Panel and Pseudo-Panel Estimation of Cross-Sectional and Time Series Elasticities of Food Consumption: The Case of U.S. and Polish Data Author(s): François Gardes, Greg. J. Duncan, Patrice Gaubert, Marc Gurgand, Christophe Starzec Source: Journal of Business & Economic Statistics, Vol. 23, No. 2 (Apr., 2005), pp. 242-253 Published by: American Statistical Association Stable URL: http://www.jstor.org/stable/27638815 . Accessed: 27/10/2011 05:21 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact
[email protected].
American Statistical Association is collaborating with JSTOR to digitize, preserve and extend access to Journal of Business & Economic Statistics.
http://www.jstor.org
and
Panel
Pseudo-Panel
Estimation
and
Cross-Sectional
Time
of Food
The
Consumption: and Polish Data
U.S.
Fran?ois Gardes
IPantheon-Sorbonne
Universit? de Paris
Greg J. Duncan
Elasticities
Case
Paris, 75005,
(Cermsem),
IL 60208
Northwestern University, Evanston,
Series
of
of
France
(
[email protected])
(
[email protected])
Patrice Gaubert sur Mer Cedex,
Universit? du Littoral (LEMMA), Boulogne Marc
France
62321,
(gaubert?
univ-paris1 .fr)
Gurgand
CNRS-Delta,
France
Paris, 75014,
Christophe Starzec
France
Paris, 75013,
CNRS-TEAM,
(
[email protected])
(starzec@univ-paris1
.fr)
This article addresses
the problem of the bias of income and expenditure elasticities estimated on pseudo panel data caused by measurement error and unobserved heterogeneity. We gauge these biases empirically by comparing cross-sectional, pseudo-panel, and true panel data from both Polish and U.S. expenditure surveys. Our results suggest that unobserved heterogeneity imparts a downward bias to cross-section
estimates of income elasticities of at-home food expenditures and an upward bias to estimates of income elasticities of away-from-home food expenditures. "Within" and first-difference estimators suffer less bias, but only if the effects of measurement error are accounted forwith instrumental variables.
KEY WORDS:
1.
AIDS
Individual and grouped data; Unobserved
model;
INTRODUCTION
to estimate
time
series,
panels
demand
systems. time
within-group
using
Data
types
series, and
data,
aggregated
include
as theirbirth year or the education level of the reference per son of a household. Estimation with pseudo-panel data dimin
can be
A multitude of data types and econometric models used
ishes
and
panels
using individual data. Aggregate time series data frequently produce aggregation biases because of composition effectsdue to the change of the population or the heterogeneity of price and
income
different
effect between
social
classes.
These
prob
lems have led the vast majority of empirical studies in labor economics to use individual data (Angrist and Krueger 1999). In contrast, and
periods els
individual
are
on countries
attrition can
sectors
industrial
span
generally
to nonresponse
subject or
data
panel
bias.
suffer from
short Even
time
ternative, riods
or
to estimate
data
grouping
even when
panel
to compare
data
different
on
exist, countries.
a
pseudo-panel on
to estimate
Pseudo-panel
same
reference and
In pseudo-panel to criteria
population sometimes analyses,
that do not
change
in different
but
periods,
from
one
are
the cross-equation
survey
as
of measurement,
period
imposed
by consumption
ture elasticities.
by
article
addresses
elasticities
the
issue
caused
by
to income
of bias errors
of
ex
and
mea
specification,
static
estimating
expenditure
models
using
cross-sectional,
pseudo-panel, and true panel data from both Polish and U.S.
pe
expenditure
are
parison
sometimes
timations
surveys.
between based
To
our
cross-sectional, on
the
same
?
according
to another,
time
restrictions
this
knowledge, dataset.
of Business
The
and use
com
is the first
pseudo-panel,
2005 American
Journal grouped
show
theory.Using different types of data helps reveal the nature of thebiases that they impart to estimates of income and expendi
not. individuals
but we
dimension,
surement, and omitted variables or by heteroscedasticity in grouped and individual-based models. We gauge these biases
typically constructed from a time series of independent surveys thathave been conducted under the same methodology on the consecutive
or
units
consumption as
penditure
is an al
data
across well
This
structural
longer
the cross-sectional
Static and dynamic demand models have been developed for these different types of data, with each model adopting a differ ent approach to problems caused by unobserved heterogeneity
pan
changes or composition effects thatmake itdifficult tomaintain the stationarityhypotheses for all variables. Thus,
on
dimension.
pseudo
cross-sections
efficiency
herein that it also gives rise to a heteroscedasticity in the time
aggregate
cross-sections,
heterogeneity.
panel
of one
Statistical
es our
Association
& Economic April 2005,
of
Statistics
Vol. 23, No. 2
DOI 10.1198/073500104000000587
such
242
243
of Food Consumption
Elasticities
et al.: Estimation
Gardes
Panel Study of Income Dynamics (PSID)? is motivated by the numerous expenditure studies based on it (Altug and Miller 1990; Altonji and Siow 1987; Hall and Mishkin 1982; Naik and Moore 1996; Zeldes 1989). Our sec ond dataset is fromPoland in the late 1980s, which enables us to capitalize on large income and price variations during the transitionperiod inPoland. Section 2 presents a background discussion. Section 3 two dataseis?the
the econometric
presents
and
problems
the methods
describes
results.
these
6 discusses
Section
and
evaluate
systems
static
"within"
forms.
