Weak Separability and the Estimation of Elasticities in Multistage Demand Systems David L. Edgerton American Journal of Agricultural Economics, Vol. 79, No. 1. (Feb., 1997), pp. 62-79. Stable URL: http://links.jstor.org/sici?sici=0002-9092%28199702%2979%3A1%3C62%3AWSATEO%3E2.0.CO%3B2-4 American Journal of Agricultural Economics is currently published by American Agricultural Economics Association.
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/journals/aaea.html. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission.
The JSTOR Archive is a trusted digital repository providing for long-term preservation and access to leading academic journals and scholarly literature from around the world. The Archive is supported by libraries, scholarly societies, publishers, and foundations. It is an initiative of JSTOR, a not-for-profit organization with a mission to help the scholarly community take advantage of advances in technology. For more information regarding JSTOR, please contact
[email protected].
http://www.jstor.org Fri Feb 22 14:52:22 2008
Weak Separability and the Estimation of Elasticities in Multistage Demand Systems David L. Edgerton The necessary conditions for multistage budgeting to be at least approximately justified are reviewed, and the plausibility of these assumptions (i.e., weak separability and low variability of price indices with utility level) are discussed. Using no further assumptions, simple relationships between expenditure and price elasticities in different budgeting levels are derived. These results are applied to a three-stage model for Swedish food consumption, and it is shown that the assumptions are not inappropriate. Restricting the analysis to the last stage of a multistage budgeting process is also shown to lead to considerable errors, which could well have policy consequences. Key w o r d s : elasticities, multistage budgeting, price indices, weak separability.
An economist who wishes to investigate patterns of consumption is always faced with the problems caused by the immense number of commodities and services available to the consumer. An analysis of a complete demand system, consisting of thousands of equations, would require huge quantities of data and computer memory. Within the basis of a time-series analysis, such an exercise is nigh on impossible. The usual way to address this problem is to assume a priori some sort of structure in the consumers' preferences, the most common assumption being that of weak separability. This approach implies that commodities can be partitioned into a number of "separate" groups, where a change in price of a commodity in one group affects the demand for all commodities in another group in the same manner. In addition, the practitioner will often also assume (at least implicitly) a multistage decision process, where expenditure is allocated between groups using The author is associate professor in the Department of Economics at Lund University, Sweden. This article is part of the Nordic Food Demand Study, supported by the Swedish Agricultural and Forestry Research Council. The author wishes to thank Bengt Assarsson, Kyrre Rickertsen, PerHalvor Vale, and the other members of the Nordic F O O ~Demand Study for helpful comments. The suggestions of three anonymous referees have also been of great assistance. Earlier versions of the paper were presented at the American Agricultural Economics Association meeting in Baltimore, 1992, and at the Econometric Society meeting in Uppsala, 1993.
price indices, and where within-group allocation is performed independently. Multistage budgeting is very common in agricultural economics; for example, studies of meat demand nearly always assume the separability of meats from other foods and nonfood.' The budgeting process, however, implies relationships between the price and expenditure elasticities that are calculated at different budget levels. The sensitivity of, for example, beef consumption to changes in income depends both on the reaction of the whole meat group to income changes and on the reaction of beef consumption to changes in meat expenditure. Price elasticities are also affected; a change in the price of beef influences pork consumption directly by redistributing consumption within the meat group, but also indirectly through changes in the price and consumption of the meat group as a whole. The wide use of multistage budgeting raises a number of questions, however. What conditions are necessary for multistage budgeting to be appropriate? Are these conditions plausible? Can these conditions be tested or checked? What implications do these conditions have for the calculation of elasticities?
' See, for example,
the survey of food demand literature in
Edgenon (1992).
Amer. J . Agr. Econ. 79 (February 1997): 62-79
Copyright 1997 American Agricultural Economics Association
Edgerton
In many cases these questions are ignored, because interest is concentrated on the analysis of a specific commodity group that constitutes merely one stage of a multistage budgeting process. This practice, however, is either erroneous or inefficient. It will be erroneous if the assumptions that lie behind the budgeting process are incorrect, and it will be inefficient if the assumptions are correct but their implications are disregarded. In this article we briefly review the assumptions needed for multistage budgeting to be appropriate (at least approximately). The plausibility and testability of these conditions will also be discussed. Our main purpose, however, is to answer the fourth question posed above, and we derive general formulae for the total expenditure and total price elasticities. These results need no further assumptions, and the formulae we obtain are independent of the functional form that is being used.2 Finally, we apply our findings to a three-stage budget process for food demand in Sweden.
Two-Stage Budgeting and Separability Consider the allocation of total consumption between m elementary c ~ m m o d i t i e s .The ~ unconditional (or total) Marshallian demand functions can be written in vector form as
where q represents the m x 1 vector of commodity quantities, p the corresponding vector of nominal prices, and y = q'p the total expenditure. Two-stage budgeting (TSB) implies that the allocation takes place in two independent steps. In the first stage total expenditure is allocated between n broad groups of commodities. This can be formally expressed as
where x is the n x 1 vector of group expenditures and P is an n x 1 vector of group price indices. The second stage consists of the allocation of group expenditures between the elemen-
Correctly identifying the functional form is, of course, necessary in all empirical work. To estimate the elasticities we will need to specify functional fonn, price indices, etc. The adjective "elementary" is used here to emphasize the fact that none of these commodities are aggregates.
Weak Separability and Dentand Estimation
63
tary commodities in each group. The conditional (or within group) Marshallian demand functions are given by
where q, is the m, x 1 subvector of q corresponding to the commodities in the rth group, p, is the equivalent subvector of p, and x , is the rth element of x . The restriction x , = qip, holds in each group, while %, = y and Cm, = m. If TSB is to be appropriate then the conditional and unconditional demand functions must yield the same result; that is, the following relationship must hold for all p and y,
The conditions that are necessary for equation (4) to hold are discussed in, for example, Deaton and Muellbauer (1980a, sect. 5.2). Consistency of the second-stage allocation will follow from an assumption of weak separability, which is quite a rigorous condition, but not completely implausible. Note, however, that we very rarely estimate the true last stage of a multistage process; for example, the good "beef' in a "last-stage" conditional demand system is, in fact, an aggregate of many cuts. Consistency of the first-stage allocation, however, is not implied by weak separability alone.
