This paper examines the appropriate functional form and the size of the wage returns to training. In both the National Longitudinal Survey of Youth (NLSY) and Employer Opportunity Pilot Project (EOPP) datasets a log specification fits best. In the NLSY, the full effect of training occurs with a lag as long as two years, training on previous jobs is a substitute for training on the current job, and the return to training declines with labor market experience. The EOPP data indicate that formal and informal training are perfect substitutes; however, an hour of formal training has a much greater effect on wages than does an hour of informal training.
We find very large returns to formal training in both the NLSY and EOPP. The mixed continuous-discrete nature of the training variable means that measurement error can cause estimates of the effects of short spells of training to be biased upward, but we demonstrate that the maximum upward bias in estimated returns at the geometric mean is minimal. Heterogeneity in returns is a more plausible explanation of the high estimated return to training; in the EOPP data, the return to training is significantly higher in more complex jobs. With unobserved heterogeneity in returns, our estimates can be regarded as the return to training for the trained, but cannot be extrapolated to the untrained.