Beer taxation and alcohol-related traffic fatalities.

1. Introduction

Traffic accidents cause thousands of deaths in the U.S. every year, and a substantial proportion of those deaths are estimated to have been alcohol related. For instance, 17,126 trafficaccident deaths in the U.S. during 1996 were alcohol related (40.9% of total traffic fatalities) as defined by the National Highway Traffic Safety Administration (NHTSA) (accidents in which one or more of the drivers involved had been drinking, i.e., had a blood-alcohol content [BAC] of 0.01 or above). Economists have explored the relative effectiveness of policies intended to reduce this carnage on the roads, generally by estimating the deterrence effects on traffic fatalities of various laws aimed at drunk drivers, age restrictions on alcohol consumption, and alcohol taxes. Forced to select among many policy variables in order to keep reduced-form regressions manageable, however, researchers have virtually always controlled for beer taxes and drinkingage laws in reduced-form models, while the selection of other variables has been much less consistent. Therefore, even though most studies find that some laws significantly curb driving under the influence of alcohol (DUI), the literature has not provided consistent evidence of any particular source of direct deterrenee due to the probability and/or severity of punishment. Increasing the legal drinking age from 18 to 21 is generally shown to significantly reduce auto deaths, a conclusion that offers little guidance for public policy since 21 is now standard. However, raising alcohol taxes, and particularly beer taxes, also appears to be an important policy tool for reducing DUI.

Several studies support findings first reported in Cook (1981) that higher excise taxes on beer significantly reduce traffic mortality rates. Indeed, Phelps (1988, p. 19) concludes that "[t]he effectiveness of a higher alcohol tax on vehicle fatalities does not seem open to serious question (e.g., Cook 1981; Saffer and Grossman 1987a, b). The primary issue is 'how much' rather than 'whether.'" Similarly, the U.S. Department of Health and Human Services (1988, p. 18) reports that "research evidence shows that an increase in the excise tax could have the largest long-term effect on alcohol-impaired driving of all policy and program options available." Subsequent studies reinforce this interpretation (e.g., Evans, Neville, and Graham 1991; Chaloupka, Saffer, and Grossman 1993; Mullahay and Sindelar 1994; Ruhm 1995, 1996), so similar conclusions to those quoted above are drawn in more recent reviews such as Chaloupka (1993) and Grossman et al. (1993). The strong and consistent findings regarding the effectiveness of alcohol taxes are somewhat surprising, however, because beer market studies suggest that beer taxes only have a small impact on consumption, and studies of drinking behavior using survey data indicate that heavy drinkers are the least responsive to prices.

Section 2 outlines a model of the relationship between alcohol taxes and DUI fatalities and discusses related literature on alcohol markets and drinking behavior in order to suggest why alcohol taxes should not be as important as they appear to be. Then, given this contention that a strong relationship between alcohol taxes and traffic fatalities seems unlikely, the question becomes, why do beer taxes appear to be such an important deterrent to traffic fatalities in many studies that use state-level data? Possible answers are proposed in section 2, and the primary focus of this study is an empirical exploration of two of those potential answers. Section 3 discusses the data employed for many of the previous empirical tests and for those presented in sections 4 and 5. Section 4 explores and supports the first potential answer: that the results derived from the data periods used in many studies may not be replicated for other time periods. Section 5 considers another potential answer: possible missing variable biases. Three potential missing variable issues are explored. One, a lack of control for law enforcement effort, does not appear to bias the tax results. The other two, a lack of consideration for determinants of alcohol price and consumption other than taxes and drinking age and frequent failure to control for factors that may simultaneously determine drinking behavior and support for alcohol taxes, may bias the tax coefficient, however. Therefore, even though data limitations prevent definitive conclusions with regard to some issues raised here, the results suggest that the focus on taxes as a primary policy tool to control drunk driving deserves additional consideration. Concluding remarks in section 6 elaborate on this point.

2. Modeling the Relationship Between Alcohol Taxes and DUI Fatalities: A Reevaluation of the Impact of Taxes on DUI

No direct measure of DUI offenses exists. Therefore, studies of DUI use various measures of traffic fatalities to explore the impacts of potential DUI deterrents. Theoretically, alcoholrelated traffic deaths in a state, R, are a function of unmeasurable drunk driving offenses (D) and a vector T containing measures of traffic, vehicle safety, and driver safety and is expressed as

R = f(= D, T). (1)

Drunk driving and the amount of traffic should have positive effects on R, while vehicle safety and driver safety should reduce R. Drunk driving is in turn a function of alcohol consumption ([Q.sub.a]), the expected punishment from drunk driving, determined by the probability of being arrested and convicted (P) and the expected severity of punishment (S), and a vector N of variables measuring the likelihood of driving (drunk or not) for people who drink, that is,

D = f(= [Q.sub.a], P, S, N). (2)

[Q.sub.a] and N should have positive effects on D, but P and S should be negatively related to D. Since no data exist to measure the drunk driving offense rate, the driver-involvement equation must be estimated in reduced form by substituting Equation 2 into Equation 1 as

R = f(= [Q.sub.a], P, S, T, N). (3)

[Q.sub.a] should be positively related to R. It is also hypothesized that deterrence works; that is, as P and S rise, the driver-involvement rate should fall. An empirical model of Equation 3 appears in Benson, Mast, and Rasmussen (1999b). Most studies have not considered [Q.sub.a] directly, however, but instead have included various determinants of [Q.sub.a], primarily taxes and drinking age, in the reduced-form model.

Alcohol consumption, [Q.sub.a], is determined by the interaction of supply and demand. The quantity demanded, [Q.sub.d], depends on the price ([P.sub.a]), a vector of laws that affect alcohol availability (L), including the legal drinking age, and a vector of nonprice determinants of demand (M) such as income and population characteristics, which include attitudes toward alcohol consumption, such that

[Q.sub.d] = f(= [P.sub.a], L, M). (4)

Quantity supplied, [Q.sub.s], also depends on price ([P.sub.a]), a vector of variables influencing the level of competition (C) such as laws regarding entry and market practices, and costs of supplying alcohol. Assuming that production costs are roughly equal for a particular alcohol type (e.g., beer), differences in the unit costs across states should reflect transportation costs (Tr) and taxes (Ta) so that

[Q.sub.s] = f(= [P.sub.a], C, Tr, Ta). (5)

In equilibrium, [Q.sub.s] = [Q.sub.d], and because the price is endogenous, the equilibrium quantity, [Q.sub.a], can be estimated in reduced form, including only exogenous variables from Equations 4 and 5 (Sass and Saurman 1993) as

[Q.sub.a] = f(= L, M, C, Ta, Tr). (6)

Laws limiting availability, policies reducing the intensity of competition, and taxes should be negatively related to [Q.sub.a]. The effects on [Q.sub.a] of various nonprice determinants of demand are discussed below when the empirical model is specified. The standard model of DUI estimates the driver-involvement equation in reduced form, as noted above, by implicitly substituting Equation 6 into Equation 3 as

R = f(= L, M, C, Ta, Tr, P, S, T, N). (7)

A large and significant impact of alcohol taxes on highway fatalities is surprising in such a model for several reasons.

Why Are Strong Tax Effects Surprising?

Drunk driving obviously requires alcohol consumption, which is in turn at least partially determined by alcohol price, and one determinant of price is alcohol excise taxes. Recent studies of alcohol markets indicate that excise taxes have only a relatively small impact on the money price of alcohol, however, and that money price in turn has only a relatively small impact on consumption decisions (e.g., Sass and Saurman 1993). The law of demand certainly holds, but the price elasticity is relatively low (particularly for heavy drinkers, as noted below) as other factors such as transactions costs due to market structure characteristics and regulations, religious beliefs, and age distribution are also important determinants of demand. Most DUI studies imply price elasticities of alcohol demand that are much higher than direct studies of alcohol consumption suggest exist, perhaps because many determinants of demand are not included in most of the reduced-form DUI models. Indeed, as Dee (1999) notes, the magnitude of the beertax elasticities of traffic fatalities reported in previous DUI studies is inconsistent with the evidence from studies of drinking.

Recent empirical evidence from studies using survey data also indicates that the price elasticity of demand for alcohol may be lowest among heavy drinkers (Sloan, Reilly, and Schenzler 1994b; Chaloupka and Wechsler 1996; Kenkel 1996). Chaloupka and Wechsler (1996) find that male college students apparently do not respond significantly to money price changes, for instance, although female students do. In this same vein, Sloan, Reilly, and Schenzler (1994b) do not find a significant impact of alcohol price on motor vehicle fatality rates among 21-24year-old drivers (or 25-64-year-old drivers, for that matter), although fatalities among 18-20 year olds were reduced by a higher price of alcohol. Dee (1999) examines a cross-section timeseries pooling of data from surveys of high school seniors, however, and finds no statistical relationship between beer taxes and the level of teen drinking. Similarly, Kenkel (1996) suggests a relatively low price elasticity of the frequency of heavy drinking by males who have relatively little information about the dangers of drinking. His measure of the frequency of heavy drinking is the number of days in the past year a person had five or more drinks, and a substantial portion of male drinkers in the survey sample are in his category of relatively uninformed heavy drinkers. If the college males, 21-24 year olds, high school seniors, and/or relatively uninformed heavy-drinking males are particularly prone to drink and drive, findings that beer taxes are an effective way to combat drunk driving fatalities are, once again, surprising. After all, given the degree of product variety in alcohol markets and even in beer markets, an individual can be quite responsive to price changes in one segment of the market without actually reducing alcohol consumption. A beer drinker who consumes a relatively high-priced import or premium domestic beer may buy less of it after a tax increase, for instance, but maintain consumption of alcohol by substituting into a lower priced domestic brand. Similarly, consumers of highly advertised domestic brands can substitute into even lower priced brands or generic beers. Thus, tax increases could lead to considerable changes in consumption patterns across brands (and even across types of alcohol - prohibition led to a substitution of liquor for beer, e.g. [Rasmussen and Benson 1994]) without actually reducing consumption of alcohol at all, at least for heavy drinkers.

The importance of taxes as a DUI deterrent is also surprising because drinking at onpremise sites before driving accounts for 40-69% of drunk drivers (O'Donnell 1985; Wieczorek, Miller, and Nochajski 1989), leaving only 31-60% of drunk drivers who purchase their alcohol at retail outlets. The retail price of alcohol may not be a good measure of the cost of drinking in bars and restaurants, however, and this may be even more true for excise taxes as a determinant of or proxy for price. Indeed, while the relative prices of at-home and on-premise drinking might be an important determinant of DUI, excise taxes clearly do not account for the fact that the price of alcohol consumed in bars and restaurants increased rather dramatically during the 1980s relative to the retail price of alcohol purchased for at-home consumption (Sloan, Reilly, and Schenzler 1994b).

So Why Do Taxes Appear to Be So Important?

Cross-section and cross-section time-series pools of state-level data cannot directly control for the impact of policies on heavy versus light drinkers or at-home versus on-premise drinking, of course, so such studies might be relatively ineffective at detecting the effects of specific policies.(1) Given this aggregation problem and our contention that a strong relationship between beer taxes and traffic fatalities is unlikely, why do beer taxes appear to be such an important deterrent to traffic fatalities in many studies that use state-level data?

One possible answer is suggested by Laixuthai and Chaloupka's (1993) analysis of the impact of beer taxes on various measures of consumption among high school seniors. They report that the beer tax coefficient is significantly smaller for their 1989 sample than for their 1982 sample. More relevant for the frequency of drunk driving, their estimates using 1989 data indicate that beer taxes are not significantly related to the probability of at least one heavy drinking episode in the past two weeks. These findings are of potential importance because most previous DUI studies using state-level aggregate data considered periods ending in 1988 or earlier and found significant tax effects. Some of the most recently published studies focus on the 1982-1988 period (Chaloupka, Saffer, and Grossman 1993; Ruhm 1996), for instance, and Ruhm (1995) uses the 1975-1988 period. Therefore, it may be important to examine the tax-DUI relationship using more recent data (Young and Likens 1998).

