TO POOL OR NOT TO POOL: HOMOGENEOUS VERSUS HETEROGENEOUS

ESTIMATORS APPLIED TO CIGARETTE DEMAND

Badi H. Baltagi, James M. Griffin, and Weiwen Xiong*

Abstract —This paper reexamines the beneŽ ts of pooling and, in addition, But, even if one agrees that a pooled estimator is preferable
contrasts the performance of newly proposed heterogeneous estimators.
The analysis utilizes a panel data set from 46 American states over the
to individual time-series or cross-section estimates, there
period 1963 to 1992 and a dynamic demand speciŽ cation for cigarettes. remains the question of which pooled estimators yields the
Also, the forecast performance of the various estimators is compared. most plausible estimates.
More recently, the fundamental homogeneity assumption
I. Introduction underlying pooled models has been called into question.
Robertson and Symons (1992), Pesaran and Smith (1995)
T here has been a recent proliferation of pooled and
heterogeneous estimators with the latter questioning the
beneŽ ts of pooling. We place this debate within the context
and Maddala et al. (1994) argue in favor of heterogeneous
estimates rather than the traditional homogeneous estimates.
of cigarette demand because of the policy importance of the Depending on the extent of interstate heterogeneity in
long-run price elasticity of demand in affecting tax revenues parameters, researchers may prefer these heterogeneous
and discouraging consumption. Both from a methodological estimators to the traditional pooled homogeneous parameter
as well as a policy perspective, the long-run price elasticity estimators.
of cigarette demand remains an important unresolved issue. Our objective in this study is to compare the performance
An interesting methodological question is to what degree of both these homogeneous and heterogeneous estimators
can elasticity differences be attributable to the manner in applying them to a well-researched good (cigarette consump-
which applied econometricians analyze a given body of tion), a common data set (a panel of 46 states), and a
data. SpeciŽ cally, this study standardizes on the frequently routinely applied dynamic demand speciŽ cation. In compar-
utilized dynamic demand speciŽ cation and a common data ing pooled homogeneous estimators and their heterogeneous
set (annual data for 46 states covering the period 1963– rivals, we examine the plausibility of alternative estimates of
1992) and asks the following three questions. the price and income elasticities as well as the speed of the
adjustment path to long-run equilibrium. Admittedly, ours is
(1) How do the price elasticity estimates differ depend- a ‘‘case study’’ for cigarette demand. The fact that these
ing on the estimation approach? Ž ndings corroborate similar results using an international
(2) Based on predictive performance, which estimates panel for gasoline consumption (Baltagi and Griffin (1997))
are more plausible? suggests that the Ž ndings offer useful general guidelines to
(3) Do these Ž ndings suggest useful rules of thumb to applied researchers facing the dilemma ‘‘to pool or not to
applied researchers facing the decision to pool or not pool.’’
to pool? A distinctive characteristic of this paper is that we
compare the forecast performance of these alternative ap-
Armed with a panel data set, the researcher faces a wide proaches using the model to provide forecasts of cigarette
variety of estimation options. For example, if heterogeneity consumption over a ten-year horizon. Section II brie y
between states is viewed as pervasive, one can simply reviews the standard habit-persistence type of dynamic
forsake pooling and apply individual time series to each demand speciŽ cation. Section III presents the pooled homo-
state. Alternatively, if one believes that the long-run re- geneous parameter results, while section IV presents the
sponse is best captured by cross-sectional variation, a heterogeneous model results. Section V compares the plausi-
between state regression approach can be employed. The bility of the various estimates and their forecasting perfor-
Baltagi and Levin (1986) cigarette-demand study argues mance over horizons of one, Ž ve, and ten years. Section VI
persuasively for pooling the data as the best approach for recapitulates our major Ž ndings.
obtaining reliable price and income elasticities. It also points II. Model SpeciŽ cation
out that pure cross-section studies cannot control for state-
speciŽ c effects, whereas pure time-series studies cannot Following Laughhunn and Lyon (1971), Hamilton (1972),
control for unobservable taste changes occurring over time. Doron (1979), and Baltagi and Levin (1986), it is reasonable
to model cigarette demand as follows.

Cit 5 f (Pit , Yit , Pnit , Zi , Zt ), (1)

Received for publication July 15, 1997. Revision accepted for publica-
tion February 8, 1999.
* Texas A&M University. where Cit is real per capita sales of cigarettes (measured in
An earlier version of this paper was presented at the North American packs of cigarettes per head) by persons of smoking age
Summer Meeting of the Econometric Society held at Caltech in June of
1997. We would like to thank the referees for helpful comments and (sixteen years and older), Pit is the average retail price of a
suggestions. pack of cigarettes measured in real terms, Yit is real per

The Review of Economics and Statistics, February 2000, 82(1): 117–126

r 2000 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology
118 THE REVIEW OF ECONOMICS AND STATISTICS

capita disposable income, and Pnit denotes the minimum real advantages of a panel is its ability to control for all
price of cigarettes in any neighboring state.1 This enables time-invariant variables or state-invariant variables whose
controlling for possible ‘‘bootlegging effects,’’ which Baltagi omission could bias the estimates in a typical cross-section
and Levin (1986) found important in explaining why some or time-series study. Both effects can be assumed to be either
very low tax states enjoy much higher cigarette sales than Ž xed or random. We assume that the time-period effects (the
neighboring states with higher taxes.2 The subscript i l t’s) are Ž xed parameters to be estimated as coefficients of
denotes the ith state (i 5 1, . . . , 46), and the subscript t time dummies (Dt ) for each year in the sample. This can be
denotes the tth year (t 5 1, . . . , 30). This study updates the justiŽ ed given the numerous policy interventions as well as
original Baltagi and Levin data twelve more years from health warnings and Surgeon General’s reports that previous
1981 to 1992, so that the panel covers 46 states over thirty studies accounted for using time-dummy variables.3 Major
years (1963–1992). Zi denotes a vector of state-speciŽ c, policy interventions that can change l t include
time-invariant variables that include religion, race, educa-
tion, tax-free Indian reservations, and tourism. Zt denotes a 1. the imposition of warning labels by the Federal Trade
vector of year-speciŽ c, state-invariant variables that include Commission effective January 1965,
health warnings due to the Surgeon General, warning labels 2. the application of the Fairness Doctrine Act to ciga-
by the Federal Trade Commission, national advertising rette advertising in June 1967, which subsidized
expenditures on TV and radio, and the ban of broadcast antismoking messages from 1968 to 1970,
advertising of cigarettes effective January, 1971. 3. the Congressional ban of broadcast advertising of
Following Houthakker and Taylor (1970) and McGuiness cigarettes effective January 1971, and
and Cowling (1975), we assume that cigarette consumption 4. ‘‘clean air laws’’ restricting smoking in the work
is governed by a partial adjustment or habit-persistence place, public places, and commercial  ights within the
model; that is, U.S.4
ln Cit 2 ln Ci,t2 5 d (ln C *
it 2 ln Ci,t2 1 ) 1 uit , (2)
1 Similarly, the µi’s are state-speciŽ c effects that can represent
where ln C *it is the expected or desired level of consumption any state-speciŽ c characteristic such as:
of cigarettes which is given by
1. Indian reservations selling tax-exempt cigarettes,
ln C *it 5 a *1 b *1 ln Pit 1 b *2 ln Yit 1 b *3 ln Pnit 2. tax-exempt military bases in certain states,
(3) 3. a high percentage of Mormon population (a religion
1 Z8ig * 1 Z8th *.
that forbids smoking), and
Substituting equation (3) into (2) produces the following 4. a highly touristic state, such as Nevada, with a per
log-linear dynamic demand model: capita consumption of twice the national average.

