xml version 1.0 encoding UTF8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchemainstance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam Ka
controlfield tag 001 001915091
003 fts
005 20071031132737.0
006 med
007 cr mnuuuuuu
008 071031s2006 flu sbm 000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14SFE0001775
040
FHM
c FHM
035
(OCoLC)180122253
049
FHMM
090
HF71 (ONLINE)
1 100
Austin, Wesley A.
0 245
Adolescent alcohol use and educational outcomes
h [electronic resource] /
by Wesley A. Austin.
260
[Tampa, Fla] :
b University of South Florida,
2006.
3 520
ABSTRACT: There is some controversy over whether adolescent alcohol use has deleterious causal effects on educational outcomes. In particular, does drinking reduce academic performance and school enrollment rates and increase truancy, or does the observed negative correlation between drinking and educational outcomes merely reflect common unobservable factors? This dissertation sheds further light on the issue by estimating the causal impacts of alcohol use on various educational outcomes. Specifically, an instrumental variables model is estimated to study the effects of several drinking measures on grades, school enrollment and absenteeism.
502
Dissertation (Ph.D.)University of South Florida, 2006.
504
Includes bibliographical references.
516
Text (Electronic dissertation) in PDF format.
538
System requirements: World Wide Web browser and PDF reader.
Mode of access: World Wide Web.
500
Title from PDF of title page.
Document formatted into pages; contains 90 pages.
Includes vita.
590
Adviser: Gabriel Picone, Ph.D.
653
Economics.
Health.
Drinking.
Grades.
Human capital.
690
Dissertations, Academic
z USF
x Business Administration
Doctoral.
773
t USF Electronic Theses and Dissertations.
4 856
u http://digital.lib.usf.edu/?e14.1775
PAGE 1
Adolescent Alcohol Use and Educational Outcomes by Wesley A. Austin A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy Department of Economics College of Business Administration University of South Florida CoMajor Professor: Jeffrey DeSimone, Ph.D. CoMajor Professor: Gabriel Picone, Ph.D. Don Bellante, Ph.D. Richard Smith, Ph.D. Date of Approval: August 17, 2006 Keywords: economics, health, drinking, grades, human capital Copyright 2006, Wesley A. Austin
PAGE 2
i Table of Contents List of Tables................................................................................................................. ....iv Abstract....................................................................................................................... .........v Chapter One: Introduction...................................................................................................1 Chapter Two: Background and Literature Review..............................................................7 Human Capital Theory.............................................................................................7 A. Completed Schooling..........................................................................................9 B. Grade Point Average.........................................................................................12 C. Schoolrelated Behaviors..................................................................................15 Chapter Three: Research Methodology.............................................................................17 Chapter Four: Data and Empirical Specification...............................................................24 Data Description....................................................................................................24 Empirical Specification..........................................................................................28 Chapter Five: Empirical Results........................................................................................34 A. Drinking and Grades.........................................................................................38 First Stage Regression Results...................................................................39 The Effects of Drinking on the Proba bility of Obtaining an A ..............41 Instrument Robustness and the Pr obability of an A Average.................42 The Effect of Drinking on the Probability of a C or lower average........44 Instrument Robustness and the Pr obability of a C Average....................46
PAGE 3
ii B. Drinking and School Enrollment......................................................................47 First Stage Regression Results...................................................................48 The Effects of Drinking on School Enrollment (1619 sample) ...............50 Instrument Robustness and School Enrollment (1619 sample) ..............52 The Effects of Drinking on School Enrollment (1825 sample) ..............53 Instrument Robustness and School Enrollment (1825 sample)................55 C. Drinking and Absenteeism................................................................................56 First Stage Regression Results...................................................................56 The Effects of Drinking on Absenteeism (1619 sample) ........................59 Instrumental Variable Robustness and Absenteeism (1619 sample)........61 The Effects of Drinking on Absenteeism (1825 sample).........................63 Instrumental Variable Robustness and Absenteeism (1825 sample)........64 Chapter Six: Summary and Conclusions...........................................................................67 Limitations.................................................................................................70 Policy Implications....................................................................................71 References..........................................................................................................................74 Appendices.........................................................................................................................78 Appendix 1: Probit estimates for the pr obability of A and C or below............79 Appendix 2: All IV estimates on th e probability of an A for binge drinking..................................................................................................................80 Appendix 3: All IV estimates on the probability of a C or lower for binge drinking..................................................................................................................81 Appendix 4: Probit estimates for enrollment (1825 years old)............................82
PAGE 4
iii Appendix 5: Probit estimates for enrollment (1619 years old)............................83 Appendix 6: All IV estimates on the probability of enrollment for binge drinking (1619 sample).........................................................................................84 Appendix 7: All IV estimates on the probability of enrollment for binge drinking (1825 sample).........................................................................................85 Appendix 8: Probit estimates for absenteeism (1825 years old)..........................86 Appendix 9: Probit estimates on absenteeism (1619 years old)...........................87 Appendix 10: All IV estimates on absenteeism for binge drinking (1619 sample)...................................................................................................................88 Appendix 11: All IV estimates on absenteeism for binge drinking (1825 sample)...................................................................................................................89 About the Author...................................................................................................End Page
PAGE 5
iv List of Tables Table 1: Descriptive statistic s (Youth Experience sample) ..............................................36 Table 2: Descriptive sta tistics (1825 years old)................................................................37 Table 3: Descriptive sta tistics (1619 years old)................................................................38 Table 4: Probit/ First stage regressi on estimates for grade outcomes................................39 Table 5: IV/ OLS estimates of drin king on the probability of an A................................41 Table 6: IV estimates of drinking on the probability of an A using IV pairs..................43 Table 7: IV/ OLS estimates of drinking on the probability of a C or below...................44 Table 8: IV estimates of drinking on th e probability of a C using IV pairs....................46 Table 9: Probit/ First stage estimates for school enrollment (1825 years old).................48 Table 10: Probit/ First stage estimates fo r school enrollment (1619 years old)...............49 Table 11: IV/ OLS estimates of drinki ng on school enrollment (1619 years old)...........51 Table 12: IV estimates of drinking on enrollment using IV pairs (1619 years old).........53 Table 13: IV/ OLS estimates of drinking on school enrollment (1825 years old) ..........54 Table 14: IV estimates of drinking on enrollment using IV pairs (1825 years old).........55 Table 15: Probit/ First stage estimates for absenteeism (1825 years old)........................57 Table 16: Probit/ First stage estimates for absenteeism (1619 years old)........................58 Table 17: IV/ OLS estimates of drinking on absenteeism (1619 years old).....................60 Table 18: IV estimates of drinking on ab senteeism using IV pa irs(1619 years old)........62 Table 19: IV/ OLS estimates of drin king on absenteeism (1825 years old).....................63 Table 20: IV estimate of drinking on absent eeism using IV pairs (1825 years old)........66
PAGE 6
v Adolescent Alcohol Use a nd Educational Outcomes Wesley Austin ABSTRACT There is some controversy over whether adolescent alcohol use has deleterious causal effects on educational outcomes. In particular, does drinking reduce academic performance and school enrollment rates a nd increase truancy, or does the observed negative correlation between drinking and e ducational outcomes me rely reflect common unobservable factors? This dissertation sheds further light on the i ssue by estimating the causal impacts of alcohol use on various educational outcomes. Specifically, an instrumental variables model is estimated to study the effects of several drinking measures on grades, school enrollment and absenteeism.
PAGE 7
1 Chapter One: Introduction In many healthrelated and social scien ce fields, there has long been concern about various harmful effects of alcohol use. One specific potential consequence in which economists have recently shown interest is the reduction of human capital accumulation. This issue is particularly re levant during adolescence and early adulthood, during which decisions regarding high school completion and college attendance are first considered and academic performance realiza tions that affect l ongerterm educational and economic outcomes are observed. Excessive drinking is associated with this age group despite its illegality until the age of 21. For instance, the 2003 National Survey on Drug Use and Health (NSDUH) reports that bi nge drinking, i.e. the consumption of at least five alcoholic beverages in one episode, occurred in the past month among 17 percent of high school students and 35 percent of college students. There are several ways in which heavy drinking could potentially impair human capital formation. Intoxicati on could interfere with class attendance and learning, and the time spent in activities where drinking occurs could substitute away from time allocated to studying. This could hur t academic performance in the short term, which might diminish the ability or in centive to continue schooling over the longer term. Risks stemming from intoxication such as injury from accidents or fights, pregnancy and disease from unsafe sex, conflicts with parents or the law, and a tarn ished reputation with school authorities might also limit the capabilit y or motivation of a student to remain in school (Cook and Moore 1993). Alternatively, social interactions associated with
PAGE 8
2 drinking might improve academic achievemen t by providing a means of relieving stress (Williams et al. 2003). Why is the potential impact of alcohol use on educational outcomes relevant for the discipline of economics? Substantivel y, human capital accumulation bears directly on the fields of labor economics, given that estimating the returns to schooling has been one of the most prominent endeavors in the fi eld; health economics, within which a large literature on the complex relationship between schooling and health has evolved; public economics, since education is the archetypal example of a good that conveys a positive externality; and macroeconomics, because of the importance of human capital to economic growth. Moreover, investigating the various causes and consequences of alcohol use has been a foundational topic in health economics, and understanding the impacts of alcohol policies such as excise taxes, minimum legal drinking ages (MLDAs) and zerotolerance laws is perennially of great pub lic policy concern. Perhaps even more importantly, economics is relevant because estimating the effect of drinking on educational outcomes is inherently an empirical matter for which the tools of econometrics can be effectively us ed. In particular, much evidence, ranging from anecdotal to academic research in disc iplines outside of economics, has established a strong negative relationship between the re gularity and intensity of drinking and human capital measures such as educational attain ment and academic performance. But from a public policy perspective, distinguishing whether this rela tionship is causal, such that increased alcohol consumption directly reduces completed schooling or lowers grades in school, or merely correlational, with changes in other confounding variables simultaneously leading to drinking and wors e educational outcomes, is critical. If
PAGE 9
3 frequent or heavy drinking truly compro mises academic achievement, programs that decrease such alcohol use should have both the private and social valu es of the resulting educational gains included as benefits when optimal budget allocations are calculated. This is not true if the association betw een drinking and poor school performance is merely spurious. Obtaining an estimate of the magnitude of the causal effect that alcohol use has on educational outcomes should thus be of inte rest to economists as well as those who analyze or participate in the formation of public policies. The role of econometrics is to identify this causal effect in the context of a broader relationship in which various third factors that are difficult to measure might create an inverse covariance between alcohol use and human capital accumulation, or whet her academic performance might have a reverse causal impact on drinking. This task is a natural one to tackle using econometric techniques, because one of the mainstays of econometrics is instrumental variables (IV) regression, a method specifically designed to esti mate the causal impact of a variable that does not necessary otherwise vary independent ly with other unobser ved determinants of the outcome being examined. Yet, only within the past 15 years has the relationship betw een alcohol use and human capital accumulation been addressed by economists, and even during that time research on the topic had been limited in both quantity and scope. In terms of quantity, fewer than a dozen relevant studies have b een published in economics journals, many do not use IV or any other method to specifically account for the possibility that drinking is endogenous in educational outcome equations and even some that do have used approaches that have since been criticized as unconvincing. Rega rding scope, most of
PAGE 10
4 the literature has focused on completed schooling, with little attenti on paid to academic performance, especially among precollege students, and other aspects of inschool behavior. Furthermore, while much research has found evidence that at least part of the negative relationship between alcohol use and ac hievement represents a causal impact of drinking, a couple of recent investigations that were quite thorough have disagreed. This dissertation will estimate the cau sal impact of various frequencies and intensities of drinking among adolescents and yo ung adults on three sets of human capital accumulation measures. In the results chapte r, I begin by examining the impact on recent academic performance, specifically the probabi lity of obtaining an A as well as the probability of receiving a C or lower grade. Grades are an important intermediate outcome that are related to longerterm labor market experiences through their impact on both the quantity and quality of schooling recei ved, but have largely been ignored in the literature. Next, I analyze the effects on school enrollmen t. This is a commonly examined education outcom e among both related studies and broader literatures on human capital accumulation, given that school attendance is easily measured and has a clear marginal impact on future wages th at labor economists have long focused on estimating. Finally, I investig ate effects on truancy, i.e. cl asses missed due to skipping and illness, which has not been widely studied despite clearly affec ting the acquisition of human capital. A major contribution of my research is th e use of data that have not previously been examined in the litera ture. The aforementioned NSDU H contains extensive annual data on alcohol use and the types of edu cational outcomes outlined above among large nationally representative samples of U.S. residents aged 12 and above. This survey,
PAGE 11
5 which in different permutations dates back to 1979, is a primary source of information on U.S. substance use trends from which data are disseminated widely among both researchers and the media, and have been us ed by several studies on various aspects of drug use that have been published in economics journals. However, none of the previous research on the relationship between drinking and human capital accumulation has utilized these data. Data from the NSDUH allow for both breadth and depth of coverage on the topic. Breadth comes from the ability to study aspect s of all three types of educational outcomes outlined above using data from an elaborate questionnaire administered to 12 year olds on a wide array of youth experiences relating to education and alcohol use, and questions asked of older respondents that pe rtain to schooling and drinking behaviors. Depth is provided by additional questions on education and drinking intensity, which allow for a more thorough anal ysis of the main outcome variables this dissertation addresses. An equally important facet of the NSDUH da ta is that they are conducive to the use of the IV regression me thodology to estimate the causal effect of alcohol use on human capital accumulation. Abundant inform ation is collected on preferences and experiences related to alcohol consumption, including measures of parental disapproval, peer use, and the perceived risks involved, as well as religious sentiment. An assortment of variables are therefore observed that have the potential to serve as instruments for drinking in educational outcome equations, in th e sense that they are very likely to be highly correlated with alcohol us e but would not have any obvi ous reason to be otherwise associated with educational outcomes. Current econometric techniques that are
PAGE 12
6 straightforward to implement will be applied to empirically test for the appropriateness of these identification restrictions Moreover, instrumental variab les models are estimated in the context of endogenous drinking measures th at take the form of binary indicators. The remainder of this dissertation is stru ctured as follows. Chapter 2 offers an overview of the relevant literature and is di vided into three parts, each corresponding to one of the distinct types of educational out comes under investigation. Chapter 3 explains the research methodology employed to obtain estimates that convincingly represent causal effects of drinking despite its proba ble endogeneity in the academic achievement equations. Chapter 4 describes the NSDUH da ta that are analyzed and describes the empirical specification. Chapter 5 offers es timation and specification test results and chapter 6 concludes by highligh ting particularly relevant findings and outlining some implications for alcohol policy.
