Keywords: causal inference, cross-validation, model selection, treatment effect heterogeneity
Estimation of (potentially) heterogeneous treatment effects is a popular area of research, and several methods have been recently proposed to estimate the effect of a treatment conditional on covariates. A data analyst who wants to estimate such conditional treatment effects must answer the question, “Which estimation procedure is best for the data I have?” Model selection methods typically focus on estimation of the regression function; however, within a set of candidate models, those that best estimate the regression function may not be the best for estimating the conditional effect of a treatment. The usual model selection methods may therefore be unsuitable for the purpose of treatment effect estimation. In this talk, we propose a modification of cross-validation that can be used for model selection targeted to estimating (potentially) heterogeneous treatment effects. Areas of future research also will be discussed.