear relations while "excess sensitivity"
tests can be particularly useful to detect other sources of aggregation bias.
In Section 4 we discuss different aggregation problems. We first investigate the role
of non-linearities. Equation (1) requires that we take the sum ofthe logarithm of individual
expenditures, not the logarithm of the sum. In so far as the cross-sectional distribution
of expenditure varies over time, the two variables will differ in a non-constant way. We
show that in our data this difference, which is also known as Theil's entropy measure, is
strongly serially correlated, and thus responsible for high-order serial correlation in the
disturbances of equations based on arithmetic averages of the data. Another source of
aggregation bias is given by omitted demographic factors. In general, the marginal rate
of intertemporal substitution will depend on such demographic characteristics as family
composition, education and, possibly, employment status. This would be reflected either
in a demographic-dependent intercept or in a time-varying a. Hence, at the individual
level we can either control for such factors or exclude those individuals whose circum-
stances are likely to vary over the sample period (e.g. those households whose head is
close to retirement age). At the aggregate level, such factors may be unobservable or
hard to identify.' Aggregation problems may also arise if the same coefficients are imposed
across different cohorts. In subsection 4.3 we show that the effects of socio-demographic
factors are not the same for different cohorts. This may reflect differences in preferences
or in opportunity sets: the fact that for cohorts who are young or old within the sample
period there are signs of misspecification could be interpreted as evidence of financial
markets imperfections. Finally, we discuss the effects of finite lives and missing markets
on the estimation of an equation like (1).

2. DATA AND ECONOMETRIC ISSUES


In our empirical application we use British data drawn from the National Accounts (eSO
Data Bank) and from the Family Expenditure Survey. While the first source is commonly
employed in applied macroeconomic work, the latter has been employed mainly by
micro-researchers, thus deserving some discussion.
The British Family Expenditure Survey (FES) is run every year on a randomly
selected sample of around 7000 households. The original sampling design covers about

3. Blundell, Browning and Meghir (1989) thoroughly investigate the role of demographics and employment
status variables in Euler equations for consumption, by making (T a function of time-varying characteristics.
This approach leads to even greater aggregation problems, as one needs to take the sums of expenditures across
different demographic groups.
ATIANASIO & WEBER CONSUMPTION GROWTH 633

10,000 households, hence little less than a third of the households fail to respond. The
FES has been extensively studied and used by economists in the past. Atkinson and
Micklewright (1983), for example, thoroughly addressed the issue of grossing up income
data from the FES, and reached the conclusion that the only types of income where
under-reporting is substantial are investment and self-employment income. On consump-
tion data, under-reporting is noticeable only on alcohol, a relatively small item. Expen-
diture on other items is thought to be accurately recorded, thanks to the careful sampling
design. Each household is in fact interviewed twice over a two-week period and is asked

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
to produce the latest gas, electricity and telephone bills, records of rent paid, of lumpy
purchases and of any credit-financed expenditure. On top of answering questions on
work-related matters and on expenditure on durable goods, each adult household member
fills in a detailed diary.
In our application we use 17 years (1910-86) of FES data to construct quarterly
series on consumption, C, income, ~ and prices P. For each household we take total
expenditure, PC, on non-durable goods and service (exclusive of housing services), net
household income PY (inclusive of earnings, self-employment income, pensions, social
security; exclusive of investment income and imputed rent from owner occupation) and
a Divisia price index based on the same eight broad groups of commodities included in
the consumption measure and on their published retail price indices, v: We then take
geometric means of each variable over each quarter and construct the following time-
series data:

(1) average cohort data, where the cohort includes all households whose head was
born in the 1930-40 interval (the mean sample size is 309 households per quarter);
(2) average FES data, including all participant households (mean sample size: 1710).

For consistency with aggregate National Accounts (NA) data, we also produce
average FES data by taking simple arithmetic means.'
In all cases, we also retrieve some socio-demographic variables. For each household,
we compute the total number of adult members, Adlt, and construct a multiple adult
indicator for households with more than two adults, Ma; we also create two variables
relating to children: the number of all members aged 0-17, nehil, and the number of
children of pre-school age (0-5), ykids. Two labour market variables are also used: an
out-of-work indicator for the head (taking value 1 if the head of household is out of
work), DOW, and an employment indicator for the second adult (a "working wife" dummy),
ww. We would also like to control for effects of education on preferences. Unfortunately,
the Family Expenditure Survey contains an educational code only since 1979. However,
given the high correlation between education and non-manual jobs, we construct an
educational indicator based on the occupation of the head of household. If the head is
a "white collar" worker (defined as professional, managerial, administrative, teaching or
clerical), the variable Whe takes the value 1 (zero otherwise).
The aggregate data we use are as close as possible to the survey data available. In
particular, we take consumers' expenditure over the same set of commodities as the

4. The Divisia index is defined recursively: its growth rate is the budget-share weighted average of the
commodity-specific inflation rates. Because the weights refer to the previous period and the formula is recursive.
it cannot be computed at the household level. except on panel data. However. it has desirable theoretical
properties. including base independence.
5. The Section 4 of this paper also uses data for two other cohorts. as well as geometric averages of
expenditure by household whose head's age falls within certain bands. We leave a detailed discussion of these
data until later.
634 REVIEW OF ECONOMIC STUDIES

consumption variable and use personal disposable income as the income variable. The
price index we adopt is the appropriate implicit deflator; we also experimented with an
RPI-based index and found no substantial differences.
As for asset returns, we use the interest rate on building society deposits. This type
of interest-bearing deposit is particularly attractive on a number of grounds. First of all,
it was the most commonly held asset over the sample period (over 50% of FES respondents
indicate interest income of this type). Secondly, the quoted interest rate is net of tax for
standard rate taxpayers (the great majority of households). Thirdly, it is an asset where

