Page 1

Economics of Education Review 25 (2006) 482–497

Does alcohol use during high school affect educational

attainment?: Evidence from the National Education

Longitudinal Study

Pinka Chatterji?

Center for Multicultural Mental Health Research, Cambridge Health Alliance/Harvard Medical School,

120 Beacon Street, 4th Floor, Somerville, MA 02143, USA

Received 25 March 2004; accepted 19 May 2005

Abstract

This paper uses data from the National Education Longitudinal Study to estimate the association between high

school alcohol use and educational attainment measured around age 26. Initially, the effect of alcohol use on

educational attainment is estimated using baseline probit models, which ignore the possibility that unmeasured

determinants of alcohol use and educational attainment are correlated. A bivariate probit model is used next to estimate

the equations jointly, with alcohol policies as identifying variables. Because these identifying variables are problematic,

the bivariate probit model is then re-estimated without any identifying exclusions but with the correlation coefficient

fixed at various levels. This part of the analysis allows one to gauge the sensitivity of the estimates to correlation

between the unobservable determinants of both outcomes. The results suggest that alcohol use is associated with

reductions in educational attainment, but there is little evidence that this association represents a causal relationship.

r 2005 Elsevier Ltd. All rights reserved.

JEL classification: I2

Keywords: Human capital; Demand for schooling

1. Introduction and background

Alcohol is the most widely used and the most

intensely used drug among high school students. In the

2003 Monitoring the Future Study (MTF), a national

survey of adolescent substance use, about 48 percent of

high school seniors reported alcohol use in the past

month (Johnston, O’Malley, & Bachman, 2003). This

high prevalence of drinking among high school seniors is

consistent with the fact that a large percentage of

younger adolescents believe that regular and intense use

of alcohol is not very harmful. Among 8th grade

respondents in the 2003 MTF, only about 30 percent

felt that having one or two alcoholic drinks every day

was potentially very harmful (Johnston et al., 2003).

At first glance, the widespread belief among adoles-

cents that regular and intense alcohol use is not harmful

is at odds with research on the educational consequences

of alcohol use during youth. Adolescent alcohol users

earn lower grades, are more likely to report acade-

mic difficulties, and are less likely to graduate from

high school compared to their non-using peers (Cook

& Moore, 1993; Ellickson, Tucker, & Klein, 2003;

ARTICLE IN PRESS

www.elsevier.com/locate/econedurev

0272-7757/$-see front matter r 2005 Elsevier Ltd. All rights reserved.

doi:10.1016/j.econedurev.2005.05.005

?Tel.: +16175038449; fax: +16175038430.

E-mail address: pchatterji@charesearch.org.

Page 2

Yamada, Kendix, & Yamada, 1996). Existing research,

however, offers only limited, conflicting information

regarding whether or not this association between

alcohol use and negative educational outcomes is causal.

Although alcohol dependence is clearly linked to serious

cognitive deficits (see National Institute on Alcohol

Abuse and Alcoholism (1989) for a review) that would

be expected to affect schooling, there is only mixed

evidence that ‘‘social drinking’’ impairs cognition, and

these studies focus on adults rather than teenagers

(Parsons, 1986). In most studies focused on adolescents,

the hypothesized causal mechanism linking alcohol use

to later alcohol-related problems (such as poor educa-

tional attainment) is the progression from experimenta-

tion to more intense alcohol use (Bonomo, Bowes,

Coffey, Carlin, & Patton, 2004; Ellickson, Tucker,

Klein, & McGuigan, 2001; Guilamo-Ramos, Turrisi,

Jaccard, Wood, & Gonzalez, 2004). However, this

progression does not necessarily imply a causal relation-

ship. Teenage alcohol use is associated with a range of

individual and family risk factors, including genetic

factors, behavioral and family problems, low levels of

parental monitoring, parental substance use, and weak

connection to school (Borawski, Ievers-Landis, Love-

green, & Trapl, 2003; Diego, Field, & Sanders, 2003;

Ellickson et al., 2001; Maney, Higham-Gardill, &

Mahoney, 2002; Sale, Sambrano, Springer, & Turner,

2003; Silberg, Rutter, D’Onofrio, & Eaves, 2003). These

characteristics typically are difficult to measure when

using a secondary data set, and they have the potential

to directly affect educational attainment. This source of

endogeneity may bias estimates of the effect of alcohol

use on educational attainment.

In implementing methods that address the potential

endogeneity of alcohol use, a main source of concern has

been the validity and quality of the variables used to

identify the outcome equation (e.g. the validity and

quality of identifying instruments in an instrumental

variables context). In previous work, state-level alcohol

policies have been used as identifying variables because

they are expected to be good predictors of adolescent

alcohol use, but they are not expected to directly affect

educational attainment or to be correlated with the

disturbance term. For example, Cook and Moore (1993)

use data on high school seniors from the National

Longitudinal Survey of Youth (NLSY) and an instru-

mental variables methodology to study the impact of

frequent drinking (drinking on at least 2 days in the past

week) on the number of years of education completed.

Using state-level alcohol policies as instruments for

frequent drinking, they find that frequent drinkers

complete 2.3 fewer years of college compared to seniors

who are not frequent drinkers (Cook & Moore, 1993).

However, state policy variables may not be good

predictors of adolescent alcohol use. Bollen, Guilkey,

and Mroz (1995), Bound, Jaeger, and Baker (1995),

Nelson and Startz (1990), Staiger and Stock (1997), and

others show that a low first stage F-statistic for the

identifying instrumental variables may suggest that IV

estimates are no better than biased OLS estimates.

Rashad and Kaestner (2004) also show that these policy

variables may be problematic when used as identifying

variables in the bivariate probit model, where equations

modeling alcohol use and the consequences of alcohol

use are estimated jointly. In addition to these concerns,

Dee (1999) suggests that state policies may be associated

with unobserved state sentiments that underlie both

alcohol use and educational attainment.

Because of these issues, two recent studies on alcohol

use and educational attainment apply alternative em-

pirical approaches to address the problem of endogene-

ity. Koch and Ribar (2001), using a sample from the

NLSY 1979, estimate the effect of the age of initiation of

alcohol use on the number of years of schooling

completed by age 25. Using data on siblings, they

estimate: (1) family fixed effects models; and (2)

instrumental variables models using sibling age of

alcohol use initiation as an instrument. The findings

suggest that at most, the age of initiation increases years

of education by 0.47 years for men and by 0.36 years for

women.

Dee and Evans (2003) use pooled data from the

1977–1992 MTF surveys to estimate the impact of

minimum drinking ages on drinking, and data from the

Census Bureau’s 1990 Public Use Sample to estimate

reduced form equations modeling the effect of drinking

ages on schooling. Using a two-sample IV approach

(TSIV), they draw on both sets of results to generate

estimates of the impact of drinking on educational

attainment. The results indicate that alcohol use has no

statistically significant impact on high school comple-

tion, college entrance or college persistence.

The present study builds on recent efforts to better

understand the nature of the association between high

school alcohol use and educational attainment. Data

come from the Fourth Follow-Up to the National

Education Longitudinal Study (NELS), which allows an

assessment of the effects of high school alcohol use on

educational attainment measured around age 26. Four

educational attainment indicator variables are of interest:

(1) graduation from high school on schedule; (2) receiving

any type of high school diploma; (3) entering a 4 year

college; and (4) college graduation. Initially, the analysis

focuses on estimating the effect of a binary measure of

alcohol use on each educational attainment indicator

using baseline probit models. These models ignore the

possibility that the unmeasured determinants of alcohol

use and educational attainment may be correlated with

each other. To address this issue, a bivariate probit model

with alcohol use policies as identifying variables is used

next to estimate the educational attainment and alcohol

use equations jointly.