Prior to 1988, driver-involvement rates (the number of drivers in fatal accidents with a blood-alcohol content [BAC] greater than or equal to 0.01 divided by the total number of drivers) did not change much, while real beer taxes fell somewhat (Kenkel 1996). The 19841992 period examined below provides greater variation in drunk driving and beer taxes than the 1982-1988 period: The driver-involvement rate began to fall sharply in 1989 and federal beer taxes doubled in 1990 from $9 to $18 per barrel. Since state taxes are added on top of the federal taxes in most empirical work on DUI (e.g., they are marginal to the federal taxes), the marginal impact of taxes presumably should rise after this federal tax increase (i.e., if taxes have an important impact on price, this doubling of federal taxes should push price up pretty dramatically and demand should be more elastic, assuming that elasticity rises as price rises). Table 1 shows the trends in driver involvement per 1,000,000 drivers from 1984 to 1992. Annual rates of driver involvement in traffic fatalities fluctuated modestly from 1984 to 1988. After 1989, this rate declined substantially in three of the four years. Driver-involvement rates declined by about 30% between 1984 and 1992, with the 1989-1992 period accounting for most of the decline. If beer tax increases are an important cause of the decline in driver-involvement rates, a data set that includes these recent changes in driver-involvement rates and taxes clearly should reveal a significant tax effect even more readily than the data sets limited to pre-1989 observations. However, if the price elasticity of demand is lower after 1989, as suggested by Laixuthai and Chaloupka's (1993) findings, then the opposite result might be anticipated. Indeed, two recently completed studies using relatively more recent data support this possibility. Young and Likens (1998) use data from 1982 through 1990 to replicate earlier studies and conclude, in contrast to the general literature, that "while earlier findings seemed to unambiguously suggest that the beer tax would be a powerful and certain tool for reducing fatalities, the results of this study suggest that the effects of raising beer taxes cannot be predicted with much confidence." Similarly, Dee (1999) uses data from 1977 through 1992 and finds that the beer-tax/trafficfatality results of earlier studies are not robust. Data for the 1989-1992 period is included in the analysis in section 4 to test the robustness of the existing evidence. This is simply the starting point for the analysis, however, because there are other potential explanations for the robust relationship between beer taxes and various measures of traffic fatalities in most previous DUI studies.

Table 1. Rates of Driver Involvement; All Drivers in Fatal Accidents
with BAC [greater than or equal to] 0.01

Year                           Rate                   % Change

1984                          138.14
1985                          130.89                    -5.25
1986                          138.66                     5.94
1987                          132.13                    -4.71
1988                          133.56                     1.08
1989                          121.12                    -9.31
1990                          122.53                     1.16
1991                          108.03                   -11.83
1992                           96.39                   -10.77

% Change 1984-1992 -30.22

Rates are expressed as drivers involved in fatal accidents with a
BAC of 0.01 + per 1,000,000 drivers.

Since there are so many factors that might reasonably be hypothesized to affect traffic fatalities, a reduced-form model of Equation 7 attempting to control for them all would be both unmanageable and uninterpretable due to substantial collinearity between policy variables. In an effort to keep such a regression manageable and to avoid severe multicollinearity problems, researchers have been forced to present relatively lean specifications. The question inevitably becomes which variables to omit? But this choice means that some studies may suffer from missing variable biases (Dee 1999). Three potential sources of biases are considered in section 5.

The literature has produced a large number of specifications, but there appears to be little consensus as to what policy variables other than taxes and drinking age should be included. State laws mandating DUI penalties are often used as measures of the severity of punishment, S in Equation 7, for instance, although the combination of variables varies from study to study. Perhaps more significantly, control for enforcement effort has been largely ignored. This probably reflects the fact that a direct measure of the probability of arrest and punishment, P in Equation 7, does not exist. Nonetheless, this missing variable problem could result in a biased estimate of the tax coefficient, so it is worthy of additional analysis. The results presented below, which control for potential determinants of the probability of arrest, suggest that this may not be a significant problem for previous studies, however.

Another potential source of missing variable bias arises because some determinants of alcohol availability and consumption ([Q.sub.a]) revealed to be important in the alcohol demand literature have not been systematically considered in the DUI literature. Thus, if taxes are correlated with factors such as determinants of market competition or transportation costs, the tax coefficient may be biased in the negative direction. The estimates in section 5 suggest that this missing variable problem could be an important source of bias for tax coefficients in many previous studies.

A third missing variable problem could exist in some (but not all) DUI studies if some determinants of the propensity to drink and drive are correlated with the propensity to support politicians who advocate alcohol taxes. A state with a large population that follows a religion that prohibits drinking may have high alcohol taxes and low numbers of DUI fatalities, for instance, but the high taxes are not likely to be the cause of the low fatalities. In this same vein, a state with a heavy drinking population may support politicians who oppose alcohol taxes, but the low taxes would not be the cause of the heavy drinking. Some studies have used panel estimation techniques to control for fixed effects that may alleviate this missing variable problem to a degree, while others have not (fixed-effect dummies also may be controlling for other factors, as explained below, so their inclusion may be warranted for other reasons as well). Similarly, some studies have controlled for indicators of general attitudes toward drinking such as religious sentiment, while others have not. If factors that are not controlled for actually cause both high taxes and low rates of alcohol consumption or alcohol abuse, the tax coefficient may be biased, and the analysis presented below suggests that this is the case (see also Dee [1999]). In this regard, for instance, Coate and Grossman (1988) find that the price responsiveness of teenage drinking is not robust when proxies for drinking sentiment (population concentrated in various religious denominations) are included in a regression.

These potential missing variable biases reflect both weaknesses in the data and the restrictions imposed by using a reduced-form model, of course. While data problems cannot be resolved directly, the restrictions imposed by using a reduced-form model can be alleviated, at least to a degree, since Equations 3 and 6 can be estimated separately. In particular, if alcohol consumption is not affected by driver involvement in alcohol-related fatalities, then the driverinvolvement and alcohol consumption equations (Eqns. 3 and 6) can be estimated in a recursive model.(2) Indeed, Young and Likens (1998), who follow the literature's practice of focusing on alternative reduced-form specifications, suggest that "much work remains to be done on the connections between taxes, prices, consumption, abuse and fatalities. Current work, which largely focuses on the reduced form relationship between taxes and fatalities, would be greatly improved if efforts were made to verify other links in the theoretical chain - that is, that taxes do indeed significantly affect prices, that prices affect consumption among alcohol abusers, and how consumption is related to drunk driving behavior" (pp. 17-18). As suggested above, studies using survey data are exploring the relationship between price and consumption among alcohol abusers (e.g., Chaloupka and Wechsler 1996; Kenkel 1996; Dee 1999), and the alcohol market literature (e.g., Sass and Saurman 1993) is considering the tax-price relationship. Section 5 offers a two-equation recursive model that separates the determinants of alcohol consumption from the other determinants of DUI in order to consider variables missing from reduced-form models and perhaps to avoid some of the collinearity between these two sets of determinants.(3) Now let us turn to the data employed for such analysis.

3. Data for Empirical Analysis of DUI Deterrence

There is no direct measure of drunk driving, so studies employ various state vehicle mortality rates, the most serious consequence of DUI, as proxies for drunk driving offenses. Examples of these proxies are the total fatality rate (Chaloupka, Saffer, and Grossman 1993; Ruhm 1996; Young and Likens 1998; Benson, Mast, and Rasmussen 1999b), motor vehicle death rate for ages 18-20 (Saffer and Grossman 1987a, b; Ruhm 1996; Young and Liken 1998; Dee 1999), night-time fatality rate (Chaloupka, Saffer, and Grossman 1993; Ruhm 1996) or night-time fatality rate for ages 18-20 (Dee 1999), the single-vehicle crash occupant fatality count (Evans, Neville, and Graham 1991), and a driver-involvement rate accounting for driver fatalities in accidents involving drivers who test positive for alcohol (Chaloupka, Saffer, and Grossman 1993; Young and Likens 1998; Benson, Mast, and Rasmussen 1999b). Each measure has advantages and disadvantages. Measures that closely correspond with alcohol involvement, such as night-time single-vehicle occupant fatalities, will not result in accurate estimates of total lives saved from certain policies since alcohol-related deaths occur to some degree for multivehicle and daytime fatalities as well. As Evans, Neville, and Graham (1991) note, such alcohol-sensitive measures provide lower-bound estimates of the total lifesaving potential of different policies. Using total fatalities could conceptually give the best estimate of total lifesaving potential of antidrunk driving measures since alcohol is involved to some degree in all categories of fatal motor vehicle accidents. Evans, Neville, and Graham (1991) refer to total fatality measures as inclusive, and Cook and Tauchen (1984) advocate using a measure of total fatalities for this reason. On the other hand, if the target is DUI deterrence, an analysis of the factors influencing this policy goal is best served by a fatality measure that closely corresponds with drunk driving. For instance, in 1992, driver fatalities were alcohol related (a BAC of 0.01 or greater) as defined by the NHTSA in 73.8% of night-time, single-vehicle accidents, while only 12.1% of fatally injured drivers in multiple-vehicle, daytime crashes had a BAC of 0.01 or greater. Thus, many studies in the literature estimate equations with both exclusive and inclusive measures, and one of each type is employed here.

As an inclusive measure, the total vehicle fatality rate per 1000 population (the total number of vehicle deaths in each state divided by the population ages 16 and above: F = [fatalities]/[population age 16 and older in thousands]) is the dependent variable in the first column of Table 3. The minimum and maximum values for this variable are 0.08 and 0.38, so as defined, it falls within the unit interval (means and standard deviations for this and all other variables used in the various estimations are reported in Table 2, where the variables are also defined).(4)

The exclusive measure employed below is the driver-involvement rate per 1000 drivers (R), which is computed by dividing the number of drivers in fatal accidents with a BAC greater than or equal to 0.01 by the total numbers of drivers (in thousands) in each state (this variable also falls within the unit interval, with minimum and maximum values of 0.04 and 0.27). These data are preferred to the other alcohol-sensitive measures because they identify fatal accidents that involve drivers who consumed alcohol and they relate to all traffic fatalities as opposed to only those that occur at night, those in which a driver is killed, and those involving only singlevehicle crashes. After all, the role of alcohol in accidents is indicated by the BAC of drivers. These data have not been widely used in the literature, however (Chaloupka, Saffer, and Grossman [1993], Benson, Mast, and Rasmussen [1999a, b], and Young and Liken [1998] appear to be the only studies using them), in part because only 15 states consistently tested fatally injured drivers' BAC from 1980 to 1985, with testing of surviving drivers involved in fatal accidents done even less often. Testing has increased over time but, in 1990, BAC test results were still missing for 27% of dead drivers and 75% of surviving drivers involved in accidents producing fatalities. However, the NHTSA has introduced a methodology for estimating BAC values for drivers and nonmotorists for whom test results were lacking. This method

utilizes discriminant analysis to form linear combinations of variables associated with alcohol involvement in drivers and nonoccupants, and uses these linear functions to estimate posterior BAC distributions on various person, vehicle, and accident attributes. . . . Variables found most useful in estimating BAC are: police-reported alcohol involvement, accident hour, person age, vehicle role, injury severity, weekday/weekend, use of occupant restraint, driver license status, number of entries on driver record, person sex, location of nonoccupant in relation to roadway, and whether or not the driver could drink (minimum drinking age in accident state). (Klein 1986, p. i)

Validation tests were conducted using cases with known BAC test results from the 1984 and 1985 FARS. Klein (1986) reports that, for drivers involved in fatal accidents in 1985 with known BAC, the actual (estimated) percent with BAC between 0.01 and 0.09 was 13 (12). The actual (estimated) percent of drivers with BAC levels of 0.10+ was 45 (45) in 1985. Since the actual observations of BAC are probably the best data available as an exclusive measure and the estimates for missing observations appear to be quite accurate, state-level data from 1984 to 1992 for all drivers involved in alcohol-related traffic fatalities serve as a second dependent variable in the following analysis.