PAGE 13
7 Chapter Two: Background and Literature Review This chapter begins by discussing the human capital theory upon which my dissertation is based. It then provides details regarding previous studies on the topic of alcohol use and educational outcomes, concen trating on those appearing in the economics literature. First it covers research on comp leted schooling, the most often investigated human capital consequence of drinking. The chapter then proceeds to discuss academic performance, which has been addressed by only a handful of studies, most of which focus on students in college rather than high school The review conclude s with schoolrelated behaviors, which have largely been ignored de spite their clea r potential to be impacted by the consumption of alcohol. Human Capital Theory Human capital theory asserts that increases in a persons stock of knowledge, i.e. human capital, raises that i ndividuals productivity in the market sector of the economy, where he/she produces money earnings, and in the nonmarket or household sector, where he/she produces commodities that enter the utility function. To obtain gains in productivity, which raises subsequent earnings, a person must invest in formal schooling. Several seminal studies (Becker (1964), BenPorath (1967), Grossman (1972) and Mincer (1974) have well developed this theory and its implication. Wit hout loss of generality this section only provides an overview of Gr ossman (1972) and Mincer (1974). Grossman (1972) argues that health capit al is an important component of an individual utility function. In his model, an individuals stock of knowledge influences
PAGE 14
8 her market and nonmarket productivity, while he r stock of health de termines the total amount of time she can spend producing mone y earnings and commodities. An individual is assumed to inherit an initial stock of h ealth that depreciates with age and can be prolonged by health investments. Education may increase the health stock if more educated people produce more health. Alc ohol consumption may impact health by direct negative effects on cognitive functio ning and other bodily da mage, or by reducing educational acquisition and thus making production of health less efficient. Theoretically, alcohol use can reduce huma n capital formation in two respects: 1) the direct negative impact of alcohol c onsumption on cognitive functioning and other health measures, 2) the time and effort devoted to obtaining and consuming alcohol, which takes time away from activities that augment human capital. Generally, if drinking has negative health conse quences, the resulting reducti on in human capital lowers productivity and therefore ear nings and overall utility. According to Mincer (1974), the positive relationship between market earnings and human capital investments provide in centives for obtaining higher levels of education. The theoretical specification of the model suggests that it is human capital utilized during working hours that generates ea rnings. Mincer expresses earnings directly as a function of years of schooli ng completed and experience. Accumulation of human capital is subj ect to optimization, given costs of acquisition and returns to human capital investment. Al cohol use could serve to reduce the rate of return to human capital investme nts. Thus, the overall optimal level of human capital investment is reduced. One of the impli cations of the Mincer approach is that the larger the human capital investment, the higher the wage rate tends to be and the more
PAGE 15
9 rapid the rate of increase in wages over ones working lif e. Youth drinking therefore theoretically lowers the wage and its rate of increase over time. A. Completed Schooling Several articles published by economists have obtained estimated effects of alcohol use on educational achie vement, with measures of drinking and schooling as well as conclusions varying across studies. Comparatively early research produced evidence of a negative relationship, but either made no attempt to econometrically deal with the potential endogeneity of drinking in attainment equations, or did so in a way that has since been criticized as unsatisfactory, so that it is unclear whether this negative correlation indeed represents declines in co mpleted schooling that are caused by drinking. Two more recent and relatively thorough studies found that the causal impact of alcohol use on educational attainment is either small or nonexistent. However, another recently completed and equally thorough analysis di sagrees, finding evidence of a sizable reduction in the probability of high school comple tion attributable to previous frequent or heavy alcohol consumption. The first study to appear in the liter ature, Cook and Moore (1993), analyzed National Longitudinal Survey of Youth (N LSY) data on the 753 members of the two youngest cohorts (those ages 14 or 15 in 1979) who were enrolled in 12 th grade as of the 1982 interview. They estimated IV models in which the effect of current alcohol use on postsecondary attainment was id entified by the state excise tax on beer and an indicator for whether the student could legally drink based on the states MLDA. Results from three separate specifications s howed that heavy drinking in 12 th grade decreased subsequent schooling, by 0.13 years for each drink consumed in the preceding week, 2.3
PAGE 16
10 years for students drinking on at least two occasions in the previous week, and 2.2 years for respondents who had at least six drinks on at least four occasions in the preceding month. Direct regressions of educational attainment on the alcohol policy measures similarly implied that students from states with higher beer taxes continued further in school and were more likely to graduate from college. Mullahy and Sindelar (1994), using ordinary least squares (OLS) regressions in data on males from Wave 1 of the New Haven site of the Epidemiological Catchment Area survey, found that the onset of alcoholis m symptoms by age 22 was associated with a five percent reduction in completed schoo ling. The authors emphasize the typically overlooked role that this advers e impact of drinking on educat ional attainment, if causal, would have in indirectly reduc ing the incomes of alcoholics. Yamada et al. (1996) similarly analyzed data on NLSY respondents who were 12 th graders during the 1981 academic year using single equation probit models that did not account for the possibility that alcohol use is endogenous. They estimated that the probability of hi gh school graduation was 6.5 percent lower for students who consumed alcohol on at least two occasions in the previous week and 2.0 percent lower for those who drank wine or liquor. In addition, drinking was found to be inversely related to beer taxes, liq uor prices, MLDAs and marijuan a decriminalization, meaning that each was positively associated with high school gradua tion rates through its covariance with alcohol use. Koch and Ribar (2001) examined the rela tionship between age of drinking onset and educational attainment by age 25 in data on approximately 650 samesex sibling pairs of each gender from the 1979 waves of the NLSY. Estimates from an IV
PAGE 17
11 model that specified sibling onset age as the instrument for res pondent onset age imply that delaying alcohol initiation by a year increases subsequent schooling by 0.22 years regardless of gender. However, they argued that this represents an upper bound for the effect size based on the sign of the bias if th e assumptions needed for consistency are not met, and indeed OLS and family fixed effect s models produce estimates that are three to four times smaller for males, and still smaller and sometimes insignificant for females. Dee and Evans (2003) called into question the causal effect in terpretation of the results from Cook and Moore (1993), arguing th at the use of crossstate alcohol policy variation to identify the eff ects of drinking on other outcomes in an IV framework is potentially problematic because such varia tion might be correlated with unobservable attributes that affect both alcohol use and th e outcome measure, in this case educational attainment. They estimated models that include state fixed effects in order to isolate the effects of withinstate policy variation on drinking. In pooled cross sections from the 1977 Monitoring the Future (MTF) surveys, alcohol use declined when MLDAs increased, but did not respond to beer tax ch anges. Moreover, in 1990 Census data on over a million members of the 1960 birth cohorts, not only did educational attainment fail to rise after MLDAs were in creased, but twosample IV estimates of the effects of drinking on high school completi on, college entrance and college completion were all small and insignificant. The most recent evidence on the subject comes from Chatterji and DeSimone (2005), who estimated the effect of binge and frequent drinking by adolescents on subsequent high school dropout in data from the NLSY Young Adults using an IV model with an indicator of any past month alcohol use as the identifying in strument, while also
PAGE 18
12 control for a wide array of potentially confounding variables in cluding maternal characteristics and dropout risk factors. In contrast to the last two studies cited above, the authors found that OLS yielded conservative estimates of the causal impact of heavy drinking on dropping out, such that binge or frequent drinking among 15 year old students lowered the probability of having graduated or being enrolled in high school four years later by at least 11 percent. The results of overi dentification tests using two measures of maternal youthful alcohol use as additional instruments provided support for their empirical strategy. B. Grade Point Average The only previous study that attempts to identify the causal impact of drinking on academic performance among precollege students is DeSimone and Wolaver (2004), who analyze 2001 and 2003 Youth Risk Behavi or Survey (YRBS) data on high school students. They estimated regressions that in cluded proxies for myriad potential sources of unobserved heterogeneity, part icularly risk and time prefer ence, mental health, selfesteem, and tastes for substance use. Estimated effects of alcohol use on grades are substantially reduced in magnitude when these additional covari ates are added, but typically remain significantly negative. The impact on the extensive margin impact (i.e. whether or not drinking occurred) was over tw ice as large for binge drinking than for nonbinge drinking, and binge drinking also has effects on the intensive margin, in terms of consumption frequency, that nonbinge drinking did not. Drinkingrelated grade reductions were larger among st udents who are observably more risk averse and futureoriented, and effects on several outcomes with which drinking is likely associated in a noncausal way were insignificant.
PAGE 19
13 Two additional studies estimated that heavy drinking reduced grades among college students surveyed in Harvards College Alcohol Study (CAS). In the first wave of the CAS, from 1993, Wolaver (2002) used generalized met hod of moments to estimate a threeequation IV model in which alc ohol consumption, study hours and academic performance are simultaneously determined. Instruments included measures of the ease of obtaining alcohol, parents drinking behaviors, family attitudes about drinking and religiosity (for alcohol use) and peer studying and drinking behavior (for study hours). Results indicated that heavy dr inking, in the form of any or frequent binge drinking or drunkenness during the previous month, reduced the probability of an A average by 12 to 18 percentage points, with commensurate in creases in the likelihood of receiving Bs and Cs. These mostly represented a direct impact on grades, though indirect effects through a decrease in study hours were also significant. Effects were larger for the underage than for students age 21 and above. However, overide ntification tests uniformly indicated that the exclusion restrictions as a whole were invalid, and drinking significantly affected neither study hours nor GPA in models that did not specify religiosity as an instrument. Williams et al. (2003) used data from the first three CAS waves, from 1993, 1997 and 1999, and used twostage least squares (2SLS) to estimate a similar threeequation IV model for drinking, GPA and study hours. Vari ables reflecting the full price of alcohol, including the beer tax, statelevel variable s related to access and opportunity to use and the costs of drinking and driv ing, and religiosity, served as instruments for past month alcohol use, which was measured as fre quency of consumption, number of drinks consumed per episode, or total number of drinks. Empirically, small positive direct effects of drinking on GPA were outweighed by larger negative indi rect effects that
PAGE 20
14 operated through reductions in study time. For instance, the estimates implied that an additional drink each day would directly raise GPA by 0.09, but indirectly lower GPA by 0.27 because of a 40 minute fall in daily study time. Again, though, overidentification statistics corresponding to the drinking equatio ns were always signifi cant at the 5 percent level. Also, the use of cro ssstate alcohol policy variati on to identify alcohol use is subject to the aforementioned Dee and Evans (2003) criticism. Kremer and Levy (2003) offered evidence on the topic from a natural experiment in which students at a large state univers ity were randomly assigned roommates through a lottery system. Males assigned to roommates who reported dr inking in the year prior to entering college had GPAs that were lo wer than those assigned to nondrinking roommates by onequarter of a point on average, onehalf of a point at the 10th percentile, and an average of twothirds of a point among those who drank frequently prior to college. These effects persisted over time and appear even more important in the context of the lack of any effect on GPA of roommates high school grades, admission test scores or family background. In contrast, prior drinking of roommates had no affect on female GPAs. One recent study from outside economics that warrants mention is Jeynes (2002), who examined a sample of 18,726 12 th graders from the 1992 National Education Longitudinal Survey. He found that two measur es of drinking were negatively related to achievement when simultaneously included in regression equations: increases of one standard deviation in ever having binge drank and ever having been drunk at school were associated with reductions in the composite sc ore from standardized tests in reading and math of 0.25 and 0.09 standard deviati ons, respectively. Although parental
PAGE 21
15 socioeconomic status, daily cigarette smoki ng, and ever having been under the influence of marijuana and cocaine at school were also held constant, race and gender were the only other variables included in the model, which was estimated by OLS. Thus, typical econometric standards for establis hing causation were not met. C. Schoolrelated behaviors Two studies concerning the impact of dr inking on the schoolrelated behaviors that have appeared in the economics literature and merit some attention. Roebuck et al. (2004) examined the likelihood of qu itting school and truancy among NHSDA respondents, albeit those in terviewed in 1997 and 1998, before questions on grades in school became part of the survey. In a sample of 15,168 12 year olds who had not yet completed high school, a probit regression show ed that those who consumed any alcohol over the previous year were 0.6 percent mo re likely to not be enrolled in school, representing a semielasticity of nearly 0.2 at the mean dropout rate of 3.1 percent, but a negative binomial regression f ound no relationship between days truant and any past year drinking among those enrolled. Although measures of illega l drug use were included in the models, because the focus of the study was on marijuana rather than alcohol, no attempt was made to account for the potential endogeneity of drinking. Markowitz (2001) estimated effects of the number of days the respondent drank and binge drank over the prior 30 days on fighting and weapons carrying in the 1991, 1993 and 1995 waves of the YRBS, using a 2SLS procedure in which three state level price measures, the beer tax, the cocaine pri ce and an indicator of whether marijuana is decriminalized, served as instruments. Analysis of these behavior variables has connecting implications for school attendance in that disciplinary sanctions for such
PAGE 22
16 actions may include suspension or expulsion from school. Her results showed that the probability of having been in a physical fight during the previous year rose by about six percentage points with each day of drinki ng and 11 points with each binge drinking day, but neither drinking variable was related to carrying a gun or other type other weapon in the past 30 days. An important caveat is th at the IV methodology is subject to the same criticism as that of Cook and Moore (1993), b ecause state fixed effect s were not included and thus crossstate price va riation contributed to identi fication. Indeed, when census division indicators were added, both drinking measures became negative and insignificant in the fighting equations, but significantly positiv e in the gun carrying equation, while the Fstatistics for the joint significance of the instruments fell from around four (significant though not particularly large) to below two and insignificant.
PAGE 23
17 Chapter Three: Research Methodology The purpose of this dissertation is to investigate whether alcohol use among adolescents and young adults causally influe nces various measures of human capital accumulation. In determining causation, the primary methodological question that must be resolved is whether drinking is properly specified as an exogenous variable with respect to educational outcomes, or should in stead be treated as endogenous. To frame the discussion, consider the following equatio ns, in which drinking (D) is a function of exogenous factors while educational attainment or achievement (E) is a function of some (but not all) of the same exoge nous determinants as well as D, (1) D = 0 + Z 1 + X 2 + (2) E = 0 + 1 D + X 2 + In the above equations, which apply to individual NSDUH respondents (with the corresponding observationlevel subscrip t suppressed), vect ors are in bold, X represents a set of exogenous variables that could aff ect both drinking and educational outcomes, Z represents another set of exogenous variables that coul d effect drinking but not educational outcomes (Z), and are error terms that encomp ass all factors influencing the drinking and educational out comes, respectively, that are not explicitly controlled for on the right hand side of the equations, and the s and s are parameters to be estimated by the regression analysis. Econometrically, alcohol use is exogenous in equation 2 if it is uncorrelated with the error term This condition holds, by definition, if none of the unobserved
PAGE 24
18 educational outcome determinants are related to drinking. If so, there is no need to estimate equation 1; a single equation regres sion method such as OLS or probit (in the case of a binary educational outcome measur e) will produce consistent estimates of the causal effect of drinking, 1 However, two sources of endogeneity co uld possibly lead to a nonzero correlation between alcohol use and the error term in (2). One is unobserved heterogeneity, which would occur if any of the unmeasured educational outcome determinants that are subsumed in the error term are correlated with alcohol use. The resulting estimate of 1 in (2) would suffer from omitted variable bi as, which cannot be eliminated directly because the omitted variables are not recorded in the data. Characteristics such as a lack of concern for the future relative to the pres ent, or a disruptive event such as parental divorce, might simultaneously be responsible for greater alcohol c onsumption and lower attainment or achievement. Because such factors cannot be observed, though, they are not held constant in the regression, and th e negative correlati on between drinking and educational outcomes that they induce becomes embedded into the alcohol use coefficient, which is thus biased negatively as an estimate of the causal drinking effect. Conversely, unmeasured ability or socio economic background could create a positive bias in the estimated drinking effect, if higher ability students are better able to function normally after alcohol consumption or st udents who have more money to spend on alcohol are also higher achievers. The other potential source of endogeneity is reverse causation. If alcohol use and educational outcomes are simultaneously de termined, in an econometric sense, educational achievement will not only be a f unction of drinking, as specified in equation
PAGE 25
2, but also will be a contributing factor to the decision regarding whether and how much alcohol to consume, a mechanism not accommodated by the above twoequation framework. In terms of equati on 2, shocks to the error term that, by definition, influence educational outcomes will ultimately extend to drinking through the feedback effect of educational outcomes on alcohol consumption, thus creating a correlation between alcohol use and that renders the estimate of the drinking effect 1 inconsistent. Again, the resulting bias could occur in eith er direction depending on the source of the reverse causation: it would be positive if acad emic success is celebrated by drinking or leads to additional income that is spent on al cohol consumption, but negative if alcohol is used to drown academic sorrows or if academic shortcomings reduce the opportunity cost of drinking. In order to investigate th e possibility that alcohol use is endogenous as an explanatory factor for educational outcomes (t o which I will refer to as achievement, for parsimony, but without loss of generality, for th e remainder of the section) and generate estimated effects of drinking on achievement that can be interpreted as causal, this analysis will use the method of instrumental variables (IV). To use IV, the vector Z in equation 1 above must exist, i.e. there must be at least one, and pr eferably two or more, variables (i.e. instruments or IV) that affect alcohol use but have no direct impact on achievement. In the case of exactly one instrument Z the IV method works by estimating the causal drinking effect 1 as the ratio of the sample correlation between the instrument and achievement to the sample co rrelation between the in strument and alcohol use, i.e. (3) D],rr[o cE],rr[o c 1Z Z 19
PAGE 26
where the ^ symbolizes that the quantity is estimated from the data and the correlations are estimated while holding constant the vector X of explanatory factors. The idea is that because the instrument is exogenous and re lated to achievement only through drinking, the sample correlation between the instrume nt and achievement is purely a product of that between drinki ng and achievement. Thus, the sample correlation between the instrument and achievement merely needs to be standardized by that between the instrument and drinking in order to be used as an estimate for the causal effect of drinking on achievement. In the ca se of two or more instruments, the linear projection of Z onto D, takes the place of Z in equation 3, in order to reduce the dimensionality so that both correlation expressions are scalars. In fact, this is true even in the case of a single instrument, but for expository purposes it is simpler to consider (3) as written above. D Equation 3 makes transparent the two impor tant conditions that the instrument vector Z must satisfy in order fo r IV to produce consistent estimates of the drinking effect 1 : the instruments must be highly correlat ed with alcohol use but not correlated with the educational outcome under investig ation through any other mechanism besides drinking. If the correlation between the in struments and drinking is not statistically significant, the denominator in (3) is statistically equal to zero, thus rendering the expression for 1 indeterminate. The strength of this correlation is eas ily judged from the Fstatistic for the joint significance of 1 in equation 1, which is equivalent to the tstatistic on the scalar 1 when there is a single instrument. Minimally, 1 should be significant at the 1 percent le vel; beyond this, Staiger and St ock (1997) advise the more stringent requirement that the associ ated Fstatistic be at least 10. 20
PAGE 27
Meanwhile, if a direct correlation between the instruments and achievement exists outside of the pathway from the instruments to drinking to achievement, the numerator in (3) includes variation that is not part of the relatio nship between drinking and achievement, and consequently the expression is no longer a consistent estimate of the effect of drinking on achievement. The reas on multiple instruments are preferred is that this overidentifies equation 2, which allows for specification tests to determine the empirical validity of excl uding the instrument set Z from (2). In particular, under the null hypothesis that the instruments are not sepa rately correlated w ith achievement, the sample size multiplied by the R 2 from a regression of the residual in (2), on all the exogenous variables (i.e. a constant, X and Z ) is distributed as chisquare with degrees of freedom equal to one less than the number of instruments. The logic is that as the extent of any direct correlation betw een the instruments and achievement increases, the strength of the partial correlation between and Z and thus the R 2 from the above auxiliary regression, does as well. Typically, the estimator represented by equation 3 is generated by a twostage least squares (2SLS) procedure. The first stage estimates equation 1 above using OLS. From the estimated parameters, predicted values of alcohol use, (which earlier was called the linear projection of Z onto D), are constructed for each respondent using their corresponding values of the explanatory variables X and instruments Z The second stage estimates equation 2 using the fitted values in place of observed drinking D. The entire process is performed within a preprogrammed routine in econometric estimation software packages (e.g. Stata), whic h also provides corre ct estimates of the standard errors (which should be calculated using the actual rather than predicted D D 21
PAGE 28
22 drinking variable). In both equations, standard errors that are robust to arbitrary forms of heteroskedasticity will be displayed in the output tables and used to ca lculate relevant test statistics. 2SLS yields consistent estimates even when alcohol use and/or achievement are represented by a binary indicator, which occu rs frequently in my data. However, for binary drinking measures, e.g. an indicator of any past month binge drinking, I will utilize an approach, suggested by Wooldridge (2003) to improve efficiency, which is similar to 2SLS with two modifications. Firs t, before running 2SLS, a preliminary probit regression for equation 1 is estimated. S econd, the ensuing 2SLS procedure uses the predicted probabilities of drinking from the probit regression as instruments in place of Z The resulting estimates are likely to be si milar in magnitude to those that would be generated by the analogous 2SLS regression, but standard errors will be slightly smaller. The necessary conditions for using 2SLS and its desirable properties still hold. Two other methodological points should be mentioned. First, although IV estimates are consistent if the instrument strength and exogeneity conditions outlined above are satisfied, they are ine fficient relative to OLS if it turns out that alcohol use is truly exogenous with respect to achievement in which case the OLS estimates can be interpreted as causal effects. Even under ideal circumstances, i.e. very strong instruments that empirically have very little correla tion with achievement conditional on drinking, standard errors from IV re gressions tend to be much la rger than those from OLS regressions. Thus, it is desirable to economet rically test the null hypothesis that drinking is exogenous in the achievement equation. Th is is easily done using a Hausman (1978) test, which examines whether the IV and OLS estimates of 1 are significantly different
PAGE 29
23 from each other. The relevant test statistic is simply a tstatistic in which the numerator is the difference between the IV and OLS estimates of 1 and the denominator is the square root of the difference in the estimated variances of 1 under IV and OLS. Rejection of this null implies that OLS estim ates are inconsistent and hence conclusions should be based on IV estimates; failure to re ject this null means that OLS estimates are preferable because of their smaller standard errors. Finally, an additional advantage of IV is that it also addresses the issue of errors in the measurement of the drinking variables, which will prospectively be present at least to some degree because data are selfreporte d and thus, in the case of measures like alcohol use that require respondents to re member whether and how much consumption occurred during certain time periods, subjec t to recall error. Even if any resulting measurement error is random, e.g. uncorrelate d with actual or measured drinking or achievement, OLS estimates will be biased to wards zero (the potential for measurement error that varies systematically with drinking or achievement will be discussed in the following section). However, IV estimates are consistent even when alcohol use is measured with random error.
