Content uploaded by Daniel Eisenberg
Author content
All content in this area was uploaded by Daniel Eisenberg
Content may be subject to copyright.
1
Mental Health and Academic Success in College
(revised and resubmitted to the B.E. Journal of Economic Analysis & Policy)
May 2009 (first version January 2009)
Daniel Eisenberg (University of Michigan, daneis@umich.edu)
Ezra Golberstein (Harvard Medical School, golberstein@hcp.med.harvard.edu)
Justin Hunt (University of Michigan, jbhunt@med.umich.edu)
Abstract
Mental health problems represent a potentially important but relatively
unexplored factor in explaining human capital accumulation during college. We
conduct the first study, to our knowledge, of how mental health predicts academic
success during college in a random longitudinal sample of students. We find that
depression is a significant predictor of lower GPA and higher probability of
dropping out, controlling for prior academic performance and other variables.
The association between depression and academic outcomes is strongest among
students with a positive anxiety disorder screen. In within-person estimates using
our longitudinal sample, we find again that co-occurring depression and anxiety
are associated with lower GPA, and we find that symptoms of eating disorders are
also associated with lower GPA. This descriptive study highlights the policy
relevance of generating more definitive causal estimates of the effect of mental
health on college success, which will likely require a randomized trial.
2
Acknowledgments
This study was funded by the Blue Cross Blue Shield Foundation of Michigan
and the following units at the University of Michigan: the Comprehensive
Depression Center (Innovation Fund), the School of Public Health, the
Department of Health Management and Policy (McNerney Award), the Rackham
Graduate School, and the Office for the Vice President of Research. During the
writing of this paper Ezra Golberstein was funded by NIMH (T32 postdoctoral
traineeship) and Justin Hunt was funded by the Robert Wood Johnson Foundation
Clinical Scholars Program. We are grateful for helpful comments from Martha
Bailey, Jason Fletcher, Richard Frank, Tom McGuire, Ellen Meara, Kevin Stange,
Jacob Vigdor, two anonymous reviewers and participants in the University of
Michigan informal labor economics seminar. We are also grateful to Scott
Crawford and the Survey Sciences Group for implementing the web surveys, to
Sarah Gollust and Jennifer Hefner for assistance developing the Healthy Minds
Study, and to Andy Cameron for assistance acquiring the administrative data.
3
Americans are inundated with messages about success—in school, in a
profession, in parenting, in relationships—without appreciating that successful
performance rests on a foundation of mental health.
United States Surgeon General’s Report on Mental Health, 1999
1. INTRODUCTION
Among children and adolescents in the United States, mental disorders are
estimated to account for a larger burden of disease, as measured in disability-
adjusted life years (DALYs), than any other class of health conditions (Michaud
et al., 2006). One of the primary concerns in younger populations is that mental
health problems may affect human capital accumulation—in particular, the
amount and productivity of schooling—which may in turn have lifelong
consequences for employment, income, and other outcomes. Understanding the
link between mental health and academic success is therefore a crucial step
towards assessing the returns to preventing, detecting, and treating mental health
issues among young people.
In this paper we analyze the connection between mental health and
detailed measures of academic success during college. In the modern economy,
college education has become an increasingly important component of human
capital, and is associated with substantially higher earnings (Jaeger & Page, 1996;
Kane & Rouse, 1995) and better health outcomes (Cutler & Lleras-Muney, 2006;
Ross & Mirowsky, 1999). Approximately two thirds of high school graduates
attend college (U.S. Department of Education, 2006), but fewer than 50 percent of
college enrollees graduate (Knapp, 2007), and this proportion is 12-18 percent
lower among students who are black, Hispanic, American-Indian, or lower
socioeconomic status (Horn & Berger, 2004). Previous studies have considered a
range of factors—such as financial aid (Dynarski, 1999) and academic and social
involvement (Tinto, 1998)—that affect remaining in and completing college.
Another important factor may be mental health.
Mental disorders frequently have first onset shortly before or during the
typical college age range (18-24) (Kessler et al., 2005), yet relatively little is
known about the link between mental health and academic success in college.
Understanding this connection could be valuable due to the many ways in which
college settings can reach young people; college represents the only time in many
people’s lives when a single setting encompasses their main activities, social
networks, and a range of supportive services and organizations.
We examine how symptoms of mental disorders predict academic
outcomes during college using unique data collected at a large, academically
4
competitive, public university. We surveyed a random sample of approximately
2,800 undergraduate and graduate students about a range of mental health issues
in fall 2005, and we conducted a follow-up survey with a subset of the sample in
fall 2007. In this paper we focus on three of the most common types of mental
disorders among adolescents and young adults: depression, anxiety disorders, and
eating disorders. We link the survey data on mental health to academic measures
collected from the university’s administrative records.
This is the first study, to our knowledge, that examines how mental health
predicts GPA and dropping out in a random sample of college students. We find
that depression is a significant predictor of lower GPA and higher probability of
dropping out, even after controlling for symptoms of anxiety and eating disorders,
prior academic performance, and other covariates. Depression also appears to
interact with anxiety; the association between depression and academic outcomes
is particularly strong among students who also have a positive screen for an
anxiety disorder. Among the symptoms of depression, the strongest negative
predictor of academic performance is anhedonia (lack of pleasure and interest in
usual activities). By contrast, negative affect per se (feeling depressed or
hopeless) is not independently associated with a lower GPA. Finally, in fixed
effects (within-person) regressions of GPA on mental health variables using the
longitudinal sample, we find again that co-occurring depression and anxiety are
associated with lower GPA, and we also find that symptoms of eating disorders
are associated with lower GPA.
This study is best characterized as a detailed descriptive analysis of the
association between mental health and academic outcomes in college, rather than
a causal analysis. In the final section of the paper we illustrate that, if the
estimates were assumed to be reasonable approximations of causal relationships,
then they would imply sizeable economic returns, relative to the likely costs, from
programs to increase the detection and treatment of depression among college
students. This exercise underscores the policy relevance of acquiring more
definitive knowledge about the causal effect of mental health on college success,
which will likely require a randomized trial of mental health treatment that
collects detailed academic outcomes.
2. CONCEPTUAL FRAMEWORK AND RELATED LITERATURE
CONCEPTUAL FRAMEWORK
Mental health may affect college students’ academic outcomes along two
margins: 1) the decision to remain in school; 2) productivity, or performance,
5
given that one is in school.1 Regarding the first margin, in a simple economic
model of schooling attainment the individual chooses the amount of schooling, s,
to maximize the present discounted value of future income, V(s), where y denotes
earnings, r denotes the discount rate, A denotes abilities, and R denotes the time of
retirement or death (Becker, 1993).
∫−
=R
s
rt dteAsysV );()(
Poor mental health could reduce the marginal return to continuing schooling
(sV ∂∂ /) for any of the following reasons: a) decreasing one’s performance while
in school, which may reduce the accrual of both real skills and outward signals
(e.g., graduating with a high GPA) that increase expected job opportunities and
productivity; b) decreasing one’s expected future mental health, which in turns
decreases one’s expected productivity in future employment (e.g., by decreasing
one’s expected reliability); and, c) shortening the time horizon over which one
expects to be in the labor force (reducing R). Although in theory these factors
could imply that poor mental health causes an increase in schooling (due to the
income effect—the higher marginal value of income at lower levels, in this case),
we hypothesize that these factors on balance would cause a decrease in schooling
(due to the predominance of the substitution effect—the lower marginal return to
schooling) and therefore an increase in the likelihood of dropping out. This
hypothesis is also based on the additional possibility that poor mental health may
decrease one’s interest in the future (one’s discount rate), which would reduce
one’s willingness to make long-term investments like schooling.
As a simplified model of the second margin of interest—academic
performance while in school—Todd and Wolpin (2003) propose that achievement
T (e.g., test scores or grades) at age a is a function of family inputs, )(aFi, and
schooling inputs up until age a, )(aSi, and a fixed natural ability, 0i
A:
]),(),([ 0iiiaia AaSaFTT =
Cunha and Heckman (2006) supplement this type of model by emphasizing that
ability consists of cognitive and noncognitive skills that evolve over time. They
mention several examples of noncognitive skills that may affect the acquisition of
cognitive skills: persistence, motivation, consistency, patience, self-control, self-
discipline, self-esteem, and interpersonal behavior. Each of the mental health
1 We acknowledge that the discussion here does not reflect a number of factors that are likely to
affect academic outcomes in college, such as financial aid (Dynarski, 1999) and learning about
one’s ability (Stinebrickner & Stinebrickner, 2008). We omit these factors from our brief
conceptual discussion because it is difficult to predict how mental health would affect, interact
with, or be affected by them.
6
problems that we consider—depression, anxiety, and eating disorders—could
plausibly affect these noncognitive factors, in addition to having direct effects on
cognitive ability.
Specifically, a number of depressive symptoms may affect the
productivity of time in academic activities and/or the amount of time dedicated to
academic activities.2 These symptoms include reduced interest or pleasure in
usual activities (anhedonia), sleep disturbances (less or more than normal),
reduced energy, difficulty concentrating or making decisions, restlessness or
slowing of movement, and suicidal thoughts (which may impair concentration or
decrease interest in investing in the future) (Sadock & Sadock, 2000). In
addition, negative affect (feeling sad or hopeless) may decrease interest in the
future.
A common anxiety disorder, generalized anxiety, is marked by excessive
worrying and difficulty controlling this worrying. At lower levels anxiety can
actually be productive, but at higher levels it often impairs concentration and the
ability to remain on task (Sadock & Sadock, 2000). Generalized anxiety shares
many symptoms of depression (e.g., reduced energy, sleep disturbance, and
reduced concentration) and therefore could affect academic outcomes for many of
the same reasons that depression would. Another anxiety disorder that we
measure, panic disorder, consists of recurrent and unexpected panic attacks,
which include at least four of the following symptoms: palpitations, sweating,
shaking, shortness of breath, feeling of choking, chest pain, nausea, feeling dizzy,
derealization/depersonalization, fear of “going crazy,” fear of dying, numbness or
tingling sensations, and chills or hot flashes (Sadock & Sadock, 2000). The
attacks do not typically last long enough to impair productivity by themselves, but
they can lead to significant worrying and attempts to avoid attacks (e.g. avoiding
class or studying if those activities are associated with the anxiety).