The NHTSA defines an accident as alcohol related if the driver has a BAC of 0.01 or greater, so that same criteria is adopted here. Recognizing that a BAC of 0.01 does not imply alcohol impairment, however (it simply implies a very modest amount of alcohol consumption), other BAC levels were also considered (cf., Benson, Mast, and Rasmussen 1999b). The BAC cutoff does not make a difference in the results reported here. Indeed, models with estimated BAC of 0.10 and above produced results virtually identical to those for drivers with BAC of 0.01 and above. The fact is that the data are dominated by observations with BAC near 0.10 and above, so even if some range between 0.01 and 0.10 (e.g., perhaps below 0.05) are not actually impaired by the low level of alcohol involved, the added observations with lower BAC using the NHTSA's definition of alcohol related do not change the results. Now let us turn to the explanatory variables used in the literature (the actual variables employed in the empirical analysis presented below are defined in Table 2).

Probability of Punishment (P)

No direct measure of the probability of arrest or of law enforcement efforts directed at DUI is available. Sloan, Reilly, and Schenzler (1994a, b, 1995) and Benson, Mast, and Rasmussen (1999b) control for police employment rates in their studies of drinking and driving.(5) Sloan, Reilly, and Schenzler's (1995) results suggest that police employment may reduce binge drinking leading to drunk driving, but their other studies (1994a, b) indicate that these results are sensitive to the inclusion of time dummies. Benson, Mast, and Rasmussen (1999b) similarly find that this variable is sensitive to inclusion of controls for fixed effects, but they also suggest that several other variables are likely determinants of the probability of arrest.(6) One is opencontainer laws, which make it illegal to have an open container of alcohol in an automobile's passenger compartment. These laws could increase the probability of being stopped while drinking and driving if a driver is viewed by police while drinking as well as the probability of arrest once stopped for drivers who are not legally intoxicated but have open containers in their cars. Such laws could also increase the probability of conviction after an arrest and the severity of punishment if drivers are convicted of DUI and open-container violations and as a result receive harsher sentences.

Anti consumption laws ban all consumption of alcoholic beverages in automobiles. These laws could deter drunk driving in a manner similar to open-container laws. Preliminary breath test laws allow police to administer breath tests without medical supervision. The results are used as probable cause for DUI arrest. This law could increase the probability of being arrested once stopped as well as the probability of being convicted given arrest.

Implied-consent laws presume driver's license holders agree to alcohol and drug tests on request or their licenses are suspended. This law could also increase the probabilities of arrest and conviction. The minimum license suspension pursuant to implied consent laws for drivers who refuse breath tests is the variable that is used below to control for such laws.

Illegal per se laws make it a crime to drive with a BAC at or above some predetermined level. Under these laws, prosecutors do not have to show that the driver was actually impaired to get DUI convictions, so this law also could increase the probabilities of arrest and conviction. Combinations of laws may offer significant aleterrence even when individual laws do not appear to be important (Evans, Neville, and Graham 1991). Chaloupka and Wechsler (1996, p. 121) use an index developed by Mothers Against Drunk Driving (MADD) to reflect the restrictiveness of each state's drunk driving laws as a group, for instance, and conclude that strong state policies "significantly reduce all measures of drinking in both specifications for the underage and older college student samples." Benson, Mast, and Rasmussen (1999b) examine various groupings of deterrence variables and conclude that the determinants of the probability of arrest are significant [TABULAR DATA FOR TABLE 2 OMITTED] as a group, while the group of variables controlling for the severity of punishment are generally not significant.(7)

Table 3. Reduced-Form Estimates

                         Total Fatalities   Alcohol-Involved Drivers

Real income                -1.08E-4(***)            -1.13E-4(***)
                          (13.97)                  (10.03)

Over 65 mph                 5.34E-3(***)             7.73E-3(**)
                           (2.59)                   (2.54)

Vehicle miles traveled      2.72E-5(***)             8.78E-5(***)
                           (3.69)                   (8.11)

Population ages 16-24      -2.42(***)               -4.11(***)
                           (2.74)                   (3.16)

Seat belt law              -0.017                    3.61E-3
                           (0.72)                   (0.10)

Unemployment rate          -0.028(***)              -0.019(**)
                           (4.93)                   (2.28)

Dry county                  0.093                    0.372(**)
                           (0.80)                   (2.19)

Mormon                     -1.08(***)               -1.06(***)
                           (6.31)                   (3.86)

Southern Baptist            0.068                    0.157
                           (0.43)                   (0.68)

Catholic                   -0.968(***)               0.103
                           (7.96)                   (0.58)

Protestant                 -1.44(***)               -1.18(***)
                           (9.29)                   (5.22)

Legal drinking age          0.021                   -0.034(*)
                           (1.53)                   (1.72)

Real beer tax               0.104                   -0.194(*)
                           (1.51)                   (1.90)

Preliminary breath test    -0.040(**)               -0.109(***)
                           (2.17)                   (4.07)

No plea bargain            -0.013                   -0.041
                           (0.50)                   (1.05)

Dram shop                  -0.089(***)              -0.117(***)
                           (4.87)                   (4.44)

Administrative per se      -0.023                   -0.103(***)
                           (0.97)                   (2.96)

Minimum administrative      5.46E-4(**)              1.20E-3(***)
                           (2.24)                   (3.51)

Mandatory fine             -0.044                   -0.042
                           (1.04)                   (0.68)

Real minimum fine           1.06E-4                 -1.90E-5
                           (0.94)                   (0.12)

Mandatory license          -0.060(***)              -0.081(**)
                           (2.64)                   (2.43)

Minimum license             6.92E-4(***)             1.15E-4
                           (3.29)                   (0.38)

Adjusted [R.sup.2]          0.811                    0.688

F                          62.5                     32.7

N = 432. Absolute value of the t-ratios are in parentheses.
Intercepts and year dummy coefficients are not reported. Significant
at the * 0.10 level, ** the 0.05 level, and the *** 0.01 level.

Severity of Punishment (S)

Administrative per se laws allow for the automatic suspension of a driver's license at the time of a DUI arrest, thus increasing the certainty of punishment for DUI arrest whether the person is convicted or not. Two variables can be used to control for such laws: (i) a dummy equalling one if the state has such a law for a first DUI arrest and zero otherwise and (ii) the mandatory minimum suspension or revocation under the law.

No plea bargaining laws require persons arrested for DUI to be tried for DUI unless there is clearly insufficient evidence for conviction. Presumably, this increases the likelihood of conviction for DUI rather than for some lesser offense through a plea bargain.

Laws establishing minimum mandatory penalties for first DUI convictions include fines, license suspensions, and jail sentences. Such laws can be represented by dummy variables, by continuous variables represented by the minimum statutory penalty, or by some combination of the two. Several states also have laws establishing relatively severe minimum penalties for conviction on a second and third (or subsequent) DUI offense, including fines, license suspensions, and jail sentences. Given the number of potential variables to control for expected severity, most studies have selected some subset of these various possibilities.

Alcohol Consumption ([Q.sub.a])

Per capita alcohol consumption can be measured by shipments (in gallons) of beer, liquor, and wine divided by the population aged 18 and over or by ethanol per capita computed from data on shipments of beer, liquor, and wine using average alcohol contents for the three beverages (0.046 alcohol content for beer, 0.40 for liquor, and 0.11 for wine).(8) Most studies have not controlled for these variables directly, however, choosing instead to control for determinants of alcohol consumption (Benson, Mast, and Rasmussen [1999b] control for ethanol per capita).

Determinants of Liquor Availability (L)

Dram-shop laws allow those injured by an intoxicated individual to bring a suit against the public establishment that served the alcohol. Such laws are expected to reduce DUI, so a dummy variable is often included to account for their existence. Some measure of the state's minimum legal purchase age for beer with 3.2% or greater alcohol content is also included, although the precise specification of the variable varies from study to study. All states had a drinking age of 21 by the end of the period being studied, but some had lower drinking ages earlier in the period studied here. One measure used in the literature is the legal age itself, and if the drinking-age law changes during a year, this variable can equal a weighted average based on the number of days each drinking age was in effect during the year. As states changed their drinking-age laws, those that had already been allowed to drink at the younger age were generally grandfathered in, so some studies have adjusted their age variable to account for this grandfathering provision. A preferred measure is employed in studies that control for the portion of the population ages 18-20 that can legally consume beer (if the drinking age changed during a year, this variable is a weighted average based on the number of days each drinking age was in effect during the year). This variable is expected to more accurately capture the impact of drinking age since raising the drinking age in a state with a large portion of the population between 18 and 20 should have a more substantial impact than doing so in a state with only a small portion of the population in that age group.(9) Since no state-level data on age-specific alcohol consumption are available, the legal drinking age might pick up aspects of the age distribution of alcohol consumption. Also, the minimum drinking age might affect aspects of drinking behavior such as location and intensity. A higher drinking age should reduce alcohol availability for drivers between the ages of 18 and 20, so the sign of this coefficient could be negative. The impact of a higher drinking age for other ages may not be in the same direction, however, as Asch and Levy (1990) provide evidence that higher legal drinking ages increase fatality rates for drivers above the legal age by decreasing their drinking experience at younger ages. Therefore, the sign for the population as a whole cannot be predicted a priori.

Liquor Market Characteristics (C and Tr)

The DUI literature has focused on beer as the most significant source of alcohol affecting DUI behavior. Results in section 5 tentatively support this focus, as the relationships between the driver involvement rate and both wine and liquor consumption are not robust across model specifications. Therefore, when Equation 6 is estimated, only beer is considered, and the relevant market characteristics are those suggested by the beer market literature. For instance, a variable measuring the fraction of the year that states mandated exclusive territories for beer distributors is often considered in the literature, although there is some dispute as to its impact on consumption (recent studies of mandated exclusive beer territories include Sass and Saurman [1993], Culbertson and Bradford [1991], Jordan and Jaffee [1987], and Carstensen and Dahlson [1986]).(10) Some researchers contend that mandated exclusive territories reduce competition and therefore raise price and lower consumption. Others point out that mandated territories do not reduce interbrand competition, however, and that they may even increase such competition as the benefits of brand advertising and promotions can be internalized by the distributors. A dummy variable for states requiring retailers to pay immediately for beer purchased from wholesalers (so-called cash laws) is also frequently examined.(11) This tends to raise the transactions costs for beer retailers so it is expected to raise price (and perhaps reduce the number of outlets) and reduce consumption.

Forced deposits, measured as the fraction of days out of the year that the state had a law requiring all beverage containers of certain sizes to be returnable and carry deposits, are also expected to raise relative beer prices (e.g., relative to liquor and wine substitutes) by raising the transactions costs associated with both selling beer (collecting deposits, keeping records, etc.) and buying beer (to reduce the money price buyers must retain and return bottles or cans), and these higher transactions costs should reduce beer consumption.(12) Hotel, motel, and tourist court receipts as a percentage of retail sales is also a common control variable.(13) It is expected to have a positive influence on beer consumption.

Finally, distance from the nearest major brewery (Anheuser-Busch, Miller, Stroh, Heileman, Pabst, and Coors) to the most populous city in each state is frequently used in beer market studies as a proxy for transportation costs in beer delivery, as this is expected to be the primary source of cost other than taxes that might vary across geographic space.(14) This distance variable is expected to negatively affect beer consumption (it could also be correlated with state regulations that encourage or discourage the production of beer within the state or that favor local microbreweries over the large national competitors).