PAGE 30
24 Chapter Four: Data and Empirical Specification Data Description This dissertation will analyze data fr om the 2002 and 2003 waves of the NSDUH. I do not incorporate earlier data for two reasons. First, beginning in 2002 the NSDUH administrators undertook steps to improve the quality of the data gathered by the survey, which included implementation of improved da ta collection procedur es and the payment of $30 to respondents for completed surveys (which raised response rates). Second, before 2002, information utilized in this study regarding variables that serve as instruments do not consistently appear. The NSDUH, sponsored by the Substance Abuse and Mental Health Services Administration (SAMHSA), is administered annually to approximately 55,000 civilian, noninstitutionalized individuals age 12 and over, chosen so that the application of sample weights produces a nationally representative sample, with approximately equal numbers of respondents from the 12, 18 and 26 and over age groups. Geographically, eight large states cont ribute roughly 3,600 respondents each and remaining states provide about 900 respondents e ach. The sample is stratified by state, which are separated into field interviewer regi ons that are further divided into segments consisting of adjacent census blocks. Interviewers visit selected households, one or two residents of which complete the survey. The interviewer uses a computerized survey to enter some responses, but most are answer ed privately in a way that precludes interviewer knowledge of the answers supplied.
PAGE 31
25 An important aspect of the survey that partially dictates sample composition is that the Youth Experiences section, which is the only source of information in the NSDUH on academic performance, is administ ered only to 12 year olds. Academic performance is represented by grades in the most recent marking period, coded as a categorical variable with f our choices: A+ through A, B+ through B, C+ through C, or D or below. This disser tation uses this information in the form of two binary variables, one indicating whet her the grade was A and the other indicating whether it was C or below. Questions on other human capital variable s were asked to all respondents as part of the core demographics section. Curren t school enrollment is a binary variable indicating whether the respondent is curre ntly enrolled in middle or high school (including those who are home schooled) or a co llege/university. The sample for this part of the study is restricted to 1625 year olds. Approximately 99 percent of youth ages 15 and under report attending school, and individuals ages 26 and above who have not graduated from college are particularly likel y to have experienced previous gaps in school enrollment, not currently be enrolled and not return to school in the future. In addition, all enrolled respondents are as ked to report the number of days of school over the past month that were missed because of illness or injury and that were missed due to skipping, with a response range from 0 to 30 days. This dissertation will also examine the effect of drinking on absent eeism due to skipping classes and illness or injury. The latter merits attention in that alcohol consumption may be the reason for reporting illness or injury.
PAGE 32
26 Alcohol use is observed for consumption of various types and over different time periods. The three main measures on which the analysis will focus are the number of days alcohol was consumed during the previous year, the number of drinks consumed in the past month and the occurrence of binge drin king in the past month. Although the timing of the number of drinks and binge drinking vari ables is not an ideal match for some of the educational outcome measures, in the sense that past mont h consumption cannot literally affect behavior that preceded the past m onth, my work will follow that of previous studies in assuming that previous month dr inking patterns proxy those occurring in the recent period prior to the previous month. One other additional piece of alcohol consumption information will also be examined: an indicator of whether respondent s exhibited symptoms of alcohol abuse or dependence in the past year. This is retrospectively coded by SAMHSA based on responses to questions corresponding to crite ria outlined in the fourth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSMIV), the clinical standard for establishing drug abuse and dependence. Th ere are seven criteria utilized, three of which must be met for categorization as depe ndent: 1) Spent a great deal of time over a period of a month getting, using, or getting over the effects of the substance, 2) Unable to keep set limits on substance use or used mo re often than intended, 3) Needed to use substance more than before to get desired e ffects or noticed that using the same amount had less effect than before, 4) Unable to cu t down or stop using the substance every time he or she tried or wanted to, 5) Continue d to use substance even though it was causing problems with emotions, nerves, mental health, or physical problems, 6) Reduced or gave up participation in important activities due to substance use, and 7) experienced
PAGE 33
27 substance specific withdrawal symptoms at one time that lasted for longer than a day after they cut back or stoppe d using. To be categorized as abusive, the respondent must respond positively to one of more of the followi ng criteria: 1) has se rious problems due to substance use at home, work or school, 2) used substance regularly then did something where the substance use might have put them in physical danger, 3) substance use caused actions that repeatedly got them in troubl e with the law, 4) has problems caused by substance use with family or friends and continued to use substance even though it was thought to be causing problems with family and friends. A potentially problematic attribute of th e data is nonrandom measurement error emanating from the selfreported nature of responses. Although IV will eliminate bias from random measurement error, as previous ly discussed, it cannot salvage data plagued by systematic measurement error, which c ould affect both educational outcomes and alcohol use. In particular, one might expect respondents to ar tificially inflate grades, but underreport alcohol consumption. However, studies on the quality of selfreported academic performance and drinking data sugge st that such reporting bias should be minimal. Cassady (2001) found that selfrepor ted GPA values are remarkably similar to official records and therefore are highl y reliable and sufficiently adequate for research use. Similarly, Gr ant et al. (1988), Mida nik (1988) and Reinis ch et al. (1991) concluded that youth drinking selfreports are reliable, based on the consistency of responses to alcohol use ques tions from repeated interviews. Johnston et al. (1988, pg. 20) write that the considerab le amount of inferential ev idence that exists strongly suggests that the selfreport questions produ ce largely valid data. And Harrison and Hughes (1997) found that individuals tend to underreport the us e of stigmatized
PAGE 34
28 substances such as cocaine and heroin mu ch more so than alcohol, whereas survey methods not requiring subjects to verbally answer questions, as in the NSDUH where interviewers are unable to match responses with respondents, increase the accuracy of substance use selfreports. Empirical Specification As discussed in chapter three, the empirical strategy empl oyed involves a two equation model. First, a probit regression is conducted and pr edicted values of drinking are obtained. Second, a standard 2SLS regres sion is performed. Once again the equation system stated in chapter three is: (1) D = 0 + Z 1 + X 2 + (2) E = 0 + 1 D + X 2 + In terms of the NSDUH variable s, the notation is as follows. D represents one of the four drinking measures defined in the pr eceding section: 1) the number of days the respondent drank in the previous year, 2) the number of drinks the respondent consumed in the previous month, 3) whether the respondent engaged in binge dri nking in the last 30 days or, 4) whether the responde nt is categorized as alcoho l dependent and/or abusive. E denotes one of the educational outcom es examined: 1) indicators of obtaining an A and a C or lower average, 2) an indicator of school enrollment and, 3) the number of days the respondent reported abse nteeism due to skipping or due to illness or injury. Z is a set of instrumental variables th at influence drinking but not education directly. The instruments are discussed in detail below. X represents other plausibly exogeneous educational outcome determinants th at are included as explanatory variables in the regression equations. The spec ific variables utilized in the Z and X vectors depend
PAGE 35
29 upon the outcome, E under investigation and the subs ample used: some variables that serve as instruments and explanatory vari ables are only availa ble only for 12 year olds in the Youth Experience section of the NSDUH, while others are available for all age categories. When E is an indicator of obtaining an A av erage or a C or lower average, the analysis is conducted using hi gh school students ranging in age from 12 to 17 as grades are only available for the Youth Experience sa mple. The set of instrumental variables for the analysis of grades includes indicators for perceived risk of consuming alcohol, parental disapproval of dri nking and peer alcohol use, as defined below. The X vector includes measures of student age, race, grade level, family income, family size, whether each of the mother and father live in the household, whether the respondent was born in the U.S., the number of times the family relocated in the previous five years, the extent to which parents help with homework, and two population density categories, also defined below. When E is an indicator of school enrollment or the number of days missed due to skipping classes or due to illness or inju ry, two subsamples are analyzed. One sample includes high school age students ranging in age from 16 to 19 years old and another sample includes college age students ranging from 18 to 25 years old. For the high school age sample, high school graduates are exclude d. For the college samp le, only high school graduates are included, and college gra duates are excluded. For these educational outcomes, the instruments employed are indi cators of perceived risk of consuming alcohol, whether religion is important in the respondents life and whether religious believes influence decisions. The X vector includes race, family income, family size,
PAGE 36
30 whether the respondent was born in the U.S., the number of times the respondent moved in the previous five years, and two population density categories. Though the same instruments and explanatory variables are used in each case, the samples are stratified in order to allow for the effect that drinking has on enrollment and absenteeism to differ between high school and college students. The explanatory variables utilized in all the samples are defined as follows. Family income is measured in four categories: $10,000$19,999; $20,000$49,999; $50,000$79,999; and $75,000 or greater, with $10,000$19,999 as the omitted category. Population density is represented by indicat ors for two categories: an MSA with one million persons or greater and an MSA of less than one million persons, with nonMSA areas as the omitted category. For race, indicators are specified for African Americans, Native Americans, Asians, and nonwhite Hisp anics, with whites as the omitted category. Family size is measured using two variables, the number of member s if the household has one to five members and an indicator for those with over five members, for which the numerical variable is set to zero. Those explanatory variables that were only available for the youth experience sample are specified as follows: age indicat ors for 15, 16, or 17 years old (with age 12 omitted, because very few high school students are age 123), indicators of whether the mother or father reside in the household, indicators for whether parents assisted the student with homework always, sometimes or seldom in the past 12 months, with never as the omitted category, and indicators for whether the student is currently attending the 10 th 11 th or 12 th grade (with 9 th as the omitted grade).
PAGE 37
31 Several NSDUH variables conceivably influence drinking without having direct effects on educational outcomes and are thus can didates to serve as instrumental variables in the analysis. The specific variables utilized in this st udy are: parental disapproval of alcohol use; peer use of alc ohol, perceived risk of bodily harm from alcohol use; whether religion is important in the re spondents life; and whether re ligious beliefs influence the respondents decisions. The first two variables are recorded only for 1217 year olds as part of the Youth Experiences section of the NSDUH. The latte r three are available for all age groups and are employed in subsamples of 1619 and 1825 year olds. Information is reported on whether pare nts would neither approve nor disapprove, somewhat disapprove or strongl y disapprove of the respondent having one or two drinks of an alcoholic beverage nearly every day. In this study, a bi nary variable is created and coded as 0 if the parent is indifferent or somewhat disapproves and 1 if the parent strongly disapproves. Peer use information reflects a question about perceptions of the respondent regarding whether none, a few, most, or all student s in the same grade at his or her school consume alcohol. For the peer us e variable, a binary measure is defined to designate if the respondent f eels that most or all schoolma tes consume alcohol. Potential endogeneity of the peer vari able, stemming from a possibl e connection between ones own behavior and perceptions about the beha vior of others, should be mitigated by the fact that the relevant questi ons cover all classmates rather than simply friends, who are presumably chosen by the respondent. Norton et al. (1998), Gaviria and Raphael (2001) and Kremer and Levy (2003) each found evidence that increased ra tes of drinking among peers raised the propensity to consume alcohol.
PAGE 38
32 With regard to the other three instrume ntal variables utilized in this study, NSDUH respondents were asked qu estions regarding religious beliefs and perceived risks involved in drinking. Religiosi ty encompasses the strength of agreement or disagreement with statements regarding whether religious be liefs are important and influence decisions. The risk variables indicate th e extent to which respondents th ink that people risk harming themselves physically and in other ways by ha ving four or five dri nks nearly every day. For the religion factors, binary measures are constructed to indicate if the respondent agrees that religious beliefs are important in his/ her life and if the respondent agrees that religious beliefs influence his/her decisions. For risk, the binary measure indicates if the respondent feel s there is moderate to great risk of harm, physically or otherwise, from consuming four to five drinks daily. The instruments are presumed to be corre lated with youth drinking yet not exhibit a direct correlation to the educational outcome in question. However, that characteristic of the instruments needs to be empirically tested, as one can imagine other indirect avenues through which the instruments might impact educational outcomes. For instance, parents that disapprove of alc ohol use might also strictly discipline their children to succeed in school and channel more family resources to schooling, thereby raising grades and enrollment and lowering truancy. In a ddition, students who drink might overestimate the accurate amount of drinking by schoolmates and vice versa, and a correlation with education might exist if classmate drinking proxies for sc hool quality. Students that perceive greater risks in dri nking may also be more risk averse in general, therefore perceiving greater risks associated with academic failure and experiencing better educational outcomes. Finally, students who st ate religion is important may also believe
PAGE 39
33 education is more important and may be gr eater discipline to do well in school; more generally, strong religious beliefs might al so influence attitude s about education. Given these caveats, the true exogeneity of the instruments may be comprised. Overidentification tests, and te sts of the sensitivity of the IV estimates utilizing differing sets of instruments, are thus conducted to verify the hypothesis of instrument exogeneity. Theoretically the bias of OLS estimates co uld be in either direction, meaning that OLS estimates could be either larger or sma ller than their IV count erparts. OLS estimates will overestimate the negative effects of drinking if unobserved factors such as the emotional distress or personality characteris tics induce a student to drink more while commensurately lowering academic performance, or if reverse causation is such that poor grades lead to drinking. In contrast OLS will underestimate negative effects of drinking if income effects lead to more drinking and better academic performance, or good performance is celebrated with drinking. Also, random measurement error will cause attenuation bias in OLS. It is expected that the major cause of bias is unobserved heterogeneity that inflates the magnitude of OLS estimates, but that income and measurement error effects working in the oppos ite direction might also be important.
PAGE 40
34 Chapter Five: Empirical Results This chapter discusses the empirical resu lts. The causal effect youth alcohol use has on these educational outcomes is discusse d, as are the effects of other independent variables. The effect of alc ohol consumption on high school grad es is discussed first. The effect of alcohol use on the pr obability of obtaining an A aver age versus lower grades is examined, as is the probability of obtaining a C average or lower versus higher grades. Those results are based solely on the youth e xperiences sample of 12 year olds. Then, the effects of alcohol consumption on the proba bility of school en rollment are analyzed utilizing the samples of 16 year olds who have not graduated from high school and 18 year olds who have graduated from hi gh school but not from college. The effects of drinking on absenteeism (t he number of school days the student missed due to skipping class and illness or injury) are studied as well, utilizing the age 16 and 18 25 samples. The youth experience sample utilizes three instrumental variables to identify drinking in the grade equations : parental disapproval of drin king, perceived risk of harm from drinking and peer use of alcohol. The enrollment and absenteeism regressions utilize three instrumental variables: percei ved risk of harm from drinking, whether religious beliefs are an impor tant part of the respondents life, and whether religious beliefs influence how the respondent makes decisions. In order to assess the impact of instrumental variables, this chapter includes comparisons of coefficient estimates from single equation estimation using OLS with
PAGE 41
35 those from instrumental variables models. To determine if the effects of youth drinking on the educational measures are influenced by the choice of instruments, separate analyses are conducted using di fferent sets of instruments. All regressions employ all three instruments in the main specification and the robustness analyses use pairs of two instruments. Tables 1 and 2 present summary statistics fo r each of the variables utilized for the 12 and 18 year olds respectively. For the youth sample, the mean number of days drinks were consumed in the past year is 19.8 while the mean number of drinks consumed in the past month is 8.2. Approxi mately onesixth of high school students binge drink while about onetenth are classified as alcohol dependent. The mean value of reported peer drinking is 0.5. A vast majority of youths, 87 percent, report their parents discourage drinking. Family income is le ss than $20,000 for 17 percent of respondents but greater than $75,000 for 26 percent. About 72 percent of respondents live in an MSA, roughly equally split between MSAs with popul ations greater than and less than one million. Fathers are less likely to be presen t in the household than are mothers. The proportion of parents that help with homework is also very high. African Americans comprise about 12 percent of the sample while nonwhite, nonblack Hispanics account for about 14 percent.
