II. Previous Findings
Although the studies that have investigated male and female earnings are now legion, the number that actually report separate schooling coefficients is
much smaller. Many of those focusing on the returns to education
4
with data on both sexes have fitted pooled regressions, allowing for a sex differential by including a sim-
ple dummy variable with no interactive term. Where separate regressions have been run, schooling is often among the unreported controls. In the case of the subliterature
on the earnings gap, the single most plentiful source of studies with separate regres- sions, the regression coefficients are sometimes not reported at all.
Appendix 1 summarizes the 27 U.S. studies that satisfy three conditions: 1 the study reports either male and female schooling coefficients in parallel
regressions, or a joint regression that includes a female-schooling interactive term;
2 a Mincerian semilogarithmic specification is used for the wage equation; 3 the controls do not include either occupation or industry, and the sample is
not restricted narrowly by occupation or industry. The list is intended to be comprehensive, though doubtless some eligible studies have
been missed. The second condition occurs because it is generally impossible to derive compara-
ble schooling coefficients from studies that have used a linear specification for earn- ings and so a number of widely cited studies Cohen 1971; Suter and Miller 1973;
Featherman and Hauser 1976; Roos 1981; Grubb 1993 have had to be discarded. The linear model is in any case a misspecification Heckman and Polachek 1974;
Dougherty and Jimenez 1991.
The reason for the third is that much of the impact of schooling on earnings is medi- ated by occupational attainment and, perhaps to a lesser extent, by industrial recruitment.
Accordingly the use of occupational andor industrial controls, popular in the male- female earnings gap literature as a means of assessing how much of the gap is attributa-
ble to occupational segregation, strips the schooling control of much of its impact.
As can be seen from Appendix 1, most of the studies have used data from large, nationally representative data bases: the National Longitudinal Studies of Labor
Market Experience, the Panel Study of Income Dynamics, the Current Population Survey, Censuses of Population, and the National Longitudinal Study of the High
School Class of 1972 and its successor, High School and Beyond. Most of them have used years of schooling as the educational variable but some have used sets of dummy
variables. The latter approach makes male-female comparisons of the returns to schooling less straightforward, but in some cases it does provide an opportunity for
attempting to identify the schooling level at which the returns diverge. Most of the studies have used actual work experience and its square as controls. However some
4. By this is meant the proportional increase in earnings per year of schooling, following conventional usage. The expression is not intended to refer to the Fisherian internal rate of return. See Psacharopoulos
1981 for a discussion of the conditions under which the schooling coefficient might be interpreted as an approximation to the internal rate of return.
Dougherty 971
have used potential work experience and three have used age as a proxy. They vary widely in their use of other controls.
Of the 27 studies, 18 report unambiguously higher schooling coefficients for females. Six report multiple estimates where the female coefficients are mostly
higher.
5
Two report mixed results that are evenly balanced.
6
Only one reports higher schooling coefficients for males, and this study had a relatively small sample.
7
In gen- eral the studies report only point estimates, and it is thus impossible to determine
whether the differences in the coefficients are significant. However, taking the studies together, the effect is undoubtedly significant, for under the null hypothesis that the
coefficients are the same; the probability of 24 out of 27 studies finding higher coef- ficients for either sex is less than 0.01 percent.
As to which levels of schooling are responsible for the effect, there is some evi- dence that the returns to partial and complete college are higher for females than for
males; for years of high school, the evidence points the other way; and for postgrad- uate studies, the evidence is mixed.
8
Given the institutional differences between the labor market in the United States and labor markets in other countries, one would not anticipate that a feature of the U.S. mar-
ket would necessarily be encountered elsewhere. However, two recent surveys suggest that the effect may not be confined to the United States. Trostel, Walker, and Woolley
2002 estimate the returns to schooling in 28, mostly European, countries with data derived from a common survey instrument and found that the female schooling coeffi-
cient was higher in 24. Psacharopoulos and Patrinos 2002 list 95 estimates of male and female schooling coefficients from 49 countries at different dates. Of these 63 are greater
for females, three are equal, and 23 are greater for males.
III. Possible Causes of the Effect