joint work with P. Frumento, B. Pacini, D.B. Rubin

The effects of a training program on employment and wages are evaluated, using data from a randomized study, the National Job Corps Study, and the principal stratification approach to simultaneously address the complications of noncompliance, truncation of wages by nonemployment, and missing outcomes. We conduct a likelihood-based analysis using the theory of finite mixture models for identifiability and exploiting the EM algorithm. We maintain an exclusion restriction assumption on the outcomes but do not impose monotonicity of the truncation of wages. Monotonicity of compliance holds by design. We assess the robustness of our results by providing estimates under the Missing at Random assumption, as well as other nonignorable missing data mechanisms. The plausibility of meaningful restrictions is investigates by means of scaled log-likelihood ratio statistics. For compliers, results show that the effect on employment is generally positive; however, there is a subgroup of complying participants for whom the program is detrimental on employment in the short term, although this effect becomes negligible in the long term. For the subgroup of always-employed compliers, that is, those who would be employed whether trained or not, the effect of training on wages is positive in the short and medium term, while it becomes negligible in the long run.
 

VIEW PAPER