Association for Public Policy Analysis & Management

We propose a method for identifying the characteristics of individuals who have the highest and lowest responses to a treatment.  Our approach is designed for situations where theory is lacking regarding the types of persons who should be most responsive to the treatment.  The method involves exploring a large combination of specifications that interact the treatment indicator with individual covariates, predicting the individual treatment effect in a training sample, applying the resulting coefficients to a cross-validation sample, finding the specification with the lowest mean squared prediction error, and finally applying the results to make out-of-sample predictions in a test sample.  We apply this method to re-examine the results of three papers in the literature: Wolf et al. (2013) on the effects of a school voucher program in Washington, DC; Finkelstein et al. (2012) on the effects of Medicaid expansion in Oregon; and Puma et al. (2010) on the effects of Head Start.  Our method illustrates how program administrators could (or could not) increase efficiency by targeting scarce program resources and program participation towards particular groups.