Primary Submission Category: Randomized Designs and Analyses
Learning Treatment Effects while Treating under Priority Queues
Authors: JungHo Lee, Johnna Sundberg, Bryan Wilder,
Presenting Author: Johnna Sundberg*
A recurring challenge in social programs is allocating scarce resources, such as housing assistance, under uncertainty about the program’s benefits. In practice, resources are often prioritized toward individuals judged to have higher need, and applicants are commonly first categorized into priority tiers rather than randomized outright. Motivated by this, we introduce an experimental design that randomizes incoming applicants into priority queues using assignment probabilities based on their risk scores; within each queue, treatment offers are made in priority order and then first-in-first-out as budget becomes available. We formulate the choice of queue-assignment probabilities as a convex optimization problem for efficient estimation of the average treatment effect and priority-based local effects under noncompliance, formalizing the trade-off between more conservative identification assumptions and achievable statistical power. The framework will be deployed with the Allegheny County Department of Human Services in Pittsburgh, Pennsylvania, to provide housing assistance to people experiencing homelessness.