ln Cit 5 a 1 (1 2 d ) ln Ci,t2 1 1 b 1 ln Pit 1 b 2 ln Yit Econometric treatment of these types of effects can range
(4) from ignoring them altogether to incorporating a full set of
1 b 3 ln Pnit 1 Z8ig 1 Z 8th 1 uit ,
state dummy variables (Di ) for i 5 1, 2, . . . , 46. An
where a 5 d a *, b j 5 d b *j for j 5 1,2,3, g 5 d g *, and h 5 intermediate solution is to explicitly model these state
d h *. attributes by including state-speciŽ c, time-invariant vari-
We follow the usual convention (Hsiao (1986)) of assum- ables, which we denote by Zi. These include religion, race,
ing that the disturbance term in equation (1) is speciŽ ed as a education, and a state tourism index. 5
two-way error component model:
uit 5 µi 1 l t 1 n it
3
See Hamilton (1972) and Baltagi and Levin (1986).
(5) 4 See Doron (1979), appendix A, for a chronology of Federal Commis-
i5 1, 2, . . . , 46; t5 1, 2, . . . , 30 sion interventions in the market practices of the cigarette industry. See also
the 1989 Surgeon General’s report (USDHHS, 1989), which Ž nds that, in
where µi denotes a state-speciŽ c effect, l t denotes a year- 1985, one out of every six deaths in the U.S. was the result of past and
current smoking.
speciŽ c effect, and n it is white noise. One of the major 5 Religion is a group of variables giving the percentage of the state’s

population that is Mormon, Catholic, Southern Baptist, United Methodist,

United Church of Christ, Episcopalian, Jewish, and Lutheran. Race gives
1 For data sources, as well as details on the construction of the the percentage of the state’s population that is black, white, Asian,
neighboring price, see Baltagi and Levin (1986). Per capita sales of Hispanic, PaciŽ c Islander, or Eskimo/American Indian. Education gives
cigarettes and the average retail price per pack were obtained from The the percentage of the state’s population in secondary and elementary
Tobacco Institute (1993). Other studies indicating the importance of schools. The tourism variable gives the percentage of the state’s revenue
bootlegging include Manchester (1976) and Warner (1982). from tourism-related services. The religion variable is obtained from
2 The Advisory Commission on Intergovernmental Relations estimates Johnson et al. (1974). The percentage of the state’s population that is
that, in 1975, cigarette bootlegging cost the high-tax states $391 million. Jewish is obtained from the American Jewish Yearbook. The tourism and
The seriousness of the problem was officially recognized by the enactment race variables are obtained from the County and City Data Book. The
of the 1978 Federal Contraband Act which made the transportation of education variable is obtained from the Statistical Abstract of the United
contraband cigarettes in interstate commerce a federal criminal offense. States.
 HOMOGENEOUS VERSUS HETEROGENEOUS ESTIMATORS 119

TABLE 1.—CIGARETTE DEMAND: HOMOGENEOUS PARAMETER ESTIMATES 1963 –1992

Short Run Long Run

Model Type ln Cit2 l ln Pit ln Pnit ln Yit ln Pit ln Pnit ln Yit

A. Traditional Pooled Estimators

OLS 0.97 2 0.090 0.024 2 0.03 2 2.90 0.77 2 1.00
(157.7) (6.2) (1.8) (5.1) (5.04) (1.78) (3.39)
OLS with Dt 0.95 2 0.137 0.037 2 0.01 2 2.98 0.80 2 0.20
(148.5) (8.7) (2.7) (1.1) (7.76) (2.60) (1.06)
OLS with Zi and Dt 0.90 2 0.187 0.053 0.042 2 1.95 0.55 0.44
(102.7) (11.1) (3.2) (3.3) (11.79) (3.08) (3.51)
Within 0.83 2 0.299 0.034 0.10 2 1.79 0.20 0.60
(66.3) (12.7) (1.2) (4.2) (12.38) (1.22) (4.27)
GLS with Zi and Dt 0.91 2 0.178 0.050 0.03 2 2.00 0.56 0.31
(106.4) (10.8) (3.1) (2.3) (11.16) (3.03) (2.46)
GLS-AR(1) with Zi 0.96 2 0.089 0.016 2 0.03 2 2.09 0.37 2 0.63
(147.6) (6.8) (1.3) (4.0) (6.64) (1.32) (3.27)
B. 2SLS Type Pooled Estimators
2SLS 0.85 2 0.205 0.052 2 0.02 2 1.37 0.35 2 0.11
(25.3) (5.8) (3.1) (2.2) (11.97) (1.52) (1.46)
2SLS with Zi and Dt 0.63 2 0.482 0.073 0.19 2 1.31 0.20 0.51
(17.3) (11.1) (3.3) (7.5) (23.92) (3.23) (11.85)
2SLS-KR 0.71 2 0.311 0.071 2 0.02 2 1.07 0.24 2 0.05
(22.7) (13.9) (3.7) (1.5) (11.80) (3.19) (1.55)
2SLS-KR with Zi 0.61 2 0.390 0.074 2 0.04 2 1.00 0.19 2 0.11
(16.7) (15.9) (3.3) (2.9) (12.44) (2.90) (3.07)
Within-2SLS 0.60 2 0.496 2 0.106 0.19 2 1.25 2 0.04 0.47
(17.0) (13.0) (0.5) (6.4) (17.96) (0.51) (7.14)
EC2SLS with Zi and Dt 0.70 2 0.406 0.067 0.14 2 1.35 0.22 0.47
(23.3) (11.2) (3.4) (6.9) (22.00) (3.36) (10.06)
EC2SLS-AR(1) with Zi 0.67 2 0.35 0.12 0.05 2 1.07 0.37 0.14
(16.6) (11.9) (3.9) (1.8) (6.7) (3.5) (1.75)
FD2SLS 0.51 2 0.348 0.112 0.10 2 0.71 0.23 0.21
(9.5) (12.3) (3.5) (2.9) (7.49) (3.05) (2.68)
FD2SLS-KR 0.49 2 0.348 0.095 0.13 2 0.68 0.19 0.26
(18.7) (18.0) (4.7) (9.0) (15.64) (4.38) (8.70)
Numbers in parentheses denote t-statistics.