Two-Stage Budgeting and the Use of Price and Quantity Indices The problems connected with the first-stage allocation (and with all but the last stage of multistage processes) arise from the fact that in general it is not possible to replace the prices of all goods in a group with a single price index. Gorman has shown that this is only exactly possible if either (i) the subutility functions are monotonic transformations of functions that are homogenous of the first degree in q,, or (ii) the utility function is additive between groups and the indirect subutility functions are of the Gorman generalized polar form. The first condition, which is a definition of homothetic preferences for commodities in the same group, implies that all the conditional income elasticities must be unity. The second condition implies strongly separable preferences. Both conditions are thus very restrictive, and must in general be considered implausible. Notwithstanding the above argument, the use
64
February 1997
of price and quantity indices in aggregate demand functions is very widespread. If this practice is to be defended then we must be able to give it at least an approximate justification, and this can in fact be done through the following (albeit rather roundabout) four steps. (i) Let us first consider using the true cost-ofliving index (TCOL) in the allocation process. This is defined as the ratio of group cost functions at two price levels (the level of interest, p,, and base year prices, n,) and at a specific reference subutility level, u,, i.e.,
If the vector of price indices, P, does not vary at all with utility level, then Deaton and Muellbauer (1980a, pp. 131-32) have shown that allocation between groups of commodities (and thus TSB) will be consistent with the unconditional demand analysis. Unfortunately, the TCOL is only completely invariant to utility level when preferences are homothetic, and we are then back to Gorman's first condition; see Deaton and Muellbauer (1980a, p. 173). If, however, P only varies slightly with utility level, then we have the approximate justification we are looking for. (ii) Some limited empirical evidence exists that shows that the TCOL does indeed vary only slightly with utility leveL4 A more common empirical observation, however, is that most price indices are highly collinear. If this is truly the case, then the TCOL can be approximated by either a Paasche or a Laspeyres index, and since these are based on widely different utility levels (base year and current year) we can expect the variation of the TCOL with utility to be quite small. It also allows us to replace the TCOL by one of the above indices in our empirical analyses, which is very useful since they are independent of the functional form being used. In any specific application, the reliability of these approximations can be judged empirically by seeing how much the estimated TCOL varies when different utility levels are used, and by calculating correlation coefficients between different price indices. (iii) There is also a theoretical reason for expecting the Paasche and Laspeyres indices to be
"sing Swedish data, Assarsson (1984) has, for example, shown that the estimated true cost-of-living index differed by less than 1% in absolute value for widely varying utility levels.
Amer. J . Agr. Econ.
collinear. It follows from theorem 1 in Wilks that all price indices of the form Cog, will be highly correlated with each other, if the number of elementary commodities in the group is large.' In practical applications even the most disagregated budget level usually consists of quite highly aggregated commodities, and the number of elementary commodities included in each group thus will be large. (iv) Finally, there is also a theoretical reason for expecting the TCOL to be well approximated by the Paasche and Laspeyres indices. Deaton and Muellbauer (1980a, p. 174) used a Taylor expansion of the cost function around (u,, n,) to show that a first-order approximation to the TCOL will be given by the Laspeyreslike index
where q, is the reference quantity bundle given by the Hicksian demand function at (u,, x , ) . ~ The approximation error is given by 3 ( p r nr)'Sr(pr - IT,)/ IT,, where S, is the Slutsky matrix for the commodities in the rth group, evaluated at some point between p, and n,. The approximation thus will be very accurate if at least one of the following conditions holds: (a) p, - n, is small, (b) p, is nearly proportional to n r [because S,(p, - n,) = k,S,n, = k,S,p, and S:r, = SFp, = 0, SF and S: are the Slutsky matrices at p, and n,], or (c) substitution between commodities is limited. Taking a Taylor expansion around (u,, p,) yields a Paasche index with similar approximation error. To summarize the last two section; TSB will lead to an approximately correct allocation if (i) preferences are weakly separable and (ii) the group price indices being used do not vary too greatly with the utility (or, equivalently, expenditure) level. Both the above assumptions are not completely implausible. Methods of devising reasonable separability structures and testTo be more exact, Wilks shows that the correlation coefficient between two linear combinations of n intercorrelated random variables [where the number of zero correlations is at most of order O ( I ) ] has a distribution with a mean that differs from 1 by a factor of order O(n-') and a variance of order O ( r 2 ) . The theorem was proved under the assumption of unit variances, but this is not essential since a standardization factor can be absorbed into the weights of the linear combinations. In a multistage process P = Z,op,= Z,o,Z,o,,p,, = X,Z, o;,p,,, and it is thus the number of elementary commodities that form the index that ls of interest. If u, 1s taken to be the subutility level at the base year, then q, 1s the observed base year quantity bundle and the right-hand-s~de of equation (6) is the Laspeyres index.
Edgerton
Weak Separability and Demand Estimation
65
ing weak separability can be found, e.g., in Moschini, Moro, and Green and in Pudney. The variability of TCOL with utility level, and collinearity with other price indices, also can be checked after estimation.
can be defined as wi = (priqri)ly,w(,)~= (priqrr)lxr and wo = (PrQr)ly. In a similar manner we can also define the uncompensated total price elasticity for the ith commodity (in the rth group) and jth commodity (in the sth ,group) as
Implications of Two-Stage Budgeting on the Calculation of Elasticities
In fri (12) eij = In Psj
Elasticities are defined as the relative change in consumption of a commodity for an infinitesima1 change in expenditure or price. The total (or unconditional) expenditure elasticity for the ith commodity within the rth commodity group can thus be defined as
The within-group price elasticity between the ith and jth commodities within the rth cornmodity group will be denoted as
(7)
a In a
a a
( 13) e(r)ij=
a In h, a In P ,
fri
Ei = In y
and the uncompensated group price elasticity for the rth and sth commodity groups as
In a similar manner the within-group (or conditional) expenditure elasticity can be denoted as
(8)
a a
In h,, E(,)i= In X,
Defining elasticities for a group of commodities is slightly more complicated, however. The consumption of a group of commodities, in general, can only be defined in terms of quantity indices, because the elementary commodities in the group need not be measured in the same units. In TSB, indices are also used to measure the prices of groups of commodities and, as expenditure equals price times quantity, an obvious quantity index is given by
Defining the function g, = y j P r we can write the aggregate (or between-group) demand function as
where Q and P are the vectors of quantity and price indices. The group expenditure elasticity for the rth commodity group is thus defined as
Budget shares at the different levels likewise
Finally we can define the equivalent compensated price elasticities as
and
Relationships between elasticities in the different stages of a multistage budgeting process have been studied by a number of authors. Bieri and de Janvry use the idea of local group price indices, introduced by Gorman, and develop formulae for estimating total elasticities under both weak and strong separability. These estimates require both time-series and cross-section information, and for weak separability also some prior knowledge (see Bieri and de Janvry, p. 28). They also study the effect of assuming homotheticity to allow estimation of the first stage, but not assuming it when calculating the total elasticities (see also de Janvry, Bieri, and Nufiez for results concerning strong separability). Deaton takes a more pragmatic approach, and assumes a linear expenditure function and Paasche price indices. Formulae for the total
66
February I997
Amer. J . A g r . Econ.
elasticities are found under these assumptions, see Deaton (1975, p. 184). A more recent study by Michalek and Keyzer uses an analogous argument for an LES-almost ideal model, while Heien and Pompelli have proceeded in a similar manner for a model with the linearized almost ideal system in the second stage. Two further papers approach the problem under somewhat different premises. Capps and Havlicek use the specific properties of the Sbranch model while Hassan and Johnson make use of prior information and estimate both complete and incomplete systems. In the following we will merely assume that weak separability holds and that usual price indices are good approximations of ideal price indices. These are exactly the assumptions that are necessary for multistage budgeting to be (approximately) consistent with a complete demand analysis. From this point of departure, we derive formulae for total elasticities that are considerably simpler than those obtained by Bieri and de Janvry, and which can be estimated merely from time-series data. These formulae also reduce to those given by Deaton, by Michalek and Keyzer, and by Heien and Pompelli for their special functional forms.
