R

OU

a

ARRAA

JIIJ

KCHRAB

1

psdatHaaafnbc

iwhg

0d

Journal of Health Economics 30 (2011) 328–339

Contents lists available at ScienceDirect

Journal of Health Economics

journa l homepage: www.e lsev ier .com/ locate /econbase

acial disparities in the cognition-health relationship

wen Thompson ∗

niversity of Massachusetts, Amherst, United States

r t i c l e i n f o

rticle history:eceived 23 July 2010eceived in revised form 1 January 2011ccepted 26 January 2011vailable online 2 February 2011

EL classification:1218

a b s t r a c t

This paper investigates how the association between cognitive achievement and self-rated health inmiddle age differs by race, and attempts to explain these differences. The role of cognition in healthdetermination has received only limited empirical attention, and even less is known about how racemay affect this relationship. Using data from the NLSY, I find that while Whites with higher cognitiveachievement scores tend to report substantially better general health, this relationship is far weaker orwholly absent among Blacks. Further tests suggest that about 35% of this racial difference can be explainedby behavioral decisions during adulthood, and that another portion of the disparity may trace back toprenatal and early childhood experiences. The paper closes by noting that its results are broadly consistent

15

eywords:ognitionealthaceFQT

with explanations of the racial health gap that emphasize entrenched forms of racial discrimination.© 2011 Elsevier B.V. All rights reserved.

ieaco

icpail

irth weight

. Introduction

While scores on cognitive achievement tests have played arominent role in economic studies of racial wage gaps, mea-ures of cognition have received scant attention in the otherwiseeep literature on the socioeconomic determinants of health. Onlyhandful of published studies contain any empirical analysis of

he relationship between cognition and health (Grossman, 1975;artog and Oosterbeek, 1998; Gottfredson and Deary, 2004; Auldnd Sidhu, 2005; Heckman, 2007; Cutler and Lleras-Muney, 2010)nd none of these address potential racial differences in thisssociation.1 The paucity of research on this topic is surprising

or multiple reasons. First, the incorporation of controls for cog-itive achievement adds an important and novel characteristic toe analyzed alongside more traditionally used measures of socioe-onomic status such as education and earnings. Perhaps more

∗ Corresponding author. Tel.: +1 612 723 2263.E-mail address: owen@econs.umass.edu

1 In a working paper, Cutler and Lleras-Muney (2008) address racial differencesn the education-health gradient, and find that association to be stronger among

hites than among blacks. While their question is distinct from the one addressedere, their findings broadly in line with this study’s results on the cognition-healthradient.

batipeafi

bC

167-6296/$ – see front matter © 2011 Elsevier B.V. All rights reserved.oi:10.1016/j.jhealeco.2011.01.008

mportantly, cognitive achievement is strongly correlated witharly childhood experiences, and exploring its relationship withdult health may offer otherwise elusive information regardingritical pathways between socioeconomic background and healthutcomes.2

The present paper seeks to begin filling this gap by estimat-ng the determinants of adult health, among them a measure ofognitive achievement, for both Black and White samples. Therimary finding is that while the relationship between cognitionnd adult health is quite strong among Whites, this relationships much weaker or even absent entirely within the Black popu-ation. The paper will proceed in four sections. The first sectionriefly describes the data, paying particular attention to the avail-ble measures of health and cognition. The second section presentshe primary results and discusses their robustness and reliabil-ty. The third section explores two possible explanations for therimary findings, and presents additional empirical evidence on

ach explanation’s validity. The fourth section considers possiblelternative explanations and interpretations of the study’s overallndings and concludes.

2 There has recently been strong and increasing interest in the relationshipetween childhood conditions and adult health. See for example Case et al. (2002),ase et al. (2005), and Frijters et al. (2010).

lth Ec

2

tsratain

aor.chmtpaao

a1(ssosAi

ebcaac(apuir

Brmadmmfisi

rtttoa

asfhbGdavu

ticsfbribc

cvicN

3

arecasafwsW

O. Thompson / Journal of Hea

. Data

The data used in this study is drawn from the National Longi-udinal Survey of Youth (NLSY). The NLSY began in 1979 with aample of 12,686 individuals between the ages of 14 and 22, andespondents were eligible to be interviewed annually until 1994nd biennially thereafter, with the most recent wave available athe time of writing occurring in 2006. Retention rates were gener-lly high, and by the time respondents began giving detailed healthnformation in 1998 (see below), the response rate among eligibleon-deceased participants was 86.7%.

Beginning in the 1998 wave of the NLSY, respondents weresked to complete a “40 and older health module”, which amongther items included a self-rating of general health. Specifically,espondents were asked “in general, would you say your health is. .” then given 5 choices ranging from poor to excellent. I haveoded their responses so that higher numbers equate to betterealth. Members of the youngest NLSY cohort completed the healthodule during the 2006 wave of the survey, creating for the first

ime a set of detailed health variables for the entire active sam-le. Since self-rated health is the most general measure of healthvailable and not subject to biases relating to health care accessnd diagnoses, it is the dependant variable used throughout mostf this study.3

The NLSY contains a reliable measure of cognitive achievements well. In 1980, 94% of the original 1979 sample completed all0 sections of the Armed Services Vocational Aptitude BatteryASVAB) exam. The primary proxy of cognitive ability used in thistudy is the sum of the respondent’s score on the arithmetic rea-oning, word knowledge and paragraph comprehension sectionsf the ASVAB plus one half their score on the numerical operationsection of the test. This measure is commonly referred to as thermed Forces Qualification Test (AFQT), and throughout this study

t is measured in standard units, i.e. z-scores.Since NLSY respondents were of different ages and had differ-

nt levels of education at the time of testing, both of which haveeen shown to effect performance on the AFQT, unadjusted z-scoresannot credibly be inserted directly into a regression equation. Toddress this issue I follow the adjustment strategy utilized by Auldnd Sidhu (2005). Their procedure first uses an estimate of theausal effect of education on AFQT score reported by Hansen et al.2003), who estimated that each additional year of schooling causesn increase in AFQT score of .17 standard deviations, to subtract the

ortion of AFQT score due to education from each observation’snadjusted AFQT score. Then, the education-adjusted AFQT score

s regressed on age at the time of testing, and the residuals of thisegression are used as the adjusted AFQT measure.4

3 While self-rated health may be subject to considerable measurement error (seeaker et al., 2004) the harmful effects of that error are limited in this case since self-ated health is being used as a dependant variable. In any case, self-rated healtheasures may be generally preferable to specific disease measures, which typically

sk the respondent whether they have been told by a doctor that they have a givenisease, introducing a strong diagnostic bias. Even allowing for the possibility ofeasurement error, self-rated health has been shown to be a strong independentarker of future health issues. Idler and Benyamini (1997) review 27 studies which

nd a positive relationship between self-rated health and subsequent mortality, anduggest that self rated health may be such a consistent predictor of mortality becauset incorporates otherwise unobservable but critical aspects of health determination.

4 Alternative adjustment strategies do exist. For example, one could simplyegress the unadjusted AFQT score on age and education at time of testing, andhen use the residuals of this regression as the adjusted AFQT measure. Alterna-ively, unadjusted AFQT score could be inserted directly into regression equationshat also include age and education at the time of testing as covariates. The resultsf this study’s preferred specification using these alternative adjustment strategiesre reported in Table A1 of Appendix A, are similar to those in Section 2 below.

aboh

Tmhefdto

i2o

1tr

onomics 30 (2011) 328–339 329

The NLSY also contains a rich and detailed set of demographicnd socioeconomic variables that are utilized in this study, andome of these warrant a brief explanation. First, beginning with theoundational work of Grossman (1972), most health economistsave conceptualized of health as a stock variable which reflectsoth current and previous values of important health determinants.iven this, in place of current income I use the mean of each respon-ent’s annual household income between the ages of 22 and 38, andlso include the standard deviation of income to account for incomeariability and insecurity.5 Income data was inflated to 2006 dollarssing the CPI-U-RS and is expressed in $10,000 increments.

