Sie sind auf Seite 1von 18

Aggregate Consumer Behaviour without Exact Aggregation

Author(s): Christopher J. Nicol


Source: The Canadian Journal of Economics / Revue canadienne d'Economique, Vol. 24, No. 3 (
Aug., 1991), pp. 578-594
Published by: Wiley on behalf of the Canadian Economics Association
Stable URL: http://www.jstor.org/stable/135581
Accessed: 07-03-2016 12:14 UTC

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 support@jstor.org.

Wiley and Canadian Economics Association are collaborating with JSTOR to digitize, preserve and extend access to The
Canadian Journal of Economics / Revue canadienne d'Economique.

http://www.jstor.org

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

Aggregate consumer behaviour without

exact aggregation

CHRISTOPHER J. NICOL University of Regina

Abstract. A model of consumer behaviour is estimated by pooling Canadian cross-sectional

and time-series data. In contrast to a similar study by Jorgenson, Lau and Stoker (1982), exact

aggregation, homogeneity, and symmetry are not imposed but are tested. Exact aggregation

is rejected. However, these data are less sensitive to homogeneity and symmetry restrictions.

Demographic effects appear to be important determinants of demand. The model can also be

estimated using only cross-sectional data. These results indicate that exact aggregation should

not be imposed, and household demographic effects should be modelled. Homegeneity and

symmetry seem relatively less restrictive for these data.

Le comportement agrege des consommateurs en l'absence d'agregation exacte. L'auteur

calibre un modele de comportement des consommateurs en combinant des donnees tirees

d'etudes synchroniques et diachroniques. Contrairement a ce qui avait ete le cas dans une

etude similaire de Jorgenson, Lau et Stoker (1982), l'auteur n'impose pas des conditions

d'agregation exacte, d'homogeneite et de symetrie: il s'agit de conditions dont on veut

tester l'existence. On rejette la condition d'agregation exacte. Cependant, les donn6es sem-

blent moins sensibles aux restrictions imposees par l'homogeneite et la symetrie. Les effets

demographiques semblent etre des determinants importants de la demande. On peut aussi

calibrer la modele a partir des donnees synchroniques seulement. Les resultats montrent que

l'on ne devrait pas imposer la restriction d' agregation exacte et qu' il faut inclure les carac-

teristiques demographiques des menages dans le modele. Les conditions d'homogeneite et

de symetrie semblent relativement moins restrictives pour les donnees.

This analysis is based, in part, on the Statistics Canada 1978, 1982 and 1984 Surveys of Family

Expenditures Public Use Microdata Files, which contain anonymized data collected in the 1978,

1982, and 1984 Family Expenditure Surveys. All computations in this paper on these microdata

were prepared by the author. Responsibility for this use and interpretation of these data is entirely

that of the author. The author acknowledges financial support from the University of Regina

President's Fund and the Government of Canada 'Challenge '87 and '88 Summer Employment

Programmes.' which met some of the costs of data purchase and research assistance, respectively.

Lorraine Nicol and Binyam Solomon provided valuable research assistance. I also thank two

anonymous referees, whose comments led to improvements in this paper. Any remaining errors

are the responsibility of the author.

Canadian Journal of Economics Revue canadienne d'Economique, XXIV, No. 3

August aout 1991. Printed in Canada Imprime au Canada

0008-4085 / 91 / 578-594 $1.50 ? Canadian Economics Association

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

Aggregate consumer behaviour 579

I. INTRODUCTION

In an important paper by Jorgenson, Lau, and Stoker (1982) (hereafter, JLS),

an innovative model of aggregate consumer behaviour was presented. Their model

consisted of a system of micro-level demand equations, which were then aggregated

across households, giving a system of aggregate or macro-level demand equations.

The combined systems of micro- and macro-level demands were then estimated

using pooled, time-series and cross-sectional data.'

For tractability in deriving aggregate demand function, JLS incorporated three

sets of parameter restrictions in their model at the household and aggregate levels.

First, exact aggregation was imposed. Second, variations in consumer behaviour

due to differences in demographic characteristics were modelled as intercept-shift

dummy variables. Finally, homogeneity and symmetry were imposed.

The above restrictions in the JLS model, although convenient, are non the less

testable. Indeed, when estimating a system of Engel curves using a cross-section

of micro-data, Nicol (1989) found that the exact aggregation restrictions imposed

by JLS were rejected. Nicol's (1989) results also showed that micro-level demo-

graphic effects in demand functions should be more general than intercept-shift

dummy variables. Furthermore, applied demand researchers have been increasingly

interested in explicitly testing (rather than imposing) homogeneity and symmetry

(see Deaton 1986 for a survey of this literature). In a number of papers, these

restrictions have been rejected (see, e.g., Christensen, Jorgenson, and Lau 1975;

Deaton and Muellbauer 1980).

In this paper a model of consumer behaviour is estimated using pooled Cana-

dian cross sectional and time-series data. However, unlike JLS, exact aggregation,

homogeneity, and symmetry are not imposed a priori but are explicitly tested, and

the modelling of household demographic effects here differs.

Test results indicate that the exact aggregation restrictions are strongly rejected.

Homogeneity and symmetry, however, are supported by the data (with the exception

of one marginal rejection), and demographic effects are found to be statistically

significant determinants of consumer behaviour.

These results suggest that models of aggregate consumer behaviour of the type

estimated here should be implemented in a way that relaxes exact aggregation.

Imposition of homogeneity and symmetry appears, however, to be less restrictive.

