Rees and Sabia 323
is the indicator variable defined above. Our focus is on the estimate of 
2
. If 
2
is less than zero this could be interpreted as evidence that, as hypothesized, migraine
headache imposes a cost on suffers in terms of reduced human capital accumulation.
We include a large number of individual- and family-level variables in X
i
. Most are measured at Wave I, including the educational attainment of the parent who
answered the parental questionnaire, household income, parents’ marital status, the respondent’s Peabody Picture Vocabulary Test PPVT score a measure of cognitive
ability, race, ethnicity, height, body mass index BMI, and whether the respondent had an older sibling. A set of age dummies were constructed from information
available in the Wave III in-home survey.
OLS results for the full sample and by gender are presented in the first column of Table 2. Even controlling for family background and individual characteristics,
there is a strong negative relationship between migraine headache as an adolescent and educational attainment. Among male respondents, migraine headache is asso-
ciated with a 0.105 decrease in high school GPA and a 0.060 decrease in the prob- ability of attending college. Among females, migraine headache is associated with
a 0.079 decrease in high school GPA, a 0.051 decrease in the probability of gradu- ating high school, and a 0.076 decrease in the probability of attending college.
In terms of magnitude, the impact of migraine headache on educational attainment appears to be substantial. For instance, the average high school GPA of male Add
Health respondents in our sample was 2.40. Interpreting the OLS estimates as causal for the moment, they suggest that, among males, migraine headache leads to a 4.4
percent reduction 0.1052.40 in high school GPA. Or, to take another example, migraine headache is associated with an almost 12 percent decrease in the probability
that female respondents were attending college at the time of the Wave III interview.
V. Propensity Score Matching
One potential concern with regard to the OLS estimates is that mig- raineurs could look quite different from nonmigraineurs in terms of observable char-
acteristics. If respondents from the two groups lack common support, then propensity score matching should produce more reliable estimates of the effect of migraine on
educational attainment than the standard regression approach Rosenbaum and Rubin 1983. Following previous researchers who have employed PSM as an alternative
to OLS or other procedures Levine and Painter 2003; da Veiga and Wilder 2008, we begin by estimating a standard probit model of the following form:
P
Migraine ⳱1⳱1ⳮUⳮ ⳮ ⬘X ⳮ ⬘Z , 2
i 1
i 2
i
where Z
i
is a set of additional controls designed to capture other factors that could be correlated with both migraines and education. Because there is evidence of an
association between depression and migraine headache Zwart et al. 2003, we in- clude two measures of psychological well-being in the vector Z
i
: the Center for Epidemiological Studies Depression CES-D score and the Rosenberg Self-Esteem
RSE score. In addition, we include indicators for whether the respondent suffered from hypertension and whether the respondent reported being unable to get enough
324 The
Journal of
Human
Resources
Table 2 Estimates of the Relationship between Migraine Headache and Educational Attainment
Ordinary Least Squares Propensity Score Matching
Family Fixed Effects 1
2 3
Grade point average Combined sample
ⳮ 0.088
0.029 n⳱7,726 ⳮ
0.071 0.039 n⳱1,536
ⳮ 0.124
0.051 n⳱214 Males
ⳮ 0.105
0.039 n⳱3,734 ⳮ
0.078 0.090 n⳱554
ⳮ 0.255
0.113 n⳱63 Females
ⳮ 0.079
0.035 n⳱3,734 ⳮ
0.002 0.058 n⳱908
ⳮ 0.079
0.113 n⳱86 High school graduation
Combined sample ⳮ
0.039 0.014 n⳱9,591
ⳮ 0.012
0.017 n⳱1,897 ⳮ
0.042 0.041 n⳱280
Males ⳮ
0.021 0.019 n⳱4,660
ⳮ 0.011
0.029 n⳱724 ⳮ
0.167 0.059 n⳱73
Females ⳮ
0.051 0.015 n⳱4,931
ⳮ 0.025
0.023 n⳱1,113 0.014
0.084 n⳱104 College attendance
Combined sample ⳮ
0.069 0.014 n⳱9,593
ⳮ 0.045
0.023 n⳱1,901 ⳮ
0.079 0.043 n⳱280
Males ⳮ
0.060 0.019 n⳱4,658
ⳮ 0.077
0.044 n⳱728 ⳮ
0.084 0.067 n⳱73
Females ⳮ
0.076 0.019 n⳱4,935
ⳮ 0.024
0.033 n⳱1,104 ⳮ
0.167 0.069 n⳱104
Note: Estimates are based on unweighted data from Waves I and III of The National Longitudinal Study of Adolescent Health. Ordinary least squares OLS estimations controlled for parental education, parental marital status, birth order, household income, family size, BMI, height, age, race, gender, and cognitive ability. Standard errors
of the OLS estimates are corrected for clustering at the school level and are in parentheses. Propensity score matching PSM estimations included additional controls for depression, self-esteem, frequency of drunkenness, hypertension, sleep deprivation, school effects, and grade effects. These additional controls are listed in Appendix
Table A1. Bootstrapped standard errors of the PSM estimates are in parentheses. Family fixed effects FFE estimations controlled for BMI, height, age, whether the respondent had an older sibling, gender, and cognitive ability. Standard errors of the FFE estimates are in parentheses.
statistically significant at the 10 percent level; sat the 5 percent level; at the 1 percent level
Rees and Sabia 325
sleep, both of which have been found to be associated with migraine headache Kelman and Rains 2005; Scher et al. 2005.
5
Finally, we include indicators of drunkenness to capture any “hangover effects” that might have been misdiagnosed
by parents as migraine headache, as well as school and grade effects to control for community and school characteristics that could potentially be correlated with mi-
graine triggers such as noise, weather, or even florescent lighting.
6
After estimating Equation 2, we used nearest neighbor matching without replace- ment to assign migraineurs with a nonmigraineur whose estimated propensity score
was within 0.0015. In a further effort to ensure common support, we dropped mig- raineurs whose estimated propensity score was higher than the maximum or less
than the minimum estimated propensity score of nonmigraineurs, and then dropped the 10 percent of migraineurs whose propensity score was furthest from the propen-
sity score of their match. This procedure produced a matched sample that appears to be well-balanced across observables.
7
The second column of Table 2 presents the PSM estimates. When the male and females samples are combined, the PSM estimates are smaller in absolute magnitude
than the OLS estimates, but nonetheless provide evidence of a negative relationship between migraine headache and educational attainment. Specifically, migraine head-
ache is associated with a 0.071 decrease in high school GPA and a 0.045 decrease in the probability of college attendance. The estimated effect of migraine headache
on high school graduation is not statistically significant at conventional levels. When the sample is divided by gender, the PSM estimates become less precise, but suggest
that the effect of migraine headache on male educational attainment may be larger than its effect on female educational attainment. For instance, migraine headache is
associated with a 0.077 decrease in the probability of attending college for males, but a statistically insignificant 0.024 decrease for females.
VI. Sibling Comparisons