PAGE 42
Table 1. Descriptive Statistics (Youth Experience sample) (n=18,231) Standard Variable MeanDeviation Number of days drankpast year 19.78047.620 Number of drinks in previous month 8.17238.325 Binge drinking in the past 30 days 0.1530.360 Abuse/ Dependence on alcohol classification 0.0910.287 Respondent perceives risk of harm from drinking 0.8230.382 Respondent perceives schoolmate (peer) drinking 0.4960.500 Parents disapprove of alcohol 0.8670.399 Probability of an 'A' grade 0.2650.441 Probability of a 'C' or lower grade 0.3110.464 Family income ($20,000$49,999) 0.3650.481 Family income ($50,000$74,999) 0.2070.405 Family income ($75,000 or more) 0.2600.438 MSA segement with 1+ million persons 0.3610.480 MSA segment of less than 1 million 0.3580.479 Age of student (14 years old) 0.1810.380 Age of student (15 years old) 0.2680.440 Age of student (16 years old) 0.2740.446 Age of student (17 years old) 0.2750.446 Mother in household 0.9060.291 Father in household 0.7280.444 Parents help with homework (always) 0.5260.499 Parents help with homework (sometimes) 0.2350.420 Parents help with homework (seldom) 0.1250.320 Grade in (10th grade) 0.2700.440 Grade in (11th grade) 0.2500.430 Grade in (12th grade) 0.1600.370 Race (African American) 0.1200.320 Race (Native American) 0.0300.190 Race (Asian) 0.0290.169 Race (nonwhite Hispanic) 0.1410.348 Number in family 3.2671.529 Number in family (>5) 0.1330.340 year 2002 0.4850.499 For the 18 year old sample, the mean num ber of drinks consumed in the past year increases to 58.2, and the number of dri nks consumed in the past month rises to 24.0. The incidence of binging is 0.41 and the m ean of risk associated with alcohol use falls from 0.82 in the youth sample to 0.71. Among the 44 percent of respondents reporting current school enrollme nt, about one day per month of classes is skipped and 36
PAGE 43
another nearly threequarters of class days per month is missed due to illness. The racial composition is similar to th at of the youth sample. Table 2. Descriptive Statistics (1825 years old) (n=28,065) Standard Variable MeanDeviation Number of days drankpast year 58.17276.960 Number of drinks in previous month 23.98265.176 Binge drinking in the past 30 days 0.4080.490 Abuse/ Dependence on alcohol 0.1730.378 Respondent perceives risk of harm from drinking 0.7080.454 Respondent states religion is important in life 0.6950.450 Respondent states religion influences their decisions 0.6180.485 Current school enrollment 0.4370.496 Number of skipped school days (past 30 days) 0.9242.059 Number of school days missed due to illness (past 30 days) 0.7221.850 Family income ($20,000$49,999) 0.3970.341 Family income ($50,000$74,999) 0.1350.489 Family income ($75,000 or more) 0.1360.341 MSA segement with 1+ million persons 0.3400.473 MSA segment of less than 1 million 0.3920.488 Age of student (19 years old) 0.1470.354 Age of student (20 years old) 0.1380.345 Age of student (21 years old) 0.1350.342 Age of student (2223 years old) 0.2160.411 Age of student (2425 years old) 0.1960.397 Last grade completed (Freshman) 0.1450.352 Last grade completed (Sophomore/ Junior) 0.1970.398 Race (African American) 0.1350.330 Race (Native American) 0.0130.115 Race (Asian) 0.0260.160 Race (nonwhite Hispanic) 0.1670.373 Number in family 2.9481.327 Number in family (>5) 0.0800.270 year 2002 0.4870.499 Table 3 presents summary statistics for the sample of 16 year olds who have not graduated from high school. Drinking is gr eater than in the youth experience sample but not as high as in the 18 sample: abou t 24 percent reports bi nge drinking in the past 30 days and 12 percent are classified as abusing/dependent on alcohol. Those enrolled skipped school about half a day in the past month, and also missed just over a 37
PAGE 44
day per month due to illness. Household income and racial composition roughly mirrors that of the other subsamples. Table 3. Descriptive Statistics (1619 years old) (n=13,526) Standard Variable MeanDeviation Number of days drankpast year 28.69756.155 Number of drinks in previous month 12.84448.536 Binge drinking in the past 30 days 0.2390.427 Abuse/ Dependence on alcohol classification 0.1230.328 Respondent perceives risk of harm from drinking 0.7380.439 Respondent states religion is important in life 0.7130.451 Respondent states religion influences their decisions 0.5990.490 Current school enrollment 0.8630.342 Number of skipped school days (past 30 days) 0.5091.620 Number of school days missed due to illness (past 30 days) 1.0932.241 Family income ($20,000$49,999) 0.3630.480 Family income ($50,000$74,999) 0.1870.390 Family income ($75,000 or more) 0.2300.421 MSA segment with 1+ million persons 0.3470.470 MSA segment of less than 1 million 0.3660.481 Age of student (16 years old) 0.2310.421 Age of student (17 years old) 0.2300.421 Age of student (18 years old) 0.1880.390 Last grade completed (9th) 0.2310.421 Last grade completed (10th) 0.2400.427 Last grade completed (11th) 0.2290.420 Race (African American) 0.1340.341 Race (Native American) 0.1270.112 Race (Asian) 0.0280.165 Race (nonwhite Hispanic) 0.1450.352 Number in family 3.1811.493 Number in family (>5) 0.1240.330 year 2002 0.4820.499 A. Drinking and Grades This section presents results for th e effect of youth drinking on grades. Specifically, causal effects that drinking ha s on the probabilities of obtaining certain grades is estimated using the three in strumental variables listed earlier. The overidentification statistics aid in revealing whether the inst rument set is exogenous with respect to academic performance. An analys is is conducted that employs differing pairs 38
PAGE 45
of instruments to determine if some instru ment sets are more plausibly exogenous. The main results of the IV analysis are also compared with parameter estimates obtained using OLS. While the discussion here focuses mainly on the effects of alcohol consumption on grades, appendix 1 presents the probit estimates from the drinking equations. For the binge drinking measure, appe ndices 2 and 3 show the IV coefficients and standard errors of all exogenous variables for the for the probability of obtaining an A and a C or lower, respectively. For probit models, tables contain marginal effects at the explanatory variable means. First Stage Regression Results Results from the probit and first stage re gressions of the drinking measures on the instrumental variables for grade probabilities are shown in table 4. The two left columns are from OLS first stage regressions. The tw o right columns are probit marginal effects and associated standard errors except for the last row, which shows the coefficient and standard error for the predicted value of drinking obtained from the drinking probits. Table 4. Probit/ First stage regression estimates for the grade outcomes (n=18,231) number of daysnumber of drinksBingeAbuse/ Dependence exogeneous variables drank in past yearin past monthdrinkingon alcohol Risk of bodily harm from drinking 15.890 10.420 0.125 0.078 (0.894) (0.749) (0.008) (0.006) Peer use of alcohol 11.949 6.241 0.086 0.054 (0.716) (0.600) (0.005) (0.004) Parental disapproval of alcohol use 24.706 14.201 0.162 0.074 (0.716) (0.845) (0.005) (0.007) F stat/ chi2coefficient of joint significance 486.83 246.49 1104.87 612.52 Pvalue of significance level (0.0000) (0.0000)(0.0000)(0.0000) predicted drinking coefficient 0.978 0.934 (0.024) (0.032) The results demonstrate that peer alc ohol use has positive effects on drinking while parental disapproval and perceived risk have negative effects. For respondents who agreed that there is moderate to great risk of harm from consuming 45 drinks almost 39
PAGE 46
40 every day, the number of days that drinking occurred in the past year is lowered by approximately 16 days, the number of drinks consumed in the past month is reduced by 10 days, the probability of binge drinking in the last 30 days falls by 0.13, and the likelihood of being categorized as abusiv e/dependent on alcohol falls by 0.08. For respondents who report most or all their schoolmates use alcohol, the number of days drinking occurred in the past year rises by 12 days, the number of drinks consumed in the past month rises by 6, the pr obability of binge drinking in the last 30 days rises by 0.09, and the likelihood of be ing categorized as abusive/dependent on alcohol rises by 0.05. Parental objection to alcohol use has th e strongest effect. For respondents who report their parents strongly di sapprove of having one or two drinks per day, the number of days drinking occurred in the past year is lowered by 25 days. The number of drinks consumed in the past month is reduced by 14, while the likelihood of binge drinking in the last 30 days falls by 0.16. The likelihood of being categorized as abusive/ dependent on alcohol falls by 0.07. The F statistics and 2 coefficients and associated pvalues give strong evidence of joint instrument significance with respect to all the drinking measures. The predicted drinking coefficients in the first stage regressions from the Wooldridge binary endogenous variable method are 0.98 fo r binge drinking in the past 30 days and 0.93 for abuse/dependence on alcohol. The Effects of Drinking on the Pr obability of Obtaining an A Table 5 presents results for the probab ility of obtaining an A versus lower grades. Drinking has significant, negative eff ects on the probability of earning an A.
PAGE 47
Table 5. IV estimates of drinking on the probability of an 'A' All three youth experience instruments (n=18,231) Alcohol variables IV OLS number of days drankpast year 0.003*0.0009* Marginal Effect Standard Error (0.0002)(0.00006) Pvalue of overidentification test 0.106 Hausman statistic 10.008* number of drinks in past month 0.005*0.0006* Marginal Effect Standard Error (0.0005)(0.00008) Pvalue of overidentification test 0.051 Hausman statistic 9.826* binge drinking 0.351*0.125* Marginal Effect Standard Error (0.030) (0.008) Pvalue of overidentification test 0.006 Hausman statistic 8.621* abuse/ dependence on alcohol 0.557*0.096* Marginal Effect Standard Error (0.055) (0.010) Pvalue of overidentification test 0.007 Hausman statistic 8.944* *Statistically significant at 1% An additional day increase in the numb er of past year drinking days reduces the probability of achieving an A by 0.003 percen tage points. For instance, if a student reports drinking 52 days in the previous year, the probability of having an A average in the current grading period is reduced by 0.156 points compared to not drinking at all. For each drink increase in the number of drinks consumed in the past month, the probability of obtaining an A is reduced by 0.005. If the student consumes, on average, two alcoholic drinks per day in the past 30 days, the probability of the student having an A average falls by 0.30 rela tive to abstaining. For respondents that reported binge drinking in the previous 30 day period, there is an associated reduction in the probability of obtaining an A average of 0.35. For t hose categorized as abusive/ dependent on alcohol, the probability of obtaining an A is reduced by 0.56. 41
PAGE 48
42 The Hausman statistics signify that there are statistically significant differences between the OLS and IV parameter estimat es for all the drinking measures. The overidentification tests for binging and abuse/dependence on alcohol have associated pvalues that offer little evidence in support of the assumption of exogeneity. For the other two drinking measures, instrument exogeneity is not rejected at the 5 percent level. There are fairly large negative effects on the probability of acquiring an A for each drinking measure analyzed. For instance, for each additional drink consumed in the previous month, there is a 1.8 percent decline in the probability of obtaining an A. This indicates that alcohol consumption on the part of high school student s could be impairing the learning process, which in turn reduces the capability of the student to earn top grades. There is also an opportunity cost i nvolved in drinking, which includes reduced study time and possibly increased devotion of the students monetary resources to consuming alcohol that detracts from the pr ospect of receiving an A average. These results imply that those costs could be substa ntial. While the overidentification tests offer weak support for the assumption of instrument exogeneity, the following section further explores the issue by conducting the analysis utilizing differing pairs of instruments. Instrument Robustness and the Pr obability of an A Average Table 6 shows the results of regressi ons performed with varying pairs of instruments. This exercise is undertaken to determine if there is any sensitivity in the main results to changes in the instrument set. The instrument that is omitted from the IV combination is utilized as an explanatory variable and its coefficient and standard error is reported in the table.
PAGE 49
Table 6. IV estimates of drinking on the probability of a 'A' using IV pairs (n=18,231) parent disapproverisk and parent disapprove and risk peer use and peer use Alcohol variables number of days drankpast year 0.002* 0.003* 0.003* Marginal Effect Standard Error (0.0002) (0.0004) (0.0003) Pvalue of overidentification test 0.270 0.386 0.036 Hausman statistic 7.723* 7.792* 7.970* Coefficient (Standard Error) of omitted IV 0.015 (0.008)0.028 (0.014)0.004 (0.010) number of drinks in past month 0.004* 0.006* 0.005* Marginal Effect Standard Error (0.0005) (0.0007) (0.0006) Pvalue of overidentification test 0.731 0.049 0.018 Hausman statistic 8.047* 7.479* 7.953* Coefficient (Standard Error) of omitted IV 0.020 (0.008)0.021 (0.016)0.007 (0.012) binge drinking 0.295* 0.327* 0.315* Marginal Effect Standard Error (0.033) (0.041) (0.037) Pvalue of overidentification test 0.015 0.003 0.004 Hausman statistic 5.985* 5.512* 5.954* Coefficient (Standard Error) of omitted IV 0.026 (0.007)0.010 (0.012)0.016 (0.010) abuse/ dependence on alcohol 0.459* 0.468* 0.496* Marginal Effect Standard Error (0.062) (0.068) (0.070) Pvalue of overidentification test 0.002 0.001 0.001 Hausman statistic 6.354* 5.870* 6.160* Coefficient (Standard Error) of omitted IV 0.026 (0.008)0.026 (0.011)0.016 (0.010) *Statistically significant at 1% The table 6 estimates are similar to thos e in the regressions in which all three instruments are employed. For binging a nd abuse/dependence on alcohol, the overidentification tests continue to reject instrument e xogeneity. Instrument exogeneity is not rejected for the past year and past mont h drinking variables when peer use is entered into the grade equation, and this specification also yields the highest overidentification pvalues for binge drinking and abuse/depe ndence. Peer use is accordingly always significant in the grade equati on, as is parental disapproval for past year drinking and abuse/dependence. Hausman tests indicate ther e are statistically significant differences between IV and OLS estimates in all specifications. 43
PAGE 50
Overall, the results of this sensitivit y analysis offer some evidence that the identification strategy produces consistent es timates of the effect of drinking on the probability of an A average. At the 5 percent level, instrument exogeneity is not rejected in some cases. In addition, for each m easure of alcohol use, drinking coefficients are similar regardless of which instru ments are used to identify drinking. The Effects of Drinking on the Probab ility of a C or Lower Average Table 7 presents the IV regression estimates for the probability the respondent has a C or lower grade versus other grades. Th ere are significant and positive effects on the probability of earning a C or lower aver age for all the drinking variables. Table 7. IV/ OLS estimates of drinking on the probability of a 'C' or lower All three youth experience instruments (n=18,231) Alcohol variables IV OLS number of days drankpast year 0.004* 0.001* Marginal Effect Standard Error (0.0003)(0.00007) Pvalue of overidentification test 0.048 Hausman statistic 10.286* number of drinks in past month 0.003* 0.0005* Marginal Effect Standard Error (0.0005)(0.00008) Pvalue of overidentification test 0.041 Hausman statistic 9.704* binge drinking 0.429* 0.123* Marginal Effect Standard Error (0.036) (0.009) Pvalue of overidentification test 0.021 Hausman statistic 9.537* abuse/ dependence on alcohol 0.689* 0.111* Marginal Effect Standard Error (0.067) (0.011) Pvalue of overidentification test 0.063 Hausman statistic 9.124* *Statistically significant at 1% Each daily increase in the num ber of past year drinking days raises the probability of having a C or lower average by 0.004, wh ile the probability is raised by 0.003 for each extra drink consumed in the previous month. If the respondent drinks 52 additional 44
PAGE 51
45 days in the past year, the probability of a C or lower average rises by 0.21 points. And if 30 more drinks are consumed by the student in the past month, the probability of having a C or lower grade is raised by 0.09. For those engaging in binge or abusive/ dependent drinking, there is a significant positive effect on the probability the student has a depressed grade point average in the current period. For binge drinkers there is an associated elevation in the probability of having a C average of 0.43. For those categori zed as abusive/dependent on alcohol, the probability of obtaining a C or lo wer average is raised by 0.69. The pvalues of the overidentification test s afford little support for the assumption of exogeneity. Only for the abuse/ dependen ce indicator is instrument exogeneity not rejected at the 5 percent level. The Hausman coefficients, however, show that there are statistically significant differences between IV and OLS estimates. The estimated effects for binge drinking and abuse/dependence are quite large. There may be large opportunity costs associated with this intense drin king, especially at abuse and dependence levels, which drastica lly undercut academic achievement. Thus, grades are dramatically lower, and possibl y high failure rates ma y account for some of this.
PAGE 52
Instrument Robustness and the Probab ility of a C or Lower Average Table 8 shows the results of regressi ons performed with varying pairs of instruments. The analysis is conducted to determine if there is any sensitivity in the results to changes in the instrument set. Table 8. IV estimates of drinking on the probability of a 'C' using IV pairs (n=18,231) parent disapproverisk and parent disapprove and risk peer use and peer use Alcohol variables number of days drankpast year 0.004* 0.004* 0.003* Marginal Effect Standard Error (0.0003) (0.0004) (0.0003) Pvalue of overidentification test 0.033 0.860 0.055 Hausman statistic 7.912* 8.236* 7.311* Coefficient (Standard Error) of omitted IV 0.008 (0.009)0.038 (0.018)0.017 (0.012) number of drinks in past month Marginal Effect Standard Error 0.003* 0.006* 0.005* Marginal Effect Standard Error (0.0006) (0.0007) (0.0006) Pvalue of overidentification test 0.216 0.275 0.028 Hausman statistic 6.905* 8.231* 8.010* Coefficient (Standard Error) of omitted IV 0.007 (0.008)0.031 (0.017)0.005 (0.013) binge drinking 0.386* 0.437* 0.376* Marginal Effect Standard Error (0.040) (0.050) (0.044) Pvalue of overidentification test 0.018 0.009 0.005 Hausman statistic 7.325* 6.981* 6.423* Coefficient (Standard Error) of omitted IV 0.020 (0.008)0.003 (0.015)0.024 (0.012) abuse/ dependence on alcohol 0.619* 0.634* 0.602* Marginal Effect Standard Error (0.077) (0.084) (0.084) Pvalue of overidentification test 0.014 0.004 0.004 Hausman statistic 7.028* 6.589* 6.208* Coefficient (Standard Error) of omitted IV 0.018 (0.009)0.016 (0.015)0.022 (0.013) *Statistically significant at 1% Again, the IV estimates are similar rega rdless of instrument choice. For the drinking measures, the overidentification test s reject instrument exogeneity at the 10 percent level, save for past year drinking in the risk and peer use specification, and past month drinking except when risk is included in the grade equation. For binge drinking 46
PAGE 53
47 and abuse/dependence factors, the overidentifica tion test results offer little evidence that the instruments are exogeneous. Throughout the grade analysis, OLS parameter estimates consistently underestimate the magnitude of the negative eff ects. This could possibl y be attributed to higher ability students drinking more or highe r income students having more resources to devote both to drinking and their education. Random measurement error that IV corrects could also play a role. B. Drinking and School Enrollment This section presents results for the eff ect of youth drinking on the probability of school enrollment. This outcome variable is described in chapter f our. The causal effect that drinking has on this variable is analyzed using the three instru mental variables also described in that chapter. The enrollment anal ysis is conducted utilizing a sample of highschool age students (1619 years old) and college age students (1825 years old). To more accurately determine the validity of the excl usion restrictions for the instruments, a robustness analysis is conducted to using va rious combinations of instruments. The analysis also discusses results of comparisons between IV and OLS parameter estimates. The discussion that follows focuses primarily on the effect of drinking on enrollment. Appendices 4 and 5 show the probit estimates for enrollment for the 1825 year old sample and the 1619 year old sample respectively. Appendices 6 and 7 show the coefficients and standard errors of all e xogenous variables for the binge drinking measure for the probability of enrollment for the 1619 and 1825 samples respectively.
