Biased?An obvious objection to our interpretation of table 1 is that the AFQTis a racially biased test in the sense that its scores underpredict pro-ductivity or job performance for blacks compared to whites. For manytests, it would be impossible to judge the validity of such an assertionbecause we typically have no way of directly measuring job perfor-mance and relating it to the test scores received. However, in 1991the National Academy of Sciences (NAS) completed an exhaustivestudy with the Department of Defense of the validity of the AFQTwith special emphasis on the racial fairness of the test. The uniqueaspect of the NAS study is that job performance was measured with-out using either supervisor evaluations or written tests, two methodsthat could be seen as introducing racial bias. Instead, for several mili-tary occupational specialties, direct measures of performance on thetasks constituting the job were undertaken. As an example, the jobof infantry rifleman in the Marine Corps was broken into 15 tasksand each task further divided into subtasks. Subtasks were smallenough that performance could be evaluated by a (1, 0) yes-no scor-ing system, which ensured a high degree of consistency across evalua-tors. Military job experts designed a weighting system that translatesthe subtask scores into a composite job performance measure.5 Thenthese "hands-on" measures of job performance were regressed onthe AFQT score of the individual at the time he or she enlisted inthe military.How well does AFQT predict military job performance? For the23 military occupations studied, the correlations between AFQTscores and job performance ranged from .13 to .49, with a mediancorrelation of .38.16 The more important question, however, concernsracial bias, a key issue for the NAS panel. It concluded that AFQTdoes not systematically underpredict black job performance relativeto white performance: "for practical purposes the same regressionlines predicted performance about as well for both groups" (Wigdor15 Examples of tasks tested are land navigation, squad automatic weapons, first aid,night vision device, rifle, live fire, etc. (see Wigdor and Green 1991, vol. 1).16 These correlations are likely to understate the correlation between AFQT and ageneral skill or capacity to learn a specific task because selection into military occupa-tional specialties is accomplished in part with test scores. Hence the range of test scoresfor any particular job is truncated. Since AFQT is also used to select individuals intothe military, any observations about racial differences in the power of AFQT to predictmilitary job performance apply only to the individuals joining the military.
--------------------------------------------------------------------------------
Page 12
BLACK-WHITE WAGE DIFFERENCES879and Green 1991, p. 179). 17 If anything, test scores slightly overpredictjob performance by blacks. We view the NAS findings as strong inde-pendent verification that the AFQT can be considered a racially unbi-ased predictor of success in acquiring new skills in the military, andwe have no reason to believe that the AFQT would be a racially biasedpredictor of success in acquiring civilian job skills.Do Blacks Underinvest in Skill Because theReturn Is Lower?Models of discrimination developed by Arrow (1973) and Lundbergand Startz (1983) yield discriminatory equilibria from black-white dif-ferences in the return to acquiring skill. In both models, blacks withmore skill have more difficulty distinguishing themselves to employ-ers than high-skill whites, and therefore the payoff to acquiring skillis lower for blacks. Our results in table 1 indicate that blacks andwhites earn different wages in large part because they typically begintheir careers with different levels of human capital. These models ofdiscrimination highlight the possibility that black youths enter thelabor market with relatively few skills simply because they anticipatethat the returns from acquiring skills will be low.We investigate this possibility in tables 2 and 3. While we have nodirect evidence about the expectations of these youths, we can lookfor differences among blacks, whites, and Hispanics in the realizedeffects of AFQT scores on civilian wages. The regression equationsreported in column 1 of both tables 2 and 3 include an interactionbetween black and AFQT. For men, there is some indication thatblack men fare relatively better at the high end of the AFQT distribu-tion. For women, the opposite is true. However, for both sexes, theestimated coefficients on the interaction terms are jointly insignifi-cant.'8 The remaining results in tables 2 and 3 show the marginaleffect of AFQT on log wages for each racial group. There are small,statistically insignificant black-white differences for men in table 2,and columns 2 and 3 of table 3 show that AFQT exerts an almostidentical effect on the wages of black and white women. For bothblack and white men and women, the law of one price roughly holdsfor skill as measured by AFQT. Nonetheless, since the Cutright(1973) study found that the return to skill investment was lower for17 At the mean level of black test scores, the average overprediction of black perfor-mance, in standardized units, is .15 when the job includes at least 75 blacks tested(Wigdor and Green 1991, p. 178). Overprediction also occurs on average for jobs withsmaller samples of blacks.18 Under the null hypothesis that the coefficients on both interaction terms are zero,the F-statistics for the male and female regressions are 2.20 and 2.17, respectively.
--------------------------------------------------------------------------------
Page 13