ARTICLE IN PRESS

P. Chatterji / Economics of Education Review 25 (2006) 482–497

483

Page 3

Because the use of alcohol policies as identifying

variables is problematic in this case (as demonstrated

below), the bivariate probit model is then re-estimated

successively without any identifying exclusions but with

the correlation coefficient fixed at various levels. This

part of the analysis allows one to gauge how sensitive

estimates of the effect of alcohol use on educational

attainment are to increasingly stronger (imposed) levels

of correlation between the unobservable determinants of

both outcomes. Finally, the amount of selection on

observed variables in the model is calculated, and the

bivariate probit model is re-estimated under the

assumption that the selection on unobserved variables

is equal to the selection on observed variables (the

details of this approach, which is proposed in Altonji,

Elder, and Taber (2001), are summarized later in the

paper). This estimate is considered a conservative

estimate because it is generated under the assumption

that individuals sort themselves into alcohol use and

educational attainment along observed and unobserved

factors that are equally important.

The primary contribution of this analysis is metho-

dological in that it allows one to gauge the sensitivity of

the estimated effect of alcohol use on educational

attainment to various degrees of selection on unob-

served variables, without relying on any identifying

exclusions which recently have been shown to be

problematic (see Dee & Evans, 2003; Rashad &

Kaestner, 2004). The results suggest that the observed

association between alcohol use and reduced educa-

tional attainment does not persist when moderate levels

of selection on unobserved factors are imposed on the

models. Thus, although there is clearly an association

between alcohol use in high school and reduced

educational attainment, this study finds only weak

evidence that this relationship is causal.

2. Methods

Theoretically, a study of the effects of alcohol use on

educational attainment draws on paradigms developed

in the areas of human capital theory (Becker, 1964),

household production (Becker, 1965), and decision-

making over the life cycle (Ghez & Becker, 1975).

Adolescents making decisions about alcohol use weigh

the benefits of use, primarily utility from intoxication,

against the costs of use that may be incurred in the

present or in the future. An individual teenager

consumes alcohol such that the marginal utility gained

from an incremental increase in alcohol consumption

equals the shadow price of alcohol consumption. Based

on the psychological literature discussed above, which

suggests that alcohol use is associated with academic

problems, alcohol consumption would be expected to

decrease the marginal benefit of an additional year of

school. Thus, ignoring the potential for addiction, the

shadow price of alcohol consumption includes not only

the price of alcohol (including the money price, time

costs, and potential penalties for use) but also the

adverse effects of current alcohol use on future

educational attainment (discounted future earnings).

A rational adolescent would choose the optimal amount

of alcohol use and schooling, taking into account these

costs and benefits.

Empirically, this analysis focuses on estimating

Eq. (1). Educational attainment (E) is determined by

alcohol use (A), by observed characteristics such as

gender and race (X), and by unobserved, personal

variables such as ability or motivation (m).

E ¼ a1A þ a2X þ a3m þ ?.

Eq. (2) represents the demand for alcohol. The vector

X represents observed characteristics that affect alcohol

use, which are the same as the observed determinants of

educational attainment (X). Similarly, the vector m

represents the unobserved determinants of educational

attainment that also may influence alcohol use. Inter-

cepts are suppressed for convenience.

(1)

A ¼ b1X þ b2m þ Z.

The parameter of interest is a1, the structural effect

of alcohol use on educational attainment. Estimating

Eq. (1) by standard methods can lead to biased and

inconsistent coefficients if alcohol use is determined by

the same unmeasured characteristics that determine

one’s educational attainment (a3a0 and b2a0) (Greene,

2003).

(2)

2.1. The baseline probit model

Eq. (1) is initially estimated using probit regression

models that attempt to include all of the important

determinants of educational attainment, but that do not

address directly the potential endogeneity problem.

First, the probit model is estimated with a limited set

of covariates. Next, a more fully specified probit model

is estimated including controls for adolescent academic

performance and smoking behavior measured prior to

high school, in 8th grade. Comparing results from the

two models allows one to evaluate how much of the

association between high school alcohol use and later

educational attainment appears to be driven by selection

into these two outcomes along observed factors (see

Painter & Levine, 2000). Finally, to account for the

possibility that students at the same school may share

unobserved determinants of outcomes, the models are

estimated with school-specific random effects in addition

to the 8th grade characteristics.

Despite the inclusion of a detailed set of covariates

including 8th grade characteristics and school-level

random effects, the observed correlation between

ARTICLE IN PRESS

P. Chatterji / Economics of Education Review 25 (2006) 482–497

484

Page 4

alcohol use and educational attainment still may be

influenced by selection bias. Accounting for selection

bias empirically is likely to reduce the magnitude of the

effect because most important unobserved factors,

such as ability level, are likely to be negatively correlated

with alcoholuseand positively

educational attainment. Therefore, the estimated coeffi-

cients on alcohol use from the models that ignore

endogeneity (but that include 8th grade measures and

school-level effects) may be considered baseline esti-

mates of the impact of alcohol use on educational

attainment.

correlatedwith

2.2. The bivariate probit model

One approach to empirically addressing the problem

correlated unobserved

this correlation explicitly using full information max-

imum likelihood strategies, such as the bivariate

probit model (relevant to this case, since both the

educational attainment and the alcohol use measures are

binary). The bivariate probit model assumes that the

disturbance terms in Eqs. (1) and (2) are jointly normally

distributed, and the equations are estimated simulta-

neously using maximum likelihood (Greene, 2003). In

this study, the same vector of covariates is included in

both the alcohol use and the educational attainment

equations. Although identification could come from

functional form restrictions, such restrictions generally

are difficult to defend in practice. Thus, practical

implementation of the bivariate probit model requires

valid exclusion restrictions—variables that affect alcohol

use but do not directly affect educational attainment.

Following previous work in this area, the analysis uses

state-level alcohol policies as identifying variables, and

the model is estimated taking account of the possibility

that observations from the same state may not be

independent.

If the identifying variables are poor predictors of

alcohol use, the bivariate probit model does not work

well, yielding very imprecise estimates (see Rashad &

Kaestner, 2004). Moreover, the exclusion of the state-

level policies from the educational attainment equation

must be a valid restriction. In the present study, the

identifying variables are problematic, making it difficult

to determine whether or not the bivariate probit model

estimates are trustworthy. Consequently, another em-

pirical method is considered that does not rely on

identifying exclusions but that does allow an evaluation

of how sensitive the alcohol use estimates are to

correlation between unobserved factors.

of variablesis tomodel

2.3. The constrained bivariate probit model

Altonji et al. (2001) propose an approach to the

problem of questionable identifying variables that is

based on constrained bivariate probit models.1The

method is based on estimation of a bivariate probit

model without any identifying restrictions but with a

constrained correlation coefficient, r. In this study, r is

set at ?0.10 initially and then the absolute value of r is

increased in increments of 0.10 to ?0.20, ?0.30, and

?0.40. In this way, increasingly stronger, negative

correlation between the unobservables is imposed on

the model, which allows one to examine whether or not

the effect of alcohol use on educational attainment is

robust to such changes. This analysis can uncover the

threshold of selection on unobservables, if any, at which

alcohol use no longer has a statistically significant effect

on educational attainment. The correlation between the

unobserved determinants of alcohol use and educational

attainment is constrained to negative values based on

prior research (Jessor & Jessor, 1977).

Altonji et al. (2001) argue that if the observable

determinants of an outcome are truly just a random sub-

set of the complete set of determinants, selection on

observable characteristics must be equal to selection on

unobservable characteristics.

showthatthis condition

var(X0g) ¼ cov(H*,?)/var(?), where H* is an unob-

served, continuous measure of the net benefits from

high school graduation, X0g is the vector of observed

variables that affect H* weighted by their corresponding

coefficients, and ? is the unobserved determinants of

variables that affect H* weighted by their relevant

coefficients. In words, this condition means that the data

collected in a survey are no more relevant to the

outcome being studied than the data that were not

collected.

