Do drinkers know when to say when? An empirical analysis of drunk driving.

Mullahy, John ; Sindelar, Jody L.

Drunkenness, for example, in ordinary cases, is not a fit subject for legislative interference; but I should deem it perfectly legitimate that a person, who had once been convicted of any act of violence to others under the influence of drink, should be placed under a special legal restriction, personal to himself; that if he were afterwards found drunk, he should be liable to a penalty, and that if and when in that state he committed another offence, the punishment to which he would be liable for that other offence should be increased in severity.

John Stuart Mill, On Liberty

I. INTRODUCTION

It is estimated that about 40 percent of the U.S. population will be involved in an alcohol-related traffic accident sometime during their lifetime, according to Vegega and Klein [1990]. Traffic accidents are the leading cause of death in the U.S. for persons aged one to thirty-four, and during 1989, approximately one-half of the 45,555 traffic fatalities in the U.S. were estimated to be alcohol-related.

It is clear that the negative externalities drunk drivers impose on others account for a large part of the interest in reducing drunk driving. The enormous social interest in reducing the adverse consequences of drunk driving is apparent in the mass media as well as in the scholarly literature in economics and other disciplines.(1) Reducing drunk driving and its attendant social costs is a widely articulated goal of public policies and of groups like MADD and SADD as well as major insurance companies.

Public policies have, in large measure, adopted the economic view that a more stringent penalty structure, by increasing the expected "full price" of drunk driving, should reduce drunk driving and its attendant externalities. Such policies include license revocation, fines, and imprisonment as well as strategies to increase the social stigma attached to drunk driving, the costs of purchasing alcoholic beverages, and the awareness of these costs.

The objective of this paper is to analyze the determinants of drunk driving behavior from an economic perspective.(2) Our interest is in the roles of key sociodemographic factors (e.g., schooling, family structure, race, sex, etc.) and various public policies directly or indirectly oriented toward reducing drunk driving behavior. By characterizing such behavior as an economic choice, we describe a "demand for drunk driving" and empirically estimate its determinants. To this end, we utilize data from the 1988 National Health Interview Survey, which surveys individuals about "driving after drinking too much." Such individual data provide what we feel is an interesting alternative to the state-level data often used to analyze drunk driving by, for example, Chaloupka, Grossman, and Saffer [1993].(3) Our empirical results largely confirm the major conclusion of studies that have used state-level data with respect to driving while intoxicated (DWI): the demand for drunk driving is negatively related to its full price.

This paper proceeds as follows. Section II presents a conceptual model of an individual's decision to drive drunk. Section III describes the data. Section IV presents estimates of the determinants of individuals' propensities to "drive after drinking too much." Section V summarizes our results. The appendix discusses possible reporting biases in drunk driving behavior and their implications for inferences.

II. CONCEPTUAL BACKGROUND

Becker [1968] revolutionized social scientists' thinking about criminal behavior by suggesting that even the decision to engage in illegal behavior is likely to be based on a rational comparison of expected marginal costs and benefits. The expected "full price" of such illegal or criminal behavior is posited to be negatively related to this behavior. The full price depends on the pecuniary and psychic costs relating to crime. It is in this intellectual tradition that we undertake an analysis of the economic determinants of drunk driving qua illegal behavior.

The structure of the criminal justice system thus affects, in part, the decision to drive drunk. The probability of apprehension and conviction as well as the severity of the penalties (e.g., fines and/or prison sentences) affect the decision to drive drunk. In addition, social norms, as described by Coleman [1990], that establish acceptable social behavior, proscribe certain activities (e.g., drunk driving), and affect psychic costs of deviating from the norms, influence driving while intoxicated. Personal characteristics may also affect the full price and thus the decision to drive drunk. For example, the relevant social norms as well as the impact of the stigma may depend on one's education, race, and gender.

The following simple consumer choice model suggests the basic economic elements of our empirical analysis.(4) An individual is assumed to have preferences representable by a utility function

(1) U = U(DD, S, X; Z),

whose arguments are driving while drunk (DD), the "services" provided by the use of alcoholic beverages (S), a Hicksian composite commodity (X), and a vector of household and individual characteristics (Z). This formulation of the individual's maximand follows Lancaster [1966] and Becker [1965] in specifying produced commodities as the ultimate source of consumer satisfaction. Goods, services, and time are the inputs that produce the commodities. Demands for the inputs are derived from the demands for these commodities.

