Alcohol Availability and Crime: Evidence from Census Tract Data.

Gyimah-Brempong, Kwabena

Kwabena Gyimah-Brempong [*]

Using census tract data from the city of Detroit and a reduced-form crime equation, this article finds that alcohol availability is positively and significantly related to total, property, and violent crime rates and homicides. The elasticity of crime rates with respect to alcohol availability calculated in this study are 0.92, 0.82, 0.87, and 0.12 for total crime, violent crime, property crime, and homicide, respectively. These elasticities do not change qualitatively across estimation methods for the various measures of crime rates. I find that ordinary least squares estimates impart a downward bias to the effects alcohol availability has on crime rates. Failure to account for the endogeneity of alcohol outlets will therefore result in an underestimate of crime elasticities with respect to alcohol availability. The estimates imply that reducing alcohol availability may decrease crime rates and improve social welfare.

1. Introduction

This article uses census tract data from the city of Detroit and a reduced-form crime equation to investigate the effects of alcohol availability on crime. A fact of urban life in the United States is high crime rates, rates that differ across neighborhoods in cities and easy availability of alcohol. A disproportionately large number of crimes are committed by people who have just consumed alcohol (Cook and Moore 1993a). The homicide rate among youths in poor inner cities, where alcohol consumption tends to be high, is four times as high as the national average among all youths (FBI 1996). Parker (1993) reports that, all things equal, homicides are higher in high alcohol consumption neighborhoods than in low alcohol consumption neighborhoods. According to the National Institute of Criminal Justice statistics, 40% of all violent crime victimization, 40% of all fatal motor vehicle accidents, and 67% of all domestic violence in 1995 were alcohol related. In addition, 40% of all violent offenders in jail reporte d using alcohol just before committing the crime (Greenfeld 1998). Similarly, a disproportionately large share of crimes in the United Kingdom occurs in or near pubs during peak hours of operation (Hutchinson, Henderson, and Davis 1995). Besides being perpetrators, a large number of victims of crime had used alcohol at the time of their victimization. In this study, I use the term commission of crime more broadly to include perpetrators as well as the victims of crime.

In spite of these statistics, easy availability of alcohol and its use continue to be part of the American culture. A Harris poll reports that 38% of adult males think they drink too much but will not change their behavior and 28% of high school students think that adults should be able to drink all the alcohol they want (FBI 1996). However, a recent study conducted by the University of Michigan for the National Institute of Drug Abuse found that over 80% of Americans disapprove of youth drinking (University of Michigan Institute for Social Research 1998). In spite of these statistics linking alcohol to crime, with the exception of the relationship between alcohol control policy and drunk driving, economists have not investigated the effects of alcohol availability on crime rates generally.

Given the high correlation between alcohol availability and crime, DiIulio (1995) suggests using zoning laws to decrease alcohol availability as a means of decreasing crime. It is not clear whether there is a causal relationship between alcohol availability and crime rates. Does alcohol availability increase crime? If increased alcohol availability leads to increased crime, what is the mechanism through which alcohol availability affects crime? While answers to these questions may be important for policy formulation and implementation, my concern is with the investigation of the effects of alcohol availability on crime rates. If alcohol availability increases crime rates, it may be possible to formulate alcohol control policies that also depress crime. While there are few theoretical models that link alcohol availability to specific crimes, empirical studies of the effects of geographical availability of alcohol on crime rates generally are few and far between in the economics literature. More empirical studi es could help establish the link that has been made between alcohol control policies and crime.

Results of studies investigating the connection between alcohol availability and crime have implications for further research and policy. For example, if alcohol availability increases crime rates, then decreasing alcohol availability may decrease crime and improve social welfare. If there is a positive link between alcohol availability and crime rates, then alcohol availability should be included in the supply of offense equation as an added explanatory variable. There is, therefore, the need to better understand the relationship between the geographical availability of alcohol and crime rates in the economics of crime literature. This article attempts to contribute to the literature on the effects of alcohol availability on crime rates by linking the availability of alcohol directly to crime rates instead of looking at the relationship between alcohol control policies and crime. To the extent that alcohol control policies affect the availability of alcohol, this article is in the same vein as those that inv estigate the effects of alcohol control policies on some crimes.

Criminologists and sociologists find a positive correlation between alcohol and crime (Homel, Tomsen, and Thommeny 1992; Stitt and Giacopassi 1992; Parker 1993, 1995; Van Oers and Garretsen 1993; Blount et al. 1994; Scribner, MacKinnnon, and Dwyer 1995; Valdez et al. 1995; Roizen 1997; Parker and Cartmill 1998; Scribner et al. 1999). Some researchers argue that alcohol affects criminal behavior in a pharmacophysiological way by either impairing reasoning or reducing inhibition (Lang and Sibrel 1989; Sprunt et al. 1994) after consumption. In effect, alcohol acts as a catalyst rather than as the cause of crime. Other researchers argue that alcohol use is related to crime because of the need to obtain resources to purchase alcohol (Rush, Glickman, and Brook 1986). A third group argue that drinking per se does not result in crime; it is rather the environment in which drinking takes place that encourages criminal behavior (Homel, Tomsen, and Thommeny 1992). Gruenewald, Madden, and Janes (1992) and Gruenewald, Pon ick, and Holder (1992) use state data to investigate the effects of alcohol outlet density on alcohol consumption and find a significantly positive correlation between density and consumption after controlling for the endogeneity of density. They did not, however, investigate the effects of alcohol outlet density on crime. Second, they provide no detailed economic rationale for the hypothesized endogeneity of alcohol availability.

Economists mostly investigate the relationship between alcohol pricing, regulation, and drunk driving and have not generally investigated the relationship between alcohol availability and other crimes. There are, however, a few exceptions. Using census tract data from Milwaukee, Wisconsin, DiIulio (1995) finds that alcoholic outlet density has a positive impact on all indices of crime after controlling for socioeconomic characteristics of communities. Brown, Jewell, and Richer (1996) find that alcohol control policies decrease alcohol-related motor vehicle fatalities in Texas counties. Chaloupka and Weschler (1996) find that lower alcohol prices and availability are positively correlated with binge drinking and crimes among U.S. college students. Further, they find that crime rates are substantially higher when alcohol is available on campus than when it is not. In two recent articles, Markowitz and Grossman (1998a, b) investigate the effects of state alcohol control policies on child abuse in the United Sta tes. Using state beer taxes and prohibition of billboard advertising of alcohol as their measures of alcohol control policies, they find a negative and significant relationship between state alcohol control policies and child abuse. Saffer and Chaloupka (1995) find that decriminalization of marijuana increases marijuana consumption by 4-6%. If a given proportion of those who consume drugs (including alcohol) commit crime or become victims of crime, then increased consumption of the drug due to increased availability may increase crime rates. Cook and Moore (1993b), using pooled time series and cross-section data, find that alcohol availability increases alcohol consumption, which in turn increases violence in the United States.

Most of the studies reviewed above use either statewide or countywide data to investigate the relationship between alcohol availability and crime. The use of data from large geographical areas such as a state or county implies that one may not be able to control for many environmental variables that affect crime. For example, outlet density, defined as the concentration of alcohol outlets in a geographical area, measured at the state level may have different crime implications in a densely populated urban area in the state than in a sparsely populated rural part of the state. It is also well known that the distribution of alcohol outlets is not uniform within a state or a city. Within a city, alcohol outlets tend to be concentrated in a few neighborhoods. Citywide alcohol outlet density will not capture these neighborhood differences. By using census tract data from a large city, I hope to minimize the influence of such intervening variables. [1] The use of the census tract as the unit of analysis is similar to the approach adopted by Scribner et al. (1999). While studies investigating the connection between alcohol and crime assume that alcohol availability is exogenous, I test for its exogeneity in this study. I find it endogenous and treat it as such in my investigation.

