Revisiting Covariate and Hypothesis Roles in ITE Estimation: A New Approach Using Laplacian Regularization

David Cohen, Ronen Talmon, Uri Shaham

Keywords: Individual Treatment Effect (ITE), Conditional Average Treatment Effect (CATE), Covariate Shift, Laplacian Regularization, Causal Inference

Abstract: The recent surge in data availability across many fields, such as medicine, social science, and marketing, has brought to the forefront the problem of estimating Individual Treatment Effect (ITE) from observational data to effectively tailor treatment to personalized characteristics. ITE estimation is known to be a challenging task because we can only observe the outcome with or without treatment, but never both. Moreover, observational datasets exhibit selection bias induced by the treatment assignment policy. In this paper, we present a new approach consisting of two novel aspects. First, we depart from conventional approaches that minimize the covariate shift. Instead, we incorporate it as a crucial element in ITE estimation, recognizing that it stems from highly predictive features that exhibit significant imbalance in observational data. Second, unlike existing methods, our approach utilizes hypothesis functions to directly estimate outcomes under covariate shift, enhancing reliability across observed and unobserved outcomes. To support this approach theoretically, we derive a new upper bound of the expected ITE loss and show that it explicitly depends on the discrepancy between the hypothesis functions, which are absent from the objectives of existing methods. Based on this new approach, we present LITE: Laplacian Individual Treatment Effect, a novel method that leverages Laplacian-regularized representation and incorporates both the covariate shift and the hypothesis functions for ITE estimation, effectively bridging observed and unobserved outcomes. We demonstrate LITE on illustrative simulations and two leading benchmarks, where we show superior results compared to state-of-the-art methods.

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.

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

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: 10099