The Price of Alcohol, Wife Abuse, and Husband Abuse.

Sara Markowitz [*]

Alcohol consumption has been frequently linked to violence. This paper examines the direct relationship between the price of alcohol, which determines consumption, and violence toward husbands and wives. The data come from the 1985 cross section and the 1985-1987 panel of the National Family Violence Survey. A reduced form violence equation is estimated, and individual-level fixed effects are used to control for unobserved characteristics in the panel. Results indicate that an increase in the price of pure alcohol, as measured by a weighted average of the price of alcohol from beer, wine, and liquor, will reduce violence aimed at wives. The evidence on the propensity of an increase in the price of alcohol to lower violence toward husbands is mixed.

1. Introduction

The problem of violence between spouses has been a characteristic of families for many generations. Only since the 1960s has this problem gained national attention as a serious threat to the health and welfare of women, and only since the late 1970s has the problem of violence against men gained any notoriety. Estimates show that 30 out of every 1,000 females and 45 out of every 1,000 males are victims of severe violence committed by their spouses. [1]

Alcohol consumption has commonly been associated with incidents of spousal abuse. In reviews of the literature, Gelles and Cornell (1990) and Leonard (1993) note that virtually every study of aggression in families shows that alcohol consumption is a strong correlate of violence. In nationally representative samples, alcohol use is frequently found to accompany violence (Straus, Gelles, and Steinmetz 1980; Kantor and Straus 1987). Studies of violent families also find a similar association (Wolfgang 1958; Byles 1978). For example, in a sample of violent families, Gelles (1974) found that drinking accompanied violence in almost half of the cases. In samples of battered women, many different studies also find that the husbands had frequently been drinking prior to the assault (Walker 1979, 1984; Fagan, Stewart, and Hansen 1983). Walker (1984) interviewed 400 self-identified battered women on the drinking habits of male batterers and nonbatterers and found that 67% of male batterers were reported to frequently drink alcohol (as compared to never drinking or occasionally drinking) versus 43% of nonbatterers. Drinking by the victim is also observed in some studies of violence between spouses, although the percentage of victims reported to have been drinking prior to the attack is often much lower than the percentage of perpetrators who had been drinking (Collins 1981; Walker 1984; Bard and Zacker 1974).

In addition to the well-established link between alcohol consumption and violence, alcohol consumption has been shown to be negatively related to the price of alcoholic beverages (see for example Kenkel 1993; Leung and Phelps 1993; Manning, Blumberg, and Moulton 1995; Grossman, Chaloupka, and Sirtalan 1998). Given these two relationships, the purpose of this paper is to examine the direct relationship between the price of alcohol and the incidence of violence. In other words, the purpose of this paper is to answer the question of whether an increase in the price of alcohol will lower violence toward spouses through lower consumption.

2. Related Studies

There are numerous studies on domestic violence in the literature of sociology, psychology, and other disciplines. However, the topic of domestic violence is a relative newcomer to the economics literature. The most closely related studies are those on the effects of alcohol regulation on violence aimed at children by Markowitz and Grossman (1998a, 2000). These studies use the 1976 and 1985 National Family Violence Surveys and show that increasing the beer tax is an effective policy tool in reducing both the probability and frequency of violence toward children. [2] The former study looks at the 1976 cross section only, while the later study pools the two years of data and controls for state-fixed effects over time. Fixed effects are important in determining whether the effects of the state-level alcohol regulation variables in the cross sections are reflecting unobserved sentiment toward regulation and violence rather than true policy effects.

Other notable economic studies on domestic violence include those by Long, Witte, and Karr (1983), Tauchen, Witte, and Long (1991), and Farmer and Tiefenthaler (1997). These papers focus only on wife abuse and model violence as a good that can be bought or avoided with income transfers. One drawback of these studies is that their frameworks do not allow for the effects of alcohol consumption and the determinates of consumption on violence. Secondly, the data sets used by Tauchen, Witte, and Long and Farmer and Tiefenthaler are nonrandom samples of battered women. This paper adds to the current spouse abuse literature in that it addresses both of these limitations. The analysis presented here accounts for the effects of alcohol on violence and it presents the empirical estimation using a nationally representative sample. The use of the 1985-1987 panel of the National Family Violence Survey is also novel in that it allows unobserved individual characteristics to be controlled for.

3. Analytical Framework

The studies cited in section 1 on the link between alcohol consumption and violence do not posit a causal mechanism from alcohol to violence, rather, they highlight an association. Within this literature, there are a variety of hypotheses of why alcohol and violence are linked. One theory states that there may exist a psychopharmacological relationship in which alcohol can alter behavior by increasing excitability and/or boosting courage (see Pernanen 1981 and Fagan 1993 for a complete discussion). Under this theory, people may be more likely to commit a violent act when under the influence of alcohol than they would be otherwise. A second theory asserts that people use alcohol as an excuse for aberrant behavior. It is commonly believed that alcohol use may cause people to lose their inhibitions and/or release violent tendencies, and thus users cannot be fully blamed for their actions. In other words, drunkenness may give people an excuse for violence, despite whether or not actual pharmacological effects exist (Fagan 1990; Gelles and Cornell 1990). Finally, there is the "third factor" theory in which there exists some unknown common cause that results in both drinking and violent behaviors (Fagan 1990). These are only a few of a wide variety of possible explanations on the link bet ween alcohol consumption and violence; however, there is no agreement in the literature on the true cause of the association (see Reiss and Roth 1993 for more details).

Given the variety of explanations for the link between alcohol consumption and violence, there are many ways in which one can model a perpetrator's choice level of violence. Two simple models are presented here. A more formal derivation of these models is outlined in Markowitz and Grossman (1998b). The first encompasses the idea that the pharmacological properties of alcohol consumption cause unplanned violence, while the second one models alcohol as a facilitator in committing planned violence. In this second version, it is assumed that alcohol lowers the costs of violence by creating an excuse for the behavior. Even though the mechanism through which alcohol promotes violence is different, both frameworks presented below result in the same estimating equation and predictions. The "third factor" theory mentioned above can be accounted for by including a variety of individual characteristics and by using a fixed-effects estimation that controls for unobserved heterogeneity across individuals.