The contribution of this article to the literature is twofold. First, I combine the criminology and economics of crime literature to explore the connection between alcohol availability and crime. Second, the use of census tract data from one city reduces the confounding effects of other factors on the relationship between alcohol availability and crime. The use of Detroit as the study city is interesting because of the city's reputation as a high crime, aging industrial city with a predominantly African-American population. [2] Most researchers who investigate the effects of alcohol on crime have used ordinary least squares (OLS) estimation methodology in their investigation. It is most likely that alcohol availability is correlated with some unmeasured community characteristics, hence the error term. Using the OLS estimator under such circumstances will result in biased coefficient estimates. I use an instrumental variables (IV) estimator to estimate the crime equation and compare the IV estimates to OLS e stimates. I find that alcohol availability is positively and significantly related to all crime rates whether the IV estimator or the OLS estimator is used. However, there is evidence that the OLS estimator imparts a small downward bias to the effects of alcohol on crime rates.

The rest of the article is organized as follows. Section 2 introduces the econometric model used to investigate the relationship between alcohol availability and crime. Section 3 discusses the data used to estimate the model, while section 4 presents and discusses the statistical results. Section 5 concludes the article.

2. Model

Any sale or transfer of alcohol in the state of Michigan requires a license. All aspects of alcohol distribution in the state, including licensing, taxation, advertising, and hours of operation are controlled by the Michigan Liquor Control Board (MLCB). Although the Board grants several types of licenses, there are four groups of commercial liquor licenses granted by the MLCB [3] Any person over the age of 21, business, or organization, can apply for a liquor license. The application is submitted to the MLCB in Lansing on the Board's application form and is approved or denied after investigation and an open public hearing at which the application can be challenged on many grounds. The applicant should demonstrate that she/he/it has strong financial strength and a good physical plant in which to serve alcohol and that the location "adequately services the public." Other things the MLCB considers in granting or refusing a license include, but are not limited to, the opinions of local residents, legislative bod ies, and law enforcement; effects of the liquor license on business development at the location; and its effects on the health, welfare, and safety of citizens at the location. Although licenses are granted for a three-year period, a license can be revoked by the MLCB at any time. These considerations imply that an applicant for a liquor license will have to incur cost to overcome opposition, if any, of local residents to the license. It is reasonable to assume that the stronger the opposition, the greater the cost incurred to acquire the license, all things equal.

U.S. data suggest a strong positive relationship among alcohol use, alcohol availability, and crime rates in the United States (Parker 1993; Scribner, MacKinnon, and Dwyer 1995; Chaloupka and Weschler 1996; Greenfeld 1998; Parker and Cartmill 1998). While alcohol consumption could lead to criminal behavior or criminal victimization through its pharmacological effects or through overvaluation of the net benefit from criminal behavior, it is not clear why alcohol availability increases crime. One possible avenue to explain the positive correlation between alcohol availability and crime rate is that alcohol availability increases alcohol consumption through a reduction of its effective price even though the money price may not be lower. Assuming that a constant proportion (k) of those who consume alcohol commit crime or become victims of crime, the increased consumption of alcohol generated by increased availability may lead to higher crime rates. The relationship between alcohol availability and crime may ther efore be an indirect one. My approach to investigating the relationship between alcohol availability and crime rates follow this line of thinking.

The model used to investigate the effect of alcohol availability on crime rate is a reduced-form crime equation. The crime equation is the traditional Becker supply of offense model that is expanded to include alcohol availability as an explanatory variable (Becker 1968; Erhlich 1973; Eide 1998). [4] My model is similar to the one employed by Markowitz and Grossman (1998a, b) and Cook and Moore (1993b). As argued in the criminology literature, alcohol use or the environment in which alcohol is used is an intervening variable in criminal behavior or criminal victimization. The basic assumption in this article is that a constant proportion of those who drink alcohol will be involved in a criminal situation either as offenders or as victims. Of those who drink and are involved in criminal situations, I further assume that the more alcohol they drink, the more they are likely to be involved in a crime. [5]

I recognize that not all consumers of alcohol commit crimes or will be victims of crime and not all criminals or crime victims use alcohol. I assume that a fraction of those who consume alcohol will commit a crime or will be the victims of crime either because alcohol consumption clouds their judgment, hence leading them to overestimate the net benefit from criminal activity, or because of the pharmacological effects of alcohol. This implies that the crime rate is, in part, dependent on alcohol consumption as well as the expected net benefits from criminal behavior and personal characteristics of the decision maker. The general crime-generating equation can be written as

[CR.sub.i] = [CR.sub.i](A, X, Z) [partial][CR.sub.i]/[partial]A, [partial][CR.sub.i]/[partial]X [greater than] 0, (1)

where [CR.sub.i] is crime rate for crime i, A is alcohol consumption, X is expected net returns to criminal activity, and Z is a vector of socioeconomic characteristics. An increase in alcohol consumption increases criminal behavior and criminal victimization and, consistent with the economics of crime literature, an increase in expected benefits from crime will increase criminal activity, all things equal.

I assume that the demand for alcohol consumption depends on the price of alcohol, alcohol availability as indicated by the number of alcohol outlets, and consumers' income. There could be several reasons why the number of alcohol outlets will have a positive impact on alcohol consumption. First, accessibility of alcohol, as measured by geographical distance, decreases the time cost of finding and consuming alcohol. Second, alcohol outlets act as a form of advertisement that might tip the balance of a marginal consumer toward the purchase and consumption of alcohol. Third, easy availability of alcohol may give youths the impression that alcohol consumption is desirable, hence increasing the possibility of alcohol addiction. These youths then grow up to be alcohol users. It is also likely that people who live in a neighborhood with easy availability of alcohol accept alcohol use as part of the social cultural norms and use alcohol. This argument is consistent with the results of Moore and Cook (1995) as well a s those who find that youths who live in households or counties with easy availability of alcohol are more likely to be addicted to alcohol than their counterparts who live in 'dry' homes or counties (see Brown, Jewell, and Richer 1996). In view of these considerations, I write the alcohol demand equation as

[A.sup.D] = A([P.sub.A], S, y) [partial][A.sup.D]/[delta][P.sub.A] [less than] 0, [partial][A.sup.D]/[delta]S [greater than] 0, (2)

where [P.sub.A] is the money price of alcohol, S is the number of alcohol outlets, y is income, and [A.sup.D] is the demand for alcohol. The effects of income on the demand for alcohol will depend on whether alcohol is considered a normal good or an inferior good.

Although the MLCB grants licenses for the sale of alcohol, the Board does not set the retail price of alcohol. There is spatial variation in the retail price of alcohol in the state of Michigan and in the city of Detroit. I assume that the supply of alcohol is [A.sup.S] = h([P.sub.A]), h' [greater than] 0. The general solution for the equilibrium price and quantity of alcohol consumed is a function of the number of alcohol outlets and income,

[A.sup.*] = [A.sup.*](S, y) [[P.sup.*].sub.A](S, y)

where [A.sup.*] is the equilibrium quantity of alcohol purchased and sold at a location. Assuming that the fixed cost of opening an alcoholic outlet is F and k(n) is the variable cost of selling n units of alcohol, the equilibrium number of alcoholic outlets is a decreasing function of the fixed cost of opening an outlet (F). [6] F includes the cost of obtaining and maintaining a license and preparing the physical premise for the sale of alcohol. Formally, the number of alcohol outlets is given as

[S.sup.*] = [S.sup.*](F)[partial][S.sup.*]/[partial]F [less than] 0. (3)

Substituting the equilibrium number of outlets into the equilibrium alcohol demand and substituting the resulting alcohol demand into the crime equation, a reduced-form crime equation can be written as

[CR.sub.i] = [psi]([S.sup.*](F), y, X, Z) [partial][CR.sub.i]/[partial][S.sup.*] [greater than] 0. (4)

This reduced-form crime equation indicates that alcohol outlets increase crime but they do so indirectly through increased consumption of alcohol.