III. Pooled Results 6. Finally, we assume that the µi’s are random and the
n it’s follow an AR(1) process, and we estimate the
A. Results Using Traditional Homogeneous Panel
model by GLS as described in Baltagi and Li (1991).
Estimators
We begin by comparing the results of six traditional As shown in section A of table 1, in the Ž rst regression,
homogeneous estimators, all of which make slightly differ- OLS—which ignores intertemporal and interstate taste differ-
ent assumptions about the error term in equation (4). These ences—Ž nds a very low short-run own price elasticity of
six traditional estimators are 2 0.090 but a very large long-run price elasticity of 2 2.90.
Inclusion of the time-period dummies in OLS (OLS with Dt )
1. OLS (without Dt and Di ), which ignores state-speciŽ c in table 1 does not materially affect the short-run and
and time-speciŽ c effects; long-run price elasticities. The 0.95 coefficient on lagged
2. OLS with Dt, which assumes time-period dummies consumption is no doubt biased because it is correlated with
re ecting structural shifts over time but ignores state- the omitted µi effects.6 To illustrate the importance of
speciŽ c effects; explicitly modeling interregional taste variables, the third
3. OLS with Zi and Dt, which assumes that the state
regression in table 1A—OLS with Zi and Dt—introduces Zi’s
effects can be explicitly represented by the Zi variables
for race, religion, education, and tourism to explain inter-
(such as race, religion, education, and tourism);
state taste differences. As expected, controlling for interre-
4. the Within estimator, which assumes the µi’s and l t’s
gional effects with the Zi’s causes the coefficient on lagged
are Ž xed effects that are modeled by time-period and
state dummies (Di and Dt ); consumption to decline from 0.95 to 0.90 and markedly
5. the GLS estimator, which treats the µi’s as random and reduces the long-run price elasticity to 2 1.95. Both the
the l t as Ž xed-effects, time-period dummies. (We also long-run neighboring price elasticity (0.55) and the income
include the Zi’s and estimate the model using the
instrumental variable method proposed by Hausman 6 In addition, OLS has biased standard errors because it does not account

and Taylor (1981); for state-speciŽ c effects. See Moulton (1986).

120 THE REVIEW OF ECONOMICS AND STATISTICS

elasticity (0.44) are statistically signiŽ cant, and the esti- (a) there exist interstate random effects (µi ) that will
mates of the Zi coefficients are generally plausible. 7 surely be correlated with lagged consumption and
But the list of Zi’s is most likely incomplete. The Within possibly correlated with the other explanatory vari-
estimator in table 1 may be preferred in principle because it ables, and
completely controls for time- and state-speciŽ c effects.8 (b) autocorrelation in the disturbance term n it in equation
However, the cost of this is to lose all between-state (4) introduces endogeneity between the n it and the
variation and thus information on the source of interstate lagged dependent variable.
taste differences, which serve to identify the Zi coefficients.
Most signiŽ cantly, the coefficient on lagged consumption In contrast, the Within estimator avoids the inconsistency
drops even further to 0.83, dropping the long-run own price arising in (a) from the µi’s being correlated with the
elasticity to 2 1.79. Note that another casualty of eliminating explanatory variables, because it sweeps away the µi’s.
between-state variation was the statistical insigniŽ cance of However, the Within estimator is still subject to the inconsis-
the neighboring price effect (Pn), which depended on tency in (b) arising from serial correlation of the n it’s (see
long-standing interstate tax differentials to identify its ef- Nickell (1981) and, more recently, Kiviet (1995)). Whereas,
fects. Finally, the GLS estimates in table 1A treat the µi’s as the Hausman and Taylor GLS estimator recaptures estimates
random effects and thereby incorporate between-state varia- of the Zi’s lost by the Within estimator, it will be inconsistent
tion. For the Hausman and Taylor instrumental variable due to either (a) or (b).9 In fact, Hausman’s (1978) test for
estimation (GLS with Zi and Dt ), the lagged consumption- speciŽ cation error based on the difference between GLS and
coefficient estimate is 0.91, and the own long-run price the Within estimator yields a x 24 5 58.29. This has a p-value
elasticity is 2 2.00. Adjusting for Ž rst-order serial correla- of 0.0001 and decisively rejects the independence of the µi’s
tion on the remainder disturbances (GLS-AR(1)), the lagged and the explanatory variables. Consequently, we turn to
consumption-coefficient estimate is 0.96, and the long-run instrumental variable methods that will correct for the
own price elasticity is 2 2.09. inconsistency caused by reasons (a) or (b).
Interestingly, the traditional estimators as a group enjoy
certain areas of conformity, but also raise troubling differ- B. Results Using 2SLS Type Panel Estimators
ences. For examples of conformity, the range of estimates on
A number of alternative instrumental variable estimators
lagged consumption from 0.83 to 0.97 imply strong habit
are designed to deal with the lagged consumption variable.
persistence. Furthermore, the long-run price elasticity is
The simplest is 2SLS, shown in table 1B, which differs from
large, ranging from 2 1.79 to 2 2.98. On the other hand,
OLS only in that it assumes lagged consumption endog-
estimates of long-run income elasticity range from statisti-
enous. Comparison of the OLS results in table 1A with the
cally signiŽ cant values of 2 1.00 to 1 0.60. An explanation
2SLS results in table 1B show a substantial drop in the
for the anomalous results for income is that the between
lagged-consumption coefficient from 0.97 to 0.85 and a drop
variation implies that low-income states, particularly in the
in the long-run price elasticity from 2 2.90 to 2 1.37. The
South, are associated with high cigarette consumption.
problem with simple 2SLS is that it does not account for the
Estimators that emphasize between variation are more likely
Zi variables or individual state effects, any of which could
to produce negative income elasticities.
result in omission bias. As illustrated by the 2SLS with Zi
Traditional panel data estimators are subject to simulta-
and Dt results, these variables matter. With their inclusion,
neous equation bias due to the presence of lagged cigarette
the lagged coefficient on consumption drops even further to
consumption. Interestingly, depending on the traditional
0.63 and the long-run own price elasticity is 2 1.31.
estimator chosen, there are alternative ways by which
But even 2SLS with Zi and Dt can yield potentially
simultaneous equation bias can arise. The OLS results are
inconsistent estimates of our model if the Zi’s are an
biased and inconsistent for either of two reasons:
incomplete representation of the µi state effects. In addition,
7 Appendix A reports the estimated coefficients of the Z variables in
i serial correlation of the n it’s will render our estimators
these and subsequent regressions. Among the various racial variables, asymptotically inefficient. Keane and Runkle (1992) (hereaf-
blacks appear to have lower cigarette consumption than do whites. Among
the religious variables, the Mormon ban on smoking seems supported by ter denoted by KR) suggest a modiŽ cation of this 2SLS
the data. Tourism is positively related to cigarette consumption, while estimator that allows for any arbitrary type of serial correla-
education is statistically insigniŽ cant. tion in the n it’s. Applying 2SLS-KR, the lagged coefficient
8 The speciŽ c state and time dummies were tested for their signiŽ cance

jointly as well as separately. Both state and time dummies were signiŽ cant on consumption decreases from 0.85 (for 2SLS) to 0.71, and
with an observed F-statistic of 7.39 and a p-value of 0.0001 under the null the long-run own price effect decreases in absolute value
distribution of F(73, 1256). The observed F-statistic for the signiŽ cance of from 2 1.31 (for 2SLS) to 2 1.07. The Within-2SLS Ž nds an
state dummies (given the existence of time dummies) is 4.16, which has a
p-value of 0.0001 under the null distribution of F(45, 1256). The observed even smaller lagged consumption-coefficient estimate, a
F-statistic for the signiŽ cance of the time dummies (given the existence of higher short-run own price elasticity in absolute value, and
state dummies) is 16.05, which has a p-value of 0.0001 under the null
distribution of F(28, 1256). These F-tests perform well even when the state
effects are random. See Moulton and Randolph (1989). The results of these 9 The Hausman/Taylor (1981) GLS estimator is an instrumental variable

