Learning Identifiable Balanced Prognostic Score for Treatment Effect Estimation Under Limited Overlap

Tai-Wei Chang, Jiahui Li, Kun Kuang, Zhengqing Fang, Jianmin Huang, Longfei Li, JUN ZHOU, Wei Chu

Primary Area: causal reasoning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: causal inference; prognostic score; selection bias; limited overlap

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Understanding individual-level treatment effects is a fundamental and crucial problem in causal inference. In this paper, our objective is to tackle the issue of limited overlap, where certain covariates only exist in a single treatment group. We demonstrate that, under weak conditions, it is possible to simultaneously recover identifiable balanced prognostic scores and balancing scores. By leveraging these scores, we relax the requirement of overlapping conditions in a latent space, enabling us to generalize beyond overlapped regions. This approach also allows us to handle out-of-distribution treatments with no overlap. Additionally, our approach is adaptable to various tasks, including both binary and structured treatment settings. Empirical results on different benchmarks demonstrate that our method achieves state-of-the-art performance.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1752