Factors Influencing Both the Demand for Alcohol and the Likelihood of Driving (M and N)

Measures of income (e.g., real per capita income, real disposable per capita income) and/or the unemployment rate are included in DUI studies to capture differences in economic conditions across states.(15) While income is likely to be positively related to car ownership and driving, it probably is also positively correlated with car safety and with the opportunity cost of drunk driving (e.g., the possibility of being apprehended, injured, or killed). In addition, income is also expected to be a determinant of alcohol demand, of course, and its relationship depends on whether alcohol products are normal or inferior goods. Therefore, no a priori signs can be predicted for its coefficient. As with income, the impact of the unemployment rate on fatalities is not certain. Unemployed people presumably have more leisure time to spend drinking and driving, but they also have less income, so if drinking and driving are both normal goods, they could drink and drive less. Some studies also control for urbanization. For instance, the percent of the population residing in metropolitan areas within a state and/or the population density of a state may imply relatively less driving (e.g., due to the availability of public transportation and taxis), and they also may be related to urban poverty and other characteristics of urban areas that influence driving and or drunk driving behavior.(16) Less driving should reduce DUI, but the total impact of urbanization on DUI is unclear. After all, urbanization and densely populated areas probably increase alcohol availability (e.g., proximity to and numbers of outlets) but they also provide both substitutes for and complements to alcohol consumption.

To capture attitudes toward drinking, some DUI studies have used state level dummies to control for fixed effects, assuming that such attitudes are fixed within states over time. Of course, such dummies may control for other unmeasurable state characteristics that are fixed, such as the quality of roads, the relative length and severity of winters, which imply fewer daylight hours as well as more hazardous road conditions, perhaps the propensity to drink beer due to hot weather, and so on (in this same vein, several studies also use year dummies to control for temporal variation in unmeasured variables and any other time trends). Therefore, they may be appropriate to include even if they do not control for fixed attitudes (Young and Likens 1998; Dee 1999). Furthermore, such dummies will not control for attitudes if those attitudes are changing over time. Other studies have controlled for the fraction of the population living in counties dry for beer, a variable that also clearly influences the availability of beer. Some studies have also used the fraction of the population that is Catholic, Mormon, Southern Baptist, and Protestant (other than Mormon or Southern Baptist).(17) Residents of dry counties might drink less than people in wet counties, but when they do drink in bars and restaurants, they can be expected to drive more. For this reason, the coefficient on this variable has no a priori predicted sign. While Catholics and Protestants do not prohibit alcohol use, Mormons and Southern Baptists explicitly forbid its consumption. Thus, it might seem reasonable to assume a priori that the signs for the coefficients of at least these two variables will be negative. Evidence suggests otherwise, however, at least for beer. Ornstein and Hanssens' (1985) findings indicate that Mormons and Southern Baptists as a group (as well as other Protestants and Catholics) tend to substitute beer for other kinds of liquor, so they are associated with increasing beer consumption and decreasing wine and liquor consumption. These variables may also have an independent impact on fatalities, perhaps by affecting the manner or location of drinking (e.g., drinking in public at bars versus drinking at home) and/or the type of drinking, given a quantity (heavy but infrequent versus frequent but moderate). Therefore, no a priori predictions are made for the signs of these coefficients.

Traffic Conditions (T)

Vehicle miles traveled per licensed driver and the percentage of the highway traffic exceeding 65 miles per hour should both be positively related to traffic fatalities.(18) Use of seat belts should reduce fatalities, so laws mandating seat belt use can be expected to have a negative coefficient.(19) Some studies have also controlled for mandated vehicle inspections, which might be expected to improve automobile safety and reduce fatalities (these data were not regularly available for all of the years of this study, however). A larger portion of drivers who are young should increase fatalities. Some studies have used the fraction of licensed drivers age 24 or under to control for this effect. Licensed drivers by age are not available for all states for some of the years in our sample, however, and the available data are problematic. For instance, the NHTSA estimates the percentage of individuals by age who are licensed and, for many states, their estimates suggest that more than 100% of the young drivers are licensed. We believe that the data are too unreliable to use and therefore we substitute a measure of young population the percentage of the population over the age of 16 that is in the 16-24 range.

Taxes (Ta)

The key variable of interest is the sum of real federal plus state excise tax rates on beer.(20) The use of beer taxes is common in the literature, as beer appears to be the dominant type of alcohol consumed by those prone to drink and drive. Beer taxes are sometimes used as proxies for price because available beer price data are considered unreliable. The model outlined above includes beer taxes as one determinant of price and quantity, however.

4. Replication of a Representative Reduced-Form Model

As a starting point for the empirical analysis, the effect of taxes on traffic fatalities is estimated in reduced form, as in previous studies. This is done in order to determine whether the first possible reason for the consistent findings of a significant tax coefficient is relevant: the data period used in most studies. A model using panel data for the 48 contiguous states in Chaloupka, Saffer, and Grossman (1993) is replicated as closely as possible, given data availability and related considerations, except the 1984-1992 period replaces the 1982-88 period they use. Chaloupka, Saffer, and Grossman (1993) is chosen because it is clearly the most comprehensive of all of the DUI studies using state-level data. Therefore, if the relationship between taxes and traffic fatalities is not robust across data periods using their model, the implication is that the findings from less comprehensive studies probably should not be generalized either. The first regression in Table 3 represents this replication (summary statistics for all variables are reported in Table 2, where variables are also defined). It uses the total fatality rate and, as in Chaloupka, Saffer and Grossman (1993), this variable lies within the unit interval for each state, so the equation is estimated using the minimum chi-square method with the dependent variable a logistic transformation of F = ln[F/(1 - F)]. Weighted least squares is employed with weights [[N x F(1 - F)].sup.0.5], as suggested by Maddala (1983).(21)

Like most studies of alcohol-involved traffic fatalities, Chaloupka, Saffer, and Grossman (1993) employ a number of explanatory variables in a reduced-form equation, although in recognition of the problems of interpretation that arise when large numbers of variables are included, they appropriately provide a limited specification. They select the variables for their limited specification (the one replicated here) using an empirical criteria - choosing variables that worked as expected in other regressions using fewer or more variables. Dummy variables for each year are included in their regressions and those reported in Table 3 in order to control for temporal variation in unmeasured variables and any other time trends, but the coefficients are not reported here or in Chaloupka, Saffer, and Grossman (1993). The hypothesized relationships between the variables and total fatalities are implied in the discussions of the model above, but more detailed discussion can be found in Chaloupka, Saffer, and Grossman (1993).

The first regression in Table 3 is a fairly faithful reproduction of the Chaloupka-SafferGrossman (1993) limited-specification model, although there are some differences that should be noted. For instance, Chaloupka, Saffer, and Grossman control for grandfather clauses in their legal drinking-age variable that exempted state residents who were of legal age prior to the increase, while our data do not (both coefficients are positive and neither is significant, however). Similarly, Chaloupka, Saffer, and Grossman (1993) use per capita income, while our variable is disposable per capita income, but again the results are consistent in sign and significance. They also use the fraction of licensed drivers age 24 or under, but data on licensed drivers by age are not available for all states for some of the years in our sample, and the available data are problematic for reasons suggested above. Therefore, we substitute a measure of young population - the percentage of the population over the age of 16 with ages in the 16-24 range. This substitution may capture something more than the young driver's relationship with total fatalities, however, since Chaloupka, Saffer, and Grossman (1993) report a positive and significant coefficient while the results in Table 3 show a negative and significant relationship. Chaloupka, Saffer, and Grossman (1993) also include a dummy indicating that a state requires annual safety inspections. We do not have ready access to the safety-inspection law data for the post1988 years, so this variable is not included (the only variable from Chaloupka, Saffer, and Grossman [1993] that is not represented). Finally, Chaloupka, Saffer, and Grossman (1993) use taxes per case of beer, while taxes per six pack are used here, so the size of the coefficients are not directly comparable (our coefficient could be multiplied by four to compare them). Furthermore, our tax variable is measured as of July 1 of each year, while Chaloupka, Saffer, and Grossman (1993) use a weighted average if taxes change during a year.(22) These are the only differences between the specification in Table 3 and Chaloupka, Saffer, and Grossman (1993) other than the data periods.

For the most part, the results in Table 3 are consistent with Chaloupka, Saffer, and Grossman (1993), although there are a few differences. A variable-by-variable comparison is not necessary or even appropriate given our focus on taxes. Instead, note that the coefficient on taxes is positive and insignificant, in contrast to the general implications in the DUI literature. Since Chaloupka, Saffer, and Grossman's (1993) specification is arrived at after consideration of the most comprehensive list of policy variables that appears in the literature, the implication would appear to be that the literature's findings regarding taxes are not robust across time periods. Something must have changed around 1988 (perhaps the change began earlier and was gradual, of course) that eliminated the negative correlation between taxes and total fatalities. Examination of the coefficients for the year dummies support this conjecture since both the 1991 and 1992 dummies have significant negative coefficients that are relatively large. These coefficients are not reported in Table 3, but only three are significant (1984 is the omitted year): 1986 with a coefficient of 0.079, 1991 with a coefficient of -0.155, and 1992 with a coefficient of -0.182. All of the others are smaller than the 1986 coefficient and are insignificant. A restriction test was also performed to determine if the beer tax coefficient changes significantly after 1988. The coefficient for 1989-1992 was positive, larger than the coefficient reported in Table 3, and relatively more significant, although the difference was not statistically significant at the 0.10 level.

What might account for the reduction in the tax coefficient when more recent data are considered? Laixuthai and Chaloupka (1993) explain the decline in significance of the beer tax coefficient in their study by pointing to the increase in the average minimum drinking age. The average drinking age was relatively high by 1989, so they conclude that a given increase in alcohol taxes has a smaller impact on the full price of alcohol because of the higher indirect costs of obtaining alcohol for young drivers. In this regard, Sloan, Reilly, and Schenzler's (1994b) findings may be particularly revealing. They do not find a significant impact of alcohol price on motor vehicle fatality rates among 21-24-year-old drivers or 25-64-year-old drivers. Only fatalities among 18-20 year olds are reduced by a higher price of alcohol, so as legal drinking ages have risen, those drivers who are most likely to be impacted by taxes apparently are less likely to be drinking and driving. Young and Likens (1998) make similar arguments in explaining their conclusions regarding the lack of significance of the tax coefficient when they analyze the 1982-1990 period, noting that states have also taken a number of other actions to reduce drunk driving so that the marginal impact of taxes may have diminished. Yet another possibility is that factors not controlled for in most studies that are correlated with taxes have changed. Moderate drinkers are more price sensitive than heavy drinkers (Laixuthai and Chaloupka 1993; Kenkel 1996), for instance, so the antidrinking campaigns that have provided political support for higher taxes may have also reduced the price elasticity of aggregate demand as the consumers most responsive to price would be most likely to have ceased consumption altogether. Furthermore, as Young and Likens (1998) and Dee (1999) note, the 1980s witnessed the rise of a number of grass-roots organizations such as MADD and Students Against Drunk Driving (SADD) that have developed extensive media campaigns to educate potential drinkers about the consequences of DUI (as well as being involved in political campaigns to establish more powerful deterrents). If people have responded to such information (e.g., by using designated drivers when they drink away from home, etc.), then the marginal impact of taxes (and other deterrents) could be decreasing. These suggestions are only tentative, of course, but the lack of robustness in the tax coefficient is clear. It should also be noted that several other alcoholrelated variables are significant in the equation, such as preliminary breath test laws, some of the minimum mandatory penalties for DUI, dram-shop laws, and several of the religion variables intended to control for alcohol sentiment, so not all of the alcohol-related variables are as sensitive as taxes to the change in data period (some of the other coefficients also differ, however, including administrative per se laws and some of the minimum mandatory punishments).

