Towards Principled Causal Effect Estimation by Deep Identifiable Models

TL;DR

This paper introduces Intact-VAE, a novel deep generative model for causal effect estimation that leverages latent confounder representations, achieving state-of-the-art results and theoretical identifiability under unconfoundedness.

Contribution

It proposes Intact-VAE, a variational autoencoder variant that ensures identifiable and balanced confounder representations for causal inference.

Findings

State-of-the-art performance on synthetic datasets

Proves identifiability of treatment effects under unconfoundedness

Demonstrates robustness in settings with unobserved confounding

Abstract

As an important problem in causal inference, we discuss the estimation of treatment effects (TEs). Representing the confounder as a latent variable, we propose Intact-VAE, a new variant of variational autoencoder (VAE), motivated by the prognostic score that is sufficient for identifying TEs. Our VAE also naturally gives representations balanced for treatment groups, using its prior. Experiments on (semi-)synthetic datasets show state-of-the-art performance under diverse settings, including unobserved confounding. Based on the identifiability of our model, we prove identification of TEs under unconfoundedness, and also discuss (possible) extensions to harder settings.