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
negative holdings are common, in the form of mortgages: tax-exemption rules imply that
for many households, after-tax lending and borrowing rates are very close to each other
(in 1984 the interest paid on the first £30,000 of most mortgages was tax-deductable;
typically, only young households would have a larger outstanding mortgage).
Finally, we also use NA data to construct a productivity growth measure as close as
possible to the ones normally adopted in the macroeconomic literature. We follow Layard
andNickell (l986) in using time series data on value added, employed labour force and
total captital stock to compute labour-augmenting Solow residuals. We also tried a more
naive output per worker measure, and a more sophisticated residual which takes into
account (seasonally adjusted) hours of work, but found that our empirical results were
fairly robust.
Some econometric issues are worth discussing. Because the error term in (1) contains
an expectational component, uncertain current variables are to be treated as endogenous.
Such an equation is therefore typically estimated by an Instrumental Variables technique.
No matter how good the quality of the data used, data on individual households are
likely to be affected by measurement error. Aggregate data are the result of averaging
over a very large number of households and measurement error is likely to disappear.
In average survey data, measurement error is likely to persist, unless the number of
included households is very large. Furthermore, differencing data affected by measure-
ment error induces an MA (1) residual. In equation (1) the disturbance is the sum of an
expectational error (which is either white noise or an MA (l) with positive coefficient,
Hall (l988» and the measurement errors on consumption growth and the inflation rate,
which are MA (1) processes with negative unitary coefficient. The sum of two MA (1)
processes (or of an MA (l) and a white noise) is still an MA (I), Its first-order
autocorrelation depends on the relative magnitude of the variances of its two components.
OUf data indicate negative first-order autocorrelation, thus suggesting that measurement
error is more important than time aggregation in this data set. Because of the presence
of MA (1) residuals, the use of instruments dated t - 1 is invalid, but instruments lagged
2 and more yield consistent estimates. However, this statement is only true if higher-order
serial correlation can be ruled out. For this reason, we routinely test for correlation of
order 2-5 by means of a Lagrange Multiplier test procedure proposed by Breusch and
Godfrey (1981), and modified by us to take into account that under the null the error
term is an MA (l) rather than white noise."

6. Given the (log-)linearity of equation (1), the sample information on within-cohort variances and
covariances of the relevant variables could be used to implement a more efficient estimation technique, as
suggested by Deaton (1985). One can show that in the case at hand, the IV estimator requires the selection of
instruments lagged 2 or more and involves simple reweighing of the projection matrix so as to give more weight
to the instruments less affected by measurement error. Despite the amount of noise in our data, we found that
the efficiencygain obtained using this method is negligible, while small-sample numerical problems (noted by
Fuller (1987» often require ad hoc adjustments of the reweighting matrix. For these reasons, we decided not
to use the measurement error corrected estimator.
AITANASIO & WEBER CONSUMPTION GROWTH 635

3. EMPIRICAL RESULTS
Equation (1) represents the basic relationship analysed in this paper. Its empirical
specification is, however, more complex.
In the empirical implementation reported in this section, we try to take into account
two basic facts: that consumption exhibits strong seasonal fluctuations (our data are
seasonally unadjusted) and that utility derived from a given amount of expenditure
probably depends on the number of individuals sharing it and the labour market status.
Seasonality is modelled along the lines proposed by Miron (1986), i.e. by specifying

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
a consumption technology model with seasonal shift factors.
The second problem is tackled in different ways in aggregate and cohort data. At
the aggregate level, following the literature, we work with per capita expenditure data.
At the cohort level, after trying a simple per capita specification, we also try some
household technology models (i.e, we allow demographic shift factors as well as the
seasonal factors to enter the consumption technology model).

3.1. Aggregate data


Our instruments are the second, third and fourth lags of consumption growth (in per
capita terms), disposable income growth, inflation and the nominal rate of return (on
building society deposits), plus three seasonal dummies and a constant term (which we
also include in the equation).
On the 63 available observations on NA data we obtain:
u= 0'3541; SC (11) = 18·71, BG (4) =9'09,
(0'1314)
where SC denotes the Sargan criterion, i.e. the test for overidentifying restrictions, which
is a X 2 with 11 degrees of freedom; BG (4) is a test for serial correlation of order 2 to 5,
and is distributed as a X 2 with 4 degrees of freedom. Finally, the MA (1) corrected
standard error is reported in parentheses. In this specification, both Sargan and Breusch-
Godfrey tests are close to or above conventional significance levels. An analysis of the
estimated residuals reveals the presence of fourth- and fifth-order autocorrelation.
As a further test of correct specification, we want to see whether the equation displays
excess sensitivity to current income growth (see Campbell and Mankiw (1989)). When
we introduce real disposable income growth on the right-hand side of our equation, we
obtain:
d log (C,+I) = 0·3281r,+1 + 0·2091d log (Yt+l) + seas.; SC (10) = 19·93
(0,0960) (0·1120) BG (4) =9,67
which gives some indication of excess sensitivity, as well as general misspecification
(detected in both specifications by the Sargan and the Breusch-Godfrey criteria). A real
earnings measure prod .... '.ed similar results. Income growth in this equation is instrumented
using the same variables as in the first equation. This explains the decline in the number
of degrees of freedom in the Sargan criterion. The results reported here are slightly
different from those reported by Campbell and Mankiw (1989) for U.K. aggregate data.
The main difference seems to be caused by our different (and shorter) sample period.
However, the CSO has recently revised the consumption series for the U.K.: this could
also account for some of the difference.'
7. We tried different instrument sets, replacing lagged income as an instrument with the FTIOO average
return on shares or the Treasury Bill rate, but obtained similar results.
636 REVIEW OF ECONOMIC STUDIES

3.2. Cohort data


The use of aggregate data to estimate (1) can be criticised on a number of grounds.
Individual heterogeneity may prevent aggregation over wider groups of the population,
particularly in view of possible non-linearities of the marginal rate of intertemporal
substitution. Furthermore, if individuals are life-cycle maximizers, the equation fails at
the aggregate level because of entries and exits (births and deaths). Finally, individuals
may be willing but unable to borrow in the early part of their lives (when labour income
uncertainty is at its highest, with its moral hazard/adverse selection implications), and