In order to see if a more exclusive dependent variable may be related to taxes, Table 3 provides a second reduced-form regression with the driver-involvement rate as the dependant variable.(23) Again, the equation is estimated using the minimum chi-square method with the dependent variable a logistic transformation of R, employing weighted least squares, as in the total fatality equation. The same set of independent variables is included as in the first regression, and the estimates are reported in the second column of Table 3.(24) In this case, the tax coefficient has the expected negative sign and, using a two-tailed t-test, the coefficient is significant at the 0.10 level. These results are offered as a starting point for further exploration of the relationships between alcohol taxes, alcohol consumption, and drunk driving fatalities, and to evaluate the suggestions about missing variable biases raised above.

5. Alcohol Consumption, Alcohol Taxes, and Drunk Driving

It is possible that the tax coefficients in Table 3 and the tax coefficients in some other studies are biased because of missing variables that could control for (i) determinants of the probability of DUI arrest, (ii) determinants of alcohol availability and consumption other than taxes and drinking age, and/or (iii) determinants of the propensity to drink and drive that may also affect the propensity to support politicians who advocate alcohol taxes. The purpose of the following analysis is to explore these possibilities.

Empirical Estimates of the Driver-Involvement Model

Estimates of several driver-involvement equations (Eqn. 3 above) are presented in Table 4. Four increasingly larger sets of independent variables are employed, and regressions for each is reported with and without state fixed effects (year dummies are included in all eight regressions, but year and state dummy coefficients are not reported) in the ordinary-least-squares (OLS) and fixed-effects (FE) columns. Because studies using pooled cross-section data contain multiple observations for each state, it is likely that the error terms for each state are correlated over time. Including state fixed effects alleviates any omitted-variable bias due to unmeasured time-invariant state factors.(25)

Consumption (in gallons) of beer, liquor, and wine divided by the population ages 18 and over are employed as measures of per capita alcohol consumption in all eight equations in Table 4.(26) Other variables discussed above, including unemployment, seat belt laws, and vehicle miles traveled, are in all of the regressions, as is real disposable income, but in this case, partially adjusted for geographic cost-of-living (COL) differences.(27) In addition, the portion of the population over the age of 16 that is male between the ages of 16 and 44 is included since members of this group are expected to be the most likely to drink and drive. Rather than using the minimum drinking age, the portion of the population ages 18-20 that can legally consume beer is employed (if the drinking age changed during a year, this variable is a weighted average based on the number of days each drinking age was in effect during the year). This variable is expected to more accurately capture the impact of drinking age, as explained above (and note that it should have the opposite sign from the results in Table 3 - if a higher drinking age reduces drunk driving fatalities, then a higher portion of the population 18-20 who can legally drink should increase fatalities). Another control variable is also added to all eight regressions - the percent of the population residing in metropolitan areas.

Since the particular focus is on the impact of taxes as they work through consumption and collinearity between alcohol consumption and the laws intended to deter drinking and/or DUI could be a problem, the first specification (the OLS and FE regressions in columns 1 and 2) in Table 4 is provided with only one deterrence variable, one that is not measured by the existence of a DUI-related law - the number of sworn officers per 100,000 population (as used by Sloan, Reilly, and Schenzler [1994a, b, 1995] and Benson, Mast, and Rasmussen [1999b]) - in order to observe the impact of adding more variables in the second. The second set of OLS and FE equations (columns 3 and 4) adds several more policy variables. First, dram-shop laws, which are intended to reduce excess consumption away from home, are controlled for. Second, additional deterrence variables are considered. There is a large array of additional deterrence variables that might be employed, of course. As a selection criteria, we draw upon the broader economics-of-crime literature, which suggests that the probability of punishment is more important as a deterrent than the severity of punishment. No direct measure of enforcement exists, but in addition to sworn officers, several laws are expected to make arrests for DUI more likely, as explained above, although they might be collinear with DUI behavior (e.g., endogenous). Such laws include open-container laws, anticonsumption laws, implied-consent laws, illegal per se laws, and preliminary breath test laws, which are represented by dummy variables. To avoid too much clutter and collinearity, these variables are added first, and potential controls for the severity of punishment are added later. This selection process is supported by empirical evidence in Benson, Mast, and Rasmussen (1999b), where a systematic treatment of deterrence variables, [TABULAR DATA FOR TABLE 4 OMITTED] including tests for group effects, suggests that, as a group, only the variables expected to affect the probability of being stopped and arrested were consistently significant, while severity variables as a group were not. Various combinations of severity variables were also included in different specifications, however. They do not influence the key alcohol consumption coefficients, so only one specification is reported here (the final two columns of Table 4) in order to illustrate this. After all, the coefficients on the alcohol consumption variables are the ones that are relevant for consideration of tax impacts.(28)

The third set of regressions (columns 5 and 6) in Table 4 includes all of the variables from the first four columns as well as variables representing religious affiliation and dry population. There may be some question as to the appropriateness of including these variables in the driverinvolvement model since they measure, at least in part, the sentiment toward alcohol and therefore influence consumption (they are also included in the beer consumption equation estimated below). This third specification is offered because it seems reasonable to expect that these variables may have an independent impact on fatalities, however, as explained above (e.g., drinkers who live in dry counties may be more likely to drink and drive, religious beliefs may affect the location and/or type of drinking [Ornstein and Hanssens 1985]). Furthermore, adding these variables does have an impact on the relevant coefficients. However, since some may question the inclusion of these variables, calculations of elasticities for beer taxes are made below with results from the fixed-effects versions of both column 4 (or column 2 since the coefficients are very similar) and column 6. The final set of regressions (columns 7 and 8) in Table 4 adds four severity of punishment variables to illustrate that they do not impact the coefficient on the key beer consumption variable (i.e., the estimated coefficient lies within the interval represented by estimates from columns 4 and 6).

Results

Coefficients of the control variables included in all three equations illustrate the consequences of adding fixed effects. The unemployment rate coefficients are insignificant without fixed effects but become negative and significant in each of the fixed-effect specifications, while income consistently loses significance. The degree of urbanization as represented by the percent metropolitan population appears to be positive without fixed effects, but when the state dummies are added, the variable significantly reduces drunk driving.

The number of drunk driving fatalities appears negatively related to the proportion of the population above the age of 16 that is composed of males between the ages of 16 and 44 without fixed effects but becomes positive and insignificant with fixed effects. A higher minimum drinking age apparently reduces the driver-involved fatality rate. While this is not revealed in the OLS specifications, adding fixed effects makes the portion of the population ages 18-20 who can legally consume beer positive and significant at a 0.05 level of confidence in the second and eighth columns and at the 0.10 level in the fourth and sixth columns.

Police employment has the anticipated sign but is not significant after fixed effects are added, and seat belt laws appear to significantly reduce drunk driving fatalities when fixed effects are added to the first regression (column 2) but lose their significance when DUI deterrence variables are added.

The number of miles traveled is consistently positive and significant. Dram-shop laws apparently are also effective means for reducing drunk driving, supporting findings in Sloan, Reilly, and Schenzler (1994a, b) and Benson, Mast, and Rasmussen (1999b). Results in the fourth, sixth, and eighth columns may appear to suggest that the most effective deterrence policy is the open-container law, but as noted above, control for fixed effects biases coefficients toward zero, so the insignificance of illegal per se laws, implied-consent laws, and anticonsumption laws may be misleading. Therefore, little weight should be put on the insignificance of these direct deterrence coefficients. The only purpose for adding deterrence variables in this analysis is to observe the impact on the coefficients for alcohol quantities (and in particular, beer quantity) in order to consider the potential tax effects, however, and since adding deterrence variables does not affect the significance of the alcohol consumption coefficients (even when more variables are added, as in column 8), we shall not address the issue further.

In the regressions reported in the sixth and eighth columns, both the Southern Baptist and the other Protestant variables are significantly and positively related to the driver-involvement rate, suggesting that they do have an independent impact on fatalities (after adding fixed effects) beyond their influence on consumption. These signs are consistent with findings by Ornstein and Hanssens (1985) and with the hypothesis suggested in section 3 that religious sentiment influences the type and/or location of drinking as well as the quantity consumed.

The most important variables in terms of determining the impact of alcohol taxes are those representing the quantities of alcohol consumed. Liquor consumption has a surprising negative sign without fixed effects but becomes insignificant in each of the regressions when fixed effects are added. Wine consumption is significant in the first five columns and in the seventh column, but it is not significant in the sixth and eighth columns where controls for religion and dry counties are combined with fixed effects. Beer consumption is significantly related to the driverinvolvement rate in all regressions (note that this appears to provide some justification for the literature's focus on beer taxes rather than taxes on other types of alcohol, although it may be that factors influencing wine consumption should also be considered), but its coefficient is much smaller in the sixth and eighth columns than in the other regressions. Given the issue raised above regarding the appropriateness of including the religion and dry county variables in this model, we shall treat the coefficient estimate in the sixth column (0.027) as a lower bound and compare its implications for tax effects with the upper-bound estimates implied by the highest coefficient (0.047) that arises in both columns 2 and 4. After all, the beer quantity coefficient in a model without fixed effects ranges between 0.031 and 0.038, and it is 0.032 in a model without liquor and wine quantities (regressions not reported but available from the authors on request). Furthermore, quite comparable beer-consumption coefficients also arise with additional deterrence variables to control for severity of punishment (e.g., 0.029 in column 8), so treating 0.047 and 0.027 as upper- and lower-bound estimates seems reasonable.

These results suggest that missing variables to control for the probability of arrest that characterizes many previous studies probably do not introduce substantial biases in the tax coefficient but that controlling for religious characteristics of the population may reduce tax impacts relative to those discovered in studies that have not done so (an implication that is reinforced below). In order to consider other possible missing variable biases, let us turn to an estimation of Equation 6. Liquor consumption does not significantly increase the level of drinking-related fatalities, so a liquor consumption equation need not be estimated. Results here suggest that wine consumption may influence drunk driving, as noted above. However, since the quantity of wine consumed is small relative to beer (see Table 2), wine consumption is not always significant while beer consumption is, and the literature has focused on beer taxes rather than liquor and wine taxes, only an empirical model of beer consumption is provided.

Empirical Estimates of a Beer-Consumption Model

To better understand the relationship between beer taxes and consumption (and therefore between taxes and DUI), three sets of independent variables are employed in regressions estimating Equation 6, with both OLS and fixed-effect versions of the regressions reported in Table 5.29 Before discussing the differences between the equations, let us consider the similarities. In particular, real COL-adjusted disposable income, the unemployment rate, the portion of the population above the age of 18 who are males between the ages of 18 and 44, the ratio of hotel, motel, and tourist court receipts as a percentage of retail sales, the percent of the state population residing in metropolitan areas, and population density are all included in each of the regressions. Several of these variables are intended to capture differences in socioeconomic determinants of beer consumption across states, and some of them appear in the driver-involvement rate equation as well, but this is because they are expected to have independent effects on driving behavior and drinking behavior, as suggested in section 3. All of the equations also contain controls for year fixed effects, although coefficients for year and state dummies (in the FE columns) are not reported.

The percentage of the 18-20 age population that can legally drink beer with alcohol content 3.2% and above is always included as well; the coefficient of this variable is expected to be positive. The real total excise tax on beer measured in dollars per six pack adjusted for geographic differences in COL is in each equation.(30) These are the only policy variables included in the first set of two columns, so these equations are, in essence, replicating the controls employed in typical reduced-form equations in the DUI literature. The second set of equations (columns 3 and 4) adds several other state laws and competitive factors discussed above that are expected to influence beer availability and price. The third set (columns 5 and 6) adds a control for the population in dry counties, which some other studies have used as a control for alcohol sentiment but which also might be interpreted as a determinant of availability, and for religious determinants of alcohol sentiment. It is the third set of regressions and, in particular, the fixed-effect specification of that regression (column 6) that we contend provide the best indicator of the impact of beer taxes on beer consumption, but first consider the regressions for the other models.

