Estimating Elasticities with the Linear Approximate Almost Ideal Demand System: Some

Monte Carlo Results

Author(s): Julian M. Alston, Kenneth A. Foster, Richard D. Gree
Source: The Review of Economics and Statistics, Vol. 76, No. 2 (May, 1994), pp. 351-356
Published by: The MIT Press
Stable URL: http://www.jstor.org/stable/2109891 .
Accessed: 13/04/2011 15:22

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. 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/action/showPublisher?publisherCode=mitpress. .

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.

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.

The MIT Press is collaborating with JSTOR to digitize, preserve and extend access to The Review of
Economics and Statistics.

http://www.jstor.org
 NOTES 351

ESTIMATING ELASTICITIES WITH THE LINEAR APPROXIMATE ALMOST

IDEAL DEMAND SYSTEM: SOME MONTE CARLO RESULTS

Julian M. Alston, Kenneth A. Foster and Richard D. Green*

Abstract-The AIDS model has many desirable theoretical
propertiesbut usuallyit is estimatedusinga linearapproxima-
tion. The quality of the approximationto the true AIDS
depends on the parametersand the collinearityamong the
exogenousprice variables.In the literaturefour alternative
formulashave been used to computeelasticitiesof a demand
systemthat is assumedto be of the AIDS form,usingparame-
ters estimatedin the linear approximateAIDS. Monte Carlo
experimentsindicate that two of those four alternatives,as
typicallyapplied,are highlyinaccuratebut the other two are
quite accurate.
The Almost Ideal Demand System (AIDS) of Deaton
and Muellbauer is one of the most widely used flexible
demand specifications. While the AIDS possesses many
desirable properties, it may be difficult to estimate.
To simplify the estimation problem, Deaton and
Muellbauer (p. 316) suggested using a linear approxi-
mation. The linear approximate almost ideal demand
system (LA/AIDS) has been employed in the vast
majority of empirical applications of the AIDS model,
with a variety of formulas to compute elasticities. This
paper investigates how well the AIDS price elasticities
Addingup, homogeneity,and symmetryconditionsare
(a) Y2iai = 1, Ei^yij = 0, Eii4 = 0; (b) Ejyij = 0; and
(c) yij = yji. Using the price index in (2) maymake the
estimation of the AIDS difficult.Thus, Stone's price
index (P*) is often used insteadof P where
n
lnPt= E wktln Pkt (3)
k=1
The resulting linear approximatealmost ideal de-
mand system(LA/AIDS) is not an integrabledemand
systemin general.Its widespreadpopularityappearsto
be based on the fact that it is comparativelyeasy to
estimate,combinedwith a belief that it is a reasonably
good approximationto the true AIDS. The purposeof
this paper is to evaluatewhetherthat belief is justified.
This hinges in part on how well Stone's index approxi-
mates the AIDS price index.The relationshipbetween
the two price indexesmay be representedas
In P, In P,* + 6t, (4)
where is a randomvariablewith E(61) = (0. Using
t

are estimated using the LA/AIDS by conducting P,* instead of the unobservableP, causes an errorsin
Monte Carlo experiments in which data are generated variablesproblemand the estimates of the AIDS pa-
by the AIDS. rameters(yij and pi) obtainedby SUR or OLS will be
inconsistent.This can be seen in the LA/AIDS esti-
I. The AIDS and LA/AIDS Models matingequation
n
The AIDS budget share equations for estimation are Wit=a + ln Pj1 + f3i ln(x,/Pt*) + u*, (5)
given by (e.g., Deaton and Muellbauer) j=1

n with u* =
ui,- P3i(6 -
60) and ac*= ai - pi6o and
Wit= ?ai + E yij ln pjt + pi ln(xt/Pt) + uit, (1) cov(u*,lnPt*) o 0.
j=1 Equations(1) through(4) can be solvedto express6t
where, in time t, wit is the budget share of good i, uit
as a functionof parameters,prices, total expenditure,
is a random disturbance, pjt is the price of commodity and the disturbancesin the AIDS model:
j, xt is total expenditure and Pt is a price index defined
by 6t ao 1 +n Epk lnPktk
n n n
k=1
ln Pt = ao + 52 ak ln Pkt + , k In Pkt xt
k=1 k=1 (Ek=1
ak lnPkt ln
n n ( n n
+2 E E2 YjklnpJtln pkk, (2) --E E oYjk In Pjt In Pkt
j=1 k=1 2j=1 k=1
n n
Received for publicationJune 18, 1991. Revision accepted X 1 - F, 1k ln Pkt| - Ukt ln Pkt (6)
for publicationDecember16, 1992. k=1 / k=1
* Universityof California,Davis, Purdue University,and There would be no problemif 6t were a constant,but
Universityof California,Davis,respectively.
The authorsthankProfessorA. Buse for criticalcomments from equation (6) it is clear that t can never be
and suggestionsthat have led to a considerableimprovement constant since the disturbancesfrom all of the equa-
in this paper.Senior authorshipis not assigned. tions influenceit; even the expected value of 6t need