Another important issue is controlling for childhood conditionshat may impact adult health. Unfortunately, since NLSY partic-pants were first interviewed at ages 14–22, direct measures ofhildhood health and other important childhood variables are onlyparsely available. To proxy for relevant genetic and childhoodactors, I use variables giving the highest grade completed byoth of the respondent’s parents, the number of siblings in theespondent’s household at age 14, and dummy variables indicat-ng whether either or both of the respondent’s parents had diedefore the age of 60. In the results tables below, these variables areollectively referred to as “childhood controls.”

Additionally, I make use of variables indicating highest gradeompleted, insurance coverage and marital status. For all of theseariables I use their values at the time of health reporting, whichs approximately 40 for each respondent. Finally, the analysis wasonducted using sampling probability weights generated by theLSY’s online custom weighting program.

. Results

One difficulty in analyzing the relationship between cognitivechievement and health is that cognition is known to be highly cor-elated with other important controls commonly included in healthquations, particularly education and income. To help develop alear understanding of the nature of the correlation between healthnd cognition, I begin by presenting simple bivariate regressions ofelf-rated health on AFQT score for White and Black samples. Thesere reported in columns 1 and 2 of Table 1, and have two importanteatures. The first is that the relationship is strong and significantithin both racial groups, while the second is that the relation-

hip is substantially stronger for Whites than for Blacks. Amonghites a standard deviation increase in adjusted AFQT score is

ssociated with an improvement in self-rated health of .273 points,ut among Blacks the comparable figure is just .168, a differencef .105 points.6 The reported p-value shows that this difference isighly significant.

Next, I estimate specifications with progressively more controls.he model estimated in columns 3 and 4 of Table 1 adds controls forarital status, health insurance coverage, gender, and the “child-

ood controls” noted above. Again, since measures of income andducation are likely strongly related to AFQT, they remain excluded

rom this second specification. The results show that the addition ofemographic and childhood controls does not substantively reducehe racial gap in AFQT coefficients. The new White AFQT coefficientf .170 is twice as large as the new Black coefficient of .0805, so

5 Since the oldest NLSY respondents were 22 during the first wave of the surveyn 1979 and the first respondents to report self-rated health did so at age 39, ages2–38 is the largest age range for which all active respondents have valid incomebservations.6 The standard deviation of self-rated health over the whole sample is equal to

.0093. Thus while I refer to the effect of independent variables in terms of “points”,hese are approximately equal to standard deviation changes in the dependant self-ated health variable.

330O

.Thompson

/JournalofHealth

Economics

30 (2011) 328–339

Table 1Determinants of self-rated health by race.

Bivariate Additional controlsadded

Education andincome added

Health behaviorsadded

White Black White Black White Black White Black

AFQT z-score .273*** .168*** .170*** (.0257) .0805** (.0367) .0776*** (.0272) −.0123 (.0404) .0732** (.0290) .0144 (.0434)Highest grade completed (.0220) (.0289) .0576*** (.00813) .0637*** (.0143) .0418*** (.00862) .0605*** (.0159)Average adult income .0380*** (.00816) .0513*** (.0147) .0397*** (.00841) .0322** (.0158)Standard deviation of average adult

income−.0110*** (.00428) −.0186** (.00814) −.0110** (.00459) −.0143* (.00867)

Currently married .141*** (.0486) .150** (.0644) .0787* (.0472) .0823 (.0647) .0771 (.0489) .0569 (.0722)Never married .0337 (.0681) −.0292 (.0692) .0251 (.0670) −.0228 (.0681) .103 (.0706) −.0207 (.0750)Insured .224*** (.0552) .0355 (.0655) .136** (.0561) −.0473 (.0659) .119** (.0592) −.0321 (.0752)Female −.0204 (.0327) −.294*** (.0513) −.0230 (.0322) −.297*** (.0517) −.0563* (.0340) −.303*** (.0577)Ever been a smoker −.175*** (.0363) −.0614 (.0594)Heavy drinker −.0372 (.0659) −.136 (.115)Obese −.424*** (.0391) −.272*** (.0576)Childhood controls N N Y Y Y Y Y YConstant 5.275*** (.673) 4.208*** (.877) 4.703*** (.712) 4.667*** (1.065) 4.294*** (.698) 3.894*** (1.033) 3.930*** (.838) 5.309*** (1.270)Observations 3940 2450 3451 1585 3437 1576 2896 1287R-squared .045 .016 .064 .049 .096 .074 .145 .093Point difference in magnitude of White

and Black AFQT coefficients.1050 .0895 .0899 .0588

p-Value from Chow test of thehypothesis that White and BlackAFQT coefficients are equal

.0031 .0124 .0184 .1459

Notes: Robust standard errors are in parenthesis. Omitted marriage category is divorced. “Childhood controls” include variables indicating the highest grade completed by both of the respondent’s parents, the number ofsiblings in the respondent’s household at age 14, and dummy variables indicating whether either or both of the respondent’s parents had died before the age of 60. Heavy drinking is defined as having drank 10 or more daysin the past month and averaging 3 or more drinks each time. All regressions control for age when health was reported, which is usually but not always 40.

* Significance at the 10% level.** Significance at the 5% level.

*** Significance at the 1% level.

O. Thompson / Journal of Health Ec3.5

3.7

3.9

Pre

dic

ted

He

alth

-2 -1 0 1 2

AFQT z-score

White Black

other variables held constant at overall population means

Fu

tp

fattthswpdna.vim

tgTppe1his

rrAohcoapoAagii

uirwrarh

4

bttaTcetar

4

babtorewaKvaa

smaoOsd

ihhtfication: whether the respondent has ever been a smoker, whetherthey reported heavy drinking in the last month,7 and whether theyhad a body mass index (BMI) over 30, which is a standard definitionof obesity.8 The smoking and drinking variables are direct measures

7 Heavy drinking is defined as having drank 10 or more days in the past month andaveraging 3 or more drinks per drinking occurrence. Reasonable changes in eitherpart of this definition do not significantly change the results.

ig. 1. AFQT-health gradient by race. Other variables held constant at overall pop-lation means.

hat the difference between the two decreases only slightly to .0895oints and remains highly significant.

The fifth and sixth columns of Table 1 present results from aully saturated model that includes educational attainment as wells average annual adult income and its standard deviation. Sincehe inclusion of these controls further purges possible omitted fac-ors that could influence both health and AFQT outcomes, this ishe preferred specification. The racial differences in the cognition-ealth association are now even more striking than in the previouspecifications. While the coefficient on AFQT score is reduced some-hat within the White population and is now .0776, in the Blackopulation the AFQT coefficient is not substantively or statisticallyifferent form zero, with a point estimate that is actually slightlyegative at −.0123. The .0899 point difference between the Blacknd White coefficients is statistically significant, with a p-value of

0184. Interestingly, the coefficients on the education and incomeariables are somewhat higher for Blacks than for Whites, suggest-ng that whatever is causing the differences in the AFQT coefficients

ay be unique to cognitive achievement.Fig. 1 provides a graphical representation of the results from

he preferred specification by using the regression coefficients toenerate predicted AFQT-health gradients for Blacks and Whites.he values of other variables are held constant at their overallopulation means, and to avoid over-extrapolating into sparselyopulated AFQT values, the gradients are graphed over a domainqual to the mean AFQT score of each racial group plus and minus.5 standard deviations. While the gradient for Whites (solid line)as the expected upward slope, the Black gradient (dashed line)

s flat, which reflects the lack of an association between AFQT andelf-rated health among Blacks.

