Policy Learning via Quantile Partial Effects under General Interventions

Zhengyu Zhang is Professor of Econometrics in the School of Economics at Shanghai University of Finance and Economics. His research focuses on micro-econometrics, causal inference, and policy evaluation. In particular, he specializes in applying causal inference models to the empirical analysis of China's economy. He was selected for the National Young Talents Program (Ministry of Education). His work has received the Third Prize of the Ministry of Education Award for Outstanding Achievements in Humanities and Social Sciences, the Multa Scripsit Award conferred by the Econometric Theory, and the First Prize of the Shanghai Philosophy and Social Sciences Outstanding Achievement Award, among other honors. To date, he has chaired five projects funded by the National Natural Science Foundation of China and the National Social Science Fund of China.

This paper develops a policy learning framework for optimizing continuous treatment interventions via marginal perturbations, with a focus on unconditional outcome quantiles. In contrast to standard approaches that prioritize binary assignment rules or global mean welfare, we formulate the optimization problem as maximizing the Average Quantile Partial Effect (AQPE) subject to an L2-norm constraint. This constrained formulation ensures the statistical well-posedness of the optimization problem and controls the complexity of the policy class without requiring global estimation of the response surface. We implement a computationally efficient first-order algorithm to solve this problem and establish uniform asymptotic theory for the estimator. Specifically, we show that the estimated policy value function converges weakly to a zero-mean Gaussian process and that the learned policy achieves an asymptotically optimal regret rate of O(n−1/2). An application to the Job Training Partnership Act (JTPA) data indicates significant distributional heterogeneity, showing that policies targeting lower-tail outcomes differ from those maximizing the mean or upper quantiles.