Copyright? 1994
352 THE REVIEW OF ECONOMICS AND STATISTICS

not be constant, because it depends on the values Several authors have used this equation to compute
taken by prices. elasticities with LA/AIDS parameter estimates. How-
What are the determinants of the quality of the ever, Green and Alston (1990) have shown that substi-
approximation? Suppose we ignore the disturbance tuting
component and conduct the argument "on average."
Deaton and Muellbauer (p. 316) asserted that the dlnP* n dlnwkn
approximation would be better if prices of the goods dln1= WJ + E Wk n Pkdl
making up the group were highly collinear, but they
provided no analytic results to justify the claim. Mini- n
mally one needs perfect collinearity of the form = Wi + E Wk lnpk(Eki + 8k])' (10)
k=1
Ekf3k ln Pkt = 0 and EkYki ln Pkt = 0 to produce
E(Qt) = ao. Similarly, in the special case of homothetic
separability (,3i = yij = 0 Vi, i), 6t = ao - EkUkt ln Pkt, into (7) yields n2 simultaneous equations for uncom-
and E(6t) = ao. Extrapolating from these results, as pensated LA/AIDS demand elasticities:'
either Bk or Ykj tend to zero, or the correlation among
the prices of goods increases, we would expect the Eij(LA) = -Sij + yij -
approximation of ln Pt by ln Pt* to get better and our W. W.

estimates of parameters and elasticities should im- n

prove. The Monte Carlo experiments below are de- W
Wj+ ln Pk(
X , Wk Eki + Ski)
signed to explore this suggestion. k=1
(11)
II. Price Elasticities
Another common approach (e.g., Chalfant) is to use
A general definition of the uncompensated price a special case of this formula, which obtains when
elasticities of demand from the AIDS and LA/AIDS d ln P*/d ln pj = wj:
(eij), suppressing t subscripts for the time being, is:
d ln Qi d ln w Eij(LA')
=
-8j W!!w (12)
wi
Eij= d ln pj -
ij +d ln pj

Eales and Unnevehr used a further special case of

= ij + W w dlnp1' (7) equation (11) that is correct only when preferences are
homothetic (,Bi = 0 Vi):
where these elasticities refer to allocations within the
group holding constant total group expenditures (x) Eij(LA") = -8jj + yij (13)
and all other prices (Pk, k jj), Sij is the Kronecker Wi

delta (8ij = 1 for i = 1; ij = 0 for i j)] and for the

LA/AIDS we use ln P* from (3) instead of ln P Equation (9) (hereafter AI) is correct when the
from (2). AIDS price index is used. Equation (11) (hereafter LA)
Four alternative formulas for price elasticities using is correct when the LA/AIDS has been estimated, so
LA/AIDS parameter estimates have appeared in the long as consistent LA/AIDS parameter estimates have
literature. The differences can be represented in terms been obtained. Equations (12) (hereafter LA') and (13)
of different expressions for the elasticity of the group (hereafter LA") are approximations to the LA formula.
price index with respect to the jth price (i.e., Corresponding formulas for compensated price elastic-
d ln P/d ln pj in the AIDS or d ln P*/d ln pj in the ities (Ei) are obtained by using the Slutsky equation
LA/AIDS). In the AIDS model, (Eij = 1ij + WjEi, X) and formulasfor the expenditure
elasticities (Ej, x) For the AI, LA' and LA", Ej x=
dlnP n 1 +,Bi/wi is used. Green and Alston's (1991) result is
d In pj
dlnp1 j E ykl ln Pk. (8)
~ k=1
1
Equation(11) can be expressedin matrixform as: E =
Substituting (8) into (7) yields: A - (BC)(E + I), where the typical elements are a = - Ij
+ (yij - 8iwj)lwi in A (an n X n matrix); bi = (lp/wi) in B
(an n x 1 vector); cj = wj ln pj in C (a 1 x n vector); and E1j
Ei1(AI) = ij + W WaJ + E Yyk ln Pkl in E (an n X n matrix). Solving for the elasticities [Eij] yields,
w. w. k=1 after some simplifications: E = [BC + I]-1[A + 1] - I. Buse
(9) has derivedan equivalentsolutionthat is simpler.
 NOTES 353