While alcohol use and crime may be positively related for individuals, this relationship may not be a linear one. Crime may increase with alcohol use up to the point where the individual is incapable of functioning, let alone being involved in a crime. Of course, victims of crime become easier prey the more alcohol they consume, all things equal. The relationship between alcohol use and crime for the individual may be a quadratic one, reaching a maximum with alcohol consumption and then falling again. I assume that different people have different levels of alcohol tolerance beyond which they cease to function. It is reasonable to assume that, within reasonable limits of alcohol availability and consumption, there are more people who can function than those who cannot with alcohol use. As alcohol consumption increases, more and more people exceed their normal operating capacity and get involved in criminal situations. At the aggregate level, I expect crime to be positively related to alcohol availability and use.

I note that my specification of the effects of alcohol on crime has some similarities as well as differences in comparison with the specification developed by other researchers. It is different from the models used by criminologists to investigate the effects of alcohol availability on crime in the sense that I am not seeking to investigate any bidirectional relationship between alcohol and crime. This article focuses on investigating the effects of alcohol availability rather than control policies on crime, as has been the case with studies conducted by economists. However, to the extent that alcohol control policies, such as taxation on alcohol, age limitations on purchase and use of alcohol, licensing, limits on hours of operation of alcohol outlets, and the use of zoning laws to limit the location of outlets, affect the availability of alcohol, this model is similar in spirit to models that investigate the effects of alcohol control policies on crime. In addition, I control for other variables that affec t crime.

To estimate Equation 3, I need to specify the functional form of the equation as well as define the elements of X and Z. Although different functional forms of the crime equation have been estimated by different researchers with different results (Eide 1998), I choose a linear functional form for the sake of simplicity without sacrificing explanatory power. The model developed above indicates that crime rate depends positively on the expected net benefits from crime, socioeconomic characteristics, and the availability of alcohol. The net benefit from crime depends on the actual gains from crime, the opportunity cost of committing crimes, the probability of punishment, and the size of the punishment. There are potentially several benefits from criminal activity, some psychic, others emotional, while others may be economic in nature. I do not have variables to measure the psychic or emotional benefits and costs of crime to either the perpetrator or the victim. I use per capita income (INC) to proxy the size of economic gain from criminal activity (the loot) since the size of the loot will depend on the average income in the community. I proxy the opportunity cost of engaging in criminal activity by the average level of educational attainment (ED UC) in the community. [7]

I had no data on the probability of punishment or the size of punishment. Besides, criminal sanctions are imposed at the state, county, or municipality level. Since my unit of analysis is the census tract from one city, the size and probability of punishment is not likely to vary across census tracts. I therefore do not include sanctions variables in the reduced-form equation I estimate. [8] The elements of Z included in the crime equation are the proportion of the population in the prime crime age group (YOUTH), population density (DENS), proportion of owner-occupied houses in a census tract (OWN), and the proportions of the population that are African-American (BLACK) and Hispanic (HISPANIC). [9,10] These variables are derived from the economics of crime literature (Erhlich 1973; Myers 1980; Gyimah-Brempong 1997; Eide 1998). I proxy alcohol availability by the number of alcohol licenses (LICENSE) in a census tract.

The reduced-form crime equation I estimate is given as

[CR.sub.i], = [[alpha].sub.0] + [[alpha].sub.1] YOUTH + [[alpha].sub.2]BLACK + [[alpha].sub.3]HISPANIC + [[alpha].sub.4]LICENSE + [[alpha].sub.5]INC + [[alpha].sub.6]EDUC + [[alpha].sub.7]DENS + [[alpha].sub.8]OWN + [epsilon], (5)

where [epsilon] is a stochastic error term and all other variables are as defined in the text above. In accordance with the criminology and economics of crime literature, I expect the coefficient of YOUTH to be positive while that of EDUC is expected to be negative. The coefficient of INC is expected to be positive if it reflects the potential gains from criminal activity in the sample; it will not be positive otherwise (Eide 1998). In accordance with my hypothesis that alcohol availability increases crime rates, I expect the coefficient of LICENSE to be positive. The coefficients of the other variables cannot be signed a priori.

3. Data

The dependent variable in this model is CRIME. I measure CRIME as the Federal Bureau of Investigation (FBI) part 1 index crime rate. Ideally, I should use each of the seven FBI part 1 index crimes as a dependent variable and hence estimate seven equations. However, because my unit of analysis--census tract--is so small, disaggregating the index crimes to all seven components may leave a lot of tracts with zero observations for a large number of these crimes. Because of this, I use four alternative measures of FBI index crime as my measures of the crime rate. The first measure of crime used is the total crime index (TOTCRIM) encompassing all the seven FBI part 1 index crimes--homicide, rape, aggravated assault, robbery, burglary, larceny, and motor vehicle theft. In addition to this measure of crime, I use three other crime indices--the index of violent crimes, defined as an aggregate of homicide, rape, aggravated assault, and robbery (VCRIME); property crimes index, defined as an aggregate of burglary, larce ny, and motor vehicle theft (PCRIME); and homicide rates (HOME)--as dependent variables. Each crime index is measured as the number of crimes per 10,000 people.

The FBI index crimes are based on data supplied by various law enforcement agencies to the FBI. Unlike the National Crime Victimization Study (NCVS) data, which are based on a survey of individuals, the FBI index crimes data only measure crimes known to law enforcement agencies. Changes in the rate at which crimes are reported to the police change the index crime rates even though actual crimes may not have changed. A comparison of crime data from the FBI Uniform Crime Reports to the NCVS data indicates that crimes are seriously underreported to the police. In 1994, it was estimated that about 60% of all crimes were not reported to the police. [11] Moreover, the degree of underreporting differs across crimes. This measurement error implies that using the FBI index crimes as my measure of crime could potentially lead to biased and inconsistent estimates. Although the NCVS data do not suffer from this serious underreporting, they are only available for broad geographic areas; they are not available at the cens us tract level. I note, however, that Myers (1980) find no evidence of biased or inconsistent estimates when one uses the FBI index crimes as the relevant measure of crime. He also found that it does not matter for consistent estimates whether or not one corrects for underreporting.

I measure LICENSE as the total number of alcohol licenses of all types granted per 1000 people in a census tract. Adjusting LICENSE for population eliminates the possibility that it is correlated with population density. The measure of LICENSE used here aggregates beer, wine, and liquor licenses together and it makes no distinction between licenses for on-premise consumption or carry-out drinks. Ideally, alcohol licenses should be disaggregated to indicate what type of alcohol can be sold--beer, wine, or liquor--and under what conditions alcohol is sold-- on-premise consumption versus off-premise consumption-since these conditions could have a significant effect on crime rate. For example, studies in Australia and the United Kingdom indicate that more violence occurs in or around drinking bars and pubs than around stores where alcohol is sold on a take-out basis (Hutchinson, Henderson, and Davis 1995). Unfortunately, the Detroit alcohol outlet data do not disaggregate licenses in any way. My measure of alcoh ol licenses also does not show whether or not the license is currently being utilized to sell alcohol. The inability to disaggregate alcohol licenses by type is probably the weakest part of my data. Second, most of the socioeconomic variables are for 1990 while the crime and license data are for 1992. If the socioeconomic variables changed dramatically during the two-year interval, my estimates would not be valid. These data problems, of course, imply that the results should be interpreted with caution and also highlight the need for more disaggregated data collection.

I proxy YOUTH by the proportion of the population that is 14-34 years old. BLACK and HISPANIC are measured as the proportions of the population that are Black and Hispanic, respectively. INC is measured as the per capita income from all sources of income in 1991 constant dollars, while OWN is the proportion of owner-occupied homes in a census tract. Education (EDUC) has been measured in several ways, including average years of schooling of the adult population, the proportion of the adult population that has high school or more education, and the proportion of the adult population that has a college education, in the economics of crime literature (Erhlich 1973; Myers 1980; Eide 1998). The census data provide the raw numbers of the adult population (25 years and older) that had fewer than 9 years of education, had between 9 and 12 years of education, had graduated from high school, had some college education, or had received a bachelor's degree or more education. I measure EDUC as the proportion of the adult population with a bachelor's degree or more of education. DENS is population density measured as the number of persons per square mile in a census tract.