However,
rather are
than
No
matter
how
survey
complete,
come
and
For
elasticities.
price
the value
example,
of time dif
fers across households and is positively related to a household's Because
income.
observed
activities
consumption
(e.g.,
rise
give
1996). We estimated instrumen
usual
using
to those
estimated
However,
these
to consider
prefer
and pseudo-panel
for the static models are
versions
dynamic
panel
the static
only
each
methods
specification.
offers
advantages
and disadvantages for handling the estimation problems inher ent
in expenditure models. error. Survey
concerns
set of
first
A
measurement
eating
dynamic
of
variety
factors
in terms of the specification and the econometric
so we
problems, True
similar
quite
questionable
expen
ditures and demographic characteristics lack explicit measures of all of thepossible factors thatmight bias the estimates of in
observations.
(Arrelano 1989) and found that elasticity esti
in this article.
presented
on household
data
were
and other
as
from
and
a wide
from
the static models
of
and
treated
be
household-level
(e.g., Naik andMoore
versions
tationmethods
mates
also
cross-sections
estimates
persistence
to dynamic models
needed
data
in first-difference can
data
independent
individual
to compare
able
types. Habit
dynamic
2. BACKGROUND
from
ex
panel
us with
models these
to esti
approaches
sets of household
provide
panels
expenditure
came
they
though grouped Thus we
two
using
two
The
data.
penditure to estimate
of alternative
implications
demand
mating
data
used in this study. Section 4 describes the data. Section 5 gives the results,
We
reports
on
centers
are
income
of household
meals) often involve inputs of both goods (e.g., groceries) and measured with error; differencing reports of household income across waves increases the extent of error. Instru time (e.g., time spent cooking and eating), households will face undoubtedly can to caused the biases mental variables be address used even if the prices of the different (full) prices of consumption by error instrumenta measurement Siow Like and case are the identical. If?as is likely in 1987). (Altonji goods-based inputs are asso at home?these tion, aggregation in pseudo-panel data helps reduce the biasing of meals positively prices prepared
and
income
with
ciated
a
have
themselves
effect
negative
on
consumption, then theomission of explicit measurement of full prices
a negative impart same The argument
bias
will
ticities.
to the estimated
income
to the case
can be applied
elas
effects
of measurement
those
prices arising from, say, liquidity constraints thataremost likely in low-income households (Cardoso and Gardes 1996). Taking
case
on
estimated
elastic
the income
expect
error
of variables
like
we
is not
location,
confine
true panel using to be serious likely
social
age,
our
data.
income
instrumented
measurement
Because
of virtual
so we
error,
ityparameters estimated with pseudo-panel data tobe similar to
and
category, variables
instrumental
in the family
adjustments
into account
the virtual
composition, to our income
sources
as
errors in our dependent expenditure variable are included in
such
re from nonmonetary prices appearing con to or due the choice time, space restricting
straintsapplying only to subpopulations, or changing from one period to another,may help us better identify and understand and
cross-sectional
series
time
differences.
estimation
because
biases,
on
information
they contain
in
changes
expenditures and income for the same households. Differencing successive
panel
waves
nets
out
effects
the biasing
of unmea
sured persistent characteristics. But while reducing bias due to omitted variables, differencing income data is likely tomagnify source
another
error. Altonji
measurement
of bias:
and
Siow
(1987) demonstrated the likely importance ofmeasurement er ror in the context of first-differenceconsumption models by showing
that estimates
of
elasticities
income
are
several
times
higherwhen income change is instrumented thanwhen it is not. Deaton (1986) presented the case for using "pseudo-panel" data
to estimate
searcher
has
demand independent
He
assumed
cross-sections
with
systems.
expenditure and demographic cross-sections
in successive
years
that the
the
re
required
information and shows how can
be
grouped
into compa
rable demographic categories and then differenced to produce many of the advantages gained from differencing individual panel data. Grouping into cells tends to homogenize the in dividual effects among individuals grouped in the same cell, so that the average specific effect is approximately invariant between two periods and is efficiently removed by within or first-difference
transformations.
residuals
total
and,
errors
data when
expenditure
correlated
should
not bias
of our
the levels
appear
do not contain
cells
corresponding
with
in
estimates.
the coefficient
can
in measurement
Measurement
predictors.
unless
variables,
dependent Special
Panel data on households provide opportunities to reduce these
model
and
in pseudo-panel individ the same
uals in two differentperiods. Thus if the firstobservation for cell 1 during the firstperiod is an individual A, then itwill be paired with a similar individual B observed during the second period,
error arises
so that measurement
this observa
between
tion of B and the truevalues forA ifhe or she had been observed during the second period. Deaton (1986) treated this problem as a measurement error; cohort
sample-based cohort
averages.
