Abstract
Using a recent result from the program evaluation literature, the author demonstrates that the interpretation of regression estimates of between-group differences in wages and other economic outcomes depends on the relative sizes of subpopulations under study. When the disadvantaged group is small, regression estimates are similar to the average loss for disadvantaged individuals. When this group is a numerical majority, regression estimates are similar to the average gain for advantaged individuals. The author analyzes racial test score gaps using ECLS-K data and racial wage gaps using CPS, NLSY79, and NSW data, and shows that the interpretation of regression estimates varies substantially across data sets. Methodologically, he develops a new version of the Oaxaca-Blinder decomposition, in which the unexplained component recovers a parameter referred to as the average outcome gap. Under additional assumptions, this estimand is equivalent to the average treatment effect. Finally, the author reinterprets the Reimers, Cotton, and Fortin decompositions in the context of the program evaluation literature, with attention to the limitations of these approaches.