The two main types of eating disorders are anorexia nervosa and bulimia
nervosa. People suffering from anorexia nervosa are often debilitated by physical
symptoms such as fatigue, cardiac problems, and electrolyte disturbances (Sadock
& Sadock, 2000). These symptoms, as well as any associated hospitalizations,
could negatively affect academic productivity and time available. In addition,
obsessions with weight and food could limit the time or concentration students
have for studies. For people suffering from bulimia nervosa, frequent binging and
purging can also consume time and energy. Eating disorders may also impair the
2 It is also important to acknowledge that depression and other disorders may affect the marginal
product and utility of time spent in non-school activities (i.e., leisure and employment). Therefore,
the premise of our conceptual framework, more precisely, is that mental health problems have a
substantially larger effect on the marginal product and utility of time in school activities than they
do on other common uses of time. This seems plausible due to all of the possible channels
described, but is a fundamental assumption that merits further study.
7
productivity of studying to the extent that they cause cognitive deficits, such as
poor attention and working memory (Tchanturia, 2004).
Finally, depression, anxiety, and eating disorders can be especially
impairing when they co-occur. For example, co-occurring depressive and anxiety
disorders is associated with high severity of illness (Joffe, 1993), functional
impairment (Joffe, 1993; Kessler, 1999), recurrence (Van Valkenburg, 1984), and
poorer treatment outcomes (Brown & Madonia, 1996).
EMPIRICAL LITERATURE
While many studies assess the effect of physical health on human capital
outcomes (see Currie (2008) for a review), we focus here on the studies that
investigate how mental health affects human capital. Several studies describe the
association between mental health early in life and subsequent educational
attainment. Among studies that assess early-life mental health retrospectively,
one study finds that early-onset (before adulthood) depression, is associated with
less schooling (Berndt et al., 2000), whereas another study finds that a number of
early-onset psychiatric disorders (although not major depression) are associated
with early termination of schooling (Breslau et al., 2008). Studies that use
longitudinal data also find mixed evidence on the relationship between early-life
emotional and mental health and subsequent educational attainment. For
example, two studies find that early-life externalizing behavioral problems (e.g.,
conduct disorders or ADHD), but not early-life internalizing behavioral problems
(e.g., depression or anxiety), are associated with lower subsequent education
(McLeod & Owens, 2004; Miech et al., 1999). Two other studies focus
specifically on the long-term consequences of adolescent depression: a study of
New Zealand adolescents finds no association between early-adolescent
depression and subsequent educational attainment after controlling for socio-
demographic characteristics (Fergusson & Woodward, 2002), while a study using
U.S. data finds that adolescent depression is positively correlated with high school
drop-out and negatively correlated with college enrollment (Fletcher, 2008). An
additional two studies investigate a variety of psychiatric disorders, and find
negative associations between mental illness during adolescence and graduating
from high school (Marcotte et al., 2004; Vander Stoep et al., 2003).
In addition to these descriptive studies, a small number of recent studies
have attempted to address explicitly the endogeneity of mental health with respect
to academic outcomes (e.g., mental health may be correlated with unobserved
variables related to the ability vector A noted earlier). First, two complementary
studies use sibling fixed-effects models to control for family-level unobservable
factors that might be correlated with both ADHD and academic outcomes (Currie
& Stabile, 2006; Fletcher & Wolfe, 2007). Both studies suggest that ADHD has a
8
strong effect on academic outcomes in secondary school, including standardized
test scores, grade repetition, and special education use. One of these two studies
also examines longer-term effects of ADHD, and finds no evidence of an effect
on total years of education and college attendance (Fletcher & Wolfe, 2007).
Second, researchers have used specific genotypes as instruments for mental
health, positing that variation in genotypes affects mental health but does not
directly affect educational attainment (Ding et al., 2007; Fletcher & Lehrer,
2008). The two studies are suggestive of effects of mental health on academic
outcomes, but come to somewhat different conclusions. Ding and colleagues find
that depression leads to significantly lower GPA among high school students, and
some evidence that attention deficits (without hyperactivity) reduce GPA,
although the latter results are sensitive to the model specification. Fletcher and
Lehrer’s findings suggest an effect of ADHD and depression on verbal test scores,
although the IV estimates are not significant at conventional levels. This latter
study also estimates instrumental variable models with sibling fixed effects and
finds marginally significant effects of attention deficits (but not hyperactivity) on
verbal test scores.
The aforementioned studies are important in advancing the understanding
of how mental health may affect academic outcomes among children and
adolescents, but they generally do not address the relationship between mental
health and human capital accumulation in higher education. The roles of
depression, anxiety, and eating disorders in college are particularly important to
examine, as the incidence of these conditions during late adolescence and young
adulthood greatly exceeds that of most other mental disorders including ADHD
(Kessler et al., 2005). In addition, although severe mental illness such as bipolar
disorder is somewhat less prevalent among college students as compared to same-
aged non-college students, depression and anxiety disorders are equally prevalent
across the two groups (Blanco et al., 2008).
Only two studies, to our knowledge, specifically examine the relationship
between mental health and academic outcomes during college. One recent study
compares the GPA of 121 students during six months following a diagnosis of
depression at the university’s student health center to the GPA of a control group
selected from the overall student population (Hysenbegasi et al., 2005). This
study finds a significant, negative association between GPA and untreated
depression (whereas treated depression is not associated with a significant
difference in GPA). An important limitation is that the study only includes
students who presented to the student health center, and it is unclear how this
group might differ from the overall population of students with significant
depressive symptoms. Another study uses data on 351 students at a British
university and finds that depression (but not anxiety) measured midway through
9
the second year is negatively related to exam scores at the end of the second year
(Andrews & Wilding, 2004).
Our study contributes to the literature on mental health and academic
outcomes due to a number of features. First, as noted above, we focus on the
important but relatively understudied setting of postsecondary education. Second,
our analyses utilize clinically-validated measures of self-reported mental health
status, detailed measures of academic outcomes, and a rich set of control variables
including multiple measures of prior academic performance. Third, this is the
first study to our knowledge to estimate the relationship between mental health
and GPA during college using mental health data at two time points, which enable
us to control for time-invariant individual characteristics that may otherwise bias
the estimated relationship between mental health and human capital accumulation.
Fourth, we consider independent associations between academic outcomes and
three of the most common types of mental disorders among young adults:
depression, anxiety, and eating disorders.
3. DATA AND EMPIRICAL FRAMEWORK
SAMPLE
Our data are from a randomly selected sample of undergraduate and graduate
students enrolled in fall 2005 at a large, public, academically competitive
university. Mental health measures and a range of other variables were collected
via web-based surveys as part of the Healthy Minds Study, a survey study
examining mental health and help-seeking behavior among college students
(Eisenberg et al., 2007a, 2007b). All participants gave informed consent and the
study was approved by the university’s Health Sciences IRB. In fall 2005, a
random sample of 5,021 students was recruited for the survey using an
introductory note via regular mail followed by email invitations, and 2,798
students (56%) completed the survey. Students completed the survey during a
three week period around the middle of the semester, during late October and
early November (the fall semester runs from early September to mid-December).
We fielded the survey during the middle of the semester because we wanted to
minimize the influence of transitions (e.g., moving into a new residence) and
high-stress periods (e.g., final exams) on self-reports of recent symptoms of
mental health problems.
Of the 2,798 students who completed the baseline survey, 1,272 were still
enrolled in fall 2007 and were invited for a two-year follow-up survey with a
nearly identical set of questions. Among those who were no longer enrolled as of
fall 2007, approximately 90% had graduated and 10% had left the university
without graduating. This indicates the predominant reason for attrition from our
10
study was normal academic progress; the other reason, dropping out, is one of the
outcome variables in our analysis, as explained below. Of the 1,272 students
remaining at the university as of fall 2007, 747 (59%) completed the follow-up
survey. As described in more detail later, we construct survey nonresponse
weights in order to adjust for differences between responders and nonresponders.
ACADEMIC MEASURES
Our dependent variables come from students’ academic records while at the
university. First, we examine GPAs during specific terms (semesters), as a
measure of human capital accumulation conditional on being in school. We
compute GPAs as weighted averages of course grades in those terms, where the
weights are equal to the number of credit-hours for each course. GPAs are
measured on a 0-4.3 scale, where A+ equals 4.3, A equals 4.0, A- equals 3.7, and
so on. Students with missing GPA values are excluded from these analyses.3 As
shown in Table 1, the average GPA in our sample is 3.38.
Our second main outcome variable is whether a student dropped out of the
university before graduating. For each term following the baseline (fall 2005), we
define a variable equal to 1 if the student dropped out by that term and 0
otherwise. We define a student as having dropped out as of term X if she or he
meets each of the following conditions: a) not enrolled in term X; b) not enrolled
in any subsequent term that we observe (through winter 2008); and, c) not
graduated from a degree program since the baseline semester (fall 2005).4 As
shown in Table 1, among our baseline sample 2% of students dropped out by
winter 2006, 4% by winter 2007, and 8% by winter 2008. These proportions are
significantly lower than national averages, reflecting the academically
competitive profile of students who attend this institution.5
3 Some students do not have GPAs for some or all terms (e.g., 21% of the sample during the fall
2005 term), because they chose to take courses as pass/fail or because their courses were only
offered as pass/fail (we cannot distinguish between these two possibilities). The vast majority
(96%) of students with missing GPA values are graduate students, of which 80% are in their 2nd
year or higher. Therefore having a missing GPA appears to be primarily a function of one’s
academic program as opposed to one’s decision to take classes pass/fail.
4 All students in our sample were enrolled in degree programs (i.e., there were none taking a few
classes for non-degree purposes), so it is reasonable to think of people who leave as dropping out.
5 For example, nationally 34% of students beginning college leave their first school within three
years without having graduated (Berkner & Choy, 2008), as compared to 12% in our data. We
cannot determine whether someone who is no longer enrolled has left college entirely or has
transferred. We would ideally distinguish between transfers and people who leave college, but
both are significant outcomes in that they often involve difficult transitions, in terms of academic
progress, social networks, and other factors (Laanan, 2007; Skahill, 2002).
