Statistical Treatment Rules for Heterogeneous Populations: With Application to Randomized Experiments

This paper uses Wald's concept of the risk of a statistical decision function to address the question: How should sample data on treatment response be used to guide treatment choices in a heterogeneous population? Statistical treatment rules (STRs) are statistical decision functions that map observed covariates of population members and sample data on treatment response into treatment choices. I propose evaluation of STRs by their expected welfare (negative risk in Wald's terms), and I apply this criterion to compare two STRs when the sample data are generated by a classical randomized experiment. The rules compared both embody the reasonable idea that persons should be assigned the treatment with the best empirical success rate, but they differ in their use of covariate information. The conditional success (CS) rule selects treatments with the best empirical success rates conditional on specified covariates and the unconditional success (US) rule selects a treatment with the best unconditional empirical success rate. The main finding is a proposition giving finite-sample bounds on expected welfare under the two rules. The bounds, which rest on a large-deviations theorem of Hoeffding, yield explicit sample-size and distributional conditions under which the CS Rule is superior to the US rule.