GDR-learners: Orthogonal Learning of Generative Models for Potential Outcomes

Valentyn Melnychuk, Stefan Feuerriegel

Keywords: causal machine learning, orthogonal learning, deep generative models, potential outcomes estimation

TL;DR: we introduce a general framework of generative Neyman-orthogonal (doubly-robust) meta-learners that target at the covariate-conditional distributions of potential outcomes (GDR-learners)

Abstract: Various deep generative models have been proposed to estimate potential outcomes distributions from observational data. However, none of them have the favorable theoretical property of general Neyman-orthogonality and, associated with it, quasi-oracle efficiency and double robustness. In this paper, we introduce a general suite of generative Neyman-orthogonal (doubly-robust) learners that estimate the conditional distributions of potential outcomes. Our proposed generative doubly-robust learners (GDR-learners) are flexible and can be instantiated with many state-of-the-art deep generative models. In particular, we develop GDR-learners based on (a) conditional normalizing flows (which we call GDR-CNFs), (b) conditional generative adversarial networks (GDR-CGANs), (c) conditional variational autoencoders (GDR-CVAEs), and (d) conditional diffusion models (GDR-CDMs). Unlike the existing methods, our GDR-learners possess the properties of quasi-oracle efficiency and rate double robustness, and are thus asymptotically optimal. In a series of (semi-)synthetic experiments, we demonstrate that our GDR-learners are very effective and outperform the existing methods in estimating the conditional distributions of potential outcomes.

Primary Area: causal reasoning

Submission Number: 11180