The paper is organised as follows. In section ii a model of household behaviour is

specified, based on an indirect utility function. Systems of household and aggregate

demands are then derived, making clear the role of the exact aggregation restric-

tions. In section iII the data used in this study are discussed. Section iv presents the

empirical results of estimation and hypothesis testing. Finally, section v provides a

summary, some potential applications and concluding remarks.

1 Stone (1953) used budget-study income elasticities as prior information in time-series estimation

with the same objective in mind: estimation of price responses via time series data and estimation

of income responses via the cross-sectional data.

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

580 Christopher J. Nicol

II. HOUSEHOLD AND AGGREGATE DEMAND SYSTEMS WITHOUT EXACT

AGGREGATION

To construct a system of demand equations that can be aggregated across consumers

in the way required here, an appropriate household indirect utility function will

first be specified. This function should be conditioned on certain demographic

characteristics.2 All parameters of households' budget-share equations can then

vary, conditional on these prespecified characteristics.

To maintain a degree of parsimony in the model specification, the indirect

utility function is conditioned on household size. Other less important demographic

variables enter the model, so that thy act as intercept-shift dummy variables in the

demand functions. An increased degree of flexibility is thus incorporated in the

demand system, compared to the JLS model. However, the number of observations

relative to estimable parameters is not reduced by as much as if the data were

stratified by all demographic variables in the model.3

If there is a total of J household sizes, the indirect utility function for the kjth

household of size j can be written

In VI = J(lnpT,A), k =1, ...,Kj; j = 1, ...,J, (1)

where ln pT- (lnP1, ..., lnPN); P (p1/Mkj, ..., PN/Mk)T is an N x 1 vector

of normalized prices of N goods and services in the complete system; Mkj is total

expenditure of the kjth household on all N categories of goods and services; and

Akj is an Rj x I vector of household attributes. Writing this indirect utility function

as a variant of the translog indirect utility function of Christensen, Jorgenson, and

Lau (1975) implies the following functional form

nVkJ = F(Aj) In npTQ+ 2n pBiInp +InpI CAkj, (2)

where Fj is a function depending only on Akj, and /i is an N x 1 vector of

parameters, which can be interpreted as first derivatives of Ti, evaluated at a point

of approximation. Also, BJ and CJ are matrices of parameters of dimensions N x N

and N x Rj, respectively. These matrices can be interpreted as having elements that

are combinations of second derivatives of Ti, evaluated at a point of approximation.

A logarithmic form of Roy's identity applied to (2) yields the kjth household's

system of budget-share equations

Pn qnk -*

Mk = Wk (3)

2 Barnes and Gillingham (1984) and Nicol (1989) found in earlier studies that modelling demo-

graphic effects as intercept-shift dummy variables was inferior to stratifying households by demo-

graphic characteristics.

3 A detailed discussion of the justifiability of this approach (apart from parsimony) will be given in

section iII, where the demographic variables are defined.

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

Aggregate consumer behaviour 581

I in Vk/ in Pn

nkj = lnVi/alnMk (4)

In Vk. laInM

_ + BJ In p + CnAkj (5)

l T(l + Bi ln p + CiAkj)

where BJn and CJn are the nth rows of BJ and CJ, respectively; /n is the nth element

of /i and t is a vector of ones of appropriate dimension.

By construction, the vector of household attributes, Ak, contains elements that

are either unity or zero, depending on whether a household does or does not exhibit

a given characteristic. The elements of Ak. are thus dummy variables. This implies

different constants (f3n + CnAkj, in general) appear in the numerator of a given

budget-share equation, depending on the household's demographic characteristics.

Furthermore, the equations in (5) are homogeneous of degree zero in parameters.

Consequently, a normalization of the parameters is required prior to estimation. A

convenient normalization is tT(/31 + C]Akj) = -1 (it is still assumed that LTCi - 0,

as in JLS). The budget-share system (5) then reduces to

* /3 + Bin In - + CnAkj 6
* f +B np+CA (6)

Wnkj -1 +tTBjInp

Simmons and Weiserbs (1979) and Deaton (1986), however, point out that normal-

izing prices by Mkj imposes homogeneity in models such as (6) above. Relaxing

homogeneity yields the following unrestricted model

OJ13 + Bin In p + CnAkj + LJn In Mk


n 3+~n+~A9IlM9(7)

Wnk. -1 + tT(BI ln p+Li lnMI )

in which Li is a vector of parameters. The homogeneity restrictions, Bknt =-LJ,

for all n, and the symmetry restrictions, Bini = Bjin, i =/ n, are, however, relaxed.

Consider now the aggregation of the equations in the system given by (7) across

households of size j. Lewbel (1987) recently suggested that the specification of

aggregate budget shares could be approximated as follows:

* 3J + Bin n p + CAj + LJ E (Mkj In Mkj)/Mj

nj 1 + tLT{Bi ln p + Li E (Mk - ln Mkj)/Mj}

where the expectation operator implies averaging across households of size j; Mj =

E (MkJ); Aj = E (Ak * Mk )/Mj; and the approximation arises in (8) because it is

assumed that

E (ln Mkj - qnkj) iv E (In Mkj) - E (qnkj) * E (Mkj * ln Mkj)/[My * E (ln Mkj)] (9)

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

582 Christopher J. Nicol

(Lewbel 1987, 319).4 The above approximation is chosen by Lewbel (1987), since

it is then possible to rewrite the statistic E (Mkj i In Mkj)/Mkj as follows:

E (Mkj - In Mkj)lMj In Mj + pi + Ti E (yklj/Y) - In (yklj/Y) (10)

= ln Mj +p' +r'vj,

where pJ and Ti are dependent on the distribution of households' (of size j) propen-

sities to consume out of income. These propensities to consume are assumed to be

constant over time. Also, ykj is the income of the kjth household, and yj = E (ykj).

