Does drinking affect grades more for women? Gender differences in the effects of heavy episodic drinking in college.

Wolaver, Amy M.

I. Introduction

Are male and female brains affected differently by alcohol use? Do these possible differences result in any meaningful differences in academic performance in college? Men and women certainly behave very differently with respect to their drinking behaviors, with men being more likely to engage in heavy episodic, or binge, drinking. If binge drinking has a negative, causal impact on grades and study effort, and academic costs greater for women, could this factor partially explain why women are less likely to binge than men?

Typically, the effects of drinking on education have not been measured for men and women separately, despite the fact that men and women metabolize alcohol differently. Further, recent studies using magnetic resonance imaging (MRIs) have found conflicting evidence about whether male alcoholics lose more or less gray matter than their female counterparts (Wuetrich 2001), leaving open the question of whether alcohol affects cognition differently for men and women. Male and female bodies process alcohol differently, but it is unclear whether these metabolic differences could result in different effects on academic success. Men and women may have systematically different preferences for drinking and studying. Differential labor market rewards for college academic performance after graduation imply that men and women will face different cost-benefit calculations related to alcohol consumption. Given the unresolved research into the physical effects of drinking on men and women and important differences in the labor market returns to education by gender, more research into gender differences in the outcomes of college drinking is clearly needed. This study fills this gap by examining whether the impact of alcohol on the behavioral outcomes of grades and study outcomes differs by gender.

This study uses the 1993 and 1997 College Alcohol Studies (Wechsler 2001, 2003) and separates the causal effects of drinking on study effort and grades from possible selection and endogeneity biases through instrumental variable techniques. The analyses are conducted separately for males and females to determine if there are significant differences between the sexes in the effects of heavy drinking on grades. Throughout this study, the behavior of interest is heavy episodic drinking, sometimes referred to as "binge" drinking. (1) An alternative measure to binge drinking, self-reported intoxication, is also modeled.

Despite the fact that men and women drink differently, study differently, and have different average GPAs, the marginal impact of heavy episodic drinking on grades is essentially the same for both genders. Men and women face similar predicted decreases in GPA for heavy episodic drinking, dropping about a third of a point on a 4 point scale. A second interesting result is that the magnitude of predicted effect on grades increases when instrumental variable techniques are used compared to estimates which assume that drinking is exogenous. This finding suggests a positive, rather than the usually conjectured negative, endogeneity bias between grades and drinking. While males and females face similar drops in grades due to drinking, the impact of drinking on studying behavior may be quite different for men and women. While the academic penalties appear to be similar for men and women, the costs to them in lost earnings in the labor market may be very different. The higher return to women in the labor market for higher GPAs is therefore a possible candidate to explain why college women are less likely to drink than men. Other candidates, not explored in this study, may be issues of difference in safety by gender or differences in the normative judgments about male and female drunkenness.

II. Literature Review

II. 1 General Models of Drinking

While clinicians and researchers in biomedical fields tend to treat alcohol abuse and alcoholism as exogenous disease states, economists take a behavioral approach. At one extreme, Becker and Murphy's (1988) rational addiction model theorizes that even addictive use of alcohol is purely behavioral. Other economic models lie somewhere in between the disease model and the rational addiction model, but choice is a factor. In the disease model, one can take the drinking behaviors as exogenous, and use simple ordinary least squares regression techniques to identify causal effects of drinking on educational and labor market outcomes. In the pure behavioral model, one needs an additional form of identification to separate out possible selection biases in the measurement of causal effects of drinking on other outcomes. Further, there may be other endogeneity issues, such as reverse causality. This identification is often accomplished with instrumental variable techniques.

Even if alcoholism is a disease, heavy episodic drinking in college does not necessarily indicate alcoholism or even clinical alcohol abuse, therefore the disease model may not apply with respect to the study of college drinking. The empirical' methods will therefore take into account and test for the possible endogeneity of drinking and the academic outcomes.

II. 2 Empirical Analyses of Drinking

College males consume more drinks in a sitting on average than college females, but because of the biological differences in the way male and female bodies process alcohol, women may experience comparable effects with fewer drinks (Lo 1995). Some evidence exists that the correlation between risk factors and drinking intensity and frequency is different for college men and women, with greek status and living in greek housing more strongly correlated with heavy drinking for men than women (McCabe 2002). Women may be affected differentially by parental and peer attitudes as well and often experience different alcohol-related problems (see Lo 1995 and 1996, for a discussion of the sociology literature on this subject).

Labor market studies of drinking typically examine direct effects of drinking on employment and wages; further, there is no clear consensus on whether alcohol has a positive or negative impact on these measures. Evidence is accumulating that there is a inverse U-shaped relationship between earnings and alcohol consumption (French and Zarkin 1995, Heien 1996, Mullahy and Sindelar 1997). That is, earnings rise with initial alcohol consumption over abstention because of positive health or social benefits of moderate drinking but begin to fall as consumption increases to problematic levels. The picture is further clouded by endogeneity issues. If alcohol is a normal good, then higher levels of income will be associated with more alcohol consumption.

Mullahy and Sindelar (1994) stress the importance of examining indirect effects, keying in on the effects of alcohol on marital status and education attainment as examples of how alcohol consumption could have additional, usually unmeasured, effects on income and earnings. One such indirect avenue is through occupation choice. Kenkel and Wang (1999) find that occupational choice is affected by alcohol use, which in turn affects wages. Mullahy and Sindelar (1997) show that the indirect effects may be more important for females than males over their lifetime. White, female alcoholics are more likely to be employed than non-alcoholics, perhaps because alcoholism is also correlated with a lower probability of marriage and fewer children, factors which may increase labor force participation in women (Mullahy and Sindelar 1997). Because of the importance of indirect effects, it is therefore important to examine the effects of drinking on educational outcomes. This literature on labor market outcomes also suggests that alcohol's effects on earnings may manifest themselves very differently by gender.

Education is another important indirect avenue through which alcohol consumption could affect earnings. The debate on whether alcohol depresses years of education is far from resolved. Alcohol consumption is generally assumed to be endogenous with educational outcomes, but there is some controversy over the appropriate instruments to use to identify the models. Cook and Moore (1993), Mullahy and Sindelar (1989) and Yamada, Kendix and Yamada (1996) find negative effects of youth drinking on total years of education and graduation rates, while Dee and Evans (2003) and Koch and Ribar (2001) finding little or no effects on those outcomes. The use of state policy variables as instruments to identify the effects of drinking on educational attainment is criticized in Dee and Evans (2003) and may be the primary source of the difference in the results in these studies.

While the effects of heavy episodic drinking on total years of education are important to establish, what are the impacts of binging among students who will eventually graduate from college and enter the work force? College grade point average is an important predictor of post-graduate earnings (Jones and Jackson 1990, Loury & Garman 1995, and Loury 1997), and the returns to GPA differ by gender. Loury (1997) finds that the earnings returns to GPA and to college graduation were higher for women than for men, indicating that any drop in GPA and probability of graduation is likely to cost females more than males. These studies highlight the value of examining the effects of alcohol consumption on college grades, and the critical importance of examining these effects separately for men and women.

As with the literature on total educational attainment, there is also mixed evidence on the effects of drinking on grades and study hours. Wood et al. (1997), using longitudinal data, find very small, statistically insignificant effects of freshmen drinking on later academic problems. Their sample, however, is drawn from students who are at-risk for alcoholism, as measured by family history, and may not be generalizable to all college students. In particular, these students may have observed techniques from their alcoholic family members that enable them to function better with alcohol use than other students. A large portion of the effect of alcohol use disappears when prior academic achievement and academic aptitude are included as controls, but these variables could also have been affected by prior alcohol use themselves. Wolaver (2002) uses instrumental variable procedures to disentangle causal from selection effects using the 1993 Harvard College Alcohol Study (CAS) and finds that heavy drinking, as measured by binge drinking and self-reported intoxication, does appear to have a detrimental effect on grades, both directly and indirectly through study effort. In contrast, using the 1993-1999 CAS data, Williams et al. (2003) find that continuous measures of drinking (average drinks per occasion, total drinks, and number of occasions drinking in the past 30 days) have a positive direct effect on GPA, negative indirect effect through reduced studying, and a small, negative total effect on GPA. The instruments, measures of drinking, and functional forms of the analysis differ in these studies, which may explain these discrepancies.

