Substance abuse treatment and motor vehicle fatalities.

Freeborn, Beth A. ; McManus, Brian

1. Introduction

In recent years, motor vehicle accidents have led to over 40,000 deaths annually in the United States, and alcohol-related accidents account for about 40% of these deaths. Legally drunk drivers are much more dangerous on the road than sober drivers, (1) so considerable effort has been devoted to reducing the incidence of alcohol-impaired driving. Policies targeting drunk driving include alcohol taxes, increases in the legal drinking age, educational efforts, more stringent blood alcohol content (BAC) limits, and increased punishments for those arrested for driving under the influence of alcohol (DU1). (2) In this article we evaluate the impact of an additional policy instrument for reducing the incidence of drunk driving: the supply of substance abuse treatment (SAT).

In 2006, an estimated 22.6 million U.S. residents were classified with substance dependence or substance abuse problems. Only 30% of this group, however, received treatment for alcohol abuse or drug addiction. (3) There is abundant evidence that SAT reduces drug and alcohol abuse, specifically among heavy users. (4) A reduction in dangerous driving behavior is just one of the positive effects of successful SAT; others include improvements in physical health, employment performance, and happiness at home. Substance abuse treatment also has the advantage of being a lower-cost approach to consumption reduction compared to criminal justice interventions in alcohol and drug abuse. (5)

In order to assess the impact of increased SAT on reductions in traffic fatalities, we would ideally observe data on the individuals who desire or require treatment, which individuals do and do not receive treatment, the subsequent drug and alcohol consumption of the treated and untreated, and finally the differences in driving behavior of the treated and untreated populations. Data at this level of detail on substance abuse, treatment, and driving are simply unavailable, so in this article we employ coarser measures of SAT in the United States. (6) We observe the numbers of SAT clinics and traffic fatalities in non-metropolitan U.S. counties, where the extent of treatment can be measured by the number of local SAT clinics. We effectively assume that an increase in the number of SAT clinics reduces the costs or inconvenience for the local population to receive treatment. As more individuals receive treatment, the safety of driving behavior increases on local roads. Our approach, therefore, is conservative in the sense that we do not capture the opportunity of one county's residents to take treatment in a neighboring county, or the beneficial impact of successful SAT on driving in areas other than a person's home county.

We face a considerable challenge in establishing a causal link between SAT and traffic fatalities. Even if increasing SAT truly reduces dangerous driving, many factors that are difficult or impossible to observe can interfere with correctly making this inference. For example, counties with populations that, in unobserved ways, are especially prone to alcohol and drug abuse could have both a high rate of auto fatalities and a large number of SAT clinics. Alternatively, it could be the case that some counties are particularly aggressive in treating substance-related disorders and minimizing their impact on drivers, and these counties would have both more SAT clinics and fewer deaths. In addition to issues related to local unobservables, the nature of SAT creates challenges that are not present in many alternative alcohol-related policies, such as revisions to BAC limits. SAT enrollment is often voluntary and, if successful, the impact of treatment is long-lasting. Even if SAT does reduce drunk driving, identifying this relationship from year-to-year changes in the number of SAT clinics may be difficult and perhaps inappropriate. We address these challenges with fixed effects where appropriate and with instrumental variables estimation.

Our results indicate that the number of total clinics and clinics offering outpatient treatment are negatively and significantly related to the number of alcohol-related motor vehicle deaths. The relationship between treatment and non-alcohol fatalities is not significantly different from zero in our analysis. On average, increasing the number of clinics by one in all of our sample counties would reduce the number of alcohol-related motor vehicle deaths by 15% each year. Increasing the number of clinics that specifically offer outpatient treatment services would reduce the number of alcohol-related accidents by 26% per year. This reduction in alcohol-related deaths due to outpatient SAT amounts to 0.66 lives saved per county per year in our sample of smaller U.S. counties. With 70% of the addicted population untreated, local policy makers have an opportunity to increase road safety by supporting increased substance abuse treatment.

2. Previous Research

Federal and state governments have utilized a variety of policies to reduce the incidence of fatal traffic accidents. Several of these policies, such as seat belt laws, speed limit restrictions, and insurance regulations are not limited to alcohol-impaired drivers. (7) For the purposes of the present article, however, we focus on efforts to reduce drug- and alcohol-related accidents. (8)

Policies to reduce alcohol consumption, especially before driving, comprise one important set of efforts to reduce alcohol-related accident fatalities. Some U.S. counties have declared themselves to be "dry" and generally prohibit the sale of alcohol. Despite the strong nature of these restrictions, the effectiveness of this designation in reducing alcohol-related fatalities is unclear. Brown, Jewell, and Richer (1996) find that dry counties in Texas have fewer fatal motor vehicle accidents each year than wet counties, but Baughman et al. (2001) report that these differences across counties are more likely due to county-specific heterogeneity rather than the alcohol restrictions. Miron and Tetelbaum (2007) challenge previous beliefs that minimum legal drinking age laws have a significant impact on reducing traffic fatalities.

In addition to alcohol prohibitions, policy makers might also use prices to reduce alcohol consumption. Several studies have found that beer taxes are associated with reductions in motor vehicle deaths. (9) The apparent success of these tax policies, however, may be specious. Once state fixed effects are included in the analysis, beer taxes appear to have a negligible effect on alcohol consumption. (10)

A second type of policy intervention is the strengthening of laws against impaired driving. Chaloupka, Saffer, and Grossman (1993) compare the effectiveness of all major drunk-driving laws, and they find that punishments that include license revocation are the most successful in reducing motor vehicle deaths. By contrast, Benson, Rasmussen, and Mast (1999) report that the only effective enforcement-oriented laws are those that increase the probability that a drunk driver will be pulled over by police.

In total, the empirical findings have been mixed on the effect of policy interventions on drunk driving accidents. (11) One lesson to draw from these results is that there exists an inherent difficulty in inferring a causal relationship when several layers of activity separate a policy intervention from its desired outcome (e.g., a beer tax's effect on consumption, which affects whether a potential driver is impaired, which then affects safety conditional on the decision to drive). A second lesson concerns the importance of potentially unobserved local characteristics that affect drinking behavior and road safety conditions. In our analysis below, we employ a variety of empirical strategies in order to establish that the relationship between SAT and driving deaths is robust.

3. Data

Our data are a panel of county-level variables for the years 1998, 2000, and 2002 2004. For a complete description of all variables and their sources, please see the data appendix. We limit our analysis to 1926 counties that had populations between 5000 and 80,000 in 1998 and are not in Metropolitan Statistical Areas (MSAs), resulting in 9630 county-year observations. This sample covers over 60% of all U.S. counties. There are two main reasons why we focus on this set of relatively small counties. First, we are focusing on situations in which both substance abuse treatment and vehicle travel are local. Data from counties in MSAs, which are characterized by frequent travel across county borders within a metropolitan area, would make it more difficult to uncover the true relationship between treatment and accidents. Second, we measure the supply of substance abuse treatment through the count of clinics. This approach to treatment supply is likely to be most informative when the number of clinics is small, as it is in non-metropolitan counties. A large percentage of our sample markets (68%) have either zero or one clinic, and in this situation moving between these numbers of clinics must change the amount of treatment taken by the addict population. By contrast, in a large city with dozens of clinics, adding a new clinic may be offset by small (and difficult to observe) reductions in incumbent clinics' capacities.

