To generate persistence we augment the standard real business cycle (RBC) model with a "learning by doing" (LBD) mechanism, where current labor supply affects workers' future labor productivity. Our econometric analysis shows that the LBD model fits aggregate data much better than the standard RBC model. We calculate posterior odds for the structural models and formally show that the LBD model more closely mimics the autocorrelation and impulse response patterns that we found in a bivariate VAR analysis.