The literature on alcohol and human capital effects provides several relevant lessons for this study. First, possible endogeneity between alcohol consumption and other educational measures is the usual presumption by economic studies which are based in the behavioral, rather than medical, model of alcohol consumption. Second, the value of appropriate instruments in treating the endogeneity bias should not be underestimated. The empirical methods outlined below will take into account the probable endogeneity of drinking and human capital formation. Finally, the results strongly suggest that one should examine the possibility of gender differences in the impact of drinking on grades and study effort.

III. Empirical Methods

Economic theory highlights the importance of accounting for the possible endogeneity of drinking and academic effort and outcomes. Theory suggests the following general functional form for the empirical analysis:

[Y.sub.i] = f([X.sub.i,Y], [G.sub.i], [S.sub.i]) (1)

[G.sub.i] = g([X.sub.i,G,] [Y.sub.i], [S.sub.i]) (2)

[S.sub.i] = h([X.sub.i,S,] [Y.sub.i] (3)

where [Y.sub.i] represents the drinking behavior of the individual, i, [G.sub.i] is college GPA, [S.sub.i] is study hours, f, g, and h are general functions, and [X.sub.i,Y], [X.sub.i,G] and [X.sub.i,S] are vectors of personal characteristics that are related to drinking behavior, GPA, and study hours, respectively; there may be some overlap in the members of these vectors of characteristics. GPA and study hours appear as determinants of drinking behavior because there may be reverse causality. Several examples illustrate the possible relationships; good grades may provide a celebratory reason for students to drink more. Alternatively, after students receive bad grades they may drink more to "drown their sorrows" or because they decide to shift efforts away from education and into consumption behavior. Similar stories could be told about the relationship between study hours and drinking. Another source of endogeneity bias may be through omitted variables; there may be unmeasured factors that affect both the probability of heavy drinking and the academic outcomes. If students smooth their consumption over the lifetime and if good grades lead to higher incomes later in life, then the higher income associated with good grades could also introduce a bias into the results. If one does not account for this possible endogeneity, then the relationship of interest, the causal impact of drinking on grades may be either over- or underestimated, depending on the nature of the reverse causality. Durbin-Wu-Hausman tests for the endogeneity of drinking and the academic outcomes are conducted and reported.

Given the categorical nature of the grade and drinking variables, the specific forms are as follows:

[y.sup.*.sub.i] = [alpha]'[X.sub.i,Y] + [[epsilon].sub.i,Y] [Y.sub.i] = 1 if [y.sup.*.sub.i] [greater than or equal to] 0, [Y.sub.i] = 0 if [y.sup.*.sub.i] < 0, (4)

[S.sub.i] = [beta]' [X.sub.i,S] + [[beta].sub.Y][Y.sub.i][[epsilon].sub.i,S] (5) [G.sub.i] = [gamma]' [X.sub.i,G] + [[gamma].sub.Y][Y.sub.i] + [[gamma].sub.S][S.sub.i] + [[epsilon].sub.i,G]

Where [GPA.sub.i] = D if [G.sub.i] [less than or equal to] [[mu].sub.0] [GPA.sub.i] = C if [[mu].sub.0] < [G.sub.i] [less than or equal to] [[mu].sub.1] [GPA.sub.i] = B if [[mu].sub.1] < [G.sub.i] [less than or equal to] [[mu].sub.2] [GPA.sub.i] = A if [[mu].sub.2] < [G.sub.i] (6)

Grades are certainly ordinal, but they may not be cardinal. (2) A linear specification would assume cardinality (that is, the difference between an A and a B is the same as the difference between a C and a D). Instead, grades are modeled as an ordered probit to allow flexibility in the treatment of grade differences.

For purposes of this study, (5) and (6) are the equations of interest, while equation (4) represents the reduced form determinants of drinking. Two sets of nonlinear simultaneous equations are estimated with Generalized Methods of Moments (GMM), using SAS version 8.02. Equations (4), (5), and (6) are modeled simultaneously to determine the effect of drinking on study hours and grades. Using GMM is analagous to two-stage least squares and allows for the Hansen (1982) test of overidentification. An alternative would be to use full-information maximum likelihood (FIML), but errors in the specification tend to be magnified with FIML (Greene 1993). Because omitted variable bias is a very real concern given the data limitations, GMM is preferred to FIML. The drinking variables are also discrete indicator variables and modeled as probit specification as shown in equation (4). Study hours (plus 0.01) are logged and modeled linearly.

Instrumental variable techniques offer an avenue of identification to control for the endogeneity bias; if at least one characteristic in the vector [X.sub.i,Y] is not also included in the vectors of [X.sub.i,S] and [X.sub.i,G] then [[beta].sub.Y] and [[gamma].sub.Y] can be consistently estimated. The dollar price of alcohol is the most theoretically defensible instrument. Unfortunately, data on monetary prices paid by students were not collected in the 1993 wave of the survey. The full price, especially for underage students, will include nonpecuniary aspects as well as the monetary price (Chaloupka and Wechsler 1996). Alternative instruments could be state or local drinking policies, but the public use data do not contain geographic identifiers. Additionally, the use of state-level policy variables has been criticized in Dee and Evans (2003) as being correlated directly with educational outcomes, biasing estimates using these variables as instruments. Because of the difference in the availability of price data, two sets of estimates are performed. Those that include both monetary and nonmonetary measures use 1997 data only; those that include only the nonmonetary price measures are run using both sample years. The nonpecuniary costs of drinking are captured by a set of variables to capture religious attitudes, a set of variables indicating parental use of alcohol, and variables created from student responses to indicate the ease of availability of alcohol for underage students at the college. Because multiple instruments are used, the model is overidentified; various specification tests are performed. (3)

The vectors of coefficient estimates [[beta].sub.Y], [[gamma].sub.Y], and [[gamma].sub.S] are biased if unobservable characteristics contained in [[gamma].sub.i,S], and [[gamma].sub.i,G] are correlated with the drinking behavior variable. Ideally, panel data with information about previous study habits and GPA would allow a fixed effects model to control for much of the unobservable characteristics, but the data are cross-sectional. Further, the data do not include other potentially useful information such as SAT score, high school GPA, or family income, which could lead to omitted variable bias. The best strategy to mitigate this problem is to include high school drinking behavior (hsbinge) as a covariate for the college academic outcomes, to control for student "type."

This latter strategy produces another potential bias to the results. Prior drinking is highly correlated with current drinking, and if prior drinking has long-lasting effects or is an indicator of student preferences, it should be included as a determinant of the academic outcomes. Adding high school drinking to the model may bias the results, however, if the rational addiction model accurately describes drinking behavior (Becker and Murphy 1988). In the rational addiction model, previous drinking will be affected by current drinking and the current price, which may include detrimental effects on study hours and GPA, and is therefore also an endogenous variable. It may also be that heavy drinking has both long and short term effects on grades; high school binging may be picking up some of the long-term impact of youth drinking.

Because study hours may also be endogenous with GPA, instruments are necessary to identify its effect on grades. Peer study hours and the average GPA of other students at the school are used as instruments to identify this effect. Peer study behavior may affect an individual's academic effort but should not directly affect her GPA. Likewise, one's motivation to study may be affected by the GPA of one's college classmates, but one's own GPA should not be directly affected by others' grades. However, the presence of varying degrees of grade inflation or other institutional variations in grading procedures could produce some omitted variable bias to this analysis. The results are tested for potential bias due to institutional factors in several ways; first, by adding average GPA to the grade equation, second, by modeling grades as the difference between the respondent's GPA and the college average, and third, by using an instrumental variable fixed effects model. These specification tests are shown in Table 5. Various other robustness checks were also performed, but, for brevity's sake, are not shown here. They are described below and the results are available from the author upon request.