The statistic, vj, however, depends on the distribution of income and consequently

will be time varying. With these relationships in mind, (10) can be substituted in

(8) to yield

* J + BJ In p + CJAn + LJ (ln Mj +fiV;j)

_-+LT{Bilnp+L(InMj+fiV1j)} (11)

Note that since pJ is constant over time, it is absorbed in the normalization of the

equation.

The relationship between (7) and (11) above and the JLS model of aggregate

consumer behaviour can now be highlighted (in addition to the homogeneity and

symmetry restrictions, mentioned earlier). In the former case, the coefficients on

ln p and ln Mkj or ln Mj are distinct for households of different sizes, compared

with JLS, where such parameters are identical across households of all sizes. Fur-

thermore, if tTLi is set equal to zero, the system in (7) would aggregate exactly

across household of size j, in the sense of JLS. Exact aggregation within a group

of households of a given size thus requires imposition of one restriction on the

parameters of the joint system, (7) and (1 1).5

III. DATA

The cross-sectional expenditure data for this study were drawn from the 1978,

1982, and 1984 Survey of Family Expenditures Microdata Files for Canada. As

indicated in the previous section, ln Vk'j is conditioned on household size. Also,

4 Sufficient conditions for (9) to hold with equality are that E (In Mk. qnk.) = E (In Mk.)

E(q,k-) and E(Mk. i In Mkj) = E(Mkj) * E(ln MkA). It is not possibie to verify this directly, as

all the required data are not available. However, an analysis of data from the 1978, 1982, and

1984 Canadian Family Expenditure Surveys - the cross-sectional microdata used in this study,

discussed later - indicates E (In Mkj) * E (qk.) underestimates E (In Mk. qnk.) by between 2-5 per

cent, whereas E (Mk* E (In Mk ) is above E (Mkj In MkA) in two of t1ie years and below it in the

other.

5 It should be pointed out that the treatment of Mj in the macro-level demands in the model

presented here differs from the treatment in JLS. In the former model, reliance is placed on

the approximation in (8) to rewrite the macro-level demands (with or without exact aggregation)

in terms of the variables vj and In Mj. In the latter model, the variable included to model the ef-

fect of Mj on aggregate consumer behaviour reflects the movement of the distribution of Mj over

time. This statistic is estimated using information on the distribution of income. See JLS, 209.

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

Aggregate consumer behaviour 583

the elements of Akj are intercept-shift dummy variables, taking values of one if

the household exhibits a given characteristic and zero otherwise. This is partly for

parsimony. However, using Nicol's (1989) data, it was found that intercept-shift

dummy variables provided an adequate representation of the data for all demo-

graphic characteristics except family size.6 In the case of family size, the data

indicated that stratification was necessary, rather than modelling this characteristic

by intercept-shift dummy variables.

The most convenient household size to condition this empirical implementation

on is one-member households, since it is relatively easier to obtain time-series data

on the other demographic variables for such a group. All observations extracted

from the microdata files therefore pertain to single-member households.

The intercept-shift demographic variables are defined in table 1. To avoid linear

dependence of the columns of Akj for all households combined, A1 and R1 are

excluded, yielding a vector with a column dimension of Rj- 10.

Given the nature of the data and the difficulty of estimation, it is not possible to

identify all the parameters of a large system with the data used here. Consequently,

only three equations are included in the demand system estimated. As the system

is fairly small, it is important to choose a group of goods that is approximately

weakly separable from goods outside the system. The feasibility of approximate

weak separability of the group of goods in this system from other goods is discussed

in section iv.

Owing to the possibility of specification error associated with aggregation across

goods and the effect of this on hypothesis tests (see Nicol 1991), a fairly disag-

gregated group of goods was selected. This group comprised expenditures on food

(F), alcoholic beverages (A), and smokers' supplies (T). The expenditures included

in each category are defined in table 2. This choice of categories is made, since they

pertain to non-durable goods expenditures. Problems associated with the inclusion

of durable goods, where expenditures do no necessarily reflect consumption, are

therefore avoided. Also, the chosen categories are as disaggregated (by type of

good) as possible, given the available data.

It is most likely that zero observations on A and T mean that the householder is a

non-smoker and/or a non-drinker.7 Changes in prices and income will not affect the

behaviour of a household of this type. This situation is quite different from those

where a zero observation is recorded but some consumption does occur, as in latent

variable problems. Also, policy makers are usually interested in the responsiveness

of those consuming a good to changes in, for example, its price. Only householders

with positive expenditures in all categories are therefore included in the cross-

sectional data set, which has a total of 1,733 observations. These households do

include a number with annual expenditures on A and T of below $50, likely those

whose expenditures on A and T are for entertaining. Even though these expenditure

6 In Nicol (1989) the demographic characteristics modelled, other than family size, were age of

head of household; region of residence; and size of area of residence (urban or rural). This was

for comparability with the JLS study.

7 This is entirely plausible, since the data in these surveys cover expenditures over a whole year.

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

584 Christopher J. Nicol

TABLE 1

Definitions of demographic variables

Variables Definitions

Age of householder Householder aged: 15-24, 25-34, 35-44, 45-54, 55-65, and

65 and over (A-A6, respectively).

Region of residence of householder Householder risiding in: Atlantic Provinces (New Brunswick,

Newfoundland, Nova Scotia and Prince Edward Island);

Quebec; Ontario; Prairie Provinces (Alberta, Manitoba and

Saskatchewan); and British Columbia (R1-R5, respectively).