A number of specification checks were conducted to assess theobustness of the results in Table 1, and are described here theneported in Appendix A. First, one may suspect that even thoughFQT has no direct impact on the health of Blacks, it may still bef importance as a mediating variable. For example Blacks withigh AFQT scores may be better able to convert increases in edu-ation or income into health. But interactions of AFQT and thesether variables within the Black sample return uniformly smallnd insignificant results. Alternatively, it may be that there is aositive relationship between AFQT and health among Blacks, butnly at high levels of AFQT. However, when quadratic and cubicFQT terms are added to the specification, they are individually

nd jointly insignificant within the Black sample, indicating a flatradient throughout the AFQT distribution. Also, AFQT remainsnsignificant within the Black population even when the samples restricted to individuals with AFQT scores above the overall pop-

r6co

onomics 30 (2011) 328–339 331

lation mean. Since self-rated health may vary in a proportionatenstead of an absolute manner, the preferred specification was alsoe-estimated using the log of self-rated health, and the key resultsere similar to those in the linear model. Finally, the main result is

obust to estimating the equations as ordered probits and to usingn un-weighted version of the sample. The consistency of theseesults suggests it is unlikely that the racial difference in AFQT’sealth impact is an artifact of specification problems.

. Explaining the non-effect of AFQT score among Blacks

Two possible explanations of the previous section’s results wille analyzed here. The first is that health behaviors are an impor-ant intermediary between cognition and self-rated health, andhat the observed racial differences in the AFQT-health gradientre the result of differences in the relevant behavioral relationships.he second is that the racial difference in the association betweenognition and adult health is rooted in childhood and prenatalxperiences, and in particular may derive from a weaker connec-ion between cognition and health endowments among Blacks thanmong Whites. Additional evidence on these two possibilities iseported below.

.1. The role of health behaviors

Perhaps the most intuitive reason why higher AFQT score woulde associated with improved health is that AFQT may be positivelyssociated with knowledge of health determinants and in turn withetter health behaviors. The basic argument is that those with bet-er cognitive skills are more capable of acquiring and processingften complex information about the precise behaviors which willesult in good health, so that they are able to produce health morefficiently and optimally maintain higher levels of health. This ideaas first put forth in the economics literature by Grossman (1972)

nd some empirical evidence in support of it has been found byenkel (1991) and by Gottfredson and Deary (2004). If the rele-ant behavioral relationships were stronger among Whites thanmong Blacks, it would help explain why the overall AFQT-healthssociation is stronger as well.

Since the measure of general health used in this study iself-rated health, there are actually two distinct behavior relatedechanisms through which the observed racial differences could

rise. On the one hand, there may be racial differences in thebserved behaviors of individuals with varying cognitive abilities.n the other hand, and somewhat less intuitively, the relation-

hip between given health behaviors and self-rated health may beifferent for Blacks than for Whites.

As a first step in analyzing the importance of health behaviorsn generating this paper’s main result, I assess how controlling forealth behaviors effects the overall racial difference in the AFQT-ealth association. To do so, the final two columns of Table 1 addhree measures of health related behaviors to the preferred speci-

8 While adding these variables is useful for the current purpose of assessing theole played by racial differences in behavioral factors, the model in columns 5 andof Table 1 remains the preferred specification. This is because health behaviors

ould be viewed as legitimate health outcomes themselves, and thus conditioningn them may obscure the primary relationship of interest.

3 lth Ec

of

aewocoetiftAbi

hertudtaa

atwdcsibzbs

hIoseaeitrtc

srtsodst

g

afisWwmiAb

hfeatcbmc(ntSaciTsWfld

btoasiArtwgoh

4

wccfuaa

32 O. Thompson / Journal of Hea

f health behaviors, while the obesity variable is intended to proxyor behaviors related to diet and exercise.

The results suggest that differences in health related behaviorsre a major contributor to the racial gap in the AFQT-health gradi-nt. Whereas before the inclusion of behavioral variables this gapas .0899, afterwards it was reduced to .0588, an overall reduction

f .0311 points or 35%, and the difference is no longer statisti-ally significant. Methodologically, this approach is similar to thene used by Cutler and Lleras-Muney (2010) to decompose theducation-health behaviors gradient, and a reasonable interpreta-ion of the reduced gap is that 35% of the overall racial differencen the association between AFQT and self-rated health is due to dif-erences in behavioral factors. This overall reduction could be dueo either a larger response of adverse health behaviors to increasedFQT among Whites, or to racial differences in how a given set ofehaviors affects self-rated health. I address these two possibilities

n turn.Table 2 presents the results from regressions relating observable

ealth behaviors to AFQT by race. For each health behavior variable,stimates from both bivariate and more fully specified models areeported for White and Black subsamples. To ease the interpre-ation of the coefficients, all models are estimated with OLS, butsing other estimation strategies such as logit or probit modelsoes not change the basic nature of the results. The covariates inhe fully specified models are highest grade completed, averagennual adult household income and its standard deviation, gendernd childhood controls identical to those used in Table 1.9

With respect to smoking, the bivariate results indicate thatmong Whites a standard deviation increase in AFQT score reduceshe probability of ever having smoked by approximately 11.3%,hile the comparable figure among Blacks is only 7.4%, and thisifference is statistically significant. After adding the additionalovariates, the White coefficient falls substantially but remainstatistically significant and indicates that a standard deviationncrease in AFQT reduces the probability of ever having smokedy approximately 2.7%, while the Black coefficient falls almost toero and is no longer statistically significant. Also, the differenceetween the Black and White coefficients is no longer statisticallyignificant.

The effects of AFQT score in the models with an indicator ofeavy drinking as the dependant variable yield less clear results.

n the bivariate models, higher AFQT score reduces the likelihoodf heavy drinking among both racial groups, but these effects aremall and statistically insignificant. If anything, the bivariate pointstimates suggest that AFQT score reduces heavy drinking moremong Blacks than among Whites. In the more fully specified mod-ls, a one standard deviation increase in AFQT score unexpectedlyncreases the probability of heavy drinking among Whites by a sta-istically significant margin of about 2%, while the effect for Blacksemains small and insignificantly negative. The difference betweenhe White and Black AFQT coefficients in these models is statisti-ally significant.

Turning to the obesity equations, the bivariate specificationshow that increased AFQT score is associated with a substantiallyeduced probability of obesity within both racial groups, but thathe effect is considerably stronger for Whites. Among Whites, atandard deviation increase in AFQT score reduces the probability

f obesity by approximately 4.3%, while among Blacks a standardeviation increase in AFQT score reduces the probability of obe-ity by only approximately 2.9%, although the difference betweenhe White and Black coefficients is not statistically significant. In

9 The marriage and insurance variables are left out of these models due to endo-eneity concerns.

etecCmbC

onomics 30 (2011) 328–339

ddition to lacking statistical significance in the bivariate speci-cation, racial differences favoring Whites disappear in the moreaturated model. After conditioning on the additional controls, the

hite AFQT-obesity relationship is effectively reduced to zero,hile the relationship for Blacks actually increases somewhat inagnitude, and now indicates that a standard deviation increase

n AFQT reduces the probability of obesity by approximately 3.7%.lso, the difference between the White and Black AFQT coefficientsecomes significant at the 5% level.