The first framework considers violence as a by-product of alcohol consumption. A perpetrator maximizes his or her utility, which is a function of alcohol consumption (A), violence (V), and other consumption goods (X), all of which are affected by tastes (t):

U = U[A, V(A, t), X, t]. (1)

Violence is present in the utility function but is not a choice variable per se. Rather, violence is an expected or unexpected consequence of alcohol consumption and may cause negative untility. [3] In effect, violence is produced by the chemical effects of alcohol consumption and by other factors that account for a person's propensity for violence. [4] These factors can be socioeconomic or demographic factors or a person's history of family violence and may be the same ones that affect the taste for alcohol consumption.

A perpetrator will maximize utility subject to a budget constraint:

I = PA + X, (2)

where total income (I) is equal to the price of alcohol (P) times the amount of consumption of alcohol plus the total amount spent on other consumption. The price of other consumption is normalized to $1.00.

Maximization will yield the following first-order conditions:

[partial]U/[partial]A + [partial]U/[partial]V x [partial]V/[partial]A = [lambda]P, (3a)

[partial]U/[partial]X = [lambda], (3b)

where [lambda] is the marginal utility of income. The first-order conditions imply a demand function for alcohol that shows that consumption is a function of the price of alcohol, income, and tastes:

A = A(I, P. t). (4)

In choosing the level of alcohol to consume, a person will account for the effect of violence on utility if violence as a by-product is anticipated. If no violence is anticipated then the level of alcohol will be chosen without regard to the consequences of violence on utility.

Because violence is a function of alcohol consumption, substituting the demand equation for alcohol into the violence production function gives violence as a function of the price of alcohol, income, and tastes:

V = V(I, P, t). (5)

By the law of downward sloping demand, increasing the price of alcohol will decrease the quantity of alcohol demanded. When the quantity of alcohol decreases, violence will decrease, thus leading to a negative relationship between the price of alcohol and violence. In addition, if alcohol is a normal good then an increase in income would also increase violence. [5]

A second framework is established to account for the costs of violence that can be lowered by alcohol consumption. In this model, a perpetrator maximizes utility, which is a function of alcohol consumption, violence, and other consumption, all of which are affected by tastes:

U = U(A, V, X, t). (6)

Here, all three components of utility are choice variables. The budget constraint faced by the perpetrator equates total income to the price of alcohol times the quantity of alcohol consumed, plus the price of violence ([pi]C) times the amount of violence, plus the total amount spent on other consumption:

I = PA + [pi]CV + X. (7)

The price of violence is the expected value of any monetary costs of violence. Monetary costs may include legal fees, fines, or lost wages from prison sentences or even divorce settlements resulting from a dissolution of the relationship. These costs would only be incurred if the perpetrator is caught and/or punished. [6] If the monetary costs are denoted by C, and the probability of facing the costs is denoted by [pi], the price of violence is [pi]C. By assumption, an increase in alcohol consumption will lower the price of violence because it lowers [pi]. In the case of domestic violence, it is reasonable to assume that alcohol consumption will lower the probability of facing costs of violence because victims or witnesses often excuse aberrant behavior as the result of alcohol consumption and thus choose not to report their loved ones to authorities. By contrast, in other situations alcohol consumption may raise the probability of facing costs of violence. For example, a violent criminal who is drunk may be less careful than the calculating violent criminal who tries to protect himself or herself from being caught. [7,8]

Maximizing utility subject to the budget constraint yields the following first-order conditions:

[partial]U/[partial]A = [lambda][P + [partial][pi]/[partial]A X CV], (8a)

[partial]U/[partial]V = [lambda][pi]C], (8b)

[partial]U/[partial]X = [lambda]. (8c)

These first-order conditions imply a demand function for violence that depends on the monetary cost of violence, the price of alcohol, income, and tastes:

V = V(C,P,I, t). (9)

With the exception of the monetary costs of violence, this reduced form is the same as the one derived in the first model and yields the same prediction that an increase in the price of alcohol will decrease violence. In this framework, however, the mechanism works through a decrease in the price of alcohol that will increase consumption. The increase in consumption will then lower the price of violence (through a lower probability of facing penalties) and increase the amount of violence. [9] Empirically, Equation 9 will be indistinguishable from Equation 5 when the monetary value of the costs of violence for each individual is unobserved.

Equation 9 also shows that income is a determinant of violence. In the first framework, the effect of income on violence works indirectly through the purchase of alcohol. Here, however, income can affect violence both directly and indirectly. There will be a positive direct effect of income on violence if income is used by the perpetrator to purchase more violence. The opposite effect might hold if income gained by the victim is used to avoid violence. These are also the predictions that arise from the models of domestic violence as presented by Long, Witte, and Karr (1983), Tauchen, Witte, and Long (1991), and Farmer and Tiefenthaler (1997). The indirect effect occurs when an increase in income also increases the quantity of alcohol, which in turn lowers the shadow price of violence causing an increase in violence.