This paper uses a specialized survey that was designed

to study the determinants of educational outcomes. It

seems unlikely, therefore, that selection on observable

factors is equal to selection on unobservable factors in

this case; on the contrary, one would expect that

selection on observable factors would be more impor-

tant than selection on unobservable factors. Therefore,

the conservative estimate is this estimate obtained under

the assumption that selection on unobservable variables

is equal to selection on observables. This estimate can be

compared to the baseline estimate, which is the estimate

from the single equation probit model that includes 8th

grade covariates and school-level effects, but assumes no

selection on unobservable variables. The approach

demonstrates how sensitive the baseline estimates are

to a relatively stringent assumption about the degree of

selection on unobserved variables.

Altonji

implies:

et al.

cov(H*,X0g)/

(2001)

ARTICLE IN PRESS

1See Grossman, Kaestner, and Markowitz (2004), which

studies the relationship between adolescent alcohol use and

sexual behavior using bivariate probit techniques similar to

those used in the present study.

P. Chatterji / Economics of Education Review 25 (2006) 482–497

485

Page 5

3. Data and measures

Data come from the Fourth Follow-Up to the NELS.

This education survey was initiated in the spring

semester of 1988, when about 25,000 8th grade students

completed surveys along with their parents, teachers and

school administrators. The respondents are a clustered,

stratified probability sample of over 1000 public and

private schools. Extensive information was collected

about the students’ school experiences, family back-

ground, activities, attitudes, cognitive test scores, and

alcohol use. After the baseline survey was conducted in

1988, the students completed surveys four more times, in

1990, 1992, 1994, and finally, in 2000, when most

respondents were about 26 years old and had been out

of high school for eight years (US Department of

Education, 2002). The survey continued to track and

interview students who dropped out of high school. The

weighted response rate for 8th grade respondents at the

time of the 2000 survey was 83.8 percent (US Depart-

ment of Education, 2002).

This study is based on data from the 2000 follow-up

of the NELS, which included 12,144 respondents from

the three previous surveys. Two samples are used in the

analysis: (1) a 10th grade sample, comprised of 7604

students who were interviewed in 8th grade and who

were still in school by 1990; and (2) a 12th grade sample,

comprised of 5421 students from the 10th grade sample

who were still in school by 1992. The 10th grade sample

is used to examine the effect of 10th grade alcohol use on

educational attainment among students who were still in

school (but not necessarily at grade level) 2 years

after the baseline interview. The 10th grade sample

includes students who dropped out of school after

being interviewed in 10th grade, but it excludes students

who dropped out prior to being interviewed in 10th

grade.

Of the 12,144 respondents interviewed in 2000, the

10th grade analysis sample was limited to eligible

respondents who had been interviewed as students in

8th and 10th grade, who had available information from

the school and parent interviews in the base year, and

who had test score and school information available for

the first follow-up survey. These restrictions were needed

to ensure that respondents in the 10th grade sample had

relatively complete data and sufficient longitudinal

information to assess the effects of high school alcohol

use on later educational outcomes. Thus, of the 12,144

respondents, the 10th grade sample ðN ¼ 7604Þ ex-

cluded: (1) 585 respondents who were ineligible in the

base year or were freshened in 10th grade; (2) 647

respondents who were not in school in 10th grade; (3)

335 respondents who did not complete baseline or first

follow-up surveys; (4) 923 respondents who did not have

base year parent, test or school information available;

(5) 559 respondents who did not have first follow-up test

or school information available; and (6) 157 respondents

who had missing information on state of residence. The

sample also excluded: (7) 1178 respondents who had

missing information on substance use; and (8) 125

observations with missing information on educational

attainment in 2000. Models were estimated by gender; as

a result, 31 observations with missing gender also were

dropped from the analysis sample. Missing values for

other data elements were imputed using sample means.

To assess the effect of these sample restrictions, the main

10th grade models were re-estimated using a sample that

included respondents without school or parent informa-

tion available, and respondents with state of residence

missing. This new sample included 7676 respondents

(instead of 7604). These models yielded almost identical

results to those shown in the paper.

The 12th grade sample was created by restricting the

sample based on conditions (1)–(8) above, and then

limiting the sample to: (1) students in school (but not

necessarily in 12th grade) at second follow-up (387

observations excluded); (2) students who have second

follow-up student, test and school survey information

available (1275 observations excluded); (3) students with

12th grade substance use information available (529

observations deleted). As in the 10th grade analysis,

respondents also were dropped if gender was missing

ðN ¼ 23Þ. This sample is used to explore the effect of

alcohol use in 12th grade among students who were still

in school (but not necessarily at expected grade level) 4

years after the baseline interview. By design, the sample

excludes students in the 10th grade sample who dropped

out between 10th and 12th grade.

The educationaloutcome

dummy variables indicating: (1) whether or not the

student graduated from high school on schedule;

(2) whether or not the student received his/her high

school diploma by 2000, including GEDs; (3) whether or

not the student had entered a 4-year college by 2000; and

(4) whether or not the student had graduated from a

4-year college by 2000. The effects of alcohol use

on educational attainment are assessed separately for

males and females because the epidemiology of adoles-

cent substance use is known to vary significantly by

gender (Hopfer, Crowley, & Hewitt, 2003). Estimation

based on gender-specific samples is empirically sup-

ported by F-tests of all covariates interacted with gender

from fully interacted models (results available upon

request).

The analysis uses the following two alcohol use

indicators, measured in both 10th and 12th grade: (1)

respondent drank at least once in the past month; and

(2) respondent had five or more drinks in a row at least

once in the past 2 weeks. Because of multicollinearity

between the alcohol use measures, the models include

each alcohol use measure separately. Because over 99%

of the 12th grade samples completed high school by

measures are four

ARTICLE IN PRESS

P. Chatterji / Economics of Education Review 25 (2006) 482–497

486

Page 6

2000, the analysis does not include estimation of the

effect of 12th grade alcohol use on high school

completion by 2000.

In addition to the alcohol use measures, all models

include: (1) race/ethnicity indicator variables; (2) a set of

religion indicators; (3) region of residence; (4) mother’s

and father’s education levels (high school graduate as

the baseline, high school dropout, college graduate);

(5) family structure in 8th grade (two parents as the

baseline, step-family, single parent); (6) family income

(lowest quartile, low-middle quartile and highest quar-

tile dummy indicators, with high-middle quartile as the

baseline); (7) number of siblings; (8) whether or not the

school is located in a suburban (baseline), rural or urban

area; (9) whether or not the student is currently a daily

smoker; and (10) whether or not the student’s current

(10th or 12th grade) math achievement test score is in

the lowest quartile. Additionally, some models include

8th grade measures to control for preexisting differences

between youth that may affect both high school alcohol

use and educational attainment. These measures are:

(1) whether or not the respondent was a daily smoker in

8th grade; (2) whether or not the respondent’s math test

scores in 8th grade was in the lowest quartile; and

(3) whether or not the student had repeated a grade by

8th grade. All models are estimated with and without

school-level random effects, based on the school the

respondent attended in 10th grade. The following two

state-level alcohol use policies and prices are used as

identifying variables in the standard bivariate probit

models: (1) the real state-level excise tax on beer; and

(2) the percentage of each state’s population living in

counties dry for beer.

