Active Causal Learning for Conditional Average Treatment Effect Estimation

ICLR 2025 Conference Submission13594 Authors

Keywords: conditional average treatment effect, dynamic sampling, partially observed Markov decision process

Abstract:

Estimating conditional average treatment effects (CATE) from observational data is an important problem and is of high practical relevance for many domains. Despite the great efforts of recent studies to accurately estimate CATE, most methods require complete observation of
all covariates of an individual. However, in real-world scenarios, the acquisition of covariate information is usually done in a active manner, which motivates us to develop methods to minimize the total measurement cost by actively selecting the most appropriate covariates to measure while guaranteeing the CATE estimation accuracy. To this end, in this paper, we first extend the existing methods for estimating CATE to allow accurate estimation in the presence of unmeasured covariates. Next, we theoretically show the advantage of dynamically adjusting the sampling strategy based on an evolving understanding of the information measured in the covariates. Then, we formulate the dynamic sampling strategy learning as a partially observed Markov decision process (POMDP) and further develop a policy gradient method to solve the optimal dynamic policy. Extensive experiments conducted on three real-world datasets demonstrate the effectiveness of our proposed methods.

Primary Area: other topics in machine learning (i.e., none of the above)

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 13594