4. Data

The data come from the 1985 National Family Violence Survey (NFVS) and the 1986 and 1987 followups to the 1985 survey. The 1985 NFVS was designed to collect information about violence in the home and has detailed information on how conflicts are resolved. The 1985 data are a nationally representative sample of 4,990 individuals who are either married or cohabiting, are single parents living with children under 18, or are individuals who had been married or cohabiting within the past two years. Only married or currently cohabiting couples are included in the analysis presented here. This reduces the potential sample size to 4,372 individuals. The 1986 and 1987 followups were designed specifically to track information on violence between spouses. Individuals were picked to be in the followup surveys based on their answers to the intercouple violence questions in the 1985 survey. Specifically, all married or cohabiting individuals who reported any violence between the partners in the past year were chosen to be reinterviewed. A sample of the individuals reporting intercouple violence not within the past year but sometime before then were also included in the followups. Finally, a sample of individuals who reported no violence were included. The panel used in this study is comprised of 1,541 married or cohabiting individuals, all of whom appear in the initial 1985 survey, have stayed with the same partner, and were included in at least one of the followups. [10] Of the total, 506 individuals appear in only the 1985 and 1986 surveys, 408 are in the 1985 and 1987 surveys only, and 627 are in all three years. Thus, the panel data set contains 3,709 potential observations." [11]

Dependent Variables

Measures of domestic violence in the NFVS are collected by use of the Conflict Tactic Scale (CTS) developed by Straus (1979). The CTS gathers information on the number of times in the past year a respondent has committed or has been the victim of a violent act. Two dependent variables are constructed: an indicator for whether the husband was violent toward the wife (termed wife abuse) and an indicator for whether the wife was violent toward the husband (termed husband abuse). Both variables are dichotomous indicators that equal 1 if the respondent committed any of the following acts toward their spouse in the past year: kicked, bit, or hit with fist; hit or tried to hit with something; beat up the other; choked him or her; threatened with a gun or knife; or had used a gun or knife. Three percent of the 1985 sample reported violence toward women and 4.5% of the 1985 sample reported violence toward men in the past year. These percentages are 3.7% and 3.5% for men and women, respectively, in the panel. [12] Tabl e 1 shows the means and standard deviations for these variables in the various years of the surveys. The statistics presented for the panel are weighted to account for the oversampling of violent individuals.

Note that the survey is only given to one member of each household. The respondent can be either male or female, and the respondent is asked about both violence committed by him or her and violence aimed at him or her. Therefore, the indicator for wife abuse, for example, can be shown with responses from both husbands and wives (perpetrators and victims) or can be shown as the response of only the wives (the potential victim). Eliminating responses from the perpetrator reduces the sample size by about half. Specifically, 56% of the respondents are females.

In both the 1985 cross section and the panel, women report being victims at greater rates than men report victimizing the women. The percentage of violence toward women as reported by women is about 0.05 in both the cross section and the panel but is 0.01 in the cross section and 0.04 in the panel for violence aimed at women as reported by men. In addition, women in the panel report victimizing men more than men report being victims, but there is no statistically significant difference in male victimization rates as reported by men and women in the 1985 cross section.

Reliability of the Data

One criticism of the NFVS focuses on the reliability of the respondents' answers to the occurrence of violence. The survey seeks to gain information about sensitive and possibly deviant types of behavior that often arouses antagonism, high refusal rates, and distorted answers from the respondents, thereby bringing into question the reliability of the results. The principal investigators of the survey discuss this criticism at length (see Straus and Gelles 1990 for complete discussion of this issue). First, they claim that the antagonistic aspects are minimized by presenting the questions in the context of resolving family conflicts. The possible responses to the question on conflicts between spouses begin with resolution tactics such as "discuss the issue calmly," which are generally viewed as positive methods of dealing with problems. The scale gradually increases to questions about more socially unacceptable behavior. Through this method of getting to the violence questions, the respondent has first been g iven a chance to give the "socially correct" answers and is less apprehensive about discussing incidence of violence.

Currently, the CTS seems to be the best available technique for collecting truthful information on domestic violence and has been used in over 200 studies to date (see Straus and Gelles 1990). Nevertheless, because of the potential for underreporting violence, the dependent variables are considered to be conservative estimates of violence. This poses no problem for the conclusions because so long as the measurement error in the dependent variable is random, measurement error only serves to raise the standard errors, leaving the coefficients as unbiased estimators. However, if for example, drinkers systematically underreport violence, then the coefficient on the price of alcohol will be biased toward zero.

A related criticism is how well the NFVS reflects the reported national incidence of violence aimed at spouses. Most of the data used in studies of domestic violence rely on small samples of battered women, or studies where the participants identify themselves as being violent. However, a few nationally representative samples do present estimates of rates of violence. Using the CTS, the 1976 NFVS reports violence rates similar to those in the 1985 survey--about 4% of women are victims of severe violence as are about 5% of men. The Violence Against Women Survey (VAWS) is a random sample of 12,300 women across Canada. Information on violence is gathered using a modified version of the CTS. The high severity index in the VAWS is comparable to the abuse indicator utilized in this paper because it includes the acts of kicking, biting, beating, choking, threatening to use or using a knife or gun, and sexual assault. The percentage of married women in the VAWS who were abused is 0.035, which is very similar to the wife abuse figures cited above. Violent acts toward men are not reported in the VAWS. Finally, the British 1994 CTS Domestic Violence Survey uses a modified version of the CTS in a sample of 1978 men and women in Great Britain. In a current relationship, about 3% of women and 6% of men reported being punched or kicked, had an object thrown at them or were hit with an object, or were stuck with a sharp object.

Other nationally representative samples in the U.S. show much lower rates of violence. The National Crime Victimization Survey (NCVS) conducted by the Bureau of Justice Statistics reports information on violence by intimates (spouses, ex-spouses, or boy/girlfriends). The rates from the NCVS are much lower than the domestic violence studies because the victim must perceive the violence as a crime in order for the violent act to be counted in the survey. Violence in the NCVS is defined as aggravated or simple assault. The average rate of violence by intimates from 1992 to 1993 is 0.76% for women and 0.13% for men. Similarly, in a 1995-1996 survey of 8000 men and 8000 women, the National Violence against Women (NVAW) survey found that only 1.3% of women and 0.9% of men were victims of physical assault. As with the NCVS, the victims in this survey do not have to be married or living with someone, and the perpetrator can be a current or former spouse, a date, or a boyfriend or girlfriend.