4. Results

4.1. Baseline probit estimates

Almost all (97–98 percent) of survey respondents

eventually graduated from high school, with 90–91

percent graduating on schedule in 1992 (Table 1). About

61 percent of 10th grade students entered a 4-year

college or university by 2000, and 38–42 percent

eventually graduated from college or university by

2000. These college entrance and graduation rates are

slightly higher for students who were in 12th grade in

1992 (results not shown).

Among 10th graders, 42 percent of boys and 38

percent of girls reported alcohol use in the past month.

Binge drinking also was common among 10th graders,

with about 25 percent of boys and 20 percent of girls

reporting at least one binge-drinking episode in the last

2 weeks.

Table 2 shows results from all models that focus on

the association between 10th grade drinking and

educational attainment measured 10 years later in

2000. Results are presented by gender for each of the

four educational outcomes. The top panel shows results

from models with an indicator any past month alcohol

use included, while the bottom panel shows results from

models with an indicator of any binge drinking in the

past 2 weeks included. Three models are presented for

each educational outcome: (1) a standard probit model

with basic covariates included; (2) the standard model

with additional 8th grade covariates included; and (3)

the standard model with both 8th grade covariates and

school-level random effects included.

As a group, all of these models show evidence of a

robust, negative association between 10th grade drink-

ing and educational attainment in 2000 (Table 2).

Among boys, any past month alcohol use is associated

with a 2 percentage point reduction in the probability of

graduating on schedule, a 7 percentage point reduction

in the probability of entering college, and a 5 percentage

point reduction in graduating from college (Table 2).

These results are similar for binge drinking in 10th grade

among boys. Girls who use alcohol in 10th grade have a

4 percentage point reduction in entering college com-

pared to girls who do not use alcohol in 10th grade.

Binge drinking among girls detracts from college

graduation, but there is no statistically significant

association between any alcohol use in the past month

and college graduation among girls.

Three overall findings in Table 2 are notable. First,

although alcohol use among boys detracts from

graduating on schedule, alcohol use in 10th grade has

more important impact on college entrance and college

completion than on high school graduation for both

boys and girls. Second, the magnitude of the association

between alcohol use and educational attainment varies

considerably by gender (Table 2). Third, in most models,

adding 8th grade measures and school random effects

does not appreciably change the magnitude of the

coefficients on alcohol use. This result suggests that

there is not important selection into educational

attainment and alcohol use along observed preexisting

characteristics.

Table 3 presents all baseline models that include the

12th grade drinking measures. Among girls, most

associations between 12th grade drinking and educa-

tional attainment are not statistically significant. How-

ever, among boys, any alcohol use in 12th grade is

associated with a 7 percentage point reduction in college

entrance and binge drinking is associated with a 9

percentage point reduction in college entrance (Table 3,

Males). These are 11 and 15 percent reductions in the

probability of entering college, measured at the sample

means for boys.

Changes in sample composition could explain some of

the differential effects of 10th versus 12th grade alcohol

use since the 12th grade sample is a sub-set of the 10th

ARTICLE IN PRESS

P. Chatterji / Economics of Education Review 25 (2006) 482–497

487

Page 7

ARTICLE IN PRESS

Table 1

Sample means and standard deviations

Variable

Definition

Mean (standard deviation)

Females

ðN ¼ 4126Þ

Males

ðN ¼ 3478Þ

10th grade sample

Graduated from high school

Dummy variable ¼ 1 if respondent graduated from high school by 2000, 0

otherwise

0.978

0.973

Graduated from high school on schedule

Dummy variable ¼ 1 if respondent graduated from high school on schedule

(in 1992), 0 otherwise

0.908

0.892

Entered college

Dummy variable ¼ 1 if respondent entered a 4 year college or university by

2000, 0 otherwise

0.608

0.606

Graduated from college

Dummy variable ¼ 1 if respondent graduated from a 4 year college or

university by 2000, 0 otherwise

0.406

0.376

Used alcohol in past month

Dummy variable ¼ 1 if respondent reports alcohol use in the past month, 0

otherwise

0.376

0.418

Binged in past 2 weeks

Dummy variable ¼ 1 if respondent reports binge drinking episode in the past

2 weeks, 0 otherwise

0.200

0.253

Smokes daily

Dummy variable ¼ 1 is respondent reports using cigarettes every day, 0

otherwise

0.170

0.142

Demographic characteristicsAfrican-American

Dummy variable ¼ 1 if respondent is African-American, 0 otherwise

0.085

0.067

Hispanic

Dummy variable ¼ 1 if respondent is Hispanic, 0 otherwise

0.107

0.099

Asian

Dummy variable ¼ 1 if respondent is Asian, 0 otherwise

0.068

0.068

Native American

Dummy variable ¼ 1 if respondent is Native American, 0 otherwise

0.036

0.035

Central

Dummy variable ¼ 1 if respondent lives in Central region, 0 otherwise

0.310

0.309

West

Dummy variable ¼ 1 if respondent lives in West region, 0 otherwise

0.179

0.181

South

Dummy variable ¼ 1 if respondent lives in South region, 0 otherwise

0.337

0.326

Urban

Dummy variable ¼ 1 if respondent lives in urban area, 0 otherwise

0.245

0.256

Rural

Dummy variable ¼ 1 if respondent lives in rural area, 0 otherwise

0.344

0.332

Family and personal characteristics

Mother dropout

Dummy variable ¼ 1 if respondent’s mother is a high school dropout, 0

otherwise

0.242

0.202

Mother college graduate

Dummy variable ¼ 1 if respondent’s mother is a college graduate, 0

otherwise

0.226

0.269

Father dropout

Dummy variable ¼ 1 if respondent’s father is a high school dropout, 0

otherwise

0.272

0.228

Father college graduate

Dummy variable ¼ 1 if respondent’s father is a college graduate, 0 otherwise

0.275

0.319

Single parent family

Dummy variable ¼ 1 if respondent lives with one biological parent or

relative only, 0 otherwise

0.169

0.158

P. Chatterji / Economics of Education Review 25 (2006) 482–497

488

Page 8

ARTICLE IN PRESS

Step-family

Dummy variable ¼ 1 if respondent lives with one biological parent and

another non-biological parent figure, 0 otherwise

0.115

0.098

Family income

Family income measured in 8th grade

41.423 (36.065)

43.227 (35.593)

Number of siblings

Number of siblings

2.20 (1.51)

2.13 (1.48)

Catholic

Dummy variable ¼ 1 if respondent is Catholic, 0 otherwise

0.280

0.295

Baptist or Methodist

Dummy variable ¼ 1 if respondent is Baptist or Methodist, 0 otherwise

0.282

0.249

Other Christian

Dummy variable ¼ 1 if respondent is of other Christian denomination, 0

otherwise

0.287

0.277

Other Religion

Dummy variable ¼ 1 if respondent is of another religion, 0 otherwise

0.081

0.076

Low math score

Dummy variable ¼ 1 if standardized achievement test score in mathematics

is in lowest quartile, 0 otherwise

0.262

0.236

8th grade personal characteristics

Repeated a grade

Dummy variable ¼ 1 if respondent reports that s/he repeated a grade before

8th grade, 0 otherwise

0.085

0.135

Low math score in 8th grade

Dummy variable ¼ 1 if standardized achievement test score in mathematics

in 8th grade is in lowest quartile, 0 otherwise

0.257

0.242

Smoked daily in 8th grade

Dummy variable ¼ 1 if respondent smoked daily in 8th grade, 0 otherwise

0.041

0.045

State alcohol policies Percentage dry counties

Percentage of counties in state where alcohol sales are banned in 1990

4.67 (8.79)

4.78 (9.05)

Beer tax

Real state excise tax on beer in 1990

0.171 (0.175)

0.165 (0.170)

P. Chatterji / Economics of Education Review 25 (2006) 482–497

489

Page 9

grade sample. To test this possibility, the 10th grade

models were estimated using the 12th grade sample.