Area of residence of householder Householder residing in an area where population exceeds

30,000 persons, UR = 1. Otherwise, UR = 0.

TABLE 2

Expenditure in each expenditure category, Family Expenditure Surveys, 1978, 1982, and 1984

Expenditure

category Expenditures included

F Food prepared at home, lunches carried from home, bulk food purchases and food

consumed at home from carry-out shops (1978:81; 1982:80; 1984:96)

A Beer, liquor (including liqueurs) and wines (1978:102; 1982:153; 1984:184)

T Tobacco products and smokers' supplies (1978:20 minus 1978:102; 1982:151;

1984:182)

NOTE: Further details of the above expenditures are contained in the documentation issued with the public-

use data tapes by Statistics Canada. Numbers in parentheses indicate the survey and 'field no.'

where variables are located on the tapes.

levels are low, it is still likely that such householders' behaviour is influenced by

their income and the price of these goods, so it is reasonable to include such low

expenditure households.

The households in the foregoing data set have then to be matched with appro-

priate price vectors. Intercity indices of retail price differentials are published for

the expenditure categories in this study in Consumer Prices and Price Indexes

(Statistics Canada, Cat. No. 62-010). To preserve anonymity of respondents, how-

ever, public-use data from the Family Expenditure Survey Microdata Files include

only 'region of residence' (not city of residence) information on households. These

regions of residence are: Atlantic Provinces (New Brunswick, Newfoundland, No-

va Scotia, and Prince Edward Island); Quebec; Ontario; Prairie Provinces (Alberta,

Manitoba, and Saskatchewan); and British Columbia. Using the intercity indices

referred to above, 1978, 1982, and 1984 regional price indices were constructed

for each expenditure category.8 Householders are then matched to one of these

8 Details of the method of construction and the data series are available from the author on request.

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

Aggregate consumer behaviour 585

fifteen price vectors, depending on which cross-section and region of residence the

household is drawn from.

The time-series data used covers the period 1951-86 and consists of four

types. There are data on expenditures and prices of F, A, and T. Also required

are observations on the variables Aj and vj, which were defined earlier.

The expenditure data required are mean expenditures by one-member house-

holds on each of the three categories. These can then be used to compute to-

tal expenditures and the budget shares. Total expenditures on the F, A, and T

series for all households and 'unattached individuals' (one-member households) in

Canada are available for the required period in National Income and Expenditure

Accounts, Annual Estimates, 1926-86 (Statistics Canada Cat. No. 13-531). Per

household estimates are constructed by deflating the F, A, and T series with the

total estimated numbers of families and unattached individuals in Canada in each

year.9

A fairly stable relationship exists over time between the proportions of mean

expenditures on each good spent by unattached individuals on the one hand, and

by all households on the other. These proportions are not constant, nor are they the

same for each expenditure category. Data on these proportions are directly avail-

able for some of the required years from 1971-86.10 For other years, proportions

were estimated by interpolation. Applying these proportions to per household mean

expenditures yields estimates of mean expenditures on F, A, and T by unattached

individuals.

The price data are computed using data on real and nominal expenditures on

F, A, and T, published in National Income and Expenditure Accounts, Annual

Estimates, 1926-86. The real and nominal series are used to generate implicit

price deflators, which are then rebased to 1978, to match the cross-sectional data.

Consider now the variable Aj. Given that the elements of Akj take values of

either zero or one, Aj is a vector with the proportions of total expenditures spent

by agents with a specific demographic characteristic as its elements. It is common

to use the proportions of households exhibiting each demographic characteristic as

proxies for the elements of Aj (see, e.g., Lewbel 1987). These data are published

for most of the required years by Statistics Canada.11

A problem with using the proxy Aj in place of the true Aj is that a given group

of agents could remain the same relative size, but become richer.12 In that case,

9 The 1971-86 observations on these totals are available in Income After Tax, Distributions by

Size in Canada (Statistics Canada Cat. No. 13-210). Some of the observations from 1951-70 are

contained in Income Distributions by Size in Canada (Statistics Canada Cat. Nos 13-528, 13-529,

13-534 and 13-544. The remaining elements of the series (1952, 1953, 1955, 1956, 1958, 1960,

1962-4, 1966, 1968, and 1970) were estimated by interpolation.

10 See Statistics Canada Cat. Nos 62-509, 62-510, 62-513, 62-521, 62-525, 62-527, 62-530, 62-537,

62-541, 62-544, 62-555 and 62-557 for details.

11 The 1951, 1954, 1957, 1959, 1961, 1965 and 1967 data are reported in Catalogues 13-528, 13-

529, 13-534 and 13-559. The 1969, 1972, 1974, 1976, 1978, 1980, 1983 and 1986 data are

reported in Catalogue 13-207, for the years stated. The other data from 1971-1985 are reported

in Catalogue 13-210 for the years stated. Data for 1952, 1953, 1955, 1956, 1958, 1960, 1962-64,

1966, 1968 and 1970 were estimated by interpolation.

12 I am grateful to a referee for highlighting this.

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

586 Christopher J. Nicol

the proxy Aj remain constant, but the true Aj rise. It is not possible to discount this

feature of the data directly for every year, since the necessary data are not available.

However, it is possible to conduct a comparison between the proxy and true Aj for

households in the cross-sectional data set for this study, from the 1978, 1982, and

1984 Family Expenditure Survey years. The age of head variables proxy and true

Aj exhibit a correlation of 0.962, and an R2 of 0.926. The regional variables proxy

and true Aj exhibit a correlation and R2 of 0.999. For data that can be compared,

the proxy and true Aj are therefore more or less proportional. If the other years

follow this pattern, using the proxy Aj will not have an important effect on the

empirical results.