Independent Variables

Alcohol Control Variables

The price of alcohol is a composite price of 1 ounce of pure alcohol. This price is a weighted average of the price of pure alcohol from beer, liquor, and wine. The weights are the percentage of total pure alcohol consumption accounted for by each of the three types of beverages in 1985. Each person is assigned a price based on the state in which they live.

Prices for beer, wine, and liquor come from the Inter-City Cost of Living Index, published quarterly by the American Chamber of Commerce Researchers Association (ACCRA) for anywhere from 220 to 260 cities. Each quarterly city-level alcoholic beverage price is deflated by the CPI for the U.S. as a whole (1982-1984 = 1) and by the ACCRA city-specific cost of living index. A state-level quarterly price is computed as a population weighted average of the price from each city within a state. The 1985 NFVS was conducted in June 1985 so the annual price is taken as a simple average of the prices that existed in the first two quarters of 1985 and the last two quarters of 1984. The followups were both conducted during the third quarter of the survey year so the annual prices for 1986 and 1987 are taken as simple averages of the prices that existed in the first three quarters of the survey year and the fourth quarter of the preceding year. Average annual prices are converted into pure alcohol prices based on the perce ntage of alcohol in each beverage. The liquor price is the price of Seagram's 7 Crown Whiskey, which is 40% alcohol. The wine price is given for a bottle of Paul Masson Chablis, which is 10.5% alcohol, and the beer price is for a six-pack of Budweiser or Schlitz, which is 4.5% alcohol. The average pure prices for each beverage are then weighted by consumption to form the composite price.

Most of the variation in alcoholic beverage prices across states arise from variations in taxes. The beer tax and the composite price are positively correlated (r = 0.49). Models were also tested that include only the prices of beer, wine, or liquor separately, and again, results are very similar to those presented here.

There is substantial variation in the composite price of alcohol across states. Table 1 shows the means and standard deviations of the alcohol price and the other alcohol control variables. The average composite alcohol price in 1985 across the 49 states included in the sample plus the District of Columbia is $0.76 with a standard deviation of $0.07. [13] In 1985, California has the lowest price of $0.60, and Alabama has the highest price of $0.91. In general, prices are highest in the southern states and lowest in the Midwest. Even though there is a downward trend in the real price from 1985 to 1987, the variation in the real price of alcohol over time is small. Twenty of the 50 states in the sample had price changes of over $0.03 between 1985 and 1987, with an average decrease of about $0.08 for those 20 states.

Retail availability of alcohol factors into the full price of alcohol faced by individuals. To capture the availability effects, two measures are employed in some models. First, the percentage of each state's population living in counties dry for beer in each of the survey years is included. These data come from the Beer Institute's Brewers' Almanac (1996). With larger percentages of populations living in dry counties, travel time to obtain alcohol increases, adding to the full price of alcohol. In addition, this measure serves to capture some of the unobserved state sentiment toward drinking that may be reflected in the drinking habits of the state's residents. Secondly, the number of retail outlets per 1000 population that are licensed to sell alcoholic beverages for on-premise or off-premise consumption is included. These data come from Jobson's Liquor Handbook (various years).

Individual Characteristics

Literature on domestic violence provides insight into the personal characteristics that lead to a predisposition toward violence (see Gelles and Cornell 1990 for profiles of domestic abusers and their victims). People who were abused by their parents or saw their parents fight a lot, for example, are more likely to be violent. In order to proxy for these two factors, dichotomous indicators are included to represent whether or not the respondent's parents used physical punishment on the respondent and if the respondent's parents hit or threw things at each other during the respondent's teenage years.

Socioeconomic and demographic characteristics may also play a role in determining an individual's propensity toward violence. Three indicators of the respondent's race are included: black, not Hispanic; Hispanic; and other race. The omitted category is white, not Hispanic. Also included in the models are household income, an indicator for whether the wife was pregnant or had a child in the last 12 months, the number of children under age 17 in the home in 1985, the respondent's age, gender, and a measure of the respondent's level of stress. The question measuring stress asks how often the respondent felt nervous or stressed in the past year, with answers ranging from never to very often. Educational attainment and employment status are included for both the husband and the wife. Dummies indicating the religion of the wife are included. The husband's religion is excluded because it is highly correlated with the wife's religion. Finally, age squared, income squared, and education squared were added to allow fo r nonlinear effects of these variables.

In some models, the respondent's gender is included to act as an indicator of whether the respondent was the victim or the perpetrator. For example, when wife abuse is the dependent variable and gender is equal to one, the respondent is female, thus indicating that the victim is answering the question. When gender is equal to zero, the respondent is male, indicating that the perpetrator is answering the question.

Any observations with missing values are dropped from the regressions. If an individual in the panel has missing values for one year, but not other years, only the observations from the incomplete years are excluded. Deleting the incomplete observations results in the loss of about 15% of the cross section and about 8% of the panel. Observations with missing values on income in the panel are not necessarily excluded if some information on income is available. That is, individuals in the panel with missing observations on income were assigned a value if income was reported in at least one of the years in the panel (132 observations). The procedure for filling in these missing values differs slightly for two cases: The first case is if a respondent is represented in all three years, and income is available for two of the three years. The missing income is assigned based on the rate of growth between the years with values. The second case is if a respondent is in only two years of the panel and has one missing year for income, or if the respondent is in all three years, but has two missing values for income. An average rate of growth for income over the known observations in the whole sample was applied to the individual's missing values. Any individuals with all years of missing income were deleted from the sample.