IV. Data

The 1993 and 1997 Harvard College Alcohol Studies (CAS) (Wechsler 2001, 2003) are a large, nationally representative sample of college students and are primarily focused on drinking behaviors and attitudes. The data contain demographic information as well as students' academic, extracurricular, and social activities. It is possible to construct school-level variables based on the student population sample analogs at each college. In 1997, the data also include information on the price paid for alcoholic beverages by the respondent. Table 1 shows the names, definitions, and the gender-specific means and standard deviations for selected variables.

See Wechsler et al. (1994) for a discussion of the sample. Some of the colleges have been dropped from the 1997 sample, which may also affect the results. Female college students are slightly over-represented in the sample- women are roughly 59% of the CAS sample, but only 55 percent of college enrollees were women in 1993 & 1997 (U.S. Census Bureau 2004). Some of the difference may be due to the survey's inclusion of some traditionally women's only colleges, which account for an extra 1297 of the female respondents. There are more females in the 1997 data than in the 1993 data (57.6% in 1993 compared to 59.9% in 1997). The difference in the percentage of females between 1993 and 1997 is statistically significant at the 99% significance level. (4) The 1997 data include a higher percentage of respondents from traditionally female colleges, which indicates that coeducational institutions were more likely to drop from the sample than women's colleges. The main concern with these sample issues is because purpose of the study is to compare the effects for females to males. The estimated differences could be biased if females select into responding to the survey differently than males select into responding to the survey. It is not immediately obvious whether these different participation rates for males and females would produce a negative or positive bias in the differences in the effects of drinking between the genders.

Consistent with most available data on gender-specific drinking habits (Engs and Hanson 1990), men are more likely to drink heavily than women under any definition. Men are 40% more likely to binge than women, and 140% more likely to binge more than twice in the previous two weeks. The differences are smaller when self reported intoxication (drunk) is used as a measure; men are 20% more likely to report at least one episode of intoxication in the previous month than women, but they are twice as likely to report six or more episodes of intoxication (freqdrunk). These differences are statistically significant at the 1% confidence level. Men and women also differ in the type of alcohol they prefer to drink. While beer is the mode among those who drink for both genders, men are almost twice as likely as women to report it as their usual drink. Female drinkers are three times as likely as male drinkers to prefer wine and wine coolers combined, and almost twice as likely to prefer mixed liquor drinks. Women spend on average 19.2 more minutes a day studying than men and have slightly higher grades on average (3.19 compared to 3.08 on a four point scale). These differences are statistically significantly different from zero at the 99% significance level.

Other important variables are described here. Because nondrinkers did not respond to the price and availability questions, price information needs to be constructed for these students. Further, the price questions include other information besides a dollar value for the prices paid for one drink. Therefore, a set of college price averages among drinkers is constructed. In 1997, avgpriee is defined as the average reported nonzero price among drinkers at the college, excluding the respondent. Additionally, some drinkers report paying nothing for alcohol and others report paying a fixed fee for drinks. The fraction of drinkers (excluding the respondent) at the college reporting these pricing mechanisms (drnkfree and fixfee) are included in the models also to represent zero marginal costs. Alcohol is relatively inexpensive-14 to 15 percent of men and women report drinking for free, with an additional 5 percent paying a fixed fee (and therefore facing zero marginal costs per drink); among those who do report a price, the cost is less than $2 per drink on average, although it varies from a low of $0.25 to a high of $3.01. Because these variables are college averages based on student responses, they include measurement error, and the coefficient estimates on the effect of price on drinking will be biased towards zero.

Because of the relatively low monetary prices paid by college students for alcohol, other costs of student drinking, particularly for underage drinkers, are likely to be important, perhaps more important than the pecuniary price, given the low average reported prices. Strength of religious beliefs (strongrelig) has been tied to drinking behaviors, but should not affect study hours or GPA. Indicator variables for affiliation with different religious denominations are also included as instruments (catholic, jewish, moslem, protestant, and other) Parental use and family attitudes about drinking are also employed as instruments, as students' preferences for alcohol may be affected by their upbringing. These full costs are proxied by religious denomination, strength of religious beliefs (strongrelig), parents' drinking behaviors (momever, dadever, momheavy, dadheavy), family attitudes about drinking (famdisp), and two variables representing the ease of obtaining alcohol for underage drinkers at the college. The latter are proxied by fraction of drinkers at the college that have purchased alcohol without identification (noid) and an interaction term between an indicator for the respondent being under 21 years of age and noid (noidu21). Relative preferences for time discounting and risk taking are proxied with cigarette use within the past year (cigyr). Although, as mentioned above, monetary price variables are the most theoretically valid instruments, they are not statistically significant and may produce bias due to weak instruments. Therefore, the specifications presented here include the full set of instruments, which are jointly significant, and tests of overidentification are presented below.

Finally, to control for the average grades of other students at the college, avggpa is constructed as follows. Each letter grade is assigned a numerical value on a 4-point scale (A = 4.0, A- = 3.67, B+ -3.33 etc ...) and the numerical average is constructed for each college in each sample year. The survey questions regarding study hours represent the average hours per day of study from the past 30 days reported by students. The responses can range from 0 to 8 or more hours.

V. Results

The results for the multivariate regressions are presented in Tables 2 and 3. Table 2 shows the reduced form probit results for selected variables on the probability of binging and the probability of self-reported intoxication measures. The dependent variable in the top panel of the table equals 1 if the student reported having engaged in binge drinking (5 drinks in one sitting for men and 4 drinks for women) and zero otherwise. The bottom panel presents the result for the analogous variable defined as being intoxicated in the previous 30 days. Table 3 shows the simultaneous equation results for the effect of selected variables on study hours and grade point average. These latter estimates are corrected for the endogeneity bias.

V. 1 Drinking Determinants

Most of the predictors for drinking behavior perform as expected, with a few exceptions. The price variables, in particular, are rarely statistically significant and the signs are often the wrong direction. The variables indicating zero marginal cost do not statistically significantly raise the probability that a student drinks heavily and some estimates are negative. That is, the more students at the college who report drinking for free or paying a fixed fee for alcohol, the less likely the respondent is to drink heavily, contrary to predictions. The relatively poor performance of these variables may be due to several factors. First, since these variables are averages for the college, they are measured with some error and are therefore subject to some attenuation bias, which will bias the coefficients towards zero. Second, the prices may also be correlated with other policies or institutional factors which are not conducive to heavy drinking. Third, since alcohol is so inexpensive, the monetary price may have a smaller impact among these relatively wealthy individuals. Fourth, the operational margin for these variables may not be adequately reflected in the dependent variables. For example, the monetary price may lower the total number of drinks consumed for students over the course of a month or a year, but does not affect their probability of drinking four or more (or five or more) drinks or becoming intoxicated in one sitting. For underage students, the legal and moral impediments to obtaining alcohol may be more important than the price.

The results provide some support for the latter hypothesis, at least for males. The interaction term of fraction of students obtaining alcohol without an id and the indicator for being underage performs better- it is usually positive, as expected, and statistically significant for both genders when the dependent variable is self-reported intoxication. The results for other variables that may reflect other costs of heavy drinking to the student-family attitudes and religious beliefs, provide additional support for the hypothesis that nonpecuniary price is relatively important for college students. Strong religious beliefs are always negatively and statistically significantly correlated with heavy episodic drinking, as expected, and the religion variables are all always jointly statistically significant. Being Catholic increases and being a Moslem decreases the probability of heavy drinking relative to no religious affiliation. The other religious denominations have mixed effects, probably due to error in the variable definition. For example, some Protestant denominations strongly discourage alcohol consumption while others have more liberal attitudes toward alcohol; the data do not allow one to distinguish between these denominations.