Expenditure (Income) Elasticities The total expenditure (income) elasticity must obviously be affected by both the within-group expenditure elasticity and the relevant betweengroup expenditure elasticity. Consider, as above, the ith commodity within the rth commodity group. Differentiating equation (4) yields
Using the relation In y, = In g, write
(19)
we are assuming that P, is approximately independent of the level of expenditure, and thus that
Similar formulae can be found in Bieri and de Janvry, in de Janvry, Bieri and Nufiez, and in Manser. The relative approximation error caused by using equation (21) instead of equation (20) is given by
where the second equality follows from the fact that P, can be written as a function of x,, p, and n, only. In spite of the approximation error entailed, there are several reasons for preferring the general formula (21) to an explicit calculation of equation (18) for the functional form being used. Firstly, most published results are given in terms of elasticities of consumption, not elasticities of expenditures, and equation (21) thus allows us to combine results from different sources. Secondly, the practitioner also will usually have a better intuition about the elasticities of consumption than about the elasticities of expenditures (if quantity indices are being used this implies that we feel happier judging changes in real expenditure than judging changes in nominal expenditure). Thirdly, the variation of P, with x, means that the calculation of the last term in equation (18) is not completely standard, as will be exemplified in equations (40)-(43). Finally, the two formulae only will differ to any extent if TSB is inappropriate.
+ In P, we can
a In fri- -.a In h,, a ln a ln xr
-
that is,
The fact that we are using TSB implies that
Price Elasticities A price change in a commodity will cause a direct effect on the quantities purchased within the same commodity group, given an unchanged group expenditure. The price change will, however, also affect the group price index, and thus the allocation of expenditures between groups. This latter effect will influence all commodities, both within and without the same group. Separability does not imply that price changes for commodities in different groups do not affect each other, but merely that such effects are channeled through the group expenditures.
Weak Separability and Demand Estimation
Edgerton
Let us consider two commodities i and j that belong to commodity groups (r) and (s). Using the chain rule on equation (4) yields
while using In y, = In g, rule once again yields
+ In P, and the chain
where 6,, is Kronecker's delta, equal to one when r = s and zero elsewhere. Equation (24b) follows from the fact that only prices in goods belonging to the kth group are to be found in P,, and that In Pda In psj thus will be zero if k # s. Bieri and de Janvry [equations (58) and (59)], de Janvry, Bieri and Nuiiez, and Heien and Pompelli have similar expressions to equation (24a), but do not develop equation (24b). To proceed further we need to evaluate In P j a lnp,, in equation (24b). If P, is defined as the weighted price index ( 6 ) , then differentiating that expression with respect top, will yield
67
term equals the given budget share. Equation (26) is therefore valid for all the price indices we are considering in this paper. Substituting it into equation (24b) and then into equation (23) will now yield the following general result:
If we now choose the current year to be the reference year, which implies using a Paasche index as an approximation to the true cost-ofliving index, then W(s,jbecomes the observed budget share w(,,,. The total uncompensated price elasticities therefore can be written as
because 6,w(,) = 6,w(,),.By rearranging we can obtain the compensated elasticities
a
a
which implies that
where W(,,, is the budget share (in current prices) of the reference quantity bundle. This result is equally applicable for the Laspeyres index (where the reference bundle is defined for the base year) and for the Paasche index (where the reference bundle is defined for the current year). The same result also can be obtained if P, is defined as the true cost-of-living index, since from equation (5) we can see that In Ps = In c(u, ps) - In c(u, 7 ~ ~The ) . derivative of the second term is zero, while it follows from the fact that aclap equals the Hicksian demand function that the derivative of the first
Note that the calculations leading to equations (28)-(30) do not depend directly on the assumption that the price indices are approximately independent of income (although the fact that we are using TSB at all does). How should we interpret these results? Looking at equation (28) we can see that for two commodities in the rth group, the total price elasticity is equal to the within-group price elasticity, plus a factor that is equal to the relative change in the price index [equals w ( , , f~rom equation (26)l times the effect this has on the group expenditure [equals 1 + e(,,,,,]times the within group expenditure elasticity [E,,,i]. If the own between-group price elasticity e,,,,,, = -1, then the group expenditure is unaffected by the price change and eij = e,,,ij.7If, on the other hand, e(,,(,,= 0, then the price change produces a proportional effect on the group expenditure and eij = i?(,,ij. Note also that the difference between the within-group own-price elasticity and the total own price elasticity will be small if the commodity's within-group budget share or expenditure elasticity is small, or if e(,,(,,is close to -1. The group budget shares and expenditure
' This has also been pointed out by Heien and Pompelli for their model.
68
February 1997
elasticities are, however, of no importance. The results of this section can be summarized in the following theorem.
TSB will lead to an approximately THEOREM. correct allocation if ( i ) preferences are weakly separable and (ii) the group price indices being used do not vary too greatly with the utility (ol; equivalently, expenditure) level. Under these conditions the following relationships between elasticities also will hold,
Amer. J . Agr. Econ.
within the meat group, but also of the reallocation effects of meat within food and food within total consumption. With a little notational complexity (but no conceptual difficulty) these results can be extended to any number of stages.
Aggregation Over Households
The definition of demand functions presented in the previous sections is based on microeconomic theory. A common practice, however, is to estimate models using time series of data that have been aggregated over households. In empirical studies of such macro models it is often found that the laws of demand (homogeneity, symmetry, etc.) are rejected, and it is reasonable to ask what happens to our results when they are applied in such situations. Elasticities in Three-Stage Budgeting Keuzenkamp and Barten discuss testing the homogeneity hypothesis in macro models, and Consider now a three-stage, weakly separable come to the conclusion that misspecification of budgeting process, where the first stage is com- the aggregation mechanism is the most usual prised of main commodity groups a = 1, ... , n. reason for rejection. This does not mean that In the second stage the ath group comprises of aggregate models are useless, however. If the subcommodity groups r = 1, ..., na, and, in the macro model is itself well s p e ~ i f i e d then , ~ any third stage, the rth subgroup within the ath elasticities derived from it will be valid as long main group comprises of commodities i = 1, ... , as they are interpreted in terms of relative mar.We obviously can generalize the results of changes in aggregate consumption. These elasthe previous section by first considering stages ticities will not have an obvious interpretation one and two as a two-stage budgeting process, at the individual or household level, however, and then repeating this for stages two and three. unless the aggregation mechanism has been If commodity i is within subcommodity correctly specified. group ( r ) within main-commodity group [ a ] , If the macro between-group and within-group and commodity j is within subcommodity group allocation functions are well specified in equa( s ) within main-commodity group [ b ] , then re- tions (2) and (3), then a macro unconditional peated substitution into equations (21) and (30) demand function will be defined by equation leads to (4). If, furthermore, price indices of the form in equation (6) do not vary too greatly at different expenditure levels, then all our earlier results will all carry through. It will therefore be quite proper to use equations (21) and (28) in an and econometrically well-specified macro model when calculating total expenditure and price elasticities, even if the model violates the laws of demand, as long as these elasticities are interpreted in aggregate terms and not in terms of a representative consumer. The income elasticity of beef will thus be a product of the income elasticity of food, the food expenditure elasticity of meat, and the meat expenditure elasticity of beef. The price elasticity will be the result of a direct effect
One problem with the use of macro models is that they may be unstable with regard to changes in the distribution of consumption over households. Tests for parameter stability, and other econometric misspecification tests, are therefore an essential part of the evaluation of any macro model. The concept "well specified" should be interpreted in this light.