The marginal utility of DD, driving while drunk, could be either positive or negative; for instance, some may consider drunk driving an unwelcome component of otherwise desirable activities (e.g., drinking with friends at a bar) while others may derive direct satisfaction from the process of drunk driving (e.g., a joy ride). The marginal utility of S, the "service" provided by alcohol, is likely to be positive or zero. The commodities DD and S are not purchased in the market, but rather are jointly produced according to some (perhaps complex) transformation function

(2) T(DD, S, A, DR; Z) = O,

where the level of consumption of alcoholic beverages (A) and time spent driving a motor vehicle (DR) are "inputs" and drunk driving (DD) and intoxication (S) are the "outputs."

The individual is assumed to exhaust his or her monetary budget (B) by expenditures on alcohol consumption, driving, and other commodities (A, DR, X), and any financial penalties that are associated with drunk driving. The prices of these variables are given by [P.sub.X], [P.sub.A],(5) and [P.sub.DR].(6) The expected full price of one unit of drunk driving is [[Pi].sub.DD], whose level is determined by, among other things, the dollar value of fines, the probability of apprehension and conviction, and the social stigma and other psychic costs associated with DWI arrests and convictions.

Consumer optimization results, among other things, in a reduced-form choice function for drunk driving:

(3) DD = DD([P.sub.X], [P.sub.A], [P.sub.DR], [[Pi].sub.DD], B; Z),

which is a hybrid of preference and production relationships. The empirical analysis reported below focuses on the decision to drive drunk; this decision is considered to arise from individuals making constrained choices while facing expected "full prices" for different activities.

The probability of having an accident and getting injured, and state drunk driving policies and their implementation, among other things, affect the expected "full price" of drunk driving.(7) Individuals' knowledge of these full prices can come from a variety of sources, including one's own experience, experiences related by others, and the media. Media coverage would include public service announcements (e.g., "Friends don't let friends drive drunk," "Know when to say when," etc.) as well as news reports on the enactment of or changes in laws.

Ross [1982; 1990] has suggested that knowledge of laws and penalties is heightened when changes in laws are made. Considerable media coverage of the political process and debate typically accompany changes in laws, as do public service announcements to inform the population of the changes. Ross has hypothesized, and empirically confirmed to some degree, that changes in laws have a greater deterrent effect than the mere existence of laws, presumably because the changes shock individuals' information sets.(8)

III. DATA: 1988 NATIONAL HEALTH INTERVIEW SURVEY

The 1988 National Health Interview Survey, a stratified, multi-stage probability sample of the U.S. population, supplied the data used in this study.(9) The 1988 National Health Interview Survey includes a detailed Alcohol Survey funded and structured by the National Institute on Alcohol Abuse and Alcoholism. The Alcohol Survey obtained data from over 43,000 individuals on drinking patterns, symptoms of alcoholism, family history of problem drinking or alcoholism, as well as standard socioeconomic and demographic data.

The Alcohol Survey asked each respondent considered a current drinker if he or she had "driven a car after having too much to drink" in the past year.(10) We used this information to form a binary driving and drinking variable (i.e., DRIVE DRUNK = 1); 24 percent of male drinkers and 12 percent of female drinkers reported having driven after drinking too much over the previous twelve months.(11) It is important to stress that the individuals surveyed determined what constitutes "drinking too much." The appendix considers the implications of such behavior being underreported.

Socioeconomic and demographic variables are hypothesized to affect driving while intoxicated. Table I displays the definitions of these and the drunk driving policy variables used in the empirical analysis; Table II presents sample descriptive statistics. Our analyses are conducted on current drinkers only.(12)

TABLE I

Variable Definitions

DRIVE DRUNK: In past twelve months, driven a car after
 having had too much to drink (1/0 dummy
 variable)

BEER TAX: State beer tax, in cents

FINE: Mandatory minimum fine, DWI conviction,
 first offense, in dollars

REVOKE: Implied consent/mandatory licensing
 revocation, first offense (1/0 dummy variable)

LAW CHANGE: 1 if any state law related to drunk driving
 changed in past year, 0 else

APPARENT ETHANOL: State apparent ethanol consumption, per
 capita

LIVED WITH ALCOHOLIC: 1 if lived with a problem drinker/alcoholic
 during first eighteen years, 0 else

AGE: Age in years

WHITE: 1 if white, 0 else

MARRIED: 1 if married, 0 else

EDUCATION: Years of completed schooling

WORKS: 1 if main activity is working, 0 else

FAMILY SIZE: Size of family

VIETNAM VETERAN: 1 if Vietnam-era veteran, 0 else

Matching state-level data on beer taxes, apparent ethanol consumption, and drunk driving-related regulations augmented the individual level data. The variations in state policy instruments designed to reduce the adverse effects of drunk driving allowed us to explore empirically their effects on reported drunk driving.

