|
Abstract:
|
In this paper we consider inference on the treatment effect heterogeneity from experimental data. We take as a basic premise that the goal of such inference is to guide policy decisions regarding targeting of treatment assignment. We suppose that the researcher has access to data from a randomized experiment that, in addition to outcomes and treatment status, contains (possibly a large number of) pre-treatment covariates. We propose a data-driven method, based on lasso variable selection, to discover those subgroups that show the largest gains from treatment (allowing for a targeting cost). Our goal is thus quite different from what has been considered in the literature, which largely focuses on either selecting variables to efficiently or unbiasedly estimate the average treatment effect, or else comparing heterogeneity across subgroups. We illustrate our method on both real and simulated data.
|