5. Estimation

Tables 2-5 show estimates of the reduced form violence equation. In each table, all models show the effects of the price of alcohol and the household and individual characteristics on the probability of abuse. An alternative model is shown in each table which adds to the first model two other regulatory variables, the number of outlets licensed to sell alcohol and the percentage of the state's population living in counties that are dry for beer. All equations are estimated by linear probability models. [14] Even though there is oversampling of violent respondents in the panel, weights are not used because the distribution of violence in the panel is very similar to the distribution of violence in the cross section. The unweighted proportion of wife abuse victims is 4.4% in the panel and 3% in the cross section. The same numbers for males are 4.7% and 4.5%, respectively. In addition, DuMouchel and Duncan (1983) show that weighted regressions are not appropriate if averages of strataspecific regression coeffic ients are desired. Nevertheless, weighted regressions were tested, but the results were very similar to the Unweighted regressions.

Tables 2 and 3 show the effects of the composite price of alcohol on the probability of wife abuse and husband abuse, respectively, in the full 1985 cross section. The first two columns include both male and female respondents while the second two columns include only the victim as the respondent. Tables 4 and 5 show the results of the panel of respondents. The first two columns of each of these tables show the results of the price of alcohol on the probability of abuse using the full set of independent variables. The third and fourth columns use a limited set of independent variables--only those that vary across time. These variables are the same set of variables used in fixed effects models in columns 5 and 6. A larger sample is used in the limited specification of columns 3 and 4 than in the full specification (columns 1 and 2) because of missing values on some of the time-invariant variables. Regressions were run on the limited specification using the smaller sample from columns 1 and 2, but the results were no different.

Columns 5 and 6 of Tables 3 and 4 utilize the panel nature of the data to include individual-level fixed effects. These models are the preferred specifications because fixed effects serve as a control for any unobserved individual characteristics that may predict violence and that may be correlated with some righthand-side variables. For example, individual fixed effects may help separate out the effects of income, labor force, and other time-varying variables on violence from effects due to personality. The fixed effects models are estimated by transforming variables into deviations from person-specific means. All time-invariant variables are dropped from the regressions. This technique is equivalent to adding dummy variables for every individual.

As well as being correlated with some of the individual's righthand-side characteristics, the unobserved personality traits may be correlated with the state-specific price or regulatory variables if these variables reflect state sentiment toward alcohol consumption and the individual shares the same sentiment toward alcohol as the rest of the state. When this sentiment also determines the individual's propensity toward violence, such a correlation might imply that the alcohol regulatory variables are endogenous. A common method of accounting for an unobserved, time-invariant, state-level sentiment toward alcohol is to include state-level fixed effects. Models were tested that include state-level fixed effects for both males and females, and the magnitude of the coefficients on the price of alcohol are very similar to those resulting from individual-level fixed effects, although the standard errors are larger. The advantage to the individual-level fixed effects is that the individual-level effects guarantees that the individual's sentiment toward drinking is captured rather than the state-level sentiment that may or may not be shared by the individual. [15] In addition, if this is the appropriate model, state fixed effects drop out when individual fixed effects are included if people do not move.

There are a few potential problems with the two models presented in each table. First, the specification that includes only the price on alcohol is prone to omitted variable bias if the availability measures are predictors of violence. However, including all the relevant control variables may lead to the problems of multicollinearity. This problem may arise because states that have high anti-drinking sentiment may impose both higher taxes on alcohol and more restrictions on availability.

When individual-level fixed effects are excluded, another potential problem is that many of the individual characteristics may be correlated with the error term in the violence equations. That is, there may be some unmeasured factor that affects the outcomes of both the propensity to commit violence and the individual characteristics. The characteristics most likely to be endogenous are the measure of stress, pregnancy, the number of children at home, education, income, occupation, employment status, and religion. The coefficients on these potentially endogenous variables are likely to be biased if not instrumented for. However, including these variables will not bias the coefficients on the state-level regulatory variables (the variables of interest in this paper) so long as the individual and state-level variables are not correlated. Models were tested that exclude the above-mentioned variables from the 1985 cross section and the panel. The results are not shown, and the coefficients on the price and avail ability measures are unaffected by the exclusion of the potentially problematic individual characteristics.

A final problem with the data is that the price of alcohol is measured with error for several reasons. First, the price data pertain to the state the respondent lives in rather than a city or county price. Random measurement error in an independent variable biases its coefficient and t-ratio toward zero. Thus, the price coefficients and associated t-ratios are conservative lower-bound estimates. Secondly, there may be some measurement error built into the alcohol regulatory variables in the 1986 and 1987 waves of the survey because the survey data do not contain information on whether or not the respondent moved during followups. The problem can be especially serious in the fixed effects models because the downward bias in regression coefficients and t-ratios due to measurement error in regressors are exacerbated (Griliches and Hausman 1986). In matching the prices to the individuals, it is simply assumed that the individual has not moved. This may not be a bad assumption because the survey methodology was t o attempt to contact respondents through the phone number as given in the 1985 survey. Unless a new phone number was provided, it is likely that an individual who moved was excluded from the survey. In addition, a move within the same state would not change the alcohol regulatory variables assigned to the individual.

6. Results

Table 2 shows the effectiveness of an increase in the price of alcohol in reducing the probability of wife abuse in the 1985 cross section. The coefficient on the price of alcohol is negative and significant in all models. [16] This result holds for the regressions that include all respondents regardless of gender as well as the regressions that include only female respondents. The magnitude of the coefficient on the price is similar in all models in Table 2 and indicates that a 1% increase in the price of alcohol will reduce the probability of wife abuse by a range of 3.1-3.5%. Elasticities are calculated by multiplying the price coefficients by the ratio of the average price to the proportion of spouses reported to have been victimized in the relevant sample. Note that these coefficients are measured with some imprecision, with a 95% confidence interval showing that the elasticities fall anywhere in the range of -1.0% to -5.5%.

The availability measures are not statistically significant predictors of the probability of violence. In addition, the coefficients on the percent dry (in column 4) and the number of outlets licensed to sell alcohol (in columns 2 and 4) do not display the anticipated signs.