When the ‘‘graduated by 1992’’ standard probit models

were estimated this way, the estimated effects of alcohol

use became considerably smaller and lost statistical

significance at conventional levels, particularly for

females. When the college entrance and college gradua-

tion models were estimated this way, however, the

results were not appreciably different. These findings

suggest that changes in sample composition, as well as

the overall reduction in sample size, may explain some

portion of the differential effects of alcohol use in 10th

grade versus 12th grade in the case of the high school

graduation by 1992 outcome.

4.2. Standard bivariate probit model estimates

Table 4 shows results from a standard bivariate probit

model, where the educational attainment and alcohol

use equations are estimated jointly, while accounting for

the fact that observations in the same state may not be

independent. The model assumes that the error terms

have a joint probability distribution that is bivariate

normal. The model is identified by two state-level

ARTICLE IN PRESS

Table 2

Effect of 10th grade drinking on educational attainmenta

Graduated from high

school on schedule

Graduated from high

school by 2000

Entered 4-year

college by 2000

Graduated from 4-

year college by 2000

Males FemalesMalesFemales MalesFemales Males Females

Effect of past month alcohol use in 10th grade

Basic modelb

Probit coefficient

Huber/White standard error

Marginal effect

?0.148

0.068

?0.021

?0.131

0.068

?0.015

0.009

0.116

0.000

?0.108

0.112

?0.001

?0.165

0.053

?0.062

?0.087

0.051

?0.033

?0.127

0.054

?0.045

?0.052

0.051

?0.020

w/8th grade measuresc

Probit coefficient

Huber/White standard error

Marginal effect

?0.157

0.068

?0.021

?0.137

0.068

?0.016

?0.008

0.116

?0.000

?0.137

0.115

?0.002

?0.177

0.053

?0.066

?0.094

0.051

?0.035

?0.127

0.054

?0.045

?0.056

0.051

?0.021

w/school random effectsd

Probit coefficient

Huber/White standard error

Marginal effect

?0.157

0.068

?0.021

?0.135

0.074

?0.012

?0.008

0.119

?0.000

?0.137

0.117

?0.002

?0.195

0.057

?0.072

?0.103

0.053

?0.038

?0.127

0.054

?0.045

?0.056

0.051

?0.021

Effect of binge drinking in past 2 weeks in 10th grade

Basic modelb

Probit coefficient

Huber/White standard error

Marginal effect

?0.293

0.072

?0.044

?0.141

0.078

?0.017

?0.155

0.119

?0.002

?0.117

0.123

?0.002

?0.181

0.059

?0.068

?0.109

0.061

?0.041

?0.232

0.063

?0.081

?0.150

0.063

?0.056

w/8th grade measuresc

Probit coefficient

Huber/White standard error

Marginal effect

?0.281

0.073

?0.041

?0.130

0.080

?0.015

?0.127

0.119

?0.001

?0.117

0.126

?0.001

?0.169

0.060

?0.064

?0.100

0.062

?0.038

?0.213

0.063

?0.074

?0.150

0.064

?0.054

w/school random effectsd

Probit coefficient

Huber/White standard error

Marginal effect

?0.281

0.071

?0.041

?0.122

0.081

?0.011

?0.127

0.122

?0.001

?0.117

0.124

?0.001

?0.187

0.065

?0.070

?0.111

0.064

?0.042

?0.213

0.064

?0.074

?0.150

0.063

?0.054

aFemales N ¼ 4126; Males N ¼ 3478.

bThe basic model includes race/ethnicity, region, religion, parents’ education levels, family structure, family income quartiles,

number of siblings, urban school location, rural school location, currently smokes daily, and current math score in lowest quartile.

cThe 8th grade measures are math score in lowest quartile, daily smoker and repeated a grade.

dThese models were estimated using the random effects probit model. The random effect specified was the school the respondent

attended in 10th grade. The sample attended about 978 different schools in 10th grade. The standard errors are not Huber/White

corrected.

P. Chatterji / Economics of Education Review 25 (2006) 482–497

490

Page 10

alcohol policies (state beer tax and percent of state that

is dry for alcohol) that are included in the alcohol use

equation but not in the educational attainment equa-

tion. The success of the bivariate probit method relies on

the identifying variables being good predictors of

alcohol use, and the exclusion of these variables from

the educational attainment being valid. These two issues

are explored below (this part of the analysis draws on

Rashad and Kaestner (2004)).

The instruments as a group are statistically significant

predictors of alcohol use in all models estimated using

the two female samples (results not shown). However,

this is not true for the two male samples-test of the joint

significance of the instruments could not reject the null

hypothesis of no effect in 14 of the 16 models estimated

using the two male samples (results not shown). It is

interesting that Sen (2002) finds the same inconsistency

in her analysis of substance use and sexual behaviors

based on the National Longitudinal Survey of Youth—a

set of identifying variables, which include state-level

alcohol policies, are statistically significant predictors of

alcohol use for females but not for males. Rashad and

Kaestner (2004) caution that this discrepancy in Sen

(2002) may indicate that the correlation for females is

spurious. This possibility is equally relevant in the

present case, casting doubt on the identification strategy.

To examine whether or not the alcohol policies can be

legitimately excluded from the educational attainment

ARTICLE IN PRESS

Table 3

Effect of 12th grade drinking on educational attainmenta

Graduated from high

school on schedule

Graduated from high

school by 2000

Entered 4-year

college by 2000

Graduated from 4-

year college by 2000

MalesFemales MalesFemales MalesFemales MalesFemales

Effect of past month alcohol use in 12th grade

Basic modelb

Probit coefficient

Huber/White standard error

Marginal effect

?0.234

0.106

?0.017

0.031

0.107

0.001

?0.193

0.064

?0.065

?0.054

0.059

?0.019

?0.119

0.062

?0.045

?0.103

0.057

?0.040

w/8th grade measuresc

Probit coefficient

Huber/White standard error

Marginal effect

?0.243

0.107

?0.017

0.019

0.108

0.001

?0.196

0.064

?0.066

?0.049

0.060

?0.017

?0.118

0.063

?0.045

?0.096

0.057

?0.038

w/school random effectsd

Probit coefficient

Huber/White standard error

Marginal effect

?0.243

0.104

?0.017

0.027

0.139

0.000

?0.204

0.068

?0.068

?0.049

0.060

?0.017

?0.118

0.063

?0.045

?0.102

0.060

?0.040

Effect of binge drinking in past 2 weeks in 12th grade

Basic modelb

Probit coefficient

Huber/White standard error

Marginal effect

?0.206

0.099

?0.016

?0.141

0.113

?0.007

?0.259

0.065

?0.090

?0.076

0.070

?0.026

?0.107

0.065

?0.041

?0.019

0.068

?0.007

w/8th grade measuresc

Probit coefficient

Huber/White standard error

Marginal effect

?0.205

0.100

?0.015

?0.151

0.114

?0.007

?0.252

0.065

?0.087

?0.057

0.070

?0.020

?0.095

0.066

?0.036

0.006

0.069

0.002

w/school random effectsd

Probit coefficient

Huber/White standard error

Marginal effect

?0.205

0.101

?0.015

?0.151

0.146

?0.002

?0.267

0.070

?0.091

?0.057

0.071

?0.020

?0.095

0.066

?0.036

0.008

0.074

0.003

aFemale N ¼ 2949; Male N ¼ 2472.

bThe basic model includes race/ethnicity, region, religion, parents’ education levels, family structure, family income quartiles,

number of siblings, urban school location, rural school location, currently smokes daily, and current math score in lowest quartile.

cThe 8th grade measures are math score in lowest quartile, daily smoker and repeated a grade.

dThese models were estimated using the random effects probit model. The random effect specified was the school the respondent

attended in 10th grade. The sample attended about 978 different schools in 10th grade. The standard errors are not Huber/White

corrected.

