讲座信息
题目:Generalized Balke–Pearl Bounds under Relaxed Exclusion Restrictions
报告人:刘越(中国人民大学 副教授)
时 间:2026.09.16 19:30
联系人:吴鹏(北京工商大学)
报告摘要
Instrumental variable methods rely on the exclusion restriction, which is often hard to justify in practice. We propose generalized Balke--Pearl bounds, sharp bounds for average treatment effects under exogeneity and a relaxed exclusion restriction, with the sensitivity parameter given by the proportion of units who violate the exclusion restriction. The proposed bounds reduce to the classical Balke--Pearl bounds when the violation proportion is zero and widen monotonically as it increases. We further show that exogeneity and the full exclusion restriction may not jointly hold given the observed data, and derive the sharp minimum violation proportion required for compatibility. This minimum is zero if and only if the instrumental inequalities hold, linking our sensitivity analysis directly to the testable implications of the standard IV model. When the treatment effect is sign-identified under the exclusion restriction, we propose an E-value that quantifies the minimum violation proportion needed to overturn the qualitative conclusion. We also examine whether various monotonicity assumptions can sharpen the proposed bounds. We illustrate the proposed method using four real-world datasets.
报告人简介
中国人民大学统计学院副教授,2019年博士毕业于北京大学。多篇文章发表于Journal of the Royal Statistical Society Series B(JRSS-B),Journal of Machine Learning Research(JMLR), IEEE Transactions on Pattern Analysis and Machine Intelligence(PAMI),Artificial Intelligence(AIJ) International Conference on Machine Learning(ICML),Neurips等机器学习与统计学期刊及会议。研究兴趣主要包括敏感性分析,因果图模型以及可信机器学习。