Treatment effect estimation with disentangled latent factors

A common challenge of many scientific studies is to determine whether a treatment is effective for an outcome. When considering a binary treatment, this problem can be addressed by estimating the average treatment effect using the potential outcome framework. Moreover, since different individuals often respond differently to the same treatment due to their distinct characteristics. In order to understand the heterogeneous treatment effect for different individuals, practitioners need to estimate the conditional average treatment effects conditioning on the variables describing the distinct characteristics of individuals. Much research has been devoted to the estimation of treatment effects from observational data; however, most of them assume that the set of observed variables contains exactly all the confounders that affect both the treatment and the outcome. Unfortunately, this assumption is frequently violated in real-world applications not only because some of the observed variables only affect the treatment or the outcome, but also due to the fact that in many cases only the proxy variables of the underlying confounding factors can be observed. In this work, we first show the importance of differentiating confounding factors from instrumental and risk factors for average and conditional average treatment effect estimation, and then we propose a variational inference approach to simultaneously infer latent factors from the observed variables and disentangle the factors into three disjoint sets corresponding to the instrumental, confounding, and risk factors. Experimental results demonstrate the effectiveness of the proposed method on synthetic, benchmark, and real-world datasets for treatment effect estimation.