P. Chatterji / Economics of Education Review 25 (2006) 482–497

491

Page 11

equation, a just-identified version of the bivariate

probit model is estimated, with the percentage dry

counties variable included in the educational attainment

equation.In almostall

females, percentage dry is not a statistically significant

predictor of educational attainment (results not shown).

Therefore, this exclusion restriction appears to be

appropriate.

Given that the instruments have questionable pre-

dictive power, it is not surprising that the bivarate

probit models yield puzzling results about the effects

of alcohol use on educational attainment. The estimat-

ed coefficients on the alcohol use measures and

the estimated correlation coefficients are inconsistent

modelsfor malesand

in sign and statistical significance. These unstable results

are likely to be related to the weak identification

strategy.

4.3. Constrained bivariate probit model estimates

Because of these problems in implementing the

standard bivariate probit model with a poor set of

identifying variables, the analysis moves to a constrained

bivariate probit model approach. Tables 5 and 6 present

results from bivariate probit models with imposed

correlation coefficients ranging from ?0.10 to ?0.40,

and then with an imposed equal selection assumption.

The equal selection assumption constrains the model to

ARTICLE IN PRESS

Table 4

Effect of 10th grade and 12th grade drinking on educational attainment

Graduated from

high school on

schedule

Graduated from

high school by

2000

Entered 4-year

college by 2000

Graduated from

4-year college by

2000

MalesFemales MalesFemalesMalesFemalesMales Females

Effect of alcohol use in 10th grade

Alcohol use

Standard bivariate probit coefficient

Standard error

Marginal effect

Rho

p-value on Wald test of rho

?0.083

0.642

?0.011

?0.044

0.907

?0.531

0.429

?0.068

0.233

0.377

0.146

0.410

0.001

?0.092

0.682

?0.188

0.317

?0.002

0.030

0.868

1.02

0.231

0.364

?0.727

0.003

0.344

0.382

0.126

?0.264

0.264

0.999

0.242

0.365

?0.680

0.004

0.758

0.586

0.287

?0.491

0.252

Binge drinking

Standard bivariate probit coefficient

Standard error

Marginal effect

Rho

p-value on Wald test of rho

0.472

0.282

0.060

?0.445

0.026

0.400

0.243

0.039

?0.313

0.050

0.277

0.361

0.002

?0.241

0.308

0.198

0.283

0.002

?0.188

0.320

1.29

0.071

0.407

?0.871

0.000

0.895

0.224

0.288

?0.571

0.001

?0.210

—

?0.073

?0.005

0.887

1.05

0.169

0.401

?0.659

0.000

Effect of alcohol use in 12th grade

Alcohol use

Standard bivariate probit coefficient

Standard error

Marginal effect

Rho

p-value on Wald test of rho

?0.578

0.197

?0.041

0.206

0.124

?0.114

0.793

?0.005

0.080

0.863

?1.51

0.100

?0.490

0.849

0.00

0.718

0.664

0.245

?0.467

0.338

?0.118

—

?0.045

?0.001

0.980

?0.096

—

?0.038

?0.000

0.996

Binge drinking

Standard bivariate probit coefficient

Standard error

Marginal effect

Rho

p-value on Wald test of rho

0.877

0.894

0.083

?0.639

0.369

?0.151

—

?0.007

0.000

0.994

0.662

0.382

0.217

?0.546

0.058

1.25

0.128

0.349

?0.779

0.000

1.07

0.342

0.404

?0.701

0.028

1.51

0.120

0.531

?0.874

0.000

Notes: (a) Female 10th Grade N ¼ 4126; Male 10th Grade N ¼ 3478; Female 12th Grade N ¼ 2949; Male 12th Grade N ¼ 2472.

(b) Models include race/ethnicity, region, religion, parents’ education levels, family structure, family income quartiles, number of

siblings, urban school location, rural school location, currently smokes daily, current math score in lowest quartile, whether 8th math

score is in lowest quartile, whether respondent was a daily smoker in 8th grade and repeated a grade by 8th grade.

(c) The identifying variables are percent of state that is dry for alcohol and state excise tax on beer.

(d) The marginal effects are calculated as the univariate marginal predicted probability that the education outcome equals 1.

P. Chatterji / Economics of Education Review 25 (2006) 482–497

492

Page 12

ARTICLE IN PRESS

Table 5

Effect of 10th grade drinking on educational attainment

Graduated from

high school on

schedule

Graduated from

high school by

2000

Entered 4-year

college by 2000

Graduated from

4-year college by

2000

Constrained RhoMales FemalesMales FemalesMales FemalesMales Females

Effect of past month alcohol use in 10th grade

?0.1

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.009

0.068

0.001

0.030

0.068

0.003

0.158

0.119

0.002

0.031

0.116

0.000

?0.012

0.053

?0.005

0.072

0.051

0.027

0.037

0.054

0.013

0.110

0.051

0.041

?0.2

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.174

0.067

0.024

0.197

0.067

0.021

0.325

0.118

0.003

0.200

0.115

0.002

0.153

0.053

0.057

0.239

0.050

0.088

0.202

0.053

0.073

0.277

0.050

0.105

?0.3

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.340

0.066

0.047

0.363

0.066

0.040

0.493

0.116

0.006

0.369

0.113

0.004

0.318

0.052

0.118

0.405

0.049

0.147

0.368

0.053

0.133

0.444

0.049

0.168

?0.4

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.506

0.064

0.072

0.530

0.065

0.060

0.662

0.114

0.009

0.540

0.111

0.007

0.483

0.050

0.177

0.571

0.048

0.205

0.534

0.051

0.194

0.610

0.048

0.231

Equal selection

Constrained bivariate probit coefficient

Standard error

Marginal effect

Rho

p-value

0.500

0.108

0.023

?0.281

0.680

0.323

0.066

0.036

?0.276

0.726

?0.044

0.296

?0.000

?0.004

0.949

0.061

0.116

0.001

?0.118

0.885

0.081

0.060

0.028

?0.063

0.847

0.113

0.051

0.042

?0.124

0.711

0.066

0.058

0.026

?0.084

0.801

0.172

0.050

0.065

?0.137

0.5161

Effect of binge drinking in 10th grade

?0.1

Constrained bivariate probit coefficient

Standard error

Marginal effect

?0.110

0.071

?0.015

0.043

0.074

0.005

?0.174

0.119

?0.002

?0.156

0.122

?0.002

0.004

0.060

0.002

0.076

0.060

0.028

?0.039

0.064

?0.014

0.034

0.063

0.013

?0.2

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.060

0.070

0.008

0.214

0.074

0.022

?0.008

0.118

?0.000

0.015

0.120

0.000

0.177

0.059

0.065

0.253

0.060

0.091

0.138

0.063

0.050

0.214

0.062

0.082

?0.3

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.229

0.069

0.029

0.381

0.072

0.037

0.159

0.116

0.002

0.184

0.118

0.003

0.349

0.058

0.126

0.428

0.059

0.150

0.316

0.062

0.117

0.396

0.061

0.153

?0.4

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.397

0.067

0.050

0.547

0.070

0.052

0.325

0.113

0.004

0.351

0.115

0.005

0.521

0.057

0.183

0.457

0.053

0.155

0.495

0.061

0.185

0.578

0.060

0.224

Equal selection

Constrained bivariate probit coefficient

Standard error

Marginal effect

Rho

p-value

0.489

0.064

0.071

?0.384

0.221

0.677

0.069

0.064

?0.479

0.038

?0.043

0.118

?0.001

?0.179

0.700

?0.057

0.121

?0.001

?0.158

0.835

0.119

0.053

0.044

?0.166

0.525

0.342

0.059

0.121

?0.251

0.050

0.152

0.054

0.055

?0.152

0.757

0.327

0.062

0.126

?0.262

0.038

Notes: (a) Females N ¼ 4126; Males N ¼ 3478. (b) The marginal effects are calculated as the univariate marginal predicted probability

that the education outcome equals 1.