Some of the individual and family characteristics are significant predictors of violence. The interpretation of a few of the coefficients in the full sample (columns 1 and 2) is not straightforward in that the age, race, the measure of stress, and the indicators of a history of violence all refer to the respondent who can be either the husband or the wife. For age and race it is likely that the spouses are close in age and share the same race so the interpretation of these variables is not problematic. With this caveat in mind, the results indicate that older people are less likely to be abused or to abuse, while those who are more stressed, experienced a family history of violence, or are black are more likely be victims or be violent toward the wife. [17] With the exception of being black, similar results hold in the sample of female respondents (columns 3 and 4). Here, the interpretation of the coefficients on the variables is straightforward in that older women are less likely to be abused, whereas women with more stress in their lives or those who have a history of parental violence are more likely to be abused.

As shown in Table 3, violence by women directed at their husbands is not sensitive to changes in the price of alcohol. The coefficients are always negative but are not significant at conventional levels. The only other alcohol regulatory variable that may have an effect on violence is the number of outlets. The coefficients on this variable in columns 2 and 4 are negative and significant, yet the negative sign is contrary to what is expected a priori. A negative sign indicates that more licensed outlets would reduce violence. In the full sample, stress, the respondent's parents having fought, being black or Hispanic, and part-time employment by the wife (versus full time employment) all increase the probability of violence at at least the 10% significance level. Having more children living at home reduces the probability of violence toward the husband, while being older than the mean age (42 years) actually increases the probability. In the sample of male respondents only, being black or part-time employment by the wife increases violence, while age, having more children, or the wife being Protestant reduces violence.

Table 4 utilizes the panel to examine the effects of the price of alcohol on the probability of violence toward females. Column 1 includes the price of alcohol and the individual characteristics, column 2 adds the availability measures, and columns 3 and 4 limit the set of independent variables to those that vary over time. Because the results from the 1985 cross section indicate that there is little difference in the results of the alcohol control variables whether or not the sample is limited to respondents who are victims, only models that include the full sample are shown. The coefficients in the first four columns of Table 4 all show that the price of alcohol is negatively related to the probability of wife abuse. However, this result is only statistically significant when the availability measures are excluded. These models may suffer from multicollinearity in that the other state-level variables may be correlated with the price. The price elasticities are - 2.14, - 1.80, - 2.35, and - 1.73 for columns 1-4, respectively.

The results of the other alcohol control variables and the individual characteristics in columns 1 and 2 of Table 4 are very similar to those in the 1985 cross section. The coefficients on percent dry and number of outlets are statistically insignificant but this time display the anticipated signs. As for the individual characteristics, age, stress, parental history of violence, black, Hispanic, and number of children at home all affect the probability of wife abuse.

Columns 5 and 6 of Table 4 display the results of the fixed effects estimation. Again, the coefficients on the price of alcohol are negative and statistically significant, and the price elasticity is -5.41 for the model in column 5 and -5.26 for the model in column 6. A 95% confidence interval around these elasticities lies between -1.0 and -9.7. These results indicate that even after controlling for the unmeasured individual traits, increases in the price of alcohol serve to reduce the probability of violence toward wives. Although not unexpected, a second finding of the fixed effect estimation is that almost none of the individual characteristics are significant predictors of violence. The one exception is that husbands who are not in the labor force are more likely to abuse their wives. The majority of husbands who are not working in this sample are retirees, but this classification can also include students, homemakers, and disabled individuals.

The results for violence toward males in the panel as shown in the first four columns of Table 5 are similar to those of the 1985 cross section. The results show that the price of alcohol has no statistically significant relationship with the probability of husband abuse. In fact, the coefficients on the price are only negative in one of the four specifications (see column 3). However, unlike the 1985 results, increases in the percentage of a state's population living in dry counties may decrease violence toward men. This result is significant at the 10% level in column 1 and at the 5% level in the limited specification in column 4.

As for the individual characteristics that determine whether a husband is abused or not, family history of violence increases the probability of abuse, as does stress, black, Hispanic, and the wife identifying herself with no religion. Age reduces the probability of abuse. Income and income squared are jointly significant, and calculating the marginal effect at the mean level of income shows that the probability of abuse increases with family income. This last result is interesting because as shown in Table 4, income is not a statistically significant predictor of wife abuse. This result, however, is consistent with the prediction that arises from the theoretical models showing that higher income will indirectly result in more violence if alcohol is a normal good. It is also consistent with the prediction from the second theoretical model which shows that more violence can be purchased directly with income, although as Farmer and Tiefenthaler (1997) demonstrate, it matters who holds the income (the perpetrat or or the victim). The owner of the income is unknown in these data because only household income is reported. The effects of income on violence may be a result of the combination of the direct and indirect effects, although it is not possible to empirically distinguish between the two.

Finally, columns 5 and 6 of Table 5 show the results of the fixed effect estimation of the reduced form violence equation. Surprisingly, the price of alcohol is both a negative and statistically significant predictor of violence at about the 10% level. The availability measures in column 6 display the anticipated signs but are not statistically significant. None of the individual characteristics are predictors of violence in the fixed effects models, with the exceptions of income and income squared that are jointly significant, and age and age squared. Higher income is associated with an increased probability of violence against men, as is age, when evaluated at the mean.

7. Discussion

A consistent result that emerges from this paper is that increases in the price of 1 ounce of pure alcohol, as measured by a weighted average of the prices of alcohol from beer, wine, and liquor, will serve to reduce the probability of severe violence aimed at wives. Severe violence includes the acts of kicking, biting, or hitting with a fist; hitting or trying to hit with something; beating up the other; choking him or her; or threatening to use or using a gun or knife. Using an average of the two estimates from the fixed effects specification, a 1% increase in the price of an ounce of pure alcohol would decrease the probability of being a victim of wife abuse by 5.34%. I caution that the magnitude is estimated with some imprecision, for a 95% confidence interval around this elasticity lies in the range of - 1.0% to -9.7%. Nevertheless, in 1985, there were 54.4 million married women in the United States. If 3.6% were abused a 1% increase in the price of pure alcohol will, on average, decrease the number of abused married women by about 104,600. Using the results from the 1985 cross section, a 1% increase in the price of pure alcohol will decrease the number of abused women by about 53,500.

