报告人：Kang Shuai
时间： 2026-06-24 14：00-15：00
地点： 智华楼313
摘要：
Practitioners (e.g, doctors) will usually report an Average Treatment Effect (ATE), which fails to reflect the clinical experience and expectation that differences in patient prognostic characteristics will lead to heterogeneous responses to therapy. This is characterized by the so-called heterogeneous treatment effect (HTE). It is also a key question of great interest in the presence of survival outcome. The presence of censored data complicates the analysis of HTE due to the inclusion of the censoring probability. The existing methods mainly focus on directly estimating the so-called conditional average treatment effect (CATE), which represents the treatment effect for individuals with the same characteristics. This will encounter practical difficulties with many covariates and small sample size. Inference of the heterogeneity using the estimated CATE will be a further problem due to the multiple testing issue and small samples for certain survived subpopulations. In this paper, we propose a general approach based on the Neyman's repeated sampling framework and a pre-specified effect scoring rule, which can either be given in advance or estimated using the observed dataset. Using the effect score, we can further split the dataset into three subsets, with each estimating censoring probability, scoring rule and group average treatment effect (GATE). Then we conduct the test of homogeneous treatment effects across the groups and the rank-consistency of the proposed scoring rule, with valid confidence intervals being constructed. We further apply the cross-fitting technique to improve the estimation performance of the GATEs. The proposed method can be especially useful from the exploratory analysis perspective. Its validity does not rely on the specific form of scoring rule because the randomness comes solely from the treatment assignments, sampling variability of individuals and censoring probability estimation. Simulation experiments and an application are performed to illustrate our methodology.
报告人简介：
Kang is now a postdoc at The Dartmouth Institute at Geisel School of Medicine at dartmouth. He completed his PhD at Peking University in 2025. Before that, he obtained a bachelor degree in mathematics and applied mathematics from Zhejiang University in 2020. From January to May in 2024, he visited the statistical department at UC Berkeley. Kang mainly focuses on causal inference researches, including mediation analysis, peer effect, instrumental variables, factorial instrumental variables and effect score analysis. His work has been published in some reputable statistical journals including Statistica Sinica and Electronic Journal of Statistics.