The effect of family disapproval of drinking is not consistent with prior expectations. The estimates are positive, and are statistically significant in one case. The measures of any parental use (momever, dadever) are usually positive as expected. Mothers' drinking behaviors appear to be a stronger determinant of drinking than fathers' behaviors, particularly for sons. There are mixed effects of parental heavy drinking. These variables are also likely to be subject to some measurement error, since they are reported by the student and no definition of what constitutes "heavy" drinking is given in the questionnaire--students are therefore making subjective judgments and these judgments may differ across individuals.

V. 2 Study Hours

The effects of current and past drinking on study hours, as well as the influence of peer study habits and peer grade point average, are shown in Table 3A. If the instruments are valid, these effects can be interpreted as causal effects and not mere correlations. A distinct, but not universal, pattern emerges. For men, previous binging increases study hours and current drinking decreases study hours in the 1997 sample, but current drinking does not have a statistically significant effect in the 1993 & 1997 joint data which do not include the monetary price identifiers. The pattern is much more consistent for women, however. Previous drinking always decreases current study hours and current drinking always increases study hours. In addition, drinking has a bigger effect on females' study habits compared to males, as the coefficient estimates are usually much larger in magnitude. The impacts are more precisely measured for females also; the coefficient estimate for high school binging is always statistically significant. The positive effect of current drinking on study hours may have an efficiency explanation. Female binge drinkers may be less efficient in their studying because of alcohol's effect on cognitive function or some other mechanism. Another possible interpretation is that women are more aware or value the costs of binging more than men, and female heavy drinkers attempt to mitigate those costs by studying more to make up for their heavy drinking. The differences in the estimated effects for males and females are statistically significant. Poor study habits of drinkers may have been formed in high school, or the high school binging variable may be capturing unobserved preferences for studying and discount rates. It may also be that drinking has both short and long term effects, and the high school binge variable is capturing the long term effects of previous heavy drinking.

Both men and women respond positively to the average study hours of their peers- these coefficient estimates are always statistically significant, with males reacting slightly more strongly than females, in most cases. Students do not respond strongly to the average GPA of their peers, with the exception of males in one of the specifications. Perhaps two countervailing pressures are captured by the average peer GPA; grade inflation on the one hand may tend to depress student effort, while competing against able peers may tend to increase student effort. The overall impact on study effort may be a wash.

V. 3 Grade Point Average

The direct effect of drinking on GPA is uniformly negative. The coefficient estimates presented here are not easily translatable into marginal effects because of the nature of the ordered probit and because drinking may have indirect effects on grades through study hours. Negative estimates in an ordered probit do indicate a decreased probability of falling in the top category, in this case the "A" GPA, relative to the other possibilities but it is unclear whether the probability of falling in another category will increase or decrease. The direct and indirect marginal impacts are computed and presented in Table 4. Both current and high school binging negatively impact the probability of an "A" GPA. These estimates are statistically significant in just under half of the male specifications and in all but one of the female specifications. The coefficient estimates for college drinking's effects on GPA are usually larger for males than for females. Unlike the results for study hours, the gender differences in the impact of drinking on GPA are not statistically significant, nor, as shown below, do they appear to be economically important. However, similar marginal drops in GPA do present different labor market costs by gender.

V. 4 Marginal Effects

To give a better sense of the importance of the effects of heavy episodic drinking on grades, Table 4 shows the predicted average study hours and GPAs for bingers versus non-bingers for each gender. Because the coefficient estimates for a single-equation ordered probit are not directly interpretable and drinking effects grades both directly and indirectly through its effect on study hours, additional calculations must be made to show the marginal effects. The binge variable is artificially set to zero and one, alternatively, with individuals retaining their other characteristics, and the predicted study hours and probabilities of receiving each grade are calculated under each of these alternatives (See Greene 1993, for the formulas for the predicted probabilities in ordered probits).

The effects are quite large; remember that GPA here is cumulative over the year to date, not for a single course. For both males and females, the probability of receiving an "A" grade point average decreases with heavy drinking. Since study hours are positively affected by drinking for women and men, the total decrease in the "A" GPA probabilities are mitigated somewhat when the indirect drinking effects are included. There are some interesting qualitative differences in the predicted total impact on GPA between men and women. Both experience similar decreases in the probability of an "A" GPA (17 percentage points for women and 16 percentage points for men) with binging. More women shift into "B" averages than men, and slightly more men shift into "C" averages. Taking the changes in probabilities and multiplying them by a numerical value for A's, B's, C's and D's produces an estimated marginal effect on GPA. Using this methodology, these disparities translate into similar drops in expected GPA, almost a third of a point, or the difference between a "B" and a "B-" for women, and about a "B-" and a "C+" for men.

Although average GPA falls by roughly the same amount for men and women from binge drinking, the negative effects on postgraduate earnings are larger for women. Table 4 also provides some back of the envelope estimates of earnings losses by gender. These are not intended to be exact measures of the earnings losses from drinking, but are provided to give a rough sense of the scale of the costs in the labor market for men and women. Results from Loury (1997), who examines the earnings returns to grade point average separately for men and women, are used to estimate the expected drop in GPA into expected weekly and annual earnings losses. Women bingers can expect a 4.2 percent drop in earnings, which translates into losses of $35.83 weekly or $1863.03 annually. Men, on the other hand, can expect to lose only 0.47 percent in earnings from the associated drop in GPA, or $8.51 weekly and $276.68 annually, from binging. These are based on the median weekly wages of college graduates, which will tend to overestimate the initial wage loss. Since the drop in expected GPAs are similar for men and women, despite the slight qualitative differences in the way grades fall, the earnings differences are largely due to the way the labor market differentially rewards male and female college graduates. Women get a much higher boost in earnings from higher GPAs than men, therefore the impact of college drinking on their post-graduation outcomes is much larger. Women may be less likely to drink heavily and those women who binge may increase their study effort to mitigate the grade impact simply because it is more costly for them to do so. The estimates of returns to college GPA are becoming dated, and the gender composition of college graduates is changing (Sum et al. 2003) which may change the labor market responses to grades in the future, but they provide a compelling possible explanation for the observed gender gap in heavy drinking in college.

V. 5 Specification Tests

Instrumental variables are used to remove endogeneity bias from the estimates and to produce estimates of the causal impact of drinking on academic outcomes. There are several areas of potential concern with instrumental variable results in general that should be addressed. First, one should have theoretically defensible instruments; in this case, variables that have a direct effect on drinking but no direct effect on study hours or GPA. The strategy chosen in this paper is to use variables proxying for the "full" price of alcohol. These variables include information on the monetary prices paid by students at the college, including those who pay only a fixed fee and those who report paying nothing for alcohol, as well as nonmonetary variables, including ease of access measures (noid, noidu21), family attitudes and parental drinking measures, and religious beliefs. Second, these instruments should have statistically significant effects on binge drinking, because using weak instruments may also bias the results (Stock and Staiger 1997). The full set of instruments are jointly statistically significant, but the subset of monetary price variables are not, probably partly because of attenuation bias. Third, in this analysis, multiple possible instrumental variables exist, leading to the possibility that if each instrument were used in turn singly in the analysis, each of these separate estimations might produce different point estimates of the causal effect of drinking on studying and grades. Overidentification tests are performed. The Objective x N test statistics from the GMM estimates are shown in Table 3; overidentification cannot be rejected in only two cases in the study hours estimates. These cases are for both men and women in the 1997-only-sample estimates when intoxication is used as the drinking measure. The test fails in all of the grade point average estimates. Therefore, one can be confident that the point estimates presented are "correct" only for estimates of the impact of intoxication on study hours in the 1997 samples.