used for the LA formula.The formulasfor compen- the prices, Zi is a vector of standardnormalrandom
sated elasticitiesare: variablesand A is vector of mean prices.
The "high," "medium," and "low" values of
E,J(AI)
tj =-8ij + Li (18 182' /3) were (0.70, -0.40, -0.30), (0.40, -0.25,
Wi~~~~I -0.15) and (0.05, -0.03, -0.02), respectively. The pa-
rametera0 was set at unity. The ai were determined
+ a +
(Oj Ykiln Pk wi).
wj usingthe share equationand the medianvaluesfor the
shares(defineda priorias 0.4, 0.3, and 0.3 for goods 1,
(9"
2 and 3). Imposingsymmetryand homogeneityrestric-
yii
tions leaves three free (yij) parametersto be defined
WW (Y12' 713' Y23). These were determinedby definingthe
correspondingelasticities of substitution (ij). The
+ Wj- -(kElWk ln Pk(ekj + Skj Wj))- "high,""medium,"and "low"values of (0`12, r13', 023)
(11')
were (1.70, 1.70, - 1.70), (1.00, 1.00, -1.00) and
(0.30,0.30, -0.30), respectively.The logarithmof ex-
E,](ILzA')= -(S. + - + wi. (12') penditureswas generatedfrom the AIDS expenditure
7-- function(e.g., see Deaton and Muellbauer,p. 313) with
Wi3
the utility index set at 0.5 and using the generated
E1J(IA) = -jJ + W Wj
+ - . (13') prices and established"true" parameters.One thou-
sand sets of normalrandomdisturbanceswere gener-
ated for each designpoint and added to the generated
III. ExperimentalDesign shares.The disturbancevectorswere definedby: U2 -
Data were generatedfor three commoditiesallowing N(0,.007), U3 N(0,.007), and U1 = -U2 - U3. The
three different settings each for the degrees of precision level of 0.007 was chosen to coincide with
(i) multicollinearity(pij's), (ii) homotheticity(f38's),and results obtained from previous estimations of the
LA/AIDS (e.g., Blanciforti,Green, and King).3The
(yij's).The sets of parametervalues
(iii) substitutability
were chosen to satisfy theoretical restrictionsand to LA/AIDS was estimatedby iterated SeeminglyUnre-
result in a typicalrangeof elasticitiesas found in food lated Regression(ITSUR) and, considering(11) treats
demandstudies.2At each of the 27 designpoints, 1,000 the shares as endogenous, by iterated Three Stage
replicationswere performed,each with 35 (and also Least Squares(3SLS). For each set of estimates, de-
70) observations. mand elasticitieswere computedusing the four formu-
Priceswere drawnfrom a multivariatenormaldistri- las and we computedan overall"mean percent error"
butionwith a covariancematrix: for each formula:

2
1 P120VTi72 P131l03 Mean PercentError= 100 E 9[ i
i=1 j=1 9 EIJ
P12U2U1 2 P2323 (14) (16)
[P13of1o3 P230f20'3 0`3

where the o-iare standarddeviationsand the pij repre- IV. SimulationResults

sent the simple correlationsbetween the prices. The
Table 1 shows the results for compensatedelastici-
"high," "medium"and "low" values used for the pij
ties at one of the 27 design points (HMM): "high"
are (pij = 0.99), (pij = 0.75) and (pij = 0.50). The final
values for 8ji'sand "medium"values for yij's and p.
price series were generatedby factoring I (the price
This case was typical in that the LA and LA' were
covariancematrix)into a matrix A such that X = A '
quite accurate,very similarto one another, and both
and computing
more accuratethan the LA"and, especially,the AI. As
P, = AZi + . (15) shown in equation(5), the LA/AIDS estimatesof the
where the vector Pi representsthe ith observationon
3Although u2 and U3 were drawnindependently,the sam-
ple covarianceswere not exactlyzero (i.e., 0'23 = COV(U2,U3)
2
Desiredvalues for the elasticitiesof substitution,expendi- # 0) and, by construction,Cr12 = - C2 - O23 and 013 = -32
ture elasticities,and simplecorrelationsbetweenprices,were - 0T23. Some limitedexperimentationwith a bivariatenormal
specifiedin order to define the parameterscomputedat the distributionfor u2 and U3 indicated that the size of the
mean shares and prices. Since the regressorsare Inpj, the covariancebetween ul and u2 did not affect results much,
correlationsbetween regressorswill be different from the and introducingthis as an additionaldesign dimensionap-
correlationsbetweenthe originalprice series. pears to add unnecessarycomplications.
354 THE REVIEW OF ECONOMICS AND STATISTICS