Weak Separability and Demand Estimation
Edgerton
69
PRIVATE CONSUMPTION IN
NONDURABLES
FOOD-AT-HOME
FOOD-NOT-ATHOME
GOODS Excluding Food
SERVICES
ExclIksmrants
MIS(=ELLANEOUS
r
ANIMALLA
BEVERAGES
VEGETABLIA
h4EK, CHEESE AND EGGS
ALCOHOLIC BEVERAGES
POTATOES
Figure 1. Utility tree
The Nordic Food Demand Study is a project aimed at estimating and comparing the demand for fairly disaggregated food groups in Denmark, Finland, Norway, and S ~ e d e n This .~ study has utilized a three-stage budgeting model, with preference structure assumed as given in figure 1. In the first stage, total private consumption is divided into four commodity groups, one of which is food-at-home.I0 In the second stage, food-at-home is divided into four food commodity groups, and finally, in the third stage, each food group is divided into three more
disaggregated food commodities. The first two stages are thus four-equation systems, while the last stage consists of three-equation systems. The model has been estimated for Swedish data using various forms of Deaton and Muellbauer's almost ideal demand system. Preliminary studies showed that static models were poorly specified regarding such aspects as autocorrelation, heteroskedasticity, omitted variables, etc. Various forms of habit-forming mechanisms, therefore, have been tried, and the model estimated in this report is a dynamic form of the linearized almost ideal system (LAIDS)." For a complete system this is given by
A full account of the project is given in Edgerton et al., where a detailed sensitivity analysis for Swedish food demand also can be found. 'O We have excluded durables and semidurables from private consumption to avoid having to construct intertemporal models. The model used accounts for about 70% of domestic private consumption (or disposable income).
" An alternative dynamic specification, suggested by Ray, is to set tl,, = tl and 0 , = 0 if i # j. This specification was rejected for our data using likelihood ratio tests. Our results also have revealed only small empirical differences between the LAIDS model, using Stone's price index, and the nonlinear AIDS model using the full price index-see also Edgerton et al. This supports our argument from earlier in the paper that the choice of price index is not crucial to estimation.
Demand for Food in Sweden-Model
70
February 1997
Amer. J . Agr. Econ. n
(33) wir = ai +
x 8,jw,,,-l j=1
rI
+
C y,
j=l
In P, + P,(ln x, - In
e*>
for i = 1, . . ., n, where In P,' = XkwkIn p,,. Some restrictions must be imposed to enable identification of the parameters of the lagged budget shares, and we have chosen Cjeij = O.I2 The restrictions Xipi = Xi?,. = XieiJ= 0 and Xia, = 1 follow from the adding-up condition, while homogeneity and symmetry can be enforced by imposing the restrictions Xjyij = 0 and y, = y,,. We have not imposed homogeneity and symmetry in this study, because these will only apply to the macro-model if individual preferences are of the form envisioned by Deaton and Muellbauer. This is a point we wish to test. The dynamic part of the model is a simplified form of the interrelated partial adjustment model proposed by Anderson and Blundell, which has been used by Alessie and Kapteyn and by Assarsson (1991). The short-run elasticity formulae used in this paper are adapted from the approximations suggested by Goddard and by Chalfant (see also Buse), namely,
The dynamic LAIDS mode1 (33) is an approximate version of the nonlinear dynamic almost ideal system, which has a cost function given by (37) In c(u, pt) = 4 p t ) + M p t ) where
and
This is an adaptation of the usual AIDS cost function, where the third term in equation (38) is the cost of habit forming. The TCOL is thus given by
while solving equation (37) for u yields
Substituting equation (41) into equation (40) and differentiating with respect to In y leads to and
where for the sake of simplicity we have dropped the subscript t on the variables and elasticities. Equations (33)-(36) have been applied separately to each of the six subsystems for private consumption {stage I ] , food-at-home { 2 ) , animalia ( 3 1, beverages { 3 ), vegetablia { 31, and miscellaneous ( 3 ) . This yields, with some notational adjustment, between and within group elasticities, and equations (31) and (32) then can be used to obtain the total elasticities. " Other alternat~vesare a , = 0 or €I, =, 0. This is merely a choice of normalization, which does not affect the other results.
This formula applies equally well to all stages of the budgeting process. The approximation error (22), which uses relations for group TCOLs, thus can be written as
where the subscript (r) indicates, as usual, within-group variables and parameters. Note that if p, is proportional to n, then the addingup condition Xpk = 0 implies that A, = 0. If our only approximation was to assume that
Edgerton
the TCOL was independent of expenditure level, then we could use equation (43) in equation (20) to obtain an exact result. However, we are also making a number of other approximations, namely, we are (i) using the Paasche index instead of the TCOL when allocating between groups and (ii) using Stone's index instead of the true AIDS price index when estimating the system. These two approximations mean that we still need to look at the correlation between the TCOL, Paasche index, and Stone's index when assessing the validity of our approach. We have followed the usual practice of normalizing base year prices to x: = 1, and in this case the group TCOLs are given by
where a: (p) = a,(p) - a, and b: (p) = p;Ab, (p) - 1 are independent of a, and p,. From equation (37) we see that In c,(u, l ) = a, + up,, and a, thus measures the logarithm of subsistence real group expenditure at base year prices, while a, + P, measures the logarithm of bliss real group expenditure. If we denote group real expenditure by x,(u), then it follows that
Equation (45) allows us to estimate the group TCOLs for different utility levels.
Demand for Food in Sweden-Estimation The dynamic LAIDS model has been estimated using annual data for aggregate private consumption, obtained from the Swedish national accounts for the period 1963-90.13 Because these data are aggregated over households, the expenditures used in equation (33) were defined per capita. Price data have been obtained from the implicit price indices formed by dividing current expenditures by real expenditures
l 3 Statistics Sweden publish yearly data for private consumption, disaggregated to the level given in figure 1, formed by adjusting production data for home production, industrial consumption, public consumption, balance of trade, etc. Some adjustments have to be made, however, for changes of base year, and these adjusted series can be obtained from the author. Major revisions in the organization of the national accounts have taken place in 1963 and 1991, which is the reason for choice of estimation period. Quarterly data series exist since 1980, but at this level of disaggregation they are merely extrapolations. More highly disaggregated series are also available, but then only for total consumption (that is, prlvate, public and food-not-at-home).