The model developed in section 2 indicates that the number of alcohol outlets in a census tract depends on the fixed cost of opening an outlet. Alcohol license fees are set by the state and the city; hence, they do not vary across census tracts in a city. I assume that the more residential a census tract is, the more opposition there will be to the operation of an alcohol outlet in that neighborhood, necessitating investors to invest more time and resources to obtain an alcohol license. I therefore assume that the fixed cost of operating an alcohol outlet in a commercialized census tract will be lower than in a residential census tract. I proxy the level of commercialization by the number of gas stations in a census tract (GAS). I also use the median rent in the census tract (MRENT) as an additional indicator of fixed cost of opening and operating an alcoholic outlet in a census tract.

The socioeconomic variables (INC, BLACK, HISPANIC, DENS, EDUC, OWN, GAS, MRENT) were obtained from Census of Population and Housing, 1990: Summary Tape Files 1 and 3, Michigan, (Washington, D.C., Bureau of the Census, 1991). The crime rates (TOTCRIM, VCRIME, PCRIME, HOME) were obtained from the city of Detroit Police Department and were compiled by the Michigan Metropolitan Information Center (MIMIC) and were convet-ted to the census tract level using GIS methodology. The crime data were for 1992. The alcohol license data were obtained from MLCB (Lansing, Michigan, State of Michigan, 1994) and they are for 1992. There were a total of 323 census tracts in the city of Detroit data. However, there were a few observations that had some missing variables. Excluding those observations left me with a total of 315 usable observations. I limited the sample to the city of Detroit in part because of data availability, that is, crime data by census tract were not available for the entire metropolitan area.

All data were collected at the census-tract level. There are some advantages to using the census tract as the unit of analysis. For example, the relationship between alcohol availability and crime is less likely to be confounded by other factors, such as policy differences, than in the case where a large geographical area, such as the state, is the unit of analysis. However, the use of census tract as the unit of analysis has its own disadvantages. Apart from the problem of getting the appropriate data for many of the theoretically relevant variables, there is the problem of spillover effects. It is more likely that an individual may buy and consume alcohol in one census tract and commit a crime in another census tract than it is for an individual to consume alcohol in one city (state) and commit crime in another. In spite of these problems, I believe that the use of census tract data adds to our understanding of the relationship between alcohol availability and crime. At least any relationship between alcoh ol availability and crime rate I find is not driven by policy differences or differences in the nominal price of alcohol.

Summary statistics of the sample data are presented in Table 1. It is clear from Table 1 that the average crime rate in Detroit is high but variable across the city's census tracts, as shown by the large standard errors relative to the mean crime rates. Alcohol license density in the city of Detroit is very high and variable, as the standard errors of LICENSE indicate. The number of alcohol licenses across census tracts ranges from 0 to 67-an unusually wide range. The correlation between alcohol density and crime rates across the city is relatively high. The Pearson correlation coefficients between LICENSE and TOTCRIM, PCRIME, VCRIME, and HOME are 0.49, 0.51, 0.39, and 0.36, respectively, and all these correlation coefficients are significant at a 99% confidence level. The socioeconomic variables similarly show wide variations across census tracts in the city. One characteristic of the city of Detroit that stands out from the sample statistics is that racial minorities make up an unusually high proportion of its population--African-Americans and Hispanics make up about 78% of the city's population. Per capita income is very low and poverty and unemployment rates are very high for a declining industrial city such as Detroit.

4. Results

Because LICENSE is not strictly exogenous, I use an IV estimator to estimate Equation 4. [12] Another reason for using the IV estimator to estimate the crime equation is that both the dependent variable and alcohol use may be measured with error. Given that the measurement error in the dependent variable may be correlated with some of the regressors, the IV estimator is the appropriate estimator to use. Staiger and Stock (1997) argue that IV estimates are biased toward OLS estimates when the instruments are weak but that maximum likelihood (ML) estimates are not so affected. I therefore present a limited information maximum likelihood (LIML) estimate of the crime equations to see if the estimates differ from the IV estimates. I regressed the log of crime on the logs of the explanatory variables, allowing me to interpret the coefficient estimates as elasticities. [3] My approach to presenting the results is as follows: I first present and discuss the IV estimates of the crime equation. I then present the LIML estimates and compare them to the IV estimates. Finally, I present OLS estimates of the crime equation and compare the results with those of the IV estimates to see if using OLS imparts substantial bias to the coefficient estimates.

In estimating Equation 4, I used the median rent (MRENT) and the number of gas stations in a census tract (GAS) as instruments for LICENSE. The F-statistic to test the null hypothesis that the coefficients of MRENT and GAS are jointly equal to zero in the first stage regression of LICENSE is 21.289. I therefore conclude that these instruments are reasonably strong instruments for LICENSE (Staiger and Stock 1997). The IV estimates are presented in Table 2. Generally, all four equations have relatively good fit, as indicated by the regression statistics. The null hypothesis that all slope coefficients are jointly equal to zero is rejected at [alpha] = 0.01 for all four equations. I used White's test to test for heteroskedasticity but could not reject the null hypothesis of homoskedastic disturbances at [alpha] = 0.01. It is possible that some variables that should belong to the crime equation have been excluded from it and instead are treated as instruments. Such restrictions will result in model misspecificat ion. I use Basmann's overidentifying restriction test (Basmann 1960), which is a test of the null hypothesis that none of the instruments should be included in the crime equation, to test the null hypothesis that the correct overidentification restrictions have been imposed. Basmann's F-statistics are presented at the bottom of the estimates in Table 2. I cannot reject the null hypothesis that the overidentifying restrictions I have imposed are correct at [alpha] = 0.05 in all four crime equations.

In the TOTCRIM equation presented in column 2, the coefficients of DENS and OWN are negative and statistically significant at conventional levels of significance, although the absolute magnitudes of these coefficients are low. The negative and significant coefficients of OWN and DENS indicate that increased home ownership and population density are negatively correlated with total crime rate. The negative coefficient of OWN could come from two sources. Home ownership implies higher stakes in the welfare of the community; hence, home owners work to reduce crime through preventive measures. It is also possible that home owners use their political clout to reduce the location of alcohol outlets in their neighborhoods as a means of holding up their property values. Additionally, a high percentage of ownership is positively correlated with income, suggesting that the neighborhood will have the resources to prevent crime, all things equal. The coefficient of YOUTH is positive and significant at [alpha] = 0.01, ind icating that the proportion of the population that is young in society is positively correlated with crime in the sample. This result is similar to the results obtained by earlier researchers who found crime rates to be positively correlated with the young population (Erhlich 1973; Myers 1980; Eide 1998).

The coefficient of INC is positive and significantly different from zero at [alpha] 0.01, with an estimated coefficient of 0.520. The positive coefficient of INC in the crime equation is consistent with my hypothesis that income is a proxy for criminal opportunities in this sample. The positive coefficient of INC is also consistent with the results of some researchers who have used income as a proxy for criminal opportunities. The coefficient of EDUC is negative and significant. The coefficient of BLACK is insignificant while the coefficient of HISPANIC is negative and significant at [alpha] = 0.01, indicating that the proportion of a census tract's population that is Hispanic is negatively correlated with total crime rate in the city of Detroit.

The coefficient of LICENSE in the TOTCRIM equation is positive, relatively large, and significantly different from zero at [alpha] = 0.01 or better. The positive and significant coefficient of LICENSE indicates that alcohol availability has a positive and significant effect on total crime rate. A 10% increase in the number of alcohol licenses in a census tract increases the total crime rate by about 9.2%, a relatively large response, all things equal. The positive and significant coefficient of LICENSE I find in the TOTCRIM equation is consistent with the results obtained by researchers who find that alcohol availability is positively related to crime rates. My results are also consistent with the results of studies that find that alcohol control policies have depressing effects on crime (Cook and Moore 1993b; Brown, Jewell, and Richer 1996; DiIulio 1995, 1996; Gruenewald, Ponicki, and Holder 1992; Parker and Cartmill 1998; Scribner, MacKinnon, and Dwyer 1995; Scribner et al. 1999).

