A Generative Approach for Treatment Effect Estimation under Collider Bias: From an Out-of-Distribution Perspective

This paper addresses the challenge of estimating incremental value in advertising by tackling selection bias through an out-of-distribution (OOD) perspective. The authors propose a coupled generative adversarial model that combines biased observational data with unbiased experimental data to achieve unbiased and robust treatment effect estimation. The model consists of two generators and two discriminators that work together to generate missing samples and labels, enabling the creation of a bias-free dataset for causal inference. Experimental results on three public datasets demonstrate the effectiveness of the proposed method in reducing estimation bias compared to using observational or experimental data alone.