He
sure convergence
averages proposed
measures
are error-ridden a
Fuller-type estimates.
of pseudo-panel
of true to en
correction However,
Verbeek
and Nijman (1993) showed thatDeaton's estimator converges with thenumber of timeperiods. Moreover, Le Pellec and Roux (2002) matrix
showed used
that the measurement in Deaton's
estimator
error is often
correction not
variance
positive
defi
nite in the data. The simpler pseudo-panel estimator used in this article has been shown to converge with the cell sizes (Verbeek and Nijman 1993; Moffit 1993), because measure ment
error
becomes
negligible
when
cells
are
large.
Based
on
simulations, Verbeek and Nijman (1993) argued that cells must contain about 100 individuals, although the cell sizes may be smaller if the individuals grouped in each cell are sufficiently homogeneous.
244 error
the measurement
Resolving
problem
by
using
large
samples within cells creates another problem: loss of efficiency of the estimators. This difficultywas shown by Cramer (1964) and Haitovsky (1973) with estimations based on grouped data and by Pakes (1983) with the problem induced by an omitted variable with a group structure,which is similar to the problem error.
of measurement
to a minimum
but
also
error
measurement
keeping
ignorably
small (Baltagi 1995). Grouping methods were developed by Cramer (1964), Haitovsky (1973), Theil and Uribe (1967), and Verbeek and Nijman (1993) that involve carefully choosing co horts to obtain the largest reduction of heterogeneity within each
cohort
but
same
at the
time maximizing
the heterogene
itybetween them. Following these empirical principles, using leads to consistent pseudo-panels out the problems with associated
and
efficient
true panels.
with
estimators Our
work
groups
individuals into cells thatare both homogeneous and large. a
duces
inherent
the aggregation
Second,
systematic
heteroscedasticity.
in pseudo-panel can be This
data
corrected
exactly by decomposing the data into between and within di mensions and computing the exact heteroscedasticity on both dimensions.
But
because
the heteroscedastic
factor
on
depends
can
result
in serious
errors
estimation
to, cell
size.
Thus
the
least
squares
coefficients
on
computed
the grouped data may differ slightly from those estimated on individual data. As described in the next section, this approxi mate and easily implemented correction uses GLS on thewithin and
between
dimensions
as computed structure eroscedastic
with
a common
the between
matrix,
variance-covariance of
transformation
the het
to aggregation.
due
Third, unmeasured heterogeneity is likely to be present in both
panel
and
pseudo-panel
data.
In
the case
of panel
data,
the individual-specific effect for household h is a(h), which is assumed
to be
constant
through
time.
In
the case
of pseudo
panel data, the individual-specific effects for a household (h) belonging to the cell (H) at period t can be written as the sum of two orthogonal effects,a(h,t) = ?(H) + v(h, t). Note that the second component depends on time,because the individuals composing the cell H change through time. The specific effect ?jlcorresponding to the cell H [?(H)] rep resents the influence of unknown explanatory variables W(H), constant through time, for the reference group H, which is defined here by the cell selection criteria. Here v(h,t) are individual-specific variables
planatory
effects Z(h,t).
effects containing In the pseudo-panel
gregated specific effect ? (H) for the cell H aggregation of individual-specific effects,
of unknown data,
ex
the ag
is defined as the
a part
cancels
aggregation
of
Therefore, with panel data, thewithin and thefirst-difference operators suppress all of the endogeneity biases. With pseudo data,
panel
the same
operator
the endogeneity
suppresses
due
to ?i,but not thatdue toJ2 Y (h, t) x v(h, t). For each individual, this part of the residual may be smaller relative to ?i, because cell homogeneity is increased. Conversely, the aggregation into cells is likely to cancel this same component v across individ uals, so it is not easy to predict the effectof the aggregation on the endogeneity bias. Our search for robust results is facilitated by the fact that two
that we
dataseis
panel
use
cover
different
extremely
societies and historical periods. One dataset is from theUnited States for 1984-1987, a period of steady and substantialmacro economic growth. The other is from Poland for 1987-1990, a turbulentperiod that spans the beginning of Poland's tran sition
from a command
to a free-market
economy.