Overall, the results from Table 2 suggest that AFQT does notave a more favorable impact on health behaviors for Whites than

or Blacks. Given this, I now turn to investigating racial differ-nces in the relationship between a given set of health behaviorsnd self-rated health. In particular, I analyze the coefficients onhe behavioral variables in the final two columns of Table 1. Theoefficients on the indicator of heavy drinking are insignificant foroth racial groups, although it is noteworthy that the point esti-ate in the Black sample of −.136 is considerably larger than its

ounterpart of −.0372 in the White sample. However, a Chow testnot shown) indicates that this difference is not statistically sig-ificant. For the smoking and obesity indicators, the coefficients inhe White sample are much larger than those for the Black sample.pecifically, the White and Black smoking coefficients are −.175nd −.0614, respectively and the White and Black obesity coeffi-ients are −.424 and −.272, respectively. Chow tests (not shown)ndicate that both of these differences are statistically significant.hese results demonstrate that when Blacks avoid smoking or obe-ity, they report moderately better general health, but that when

hites avoid these same unhealthy behavioral traits they reportar greater improvements in their general health. These differencesikely account for a substantial portion of the overall Black-Whiteifference in the AFQT-health gradient.

Taking into account all of the evidence presented above, it maye most accurate to say that racial differences in behavioral fac-ors are likely to account for some, but certainly not all, of theverall racial difference in the AFQT-health gradient. Addition-lly, it appears that racial differences in the relationship betweenelf-rated general health and given health behaviors are a moremportant factor than racial differences in the relationship betweenFQT score and specific health behaviors. Taking into account theole of both mechanisms, behavioral factors can explain about one-hird of the overall racial gap in the AFQT-health gradient, whichhile certainly substantial, still leaves and a sizeable portion of the

ap to be explained by other causes. The next subsection considersne such additional explanation, the role of early experiences andealth endowments.

.2. The role of early experiences

While the behaviorally driven explanations explored aboveere causal in nature, an alternative explanation is that both high

ognition and good health are associated with a third factor notontrolled for in the regressions from Table 1. One possible thirdactor linking cognitive achievement and adult health is an individ-al’s health endowment, which for present purposes I will defines fixed characteristics present at the time of birth or shortly there-fter which impact health through the life cycle. It is now wellstablished that health endowments determine a substantial por-ion of adult health, as evidenced by the power of prenatal andarly childhood environmental factors to predict later health out-

omes (Wadsworth and Kuh, 1997; Barker, 1997; Case et al., 2002;ase et al., 2005). A large portion of cognitive ability is also deter-ined early in life, and again early childhood experiences have

een shown to play a particularly important role (Heckman, 2006;unha et al., 2010).

O.Thom

pson/JournalofH

ealthEconom

ics30 (2011) 328–339

333

Table 2Effect of AFQT on health behaviors by race.

Ever been a smoker Heavy drinking in previous month Obese

White Black White Black White Black White Black White Black White Black

AFQT z-score −.113***

(.0110)−.0738***

(.0142)−.0266*

(.0147)−.00278(.0204)

−.00202(.00624)

−.0111(.00736)

.0202**

(.00801)−.00630(.0118)

−.0427***

−(.0112).0292**

(.0147).00504(.0144)

−.0374*

(.0147)Highest grade

completed−.0477***

(.00436)

*.0501***

(.00714)−.0147***

(.00246)

*.000461(.00382)

−.0172***

(.00414)−.00462(.00805)

Average adultincome

−.0275***

(.00417)

*.0314***

(.00769)−.000201(.00189)

1.00992**

(.00503)−.00324(.00340)

.0106(.00812)

Standard deviationof average adultincome

.0129***

(.00232).0105***

(.00390).000212(.00108)

.00573(.00357)

−.00205(.00181)

−.00621(.00397)

Female .0237(.0173)

−.114***

(.0271)−.0874***

(.0100).0793***

(.0144)−.0666***

(.0167).0900***

(.0286)Childhood controls N N Y Y N N Y Y N N Y YConstant .567***

(.00945).428***

(.0142)1.154***

(.0677)1.148***

(.113).0782***

(.00551).0553***

(.00700).270***

(.0371).0494(.0585)

.299***

(.00966).400***

(.0147).694***

(.0648).481***

(.125)Observations 3960 2385 3282 1469 3556 2247 3114 1450 3438 2155 3019 1385R-squared .027 .013 .097 .096 .000 .001 .042 .042 .005 .002 .032 .017Point difference in

magnitude ofWhite and BlackAFQT coefficients

.0392 .0238 −.0091 .0265 .0135 .0424

p-Value from Chowtest of thehypothesis thatWhite and BlackAFQT coefficientsare equal

.0155 .2402 .3165 .0185 .4127 .0360

Notes: Robust standard errors are in parenthesis. “Childhood controls” include variables indicating the highest grade completed by both of the respondent’s parents, the number of siblings in the respondent’s household atindicating whether either or both of the respondent’s parents had died before the age of 60. Heavy drinking is defined as having drank 10 or more days in the past month and averaging 3 or more drinks each time. To easeinterpretation, all models are estimated via OLS. Estimation using logits or probits does not substantively change the results.

* Significance at the 10% level.** Significance at the 5% level.

*** Significance at the 1% level.

3 lth Ec

vihpwsaslgro

qtcssbasRr

uw2gbondowc

rgweawhjmbmni

2awdbbct

aanmt

steobac(rql1b

mnfvhTiwlswacpc

m2rdraWBecbio

dbatotionship between maternal cognition and maternal behaviors dur-ing pregnancy. To test whether this is the case, the third and fourthcolumns of Table 3 add controls for four important maternal behav-iors during pregnancy. These measures are indicators of whether

10 The birth weight variable, as well as the maternal behavior variables in sub-sequent models, are measured in reference to the first live birth of female NLSYparticipants who reported at least one live birth. The rationale for restricting theanalysis to first live births is that higher order births may vary in systematic waysfrom first births, and the only order of birth that is universal to all women who have

34 O. Thompson / Journal of Hea

The prenatal and early childhood experiences of a given indi-idual depend mostly on the decisions or circumstances of thatndividual’s parents, and the parental practices which lead to highealth endowments seem likely to overlap considerably with thearental practices that encourage high AFQT scores. Given this,e might expect to observe a positive association between AFQT

core and health endowment that is driven by their shared associ-tion with a particular set of parental practices. The existence andtrength of such an AFQT-health endowment association early inife could help to explain the existence and strength of the moreeneral AFQT-health association later in life as well. Of course, anyacial differences in the AFQT-health endowment association aref particular interest in the present paper.

The empirical strategy pursued in this subsection is to firstuantify any racial differences in the association between cogni-ive achievement scores and health endowments, and then to addontrols for relevant parental behaviors and observe whether anyuch racial differences are attenuated. In order to implement thistrategy, two important issues must be resolved. First, a credi-le measurement of health endowment must be obtained. Second,mong the same individuals for whom health endowment is mea-ured, an indicator of cognitive achievement is needed as well.esolving these issues required using data on the children of NLSYespondents, and I explain my approach in detail below.