11
Table 1: Sample characteristics (weighted means)
Baseline
sample, fall
2005
(N=2,798)
Longitudinal
sample, in fall
2005(N=747)
Longitudinal
sample, in fall
2007 (N=747)
Dependent variables Mean SD Mean SD Mean SD
GPA in…
Fall 2005 3.38 0.56 3.43 0.54
Fall 2007 3.38 0.58 3.47 0.54
Credit hours in…
Fall 2005 13.2 3.6 13.4 3.6
Fall 2007 12.4 4.3 12.4 4.1
Dropped out* by:
Winter 2006 0.02
Winter 2007 0.04
Winter 2008 0.08
Mental health variables
PHQ depression score (0-27) 5.16 5.13 4.22 5.61 4.68
PHQ score: 0-4 0.56 0.55 0.49
PHQ score: 5-9 0.30 0.30 0.34
PHQ score: 10-14 0.11 0.11 0.13
PHQ score: 15-19 0.02 0.03 0.02
PHQ score: 20-27 0.01 0.01 0.02
Change PHQ of 5+ ('05 vs '07) 0.08
Panic disorder (positive screen) 0.02 0.03
Change in status ('05 vs '07) 0.03
Generalized anxiety (positive screen) 0.03 0.04
Change in status ('05 vs '07) 0.06
Eating disorder (positive screen) 0.08 0.09 0.08
Number of ED symptoms (0-5) 0.62 0.64 0.66
Change in ED screen ('05 vs '07) 0.1
Dep. (PHQ>=10) & anxiety 0.03 0.02 0.05
Dep. & ED (positive SCOFF) 0.03 0.03 0.03
Anxiety & ED 0.01 0.01 0.02
12
Table 1 (cont'd)
Baseline
sample, fall
2005
(N=2,798)
Longitudinal
sample, in fall
2005(N=747)
Other independent variables Mean SD Mean SD
Female 0.48 0.48
Age
18-22 0.65 0.71
23-25 0.14 0.12
26-30 0.14 0.11
31+ 0.08 0.06
Race/ethnicity
Asian 0.20 0.18
Black 0.06 0.04
Hispanic 0.04 0.04
White 0.62 0.65
Multi 0.05 0.07
Other 0.03 0.02
Degree program
Bachelors 0.63 0.66
Masters 0.20 0.14
JD 0.04 0.02
MD 0.04 0.04
PhD 0.14 0.21
Finances while growing up:
Very poor 0.02 0.02
Enough to get by 0.26 0.25
Comfortable 0.56 0.58
Well-to-do 0.16 0.15
Undergrad admissions variables
SAT score (max=1600) 1292 143 1309 143
ACT score (max=36) 28.2 3.7 28.6 3.2
HS GPA (undergrads) 3.93 0.18 3.94 0.17
Grad student admissions variables
GRE (verb. + quant.) 1260 167 1287 154
LSAT (max=180) 163 7.6 165 7.4
GMAT (max=800) 663 67 671 60
MCAT (max=13) 11.3 1.1 11.6 0.85
Undergrad GPA (grad studs.) 3.47 0.41 3.55 0.38
Cumulative GPA (pre fall '05) 3.35 0.47 3.43 0.47
13
As a third dependent variable we examine the number of credit hours taken by
students during specific terms. The average number of credit hours is just over 13
in our fall 2005 sample. In our data we only observe the number of credit hours
completed in a semester, and not the number of credit hours that the student
signed up for at the beginning of the semester. Therefore we cannot examine
dropping classes during the semester as an outcome. Nevertheless, to the extent
that mental health problems cause students to drop courses during the semester,
we expect to observe a negative relationship between these problems and
completed course hours. Of course, this negative relationship may also stem in
part from students who anticipate that they will have mental health problems
during the semester and sign up for fewer courses from the beginning.
In addition to the dependent variables, a few key covariates are taken from
the university’s administrative records. The purpose of these covariates is to
control for academic performance prior to when we measure mental health in fall
2005. First, we include cumulative GPA at the university, prior to fall 2005, as a
covariate. To allow for a nonlinear relationship between this measure and the
dependent variables (academic outcomes during and after fall 2005), we construct
categorical dummy variables: none/missing (for first year students), 0.00-3.30,
3.30-3.69, 3.70-3.99, and 4.00-4.30.6 Second, we use admissions records as
additional covariates. For undergraduates, the admissions data include high
school GPA as well as SAT or ACT score (most students took one or the other).
We code high school GPA using categorical dummies as described above, and we
code admission test scores as quintile dummies (based on the distribution within
our sample). Similarly, for graduate students the admissions data include college
GPA and test scores (GRE, GMAT, LSAT, or MCAT), which we code as
categorical dummies in analogous fashion. The means of these and other
variables are shown in Table 1.
MENTAL HEALTH MEASURES
We measure symptoms of depression, anxiety, and eating disorders, which are
three of the most common types of mental disorders among adolescents and
young adults (Kessler et al., 2005). These disorders may affect academic
outcomes for reasons discussed earlier. In the surveys we measure these
6 These numerical intervals split the sample with nonmissing cumulative GPAs roughly into
quartiles; in sensitivity analyses we also use smaller intervals (dividing the sample roughly into
deciles) and find that the main results do not change. In additional sensitivity analyses, we control
for both cumulative GPA and GPA during the semester prior to baseline (i.e., winter 2005), to
account for the possibility that students with mental health problems in fall 2005 were already
experiencing declines in performance in the prior semester. Again, our main results remain the
same under this specification.
14
symptoms with widely used brief screens that have been validated in a range of
populations including young adults. To measure depression, we use the Patient
Health Questionnaire-9 (PHQ-9), a nine-item screening instrument based on the
nine DSM-IV criteria for a major depressive episode. This instrument asks the
respondent to indicate the frequency of various symptoms over the past two
weeks. Following previous studies (Huang, 2006; Weiss, 2006) as well as
common clinical use of this screen,7 we convert the responses to these nine items
to a continuous score on a 0-27 scale, with higher scores indicating higher
severity of depressive symptoms. This screening tool has been validated as
highly correlated with diagnosis by mental health professionals and more detailed
assessment tools in a variety of populations and settings (Diez-Quevedo et al.,
2001; Henkel et al., 2004; Kroenke et al., 2001; Martin et al., 2006; Spitzer et al.,
1999). To measure anxiety, we use the PHQ screens for panic disorder and
generalized anxiety disorder. These screens ask about symptoms over the past
four weeks, and have been validated as being highly correlated with clinical
diagnoses of these conditions (Spitzer et al., 1999). These anxiety screens do not
translate to continuous measures, so we simply use indicators for whether a
student has a positive screen for each anxiety disorder or not. To measure eating
disorders, we use the SCOFF screening instrument, a 5-item questionnaire
designed to identify subjects likely to have an eating disorder (Cotton et al., 2003;
Morgan et al., 1999; Parker et al., 2005). Each item is a yes/no question about a
current symptom, and we convert each student’s answers to a 0-5 score based on
the number of yes’s.8
Table 1 shows the mean values of the mental health measures in our
analytic samples (both baseline and longitudinal). Based on these measures,
depression is the most prevalent mental health condition, with 14% of students at
baseline scoring 10 or higher on the PHQ-9 (10 is often used as the threshold for a
positive screen). Eight percent of students have a positive screen for eating
disorders, 3% for generalized anxiety, and 2% for panic disorder. The prevalence
of these conditions in this student population is similar to that at other colleges
and universities we have studied.9 The prevalence of co-occurring conditions is
7 For example, in the National College Depression Partnership, a recently developed initiative to
improve depression care on college campuses, participating campuses are screening students and
monitoring their depressive symptoms using the continuous PHQ scores.
8 Some studies have used 2 yes’s as the cutoff for a positive screen (Parker et al., 2005) whereas
others have used 3 (Cotton et al., 2003). We focus instead on a continuous measure in order to
approximate severity.
9 In fall 2007, in addition to the follow-up survey used in the present study, a new data collection
was conducted with random samples at 13 colleges and universities nationwide, including the
large public university that is the setting for the present study. This set of schools represents a
relatively diverse mix in terms of geographic location and enrollment size: California State-Chico,
Emory, Miami of Ohio, New Mexico State, Penn State, Tufts, University of Michigan, UNC-
15
3% for depression and anxiety, 3% for depression and eating disorders, and 1%
for anxiety and eating disorders. As shown in Table 1, the longitudinal sample is
similar at baseline to the overall sample in terms of mental health, and the
longitudinal sample’s symptoms of depression and anxiety increase slightly
between 2005 and 2007.
OTHER COVARIATES
Our analysis includes several additional covariates measured in the survey, as
shown in Table 1: gender, age, race/ethnicity, degree program, and financial
situation while growing up. In previous work we have found each of these
variables to be independently associated with at least one of the mental health
conditions examined in the present study (Eisenberg et al., 2007b). In addition to
these variables, in our regressions we control for a vector of dummy variables
corresponding to the student’s year in degree program, as well as a vector of
dummies corresponding to the student’s field of study (divided into 22 different
categories, such as humanities, social science, natural science and math, business,
medicine, engineering, etc.).
EMPIRICAL FRAMEWORK
Our analysis of the baseline sample uses variants of the following regression
equation, which is analogous to the commonly used “value-added” empirical
model of academic achievement described by Todd and Wolpin (2003):
)()(3,...)2,1,(2),(10),( iittitinti XAcademicsMHAcademics
ε
β
β
β
β
+
+
+
+= −−+
Academics(i, t+n) refers to an academic outcome (GPA, credit hours, or dropping
out) for individual i as of semester t+n, where n>=0. MH(i, t) refers to mental
health during semester t (the baseline semester, fall 2005, in most regressions).
Academics(i, t-1, t-2, …) refers to academic outcomes prior to semester t (i.e., pre-
university GPA and test scores, and GPA at the university prior to semester t).
X(i) refers to other covariates measured in the survey (gender, age, race/ethnicity,
degree program, and financial situation while growing up). In our analyses of
GPA and credit hours we use ordinary least squares regressions. For the binary
outcome of dropping out, we find that the results remain similar across linear,
probit, and logistic specifications, and for ease of interpretation we present the
marginal effects for the probit specification.
Chapel Hill, UNC-Greensboro, UI-Springfield, UI-Chicago, UI-Urbana Champaign, and Yeshiva.
In preliminary analysis of these data, we have found that the estimated prevalence rates of
depression, anxiety, and eating disorders at the large public university in the present study are not
statistically different from the means for the overall sample of 13 schools.
16
In this framework, three factors could bias our estimates from the true
causal effect of mental health on academic outcomes. First, the causal path may
be bidirectional—the dependent variable (academic outcome) may affect our key
independent variables (mental health). This seems particularly plausible for
regressions in which the dependent variable is GPA or credit hours in the same
semester that we measure mental health. Even though course grades are
determined at the end of the semester (one to two months after we measure mental
health), these grades partly reflect midterms and assignments that may have taken
place prior to our measure of mental health. If poor grades on these earlier
assignments and midterms influence mental health, then this would bias our
estimates (presumably in a negative direction). On the other hand, this source of
bias would not apply to our estimates of how mental health in the baseline
semester relates to GPA in subsequent semesters. It is worth noting that our
results are not sensitive to whether we control for not only cumulative GPA prior
to baseline but also GPA during the semester prior to baseline (winter 2005),
which would at least account for the possibility that academic performance is
already declining prior to the baseline semester. Also, we find that GPA in the
previous semester (winter 2005) is not a significant predictor of mental health in
the baseline semester (fall 2005), conditional on the other covariates.
