This paper derives a statistical discrimination model that includes the self selection that results when employees optimally choose which jobs to apply for. We show that in such a model important theoretical results in the statistical discrimination literature are overturned. For example, a simple yardstick like differences in average qualifications does not guarantee that members of the worse qualified group are always discriminated against. Strong conditions on group differences (MLRP must hold) are required to ensure that statistical discrimination against members of a single group will result. Furthermore, the resulting statistical discrimination is smaller than what would result in an economy in which employees do not select the jobs they apply for. When employers’ ability to measure qualifications differs from one group to another, we show that the group employers know least about is favored. Consequently any endogenous quality differences that might result from employee investment decisions favors the less familiar group. Finally, we show how our results can be used to explain a number of empirical puzzles that are documented in the literature.