This paper studies the estimation and evaluation of targeted treatment assignment policies, where a data-driven prioritization rule is used to determine to whom the treatment is either offered or delivered. We consider a setting where the data come from a randomized experiment with non-compliance.
Consider first the motivation for analyzing heterogeneity in treatment effects with respect to the observed characteristics (for example, how treatment effects depend on pre-treatment characteristics such as pre-existing medical conditions) in experiments without compliance concerns. Traditionally, analyses of randomized experimental data have relied on pre-specified subgroup analyses to examine treatment effect heterogeneity. However, because of concerns about statistical power and multiple hypothesis testing, these analyses often examine only a small number of strata and subgroups, for example stratifying analyses by age or gender, preventing a comprehensive investigation of treatment effect heterogeneity. These analyses often do not consider subgroups defined by interactions among covariates, or nonlinear functions of covariates. This limitation has two consequences.