Model One (Columns 1 and 2): Taxes and Drinking Age as the Only Policy Determinants of Beer Consumption. Before considering the coefficient on taxes, note that income is significantly and positively related to beer consumption and unemployment has a significant negative coefficient when fixed effects are included, suggesting that beer is a normal good. The metropolitan population variable is negative and insignificant without fixed effects but becomes positively and significantly related to beer consumption when state dummies are added in column 2, while population density consistently has a significant and negative relationship. The coefficient on the male population between 18 and 44 has the expected sign but is insignificant (it becomes significant when alcohol-control variables and state dummies are added, as noted below). Finally, a higher portion of the population between 18 and 20 of legal drinking age apparently increases beer consumption, as expected, while taxes apparently have a large negative impact. Note in this regard that, when fixed effects are included in this regression, the tax coefficient actually becomes substantially larger, so in contrast to Saffer and Grossman (1987b), results regarding the impact of taxes might actually be sensitive to the inclusion of these state dummy variables (a point that becomes more apparent when columns 5 and 6 are considered below). The coefficient on taxes in this equation may exaggerate the tax impact, however, if taxes are a reflection of the general policy environment that characterizes beer markets and therefore determines, in part, beer consumption. To see if this is the case, consider the model in columns 3 and 4.

Model Two (Columns 3 and 4): Policy Determinants of Beer Consumption. The beer market literature suggests that several policy options can influence beer consumption. The regressions for this version of the model include a variable measuring the fraction of the year that states mandated exclusive territories for beer distributors, a dummy variable for states requiring retailers to pay immediately for beer purchased from wholesalers (cash laws), the fraction of days out of the year that the state had a law requiring all beverage containers of certain sizes to be [TABULAR DATA FOR TABLE 5 OMITTED] returnable and carry deposits, and distance from the most populous city in each state to the nearest major brewery (Anheuser-Busch, Miller, Stroh, Heileman, Pabst, and Coors). The OLS version of the regression in column 3 implies that mandated territories reduce beer consumption (while the relationship is insignificant in column 3, it is significant in the OLS regression in column 5), suggesting that they may reduce competition, but the fixed-effects specification in column 4 (and in column 6) suggests the opposite, as mandated territories appear to significantly increase consumption, implying that they may enhance interbrand competition. Cash laws consistently appear to reduce consumption, as expected. The distance to the nearest brewery has the expected sign, but its coefficient is not significant (it becomes significant when the controls for dry county population and religious sentiment are added in both the OLS regression and FE regression in columns 5 and 6, however). The coefficients on forced deposits are negative in OLS regressions, but they become insignificant when fixed effects are added. Also note that the males 18-44 variable becomes significant when fixed effects are added to this model. More importantly in regard to the issue being addressed here, adding these four variables reduces the size of the tax coefficient by almost half with the fixed-effects model while leaving the legal drinking-age coefficient virtually unchanged. The estimated elasticity of per capita beer consumption with respect to excise taxes is approximately -0.073 in the fixed-effects version (column 4), a result quite consistent with findings in recent beer market studies (e.g., Sass and Saurman 1993). This suggests that the tax coefficients in previous studies of DUI deterrence may suffer from missing variable bias due to lack of controls for such alcohol market characteristics. Still other determinants of beer consumption should be considered, however.

Model Three (Columns 5 and 6): Controlling for Alcohol Sentiment. Other determinants of alcohol consumption are added to the model in columns 5 and 6. As Chaloupka, Saffer, and Grossman (1993, p. 170) explain, "[I]f sentiment is excluded from the fatality equations, the estimated coefficients on taxes and the drunk-driving laws both overstate these variables' effects" in a reduced-form model. They employ variables controlling for the population living in counties that are dry for beer and the fractions of the population that are Catholic, Mormon, Southern Baptist, and other Protestant in an effort to deal with this issue, and the same is done here.(31) In this case, the final regression in Table 5 indicates that Mormon church membership is estimated to reduce beer consumption, while Catholic, Southern Baptist, and Protestant (other than Morman and Southern Baptist) affiliation have the opposite effect (the coefficients on the Southern Baptist and other Protestant variables shift from positive and significant to negative and significant with the addition of state dummies). The findings with fixed effects are consistent with the Ornstein and Hanssens (1985) conclusions that Southern Baptists, Catholics, and other Protestants tend to substitute beer for other types of alcohol.

For the most part, the other coefficients in columns 5 and 6 are very similar to those in columns 3 and 4 (note that the mandated territories variable changes from negative and significant to positive and significant when fixed effects are added), but there is one key exception - real beer taxes. With religion and dry population accounted for, the effect of beer taxes on consumption is dramatically smaller. Indeed, without state dummies, the coefficient in column 5 is significant and positive! Adding fixed effects re-establishes the anticipated sign for the tax coefficient, although it is insignificant and much smaller than it appears to be in any of the previous regressions, with an estimated elasticity of -0.009. The implication is that the lack of control for determinants of attitudes toward drinking that characterizes some (but certainly not all) studies is another potential source of bias for tax coefficients. This bias may arise because such population characteristics also influence attitudes toward beer taxation policy (negative attitudes toward drinking could also induce positive support for beer taxes and vise versa, e.g.). Additional research on the determinants of beer taxes is called for in order to flesh out these relationships, but the results presented here are sufficient to suggest that beer tax coefficients may be biased when controls for such factors are missing.

Taxes and DUI

Direct comparison of the results from this model and the rest of the literature is not possible because the model employed here is quite different from most other models. Indeed, the primary purpose for developing this recursive model is to explore potential missing variable biases in reduced-form models that typify the literature. Some insight can be gained, however, by estimating the impact of taxes on the driver-involvement rate and comparing the estimate to others in the literature. Therefore, even though the beer tax variable is insignificant in the fully specified model of beer consumption, upper- and lower-bound estimates of the relationship between beer consumption and the driver-involvement rate from the upper- and lower-bound estimates in columns 2 (or 4) and 6 in Table 4 and estimates of the relationship between taxes and beer consumption from the final column in Table 5 are used to obtain upper- and lower-bound estimates of the impact of beer taxes on alcohol-related traffic fatalities. Ignoring measures of religion and dry population in the driver-involvement equation (either column 2 or 4 in Table 4), the elasticity of the driver-involvement rate with respect to the beer excise tax is estimated to be about -0.012. After controlling for these variables in Table 4's sixth column, this estimated elasticity falls to -0.007. The estimates in this study are smaller than many previous estimates of the effect of beer taxes on fatalities related to alcohol, including previous estimates when religious affiliation and dry population are held constant.(32) For instance, Ruhm (1996) estimates elasticities of night-time vehicle fatalities with respect to beer taxes of between -0.21 and -0.18. Chaloupka, Saffer, and Grossman (1993) estimate that doubling the federal beer tax during their 1982-1988 study period would have reduced night-time driver fatalities by about 8.4% and alcohol-involved driver fatalities by 9.7%. Note that such a doubling actually occurred during the data period examined here, as the federal tax rose from $9 to $18 per barrel in 1990, but the expected impact apparently did not materialize given the results presented above. Evans, Neville, and Graham (1991) estimate an elasticity of single-vehicle, night-time occupant fatalities with respect to beer taxes of -0.12.

These differences apparently reflect the fact that other studies using state-level aggregate data have focused on pre-1989 data, as suggested by the discussion in section 4. Relationships clearly have changed, but the precise timing of this change is not clear. Testing the restriction that the coefficient on beer taxes was the same during the 1984-1988 and the 1989-1992 periods provides some evidence that the coefficient is smaller in the second period, but there is not a statistically significant difference, suggesting that the change may have been a gradual one or that it may have occurred either later or earlier. The recursive model employed here also allows us to control for additional determinants of beer consumption such as mandated exclusive territory and cash laws (and for some studies, determinants of attitudes toward alcohol such as the religious makeup of the population) that have not been considered in the reduced-form models used in other DUI studies.

Some of our specifications produce elasticity estimates much more in line with previous studies, of course. If the estimates in column 4 of Table 5 are employed rather than those in column 6, the estimated elasticity range is between 0.057 and 0.097, for instance. The point is that the relationship between beer taxes and alcohol-involved traffic fatalities is very sensitive to specification, however, because beer taxes clearly are correlated with other variables that can reasonably be hypothesized to influence beer consumption. When this is the case, a scaled down model that includes taxes but not the other variables implies that the coefficient on the tax variable cannot be interpreted as a pure tax impact, as it may be picking up the causal effects of the left-out variables. Therefore, if a leaner specification is appropriate, the question becomes which variables should be omitted, and any procedure that excludes some variables simply because they are correlated with taxes is clearly ad hoc. In fact, the arguments made above suggest that there are reasons to expect that taxes may not be a particularly important determinant of fatalities and therefore that the tax variable should be the one that is dropped. Furthermore, taxes appear to be the only policy variable in the recursive model that is highly sensitive to specification, suggesting that in some models it has drawn explanatory power from left-out variables.

6. Conclusion

Many studies using data from the period falling between 1975 and 1988 conclude that raising beer taxes is the most effective policy to reduce traffic fatalities. A reduced-form model (the common modeling procedure in the literature) using data from 1984 through 1992 is unable to replicate these results in a total fatalities equation, however. Therefore, an alternative dependent variable, the driver-involvement rate (the number of drivers in fatal accidents with a blood-alcohol content greater than or equal to 0.01 divided by the total numbers of drivers) is considered where the expected relationship appears to hold, at least weakly. However, this relationship also tends to disappear when missing variable biases are alleviated in a two-equation recursive model that considers the impact of taxes on alcohol consumption and of alcohol consumption on traffic fatalities. Indeed, this approach proves to be quite useful in sorting out some of the potential missing-variable biases that may be explanations for the significance of taxes in studies using reduced-form models.

The use of a reduced-form model may also explain why the literature has generally not provided consistent evidence of any particular source of deterrence other than taxes and drinking age. Forced to select among many policy variables, researchers have consistently considered these two, but the selection of other variables (laws establishing sanctions for DUI, law enforcement effort, the potential for civil action against drivers or against bar owners who sell to drunk drivers, determinants of liquor price and availability, attitudes toward drinking, etc.) has been much less consistent. Thus, some studies find some policy variables significant while others choose a different mix of variables and reach different conclusions. A recursive model allows somewhat better control for potential policy variables directed at drinking as well as variables directed at drunk driving (an issue that is not explored at length here, but see Benson, Mast, and Rasmussen [1999b]).

When other policies that have significant impacts on beer consumption are not controlled for, taxes do indeed appear to be an important determinant of consumption. However, adding these policy variables cuts the coefficient on taxes in half. Furthermore, adding variables that control for alcohol sentiment makes the tax variable insignificant and reduces the size of the coefficient even more, suggesting that taxes are correlated with (and probably caused by) factors that also influence the propensity to drink. While these results are not necessarily conclusive, they do suggest that the consistent findings of large tax impacts on DUI fatalities in previous studies may be a result of a combination of factors, including the time period from which the data have been drawn, the failure to systematically control for other policies that determine drinking and driving behavior, and/or the possibility that both state alcohol policy and the drinking behavior of state residents are determined by some of the same factors.

The appropriate specification for determining the relationships between beer taxes, beer consumption, and traffic fatalities is certainly debatable, but the evidence presented here suggests that the relationship is not nearly as robust as much of the previous literature suggests, and its sensitivity to specification clearly implies a much more cautious interpretation than is frequently offered. For instance, the U.S. Department of Health and Human Services' (1988, p. 18) conclusion that "research evidence shows that an increase in the excise tax could have the largest long-term effect on alcohol-impaired driving of all policy and program options available" clearly warrants re-evaluation. After all, even though the lack of robustness of alcohol taxes as a determinant of traffic fatalities is in sharp contrast to much of the DUI literature (probably all studies prior to Dee [1999] and Young and Likens [1998]), it is actually not surprising. Theoretically, the price of alcohol should affect consumption, which in turn influences drunk driving. This expectation is supported here to the extent that beer market factors that are expected to affect price are significant. Variations in taxes only have a small impact on alcohol price (Sass and Saurman 1993), however, and therefore, the tax impact on consumption is relatively minor. Furthermore, heavy drinkers who might be most prone to drive drunk appear to be the least responsive to alcohol price changes (Chaloupka and Wechsler 1996; Kenkel 1996).

