Alcohol prices, consumption, and traffic fatalities.

• Author: Young, Douglas J.

1. Introduction

   Traffic fatalities are a leading cause of premature death, particularly among people under 35 years of age. (1) Alcohol is involved in about 40% of all traffic fatalities, and hence alcohol policies have the potential to significantly reduce fatality rates. (2)

   This paper focuses on the links between alcohol prices, consumption, and traffic fatalities. The fundamental question is, How does the price of alcohol affect fatalities? Price is an important policy variable, since it is affected by taxes and other measures, and for some beverages in some states, price is actually set by alcohol control authorities. Economic theory predicts that alcohol consumption will be negatively related to price, and thus increases in price are expected, ceteris paribus, to reduce fatalities. Schematically, the hypothesized relationships are Tax [right arrow] Price [right arrow] Consumption [right arrow] Fatalities.

   Although there is little dispute about the qualitative nature of these relationships, there is a wide range of quantitative estimates of the magnitudes involved at each step and a further question about whether, taken together, the estimated magnitudes make sense. Some researchers, for example, have estimated reduced form relationships based on tax and fatality data, ignoring the intermediate relationships between taxes and prices, prices and consumption, and consumption and fatalities. Even these studies have produced markedly different estimates. (3) There are relatively few studies taking a structural approach. Young and Bielinska-Kwapisz (2002) find that beverage prices rise more than one for one with alcohol taxes, suggesting that taxes may indeed have substantial impacts on consumption and fatalities. However, the few studies that estimated a price-fatality relationship yielded inconsistent results. (4)

   A key empirical concern is the quality of the price data that is typically employed in analyses of consumption, fatalities, and other behaviors. Many studies have employed beer, wine, and spirits prices collected by the American Chamber of Commerce Researchers Association (ACCRA), but these data are likely to suffer from substantial measurement error (Young and Bielinska-Kwapisz 2003). (5) In addition, beverage prices may be endogenous in the sense that higher demand may result in higher market prices (Manning, Blumberg, and Moulton 1995).

   Measurement error in the price data implies that the ordinary least squares (OLS) estimator is biased and inconsistent. Similarly, endogeneity of prices also renders the OLS estimator biased and inconsistent. In simple models, both problems bias the estimated price elasticity toward zero. That is, OLS may substantially underestimate how much higher prices discourage consumption and traffic fatalities.

   However, alcohol taxes--including beer, wine, and spirits per unit excise taxes, percentage excise taxes, and state markups in control states--provides a set of instrumental variables that, in principle, can resolve the problems with the price data. This paper applies instrumental variable (IV) techniques to the estimation of the price--fatalities and consumption--fatalities linkages. The results provide substantial evidence of measurement error and/or endogeneity in both the price and consumption data, and IV estimates of price and consumption effects on fatalities are substantially larger and more significant than those from OLS. However, a priori considerations and some evidence suggest that taxes are not fully suitable as instruments.

2. Methods

   We estimate regression models that express fatality rates as functions of alcohol prices (or consumption), socioeconomic characteristics, the legal environment, and state and year fixed effects:

   [Y.sub.it] = ln [P.sub.it] [alpha] + [x.sub.it] [beta] + [u.sub.i] + [v.sub.t] + [W.sub.it]. (1)

   The dependent variable, [Y.sub.it], is the logistic transformation of the observed fatality rate, [r.sub.it], that is fatalities per capita. The logistic transformation restricts the predicted value of fatalities to be nonnegative and has been widely used in previous studies. (6) However, the logistic transformation implies that the disturbance [w.sub.it] is heteroskedastic, and thus Equation 1 is estimated by weighted least squares (Greene 2003, section 21.4.6).

   The price of alcohol, P, is the Stone price index, a geometric weighted average of the prices of beer, wine, and spirits, with weights equal to each beverage's share, [[theta].sub.j], in national expenditure on alcohol: (7)

   lnP [equivalent to] [summation over (j)] [[theta].sub.j] ln [p.sub.j]. (2)

   The coefficient, [alpha] on the price variable in Equation 1 is approximately the elasticity of fatalities with respect to the price of alcohol. (8) For example, a coefficient of -0.1 implies that a 10% increase in alcohol prices is associated with a 1% decline in fatalities. Similarly, when the natural logarithm of per capita alcohol consumption is entered instead of price, the coefficient is approximately the elasticity of fatalities with respect to consumption.

