Unexplained Gaps and Oaxaca-Blinder Decompositions

We analyze four methods to measure unexplained gaps in mean outcomes: three decompositions based on the seminal work of Oaxaca (1973) and Blinder (1973) and an approach involving a seemingly naive regression that includes a group indicator variable. Our analysis yields two principal findings. We first show that a commonly-used pooling decomposition systematically overstates the contribution of observable characteristics to mean outcome differences, therefore understating unexplained differences. We also show that the coefficient on a group indicator variable from an OLS regression is an attractive approach for obtaining a single measure of the unexplained gap. We then provide three empirical examples that explore the practical importance of our analytic results.