3. SPECIFICATION AND ECONOMETRICS OF THE CONSUMPTION MODEL
and
Gardes,
(Gurgand,
because
data,
aggregated
this effect.
time, correcting it by generalized least squares (GLS) makes individual-specific effects vary with time, thus canceling the spectral decomposition inbetween and within dimensions. This Bolduc 1997). The approximate correction of heteroscedastic ity thatwe use involves weighting each observation by a het eroscedasticity factor that is a function of, but not exactly equal
April 2005
Statistics,
operators, whereas the individual effect v(h,t) may be largely eliminated by the aggregation. Thus it can be supposed that the endogeneity of the specific effect is greater on individual
the
pro
& Economic
The within and first-difference operators estimated with panel data cancel the individual-specific effectsa(h). The com ponent ?i(H) is also canceled on pseudo-panel data by the same
than on
The answer to the efficiency problem is to define group ings that are optimal in the sense of keeping efficiency losses
Journal of Business
Data
constraints
force
us
a demand
to estimate
system
on
only two commodity groups over a period of 4 years: food con sumed
at home
from-home estimates
and
food are
food
consumed
not
very
reliable,
from
away
are rare
expenditures
home.
in the Polish
but we
keep
them
data,
Away so our
to compare
estimates. We use the almost-ideal demand system (AIDS) developed by Deaton and Muellbauer (1980), with a quadratic form for the natural logarithm of total income them with PSID
or
expenditures
to take nonlinearities
into account.
that the true quadratic system proposed by Banks, Blundell, and Lewbel (1997) implies much more sophisticated econometrics if the nonlinear effect of prices is taken into ac count. Itmay be difficult to precisely estimate the price effect because of the short estimation period in the PSID data. On the other hand, our Polish dataset contains relative prices for food that change both across 16 quarters and between 4 so Note
cioeconomic
classes.
Thus
we
estimated
the linearized
version
of Quadratic Almost Ideal System (QAIDS) (with the Stone index) on the Polish panel, using the convergence algorithm proposed by Banks et al. (1997) to estimate the integrability parameter e(p) in the coefficient of the quadratic log income [see (1)]. We obtained very similar income elasticities for food consumed at home and food consumed away from home to those obtained by the linearAIDS model. The AIDS estimates for both countries are presented in Tables 1 and 2. The ad ditivity constraint is automatically imposed by ordinary least
squares (OLS). The possible correlation between the residuals of food con t) xa(h, Y?,Y(h, sumed at home and food consumed away from home would where t indicates theobservation period and y is theweight for suggest the use of seemingly unrelated regression (SUR). the aggregation of h within cells. Note that the aggregate, but We tested this possible correlation on Polish data and found no significant difference between OLS and SUR estimations. not individual-specific, effects depend on time. = ?(H, t) ^y(h,
= t) ?(H)+
t) x v(h,t),
et al.: Estimation
Gardes
245
of Food Consumption
Elasticities
1. Income Elasticities
Table
at Home
forFood
Food away Not instrumented Instrumented Whole population Robust estimation
Pseudo-panel Food at home
Age,
.963 (.019)
.872 (.039)
and
away
1.387
.265 (.056)
1.265
(.068) (.060)
1.384
1984-1985-1986-1987 90 Log of age and itssquare,
Surveys n
(.043)
.062 (.063)
.375 (.166) .387 (.166)
.431 (.260) .447 (.258)
.240 (.095) .460 (.062)
.382 (.114)
(.083) 1.125
.800 (.150) (.127)
.847 (.172)
scale
and
its square
Control
variables
NOTE:
of totalexpendituresby theusualmethod. The values inparentheses are standarderrorsadjusted forthe instrumentation
relative
do
prices
not
waves
within
vary much
of
equivalence
of prices, shocks
variations
other macroeconomic
we
account
with
for price
survey
year
he or she was
Table 2.
Elasticities
forFood
at Home
Between Panel Food at home
Not instrumented Instrumented
Food
away
Not instrumented Instrumented
and Away From Home: Cross-sections
Polish
Surveys
(1987-1990)
Within
First differences
.466 (.005) .755 (.013)
.451 (.010) .788 (.028)
(.073) (.143) 14,520
1.239 (.164) 1.326 (.329) 14,520
2.618 (.556) 4.195(1.080) 14,520
1.460 (.185) 1.315 (.393) 10,890
.583 (.011) (b) .591 (.010) (c) .591 (.011) (d) .589 (.013)
.572 (.017) .584 .581 (.018) .581 (.018)
.549 (.020) (.022) .526 (.020) .965 (.023)
.864 (.033)
.820 (.203)
.890 (.258)
-.218 (.318)
.696 (.331)
level, location, Log of age, proportion of children, education and yearly dummies price forall commodities, cross-quarterly
log of relative
(not instrumented)
(a)
Food
is the expenditure budget share on good /by house wlht at hold h time t, Y ht is its income in the case of theU.S. data or logarithmic total expenditure in the case of the Polish data,
where
.536 (.009) .567 (.023)
1.119 1.216
variables
Pseudo-panel Food at home
(1)
(.004) (.010)
.579 .494
n Control
= d + blMYht/p,)+ + ZhA1+ u\v c'/eipnHYht/pt)]2
effects
surveyed.
