Abstract

We consider the problem of inferring the causal effect of an exposure variable X on an outcome variable Y. Besides X and Y, a high-dimensional covariate V is also measured. It is assumed that confounding variables that may cause bias in the desired causal inference are low-dimensional features of V. Under this assumption, we propose an encoding generative modeling (EGM) approach for the estimation of the average dose response function, a function that captures, in an average sense, the dependency of Y on X when confounders were held fixed. We show that EGM provides a framework for us to develop deep learning-based estimates for the structural equations that describe the causal relations among the variables. We will present numerical and theoretical evidence to demonstrate the effectiveness of our approach.