Alternative Panel Estimates of Alcohol Demand, Taxation, and the Business Cycle.

• Author: Freeman, Donald G.
• Position: Statistical Data Included

Donald G. Freeman [*]

This paper uses a new technique of estimating dynamic heterogeneous panels developed by Pesaran, Shin, and Smith (1999) on state-level alcohol consumption as a function of income, taxes, and cyclical variables. Pooled mean group (PMG) estimators provide an alternative to extremes of pooling the data assuming slope homogeneity and estimating individual states assuming complete heterogeneity. Postsample tests indicate that a conventional fixed-effects model outperforms both the PMG estimator and individual state estimators, despite the heterogeneity of the sample. Current levels of taxation appear to have little effect on alcohol consumption, and alcohol is found to be a procyclical good.

1. Introduction

   Perhaps no other commonly used substance has as controversial a legacy as alcohol. Even those who drink occasionally or moderately may place themselves at risk for adverse events, and the cost of heavy drinking can be high. Alcohol is involved in one-third of the 40,000 traffic fatalities annually in the United States; of the 13,693 alcohol-related traffic deaths in 1994, 4265 were passengers or nonoccupants of the vehicle driven by the drunk driver. A recent study by the National Institutes of Health (NIH) (U.S. Department of Health and Human Services 1998) estimated the annual costs of alcohol abuse at $148 billion; of this amount, the NIH estimates that 45% is borne by abusers and members of their households. The remainder is borne by federal, state, and local governments, by insurance carriers, and by victims of abuse such as those injured or killed in alcohol-related accidents. Thus the majority of the costs of alcohol abuse are external to the user or his or her immediate family, and the control of alc ohol abuse is an appropriate public policy concern.

   The vast majority of alcohol users suffer no ill effects, however, and unlike tobacco, another product subject to much debate, alcohol appears to have health benefits. Used in moderation, alcohol acts as a relaxant, a social lubricant, and even, as some studies have concluded, a prophylaxis against some forms of heart disease (U.S. Department of Health and Human Services 1993). Alcohol is also a significant source of tax revenues; excise and license taxes in 1997 totaled just over $12 billion for the states and the federal government.

   Policies to mitigate the social costs of alcohol usually involve some combination of education, sanctions (such as penalties for drunk driving and minimum-age drinking laws), and taxation. Optimal taxation of alcoholic beverages is an especially complex issue, involving considerations of revenue, deterrence, and potential regressivity (Lyon and Schwab 1995), and virtually all aspects of alcohol taxation are subject to contentious debate (see especially Grossman et al. 1993, and the ensuing exchange between Grossman et al. 1995 and Heien 1995), many of which are beyond the scope of this paper.

   The present analysis uses aggregate data to investigate alcohol taxation primarily as a source of revenue generation. For revenue considerations, microsurvey data have troubling issues with persistent underreporting of alcohol consumption and may not be reliable guides for state or federal revenue estimates (Pogue and Sgontz 1989). Aggregate data also have limitations, however, due to the potential for cross-border shipping from low- to high-tax states, gift-giving and the like, and differing levels of abstention across states. As Leung and Phelps (1993) point out, it is virtually impossible, barring strong assumptions on the distributional effects of taxes on categories of drinkers, to determine from aggregate data whether tax-induced increases in alcoholic beverage prices have significant effects on problem drinkers. Also, states with high levels of abstention may report lower average levels of consumption, even though the number of problem drinkers in those states may be no fewer than states with lower le vels of abstention.

   Reliable estimates of alcohol's demand elasticities, upon which optimal tax rates rely, remain elusive. Early surveys by Orustein (1980) and Ornstein and Levy (1983) and more recent work by Leung and Phelps (1993) and Blake and Nied (1997) produce estimates of price elasticities ranging from -1.39 to +0.22 for beer, -2.94 to -0.05 for spirits, and -2.67 to -0.13 for wine. Estimates of income elasticities produce similar variation, depending on choice of econometric technique, sample size and period, and use of aggregate versus survey data.