Because the results of tests of overidentifying restrictions indicate that different sets of instruments might produce different point estimates of the impact of drinking on studying and grades and because the monetary price variables are statistically insignificant, the robustness of the results is in question. Each (set) of the non-monetary instruments used are debatable--plausible arguments that they may directly affect study hours and GPA could be made in many cases. Kenkel and Ribar (1994), for example, include religiosity in their earnings estimates, arguing that it is a personal characteristic that may affect preferences for earnings versus other non-pecuniary job aspects. Parental use of alcohol may affect the student's concentration or mother's use during pregnancy may have affected cognitive abilities. The price variables may be correlated with other school policies or with factors associated with the academic environment which are not available and may therefore be poor instruments (see Dee and Evans 2003 for a discussion of these factors for high school students). As discussed above, the inclusion of high school binge drinking in the study hours and GPA equations may also bias the results. Finally, the peer rates of drinking, study hours and grades may themselves be endogenous (Manski 2000), which could bias the coefficient estimates not only for those variables, but for all of the others in the analysis.

Are the overall implications of the estimates affected by this overidentification? To test whether these potential problems affect the conclusions reached, the estimates were re-run using different sets of variables as instruments. Single equation estimates which do not correct for endogeneity are also performed. The results are not shown, but available from the author upon request. The general flavor of these robustness checks are described below.

The results for study hours do vary depending upon the specification chosen. The coefficient for binge drinking is usually positive and statistically significant for females but does switch sign in one specification. The results for men are much more variable. The impact flips from positive to negative and is statistically significant in only about half of the specifications. Because study hours have an impact on grades, these estimates could have some impact on the total impact of drinking on GPA. However, the indirect effect of drinking on grades through study hours is relatively small and this variation does not change the overall direction of the effect of heavy episodic drinking on grades.

The picture for the direct effect of drinking on grades is much more consistent. Binge drinking always has a statistically significantly negative impact on grades for both men and women, regardless of the set of instruments or the subsample used. The estimates do vary somewhat in the magnitude of the effect of drinking on grades; the lowest estimated impacts are those which treat binge drinking as exogenous to grades. Finally, in most of the estimates, the differences between males and females are nonexistent or are very small. The consistency of these estimates provide reassurance that, despite the failure of the overidentification tests, the overall conclusions shown here are valid. One should not place too much importance on the point estimates presented here as an exact measure of the impact of heavy episodic drinking on grades but can be relatively confident that binge drinking has an independent, negative impact on grades.

One final econometric issue should be addressed. It is possible that a student's GPA at one school is very different than a GPA at a different school. There could be institutional differences in grading. For instance, different levels of grade inflation or curving practices could exist across schools. Several strategies are used to address this possibility. Further, the average GPA at the school might have a direct impact on grades and therefore not be an appropriate instrument to separate the causal effect of study hours on grades from endogeneity bias between these two outcome variables. This issue is dealt with using several different methods. The coefficient estimates for the effect of binge drinking on grades from these additional robustness checks are presented in Table 5. First, average GPA at the college is included in equation (6) as well as equation (5), that is, average GPA at the college is no longer used as an instrument for study hours. These results for the binge drinking effect on grades are shown as Test A. Second, the respondents' grades were assigned a numerical value and the dependent variable in equation (6) is replaced by the difference between the respondent's GPA and the college average. Instead of using an ordered probit specification, this test uses a linear specification for equation (6). These results are shown with Test B. Finally, the respondent's numerical GPA was used as the dependent variable in an instrumental variables, college fixed effects model; shown in Table 5 as Test C. The Test A results are in line with those presented above in the baseline marginal effects. Since endogeneity is accounted for, and since the binge drinking of one respondent should not affect the college average GPA, the Test B results may also be interpreted as a marginal effect of heavy episodic drinking on individual GPA. In each of these tests, binge drinking continues to negatively affect GPA. Further, the effects are economically similar for both males and females across all specification changes, bolstering the conclusion that binge drinking has the same effect on college men and women in terms of their academic productivity. The size of the predicted effect of binge drinking on grades, however, is somewhat variable. In one specification the magnitude of the effect drops to a marginal effect of about a tenth of a point on a four point grade scale but remains statistically significantly different from zero.

VI. Conclusions

Several lessons can be culled from this analysis. The research highlights the importance of careful econometric analysis when examining the impact of drinking on educational outcomes. Simple correlations between heavy drinking and academic problems may actually understate the true, causal effect on grades, a somewhat surprising result, given the usual stated assumption that the poorer students are more likely to select into heavy drinking. The estimates fail tests of overidentifying restrictions, but robustness checks provide strong assurances that the direct effect of binge drinking on grades is negative, although the exact magnitude of the effect may be imprecisely measured. The results discussed represent a middle point estimate of the true underlying causal impact of drinking on grades.

The impact of college drinking on study hours is much less clear for men but appear to be positive for women. Significant differences are evident in the effects of binging and intoxication episodes between men and women. Study hours are a behavioral outcome that is more clearly under the control of the respondent. The study hours' estimates are consistent with Mullahy's and Sindelar's (1997) evidence that indirect effects of drinking on earnings may be more important for females than for males and warrant further investigation.

Finally, and most importantly, the study addressed the question of whether there are gender differences in the impact of drinking on college performance. The scientific literature does not give us a definitive answer about whether mens' and womens' brains are affected differently by alcohol from a physical standpoint, but the results presented here indicate that the marginal impacts of heavy episodic drinking on grades between men and women are statistically indistinguishable from one another. The similarity in results is surprising given the much different propensity of college men and women to engage in these drinking behaviors. Furthermore, even if the coefficient estimates between men and women were statistically significant, they do not produce differences in the marginal effects on grades that are economically important. Rough estimates of the impact of this grade drop on earnings suggest possible explanations for the variation in heavy episodic drinking by gender, but future research with additional data should try to measure the earnings impacts of youth drinking more directly.

The insight that binge drinking has similar impacts on college grades for men and women eliminates it as a possible explanation for the observed variation in the rates of binge drinking/self-reported intoxication between men and women. But, while heavy drinking is predicted to have essentially the same negative effect on grades for males and females, females stand to lose roughly ten times the labor market earnings relative to otherwise identical males from the same drop in GPA. This increased cost of binging for females is a plausible contributor, although not the only plausible one, to the observed lower rates of binging for men versus women.

References

Becker, G.S., and K.M. Murphy. 1988. A Theory of Rational Addiction. Journal of Political Economy 96(4): 75-700.

Chaloupka, F.J. and H. Wechsler. 1996. Binge Drinking in College: The Impact of Price, Availability, and Alcohol Control Prices. Contemporary Economic Policy 14(4): 112-114.

Cook, P., and M.J. Moore. 1993. Drinking and Schooling. Journal of Health Economics 12(4): 411-29.

Dee, T.S., and W.N. Evans. 2003. Teen Drinking and Educational Attainment: Evidence from Two-Sample Instrumental Variables Estimates. Journal of Labor Economics 21 (1): 178-209.

Engs, R.C., and D.J. Hanson. 1990. Gender Differences in Drinking Patterns and Problems Among College Students: A Review of the Literature. Journal of Alcohol and Drug Education 35(1): 36-47.

French, M.T., and G.A. Zarkin. 1995. Is Moderate Alcohol Use Related to Wages? Evidence from Four Worksites. Journal of Health Economics 14 (3): 319-345.

Greene, W. 1993. Econometrics 2nd ed. New York: Macmillan Publishing Company.

Hansen, L. P. 1982. Large Sample Properties of Generalized Method of Moments Estimators. Econometrica 48(3): 1029-1054.

Heien, D.M. 1996. Do Drinkers Earn Less? Southern Economic Journal 63(1): 60-69.

Jones, E.B., and J.D. Jackson. 1990. College Grades and Labor Market Rewards. Journal of Human Resources 25(2): 253-66.

Kenkel, D. and D.C. Ribar. 1994. Alcohol Consumption and Young Adults' Socioeconomic Status. Brookings Papers on Economic Activity, Microeconomics 1994:119-61.