Finally, the vj statistics can be estimated from income quintile and mean income

data, using the approximation

15

M}-5~ ~ q In ( ,}( (Yj ) (12)

q=

where yqj represents the mean income of an unattached individual (a household of

size j = 1) in the qth quintile. Again, mean income and income quintile data are

available in Statistics Canada publications.13

The pooled data set to be used thus contains a total of 1,769 observations: 1,733

cross-sectional observations and thirty-six time-series observations.

IV. ESTIMATION AND RESULTS

To implement estimation, the stochastic structure of the unrestricted model, (7)

and (11), has first to be specified. It is assumed that a random error enters the

micro-level demands, (7), as follows:14

Wnkj = nk Wnkj + nj (13)

It is further assumed that Ekj = (6uj, .. . ) '--' NI (0, Q) for all kj, where rank

(Q) = N - 1 is assumed, due to adding up.

JLS argue that the random errors in the aggregate budget-share functions must

be heteroscedastic. This arises (with or without exact aggregation imposed) be-

cause of the aggregation across agents of the random errors entering (13). That is,

aggregating the stochastic budget-shares, (13), across households yields

Wn jtWt + 6njtj Vt =, ...IT, (14)

13 The data use here are before-tax income quintile and mean incomes of unattached individuals.

They are reported in Statistics Canada Catalogue Numbers 13-529, 13-207 (1971, 1976, 1978 and

1986).

14 This is how the micro-level stochastic structure of the budget-share functions is handled by JLS

also.

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

Aggregate consumer behaviour 587

where w* t = (Pnt Ekj qnkjt)/(Ekj Mkjt); Enjt = (jk MAjt6nk,t)/ sj7k Akj; and t in-

dexes time periods. Based on the assumed properties of Enkj,, Ejt has the properties

.jt)TNl (?, ( MY N2 .) (15)

The estimation procedure employed here will not, however, take account of this

heteroscedasticity, since the data required for the adjustment are not available. The

pooled model will be estimated based on the assumption that the enjt have the same

distribution and parameters as Enkj. Since the proportion of time-series observations

is small relative to the total sample size, this assumption is unlikely to influence

the outcome of the hypothesis tests to be conducted.

The unrestricted N -1 = 2 equation system'5 was estimated with a Broyden-

Fletcher-Goldfarb-Shanno algorithm, as implemented in the non-linear regression

methods in SHAZAM (White 1978), using the pooled, cross-sectional, and time-series

data. Parameter estimates and their t-statistics are reported in table 3. These esti-

mates demonstrate that the demographic variables included as intercept-shift dum-

my variables are important explanatory variables. Residuals diagnostics indicated

the absence of autocorrelation but some evidence of heteroscedasticity. The pattern

of this heteroscedasticity seemed to be related to the ln M,j variable. However, the

correlations between the squared residuals and ln Mkj were close to zero, indicating

that the outcomes of hypotheses tests would be unaffected if uncorrected residu-

als are used in the calculation of the estimated covariance matrix for the system

(Greene 1990, 403). It would, in any event, be extremely complicated to compute

a 'heteroscedasticity-consistent' estimate of the covariance matrix for this model.

Estimation with pooled data of the unrestricted model represents the alternative

hypothesis against which a number of null hypotheses are to be tested. Interest is

first focused on a test of the exact aggregation restriction (requiring that LTL- 0).

Also, the homogeneity and homogeneity plus symmetry restrictions are tested, in

the absence of exact aggregation. The results of these tests are presented in table

4.

The exact aggregation restriction is strongly rejected by the data. The upper-

tail prob. value for this test is approximately zero, as indicated in table 4. This

supports Nicol's (1989) earlier result, rejecting exact aggregation when employing

cross-sectional data alone. It is also found that homogeneity is not rejected at a

significance level of 0.001, but homogeneity plus symmetry is marginally reject-

ed at a significance level of 0.005 (the overall significance level for the test of

homogeneity plus symmetry, in the absence of exact aggregation). The upper-tail

prob. values of these test are also given in table 4. Note, however, that the rejection

15 This reduced system is estimated owing to singularity of 52. The parameters of the Nth equation

can be recovered via the adding-up property of the budget shares, which requires that En W=kj

En wnj = 1. In these and other results reported below, the dropped equation is T.

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

588 Christopher J. Nicol

TABLE 3

Estimated parameters and t-ratios, pooled estimation

Parameters Estimates t-statistics

Parameters Estimates t-statistics

Pi -0.6568 -17.8830

P2 -0.1697 -3.9525

B1,1 -0.1255 -0.6178

B2,1 0.0653 1.1091

B2,2 0.0342 0.4699

B1 2 -0.0112 -0.0878

B2,3 -0.0636 -2.0621

B1,3 -0.1040 -1.7074

C2,1 0.0073 1.8737

Cl,1 -0.0115 -2.1621

C2,2 0.0244 2.6768

C1,2 -0.0274 -2.6335

C2,3 0.0272 2.6703

C1,3 -0.0263 -2.5810

C2,4 0.0358 2.7459

C1,4 -0.0341 -2.6924

C2,5 0.0445 2.8076

C1,5 -0.0544 -2.7818

C2,6 0.0001 0.0346

C1,6 0.0007 0.1893

C2,7 -0.0049 -1.2357

C1,7 0.0018 0.4179

C2,8 -0.0188 -2.5820

C1,8 0.0157 2.3292

C2,9 -0.0140 -2.2644

Cl,9 0.0102 1.7370

C2,10 0.0002 0.0687

C1,10 0.0002 0.0687

L2 -0.0094 -2.2941

LI 0.0845 2.5359

T 2.2419 2.3122

BM1 -0.0736 -0.2483

BM2 -0.0221 0.1010

BM3 0.0302 0.4686

LM 0. 1066 2.4923

NOTES

The superscript j on all parameters reported here and in table 6 is omitted, since all results relate to one-

person households.