On the basis of the hypotheses discussed at the end of section II, the policy variables used in the econometric analysis are FINE, REVOKE, and LAW CHANGE, defined in Table I.(13) The variable REVOKE comes from the implied consent and mandatory minimum licensing action laws that can result in a driver's license being suspended or revoked.(14) The variable REVOKE indicates whether the state has greater than a thirty-day minimum for the number of days a license is revoked for the first offense. The variable FINE is the minimum mandatory pecuniary fine that must be imposed after the first offense.(15)

TABLE II

Main Estimation Samples: Descriptive Statistics

 Males (N = 6,865) Females (N = 7,934)
Variable Mean Min Max Mean Min
Max

DRIVE DRUNK(*) .245 0 1 .124 0 1
BEER TAX 15.7 2 81 15.4 2 81
FINE 121.6 0 575 120.7 0 575
REVOKE .227 0 1 .228 0 1
LAW CHANGE .180 0 1 .177 0 1
APPARENT ETHANOL 2.03 1.01 3.99 2.02 1.01 3.99
LIVED WITH ALCOHOLIC .175 0 1 .223 0 1
AGE 41.5 18 94 40.7 18 94
WHITE .862 0 1 .868 0 1
MARRIED .454 0 1 .464 0 1
EDUCATION 13.2 0 18 13.2 0 18
WORKS .740 0 1 .603 0 1
FAMILY SIZE 1.98 1 10 2.32 1 14
VIETNAM VETERAN .108 0 1 - - -

* The means for drunk driving by gender and race are as follows: .255 for
white males, .155 for nonwhite males, .060 for nonwhite females, and .131 for
white females.

The binary variable LAW CHANGE indicates whether any substantive changes in the drunk driving laws occurred from 1987 to 1988. During this period, ten states changed at least one law. Some states (Kansas, for instance) implemented many changes between 1987 and 1988. Minimum licensing sanctions were the most frequently changed laws, followed by administrative per se laws.

Among our policy measures we also include the state excise tax on beer (BEER TAX), which affects the full price of drinking and driving. While not a policy designed to deter drunk driving, mounting evidence suggests the beer tax may have important indirect effects on drunk driving behavior as an inhibitor of consumption, particularly among younger individuals, as documented in Grossman [1989], and Saffer and Grossman [1987]. State per capita apparent consumption of ethanol (APPARENT ETHANOL) is also used as a control variable on the assumption that an individual's "alcohol consumption environment" may have important effects on drunk driving.

IV. EMPIRICAL RESULTS

We report here estimates of the models of drunk driving propensity (DRIVE DRUNK) with the objective of uncovering both the basic sociodemographic and policy determinants of such reported behavior. The models are estimated by probit regression for separate samples of males and females, with the results presented in Table III.(16)

Column 1 displays the basic model's results for males. Here it is seen that the estimated coefficients on three of the key policy variables, FINE (the minimum mandatory fine for an offense), BEER TAX (the state excise tax on beer), and REVOKE (mandatory license revocation for a drunk driving offense), all have negative coefficients, suggesting "downward-sloping demands" for drunk driving. That is, all three policy instruments raise the expected full price of drunk driving. All are significant by conventional standards;(17) the coefficients on FINE, BEER TAX, and REVOKE are jointly significant at p[is less than].0001 with a [[[Chi].sup.2].sub.(3)] statistic of 21.2. The publicity effect of changes in legislation appears unimportant: the LAW CHANGE coefficient estimate has the "wrong" sign, but is not significant either in this model or in any of the others reported below;(18) the APPARENT ETHANOL (per capita consumption of alcohol) estimate has the expected sign, but it is also not significant in either this specification or most other specifications estimated.(19)

The results for the sociodemographic variables for males, all of which are significant at conventional levels, indicate that one's age, martial status, family size and level of eduction are negatively related to reported drunk driving, as would be expected. On the other hand, being white, a Vietnam veteran, employed and living with an alcoholic are positively related to drunk driving.(20) While theory cannot lead to sharp predictions about the relationships between these variables and the propensity to drive while drunk, all the estimated signs seem reasonable. Living with an alcoholic, for example, may increase the propensity to drive drunk by lessening the individual's perception of the stigma attached to drunk driving. Working may increase drunk driving by raising the "exposure," e.g., drinking on the way home from work.

