Recurrent Continuous Adversarial Balancing for Estimating Individual Treatment Effect Over Time

Abstract

Estimating individual treatment effects (ITE) from observational data have become an important topic in various fields. In healthcare, determining the optimal time and amount of the treatment is crucial for improving each patient’s quality of life. However, existing methods mainly focused on estimating treatment effect under static setting with categorical treatments. In this work, we proposed Recurrent Continuous Adversarial Balancing (RCAB), a method that incorporates adversarial learning into a sequence-to-sequence recurrent network to reduce selection bias in dynamic data. Moreover, we reformulated imbalance loss to be appropriate for continuous scenario. RCAB is empirically validated in both synthetic and real-world datasets. The results demonstrate that RCAB outperforms various state-of-the-art methods, and it is applicable to time-series data with any types of treatments.