PAGE 54
First Stage Regression Results Table 9 presents the results of the probit and first stage regres sions of the drinking measures on the instruments for the 1825 year old age group. Table 9. Probit/ First stage estimates for enrollment (1825 years old) (n=28,065) number of daysnumber of drinksBingeAbuse/ Dependence exogenous variables drank in past yearin past monthdrinkingon alcohol Risk of bodily harm from drinking 34.717 19.622 0.242 0.096 (0.971) (0.843) (0.006) (0.005) Respondent states religion is important in life 3.598 0.954 0.010 0.008 (1.221) (1.060) (0.008) (0.006) Respondent states religion influences decisions 13.276 8.946 0.106 0.041 (1.156) (1.044) (0.008) (0.007) F stat/ chi2coefficient of joint significance 564.78 247.24 1671.970 497.940 Pvalue of significance level (0.0000) (0.0000) (0.0000) (0.0000) predicted drinking coefficient 1.006 0.99 (0.022) (0.040) Of those who perceive that there is moderate to gr eat risk of harm from consuming 45 drinks almost ev ery day, the number of days drinking occurred in the past year is lowered by about 34 days. The number of drinks consumed in the past month is reduced by 20, while the likelihood of binge drinking in the last 30 days falls by 0.24. The likelihood of being categorized as abusive/dependent on alcohol falls by 0.10. Importance of religious beliefs reduces al l alcohol use measures. For those that report that religion is important in life, the number of days drinking occurred in the past year is lowered by 3.6 days. The number of dri nks consumed in the past month is reduced by 0.95, while the probability of binge drinki ng in the last 30 days falls by 0.01. The likelihood of being categorized as abus ive/dependent on alcohol falls by 0.008. When religiosity impacts decisions, the eff ects on the drinking measures are more pronounced. The number of days drinking occurr ed in the past year is lowered by 13 days. The number of drinks consumed in the past month is reduced by about nine, while 48
PAGE 55
the probability of binge drinking in the last 30 days falls by 0.11. The likelihood of being categorized as abusive/depende nt on alcohol falls by 0.04. The 2 coefficients and associated pvalues indicate that the instrume nts are jointly significant for all the drinking measures. The predicted drinking coefficients in the binary drinki ng measure first stage regresssions are 1.01 for binge drinking in the past 30 days and 0.99 for abuse/ dependence on alcohol. Table 10 presents the probit and first stag e results for the instruments for the 1619 year old age group. Table 10. Probit/ First stage estimates for enrollment (1619 years old) (n=13,526) number of daysnumber of drinksBingeAbuse/ Dependence exogeneous variables drank in past yearin past monthdrinkingon alcohol Risk of bodily harm from drinking 20.650 12.721 0.170 0.090 (1.092) (0.961) (0.009) (0.007) Respondent states religion is important in life 1.650 2.780 0.013 0.007 (1.346) (1.180) (0.009) (0.007) Respondent states religion influences decisions 10.841 4.908 0.100 0.049 (1.260) (1.109) (0.009) (0.007) F stat/ chi2coefficient of joint significance 185.76 87.46 664.860 299.180 Pvalue of significance level (0.0000) (0.0000)(0.0000) (0.0000) predicted drinking coefficient 1.006 0.989 (0.034) (0.052) For this age group, if moderate to great ri sk of harm from consuming 45 drinks almost every day is perceived, the number of days in which drinking occurred in the past year is lowered by 21 days. The number of dri nks consumed in the past month is reduced by roughly13, while the probability of binge dr inking in the last 30 days falls by 0.17. The likelihood of being categorized as abus ive/dependent on alcohol also falls by 0.09. Importance of religious beliefs and relig iously influenced decisions reduce all alcohol use measures. For those who report that religion is important in life, the number of days in which drinking occurred in th e past year is lowered by 1.7 days. The number 49
PAGE 56
50 of drinks consumed in the past month is reduced by 2.8, while the probability of binge drinking in the last 30 days falls by 0.01. The likelihood of being categorized as abusive/ dependent on alcohol falls by 0.007. When relig iosity impacts decisions, the number of days in which drinking occurred in the past year is reduced by 11. The number of drinks consumed in the past month is reduced by five, while the probability of binge drinking in the last 30 days falls by 0.10. The likelihood of being categorized as abusive/dependent on alcohol falls by 0.05. The F statistics and 2 pvalues signify support for the hypothesis of joint instrument significance for all the drinking measures. The predicted drinking values are 1.01 for binge drinking in the past 30 days and 0.99 for abuse/ dependence on alcohol. The Effects of Drinking on Sc hool Enrollment (1619 sample) Table 11 presents findings for school enro llment using all three instruments. The analysis is conducted utilizi ng the subsample of 16 year olds who have not graduated from high school. For each daily increase in reported drinki ng, the probability of being enrolled is subsequently lowered by 0.001. If, for instan ce, the respondent reports drinking 52 days in the previous year, the likelihood of enro llment is diminished by approximately 0.052 compared to not drinking at all. For each drink consumed in the prior month the probability of enrollment is lowered by 0.001 percent. If the student reports consuming 30 drinks in the previous month, the proba bility of enrollment d ecreases by 0.03 points. Binge drinking and abuse/dependence on alc ohol further reduce the probability of enrollment. Binging reduces the probability of enrollment by 0.08. For students who have
PAGE 57
engaged in binge drinking, the probabili ty of school enrollment declines by approximately 10 percent compared to not binging. Table 11. IV/ OLS estimates of drinking on school enrollment (1619 years old) (all three instruments) (n=13,526) Alcohol variables IV OLS number of days drankpast year 0.001*0.0003* Marginal Effect Standard Error (0.0001)(0.00003) Pvalue of overidentification test 0.014 Hausman statistic 2.339** number of drinks in past month 0.001*0.0002* Marginal Effect Standard Error (0.0003)(0.00004) Pvalue of overidentification test 0.011 Hausman statistic 3.106* binge drinking 0.083*0.034* Marginal Effect Standard Error (0.023)(0.005) Pvalue of overidentification test 0.021 Hausman statistic 2.259** abuse/ dependence on alcohol 0.178*0.018* Marginal Effect Standard Error (0.046)(0.006) Pvalue of overidentification test 0.017 Hausman statistic 3.545* *Statistically significant at 1% **Statistically significant at 5% For those classified as abusive/dependent with respect to alcohol, the probability of enrollment decreases by 0.18. Categoriza tion as abusive/dependent reduces the probability of school enrollment by 21 percent. There is again little evid ence to support the hypothesis of instrument exogeneity for the three instrument specification. The pva lues associated with the overidentification tests for all drinking measures i ndicate that instrument exogene ity is rejected even at the 5 percent level. For each drinking measure, the Hausman coefficient signifies that statistically significant differences pr evail between IV and OLS estimation. Overall, in the high school sample, th ere is a strong indi cation that drinking, possibly by raising the opportuni ty cost of high school educat ion and impairing cognitive 51
PAGE 58
52 functioning, reduces enrollment in high school And, considering the additional resources the student devotes toward dri nking if the student binge drinks or is abusive/dependent on alcohol, there is compelling evidence that the probability of high school enrollment is largely and negatively impacted. Though overiden tification tests cast doubt on instrument exogeneity, drinking effects ar e significant and rather si zeable. The following section further investigates exogeneity by conducti ng an analysis using differing pairs of instruments. Instrument Robustness and School Enrollment (1619 sample) Table 12 shows the results of regressi ons performed with varying pairs of instruments. This is undertaken to determine if there is any sensitivity in the main results to changes in the instrument set. The excl uded instrument is used as an explanatory variable. Models in the first and third co lumns are preferred, based on large overidentification test pvalues, to those in the 2 nd column and in table 11. This means that religion being important should be used as an instrument with either of the other two instruments, but not both of them. The IV estimates vary widely across these two models, however, with coefficients in the thir d column being three to four times as large as those in the first column. Because the Ha usman statistic in the first column, which offers the more conservative estimate, is never significant at even the 10 percent level, but is always larger than magnitude than th e OLS estimate, one would have to conclude that the OLS estimate gives the best estim ate of the causal effect of drinking on enrollment. Table 11 shows that the OLS estimates are significantly negative, but relatively small.
PAGE 59
Table 12. IV estimates of drinking on enrollment using IV pairs (1619 years old) (n=13,526) religion importantreligious decisionsreligion important & and risk and risk religious decisions Alcohol variables number of days drankpast year 0.0004* 0.0005* 0.001* Marginal Effect Standard Error (0.0002) (0.0002) (0.0005) Pvalue of overidentification test 0.599 0.029 0.876 Hausman statistic 0.199 1.293 3.448* Coefficient (Standard Error) of omitted IV 0.015 (0.005) 0.010 (0.005) 0.027 (0.011) number of drinks in past month 0.0005 0.001* 0.002* Marginal Effect Standard Error (0.0004) (0.0003) (0.0007) Pvalue of overidentification test 0.723 0.015 0.514 Hausman statistic 0.999 1.921** 3.550* Coefficient (Standard Error) of omitted IV 0.015 (0.005) 0.009 (0.005) 0.029 (0.011) binge drinking 0.035 0.062* 0.145* Marginal Effect Standard Error (0.029) (0.026) (0.040) Pvalue of overidentification test 0.572 0.033 0.336 Hausman statistic 0.116 1.206 2.836* Coefficient (Standard Error) of omitted IV 0.016 (0.005) 0.010 (0.005) 0.017 (0.008) abuse/ dependence on alcohol 0.104** 0.145* 0.371* Marginal Effect Standard Error (0.056) (0.050) (0.091) Pvalue of overidentification test 0.468 0.028 0.905 Hausman statistic 1.570 2.577* 3.879* Coefficient (Standard Error) of omitted IV 0.013 (0.005) 0.009 (0.006) 0.028 (0.010) *Statistically significant at 1% **Statistically significant at 10% The Effects of Drinking on School Enrollment (1825 sample) Table 13 shows the findings for enro llment for the 1825 age group while employing all three instruments. For each daily increase in reported past year drinking, the probability of being enrolled is subsequently lowered by 0.0004. While for each additional drink increase in the number of drinks the respondent consumed in the past month, the probability of enrollment is lowered by 0.0008. The probability of enrollment is reduced by 0.06 percentage points for those that report bi nge drinking in the previous 30 day period. For binge drinkers in this sample, the probability of school enrollment is reduced by 14 53
PAGE 60
percent compared to nonbinge drinkers. Fo r those categorized as abusive/dependent on alcohol, the probability of enrollment fa lls by approximately 0.10 points, i.e. 23 percent compared to those not abusive/dependent. Table 13. IV/ OLS estimates of drinking on school enrollment (1825 years old) (all three instruments) (n=28,065) Alcohol variables IV OLS number of days drankpast year 0.0004*0.0001* Marginal Effect Standard Error (0.0001)(0.00003) Pvalue of overidentification test 0.000 Hausman statistic 2.865* number of drinks in past month 0.0008*0.0003 Marginal Effect Standard Error (0.0002)(0.00003) Pvalue of overidentification test 0.000 Hausman statistic 3.609* binge drinking 0.063* 0.002 Marginal Effect Standard Error (0.019) (0.005) Pvalue of overidentification test 0.000 Hausman statistic 3.522* abuse/ dependence on alcohol 0.099** 0.003 Marginal Effect Standard Error (0.043) (0.006) Pvalue of overidentification test 0.000 Hausman statistic 2.457** *Statistically significant at 1% **Statistically significant at 5% The pvalues associated with the overi dentification tests indicate that the assumption of exogeneity is not at all supporte d when all three inst ruments are utilized. Results of the Hausman tests revel that there are statistically significant differences between IV and OLS estimates. As with the high school sample, there is a strong indication that alcohol consumption, plausibly by raising the opportu nity cost of post high school education, causally and negatively impacts college level enrollment. And, considering the resources the student devotes toward drinking, particul arly if he/ she is abusive/ dependent on alcohol, the probability of post high school enroll ment is also lessened to some degree. 54
PAGE 61
Instrument Robustness and School Enrollment (1825 sample) Table 14 shows the results of regressi ons performed with varying pairs of instruments, which parall el those in table 12. Table 14. IV estimates of drinking on enro llment using IV pairs (1825 years old) (n=28,065) religion important religious decisionsreligion important & and ris k and ris k religious decisions Alcohol variables number of days drankpast year 0.0001 0.0003** 0.001*** Marginal Effect Standard Error (0.0001) (0.0001) (0.0003) Pvalue of overidentification test 0.684 0.000 0.220 Hausman statistic 1.790*** 1.081 4.924* Coefficient (Standard Error) of omitted IV 0.025 (0.006) 0.012 (0.005) 0.055 (0.013) number of drinks in past month 0.0002 0.0006** 0.002** Marginal Effect Standard Error (0.0002) (0.0002) (0.0005) Pvalue of overidentification test 0.710 0.000 0.494 Hausman statistic 0.768 2.578* 3.609* Coefficient (Standard Error) of omitted IV 0.025 (0.005) 0.012 (0.005) 0.051 (0.013) binge drinking 0.012 0.046** 0.211** Marginal Effect Standard Error (0.023) (0.020) (0.045) Pvalue of overidentification test 0.642 0.0002 0.186 Hausman statistic 0.685 2.458** 4.711* Coefficient (Standard Error) of omitted IV 0.025 (0.006) 0.013 (0.006) 0.044 (0.012) abuse/ dependence on alcohol 0.026 0.053 0.202** Marginal Effect Standard Error (0.049) (0.046) (0.090) Pvalue of overidentification test 0.171 0.000 0.000 Hausman statistic 0.403 1.311 2.309** Coefficient (Standard Error) of omitted IV 0.027 (0.006) 0.014 (0.006) 0.014 (0.010) *Statistically significant at 1% **Statistically significant at 5% ***Statistically significant at 10% The first column, which uses religion bei ng important and risk as instruments and includes religious decisions in the grade equation, is the only model for which overidentification tests ar e always insignificant. It also yields the most conservative IV estimates. These again suggest that OLS is consistent and efficient, which in turn suggests that drinking does not significantly impact college enrollment for high school graduates. 55
PAGE 62
56 C. Drinking and Absenteeism This section presents results for the e ffect of youth drinking on the number of school days the student missed due to ski pping classes and the days missed due to illness or injury for students currently enro lled in school. This outcome variable is described in more detail in chapter four. The analysis is conducted using a sample of high school age students of 16 to 19 years old and college age students of 18 to 25 years old. The causal effect that drinki ng has on this variable is analyzed using the three instrumental variables listed a bove. The main results of the IV analysis are also compared with parameter estimates obtained using OL S methodology. The discussion that follows concentrates on the effects of alcohol cons umption. Appendices 8 and 9 present all probit estimates for the 1825 and 1619 year old age groups respectively. Appendices 10 and 11 present coefficients and standard erro rs of all explanator y variables for both absenteeism factors, with binge drinking as the selected alcohol use measure for both the 1619 and 1825 samples respectively. First Stage Regression Results Table 15 presents the probit and first stag e results for the sample of 1825 year old respondents who are enrolled in college.