We select our population cutoffs of 5000 and 80,000 by considering the counties in which small numbers of SAT clinics are most frequently observed, including the possibility that a market has no clinics. Of the counties with populations greater than 80,000, 99% have one or more clinics and 90% have at least two clinics. Only 15% of counties with fewer than 5000 people have one or more clinics. In "Robustness", we investigate whether our main findings are sensitive to these population thresholds.

Fatal Accidents

Our data on traffic fatalities are from the Fatal Accident Reporting System (FARS), which is administered by the National Highway Traffic Safety Administration. During the years of our panel, approximately 37,000 fatal accidents occurred per year, resulting in 42,000 deaths. (12) Across the 3140 counties in the United States, this implies an average of about 12 fatal accidents per county per year. In our sample of smaller counties, there are 6.1 fatal accidents per county per year resulting in 6.7 deaths. In about a third of these accidents FARS indicates that a "drinking driver" was involved. For the purposes of this article's analysis, we define an accident as alcohol-related if any of the variables in the FARS system indicate alcohol or drug involvement. An accident is alcohol-related if FARS reports any of the following: a BAC test result of at least 0.08, a positive drug test result, that a driver is charged with a drug or alcohol violation, or that police identify a driver as being impaired by alcohol or drugs. (13) While the number of vehicular deaths per year increased slightly between 1998 and 2004, total alcohol-related fatalities have declined since 2002.

Substance Abuse Treatment

The primary explanatory variables of interest are the numbers of SAT clinics and outpatient-SAT clinics. In our sample, 86% of clinics offered outpatient services. Patients who receive outpatient treatment attend several hours of therapy per week that is scheduled around the patient's normal activities. Outpatient clinics usually offer a combination of individual and group counseling sessions. By contrast, inpatient treatment is residential, and patients who receive inpatient treatment are removed from their former surroundings. Inpatient treatment patients often travel substantial distances from their homes to the clinic.

Data on the location and characteristics of treatment facilities are from the National Survey of Substance Abuse Treatment Services (N-SSATS), an annual census of substance abuse treatment facilities conducted by the Substance Abuse and Mental Health Services Administration (SAMHSA). Due to temporary suspensions of the N-SSATS survey, data from 1999 and 2001 are not available.

In our sample the average number of clinics per county is 1.32, and about 70% of counties have at least one clinic. (14) A small number of counties have a relatively large number of clinics, so we truncate at eight the number of clinics per county. (15) This truncation affects less than 1% of the county-year observations in our sample. (16) In Table 1 we summarize SAT supply by year in our sample's counties. Much of the variation in treatment supply is cross-sectional, but around a quarter of counties experience a year-to-year change in the number of clinics. While this amount of variation is not negligible, a substantial fraction (50%) of the variation occurs in counties that begin or end a transition with three or more clinics.

More importantly, the usefulness of this year-to-year variation in clinic counts is limited for us because of the voluntary and durable nature of SAT. Unlike a BAC limit or a beer tax that applies to all people immediately after it is enacted and disappears completely if it is removed, a change in SAT supply is unlikely to have an immediate effect on substance abuse in a market. For example, if the sole clinic in a market exits after several years of successful patient treatment, we would expect the clinic's market to continue having fewer problems with substance abuse and its consequences than a market that never had a clinic but shares all of the same observed and unobserved characteristics. This attribute of SAT prevents empirical models with county fixed effects from providing informative results in the analysis below.

County Demographics and Road Safety

We employ two additional sets of explanatory variables to describe variation in traffic fatalities across counties. First, we construct a set of control variables that describe a county's demographic characteristics. We do this to capture local economic conditions and tendencies for risky behavior, which can affect driving habits, drug and alcohol consumption, and the combination of drugs or alcohol with driving. For each year in our data we observe a county's median age, median income, unemployment rate, percentage of the population living below the poverty line, and percentages of blacks and Hispanics. Using data from the 2000 census, we also observe the fraction of divorced females and the percentages of adults who have completed high school and college. Because the data on divorce and education are not updated for each year of our panel, they act as fixed county-level characteristics in the analysis below.

We use a second set of control variables to describe road safety conditions. (17) Due to data limitations and the level of government at which traffic laws are typically written, most of these variables are available at the state level rather than by county. We note whether a state has a standard seat belt enforcement law (i.e., drivers may be stopped for not wearing a seat belt without committing another offense), the state's level of vehicle miles traveled (VMT), traffic density, blood alcohol content (BAC) limits, and maximum speed limits. (18) The maximum BAC limit is 0.08 in 63% of the states over the years 1998-2004. (19) We also include the state excise tax rate on packaged beer per gallon. Other state laws on alcohol and driving, reviewed briefly in section 2, varied only minimally over time in our panel, so the impacts of these laws are generally captured by state-level fixed effects, which we employ in our empirical analysis.

At the county level, we record the numbers of emergency medical personnel and hospitals to describe medical treatment quality conditional on an accident. These variables can also proxy for the county's overall disposition toward medical treatment, which may be correlated with local attitudes toward treating substance abuse disorders.

In Table 2 we provide summary statistics for motor vehicle fatalities and the control variables. For the county-year observations in this study, the average number of deaths per county is 6.7 and the average number of deaths per 10,000 residents is 3.07. The columns of Table 2 illustrate how county demographics and motor vehicle fatality rates vary with the number of clinics in the market. The numbers of total deaths and alcohol-related deaths both fall as the number of clinics increases, but non-alcohol vehicle deaths fall as well. This could be due to counties with larger and denser populations being able to support more clinics as well as reducing the average number of miles driven. Clinics are more common in counties with relatively high education attainment, income levels, and divorce rates.

Instruments

We are concerned that unobservable local characteristics may be correlated with both the number of SAT clinics and the number of traffic fatalities. As we argued in the Introduction, SAT supply could be negatively or positively correlated with the unobservables that affect motor vehicle deaths. To address this complication, in our preferred specifications below we employ instrumental variables (IV) estimators of SAT clinics' effects on motor vehicle deaths. In considering possible instruments to use within our analysis, we focus on factors which can shift the number of SAT clinics in a market without being related to local (unobservable) attitudes and activities that affect both treatment supply and driving behavior.