TABLE l.-MEANS AND ROOTMEAN SQUARE ERRORS OF LA / AIDS COMPENSATED

PRICE ELASTICITIES, HMM DESIGN POINT (N = 35, ITSUR)

Mean Calculated Values

Elasticity True Value AI LA" LA' LA

-0.58 - 0.39 - 0.52 -0.61 -0.61

(0.35) (0.20) (0.03) (0.03)
0.24 0.20 0.31 0.26 0.27
(0.28) (0.13) (0.03) (0.04)
61*3 0.35 0.20 0.42 0.35 0.34
(0.34) (0.19) (0.02) (0.03)
62*1 0.38 0.32 0.74 0.42 0.42
(0.45) (0.50) (0.05) (0.05)
22 -0.02 0.59 0.11 -0.10 - 0.05
(0.98) (0.21) (0.09) (0.04)
62*3 -0.36 -0.91 -0.03 - 0.32 - 0.37
(0.83) (0.47) (0.05) (0.03)
0.39 0.22 0.07 0.39 0.39
(0.38) (0.36) (0.02) (0.02)
63*2 -0.26 -0.64 -0.43 -0.23 -0.27
(0.59) (0.19) (0.04) (0.03)
63*3 --0.13 0.41 - 0.45 -0.18 -0.12
(0.72) (0.35) (0.04) (0.02)
Mean Percent Error 434.74 142.96 47.63 21.48
Notes: The HMM design point consists of "high" values of /3j's (0.7, -0.4, -0.3), a "medium" degree of
multicollinearity (simple correlation coefficients between prices of p,, = 0.75), and "medium" values of elasticities
of substitution (1.0, 1.0, - 1.0). The estimated correlation between In P* and In P is pp = 0.960. Entries are
averages of 1,000 values of compensated price elasticities and numbers in parentheses are root mean square errors.

intercepts are related to the intercepts in the true
AIDS share equations according to ac* = ai - P8i. If
we denote the elasticity using ac by Eij(AI*) and the
elasticity using ai by Eij(AI), then we can write
Eij(AI*) = Eij(AI) + p8pj8/wj. At unit prices 6 = a0
> 0 (the subsistence level), and Eij(AI*) will systemat-
ically over-estimate the own-price elasticities whereas
the direction of bias in the cross-price elasticities will
be determined by the signs of p1 and pj. This effect
can be seen in the results in table 1.4
Table 2 summarizes the results across the 27 points
in terms of Mean Percent Errors. Several patterns are
apparent. First, the LA and LA' formulas are consis-
tently more accurate than the LA" and AI formulas.
Second, the LA' (as used by Chalfant) provides a very
good approximation to the LA (advocated by Green
and Alston, 1990) and, in fact, in twelve of the twenty-
seven cases outperformed the LA. Third, as the degree
of multicollinearity increases, accuracy decreases for
all of the formulas, but especially for the LA" and AI.
While higher multicollinearity is expected to lead to an
increase in the accuracy of the Stone's index as an
approximation to the AIDS price index, it also makes
estimation of parameters more difficult generally.
Fourth, as preferences depart further from homoth-
eticity (as the pj's increase) precision falls, and as
4 This result, amnongothers, is derived by Buse.
substitutability among the goods rises (as the yij's
increase) precision improves.
This last result corresponds to the idea that the
precision of elasticity estimates would depend on the
accuracy of the approximation using the Stone's index.
At every level of multicollinearity, the errors were
greatest when low yij's were combined with high fj's,
the conditions under which precision of the Stone's
index was lowest. The simple correlations (pp) between
the AIDS price index (ln P) and Stone's price index
(lnP*) at each of the 27 design points are shown in
table 2. Our analytic results suggested that low yij's,
low f3j's, or high multicollinearity among the goods'
prices would improve the precision of the approxima-
tion. The results bear this out. The correlations were
greater than 96% in two-thirds of the cases; greater
than 99% in all of the cases when multicollinearity was
"high" level and in all of the cases when f3j's were
"high." With "low" or "medium" multicollinearity, the
correlations fell with increases in pj's and with de-
creases in yij's.
In summary, the LA and LA' formulas for compen-
sated elasticities are always more accurate than the
alternatives as they are commonly applied. The LA'
provides a very good approximation, and it might not
be worth making the additional effort required to com-
pute the LA formulas. Increasing the sample size from
35 to 70 increased precision slightly but did not change
the overall picture. Using 3SLS rather than SUR, in
 NOTES 355

