Title:Generative Learning of Partially Identified Treatment Effects
Speaker: 解海天（北京大学）
Time: Thursday, October 15, 2026, 15:00-16:00
Venue: Gewu Building, Room 315
Abstract:
We introduce a flexible generative learning framework to characterize the sharp identified set of causal parameters in instrumental variable models with heterogeneous treatment effects. The problem is framed as an inference task on whether a candidate parameter value can be produced by a generative model whose implied distribution is observationally indistinguishable from the observed data. We establish sharp identification, uniform size control, and consistency of the proposed procedure. The method accommodates a wide range of identification assumptions, does not rely on instrumental variable monotonicity or convexity of the space of heterogeneous treatment effect distributions, and applies to both discrete and continuous data. We use reproducing kernel Hilbert space methods to reduce computation and implement the generative model using neural networks with PyTorch for efficient optimization. We present empirically calibrated simulations and an empirical application on the returns to college.
About the Speaker：解海天，现任北京大学光华管理学院商务统计与经济计量系助理教授、博士生导师。2023年于加州大学圣地亚哥分校获得经济学博士学位，本科毕业于武汉大学。主要研究方向为经济学中的因果推断方法，包括工具变量、断点回归、双重差分、合成控制以及政策学习等。研究成果发表于Journal of Business & Economic Statistics、Journal of Econometrics、Theoretical Economics、《数量经济技术经济研究》、《经济学报》等。主持国家自然科学基金青年基金项目（C类），参与国家自然科学基金面上项目、重大项目。多篇文章被人大复印报刊资料全文转载。