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
may be affected by annuity markets imperfections at a later stage in life.
To check for the empirical importance of these sources of aggregation bias, we use
the Family Expenditure Survey to select all those households whose head was born in
the 1930-40 decade and compute the geometric average of their per capita expenditure.
This implies that nobody younger than 30 or older than 56 is included in our cohort. We
also compute a cohort-specific Divisia price index to compute inflation, and estimate (1)
by Instrumental Variables. We start by estimating equation (1) using per capita consump-
tion to have an equation which is comparable to the one estimated on NA data. By doing
so we neglect most of the information on family composition available in the FES. Our
results are:
a > 0'5961; SC (11) = 9,18, BG (4) = 6,71,
(0,3282)
where all statistics are robust to an MA (1) error. The point estimate of a is higher, even
though it is estimated without much precision. The most interesting feature, however, is
the failure of the two specification tests to reject the null. The Sargan criterion is very
close to its mean, while the correlogram of the estimated residuals looks like a textbook
example of an MA (1) process.
To test for the presence of excess sensitivity to expected labour income, we add its
rate of growth to equation (1). In this case we obtain:
~ log (C,+l) = 0'6251r,+1 + 0'3714~ log (~+1) + seas.; SC (10) = 10·83
(0'2773) (0'1500) BG (4) =4,95
Therefore, even on average cohort data, we find, at least for a simple specification,
evidence of excess sensitivity of consumption to labour income. This could be interpreted
along the lines of Campbell and Mankiw (1989), who suggest the presence of liquidity-
constrained households that consume their labour income, or could be due to mis-
specification of our simple equation. One possibility is that the per capita formulation
used so far is not appropriate. Another possibility is that leisure and goods are non-
separable (as argued by Heckman (1974) and Bean (1986». A simple way to account
for lack of separability is to let employment status affect the utility derived from goods
consumption.
The availability of household data on demographic and labour supply variables
allows us to test these hypotheses directly. We modify equation (1) to make consumption
services a function of consumption expenditure and a number of household characteristics.
In a first attempt, we introduced four household composition variables and two labour
market and one educational indicators (the estimates for this broad specification are
reported in column 2 of Table I, in the Appendix). After dropping those variables that
were least statistically significant (those based on the number of children, which exhibit
little time-series variability for this cohort), we obtained a more parsimonious specification,
which includes the number of adults and a multiple adult indicator, and the educational
ATfANASIO & WEBER CONSUMPTION GROWTH 637

dummy and two employment status indicators described in Section 2. The last three
variables are treated as endogenous, and instrumented with their own third and fourth
lags. The employment status indicators capture the fact that preferences presumably
change when the first and/ or second adult moves in and out of work." The demographic
variables are treated as deterministic and therefore are not instrumented. Their inclusion
allows a household technology more flexible than that based on per capita expenditure.
The equation that is eventually estimated is the following:
d log (Ct +1 ) = 0·7753rt+ 1 + 0·3504d WhCt+l + 0'2379~dltt+l

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
(0'2796) (0'1784) (0'1398)
+ 0'7149~Mat+l - 1'0718~ooWt+l + 0·2115d WWt+l + seas.;
(0,2648) (0'3785) (0'1738)
SC (14) =9'21, BG (4) = 2·75
where Ct+1 now denotes total household consumption, Whct+ 1 is the proportion of
non-manual workers (which is closely related to the education of the head), Adltt+ 1 is
the number of adults, Ma t+1 the multiple adult indicator, OOWt+l the proportion of
households whose head is out of work and WW t + 1 the proportion of households where
the second adult works (based on a "working wife" indicator). When we add labour
income growth to this specification, we obtain:
d log (Ct +1 ) = 0·7261 rt+ 1 + 0·2797 ~ WhCt+l + 0·1905dAdltt+ 1
(0'2690) (0'1834) (0'1372)
+0'6956dMa t+ 1 - 0'8393dooWt+l + 0·2183dwwt +1
(0'2430) (0'4482) (0'1596)
+ 0·1191d log (Yt+l) + seas.;
(0'1454)
SC (13) = 10,17, BG (4) = 3·08.
These results indicate that, once we allow for changes in preferences due to family
composition, education and employment status, consumption growth does not exhibit
excess sensitivity to labour income. This could be interpreted as showing that labour
income is measured with much error, or is otherwise a poor indicator of the presence of
liquidity constraints (because of averaging). In this respect, the proportion of heads who
are out of work may be a better proxy for the severity of borrowing restrictions. However,
as we shall see in the next subsection, labour income growth does enter significantly
alongside the employment status indicator when we extend the sample to all FES
participants. Therefore, we shall interpret the significant coefficient on the head out of
work variable as reflecting the non-separability of non-durable goods and leisure.
In summary, these are our main findings from estimating the equation on average
cohort data: the elasticity of intertemporal substitution is estimated at almost 0'8, which
is substantially higher than the point estimates obtained with aggregate data; the model
does not exhibit any sign of misspecification (both statistical tests fail to reject the null);
the excess sensitivity test fails to reject the null of a zero coefficient on labour income
growth.
8. If we specify the equation without the labour market variables, we find that (T is estimated to be
between 0·6 and 0·7 and that there is no excess sensitivity to labour income. However, as soon as lagged labour
market variables are used as instruments, the broader specification presented in the text appears necessary in
statistical terms. From an intuitive point of view, it seems natural that the utility obtained from expenditure
on, say, gasoline is different if somebody drives to work. Browning and Meghir (1991) firmly establish that
leisure and goods are non-separable, using U.K. FES data to estimate a demand system for non-durable goods.
638 REVIEW OF ECONOMIC STUDIES

3.3. Ruling out simple explanations


Before exploring theoretical hypotheses for the empirical results reported in Subsections
3.1 and 3.1, we must rule out some simpler explanations:
(1) The FES consumption and income data could be fundamentally different from
the corresponding NA aggregates;
(2) a different procedure is followed when the constructing average log consumption
and log income on FES data from that followed using NA data;

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
(3) the excess sensitivity test could, by its very nature, lose all its power when
demographic and employment status variables are added to the equation. Also,
the inclusion of these variables could, as a general rule, increase the estimated
coefficient on the real interest rate.
The first issue concerns the comparability of the two data sources. We do not attempt
a descriptive analysis of the Family Expenditure Survey, but simply show that its design
is not responsible for the different results (both in terms of point estimates and mis-
specification tests). If the two data sources are not fundamentally different, by averaging
across the whole FES sample, we should get a point estimate of the elasticity of substitution
similar to that obtained on NA data, and (perhaps more importantly) find that our
specification tests detect rejections of the overidentifying restrictions and of lack of serial
correlation of order higher than one.
If we take arithmetic averages of expenditure and income data over the whole FES
sample, express both series in per capita terms and then log transform, we have:
a = 0·3661; SC (11) = 24·36, BO (4) = 8·13.
(0·1848)
Remarkably, the estimate of a is very similar on this sample to that on NA data.
As with aggregate data, the Sargan criterion rejects the null of correct specification and
the Breusch-Godfrey test signals correlations of the fourth and fifth order."
The second issue concerns the procedure followed to construct average log C and
average log Y from the two data sources. While taking logarithms of the arithmetic
average of expenditure is what is normally done when using NA data, one should take
the average of the logarithm. This is impossible on NA data, but can be done while
aggregating the FES data (and is what we did to construct average cohort data). When
we average the logarithm of consumption for all the households in the FES we obtain:
u= 0·2908; SC (11) = 12·92, BO (4) = 2·73.
(0·1487)
Thus, the non-linearity in (1) seems to be responsible for the strong serial correlation
of the error (which can in tum lead to rejecting the overidentifying restrictions test). The
point estimates of a, however, are not substantially different.
In the next section we shall investigate in greater detail what moments of the
distribution of expenditure across households are responsible for the serial correlation
and the rejection of the overidentifying restrictions we find while working with arithmetic
means.
9. It is worth pointing out that the presence of serial correlation of order higher than one implies that
the formulae used to compute the variance-covariance matrix of the residuals (and therefore the standard errors
and the test statistics) are not correct. We therefore also computed standard errors and test statistics using the
Newey and West correction and a five-period window. The S.E. on o was 0·1727, the SC just 6·70 and the
BG rose to 9·02, with very strong indications of serial correlation of the fifth order. It is worth noting that the
overidentifying restrictions test would convincingly reject the null if lagged-S instruments were used.
ATIANASIO & WEBER CONSUMPTION GROWTH 639