TABLE 2.-MEAN PERCENT ERRORS IN COMPENSATED ELASTICITIES, ALL DESIGN POINTS

(N = 35, ITSUR estimates)
Mean Percent Errors
Design Point
(pi's, Pij, Yij's) LA A LA AI pp

LLL 6.16 3.94 77.26 325.55 0.998

LLM 2.46 3.97 31.83 28.21 0.992
LLH 0.57 0.70 8.01 29.90 0.996
MLL 13.78 14.09 34.44 51.07 0.935
MLM 3.64 7.06 16.78 18.39 0.905
MLH 0.97 0.99 19.62 22.91 0.949
HLL 19.36 9.49 288.62 373.09 0.915
HLM 6.15 5.30 29.81 75.76 0.939
HLH 1.02 1.44 53.67 82.49 0.830
LML 9.75 5.03 275.72 595.57 0.999
LMM 2.25 10.29 102.33 300.48 0.999
LMH 0.85 0.48 22.59 93.18 0.999
MML 1.33 0.95 42.29 177.29 0.967
MMM 1.75 2.27 14.48 27.55 0.955
MMH 2.66 3.42 26.44 49.19 0.983
HML 43.68 68.22 820.69 2322.74 0.954
HMM 21.48 47.63 142.96 434.74 0.960
HMH 2.97 3.33 37.91 145.73 0.920
LHL 32.06 2.64 3208.50 67928.71 0.999
LHM 5.75 0.75 551.09 10257.75 0.999
LHH 19.10 22.13 681.58 13147.08 0.999
MHL 27.46 21.54 1328.24 15867.05 0.999
MHM 5.62 4.32 246.41 7996.77 0.999
MHH 1.65 1.24 365.42 6766.09 0.999
HHL 103.51 169.59 4119.20 88055.89 0.994
HHM 28.40 7.46 1644.05 42653.96 0.997
HHH 7.58 7.82 254.90 3180.52 0.997
Notes: See notes to table 1. L, M, and H denote "low," "medium," and "high" settings of parameters. For
example, LMH denotes "low" values for /3,'s, "medium" values for pq's, and "high" values for -y,'s. pp is the
estimated correlation between In P* and In P.

