Recent papers in empirical economics (Giné et al. (2012), Dynarski et al. (2011), and others) have been interested in examining if a treatment has more effect among those who are predicted to have the worst response in the absence of any treatment. The statistical validity of this method has not been verified. There has been some work by Chingos and West (2012) showing simulationally that this method introduces bias into the estimate of the interaction between predicted response and treatment effect. We offer an analytic examination of the bias. We hope to produce a more defensible alternative.
A second topic provides a contribution to optimal matching in observational studies. We believe that we can gain speedups in algorithms towards that goal by leveraging parallel computation resources that are becoming widely available. We believe that by leaning on propensity score matching, we can offer a more structured approach to the matching problem that more readily allows such parallelized algorithms.