It is possible that the motivation for committing property crimes differs from those for violent crimes. To test the possibility of a differential impact of alcohol availability on violent and property crimes, I partitioned the crime data into violent crime (VCRIME), property crime (PCRIME), and homicide (HOME); estimated the model; and compared the estimates from the three crime equations with the estimates from the total crime equation. The results of these estimates are presented in columns 3, 4, and 5, respectively, in Table 2.

The coefficient of YOUTH is positive and significantly different from zero at [alpha] = 0.01 in the PCRIME equation, but it is not significant in the VCRIME and HOME equations, suggesting that the proportion of the population that is young is positively correlated with property crimes but not with violent crimes or homicides. As in the TOTCRIM equation, the coefficient of BLACK is insignificant in the VCRIME equation while HISPANIC has a negative and significant coefficient in that equation. However, the coefficient of BLACK is positive and significant in the PCRIME and HOME equations while the coefficient of HISPANIC is negative and significant in both equations. The coefficient of INC is positive and significant in the PCRIME equation, but it is insignificant in the VCRIME and HOME equations. The coefficient of EDUC is insignificant in the VCRIME and PCRIME equations, but it is negative and significant in the HOME equation. The coefficients of DENS and OWN are negative and significantly different from zero at [alpha] = 0.01 in the PCRIME and VCRIME equations. These coefficients are, however, insignificant in the HOME equation. The coefficient estimates in the PCRIME and VCRIME equations are similar to the estimates in the TOTCRIM equation. There are, however, some differences between the estimates for HOME and those for TOTCRIM, suggesting that the motivation for homicide may be different from those for other crimes.

The coefficient of LICENSE is positive, relatively large, and statistically significant at a = 0.01 or better in both the VCRIME and PCRIME equations. The estimated coefficient of LICENSE in the VCRIME equation is about 0.83, while it is about 0.89 in the PCRIME equation. There is not a statistically significant difference between the estimated coefficient of LICENSE in the VCRIME and PCRIME equations and its counterpart in the TOTCRIM equation. The coefficient of LICENSE in the HOME equation is positive and significantly different from zero at [alpha] = 0.01. However, the absolute value of this coefficient estimate of about 0.12 is significantly far less than the coefficient estimate of LICENSE in the TOTCRIM, VCRIME, and PCRIME equations. The estimated coefficients in these equations imply that alcohol availability has a positive and significant impact on violent and economic crimes as well as on homicides. If one compares the estimated coefficient of LICENSE in the VCRIME and PCRIME equations with the coe fficient of LICENSE in the TOTCRIM equation, one observes that all three estimates are virtually identical. This implies that alcohol availability has similar effects on total, property, and violent crimes. There is, however, a quantitative difference between the effects of alcohol availability on homicides and on the other three crimes.

What is the effect of alcohol availability on crime rates? The estimates presented in Table 2 indicate that alcohol availability is positively correlated with all four crime rates investigated here. This positive relationship may be due to more people drinking alcohol and committing crimes as availability of alcohol increases or it may come from the same number of people drinking more alcohol as availability increases and committing more crimes after alcohol consumption. It may also reflect the possibility that crime victimization increases with alcohol availability and use in a complex way. [14]

The relationship between alcohol availability and crime is likely to be a complex one. While alcohol consumption may be positively correlated with its availability (Gruenwald, Ponicki, and Holder 1992; Cook and Moore 1993a; Van Oers and Garretsen 1993; Scribner et al. 1999), it is also possible that high-income residential neighborhoods are able to keep alcohol outlets away. These neighborhoods are also likely to be the ones with enough resources to fight crime, keeping crime rates low. On the other hand, low-income neighborhoods may not have the resources to prevent a high density of outlets or to prevent crime. This will imply that neighborhoods with high alcohol outlet densities are also the neighborhoods with high criminal capital (Gyimah-Brempong 1997). Indeed, there is evidence that alcohol and income interact in some ways to affect crime. [15]

Staiger and Stock (1997) and Angrist, Imbens, and Krueger (1999) show that, in the presence of weak instruments, IV estimates are biased toward OLS estimates while ML estimates are not subject to such bias, although the ML estimates have larger confidence bands. Although the instruments I use are relatively strong, I nevertheless use LIML to estimate the crime equations to see if there are any qualitative differences in the relationship between alcohol availability and crime rates using the two estimators. The LIML estimates are presented in Table 3. The coefficients of the TOTCRIM equation are presented in column 2; column 3 presents the estimates for VCRIME; column 4 presents the estimates for PCRIME; while the estimates for HOME are presented in column 5. The coefficient estimates are generally similar in sign, absolute magnitude, and statistical significance as their counterparts in Table 2. The coefficient of LICENSE in the LIML estimate is not significantly different from the IV estimate presented in T able 2. These estimates suggest that the IV estimates may not be biased. The LIML estimates confirm my results that alcohol availability is positively related to crime rates.

The coefficient estimates presented in Table 2 are predicated on the assumption that LICENSE cannot be treated as exogenous. It is likely that the instruments I use for LICENSE are also correlated with cnme rates; hence, the error term. This his renders the IV estimator inconsistent. [16] The solution to this problem is to find better instruments. I do not have any other reasonable instruments from the sample data. An alternative strategy is to include as many environmental variables as possible in the crime equation and estimate it using OLS. I follow this strategy and estimate Equation 5 by OLS and compare the results to the IV estimates presented in Table 2. Because of the possibility that the instruments I use may not be valid instruments, I include the instruments as additional regressors. If the instruments I use are valid, they should not have significant coefficients in the crime equations that also include LICENSE as an explanatory variable.

The results of the OLS estimates are presented in Table 4. Column 2 presents the estimates for the TOTCRIM equation; column 3 presents the estimates for the VCRIME equation; column 4 presents the estimates for the PCRIME equation; while column 5 presents the estimates for the HOME equation. The OLS estimates indicate that the model explains a relatively large proportion of the variance in crime rates across census tracts in Detroit. The coefficient of MRENT is significant in most of the equations while that of GAS is not significant in any of the crime equations. An F-test to test the null hypothesis that the coefficients of these two variables are jointly equal to zero in each of the crime equations is soundly rejected at [alpha] = 0.01 for each of the crime equations. [17] However, I note that the coefficients of INC and OWN are insignificant when these instruments are added as regressors in the crime equations. The significant coefficient on the environmental variables may stem from the high correlation between these variables and some of the regressors. [18] These environmental variables may be preempting INC and OWN in the crime equations. This result could also mean that these environmental variables interact with some of the explanatory variables to influence crime rates in ways that have not been captured by the model I use here. [19]

The coefficient of LICENSE in the OLS estimates remains positive and significantly different from zero at [alpha] = 0.01 in all four crime equations, with estimated coefficients of 0.2576, 0.2003, 0.1788, and 0.0993 for TOTCRIM, VCRIME, PCRIME, and HOME equations, respectively. I note, however, that the OLS estimates are very low in absolute magnitude and are less precisely estimated compared with their IV counterparts partly because of collinearity among the regressors. The OLS estimates indicate that treating LICENSE as exogenous does not qualitatively change my conclusion that alcohol availability has a positive and significant effect on crime rates. However, there is a downward bias of the coefficient of LICENSE in the OLS estimates compared with the IV and LIML estimates. A Hausman specification test rejects the null hypothesis that the OLS estimates are equal to the IV estimates at any reasonable confidence level for all four crime equations. [20] I therefore do not consider the OLS estimates as being the same as the IV estimates.