Column 4 of Table III presents the corresponding results for females. Despite the known different drinking patterns of males and females (and the possibly different driving patterns), the point estimates of every coefficient, except the constant, have the same signs as those in the corresponding model for males. The point estimates of the coefficients on drunk-driving fines and beer taxes are strikingly similar, albeit less significant for females. The sociodemographic variables are all significant again and are of the same sign and of reasonably similar magnitude for females as for males.(21)

Given the significance of the white race coefficient estimate for both males and females, we split the sample and estimated separate models for nonwhites and whites.(22) These results are reported in columns 2 and 3 of Table III for males and in columns 5 and 6 for females. It is particularly interesting that, for both sexes, the point estimates of the policy instruments beer taxes, license revocation, and legislative changes in drunk driving laws are significant (or nearly so) for nonwhites while for whites fines are also significant; license revocation is significant for male whites also.

Among the sociodemographic variables, living with an alcoholic and age are significantly related to drunk driving for TABULAR DATA OMITTED both sexes and for whites and nonwhites alike; living with an alcoholic makes one more likely to drive drunk whereas increasing age makes one less likely to. Being a Vietnam veteran has a positive and significant effect on males of both races. The remainder of the sociodemographic variables are always significant for whites of each sex, but rarely significant for nonwhites. The lower levels of significance for nonwhites may be attributable to different underlying relationships or to their smaller sample sizes in combination with their lower propensity to drive drunk. Whites are nearly twice as likely to drive drunk as are nonwhites.

V. SUMMARY

This paper has utilized a unique data set containing information on self-reported drunk driving as well as other individual-level data. We matched this data to various state-level driving while intoxicated (DWI) policy instruments that both directly as well as indirectly (i.e., the excise tax on beer) increase the expected "full price" of drunk driving. Our estimates take advantage of data quite different from the state-level data usually used in studies of the impact of driving while intoxicated laws.

We show that socioeconomic and demographic variables as well as drunk driving laws affect the decision to drive drunk. We find some evidence that racial differences play a part in how some socio-economic and demographic variables affect the decision to drive drunk. Furthermore, although the effects of the policy variables are qualitatively similar by sex, they differ in interesting ways by race.

Our results confirm others' findings that state-level policy variables are significant in deterring drunk driving: the demand curve for drunk driving is negatively sloped. Taken together, the evidence from household surveys and state data provides a clearer understanding of policy variables' impacts on drunk driving and gives policymakers a clearer indication of the most effective policy instruments in preventing drunk driving.

While results from state-level data have a potential for bias, our results too must be viewed cautiously because the drunk driving is self-reported. (See the appendix for a discussion of the potential for self-reporting bias.) That the concept of downward sloping demand for drunk driving depends to some degree on an individual's knowledge of the full price of drunk driving provides another caveat to our results. Individuals may have information gaps about the potential social consequences of driving while intoxicated.(23) Whether policies to reduce information failures may be productive in deterring drunk driving may be a productive line of future research.

APPENDIX

Implications of Potential Reporting Bias

Drunk driving, like many behaviors such as tobacco use, welfare receipt, etc., may be socially stigmatized. Thus it is certainly plausible that individuals may misreport their drunk driving behavior. We sketch a simple model that describes the propensity to report "driving after drinking too much" and suggest some implications of such potential underreporting for our empirical results.

Define [y.sub.D] to be the binary variable indicating whether or not the individual actually believed he "drove after drinking too much" and [y.sub.R] to be the binary variable indicating whether the individual reports "driving after drinking too much." The variables [y.sub.D] and [y.sub.R] will have some joint distribution in the population, for which the probability of an individual reporting that he did not drive after drinking too much is

(A1) PR([y.sub.R] = 0) =

PR([y.sub.R] = 0 [where] [y.sub.D] = 0)Pr([y.sub.D] = 0)

+ Pr([y.sub.R] = 0 [where] [y.sub.D] = 1)[1 - Pr([y.sub.D] = 0)].

The econometric analysis and subsequent interpretation of results for purposes of policy formulation would be simplified greatly if Pr([y.sub.R] = 0 [where] [y.sub.D] = 1) = 0, i.e., people were completely honest in their reporting. Clearly, this cannot be maintained, however. It might be reasonable to maintain that Pr([y.sub.R] = 0 [where] [y.sub.D] = 0) = 1, i.e., if an individual does not drive after drinking too much, then he will not report having done so. This would give

(A2) Pr([y.sub.R] = 0) = Pr([y.sub.D] = 0)

+ Pr([y.sub.R] = 0 [where] [y.sub.D] = 1)[1 - Pr([y.sub.D] = 0)],

the sum of the true effect whose determinants would be of primary interest (the first term) and what might be thought of as the biased reporting effect that results from stigma or whatever (the second term). The data contain information only on [Y.sub.R], reported drunk driving, so that all one can hope to identify is the sum of the two right-hand-side terms.