As my basic measure of health endowment I use an individ-al’s birth weight, which has been shown to be strongly associatedith of a number of adult health outcomes (Godfrey and Barker,

001; Behrman and Rosenzweig, 2004; Black et al., 2007). There isood reason to believe these associations are at least in part causal,ecause adverse prenatal environments can exert a direct impactn organ development. This is due to the fact that when a fetus doesot receive adequate nutrition for normal rates of cell division andevelopment, it tends to focus on protecting normal brain devel-pment even when it comes at the expense of organ development,hich in addition to reducing birth weight may ‘pre-program’ spe-

ific organs to fail later in life (Barker, 1997; Smith, 1999).Beyond being a consistent marker of later health outcomes, it is

easonable to believe that parental behaviors, as opposed to solelyenetic or stochastic factors, have an important impact on birtheight. The relationship between birth weight and intrauterine

nvironment is quite strong and intrauterine environment gener-lly plays a much larger role than genetics in determining birtheight, as evidenced by data showing that the birth weights ofalf siblings who share a father have a correlation coefficient of

ust .1, whereas the comparable statistic for those who share aother is .58 (Barker, 1997). While some of this relationship may

e driven by non-genetic physiological traits such as the size of theother’s uterus, behavioral factors like smoking cessation, mater-

al nutrition, and prenatal medical care also seem likely to play anmportant role.

NLSY participants were first interviewed between ages 14 and2, and their birth weights were never collected. Also unavailablere proxies of the prenatal environment that NLSY respondentsere subject to prior to birth, such as the behaviors of their mothersuring pregnancy. However, as the NLSY participants themselvesecame parents, the survey did collect detailed data on their ownehaviors during pregnancy and recorded the birth weights of theirhildren. It is this information on the children of NLSY respondentshat I use in my analysis.

But in order to measure the association between cognition

nd health endowment among the children of NLSY respondents,measure of the children’s cognitive achievement is of course

eeded as well. One available proxy for the cognitive achieve-ent of the children of NLSY respondents is the AFQT score of

heir mothers. But should we expect parental AFQT to do a rea-

h

raot

onomics 30 (2011) 328–339

onably good job of predicting the cognitive achievement levels ofhe next generation? There is a large literature on the intergen-rational transmission of cognitive achievement which spans allf the social sciences as well as research by neuroscientists andiologists, and it strongly suggests that parental AFQT scores arereasonable proxy for children’s AFQT scores. Despite extensive

ontroversy over the exact portion of IQ or general intelligence ‘g’both of which strongly correlate to AFQT score) which is envi-onmentally as opposed to genetically determined, there is littleuestion that performance on achievement tests is strongly corre-

ated between generations. Flynn (2000) averages estimates from41 studies and reports a mean estimated correlation coefficientetween average parent IQ and child IQ of .71.

Given this, the first two columns of Table 3 regress birth weighteasured in grams on maternal AFQT score and controls for mater-

al income, educational attainment and age at the time of birthor samples of White and Black mothers.10,11 The results pro-ide some evidence that the relationship between cognition andealth endowment varies by race. The results in the first column ofable 3 indicates that among Whites, a standard deviation increasen maternal AFQT score is associated with an increase in birth

eight of 56 g, and that this relationship is significant at the 5%evel. In contrast, the second column shows that among Blacks atandard deviation increase in maternal AFQT score is associatedith an increase in child birth weight of only 31 g, 25 g less than

mong Whites, and that this relationship fails to approach statisti-al significance at conventional levels. Despite this discrepancy inoint estimates, the difference between the White and Black AFQToefficients is not statistically significant.

To put these estimates in context, the mean population-wideean birth weight in the present sample is 3256 g, so that the

5-g disparity between the White and Black AFQT coefficientsepresents .77% of average birth weight. The overall White-Blackifference in birth weight may be a more appropriate point ofeference, given that this gap is a major area of policy concernnd research (Kleinman and Kessel, 1987; Nepomnyaschy, 2010).ithin the present data, the mean birth weights among White and

lack mothers are 3296 g and 3068 g, respectively. This is a differ-nce of 228 g, so the 25-g discrepancy in the White and Black AFQToefficients is equivalent to about 11% of the overall racial gap inirth weights. While modest, these magnitudes are not trivial, and

t is certainly plausible that they could have a discernable impactn the strength of the cognition-health relationship in adulthood.

Of course, the models estimated in columns 1 and 2 of Table 3id not control for maternal behaviors during pregnancy, and suchehaviors may be an important intermediary factor leading to anssociation between cognition and health endowment. Given this,he racial differences in the strength of the overall relationshipbserved thus far may simply be due to racial differences in the rela-

ad a live birth is the first.11 It may be the case that eventual educational attainment and income are betterepresentations of socioeconomic status than educational attainment and incomet the time of first birth. Using these alternative measures preserves the basic naturef the results reported in Table 3, and in fact non-trivially widens the racial gap inhe impact of AFQT.

O. Thompson / Journal of Health Economics 30 (2011) 328–339 335

Table 3Effect of maternal AFQT on child birth weight by race.

White Black White Black

AFQT z-score 56.00** (26.99) 31.41 (29.68) 48.76* (26.91) 31.88 (30.14)High grade at time of birth 12.05 (9.017) 22.96* (13.93) .867 (9.224) 20.45 (13.87)Average adult income at time of birth 1.355 (2.063) 6.878 (10.95) 1.273 (2.100) 8.593 (10.83)Age at time of birth −11.45*** (3.634) −12.08** (6.069) −11.21*** (3.578) −12.02** (5.980)Controls for behaviors during pregnancy N N Y YConstant 3402*** (95.84) 3080*** (141.0) 3532*** (137.8) 2824*** (187.1)Observations 1395 857 1392 853R-squared .012 .010 .032 .032Point difference in magnitude of White and Black AFQT

coefficients24.59 16.88

p-Value from Chow test of the hypothesis that White andBlack AFQT coefficients are equal

.5234 .6603

Notes: Controls for behaviors during pregnancy include indicators of whether each individual reduced smoking during pregnancy, reduced drinking during pregnancy,reduced salt consumption during pregnancy, and received routine prenatal medical care during pregnancy. Robust standard errors are in parenthesis.

* Significance at the 10% level.

tspdb

raeottiotiiW

fbitnrtWgwetbc

5

ralttvi

esthdddioenitt

rtcim

oSaBoGotcascore on wages is nearly identical for Blacks and Whites, which tothe extent that wages reflect productivity or general ability indi-cates an absence of racial test score bias.12 After a review of therelevant literature, Heckman (1995) concludes that “the best avail-

12 For example, Neal and Johnson (1996) find that for black and white men thecoefficients on standardized AFQT score in log wage equations are .208 and .183,

** Significance at the 5% level.*** Significance at the 1% level.

he mother reduced or eliminated drinking, smoking and salt con-umption during pregnancy, and whether she received regularrenatal medical care. These variables are all potentially importantirect determinants of child health endowment, and are likely toe associated with other important unobserved behaviors as well.

The results in columns 3 and 4 show that racial differences in theelationship between maternal AFQT and child birth weight remainfter controlling for maternal behaviors during pregnancy. Theffect of a standard deviation increase in AFQT on the birth weightf White children falls modestly to 49 g and remains significant athe 10% level, while the analogous estimate for Black children is vir-ually unchanged at 32 g and remains insignificant. The differencen AFQT coefficients is now 17 g, which is equivalent to about 7.5%f the overall racial gap in birth weight. These results suggest thathe stronger AFQT-health endowment relationship among Whitess not simply an artifact of racial differences in maternal behav-ors during pregnancy, although again the difference between the

hite and Black AFQT coefficients is not statistically significant.In addition to the noted lack of statistical significance in the dif-

erences between the Black and White AFQT coefficients, it shoulde pointed out that the coefficients on educational attainment and

ncome in the Black sample are considerably larger than those inhe White sample. This is especially true after controlling for mater-al behaviors. Since cognition, education and income are all highlyelated, the results leave open the possibility that despite cogni-ion playing a different role for the two racial groups, Black and

hite individuals with generally comparable socioeconomic back-rounds may still be reasonably expected to have similar birtheight outcomes as well. While these considerations make the

vidence in Table 3 less than conclusive, the results are still sugges-ive of a potentially important racial difference in the relationshipetween cognition and health endowments, and such a differenceould be a contributing factor to this paper’s main result.