PAGE 63
Table 15. Probit/ First stage estima tes for absenteeism (1825 years old) (n=8,817) number of days number of drinks BingeAbuse/ Dependence exogeneous variables drank in past year in past month drinkingon alcohol Risk of bodily harm from drinking 34.537 22.212 0.283 0.128 (1.523) (1.398)(0.011) (0.009) Respondent states religion is important in life 4.510 0.517 0.020 0.006 (1.983) (1.821)(0.015) (0.011) Respondent states religion influences decisions 13.585 8.587 0.126 0.054 (1.874) (1.726)(0.015) (0.011) F stat/ chi2coefficient of joint significance 253.33 104.62668.240 255.250 Pvalue of significance level (0.0000) (0.0000)(0.0000)(0.0000) p redicted drinking coefficient 1.011 0.999 (0.035) (0.056) For this age group, of those w ho perceive that there is moderate to great risk of harm from consuming 45 drinks almost ever y day, the number of days drinking occurred in the past year is lowered by 34 days. The num ber of drinks consumed in the past month is reduced by 22, while the likelihood of binge drinking in the last 30 days falls by 0.28. The likelihood of being categorized as abusive/dependent on alcohol falls by 0.13. Importance of religious beliefs reduces all alcohol consumption measures. For respondents who report that religion is importa nt in life, the number of days drinking occurred in the past year is lowered by 4.5 days. The number of drinks consumed in the past month is reduced by 0.5, while the probabi lity of binge drinking in the last 30 days falls by 0.02. The likelihood of being categorized as abusive/ depende nt on alcohol falls by approximately 0.006. When religiosity impacts decisions, the effects on drinking are more pronounced than the effects when importanc e of religious beliefs is utilized as an IV. The number of days drinking occurred in the past year is lowered by 13.6 days. The number of drinks consumed in the past month is reduced by 8.6, while the probability of binge drinking in the last 30 days falls by 0.13. The likelihood of being categorized as abusive/dependent 57
PAGE 64
on alcohol declines by 0.054. The F statistics and 2 coefficients and associated pvalues indicate the instruments are jointly signi ficant for all the drinking measures. The predicted drinking coefficients in the binge drinking and abuse/dependence first stage regressions are 1.01 and 1.00, resp ectively. Table 16 presents the probit and first stag e results for the sample of 1619 year old high school enrollees. Table 16. Probit/ First stage estima tes for absenteeism (1619 years old) (n=10,039) number of days number of drinksBi ngeAbuse/ Dependence exogeneous variables drank in past year in past month drinkingon alcohol Risk of b odily harm from drinking 19.795 13.035 0.171 0.091 (1.166) (1.033) (0.010) (0.008) Respondent states religion is important in life 9.798 3.138 0.020 0.013 (1.346) (1.277) (0.010) (0.008) Respondent states religion influences decisions3.020 4.301 0.090 0.049 (1.442) (1.192) (0.010) (0.008) chi2coefficient of joint significance 153.670 77.030 515.080 260.640 Pvalue of significance level (0.0000) (0.0000)(0.0000)(0.0000) p redicted drinking coefficien t 1.027 1.007 (0.038) (0.055) For respondents who agreed that there is moderate to great risk of harm from consuming 45 drinks almost ev ery day, the number of days drinking occurred in the past year is lowered by about 20 days. The number of drinks consumed in the past month is reduced by 13, while the probabili ty of binge drinking in th e last 30 days falls by 0.17 percentage points. The likelihood of being ca tegorized as abusive/dependent on alcohol also falls by 0.09 points. Importance of religious beliefs and relig iously influenced decisions reduce all alcohol use measures. For respondents who report that religion is im portant in life, the number of days drinking occurred in the past year is lowered by about 9.8 days. The number of drinks consumed in the past month is reduced by 3.1, while the number of 58
PAGE 65
59 binge drinking episodes in the last 30 days falls by 0.02. The likelihood of being categorized as abusive/ depe ndent on alcohol falls by 0.013. When religiosity impacts decisions, the number of days drinking occurs in the past year is reduced by three days. The number of drinks consumed in the past month is reduced by about 4.3, while the probability of binge drinking in the last 30 days falls by 0.09. The likelihood of being categorized as abusive/dependent on alcohol falls by 0.05. Again, the F statistics and 2 pvalues indicate support for the hypothesis of joint instrument significance for all the drinking m easures. The predicted drinking coefficients in the binary measure first stage models ar e 1.03 for binge drinking in the past 30 days and 1.01 for abuse/dependence. The Effects of Drinking on Absenteeism (1619 sample) The findings in table 17 show that yout h drinking among 16 to 19 year old high school students leads to increases in abse nteeism reported in the past 30 days. The regression results show that an additional day increase in the number of past year drinking days elevates days skipped by 0.013 and days missed because of illness by approximately 0.007, relative to refraining from drinking. For each additional drink increase in the number of drinks the responde nt consumed in the past month, days missed because of skipping rise by 0.02 and days mi ssed due to illness increase by 0.01. An additional day of drinking in the prior mont h leads to a 2.3 percent increase in days skipped and an additional one pe rcent increase in sick days. Binge drinking and abuse/dependence on al cohol further increase truancy. For students who have engaged in binge drinking, the number of days missed due to skipping is elevated by 1.42 days. For those classified as abusive/dependent w ith respect to alcohol
PAGE 66
use, the amount of days missed due to skippi ng escalates to 2.6 days per month compared to those not abusive/dependent. School days missed due to illness rise by 0.62 days for binge drinkers and 1.2 days for thos e who are alcohol abusive/dependent. Table 17. IV/ OLS estimates of drinking on absenteeism (1619 years old) (all three instruments) (n=10,039) days missed due to skippingdays missed due to illness Alcohol variables IV OLS IVOLS number of days drankpast year 0.013*0.004*0.007*0.002* Marginal Effect Standard Error (0.001)(0.0003)(0.002)(0.0004) Pvalue of overidentification test 0.008 0.004 Hausman statistic 5.102* 2.278** number of drinks in past month 0.020*0.003*0.010*0.002* Marginal Effect Standard Error (0.003)(0.0003)(0.003)(0.0005) Pvalue of overidentification test 0.000 0.001 Hausman statistic 5.489* 2.548** binge drinking 1.418*0.630*0.622*0.332* Marginal Effect Standard Error (0.186)(0.041)(0.197)(0.056) Pvalue of overidentification test 0.007 0.003 Hausman statistic 4.681* 1.978** abuse/ dependence on alcohol Marginal Effect Standard Error 2.616*0.727*1.202*0.441* Marginal Effect Standard Error (0.360)(0.051)(0.368)(0.070) Pvalue of overidentification test 0.016 0.005 Hausman statistic 5.412* 2.183** *Statistically significant at 1% **Statistically significant at 5% The overidentification test results are somewhat weak for this specification. For all drinking measures, instrument exogeneity is rejected at the 5 percent level. The Hausman tests all afford the same genera l result for all drinki ng measures and both truancy variables: statistically significant differences are present between IV and OLS estimates. 60
PAGE 67
61 Instrumental Variable Robustness and Absenteeism (1619 sample) The effects of drinking on absenteeis m using differing combinations of instruments are shown in table 18. The estima tes in the first column are uniformly, and uniquely, reliable with regard to having overi dentification test pva lues well above 0.1. This again points to religious decisions bei ng a poor choice for use as an instrument. With religious decisions included in the absenteeism equations, the IV estimates are again conservative with respect to the others in table 18 as well as those in table 17. For skipping, these estimates are all significantly positive and distinct from the corresponding OLS estimates. This implie s that drinking among high school students raises the propensity to skip classes. OLS estimates appear to have a positive bias. Again, it could be that higher ability and in come students drink more and skip school less, or that measurement error in the OL S estimates imparts severe downward bias. For illness, the estimates in the first column are insignificant for all drinking measures. The days and drinks coefficients are identical to th ose under OLS, so one could still legitimately conclude that additional drinking days or monthly drinks induces school absences due to illness, though not near ly as much as they increase days skipped. In contrast, the IV estimates for the binary drinking variables are smaller in magnitude than are the OLS estimates and statistically insignificant, which suggests that binge drinking and abuse/dependence do not increase illnessinduced school absences. A lack of effect of heavy drinking, though, is inconsistent with a significant effect of an additional day of drinking or dr ink. Thus, the safest inferenc e to make is that drinking does not necessarily causally influence school absences arising from illness.
PAGE 68
62 Table 18. IV estimates of drinking on absenteeism using IV pairs (1619 years old) (n=10,039) days missed due to skipping religion importantreligious decisionsreligion important & and risk and risk religious decisions Alcohol variables number of days drankpast year 0.009* 0.012* 0.019* Marginal Effect Standard Error (0.002) (0.002) (0.003) Pvalue of overidentification test 0.233 0.002 0.092 Hausman statistic 2.692* 4.107* 4.769* Coefficient (Standard Error) of omitted IV 0.126 (0.045)0.018 (0.047) 0.188 (0.080) number of drinks in past month 0.014* 0.019* 0.033* Marginal Effect Standard Error (0.003) (0.003) (0.006) Pvalue of overidentification test 0.137 0.001 0.021 Hausman statistic 3.267* 4.453* 4.322* Coefficient (Standard Error) of omitted IV 0.151 (0.044)0.026 (0.050) 0.230 (0.111) binge drinking 1.059* 1.365* 1.862* Marginal Effect Standard Error (0.223) (0.220) (0.293) Pvalue of overidentification test 0.247 0.001 0.007 Hausman statistic 2.249** 3.694* 4.681* Coefficient (Standard Error) of omitted IV 0.129 (0.044)0.026 (0.045) 0.127 (0.068) abuse/ dependence on alcohol 2.015* 2.583* 3.484* Marginal Effect Standard Error (0.432) (0.432) (0.596) Pvalue of overidentification test 0.207 0.004 0.082 Hausman statistic 3.090* 4.412* 4.704* Coefficient (Standard Error) of omitted IV 0.115 (0.046)0.008 (0.048) 0.140 (0.074) days missed due to illness religion importantreligious decisionsreligion important & and risk and risk religious decisions Alcohol variables number of days drankpast year 0.002 0.006** 0.016** Marginal Effect Standard Error (0.002) (0.002) (0.004) Pvalue of overidentification test 0.306 0.001 0.098 Hausman statistic 0.141 1.556 3.214* Coefficient (Standard Error) of omitted IV 0.183 (0.064)0.038 (0.061) 0.277 (0.111) number of drinks in past month 0.003 0.008** 0.027** Marginal Effect Standard Error (0.003) (0.003) (0.008) Pvalue of overidentification test 0.282 0.000 0.032 Hausman statistic 0.288 1.612 3.038* Coefficient (Standard Error) of omitted IV 0.190 (0.060)0.049 (0.062) 0.304 (0.137) binge drinking 0.046 0.499** 1.066** Marginal Effect Standard Error (0.253) (0.234) (0.362) Pvalue of overidentification test 0.220 0.001 0.015 Hausman statistic 0.975 0.933 2.176** Coefficient (Standard Error) of omitted IV 0.207 (0.064)0.061 (0.059) 0.127 (0.095) abuse/ dependence on alcohol 0.143 1.003* 2.115** Marginal Effect Standard Error (0.484) (0.445) (0.679) Pvalue of overidentification test 0.240 0.001 0.030 Hausman statistic 0.563 1.343 2.520** Coefficient (Standard Error) of omitted IV 0.202 (0.066)0.051 (0.061) 0.147 (0.094) *Statistically significant at 1% **Statistically significant at 5%
PAGE 69
The Effects of Drinking on Absenteeism (1825 sample) The findings in table 19 show that yo uth drinking among respondents 18 to 25 years old leads to increases in absenteeism. Overall there are positive effects on absenteeism due to skipping classes and because of illness or injury. The regression results show that the drinking coefficien ts are statistically significant for both absenteeism measures. Table 19. IV/ OLS estimates of drinking on absenteeism (1825 years old) (all three instruments) (n=8,817) days missed due to skipping days missed due to illness Alcohol variables IV OLS IV OLS number of days drankpast year 0.006*0.003*0.003*0.001* Marginal Effect Standard Error (0.001)(0.0003)(0.001)(0.0002) Pvalue of overidentification test 0.005 0.170 Hausman statistic 2.501** 1.361 number of drinks in past month 0.011*0.003*0.004**0.0004 Marginal Effect Standard Error (0.002)(0.0003)(0.001)(0.0003) Pvalue of overidentification test 0.009 0.095 Hausman statistic 3.657* 2.298** binge drinking 0.954*0.398*0.352**0.130* Marginal Effect Standard Error (0.179)(0.046)(0.150)(0.042) Pvalue of overidentification test 0.011 0.086 Hausman statistic 3.486* 1.642 abuse/ dependence on alcohol 1.888*0.626*0.653**0.302* Marginal Effect Standard Error (0.348)(0.056)(0.270)(0.051) Pvalue of overidentification test 0.010 0.122 Hausman statistic 3.777* 1.383 *Statistically significant at 1% **Statistically significant at 5% For each extra day of alcohol use in th e past year, the number of days skipped increases by 0.006 and days missed due to illness increases by 0.003 days. For each additional drink increase in the number of drinks the respondent consumed in the past month, the days missed because of skipping and illness rise by 0.011 and 0.004 63
PAGE 70
64 respectively. Respondents that had one dr ink in the prior month experience an approximate one percent increase in days skipped and an approximate one percent increase in sick days, compared to th ose who did not drink in the past month. Binge drinking and abuse/dependence on al cohol further increase skipping and days missed because of illness. For students who have engaged in binge drinking, the number of days missed due to skipping is elevated by approximately one day and days missed due to illness rises by approximately onethird of a day. For those classified as abusive/dependent with respect to alcohol, the number of days missed due to skipping increases by approximately two days a nd days missed due to illness rises by approximately twothirds of a day. The pvalues associated with the over identification tests for the days missed due to skipping model offer little support for the assumption of instrument exogeneity. At the 5 percent level, instrument exogeneity is rejected for all drinking measures. Hausman tests generally show that there are statistically significant differences between IV and OLS estimates. For days missed due to illness, the revers e is true. There is stronger support for the hypothesis of instrument exogeneity, which is never rejected at th e 5 percent level. However, Hausman statistics show an insi gnificant difference between IV and OLS estimates at the 5 percent level excep t in the case of past month drinks. Instrumental Variable Robustness and Absenteeism (1825 sample) The effects of drinking on absenteeis m using differing combinations of instruments are shown in table 20. For ski pping, overidentificati on test results are generally unconvincing, making it difficult to draw any firm conclusions. For illness,
PAGE 71
65 overidentification test s uniformly fail to reject the hypot hesis that the instruments are valid, but the conservative first column estim ates are not significantly different from OLS. Because they are larger than thos e from OLS, however, and OLS estimates are significantly positive, we can conclude that drinking raises illnessrelated absences among college students, but that th ese effects are relatively small. Overall, the results offered in this chapter demonstrate that youth alcohol consumption impedes the acquisition of human capital. The probability of achieving an A average diminishes with youth alcohol co nsumption, while the pr obability of earning a grade of C or lower is actually elevated by youth drinking. The effect drinking has on the other educat ion outcomes is also significant. For the probability of enrollment, all drinking measures produce a negative impact for both the high school age and college age samples. The analysis also reveals that elevated alcohol consumption engenders increased skipp ing classes for thos e that are attending school. And the findings provide some eviden ce that alcohol consumption increases school days missed due to illness. The robustness analyses suggest that these described effects are sensitive, to some degree, to the choice of instruments. Identificat ion tests indicate that instrument pairs are more plausibly exogenous in some specifications than in others. Statistically significant differences exist among IV and OLS parameter estimates in some, but not all, specifications.
PAGE 72
66 Table 20. IV estimates of drinking on absenteeism using IV pairs (1825 years old) (n=8,817) days missed due to skipping religion importantreligious decisionsreligion important & and risk and risk religious decisions Alcohol variables number of days drankpast year 0.005* 0.007* 0.009* Marginal Effect Standard Error (0.001) (0.001) (0.002) Pvalue of overidentification test 0.009 0.002 0.003 Hausman statistic 1.254 2.526 ** 2.363** Coefficient (Standard Error) of omitted IV0.105 (0.052)0.052 (0.054) 0.148 (0.109) number of drinks in past month 0.008* 0.011* 0.023* Marginal Effect Standard Error (0.002) (0.002) (0.006) Pvalue of overidentification test 0.035 0.002 0.082 Hausman statistic 2.507** 3.478* 2.990* Coefficient (Standard Error) of omitted IV0.116 (0.051)0.012 (0.053) 0.328 (0.168) binge drinking 0.785* 0.963* 1.920* Marginal Effect Standard Error (0.187) (0.189) (0.430) Pvalue of overidentification test 0.043 0.002 0.106 Hausman statistic 2.371** 3.339* 3.667* Coefficient (Standard Error) of omitted IV0.109 (0.050)0.010 (0.050) 0.319 (0.127) abuse/ dependence on alcohol 1.579* 1.977* 3.165* Marginal Effect Standard Error (0.369) (0.378) (0.793) Pvalue of overidentification test 0.019 0.003 0.023 Hausman statistic 2.707* 3.701* 3.254* Coefficient (Standard Error) of omitted IV0.095 (0.051)0.040 (0.052) 0.228 (0.123) days missed due to illness religion importantreligious decisionsreligion important & and risk and risk religious decisions Alcohol variables number of days drankpast year 0.002 0.002*** 0.007* Marginal Effect Standard Error (0.001) (0.001) (0.002) Pvalue of overidentification test 0.571 0.247 0.890 Hausman statistic 0.447 0.851 2.374** Coefficient (Standard Error) of omitted IV0.086 (0.051)0.070 (0.049) 0.187 (0.107) number of drinks in past month 0.002 0.003*** 0.014* Marginal Effect Standard Error (0.003) (0.002) (0.005) Pvalue of overidentification test 0.462 0.246 0.444 Hausman statistic 1.158 1.675*** 2.280** Coefficient (Standard Error) of omitted IV0.094 (0.049)0.084 (0.046) 0.249 (0.148) binge drinking 0.199 0.278*** 0.770*** Marginal Effect Standard Error (0.171) (0.158) (0.369) Pvalue of overidentification test 0.421 0.252 0.180 Hausman statistic 0.537 1.101 1.800*** Coefficient (Standard Error) of omitted IV0.099 (0.048)0.088 (0.045) 0.137 (0.118) abuse/ dependence on alcohol 0.326 0.470*** 1.051*** Marginal Effect Standard Error (0.312) (0.293) (0.624) Pvalue of overidentification test 0.415 0.220 0.144 Hausman statistic 0.128 0.630 1.229 Coefficient (Standard Error) of omitted IV0.100 (0.048)0.084 (0.046) 0.071 (0.109) *Statistically significant at 1% **Statistically significant at 5% ***Statistically significant at 10%
PAGE 73
67 Chapter Six: Summary and Conclusions The economic implications surrounding human capital have intrigued economists for centuries, even dating back to the classical school of economic thought. More specifically, the fields of health economics and labor economics have both examined how physical and mental health may impact hum an capital formation. Issues involving the consequences of substance use on human capita l have been researched in economics only within the previous fifteen years, and the current resear ch on alcohol use and human capital suffers from two important shortcomi ngs. First, existing research focuses largely on alcohol use among college students, leaving the impact of drinking among high school students largely unaddressed. Th e literature on adolescence ha s provided some evidence that alcohol use begins in th e early teen years for many students. Failure to examine the experience of high school students leaves an important gap in understanding the relation between alcohol consumption a nd educational achievement. Second, while past research has establis hed a negative link be tween drinking and educational achievement, many of these studie s have not accounted for the possibility that the negative correlation between drinking and educati onal achievement may be the result of unobserved variables that cause simultaneous increases in drinking and reductions in educational achievement. And, fo r those studies that have incorporated adjustments for endogeneity, the analyses ha ve been conducted utilizing instrumental variable procedures that have been subject to criticism.