The first subscript on each parameter indicates the equation in which the parameter appears. That is,

F = 1; A = 2. The second subscript, where appropriate, indicates which variable the parameter applies to.

For example, B1 ,1 is the F log-price variable appearing in equation F = 1. With respect to demographic

variable parameters, C,,l-C, lo, apply to the variables A2-A6, R2-R5 and UR, respectively.

The parameters denoted as BM1-BM3, and LM are the parameters of the vector LTB and tTL, respectively.

for homogeneity plus symmetry is marginal, compared with the rejection of exact

aggregation.

It is also of interest to test homogeneity and homogeneity plus symmetry when

exact aggregation has been imposed. This will give an impression of how subse-

quent test results are affected, conditional on the imposition of restrictions which

are known to be false. These results are presented in table 5. Neither of these sets

of restrictions were rejected at significance levels of 0.001 and 0.005, respectively.

(Again, the overall level of significance for homogeneity and symmetry combined

is 0.005). In conjunction with the results in table 4, these results indicate that these

data are not too sensitive to the imposition of homogeneity and homogeneity plus

symmetry, compared to exact aggregation. This finding could, of course, differ for

other data sets.

An interesting supplement to estimating the pooled model is to ascertain whether

the unrestricted parameters of (13) can be estimated using the cross-sectional data

alone. The limited variability in prices at the cross-sectional level would seem

to preclude this, which is one reason why pooling is attractive. The unrestricted

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

Aggregate consumer behaviour 589

TABLE 4

Tests of exact aggregation, homogeneity, and homogeneity plus symmetry, pooled estimation

Log. of No. of No. of Test Prob.

Model likelihood parameters restrictions statistic value

Unrestricted 2107.828 35 - - -

Exact aggregation 2072.939 34 1 69.898 0.0000

Homogeneity 2105.361 32 3 5.054 0.1679

Homogeneity plus symmetry 2098.166 29 6 19.444 0.0035

NOTE: The test statistic is -2(In L(H) - In L(H)) - x 2(q). Here, In L(0) and In L(H) are estimates of

the log-likelihood under the null and alternative hypotheses, respectively. The degrees of freedom,

q, of the tests are the number of restrictions required in each case. Critical values of the test statistic

at a significance level of 0.005 for 1, 3, and 6 degrees of freedom are 7.88, 12.8 and 18.5,

respectively.

TABLE 5

Tests of homogeneity, and homogeneity plus symmetry, conditional on exact aggregation, pooled

estimation

Log. of No. of No. of Test Prob.

Model likelihood parameters restrictions statistic value

Unrestricted 2072.939 34 -- -

Homogeneity 2070.375 31 3 5.128 0.1627

Homogeneity plus symmetry 2064.296 28 6 17.286 0.0083

NOTE: The critical values of the test statistic (defined as in table 4) at significance levels of 0.001 and

0.005 for 3 and 6 degrees of freedom are 16.3 and 18.5, respectively.

version of (13), however, was successfully estimated. Again the Broyden-Fletcher-

Goldfarb-Shanno implementation in SHAZAM was used. Parameter estimates and their

t-statistics are reported in table 6. Most of the parameters are estimated precisely.16

This finding is somewhat surprising, although rather useful. What it means is that

pooling might not be necessary once a threshold in price variability has been

achieved. This supports the growing view (see, e.g., Browning 1987; Blundell

1988; Blundell, Pashardes, and Weber 1988) that aggregate time-series data are

of limited use in applied demand studies, with the increasing availability of long

series of household expenditure surveys.

A comparison of corresponding parameter estimates in tables 3 and 6 indicates

that they are generally fairly close to each other. Also, the standard errors from

pooled estimation (table 3) are typically lower than from cross-sectional estimation

(table 6).

The cross-sectional data can also be used to test the hypotheses of exact ag-

gregation, homogeneity and homogeneity plus symmetry. The results of these tests

are presented in table 7. Exact aggregation is strongly rejected, the upper-tail prob.

16 This applies both to the demographic variables and the price variables.

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

590 Christopher J. Nicol

TABLE 6

Estimated parameters and t-ratios, estimation with micro-data

Parameters Estimates t-statistics

Parameters Estimates t-statistics

Pi -0.5704 -8.3817

P2 -0.2475 -4.2927

B1,I -0.2261 -0.9203

B2,1 0.1390 2.1418

B1,2 0.0718 0.4618

B2,2 -0.0369 -0.5119

B1,3 0.0991 1.6137

B2,3 -0.0492 -1.6805

C2,1 0.0073 1.8452

Cl,l -0.0116 -2.1218

C1,2 -0.0277 -2.5363

C2,2 0.0246 2.5706

C1,3 -0.0265 -2.4731

C2,3 0.0274 2.5661

C1,4 -0.0343 -2.5903

C2,4 0.0360 2.6349

C1,5 -0.0548 -2.6627

C2,5 0.0448 2.6898

C1,6 -0.0001 -0.0054

C2,6 0.0008 0.2568

C1,7 0.0079 1.2434

C2,7 -0.0104 -1.7477

C1,8 0.0188 2.3019

C2,8 -0.0216 -2.4575

C1,9 0.0184 1.9461

C2,9 -0.0215 -2.2176

C2,10 0.0006 0.2490

C1,10 0.0032 1.0350

LI 0.0864 2.4370

L2 -0.0093 -2.1649

BM1 -0.0958 -0.3072

BM2 0.0284 0.1269

BM3 0.0414 0.6430

LM 0. 1090 2.4019

TABLE 7

Tests of exact aggregation, homogeneity, and homogeneity plus symmetry, estimation with micro-data

only

Log. of No. of No. of Test Prob.