. Discussion and conclusion

The most important finding of this paper has been that theelationship between cognitive achievement and self-rated healtht middle age is much stronger for Whites than for Blacks. Two

ines of explanation for this result have been proposed and, tohe extent possible, tested empirically. The first was that the rela-ionship between behavioral factors and cognitive achievementaries by race. Tests using smoking, heavy drinking and obesityndicated that behavioral factors play a substantial role in gen-

rrswit

rating the paper’s main finding, and that most of their effecteems to be due to smoking and obesity being less closely linkedo the self-rated health of Blacks than they are to the self-ratedealth of Whites. The second proposed explanation was that racialifferences in the adult AFQT-health gradient may derive fromifferences in that gradient early in life, and in particular mayerive from a weaker association between maternal cognitive abil-

ty and children’s health endowment among Blacks. Some evidencef a weaker relationship between maternal AFQT and child healthndowment for Blacks was found, and that weaker association didot appear to be an artifact of differences in the maternal behav-

ors of Black and White mothers with similar AFQT scores. However,hese racial differences were modest in scale and were not statis-ically significant.

In general though, a substantial portion of the core resultemains unexplained. Since the two most important variables ofhe analysis (AFQT score and self-rated health) may have signifi-ant subjective components that could vary across cultural groups,t is natural to hypothesize that racial bias in one or both of these

easures accounts for the remaining racial gap.The question of racial bias in the AFQT was addressed as part

f a large validation study conducted by the National Academy ofciences and the Department of Defense (Widgor and Green, 1991)nd overall there is little if any evidence that the AFQT scores oflacks and Whites systematically differ when reasonable measuresf underlying ability are held constant. For example, Widgor andreen (1991, p. 179) regress objective measures of job performancento AFQT score for individual workers within a variety of mili-ary occupations, and find that AFQT has nearly identical regressionoefficients and predictive power for the job performance of Blacksnd Whites. Multiple studies have also found that the effect of AFQT

espectively, while the coefficients for black and white women are .223 and .189,espectively. Similarly, Rogers and Spriggs (1996) used adjusted AFQT percentilecore in a log wage equation and found coefficients of .0053 and .0052 for black andhite men, and coefficients of .0079 and .0069 for black and white women, although

t should be noted that Rogers and Spriggs emphasize the fact that the constants inhe two equations typically differ by a statistically significant amount.

336 O. Thompson / Journal of Health Economics 30 (2011) 328–339

0.1

.2.3

.4

De

nsity

1 2 3 4 5

Health Rating

Black0

.1.2

.3.4

De

nsity

1 2 3 4 5

Health Rating

White

alth di

aA

nhatrfsrurcaas

iaFttatpafteB

ano(w

nc

hae

iid(bwaep1

faiddtwowt

whnBtebshda

Fig. 2. Self-rated he

ble evidence indicates that IQ and achievement tests [includingFQT] are not culturally biased.”

While racial bias in the AFQT does not appear to be a likely expla-ation for this paper’s main result, the reliability of the self-ratedealth variable for comparisons across racial groups remains anrea of potential concern. Perhaps it is the case that due to cul-ural differences in health perception, self-rated health varies alongacial lines even when true underlying health does not. One specificorm of health reporting bias that could erroneously produce thistudy’s main finding is a “ceiling effect” among Blacks, in which self-ated health measures rarely exceed a given level, irrespective ofnderlying health. While there is some previous evidence that self-ated health is a non-racially biased predictor of observable healthonditions and mortality,13 this issue has not received enormousttention and there is currently no consensus among researcherss to whether Blacks and Whites exhibit systematically differentelf-ratings of health.

One simple method of assessing racial health reporting biases,ncluding a possible ceiling effect, is to visually examine the Whitend Black distributions of self-rated health, which are presented inig. 2. The figure shows that while the White distribution is cer-ainly more favorable overall than the Black distribution, most ofhis racial difference occurs in the middle range of health ratings,s opposed to in the tails. In particular, while Blacks are less likelyo report the highest possible health rating of 5 (18.4% as com-ared to 23.5% for Whites) Fig. 2 displays no obvious ceiling effectmong Blacks.14 Furthermore, even if we assume that meaning-ul culturally generated reporting biases exist and that they reducehe overall level of self-rated health among Blacks, that would notxplain why the effect of a trait like AFQT score is negligible in thelack sample.

Although racial biases in either the AFQT or self-rated healthppear unlikely to account for this study’s main results, it is

oteworthy that the results are broadly consistent with a the-ry of racial health gaps known as “the weathering hypothesis”Geronimus et al., 2006). The basic underlying concept of theeathering hypothesis is allostatic load. Allostasis refers to the abil-

13 For example, Chandola and Jenkinson (2000) report finding no evidence of eth-ic group differences in the association between self-rated health and hypertension,ardiovascular disease or diabetes.14 The racial difference in the proportion of individuals giving the best possibleealth rating becomes smaller when the analysis is restricted to individuals in rel-tively good health. For example, among respondents who gave health ratings ofither 4 or 5, 36.2% of whites and 34.8% of blacks gave the highest possible rating.

Bbidbm

tarioa

stributions by race.

ty to achieve stability through change, and allostatic load is usedn a medical context to describe the cumulative health burden ofefensive physiological reactions to external threats and stressorsMcEwen, 1998). Examples of allostatic responses are when theody elevates epinephrine (adrenaline) levels or releases cortisolhen confronted with an external threat. While these reactions

re ideal for short-term survival, regular or prolonged periods oflevated allostatic demands may take a cumulative toll on bloodressure, heart rate, and immune system functioning (McEwen,998; Smith, 1999).

It has been shown that allostatic load levels are generally higheror Blacks than for Whites, even after controlling for educationnd income, and the weathering hypothesis holds that the remain-ng racial gap is due to the cumulative effects of subtle forms ofiscrimination and from relatively immutable features of day toay life for Blacks in the US (Geronimus et al., 2006). These fea-ures may include residential segregation, overexposure to air andater toxins, differential treatment within the health care system,

r at the most basic level simply the increased stress associatedith being part of a historically marginalized minority popula-

ion.If the institutionalized forms of discrimination posited by the

eathering hypothesis do play an important role in overall racialealth disparities, they may also help to explain the irresponsive-ess of health to increased cognitive achievement scores amonglacks documented here. As one illustrative example, it is unlikelyhat marginal increases in AFQT score would help Blacks avoidnvironmental hazards if those hazards are linked not to dailyehavioral decisions but rather to historically rooted residentialegregation. Similar reasoning can be applied to my findings onealth behaviors and health endowments. In both cases, the evi-ence suggested that the relevant behavioral responses of Blacksnd Whites to increased cognitive scores were similar, but that forlacks those behavioral changes were less strongly associated withetter self-rated health or higher health endowments. A weather-

ng based interpretation of these findings is that institutionalizediscriminatory factors may prevent Blacks from converting healthyehaviors into enhanced general health or higher health endow-ents in the next generation.While generally consistent with the evidence presented in

his paper, the weathering hypothesis clearly does not constitutecomplete explanation of the residual racial difference in the

elationship between cognitive achievement and health. A fully sat-

sfactory explanation of that residual will require further researchn the mechanisms through which cognition is related to healthnd on how those mechanisms may be interrupted.

O.Thom

pson/JournalofH

ealthEconom

ics30 (2011) 328–339

337

Table A1Specification and robustness checks.

