Towards Principled Causal Effect Estimation by Deep Identifiable Models

As an important problem of causal inference, we discuss the estimation of treatment effects (TEs) under unobserved confounding. 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 representation balanced for treatment groups, using its prior. Experiments on (semi-)synthetic datasets show state-of-the-art performance under diverse settings. Based on the identifiability of our model, further theoretical developments on identification and consistent estimation are also discussed. This paves the way towards principled causal effect estimation by deep neural networks.