Kenkel, D., and E Wang. 1999. Are Alcoholics in Bad Jobs? In The Economic Analysis of Substance Use and Abuse, ed. Chaloupka, F.J., Grossman, M., Bickel, W.K. and H. Saffer Chicago: University of Chicago Press.

Koch S., and D.C. Ribar. 2001. A Siblings Analysis of the Effects of Alcohol Consumption Onset On Educational Attainment. Contemporary Economic Policy 19(2): 162-174.

Lo, C. 1995. Gender Differences in Collegiate Alcohol Use. Journal of Drug Issues 25(4): 817-836.

Lo, C. 1996. Are Women Heavier Drinkers than We Thought They Were? Journal of Studies on Alcohol 57(5): 531-535.

Loury, L.D. 1997. The Gender Earnings Gap Among College-Educated Workers. Industrial and Labor Relations Review 50(4): 580-593.

Loury, L.D., and D. Garman. 1995. College Selectivity and Earnings. Journal of Labor Economics 13(2): 289-308.

Manski, C.F. 2000. Economic Analysis of Social Interactions. Journal of Economic Perspectives 14(3): 115-137.

McCabe, S.E. 2002. Gender Differences in Collegiate Risk Factors for Heavy Episodic Drinking. Journal of Studies on Alcohol 63(1): 49-56.

Mullahy, J., and J.L. Sindelar. 1989. Life-Cycle Effects of Alcoholism on Education, Earnings, and Occupation. Inquiry 26(2): 272-283.

Mullahy, J., and J.L. Sindelar. 1991. Gender Differences in Labor Market Effects of Alcoholism. American Economic Review: AEA Papers and Proceedings 81 (2): 161-65.

Mullahy, J., and J.L. Sindelar. 1994. Alcoholism and Income: The Role of Indirect Effects. The Milbank Quarterly 72(2): 359-375.

Mullahy, J., and J.L. Sindelar. 1997. Women and Work: Tipplers and Teetotalers. Health Economics 6(6): 533-537.

Perkins, H.W., W. DeJong, and J. Linkenbach. 2001. Estimated Blood Alcohol Levels Reached by "Binge" and "Nonbinge" Drinkers: A Survey of Young Adults in Montana. Psychology of Addictive Behaviors 15(4): 317-320.

Staiger, D., and J.H. Stock. 1997. Instrumental Variables Regression with Weak Instruments. Econometrica 65(3): 557-587.

Sum, A., N. Fogg, and S. Palma. 2003. The Growing Gender Gaps in College Enrollment and Degree Attainment in the U.S. and Their Potential Economic and Social Consequences. Prepared for the Business Roundtable web accessed at: http://www.manoa.hawaii.edu/ovcafo/ DataSupport/VCAFO%20Bulletins/gender_gap. pdf on January 30, 2007.

U.S. Census Bureau, 2004. School Enrollment: Table A7 College Enrollment of Students 14 Years Old and Over, by Type of College, Attendance Status, Age, and Gender: October 1970 to 2002. Washington, D.C.: U.S. Census Bureau Population Division, Education & Social Stratification Branch, web accessed at: http://www. census.gov/population/www/socdemo/school. html on May 11, 2005.

Wechsler, H. 2001. Harvard School of Public Health College Alcohol Study II [United States]: A Search for Change And Its Correlates, 1997 [Computer File]. ICPSR Version. Boston, MA: Harvard School of Public Health [Producer], 1997. Ann Arbor, MI: Inter-University Consortium for Political And Social Research [Distributor].

Wechsler, H. 2003. Harvard School Of Public Health College Alcohol Study, 1993 [Computer File]. 2nd ICPSR Version. Boston, MA: Harvard School of Public Health [Producer], Ann Arbor, MI: Inter- University Consortium for Political And Social Research [Distributor].

Wechsler, H., Davenport, A., Dowdall, G., Moeykens, B., and Castillo, S., 1994. "Health and Behavioral Consequences of Binge Drinking in College: A National Survey of Students at 140 Campuses," Journal of the American Medical Association, 272(21), 1672-1677.

Williams, J., L.M. Powell, and H. Wechsler. 2003. Does Alcohol Consumption Reduce Human Capital Accumulation? Evidence from the College Alcohol Study," Applied Economics 35(10): 1227-1240.

Wolaver, A.M., 2002. Effects of Heavy Drinking in College on Study Effort, Grade Point Average, and Major Choice. Contemporary Economic Policy 20(4): 415-429.

Wood, P.K., K.J. Sher, D.J. Erickson, and K.A. DeBord. 1997. Predicting Academic Problems in College from Freshmen Alcohol Involvement. Journal of Studies on Alcohol 58(2): 200-210.

Wuetrich, B., 2001. Does Alcohol Damage Female Brains More? Science 291 (5511): 2077-2079.

Yamada, T., M. Kendix, and T. Yamada. 1996. The Impact of Alcohol Consumption and Marijuana Use on High School Graduation. Health Economics 5(1): 77-92.

Footnotes

(1.) Binge drinking is typically defined as four drinks in one sitting for females and five drinks for males. The definition is controversial, since it is unclear whether the respondents are drinking to impairment, and may overstate "problem" drinking. Perkins, et al. (2001) estimate blood alcohol for "binge" drinkers. They find that level of impairment varies widely in this group. Binge drinking is used as a measure here because it is comparable to existing literature but self-intoxication measures are also used for comparison.

(2.) The survey asks for an approximate GPA for the year to date using letter grade categories, including pluses and minuses, i.e. A, A-, B+, B, etc ... An "F" average is not an option, most likely because of minimum GPA requirements to remain a student in good standing at most colleges universities. Therefore, the ordered probit, while more complex econometrically, fits the nature of the data from the survey. One can artificially create a linear specification for grades by assigning a numerical value to each of the categories and running the estimates, but this variation on the estimates does not produce markedly different conclusions from those presented here. These results are similar qualitatively to those presented here and are available from the author upon request.

(3.) In addition, results may be biased if the instruments are weak (Staiger and Stock 1997)- [chi square] tests of the joint significance of subsets of instruments were conducted. These results are not shown for brevity's sake but are available from the author upon request. They are summarized here: the joint set of religion variables, the set of family use measures, and the full set of instruments are always jointly statistically significant, while the set of price variables are statistically significant for three of the four specifications conducted on the male sample and for none of the female sample specifications.

(4.) The 95% confidence levels for 1993 are 56.9 to 58.3% and for 1997 are 59.1% to 60.7%.

Amy M. Wolaver, Department of Economics, Bucknell University, Lewisburg, PA 17837. Phone: (570) 577-1699, Fax: (570) 577-3451, email: awolaver@bucknell.edu

TABLE 1
Descriptive Statistics, Selected Variables, by Gender

 Males Females

 Std. Std.
Variable, Definition Mean Dev. Mean Dev.