Specification: Age and education at timeof test as covariates

AFQT “Cleansed” of age andeducation at time of test

Interaction terms Quadratic AFQT Cubic AFQT AFQT abovepopulation mean

White Black White Black Black Black Black Black

AFQT z-score .0943*** (.0275) −.00731 (.0414) .0377 (.0254) −.0182 (.0377) .191 (.212) −.00980 (.0487) .0319 (.0645) −.0357 (.133)AFQT z-score squared – – – – – .00251 (.0365) −.0542 (.0481) –AFQT z-score cubed – – – – – – −.0432 (.0361) –(AFQT

z-score) × (highestgrade completed)

– – – – −.0141 (.0165) – – –

(AFQTz-score) × (averageadult income)

– – – – −.00349 (.0110) – – –

Highest gradecompleted

.0391*** (.00880) .0569*** (.0166) .0632*** (.00789) .0633*** (.0137) .0593*** (.0155) .0636*** (.0143) .0639*** (.0143) .0452* (.0271)

Average adult income .0344*** (.00809) .0495*** (.0149) .0392*** (.00814) .0515*** (.0146) .0503*** (.0147) .0513*** (.0147) .0517*** (.0147) .0389* (.0234)Standard deviation of

average adult income−.00941** (.00423) −.0177** (.00820) −.0115*** (.00427) −.0186** (.00813) −.0185** (.00815) −.0186** (.00814) −.0187** (.00814) −.0132 (.0133)

Currently married .0760 (.0471) .0838 (.0649) .0791* (.0473) .0827 (.0647) .0810 (.0648) .0825 (.0649) .0809 (.0648) −.0194 (.130)Never married .0202 (.0668) −.0221 (.0682) .0229 (.0670) −.0226 (.0681) −.0217 (.0682) −.0228 (.0682) −.0253 (.0679) −.0638 (.141)Insured .126** (.0558) −.0497 (.0662) .141** (.0562) −.0474 (.0659) −.0526 (.0663) −.0474 (.0660) −.0550 (.0662) .299* (.169)Female −.0312 (.0322) −.299*** (.0519) −.0228 (.0322) −.297*** (.0518) −.299*** (.0517) −.297*** (.0519) −.298*** (.0519) −.387*** (.108)Age at time of AFQT −.0366*** (.0132) −.00535 (.0186) – – – – – –Highest grade

completed at time ofAFQT

.0609*** (.0172) .0197 (.0267) – – – – – –

Constant 4.538*** (.715) 3.937*** (1.070) 4.249*** (.699) 3.903*** (1.027) 3.977*** (1.043) 3.898*** (1.025) 4.031*** (.969) 5.454*** (1.986)Observations 3437 1576 3437 1576 1576 1576 1576 370R-squared .103 .075 .094 .074 .075 .074 .076 .093

Notes: Robust standard errors are in parenthesis. Omitted marriage category is divorced. All models include “Childhood controls” indicating the highest grade completed by both of the respondent’s parents, the number ofsiblings in the respondent’s household at age 14, and dummy variables indicating whether either or both of the respondent’s parents had died before the age of 60. All models also control for age when health was reported,which is usually but not always 40.

* Significance at the 10% level.** Significance at the 5% level.

*** Significance at the 1% level.

3 lth Economics 30 (2011) 328–339

A

owca

A

pitBscsia.

eretwifcTt.

flesputNtTWii

fisaiaawsTfiscss

ifica

tion

and

robu

stn

ess

chec

ks.

Log

ofse

lf-r

ated

hea

lth

Un

wei

ghte

dO

rder

edp

robi

ts

Wh

ite

Bla

ckW

hit

eB

lack

Wh

ite

Bla

ck

.024

5***

(.00

883)

−.00

0978

(.01

32)

.088

3***

(.02

67)

.005

72(.

0386

).0

845**

*(.

0317

)−.

0216

(.04

59)

com

ple

ted

.018

5***

(.00

261)

.020

3***

(.00

462)

.060

4***

(.00

777)

.063

8***

(.01

38)

.069

2***

(.00

997)

.073

6***

(.01

66)

inco

me

.012

6***

(.00

257)

.017

3***

(.00

452)

.037

8***

(.00

797)

.052

1***

(.01

42)

.046

4***

(.01

07)

.056

7***

(.01

71)

atio

nof

aver

age

adu

ltin

com

e.0

0416

***

(.00

132)

.006

64**

*(.

0024

3)−.

0107

***

(.00

403)

−.01

89**

(.00

805)

−.01

18**

(.00

544)

−.01

98**

(.00

952)

ried

.031

6*(.

0161

).0

342

(.02

11)

.072

6(.

0451

).0

707

(.06

24)

.076

9(.

0545

).0

910

(.07

28)

.017

7(.

0217

).0

0161

(.02

31)

.021

9(.

0640

)−.

0537

(.06

61)

.017

6(.

0777

)−.

0333

(.07

58)

.038

6**(.

0194

)−.

0165

(.02

20)

.103

*(.

0543

)−.

0462

(.06

43)

.151

**(.

0624

)−.

0475

(.07

38)

−.01

34(.

0103

)−.

0940

***

(.01

66)

−.03

77(.

0310

)−.

318**

*(.

0496

)−.

0203

(.03

85)

−.33

8***

(.05

93)

1.51

0***

(.23

9)1.

276**

*(.

317)

4.08

6***

(.67

9)3.

927**

*(.

932)

––

3437

1576

3437

1576

3437

1576

.091

.075

.099

.079

––

tan

dar

der

rors

are

inp

aren

thes

is.O

mit

ted

mar

riag

eca

tego

ryis

div

orce

d.A

llm

odel

sin

clu

de

“Ch

ild

hoo

dco

ntr

ols”

ind

icat

ing

the

hig

hes

tgr

ade

com

ple

ted

bybo

thof

the

resp

ond

ent’

sp

aren

ts,t

he

nu

mbe

rof

esp

ond

ent’

sh

ouse

hol

dat

age

14,a

nd

du

mm

yva

riab

les

ind

icat

ing

wh

eth

erei

ther

orbo

thof

the

resp

ond

ent’

sp

aren

tsh

add

ied

befo

reth

eag

eof

60.A

llm

odel

sal

soco

ntr

olfo

rag

ew

hen

hea

lth

was

rep

orte

d,

but

not

alw

ays

40.

atth

e10

%le

vel.

atth

e5%

leve

l.at

the

1%le

vel.

38 O. Thompson / Journal of Hea

cknowledgments

Michael Ash and Arindrajit Dube provided invaluable feedbackn multiple drafts. The paper also benefited from conversationsith Nancy Folbre and Samuel Bowles, as well as from detailed

omments provided by associate editor David Cutler and twononymous referees.

ppendix A.

The first four columns of Table A1 present versions of theaper’s preferred specification using alternative methods of adjust-

ng AFQT score to account for age and education when the AFQT wasaken. The first and second columns show results for the White andlack samples when age and education at the time of testing areimply added as covariates. Within the White sample, the AFQToefficient is .0943 and is highly significant, but within the Blackample the AFQT coefficient is slightly negative at −.00731 ands not statistically significant. The difference between the Whitend Black AFQT coefficients is highly significant, with a p-value of

0086.The third and fourth columns of Table A1 show results where

ach respondent’s adjusted AFQT measure is the residual from aegression with AFQT score as the dependant variable and age andducation at the time of testing as the independent variables. Withhis alternative adjustment technique, the health impact of AFQTithin the Black sample remains slightly negative and statistically

nsignificant, while within the White sample the AFQT coefficientalls to .0377 and just misses the conventional cutoff for statisti-al significance, with a t-statistic of 1.57 and a p-value of .1160.he difference between the White and Black coefficients is just onhe threshold of statistical significance as well, with a p-value of1163.