The estimates indicate that LICENSE has a positive and significant impact on crime rates. However, it is possible that alcohol availability, as proxied by LICENSE, does not have any explanatory power of its own in the crime equation; it may be preempting one or more of the explanatory variables with which it is correlated. To further investigate the effects of alcohol availability on crime, I reestimate the crime equations without LICENSE as a regressor and test to see if the truncated equations are different from the full equations. The coefficient estimates of the truncated equations are quantitatively different from those of the full model. [21] In particular, the absolute magnitude of the coefficient estimates from the truncated equations are lower and less precisely estimated than their counterparts in the full equations. This suggests the existence of possible omitted variable bias. In addition to differences in the magnitude in the coefficient estimates, the calculated F-statistics to test the null hyp othesis that LICENSE has no effect on crime rate are 39.61, 27.78, 26.47, and 16.18 for the TOTCRIM, PCRIME, VCRIME, and HOME equations, respectively. I conclude from these statistics as well as the highly significant coefficients of LICENSE in Tables 2-4 that alcohol availability has an independent positive and significant effect on all crime rates.

The results presented here are consistent with the results of previous research on the relationship between alcohol availability and crime. In particular, the positive correlation between alcohol license density and crime rate I find here is similar to the results obtained by DiIulio (1995), Sloan, Reilly, and Schenzler (1994), and Cook and Moore (1993b) and those found in the criminology literature (Homel, Tomsen, and Thommeny 1992; Van Oers and Garretsen 1993; Parker 1995; Scribner, MacKinnon, and Dwyer 1995; Valdez et al. 1995; Scribner et al. 1999, among others). The positive effect of alcohol availability on crime rates is consistent with the results obtained by Chaloupka and Weschler (1996), Jewel and Brown (1995), Stitt and Giacopassi (1992), among others. It is also consistent with the results of studies that find that alcohol control policies have effects on crime (Brown, Jewell, and Richer 1996; Cook and Moore 1993a; Moore and Cook 1995; Markowitz and Grossman 1998a, b; Joksch and Jones 1993).

My results indicate that alcohol availability has a positive and statistically significant impact on crime rates, all things equal. The results imply that public policy that decreases the availability of alcohol may decrease the crime rate in communities. Society could decrease crime rates by decreasing the availability of alcohol or at least decreasing the concentration of alcohol outlets through differential taxation based on density of outlets at a location or through zoning laws. Alcohol availability could also be decreased through increased taxes on alcohol or increases in the legal drinking age. The policy implications derived from my results are similar to those implied by the results of earlier investigations of the effects of alcohol control policies on various forms of crime. The research implication flowing from this result is that researchers who estimate the supply of crime equations should include alcohol availability as one of the explanatory variables in their equations.

The results presented in this article should be interpreted with caution. It is likely that crime victims who have consumed alcohol may not report the crimes to the police for various reasons. Since I postulate that alcohol consumption is positively correlated with alcohol availability (LICENSE), the implication is that the measurement error in crime rate is correlated with LICENSE. This situation would bias the coefficient of LICENSE in the crime equation toward zero. Given the possibility that such measurement error exists, the estimates presented here should be considered as a lower bound of the effects of alcohol availability on crime. It is gratifying to note that my estimates are significantly different from zero at any reasonable level of confidence. Because of the geographical proximity of census tracts in a city, it is likely that the census tract data are spatially autocorrelated. Millar and Gruenewald (1997) argue that failure to correct for spatial autocorrelation will not result in biased coeffi cient estimates but will bias the standard errors of the estimates. They show that OLS estimates increase the standard errors of the estimates. In my estimates, I find that the coefficient estimates of LICENSE to be significant, indicating that correcting for spatial autocorrelation will only strengthen my results. [22]

5. Conclusion

This article uses census tract data from the city of Detroit and a reduced-form equation to investigate the relationship between alcohol availability and crime. Measuring alcohol availability by alcohol license density, I find that alcohol availability has a significantly positive effect on total crime rate, violent crime rate, property crime rate, and homicide rate in the city of Detroit. The calculated alcohol elasticity of crime rates are 0.92, 0.82, 0.87, and 0.12 for total crime rate, violent crime rate, property crime rate, and homicide rate, respectively. These effects do not change whether one uses an IV estimator or an LIML estimator. However, the estimated crime elasticities are quantitatively low when I use the OLS estimator to estimate the crime equations. This may be an indication that the OLS estimator underestimates the effects of alcohol availability on crime rates. The results confirm the positive relationship between alcohol availability and crime estimated by criminologists and economists. The research implication is that researchers who estimate crime equations should consider alcohol availability as one of the explanatory variables. The policy implication of the results is that policy makers could use alcohol control policies as a means of fighting crime.

The results of this article should, however, be interpreted with caution. The model used to investigate the relationship between crime and alcohol availability is very rudimentary. Second, there is no control variable for deterrence in the crime equation, which may possibly bias some of the coefficient estimates. Third, LICENSE itself was measured at a highly aggregated level. Finally, the model does not control for several environmental variables that could affect the crime rate. It is possible that alcohol availability is only a proxy for unmeasured heterogeneity among census tracts. Hopefully, future studies will correct these weaknesses and cover a wider geographic area, such as the MSA. With these caveats, my results should be treated as indicative rather than definitive.

(*.) Department of Economics, University of South Florida, Tampa, FL 33620, USA; E-mail kgyimah@coba.usf.edu.

Financial support for this research was provided by a Creative Research Grant, Office of Research, University of South Florida. An earlier version of this article was presented at the Southern Economic Association Annual Meeting in Atlanta, Georgia, November 1997. Gabriel Picone and John Swinton provided helpful comments on an earlier draft. I thank two anonymous referees for providing comments that greatly improved the article. I bear the sole responsibility for any remaining errors.

(1.) See the discussion of the weaknesses of census tract data in section 3. Also see Scribner et al. (1999) for some of the advantages of using the census tract as the unit of analysis.

(2.) In this article, I use the terms African-American and Black interchangeably. They refer to Americans of African descent.

(3.) Two types on-premise and two types off-premise licenses are grant by the MLCB. Class C, B hotel, and club licenses are on-premise retail licenses granted for the sale of beer, wine, and liquor. License fees for this group of licenses consist of a fixed amount plus an additional charge per bar or number of rooms per year. Tavern and A hotel licenses are on-premise licenses for the sale of beer and wine only. License fees are a fixed amount for the year. The two types of off-premise licenses are SDD licenses, authorizing one to sell liquor only, and SDM licenses, authorizing one to sell beer and wine only. The fee for an SDM license is a fixed amount annually while the fee for an SDD license consists of a fixed amount plus an ad valorem tax on liquor sales.

(4.) Becker's model (Becker 1968) sees criminal behavior as the result of a constrained optimization decision on the part of the criminal. It has become the base model of the economics of crime literature. For a good review of the economies of crime literature, see Eide (1998).

(5.) It is possible that an individual can drink up to the point where he/she cannot function, let alone commit a crime. Of course this individual will be an easy target for the criminal, hence a victim of crime. It is therefore possible that the relationship between alcohol use and crime is a quadratic one. One should note that this article investigates the effects of alcohol availability rather than alcohol consumption on crime.

(6.) I note that this fixed cost may include the cost of obtaining a license at a particular location. Although the nominal license fee does not vary by location, the cost of overcoming opposition to the license application and the cost of preparing the physical facility will vary by location.

(7.) See Eide (1998) for a discussion of the variables that have been used in crime-generating equations.

(8.) I tried, without success, to obtain data from police patrol patterns at the census tract level from the Detroit Police Department.

(9.) I note that the inclusion of racial minorities does not imply that race per se causes crime. Race is only a proxy for some unobserved variable that may be highly correlated with race. For more on the correlation between race and crime, see Gyimah-Brempong (1997).

(10.) Other socioeconomic variables that have been included in crime-generation equations are the unemployment rate, poverty rate, and percent on public assistance. These variables are, however, all correlated with income and education. In the interest of parsimony, I do not include these variables in our model.

(11.) See National Crime Victimization Survey, 1991-1996, ICPSR 6406.

(12.) A Hausman exogeneity test rejects the null hypothesis that LICENSE is exogenous in all four crime equations.

(13.) In order to avoid taking the log of zero, in case LICENSE had a value of zero, I defined the log of license as LICENSE = log(LICENSE + 0.5).