Second, there may be omitted variables bias. Mental health is of course
not randomly distributed (as we discuss in more detail in the next section), and
students with mental health problems may be different in ways that are correlated
with unobserved factors (e.g., the ability vector A noted earlier) that affect
academic performance. We control for several individual characteristics,
including multiple measures of prior academic performance, but we cannot rule
out this source of bias.
Third, the “value-added” framework may misspecify the true relationship
between inputs (mental health in our context) and academic achievement, if past
inputs affect current achievement even after conditioning on current inputs and
achievement as of the previous period. As a specification check for this issue,
Todd and Wolpin (2003) suggest adding past inputs as a covariate and examining
whether they have significant coefficients. We perform this check by including
previous diagnoses of any mental disorders and of depression specifically as
proxies for past mental health, and we find that these variables are not significant
and their inclusion does not change the coefficients on the current mental health.
To control for time-invariant individual characteristics, we turn to our
longitudinal data, with baseline and follow-up survey data from fall 2005 and fall
2007 respectively. We estimate the following linear regression of GPA on mental
health and individual and time (semester) fixed effects, which is analogous to the
within-person empirical approach discussed by Todd and Wolpin (2003):
17
),()(3)(2),(10),( titititi SemesterIndividualMHGPA
ε
β
β
β
β
+
+
++=
In this framework, identification of the effect of mental health on GPA (or credit
hours) depends on assumptions that bidirectional causality (as described above) is
not present and there are no time-variant omitted variables that are correlated with
mental health and significantly affect grades. These assumptions cannot be
definitively tested, of course, but are important to keep in mind in interpreting the
results of this study.
Another empirical issue is how to account for treatment of mental health
problems. In this study we omit treatment from our primary analyses, because it
is likely to be highly endogenous with respect to mental health and academic
outcomes (e.g., correlated with unmeasured or imperfectly measured factors such
as severity of symptoms and motivation to succeed in college). Including
treatment in the analysis would therefore add another layer of uncertain
assumptions for interpreting the results for mental health variables. In the final
section of this paper, however, we briefly mention results from analyses that
include treatment variables.
NON-RESPONSE ADJUSTMENTS
To account for survey non-response in our analyses, we include sample
probability weights in all regressions, although it is important to note at the outset
that none of our main results are sensitive to whether we include these weights.
The weights are equal to one divided by the predicted probability of survey
response, which is estimated using logistic regressions of survey response
(yes/no) on variables that are available for both responders and non-responders.
For the baseline sample, these variables come from administrative data on all
students randomly selected to be invited to the study, which include gender,
degree program, race/ethnicity, international/domestic citizenship, and GPA. The
weights are further adjusted using mental health data from an abbreviated survey
of a random sample of non-responders to the main survey, which indicates that
people with mental health problems were somewhat more likely to respond to the
main survey, as detailed in the appendix to Eisenberg et al (2007a). Other
significant positive correlates of responding to the survey are female gender and
graduate student status, and significant negative correlates are African-American
race and age (controlling for academic level). Responding at baseline is also
positively correlated with measures of prior academic performance, although this
relationship is not entirely consistent (response is lower among students in the
lowest admission test and admissions GPA categories, but not significantly
different across the other groups; also, response is highest among students in the
highest category of cumulative GPA at the university, but not otherwise different
18
across groups). Importantly, however, responding at baseline is not significantly
associated with the key outcome variables (GPA during fall 2005, and dropping
out subsequent to fall 2005), conditional on other covariates.
For the longitudinal sample, the weights are constructed based only on
information about students who were eligible for the follow-up survey (those who
completed the baseline survey in fall 2005 and were still enrolled in fall 2007).
The predicted probability of response at follow-up is estimated using
demographic and mental health variables from the survey at baseline. Thus, the
panel weights are intended to account only for non-response at follow-up.10 The
significant baseline predictors of nonresponse at follow-up are being female and
being a PhD student (both of which are positive predictors), and none of the
baseline mental health variables significantly predict response at follow-up.
4. RESULTS
In Table 2 we compare the means of covariates, including demographic
characteristics and prior academic measures, across mental health status. The
purpose of these comparisons is to acknowledge that mental health is not
randomly distributed across students; given this fact, examining differences in
observable characteristics by mental health status is useful for thinking about how
unobservable differences might bias our estimates from true causal effects.
Female students are more likely to experience symptoms of each mental
health problem of interest (depression, anxiety, and eating disorders). Younger
students are more likely to experience symptoms of depression and eating
disorders, but not anxiety. For the most part, the racial/ethnic composition of
students across mental health categories is consistent; the only exception is that
Asian students are less likely to report elevated anxiety.11 Table 2 also shows that
depressed or anxious students are more likely to report having grown up in a poor
10 We construct weights in this way because we are mainly concerned about accounting to the
extent possible for attrition from the baseline sample, rather than weighting the sample to be
representative of all students who were enrolled in both fall 2005 and fall 2007 (which does not
correspond to a population of any particular policy relevance).
11 It is important to note that while some of these differences may be related to how people report
their symptoms, as opposed to “true” differences in symptoms, the screening tools were designed
to minimize such biases and have been validated in settings and populations with a wide range of
demographic characteristics (Huang et al., 2006).
19
Table 2: Sample characteristics, by MH status (Baseline sample, N=2,798)
Depression (PHQ-9) Anxiety
(PHQ) ED (SCOFF)
Low
(0-9)
Med.
(10-
14)
High
(15-
27) Neg. Pos. Neg. Pos.
N 2,343 311 144 2,642 146 2,550 224
Female 0.47 0.49 0.60 * 0.47 0.68 * 0.46 0.76 *
Age
18-22 0.64 0.67 0.71 * 0.65 0.66 0.64 0.75 *
23-25 0.14 0.15 0.15 0.14 0.14 0.14 0.13
26-30 0.14 0.12 0.10 * 0.14 0.15 0.14 0.09 *
31+ 0.08 0.07 0.04 * 0.08 0.05 0.08 0.04 *
Race/ethnicity
Asian 0.20 0.21 0.18 0.21 0.13 * 0.20 0.20
Black 0.06 0.07 0.05 0.06 0.06 0.07 0.04
Hispanic 0.04 0.04 0.02 0.03 0.06 0.03 0.04
White 0.62 0.59 0.66 0.62 0.64 0.62 0.63
Multi 0.05 0.06 0.06 0.05 0.08 0.05 0.07
Other 0.03 0.03 0.03 0.03 0.03 0.03 0.03
Finances growing up:
Very poor 0.02 0.03 0.06 * 0.02 0.06 * 0.02 0.03
Enough to get by 0.25 0.30 0.29 0.26 0.28 0.26 0.22
Comfortable 0.57 0.47 0.57 0.56 0.54 0.56 0.56
Well-to-do 0.16 0.20 0.08 * 0.16 0.12 0.16 0.19
Admissions test percentile 0.52 0.51 0.51 0.52 0.45 * 0.52 0.49
HS GPA (undergrads) 3.94 3.93 3.89 3.94 3.92 3.93 3.95
Undergrad GPA (grad
studs.) 3.47 3.42 3.49 3.47 3.46 3.47 3.43
GPA at univ. (pre fall '05) 3.36 3.29 3.29 3.36 3.22 * 3.35 3.35
PHQ depress. score >= 10 0.00 1.00 1.00 0.12 0.71 * 0.13 0.35 *
Anxiety (positive screen) 0.01 0.12 0.46 * 0.00 1.00 0.04 0.08 *
ED (positive screen) 0.06 0.18 0.21 * 0.08 0.15 * 0.00 1.00
"*" denotes that the variable means are significantly different by MH categories to the left at p<0.05
20
family, whereas symptoms of eating disorders are not correlated with one’s
financial situation while growing up.
Measures of academic performance prior to the baseline survey differ
significantly by mental health status only in the case of anxiety disorders.
Students with positive screens for anxiety disorders have slightly lower
admissions test scores and cumulative GPAs at the university. Finally, the bottom
rows in Table 2 show that the three mental health conditions are significantly
correlated with each other, which is consistent with an extensive mental health
literature documenting the co-occurrence of disorders (Kessler, 2008). This co-
occurrence highlights the value of examining depression, anxiety, and eating
disorders simultaneously in our analyses, in order to disentangle their independent
associations with academic outcomes.
Overall, the comparisons of covariates by mental health status indicate
that, while mental health problems are far from randomly distributed in the
student population, they are relatively prevalent among nearly all types of
students that we examine. Therefore, the central comparisons in this study—
academic outcomes across mental health status—are based on comparisons of
students who differ somewhat but not drastically in terms of observable
characteristics.12 At the same time, the fact that students with mental health
problems are more likely to come from poor families and students with symptoms
of anxiety disorders in particular have lower prior academic performance raises
the question of how other, unmeasured differences across mental health status
might affect academic outcomes.
Our first set of main results are shown in Table 3, which reports the
association between mental health measured in fall 2005 and the GPA that
students obtained in that semester.13 Perhaps the most notable findings are that
depression has a significant negative association with GPA and that the co-
occurrence of depression and anxiety is associated with a significant additional
drop in GPA. The magnitude of the coefficients in column 7 indicates, for
12 To complement the simple comparisons in Table 2, we also examine the independent
associations between these covariates and mental health status using regressions. The results of
this analysis are largely consistent with the comparisons in Table 2. Female, age, and past
financial situation are independently associated with mental health (significant at p<0.05), whereas
previous academic performance generally is not (the exceptions are that admission test scores are
negatively associated with anxiety and eating disorders, and previous GPA at the university is
negatively associated with anxiety).
13 Among the estimated coefficients for covariates other than mental health (not shown in Table
3), the following were positive and significant at p<0.05: female, growing up in a “well-to-do”
family, admissions test score, and cumulative GPA at the university prior to the baseline survey.
The following covariates were negative and significant at p<0.05: being black or Hispanic, and
being a bachelor’s or JD student. Also, many of the 22 fields of study differed significantly from
each other in terms of mean GPA conditional on other covariates. These additional results are
available on request.
21
Table 3: Association between mental health and GPA in same semester
Linear regressions of GPA in fall 2005 on MH in fall 2005, with SEs in parentheses.
