When Treatment Effect Estimation Meets Collider Bias: A Dual Counterfactual Generative Approach
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, Treatment Effect Estimation, Collider Bias, Selection Bias, Generative Models
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Collider bias poses a great challenge in estimating the treatment effect from observational data due to the sample selection mechanism on both treatments and outcomes. Previous works mainly focused on addressing confounding bias and selection bias caused by covariates only. However, they failed to accurately estimate the causal effect with collider bias, which is known to be an unidentifiable problem without further assumptions on the observational data. In this paper, we address collider bias in the observational data by introducing small-scale experimental data. Specifically, we treat the collider bias problem from an out-of-distribution perspective, where the selected observational data comes from an environment labeled with $S=1$, and the unselected data comes from another environment labeled with $S=0$. The experimental data comes from the entire data space, but the environment labels are unknown. Then, we propose a novel method named Dual Counterfactual Generative Model (DCGM), which consists of two generators that respectively generate the unselected data and the missing $S$ labels, and two discriminators that discriminate between the observational data and data with generated $S=1$ labels, as well as between the generated unselected samples and data with generated $S=0$ labels for training the generators. Combining the observational data with the unselected samples generated by DCGM, the treatment effect can be accurately estimated using the existing approaches without considering the collider bias. Extensive experiments on synthetic and real-world data demonstrate the effectiveness and the potential application value of the proposed method.
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: 5504
Loading