Understanding gender and racial wage gaps among the highly educate

Hundreds of labor market studies investigate the wage gaps between white men and other demographic groups. Virtually all of these studies use a common approach: linear regression is used to adjust for factors that reflect differences in preferences and human capital between groups. Disparities that are not accounted for by these factors are potentially "due to discrimination." This research begins by discussing the assumptions necessary for making causal inferences about discrimination from observational data where the proxy for treatment, demographic group membership, is not manipulable. In light of this discussion I focus on estimating what is known as "the effect of treatment on the treated" instead of the usual parameter of interest in the parametric models. This new parameter estimates the effect of discrimination on those potentially discriminated against, accounting for their distribution of covariates. I use the 1993 National Survey of College Graduates (NSCG), to illustrate the methodology and investigate questions of discrimination. I propose estimating the effect of treatment on the treated using nonparametric methods new to the wage gap literature; exact matching and kernel smoothed matching. The nonparametric methods are carefully adapted to the specifics of this complex survey sample including issues with sampling weights and finite versus superpopulation inferences. Nonparametric bootstrap methods yielding random effective sample sizes and effective strata sizes are employed to obtain appropriate estimates of the standard errors.;Applying these methods, I present an extensive analysis of gender, racial/ethnic, and sexual orientation wage gaps in the United States among the highly educated. My research argues that among the highly educated, pre-labor market factors are associated with more than half of the wage gap for all groups and all of the wage gaps for some groups or subgroups. The non-parametric analysis I employ here benefits both from allowing flexible functional forms and from making transparent the lack of overlapping support between demographic groups. I find that without careful attention to these two issues, especially for women, for whom the distributions of pre-labor market factors and labor market experience are quite different than for white men, wage disparity estimates can be substantially biased.