We use a county's number of practicing psychiatrists as an instrument for the county's SAT supply. (20) SAT clinics use psychiatrists to provide services, so an increase in the number of psychiatrists reduces the cost of operating a clinic. We assume, however, that the supply of psychiatrists is uncorrelated with the unobserved aspects of the local culture that contribute jointly to substance abuse, its treatment, and dangerous driving, after controlling for the observable measures of personal and economic stress that are captured in our demographic variables. A local population of psychiatrists cannot depend on SAT clinics alone for employment, as only 13% of clinics have a full-time physician on staff, so psychiatrists must be drawn into markets by more general mental health needs. (21)

4. Empirical Analysis

The probability of an alcohol-related fatal motor vehicle accident is affected by several stages of choices. First, an agent decides whether to consume alcohol or illicit drugs. Second, the agent decides whether to operate their vehicle. Third, conditional on driving, the agent makes choices while driving (e.g., at what speed to travel), while law enforcement officials decide how vigorously to patrol for impaired drivers. While policy interventions may occur at any of these stages, we focus primarily on the demand decisions in the first stage. We view SAT programs as a method to reduce demand for alcohol and drugs among agents whose consumption of these products may be excessive or lead to poor choices. It is also possible that while in SAT a consumer improves his ability to resist driving if he does ingest drugs or alcohol.

Our general empirical approach is to regress measures of motor vehicle deaths on measures of substance abuse treatment availability, controlling for local travel conditions and demographic characteristics. In these regressions we transform the total count of vehicle deaths into a measure of the number of deaths per 10,000 county residents. We refer to this measure as the rate of motor vehicle deaths. (22)

Let [Clinics.sub.it] represent the total number of SAT clinics in county i during year t. [Deaths.sub.it] is the motor vehicle death rate in i during t. The vector Dit contains the demographic characteristics of the county plus a dummy variable for each year to capture national trends in SAT and road safety. The vector [R.sub.it] includes information on state- and county-level road usage, road safety, and driving laws. The unobservable characteristics of county i during t are captured by the error term [[epsilon].sub.it]. Our empirical approach is to estimate models of the form:

[Deaths.sub.it] = [alpha] + [beta][Clinics.sub.it] + [D.sub.it][delta] + [R.sub.it][rho] + [[epsilon].sub.it]. (1)

In some models below we replace the variable [Deaths.sub.it] with a similarly constructed measure of the alcohol-related death rate in county i during year t. In addition, we perform some of our analysis with the number of outpatient clinics in county i during t instead of [Clinics.sub.it].

Throughout the analysis below, a central concern is the nature and content of the error term [epsilon]. As we argued above, e may be correlated with Clinics. We take several approaches to estimating the empirical model in order to recover estimates of [beta] that are robust to a variety of concerns about [epsilon]. Throughout the analysis we allow [[epsilon].sub.it] to contain a state-level fixed effect to control for unobserved characteristics of states that may drive correlation between SAT and motor vehicle safety. That is, we specify [[epsilon].sub.it] = [[mu].sub.s] + [v.sub.it], with 11 as the fixed effect. In addition, we cluster standard errors at the state level to capture heterogeneous effects of Clinics on Deaths at the state and county levels.

Estimation with Exogenous SAT

We begin by estimating Equation 1 under the assumption that [v.sub.it], is uncorrelated with Clinics and the remaining explanatory variables. We consider the effect of Clinics on all motor vehicle deaths, alcohol-related deaths, and non-alcohol deaths. Our results are in Table 3. In the "All deaths" specification, the estimated coefficients [delta] and [rho] contain some expected results and some surprises. Deaths falls with the unemployment rate, which could reflect reduced travel, and it also falls with educational attainment, which could reflect the opportunity cost of a traffic injury. Seat belt requirements have no significant impact on the death rate. In specifying the dependent variable as a rate, we interpret the population variables as measures of congestion or opportunities to travel within a county.

The relationship between SAT clinics and deaths appears weak. In each specification we find a small negative relationship between the number of clinics and the death rate, and in all three cases the estimate is insignificantly different from zero. For all deaths and alcohol-related deaths, however, we note that the Clinics coefficients have t-statistics with magnitudes over 1.5. (23) We obtain similar results if we replace Clinics with the number of outpatient facilities in a county.

We also present results from estimating Equation 1 with the error term specified as [[epsilon].sub.it] = [[mu].sub.c] + [v.sub.it], with [mu] as a county-level fixed effect. These results are in Table 4, and the SAT-related point estimates are similar to those in Table 3. There appears to be a small, negative relationship between motor vehicle deaths and the number of clinics, but the coefficients are not statistically significant and the relevant t-statistics are less than 1. This is not surprising, given that the durable and voluntary nature of SAT may imply that models utilizing county fixed effects are poorly suited to this analysis.

While our initial models control for some unobserved local differences in substance abuse patterns and traffic fatalities, it is likely that correlation between Clinics and v remains. (24) Our intuition on the relationship between unobserved county characteristics and Clinics suggests that the correlation is most likely to be positive. Counties with more pervasive (unobserved) alcohol problems will have both more SAT clinics and more auto deaths. This implies that the estimates in Table 3 are positively biased, and addressing the problematic correlation could increase the magnitude of our SAT-related coefficients.

Instrumental Variables Estimation

Our preferred approach to the empirical problem involves instrumental variables estimation of SAT clinics' effects on fatalities. We begin this analysis by estimating the first-stage relationship between the number of psychiatrists and the numbers of clinics and outpatient clinics. These results, which are presented in Table 5, show a positive and significant impact of psychiatrists on SAT. (25)

Table 6 contains our IV estimates of the effect of SAT clinics on auto fatalities. We find that the alcohol-related death rate falls significantly with the addition of a SAT clinic (Models 1-3). When we focus on outpatient clinics (Models 4-6), which are more likely to treat the local population, the reduction in the alcohol-related death rate is larger in magnitude and again significantly different from zero. In interpreting these results relative to those in Table 3, we conclude that some counties are unobservably "dangerous," and the relatively high level of substance abuse in these counties attracts more clinics while also causing more vehicle deaths.

To confirm that we are recovering a treatment relationship between SAT clinics and driving under the influence, we estimate the impact of Clinics and outpatient clinics on the non-alcohol death rate, and we find no significant reduction in this rate with an increase in the number of clinics of either type. The negative coefficient on Clinics in the model of non-alcohol deaths may be due to data difficulties in classifying alcohol-related deaths appropriately.

The results in Table 6 are from models that include state fixed effects, which are preferable to county fixed effects because of the durable nature of substance abuse treatment. When we include county fixed effects in an IV model of outpatient clinics' effect on the alcohol-related death rate, however, we obtain a similar point estimate to the result in Table 6. The coefficient estimate in this case is -0.211, but the standard error of 1.374 is an order of magnitude larger than the corresponding value in Table 6.

