Treatment Effect Estimation from Observational Network Data using Augmented Inverse Probability Weighting and Machine Learning

Abstract

Causal inference methods for treatment effect estimation usually assume independent experimental units. However, this assumption is often questionable because experimental units may interact. We develop augmented inverse probability weighting (AIPW) for estimation and inference of causal treatment effects on dependent observational data. Our framework covers very general cases of spillover effects induced by units interacting in networks. We use plugin machine learning to estimate infinite-dimensional nuisance components leading to a consistent treatment effect estimator that converges at the parametric rate and asymptotically follows a Gaussian distribution.

Unconfoundedness with Network Interference

“…, which are of the same type as those in the simulations of Emmenegger et al (2022) and Forastiere et al (2021). For these, we use generalized linear models (GLMs), meaning logistic and linear regression, with polynomial sieves of order 1, 2, and 3.…”

“…1.1 Unconfoundedness Emmenegger et al (2022), Forastiere et al (2021), and Ogburn et al (2022) study estimation of (2) under the unconfoundedness condition…”

“…This does not require the econometrician to specify a low-dimensional function T i of pD, Aq to restrict interference, as in (1), or a low-dimensional function W i of pX, Aq that summarizes confounding, as in (3). Forastiere et al (2021) and Sánchez-Becerra ( 2022) utilize semiparametric models of potential outcomes and/or treatment selection, while Emmenegger et al (2022) take a nonparametric approach. We also study nonparametric estimation and, like Emmenegger et al (2022), utilize a doubly robust estimator.…”

“…Forastiere et al (2021) and Sánchez-Becerra ( 2022) utilize semiparametric models of potential outcomes and/or treatment selection, while Emmenegger et al (2022) take a nonparametric approach. We also study nonparametric estimation and, like Emmenegger et al (2022), utilize a doubly robust estimator. Unlike their paper, we specifically make the case for estimating the unknown functions using graph neural networks (GNNs).…”

“…This would generally preclude consistent estimation of the nonparametric function µ t pi, X, Aq in the doubly robust estimator defined in §2.3. Ours is therefore a superpopulation framework, which is common for observational data and followsEmmenegger et al (2022) andForastiere et al (2021).…”