One caution to note is that while increasing the price of alcohol (through tax increases) would lower violence aimed at women, any policy decisions must weight the cost of raising the price versus the benefits of the reduction in violence. Raising the price would serve to penalize people who consume alcohol but who are not violent. According to the National Household Survey on Drug Abuse, the percentage of adults age 26 or over who were current users of alcohol in 1985 was 60.7 (about 89.5 million people).

By contrast to violence against women, the, evidence on the propensity of increases in the price of alcohol to lower violence toward husbands is mixed. When individual-level characteristics are not controlled for, the coefficients on price of alcohol are not statistically different from zero and hence cannot be considered as predictors of violence toward men. However, once the individual traits are controlled for, a negative and statistically significant relationship emerges. It would be premature to draw firm conclusions on the propensity of increases in the price of alcohol to reduce violence aimed at men solely from the fixed effects models. One possible explanation for the mixed results is that the unobserved individual traits are correlated with the females' (the perpetrators') attitudes toward drinking, which in turn are reflected by the statelevel price of alcohol. Once this correlation is accounted for in the fixed effects models, the negative relationship between the price of alcohol and violence em erges. It is clear that further research is needed on this issue.

Secondary findings from this paper show no relationship between the availability of alcohol and the probability of violence toward women. There may be some reduction in violence aimed at men from increases in the percentage of dry counties in a state; however, this finding does not hold across all specifications. The individual character traits that generally serve to increase violence at either spouse are stress, family history of violence, and being black, while age decreases the probability of violence.

(*.) Department of Economics, Rutgers University, University Heights, Newark, NJ 07102, USA; E-mail [e-mail]. Present address: National Bureau of Economic Research, 365 Fifth Avenue, 5th Floor, New York, NY 10016.

Research for this paper was supported by grant number 1 R01 AAl0817 from the National Institute on Alcohol Abuse and Alcoholism (NIAAA) to the NBER. I thank Kirk Williams at the Center for the Study and Prevention of Violence at the University of Colorado for providing assistance with the data and Michael Grossman, Sara McLanahan, and two anonymous referees for helpful suggestions and comments. This paper has not undergone review accorded official NBER publications; in particular, it has not been submitted for approval by the Board of Directors. Any opinions expressed are those of the author and not that of NIAAA or the NEER.

Received February 1999; accepted January 2000.

(1.) Estimates are from the 1985 National Family Violence Survey.

(2.) The beer tax is used in these studies because prices of beer are not available in 1976.

(3.) No a priori sign is assigned to the marginal utility of violence. Positive utility may arise if violence provides relief from stress or control or power over the victim.

(4.) In the violence equation, what are described as taste parameters are technically efficiency parameters. However, in the reduced form the two are indistinguishable from each other.

(5.) This latter prediction does not allow for the possible direct effect of income on violence. The second framework presented below does allow for this direct effect.

(6.) Nonmonetary costs may also be incurred and might include the emotional costs of the dissolution of the relationship or the loss of respect by the victim, family members, or friends. The nonmonetary costs are assumed to be factors that would affect the choice level of violence and for simplicity are included in the taste term.

(7.) A positive coefficient an the price in the reduced form would provide evidence that alcohol consumption raises the probability of facing penalties.

(8.) An alternative way to model the probability of facing costs arises if the cost of violence does not increase with the level of violence and [pi] depends on both A and V. That is, [pi] can be lowered with increased alcohol consumption but raised with increased violence. With reasonable functional forms (such as the logit function) and a small value of [pi], the cross effects of violence and alcohol on [pi] will be negative. Then, increasing the price of alcohol will decrease consumption, thereby increasing the marginal effect of violence on [pi] and raising the shadow price of violence, which in turn will lower violence. In this case, as in the model outlined below, the price of alcohol and violence are negatively related.

(9.) Theoretically, there is also a cross-price effect associated with an increase in the price of alcohol. This causes violence to rise if alcohol and violence are substitutes in consumption and causes violence to fall if the two items are complements in consumption. I make the plausible assumption that if the cross-price effect is positive, it is outweighed by the effect generated by the negative relationship between alcohol consumption and the probability of facing the costs of violence.

(10.) Couples who were together during the 1985 survey are not required to stay together to be included in the followups. However, the 80 respondents whose initial relationship ended are excluded from the analysis presented here for two reasons. First, the violence questions refers to violence in the last 12 months the couple was together. Thus, a respondent without a new partner may give the same answer in two different survey years. Secondly, no information is given for when any new relationships may have begun making it difficult to merge the appropriate price with the individual data.

(11.) This total includes observations for which there are missing values. The treatment of missing values is discussed below.

(12.) Comparisons of the rates of violence by men and women may be inappropriate because the rates do not contain any information about the number of times violence occurred in the past year or the potential for harm.

(13.) Hawaii is not included because alcoholic beverage prices for Hawaii are not available from ACCRA.

(14.) Logit models were tested, but results are very similar. Linear probability models corrected for heteroskedasticity were also tested, but the standard errors remain unchanged.

(15.) When unmeasured tastes are not time invariant, both state- and individual-level fixed effects are inadequate in capturing sentiment. Strumpf and Oberholzer-Gee (1999) examine the bias resulting from ignoring the potential endogeneity of liquor prices in a demand equation. The authors find that the coefficient on price in a model that ignores a measure of sentiment--the mean taste for legalizing liquor in a state--is biased upward; however, the magnitude of this bias is small. The elasticity that accounts for sentiment is -0.57 versus -0.7 when sentiment is not included. Given this, if a person's tastes toward violence and drinking change during the sample period, the estimates presented here may be slightly biased upward.