Finally, the third issue can be tackled by introducing demographic and employment
variables in an equation for the whole FES, as a way of finding out how an "aggregate"
equation would work once we add these variables. Again, this would be problematic
with aggregate data, but is straightforward with FES data. We therefore estimate broader
specifications on FES data averaged over the whole sample (a full list of the estimates
is given in colume 1 of Table I, in the Appendix):
a log (Ct+I) = 0·3792rt+1 + seas. + demo + empl.var.;
(0·2125)

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
SC (14) = 19·81, BG (4) = 6·13.
Even though some of the demographic variables are strongly significant, the estimate
of a is not affected substantially. These estimates differ from those obtained on average
cohort data, in that the' coefficient on the interest rate is still substantially lower. When
we add income growth to this specification we get:
~ log (Ct+ l ) = 0·2851rt+1 + 0·4098~ log (~+I)+seas.+dem. -l-empl.var.;
(0·2506) (0·2297)
SC (13) = 16·48, BG (4) = 1·30.
Unlike the specification on average cohort data, consumption growth is significantly
affected by income growth (on the basis of a one-sided r-test at the 95% significance
level). This may reflect the presence of consumers affected by credit and annuity markets
imperfections, the omission of demographic variables (over and above those already
included-see Table I in the Appendix), or the restriction implicitly imposed of equal
coefficients on the demographic factors. However, it does indicate that excess sensitivity
tests can retain their power even in the presence of employment status indicators.

3.4. Concluding remarks


We believe the results presented in this section are remarkable in three respects.
(1) They show that estimates of the elasticity of intertemporal substitution in con-
sumption obtained from aggregate and average cohort data differ substantially.
These differences are not due to a peculiarity of the data set but are a genuine
result of aggregation across cohorts.
(2) The difference between the two sets of results, and in particular the rejection of
the overidentifying restrictions found in National Accounts data, can partly be
explained by non-linearities. The difference between geometric and arithmetic
averages has important effects on the serial correlation properties of the error
term. This may in tum make some of the instruments invalid.
(3) Once we go beyond simple specifications for the Euler equations based on per
capita expenditure and allow preferences to change with changes in family
composition and employment status, the theoretical model performs better on
average cohort data than on aggregate data (on the basis of an excess sensitivity
test to labour income growth).

4. SOME EXPLANATIONS
The evidence presented in the previous section shows that aggregation across different
cohorts matters in a fundamental way. On the one hand, the estimates of the elasticity
640 REVIEW OF ECONOMIC STUDIES

of intertemporal substitution differ depending on whether we use aggregate or average


cohort data. On the other hand, the life-cycle model performs better on cohort than on
aggregate data.
In this section we discuss a number of different explanations for our findings, and
produce empirical evidence relating to each. Our investigation is unlikely to be exhaustive,
but should help us understand the empirical relevance of some theoretical explanations
on aggregation bias in Euler equations for non-durable consumption.
In Subsection 4.1 we look into the importance of non-linearities. We do this by

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
analysing the properties of the higher moments of the cross-sectional distribution of per
capita non-durable expenditure. This analysis tries to explain why the restrictions implied
by the model are violated when we use arithmetic means, but not when we use geometric
means. In Subsection 4.2 we discuss the role of demographic factors in aggregation,
while Subsection 4.3 considers the Euler equation for different cohorts. With this exercise
we hope to explain the differences obtained using aggregate FES and average cohort
data. Finally, in Subsection 4.4, we explore the importance of finite lives and missing
markets.

4.1. The role of non-linearities


The individual Euler equation behind expression (1) is linear in logs, so that it suggests
aggregation by geometric rather than arithmetic means. While National Accounts data
are only available as the arithmetic mean, it is straightforward to aggregate the FES data
in different ways.
An interesting finding of the empirical exercise carried out in Section 3 is the
discrepancy between the results obtained when working with the logarithm of the arith-
metic mean and those derived from the mean of the logarithm (i.e. the logarithm of the
geometric mean). Aggregating all FES households, we found that the simple relation in
per capita terms produced no signs of misspecification when we used the logs of the
geometric means, but exhibited serial correlation of order higher than one when we
considered the logs of the arithmetic means. Furthermore, in the latter case, we rejected
the overidentifying restrictions implied by the model.'?
This evidence suggests that non-linearities may play an important role in the rejection
of the overidentifying restrictions implied by the model typically found in aggregate time
series. It is therefore interesting to analyse the differences between geometric and arith-
metic means, also known as Theil's entropy measure."
Let us define changes in Theil's entropy measure by dDiff:

~Diff = dE (log (Ct + 1,h)) - dlog (E( C t + 1,h)), (2)

where E denotes the cross-sectional mean operator and h the hth household. If dDiff

10. The rejection of the overidentifying restrictions implied by the Sargan criterion is based on a formula
for the variance-covariance matrix of the residuals that assumes an MA (1) process. If the residuals are
characterized by higher-order serial correlation, this formula is not valid. Indeed, when we force an MA (1)
process on these residuals we sometimes obtain point estimates of first-order autocorrelation below -0'5, which
are inconsistent with an MA (1) process. In these cases we compute the variance-covariance matrix of the
residuals using the Newey-West formulae with windows of various length. Then the Sargan criterion does not
reject the overidentifying restrictions. Notice, however, that the presence of high-order autocorrelation is, per
se, a sign of misspecification.
11. It is interesting to notice that Modigliani and Ando (1963) refer to the constancy of Theil's entropy
measure as a necessary condition to estimate the life-cycle model using aggregate time-series data.
ATIANASIO & WEBER CONSUMPTION GROWTH 641

is constant or uncorrelated over time and with the instrument set, the log of the arithmetic
mean can be used in place of the log of the geometric mean. In practice, it exhibits serial
correlation and is also correlated with the instruments used in estimating the model.
It may be of some interest to decompose Theil's entropy measure into different
central moments and establish which of these are responsible for our findings. If the
cross-sectional distribution of household consumption was log-normal it would be possible
to express the difference in (2) as a function of the cross-sectional variance only (plus
a measurement error term that would reflect small sample bias). A Kolmogoroff normality