Model likelihood parameters restrictions statistic value

Unrestricted 2031.867 34 - - -

Exact aggregation 1996.519 33 1 70.696 0.0000

Homogeneity 2029.472 31 3 4.790 0.1878

Homogeneity plus symmetry 2023.060 28 6 17.614 0.0073

NOTE: The test statistic is as defined in table 3.

value of the test being approximately zero. Neither homogeneity nor homogeneity

plus symmetry, however, are rejected at significance levels of 0.001 and 0.005,

respectively.

The results presented above therefore indicate that exact aggregation should not

be imposed. Homogeneity and homogeneity plus symmetry, on the other hand, seem

to be more acceptable restrictions. This result is important for two reasons. First,

the vehement rejection of exact aggregation indicates that the data are sensitive

to certain sorts of restrictions. Second, the homogeneity and symmetry restrictions

have been rejected in many applied demand studies, as mentioned earlier. However,

for the most part, these rejections have been observed when aggregate, time-series

data alone were used to estimate demand systems. Since such estimation procedure

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

Aggregate consumer behaviour 591

implicitly impose exact aggregation, tests of homogeneity and symmetry could be

sensitive to this for such data sets, yielding apparent rejections of homogeneity and

symmetry.

The fact that homogeneity is not rejected in any of the test situations also

supports the contention that the F, A, and T expenditures are approximately weakly

separable from expenditures on other goods. One would expect a rejection of

homogeneity at least, if other excluded goods' prices were important determinants

of demand.

An interesting additional finding is that it is not essential to pool data to ensure

empirical identification of the parameters of (13). This is a useful result, since it

is not always possible to obtain all the time-series data required to estimate the

pooled model. In this application, for example, some of the observations in the

time-series data set had to be estimated.

The last set of results which will be reported are estimated elasticities. These

are computed based on parameter estimates generated by the pooled results

only. Two types of estimates are presented: homogeneity plus symmetry con-

strained elasticities, without the imposition of exact aggregation, and exact aggrega-

tion, homogeneity and symmetry constrained elasticities. As usual in such models,

estimated elasticities depend on the data. Consequently, the estimates are

evaluated at 1982 Canada average prices, and mean cross-sectional data values

for wn and M.17

Table 8 contains the estimated elasticities.18 It can be seen that F and T are own-

price and income inelastic, relative to A, which is own-price and income elastic.

However, although qualitatively similar, estimates differ depending on restrictions

imposed. For example, the own-price elasticity for F is -0.337, assuming homo-

geneity plus symmetry without exact aggregation, but is -0.518 if exact aggregation

is also imposed.

With respect to cross-price elasticities, it appears that F is complementary to A

and T; while A and T are substitutes. Again, however, the elasticity estimates differ

significantly, depending on whether or not exact aggregation is imposed. Thus, if

estimated elasticities are required for policy purposes, these results suggest that

care should be exercised in the imposition of model restrictions, since estimated

elasticities can be significantly affected.

V. SUMMARY AND CONCLUSIONS

In this paper a model of consumer behaviour was estimated by pooling cross-

sectional and time-series data. This was similar in spirit to the work of JLS, except

17 The 1982 data values are chosen, since this is the mid-year for the cross-sectional data. Estimates

using 1978 and 1984 data values have similar magnitudes and are available from the author on

request.

18 Although the elasticity estimates are of reasonable magnitudes and their signs are not contra-

dictory, both estimates of BJ on which the elasticity estimates are based do not meet the quasi-

convexity requirements for the indirect utility function. Two of the eigenvalues of each of the

estimated BJ matrices are negative.

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

592 Christopher J. Nicol

TABLE 8

Estimated elasticities evaluated at 1982 data values

Homogeneity plus symmetry constrained

Homogeneity, symmetry and exact aggregation

elasticities

constrained elasticities

Elasticities

Equation F A T M

Elasticities

Equation F A T M

F -0.337 -1.186 -0.668 0.978

F -0.518 -0.764 -0.865 0.916

A -0.741 -1.636 1.448 1.879

A -0.281 -1.722 1.055 1.616

T -0.540 1.614 -1.004 0.134

T -0.281 1.139 -0.953 0.658

that here, exact aggregation, homogeneity, and symmetry were not imposed. The

modelling of houwchold demographic effects also differed in this paper from that

of JLS.

Since the cross-sectional data set covered three Family Expenditure Surveys, an

attempt was also made to estimate the unrestricted model using only the cross-

sectional data. This objective was achieved, which is a useful finding. That is,

required time-series data are not always available to facilitate estimation of a pooled

model, particularly for demographic variables, so it is important to be able to rely

on cross-sectional data by itself.

It was also found that demographic variables were important determinants of

consumer behaviour, with both pooled and cross-sectional data. Indeed, the use of

cross-sectional data greatly enhances the degree to which demographic effects can

be modelled.

Hypotheses test of exact aggregation, homogeneity, and homogeneity plus sym-

metry were conducted, using both pooled and cross-sectional data. Exact aggrega-

tion was always strongly rejected, but only one marginal rejection of homogeneity

plus symmetry was observed.

