Seminario Affordable Decisions in a Confounded World

Abstract

A policymaker observes K populations that adopted an innovation and a target population in the status quo. She decides whom to innovate in the target population and is accountable for any wrong decision: a finite compensation budget must cover the worst-case compensation cost. Standard policy learning assumptions (unconfoundedness, strict overlap, and common population) are violated. I define as certified decision rules that control the sign-error probability uniformly over a coverage set and show that certification implies an affordability guarantee. For the class of matching estimators with positive weights, I derive a finite-sample sufficient condition for certification. I show that in large samples the geometry of the estimating data governs affordability, simplex-based estimators provide the best affordability guarantees, and Delaunay matching is the optimum. I leverage this result to characterize data collection plans that achieve affordability at a given budget with finite collection effort.