We can use the coefficient estimates to make predictions regarding the number of lives that would be saved by increasing or introducing substance abuse treatment facilities,. In our sample counties, the average number of alcohol-related deaths is 2.52 per year. When one SAT clinic is added to each county in our sample, the annual number of alcohol-related deaths falls by 0.37 (15%). (26) Specifically adding one outpatient clinic leads to a decrease in alcohol-related deaths by 0.66 (26%). (27) Across the 1926 U.S. counties in our sample, this implies 1271 fewer deaths in alcohol-related traffic accidents per year.

Although this reduction in deaths may seem large, SAT works by reducing a population of disproportionately dangerous drivers. Levitt and Porter (2001) use data on accidents involving two drunk drivers to estimate that impaired drivers are 13 times more likely to cause a fatal accident than a sober driver. The FARS data offers further evidence of these dangers, as about 10% of alcohol-related fatal accidents involve drivers with a DUI conviction in the last three years while only 0.6% of drivers receive a DUI each year (National Survey on Drug Use and Health 2003).

Robustness

To assess the robustness of our results we now consider several variations on the models discussed in "Instrumental Variables Estimation." We first generalize the error structure to include a separate fixed effect for every state-year pair (i.e., [epsilon].sub.it] = [[mu].sub.st] + [v.sub.it]). This captures all state-level trends and policy changes that may have affected SAT services and driving behavior. Our results, which are reported in Table 7, are very similar to those in Table 6 with state fixed effects. An increase in the number of clinics or the number of outpatient clinics results in a significant reduction in alcohol-related deaths but no significant change in overall or non-alcohol deaths.

We also consider a potential concern that Clinics and (separately) psychiatrists are positively correlated with the quality of medical care in a market. While we already include controls for a county's numbers of hospitals and emergency medical personnel, other differences across counties may remain. If this is the case, then lower death rates associated with more SAT may be due to the medical care accident victims receive, and not a reduction in impaired drivers on the road. We investigate this possibility by focusing on fatalities in which a victim is declared dead at the scene of an accident, which is the case for 57% of all accident deaths in our sample. We re-estimate the models of Table 6 with our original error structure ([[epsilon].sub.it] = [[mu].sub.s] + [v.sub.it]) and with psychiatrists as an instrument, and we report results on Table 8. All estimates in Table 8 are closer to zero than those in Table 6, with the treatment coefficients for the models of alcohol-related deaths 57% smaller, matching exactly the proportion of deaths that occurred at the scene of an accident. The impact of SAT clinics on these alcohol-related deaths is significantly different from zero at the p = 0.10 level.

Next, we consider the impact of unobserved local travel conditions on death rates. It may be the case that counties with unusually dangerous roads have high vehicular death rates of all sorts, including alcohol-related deaths. For example, a county could have road construction, weather fluctuations, or economic conditions that are difficult to observe but have a substantial effect on road safety. We account for this by including a county's non-alcohol death rate as an additional control variable in our models of alcohol-related deaths. Our results are presented in Table 9, in which we show that the main results of Table 6 are largely unchanged. In addition, we find a strong positive correlation between alcohol and non-alcohol death rates.

Finally, we consider the impact of altering the population thresholds of 5000 and 80,000, which we use to define the relevant markets for our sample. In Table 10 we display how our main results on outpatient clinics are affected by changing the population threshold. The relevant results from Table 6 are repeated as the "Base Models" for convenience. Models 1-3 show that when the lower population threshold is removed, the impact of adding another outpatient clinic to the market becomes larger in magnitude. This is likely because an additional clinic in these markets is usually the first or second outpatient clinic that a county receives, and previously there would have been little or no treatment opportunities for the worst-off patients. In Models 4-6 we reduce the upper population limit to 70,000 while retaining the lower limit of 5000. There are few notable differences between these results and those in the base models, in part because the sample size fell by less than 2%. The change in sample size is slightly smaller when we increase the population threshold to 90,000, but we see in Models 7-9 that this change eliminates the statistical significance of our main result by reducing the magnitudes of the coefficients. This is probably due to the nature of additional clinics in the markets with populations between 80,000 and 90,000. In contrast to the markets on which we focus in our main analysis, counties in this population range have an average of 4.75 clinics, so adding another clinic to these markets may have little marginal value, thus reducing the magnitude of the coefficients. This reinforces our conjecture that increasing the clinic count is most closely related to increasing effective SAT in markets with very few clinics. In larger markets, increasing treatment may be equally effective in reducing deaths, but it is more difficult to detect through changes in the number of clinics.

5. Conclusion

Drug and alcohol consumption leads to motor vehicle fatalities. One policy for reducing drug and alcohol consumption is increasing the provision of local substance abuse treatment. In this article we find that a policy of increased treatment for substance abuse problems results in a statistically significant decrease in alcohol-related motor vehicle deaths.

Uncovering this relationship is challenging because of the local unobservables that are likely correlated with both SAT and vehicle deaths, as well as the voluntary and long-lasting nature of SAT. We address these challenges by including state-level fixed effects and using instrumental variables estimation. We control for local demographic and driving conditions, and we use the number of practicing psychiatrists as an instrument. The results are robust to a variety of approaches to the data, and we show a consistent and strong relationship between the number of treatment facilities and the alcohol-related motor vehicle death rate.

Our results imply that adding one more outpatient treatment facility to each county in the sample will reduce the average number of alcohol-related deaths by 0.66 (26%) per year. SAT is effective in reducing alcohol related deaths because it impacts the most dangerous drivers. Considering the population of individuals with drug and alcohol dependence problems that are currently untreated and the potential for these individuals to become dangerous drivers, policy makers have an opportunity to increase road safety while also enhancing local health by supporting increased substance abuse treatment.

Appendix

1. Substance Abuse Treatment (SAT). County level data on SAT are from the National Survey on Substance Abuse Treatment Services (N-SSATS).

* Clinics: Number of SAT clinics

* Outpatient Clinics: Number of SAT clinics offering outpatient services

2. Motor Vehicle Deaths. County-level data on deaths are from the Fatalities Analysis Reporting System (FARS).

* Motor Vehicle Deaths: The number of traffic fatalities of drivers and passengers

* Alcohol-Related Motor Vehicle Deaths: The number of traffic fatalities where alcohol or drugs may have been a factor

3. Instruments. County-level data on practicing psychiatrists are from the Area Resource File (ARF).

* Number of Psychiatrists

4. Control Variables

(a) County-level demographic data are from the U.S. Census and the ARF.

* Population: County population, divided by 10,000 in regression analysis

* % Black

* % Hispanic

* Median Age (in years)

* Median Income: In dollars, divided by 10,000 in regression analysis

* % Poverty

* % Unemployment

* % College Graduate: Percentage of adults with a college degree (2000 Decennial Census)

* % HS Graduate: Percentage of adults with a high school degree (2000 Decennial Census)

* % Females Divorced: Percentage of women divorced (2000 Decennial Census)

* Emergency Personnel: Number of emergency medical personnel

* Hospitals: Number of hospitals

(b) State-level traffic variables are obtained from the publication "Highway Statistics" from the U.S. Department of Transportation, Federal Highway Administration.

