Continuous treatment effect estimation via generative adversarial de-confounding

One fundamental problem in causal inference is the treatment effect estimation in observational studies, and its key challenge is to handle the confounding bias induced by the associations between covariates and treatment variable. In this paper, we study the problem of effect estimation on continuous treatment from observational data, going beyond previous work on binary treatments. Previous work on binary treatment focuses on de-confounding by balancing the distribution of covariates between the treated and control groups with either propensity score or confounder balancing techniques. In the continuous setting, those methods would fail as we can hardly evaluate the distribution of covariates under each treatment status. To tackle the case of continuous treatments, we propose a novel Generative Adversarial De-confounding (GAD) algorithm to eliminate the associations between covariates and treatment variable with two main steps: (1) generating an “calibration” distribution without associations between covariates and treatment by randomly perturbation on treatment variable; (2) learning sample weights that transfer the distribution of observed data to the “calibration” distribution for de-confounding with a Generative Adversarial Network. We show, both theoretically and with empirical experiments, that our GAD algorithm can remove the associations between covariates and treatment, hence, precisely estimating the causal effect of continuous treatment. Extensive experiments on both synthetic and real-world datasets demonstrate that our algorithm outperforms the state-of-the-art methods for effect estimation of continuous treatment with observational data.