binge =1 if 5 drinks for men,
 4 for women in past 2 wks 0.49 0.50 0.35 0.48
freqbinge =1 if binge 2 +
 times in past 2 weeks 0.24 0.42 0.13 0.33
hsbinge = 1 if binged in
 high school 0.34 0.48 0.28 0.45
drunk = 1 if self-reported
 intoxication in past month 0.52 0.50 0.45 0.50
freqdrunk = 1 if drunk 6 +
 times in past month 0.12 0.32 0.06 0.24
greek = 1 if in fraternity
 or sorority 0.15 0.36 0.14 0.35
gpa - grade point average,
 4 point scale 3.08 0.60 3.19 0.57
studyhrs - usual hours
 study per day 2.86 1.69 3.24 1.61
avgprice (1997 data only)
 college avg. reported price
 of an alcoholic beverage
 of those reported non-zero
 price, respondent excluded 1.91 0.28 1.97 0.32
drnkfree (1997 data only)
 fraction of drinkers at
 college who report usually
 drink free, respondent
 excluded 0.14 0.06 0.15 0.08
fixfee (1997 data only)
 fraction of drinkers at
 college who report usually pay
 fixed fee for unlimited
 quantity of alcohol,
 respondent excluded 0.05 0.05 0.05 0.05
noid - % of drinkers at college
 that report alcohol can be
 obtained at off campus bar,
 campus pub, or liquor store
 without ID, respondent excluded 0.49 0.16 0.51 0.17
dadever = 1 if father ever
 consumes alcohol 0.73 0.45 0.74 0.44
momever = 1 if mother ever
 consumes alcohol 0.59 0.49 0.63 0.48
dadheavy = 1 if father heavy
 consumer of alcohol 0.09 0.29 0.13 0.34
momheavy = 1 if mother heavy
 consumer alcohol 0.02 0.14 0.03 0.17
strongrelig = 1 if religion is
 important/very important
 to respondent 0.26 0.44 0.31 0.46
Catholic = 1 if respondent
 indicated were raised in
 Catholic religion, relative
 to no religious background 0.35 0.48 0.35 0.48
Jewish = 1 if respondent indicated
 were raised in Jewish religion,
 relative to no religious
 background 0.02 0.18 0.03 0.18
Moslem = 1 if respondent indicated
 were raised in Moslem religion,
 relative to no religious
 background 0.01 0.11 0.01 0.08
Protestant = 1 if respondent
 indicated were raised in
 Protestant religion, relative
 to no religious background 0.34 0.47 0.35 0.48
Other= 1 if respondent indicated
 were raised in another religion,
 relative to no religious
 background 0.10 0.30 0.12 0.32
famdisp = 1 if family
 disapproves of drinking 0.18 0.38 0.19 0.39
Among Drinkers:
Fraction who usually drink beer 0.67 0.47 0.37 0.48
Fraction who usually drink wine 0.03 0.16 0.11 0.31
Fraction who usually drink
 wine coolers 0.04 0.20 0.10 0.30
Fraction who drink mixed liquor 0.16 0.36 0.28 0.45
Fraction with no usual drink 0.10 0.30 0.14 0.35

N 13,080 18,079

Source: Author's calculations from 1993, 1997 College Alcohol Surveys

TABLE 2
Selected Determinants of Drinking, By Gender (Coefficients/standard
error)

 Males

 1993 & 1997 data

Variable Std.
 [beta] Error

Binge (N=12,876 & 5,475 for males, 16,941 & 8,119 for females)

avgprice
fixfee
dmkfree
noid 0.10 0.11
noidu21 0.17 0.16
famdisp 0.03 0.04
momever 0.13 0.03 *
dadever 0.08 0.03 *
momheavy -0.02 0.09
dadheavy 0.03 0.04
strongrelig -0.40 0.03 *
percbinge 1.26 0.13 *
hsbinge 0.86 0.03 *
catholic 0.25 0.04 *
jewish 0.07 0.07
moslem -0.24 0.14 ([dagger])
protestant 0.04 0.04
other 0.09 0.05
Durbin-Wu-
Hausman test 4.42 P-value = 0.00
log likelihood -6472.7

Drunk (N=10,799 & 4,881 for males 14,526 & 7001 for females)

avgprice
fixfee
dmkfree
noid 0.42 0.17 *
noidu21 0.72 0.03 *
famdisp 0.03 0.04
momever 0.15 0.03 *
dadever 0.06 0.03 ([dagger])
momheavy -0.02 0.09
dadheavy -0.03 0.05
strongrelig -0.36 0.03 *
percdrunk 1.25 0.12 *
hsbinge 0.69 0.03 *
catholic 0.14 0.04 *
jewish 0.09 0.08
moslem -0.67 0.18 *
protestant -0.01 0.04
other -0.03 0.06
Durbin-Wu-
Hausman test 5.54 P-value = 0.019
Log likelihood -5937.18

 Males

 1997 data only

Variable Std.
 [beta] Error

Binge (N=12,876 & 5,475 for males, 16,941 & 8,119 for females)

avgprice -0.11 0.10
fixfee 0.23 0.43
dmkfree -0.85 0.36 *
noid 0.12 0.17
noidu21 0.22 0.24
famdisp 0.04 0.05 *
momever 0.12 0.04
dadever 0.07 0.05
momheavy -0.12 0.14
dadheavy -0.004 0.07
strongrelig -0.55 0.06 *
percbinge 1.30 0.25 *
hsbinge 0.92 0.04 *
catholic 0.23 0.05 *
jewish 0.09 0.11
moslem -0.25 0.23
protestant 0.03 0.06
other 0.11 0.06
Durbin-Wu-
Hausman test 4.42 P-value = 0.00
log likelihood -2858.4

Drunk (N=10,799 & 4,881 for males 14,526 & 7001 for females)

avgprice 0.03 0.10
fixfee -0.69 0.45
dmkfree -0.08 0.37
noid 0.80 0.25 *
noidu21 0.70 0.04 *
famdisp 0.08 0.05
momever 0.14 0.05 *
dadever 0.13 0.05 *
momheavy 0.05 0.15
dadheavy -0.07 0.07
strongrelig -0.54 0.06 *
percdrunk 1.58 0.22 *
hsbinge 0.78 0.05 *
catholic 0.08 0.06
jewish 0.17 0.12
moslem -0.97 0.33 *
protestant -0.04 0.06
other 0.0003 0.07
Durbin-Wu-
Hausman test 5.54 P-value = 0.019
Log likelihood -2615.42

 Females

 1993 & 1997 data

Variable Std.
 [beta] Error

Binge (N=12,876 & 5,475 for males, 16,941 & 8,119 for females)

avgprice
fixfee
dmkfree
noid 0.03 0.10
noidu21 -0.01 0.13
famdisp 0.05 0.04
momever 0.04 0.03
dadever 0.06 0.03 ([dagger])
momheavy 0.005 0.03
dadheavy -0.001 0.03
strongrelig -0.27 0.03 *
percbinge 1.46 0.11 *
hsbinge 0.54 0.02 *
catholic 0.11 0.03 *
jewish -0.007 0.06
moslem -0.30 0.20
protestant -0.07 0.03 *
other -0.06 0.04
Durbin-Wu-
Hausman test 114.94 P-value = 0.00
log likelihood -9024.24

Drunk (N=10,799 & 4,881 for males 14,526 & 7001 for females)

avgprice
fixfee
dmkfree
noid -0.06 0.13
noidu21 0.77 0.02 *
famdisp 0.004 0.03
momever 0.07 0.03 *
dadever 0.07 0.03 *
momheavy 0.10 0.07
dadheavy -0.07 0.04 ([dagger])
strongrelig -0.3 0.03 *
percdrunk 1.47 0.11 *
hsbinge 0.55 0.03 *
catholic 0.10 0.04 *
jewish -0.08 0.07
moslem -0.34 0.19 ([dagger])
protestant -0.02 0.04
other -0.11 0.05 *
Durbin-Wu-
Hausman test 5.45 P-value = 0.020
Log likelihood -7982.93

 Females

 1997 data only

Variable Std.
 [beta] Error

Binge (N=12,876 & 5,475 for males, 16,941 & 8,119 for females)

avgprice 0.03 0.07
fixfee -0.11 0.31
dmkfree -0.10 0.14
noid -0.16 0.14
noidu21 0.08 0.17
famdisp 0.08 0.04 *
momever 0.03 0.04
dadever -0.01 0.04
momheavy -0.03 0.09
dadheavy -0.05 0.05
strongrelig -0.33 0.05 *
percbinge 1.47 0.19 *
hsbinge 0.47 0.03 *
catholic 0.06 0.04
jewish 0.04 0.09
moslem -0.22 0.26
protestant -0.09 0.05 ([dagger])
other -0.07 0.05
Durbin-Wu-
Hausman test 114.94 P-value = 0.00
log likelihood -4288.79

Drunk (N=10,799 & 4,881 for males 14,526 & 7001 for females)

avgprice -0.03 0.08
fixfee 0.36 0.35
dmkfree 0.11 0.24
noid -0.19 0.19
noidu21 0.85 0.03 *
famdisp 0.03 0.04
momever 0.11 0.04 *
dadever 0.03 0.04
momheavy -0.03 0.10
dadheavy -0.08 0.05
strongrelig -0.34 0.05
percdrunk 1.62 0.17 *
hsbinge 0.56 0.04 *
catholic 0.09 0.05 ([dagger])
jewish -0.16 0.09 ([dagger])
moslem -0.40 0.29
protestant -0.04 0.05
other -0.35 0.05 *
Durbin-Wu-
Hausman test 5.45 P-value = 0.020
Log likelihood -3646.16

Source: Author's calculations from 1993 & 1997 Harvard College
Alcohol Studies.

