Primary Submission Category: Machine Learning and Causal Inference
Taming Individualized Treatment Effects under Interference via Representation Learning with Graph Neural Networks
Authors: Mauricio Tec, Claudio Battiloro,
Presenting Author: Mauricio Tec*
Causal inference is used to estimate the impact of a treatment on a specific outcome. A widely accepted assumption is that one unit’s treatment does not impact other units’ potential outcomes, which can be violated in practical settings and lead to biased estimates. While increasing research focuses on this issue, many studies have concentrated on average direct and spillover effects. Instead, this study considers individualized treatment effects (ITE) under interference. To this end, we propose and evaluate graph neural network (GNN) methods with the minimal assumption that the potential outcomes can be written as a symmetric function of the interfering units’ treatments and covariates. We formulate our learning task using the general framework of bipartite interference, which contains standard network interference as a special case. We formalize the estimand of interest and analyze the counterfactual generalization error based on distributional shift. Our experiments with synthetic data evaluate architectures and representation learning methods previously proposed for ITE estimation (under no interference). These results provide directions for designing GNN-based estimators of ITEs under interference. In addition, we demonstrate the utility of our approach in an application to estimate the health effects of implementing an intervention to reduce emissions in selected US coal-fired power plants.