It does not follow that nothing can be done about drunk driving, of course. Estimates of the effectiveness of other policies in Tables 4 and 5 are much more robust than those regarding beer taxes. Adding cash-only laws for beer vendors could reduce beer consumption significantly (eliminating mandated exclusive territories for beer distributors may do so as well, although results are less robust with this variable), for instance, and therefore potentially reduce traffic fatalities. While this policy is not targeted at DUI, reduced traffic fatalities may be an unintended consequence. Similarly, dram-shop liability is predicted to reduce the number of drivers in alcohol-related fatalities by 9%. More direct efforts against drinking and driving will not impose costs on beer consumers who do not drink and drive, however, and therefore they may be more equitable. In this regard, recent findings reported elsewhere suggest that a combination of laws and policies that together make arresting and punishing drunk drivers easier is likely to be more effective than any one law or policy option by itself (Benson, Mast, and Rasmussen 1999b). Furthermore, policy experiments suggest that systematic proactive law enforcement efforts that target the control of drunk driving can be very effective; they simply are not carried out in a widespread and continuous manner (Benson, Mast, and Rasmussen 1999a). Taxing sin may not have much impact but increasing the chances of punishing it may.

We would like to thank the three anonymous referees, Jonathan Hamilton, and Tim R. Sass for their helpful comments and suggestions. The paper has been substantially improved by their efforts. Research for this paper was supported by grant 1 RO1 AA10376-01A1 from the National Institute of Alcohol Abuse and Alcoholism. Data files are available from the authors on request.

1 Others have studied self-reported drunk driving information contained in microsurvey data (Kenkel 1993, 1996; Mullahy and Sindelar 1994; Sloan and Givens 1994; Sloan, Reilly, and Schenzler 1995). While microdata from surveys overcomes some of the problems associated with aggregate data, it too is imperfect. See Mast (1996) for a discussion of the benefits and potential shortcomings of using survey data to study DUI deterrence. Therefore, studies using both survey and aggregate data play important roles in this literature in recognition of the shortcomings of each.

2 It may be the case, however, that, while alcohol consumption is not directly affected by drunk driving fatalities, both are functions of unmeasured sentiment toward alcohol. In Mast (1996), estimates of driver-involvement rates were obtained treating beer consumption as endogenous, with tourism and other variables as instruments. To address the possible endogeneity of other variables (liquor and wine consumption, drunk driving laws, and the drinking age), estimates of driver-involvement rates are obtained leaving out all possibly endogenous regressors except beer consumption. Hausman (1978) tests suggest that the two-stage least squares estimates do not significantly differ from their weighted least squares counterparts. Estimates of the effect of beer consumption on driver involvement in these models are in the range of estimates presented below.

3 A multiple equation model such as the one offered below has its drawbacks due to the aggregation problems discussed in Cook (1981). After all, a substantial portion of alcohol consumption probably is not related to drinking and driving. However, reduced-form models may also be problematic due to possible missing variable biases, as suggested above. Therefore, a relatively accurate understanding of DUI deterrence probably requires consideration of results from both kinds of empirical studies, recognizing the relative shortcomings and benefits of each. Indeed, despite his concerns about using a multiple equation model, Cook employees this technique in recent work on the relationship between alcohol and violent crime (Cook and Moore 1993a, b). Furthermore, it must be emphasized that the purpose of the recursive model in section 5 is to see if the tax results using the driver-involvement rate as an independent variable are robust when some of these potential missing variable problems are considered. In doing so, however, it also becomes possible to suggest why taxes may appear to be significant in other studies when the actual impact may be much smaller than those studies imply.

4 All fatality data are from NHTSA's Fatal Accident Reporting System (various years) (hereafter, FARS).

5 These full-time equivalent police employment rates are from the U.S. Department of Commerce's Public Employment (1984, 1986-1992), although the data for 1985 are not available and are interpolated from 1984 and 1986 data.

6 All drunk driving laws discussed here and in the next section are from the U.S. Department of Transportation's Digest of State Alcohol -Highway Safety Related Legislation (various years).

7 Police expenditures are also used as a control for law enforcement effort in Brown, Jewel, and Richer (1996) and as a substitute for police employment in Sloan, Reilly, and Schenzler (1994a). Since police employment and police budgets are intended to control for the same thing, only results with employment are reported below. Evans, Neville, and Graham (1991) and Kenkel (1993, 1996) include a dummy variable for states that had laws authorizing police to use sobriety checkpoints, expecting such laws to increase the probability of DUI arrests. Chaloupka, Saffer, and Grossman (1993) reject including sobriety checkpoints in their study, however, pointing out that binary variables can not account for interstate differences in enforcement intensity. If such laws increase the likelihood of police making DUI arrests, they may be determinants of intensity, but in this case, Chaloupka, Saffer, and Grossman (1993) also note that checkpoints appear to be used in all states whether such a law exists or not (Evans, Neville, and Graham [1991] acknowledged many states use sobriety checkpoints as part of routine safety checks, e.g.). Therefore, these laws are not considered below.

8 These shipment data are from the U.S. Brewer's Association (various years). All population data used below are from the U.S. Bureau of Census (various years).

9 Data on drinking-age laws were provided by Tim R. Sass and David S. Saurman; drinking-age population data are from the U.S. Bureau of Census (various years).

10 Information on state laws regarding mandated exclusive territories is obtained from the Modern Brewery Age Blue Book (various years) and from direct examination of some state statutes.

11 Information on cash laws is from the ModeRN Brewery Age Blue Book (various years).

12 These laws are found in Beverage World Databank (various years) and the Modern Brewery Age Blue Book (various years).

13 These data are from the U.S. Department of Commerce, Census of Service Industries (various years). Values for 1984-1986 and 1988-1991 are interpolated from reported data for 1982, 1987, and 1992.

14 These data were provided by Tim R. Sass and David S. Saurman.

15 Data on per capita disposable income and unemployment are taken from the U.S. Bureau of the Census (various years) and consumer price indices (CPI) are from the U.S. President's (various years) Economic Report. Where employed, geographic cost of living indices are from McMahon and Chang (1991) with adjustments made in light of findings in Dumond, Hirsch, and Macpherson (1999).

16 These data are from the U.S. Bureau of the Census (various years).

17 Data on dry counties are from the U.S. Brewers Association (various years), while the religious variables for 1984-1989 and 1991-1992 were interpolated/extrapolated from data for 1980 from Quinn et al. (1980) and 1990 from Bradley et al. (1990).

18 Data on miles traveled and speed are from the U.S. Department of Transportation's Highway Statistics (various years), with data on drivers provided by the Federal Highway Administration.

19 Information on these laws was provided by the NHTSA.

20 Tax data are from the U.S. Brewers Association (various years) and the consumer price indices (CPI) taken from the U.S. President (various years). Again, where employed, geographic cost of living indices are from McMahon and Chang (1991).

21 In this regard, almost all recent state-level studies use logistic transformations of fatality rates as dependent variables (cf., Saffer and Grossman 1987a, b; Sloan, Reilly, and Schenzler 1994a, b; Ruhm 1995, 1996; Dee 1999; Young and Likens 1998). Asch and Levy (1990) use a closely related Cox transformation while Evans, Neville, and Graham (1991) use the log of the fatality rate. Young and Likens (1998) also experiment with alternative specifications and weights, but these alternatives do not influence their conclusions regarding taxes.

22 Tax data for this study and for Chaloupka, Saffer, and Grossman (1993) are from the U.S. Brewers Association (various years), wherein one table reports annual taxes as of July 1 and another reports the history of state taxes, therefore providing the source of information on the timing of tax changes. Unfortunately, these two tables have numerous inconsistencies for the data period of our analysis, and discussions with the publishers indicate that the errors are in the table reporting the historical changes in taxes (the historical table has not been consistently and carefully updated). The table listing taxes as of July 1 in each year is reportedly accurate, however. Therefore, while we would prefer to use the weighted average variable employed by Chaloupka, Saffer, and Grossman (1993), the lack of reliability in the data prevents us from doing so.

23 Young and Likens (1998) explore a large number of alternative specifications using a total fatality rate as well as other more exclusive dependent variables, and Dee (1999) also uses more exclusive dependent variables. Their findings also demonstrate the lack of robustness in the tax coefficient across data periods and fatality measures.

24 Note that Chaloupka, Saffer, and Grossman (1993) also use a driver-involvement rate variable, but it differs from the one used here, so the second equation in Table 3 is not as directly comparable to their results as the first equation. In particular, observations where BAC levels are actually missing but estimated by the NHTSA using Klein's (1986) method are included here, but apparently they are not in Chaloupka, Saffer, and Grossman (1993). Furthermore, Chaloupka, Saffer, and Grossman (1993, p. 165) estimate an alcohol-involved driver fatality rate "based on the fraction of dead drivers tested and the fraction of those tested with BACs of at least 0.05 percent," while we use a BAC of 0.01 or above (although other alternatives, ranging up to 0.10 were tried with no significant change in the results, as noted above).

25 In this regard, Ruhm (1996) and Dee (1999) find large differences between models estimated with and without state dummy variables. Sloan, Reilly, and Schenzler (1994a, b), Evans, Neville, and Graham (1991), Benson, Mast, and Rasmussen (1999b), and Young and Likens (1998) also use fixed-effects models. Fixed-effects models do bias coefficients toward zero, of course, making it more difficult to find meaningful results in the form of significant relationships. In fact, in this regard, Saffer and Grossman (1987a, p. 369) report estimating a model with state fixed effects, and conclude that their results "suggest a model with state dummies is overdetermined and plagued by multicollinearity." Similarly, Chaloupka, Saffer, and Grossman (1993, p. 172) note that they tried a fixed-effects model but "collinearity made it impossible to obtain meaningful results." This would be the case if one or more of the laws did not change much over time, for instance. An inability to interpret coefficients due to such biases is problematic if the objective is to draw policy inferences from a complete model of DUI deterrence. However, the purpose here is limited to exploring the relationship between the driver-involvement rate, alcohol consumption, and alcohol taxes, and in this context, Saffer and Grossman (1987b) report that estimating their model with state dummy variables does not impact the relationship between taxes and fatalities; that is, their tax coefficients are not sensitive to the inclusion of these state dummies in their model. A similar implication is drawn from the results in Table 4 (although results in Table 5 indicate that the fixed-effects specification actually raises the size of the tax coefficient, in contrast to Dee [1999]). While the coefficients for some of the other variables (and in particular, various deterrence variables) may be insignificant because of the bias towards zero in fixed-effect models, the key variable in Table 4 for considering tax effects is not affected in this way. Even though the Hausman (1978) test indicates that the fixed-effects model is preferred over the OLS model, however, both versions of the equations are reported.

26 Note that the coefficients on these variables are not directly comparable since these beverages have different levels of alcohol content (scaling for alcohol equivalence can be done using the average alcohol contents of 0.046 for beer, 0.40 for liquor, and 0.11 for wine).

27 Dumond, Hirsch, and Macpherson (1999) provide evidence that the COL indexes overestimate differences in wages across metropolitan areas due to COL by about 60% due to differences in demand across these areas, so per capita income was adjusted by 40% of COL. The 65-miles per hour variable is not included because it turns out to be correlated with unobservables and loses all explanatory power in the fixed-effects model.

