Go to AAAI 2025 Workshop AICT homepageA Reweighting Based Approach for Treatment Effect Estimation Under Unmeasured Confounding with Non-Representative Randomized Data
Keywords: treatment effect, unmeasured confounding, reweighting
Abstract: Causal effect estimation aims to measure the true causal relationship between treatment and outcome variables, which is widely applied in areas such as medicine, commerce, and sociology. A challenge in causal effect estimation is that unmeasured variables may affect both treatment and outcome variables, which are named unmeasured confounders. Traditional methods of causal effect estimation are biased in the presence of unmeasured confounding. Previous data fusion-based methods employ observational data (OBS) combined with limited-sized randomized controlled trial (RCT) data to eliminate confounding bias. However, existing methods typically assume that the OBS and RCT data come from the same target population, a relatively strong assumption given the difficulties of randomized trials. In this paper, we consider relaxing this assumption to achieve data fusion in the case where the RCT data is a biased sample of the target population, thus eliminating selection bias and obtaining unbiased estimates of causal effects. We propose a reweighting-based approach that uses OBS and RCT data successively and debiases in the second stage via reweighting. Extensive experiments are conducted to demonstrate the effectiveness of our method.
Submission Number: 34
Loading