The reduction in the White AFQT coefficient is likely due to theact that educational attainment and AFQT score are highly corre-ated, so that using only the portion of AFQT score which is notxplained by educational attainment excludes many of the unob-ervable characteristics which AFQT score otherwise indexes. Theresence of this issue is the reason that the adjustment methodsed in the main body of this paper relies on the estimate of educa-ion’s causal effect on AFQT score provided by Hansen et al. (2003).ote that while the White AFQT coefficient is indeed smaller under

his alternative adjustment method, it remains substantively large.he association between AFQT score and adult health within thehite sample is of about the same magnitude as a $10,000 increase

n average annual adult income, while within the Black sample its not meaningfully different from zero.

The fifth column of Table A1 reproduces the preferred speci-cation for the Black sample, but adds interaction terms of AFQTcore with education and income. Both interaction terms are smallnd insignificant. Columns 6–8 investigate potential nonlinearitiesn the effect of AFQT score on health within the Black sample bydding quadratic and cubic terms to the preferred specificationnd by estimating the preferred specification on a sample of Blacksith AFQT scores above the population mean. In all cases, AFQT

core remains an insignificant predictor of health among Blacks.able A2 shows results from three additional variants of the pre-erred specification. The first two columns enter self-rated health

n logs; the middle two columns use an un-weighted version of theample; and the final two columns estimate the preferred specifi-ation as an ordered probit. In all cases, AFQT score is a large andtatistically significant predictor of adult health for Whites, but ismall and insignificant for Blacks. Ta

ble

A2

Ad

dit

ion

alsp

ec

Spec

ifica

tion

AFQ

Tz-

scor

eH

igh

est

grad

eA

vera

gead

ult

Stan

dar

dd

evi

Cu

rren

tly

mar

Nev

erm

arri

edIn

sure

dFe

mal

eC

onst

ant

Obs

erva

tion

sR

-squ

ared

Not

es:

Rob

ust

ssi

blin

gsin

the

rw

hic

his

usu

ally

*Si

gnifi

can

ce**

Sign

ifica

nce

***

Sign

ifica

nce

lth Ec

R

A

B

B

B

B

C

C

C

C

C

C

F

F

G

G

G

G

G

H

H

H

H

H

I

K

K

M

N

N

R

S

Wadsworth, M., Kuh, D., 1997. Childhood influence on adult health a review of recentwork from the British 1946 national birth cohort study, the MRC national sur-

O. Thompson / Journal of Hea

eferences

uld, C., Sidhu, N., 2005. Schooling cognitive ability and health. Health Economics14, 1019–1034.

aker, M., Stabile, M., Deri, C., 2004. What do self-reported objective measures ofhealth measure? Journal of Human Resources 39 (4), 1067–1093.

arker, D., 1997. Maternal nutrition, fetal nutrition and diseases later in life. Nutri-tion 13 (9), 807–813.

ehrman, J., Rosenzweig, M., 2004. Returns to birth weight. Review of Economicsand Statistics 86 (2), 586–601.

lack, S., Devereux, P., Salvanes, K., 2007. From the cradle to the labor market? Theeffect of birth weight on adult outcomes. Quarterly Journal of Economics 122(1), 409–439.

ase, A., Fertig, A., Paxson, C., 2005. The lasting impact of childhood health andcircumstance. Journal of Health Economics 24 (2), 365–389.

ase, A., Lubtosky, D., Paxson, C., 2002. Economic status and health in childhood:the origins of the gradient. American Economic Review 92 (5), 1308–1334.

handola, T., Jenkinson, C., 2000. Validating self-rated health in different ethnicgroups. Ethnicity & Health 5 (2), 151–159.

unha, F., Heckman, J.J., Schennach, S.M., 2010. Estimating the technology of cogni-tive and noncognitive skill formation. Econometrica 78 (3), 883–931.

utler, D., Lleras-Muney, A., 2010. Understanding differences in health behaviors byeducation. Journal of Health Economics 29 (1), 1–28.

utler, D., Lleras-Muney, A., 2008 Understanding Differences in Health Behaviors byEducation. Princeton University, Mimeo.

lynn, A., 2000. IS trends over time. In: Arrow, K., Bowles, S., Durlauf, S. (Eds.),Meritocracy and Economic Inequality. Princeton University Press, Princeton, NJ.

rijters, P., Hatton, T., Martin, R., Shields, M., 2010. Childhood economic conditionsand length of life: evidence from the UK boyd orr cohort 1937–2005. Journal ofHealth Economics 29 (1), 39–47.

eronimus, A., Hicken, M., Keene, D., Bound, J., 2006. Weathering and age patternsof allostatic load scores among Blacks and Whites in the United States. AmericanJournal of Public Health 96 (5), 826–833.

odfrey, K., Barker, D., 2001. Fetal programing and adult health. Public Health and

Nutrition 4, 611–624.

ottfredson, L., Deary, I., 2004. Intelligence predicts health and longevity, but why?Current Directions in Psychological Science 13, 1–4.

rossman, M., 1975. The correlation between health and schooling. In: Terleckyj, N.(Ed.), Household Production and Consumption, Studies in Income and Wealth.Columbia University Press, New York, New York.

W

onomics 30 (2011) 328–339 339

rossman, M., 1972. The Demand for Health: A Theoretical and Empirical Investi-gation. National Bureau of Economic Research, New York, New York.

ansen, K., Heckman, J., Mullen, K., 2003. The Effect of Schooling and Ability on Cog-nitive Achievement Test Scores. National Bureau of Economic Research WorkingPaper.

artog, J., Oosterbeek, H., 1998. Health, wealth and happiness: why pursue a highereducation. Economics of Education Review 17 (3), 245–256.

eckman, J., 2006. Skill formation and the economics of investing in disadvantagedchildren. Science 312 (5782), 1900–1902.

eckman, J., 2007. The economics, technology, and neuroscience of human capa-bility formation. Proceedings of the National Academy of Sciences 104 (33),13250–13255.

eckman, J., 1995. Lessons from the bell curve. Journal of Political Economy 103 (5),1091–1120.

dler, E., Benyamini, Y., 1997. Self-rated health and mortality: a review of 27 com-munity studies. Journal of Health and Social Behavior 38 (1), 21–37.

enkel, D., 1991. Health behavior, health knowledge, and schooling. Journal of Polit-ical Economy 99 (2), 287–305.

leinman, J, Kessel, S., 1987. Racial differences in low birth weight: trends and riskfactors. The New England Journal of Medicine 317 (12), 749–753.

cEwen, B., 1998. Protective and damaging effects of stress mediators. The NewEngland Journal of Medicine 338 (3), 171–179.

eal, D., Johnson, W., 1996. The role of premarket factors in Black-White wagedifferences. Journal of Political Economy 104 (5), 869–895.

epomnyaschy, L., 2010. Race disparities in low birth weight in the US south andthe rest of the nation. Social Science & Medicine 70 (5), 684–691.

ogers, W., Spriggs, W., 1996. What does the AFQT reall measure: race, wages,schooling and the AFQT score. Review of Black Poltical Economy 24 (4),13–46.

mith, J., 1999. Health bodies and thick wallets: the dual relationship betweenhealth and economic status. The Journal of Economic Perspectives 13 (2), 145–166.

vey of health and development. Pediatric and Perinatal Epidemiology 11 (1),2–20.

idgor, A., Green, B., 1991. Performance Assessment for the Workplace. NationalAcademy Press, Washington, DC.