28 Direct deterrence is not relevant to the issues being addressed here, but for a detailed discussion of this subject, see Benson, Mast, and Rasmussen (1999b), where a wider array of deterrence variables is considered in similar equations.

29 Again, Hausman tests suggest that the fixed-effects specifications are superior, and in this case, adding the fixed effects impacts the signs of a number of coefficients and makes some significant that otherwise are not, but both OLS and FE results are reported.

30 Recall that the regressions in Table 3 do not control for COL differences in order to be consistent with other studies of drunk driving cited here. When taxes are not adjusted to reflect COL differences, the tax coefficients are slightly larger, suggesting that the general practice in the DUI literature of not controlling for COL may be one reason for consistently larger impacts than are estimated here.

31 While some previous writers have suggested that either fixed effects or religious variables, but not both, should be used to control for alcohol sentiment, these results suggest that these controls have very different effects. After all, while state dummies may control for some religious differences across states that are fixed, they may also control for other fixed effects (e.g., hot weather, which could increase beer consumption) and for various nonreligious characteristics of the population that could influence attitudes toward alcohol (e.g., ethnic background). The religious and dry county variables vary over time, however, and results in Table 5 suggest that they do not control for the same factors that the fixed effects do (dry counties can also be considered as a determinant of availability, e.g., as noted above and the significance of the religion variables suggests that within-state variation is substantial). Correlation coefficients between state dummies and the religion and dry county variables reinforce this point, as almost all of the coefficients are less than 0.10 in absolute value. Thus, any changes in the coefficients for policy variables that arise with the addition of these variables apparently implies that the coefficients in columns 3 and 4 are biased and the coefficients in the final column are likely to be more reliable. Indeed, only one state, Utah, has a correlation coefficient greater than 0.5 in absolute value between the state dummies and any of the religious and dry county variables. The correlation coefficient between the Utah dummy and the Mormon variable is 0.92. Estimating the regression in columns 6 but without the Mormon variable produces a positive tax coefficient, although it is small (0.278) and insignificant (t-ratio of 0.096). The beer-tax coefficient is the only one that differs in sign from those reported in column 6, all other coefficients have comparable magnitudes, and significance levels remain the same except for the distance to major brewery variable, which is marginally significant (at the 0.10 level) in column 6 but not in this alternative regression.

32 Mast (1996) estimates beer tax elasticities for 18-20 year olds with BAC 0.01 + to be -0.03, a result similar to Dee's (1999) estimate of -0.026. Results for 21-24 year olds were about the same as for all drivers. Also see Young and Likens (1998) for similar findings.

References

Asch, Peter, and David T. Levy. 1990. Young driver fatalities: The role of drinking age and experience. Southern Economic Journal 57:512-20.

Benson, Bruce L., Brent D. Mast, and David W. Rasmussen. 1999a. Can police deter drunk driving? Applied Economics. In press.

Benson, Bruce L., Brent D. Mast, and David W. Rasmussen. 1999b. Deterring drunk driving fatalities: An economics of crime perspective. International Review of Law and Economics 19:205-25.

Beverage world databank. 1986-1993. Beverage world databank. Dayton, OH: Keller International.

Bradley, Martin B., Norman M. Green, Jr., Dale E. Jones, Mac Lynn, and Lou McNeil. 1990. Churches and church membership in the United States 1990. Atlanta: Glenmary Research Center.

Brown, Robert W., R. Todd Jewell, and Jerrell Richer. 1996. Endogenous alcohol prohibition and drunk driving. Southern Economic Journal 62:1043-53.

Carstensen, Peter C., and Richard F. Dahlson. 1986. Vertical restraints in beer distribution: A study of the business justifications for and legal analysis of restricting competition. Wisconsin Law Review 1986:1-81.

Chaloupka, Frank J. 1993. Effects of price on alcohol-related problems. Alcohol Health & Research World 17:46-53.

Chaloupka, Frank J., Henry Saffer, and Michael Grossman. 1993. Alcohol control policies and motor-vehicle fatalities. Journal of Legal Studies 22:161-86.

Chaloupka, Frank J., and Henry Wechsler. 1996. Binge drinking in college: The impact of price, availability, and alcohol control policies. Contemporary Economic Policy 14:112-24.

Coate, P. J., and Michael Grossman. 1988. Effects of alcoholic beverage prices and legal drinking ages on youth alcohol use. Journal of Law and Economics 31:145-71.

Cook, Philip J. 1980. Research in criminal deterrence: Laying the groundwork for the second decade. In Crime and justice: An annual review of research 2, edited by Norval Morris and Michael Tonry. Chicago: University of Chicago Press, pp. 211-68.

Cook, Philip J. 1981. The effect of liquor taxes on drinking, cirrhosis, and auto accidents. In Alcohol and public policy: Beyond the shadow of prohibition, edited by Mark H. Moore and Daniel R. Gerstein. Washington, DC: National Academy Press, pp. 255-85.

Cook, Philip J., and Michael J. Moore. 1993a. Economic perspectives on reducing alcohol-related violence. In Alcohol and interpersonal violence: Fostering multidisciplinary perspectives, edited by S. E. Martin. Washington, DC: National Institute on Alcohol Abuse and Alcoholism, pp. 193-212.

Cook, Philip J., and Michael J. Moore. 1993b. Violence reduction through restrictions on alcohol availability. Alcohol Health Research World 17:151-6.

Cook, Philip J., and G. Tauchen. 1984. The effects of minimum drinking age legislation on youthful auto fatalities, 19701977. Journal of Legal Studies 13:169-90.

Culbertson, W. Patton, and David Bradford. 1991. The price of beer: Some evidence from interstate comparisons. International Journal of Industrial Organization 9:275-89.

Dee, Thomas S. 1999. State alcohol policies, teen drinking and traffic fatalities. Journal of Public Economics 72:289-315.

Dumond, J. Michael, Barry T. Hirsch, and David A. Macpherson. 1999. Wage differentials across labor markets and workers: Does cost of living matter? Economic Inquiry. In press.

Evans, William N., Doreen Neville, and John D. Graham. 1991. General deterrence of drunk driving: Evaluation of recent American policies. Risk Analysis 11:279-89.

Grossman, Michael, Judy L. Sindelar, John Mullahy, and Richard Anderson. 1993. Policy watch: Alcohol and cigarette taxes. Journal of Economic Perspectives 7:211-22.

Hausman, Jerry A. 1978. Specification tests in econometrics. Econometrica 46:1251-71.

Jordan, W. John, and Bruce L. Jaffee. 1987. The use of exclusive territories in the distribution of beer: Theoretical and empirical observations. Antitrust Bulletin 32:137-64.

Kenkel, Donald S. 1993. Drinking, driving, and deterrence: The effectiveness and social costs of alternative policies. Journal of Law and Economics 36:877-913.

Kenkel, Donald S. 1996. New estimates of the optimal tax on alcohol. Economic Inquiry 34:296-319.

Klein, Terry. 1986. A method for estimating posterior BAC distribution for persons involved in fatal traffic crashes. Washington, DC: U.S. Department of Transportation, National Highway Traffic Safety Administration.

Laixuthai, Adit, and Frank J. Chaloupka. 1993. Youth alcohol use and public policy. Contemporary Policy Issues 11: 70-81.

Maddala, G. S. 1983. Limited-dependent and qualitative variables in econometrics. New York: Cambridge University Press.

Mast, Brent D. 1996. Deterring alcohol related traffic fatalities: The role of alcohol consumption, enforcement, and punishment. Ph.D. dissertation, Florida State University, Tallahassee, FL.

McMahon, Walter W., and Chao-Chung Chang. 1991. Geographic cost of living differences: Interstate and intrastate, update 1991. MacArthur/Spencer Series No. 20.

Modern Brewery Age blue book. 1984-1992. Modern Brewery Age blue book. East Norwalk, CT: Modern Brewery Age.

Mullahay, John, and Jody L. Sindelar. 1994. Do drinkers know when to say when? An empirical analysis of drunk driving. Economic Inquiry 32:383-94.

O'Donnell, M. A. 1985. Research on drinking locations of alcohol-impaired drivers: Implications for prevention policies. Journal of Public Health Policy 6:510-21.

Ornstein, S. I., and Dominique M. Hanssens. 1985. Alcohol control laws and the consumption of distilled spirits and beer. Journal of Consumer Resources 12:200-13.

Phelps, Charles E. 1988. Death and taxes: An opportunity for substitution. Journal of Health Economics 7:1-24.

Quinn, Bernard, Herman Anderson, Martin Bradley, Paul Goetting, and Peggy Shriver. 1980. Churches and church membership in the United States 1980. Atlanta: Glenmary Research Center.

Rasmussen, David W., and Bruce L. Benson. 1994. The economic anatomy of a drug war: Criminal justice in the commons. Lanham, MD: Rowman and Littlefield.

Ruhm, Christopher J. 1995. Economic conditions and alcohol problems. Journal of Health Economics 14:583-603.

Ruhm, Christopher J. 1996. Alcohol policies and highway vehicle fatalities. Journal of Health Economics 15:435-45.

Saffer, Henry, and Michael Grossman. 1987a. Beer taxes, the legal drinking age and youth motor vehicle fatalities. Journal of Legal Studies 16:351-74.

Saffer, Henry, and Michael Grossman. 1987b. Drinking age laws and highway mortality rates: Cause and effect. Economic Inquiry 25:403-17.

Sass, Tim R., and David S. Saurman. 1993. Mandated exclusive territories and economic efficiency: An empirical analysis of the malt-beverage industry. Journal of Law and Economics 36:153-77.

Sloan, Frank A., and Penny B. Givens. 1994. Drinking, driving, and the price of automobile insurance. The Journal of Risk and Insurance 61:33-58.

Sloan, Frank A., Bridget A. Reilly, and Christoph Schenzler. 1994a. The effects of prices, civil and criminal sanctions, and law enforcement on alcohol related mortality. Journal of Studies on Alcohol 55:454-65.

Sloan, Frank A., Bridget A. Reilly, and Christoph Schenzler. 1994b. Tort liability versus other approaches for deterring careless driving. International Review of Law and Economics 14:53-71.

Sloan, Frank A., Bridget A. Reilly, and Christoph Schenzler. 1995. Effects of tort liability and insurance on heavy drinking and drinking and driving. Journal of Law and Economics 38:49-77.

U.S. Brewers Association. 1984-1992. Brewers almanac. Washington, DC: Beer Institute.

U.S. Bureau of the Census. 1983-1995. Statistical abstract of the United States. Washington, DC: Government Printing Office.

U.S. Department of Commerce. 1984-1994. Census of service industries. Geographic area series. Washington, DC: Government Printing Office.

U.S. Department of Commerce. 1986-1992. Public employment. Washington, DC: Government Printing Office.

U.S. Department of Health and Human Services. 1988. Surgeon General's Workshop on Drunk Driving: Proceedings. Washington, DC: Public Health Service.

U.S. Department of Transportation. 1985-1993. Digest of state alcohol-highway safety related legislation. Washington, DC: Government Printing Office.

U.S. Department of Transportation. 1985-1992. Fatal accident reporting system. Washington, DC: Government Printing Office.

U.S. Department of Transportation. 1984-1992. Highway statistics. Washington, DC: Government Printing Office.

U.S. Department of Transportation. 1992. Traffic safety facts. Washington, DC: Government Printing Office.

U.S. President. 1984-1992. Economic report of the President transmitted to Congress. Washington, DC: Government Printing Office.

Wieczorek, W., B. Miller, and T. Nochajski. 1989. Bar versus home drinkers: Different subgroups of problem-drinker drivers. New York State Division of Alcoholism and Alcohol Abuse Research Note 89-6.

Young, Douglas A., and Thomas W. Likens. 1998. Alcoholic beverage taxes and auto fatality rates. Unpublished paper, Montana State University.