2025/12/22 by Yue Fang, Geert Ridder, Fang, Yue +3
Mathematics · Decision Sciences · #Advanced Causal Inference Techniques #Advanced Bandit Algorithms Research #Statistical Methods and Inference
paper · doi:10.48550/arxiv.2512.19230
Scarce opportunities are often allocated by lotteries. We study how to improve such allocations by estimating optimal assignment rules that maximize welfare net of a Kullback--Leibler penalty for departing from the benchmark randomization. The framework covers discrete, continuous, and mixed treatments. Regret is asymptotically quadratic in the estimation error, so inefficient estimation raises the mean of limiting regret, not merely its dispersion. We show that inverse probability weighting with known assignment probabilities is inefficient, whereas estimated-propensity and doubly robust welfare criteria attain the efficient regret distribution. Simulations and a commitment-savings application quantify the resulting precision gains.