It is interesting to speculate as to how any determinant of these two effects might affect the probabilities in equation (A2). For instance, consider some policy instrument x (e.g., severity of drunk driving punishment) that is likely to have a positive effect on the probability of driving drunk: Pr([y.sub.D] = 0), i.e., [Delta]Pr([y.sub.D] = 0) / [Delta]x [is greater than] 0. It is plausible that the magnitude of such a policy instrument may also affect the probability of reporting having driven drunk, given that one has done so, Pr([y.sub.R] = 0 [where] [y.sub.D] = 1), and do so in the same direction, i.e., [Delta]Pr([y.sub.R] = 0 [where] [y.sub.D] = 1) / [Delta]x [is greater than] 0. Individuals residing in states with strict drunk driving penalties may, for example, have lower drunk driving propensities and sense a greater stigma against driving while intoxicated; they may therefore, be relatively less likely to report engaging in such behaviors.

If true, then estimates of a model of reported drunk driving, Pr([y.sub.R] = 0), would yield estimates of

(A3) [Delta]Pr([y.sub.R] = 0) / [Delta]x =

[1 - Pr([y.sub.D] = 0)] x [[Delta]Pr([y.sub.R] = 0 [where] [y.sub.D] = 1) / [Delta]x]

+ [1 - Pr([y.sub.R] = 0 [where] [y.sub.D] = 1)] x [[Delta]Pr([y.sub.D] = 0) / [Delta]x]

where both right-hand-side terms would have the same sign since Pr([center dot]) [is less than] 1. Thus, failure to reject a null hypothesis like [Delta]Pr([y.sub.R] = 0) / [Delta]x = 0 should give one confidence that the policy effect of interest, [Delta]Pr([y.sub.D] = 0) / [Delta]x, is statistically unimportant. Conversely, rejecting a null of [Delta]Pr([y.sub.R] = 0) / [Delta]x = 0 would suggest only that either one or both of the effects are important, not which one(s).

There is, however, at least one additional source of potentially important error. That is, regardless of what an individual reports to the survey interviewer, just because he believes he has or has not driven after drinking too much does not mean that the individual's actual behavior is consistent with any "objective" standards of driving after drinking too much (e.g., the blood alcohol content level). Moreover, one can certainly imagine that the discrepancy between belief and "fact" may depend on sociodemographic factors as well as policy instruments; e.g., a media blitz about new state DWI laws enhances awareness of the causes and effects of drunk driving.(24) This kind of error may well introduce biases that work in an opposite direction from the "stigma" biases discussed above, thus confounding further the results interpretation.

12. From the set of current drinkers, we selected only those who responded to the questionnaire themselves; that is, we eliminated those who had proxy household members respond for them for any part of the questionnaire. Seventy-six percent of the Alcohol Survey respondents self-reported to the entire supplement; 74 percent of current drinkers self-reported entirely (65 percent of the males and 84 percent of the females). After observations with key data missing were deleted, the resulting sample included 6,865 males and 7,934 females.

13. The tax and apparent consumption data come from Public Revenues from Alcohol Beverages: 1988, published by the Distilled Spirits Council of the U.S. The drunk driving regulation data are from Digest of State Alcohol-Highway Safety Related Legislation, published by the National Highway Traffic Safety Administration (NHTSA), and refer to laws as of January 1989 (i.e., laws in place in 1988). A detailed description of key variables from the NHTSA dataset is available upon request from the authors. Under an agreement with the National Center for Health Statistics, these data were matched to approximately 95 percent of the individual Health Interview Survey observations.

14. Implied consent laws presume that by acquiring a driver's license from the state, the driver has implicitly agreed to be chemically tested for alcohol (or drugs) by the police or else will have his/her license revoked or suspended. As of January 1989, thirty-nine states had laws that set the minimum number of days that an individual's license would be revoked or suspended upon first refusal to submit to a police test. More states had minimums for the second refusal and the penalties are typically more severe for the second refusal.

15. Fines typically increase after the first offense. These minimum mandatory sanctions are established by legislation and cannot be reduced by court discretion.

16. Income variables were included in some preliminary runs, but were never found to be significant. The results reported here exclude income, but interested readers can obtain the preliminary results upon request.