   This paper uses alternative panel estimation techniques on an extensive panel of alcohol consumption for the 50 states and the District of Columbia (referred to as the 51st state) spanning the years 1961-1995 to provide new estimates of alcohol's tax and income elasticities. Panel estimation offers several potential advantages over studies using only state cross-sections or national time-series aggregates. In the general case, the additional degrees of freedom available in panel data provide more efficient estimation, and the additional variability in the regressors can mitigate collinearity problems that often afflict macroeconomic data.

   In the particular case of alcohol, panel estimation can capture the legal, regulatory, safety, and health concerns that potentially affect consumer perceptions of alcohol over time. Estimation using national-level aggregate time series is likely to be misspecified by the omission of changing consumer tastes, national legislation, and other time-varying factors; and state cross-sections cannot control for state-specific effects such as regulations on times and locations of alcohol, proportion of religious denominations that eschew alcohol, and other factors that have been associated with variations in alcohol consumption (Baltagi and Griffin 1995). Panel estimation allows controls for both time-varying and state-specific effects that individual time series or cross-section estimates cannot address.

   One question that often arises with pooled estimation of panel data, however, is the validity of the restriction of identical slope coefficients. This restriction is often rejected when tested (Pesaran, Shin, and Smith 1999; Vahid 1998). In these cases, the data are often pooled anyway on the grounds that the gain in precision from pooling heterogeneous units is greater than the cost of potential bias (Baltagi and Griffin 1995). Other approaches to coping with heterogeneous panels include the "shrinkage" estimators of Maddala (1977), or averaging of individual slope coefficients (Pesaran and Smith 1995).

   Pesaran, Shin, and Smith (1999) have developed a new approach, the pooled mean group (PMG) estimator, for addressing the problem of heterogeneity in dynamic panels. The PMG estimator combines pooling and averaging to take an intermediate approach to panel estimation by assuming common long-run responses to demand determinants across states, but allowing short-run responses to vary because of state-specific factors. The PMG approach appears to be a good candidate for estimating alcohol demand. Long-term habit formation and spending characteristics may be identical across most states (especially if mobility is high), but short-run responses may be influenced by local laws or customs, which may themselves change over time, albeit at different rates across states. Previous research examining alcohol demand across jurisdictions have indicated that the null hypothesis of slope homogeneity can be rejected (Johnson et al. 1992; Baltagi and Griffin 1995).

   This paper compares the PMG estimator in postsample tests with two competitor models: a dynamic fixed effect (DFE) model that assumes slope homogeneity and intercept heterogeneity, and the mean group (MG) estimator of Pesaran and Smith (1995) that averages heterogeneous slopes and intercepts across states. Surprisingly, given the strong rejection of slope homogeneity for the full sample, the DFE estimator outperforms the other models in tests of mean-square forecast error and forecast efficiency.

   The paper also examines the influence of the business cycle on alcohol use. There are opposing theories of alcohol's cyclicality. Psychological theories of substance abuse argue that recessions and unemployment induce higher alcohol consumption in response to financial and emotional stress (Horwitz 1984). Arguments for the opposite view of reduced alcohol use during recessions include income considerations, lower employment levels, and reduced opportunities for social interaction, including business travel, tourism, entertaining, and so on.

   Besides the finding of superior predictive performance for the DFE estimator, the general results indicate that past increases in state excise taxes, controlling for income, have had very little permanent effect on alcohol consumption, suggesting that there is considerable room for revenue generation from additional taxes. Estimates of income elasticities confirm that alcohol is income inelastic, both in the short and in the long run. Also, the analysis provides weak evidence that alcohol consumption is procyclical; the conventional view that people drink more during recessions to counteract stress is not supported by aggregate data on consumption.

   The paper proceeds as follows. Section 2 offers a review of previous panel studies of alcohol demand, with discussion of the issue of pooling versus nonpooling of panel data. Section 3 provides a description of the data and an explication of the PMG estimator. Section 4 discusses special issues arising in the context of panel estimation, section 5 features the empirical results, including postsample tests of prediction, and section 6 concludes with policy recommendations.

2. Review of Literature

   There is an extensive literature estimating elasticities of alcohol demand, much of which is summarized in the surveys by Ornstein and Levy (1983) and Leung and Phelps (1993) cited earlier. For brevity, we...