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
test on log (Ct+l,h), however, rejected normality in 59 of the 63 quarters of the estimation
sample at the 99% significance level.
By expanding exp (E(log (Ct+l,h») as a Taylor series around log (E( Ct+l,h» and
taking first differences, we are able to express ~Diff as follows:
~Diff = ~ log (l +1J-t2,t+l +~J-t3,t+l + 14J-t4,t+l + R)
1 1 1 ,.,
= ~ log (l +:2J-t2.t+l +6J-t3,t+l +:24J-t4,t+l) + R, (3)
where J-ti,t is the i-th cross-sectional central moment oflog (Ct,h)' Ifwe use the approxima-
tion log (1 + x) == x we see that to approximate Theil's entropy measure we need to compute
the first differences of the second, third and fourth central moments.
The procedure just sketched involves two approximations: the truncation of moments
higher than fourth and the substitution of log (1 + x) by x. Both approximations work
remarkably well. The difference induced by the truncation is barely visible on a plot of
~Diff. When we regress aDiff on the first differences of the three moments we are
considering, we obtain an R 2 above 0·98.
Theil's entropy measure is plotted against each of the three moments (divided by 2,
6 and 24, respectively) in Figures 1-3 (all in first differences). The variance term is most
highly correlated with entropy, but all three moments exhibit strong autocorrelation,
seasonality and variability. This suggests that they are the source for the violation of the
restrictions implied by the model reported above. Indeed when we regress each of these
moments on the instrument set, we find a significant relationship for the variance and
the skewness.
As a further test of the hypothesis that it is the omission of these moments that
explains our results, we add them to the right-hand side of the equation for the arithmetic
mean of consumption. The coefficients on the second and third moment tum out to be
significantly different from zero, while the Sargan criterion to test the overidentifying
restrictions fails to reject the null, as does the Breusch-Godfrey statistic for autocorre-
lation.
Finally, we also check the effect of taking arithmetic averages over the cohort. Here,
the point estimate of o increases dramatically when arithmetic means are used (it rises
to 1'56, S.E.: 0'30), and serial correlation of third, fourth and fifth order is generated.
Similar results obtain when demographic variables are introduced in the equation.

4.2. Conditioning on micro variables


In Section 3 we saw that demographic and labour supply variables played an important
role in estimating the relationship between consumption growth and the interest rate. If
we think of equation (1) as derived from an Euler equation for a single family, it is
possible to rationalize the presence of these variables by considering a utility function
with a fairly flexible household technology. Such a technology transforms consumption
expenditure into a flow of services which enters an iso-elastic instantaneous utility function.
642 REVIEW OF ECONOMIC STUDIES

0-04 . - - - : - - - - - - - - - - - - - : : : : - - - - - - - - - - - - - - ,

0·03

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
-0·01

-0-02

-0·03

-0·04 -+-rrrT...-rl...,.,..-rr-rT"T'T"T"T"1...,.,..-rr-r-r-rT""T"T""'1rr_rT""T""T"T"'T"T...,..."rr_rT""T""T"T"'T"T""""-...--r-...._r_T""T""T""1,.-,-M

71·272·1 73·1 74·1 75-1 76·1 77·1 78·1 79·1 80·1 81·1 82·1 83·1 84·1 85·1 86·1

o .6.Entropy +.6.Variance/2

FIGURE 1
Entropy and variance
0·04 --r-------------------------,

0·03
I
0·02

0·01

-0·01

-0·02

-0·03

- O· 04 -+-r-T""T""T"T'"T"T""T'T'"1l"'T"T'"rrT"T"'T"T...,...,rr_rT"T'T...-rl...,.,..-rr-rT"T"T~rr_rT""T""T"T"'T"T"'T'T""'rr_r...._r_T,.,..,............
71·272·1 73·1 74·1 75·1 76·1 77·1 78·1 79·1 80·1 81·1 82·1 83·1 84·1 85-1 86·1

o .6.Entropy + .6.S kewness/6

FIGURE 2
Entropy and skewness
ATIANASIO & WEBER CONSUMPTION GROWTH 643

0·04 - , - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ,

0·03

0·02

0·01

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
-0·01

-0'02

-0·03

-0· 04 -t-r.......,.....,.-r-T""'C-r'T-rT",.....,.-rT""IC-r'T-rT",.....,.-"'"...,.....,...,....."-T"T"1..--r-r....,...,.....,,-,-,,..--r-r....,...,.....,,--ro-..--r-r-r-r..,,-"T""'ri
71·272·173·174·175·176·\ 77·\ 78·179·180·181·\ 82·183·\ 84·\ 85·\ 86·\

o 6Entropy + 6 Kurtosis/24

FIGURE 3
Entropy and kurtosis

To identify the elasticity of intertemporal substitution, it is then necessary to take


explicitly into account changes in the variables that determine the utility function. This
cannot be done with aggregate-time series data, but is straightforward with micro data.
This problem can be restated as a classical aggregation problem. Even ignoring
non-linearities, the aggregation equation contains some unobservable variables. There-
fore, estimates based on macro data will be consistent only if the excluded variables are
uncorrelated with the instrument set.
As an example, take the case where the instantaneous utility function for individual
i belonging to cohort a depends on consumption expenditure and a cohort-specific taste
shock s~·a:

The Euler equation for individual i will therefore be:

(4)

where i denotes the nominal interest rate and p the price index corresponding to c. Under
the assumption ofiso-elastic utility and conditional log-normality of the relevant variables,
we get:

(5)
644 REVIEW OF ECONOMIC STUDIES

where E7~1 is an expectational error uncorrelated with information available at time t.


Equation (5) can be averaged across individuals belonging to cohort a and estimated by
standard techniques.
There are at least three cases where the equation above is consistently estimated on
cohort data, but not on aggregate data: (1) when S is observable at the individual level,
but not for the whole population, and changes in S are correlated with the instruments
available at the macro level; (2) when S does not vary over time for any cohort members,
but changes systematically across cohorts; (3) when changes in S cancel out within the

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
cohort, but not for the whole population.
Suppose for instance that S is a dummy which is equal to one if the household head
(or the second adult) works. Aggregate data may not reveal the proportion of working
husbands or wives. If such a proportion is known for the cohort, the cohort equation
can be estimated consistently, while the aggregate equation cannot. It is also possible
that information on S is not available in the survey, but other characteristics are known.
These may be used to select households whose S variables are likely to stay constant
(prime-aged males are unlikely to retire, say). Finally, labour force participation may
vary unsystematically within a homogeneous group (thus washing out in the group mean),
but change over time in the population at large." In general all we need for the
non-aggregation result to obtain is that deviations of ~ log (Si,r+l) from its conditional
mean are uncorrelated with cohort-specific instruments, but correlated with macro
instruments.