Weak Separability and Demand Estimation
71
(i.e., Paasche indices are used in the allocation). The composition of the commodity groups used in this study is fairly self-evident. A few points of clarification might be needed, however. The group "Coffee and Tea" also contains other hot beverages, such as cocoa and drinking chocolate, while "Confectionery, etc." includes salt, spices, and sauces (although confectionery is by far the dominating factor here). Swedish national account data measure consumption as it reaches the consumer, and the divisions into commodities have followed this reasoning. The only real exception concerns ready made nonfrozen dishes, where estimates of the meat and vegetable parts are treated separately. The separability assumptions in figure 1 can, of course, be questioned. The intuitive reasoning behind the divisions is that the consumer, when planning food purchases, will first decide how much to spend on the main meals (animalia and vegetablia), how many beverages to consume and what other foods to buy. In a second stage, the decisions will be more finely tuned, e.g., whether meat or fish will be the main course, whether to drink coffee or wine, etc. If data of adequate quality were available, then the utility tree naturally could have been extended to more stages. One direct criticism of the assumptions used here is the inclusion of milk in the animalia group instead of the beverages group. This is only due to the fact that the Swedish national accounts do not permit the separation of milk data from those for cheese and eggs. A similar problem is the inclusion of fruit juice in the fruit and vegetables group, although this is certainly of less importance in Sweden. The results of various misspecification tests can shed some light on the adequacy of separability assumptions, and our findings are reported later in the article. The six subsystems that constitute the utility tree in figure 1 consist, in all, of twenty equations. Because one equation in each subsystem can be eliminated using the adding-up condition, we have a total of fourteen equations to be estimated. Two estimation methods are reported in this paper, ordinary least squares (OLS) and seemingly unrelated regressions (SUR). If we assume that the error terms in the different subsystems are blockwise independent of each other, then OLS will be asymptotically efficient. This follows from the fact that SUR now reduces to six separate system estimations, each with identical regressors and with no cross equation restrictions, and thus will be numerically equivalent to OLS. If we do not assume blockwise error independence, however, then
72
February 1997
using SUR on the fourteen-equation system leads to different estimates. An alternative to SUR could be to use the method of maximum likelihood (ML), found by iterating the SUR algorithm. This will obviously be equal to OLS if we assume blockwise error independence. Any attempt to use ML on the fourteen-equation system will break down, however, due to singularity in the estimated Hessian matrix. This follows, since the number of linearly independent variables in the estimated LAIDS system (41) is greater than the number of observations (27). The equivalent restriction for OLSISUR is merely that the number of linearly independent variables in each equation must be less than the number of observations.14 LaFrance has shown that multistage budgeting leads to endogenous group expenditures, and that OLS thus will be inconsistent. LaFrance's results also indicate that usual instrumental variable (IV) estimation is inconsistent if the group expenditures enter the system nonlinearly, but Edgerton (1993) showed that it can be consistent in some nonlinear systems (e.g., AIDSILAIDS). We have therefore also calculated IV estimates. These are not reported here, however, due to the results of the misspecification tests presented below, and because they in general do not differ radically from the other estimates (see Edgerton et al., chap. 6). Because the elasticities are nonlinear functions of the parameters, some form of approximation is necessary to obtain their standard errors. We have used the standard method of linearizing the function about the estimated parameter values, a technique that is easily performed by the Analyz procedure in the TSP 4.11 4.2 computer program. To calculate the OLS standard errors of the total elasticities we utilize the assumption that the error covariances between the different stages are zero. Note that both the elasticities and their standard errors are functions of the variables, and thus vary over time.
Results of Specification and Misspecification Tests
The misspecification tests reported in table l a have been calculated using Godfrey's theory of l 4 It might seem strange that SUR estimates can be obtained, but not ML estimates. Upon iteration, however, the parameter estimates converge toward values that yield a singular Hessian matrix, and thus infinite standard errors. This is a degrees-of-freedom effect similar to that which causes FIML to break down in some simultaneous equation models, even though 3SLS will work.
Amer. J . Agr. Econ.
"local equivalent alternatives" (LEA) applied to system estimation. We thus have considered regressions of the least squares residuals on linear functions of the lagged residuals (BreuschGodfrey, autocorrelation), squared predictions (RESET, functional misspecification) and instruments15 (Hausman-Wu, endogeneity). The squared residuals have been regressed on the predictions (Breusch-Pagan, heteroskedasticity) and on lagged squared residuals (ARCH). We have obtained "exact" small sample likelihood ratio tests (ELRT), using the results due to Rao (p. 556). Note the ELRT is truly exact for testing homogeneity, but will be approximate for symmetry and misspecification tests. See also McGuirk et al. and Edgerton et al. (chap. 4). Two tests, that have been applied in a singleequation environment, are given in table l b . The Jarque-Bera test of non-normality has been used (with Kendall and Stuart's small sample approximation, pp. 281-97). Finally parameter instability has been tested using the CUSUMSQ test applied to OLS residuals (see McCabe and Harrison for applicability and Edgerton and Wells for a method of calculating the P-values). The results of these tests indicate a considerable degree of misspecification in the first stage, with low P-values for autocorrelation, functional misspecification, endogeneity, nonnormality and parameter instability! The results of a sensitivity analysis in Edgerton et al. showed, however, that the food-at-home elasticities are quite stable with regard to changes in model formulation. Though not reported here, the food-at-home equation is also the least misspecified of this subsystem. Stages two and three seem to be quite well specified. Note that we would expect any errors in the choice of separability assumptions to show up in the RESET and endogeneity tests, which does not happen here. The goodness-of-fit measures for the dynamic model are given in table 2. l 6 As expected
l 5 The instruments considered have been total expenditure (whlch should reveal problems of the type discussed by LaFrance in stages two and three) and disposable Income (which will test the exogeneity of total expenditure, see for example Deaton 1986, pp. 1779-80). We have not tested for the exogeneity of prices. Note that since total expenditure is included as a regressor in the first stage, it cannot be used there as an instrument for testing exogeneity. l6 The systemwise goodness-of-fit measures are given by
= I - !/det 2,/det
e,
m being the number of equations in the nonsingular system. The adjustment for degrees of freedom is the usual, while RX, =
(R; - Ri)/(l - R:) measures the relative improvement of the dynamic (D) model over the static (S) model. In table 2, R;, is adjusted for degrees of freedom.