17. Following the suggestion of a referee, we found that the coefficient of the beer tax variable increased in both absolute value and significance when per capita consumption was omitted. This could occur because of the negative correlation between beer taxes and per capita consumption. Results are available upon request from the authors.

18. For the full sample results for both males and females, the failure to reject the null hypothesis that the [[Beta].sub.j] on the LAW CHANGE variable is zero begs the question about the power of the asymptotic t-test used to draw such an inference. Recently, Andrews [1989] has proposed a methodology that facilitates power calculations. One interesting question is how large is the interval of the true parameter (say [Theta]) such that the t-test used to test the null [H.sub.0]:[Theta] = 0 (at significance level [Alpha]) is just as likely to accept the null as to reject it. In the context of the LAW CHANGE estimates [Mathematical Expression Omitted] for males and [Mathematical Expression Omitted] for females; the formulae presented by Andrews indicate that for any true [[Beta].sub.j] in the interval [-.104, .104] for males or [-.116, .116] for females the data are such that there is a 50-50 chance that a two-sided test with significance level [Alpha] = .05 will fail to reject the null hypothesis [H.sub.0]:[[Beta].sub.j] = 0. Thus, true parameters having potentially significant magnitudes can easily fail to be "uncovered" by the standard t-tests conducted here.

19. Given the large number of state-level drunk driving regulations available in the National Highway Traffic Safety dataset, it is surely apparent to the reader that the specifications reported in Table III (that use three such measures) are not the first and only specifications we imagined estimating. Nonetheless, while we confess to some specification searching, we might note that the search was more along the lines of an "include the kitchen sink then use a 'rolling F-statistic' to eliminate insignificant variables" search than it was a "consider all estimable combinations of the policy variables and report only the results that conform with priors" search.

20. See Browning and Meghir [1991] for an interesting discussion of the use of labor supply variables as explanatory variables in commodity demand functions. For analyses of the alcoholism/labor market connection, see Mullahy and Sindelar [1989; 1991; 1993].

21. Given the similarity of the results for males and females, our concern that the results spurious is lessened to some degree. However, in both cases the results may still be fragile in that influential observations may be exerting disproportionate influence on the parameter estimates. To assess this possibility, we reestimated each model as a linear probability model using OLS and conducted a DFBETAS analysis for both males and females on the beer tax and fines variables using the statistical package STATA's dbeta proc (the DFBETAS for license revocation were quite similar insofar as identification of influential observation(s) is concerned). While the OLS results do not necessarily guarantee anything about the probit results, it would be surprising if qualitatively different conclusions emerged in the probit estimates.

This analysis showed that one observation in the males' sample and two observations in the females' sample were potentially influential observations. Positive DFBETAS for each suggest that the point estimate will decrease if the observation is deleted. We thus anticipate that these influential observations would dampen, not inflate, the estimated magnitudes of the key policy variables. Accordingly, the probit models of columns 1 and 4 were reestimated with these observations deleted. The absolute magnitudes of the fines, beer taxes, and license revocation estimates did, indeed, increase with these observations deleted, with beer taxes now significant for males and nearly so for females:

 Males Females

 Original Outlier(s) Original Outlier(s)
 Result Removed Result Removed

FINE -0.00039 -0.00042 -0.00037 -0.00041
 (3.270) (3.461) (2.743) (3.043)

BEER TAX -0.0029 -0.0034 -0.0024 -0.0033
 (1.922) (2.190) (1.438) (1.873)

REVOKE
LICENSE -0.1549 -0.1615 -0.0576 -0.0604
 (3.357) (3.488) (1.110) (1.329)

The other estimates are largely unaffected by this exercise.

1. A brief list of some of the economics literature would include Chaloupka et al. [1993], Cook [1991], Kenkel [1993a; 1993b], Manning et al. [1989], Phelps [1987; 1988; 1990], Pogue and Sgontz [1989], and Wilkinson [1987]. See USDHHS [1990] for an overview and additional references.

2. It should be noted that the "driving" considered here is land-based motor vehicle driving. We do not consider the interesting phenomena of drunk driving on the water (recall the Exxon Valdez disaster), on the rails (recall the New York City subway calamity), or in the air (recall the recent case of the Northwest Airlines pilots; also, see Modell and Mountz [1990]).

3. State-level and individual sources of data are distinctly different, each having strengths and biases. Individuals are likely to underreport their drunk driving in household surveys, as discussed further in the appendix. In contrast, state-level data on drunk driving are recorded only when drunk driving results in a fatality; thus, a large portion of drunk driving is missed, because only a small percentage of such behavior ends in a fatality. Our data thus offer an alternative source and type of information to that economists typically use to analyze such phenomena.