(16.) Unless otherwise mentioned, statistical significance refers to a two-tailed test at the 5% level..

(17.) Unless otherwise indicated, in discussing the impact of age on violence, statistical significance refers to a joint test of age and age squared, and the direction of the impact is from the marginal effect calculated at the mean age.

References

American Chamber of Commerce Research Association. Various years. Inter-city cost of living index. Louisville, KY: ACCRA.

Bard, Morton, and Joseph Zacker. 1974. Assaultiveness and alcohol use in family disputes. Criminology 12:283-92.

Beer Institute. 1996. Brewers' almanac. New York: United States Brewers Foundation.

Byles, John A. 1978. Violence, alcohol problems and other problems in disintegrating families. Journal of Studies on Alcohol 39:551-3.

Collins, James J., Jr. editor. 1981. Drinking and crime: Perspectives on the relationships between alcohol consumption and criminal behavior. New York: The Guilford Press.

DuMouchel, William H., and Greg J. Duncan. 1983. Using sample survey weights in multiple regression analyses of stratified samples. Journal of the American Statistical Association 78:535-43.

Fagan, Jeffrey. 1990. Intoxication and aggression. In Drugs and crime: Crime and justice, a review of research, volume 13, edited by Michael Tonry and James Q. Wilson. Chicago: The University of Chicago Press.

Fagan, Jeffrey. 1993. Interactions among drugs, alcohol and violence. Health Affairs 12:65-79.

Fagan, Jeffrey A., Douglas K. Stewart, and Karen V. Hansen. 1983. Violent men or violent husbands? Background factors and situational correlates. In The dark side of families, edited by David Finkelhor, Richard Gelles, Gerald Hotaling, and Murray Straus. Beverly Hills, CA: Sage Publications.

Farmer, Amy, and Jill Tiefenthaler. 1997. An economic analysis of domestic violence. Review of Social Economy 55: 337-58.

Gelles, Richard J. 1974. The violent home: A study of physical aggression between husbands and wives. Beverly Hills, CA: Sage Publications, Inc.

Gelles, Richard J., and Claire P. Cornell. 1990. Intimate violence in families. Newbury Park, CA: Sage Publications, Inc.

Griliches, Zvi, and Jerry A. Hausman. 1986. Errors in variables in panel data. Journal of Econometrics 31:93-118.

Grossman, Michael, Frank J. Chaloupka, and Ismail Sirtalan. 1998. An empirical analysis of alcohol addiction: Results from the monitoring the future panels. Economic Inquiry 36:39-48.

Jobson's Liquor Handbook. Various years. New York: Jobson Publishing Corporation.

Kantor, Glenda K., and Murray A. Straus. 1987. The 'drunken bum' theory of wife beating. Social Problems 34:213-321.

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

Leonard, Kenneth E. 1993. Drinking patterns and intoxication in marital violence: Review, critique, and future directions for research. In Alcohol and interpersonal violence: Fostering multidisciplinary perspectives, edited by Susan E. Martin. National Institute on Alcohol Abuse and Alcoholism Research Monograph 24. NIH Publication No. 93-3469, pp. 193-211.

Leung, Siu Fai, and Charles E. Phelps. 1993. 'My kingdom for a drink...?' A review of the price sensitivity of demand for alcoholic beverages. In Economic and socioeconomic issues in the prevention of alcohol-related problems, edited by Gregory Bloss and Michael Hilton. Washington, DC: U.S. Government Printing Office.

Long, Sharon K., Ann D. Witte, and Patrice Karr. 1983. Family violence: A microeconomic approach. Social Science Research 12:363-92.

Manning, Willard G., Linda Blumberg, and Lawrence H. Moulton. 1995. The demand for alcohol: The differential response to price. Journal of Health Economics 14:123-48.

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. The effect of alcohol regulation on physical child abuse. NBER Working Paper No. 6629.

Markowitz, Sara, and Michael Grossman. 2000. The effect of alcohol regulation on physical child abuse. Journal of Health Economics 19:271-82.

Pernanen, Kai. 1981. Theoretical aspects of the relationship between alcohol use and crime. In Drinking and crime: Perspectives on the relationships between alcohol consumption and criminal behavior, edited by James J. Collins, Jr. New York: The Guilford Press.

Reiss, Albert J., and Jeffery A. Roth, editors. 1993. Understanding and preventing violence. Washington, DC: National Academy Press.

Straus, Murray A. 1979. Measuring intrafamily conflict and violence: The conflict tactic scales. Journal of Marriage and the Family 41:75-88.

Straus, Murray A., and Richard J. Gelles. 1990, Physical violence in American families: Risk factors and adaptations to violence in 8,145 families. New Brunswick, NJ: Transaction Publishers.

Straus, Murray A., Richard J. Gelles, and Suzanne K. Steinmetz. 1980. Behind closed doors: Violence in the American family. Garden City, NY: Anchor Press, Doubleday.

Strumpf, Koleman S., and Felix Oberholzer-Gee. 1999. Endogenous policy decentralization: Testing the central tenet of economic federalism. Unpublished working paper, University of North Carolina, Chapel Hill.

Tauchen, Helen V., Ann D. Witte, and Sharon K. Long. 1991. Domestic violence: A nonrandom affair. International Economic Review 32:491-511.

Walker, Lenore E. 1984. The battered woman syndrome. New York, NY: Springer Publishing Company.

Walker, Lenore E. 1979. The battered woman. New York, NY: Harper & Row Publishers.

Wolfgang, Marvin E. 1958. Patterns in criminal homicide. New York, NY: John Wiley & Sons.