1 2 3 4 5 6 7
PHQ depression (0-27) -0.019 -0.022 -0.019 -0.013 -0.011
(0.003) (0.004) (0.005) (0.003) (0.004)
Panic disorder (0/1) -0.083 0.015 0.154 0.101 0.201
(0.097) (0.095) (0.124) (0.083) (0.106)
Gen. anxiety (0/1) -0.213 -0.023 0.212 -0.073 0.092
(0.078) (0.081) (0.149) (0.068) (0.133)
ED symptoms (0-5) 0.008 0.035 0.044 0.017 0.022
(0.013) (0.013) (0.023) (0.013) (0.022)
Dep. (0-27) * anx. (0/1) -0.021 -0.015
(0.010) (0.009)
Dep. (0-27) * ED (0-5) -0.002 -0.001
(0.003) (0.003)
Anx. (0/1) * ED (0-5) 0.045 0.026
(0.051) (0.049)
N 2209 2200 2189 2184 2184 1935 1935
Covariates NO NO NO NO NO YES YES
Covariates: Female, race dummies, degree program dummies (bachelors, masters, JD, MD, PhD, financial
situation growing up (categories listed in Table 2), admission test quintile dummies, admissions GPA
dummies (0-3.29, 3.3-3.69, 3.7-3.99, 4-4.3, or none/missing), cumulative GPA at university prior to fall 2005
(0-3.29, 3.3-3.69, 3.7-3.99, 4-4.3, or none/missing).
example, that a 15 point increase on the PHQ-9 scale (which would be the
difference between what are considered low levels and severe levels of depressive
symptoms) corresponds to a 0.17 drop in GPA in the absence of anxiety (p<0.01),
and a 0.40 drop in the presence of anxiety (p=0.10). These differences in GPA
are modest in absolute terms but represent, respectively, 0.3 and 0.7 standard
deviations in the GPA distribution, and would lower a student with the 50th
percentile GPA (3.61) down to the 37th and 23rd percentiles, respectively. Without
controlling for other variables, a positive screen for generalized anxiety is
significantly and negatively associated with GPA, whereas a positive screen for
panic disorder or eating disorders is not significantly associated with GPA. After
22
controlling for other mental health variables and covariates, however, generalized
anxiety is no longer significant and symptoms of eating disorders become
positively and significantly associated with GPA in some specifications. The
sensitivity of these results to the inclusion of other mental health measures
highlights again the importance of examining multiple conditions simultaneously,
given the prevalence of co-occurrence.
We also find that credit hours completed in fall 2005 are negatively related
to mental health problems during that semester (results available on request), but
these results are not significant at p<0.10. Even at the upper bound of the 95%
confidence interval, the estimated association between depressive symptoms and
credit hours is small, implying less than a one credit reduction for a 15 point
increase on the depression scale. This suggests that, although depression is
associated with lower GPA, it is not a significant contributor to people’s dropping
courses during the semester.14
Depression is a significant predictor of not only GPA but also the
likelihood of dropping out from the university (Table 4). Controlling for
covariates, each additional point on the depression scale is associated with a
0.31% increase in the probability of dropping out (p<0.01) (column 5 of the
table), which would imply that a 15 point increase on the depression scale
corresponds to a 4.7% increase in the probability of dropping out, or a 60%
increase relative to the mean probability of dropping out (8%). The coefficients
for the other mental health variables (panic disorder, generalized anxiety, and
eating disorders) are not significant at p<0.10, although the imprecise coefficient
for panic disorder implies a more than doubling of the probability of dropping out.
We also do not find any significant interactions between mental health variables
(results not shown).15 As mentioned earlier, the main pattern of results for
dropping out remain similar in linear and logistic specifications.
Given the signification association at baseline between depression and
GPA, we examine the persistence over time in this relationship in two ways.
First, we consider GPA during each semester following fall 2005 as a separate
outcome, and we estimate separate regressions of these GPAs on baseline mental
health and covariates. In Appendix Figure 1, each point in the figure represents
the estimated coefficient on depression from a regression where the dependent
14 This is likely to be related to the fact that at this university the deadline for dropping a course is
only three weeks after the beginning of the semester.
15 Among covariates other than mental health (not shown in Table 4), the only one that is
significantly associated with dropping out at p<0.05 is cumulative GPA at the university prior to
the baseline survey (as expected, lower GPAs are associated with significantly higher odds of
dropping out).
23
Table 4: Association between MH and dropping out (by winter 2008)
Probit models with "marginal effects" and SEs reported
1 2 3 4 5 6 7
Dep. score (0-27) 0.0024 0.0021 0.0005 0.0031 0.0022
(0.0011) (0.0013) (0.0016) (0.0011) (0.0014)
Panic disorder (0/1) 0.0552 0.0414 0.0387 0.0607 0.1120
(0.0508) (0.0479) (0.0643) (0.505) (0.0900)
Gen. anxiety (0/1) 0.0293 -0.0038 -0.0139 -0.0216 0.0109
(0.0312) (0.0275) (0.0504) (0.0186) (0.0572)
ED symptoms (0-5) 0.0053 0.0025 -0.0075 0.0009 -0.0067
(0.0052) (0.0052) (0.0086) (0.0048) (0.0076)
Dep. (0-27) * anx. (0/1) 0.0028 -0.0010
(0.0037) (0.0032)
Dep. (0-27) * ED (0-5) 0.0016 0.0012
(0.0010) (0.0008)
Anx. (0/1) * ED (0-5) -0.0336 -0.0297
(0.0205) (0.0168)
N 2798 2788 2774 2769 2769 2472 2472
Covariates NO NO NO NO NO YES YES
Covariates: Same as listed in note to Table 3.
variable is GPA measured for a different semester.16 The negative association
between baseline (fall 2005) mental health and semester GPA at subsequent time
points remains significant, and diminishes only slightly, over the course of 1.5
years (through winter 2006, fall 2006, and winter 2007). It is also important to
note that if depression at baseline makes students more likely to drop out, as
suggested by the results in Table 4, then this would probably bias the estimates
shown in Figure 1 towards zero (assuming that students who dropped out would
16 We also estimate analogous regressions for dropping out, in which the dependent variables are
defined as whether the student has dropped out as of each semester following the baseline
semester (fall 2005). The results indicate that depression in fall 2005 maintains a relatively
consistent relationship with the likelihood of dropping out by subsequent time points (results
available on request).
24
have obtained low grades if they had remained in school). Second, we explore
this pattern over time from a somewhat different angle. As in Appendix Figure 1,
we regress GPA during each semester on depression variables and covariates. The
difference here is that the key independent variable is a set of dummies referring
to the 2x2 combinations of depression status that one could have at baseline (fall
2005) and follow-up (fall 2007) (yes/yes, yes/no, no/yes, no/no). Therefore the
sample is restricted to students who completed both surveys. The idea is to
compare people with persistent or recurrent depression (yes/yes) to the
nondepressed (no/no) and the depressed at only one time point (yes/no and
no/yes). Although the coefficients are not all significant at p<0.05, the results
generally suggest that those with persistent or recurrent depression do
significantly worse than all three other groups.
In order to control for the effect of time-invariant individual characteristics
on GPA, we next analyze the longitudinal data using regressions with individual
and time fixed effects (Table 5). This analysis only applies to people surveyed in
both fall 2005 and fall 2007. When each mental health variable is included
separately, each has a negative association with GPA, with depression and panic
disorder significant at p<0.01 and generalized anxiety and eating disorders
significant at p<0.10. When the mental health variables are included together
(column 4), the results remain similar, except the depression coefficient declines
in absolute value and is not significant at conventional levels and generalized
anxiety is no longer significant. When we look at interactions between mental
health conditions (column 5), we see that co-occurring depression and anxiety has
a negative and significant association with lower GPA.17
Next, because our results described earlier reveal a significant relationship
between depression at baseline and subsequent GPA, we look inside the “black
box” of depression by examining the nine specific symptoms measured in the
survey (corresponding to the nine DSM-IV symptoms of major depression, as
noted earlier). In the columns of Appendix Table 2, the "separate" header
indicates that each cell in that column represents a separate regression for each
symptom, and the "together" header indicates that all nine symptoms (from the
PHQ-9) are in the same regression (i.e., the column refers to a single regression).
Following the general scoring system for the PHQ-9, each symptom is coded as 0-
3, depending on the frequency with which the symptom is reported for the
previous two weeks (0 = “not at all”, 1 = “several days”, 2 = “more than half the
days”, 3 = “nearly every day”). When entered into separate regressions, each
symptom is significantly associated with a lower GPA, which is not surprising
17 Fixed effects results for credit hours (available on request) are generally similar to the baseline
results for this dependent variable (Table 4). A positive screen for panic disorder is associated
with 1.56 fewer credit hours (p=0.03), whereas symptoms of eating disorders and depression are
negatively but not significantly associated with credit hours.
25
given that the overall depression index is significantly associated with a lower
GPA and the symptoms are correlated with each other. More importantly, when
all nine symptoms are entered into the same regression, the only symptom that
remains significant at p<0.05 is the first, anhedonia (“Little interest or pleasure in
doing things”). Also, one other symptom, psychomotor retardation or agitation, is
marginally significant (p=0.09) (“moving or speaking slowly? Or the opposite --
being fidgety or restless”).
Table 5: Fixed effects (within-person) regression of GPA on mental health
Linear regression with individual and time fixed effects, with coefficients and SE reported.
1 2 3 4 5
Depression score (0-27) -0.014 -0.009 0.002
(0.006) (0.006) (0.007)
Panic disorder (0/1) -0.470 -0.467 -0.115
(0.177) (0.167) (0.221)
Generalized anxiety (0/1) -0.148 -0.072 0.456
(0.096) (0.105) (0.244)
ED symptoms (0-5) -0.044 -0.046 -0.033
(0.026) (0.027) (0.043)
Depression (0-27) * anxiety (0/1) -0.045
(0.017)
Depression (0-27) * ED (0-5) -0.002
(0.004)
Anxiety (0/1) * ED (0-5) 0.078
(0.062)
N 1126 1139 1136 1117 1117
Note: The reported N's include two observations per student (there are 563 unique students in these regressions). Note
that the sample size is smaller than the full longitudinal sample because some students have missing GPA values, for
reasons discussed in the text.
POTENTIAL MECHANISM: TIME SPENT ON SCHOOL WORK
Although our data have little information about potential mechanisms by which
mental health could affect academic outcomes, we do have information on the
amount of time spent on school work in the fall 2005 survey. Participants were
asked, “During a typical week, about how many hours a week do you spend doing
26
work for school (includes time in class, doing homework or assignments,
studying, research)?” We find that the mean hours per week are 28 (25 for
undergraduates, 32 for graduate students). In a regression of hours of school
work per week on the mental health variables and all other covariates, we find
that panic disorder is the only mental health variable significantly related to hours
of school work (-6.0 hours, p=0.06). This suggests that the significant negative
association between depression and GPA is not due to less time spent on school
work, but rather the productivity of that time. Also, when we add time studying
as a covariate in our main GPA regression, the coefficient for depression hardly
changes (and in fact becomes slightly more negative), again suggesting that time
use is not the major mechanism.