tests emphasize the importance of individual-state and time-period effects estimator with the following set of instruments: [Pit, Yit, Pnit, Pi,t2 1, Yi,t2 1,
in the cigarette-demand equation. Pni,t2 1, Zi ].
 HOMOGENEOUS VERSUS HETEROGENEOUS ESTIMATORS 121

even lower long-run price elasticity. Clearly, completely Even though the Ž rst-difference type estimators may
controlling for state effects results in less-elastic price eliminate bias arising from state effects, it is important to
responses. Comparing the Within-2SLS estimates in table recognize that it does so at a large information cost. First
1B with the Within estimates in table 1A, we note that the differencing the data eliminates the economic structure
Within-2SLS estimator Ž nds an even lower lagged- implied by the levels of the variables within any given time
consumption estimate of 0.60 compared to 0.83, a corre- series. Thus, one worries whether these results implying
sponding lower long-run own price elasticity (2 1.25 versus implausibly weak habits persistence and dramatically lower
2 1.79), and the absence of neighboring state’s price effects. long-run price elasticities are merely the result of a sanitized
The Within estimator potentially suffers from bias in a data set. For this reason, the forecast performance simulation
dynamic model due to the correlation between lagged in section V is instructive.
consumption and n i, where n i is the average of the remainder To summarize, comparing the 2SLS type estimators to
disturbances across time. This bias disappears as T gets their traditional counterparts, the lagged coefficient on
large. However, wiping out the µi’s does not necessarily get consumption ranges between 0.49 to 0.85—much lower
rid of all endogeneity between the predetermined variables than the traditional pooled estimators which were in the
and the disturbances. Within-2SLS corrects for this by using range of 0.83 to 0.97. The long-run own price elasticity
instrumental variables. estimates are much less elastic, ranging between 2 0.68 and
Next, we apply a two-stage least squares procedure that 2 1.37—much lower in absolute value than the traditional
assumes a one-way error-component model, that is, EC2SLS pooled estimates, which were in the range of 2 1.79 to
(see Hsiao (1986)). This method transforms the error by 2 2.98. The long-run neighboring price effect ranges from an
V ˆ 2 1/2, where V ˆ is a consistent estimator of the variance- insigniŽ cant 2 0.04 to a signiŽ cant 0.37 and appears gener-
covariance matrix of the disturbances, and then applies ally lower than their traditional counterparts. For the long-
2SLS using between and within variations in the exogenous run income elasticity, there is less variation in estimates
variables and their lagged values as instruments. This among 2SLS type estimators than their traditional counter-
method yields a lagged coefficient of consumption estimate parts, giving a clear implication that cigarettes are a weakly
of 0.70 and a long-run own price elasticity of 2 1.35. normal good.
Allowing for the possibility of an AR(1) process on the
remainder disturbances (n it in equation (4)) and still preserv- IV. Results Using Heterogeneous Estimators
ing the error-component structure on the disturbances, the
Implicitly, all the pooled estimators in section III assume
EC2SLS-AR(1) procedure yields a lagged coefficient of
homogeneity of the parameters across states.12 More re-
consumption estimate of 0.67 and a long-run own price
cently, the fundamental assumption underlying pooled homo-
elasticity estimate of 2 1.07. Yet another method of control-
geneous parameters models has been called into question
ling for state effects proposed by Anderson and Hsiao (1982)
and alternative heterogeneous estimators has been pro-
amounts to Ž rst differencing the data and then applying
posed. 13 Pesaran and Smith (1995) argue that the dynamic
2SLS using lagged values of the exogenous variables as
pooled model can be biased because of heterogeneity in the
instruments, which is denoted by FD2SLS.10 Still another
parameters across each state. Furthermore, they propose that
variant would be to allow for any arbitrary form of serial
an average of the individual state regressions can lead to
correlation in the Ž rst differenced disturbances in the manner
consistent estimates of the parameters as long as N and T
of Keane/Runkle. This is denoted as the FD2SLS-KR
tend to inŽ nity. Table 2 summarizes the results of the
estimator. Although the FD2SLS and the FD2SLS-KR
individual states regressions and the Pesaran/Smith average
estimates are quite similar, they differ appreciably from
estimates. The individual state regressions yield quite a wide
other 2SLS type estimators. For example, with FD2SLS-
range of long-run price elasticities, ranging from 5.46 to
KR, the lagged coefficient of consumption is lower than that
of Within-2SLS (0.49 compared to 0.60), and the long-run
hypothesis H0; E(µi/set of instruments) 5 0. This is based on the difference
own price elasticity estimate declines to 2 0.68 versus between 2SLS and Ž rst-differenced 2SLS. The former is consistent only if
2 1.25 for Within-2SLS. 11 the µi’s are not correlated with the set of instruments, whereas Ž rst-
differenced 2SLS is consistent regardless. The x 24 statistic is 96.6, which is
also signiŽ cant. This conŽ rms the importance of controlling for these µi
10 More instrumental variables can be obtained if the n ’s are not serially
it effects.
correlated by using predetermined instruments that are not correlated with 12 A Chow-test for the equality of slope coefficients across countries
the error term. See Arellano and Bover (1995) or Ahn and Schmidt (1995). yields an F-value of 2.32. The numerator of this F-statistic is based on the
11
Underlying the instrumental-variable estimation is always the question second-stage regression of 2SLS allowing for varying intercepts and
of whether the instruments are strictly exogenous with respect to the error common slopes under its restricted version, and varying intercepts and
term. In this case, this could be due to correlation with the µi’s or the n it’s. slopes under its unrestricted version. The denominator of this F-statistic is
Keane and Runkle (1992) suggest a Hausman (1978) type test for the strict based on the unrestricted 2SLS residuals sums of squares. (See Wooldridge
exogeneity of the instruments based on the difference between Within- (1990).) Under the null hypothesis, this is distributed as F(180, 1104). The
2SLS and Ž rst-differenced 2SLS. The latter estimator is consistent whether observed F-statistic has a p-value of 0.0001, and the null is rejected.
the instruments are predetermined or strictly exogenous with respect to the 13 See, for example, Robertson and Symons (1992), who warned about

error term, whereas the former estimator is consistent only if the the bias obtained from panel data methods when the estimated model is
instruments are strictly exogenous. The x 24 statistic obtained is 118.6, dynamic and homogeneous when in fact the true model is static and
which is signiŽ cant. Keane and Runkle also suggest testing for the null heterogeneous.
122 THE REVIEW OF ECONOMICS AND STATISTICS

TABLE 2.—CIGARETTE DEMAND: HETEROGENEOUS PARAMETER ESTIMATES 1963 –1992

Short Run Long Run

ln Cit2 l ln Pit ln Pnit ln Yit ln Pit ln Pnit ln Yit

Individual Countries OLS:

Maximum 1.05 2 0.85 0.71 0.15 2 8.46 4.74 0.81
(6.67) (2.78) (3.64) (1.07) (0.31) (0.28) (0.13)
Median 0.69 2 0.24 0.06 2 0.14 2 0.57 0.10 2 0.38
(4.56) (1.86) (0.47) (1.07) (2.45) (0.24) (4.11)
Minimum 0.03 0.44 2 0.61 2 0.54 5.46 2 7.38 2 3.37
(0.15) (0.1) (1.90) (1.71) (0.56) (0.62) (0.35)
Average 0.73 2 0.28 0.09 2 0.14 2 0.79 0.25 2 0.39
(4.75) (10.1) (3.31) (8.24) (9.84) (3.18) (9.25)
Shrinkage OLS:
Maximum 0.98 2 0.16 0.09 2 0.01 2 4.84 2.62 2 0.17
(42.2) (5.70) (3.23) (0.88) (1.44) (1.16) (1.34)
Median 0.94 2 0.11 0.03 2 0.04 2 1.76 0.42 2 0.68
(46.0) (3.60) (1.00) (2.41) (2.54) (1.17) (1.74)
Minimum 0.88 2 0.04 2 0.04 2 0.08 2 0.74 2 0.71 2 1.61
(37.1) (1.32) (1.32) (5.13) (1.45) (1.13) (2.62)
Individual Countries 2SLS:
Maximum 3.35 2 1.07 0.70 1.22 2 7.46 6.53 0.38
(0.80) (2.02) (3.51) (0.54) (0.58) (0.54) (3.75)
Median 0.60 2 0.34 0.11 2 0.11 2 0.55 0.08 2 0.39
(1.35) (2.72) (0.84) (0.75) (0.97) (0.43) (5.70)
Minimum 2 0.31 0.52 2 0.55 2 0.66 5.23 2 5.54 2 1.88
(1.06) (1.24) (1.51) (3.25) (0.20) (0.24) (0.73)
Average 0.65 2 0.32 0.15 2 0.05 2 1.17 0.54 2 0.16
(0.83) (7.42) (3.36) (0.68) (2.28) (2.02) (0.68)
Shrinkage 2SLS:
Maximum 1.61 2 0.34 0.20 0.30 2 6.69 4.97 4.94
(10.5) (3.83) (2.86) (3.55) (0.94) (0.77) (0.04)
Median 0.88 2 0.14 0.04 2 0.05 2 0.85 0.18 2 0.47
(9.34) (1.71) (0.70) (0.76) (2.20) (0.13) (2.80)
Minimum 0.64 0.08 2 0.13 2 0.20 13.78 2 5.80 2 1.95
(4.34) (1.00) (1.78) (2.77) (0.04) (0.49) (1.06)
Numbers in parentheses denote t-statistics.

2 8.46 for OLS and 5.23 to 2 7.46 for 2SLS. A wide range of In summary, the heterogeneous estimators (the individual
long-run estimates is also apparent for the neighboring price 2SLS, the Pesaran/Smith average, and the Maddala et al.
and income elasticities. Pesaran and Smith’s suggestion of Shrinkage estimator) have the desirable property of allowing
using a simple average of the individual state estimates to for differences among states, but the range of individual
obtain long-run elasticity estimates implies that the long-run 2SLS estimates suggests that the individual state estimates
2SLS elasticities are 2 1.17 for own price, 0.54 for neighbor- are highly unstable and unreliable. Indeed, the instability of
ing price, and 2 0.16 for income—while the corresponding parameter estimates from individual time series has been
average OLS elasticities are, respectively, 2 0.79 for own observed quite commonly in a variety of demand studies, 14
price, 0.25 for neighboring price, and 2 0.39 for income. providing a major argument for pooling. The Pesaran/Smith
Using a quite different approach, Maddala et al. (1994) suggestion of using a simple average offers an alternative to
argue that Shrinkage estimators are superior to either the the homogeneous estimators. Likewise, the Shrinkage esti-
individual state (heterogeneous) estimates or the pooled
mator seems to provide a smaller and more plausible range
(homogeneous) estimates especially for prediction purposes.
of estimates than do individual time-series estimates. Never-
In this case, one shrinks the individual estimates towards the
theless, ranges of Shrinkage 2SLS estimates for the long-run
pooled estimate using weights depending on their correspond-
ing variance-covariance matrices. The shrinkage estimator price elasticity of 2 6.69 to 13.78 are implausible. In
suggested by Maddala et al. substantially reduces the wide contrast, the pooled estimators, which implicitly posit homo-
dispersion of estimates found in the individual countries, geneous coefficients, appear to provide, on balance, much
which demonstrates that the effect of pulling individual more plausible estimates than their heterogeneous counter-
estimates towards a common mean profoundly affects the parts.
estimates. The OLS Shrinkage estimates of the long-run
price elasticity range from 2 0.75 to 2 4.84, while those of
2SLS range from 13.78 to 2 6.69. Considerable variation in 14 For examples, see studies of gasoline demand (Baltagi and Griffin
the long-run estimates of neighboring price and income (1983)), natural gas (Balestra and Nerlove (1966)), and electricity (Taylor
elasticities are also apparent. (1975)).
 HOMOGENEOUS VERSUS HETEROGENEOUS ESTIMATORS 123

TABLE 3.—COMPARISON OF FORECAST PERFORMANCE

1st Year 5th Year 10th Year 10-Year Average
Ranking Estimator RMSE Estimator RMSE Estimator RMSE Estimator RMSE

1. FD2SLS-KR 2.80 Within-2SLS 9.29 Within-2SLS 14.08 Within-2SLS 9.93

2. FD2SLS 2.80 Within 11.57 OLS 16.11 OLS 11.80
3. Within 3.20 OLS 11.75 Within 17.21 Within 12.11
4. 2SLS 3.32 GLS 12.73 GLS 19.14 GLS 13.20
5. EC2SLS-AR(1) 3.54 FD2SLS-KR 12.81 GLS-AR(1) 23.65 FD2SLS-KR 14.44
6. Within-2SLS 3.60 FD2SLS 12.92 FD2SLS-KR 24.03 FD2SLS 14.70
7. Average OLS 3.68 Shrinkage 2SLS 13.67 Shrinkage 2SLS 24.15 GLS-AR(1) 15.04
8. OLS 3.68 GLS-AR(1) 13.89 FDSLS 24.84 Shrinkage 2SLS 15.09
9. Shrinkage OLS 3.71 Shrinkage OLS 15.02 2SLS 26.66 2SLS 16.92
10. GLS 3.73 2SLS 15.40 Shrinkage OLS 28.82 Shrinkage OLS 16.93
11. Average 2SLS 3.79 Average OLS 15.47 Average OLS 29.46 Individual OLS 16.93
12. Shrinkage 2SLS 3.88 Average 2SLS 15.49 Average 2SLS 29.66 Average OLS 17.39
13. GLS-AR(1) 3.95 EC2SLS-AR(1) 16.47 EC2SLS-AR(1) 30.10 Average 2SLS 17.45
14. Individual OLS 4.04 No Change 16.59 No Change 32.54 EC2SLS-AR(1) 18.13
15. No Change 4.50 Individual OLS 18.08 Individual OLS 34.75 No Change 19.05
16. Individual 2SLS 4.89 Individual 2SLS 27.17 Individual 2SLS 130.8 Individual 2SLS 43.84
RMSE 3 10 2.

