Comparing earnings equations in the Panel Study of Income Dynamics (PSID) and the Current Population Survey, we find that the PSID considerably underestimates the returns to education during the 1992–2007 period. Non-random selection in the PSID sample appears to be the reason.