to their analysis). The results are largely consistent with those of Cawley et al. (1996),
confirming their conclusion that “the correlations of g with wages... are modest compared to
those of education. [and] the returns to g differ significantly across race and gender: payment is
not made for ‘ability’ alone” (Cawley et al. [1996], p. 17).18 Furthermore, regression 2C shows a
similar pattern to 2A, indicating no significant change in the 1994 to 2002 period from the 1979
to 1994 period. Regressions 2B and 2D are respectively the fixed effects counterparts of 2A and
2C. The Hausman specification test is used to test for the strict exogeneity assumption. For both
time periods, we reject the null hypothesis that the strict exogeneity condition is not violated.19
Hence, for both time periods, the random effects model does not generate consistent estimators.
Therefore, Cawley et al. (1996) are not clearly justified in using the random effects model even
though it generates evidence supporting their claims.
The intuitive reason for the failure of the random effects model may be that the
composite error terms include unobservable individual characteristics, such as education quality,
which are correlated with some of the independent variables. As discussed above, education
quality may vary systematically with intelligence because colleges use intelligence-tapping tests
to differentiate and recruit students. Moreover, education quality should also be correlated with
education level. The better is someone’s schooling quality, the more likely he/she will be
qualified to further his/her education. Thus, the strict exogeneity assumption is violated and the
fixed effects model should be used instead of the random effects one.
Due to the fact that g is time-invariant in the NLSY79, the fixed effects specification pre-
empts a meaningful comparison of the respective marginal effects of g and education on the
18 The coefficient of g is 0.042 and the coefficient of education is 0.105; both are significant at the 5 percent significance level.
The coefficients of the race/gender dummy variables and their cross-terms with g are all statistically significant except for the
intercept of the black race dummy.
19 The Hausman specification test indicates that before 1994, Chi square is 1,174.86, and the probability of p>Chi square is 0.000;
after 1994, Chi square is 841.15, and the probability of p>Chi square is 0.000.
12