* / ([dagger]) Statistically significant at the 5% / 10% level.
Other variables included in regressions but not shown: year of
survey, age, indicators for type of school, religious affiliation,
class, living arrangements, under 21, greek status, working
status, participate in exercise, athletic participation, and race.

TABLE 3A
Effect of Drinking on Study Hours:
Selected Results from GMM Simultaneous Equations Estimates

 Males

 1993 & 1997 data 1997 data only

 Std. Std.
 [beta] error [beta] error

binge 0.03 0.10 -0.28 0.13 *
hsbinge -0.30 0.14 * 0.17 0.18
avgstdyhrs 0.21 0.02 * 0.26 0.04 *
avggpa 0.05 0.05 0.05 0.07

Difference in male/female drinking coefficient estimates

Objective x N 97.41 39.63

drunk -0.17 0.18 -0.60 0.24 *
hsbinge -0.09 0.25 0.71 0.34 *
avgstdyhrs 0.35 0.04 * 0.46 0.08 *
avggpa 0.23 0.09 * 0.15 0.15

Difference in male/female drinking coefficient estimates

Objective x N 62.62 26.74 *

 Females

 1993 & 1997 data 1997 data only

 Std. Std.
 [beta] error [beta] error

binge 0.59 0.12 * 0.42 0.15 *
hsbinge -0.92 0.14 * -0.58 0.16 *
avgstdyhrs 0.18 0.02 * 0.24 0.03 *
avggpa 0.04 0.04 0.03 0.05

Difference in
 male/female
 drinking coefficient
 estimates -0.56 0.16 * -0.70 0.14 *

Objective x N 25.92 44.66

drunk 0.68 0.17 * 0.62 0.21 *
hsbinge -1.13 0.22 * -1.07 0.29 *
avgstdyhrs 0.19 0.04 * 0.18 0.06 *
avggpa 0.04 0.07 -0.004 0.09

Difference in
 male/female
 drinking coefficient
 estimates -0.85 0.17 * -1.22 0.22 *

Objective x N 28.83 26.32 *

TABLE 3B
Effects of Drinking on GPA:
Selected Results from GMM Simultaneous Estimations

 Males

 1993 & 1997 data 1997 data only

 Std. Std.
 [beta] error [beta] error

binge -0.52 0.18 * -0.86 0.22 *
hsbinge -0.19 0.24 0.12 0.27
studyhours 0.23 0.02 * 0.22 0.03 *

Difference in male/female drinking coefficient estimates

Objective X N 237.92 155.38
drunk -0.58 0.18 * -0.60 0.19 *
hsbinge -0.18 0.23 -0.23 0.25
studyhours 0.12 0.01 * 0.14 0.02 *

Difference in male/female coefficient estimates

Objective X N 253.46 186.44

 Females

 1993 & 1997 data 1997 data only

 Std. Std.
 [beta] error [beta] error

binge -0.60 0.17 * -0.61 0.22 *
hsbinge -0.28 0.20 -0.43 0.24 ([dagger])
studyhours 0.12 0.02 * 0.10 0.03 *

Difference in
 male/female
 drinking coefficient 0.08 0.17 -0.25 0.22

Objective X N 285.89 228.93
drunk -0.33 0.13 * -0.28 0.15 ([dagger])
hsbinge -0.60 0.16 * -0.68 0.21
studyhours 0.07 0.01 * 0.10 0.02 *

Difference in
 male/female
 coefficient estimates -0.25 0.15 -0.32 0.17 ([dagger])

Objective X N 412.41 251.57

* / ([dagger]) Statistically significant at the 5% / 10% level.

Source: Author's calculations from Harvard School of Public
Health, College Alcohol Surveys, 1993-1997.

TABLE 4
The Predicted Impact of Binge Drinking on Study Hours and GPA

 Non- Marginal
 bingers Bingers Effect
 of Binging

 FEMALES

Average Predicted Logged
 Study Hours 0.84 1.43 0.59
Average Predicted Study Hours 2.32 4.18 1.86

Total Effect on GPA (including indirect effect through study hours)

Pr. A grade point average 0.39 0.22 -0.17
Pr. B grade point average 0.52 0.59 0.07
Pr. C grade point average 0.08 0.18 0.10
Pr. D grade point average 0.005 0.002 -0.003

Predicted GPA 3.29 3.01 -0.28

Predicted Percent Change in
 Weekly Earnings ([double
 dagger]) -0.042
Predicted Loss in Weekly Wages
 ([dagger]) $35.83
Predicted Loss in Annual Wages
 ([psi]) $1,863.03

 MALES

Average Predicted Logged
 Study Hours 0.91 0.94 0.03
Average Predicted Study Hours 2.48 2.56 0.08

Total Effect on GPA (including indirect effect through study hours)

Pr. A grade point average 0.35 0.19 -0.16
Pr. B grade point average 0.55 0.59 0.04
Pr. C grade point average 0.10 0.21 0.11
Pr. D grade point average 0.0004 0.002 0.0016

Predicted GPA 3.25 2.95 -0.30

Predicted Percent Change in
 Weekly Wages ([double dagger]) -0.00466
Predicted Loss in Real Weekly
 Wages ([dagger]) $8.51
Predicted Loss in Annual Wages
 ([psi]) $276.68

([double dagger]) Based on change in expected GPA and coefficient
estimates of effect of GPA on earnings from Loury (1997).

([dagger]) Figured at median weekly earnings in 2004 for
college graduates of $860 for females and $1143 for males from

"Table 8. Quartiles and selected deciles of usual weekly earnings
of full-time wage and salary workers by selected
characteristics, 2004 annual averages," Bureau of Labor Statistics
web accessed at http://www.bis.gov/news.release/
wkyeng.t08.htm, September 20, 2005.

([psi]) Weekly earnings loss x 52 weeks per year.

TABLE 5
Estimated Effect of Binge Drinking on Grades Under
Various Tests for Bias Due to Institutional
Variations in Grading Practices

 Males Females

 Std. Std.
Specification Check [beta] Error [beta] Error

1993 & 1997 Data (does not include monetary price of alcohol)

Test A -0.49 0.19 * -0.57 0.17 *
Test B -0.37 0.10 * -0.45 0.09 *
Test C -0.36 0.04 * -0.47 -0.04 *

1997 Data (includes monetary price of alcohol)

Test A -0.82 0.12 * -0.57 0.23 *
Test B -0.41 0.09 * -0.56 0.12 *
Test C -0.13 0.02 * -0.08 0.01 *

Source: Author's calculations from Harvard School of Public Health,
College Alcohol Surveys, 1993-1997.

* Statistically significant at the 5% level.

Test A: Specification is an ordered probit, dependent variable is
GPA letter category. These coefficients are directly comparable
to the base specification in Table 3b. Because the specifications
are ordered probits, the coefficient does not represent
a marginal effect of drinking on GPA.

Test B: Dependent variable is the difference in the respondent's
linearized GPA and the college average GPA. The specification
is a linear regression using the same instrumental variables
specification. The coefficients represent changes in
GPA relative to the college average and are also not directly
marginal effects on GPA.

Test C: Dependent variable is the respondent's linearized GPA.
The specifications are instrumental variable, fixed effects
linear regressions, with fixed effects for each college. These
coefficients do represent marginal effects of binge
drinking on GPA.