STRATIFIED ANALYSES: GENDER AND ACADEMIC LEVEL
We run all of our main analyses separately by gender, motivated by the fact that
females report a higher prevalence of mental health problems and some of the
previous studies noted earlier find different relationships between mental health
and academic outcomes by gender. For the most part, however, we find that the
results are similar by gender. For example, the relationship between depression
score and GPA (as in Table 3, column 8) is -0.011 (p=0.01) for females and -
0.013 (p=0.01) for males, and the relationship between depression score and
dropping out (as in Table 4, column 5) is 0.0032 (p=0.03) for females and 0.0028
(p=0.05). On the other hand, a few differences by gender are notable. First, the
negative relationship between anxiety disorders and credit hours appears to be
driven by females, for whom the coefficient on panic disorder is -0.91 (p=0.01)
and the coefficient on generalized anxiety is -0.96 (p=0.04), whereas the
coefficients for males are small and insignificant. Second, among the nine
symptoms of depression, for females the most significant are anhedonia (-0.038,
p=0.13), sleep impairment (-0.029, p=0.14), and appetite problems (-0.037,
p=0.09), whereas for males the most significant are anhedonia (-0.067, p=0.06)
and psychomotor retardation or agitation (-0.088, p=0.04). This suggests that the
ways in which depression impairs people may differ significantly by gender.
Finally, in the fixed effects analysis of GPA, we find that the results are stronger
for depression, panic disorder, and co-occurring depression and anxiety among
females, whereas the results are stronger for generalized anxiety among males.
We also run our analyses separately for undergraduates and graduate
students, because these two groups are different in age and academic demands,
among other potentially important factors. Two of our key results—the
significant associations between depression at baseline and both GPA and
dropping out—remain nearly identical for both undergraduates and graduate
students. On the other hand, a few results are stronger for undergraduates than
27
graduate students. First, co-occurring depression and anxiety are significantly
associated with lower GPA for undergraduates but not graduate students. Second,
panic disorder is significantly associated with higher drop-out among
undergraduates but not graduate students. Third, in the fixed effects results,
depression and panic disorder are negatively and significantly associated with
GPA for undergraduates, but not for graduate students. Collectively, these results
suggest that the general negative relationship between mental health and academic
outcomes is more robust for undergraduates than graduate students at the
institution in our study.
SENSITIVITY CHECKS
We examine the sensitivity of our results to including three additional control
variables relevant to college life: current financial situation (whether one’s
situation is a “struggle”, “tight,” or “not a problem”), frequent binge drinking
(defined as consuming at least four drinks if female, or five drinks if male, on at
least three occasions in the previous two weeks), and exercise (hours per week on
average in the previous month). We find that including these variables does not
alter any of our main findings for the mental health variables, but it is interesting
to note that binge drinking is negatively and significantly related to GPA in both
the baseline (-0.06, p=0.05) and fixed effects (-0.154, p=0.02) regressions. Binge
drinking (and other substance use) is not a focus of this paper, but appears to
warrant attention in future research on health and behavioral determinants of
college academic outcomes.
5. DISCUSSION
While our general finding that mental health problems are associated with lower
academic success is consistent with our prediction, a number of specific findings
raise additional questions.
First, the results for eating disorders are notably different between our
analysis using only baseline measures of mental health and our within-person
analysis using measures at two time points. The latter analysis indicates a
negative association with GPA, whereas the cross-sectional results indicate a
positive (though not significant) association. The cross-sectional estimates may be
confounded by the fact that people prone to eating disorders also tend to have
personality characteristics that can enhance their academic performance, such as
perfectionism and obsessive attention to detail (Halmi, 2000; Kaye et al., 2004).
In the longitudinal analysis, where we look at within-person differences, we may
be seeing that these people actually have better academic performance when they
are not actively experiencing the symptoms of eating disorders. This is just one
28
possibility, and further research that adjusts for these types of personality
characteristics would be useful to understand these discrepancies.
Another contrast between the baseline and longitudinal analysis is that the
negative association between depression and GPA is weaker in the latter analysis.
This may reflect a combination of factors, including: a) in contrast to the
explanation offered for the eating disorders results, in the case of depression time-
invariant personal characteristics might not bias the estimates towards zero; b) if
depression causes people to be more likely to drop out, as suggested by our results
in Table 4, this may bias the fixed effects estimates towards zero (as students who
would have both increased depression and lower grades at follow-up are not in the
sample); c) as suggested by the results in Appendix Table 1, the persistence of
depression may be a pivotal factor in the extent to which it causes impairment,
and our within-person fixed effects analysis cannot identify the effects of
depression for people with similarly elevated depressive symptoms at both time
points. The apparent importance of persistent depression may be related to the
fact that these people tend to have greater impairment to verbal memory than
people with first episodes of depression (Fossati, 2004). It is also possible that
some students with persistent depression are in a self-perpetuating cycle, in which
depression impairs performance, which in turns lowers one’s self-assessment of
abilities, which in turn contributes to continued depression and lower investment
in school work. These possibilities are concerning in light of the fact that
depression is often lasting and recurring (Eaton, 2008).18
Another interesting aspect of the results for depression is that anhedonia is
significantly associated with GPA, independent of other depressive symptoms,
whereas negative affect (the second symptom, “feeling down, depressed, or
hopeless”) is not. This appears to highlight the fact that many students can feel
severely depressed but still remain highly functional. Serious impairment in
academic functioning appears to arrive only once someone loses interest or
enjoyment in usual activities.
Our finding that co-occurring depression and anxiety appear to be
especially impairing is consistent with clinical and epidemiological research.
Compared to either depression or anxiety alone, co-occurring depression and
anxiety (often referred to as anxious depression) is associated with substantially
higher severity of illness (Joffe, 1993), functional impairment (Joffe, 1993;
Kessler, 1999), and chronicity (Van Valkenburg, 1984). The co-occurrence of
these two disorders also frequently predicts poor treatment outcomes (Brown &
Madonia, 1996). Given the high prevalence of both depression and anxiety
disorders among college students (Blanco et al., 2008), improving knowledge
18 We have also found in other analyses of the data sets in the present study that depression in fall
2005 is highly correlated with depression in fall 2007 (Zivin et al., 2009)).
29
about their co-occurrence and how this affects academic outcomes would be
valuable.
Although this study represents a first step, future studies are needed to
characterize the course of mental health over time within individuals in college.
In our study we do not necessarily know whether students are experiencing
mental health problems for the first time or if their symptoms are a continuation
of earlier problems. To some extent our fixed effects analysis controls for this
issue, by focusing on within-person changes, but the effects of mental health on
academic performance may vary substantially over time for a given individual,
particularly if they find treatments or other strategies to cope effectively with
these problems. In analyses not shown, we explore these possibilities in two
simple ways. First, we include variables indicating treatment (medication or
therapy/counseling in the previous year) in the regressions.19 These results,
however, indicate no significant relationship between treatment and academic
outcomes, which may be because the benefits of treatment are confounded by
negative selection (more severe cases) into treatment. Second, we conduct our
analysis with the sample restricted to students who reported being diagnosed with
a mental disorder (by fall 2005), which includes 19% of the sample (and
depression is the most common diagnosis, at 13%). Most of these students
presumably have been aware of mental health problems for some time. We find
that these results, though imprecise due to the smaller sample, are largely the
same as for the full sample.
POTENTIAL POLICY IMPLICATIONS
As noted earlier, the data in this study have unique advantages but do not permit
definitive causal estimates. Regardless of causality, however, the estimates
pertain to the question of whether including mental health criteria would be
worthwhile in efforts to screen for risk of poor academic outcomes such as
dropping out (with the aim of intervening to reduce drop-out rates among those
identified as higher risk). As a simplified example, consider the question of
whether to use a risk screen based on the previous semester’s GPA (“high risk” if
GPA<3.0, “low risk” if not) versus a screen based on both GPA and mental health
criteria (“high risk” if GPA<3.0 or positive screen for depressive or anxiety
disorder). Applying these criteria to our sample, we find that adding the mental
health criteria would substantially increase the proportion of eventual drop-outs
who are identified (from 11% to 30% of all drop-outs). However, at the
university in our study drop-out rates are low enough (less than 10% over more
19 In other work we document that about 30% of students with a positive screen for depression or
an anxiety disorder received treatment (medication or therapy) in the previous year (Eisenberg et
al., 2007).
30
than a two-year period for our baseline sample) that screening solely for the
purpose of reducing the drop-out rate may not be cost-effective. Even in the
highest risk category in the example above (those with low GPA and a positive
screen for a mental disorder), only 25% of students go on to drop out. This
implies a low specificity of screening efforts (i.e., a high “false positive” rate).
On campuses with higher drop-out rates, screening would be more likely to be
cost-effective, and our estimates suggest that mental health criteria may
substantially increase the identification of those at risk. Alternatively, mental
health criteria could be used to restrict the screening criteria and increase the
efficiency of screening: note that in the example above, we find that the drop-out
rate is 25% among students who meet both GPA and mental health criteria, as
compared to 9% among students who only meet the GPA criterion.
As a second policy consideration, we examine what our estimates would
imply about the economic returns to increasing treatment for mental disorders
among college students, if the estimates were assumed to be reasonable proxies
for the causal effect of mental health on remaining in school. This informal
exercise in cost-benefit analysis can illustrate the relevance of establishing this
relationship more definitively with a randomized trial of depression treatment
among college students. To begin, we assume that treating depression with
medication has an average effect of reducing depressive symptoms by
approximately one standard deviation.20 This translates to a reduction of -4.6
points on the PHQ depression score in our sample, which, when combined with
our estimate of the association between PHQ score and dropping out within one
year (0.003), translates to a reduction of -0.0138 in the probability of dropping out
due to treatment. We next assume that the marginal present discounted value of
earnings due to a year of college is approximately $50,000, using 2008 Current
Population Survey (CPS) earnings data with an assumption of a 7% return to a
year of schooling (Card, 1999) and a 3% discount rate. On the other hand, the
opportunity costs of a year in college would include foregone earnings (a
difference of approximately $9,000 in earnings per year for 19-22 year-olds with
a high school but no college degree, as compared to 19-22 year old college
students, according to the 2008 CPS) and tuition costs (about $10,000 per year on
average at four-year institutions) (National Center for Education Statistics, 2008).
