A large body of multidisciplinary research has documented how sentencing outcomes vary tremendously across racial and ethnic groups. The research challenge lies in establishing whether these sentencing differentials are driven by unobserved heterogeneity correlated to defendant race/ethnicity, or whether they reflect discrimination. We add to the debate by examining the robustness of racial/ethnic sentencing gaps, by gender, when allowing for selection on unobservables. We do so in the context of federal criminal cases, considering 250,000 cases, and using a dataset containing a rich set of covariates relating to defendant and legal characteristics of cases.