Deep Entire Space Cross Networks for Estimating Treatment Effect Under Unmeasured Confounding
Han Qiao · 2024
The conditional average treatment effect (CATE) estimation is crucial in fields such as healthcare, e-commerce, and economics. It is not possible to observe both potential outcomes of a given individual at the same time. Instead, we can only observe the outcome of the treatment actually received. Therefore, we rely on observed data in the training space to infer unobserved outcomes in the inference space. However, as the treatment assignment mechanism is not random, there is a distributional discrepancy between the treatment and control groups. Consequently, inferring potential outcomes solely within the corresponding treatment or control space results in inaccurate CATE estimation. An alternative approach is to use the entire sample space to jointly estimate CATE. However, existing methods fail to achieve complete coverage of the entire space and fail to consider the impact of unmeasured confounding. Our method is the first to achieve a complete match between the train and inference spaces. By combining OBS data with a propensity network, we can obtain a biased outcome estimate across the entire space. The incorporation of RCT data allows the construction of a residual network to calibrate the biased estimation. This approach enables us to achieve complete entire-space coverage and obtain accurate CATE estimation even under unmeasured confounding. Experiments on synthetic and real-world datasets demonstrate the effectiveness of our approach.