Employer Learning and Statistical Discrimination

We show that if firms statistically discriminate among young workers on the basis of easily observable characteristics such as education, then as firms learn about productivity, the coefficients on the easily observed variables should fall, and the coefficients on hard-to-observe correlates of productivity should rise. We find support for this proposition using NLSY79 data on education, the AFQT test, father's education, and wages for young men and their siblings. We find little evidence for statistical discrimination in wages on the basis of race. Our analysis has a wide range of applications in the labor market and elsewhere.