Abstract
This article provides a general estimation and inference framework to study how different levels of program participation affect participants' outcomes. We decompose differences in the outcome distribution into (i) a structure effect, arising due to the conditional outcome distributions given covariates associated with different levels of participation; and (ii) a composition effect, arising due to differences in the distributions of observable characteristics. These counterfactual differences are equivalent to the multivalued treatment effects for the treated under a conditional independence assumption. We propose efficient nonparametric estimators based on propensity score weighting together with uniform inference theory. We employ our methods to study the effects of the Workforce Investment Act (WIA) programs on participants' earnings. We find that heterogeneity in levels of program participation is an important dimension to evaluate the WIA and other social programs in which participation varies. The results of this article, both theoretically and empirically, provide rigorous assessment of intervention programs and relevant suggestions to improve their performance and cost-effectiveness.
Supplementary materials
for this article are available online.
