Three essays on empirical methods for evaluating the impact of policy interventions in education and training

This dissertation examines different empirical approaches to estimating the effects of policy interventions on individual labor market outcomes. The first essay investigates the gain from using semiparametric discrete choice methods in evaluating the effects of a job training program. Estimated probabilities of program participation play a central role in matching and generalized residual estimators of program effects. First, I show how two binary choice estimators in the literature, Ichimura's (1993) "semiparametric least squares estimator" and Klein and Spady's "quasi-maximum likelihood" estimator, can be modified for choice-based sampling designs. No semiparametric method currently exists for estimating probabilities under this type of sampling, which is often chosen in evaluation studies. In a Monte Carlo study, I investigate the practical benefit from using a semiparametric estimation method. Then I study the determinants of participation in JTPA, a government sponsored job training program and implement propensity score matching estimators of program impacts.;The second essay assesses the performance of alternative nonexperimental methods of evaluating the impact of a social program. It discusses many of the methods available and evaluates the empirical performance of different estimators. The empirical work emphasizes recently advanced semiparametric approaches (series expansion methods and local linear regression methods) of controlling for selection bias.;The third essay, which is coauthored with James J. Heckman and Anne Layne-Farrar, formulates and estimates log earnings equations that are economically interpretable as human capital pricing equations and uses these models to study the relationship between earnings and schooling quality. It demonstrates that different assumptions about the economic environment and about the substitutability of skills among workers imply different restrictions on the earnings equation. We test for many of these restrictions using Census data. A major finding of this paper is that an efficiency units representation of human capital is rejected in favor of a model with heterogeneous skill types.