PAGE 74
68 The first shortcoming is addressed by us e of data from the National Survey of Drug Use and Health (NSDUH). The survey contains many variables pertaining to the behavior and attitudes of stude nts, especially with respect to alcohol use and educational achievement. There is an entire subset of th e data devoted only to surveying 12 to 17 year old students and many data pert aining to older students is compiled as well. The NSDUH data also contain several potential variables th at can serve as instruments, which is central to the empirical strategy employed in this dissertation. Despite these advantages, or perhaps because the existing research largel y neglects high school students, the dataset has not been widely utilized by other researchers in this topic area. The second deficiency is addressed in th is study by the use of an instrumental variable estimation technique. Specifically, the technique of instrumental variables is a statistical method designed to estimate the causal impact an i ndependent variable has on a dependent variable when omitted variable bias and/ or reverse causation is present. By employing such a technique the researcher can more accurately gauge the causal effect the independent variable has on the dependent variable. Thus, the potential problem of falsely concluding that a causal negative relationship exists between drinking and education is mitigated, as are potential biases in parameter estimates. This dissertation investigat es both of these shortcomings for the purpose of improving and extending empirical knowledge about the consequences of alcohol consumption on educational achievement and to derive more accurate estimation results. The consequences of alcohol consumption fo r educational performance are illustrated first in the IV regression estimates using the probabilities of earning an A or a C or lower grade as educational outcomes. Results show a strong negative effect of drinking
PAGE 75
69 on the probability of earning an A and fairly strong positive effects on the likelihood of obtaining a C or below. Overidentificati on tests offer little evidence to support the hypothesis of exogeneity. Results derived fr om robustness analyses do provide some support for the assumption of instrument exogeneity when the outcome variable is the probability of earning an A. Evidence of exogeneity is weaker when the outcome variable is the probability of a C or lowe r average. In addition, pvalues of Hausman coefficients consistently show that there are significant differences among IV and OLS estimates, and in general OLS underestimates the impact drinking has on grade measures. This study also analyzes other determin ants of human capital development such as school enrollment and absenteeism due to skipping classes and reported illness. The effect of student drinking diminishes the pr obability the student is attending school. This result holds for both the 1619 and 1825 age groups. Drinking has positive effects on absenteeism for both samples. For both absenteeism measures, overidentification tests for the main specification of the 1619 year old sample offers weak evidence to support the hypothesis of instrument exogeneity. When IV sensitivity is evaluated however, some of the instrument specifications can afford evidence in support of the exogene ity assumption. For the 1825 year old sample, the overidentification tests in the main speci fication indicate suppor t for the hypothesis of instrument exogeneity for the days missed due to illness outcome; the support is weaker for the days missed due to skipping outcome. Wh en instrument sensitiv ity is investigated, there is virtually no support for the exogeneity hypothesis wi th respect to days missed due to skipping, though there is stronger evid ence of exogeneity with respect to days missed due to illness.
PAGE 76
70 The fundamental results in th is dissertation confirm the findings of some previous studies in the literature. For instance, the gene ral results presented he re parallel those of Wolaver (2002), Williams, et al. (2003) and DeSimone and Wolaver (2004) in that youth drinking has negative effects on grades. The negative effects on school enrollment outlined in this study confirm, at least in general terms, the Cook and Moore (1993) result that heavy drinking reduced subsequent sc hooling, and contravene the Dee and Evans (2003) conclusion that drinking had no disti nguishable effect on e ducational measures such as high school and college enrollment. This study does corroborate Roebuck (2004), as one of his results show that alcohol consumed in a previous year reduces the probability of subsequent school enrollment. Limitations Two primary limitations hamper some of the findings. First, in some of the specifications, the impact drinking has on the ed ucational outcome variable is quite large. This may give some researchers pause with rega rd to the plausibility of the magnitude of the parameter estimates. Another difficulty exis ts with respect to instrumental variable exogeneity. For grade probabilities, there is in general only weak support for the assumption of instrument exogeneity. Fo r the enrollment outcome, there is some evidence to support instrument exogeneity for the both 1619 year ol d and 1825 year old subsamples. However, support for the exogeneity assumption is dependent on the variables selected as instruments. Generall y, the same can be said for the absenteeism analysis.
PAGE 77
71 Policy Implications Overall, the results derived from the re gressions directly imply that alcohol consumption on the part of teenagers and young adults up to age 25 harms human capital attainment. Therefore, any economic or other gains to individuals and society from accruing education could also be reduced. Wh ile other factors that reduce educational achievement are certain to interact with alc ohol use, the direct im pact of drinking is shown to have large effects. The results have further implications that are applicable to policymaking. First, the estimated effects of drinking on schooling convey information regarding the external benefits on educational outcomes of policies that effectivel y inhibit alcohol consumption among youths. Identifying whether alcohol use directly leads to lower achievement or more destructive schoolrelated be havior, or is merely spuri ously correlated with worse educational outcomes through unobser ved variables that influence both sets of behaviors, have been addressed and there exists evidence that causal effects are present. Policies that reduce heavy drinking also become more attrac tive because the benefits of such policies include the value to societ y of the resulting gains in educational performance. Second, the effects of the variables that are utilized as instruments in the IV procedure provide an indication as to various specific endeav ors that might be successful in reducing alcohol consumption among youth. Parental disapprova l of drinking is negatively and strongly correla ted with youth drinking, ther efore programs encouraging parental discouragement of youth drinking c ould be beneficial. Given the discovered impact of religiosity on drinking behaviors, policies and/or social endeavors aimed at encouraging religious activity would have the added benef it of alcohol use reductions.
PAGE 78
72 The importance of perceived risk in usi ng alcohol suggests that campaigns that raise awareness of the risks inherent in consuming alcohol would be fruitful as a prospective policy tool. Strong relationships with perceived peer use of alcohol signal that the dissemination of information rega rding actual drinking prevalence among students might be warranted. The significan ce of peer drinking measures would imply that any policy able to diminish underage alcohol use could have social multiplier effects that enhance the attr activeness of such policies. Third, the finding of a significant causal effect of drinking on human capital accumulation has implications for several fields of economics such as labor and educational economics. These include, for ex ample, reducing the unexplained variation across individuals in wage and earnings equations. There are some avenues for future res earch that emanate from this study. If researchers can obtain data on other educational outcomes not covered in the NSDUH, the methodology presented in this dissertati on should prove useful. For example, an analyst could evaluate the relation be tween teen drinking and performance on standardized tests such as the SAT and AC T. While the IV strategy employed met with limited success, further investigation of the causal effects drinking may have on an economic outcome variable should be accompanied by utilizing potentially more exogenous instruments. Also, information provided by the relati onship between youth drinking behaviors and variables such as peer alcohol use, parental disapproval of teen drinking, and religious influences shown in this study s hould prove useful in conducting cost/benefit analyses of government programs that aim to curb alcohol use among youths. Use of
PAGE 79
73 instrumental variable models similar to those in this dissertation, perhaps utilizing factors such as parental divorce and sibling attitudes and behavior s not available in the NSDUH, would likely provide valuable guidance to policymakers. In sum, this study provides a plethora of information regarding the educational consequences stemming from alcohol use among teenagers and young adults. The findings presented indicate th at youth drinking significantly and negatively effects human capital accumulation. Even in areas where the results are not as strong, there remains substantial information that addresses the de terminants of youth drinking. The research questions presented in this dissertation also generate topics for future investigation in health economics and other related subject areas.
PAGE 80
74 References Becker, G. S. (1964). Human Capital : (Columbia University Press for the National Bureau of Economic Research, New York). BenPorath Y.(!967). The produc tion of human capital and the life cycle of earnings. Journal of Political Economy 75: 353:367. Bryant, A., Schulenberg, J., O'Malley, P., Bachman, J., & Johnston, L. (2003). How academic achievement, attitudes, and behavior s relate to the course of substance use during adolescence: A 6year, multiwave longitudinal study. Journal of Research on Adolescence 13(3), 361. Cassady, P. (2001). Selfreported GPA and SAT scores: A methodological note. Practical Assessment, Research & Evaluation 7(12). Chatterji, P. (2003). Illicit drug use and educational attainment. Working paper 10045, National Bureau of Economic Research. Chatterji, P. and DeSimone, J. (2005). Adolescent drinking and high school dropout. Working paper 11337 National Bureau of Economic Research. Cook, P. J., & Moore, M. J. (1993). Drinking and schooling. Journal of Health Economics, 12(4), 411. Dee, T. S., & Evans, W. N. (2003). Teen drinking and educationa l attainment: Evidence from twosample instrumental variables estimates. Journal of Labor Economics 21(1), 178.
PAGE 81
75 DeSimone, J., & Wolaver, A. M. (2005). Drinking and academic performance in high school. Working Paper 11035, National Bureau of Economic Research. Evans, W., Oates, W., & Schwab, R. (1992). Measuring peer group e ffects: A study of teenage behavior. Journal of Political Economy 100(51), 966. Gaviria, A., & Raphael, S. (2001). Schoolbas ed peer effects and juvenile behavior. Review of Economics and Statistics 83(2), 257. Grant, B. F., Harford, T. C., & Grigson, M. B. (1988). Stability of alcohol consumption among youth: a national longitudinal survey. Journal of Studies on Alcohol 49(3), 253. Grossman, M. (1972). On the concept of h ealth capital and the demand for health. Journal of Political Economy, 80(2), 223255. Harrison, L., & Hughes, A. (1997). The validity of selfreported drug use: Improving the accuracy of survey estimates. NIDA Research Monograph 167, 1. Hausman, J. (1978). Specification tests in econometrics. Econometrica 46(6), 12511271. Hoyt, G., & Chaloupka, F. (1994). Effect of su rvey conditions on selfreported substance use. Contemporary Economic Policy 12(3), 109. Jeynes, W. (2002). The relationship between the consumption of various drugs by adolescents and their academic achievement. American Journal of Drug and Alcohol Abuse 28(1), 15. Johnston, L. D., OMalley, M., & Bachman, J. G. (1988). Illicit drug use, smoking and drinking by Americas high school student s, college Students and young adults. Washington, DC: USGPO.
PAGE 82
76 Koch, S. F., & Ribar, D. C. (2001). A sib lings analysis of the effects of alcohol consumption onset on educational attainment. Contemporary Economic Policy 19(2), 162. Kremer, M., & Levy, D. (2003). Peer effect s and alcohol use among college students. Working paper 9876, National Bureau of Economic Research. Markowitz S. (2001). The role of alcohol and drug consumption in determining physical fights and weapon carrying by teenagers. Eastern Economic Journal 27(4), 409 432. Midanik, L. (1988). Validity of selfreporte d alcohol use: a lite rature review and assessment. British Journal of Addiction 83, 1019. Mincer, J. (1974). Schooling, Experience and Earnings. New York: National Bureau of Economic Research.. Mullahy, J., & Sindelar, J. L. (1994). Alcoholism and income: The role of indirect effects. The Milbank Quarterly 72(2), 359. Norton, E., Lindrooth, R., & Ennett, S. (1998). Controlling for the endogeneity of peer substance use on adolescent alcohol and tobacco use. Health Economics 36(7), 439 453. Reinisch, E. J., Bell, R. M., & Ellickson, P. (1991). How accurate are adolescent reports of drug use? Santa Monica, CA: Rand Corporation. Roebuck, C., French, M., & Dennis, M. ( 2004). Adolescent marijuana use and school attendance. Economics of Education Review 23(2), 133. Staiger, D., & Stock, J. H. (1997). Instru mental variables regression with weak instruments. Econometrica 65(3), 557.
PAGE 83
77 Williams, J., Powell, L. M., & Wechsler, H. (2003). Does alcohol consumption reduce human capital accumulation? Evidence from the college alcohol study. Applied Economics, 35(10), 1227. Wolaver, A. M. (2002). Effects of heavy drin king in college on study effort, grade point average, and major choice. Contemporary Economic Policy 20(4), 415. Wooldridge, J. M. (2003). Introductory econometrics: a modern approach (2 nd edition). SouthWestern College Publishing. Yamada, T., Kendix, M., & Yamada, T. (1996). The impact of alcohol consumption and marijuana use on high school graduation. Health Economics, 5(1), 77.
PAGE 84
78 Appendices
PAGE 85
Appendix 1. Probit estimates for the probability of 'A' and 'C' (n=18,231) Pseudo R2 = 0.11Pseudo R2 = 0.13Pseudo R2 = 0.14Pseudo R2 = 0.12 number of daysnumber of drinksBingeAbuse/ Dependence explantory variables drank in past yearin past monthdrinking on alcohol Risk of bodily harm from drinking 15.890 (0.894)10.420 (0.749)0.125 (0.008)0.078 (0.006) Peer use of alcohol 11.949 (0.716)6.241 (0.600)0.086 (0.005)0.054 (0.004) Parental disapproval of alcohol use 24.706 (0.716)14.201 (0.845)0.162 (0.005)0.074 (0.007) Mother in household 0.012 (0.014)0.006 (0.011)0.009 (0.009)0.002 (0.006) Father in household 0.051 (0.010)0.038 (0.009)0.027 (0.007)0.016 (0.005) Female 0.47 (0.007)0.010 (0.006)0.014 (0.005)0.006 (0.003) Grade in (10th grade) 0.027 (0.013)0.002 (0.011)0.002 (0.009)0.001 (0.006) Grade in (11th grade) 0.052 (0.021)0.028 (0.007)0.013 (0.009)0.008 (0.006) Grade in (12th grade) 0.073 (0.017)0.049 (0.015)0.028 (0.011)0.006 (0.007) Age of student (15 years old) 0.092 (0.035)0.054 (0.012)0.040 (0.011)0.026 (0.008) Age of student (16 years old) 0.153 (0.035)0.117 (0.015)0.096 (0.013)0.063 (0.011) Age of student (17 years old) 0.202 (0.016)0.158 (0.016)0.133 (0.014)0.083 (0.011) Race (African American) 0.153 (0.124)0.112 (0.009)0.096 (0.005)0.054 (0.004) Race (Native American) 0.015 (0.019)0.015 (0.004)0.013 (0.020)0.046 (0.002) Race (Asian) 0.227 (0.020)0.128 (0.013)0.090 (0.009)0.041 (0.007) Race (nonwhite Hispanic) 0.007 (0.124)0.014 (0.010)0.009 (0.007)0.004 (0.005) Number in family 0.016 (0.005)0.010 (0.004)0.003 (0.003)0.003 (0.002) Number in family (>5) 0.102 (0.022)0.060 (0.015)0.032 (0.013)0.011 (0.009) Family income ($20,000$49,999) 0.065 (0.012)0.054 (0.010)0.031 (0.008)0.020 (0.006) Family income ($50,000$74,999) 0.068 (0.013)0.064 (0.013)0.043 (0.017)0.020 (0.007) Family income ($75,000 or more) 0.117 (0.014)0.108 (0.013)0.056 (0.017)0.025 (0.007) MSA segement with 1+ million persons0.013 (0.010)0.013 (0.008)0.015 (0.006)0.009 (0.004) MSA segment of less than 1 million 0.014 (0.009).011 (0.008)0.009 (0.006)0.009 (0.004) Move (number of times in last 5years)0.027 (0.013)0.015 (0.002)0.008 (0.003)0.009 (0.002) Year 2002 indicator 0.059 (0.008)0.051 (0.008)0.048 (0.005)0.030 (0.004) Parents help with homework (always) 0.133 (0.013)0.088 (0.006)0.057 (0.008)0.047 (0.005) Parents help with homework (sometimes)0.081 (0.014)0.042 (0.010)0.014 (0.009)0.026 (0.004) Parents help with homework (seldom)0.032 (0.016)0.019 (0.011)0.014 (0.009)0.014 (0.005) (Standard errors are in parentheses) 79