P. Chatterji / Economics of Education Review 25 (2006) 482–497

493

Page 13

ARTICLE IN PRESS

Table 6

Effect of 12th grade drinking on educational attainment

Graduated from

high school on

schedule

Graduated from

high school by

2000

Entered 4-year

college by 2000

Graduated from

4-year college by

2000

Constrained Rho MalesFemales MalesFemales MalesFemales MalesFemales

Effect of past month alcohol use in 12th grade

?0.1

Constrained bivariate probit coefficient

Standard error

Marginal effect

?0.105

0.101

?0.007

0.079

0.105

0.004

?0.176

0.112

?0.003

0.023

0.057

0.008

?0.029

0.064

?0.011

?0.036

0.055

?0.014

?0.2

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.063

0.100

0.005

0.243

0.104

0.011

?0.009

0.062

?0.003

0.187

0.057

0.065

0.138

0.063

0.053

0.129

0.055

0.050

?0.3

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.235

0.099

0.019

0.409

0.102

0.020

0.159

0.061

0.054

0.351

0.056

0.121

0.305

0.062

0.117

0.294

0.054

0.115

?0.4

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.412

0.096

0.036

0.578

0.100

0.032

0.327

0.060

0.110

0.516

0.054

0.178

0.473

0.060

0.182

0.461

0.053

0.180

Equal selection

Constrained bivariate probit coefficient

Standard error

Marginal effect

Rho

p-value

0.813

0.089

0.073

?0.603

0.287

0.437

0.109

0.020

?0.251

0.826

?0.040

0.065

?0.014

0.095

0.530

0.015

0.060

0.005

?0.039

0.750

0.049

0.063

0.019

?0.103

0.566

?0.033

0.046

?0.013

?0.077

0.502

Effect of binge drinking in 12th grade

?0.1

Constrained bivariate probit coefficient

Standard error

Marginal effect

?0.069

0.098

?0.005

?0.073

0.112

?0.003

?0.111

0.062

?0.038

?0.006

0.067

?0.002

?0.017

0.050

?0.006

0.017

0.067

0.007

?0.2

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.096

0.097

0.007

0.095

0.111

0.004

0.054

0.062

0.019

0.167

0.067

0.056

0.146

0.050

0.054

0.191

0.066

0.075

?0.3

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.262

0.095

0.019

0.262

0.109

0.011

0.220

0.060

0.076

0.339

0.065

0.112

0.310

0.049

0.115

0.367

0.065

0.145

?0.4

Constrained bivariate probit coefficient

Standard error

Marginal effect

0.429

0.093

0.033

0.429

0.106

0.018

0.327

0.060

0.110

0.511

0.064

0.163

0.475

0.048

0.177

0.542

0.064

0.214

Equal selection

Constrained bivariate probit coefficient

Standard error

Marginal effect

Rho

p-value

0.264

0.101

0.021

?0.300

0.174

0.561

0.110

0.021

?0.422

0.220

0.156

0.064

0.053

?0.244

0.412

0.303

0.070

0.099

?0.208

0.117

?0.107

0.051

?0.039

?0.045

0.759

0.067

0.058

0.026

?0.194

0.091

Notes: (a) Female N ¼ 2949; Male N ¼ 2472. (b) The marginal effects are calculated as the univariate marginal predicted probability

that the education outcome equals 1.

P. Chatterji / Economics of Education Review 25 (2006) 482–497

494

Page 14

have equal selection on observed and unobserved factors,

which is a conservative assumption. Table 5 presents

results for 10th grade alcohol use, and Table 6 shows

results for the 12th grade alcohol use models.

Tables 5 and 6 show that for all educational

outcomes, the effect of alcohol use on schooling is very

sensitive to the imposed degree of correlation between

the unobservable determinants of these outcomes. In

every case, whenthe

constrained to relatively low levels (?0.10 or ?0.20),

the negative association between alcohol use and

schooling outcomes becomes statistically insignificant

or changes sign. These findings suggest that the observed

association between alcohol use and educational attain-

ment is not robust to moderate degrees of negative

correlation between the unobserved determinants of

both outcomes.

The equal selection model (last rows of Tables 5 and

6) is estimated under the constraint that the degree of

selection on the observed determinants of educational

outcomes must be equal to the degree of selection on the

unobserved factors. One would expect that a specialized

data set like NELS would contain information on the

most important determinants of educational outcomes,

making it unlikely that selection on data that were not

collected would be equally important. Therefore, im-

posing this constraint is though to yield conservative

estimates. The results presented in Tables 5 and 6 show

that the conservative estimates of alcohol use on

educational attainment are either statistically insignif-

icant, or, counter to intuition, positive and statistically

significant.

correlationcoefficient is

5. Discussion and conclusions

The objective of this study was to assess the strength

of the evidence supporting a causal relationship between

teenage alcohol use and reduced educational attaiment

by using a method recently introduced by Altonji et al.

(2001) for this purpose. The method involves three steps:

(1) first, estimate baseline models that do not directly

address the possibility of correlated, unobserved vari-

ables; (2) second, obtain estimates of the effect of

alcohol use on educational attainment from a bivariate

probit regression model in which the correlation

between unobservable variables is fixed at various levels;

and(3)third, calculate

into alcohol use and educational attainment on observed

variables, and obtain estimates of the effect of alcohol

use on educational attainment under the assumption

that the degree of sorting on unobserved variables is

equal to the degree of sorting on observed variables.

With this last step, one can estimate the degree of sorting

on unobservable factors using the observed data,

and identify a conservative bound on the causal

the amount ofsorting

parameter estimate without any identifying assump-

tions.2The primary strength of the Altonji et al. (2001)

method is that it allows an assessment of the possible

existence and strength of a causal relationship between

alcohol and educational attainment, without requiring

the use of identifying assumptions that have been shown

to be problematic in this context.

In this analysis, baseline findings show large and

statistically significant effects of alcohol use on educa-

tional attainment in most cases. The baseline models do

not show considerable selection into alcohol use and

educational outcomes along measured variables. Even

so, the association between alcohol use and educational

attainment still may be confounded by common,

unmeasured determinants of these outcomes. The

standard bivariate probit models account for this

correlation between the unobserved determinants of

alcohol use and educational attainment, but, in this

analysis, the results from these models are unreliable

because of poor identifying variables. The constrained

bivariate probit models, which do not rely on any

identifying variables, show that even modest amounts of

correlation between the unmeasured determinants of

educational attainment and alcohol use eliminate the

negative, statistically significant effect of alcohol use on

education. Given that it is likely that some degree of

selection on unobservable factors exists, these results

suggest that alcohol use in high school has no

appreciable, causal effect on educational attainment,

despite the strong association between these variables.

The methods used in this paper are based on

assumptions about the role of unobserved factors, and

it is true that one can never know with certainty the true

degree of selection on variables that cannot be

measured. This analysis demonstrates, however, how

important it is to assess the robustness of an association

to various degrees of selection on unobserved factors. If

an association is not robust to even modest amounts

of selection on unobserved variables, as is the case in

this study, one has less confidence that the observed

association reflects a true, causal relationship.

Although the primary contribution of this paper is

methodological, the results also have relevance at the

policy level. These findings imply that programs and

policies intended to reduce alcohol use may not be

effective in directly improving educational attainment.