If we assume that dropping out implies a loss of two years of college, on average
20 For simplicity we focus on medication, for which a recent meta-analysis indicates that the
average effect size, relative to placebo, is a 0.8 standard deviation reduction in the level of
depressive symptoms (Rief et al., 2009). Given that the placebo has therapeutic value in itself, the
true clinical benefit of medication is likely to be significantly more than 0.8 standard deviations,
so one standard deviation can be considered a conservative estimate. Certain forms of
psychotherapy such as cognitive behavioral therapy have been shown to have similar clinical
benefits as medication (U.S. Department of Health and Human Services, 1999), and would
therefore imply similar returns in this exercise (though at a somewhat higher cost of treatment).
31
(note that most drop-outs tend to occur early in a degree program, but some
people will return to college at a different school, which we cannot observe), then
the numbers would imply that the net income return from treatment in college is
about $860. This benefit is larger than the average cost of outpatient treatment
(including medication costs and physician time) (Valenstein et al., 2001), even
though it does not account for potential returns via improved learning and GPA,
nor does it account for arguably the main benefit of treatment, improved quality
of life. Colleges and universities may be interested in how these estimates would
apply to a screening program that aims to reduce the prevalence of untreated
depression. If a school conducted a universal screening program, then it might
expect to identify 7% of its students with untreated depression (out of a total of
10% with untreated depression, based on our data), given that the PHQ-9 screen
has a sensitivity in the range of 70% (Kroenke et al., 2001). If the screening
program increased the probability of receiving treatment by 20% among this 7%,
then the screening program could yield about $12 per student in the overall
population in economic returns ($860 * 20% * 7%). If we account for outpatient
treatment costs (which the school would incur for the majority of students, based
on our data on service use), and assume these costs to be in the range of $400 on
average (Valenstein et al., 2001), then the net economic benefit would be about $7
per student in the overall population. This amount compares favorably to the
costs per person of administering a brief depression screen, which are estimated to
be $5 (Valenstein et al., 2001). Of course, these estimates are imprecise and
subject to many assumptions, but they suggest that the economic returns to
depression treatment in college may be significant and would be worth
quantifying more accurately. This will probably require a relatively large
randomized trial of mental health treatment that collects academic outcomes,
because it is difficult to imagine a naturally occurring instrumental variable that
significantly affects mental health in college without affecting other factors that
contribute to academic success.21
CONCLUSIONS
This descriptive analysis shows that depression, anxiety, and eating disorders are
significantly associated with academic outcomes among college students. To the
extent that these represent causal relationships, college campuses may be able to
further their central educational missions, and generate significant economic
21 One possibility, however, would be to exploit the substantial variation in the supply of mental
health services across campuses and over time as a quasi-experiment. It remains to be seen
whether this quasi-experiment is exogenous or strong enough (particularly because the
effectiveness of treatment depends on a number of consumer and provider factors).
32
returns for society, by investing in mental health resources. Randomized studies
of mental health treatment in college populations would be valuable for further
clarifying the potential for these benefits. In addition, the association between
mental health and GPA may be relevant to improving understanding of the
broader issue of how mental health affects productivity more generally.22
Because workplace productivity is typically difficult to measure, GPA may
represent a useful proxy for studying issues that could plausibly generalize from
academic to employment settings. To the extent that productivity fluctuates as a
function of mental health, whether in an academic or workplace setting, this
suggests a wrinkle in the concept of human capital that is relevant for a sizeable
proportion of the population.
22 A number of studies suggest that mental health affects work productivity (Ettner et al., 1997;
Marcotte & Wilcox-Gök, 2001), and a recent randomized study indicated that improved treatment
of depression can increase productivity (Wang et al., 2007).
33
Appendix Table 1: Semester GPA as a fx of depression at baseline/follow-up
Linear regressions with coefficients and standard errors reported.
N NO/NO NO/YES YES/NO YES/YES
Fall 2005 575 (ref) 0.072 -0.081 -0.098
(0.067) (0.076) (0.083)
Winter 2005 550 (ref) 0.020 -0.039 -0.105
(0.072) (0.096) (0.088)
Fall 2006 499 (ref) -0.048 0.060 -0.189
(0.087) (0.106) (0.104)
Winter 2007 462 (ref) 0.057 -0.089 -0.175
(0.089) (-0.129) (0.087)
Fall 2007 453 (ref) -0.001 0.073 -0.109
(0.071) (0.107) (0.090)
Winter 2008 392 (ref) 0.024 -0.035 -0.160
(0.083) (0.089) (0.080)
Notes: Each row corresponds to a separate regression. 'NO/YES' means, for example, that the student did not have
a positive screen for depression baseline, but did at follow-up two years later.
34
Appendix Table 2: Association between GPA and depressive symptoms
1 2 3 4
SEPAR TOGETH SEPAR TOGETH
1: Little interest or pleasure in doing things -0.112 -0.091 -0.067 -0.054
(0.023) (0.032) (0.019) (0.024)
2: Feeling down, depressed or hopeless -0.057 0.039 -0.04 0.005
(0.019) (0.028) (0.017) (0.023)
3: Trouble falling or staying asleep,
or sleeping too much
-0.068 -0.036 -0.024 -0.005
(0.017) (0.020) (0.015) (0.017)
4: Feeling tired or having little energy -0.051 0.007 -0.014 0.022
(0.019) (0.024) (0.015) (0.019)
5: Poor appetite or overeating -0.052 -0.002 -0.043 -0.023
(0.016) (0.019) (0.017) (0.018)
6: Feeling bad about yourself -- or that you are
a failure or have yourself or your family down
-0.068 -0.012 -0.042 -0.012
(0.018) (0.024) (0.016) (0.021)
7: Trouble concentrating on things, such as reading the
newspaper or watching television
-0.083 -0.027 -0.053 -0.024
(0.019) (0.023) (0.018) (0.019)
8: Moving or speaking slowly? Or the opposite --
being fidgety or restless
-0.11 -0.049 -0.077 -0.047
(0.029) (0.030) (0.027) (0.028)
9: Thoughts that you would better off dead
or hurting yourself -0.1 -0.019 -0.042 0.012
(0.032) (0.041) (0.032) (0.035)
Covariates X X
Notes: Baseline sample is used (Ns are same as in Table 3); and covariates are same as listed below Table 3.
'SEPAR' denotes that the symptoms are entered into separate regressions (each row in the column is from a
separate regression), whereas 'TOGETH' denotes that all symptoms are entered into the same regression (the
column refers to a single regression).
35
Appendix Figure 1: Association between GPA and baseline depression
-0.025
-0.02
-0.015
-0.01
-0.005
0
0.005
0.01
F05 W06 F06 W07 F07 W08
Notes: “F05” refers to the Fall 2005 semester, “W06” the Winter 2006 semester, etc. Each point in the solid line refers to
the estimated coefficient for PHQ depression score (0-27), controlling for anxiety, ED symptoms, and all covariates noted
below Table 3, in a regression with the GPA in that semester as the dependent variable. The dotted lines refer to the upper
and lower bounds of the 95% confidence intervals.
REFERENCES
Andrews, B., & Wilding, J. M. (2004). The Relation of Depression and Anxiety to
Life-stress and Achievement in Students. British Journal of Psychology, 95,
509.
Becker, G. S. (1993). Human Capital: A Theoretical and Empirical Analysis, with
Special Reference to Education (3rd ed.). Chicago: University Of Chicago
Press.
Berkner, L., & Choy, S. (2008). Descriptive Summary of 2003-04 Beginning
Postsecondary Students: Three Years Later. National Center for Education
Statistics Paper #2008-174.
Berndt, E. R., Koran, L. M., Finkelstein, S. N., Gelenberg, A. J., Kornstein, S. G.,
Miller, I. M., et al. (2000). Lost Human Capital from Early-onset Chronic
Depression. The American Journal of Psychiatry, 157(6), 940-947.
Blanco, C., Okuda, M., Wright, C., Hasin, D. S., Grant, B. F., Liu, S., et al.
(2008). Mental Health of College Students and their Non-College-Attending
36
Peers: Results from the National Epidemiologic Study on Alcohol and
Related Conditions. Archives of General Psychiatry, 65(12), 1429-1437.
Breslau, J., Lane, M., Sampson, N., & Kessler, R. C. (2008). Mental Disorders
and Subsequent Educational Attainment in a US National Sample. Journal of
Psychiatric Research, 42(9), 708-716.
Brown, C., & Madonia, J. (1996). Treatment Outcomes for Primary Care Patients
with Major Depression and Lifetime Anxiety Disorders. The American
Journal of Psychiatry, 153(10), 1293.
Card, D. (1999). Chapter 30: The Causal Effect of Education on Earnings. In O.
Ashenfelter, & D. Card (Eds.), Handbook of labor economics (pp. 1801-
1863) Elsevier.
Cotton, M. A., Ball, C., & & Robinson, P. (2003). Four Simple Questions can
Help Screen for Eating Disorders. Journal of General Internal Medicine,
18(1), 53-56.
Cunha, F., & Heckman, J. J. (2006). Investing in our Young People. Unpublished
Manuscript, University of Chicago, Department of Economics.
Currie, J. (2008). Healthy, Wealthy, and Wise: Socioeconomic Status, Poor
Health in Childhood, and Human Capital Development. NBER Working
Paper, No. 13987
Currie, J., & Stabile, M. (2006). Child Mental Health and Human Capital
Accumulation: The Case of ADHD. Journal of Health Economics, 25(6),
1094-1118.
Cutler, D., & Lleras-Muney, A. (2006). Education and Health: Evaluating
Theories and Evidence. NBER Working Paper #12352.
Diez-Quevedo, C., Rangil, T., Sanchez-Planell, L., Kroenke, K., & Spitzer, R. L.
(2001). Validation and Utility of the Patient Health Questionnaire in
Diagnosing Mental Disorders in 1003 General Hospital Spanish Inpatients.
Psychosomatic Medicine, 63(4), 679-86.
Ding, W., Lehrer, S., Rosenquist, J., & Audrain-McGovern, J. (2007). The Impact
of Poor Health on Education: New Evidence using Genetic Markers.
Working Paper, Queens University,
Dynarski, S. M. (1999). Does Aid Matter? Measuring the Effect of Student Aid
on College Attendance and Completion. NBER Working Paper #7422.
Eaton, W. W. (2008). Population-based Study of First Onset and Chronicity in
Major Depressive Disorder. Archives of General Psychiatry, 65(5), 513.
Eisenberg, D., Golberstein, E., & Gollust, S. E. (2007a). Help-Seeking and
Access to Mental Health Care in a University Student Population. Medical
Care, 45(7), 594-601.