V. Comparison of Forecasts heterogeneous estimators, individual OLS and individual

2SLS perform uniformly poorly vis-a-vis both the homoge-
An important and frequently underutilized criteria is the
neous and other heterogeneous estimators. Indeed, for the
forecast properties of alternative estimators. In this section,
Ž fth- and tenth-year forecasts, individual OLS and indi-
we use the prediction-performance criteria to help us choose
vidual 2SLS have the distinction of being the only estima-
among alternative estimators with quite disparate price
tors to perform worse than the naive ‘‘no change’’ model.
elasticity implications. Given the large data set of 46 states
Such a Ž nding may seem counterintuitive because, by
for over thirty years, we estimate the above models for
deŽ nition, individual state regressions give the best Ž t over
truncated data sets and then apply each model to an
the sample period. Even in out-of-sample forecasts, it would
out-of-sample forecast period. Because the estimators imply
seem that individual state regressions would out-perform
very different long-run elasticities, it seems particularly
homogeneous panel estimators because the latter place
useful to contrast their forecast performance over time
relatively little weight on the data for any one state. Why
periods as far out as ten years.
then, particularly in Ž ve- and ten-year forecasts, do indi-
Table 3 gives a comparison of various predictors obtained
vidual state regressions perform so poorly?
by estimating the model without the last ten years of data
The explanation for this seeming paradox is that indi-
and then applying these models to out-of-sample forecasts of
vidual state regressions are highly unstable between the
cigarette consumption over these ten years. The root mean
twenty- and thirty-year samples, especially as they relate to
square errors (RMSEs) are calculated for the resulting
long-run elasticities. Table 4 reports the absolute difference
predictions. Because the ability of an estimator to character-
in the implied short- and long-run elasticities for the various
ize long-run as well as short-run responses is at issue, the
estimators. We use absolute instead of actual differences
average RMSE is calculated across the 46 states at different
because, for individual state regressions, we compute the
forecast horizons. SpeciŽ cally, each model was applied to
absolute difference in the parameter estimates and then
each state, and out-of-sample forecasts for ten years were
average it across states to provide a meaningful basis for
calculated. The relative forecast rankings are reported in
comparison. Table 4 reports the various estimators ranked by
table 3 after one year, Ž ve years, and ten years.15 The overall
the absolute change in the long-run price elasticity. Clearly,
average ranking for the full ten-year period is also reported.
individual OLS and 2SLS yield highly unstable long-run
To provide some relative comparison basis, we also include
estimates, explaining why they do so poorly in ten-year
a naive model that sequentially forecasted ‘‘no change’’ in a
forecasts. In contrast, the strength of the homogeneous panel
state’s cigarette consumption per capita. In comparing the
estimators as a group is their relative stability. Thus, as has
relative performance of the various estimators, one can
been widely conŽ rmed in other contexts, 16 there are real
analyze these results from the following perspectives: heter-
gains from pooling particularly when the time series is
ogeneous versus homogeneous, and one-year-ahead fore-
relatively short.
casts versus long-run forecasts.
The performance by the other heterogeneous estimators is
A comparison of the heterogeneous versus homogeneous
mixed. The Pesaran/Smith ‘‘Average OLS’’ and ‘‘Average
estimators reveals some very interesting patterns. Of the
15 These predictions were intercept-adjusted for each state. Additionally, 16 See Balestra and Nerlove (1966), Taylor (1975), and Baltagi and

all estimators have zero forecast errors for 1982. Griffin (1997).
124 THE REVIEW OF ECONOMICS AND STATISTICS

TABLE 4.—COMPARISON OF ABSOLUTE DIFFERENCES IN PARAMETER ESTIMATES: 20- VS. 30-YEAR SAMPLE
Short-Run Difference Long-Run Difference
Ranking* Estimator ln Pit ln Pnit ln Yit ln Pit ln Pnit ln Yit

1. FD2SLS-KR 0.04 0.11 0.05 0.06 0.14 0.12

2. Within-2SLS 0.06 0.19 0.10 0.07 0.23 0.21
3. GLS 0.03 0.03 0.00 0.16 0.54 0.12
4. Within 0.11 0.06 0.09 0.17 0.15 0.13
5. FD2SLS 0.04 0.10 0.01 0.18 0.06 0.09
6. Average OLS 0.03 0.18 0.11 0.28 0.20 0.35
7. 2SLS 0.14 0.17 0.11 0.34 0.31 0.38
8. OLS 0.03 0.05 0.03 0.38 0.83 1.00
9. EC2SLS-AR(1) 0.09 0.17 0.08 0.59 0.14 0.25
10. GLS-AR(1) 0.04 0.06 0.03 0.59 0.85 1.11
11. Average 2SLS 0.03 0.13 0.01 0.70 0.10 0.07
12. Individual OLS 0.20 0.24 0.21 1.12 1.15 0.59
13. Individual 2SLS 0.24 0.25 0.33 1.13 1.29 0.73
14. Shrinkage OLS 0.12 0.15 0.10 1.24 0.41 0.74
15. Shrinkage 2SLS 0.06 0.06 0.07 4.50 3.30 0.71
*This ranks the estimators by the absolute difference in the long-run price elasticity.

2SLS’’ estimators rank seventh and eleventh, respectively in is theoretically capable of dealing with a wider variety of
one-year-ahead forecasts. Yet, for the Ž ve- and ten-year potential biases (such as the Ž rst-difference 2SLS model
forecasts, they have dropped to eleventh and twelfth and with Keane/Runkle adjustment (FD2SLS-KR)), it does not
overall rank twelfth and thirteenth. The deteriorating perfor- necessarily provide the best forecasts. For example, FD2SLS
mance of the Pesaran/Smith average estimator arises be- and FD2SLS-KR, which control for state effects and endoge-
cause of the parameter-instability problem of the individual neity, deteriorate sharply as the forecast horizon is extended
state regressions shown in table 4. The shrinkage 2SLS beyond one year.17 The 2SLS type estimators depend
estimator appears to have the reverse problem of the critically on the quality of the instruments. One measure of
‘‘Average’’ estimator. For one-year forecasts, shrinkage the quality of these instruments is the R2 from the Ž rst stage
2SLS ranks twelfth, yet, for the Ž ve- and ten-year forecasts, of 2SLS. For 2SLS, the set of instruments is Pit, Pnit, Yit, and
it ranks seventh. (Overall, it ranks eighth.) Shrinkage OLS their lagged values. The R2 for the Ž rst stage of 2SLS is 0.31.
ranks ninth for one-year and Ž ve-year forecasts, tenth for the For Within-2SLS written as 2SLS with time and state
ten-year forecast, and tenth overall. The relatively poor dummy variables, this R2 rises to 0.89. Note that FD2SLS
performance of the shrinkage estimators can be attributed to has a different dependent variable, but the R2 for the Ž rst
its reliance upon the individual state parameter estimates. stage of FD2SLS is 0.26. Not surprisingly, Within-2SLS
Thus, what seemed an advantage to the shrinkage estimator— performs much better in forecast applications than do the
that is, placing some weight on the individual state regres- other 2SLS type estimators. Interestingly, the more tradi-
sions—becomes a liability when parameter instability is tional Within, OLS, and GLS estimators also perform quite
severe. While the coefficients of the short-run estimates were well, ranking second, third, and fourth, respectively, overall
relatively stable, the extreme long-run parameter instability in forecasting performance and yet offer no correction for
shown in table 4 could be attributed to three states for which endogeneity.
the lagged dependent-variable coefficient was close to 1. The traditional pooled estimators (OLS, Within, and
The overall RMSE forecast rankings offer a strong GLS) results of table 3 are comforting as they systematically
endorsement for the homogeneous estimators due in large perform well in forecasting for Ž ve years or longer. Further-
part to their parameter stability. Within-2SLS ranks Ž rst, more, it is noteworthy that OLS (which is potentially most
followed by OLS, Within, GLS, and FD2SLS estimators. In impacted by bias) is second only to the Within-2SLS for the
fact, the top six estimators for one-, Ž ve-, ten- and average ten-year average forecast performance. Our explanation for
ten-years forecasts are homogeneous parameter estimators. this seeming paradox is that, while OLS may be more
The Ž nding that Within-2SLS gives the lowest RMSE for susceptible to bias, it relies heavily upon between state
Ž ve-, ten-, and overall ten-year average suggests that variation. If one accepts the notion that the between
controlling for state effects and endogeneity are important. It variation tends to capture long-run responses, it should not
also provides some support for a less elastic long-run price be surprising that it would perform well in long-term
elasticity in the range of 2 1.25.
The forecast results provide comfort to researchers apply- 17 FD2SLS-KR requires estimation of the covariance matrix of error