* Traffic Density Rural: Registered vehicles per mile of rural roads

* Traffic Density Urban: Registered vehicles per mile of urban roads

* VMT Rural: Vehicle miles traveled on rural roads, divided by 100,000

* VMT Urban: Vehicle miles traveled on urban roads, divided by 100,000

(c) State level traffic and highway laws are from the Insurance Institute for Highway Safety.

* Speed Limit 65 mph: A dummy variable equal to 1 when 65 mph is the state's top speed limit

* Standard Seat Belt Enforcement: A dummy variable equal to 1 when the state has a standard-enforcement mandatory seat belt law

(d) State DUI laws are from the National Conference of State Legislatures.

* BAC of 0.08: A dummy variable equal to 1 when a state has a blood alcohol content of 0.08 to qualify as DUI

(e) County-level crime variables are from the Department of Justice.

* Violent Crimes Reported: A sum of the violent crimes (homicide, rape, robbery, aggravated assault) reported per capita

* Property Crimes Reported: A sum of the property crimes (larceny, burglary, motor vehicle theft) reported per capita

(f) State-level alcohol taxes are from the annual publication "Brewers Almanac 2007."

* Beer Tax Rate: Annual excise rate on packaged beer per gallon

References

Adams, Scott, and Chad Cotti. 2008. Drunk driving after the passage of smoking bans in bars. Journal of Public Economics 92:1288-1305.

Baughman, Reagan, Michael Conlin, Stacy Dickert-Conlin, and John Pepper. 2001. Slippery when wet: The effects of local alcohol access laws on highway safety. Journal of Health Economics 20:1089-96.

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

Brown, Robert W., R. Todd Jewell, and Jerrell Richer. 1996. Endogenous alcohol prohibition and drunk driving. Southern Economic Journal 62(4):1043-53.

Buchmueller, Thomas C., Philip F. Cooper, Mireille Jacobson, and Samuel H. Zuvekas. 2007. Parity for whom? Exemptions and the extent of state mental health parity legislation. Health Affairs 26(4):483-7.

Cartwright, William. 2000. Cost-benefit analysis of drug treatment services: Review of the literature. Journal of Mental Health Policy and Economics 3(1):11-26.

Chaloupka, Frank J., Henry Saffer, and Michael Grossman. 1993. Alcohol-control policies and motor-vehicle fatalities. Journal of Legal Studies 22(1):161-86.

Cohen, Alma, and Rajeev Dehejia. 2004. The effect of automobile insurance and accident liability laws on traffic fatalities. Journal of Law and Economics 47(2):357-93.

Cohen, Alma, and Liran Einav. 2003. The effects of mandatory seat belt laws on driving behavior and traffic fatalities. Review of Economics and Statistics 85(4):828-43.

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.

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

Hingson, Ralph, Ronda Zakocs, Timothy Heeren, Michael Winter, David Rosenbloom, and William De Jong. 2005. Effects on alcohol related fatal crashes of a community based initiative to increase substance abuse treatment and reduce alcohol availability. Injury Prevention 11:84-90.

Levitt, Steven, and Jack Porter. 2001. How dangerous are drinking drivers? Journal of Political Economy 109(6):1198-1237.

Lu, Misgshan, and Thomas McGuire. 2002. The productivity of outpatient treatment for substance abuse. The Journal of Human Resources 37(2):309-35.

Mast, Brent D., Bruce L. Benson, and David W. Rasmussen. 1999. Beer taxation and alcohol-related traffic fatalities. Southern Economic Journal 66(2):214-49.

Miron, Jeffrey A., and Elina Tetelbaum. 2009. Does the minimum legal drinking age save lives'? Economic Inquiry 47(2):317-36.

Prendergast, Michael L., Deborah Podus, Eunice Chang, and Darren Urada. 2002. The effectiveness of drug abuse treatment: A meta-analysis of comparison group studies. Drug and Alcohol Dependence 67(1):53-72.

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

Rydell, C. Peter, and Susan S. Everingham. 1994. Controlling cocaine: Supply versus demand programs, RAND Report.

Saffer, Henry, Frank Chaloupka, and Dhaval Dave. 2001. State drug control spending and illicit drug participation. Contemporary Economic Policy 19(2):150-61.

U.S. Department of Health and Human Services, Substance Abuse and Mental Health Services Administration. 2003. Alcohol and drug services study (ADSS): The national substance abuse treatment system: Facilities, clients, services, and staffing. Rockville, MD: Office of Applied Studies.

U.S. Department of Health and Human Services, Substance Abuse and Mental Health Services Administration. 2004. Results from the 2003 National Survey on Drug Use and Health: National findings. Office of Applied Studies, NSDUH Series H-25, DHHS Publication No. SMA 07-3964.

U.S. Department of Health and Human Services, Substance Abuse and Mental Health Services Administration. 2007. Results from the 2006 National Survey on Drug Use and Health: National findings. Office of Applied Studies, NSDUH Series H-32, DHHS Publication No. SMA 07-4293.

Beth A. Freeborn * and Brian McManus ([dagger])

* College of William and Mary, Department of Economics, P.O. Box 8795, Williamsburg, VA 23187, USA; Email bafree@wm.edu; corresponding author.

([dagger]) Department of Economics, CB 3305, University of North Carolina, Chapel Hill, NC 27599, USA; E-mail mcmanusb@email.unc.edu.

We thank the editor John Pepper and anonymous referees for their comments and suggestions. Any remaining errors are those of the authors.

Received June 2008; accepted May 2009.

(1) See Levitt and Porter (2001).

(2) Benson, Rasmussen, and Mast (1999) discuss the effectiveness of these policies.

(3) National Survey on Drug Use and Health (2006).

(4) For examples, see Lu and McGuire (2002) or Saffer, Chaloupka, and Dave (2002). Prendergast et al. (2002) perform a meta-analysis on the "effectiveness of treatment" literature (78 studies) and show that clients who receive treatment have significantly better outcomes than those who do not receive treatment.

(5) Specifically, substance abuse treatment is believed to be more cost effective than punishment. A prominent RAND study (Rydell and Everingham 1994) finds that treatment is seven times more cost effective than domestic law enforcement, 10 times more effective than interdiction, and 23 times more effective than the "source control" method (attacking drug supply abroad). Cartwright (2000) provides a detailed review of the literature on cost-benefit analysis of treatment and concludes that although there is a great deal of variation in the literature, the general persistent finding is that the benefits of substance abuse treatment outweigh the costs.

(6) For an example of research that is able to more closely observe activity within communities following changes to SAT, see Hingson et al. (2005). The authors study five communities which received grants under the Robert Wood Johnson Foundation's "'Fighting Back" program. These communities experienced declines in their ratios of alcohol-related to non-alcohol traffic fatalities. Due to the population sizes of these Fighting Back communities, the affected areas are not included in the data sample of the present article.

