Making the Most Out Of Social Experiments : Reducing the Intrinsic Uncertainty in Evidence from Randomized Trials with an Application to the JTPA Exp

This paper demonstrates that even under ideal conditions, social experiments in general only uniquely determine the mean impacts of programs but not the median or the distribution of program impacts. The conventional common parameter evaluation model widely used in econometrics is one case where experiments uniquely determine joint the distribution of program impacts. That model assumes that everyone responds to a social program in the same way. Allowing for heterogeneous responses to programs, the data from social experiments are consistent with a wide variety of alternative impact distribution. We discuss why it is interesting to know the distribution of program impacts. We propose and implement a variety of different ways of incorporating prior information to reduce the wide variability intrinsic in experimental data. Robust Bayesian methods and deconvolution methods are developed and applied. We analyze earnings and employment data on adult women from a recent social experiment. In order to produce plausible impact distributions, it is necessary to impose strong positive dependence between outcomes in the treatment and in the control distributions. Such dependence is an outcome of certain optimizing models of the program participation decision