ing homogeneous panel estimators, both as indicated by the terms, which is 20 3 20 for the data used for forecasting and 30 3 30 for
sizeable range of parameter estimates in table 1 and forecast the entire sample. It is unlikely that 46 cross-section observations will
generate a reliable estimate for a matrix of this size. As one of the referees
performance in table 3. But the choice of which estimator suggest, this might explain why FD2SLS-KR often performs worse than
becomes paramount. Simply because a particular estimator Within-2SLS does.
 HOMOGENEOUS VERSUS HETEROGENEOUS ESTIMATORS 125

forecasts. Thus, the researcher must keep in mind the j The Within-2SLS estimator, which Ž nds a long-run
potential trade-off between bias and the source of variation price elasticity of 2 1.25, performs best in forecast
affecting a given estimator. By the same argument, it should exercises by a signiŽ cant margin and would seem to be
not be surprising that FD2SLS type estimators perform well the preferred homogeneous estimator. We are, how-
in one-year forecasts, yet fail to capture long-term re- ever, skeptical about labelling Within-2SLS as the
sponses. ‘‘preferred’’ panel estimator, given its relatively poor
The Ž nding that the traditional pooled estimators (OLS, performance in our gasoline-demand study.
Within, and GLS) performed relatively well in this applica- j The gain from 2SLS depends critically on the quality
tion echoes a similar Ž nding in Baltagi and Griffin (1997) for of the 2SLS instruments. For cigarettes, while Within-
an international panel of gasoline consumption. In that 2SLS was the top forecast performer, other 2SLS
study, these three estimators were in the top four when estimators performed poorly. For gasoline, all of the
ranked by ten-year average forecast performance. On the 2SLS estimators performed poorly.
other hand, they found that heterogeneous estimators did not j In both studies, three of the top four forecast perform-
fare particularly well in long-run forecast applications. ers were OLS, GLS, and Within. This Ž nding should
Turning to the important question of the long-run price give comfort to applied researchers who routinely
elasticity, we note that for both the Ž ve- and ten-year employ the traditional pooled estimators.
forecasts, Within-2SLS (which implies a long-run price
elasticity of 2 1.25) signiŽ cantly outperforms the next-best
estimators.18 This result is comforting both intuitively and REFERENCES
econometrically. The Within-2SLS estimator avoids the Ahn, Seung C., and Peter Schmidt, ‘‘Efficient Estimation of Models for
problem of endogeneity of the lagged-consumption term and Dynamic Panel Data,’’ Journal of Econometrics 68 (July 1995),
the heterogeneity arising from intertemporal and interstate 5–27.
Anderson, T. W., and Cheng Hsiao, ‘‘Formulation and Estimation of
effects. Dynamic Models Using Panel Data,’’ Journal of Econometrics 18
(January 1982), 47–82.
Arellano, Manuel, and Olympia Bover, ‘‘Another Look at the Instrumental
VI. Summary and Conclusions Variables Estimation of Error-Component Models,’’ Journal of
Econometrics 68 (July 1995), 29–51.
Like the Baltagi and Griffin (1997) international gasoline- Balestra, Pietro, and Marc Nerlove, ‘‘Pooling Cross-Section and Time-
consumption panel, the results of this paper conŽ rm the Series Data in the Estimation of a Dynamic Model: The Demand for
Natural Gas,’’ Econometrica 34 (July 1966), 585–612.
value of panel data sets and justify the emphasis given to Baltagi, Badi H., and James M. Griffin, ‘‘Gasoline Demand in the OECD:
pooled estimators. Even with a relatively long time series, An Application of Pooling and Testing Procedures,’’ European
heterogeneous models for individual states tend to produce Economic Review 22 (July 1983), 117–137.
——— ‘‘Pooled Estimators vs. Their Heterogeneous Counterparts in the
implausible estimates with inferior forecasting properties. Context of Dynamic Demand for Gasoline,’’ Journal of Economet-
Our explanation for why pooled models—which posit rics 77 (April 1997), 303–327.
homogeneous parameters—outperform their heterogeneous Baltagi, Badi H., and Dan Levin, ‘‘Estimating Dynamic Demand for
Cigarettes Using Panel Data: The Effects of Bootlegging, Taxation
counterparts centers on the relative variability of the data and Advertising Reconsidered,’’ this REVIEW 48 (February 1986),
between individual time series and panels. Using an RMSE 148–155.
criterion, the efficiency gains from pooling appear to more Baltagi, Badi H., and Qi Li, ‘‘A Transformation That Will Circumvent the
Problem of Autocorrelation in An Error Component Model,’’
than offset the biases due to interstate heterogeneities. Journal of Econometrics 48 (June 1991), 385–393.
Interestingly, even with thirty years of data and dynamic Christ, Carl F., Econometric Models and Methods (New York: John Wiley
demand goods with substantial lags, individual state time and Sons, 1966).
Doron, Gideon, The Smoking Paradox: Public Regulation in the Cigarette
series perform poorly. Even panel estimators that posit Industry (Cambridge, MA: Abt Books, 1979).
heterogeneous coefficients such as the Pesaran/Smith aver- Hamilton, James L., ‘‘The Demand for Cigarettes: Advertising, the Health
age estimator and Maddala’s shrinkage estimator seem to be Scare, and the Cigarette Advertising Ban,’’ this REVIEW 54 (Novem-
ber 1972), 401–411.
inferior in forecasting performance when compared to the Hausman, Jerry A., ‘‘SpeciŽ cation Tests in Econometrics,’’ Econometrica
traditional pooled estimators. 46 (November 1978), 1251–1271.
While Ž nding rather clear evidence on the question to Hausman, Jerry A., and William E. Taylor, ‘‘Panel Data and Unobservable
Individual Effects,’’ Econometrica 49 (November 1981), 1377–
pool or not, our answer as to which pooled estimator is more 1398.
problematic. With long-run price elasticities ranging from Houthakker, Hendrik S., and Lester D. Taylor, Consumer Demand in the
2 0.68 to 2 2.98, the choice among homogeneous estimators United States: Analysis and Projections (Cambridge: Harvard
University Press, 1970).
clearly matters. We offer the following tentative Ž ndings. Hsiao, Cheng, Analysis of Panel Data (Cambridge: Cambridge University
Press, 1986).
Johnson, Douglas W., Paul R. Picard, and Bernard Quinn, Churches and
18 Interestingly, the next-closest rivals, OLS and Within, give relatively Church Membership in the United States (1971) (Washington,
similar forecast performance over the ten-year period. Yet, they imply very D.C.: Glenmary Research Center, 1974).
different long-run price elasticities (2.90 versus 2 1.79) and income Keane, Michael P., and David E. Runkle, ‘‘On the Estimation of
elasticities (2 1.0 and 0.60). Christ (1966) pointed out that, in the context Panel-Data Models with Serial Correlation When Instruments Are
of macroeconometric models, models with very different implied short- not Strictly Exogenous,’’ Journal of Business and Economics
and long-run dynamics can yield similar RMSEs in forecast applications. Statistics 10 (January 1992), 1–9.
126 THE REVIEW OF ECONOMICS AND STATISTICS