Total Expenditure
30%).
from our various
models
are
produced
in their purchasing
pseudo
of the sample.
has
heteroscedasticity
been
corrected
panel
data.
The
between
and
cross-sectional
estimates
are similar for the differentcorrection methods, especially for food
at home,
but
the within
and
first-differences
tained under the false correction, which estimations,
pseudo-panel those rection,
computed or no correction.
Thus
estimates
estimates
than
the exact
correction,
correction
ob
is currently used in
different
very
gives
for the approximate
cor
for heteroscedasticity
seems to be an importantmethodological point in the estimation
on
data. The false correction gives very different for time series estimations. in However, especially our case the exact correction rise to estimated gives parameters to those obtained so we close correction, by the approximate pseudo-panel
estimates,
principally discuss these estimates thatcan be easily compared under the spectral decomposition into the between and within dimensions.
Looking first at the PSID results for at-home food expen ditures, it is quite apparent that elasticity estimates are very sensitive
to adjustments
error
for measurement
Cross-sectional
heterogeneity.
estimations).
the cross-sections Pseudo-panel cross-sectional
in Tables
data,
pseudo-panel
robust
of very-low-income
presented
4%
represent
estimates
and
unmea in
of at-home
come elasticities are low (between .15 and .30) but statistically significantwithout or with instrumentation (when performing
5. RESULTS Estimates
observations
Rejected
by the approximate method (GLS with a heteroscedasticity fac tor 8h, that is constant through time) [Tables 1 and 2, (a)], and the exact method [Tables 1 and 2, (b)] presented inAppendix A. Also given inAppendix A are the estimates obtained without correction [Table 2, (c)] and with a false correction (GLS with a heteroscedasticity factor 2/y/n,where n is the number of ob
re
and
Journal of Business
1
(PSID) and 2 (Polish surveys). Respective columns show in come forPSID and total expenditure for thePolish data, along
The and
data
between
also
effectively
low estimates
produce similar
produces
estimates.
estimates
elasticities some
Despite
average
of elasticities. for between
variations
among
and the
different estimations, the relative income elasticity of food at home is around .20 based on this collection ofmethods. Within and first-difference estimates of PSID-based in come around
elasticities .40 with
are
around
instrumentation.
0 without A
instrumentation
Hausman
test strongly
and re
jects (p value < .01) the equality of within and between esti with estimates for between, cross-sectional (com elasticity mates (Table 3). The test compares within and GLS estimates; as the means on of cross-sectional estimates obtained puted equivalently, it can be built from thewithin and between es each cross-sectional and first-difference mod survey), within, timates (see Baltagi 1995, p. 69). The test is computed by are also for models els. Results in which the usual quadratic form,with a chi-squared distribution,with presented separately income or total expenditure is and is not instrumentedusing the V defined as the variance of the difference between the esti models detailed inTables A.2 and A.4. We expect thebetween mators tested, = or (?b j8w),where ? ?wyCV~{)(?b (?iy) = to the average to be similar of cross-sectional estimates esti on for the estimation the Polish ? quadratic panel (?ly, ?^) mates. Compared with thewithin estimates, the first-difference and V = to all the variables. V?, + Vw corresponds
explanatory
Note that a testwith V as a matrix 2x2 computed only for the specific effectsmay be better taken into account whenever they two income variables would be biased. Because thewithin and firstdifferencemodels adjust forper change within theperiod. For the PSID, we eliminated some observations to obtain sistent heterogeneity and the instrumentationadjusts formea robust estimations using theDFBETAS surement of the estimate error, .40 is our preferred explained by Belsley, approximate
estimates
may
be biased
by greater
measurement
error, but
the
et al.: Estimation
Gardes
Elasticities
of Food Consumption
249
Table 3. Hausman
Without
Instrument
instrumental
Test for Income Parameters
variables
With
and
cross-sectional
413.7 343.4
estimates,
parameter
that
suggesting
failure to adjust forheterogeneity imparts a considerably down ward
to cross-sectional
bias
estimates.
PSID-based
across
the full set of pseudo-panel
are
on
computed
total
estimates.
so
expenditures,
they must
be
multiplied by the income elasticity of total expenditures, which is around .7, to be compared to the PSID income elasticities.) Higher elasticities are to be expected for a country in which food constitutes a share of total expenditures that is three times higher than that in the United States (Tables A.l and A.3). Their consistency probably stems from the smaller degree of measurement error in the Polish expenditure than in the PSID income
data.