Edgerton
Weak Separability and Demand Estimation
73
Table la. P-Values of System Misspecification Tests (OLS Estimation)
PC FAH A B
V M
BG
BP
ARCH
RESET
HWU-DI
HWU-TX
0.013 0.073 0.154 0.698 0.362 0.496
0.687 0.580 0.109 0.219 0.583 0.618
0.133 0.748 0.956 0.778 0.682 0.720
0.017 0.088 0.650 0.074 0.625 0.335
0.001 0.922 0.565 0.553 0.353 0.985
0.361 0.359 0.677 0.234 0.864
Note: The following abbreviations have been used: Private Consumption (PC), Food-at-home (FAH), Animalia (A), Beverages (B), Vegetablia (V), Miscellaneous (M), Breusch-Godfrey (BG), Breusch-Pagan (BP) and Hausman-Wu (HWU). The instruments in the HWU tests are Disposable Income (DI) or Total expenditure (TX).
Table lb. P-Values of Single Equation Misspecification Tests (OLS Estimation) -
Jarque-Bera
PC FAH A
B V M
--
CUSUMSQ
1
2
3
4
1
2
3
4
0.670 0.898 0.810 0.946 0.915 0.939
0.050 0.927 0.796 0.772 0.675 0.544
0.022 0.619 0.553 0.970 0.971 0.874
0.416 0.783
0.178 0.079 0.407 0.457 0.204 0.360
0.000 0.051 0.707 0.777 0.170 0.138
0.039 0.521 0.363 0.187 0.726 0.325
0.254 0.210
Note: The same abbreviations have been used as in table l a . The headings "1" to "4" indicate which equation is b e ~ n gtested.
Table 2. Goodness-of-Fit Measures and Significance Tests (OLS Estimation) Systemwise R2
PC FAH A
B V M
Unadjusted
Df-adj
0.961 0.880 0.853 0.945 0.914 0.950
0.944 0.826 0.809 0.928 0.889 0.935
P-Values, ELRT D:S 0.250 0.320 -0.047 0.375 -0.025 0.531
H
HS
Static
0.001 0.137 0.008 0.060 0.001 0.829
0.001 0.030 0.018 0.004 0.002 0.232
0.006 0.001 0.748 0.000 0.461 0.000
Note: The same abbreviations have been used as in table l a plus the following: Exact Likelihood Ratio Test (ELRT), Homogeneity (H), Symmetry ( S ) , Degrees of Freedom Adjusted (Df-adj) and improvement of Dynamic over Static (D:S). See footnote 16 for formulae.
they are all close to one. From the D:S column we can see that the dynamic model yields a 25%-50% reduction in error variability over the static model for all subsystems except for animalia and vegetablia.17 The hypothesis of a static model also is rejected for the same sub-
'' A comparison with a pure vector autoregressive model gave even higher values, mostly being over 75%.
systems. Together with the results of the misspecification tests for the static model (reported verbally earlier in the text), this seems to supply firm evidence in support of a dynamic specification. It should be noted, however, that the long-run properties of the system revealed a considerable degree of instability, which would indicate that price and income changes take a long time to "work through" the system. Homogeneity and symmetry are, in general,
74
February1997
Amer. J . Agr. Ecorl.
Table 3. Adequacy of Price Index Approximations ALL
TCOL
Stage 1 2 3
PITCOL
LAIDS
min R2 (0.9990) 0.9999 0.9963
(1.OOOO) 1.OOOO 0.9998
0.9994 1.OOOO 0.9998
0.9992 0.9999 0.9986
maxlAl
maxl6,I
(0.0479) 0.0064 0.0545
(0.0913) 0.0070 0.0736
Stage
rejected for all stages of the model (and for the complete system). As previously noted, this is a very common result in macro-models and should not be taken as an indication of a failure of the laws of demand, since the Slutsky conditions are derived at the micro-level and are only invariant to aggregation under very special assumptions.18 The elasticities calculated from the model therefore must be interpreted as aggregate elasticities and not micro-elasticities coming from a "representative consumer." Aggregate elasticities are, on the other hand, exactly those required by policy makers when making decisions concerning, for example, alcohol taxation. Checking the Assumptions Concerning Price Indices We have made three assumptions concerning price indices. First, the use of TSB implies that the TCOL may not vary too greatly with utility level. Second, we have assumed that the Paasche index does not differ too greatly from the TCOL, and third we have assumed that Stone's index does not differ too greatly from the nonlinear AIDS index. These assumptions have been checked in the following manner. We have let x,(O) = +x~,,~,approximate group real subsistence expenditure at each stage and x,(l) = 2xr,,,, approximate group bliss expenditure. These levels have been used together with x,(u) = x,,~,, x,, x,,,, in equation (45) to give five values of the TCOL (denoted TCOL, to TCOL,), where TCOL, equals the nonlinear AIDS price index minus a,. '"he result is not due, however, to the statistical reasons polnted out by Laitinen and others, since we are using the ELRT which corrects for those discrepencies.
The logarithms of the five TCOLs, Stone's index, and the Paasche index have been regressed on each other, and the coefficient of determination (R2),the intercept (a) and slope (b) have been calculated for each pair. Using equation (43) we have also calculated A, the relative error in expenditure elasticities caused by variability of the TCOL, and from this we have obtained the equivalent absolute error (6,). Even if R2 from the regression of Stone's index on TCOL, is almost one, the use of LAIDS will cause an error in the estimation of the p's if the slope b differs from one. We therefore have also calculated a,, the absolute error in the expenditure elasticity caused by using LAIDS instead of AIDS. In table 3 we report the minimum value of R2 and the maximum values of Ib - 11 over different pairs of indices and commodities at each stage. The column "ALL" denotes the comparison of all indices, "TCOL" denotes comparisons of the true cost-of-living indices for observed utility levels (TCOL,, TCOL,, and TCOL,), "PITCOL" denotes regressions of the Paasche index on TCOL,, TCOL,, and TCOL,, while "LAIDS" denotes the regression of Stone's index on TCOL,. The errors A, 6,, and defined above are also reported. Note that in the first stage we do not need TCOL to be independent of utility level, and thus we include such figures in brackets. The first impression from table 3 is that price indices are indeed very collinear, since the lowest value of R2 for any index pair is 0.9963. This is not a sufficient condition for the parameter estimates to be correct, however, since the slopes must also be equal to one for this to hold. In the first stage, the slopes can vary up to 10% from unity for some index pairs, though the variation of the TCOL is less than 1% for
Edgerton
Weak Separability and Demand Estimation
75
Table 4. Expenditure Elasticities (OLS) Elasticities
Meat Fish Dairy Soft drinks Coffee Alcohol Cereals Fruitlveg Potatoes Fats Sugar Confect.