(7) See Cohen and Einav (2003) on the effects of mandatory seat belt usage laws on traffic fatalities. Cohen and Dehejia (2004) study no-fault insurance liability laws.

(8) See Adams and Cotti (2008) for an example of a policy--a smoking ban in bars that was not intended to affect road safety but ultimately increased the incidence of impaired driving. Adams and Cotti estimate that these bans result in a 19% increase in alcohol-related vehicle deaths per year.

(9) For example, see Cook (1991); Chaloupka, Saffer, and Grossman (1993); and Ruhm (1995).

(10) See Dee (1999) and Mast, Benson, and Rasmussen (1999).

(11) See Benson, Rasmussen, and Mast (1999) for a discussion on how the lack of consistency in deterrence measures across different types of studies (microsurvey, state-level aggregates) leads to a wide range of results.

(12) Fatal traffic accidents may involve more than one fatality (e.g., the driver and passenger).

(13) We search for any indication of alcohol and drug use because some of the relevant variables appear to be coded inconsistently. For example, some observations for which police record a "drinking driver" have BAC results that are zero or indicate that no test was given.

(14) The N-SSATS data include a variable containing the number of patients registered to a SAT clinic on a particular day of the year. We do not use this variable for our main analysis because it is noisy and inconsistently framed across years of the survey, but it may provide some evidence of a "first stage" effect of clinics on the amount of treatment. The median number of outpatient clients registered in a single-clinic market is 26, and the median two-clinic market has a total of 71 outpatient clients in treatment. When we regress the total number of patients in a market on the number of clinics (plus a set of appropriately-selected control variables that are a subset of the vehicle fatality controls), we find that an additional clinic increases the number of treated patients by 47%.

(15) We also truncate the number of clinics with outpatient services (separately) at 5, which affects about 1% of the observations in our sample.

(16) Our main empirical results are largely unaffected by this assumption. If we truncate at 9 clinics, the estimated effects fall slightly in our main results (in Table 6) but remain significant at the p = 0.05 level. If we reduce the truncation to 7 clinics, our coefficient magnitudes increase.

(17) Our selection of road safety variables generally follows Cohen and Einav (2003) on the impact of seat belt usage laws.

(18) During our sample period South Carolina changed its maximum speed limit from 65 to 70 mph. All other states in the sample had constant maximum speed limits during the years of the panel. Seven states instituted a standard seat belt enforcement law during 1998-2004.

(19) During this period, 35 states reduced their BAC limits to 0.08.

(20) We also investigated using state-level variation in mandated benefits under mental health parity laws. See Buchmueller et al. (2007) for an overview of these laws. Unfortunately, the variation in benefit mandate laws is insufficient to provide an effective instrument for SAT supply.

(21) SAMHSA reports that in addition to the 13% of facilities that hire a full-time physician, 21% of facilities utilize at least one part-time physician, and 37% of facilities have a contract physician on staff (Alcohol and Drug Services Study, 2003). In conversations with professionals in the field, we have confirmed that physicians involved in SAT are generally psychiatrists.

(22) While it may be preferable to calculate the number of deaths per mile traveled, the necessary data are not available at the county level.

(23) We have also estimated the models in Table 3 without state fixed effects. In all three models, the coefficient estimates on Clinics are smaller in magnitude than the point estimates reported in Table 3. The differences in magnitude are consistent with our hypothesis that unobserved local characteristics tend to increase both the number of clinics and the motor vehicle death rate.

(24) Previous research on motor vehicle fatalities has illustrated the potential for unobserved market characteristics tobias the results of OLS estimates. For example, see Mast, Benson, and Rasmussen (1999): Baughman et al. (2001); and Cohen and Einav (2003).

(25) We test for whether the number of psychiatrists is a weak instrument. In both first-stage regressions, the F- statistics were well over the recommended value of 10 (38.45 for all clinics and 22.16 for outpatient only clinics).

(26) To calculate the reduction in deaths from increasing the number of facilities, we begin by predicting the number of motor vehicle deaths per 10,000 residents using the IV coefficient estimates. Next we increase the number of clinics in each county by one and predict the motor vehicle deaths rate given the new number of clinics. We take the difference between the two predicted rates, multiply by the population, and then report the average as the number of lives saved.

(27) This is comparable to the 19% decline in alcohol-related deaths that Chaloupka, Saffer, and Grossman (1993) predict would follow from high mandatory fines for drunk driving.

Table 1. SAT Clinics in Sample Counties

           Average Number of       % Counties with
Year       Clinics in County        Clinic Present

1998             1.320                  65.32
2000             1.336                  70.25
2002             1.341                  70.51
2003             1.329                  69.57
2004             1.304                  69.06
Average          1.337                  68.94

            % Counties with     % Counties with Change
Year       Outpatient Present      in Clinic Count

1998             62.36                    --
2000             69.12                  42.06
2002             68.90                  27.99
2003             68.07                  22.38
2004             67.55                  20.20
Average          67.20                  28.15

Table 2. Summary Statistics on Motor Vehicle Deaths, SAT Supply,
and Other Control Variables

                                       All Counties   Zero Clinics

Motor vehicle deaths per 10,000             3.072          3.506
Alcohol-related motor vehicle deaths
  per 10,000                                1.173          1.291
Non-alcohol motor vehicle deaths per
  10,000                                    1.899          2.215
Population                                 23,986         14,989
Black                                       9.216         11.27
Hispanic                                    5.721          6.862
Median age                                 37.99          38.53
High school graduate                       75.25          73.10
% College graduate                         14.00          12.51
Women divorced                              9.678          8.958
% Poverty                                  14.91          15.57
Median income                              33,384         32,124
% Unemployment                              5.762          5.604
VMT urban                                   0.418          0.483
VMT rural                                   0.299          0.341
Property crimes reported                  541.8          274.6
Violent crimes reported                    59.72          34.56
Traffic density urban                     262.2          251.0
Traffic density rural                      64.51          57.18
% With BAC limit of 0.08                   63.66          65.36
% Standard seat belt enforcement           39.74          44.60
Speed limit 65 mph                         30.51          17.79
Emergency personnel                         0.973          0.346
Hospitals                                   0.838          0.633
Beer tax rate                               0.260          0.277
Psychiatrists                               0.923          0.199
N                                        9630           2991

                                        One Clinic    Two+ Clinics

Motor vehicle deaths per 10,000             3.074          2.646
Alcohol-related motor vehicle deaths
  per 10,000                                1.195          1.033
Non-alcohol motor vehicle deaths per
  10,000                                    1.879          1.613
Population                                 21,260         35,918
Black                                      10.13           6.154
Hispanic                                    4.509          6.018
Median age                                 38.20          37.21
High school graduate                       74.86          77.78
% College graduate                         13.48          16.05
Women divorced                              9.631         10.43
% Poverty                                  15.00          14.17
Median income                              33,203         34,823
% Unemployment                              5.827          5.839
VMT urban                                   0.378          0.401
VMT rural                                   0.286          0.272
Property crimes reported                  433.6          928.2
Violent crimes reported                    54.34          90.50
Traffic density urban                     267.9          266.5
Traffic density rural                      65.32          70.71
% With BAC limit of 0.08                   62.58          63.25
% Standard seat belt enforcement           37.12          38.06
Speed limit 65 mph                         34.90          37.80
Emergency personnel                         0.698          1.904
Hospitals                                   0.756          1.134
Beer tax rate                               0.263          0.238
Psychiatrists                               0.576          2.031
N                                        3570           3069

We use county-year data from 1926 counties to compute the statistics
in this table.

