experience as age minus schooling minus 6. Thus, age drops out of Regression 1C (as opposed to
1B and 1A) because it is collinear with the combination of education and labor market
experience.
Herrnstein and Murray argue that it is g that determines future social outcomes, yet
curiously they use the AFQT scores as an approximate measure of g in their analysis. To correct
for any possible error of using AFQT to approximate g, we employ normalized g in regression
1D. By using AFQT instead of g, Herrnstein and Murray may have actually under-estimated the
role of general intelligence in explaining wage differentials (the coefficient of the normalized g is
0.040, versus 0.027, the coefficient of AFQT in regression 1C; however these are both
statistically insignificant; our rerunning of equations 1A and 1B—not herein reported—also
show a similar slight upscaling in the weight placed on normalized g relative to normalized
AFQT). One important observation is that now the coefficient of g is no longer greater than those
of the social/family factors. For example, an F test between the coefficient of g and the
coefficient of normalized log of family income shows that we cannot reject the null hypothesis
that these two coefficients are equal (F=0.31, p>F=0.575). Therefore, when the omitted variables
are included in the regressions, general intelligence no longer dominates socio-economic
background in explanatory power. Moreover, the variable interacting race (Black) with AFQT
or g is significant at the 1 percent significance level in regressions 1C and 1D, and the interaction
of gender (Male) with g is significant in 1D; the noninteracted dummies for Hispanic and Male
are also statistically significant. Thus, the cross-sectional analysis indicates that the wage rate is
not independent of racial/ethnic and gender group affiliation.
The caveat of cross-sectional analysis, however, is that it does not account for
unobservable individual characteristics of the respondents. For example, the quality of one's