   If beverage prices are measured with error, then the price index will be correlated with the disturbance term and the OLS estimator of the price coefficient is biased and inconsistent (Greene 2003, section 5.6). If there is only a single variable subject to classical measurement error, then the OLS estimator is biased toward zero (attenuated). Similarly, if price is endogenous, it is correlated with the disturbance term, and OLS estimates a weighted average of the demand (negative) and supply (positive) coefficients. Thus, endogeneity of prices will also bias the OLS estimator away from negative values. Similar reasoning applies when alcohol consumption replaces price as a right-hand side variable. (9)

   The biases due to measurement error and endogeneity can be eliminated by standard two-stage estimation methods if a set of proper instrumental variables can be found. Young and BielinskaKwapisz (2002) show that state and federal excise taxes and markups explain about 30% of the variation in alcohol prices in pooled cross-section time-series data similar to that employed in this study. (10)

   Although alcohol taxes are unlikely to be correlated with errors of measurement, they may reflect unmeasured attitudes toward alcohol. In particular, taxes may be higher in states in which there is stronger anti-alcohol sentiment, or taxes may change over time in response to changes in attitudes toward alcohol. If this is the case, taxes are not proper instrumental variables. In addition, taxes may be correlated with other policies intended to reduce alcohol use, drunk driving, and/or fatalities, and data limitations may make it impossible to explicitly include these policies as control variables. In this case, taxes will be correlated with the disturbance term, and thus taxes will not be appropriately exogenous for use as instruments. (11)

   We conclude that state and federal excise taxes and markups are likely to be good instrumental variables to deal with the measurement error problem, but may not be fully satisfactory to deal with problems of endogeneity/spurious correlation. The latter problem is mitigated to the extent that the regressions include socioeconomic variables, state dummies, and/or time trends that control for sentiment toward alcohol and its evolution over the sample period.

3. Data

   The data include the contiguous states (i = 1,..., 48) and the years 1982-2000 (t= 1,..., 19), but price data are missing for various years in some states, so the total number of observations is 869. Separate regressions are estimated for fatalities in the total population and the population aged 16-20, which we term "teen" fatalities. (12) Teen fatality rates are about twice those of the total population. Both fatality rates display a significant downward trend since the early 1980s, declining by about one quarter in each case.

   Separate fatality rates are also computed based on the accident day and time. Thus, weekend night fatalities, which are most likely to involve alcohol, occur between 6 p.m. Friday and 6 a.m. Saturday, and between 6 p.m. Saturday and 6 a.m. Sunday. (13) Fatalities occurring between 6 a.m. and 6 p.m. Monday-Friday, plus those occurring 6 p.m. Sunday to 6 a.m. Monday, are much less likely to involve alcohol. We term these fatalities "other times." (14) As Table la indicates, weekend night fatalities account for one fourth to one third of fatalities, even though these times account for only one seventh of the week. For the results to "make sense," these weekend night fatalities should be more closely related to alcohol prices or consumption than are fatalities at other times.

   Per capita alcohol consumption declined steadily from the early 1980s to mid 1990s before leveling off at about 2.2 gallons of pure ethanol per capita (Nelson 1997; Nephew et al. 2002). The real price of alcohol generally declined except for an upward blip in 1991 when federal excise taxes increased. (15)

   Extensive educational efforts and legal changes in the last two decades also may have affected drinking and fatalities. Seat belt laws were eventually adopted by every state, and numerous studies have concluded that fatalities declined as a result. The effects of many other laws are harder to measure, in part because states often changed several laws simultaneously, stepped up hard-to-measure enforcement, and developed educational programs. We form an index of the legal environment by summing a series of dummy variables indicating whether or not a state has (i) an open container law, (ii) a preliminary breath test law, (iii) a dram shop law, (iv) an illegal per se blood alcohol content (BAC) level of at least 0.1, (v) mandatory licensing action upon first conviction for driving under the influence (DUI), (vi) and an administrative per se law. (16) Two more recent legal innovations are...