4. Kenkel [1993a] provides a detailed discussion of why it is fruitful to examine the phenomenon of drunk driving from the perspective of two choice margins: the demand for alcohol and the demand for driving while under the influence of alcohol. Since timing and the interplay between drinking and driving play such important roles in the analysis of drunk driving, a fairly elaborate model would have to be worked out to accommodate these considerations. The framework described here should be viewed simply as an approximation of such a detailed model.

5. If we were analyzing alcoholism itself, its addictive nature would have to be considered, as in Michaels [1988]. Past and future prices of alcoholic beverages, for example, might then be included, as in Becker and Murphy [1988] and Becker, Grossman, and Murphy [1991]. However, the concern here is with drunk driving, and both alcoholics and non-alcoholics drive drunk.

6. The empirical analysis below ignores [P.sub.DR] due to lack of data.

7. As pointed out by Jacobs [1989], the impact of these laws and the associated penalties may be small relative to the bigger risk of an accident and the attendant financial and health losses.

8. Cognitive dissonance theory, as discussed in Akerlof and Dickens [1982] provides some basis for such a hypothesis. "Shocking" individuals out of their "steady state" preconceptions is one factor that motivates substantial behavioral changes. For a related discussion, see Steele and Josephs [1990].

9. The National Health Interview Survey's core survey gathers data on socioeconomic characteristics and health conditions for about 49,000 households, yielding approximately 150,000 individual observations. One adult from each family was randomly selected to receive the Alcohol Survey, resulting in 43,809 observations.

10. "Current drinker" is defined by the National Center for Health Statistics as individuals who have had twelve or more drinks in the past year. Of the sample of 43,809 individuals, about half are current drinkers (11,727 males and 10,375 females).

11. As a basis of comparison, a 1977 Gallup poll asked two similar questions of those who drove: (1) Do you ever drive after drinking alcoholic beverages? (2) Have you ever driven when you thought that you had too much to drink to drive safely? The percentages of those questioned who responded positively are shown below by gender and overall response rate:

 Male Female All

1. Drink and Drive 48% 22% 36%
2. Drink Too Much
and Drive 26% 8% 18%

A similar set of questions asked in 1982 revealed that 40 percent of those surveyed reported that they had consumed at least some alcohol and then had driven, and 18 percent said they had driven after drinking too much, as found in Gallup [1978, 1983].

JOHN MULLAHY, Associate Professor, Trinity College, National Bureau of Economic Research and Resources for the Future

JODY L. SINDELAR, Associate Professor, Yale School of Public Health, Institution for Social and Policy Studies, Yale University and National Bureau of Economic Research.

22. The likelihood ratio test statistics for the null hypothesis that all parameters (including the constant term) are the same for nonwhites and whites are 76.98 for males and 114.64 for females. These test statistics, distributed [[Chi].sup.2] under the null, with 13 and 12 degrees of freedom for males and females, respectively, are both highly significant.

23. We produced some information relating to information failures. Each NHIS respondent was asked subsequent to the question on drinking and driving if he/she had "done things when drinking that would have caused you to be hurt?" and "done things when drinking that would have caused others to be hurt?" Current drinkers tend to report that they drive after drinking too much more often than they report that their drinking behavior might be harmful to themselves, and, in turn, they report that their drinking behavior might be harmful to themselves more frequently than they report that their drinking behavior might be harmful to others. Only about one-half of those drinkers who admit to driving after drinking too much also report that their drinking behavior may result in harm to themselves, whereas only about 30 percent of those drinkers who admit to driving after drinking too much also report that their drinking behavior may result in harm to others.

24. See Kenkel (1993a) for a recent discussion.

REFERENCES

Akerlof, G. A., and W. T. Dickens. "The Economic Consequences of Cognitive Dissonance." American Economic Review, 72(3), 1982, 307-19.

Andrews, D. W. K. "Power in Econometric Applications." Econometrica, 57(5), September 1989, 1059-1090.

Becker, G. S. "A Theory of the Allocation of Time." Economic Journal, 75, September 1965, 493-517.

-----. "Crime and Punishment: An Economic Approach." Journal of Political Economy, 76(2), March/April 1968, 169-217.

Becker, G. S., M. Grossman, and K. M. Murphy. "Rational Addiction and the Effect of Price on Consumption." American Economic Review (Papers and Proceedings), 81(2), May 1991, 237-41.