(14.) I experimented by including the square of LICENSE as an additional regressor. However, the quadratic term was highly collinear with LICENSE, leading to a deterioration of the precision of the estimates. I note that F-tests to test the null hypothesis that LICENSE and its square do not affect crime rates produced F-statistics of 22.50, 20.958, 23.42, and 18.91 for TOTCRIM, PCRIME, VCRIME, and HOME, respectively, leading to a rejection of the null. Because the coefficients in this equation were imprecisely estimated, I did not use them to calculate the effects of alcohol outlet density on crime rates.

(15.) When I included the interaction between income and license (INC X LIC) in the regression, the coefficient of this interaction was negative, relatively large, and significantly different from zero at [alpha] 0.01. Inclusion of the interaction term also increased the magnitude of the coefficient of LICENSE as well as the precision of all coefficient estimates in the equation while decreasing the absolute magnitude of other coefficients. This suggests that income and alcohol availability interact in a complex way to affect crime that may not I save been captured by my model.

(16.) I thank an anonymous referee for pointing to this error and suggesting a possible solution.

(17.) The calculated F-statistics are 8.168, 6.083, 9.858, and 7.519 for the TOTCRIM, VCRIME, PCRIME, and HOME, respectively.

(18.) For example, the Pearson correlation coefficients between INC and MRENT and between MRENT and OWN are 0.80 and 0.586, respectively.

(19.) See the effects of the interaction between INC and LICENSE discussed in footnote 23.

(20.) Hausman m-statistics, reported at the bottom of Table 2, are 986.214, 108.431, 89.634, and 34.482 for TOTCRIM, VCRIME, PCRIME, and HOME, respectively.

(21.) I do not report the coefficient estimates of these truncated equations for space considerations. They are, however, available upon request.

(22.) The reasonable thing to do is to test and correct for possible spatial autocorrelation. However, I do not have the data to calculate the connection matrix (W) that are necessary for such a test. I note that Scribner et al. (1999) do not control for spatial autocorrelation and obtained results that are similar to mine. On the other hand, Millar and Gruenewald (1997) correct for spatial autocorrelation and obtain results that are qualitatively similar to the results obtained in this study.

References

Angrist, Joshua D., G. W. Imbens, and Alan B. Krueger. 1999. Jackknife instrumental variables estimation. Journal of Applied Econometrics 14:57-67.

Basmann, R. L. 1960. On finite sample distributions of generalized classical linear identifiability test statistics. Journal of the American Statistical Association 55:650-9.

Becker, Gary S. 1968. Crime and punishment: An economic approach. Journal of Political Economy 76:169-217.

Blount, W. R., I. J. Silverman, C. S. Sellers, and R. A. Seese. 1994. Alcohol and drug use among abused women who kill, abused women who don't, and their abusers. Journal of Drug Issues 24:165-77.

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

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

Cook, Philip, and Michael Moore. 1993a. Violence reduction through restrictions on alcohol availability. Alcohol, Health, and Research World 17:148-54.

Cook, Philip, and Michael Moore. 1993b. Economic perspectives on reducing alcohol-related violence. In Alcohol and interpersonal violence: Fostering multidisciplinary perspectives, edited by Susan E. Martin. Washington, DC: National Institute on Alcohol Abuse and Alcoholism Research Monograph 24. Washington, DC: U.S. Government Printing Office.

DiIulio, John J., Jr. 1995. Broken bottles: Liquor, disorder, and crime in Wisconsin. Wisconsin Policy Research Institute Report 5(4).

DiIulio, John J., Jr. 1996. Help wanted: Economists, crime, and public policy. Journal of Economic Perspectives 10:3-24.

Eide, E. 1998. Economics of criminal behavior, In Encyclopedia of law and economics. Web Resources, European site: http://encyclo.findlaw.com/lit/index.html, No. 8100.

Erhlich, Isaac. 1973. Participation in illegitimate activities: A theoretical and empirical analysis. Journal of Political Economy 81:521-67.

Federal Bureau of Investigation (FBI). 1996. Sourcebook of criminal justice statistics, 1994, Table 2.5. Washington, DC: Department of Justice.

Greenfeld, L. 1998. Alcohol and crime: An analysis of national data on the prevalence of alcohol involvement in crime. NCJ No. 168632.

Gruenewald, Paul J., Pat Madden, and Kathy Janes. 1992. Alcohol availability and the formal power and resources of state alcohol beverage control agencies. Alcoholism: Clinical and Experimental Research 16:591-7.

Gruenewald, Paul J., William R. Ponicki, and Harold D. Holder. 1992. The relationship of outlet densities to alcohol consumption: A time series cross sectional analysis. Alcoholism: Clinical and Experimental Research 17:38-47.

Gyimah-Brempong, Kwabena. 1997. Crime and race in the U.S.: What is the connection? In Race, markets and social outcomes, edited by Patrick Mason and Rhonda Williams. New York: Kluwer Publishers.

Homel, Ross, Steve Tomsen, and Jennifer Thommeny. 1992. Public drinking and violence: Not just an alcohol problem. Journal of Drug Issues 22:679-97.

Hutchinson, G., M. Henderson, and J. Davis. 1995. Alcohol in the workplace: Cost and responses. Centre for Applied Social Psychology, University of Strathclyde, Research Monograph No. 59.

Jewell, R. Todd, and Robert W. Brown. 1995. Alcohol availability and alcohol related motor vehicle accidents. Applied Economics 27:759-65.

Joksch, H., and R. Jones. 1993. Changes in the drinking age and crime. Journal of Criminal Justice 21:209-21.

Lang, A. R., and P. A. Sibrel. 1989. Psychological perspectives on alcohol consumption and interpersonal aggression: The potential role of individual differences in alcohol-related criminal violence. Criminal Justice and Behavior 16:299-324.

Markowitz, Sara, and Michael Grossman. 1998a. Alcohol regulation and domestic violence towards children. Contemporary Economic Policy 16:309-20.

Markowitz, Sara, and Michael Grossman. 1998b. Effects of alcohol regulation on physical child abuse. NBER Working Paper No. 6629.

Millar, Alexander B., and Paul J. Gruenewald. 1997. Use of spatial models for community program evaluation of changes in alcohol outlet distribution. Addiction 92(Supplement 2):S273-S283.

Moore, Michael, and Philip Cook. 1995. Habit and heterogeneity in the youthful demand for alcohol. NBER Working Paper No. 5152.

Myers, Samuel L., Jr. 1980. Why are crimes underreported? What is the crime rate? Does it matter? Social Science Quarterly 61:23-43.

Parker, Robert N. 1993. Alcohol and theories of homicide. In Advances in criminology theory 4, edited by F Adler and W. Laufer. Somerset, NJ: Transactions Publishers, pp. 113-42.

Parker, Robert N. 1995. Bring the "booze" back in: The relationship between alcohol and homicide. Journal of Research in Crime and Delinquency 32:3-38.

Parker, Robert N. 1998. Alcohol and homicide in United States 1934-1995 or one reason why U.S. rates on violence may be going down. The Journal of Criminal Law and Criminology 88:1369-98.

Roizon, J. 1997. Epidemiological issues in alcohol-related violence. In Recent developments in alcoholism, edited by M. Galanter. New York: Plenum Press.

Rush, Brian R., Louis Glickman, and Robert Brook. 1986. Alcohol availability, alcohol consumption and alcohol related damage, I & II. Journal of Studies on Alcohol 47:1-18.

Saffer, Henry, and Frank Chaloupka. 1995. The demand for illicit drugs. NBER Working Paper No. 5238.

Scribner, Richard, D. Cohen, S. Kaplan, and S. Allen. 1999. Alcohol availability and homicide in New Orleans: Conceptual considerations for small area analysis of the effect of alcohol outlet density. Journal of Studies on Alcohol 60:310-6.

Scribner, Richard A., David P. MacKinnon, and James H. Dwyer. 1995. The risk of assaultive violence and alcohol availability in Los Angeles county. American Journal of Public Health 85:335-8.

