Identification and Estimation of Treatment Effects under Coupled Confounding and Collider Biases

Dian Jin, Baohong Li, Anpeng Wu, Yingrong Wang, Ming Ma, Kun Kuang

Keywords: Causal Inference, Instrumental Variables, Treatment Effect Estimation, Confounding Bias, Collider Bias, Selection Bias, Missing Data

TL;DR: We propose a new identification and estimation theory for treatment effects under coupled biases with an IV set.

Abstract: In causal inference, confounding bias and collider bias pose two major challenges for treatment effect estimation in observational studies. Confounding bias arises from unobserved common factors that simultaneously affect the treatment and outcome, while collider bias results from non-random sample selection caused by both variables. Existing methods focus on bias correction for a specific bias, such as using Instrumental Variables (IVs) to address confounding bias and Selection IVs (SIVs) to mitigate collider bias. However, real-world data frequently exhibit coupled confounding and collider biases, where unmeasured confounders directly affect the selection mechanism. Currently, the coupled biases problem remains an unaddressed challenge. In this paper, we propose a new identification theory for treatment effects under coupled biases with an IV set, which contains subsets serving as IV and SIV, respectively. Based on this theory, we propose a novel treatment effect estimation method, DualDebiasIV (DDIV), which decomposes the IV set to separately obtain the SIV and IV, using them for biases decoupling and correction. To the best of our knowledge, this is the first work to provide a solution for the identification and estimation of treatment effects under coupled biases. DDIV is theoretically guaranteed, with proofs provided for the correctness of the decomposition and the consistency of the estimates. Extensive experimental results on semi-synthetic and real-world datasets show that DDIV achieves significant performance improvements, further demonstrating its practical effectiveness.

Primary Area: causal reasoning

Submission Number: 8822