4.3. Aggregating across cohorts


Having stressed the importance of demographic and labour supply variables for the
estimation of the Euler equation for consumption and the potential bias that the omission
of these variables causes in aggregate time-series data, we now discuss differences across
cohorts.
In Section 3 we saw that while the introduction of demographic and labour supply
variables eliminates excess sensitivity to labour income in the cohort equation, it does
not explain the difference between the point estimates based on average cohort data and
those derived from the whole FES. Such a difference could be explained by several factors.
The first, obvious, one could be the low precision with which a is estimated on
aggregated FES data. Alternatively, it could be the case that the different estimates reflect
substantial heterogeneity across cohorts that is not captured by the micro variables we
have been considering. This heterogeneity can arise from differences in preferences,
household technology or opportunity sets.
It is therefore interesting to construct average cohort data for other groups of the
population, to establish to what extent the differences reported above reflect differences
in preferences and household technology.
We construct two cohorts in addition to the one we have used so far (which will be
referred to as the "middle" cohort): a "young" cohort, made up of all households whose
head was born between 1941 and 1948 (average number of households per quarter: 263)

12. Another source of aggregation bias may be due to information deficiencies. If we assume that
individuals know their own past (and that of other individuals belonging to the same cohort) but have limited
information about aggregate events, then cohort-specific instruments may be valid to identify the parameters
of the model, while aggregate instruments would lead to inconsistent estimates.
ATIANASIO & WEBER CONSUMPTION GROWTH 645

and an "old" cohort, consisting of all households whose head was born in the 1920-29
interval (average size: 303).13
We did not conduct a specification search for either of the new cohorts, but simply
fitted the most general specification used for the middle cohort (a complete set of results
is reported in Table I, in the Appendix). For the young cohort, we obtain a point estimate
for o of 0·56 (S.E.: 0·20). No misspecification is detected either by the Sargan criterion
or by excess sensitivity tests. The serial correlation test, however, signals some third- and
fifth-order correlation (leading to severe rejections when the test is computed on the

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
assumption of MA (1) disturbances; because of these rejections, in Table I we report
standard errors robust to fifth-order serial correlation). For the older cohort, by contrast,
the point estimate of a is 0·32 (S.E.: 0'36). We find no sign of autocorrelation from the
second to the fifth order (even though there is significant autocorrelation of order seven
and eight) and no rejection ofthe overidentifying restrictions. The test for excess sensitivity
to labour income fails to reject the null. For this last cohort, for which an increasing
number of individuals retire in the last part of the sample, the measurement of average
labour income can be misleading (despite the inclusion of some pension income in it).
For this reason we try some other tests of excess sensitivity. In particular we add to the
basic equation the expected increase in total factor productivity as measured by Solow's
residuals. In this case we do reject the null." Finally, it should be mentioned that we
have reason to doubt the specification for the oldest cohort. We already mentioned the
problem of retirement which is likely not only to affect the measurement of average
income, but also to change preferences in a dramatic way (particularly if related to poor
health). In addition to this, we have no way to control for mortality, which presumably
plays an important role for this cohort towards the end of the sample. This could have
an influence on our results if mortality is correlated with the variables we are measuring
(as it is likely to be), therefore inducing non-random attrition.
A Wald test of equality of the coefficients on the interest rate across the three cohorts
fails to reject the null. The test statistic, which is asymptotically distributed as a X 2 with
2 degrees of freedom, takes the value of 1·14. However, as soon as we impose common
coefficients on all the explanatory variables, we strongly reject the null. This is not
surprising: the multiple adult indicator, for instance, is likely to signal the presence of
elderly relatives in the household for the old cohort, while it may well indicate house-
sharing for the young one, and perhaps the presence of grown-up offspring for the middle
cohort. Similarly, the presence of a working second adult has a positive effect on
consumption for the young and middle cohorts, but a negative effect on the old cohort.
Here, the change in sign may simply reflect different attitudes to female labour participa-
tion: in older generations the wife is likely to work only because of pressing financial
need, whereas in younger ones she is more likely to work for other reasons.
These results can be interpreted in different ways. The different point estimates of
the elasticity of substitution can probably be justified by the highly imprecise estimate
13. In the FES, very few households are headed by someone aged less than 22. This is due partly to
non-response and partly to household formation. Given that household formation may be correlated with the
real interest rate (via the housing market, etc.), we did not extend the young cohort to include people born
after 1948, who would have been less than 22 years old in 1970. As for the older cohort, a relatively large
number of heads retire during the sample period. This would still be true even if we took only households
whose head was born between 1925 and 1929: we would have a large proportion of retired individuals by the
mid-1980s. This may be a problem if early retirement is induced by poor health. Further difficulties arise
because of more frequent marital breakdowns and relatively high mortality rates for males aged 55-65, which
implies that for many households the second adult becomes the head in the latter part of the sample.
14. The same variable, computed as in Layard and Nickell (1986), was tried on the other cohorts and
turned out to be insignificant. It is, however, significant on aggregate National Accounts and FES data.
646 REVIEW OF ECONOMIC STUDIES

obtained for the old cohort. However, the data strongly reject the equality of the household
technology coefficients. This could conceivably explain the difference between the average
cohort and aggregate FES data, as the latter implicitly impose that restriction. Further-
more, the middle cohort is the only one not to present any sign of misspecification.
Misspecification for the other two cohorts could be an indication of the fact that those
cohorts are observed over a period of their life cycle in which the model does not describe
accurately their behaviour. This, in tum, could be due to misspecification of the household
technology (or more generally to changes in preferences), or to the presence of imperfec-

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
tions in capital and/or annuity markets.
So far we have addressed the problem of aggregation across different cohorts. A
similar argument could be made for the aggregation of the Euler equations across
individuals within a given cohort. If we had a long time series of individual observations,
we could address this problem in the same way as we did for aggregation across cohorts.