Table 3. SAT's Impact on Traffic Fatalities When Treatment
Is Exogenous and the Models Contain State Fixed Effects

                               All       Standard
Dependent Variable           Deaths        Error

Number of clinics          -0.042         (0.026)
Population                 -0.544 ***     (0.123)
Population-squared          5.967 ***     (1.261)
% Black                     1.178 **      (0.453)
% Hispanic                 -1.016 *       (0.511)
Median age                  0.003         (0.017)
% Poverty                   0.019         (0.028)
Median income               0.060         (0.140)
% Unemployment             -0.036 *       (0.019)
% High school graduate     -0.018 **      (0.009)
College graduate           -0.030 ***     (0.011)
% Women divorced            0.049 *       (0.027)
VMT urban                   0.697         (0.874)
VMT rural                  -0.858         (1.325)
Property crimes reported   -0.236 *       (0.123)
Violent crimes reported     0.605         (0.469)
Traffic density urban       0.210 **      (0.100)
Traffic density rural      -0.133         (0.197)
BAC of 0.08                -0.145 **      (0.067)
Standard seat belt
  enforcement               0.022         (0.053)
Speed limit 65 mph          0.118         (0.077)
Emergency personnel        -0.031         (0.021)
Hospitals                  -0.070         (0.065)
Beer tax rate               0.266         (0.200)
Constant                    3.819 ***     (1.364)
[R.sup.2] (overall)                 0.181

                             Alcohol     Standard
Dependent Variable           Deaths        Error

Number of clinics          -0.021         (0.014)
Population                 -0.193 ***     (0.051)
Population-squared          2.037 ***     (0.472)
% Black                     0.045         (0.249)
% Hispanic                 -0.827 ***     (0.274)
Median age                  0.005         (0.009)
% Poverty                   0.043 **      (0.017)
Median income               0.150 *       (0.079)
% Unemployment             -0.006         (0.009)
% High school graduate     -0.006         (0.004)
College graduate           -0.006         (0.006)
% Women divorced            0.027 *       (0.015)
VMT urban                   0.393         (0.676)
VMT rural                  -0.531         (0.776)
Property crimes reported   -0.141 ***     (0.052)
Violent crimes reported     0.878 ***     (0.239)
Traffic density urban       0.078         (0.124)
Traffic density rural      -0.044         (0.153)
BAC of 0.08                -0.099 **      (0.044)
Standard seat belt
  enforcement              -0.021         (0.050)
Speed limit 65 mph         -0.292 ***     (0.042)
Emergency personnel        -0.013         (0.009)
Hospitals                  -0.034         (0.032)
Beer tax rate               0.052         (0.091)
Constant                    0.357         (0.770)
[R.sup.2] (overall)                 0.126

                           Non-Alcohol    Standard
Dependent Variable            Deaths        Error

Number of clinics          -0.021          (0.017)
Population                 -0.352 ***      (0.092)
Population-squared          3.930 ***      (1.010)
% Black                     1.132 ***      (0.267)
% Hispanic                 -0.189          (0.349)
Median age                 -0.002          (0.009)
% Poverty                  -0.024          (0.015)
Median income              -0.091          (0.083)
% Unemployment             -0.030 *        (0.016)
% High school graduate     -0.012 *        (0.007)
College graduate           -0.024 ***      (0.007)
% Women divorced            0.022          (0.016)
VMT urban                   0.305          (0.928)
VMT rural                  -0.327          (1.442)
Property crimes reported   -0.095          (0.088)
Violent crimes reported    -0.273          (0.436)
Traffic density urban       0.132          (0.100)
Traffic density rural      -0.088          (0.232)
BAC of 0.08                -0.047          (0.076)
Standard seat belt
  enforcement               0.043          (0.065)
Speed limit 65 mph          0.174 **       (0.080)
Emergency personnel        -0.018          (0.015)
Hospitals                  -0.036          (0.044)
Beer tax rate               0.214          (0.195)
Constant                    3.462 ***      (1.228)
[R.sup.2] (overall)                 0.166

N = 9628. Models include state fixed effects and a dummy variable
for each year. Standard errors are clustered by state. Statistical
significance is indicated by * for the 10% level, ** for the 5%
level, and *** for the 1% level.

Table 4. SAT's Impact on Traffic Fatalities When Treatment
Is Exogenous and the Models Contain County Fixed Effects

                                                 Standard
Dependent Variable                 All Deaths      Error

Number of clinics                 -0.020           0.034
Population                        -1.561 *         0.852
Population-squared                11.310           8.079
% Black                            2.747           5.070
% Hispanic                         2.044           3.264
Median age                         0.009           0.052
% Poverty                         -0.062 ***       0.022
Median income                      0.110           0.213
% Unemployment                     0.028           0.020
% High school graduate            dropped
% College Graduate                dropped
% Women divorced                  dropped
VMT urban                          0.536           1.052
VMT rural                         -0.841           1.139
Property crimes reported           0.060           0.124
Violent crimes reported           -0.627           0.763
Traffic density urban              0.151           0.129
Traffic density rural             -0.090           0.287
BAC of 0.08                       -0.133 **        0.067
Standard seat belt enforcement     0.034           0.119
Speed limit 65 mph                -0.008           0.376
Emergency personnel               -0.013           0.035
Hospitals                         -0.003           0.057
Beer tax rate                      0.166           0.528
Constant                           5.226 *         3.022
[R.sup.2] (Overall)                         0.041

                                    Alcohol      Standard
Dependent Variable                   Deaths        Error

Number of clinics                 -0.013           0.021
Population                        -0.412           0.533
Population-squared                 3.727           5.054
% Black                           -6.532 **        3.172
% Hispanic                        -2.741           2.042
Median age                         0.005           0.032
% Poverty                         -0.026 **        0.013
Median income                      0.083           0.133
% Unemployment                     0.011           0.012
% High school graduate
% College Graduate
% Women divorced
VMT urban                          0.155           0.657
VMT rural                         -0.635           0.712
Property crimes reported           0.022           0.077
Violent crimes reported            0.060           0.478
Traffic density urban              0.039           0.081
Traffic density rural             -0.062           0.179
BAC of 0.08                       -0.080 *         0.042
Standard seat belt enforcement     0.004           0.075
Speed limit 65 mph                -0.194           0.235
Emergency personnel                0.011           0.022
Hospitals                         -0.012           0.036
Beer tax rate                      0.037           0.330
Constant                           2.551           1.891
[R.sup.2] (Overall)                         0.000