Becker, G. S., and K. M. Murphy. "A Theory of Rational Addiction." Journal of Political Economy, 96(4), August 1988, 675-700.

Browning, M., and C. Meghir. "The Effects of Male and Female Labor Supply on Commodity Demands." Econometrica, 59(4), July 1991, 925-51.

Chaloupka, F., M. Grossman, and H. Saffer. "Alcohol, Regulation, and Motor Vehicle Fatalities." Journal of Legal Studies, 22(1), January 1993, 161-89.

Cook, P. J. "The Social Costs of Drinking." Expert Meeting on Negative Social Consequences of Alcohol Use. Oslo: Norwegian Ministry of Health and Social Affairs, 1991.

Coleman, J. S. Foundations of Social Theory. Cambridge, Mass.: Harvard University Press, 1990, 241-99.

Gallup, G. H. The Gallup Poll: Public Opinion 1972-1977. Wilmington, Del.: American Institute of Public Opinion, 1978.

-----. The Gallup Poll: Public Opinion 1982. Wilmington, Del.: The Gallup Poll, 1983.

Grossman, M. "Health Benefits of Increases in Alcohol and Cigarette Taxes." British Journal of Addiction, 84(10), October 1989, 1193-204.

Jacobs, J. B. Drunk Driving: An American Dilemma. Chicago: University of Chicago Press, 1989.

Kenkel, D. S. "Drinking, Driving, and Deterrence: The Social Costs of Alternative Policies." Journal of Law and Economics, October 1993a, 877-913.

-----. "A Note on the Optimal Amount of Drunk Driving." Manuscript, 1993b.

Lancaster, K. J. "A New Approach to Consumer Theory." Journal of Political Economy, 74, April 1966, 132-57.

Manning, W. G., et al. "The Taxes of Sin: Do Smokers and Drinkers Pay Their Way?" Journal of the American Medical Association, 261(11), March 1989, 1604-9.

Michaels, R. J. "Addiction, Compulsion and the Technology of Consumption." Economic Inquiry, 26(1), January 1988, 75-88.

Modell, J. G., and J. M. Mountz. "Drinking and Flying: The Problem of Alcohol Use by Pilots." New England Journal of Medicine, 323(7), August 16, 1990, 455-61.

Mullahy, J., and J. L. Sindelar. "Lifecycle Effects of Alcoholism on Education, Earnings, and Occupation." Inquiry, Summer 1989, 272-82.

-----. "Gender Differences in the Labor Market Effects of Alcoholism." American Economic Review (Papers and Proceedings), 81(2), May 1991, 161-65.

-----. "Alcoholism, Work and Income." Journal of Labor Economics, 11(3), 1993, 494-520.

Phelps, C. E. "Risk and Perceived Risk of Drunk Driving among Young Drivers." Journal of Policy Analysis and Management, 6(4), 1987, 708-14.

-----. "Death and Taxes: An Opportunity for Substitution." Journal of Health Economics, 7(1), March 1988, 1-24.

-----. "Control of Alcohol-Involved Driving Through Impersonal Prevention." Alcohol Health & Research World, 14(1), January 1990, 52-6.

Pogue, T. F., and L. G. Sgontz. "Taxing to Control Social Costs: The Case of Alcohol." American Economic Review, 79(1), March 1989, 235-43.

Ross, H. L. Deterring the Drinking Driver: Legal Policy and Social Control. Lexington, Mass.: Lexington Books, 1982.

-----. "Drinking and Driving: Beyond the Criminal Approach." Alcohol Health & Research World, 14(1), January 1990, 58-62.

Saffer, H., and M. Grossman. "Beer Taxes, the Legal Drinking Age, and Youth Motor Vehicle Fatalities." Journal of Legal Studies, 16(2), June 1987, 351-74.

Steele, C. M., and R. A. Josephs. "Alcohol Myopia: Its Prized and Dangerous Effects." American Psychologist, 45(8), August 1990, 921-33.

U.S. Department of Health and Human Services, National Institute on Alcohol Abuse and Alcoholism. Seventh Special Report to the U.S. Congress on Alcohol and Health. Washington: DHHS Publication No. (ADM), 1990, 90-1656.

Vegega, M. E., and T. M. Klein. "Alcohol-Related Traffic Fatalities: U.S., 1982-1989." Morbidity and Mortality Weekly Report, 39(49), 14 December 1990, 889-91.

Wilkinson, J. T. "Reducing Drunk Driving: Which Policies Are Most Effective?" Southern Economic Journal 54(2), 1987, 322-44.