嘉    宾: 解海天  助理教授 北京大学

主持人: 瞿    茜  教授 上海交通大学安泰经济与管理学院

时   间:  2024年3月20日(周三)14:00-15:30

地   点: 上海交通大学徐汇校区安泰经济与管理学院 B207

内容简介:This paper studies policy learning under the condition ofunconfoundedness with a continuous treatment variable. Our research begins byemploying kernel-based inverse propensity-weighted (PW) methods to estimatepolicy welfare. We aim to approximate the optimal policy within a global policyclass characterized by infinite Vapnik-Chervonenkis (VC) dimension. This isachieved through the utilization of a sequence of sieve policy classes, each withfinite VC dimension. Preliminary analysis reveals that welfare regret compriseof three components: global welfare deficiency, variance, and bias. This leads tcthe necessity of simultaneously selecting the optimal bandwidth for estimatiorand the optimal policy class for welfare approximation. To tackle this challengewe introduce a semi-data-driven strategy that employs penalization techniquesThis approach yields oracle inequalities that adeptly balance the threecomponents of welfare regret without prior knowledge of the welfare deficiencyBy utilizing precisemaxima andconcentration inequalities,we derive sharpelregret bounds than those currently available in the literature. In instances wherethe propensity score is unknown, we adopt the doubly robust (DR) momencondition tailored to the continuous treatment setting. In alignment with thebinary-treatment case, the DR welfare regret closely parallels the IpW welfareregret, given the fast convergence of nuisance estimators.

演讲人简介:Haitian Xie is an assistant professor at Peking UniversityGuanghua School of Management, Department of Business Statistics andEconometrics, His research focuses on econometrics, including causal inferencepolicy learning, and statistical decisions, Prior to ioining Peking University.heobtained his PhD in Economics from the University of California, San Diego