recognition of the endogeneity of shares in the Stone's
index, did not have much effect either. The results
were similar for uncompensated elasticities but in gen-
eral the accuracy was much lower than for compen-
sated elasticities, and the results across formulas were
less clear-cut, perhaps because the term involving the
income effect is also the term involving the Stone's
index in the LA/AIDS model.5
V. Conclusion
The primary conclusion of this paper is that vastly
different values can be obtained for AIDS elasticities
when the LA/AIDS parameter estimates are substi-
tuted in various elasticity expressions. Two of the for-
mulas that have been used in previous studies (one of
which is the elasticity formula for the true AIDS model)
result in very poor estimates, especially when multi-
collinearity among prices is high. The best results are
obtained from either the elasticity expression that as-
sumes that budget shares are endogenous on the right
5Results for these other cases are availablefrom the au-
thors.
hand side of the demand equations or the one that
assumes they are constant. Using these formulas for
elasticities, the LA/AIDS provides quite accurate esti-
mates of elasticities when the true data generating
process is AIDS. Demand analysts can consequently
have a certain degree of confidence when estimating
the LA/AIDS and using either of these formulas for
obtaining estimates of the "true" AIDS elasticities.
Alternatively, as shown by Buse, good estimates can be
obtained by correcting the LA/AIDS estimates of the
intercepts in the AIDS model and using the AI for-
mula.
REFERENCES
Blanciforti, Laura, Richard Green and Gordon King, "U.S.
Consumer Behavior Over the Postwar Period: An Al-
most Ideal Demand System Analysis," Giannini Foun-
dation Monograph No. 40 (Aug. 1986).
Buse, Adolf, "The Calculation of Correct AIDS Elasticities,"
University of Alberta, unpublished manuscript, May
1992.
Chalfant, James A., "A Globally Flexible, Almost Ideal De-
mand System," Journal of Business and Economic
Statistics 5 (1987), 233-242.
356 THE REVIEW OF ECONOMICS AND STATISTICS
Deaton, Angus, and John Muellbauer, "An Almost Ideal
Demand System," American Economic Review 70
(1980), 312-326.
Eales, James, and Laurian Unnevehr, "Beef and Chicken
Product Demand," American Journal of Agricultural
Economics 70 (1988), 521-532.
ESTIMATING A SYSTEM OF RECREATION DEMAND FUNCTIONS USING A
SEEMINGLY UNRELATED POISSON REGRESSION APPROACH
Teofilo Ozuna, Jr. and Irma Adriana Gomez*
Abstract-In this article, a seeminglyunrelatedPoisson re-
gressionmodel (SUPREME)is presentedas an alternativeto
usingZellner'sseeminglyunrelatedregressionmodel for esti-
matinga systemof recreationdemandfunctions.SUPREME
providesestimatesthat are asymptotically more efficientthan
equation-by-equation Poisson estimates and circumventsthe
bias and inconsistencyproblems that result when using
Zellner'sseeminglyunrelatedregressionmodel. Additionally,
SUPREMEis applied to an empiricalproblemdealingwith
the valueof recreationalboatingand the findingsindicatethat
the SUR consumersurplusestimatesare substantiallydiffer-
ent from those of SUPREME.
I. Introduction
Prior recreation demand research acknowledges the
importance of substitute sites for the estimation of
recreation demand functions. Research in this area
shows that omitted variable bias can occur if substitute
sites are not incorporated in the estimation process.
Hence, researchers either include substitute site vari-
ables in the recreation demand function or jointly
estimate a system of recreation demand functions which
include substitute sites (Burt and Brewer (1971); Sellar,
Stoll, and Chavas (1985)).
Current research also recognizes the importance of
adequately modeling the count nature of trip demand.
With respect to this issue, Creel and Loomis (1990),
Grogger and Carson (1991), and Hellerstein (1991)
suggest the use of count distributions for the esti-
mation of recreation demand functions. Failure to
account for this distributional issue will result in incon-
sistent parameter estimates (Mullahy (1986)) and, con-
sequently, invalid consumer surplus estimates.
This article adds to the aforementioned literature in
that it proposes a seemingly unrelated Poisson regres-
Received for publicationOctober 15, 1991. Revision ac-
cepted for publicationDecember16, 1992.
*Texas A & M University.
The authors are grateful to Lonnie L. Jones and Badi
Baltagifor their helpful commentsand to John R. Stoll for
providingthe data for the empiricalillustration.
Green, Richard, and Julian M. Alston, "Elasticities in AIDS
Models," American Journal of Agricultural Economics
72 (1990), 442-445.
_____, "Elasticities in AIDS Models: A Clarification and
Extension," American Journal of Agricultural Eco-
nomics 73 (1991), 874-875.
sion model (SUPREME) for the estimation of a system
of recreation demand functions. The model provides
parameter estimates that are asymptotically more ef-
ficient than equation-by-equation Poisson estimates and
circumvents the bias and inconsistency problems that
result when using Zellner's (1962) seemingly unrelated
regression model (SUR) with endogenous count vari-
ables (King (1989)). Finally, the model permits the
testing of cross equation economic hypothesis.
The rest of the article is organized as follows. The
next section briefly reviews SUPREME. The third sec-
tion applies the model to a recreational boating study.
The final section provides some concluding comments
and possible extensions of SUPREME.
II. The Seemingly Unrelated Poisson Regression
Model
Consider the following recreation demand functions:
Qli = o + 1P1i + 8p2i2 + Eli, (la)
Q2i = o + alP1i + a2P2i + e21, (lb)
where Qli and Q2i, P1i and P2j, and E1i and E2i are,
respectively, the number of trips, surrogate prices, and
disturbance terms for recreation sites 1 and 2. If the
disturbances between the functions are uncorrelated,
there is no relationship between the functions, and
separate least squares estimation is appropriate. Alter-
natively, if the disturbances are correlated, efficient
estimates are obtained using Zellner's (1962) SUR
model.
Zellner's (1962) SUR model, however, requires the
assumption of normally distributed disturbances as well
as linear or loglinear specifications for each equation.
For count endogenous variables, as encountered in
recreation demand analysis, these assumptions are im-
plausible and result in substantial sacrifices in bias and
consistency properties (King). Additionally, the SUR
Copyright ? 1994