For example, these results suggest that a state policy

change that reduces access to alcohol, such as an

increase in the legal drinking age, would not be expected

to increase educational attainment. However, even if

alcohol use does not cause reductions in schooling, the

ARTICLE IN PRESS

2It is important to note that because the degree of sorting on

observed variables was relatively moderate in this analysis, this

final model may or may not yield the most conservative

estimate.

P. Chatterji / Economics of Education Review 25 (2006) 482–497

495

Page 15

strong association between alcohol use and reduced

educational attainment suggests that alcohol users as a

group, particularly males, still is an important group to

target for intervention. Unfortunately, in the absence of

information on unobserved factors, this study does

not shed light on what causal factors should be targeted.

In order to create effective prevention programs, we

need to better understand the causal mechanism that

links alcohol use to lower educational attainment.

Future research should address the process through

which alcohol use in high school leads to reduced

educational attainment, and how this process might vary

by gender.

Acknowledgements

I gratefully acknowledge research support from Grant

K01 AA000328-03 from the National Institute of

Alcohol Abuse and Alcoholism. I also would like to

thank two anonymous reviewers for very helpful

comments.

References

Altonji, J. G., Elder, T. E., & Taber, C.R. (2001). Selection on

observed and unobserved variables. Assessing the effective-

ness of Catholic Schools. Unpublished working paper.

Becker, G. S. (1964). Human capital. New York: National

Bureau of Economic Research.

Becker, G. S. (1965). A theory of the allocation of time.

Economic Journal, 75, 493–517.

Bollen, K. A., Guilkey, D. K., & Mroz, T. A. (1995). Binary

outcomes and endogenous explanatory variables: Tests and

solutions with an application to the demand for contra-

ceptive use in Tunisia. Demography, 32(1), 111–131.

Bonomo, Y. A., Bowes, G., Coffey, C., Carlin, J. B., & Patton,

G. C. (2004). Teenage drinking and the onset of alcohol

dependence: A cohort study over seven years. Addiction,

99(12), 1520–1528.

Borawski, E. A., Ievers-Landis, C. E., Lovegreen, L. D., &

Trapl, E. S. (2003). Parental monitoring, negotiated

unsupervised time, and parental trust: The role of perceived

parenting practices in adolescent health risk behaviors.

Journal of Adolescent Health, 33(2), 60–70.

Bound, J., Jaeger, D. A., & Baker, R. M. (1995). Problems with

instrumental variables estimation when the correlation

between the instruments and the endogenous explanatory

variables is weak. Journal of the American Statistical

Association, 90(430), 443–450.

Cook, P., & Moore, M. (1993). Drinking and schooling. Journal

of Health Economics, 12(4), 411–429.

Dee, T. (1999). State alcohol policies, teen drinking and

traffic fatalities. Journal of Public Economics, 72(2),

289–315.

Dee, T. S., & Evans, W. N. (2003). Teen drinking and

educational attainment: Evidence from two-sample instru-

mental variables (TSIV) estimates. The Journal of Labor

Economics, 21(1), 178–209.

Diego, M. A., Field, T. M., & Sanders, C. E. (2003). Academic

performance, popularity, and depression predict adolescent

substance use. Adolescence, 38(149), 35–42.

Ellickson, P. L., Tucker, J. S., & Klein, D. J. (2003). Ten-year

prospective study of public health problems associated with

early drinking. Pediatrics, 111(5 Pt1), 949–955.

Ellickson, S. L., Tucker, J. S., Klein, D. J., & McGuigan, K. A.

(2001). Prospective risk factors for alcohol misuse in late

adolescence. Journal of Studies on Alcohol, 62(6), 773–782.

Ghez, G., & Becker, G. S. (1975). The allocation of time and

goods over the life cycle. New York: National Bureau of

Economic Research.

Greene, W. H. (2003). Econometric Analysis (5th ed). NJ:

Prentice Hall.

Grossman, M., Kaestner, R., & Markowitz, S. (2004). Get high

and get stupid: The effect of alcohol and marijuana use on

teen sexual behavior. Review of Economics of the Household,

2, 413–441.

Guilamo-Ramos, V., Turrisi, R., Jaccard, J., Wood, E., &

Gonzalez, B. (2004). Progressing from light experimentation

to heavy episodic drinking in early and middle adolescence.

Journal of Studies on Alcohol, 65(4), 494–500.

Hopfer, C. J., Crowley, T. J., & Hewitt, J. K. (2003). Review of

twin and adoption studies of adolescent substance use.

Journal of the American Academy of Child & Adolescent

Psychiatry, 42(6), 710–719.

Jessor, R., & Jessor, S. L. (1977). Problem behavior and

psychosocial development: A longitudinal study of youth.

New York: Academic Press.

Johnston, L. D., O’Malley, P. M., & Bachman, J. G. (2003).

Ecstasy use falls for second year in a row, overall teen drug

use drops. University of Michigan News and Information

Services: Ann Arbor, MI. [On-line]. Available: www.moni-

toringthefuture.org; accessed 3/16/04.

Koch, S. F., & Ribar, D. C. (2001). A siblings analysis of the

effects of alcohol consumption onset on educational

attainment.

Contemporary

162–174.

Maney, D. W., Higham-Gardill, D. A., & Mahoney, B. S.

(2002). The alcohol-related psychosocial and behavioral

risks of a nationally representative sample of adolescents.

Journal of School Health, 72(4), 157–163.

Monitoring the Future Study. The University of Michigan,

www.monitoringthefuture.org, accessed 6/03.

National Institute on Alcohol Abuse and Alcoholism (May

1989). Alcohol and cognition. Alcohol Alert, No. 4 PH258,

Accessed at http://www.niaaa.nih.gov/publications/

aa04.htm, 2/15/05.

Nelson, C. R., & Startz, R. (1990). The distribution of the

instrumental variables estimator and its t-ratio when the

instrument is a poor one. Journal of Business, 63(1),

S125–S139.

Painter, G., & Levine, D. I. (2000). Family structure and

youths’ outcomes: Which correlations are causal? The

Journal of Human Resources, 35(3), 524–549.

Parsons, O. A. (1986). Cognitive functioning in sober social

drinkers: A review and critique. Journal of Studies on

Alcohol, 47(2), 101–114.

Economic Policy,

19(2),

ARTICLE IN PRESS

P. Chatterji / Economics of Education Review 25 (2006) 482–497

496

Page 16

Rashad, I., & Kaestner, R. (2004). Teenage sex, drugs and

alcohol use: Problems identifying the cause of risky

behaviors. Journal of Health Economics, 23(3), 493–503.

Sale, E., Sambrano, S., Springer, J. F., & Turner, C. W. (2003).

Risk, protection, and substance use in adolescents: A multi-

site model. Journal of Drug Education, 33(1), 91–105.

Sen, B. (2002). Does alcohol-use increase the risk of sexual

intercourseamong adolescents?

NLSY97. Journal of Health Economics, 21(6), 1085–1093.

Silberg, J., Rutter, M., D’Onofrio, B., & Eaves, L. (2003).

Genetic and environmental risk factors in adolescent

Evidencefrom the

substance use. Journal of Child Psychology and Psychiatry,

44(5), 664–676.

Staiger, D., & Stock, J. H. (1997). Instrumental variables

regression with weak instruments. Econometrica, 65(3),

557–586.

US Department of Education (2002). National Center for

Education Statistics (NCES), NELS base-year to fourth

follow-up data file user’s manual, June 2002.

Yamada, T., Kendix, M., & Yamada, T. (1996). The impact of

alcohol and marijuana use on high school graduation.

Health Economics, 5(1), 77–92.

ARTICLE IN PRESS

P. Chatterji / Economics of Education Review 25 (2006) 482–497

497