Before were
10 households was
On
of food
the panel some
pressing
The
suppressed.
the prediction
mates.
were
the households
robust
data,
outliers
for suppression
in cross-sectional
estimations
similar
gave
produced
by
to those
results
a
little
smaller
than
estimates
on
based
income are
expenditures to specification timates.
elasticities
for
away-from-home even more
from
and
than the at-home
food
expenditure
quite
Between
and
cross-sectional
food
es
around
1.0
in both individual and pseudo-panel data. In contrast with the case
of
at-home
expenditures,
for heterogeneity
adjustments
through use of within and first-differenceestimates produce much lower estimates.We speculate on why thismight be the case in Section 6. Polish the survey
food-away-from-home relatively
expenditures are very low
and
rarely
are even
reported when
in
com
pared with those for other countries with comparable income levels.
teens with
about
Moreover,
served periods and
70%
cafeterias.
with
in the ob
expenditures
is spent on highly subsidized business can Thus
the estimations
the food-away-from-home
countries
these
of
should
be
compared
estimation
expenditure
in other
caution.
Price data inPoland enabled us to compute price elasticities. Quarterly
price
indices
for four
puted from themonthly GUS
social
categories
(Polish Main
from
time
were
of the
estimates
good
com
Statistical Office)
es
the cross-sectional
elasticities,
series.
As
prices
between
change
quarters
and house
levels.
a correlation
Such
is less
than
probable
the correlation
between relative income and the specific component of food consumption (which produces the endogeneity bias on income elasticities). However, if systematically high (or low) price lev are
els
correlated
with
household
type
(e.g.,
as
in segmented
markets), and if these types of households are characterized by a
systematically
positive
or
negative
specific
then
consumption,
the endogeneity bias can appear. This is not the case forPoland in 1987-1990.
6. DISCUSSION have
estimates
expenditure to adjust
Failure cases,
to assess
attempted
elasticity
food
sensitive
elasticity are
estimates
the income
error and unobserved
the microdata.
different
with
sup
obtained
elasticities.
PSID-based
provides
holds of different types, the only explanation of an endogeneity bias would be a correlation between household types and price
We
the estimations on Polish data also produced higher within and first-differenceelasticities than between and cross-sectional
categories
timates of direct price elasticities are close to those obtained
ture
the whole,
On
social
esti
for thewhole panel. Time series Polish pseudo-panel estimates
were
over
and
into cells,
grouped
criterion
consumption
time
In contrast
pseudo
Expenditure elasticities for at-home food estimated with the Polish data are much higher in value than the income elastic ity estimates based on PSID data. (Note that the Polish elas ticities
publication Biuletyn Statystyczny and imputed at the individ ual level to the dataset. The variability of the prices both over direct compensated elasticity for food at home (around ?1).
panel data also produce significantly (according to a Hausman test) higher elasticity estimates for first differences compared with between and cross-sections models, although there is little consistency
variables
46.8 .5 292.4 2.2
income elasticity of at-home food expenditures in theUnited States. This is roughly double the sizes of the corresponding be tween
_^anei_ Pseudo-panel without instrumental
instrumental variables
107.7 PSID Poland 336.6
Poland (QAIDS)
(food at home)
measurement
the bias
caused
by
direction
income
for estimates
of
expendi
adjustments operate in the
the at-home
ticityfound in thePolish data. In the case
and
to measurement
In the case of U.S. at-home heterogeneity. our estimate around .40. is elasticity, preferred some for unmeasured in and, heterogeneity error appears to impart a substantial down
ward bias to this estimate. These same
on
inattention
of U.S.
food
away-from-home
elas
expenditure
our
expenditures,
preferred elasticity estimate is less certain but similar inmag nitude to the .40 elasticity for at-home expenditures. (Note that it is considerably higher in pseudo-panel estimate.) Surprising here is themagnitude and sign (upward) of the apparent bias inherent in both individual and pseudo-panel estimates thatdo not adjust for unobserved heterogeneity. should unmeasured heterogeneity induce an upward
Why bias
in away-from-home
in the U.S.
expenditures
and
Polish
data? Earlier, we speculated that a likely downward bias for the at-home food elasticity estimates may be caused by fail ure to account for the fact that the value of time differs across households served
and is positively related to the household's
income.
The
time
input
to producing
at-home
ob
meals
leads households to face different (full) prices of consump tion even if the prices of the goods-based inputs are identi cal. If these prices are positively associated with income and themselves
have
a
negative
effect
on
consumption,
then
their
omission will impart a negative bias to the estimated income elasticities.
Note
that the bias
seems
somewhat
less pronounced
Journal of Business
250 on
The
data.
pseudo-panel
income
between
correlation
and
In the case there
rants, service
spent
food
and
restau
full-service
consuming
restaurant
fast-food households
this case higher-income to such time value positive to control for this source Failing
a more
attach
households.
low-income
in
but
meals,
well
may
of heterogeneity will probably impart a positive bias to esti mates of income elasticities. (Note that the endogeneity bias is much more pronounced for food away fromhome than for food at home.)
