Policy Learning with Unobserved Heterogeneity, Draft available [Slides] [Abstract]
This article studies the optimal assignment of policies when treatment effects vary with unobserved characteristics. Building on the empirical welfare maximization framework, I show that deciding who to treat depending on the estimated values of such characteristics introduces a new approximation-estimation error trade-off in the policy learning problem. I connect this trade-off to the classic shrinkage problem and prove that estimating unobserved factors via the James-Stein estimator decreases regret on welfare, as compared to plug-in rules. Motivated by this result, I propose a new class of assignment rules that shrink the importance of the latent variable according to the welfare gain produced by its inclusion in the decision problem, and select the shrinkage parameter via cross-validation. I show that the selected value mirrors the estimator's signal-to-noise ratio, and prove that the resulting policy rule is minimax regret. I illustrate an empirical application in development economics, where including a proxy for entrepreneurs' business skills in policy recommendations increases welfare by 6% and halves the probability of generating welfare losses.
Winner of Unicredit Young Economist Best Presentation Award
Synthetic Welfare Maximization, Work in Progress
BallotBot: Can AI Chatbots Lower Voter-Information Barriers? with Elliott Ash (ETH) and Sergio Galletta (ETH)
[CEPR WP]
[Abstract]
This study examines the potential of AI-powered chatbots to increase engagement with political information. We develop and evaluate BallotBot, an AI chatbot with access to official voter guide information from the November 2024 referendums in California. In a pre-registered three-wave survey experiment in the weeks around election day, participants (California voters) were randomly assigned to use either BallotBot or a traditional digital voter guide to answer questions about ballot initiatives. We find that BallotBot access lowered the perceived cost of acquiring information for less-informed participants, fostered greater engagement with political information, while improving participants' ability to correctly answer in-depth questions about ballot measures.
Status: Submitted