Kiviet, Jan F., ‘‘On Bias, Inconsistency and Efficiency of Some Estimators Pesaran, M. Hashem, and Ron Smith, ‘‘Estimating Long-Run Relation-
in Dynamic Panel Data Models,’’ Journal of Econometrics 68 (July ships From Dynamic Heterogeneous Panels,’’ Journal of Economet-
1995), 53–78. rics 68 (July 1995), 79–113.
Laughhunn, Dan J., and Herbert L. Lyon, ‘‘The Feasibility of Tax-Induced Robertson, Donald, and James Symons, ‘‘Some Strange Properties of
Price Increases as a Deterrent to Cigarette Consumption,’’ Journal Panel Data Estimators,’’ Journal of Applied Econometrics 7 (April/
of Business Administration 3 (Fall 1971), 27–35. June 1992), 175–189.
Maddala, G. S., V. K. Srivastava, and Hao Li, ‘‘Shrinkage Estimators for
the Estimation of Short-Run and Long-Run Parameters From Panel Taylor, Lester D., ‘‘The Demand for Electricity: A Survey,’’ Bell Journal 6
Data Models,’’ working paper, Ohio State University (1994). (1975), 74–110.
Manchester, Paul B., ‘‘Interstate Cigarette Smuggling,’’ Public Finance The Tobacco Institute, ‘‘The Tax Burden on Tobacco’’ (Washington, D.C.,
Quarterly 4 (1976), 225–238. 1993).
McGuiness, Tony, and Keith Cowling, ‘‘Advertising and the Aggregate U.S. Department of Health and Human Services, ‘‘Reducing the Health
Demand for Cigarettes,’’ European Economic Review 6 (1975), Consequences of Smoking: 25 Years of Progress,’’ A Report of the
311–328. Surgeon General. Public Health Service, Office of Smoking and
Moulton, Brent R., ‘‘Random Group Effects and the Precision of Regres- Health, Washington, DC. (1989).
sion Estimates,’’ Journal of Econometrics 32 (August 1986),
385–397. Warner, Kenneth E., ‘‘Cigarette Excise Taxation and Interstate Smuggling:
Moulton, Brent R., and William C. Randolph, ‘‘Alternative Tests of the An Assessment of Recent Activity,’’ National Tax Journal 35
Error Components Model,’’ Econometrica 57 (May 1989), 685– (1982), 483–489.
693. Wooldridge, Jeffrey M., ‘‘A Note on the Lagrange Multiplier and
Nickell, Stephen, ‘‘Biases in Dynamic Models with Fixed Effects,’’ F-Statistics for Two Stage Least Squares Regressions,’’ Economics
Econometrica 49 (November 1981), 1417–1426. Letters 34 (1990), 151–155.

APPENDIX A—ESTIMATED COEFFICIENTS OF ZI IN VARIOUS MODELS

Model OLS GLS GLS-AR(1) 2SLS 2SLS-KR EC2SLS EC2SLS-AR(1)

% Asian 2 0.0057 2 0.0049 2 0.0013 2 0.0054 2 0.0029 2 0.0133 2 0.0103

(2.8) (2.5) (0.9) (2.0) (0.4) (5.1) (1.2)
%Indian 1 Eskimo 0.0030 2 0.0030 2 0.0022 2 0.0081 2 0.0085 2 0.0074 2 0.0083
(3.7) (3.7) (3.7) (5.3) (2.5) (6.5) (2.3)
% Black 2 0.0031 2 0.0003 2 0.0002 0.0001 0.0005 2 0.0003 0.0002
(2.6) (2.4) (2.5) (0.8) (0.9) (2.3) (0.4)
% Hispanic 2 0.0006 2 0.0005 2 0.0003 2 0.0010 2 0.0025 2 0.0014 2 0.0016
(1.9) (1.7) (1.2) (2.2) (2.0) (3.7) (1.1)
% Mormon 2 0.0006 2 0.0006 2 0.0004 2 0.0020 2 0.0032 2 0.0018 2 0.0023
(4.1) (4.0) (3.3) (5.6) (5.0) (7.5) (3.4)
% Catholic 2 0.0002 0.0002 0.0001 0.0004 0.0008 0.0007 0.0005
(1.1) (1.0) (0.5) (1.6) (1.2) (3.3) (0.6)
% South Baptist 0.0003 0.0002 0.0002 2 0.0008 2 0.0019 0.0003 2 0.0007
(1.5) (1.2) (1.1) (2.7) (2.5) (1.4) (0.8)
% Methodist 0.0013 0.0004 2 0.0002 2 0.0014 2 0.0022 2 0.0002 2 0.0004
(0.2) (0.1) (0.5) (1.7) (0.9) (0.2) (0.1)
% Church of Christ 2 0.014 2 0.0012 2 0.0003 2 0.0013 2 0.0039 2 0.0040 2 0.0034
(1.2) (1.0) (0.3) (0.9) (0.8) (2.9) (0.7)
% Episcopalian 2 0.0044 2 0.0037 2 0.0006 2 0.0019 2 0.0057 2 0.0094 2 0.0064
(2.7) (2.3) (0.5) (0.9) (0.9) (4.6) (1.0)
% Lutheran 0.0019 2 0.0018 2 0.0006 2 0.0045 2 0.0063 2 0.0055 2 0.0063
(3.9) (3.6) (1.8) (4.7) (3.3) (7.1) (3.0)
% Jewish 0.0009 0.0008 0.0002 0.0023 0.0055 0.0033 0.0032
(1.4) (1.9) (0.3) (2.3) (1.9) (3.8) (1.1)
% Education 0.0003 0.0005 0.0006 0.0049 0.0005 2 0.0001 0.0047
(0.1) (0.1) (0.2) (1.0) (0.03) (0.03) (0.3)
Tourism 0.0007 0.0006 0.0002 0.0020 0.0030 0.0024 0.0027
(3.5) (3.2) (1.7) (4.9) (3.8) (7.3) (3.3)
All models in this table have time dummies.