Sloan, Frank, Bridget A. Reilly, and Christopher Schenzler. 1994. Effects of prices, civil and criminal sanctions, and law enforcement on alcohol-related mortality. Journal of the Study on Alcohol 55:454-64.

Sprunt, Barry, Paul Goldstein, Henry Brownstein, Michael Fendrich, and Sandra Langley. 1994. Alcohol and homicide: Interviews with prison inmates. Journal of Drug Issues 24:143-63.

Staiger, D., and James H. Stock. 1997. Instrumental variables regression with weak instruments. Economnetrica 65:557-86.

Stitt, B. G., and D. J. Giacopassi. 1992. Alcohol availability and alcohol-related crime. Criozinal Justice Review 17:268-79.

University of Michigan Institute for Social Research. 1998. Monitoring the future study 1997. Ann Arbor, MI: University of Michigan.

Valdez, Avelardo, Charles D. Kaplan, Russell L. Curtis, Jr., and Zenong Yin. 1995. illegal drug use, alcohol and aggressive crime among Mexican American and white male arrestees in San Antonio. Journal of Psychoactive Drugs 27:135-43.

Van Qers, J. A., and H. F. Garretsen. 1993. The geographic relationship between alcohol use, bars, liquor shops and traffic injuries in Rotterdam. Journal of Studies of Alcohol 54:739-44.

Table 1. Summary Statistics of Sample Data a
 Standard
Variable Mean Error Minimum Maximum
TOTCRIM 374.91 164.56 66.00 1088.00
VCRIME 72.31 37.19 10.00 311.00
HOME 1.3418 1.1003 1.00 7.00
POPULATION 3212.3 1348.06 62.00 6840.00
DENS 8263.92 3270.29 114.60 1677.30
BLACK (%) 75.40 29.3 1.100 99.600
HISPANIC (%) 2.99 7.81 0.16 58.20
OWN (%) 51.09 21.4 0.20 96.100
PCRIME 302.43 137.40 52.00 920.00
EDUC 6.31 8.04 0.21 50.43
LICENSE 6.35 5.6968 0.00 67.00
YOUTH (%) 33.92 4.76 14.30 63.60
INC ($) 9740.80 4464.08 6052.00 40,469.00
(a) N = 315.
Table 2. Instrumental Variables Estimates of Crime Equation a
 Coefficicnt
 Estimates
Variable TOTCRIM b VCRIME b PCRIME b HOME b
Constant 0.9821 *** 4.2891 *** 10.4097 l.0256 ***
 (2.941) (3.762) (0.332) (3.769)
YOUTH 0.0207 *** -0.0017 0.0086 * 0.0021
 (3.883) (0.638) (1.839) (0.877)
LICENSE 0.9161 *** 0.8249 *** 0.8692 *** 0.1194 ***
 (10.701) (11.492) (12.859) (4.305)
BLACK -0.0041 0.0011 0.0031 ** 0.0155 **
 (0.013) (0.958) (2.131) (2.573)
HISPANIC -0.0618 *** -0.187 *** -0.0142 *** -0.0292 ***
 (3.246) (3.206) (2.886) (2.701)
INC 0.5200 *** -0.1285 0.4214 *** 0.0438
 (6.454) (1.220) (2.738) (0.0790)
EDUC -0.1034 ** -0.1717 -0.3007 -0.0428 **
 (2.394) (0.589) (0.410) (2.468)
DENS -0.0058 *** -0.00004 *** -0.00001 *** 0.0271
 (2.819) (4.121) (2.690) (1.347)
OWN -0.0018 ** -0.0049 ** -0.0039 ** 0.0010
 (2.122) (1.970) (2.127) (0.857)
N 315 315 315 315
F 27.79 18.850 24.990 8.114
[R.sup.2] 0.4231 0.3340 0.3640 0.1655
Hauman's m 896.214 108.431 89.634 38.482
Bassman's F 1.6804 0.6138 2.0078 1.2243
 [2.302] [2.302] [2.302] [2.302]
(a) Absolute value of t-statistics in parentheses.
(b) (***), (**), (*) denote statistical significance at 0.01, 0.05,
and 0.10 levels, respectively.
Table 3. LIML Estimates of Crime Equation a
 Coefficient Estimates
Variable TOTCRIM b VCRIME b PCRIME b
Constant 2.5694 1.4440 0.1321
 (1.422) (1.491) (1.138)
YOUTH 0.0141 ** 0.0195 0.0207 ***
 (2.564) (1.591) (3.876)
LICENSE 1.0041 *** 0.8662 *** 1.0892 ***
 (9.016) (6.980) (12.694)
BLACK 0.0262 0.0129 0.0416 **
 (0.618) (0.293) (2.223)
HISPANIC -0.0675 *** -0.1074 *** -0.0527 **
 (2.576) (3.455) (2.531)
INC 0.2592 ** 0.1843 0.4181 ***
 (2.129) (1.489) (2.798)
EDUC -0.8516 ** -0.1706 *** -0.7956 ***
 (2.134) (3.206) (3.599)
DENS 0.0148 0.0952 * -0.0960 **
 (0.952) (1.717) (2.222)
OWN -0.0023 -0.0141 -0.0457 ***
 (2.522) ** (0.371) (2.731)
N 315 315 315
F 19.895 25.596 32.852
[R.sup.2] 0.2096 0.3625 0.5269
Hauman's m 867.487 111.871 91.289
Variable HOME b
Constant 2.8422 ***
 (4.048)
YOUTH 0.0018
 (0.903)
LICENSE 0.1298 ***
 (4.165)
BLACK 0.0011 **
 (2.165)
HISPANIC -0.0282 ***
 (2.746)
INC -0.0632
 (1.312)
EDUC -0.0475 **
 (2.357)
DENS 0.0264
 (1.426)
OWN 0.0009
 (1.331)
N 315
F 8.612
[R.sup.2] 0.1742
Hauman's m 30.639
(a) Absolute value of t-statisticsin parantheses.
(b) (***), (**) (*) denote statistical significance at 0.01, 0.05, and
0.10 levelsrespectively.
Table 4. OLS Estimates of Crime Equation a
 Coefficient
 Estimates
Variable TOTCRIM VCRIME PCRIME HOME
Constant -0.4658 -1.3156 1.4718 1.0461
 (0.394) (1.102) (1.121) (1.802)
YOUTH 0.2158 0.0093 * 0.2789 0.0023
 (1.1178) (1.696) (1.029) (0.555)
LICENSE 0.2576 *** 0.2003 *** 0.1788 *** 0.0993 ***
 (5.521) (5.455) (5.335) (2.852)
BLACK 0.0076 0.0213 0.0089 0.0146 ***
 (0.210) (0.477) (0.210) (6.741)
HISPANIC -0.0503 ** -0.0878 *** -0.0599 *** -0.0304 ***
 (2.190) (2.826) (3.154) (2.942)
INC 0.1679 -0.2137 0.2026 0.0526
 (1.171) (0.502) (1.360) (0.900)
EDUC -0.1521 -0.1574 *** -0.0761 ** -0.0428 **
 (1.397) (3.034) (2.389) (2.426)
DENS -0.1327 0.1542 -0.0503 0.0259
 (1.895) * (1.591) (0.821) (1.399)
OWN -0.0414 -0.0494 -0.0493 0.0243
 (1.530) (1.257) (1.223) (0.578)
GAS 0.0265 0.0340 0.0270 0.0595
 (1.238) (1.262) (1.248) (0.662)
MRENT 0.6624 0.4349 *** 0.7220 *** -0.0431
 (3.648) *** (3.863) (3.986) (0.558)
N 315 315 315 315
F 24.727 17.714 21.226 6.735
[R.sup.2] 0.4127 0.2988 0.3178 0.1501
(a) Absolute value of t-statistics in parentheses.
(b) (***), (**), (*) denote statistical significance at 0.01, 0.05,
and 0.10 levels, respectively.