PAGE 86
Appendix 2. All IV estimates on the probability of an 'A' for binge drinking (n=18,321) Explanatory variables IV coefficient (Marginal Effect SE) binge drinking 0.351 (0.030) Mother in household 0.005 (0.008) Father in household 0.010 (0.005) Female 0.031 (0.003) Grade in (10th grade) 0.045 (0.007) Grade in (11th grade) 0.085 (0.009) Grade in (12th grade) 0.109 (0.010) Age of student (15 years old) 0.042 (0.006) Age of student (16 years old) 0.062 (0.0008) Age of student (17 years old) 0.053 (0.009) Race (African American) 0.018 (0.007) Race (Native American) 0.005 (0.0020) Race (Asian) 0.012 (0.007) Race (nonwhite Hispanic) 0.006 (0.006) Number in family 0.003 (0.002) Number in family (>5) 0.015 (0.011) Family income ($20,000$49,999) 0.013 (0.007) Family income ($50,000$74,999) 0.038 (0.007) Family income ($75,000 or more) 0.054 (0.007) MSA segment with 1+ million persons 0.000 (0.004) MSA segment of less than 1 million 0.000 (0.004) Parents help with homework (always) 0.001 (0.007) Parents help with homework (sometimes) 0.000 (0.008) Parents help with homework (seldom) 0.000 (0.009) Year 2002 indicator 0.003 (0.003) Standard errors are in parentheses. 80
PAGE 87
Appendix 3. All IV estimates on the probability of an 'C' or lower for binge drinking (n=18,231) Explanatory variables IV coefficient (Marginal Effect SE) binge drinking 0.429 (0.036) Mother in household 0.020 (0.013) Father in household 0.033 (0.010) Female 0.133 (0.009) Grade in (10th grade) 0.041 (0.010) Grade in (11th grade) 0.107 (0.013) Grade in (12th grade) 0.16 (0.015) Age of student (15 years old) 0.065 (0.011) Age of student (16 years old) 0.084 (0.014) Age of student (17 years old) 0.172 (0.016) Race (African American) 0.141 (0.012) Race (Native American) 0.035 (0.019) Race (Asian) 0.033 (0.006) Race (nonwhite Hispanic) 0.055 (0.011) Number in family 0.014 (0.004) Number in family (>5) 0.059 (0.020) Family income ($20,000$49,999) 0.037 (0.011) Family income ($50,000$74,999) 0.094 (0.013) Family income ($75,000 or more) 0.157 (0.012) MSA segement with 1+ million persons 0.025 (0.009) MSA segment of less than 1 million 0.027 (0.008) Move (number of times in last 5 years) 0.025 (0.005) Parents help with homework (always) 0.045 (0.012) Parents help with homework (sometimes) 0.018 (0.013) Parents help with homework (seldom) 0.001 (0.014) Year 2002 indicator 0.005 (0.006) Standard errors are in parentheses. 81
PAGE 88
Appendix 4. Probit estimates for the probability of enrollment (1825 years old) (n=28,065) Pseudo R2 = .08Pseudo R2 = .09Pseudo R2 = .10Pseudo R2 = .06 number of daysnumber of drinksBingeAbuse/ Dependence explantory variables drank in past yearin past monthdrinkingon alcohol Risk of bodily harm from drinking 34. 717 (0.971)19.662 (0.843)0.242 (0.006)0.096 (0.005) Respondent states religion is important in life3.598 (1.221)0.954 (1.060)0.010 (0.008)0.008 (0.006) Respondent states religion influences decision s13.276 (1.156)8.946 (1.004) 0.106 (0.008)0.041 (0.007) Female 0.020 (0.004)0.089 (0. 006)0.168 (0.006)0.090 (0.005) Race (African American) 0.109 (0.007)0.130 (0.008)0.166 (0.008)0.062 (0.005) Race (Native American) 0.010 (0.021) 0.056 (0.003)0.006 (0.014)0.070 (0.021) Race (Asian) 0.190 (0.012)0.202 (0.001)0.179 (0.015)0.062 (0.001) Race (nonwhite Hispanic) 0.094 (0.007)0.109 (0.009)0.072 (0.008)0.014 (0.006) Age of student (19 years old) 0.023 (0. 008)0.036 (0.109)0.038 (0.008)0.022 (0.006) Age of student (20 years old) 0.058 (0 .007)0.070 (0.011)0.054 (0.008)0.000 (0.006) Age of student (21 years old) 0.106 (0.007)0.174 (0.010)0.136 (0.007)0.035 (0.007) Age of student (2223 years old) 0.058 (0.007)0.142 (0.007)0.096 (0.007)0.014 (0.007) Age of student (2425 years old) 0.106 (0.007)0.125 (0.007)0.068 (0.007)0.011 (0.007) Last grade completed (Freshman) 0.053 (0 .006)0.084 (0.008)0.038 (0.009)0.015 (0.005) Last grade completed (Sophomore/ Junior) 0 .071 (0.006)0.117 (0.009)0. 054 (0.008)0.025 (0.006) Number in family 0.030 (0.002)0.038 (0.003)0.024 (0.003)0.014 (0.002) Number in family (>5) 0.222 (0.015)0.240 (0.014)0.141 (0.013)0.060 (0.008) Family income ($20,000$49,999) 0.021 (0. 005)0.006 (0.0010)0.036 (0.008)0.028 (0.005) Family income ($50,000$74,999) 0.049 (0. 007)0.037 (0.007)0.017 (0.009)0.020 (0.006) Family income ($75,000 or more) 0.064 (0.007)0.072 (0.010)0.022 (0.009)0.001 (0.007) MSA segment with 1+ million persons 0. 028 (0.006)0.043 (0.007)0.002 (0.008)0.006 (0.005) MSA segment of less than 1 million 0. 034 (0.005)0.045 (0.006)0.021 (0.007)0.009 (0.005) Year 2002 indicator 0.004 (0.004)0. 005 (0.006)0.001 (0.005)0.011 (0.004) Standard errors are in parentheses. 82
PAGE 89
Appendix 5. Probit estimates for the probability of enrollment (1619 years old) (n=13,562) Pseudo R2 = 0.06Pseudo R2 = 0.08Pseudo R2 = 0.09Pseudo R2 = 0.06 number of daysnumber of drinksBingeAbuse/ Dependence explanatory variables drank in past yearin past monthdrinking on alcohol Risk of bodily harm from drinking 20.650 (1.092) 12.721 (0.961)0.170 (0.009)0.090 (0.007) Respondent states religion is important in life 1.650 (1.346)2.780 (1.180)0 .013 (0.009)0.007 (0.007) Respondent states religion influences decisions10.841 (1.260)4.908 (1.109)0.100 (0.009)0.049 (0.007) Female 0.050 (0.007)0.002 (0. 008)0.039 (0.007)0.009 (0.005) Race (African American) 0.127 (0.011) 0.112 (0.012)0.119 (0.008)0.059 (0.006) Race (Native American) 0.010 (0.040) 0.003 (0.036)0.048 (0.033)0.114 (0.002) Race (Asian) 0.224 (0.022)0.154 (0.018)0.130 (0.015)0.060 (0.010) Race (nonwhite Hispanic) 0.019 (0. 011)0.010 (0.012)0.008 (0.010)0.005 (0.014) Age of student (17 years old) 0.143 (0.001)0.144 (0.010)0.149 (0.014)0.065 (0.008) Age of student (18 years old) 0.083 (0.013)0.101 (0.010)0.102 (0.013)0.036 (0.009) Age of student (19 years old) 0.059 (0.013)0.065 (0.010)0.061 (0.015)0.027 (0.006) Last grade completed (9th grade) 0.012 (0.010)0.007 (0.008)0.018 (0.009)0.001 (0.005) Last grade completed (10th grade) 0.036 (0.012)0.16 (0.013)0.004 (0.011)0.007 (0.008) Last grade completed (11th grade) 0.045 (0.012)0.048 (0.012)0.017 (0.012)0.012 (0.012) Number in family 0.024 (0.005)0.022 (0.004)0.012 (0.007)0.091 (0.005) Number in family (>5) 0.155 (0.023)0.115 (0.016)0.073 (0.014)0.036 (0.010) Family income ($20,000$49,999) 0.055 (0.012)0.049 (0.001)0.028 (0.010)0.017 (0.005) Family income ($50,000$74,999) 0.046 (0.007)0.041 (0.012)0.039 (0.013)0.008 (0.006) Family income ($75,000 or more) 0.071 (0.014)0.084 (0.012)0.051 (0.013)0.015 (0.008) MSA segement with 1+ million persons 0.000 (0.011)0.016 (0.009)0.017 (0.008)0.012 (0.006) MSA segment of less than 1 million 0.000 (0.011)0.016 (0.009)0.013 (0.008)0.007 (0.005) Year 2002 indicator 0.021 (0.008)0. 019 (0.008)0.008 (0.007)0.001 (0.005) Standard errors are in parentheses. 83
PAGE 90
Appendix 6. All IV estimates on the probability of enrollment for binge drinking (n=13,526) 1619 sample Explanatory variables IV coefficient (Marginal Effect SE) binge drinking 0.083 (0.023) Female 0.014 (0.004) Race (African American) 0.309 (0.007) Race (Native American) 0.001 (0.027) Race (Asian) 0.033 (0.009) Race (nonwhite Hispanic) 0.001 (0.014) Age of student (17 years old) 0.681 (0.016) Age of student (18 years old) 0.548 (0.016) Age of student (19 years old) 03661 (0.018) Last grade completed (9th grade) 0.120 (0.019) Last grade completed (10th grade) 0.186 (0.019) Last grade completed (11th grade) 0.334 (0.019) Number in family 0.005 (0.002) Number in family (>5) 0.001 (0.012) Family income ($20,000$49,999) 0.036 (0.007) Family income ($50,000$74,999) 0.081 (0.008) Family income ($75,000 or more) 0.093 (0.007) MSA segement with 1+ million persons 0.020 (0.002) MSA segment of less than 1 million 0.017 (0.005) Year 2002 indicator 0.002 (0.004) 84
PAGE 91
Appendix 7. All IV estimates on the probability of enrollment for binge drinking (n=28,065) 1825 sample Explanatory variables IV coefficient (Marginal Effect SE) binge drinking 0.063 (0.019) Female 0.004 (0.006) Race (African American) 0.008 (0.008) Race (Native American) 0.017 (0.022) Race (Asian) 0.140 (0.015) Race (nonwhite Hispanic) 0.058 (0.007) Age of student (19 years old) 0.243 (0.009) Age of student (20 years old) 0.420 (0.009) Age of student (21 years old) 0.501 (0.009) Age of student (2223 years old) 0.620 (0.008) Age of student (2425 years old) 0.698 (0.008) Last grade completed (Freshman) 0.358 (0.007) Last grade completed (Sophomore/ Junior) 0.508 (0.006) Number in family 0.003 (0.002) Number in family (>5) 0.055 (0.012) Family income ($20,000$49,999) 0.108 (0.005) Family income ($50,000$74,999) 0.045 (0.008) Family income ($75,000 or more) 0.010 (0.008) MSA segement with 1+ million persons 0.047 (0.006) MSA segment of less than 1 million 0.034 (0.006) Year 2002 indicator 0.005 (0.004) Standard errors are in parentheses. 85
PAGE 92
Appenidx 8. Probit estimates for absenteeism (1825 years old) (n=8,817) Pseudo R2 = 0.08Pseudo R2 = 0.10Pseudo R2 = 0.13Pseudo R2 = 0.06 number of daysnumber of drinksBingeAbuse/ Dependence explanatory variables drank in past yearin past monthdrinking on alcohol Risk of bodily harm from drinking 34.537 (1.523)22.212 (1.398)0.283 (0.011)0.128 (0.009) Respondent states religion is important in life 4.510 (1.983)0.517 (1.821)0.020 (0.015)0.006 (0.011) Respondent states religion influences decisions13.585 (1.874)8.587 (1.726)0.126 (0.015)0.054 (0.011) Female 0.047 (0.009)0.000 (0.006)0.088 (0.011)0.046 (0.008) Race (African American) 0.120 (0.015)0.195 (0.017)0.238 (0.013)0.086 (0.010) Race (Native American) 0.005 (0.045)0.099 (0.023)0.059 (0.014)0.021 (0.042) Race (Asian) 0.191 (0.026)0.239 (0.020)0.207 (0.020)0.058 (0.016) Race (nonwhite Hispanic) 0.059 (0.016)0.079 (0.019)0.073 (0.017)0.039 (0.012) Age of student (19 years old) 0.017 (0.007)0.043 (0.016)0.033 (0.017)0.026 (0.013) Age of student (20 years old) 0.060 (0.013)0.086 (0.019)0.055 (0.013)0.016 (0.016) Age of student (21 years old) 0.129 (0.0140.230 (0.0140.192 (0.023)0.056 (0.014) Age of student (2223 years old) 0.084 (0.014)0.140 (0.020)0.093 (0.023)0.046 (0.019) Age of student (2425 years old) 0.081 (0.016)0.176 (0.022)0.063 (0.028)0.040 (0.023) Last grade completed (Freshman) 0.040 (0.012)0.069 (0.016)0.043 (0.012)0.010 (0.012) Last grade completed (Sophomore/ Junior) 0.056 (0.014)0.103 (0.017)0.032 (0.014)0.018 (0.013) Number in family 0.025 (0.002)0.039 (0.006)0.020 (0.003)0.013 (0.004) Number in family (>5) 0.239 (0.015)0.243 (0.014)0.146 (0.013)0.052 (0.018) Family income ($20,000$49,999) 0.016 (0.005)0.064 (0.001)0.093 (0.008)0.042 (0.009) Family income ($50,000$74,999) 0.005 (0.007)0.044 (0.007)0.074 (0.009)0.034 (0.006) Family income ($75,000 or more) 0.017 (0.007)0.007 (0.010)0.051 (0.011)0.023 (0.011) MSA segement with 1+ million persons 0.016 (0.012)0.014 (0.015)0.020 (0.015)0.018 (0.010) MSA segment of less than 1 million 0.016 (0.011)0.030 (0.014)0.007 (0.013)0.004 (0.010) Year 2002 indicator 0.002 (0.008)0.006 (0.011)0.013 (0.015)0.014 (0.008) Standard errors are in parentheses. 86
PAGE 93
Appendix 9. Probit estimates for absenteeism (1619 years old) (n=10,039) Pseudo R2 = 0.06Pseudo R2 = 0.07Pseudo R2 = 0.09Pseudo R2 = 0.06 number of daysnumber of drinksBingeAbuse/ Dependence explanatory variables drank in past yearin past monthdrinking on alcohol Risk of bodily harm from drinking 19.795 (1.166) 13.035 (1.033)0.171 (0.010)0.091 (0.008) Respondent states religion is important in life 9.798 (1.346)3.138 (1.277)0.020 (0.010)0.013 (0.008) Respondent states religion influences decisions3.020 (1.442)4.301 (1.192)0.090 (0.010)0.049 (0.008) Female 0.076 (0.010)0.009 (0.009)0.028 (0.008)0.001 (0.005) Race (African American) 0.120 (0.016)0.110 (0.013)0.116 (0.008)0.007 (0.006) Race (Native American) 0.003 (0.048)0.026 (0.043)0.021 (0.038)0.097 (0.038) Race (Asian) 0.215 (0.029)0.146 (0.022)0.124 (0.015)0.063 (0.010) Race (nonwhite Hispanic) 0.008 (0.011)0.007 (0.014)0.007 (0.009)0.003 (0.009) Age of student (17 years old) 0.081 (0.039)0.126 (0.010)0.111 (0.010)0.035 (0.008) Age of student (18 years old) 0.043 (0 .038)0.100 (0.032)0.084(0.013)0.017 (0.009) Age of student (19 years old) 0.015 (0.013)0.066 (0.032)0.053 (0.015)0.010 (0.006) Last grade completed (9th grade) 0.023 (0.034)0.023 (0.008)0.022 (0.009)0.009 (0.005) Last grade completed (10th grade) 0.050 (0.012)0.049 (0.013)0.044 (0.011)0.001 (0.020) Last grade completed (11th grade) 0.091 (0.035)0.091 (0.035)0.076 (0.031)0.002 (0.021) Number in family 0.025 (0.006)0.018 (0.004)0.011 (0.007)0.008 (0.005) Number in family (>5) 0.158 (0.027)0.098 (0.022)0.065 (0.014)0.025 (0.010) Family income ($20,000$49,999) 0.071 (0.015)0.059 (0.015)0.030 (0.008)0.015 (0.005) Family income ($50,000$74,999) 0.055 (0.017)0.046 (0.017)0.036 (0.010)0.010 (0.011) Family income ($75,000 or more) 0.097 (0 .016)0.099 (0.017)0.058 (0.009)0.019 (0.008) MSA segment with 1+ million persons 0.002 (0.010)0.023 (0.009)0.022 (0.008)0.012 (0.006) MSA segment of less than 1 million 0.004 (0.013)0.023 (0.009)0.019 (0.008)0.008 (0.005) Year 2002 indicator 0.020 (0.014)0.024 (0.009)0.008 (0.007)0.003 (0.006) Standard errors are in parentheses. 87
PAGE 94
Appendix 10. All IV estimates on absenteeism for binge drinking (n=10,039) 1619 sample days missed due to skipping days missed due to illness Explanatory variables IV coefficient (Marginal Effect SE)IV coefficient (Marginal Effect SE) Binge drinking 1.418 (0.186) 0.622 (0.197) Female 0.045 (0.035) 0.298 (0.047) Race (African American) 0.361 (0.062) 0.125 (0.086) Race (Native American) 0.021 (0.127) 0.483 (0.231) Race (Asian) 0.014 (0.069) 0.070 (0.125) Race (nonwhite Hispanic) 0.090 (0.161) 0.041 (0.071) Age of student (17 years old) 0.297 (0.157) 0.054 (0.148) Age of student (18 years old) 0.149 (0.153) 0.316 (0.162) Age of student (19 years old) 0.090(0.161) 0.482 (0.172) Last grade completed (9th grade) 0.190 (0.126) 0.098 (0.171) Last grade completed (10th grade) 0.313 (0.121) 0.316 (0.163) Last grade completed (11th grade) 0.404 (0.128) 0.482 (0.173) Number in family 0.009 (0.020) 0.028 (0.029) Number in family (>5) 0.031 (0.093) 0.006 (0.134) Family income ($20,000$49,999) 0.175 (0.061) 0.424 (0.082) Family income ($50,000$74,999) 0.268 (0.064) 0.582 (0.088) Family income ($75,000 or more) 0.398 (0.063) 0.701 (0.086) MSA segement with 1+ million persons 0.156 (0.045) 0.108 (0.059) MSA segment of less than 1 million 0.054 (0.040) 0.061 (0.057) Year 2002 indicator 0.019 (0.033) 0.116 (0.044) Standard errors are in parentheses. 88
PAGE 95
Appendix 11. All IV estimates for absenteeism because of binge drinking (n=8,817) 1825 sample days missed due to skipping days missed due to illness Explanatory variables IV coefficient (Marginal Effect SE)IV coefficient (Marginal Effect SE) Binge drinking 0.954 (0.179) 0.352 (0.150) Female 0.062 (0.048) 0.228 (0.045) Race (African American) 0.449 (0.083) 0.283 (0.080) Race (Native American) 0.221 (0.334) 0.069 (0.158) Race (Asian) 0.339 (0.141) 0.015 (0.171) Race (nonwhite Hispanic) 0.165 (0.090) 0.166 (0.072) Age of student (19 years old) 0.081 (0.073) 0.190 (0.065) Age of student (20 years old) 0.001 (0.101) 0.218 (0.081) Age of student (21 years old) 0.045 (0.117) 0.128 (0.087) Age of student (2223 years old) 0.088 (0.103) 0.202 (0.085) Age of student (2425 years old) 0.014 (0.119) 0.021 (0.118) Last grade completed (Freshman) 0.172 (0.079) 0.154 (0.068) Last grade completed (Sophomore/ Junior) 0.084 (0.077) 0.281 (0.081) Number in family 0.023 (0.026) 0.047 (0.022) Number in family (>5) 0.049 (0.120) 0.209 (0.113) Family income ($20,000$49,999) 0.103 (0.065) 0.027 (0.056) Family income ($50,000$74,999) 0.153 (0.077) 0.087(0.065) Family income ($75,000 or more) 0.229 (0.065) 0.109 (0.064) MSA segement with 1+ million persons 0.064 (0.062) 0.048 (0.052) MSA segment of less than 1 million 0.023 (0.054) 0.068 (0.045) Year 2002 indicator 0.020 (0.043) 0.039 (0.039) Standard errors are in parentheses. 89
PAGE 96
90 About the Author Wesley (Wes) Austin, is a Ph.D. candidate in the Department of Economics at the University of South Florida. He also holds an M.A. in economics and a bachelor degree in finance. While he is interested in seve ral fields of economic research, his primary focus is in the areas of health economi cs, education economics and labor economics. Prior to entering the Ph.D. program in economics, he held positions as a financial analyst and market researcher and he taught economic s on an adjunct basis. Among his varied hobbies are sports (especially football a nd golf) and aviation.
