Sharp Inference on Selected Subgroups in Observational Studies

Seminar Series

Friday, August 28, 2020 - 01:30
Jingshen Wang, PhD

Abstract: In the study of high-dimensional observational data (e.g., program evaluation, electronic health records data, health care claim database, and educational research), making inference on the maximum treatment effect is particularly helpful for answering some important questions in subgroup analysis, where selecting and making inference on treatment effect for the selected subgroup plays an essential role. Commonly adopted statistical inference applied to the selected subgroup typically assumes the subgroup is chosen independently of data. Such an inferential procedure can lead to an overly optimistic evaluation of the selected subgroup. In this talk, we propose a resampling framework to simultaneously address the issue of selection bias and the regularization bias induced by the high-dimensional covariates. Our procedure is computationally efficient and provides an asymptotically sharp confidence interval for the maximum treatment effect as well as the selected treatment effect. We demonstrate the merit of our proposal by intensive simulation studies and by analyzing UK Biobank data.

Jingshen Wang, PhD
Assistant Professor - Biostatistics
School of Public Health