题　目：Inference for two-stage experiments under covariate-adaptive randomization

嘉　宾：Jizhou Liu, Ph.D, University of Chicago Booth School of Business

主持人：吴瑞佳 助理教授 上海交通大学安泰经济与管理学院

时　间：2024年12月20日（周五）10:00-11:30

地　点：上海交通大学 徐汇校区 安泰浩然楼308室

内容简介:

This paper studies inference in two-stage randomized experiments under covariate-adaptive randomization. In the first stage, clusters (e.g., households, schools, or graph partitions) are stratfied and assigned to control or treatment groups based on cluster-level covariates. In the second stage, units within each treated cluster are further stratfied and assigned to control or treatment groups based on individual-level covariates. Under the homogeneous partial interference assumption, I establish conditions for the consistency and asymptotic normality of difference-in-"average of averages" estimators for primary and spillover effects. I also develop consistent estimators of their asymptotic variances, ensuring the validity of tests based on these estimators. My findings highlight that ignoring covariate information during the design stage can lead to efficiency loss, and commonly used inference methods that misuse covariate information may yield conservative or invalid results. Additionally, I explore the optimal use of covariate information in large samples and show that a generalized matched-pair design minimizes the asymptotic variance for each estimator. Finally, I discuss covariate adjustment for baseline covariates not used in the treatment assignment. The theoretical results are supported by a imulation study and an empirical application, demonstrating their practical relevance.

演讲人简介：

Jizhou Liu is a job market candidate from the University of Chicago Booth School of Business with a PhD in Econometrics and Statistics. His research focuses on the design and analysis of randomized experiments, particularly addressing challenges related to covariate-adaptive randomization and interference. His work has been published in leading journals and conferences, including Quantitative Economics, Journal of Econometrics, ICML and AISTATS.

欢迎广大师生参加!