                                  Non-Alcohol    Standard
Dependent Variable                   Deaths        Error

Number of clinics                 -0.007           0.027
Population                        -1.149 *         0.670
Population-squared                 7.584           6.345
% Black                            9.278 **        3.982
% Hispanic                         4.785 *         2.563
Median age                         0.004           0.040
% Poverty                         -0.036 **        0.016
Median income                      0.028           0.166
% Unemployment                     0.017           0.015
% High school graduate
% College Graduate
% Women divorced
VMT urban                          0.381           0.825
VMT rural                         -0.206           0.894
Property crimes reported           0.037           0.097
Violent crimes reported           -0.686           0.600
Traffic density urban              0.111           0.101
Traffic density rural             -0.028           0.225
BAC of 0.08                       -0.054           0.053
Standard seat belt enforcement     0.030           0.094
Speed limit 65 mph                 0.186           0.296
Emergency personnel               -0.024           0.028
Hospitals                          0.009           0.045
Beer tax rate                      0.129           0.415
Constant                           2.675           2.374
[R.sup.2] (Overall)                         0.055

N = 9628. Models include county fixed effects and a dummy variable
for each year. Statistical significance is indicated by * for the
10% level, ** for the 5% level, and *** for the 1% level.

Table 5. First-Stage Models Predicting SAT Services

                                      Standard   Outpatient   Standard
Dependent Variable         Clinics     Error      Clinics      Error

Number of psychiatrists   0.112 ***   (0.018)    0.064 ***    (0.013)

N = 9628. Standard errors are clustered by state. The models also
include the control variables listed in Table 3, state fixed
effects, and a dummy variable for each year. The full results are
available upon request. Statistical significance is indicated by
* for the 10% level, ** for the 5% level, and *** for the 1% level.

Table 6. IV Estimates of SAT Clinics' Effect on Traffic Fatalities

                                  All     Standard    Alcohol
Dependent Variable               Deaths    Error      Deaths

Models 1-3
  Number of clinics              -0.216   (0.132)    -0.155 **

Models 4-6
  Number of outpatient clinics   -0.381   (0.249)    -0.274 **

                                 Standard    Non-Alcohol   Standard
Dependent Variable                 Error       Deaths       Error

Models 1-3
  Number of clinics               (0.065)      -0.061      (0.094)

Models 4-6
  Number of outpatient clinics    (0.126)      -0.107      (0.169)

See Table 5 note for other details. Statistical significance
is indicated by * for the 10% level, ** for the 5% level,
and *** for the 1% level.

Table 7. IV Estimates of SAT Clinics' Effect on Traffic Fatalities
with State-Year Fixed Effects

                                  All     Standard    Alcohol
Dependent Variable               Deaths    Error      Deaths

Models 1-3
  Number of clinics              -0.220   (0.137)    -0.150 **

Models 4-6
  Number of outpatient clinics   -0.388   (0.256)    -0.265 **

                                 Standard   Non-Alcohol   Standard
Dependent Variable                Error       Deaths       Error

Models 1-3
  Number of clinics              (0.065)      -0.069      (0.098)

Models 4-6
  Number of outpatient clinics   (0.125)      -0.123      (0.176)

See Table 5 note for other details. Statistical significance
is indicated by * for the 10% level, ** for the 5% level,
and *** for the 1% level.

Table 8. IV Estimates of SAT Clinics' Effect on Deaths
at the Scene of an Accident

                                  All     Standard   Alcohol
Dependent Variable               Deaths    Error      Deaths

Models 1-3
  Number of clinics              -0.132   (0.093)    -0.088 *

Models 4-6
  Number of outpatient clinics   -0.233   (0.175)    -0.156 *

                                 Standard   Non-Alcohol   Standard
Dependent Variable                Error       Deaths       Error

Models 1-3
  Number of clinics              (0.048)      -0.044      (0.072)

Models 4-6
  Number of outpatient clinics   (0.088)      -0.078      (0.130)

See Table 5 note for other details. Statistical significance
is indicated by * for the 10% level, ** for the 5% level,
and *** for the 1% level.

Table 9. IV Estimates of SAT Clinics' Effect on Alcohol-Related
Deaths Controlling for Non-Alcohol Deaths

Dependent Variable               Alcohol Deaths   Standard Error

Model 1
  Number of clinics                -0.151 **         (0.063)
  Non-alcohol deaths                0.063 ***        (0.011)

Model 2
  Number of outpatient clinics     -0.267 **         (0.122)
  Non-alcohol deaths                0.062 ***        (0.011)

See Table 5 note for other details. Statistical significance
is indicated by * for the 10% level, ** for the 5% level,
and *** for the 1% level.

Table 10. IV Estimates of SAT Clinics' Effect on Traffic Fatalities:
The Impact of Population Thresholds

                                              Standard    Alcohol
Dependent Variable               All Deaths    Error      Deaths

Base models. Lower limit = 5000; Upper limit = 80,000. (N = 9628)
  Number of outpatient clinics     -0.381     (0.249)    -0.274 **

Models 1-3. Lower limit = 0; Upper limit = 80,000. (N = 10,846)
  Number of outpatient clinics     -0.398     (0.327)    -0.344 **

Models 4-6. Lower limit = 5000; Upper limit = 70,000. (N = 9463)
  Number of outpatient clinics    -0.329 *    (0.247)    -0.224 *

Models 7-9. Lower limit = 5000; Upper limit = 90,000. (N = 9763)
  Number of outpatient clinics     -0.158     (0.277)     -0.185

                                 Standard   Non-Alcohol   Standard
Dependent Variable                Error       Deaths       Error

Base models. Lower limit = 5000; Upper limit = 80,000. (N = 9628)
  Number of outpatient clinics   (0.126)      -0.107      (0.169)

Models 1-3. Lower limit = 0; Upper limit = 80,000. (N = 10,846)
  Number of outpatient clinics   (0.146)      -0.054      (0.228)

Models 4-6. Lower limit = 5000; Upper limit = 70,000. (N = 9463)
  Number of outpatient clinics   (0.118)      -0.105      (0.172)

Models 7-9. Lower limit = 5000; Upper limit = 90,000. (N = 9763)
  Number of outpatient clinics   (0.130)       0.026      (0.200)

Number of observations vary across models and are provided
within the Table. See Table 5 note for other details. Statistical
significance is indicated by * for the 10% level, ** for the 5%
level, and *** for the 1% level.