The use of time can be measured with the shadow price of
time. This
shadow
household
with
increases
price
so
income,
that the complete price for food, computed as the sum of the and
monetary distribution. and
time
shadow
prices,
series
increases
the difference
Therefore,
can be
elasticities
income
also
the
along
between
income
cross-sectional to the change
related
of this complete price. The argument is formalized inAppen dix B. Calibrating the price effect as half of the income effect (as suggested by Frisch) and using (B.2) inAppendix B pro duces an estimate of the income elasticity of the food shadow price (Table 4). It is remarkable that these elasticities are simi lar in the two countries, positive and around 1 for food at home straint
larger
for food
stronger
relative
and much
and negative
much
imparts
than at home,
in restaurants
because
parts
the ratio of their complete
when
a similar
changes
prices
but greater
in the expenditures,
change
cross-sectional as
im
changes
in the food away fromhome budget share. Elasticity estimates from thePolish data also proved some what sensitive to adjustments for heterogeneity. This is not surprising given the very differentprice regimes in the Polish economy during this period, with one (presumably low) set of prices faced by themany farm families who have the option of growing theirown food; official, subsidized prices set by the Polish government; and higher black market prices for the same such
grew
food
a
were
as bread, for
their own
set so low
that very
consumption,
queuing
was
not
problem.
Table
4.
Income Elasticity
of Food
PSID
Shadow
Incomeelasticity
.19 Food at home Food away from home Direct price elasticity Income elasticity of the shadow price
1.00 (FH/FA) (a) FH (b) FA
-.19 1.00 .71 -4.78 -3.13
three
from
we
model,
2-year
independent
our
treated
panels
inclusion
of the short-run
the values
changed
income
elasticity
very little. An important result of thiswork is that pseudo-panel esti are often
mates
to estimates
close
on
based
data.
panel
genuine
Large and similar apparent endogeneity biases were found in both
countries.
estimations
Cross-sectional
elasticities
produce
that are systematically higher for food at home and lower for food
from home.
away
lowers
the endogeneity
search
should
data.
expenditure
for food
bias
to verify
aim
it seems
Moreover,
Once
these
that the aggregation
re Future consumption. on a set of complete
results
for the biases,
corrected
elastic
income
ities for food at home and food away from home become very
close
to each
other
in the U.S.
a result
data,
seems
that
rea
sonable to us and highlights possible errors that can arise from estimations
using
data.
cross-sectional
ACKNOWLEDGMENTS The authors gratefully acknowledge helpful suggestions from participants at seminars at the University of Michigan, Northwestern University, Universit? de Paris I, Universit? de Gen?ve, Erudite, CREST, Journ?es de Micro?conomie Ap pliqu?e (1998), Congress of the European Economic Associ ation,
and
International
on Panel
Conference
as
as well
Data,
The research support from INRA, INRETS, and CREDOC. Polish data were made available by Professor G?recki, Uni versity
of Warsaw,
thankCREST,
of Economics.
Department
INSEE, andMATISSE
The
authors
for support.
APPENDIX A: HETEROSCEDASTICITY THE PSEUDO-PANEL MODEL Write
the pseudo-panel trix form as
model
also
presented
IN
in the text in a ma
where y is anNT column vector with N the number of cohorts
panel
and T
3,630 By social category .38 .39
ver
dynamic
y = X/? + Za + e,
1987-1990
TS
CS
Polish
a
diture and found that the coefficient on the lagged dependent variables were significant in these dynamic models, but their
Prices
(U.S.)
Period 1984-1987 n2,430 No Prices
came
they
though
few farm families
when
and
(1984-1985 through 1986-1987 for the PSID and 1987-1988 through 1989-1990 for thePolish expenditure survey).We es timated this dynamic model as a system using SURE and both with and without instrumentationfor log total income or expen
or similar products. In fact,until 1988 the official prices of sta ples,
of our dynamic
estimates
the budget
shares of food away are much smaller in both countries (espe cially in Poland, where the ratio of the price elasticities is the largest). A substitution between food at home and food away from home
model
consumption
this by estimating
investigated
in expenditures
changes
is normal
which
We
persistence.
the time con
Thus
away.
a static
April 2005
riskof bias due to theomission of dynamic factors, such as habit
data
is typically longer than time spent consuming
rant meals
than
Time
components.
on
Statistics,
sion of ourmodel that included lagged consumption. To obtain
in restau
consumed for meals of expenditures are of in the mixture variations large
are based
estimates
Our
spe
cific effectmay be decreased by aggregation, as we suggested in Section 2.