A number of estimated elasticities were presented. These estimates were reason-

able but varied fairly widely when exact aggregation restrictions were imposed.

The results in this paper have some implications for applied researchers in this

area. First, if a 'model of aggregate consumer behaviour' is to be estimated in the

way outlined by JLS, exact aggregation restrictions should not be imposed. Sec-

ondly, in some situations cross-sectional data might contain sufficient information

to identify all parameters of interest, thus precluding the need to pool data. Thirdly,

homogeneity and symmetry restrictions appear to be satisfied by the type of data

employed in this study, especially cross-sectional data, so they may be imposed in

such situations. Finally, household demographic effects should be modelled, since

they seem to be important determinants of consumer behaviour.

A possible extension of the work in this paper would be to apply the estimation

procedure suggested here to other household sizes. Also, the recent work of Jor-

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

Aggregate consumer behaviour 593

genson and Slesnick (1984, 1986, 1987) and Stoker (1986), on inequality, poverty,

household equivalence scales, and welfare employ the JLS model of aggregate

consumer behaviour as a basis for their analysis. It would be interesting to com-

pare the results of such research with those obtained using a model free of exact

aggregation restrictions, such as the one used in this paper.

REFERENCES

Barnes, R., and R. Gillingham (1984) 'Demographic effects in demand analysis: esti-

mation of the quadratic expenditure system.' Review of Economics and Statistics 66,

591-601

Blundell, R. (1988) 'Consumer behaviour: theory and empirical evidence - a survey.'

Economic Journal 98, 16-65

Blundell, R., P. Pashardes, and G. Weber (1988) 'What do we learn about consumer

demand patterns from micro-data?' Institute for Fiscal Studies Working Paper No.

88/10

Browning, M. 'Individual heterogeneity and perfect aggregation: a study of the Cana-

dian microdata, 1969-1982.' Department of Economics Working Paper No. 87-07,

McMaster University

Christensen, L.R., D.W. Jorgenson, and L.J. Lau (1975) 'Transcendental logarithmic utility

functions.' American Economic Review 65, 367-83

Consumer Prices and Price Indexes. Cat. No. 62-010 (Ottawa: Prices Division, Statistics

Canada)

Deaton, A (1986) 'Demand analysis.' In Handbook of Econometrics, ed. Z. Griliches and

M.D. Intriligator, vol. 3, chap. 30 (Amsterdam: North-Holland)

Deaton, A., and J. Muellbauer (1980) 'An almost ideal demand system.' American Eco-

nomic Review 70, 312-26

Income Dirstributions by Size in Canada. Cat. No. 13-207 (annual) (Ottawa: Statistics

Canada)

Income After Tax, Distributions by Size in Canada. Cat. No. 13-210 (annual) (Ottawa:

Statistics Canada)

Income Distributions by Size in Canada. Cat. No. 13-528 (occasional) (Ottawa: Statistics

Canada)

Income Distributions. Incomes of Non-Farm Families and Individuals in Canada Selected

Years 1951-1965. Cat. No. 13-529 (occasional) (Ottawa: Statistics Canada)

Income Distributions by Size in Canada. Cat. No. 13-534 (occasional) (Ottawa: Statistics

Canada)

Income Distributions by Size in Canada. Cat. No. 13-559 (occasional) (Ottawa: Statistics

Canada)

Jorgenson, D.W., L.J. Lau, and T.M. Stoker (1982) 'The transcendental logarithmic mod-

el of aggregate consumer behavior.' In Advances in Econometrics, vol. 1, ed. R.L.

Basmann and G.F. Rhodes, Jr (Greenwich, CT: JAI Press)

Jorgenson, D.W., and D.T. Slesnick (1984) 'Aggregate consumer behavior and the mea-

surement of inequality.' Review of Economic Studies 51, 369--92

- (1986) 'Redistributional policy and the elimination of poverty.' Harvard Institute for

Economic Research Discussion Paper No. 1227

- (1987) 'Aggregate consumer behavior and household equivalence scales.' Journal of

Business and Economic Statistics 5, 219-32

Lewbel, A (1987) 'Fractional demand systems.' Journal of Econometrics 36, 311-37

National Income and Expenditure Accounts, Annual Estimates, 1926-1986. Cat. No.

13-531 (occasional) (Ottawa: Statistics Canada)

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

594 Christopher J. Nicol

Nicol, C.J. (1989) 'Testing a theory of exact aggregation.' Journal of Business and Eco-

nomic Statistics 7, 259-65

(1981) 'The effect of expenditure aggregation on hypothesis tests in consumer demand

systems.' International Economic Review 32, 405-16

Simmons, P., and D. Weiserbs (1979) 'Translog flexible functional forms and associated

demand systems.' American Economic Review 69, 892-901

Stoker, T.M. (1986) 'The distributional welfare effects of rising prices in the United

States: the 1970's experience.' American Economic Review 75, 335-49

Stone, J.R.N. (1953) The Measurement of Consumers' Expenditure and Behaviour in the

United Kingdom, 1920-1938, vol. 1 (Cambridge: Cambridge University Press)

Survey of Family Expenditures Microdata Files (1978, 1982, and 1984) (Ottawa: Family

Expenditure Surveys Section, Statistics Canada)

White, K.J. (1978) 'A general computer program for econometric methods-sHAzAM.'

Econometrica 46, 239-40

This content downloaded from 193.226.34.226 on Mon, 07 Mar 2016 12:14:31 UTC
All use subject to JSTOR Terms and Conditions

Das könnte Ihnen auch gefallen