Evidence-Based Policy Learning

The past years have seen seen the development and deployment of machine-learning algorithms to estimate personalized treatment-assignment policies from randomized controlled trials. Yet such algorithms for the assignment of treatment typically optimize expected outcomes without taking into account that treatment assignments are frequently subject to hypothesis testing. In this article, we explicitly take significance testing of the effect of treatment-assignment policies into account, and consider assignments that optimize the probability of finding a subset of individuals with a statistically significant positive treatment effect. We provide an efficient implementation using decision trees, and demonstrate its gain over selecting subsets based on positive (estimated) treatment effects. Compared to standard tree-based regression and classification tools, this approach tends to yield substantially higher power in detecting subgroups with positive treatment effects. INTRODUCTION Recent years have seen the development of machine-learning algorithms that estimate heterogeneous causal effects from randomized controlled trials. While the estimation of average effects - for example, how effective a vaccine is overall, whether a conditional cash transfer reduces poverty, or which ad leads to more clicks - can inform the decision whether to deploy a treatment or not, heterogeneous treatment effect estimation allows us to decide who should get treated. These algorithms aim to maximize realized outcomes, and thus focus on assigning treatment to individuals with positive (estimated) treatment effects. Yet in practice, the deployment of assignment policies often only happens after passing a test that the assignment produces a positive net effect relative to some status quo. For example, a drug manufacturer may have to demonstrate that the drug is effective on the target population by submitting a hypothesis test to the FDA for approval.