& Economic
CS
TS
.49 1.22 -.38/-.18
.76 .36
the number
of periods,
X
is a matrix
of explanatory
vari
ables, ? is a vector of parameters, Z is an (NT x N) matrix that contains
cell
dummies,
a
is a vector
of cohort
effects,
and
e
is
a heteroscedastic residual. Call D a diagonal (NT x NT) ma trix such thatD = diag( 0 Budget share for food away from home Percent with away-from-home share = 0
In head's
age
In family size
-.097 (.047) -.093 (.051) .005 (.042)
Wage growth'Promotedf Wage growth*Promotedi_i Wage growth*Promotedi_2
.043 (.074) -.149 (.075) -.035 (.069) .037 (.021)
.034 (.028) -.016 (.026)
-.054 (.026) .138 (.006) .012 (.003) .022 (.003)
City size>3 head Education Wage Wagef_-|
Deviations
Level .484
(.14)
(.15) 100
100
.006
(.02)
(.03)
10.65
(.45)
3.789
(.33) 1.140
(.59)
The values inparenthesesare standarderrors.
Used
in the Polish
(.14)
(.02)
(.16)
.020
-.019
(.24)
Level
Dit
.554
.068
(.18)
(.17)
(.15)
(.17)
100 .005
(.02)
12.25
( 02)
(-62)
(.32)
( 15)
1.095
(.61)
.005
(.03)
.0002
( 03)
20.5 -.18
(.79)
3.824
100
100
-.001
26.9 .50
(.32) (.60)
1990 Dit .003
100
(.38)
1.121
Analyses
.486
-.001
(.49)
3.809
Level
-.024
29.7 11.17
Panel
1989 Dit
100
.006
28.4
of Variables
1988 Dit
.508
for
expenditure
Wage growth*Layofft Wage growtrfLayoff^-i Wage growth*Layoffi_2
Region-|_2 Region>3 City size-|_2
1987 Level
In household
error)
The values inparentheses are standarderrors.
Table A.3.
Budget share food at home
for
.001 (.015) .073 (.004) -.0007 (.00004) -.063 (.033) -.003 (.035) -.005 (.032)
Age head head squared
Age
.037 (.029) .202 (.030)
Marriage
(standard
Birth
.050 (.024) -.005 (.022) -.005 (.016)
growtht Wage Wage growthf_-| Wage growtht_2 Divorce
variable
Independent
-.271
hrsf_i hrsf_2
Unemp Unemp Unemp
Level
Coefficient
error)
.025 (.022) .047 (.021) .017 (.021)
Promoted^
NOTE:
51.3 .033
3.7801
(.013)
(.081)
.002
(.686)
.0267
(.359)
-.003
.033
10.1238
(.320)
3.7573
(.013)
.134
(.096)
10.3
.1729
10.1714
(.280)
.008
(.082) 41.5
5.5
9.6
.0731
(.657)
3.7044
head
5.7
.137
-.003
(.033)
Dit
Level
(.100)
74.0
.001
9.9254
(.086)
0
53.2
1987 Dit
Level
-.015
(.095)
.034
(.648) Inage
.129
(.084)
.033
9.5
Analyses
1986 Dit
Level
-.003
0
0
Budget share for food away from home
In household
.147
in the PSID
Used
1985 Dit
Level
(.103)
for
of Variable
Deviations
1984
1983 Level Budget share food at home
251
of Food Consumption
.014
-.026
(.21)
14.14
-.03
(.50)
(.58)
(.32)
(.15)
3.842
1.081
(.61)
.019
-.014
(.22)
Journal of Business
252 Coefficient and Standard Errors for Instrumental Regression for Level and Change for Total Expenditure Variables Equation Panel the Polish Expenditure
Table A.4.
Coefficient variable
Independent
(standard
Log Children
Log age Location
-.270
in family
percent
-.0299
squared
1.704
age
Log
(Ref: countryside) Large
.028
city Average Small city Social category . (Ref: wage earners) earners-farmers Wage
.017
-.096
Two
.004
income squared
-.028
sectional
D
matrix
=
]8GLS analysis,
between
and within
be
as
expressed which
estimators,
sum
a weighted
is called
of
decom
spectral
Gurgand et al. (1997) showed that this can be the case only ifmatrix Bf? is symmetric,which implies that theweights inD are time-invariant.The intuition of this result is that if, in the effects
individual
the variances,
scaling
no longer
transformation
a
to the mation
is no
longer
over
time,
thus
is constant
individual
resolves
heterogene
possible.
In contrast,
this argument
does
as
long
not hold,
as cell and
size
the de
obtains.
composition
A corollary is that the efficientwithin estimator cannot be based simply on the original model weighted by heteroscedas ticity
factors,
time-varying
because
this would
create
time
varying cohort effects.Gurgand et al. (1997) showed that the within estimator that is both efficient and consistent when a
are correlated
and X
?w where when
A
=
D"1
the number
= -
is
(X/WAWX)"lX/WAWy,
is de all
contains
the residual
effect
variables
an
?
cross
o?ih +
?iht- The
a correlation by
between
some
specific
effect
Sint: utht
can be biased
Znt
this
and
1978). Such a correlation is due to latent per (e.g.,
event
during
characteris
infancy,
specific of
effect
permanent
variables
the explanatory
to the between
and
transfor