4.4. Finite lives and missing markets


One further source of aggregation bias can be traced to missing markets: if some markets
are missing different groups of the population differ in their ability to hedge against
unexpected shocks. This may induce correlation between the aggregate "surprise" and
the instrument set. As an example, consider an Overlapping Generations (OG) model.
In this model, current generations cannot trade with future generations. Even though
there is an Euler equation relating consumption of agents alive in subsequent periods,
the same equation does not hold, in general, for aggregate consumption: there is no
reason for the ratio of consumption of newly-born individuals to the consumption of
those who die in the previous period to satisfy an Euler equation. This ratio will be
related to the rate of growth of resources available to new generations and therefore to
productivity growth. In so far as productivity growth correlates with the instrument set,
estimates of the standard Euler equation on aggregate data are biased."
OG models rely on one particular type of missing market, namely the inability to
trade with (as yet unborn) future generations. This is reflected in a bias in the aggregate
Euler equation because of exits from and entries into the consumption pool. Therefore
the bias introduced in considering aggregate consumption depends to a large extent on
the fraction of the population which enters and exits in each time period. On high
frequency data, this source of market failure and aggregation bias is unlikely to be
important. This is borne out by a simple empirical check. In constructing the time series
for consumption used in the estimation of the Euler equation, instead of aggregating the
whole FES, we aggregated all the households within a certain age band, and varied the
width of the band. In this way we constructed samples for which the proportion of entries
and exits can be substantially larger. If entries and exits are responsible for the lower
point estimates obtained over the whole FES (and for the rejection of the excess sensitivity
test), we should find similar, more pronounced results in the cases in which entries and
15. The basic intuition behind this statement is already contained in the seminal paper by Modigliani
and Brumberg (1954) and discussed more recently (within a steady state framework) by Deaton (1987). Recent
papers by Blanchard (1985) and Weil (1989) derive aggregate Euler equations for the case where all uncertainty
derives from the probability of death. Aggregate consumption growth depends not only on the interest rate,
but also on the ratio of wealth to consumption. In Attanasio and Weber (1992), we construct two- and
three-period overlapping generation models with productivity shocks to assess to what extent the use of aggregate
consumption introduces a bias in the estimation of the elasticity of intertemporal substitution. Within such a
framework we show that aggregate consumption growth is related not only to the real interest rate, but also to
a term reflecting capital intensity in the two periods. The omission of this term would introduce a bias whose
sign depends on the elasticity of substitution between labour and capital. If it is less than unity, the bias is negative.
ATIANASIO & WEBER CONSUMPTION GROWTH 647

exits account for a larger proportion of the sample. This did not happen. The elasticity
of intertemporal substitution was estimated without much precision and was higher than
in the whole FES estimates, and we were unable to find strong signs of misspecification,
especially for narrowly defined age bands (such as all households whose head is 35 to
45 years old).
To establish the importance of the effect of more general types of missing markets
is not easy: we can only rely on indirect tests. To test the hypothesis that different cohorts
are affected in different ways by aggregate shocks and that these shocks cannot be hedged

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
against because of missing markets, we introduced a measure of productivity growth in
the Euler equations that we had estimated so far. This term turned out to be significant
for NA and aggregate FES data and for the old cohort (detailed results are available
upon request), but not for the young and middle cohorts. This result is broadly consistent
with our findings on excess sensitivity to income growth, in pointing to the importance
of missing or imperfect markets for some groups of the population.

5. CONCLUSIONS
In this paper we have presented empirical evidence on the importance of aggregation
bias in Euler equations for consumption. We estimate the elasticity of intertemporal
substitution and test the restrictions implied by the theoretical model with aggregate and
average cohort data. In particular, we look at tests of the overidentifying restrictions and
of excess sensitivity of consumption to expected labour income, and at the autocorrelation
of the estimated residuals.
The main results are: (1) estimates of the elasticity of intertemporal substitution for
consumption are consistently lower for aggregate data than for average cohort data; (2)
the restrictions implied by the model are often rejected on aggregate data but not on
average cohort data.
The magnitude of the elasticity ofintertemporal substitution is important for a number
of different issues, ranging from the deadweight loss of interest income taxation to the
effects of changes in interest rates on savings. Our estimates range from O· 3 (on aggregate
data) to 0·8 (on cohort data). While the qualitative predictions of several models depend
on the elasticity being above or below unity, from a quantitative point of view the size
of the elasticity is obviously important." Furthermore, while a 95% confidence interval
easily includes unity for cohort-based estimates, it does not for aggregate data.
In trying to explain these differences, we have found that a major role is played by
the non-linearity of the estimatable equation and by omitted demographic factors (nor-
mally unobservable on aggregate data). However, even when these sources of aggregation
bias are corrected for, the estimates of the elasticity ofintertemporal substitution obtained
from aggregate data remain lower than those obtained from average cohort data. Further-
more, aggregate consumption growth depends on expected productivity and income
growth, and aggregate equations typically display higher-order serial correlation.
In the second half of this paper we have investigated explanations of our empirical
results. We presented empirical evidence on the importance on non-linearities, which
suggests that using arithmetic instead of geometric means leads to high-order serial
correlation. We also evaluated the role of demographic characteristics and other differen-
ces in preferences and/ or opportunity sets in biasing the aggregate estimate of the elasticity
16. Back-of-the-envelope calculations based on a two-period model without uncertainty show that an
increase in the real interest rate from 2% to 3% has almost no effect on the saving rate if (T = O·8-a 0·1%
reduction-but leads to a decrease of 1·33% when (T = O·3.
648 REVIEW OF ECONOMIC STUDIES

of intertemporal substitution. An interesting finding is that one should not force the
cofficients on socio-demographic characteristics to be the same across cohorts (this is
what is automatically done when the data are averaged across the whole population).
The differential impact of these socio-demographic characteristics (which include employ-
ment status indicators) could reflect either generational taste differences or the operation
of imperfect financial markets.

APPENDIX

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
TABLE I
Estimates with demographic factors
Estimates for the broad specification

(1) (2) (3) (4)


whole FES middle young old

r 0·3792 0·7466 0·5586 0·3167


(0,1638) (0'2108) (0·1989) (0'3586)
81 -0,0765 -0,0675 -0,1066 -0,0953
(0'0065) (0'0075) (0·0081) (0,0154)
82 0·0598 0·0583 0·0654 0·0601
(0,0060) (0,0064) (0,0076) (0-0124)
83 -0·0100 -0,0157 -0·0024 -0-0023
(0·0057) (0,0088) (0,0101) (0'0125)
S4 0·0555 0·0514 0·0700 0-0517
(0-0053) (0-0083) (0'0086) (0-0084)
dWhc 0·2389 0·4078 0·1237 0-9274
(0·4781) (0'1350) (0'1132) (0-4126)
dAdlt 0·5258 0·2201 0·6944 0·8439
(0·2565) (0·0869) (0'1800) (0'2952)
dnchil 0·1869 0·0063 0·0686 -0'0331
(0'1056) (0,0507) (0'0501) (0'1686)
dykids -0,4082 0-0446 -0,1941 0·2131
(0'2090) (0-0709) (0'1015) (0,2926)
dMa -0,7981 0·7086 -0'6343 -1-0087
(0-3801) (0-2070) (0·2678) (0,4021)
doow -0-9824 -0,9415 0·4671 -0,2593
(0'4790) (0'3213) (0,2765) (0'3185)
dWw 0-0101 0·2725 0-1911 -0,2779
(0'3642) (0'1593) (0'1831) (0·1916)
SC (14) 14·8026 11·4118 12·3984 10·7558
BG(4) 4·7205 3·0841 7·4594 0·7167