Standard Errors
Within Group
Within Food
Total
Total Group
Within Group
Within Food
Total
Total Group
1.36 0.34 0.75 0.65 0.38 1.21 1.43 0.74 -0.19 1.45 -0.55 0.99
1.45 0.36 0.80 0.72 0.42 1.33 1.38 0.71 -0.18 1.04 -0.39 0.71
0.98 0.25 0.54 0.49 0.29 0.90 0.93 0.48 -0.12 0.70 -0.26 0.48
-0.38 -0.09 -0.21 -0.16 -0.09 -0.31 -0.50 -0.26 0.07 -0.75 0.29 -0.51
0.15 0.27 0.18 0.24 0.17 0.04 0.20 0.19 0.30 0.21 0.27 0.10
0.27 0.30 0.23 0.33 0.22 0.38 0.35 0.24 0.30 0.43 0.25 0.29
0.22 0.20 0.17 0.28 0.15 0.28 0.26 0.17 0.20 0.30 0.17 0.20
0.19 0.09 0.12 0.16 0.10 0.28 0.24 0.13 0.11 0.30 0.18 0.20
observed utility levels. The error in total expenditure elasticity, caused by the fluctuation of the TCOL, is at most 5% in relative terms, or 0.1 in absolute terms (see A and 6,). The assumption of low variation in the TCOL therefore seems to be justified. The replacement of the TCOL by a Paasche index seems also to be justified (the slopes being less than 2% from unity) for the second and third stages, but somewhat more questionable in the first stage. The same is true for the use of Stone's index, though the slope discrepancy is now slightly over 3% and the absolute elasticity error (6,) is less than 0.05 in the third stage.lg We have not reported the values of the intercepts, since these only affect the estimate of the constant term. The maximum absolute value of a is less than 0.02 at all stages, however. Note also that when R2 is almost one, then the TCOL and LAIDS approximations mainly affect the estimation of p and expenditure elasticities, while the Paasche approximation mainly affects y and price elasticities. We therefore have also looked at the combined effects of the TCOL and LAIDS approximations on the expenditure elasticity, and have found the maximum discrepancy for the total expenditure elasticity to be 0.09 (in the alcohol equation). Summarizing, we can see that the assumptions made concerning price indices seem in
I9It should be noted that the discrepancy is only large for the vegetablia subsystem. For the other subsystems at this stage the b's are at most 1% from unity and the absolute elasticity error less than 0.01. Edgerton et al. compared elasticities directly for AIDS and LAIDS systems and found them to be very close.
general to be justified. The misspecification of the first stage, noted in the previous section, seems also to carry over here, however.
Results of the Elasticity Estimation
In tables 4-9 we report the values of the expenditure and own-price elasticities, and their standard errors, estimated using mean budget shares. For each elasticity, the within-group, within-food and total elasticities are given, where "within-food" implies using the TSB formulae only for stages two and three. The column "total-group" denotes the difference observed when using the correct total elasticity formulae instead of the within-group elasticities. The model seems to work quite well, to the extent of giving reasonable elasticities, with the exception of the sugar equation (the dairy and potato equations have slightly negative income elasticities and/or positive own-price elasticities, but these lie within the bounds of reasonable stochastic error). Nearly all commodities are necessities, although alcohol, meat, and cereals lie on the boundary of being luxuries, which is not the result obtained if one only looked at the within-group elasticities. Fish, potatoes, and sugar have expenditure elasticities that are within two standard errors from zero. As regards total uncompensated price elasticities, we can see that meat, fish, dairy products, coffeeltea, and potatoes are quite price inelastic. Note that the sugar equation yields a significantly positive own-price elasticity (both compensated and uncompensated). The result
76
February 1997
Amer. J . Agr. Econ.
Table 5. Uncompensated Own-Price Elasticities (OLS) -
-
- - -
-
Elasticities
Meat Fish Dairy Soft drinks Coffee Alcohol Cereals Fruitfveg Potatoes Fats Sugar Confect.
Standard Errors
Within Group
Within Food
Total
Total Group
Within Group
Within Food
Total
Total Group
-0.52 -0.24 -0.04 -0.67 -0.06 -0.96 -1.04 -0.71 0.12 -0.5 1 0.66 -0.70
-0.34 -0.23 0.03 -0.66 -0.06 -0.94 -0.86 -0.63 0.12 -0.34 0.65 -0.44
-0.22 -0.22 0.08 -0.65 -0.05 -0.83 -0.78 -0.59 0.11 -0.32 0.64 -0.41
0.30 0.02 0.12 0.02 0.01 0.13 0.26 0.12 -0.01 0.19 -0.02 0.29
0.10 0.18 0.15 0.20 0.06 0.02 0.16 0.07 0.06 0.13 0.18 0.09
0.12 0.17 0.15 0.20 0.06 0.12 0.16 0.07 0.06 0.13 0.17 0.11
0.12 0.16 0.14 0.20 0.06 0.11 0.15 0.07 0.06 0.13 0.17 0.10
0.06 0.02 0.04 0.01 0.01 0.11 0.09 0.05 0.01 0.05 0.01 0.07
Table 6. Compensated Own-Price Elasticities (OLS) -
-
--
-
-
Elasticities Within Group
Within Food
Total
-
-
-
-
-
-
Standard Errors Total Group
Within Group
Within Food
Total
Total Group
Meat Fish Dairy Soft drinks Coffee Alcohol Cereals Fruitfveg Potatoes Fats Sugar Confect.
Table 7. Expenditure Elasticities (SUR) Elasticities Within Group Meat Fish Dairy Soft drinks Coffee Alcohol Cereals Fruitlveg Potatoes Fats Sugar Confect.
Within Food
Total
Standard Errors Total Group
Within Group
Within Food
Total
Total Group
Edgerton
Weak Separability and Demand Estimation
77
Table 8. Uncompensated Own-Price Elasticities (SUR) Elasticities Within Group
Within Food
Total
Standard Errors Total Group
Within Group
Within Food
Total
Total Group
Meat Fish Dairy Soft drinks Coffee Alcohol Cereals Fruitlveg Potatoes Fats Sugar Confect.
Table 9. Compensated Own-Price Elasticities (SUR) Elasticities Within Group
Within Food
Total
Standard Errors Total Group
Within Group
Within Food
Total
Total Group
Meat Fish Dairy Soft drinks Coffee Alcohol Cereals Fruitlveg Potatoes Fats Sugar Confect.
for this compensated elasticity once more underlines the necessity of interpreting our results as merely applying to the macro-model. The differences between total and withingroup elasticities are quite large in some cases. The absolute values of the within-group price elasticities are in many cases considerably larger than for the total elasticities, which can be misleading if the former are used for making policy judgments. For example, meat has an own-price elasticity of -0.47 when animalia and food consumption are held constant, but it is merely -0.11 when changes in these consumptions are taken into account (using the SUR estimates). Similar results also apply for alcohol, cereals, fats and confectionery. The ef-
fects of, for example, taxation changes thus should be studied using total elasticities. Note also that differences between the compensated total and within-group elasticities also can be substantial. More formally, we have calculated the asymptotic standard errors of the differences between the total and within-group elasticities. It can be seen from tables 4-9 that the absolute values of the differences are greater for expenditure elasticities, but that the larger value of their standard errors implies fewer significant differences. We also have calculated the Wald statistic (corrected for degrees of freedom) for the null hypothesis that total and within-group elasticities are equal for all commodities. From
78
February 1997
Amer. J. Agr. Econ.