Eisenberg, D., Gollust, S. E., Golberstein, E., & Hefner, J. L. (2007b). Prevalence
and Correlates of Depression, Anxiety, and Suicidality among University
Students. American Journal of Orthopsychiatry, 77(4), 534-542.
37
Ettner, S. L., Frank, R. G., & Kessler, R. C. (1997). The Impact of Psychiatric
Disorders on Labor Market Outcomes. Industrial and Labor Relations
Review, 51(1), 64-81.
Fergusson, D. M., & Woodward, L. J. (2002). Mental Health, Educational, and
Social Role Outcomes of Adolescents with Depression. Archives of General
Psychiatry, 59(3), 225-231.
Fletcher, J. (2008). Adolescent Depression: Diagnosis, Treatment, and
Educational Attainment. Health Economics, 17(11), 1215-1235.
Fletcher, J., & Lehrer, S. F. (2008). Using Genetic Lotteries within Families to
Examine the Causal Impact of Poor Health on Academic Achievement.
Working Paper, Yale University.
Fletcher, J., & Wolfe, B. (2007). Child Mental Health and Human Capital
Accumulation: The Case of ADHD Revisited. Journal of Health Economics,
27(3), 794-800.
Fossati, P. (2004). Verbal Memory Performance of Patients with a First
Depressive Episode and Patients with Unipolar and Bipolar Recurrent
Depression. Journal of Psychiatric Research, 38(2), 137.
Halmi, K. A. (2000). Perfectionism in Anorexia Nervosa: Variation by Clinical
Subtype, Obsessionality, and Pathological Eating Behavior. The American
Journal of Psychiatry, 157(11), 1799.
Henkel, V., Mergl, R., Kohnen, R., Allgaier, A. K., Moller, H. J., & Hegerl, U.
(2004). Use of Brief Depression Screening Tools in Primary Care:
Consideration of Heterogeneity in Performance in Different Patient Groups.
Gen Hosp Psychiatry, 26(3), 190-8.
Horn, L., & Berger, R. (2004). College Persistence on the Rise? Changes in 5-
Year Degree Completion and Postsecondary Persistence Rates between 1994
and 2000. (U.S. Department of Education, National Center for Education
Statistics. No. NCES 2005–156). Washington, DC: U.S. Government
Printing Office.
Huang, F. Y. (2006). Racial and Ethnic Differences in the Relationship between
Depression Severity and Functional Status. Psychiatric Services, 57(4), 498.
Huang, F. Y., Chung, H., Kroenke, K., Delucchi, K. L., & Spitzer, R. L. (2006).
Using the Patient Health Questionnaire-9 to Measure Depression among
Racially and Ethnically Diverse Primary Care Patients. Journal of General
Internal Medicine, 21(6), 547-52.
Hysenbegasi, A., Hass, S. L., & Rowland, C. R. (2005). The impact of depression
on the academic productivity of university students. The Journal of Mental
Health Policy and Economics, 8(3), 145-151.
Jaeger, D. A., & Page, M. (1996). Degrees Matter: New Evidence on Sheepskin
Effects in the Returns to Education. Review of Economics and Statistics,
78(4), 733-740.
38
Joffe. (1993). Anxious and Nonanxious depression. The American Journal of
Psychiatry, 150(8), 1257.
Kane, T. J., & Rouse, C. E. (1995). Labor-Market Returns to Two-and Four-Year
College. American Economic Review, 85(3), 600-614.
Kaye, W. H., Bulik, C. M., Thornton, L., Barbarich, N., & Masters, K. (2004).
Comorbidity of Anxiety Disorders with Anorexia and Bulimia Nervosa.
American Journal of Psychiatry, 161(12), 2215-21.
Kessler, R. C. (1999). Impairment in Pure and Comorbid Generalized Anxiety
Disorder and Major Depression at 12 Months in Two National Surveys. The
American Journal of Psychiatry, 156(12), 1915.
Kessler, R. C. (2008). Co-morbid Major Depression and Generalized Anxiety
Disorders in the National Comorbidity Survey Follow-up. Psychological
Medicine, 38(3), 365.
Kessler, R. C., Berglund, P., Demler, O., Jin, R., Merikangas, K. R., & Walters,
E. E. (2005). Lifetime Prevalence and Age-of-Onset Distributions of DSM-
IV Disorders in the National Comorbidity Survey Replication. Archives of
General Psychiatry, 62(6), 593-602.
Knapp, L. G. (2007). Enrollment in Postsecondary Institutions, Fall 2005
Graduation Rates, 1999 and 2002 Cohorts and Financial Statistics, Fiscal
Year 2005. First Look. NCES 2007-154. Washington, D.C.: National Center
for Education Statistics.
Kroenke, K., Spitzer, R. L., & Williams, J. B. (2001). The PHQ-9: Validity of a
Brief Depression Severity Measure. Journal of General Internal Medicine,
16(9), 606-13.
Laanan, F. S. (2007). Studying Transfer Students. Part II: Dimensions of Transfer
Students' Adjustment. Community College Journal of Research and Practice,
31(1), 37-59.
Marcotte, D., & Wilcox-Gök, V. (2001). Estimating the Employment and
Earnings Costs of Mental Illness: Recent Developments in the United States.
Social Science & Medicine, 53(1), 21-27.
Marcotte, D., Wilcox-Gök, V., Farahati, F., & Borkoski, C. (2004). Early Onset
Depression and High School Dropout. In D. Marcotte, & V. Wilcox-Gök
(Eds.), The Economics of Gender and Mental Illness (pp. 27-51) Elsevier
Science.
Martin, A., Rief, W., Klaiberg, A., & Braehler, E. (2006). Validity of the Brief
Patient Health Questionnaire Mood Scale (PHQ-9) in the General Population.
Gen Hosp Psychiatry, 28(1), 71-7.
McLeod, J., & Owens, T. (2004). Psychological Well-being in the Early Life
Course: Variations by Socioeconomic Status, Gender, and Race/Ethnicity.
Social Psychology Quarterly, 67(3), 257-278.
39
Michaud, C. M., McKenna, M. T., Begg, S., Tomijima, N., Majmudar, M.,
Bulzacchelli, M. T., et al. (2006). The Burden of Disease and Injury in the
United States 1996. Population Health Metrics, 4, 11.
Miech, R. A., Caspi, A., Moffitt, T. E., Wright, B. R. E., & Silva, P. A. (1999).
Low Socioeconomic Status and Mental Disorders: A Longitudinal Study of
Selection and Causation during Young Adulthood. American Journal of
Sociology, 104(4), 1096-1131.
Morgan, J. F., Reid, F., & Lacey, J. H. (1999). The SCOFF Questionnaire:
Assessment of a New Screening Tool for Eating Disorders. British Medical
Journal, 319, 1467-1468.
National Center for Education Statistics. (2008). Digest of Education Statistics:
2007 No. NCES 2008-022). Washington, D.C.: National Center for
Education Statistics, Institute of Education Sciences, U.S. Department of
Education.
Parker, S. C., Lyons, J., & Bonner, J. (2005). Eating Disorders in Graduate
Students: Exploring the SCOFF Questionnaire as a Simple Screening Tool.
Journal of American College Health, 54(2), 103-107.
Rief, W., Nestoriuc, Y., Weiss, S., Welzel, E., Barsky, A. J., & Hofmann, S. G.
(2009). Meta-Analysis of the Placebo Response in Antidepressant Trials.
Journal of Affective Disorders, in press.
Ross, C. E., & Mirowsky, J. (1999). Refining the Association between Education
and Health: The Effects of Quantity, Credential, and Selectivity.
Demography, 36(4), 445-460.
Sadock, B. J., & Sadock, V. A. (2000). In Sadock B. J., Sadock V. A. (Eds.),
Kaplan and Sadock's Comprehensive Textbook of Psychiatry (Seventh ed.).
Philadelphia: Lippincott Williams & Wilkins.
Skahill, M. P. (2002). The role of Social Support Network in College Persistence
among Freshman Students. Journal of College Student Retention, 4(1), 39-52.
Spitzer, R. L., Kroenke, K., & Williams, J. B. (1999). Validation and Utility of a
Self-Report Version of PRIME-MD: The PHQ Primary Care Study. JAMA,
282(18), 1737-1744.
Stinebrickner, R., & Stinebrickner, T. R. (2008). Learning about Academic
Ability and the College Drop-Out Decision. University of Western Ontario,
Working Paper.
Tchanturia, K. (2004). Cognitive Flexibility in Anorexia Nervosa and Bulimia
Nervosa. Journal of the International Neuropsychological Society, 10(4),
513.
Tinto, V. (1998). Colleges as Communities: Taking Research on Student
Persistence Seriously. The Review of Higher Education, 21(2), 167-177.
40
Todd, P. E., & Wolpin, K. I. (2003). On the Specification and Estimation of the
Production Function for Cognitive Achievement. Economic Journal, 113, 3-
33.
U.S. Department of Education. (2006). Digest of Education Statistics, 2005. No.
NCES 2006-005). Washington, DC.: U.S. Department of Education, National
Center for Education Statistics.
U.S. Department of Health and Human Services. (1999). Mental Health: A Report
of the Surgeon General. Rockville, MD: National Institute of Mental Health.
Valenstein, M., Vijan, S., Zeber, J. E., Boehm, K., & Buttar, A. (2001). The Cost-
Utility of Screening for Depression in Primary Care. Annals of Internal
Medicine, 134(5), 345-360.
Van Valkenburg. (1984). Anxious Depressions: Clinical, Family History, and
Naturalistic Outcome: Comparisons with Panic and Major Depressive
Disorders. Journal of Affective Disorders, 6(1), 67.
Vander Stoep, A., Weiss, N. S., Kuo, E. S., Cheney, D., & Cohen, P. (2003).
What Proportion of Failure to Complete Secondary School in the US
Population is Attributable to Adolescent Psychiatric Disorder? The Journal of
Behavioral Health Services and Research, 30(1), 119-124.
Wang, P., Simon, G. E., Avorn, J., Azocar, F., Ludman, E. J., McCulloch, J., et al.
(2007). Telephone Screening, Outreach, and Care Management for Depressed
Workers and Impact on Clinical and Work Productivity Outcomes - A
Randomized Controlled Trial. JAMA, 298(12), 1401-1411.
Weiss, B. (2006). Literacy Education as Treatment for Depression in Patients with
Limited Literacy and Depression: A Randomized Controlled Trial. Journal of
General Internal Medicine, 21(8), 823.
Zivin, K., Eisenberg, D., Golberstein, E., & Gollust, S. (2009). Persistence of
Mental Health Problems and Needs in a College Student Population. Journal
of Affective Disorders, in press.