Restriction tests (columns 2-4)


equality of cofficients on r: l'14(x 2(2»
equality of all coefficients: 177'23(X 2(24»

Notes: column (1) standard errors and test statistics based on full MA (1) correction.
Columns 2-4 standard errors and test statistics based on Newey-West correction,
window length = 5. Dependent variable: d log (C,+1), where C is household con-
sumption of non-durable goods and services. Instruments: 81, 82, 83, 84, dMa,
dAd/t, dnchil, dykids, third and fourth lags of dOow, d ww and d Whc, second, third
and fourth lags of consumption growth, interest rate, inflation and income growth.

Acknowledgements. A previous version of this paper was circulated with the title: "Consumption,
Productivity Growth and the Interest Rate". Helpful discussions with C. Bean, R. Blundell, M. Browning,
W. M. Gorman, L. Goulder, T. MaCurdy, G. Mankiw, R. Manuelli, C. Meghir, J. Pencavel, N. Rau, T. Sargent,
and J_ Skinner are gratefully acknowledged. Useful suggestions were made by two anonymous referees and an
editor of this journal. We also benefited from comments received during seminars at the University of Bologna,
AITANASIO & WEBER CONSUMPTION GROWTH 649

U.C. Berkeley, Northwestern University, the University of Virginia, the Institute for Fiscal Studies, Stanford
University, the London School of Economics, CentER, Tilburg, the IFS-LBS Aggregation Conference (October
1989), U.c. Davis, Princeton University, the Wharton School, University of Pennsylvania and the University
of Warwick. Judith M. Payne provided excellent editorial assistance. We thank the UK Department of
Employment and the Institute for Fiscal Studies for providing the data used in this study, but stress that we
are solely responsible for the further analysis and interpretation of the data.

REFERENCES
ATKINSON, A. and MICKLEWRIGHT, J. (1983), "On the Reliability of Income Data in the Family
Expenditure Survey 1970-1977", Journal of the Royal Statistical Society, A. 146,33-61.

Downloaded from http://restud.oxfordjournals.org/ at University of Iowa Libraries/Serials Acquisitions on June 23, 2015
AITANASIO, O. P. and WEBER, G. (1989), "Intertemporal Substitution, Risk Aversion and the Euler Equation
for Consumption", Economic Journal, Supplement, 99,59-73.
AITANASIO, O. P. and WEBER, G. (1992), "On the Aggregation of Euler Equations for Consumption in
Simple OLG Models" (IFS Working Paper 92/12 and IGIER Discussion Paper 14).
BEAN, C. (1986), "The Estimation of 'Surprise' Models and the 'Surprise' Consumption Function", Review
of Economic Studies, 53, 497-516.
BLANCHARD, O. (1985), "Debts, Deficits and Finite Horizons", Journal of Political Economy, 93, 223-247.
BLUNDELL, R., BROWNING, M. and MEGHIR, C. (1989), "Consumer Demand and the Life-Cycle Alloca-
tion of Household Expenditures", Review of Economic Studies (forthcoming).
BREUSCH, T. S. and GODFREY, L. (1981), "A Review of Recent Work on Autocorrelation in Dynamic
Simultaneous Models", in Currie, D., Nobay, R. and Peel, D. (eds.), Macroeconomic Analysis (London:
Croom Helm).
BROWNING, M. and MEGHIR, C. (1991), "Testing for Separability of Commodity Demands from Male and
Female Labour Supply", Econometrica, 59, 925-951.
CAMPBELL, J. Y. and MANKIW, N. G. (1989), "Permanent Income, Current Income, and Consumption",
NBER Macroeconomics Annual (Cambridge: NBER).
DEATON, A. (1985), "Panel Data from Time Series of Cross Sections", Journal of Econometrics, 30, 109-126.
DEATON, A. (1987), "Life Cycle Models of Consumption: Is the Evidence Consistent with the Theory?", in
Bewley, T. F. (ed.), Advances in Econometrics, 2 (Cambridge: Cambridge University Press).
EPSTEIN, L. G. and ZIN, S. E. (1989), "Substitution, Risk Aversion, and the Temporal Behavior of Consumption
and Asset Returns. I: A Theoretical Framework", Econometrica, 57, 937-970.
FULLER, W. A. (1987) Measurement Error Models (New York: John Wiley).
GROSSMAN, S. J. and SHILLER, R. J. (1982), "Consumption Correlatedness and Risk Measurement in
Economies with Non-Traded Assets and Heterogeneous Information", Journal of Financial Economics,
10, 195-210.
HALL, R. E. (1988), "Intertemporal Substitution in Consumption", Journal of Political Economy, 96,339-357.
HANSEN, L. P. and SINGLETON, K. J. (1983), "Stochastic Consumption, Risk Aversion and the Temporal
Behavior of Asset Returns", Journal of Political Economy, 91,249-265.
HECKMAN, J. (1974), "Life Cycle Consumption and Labour Supply: An Explanation of the Relationship
Between Consumption and Income over the Life Cycle", American Economic Review, 64, 188-199.
LAYARD, R. C. P. and NICKELL, S. (1986), "Unemployment in Britain", Economica, 87, SI21-S170.
MIRON, J. A. (1986), "Seasonal Fluctuations and the Life Cycle-Permanent Income Model of Consumption",
Journal of Political Economy, 94, 1258-1279.
MODIGLIANI, F. and ANDO, A. (1963), "The Life Cycle Hypothesis of Saving: Aggregate Implications and
Tests", American Economic Review, 53, 55-84.
MODIGLIANI, F. and BRUMBERG, R. (1954), "Utility Analysis and the Consumption Function: An
Interpretation of Cross-Section Data", in Kurihara, K. K. (ed.), Post Keynesian Economics (London:
George Allen and Unwin).
WElL, P. (1989), "Overlapping Families of Infinitely-Lived Agents", Journal of Public Economics, 38, 183-198.

S-ar putea să vă placă și