Table 10. Test of Equality Between Total- and Within-Group Elasticities OLS
Expenditure Own price Compensated own price
SUR
Wald Statistic
P-Value
Wald Statistic
P-Value
15.4 41.6 69.1
0.223 0.000 0.000
20.5 68.5 100.0
0.058 0.000 0.000
table 10 we can see that this hypothesis is soundly rejected for the price elasticities, but not necessarily for the expenditure elasticities.
Conclusion In this paper, simple relationships between expenditure and price elasticities in different levels of multistage budgeting have been established. We have assumed that preferences are weakly separable and that the usual price indices are good approximations to true cost-of-living indices, which are precisely the assumptions needed to justify estimation of the model. The relation for the expenditure elasticity is well known, but rarely employed, while, as far as the author knows, the relation for price elasticities is new. The above methods have been applied to a three-stage budget model for Swedish food data, estimated using a dynamic version of the LAIDS system. The results of this study show that our assumptions regarding separability and price indices are not inappropriate. The strategy of restricting one's analysis to the last stage of a multistage budgeting process was shown to lead to considerable errors, which could well have policy consequences. The results also establish the necessity of using the same budget level when comparing results from different studies. [Received September 1995; final revision received October 1996.1
References Alessie, R., and A. Kapteyn. "Habit Forming and Interdependent Preferences in the Almost Ideal Demand System." Econ. J. 101(May 1991):404-19 Anderson, G., and R. Blundell. "Testing Restrictions in a Flexible Dynamic Demand System: An Application to Consumers' Expenditure in Canada." Rev. Econ. Stud. 50(July 1983):397-410.
Assarsson, B. "Prices and Income Distribution." Mimeo., Dept. Econ., Lund University, Sweden, 1984. -. "Alcohol Pricing Policy and the Demand for Alcohol in Sweden 1978-88." Nordic Food Demand Study SWl9, Dept. Econ., Uppsala University, Sweden, 199 1. Bieri, J., and A. de Janvry. "Empirical Analysis of Demand Under Consumer Budgeting." Giannini Foundation Monograph, No. 30, 1972. Buse, A. "Evaluating the Linearized Almost Ideal Demand System." Amer. J. Agr. Econ. 76(November 1994):781-93. Capps, O., and J. Havlicek. "National and Regional Household Demand for Meat, Poultry and Seafood: A Complete Systems Approach." Can. J. Agr. Econ. 32(March 1984):93-108. Chalfant, J.A. "A Globally Flexible, Almost Ideal Demand System." J. Bus. and Econ. Statist. S(Apri1 1987):233-42. de Janvry, A,, J. Bieri, and A. Nuiiez. "Estimation of Demand Under Consumer Budgeting: An Application to Argentina." Amer. J. Agr. Econ. 54(December 1972):422-30. Deaton, A. Models and Projections of Demand in Post-War Britain. London: Chapman and Hall, 1975. -. "Demand Analysis." Handbook of Econometrics, vol. 3, chap. 30. Z. Griliches, and M.D. Intriligator, eds. Amsterdam: North-Holland, 1986. Deaton, A., and J. Muellbauer. Economics and Consumer Behaviour. Cambridge: Cambridge University Press, 1980a. -. "An Almost Ideal Demand System." Amer. Econ. Rev. 70(June 1980b):312-26 Edgerton, D.L. "Estimating Elasticities in MultiStage Demand Systems." Working Paper, presented at the American Agricultural Economics Association meeting in Baltimore, 1992. -. "On the Estimation of Separable Demand Models." J. Agr. and Res. Econ. 18(December 1993):141-46. Edgerton, D.L., B. Assarsson, A. Hummelmose, I. P. Laurila, K . Rickertsen, and P. H . Vale. The Econometrics of Demand Systems. With Appli-
Weak Separability and Demand Estimation cation to Food Demand in the Nordic Countries, Dordrecht: Kluwer Academic Press, 1996. Edgerton, D.L., and C. Wells. "Critical Values for the Cusumsq Statistic in Medium and Large Sized Samples." Oxford Bull. Econ. and Statist. 56(August 1994):355-65. Goddard, D. "An Analysis of Canadian Aggregate Demand for Food at Home and Away from Home." Can. J. Agr. Econ. 31(November 1983):289-318. Godfrey, L.G. Misspecification Tests in Econometrics. Cambridge: Cambridge University Press, 1988. Gorman, W.M. "Separable Utility and Aggregation." Econometrica 27(July 1959):469-81. Hassan, Z.A., and S.R. Johnson. "Consistency of Elasticities from Multiple Levels of Budgeting Process." Can. J. Agr. Econ. 34(July 1986):257-62. Heien, D.M., and G. Pompelli. "The Demand for Beef Products: Cross-Section Estimation of Demographic and Economic Effects." W. J. Agr. Econ. 13(July 1988):37-44. Kendall, M.G., and A. Stuart. The Advanced Theory of Statistics, Volume I , 2nd ed. London: Charles Griffin and Company, 1963. Keuzenkamp, H.A., and A.P. Barten. "Rejection without Falsification: On the History of Testing the Homogeneity Condition in the Theory of Consumer Demand." J. Econometrics 67(May 1995): 103-28. LaFrance, J.T. "When Is Expenditure "Exogenous" in Separable Demand Models?'W J. Agr. Econ. 16(July 1991):49-62. Laitinen, K. "Why is Demand Homogeneity So Often
79
Rejected?" Econ. Letters 1 (July 1978): 187-91. Manser, M.E. "Elasticities of Demand for Food: An Analysis Using Non-Additive Utility Functions Allowing for Habit Formation." S. Econ. J. 43 (July 1976):879-91. McCabe, B.P.M., and M.J. Harrison. "Testing the Constancy of Regression Residuals over Time Using Least Squares Residuals." Appl. Statist. (J. Royal Statist. Soc., Series C) 29(1980):14248. McGuirk, A.M., P. Driscoll, J. Alwang, and H.L. Huang. "System Misspecification Testing and Structural Change in the Demand for Meats." J. Agr. and Resour. Econ. 20(July 1995): 1-21. Michalek, J., and M.A. Keyzer. "Estimation of a Two-Stage LES-AIDS Consumer Demand System for Eight EC Countries." Eur. Rev. Agr. Econ. 19(1992):137-63. Moschini G., D. Moro, and R.D. Green. "Maintaining and Testing Separability in Demand Systems." Amer. J. Agr. Econ. 76(February 1994):61-73. Pudney, S.E. "An Empirical Method of Approximating the Separable Structure of Consumer Preferences." Rev. Econ. Stud. 48(0ctober 1981):56177. Rao, C.R. Linear Statistical Inference and Its Applications, 2nd ed. New York: Wiley, 1973. Ray, R. "A Dynamic Generalisation of the Almost Ideal Demand System." Econ. Letters 14(March 1984):235-39. Wilks, S.S. "Weighting Systems for Linear Functions of Correlated Variables When There Is No Dependent Variable." Psychometrika 3(March 1938):23-40.
