Inferring individual direct causal effects under heterogeneous peer influence

Shishir Adhikari, Elena Zheleva

Abstract: Causal inference is central to understanding the effectiveness of policies and designing personalized interventions. Causal inference involves estimating the causal effects of treatments on outcomes of interest after modeling appropriate assumptions. Most causal inference approaches assume that a unit’s outcome is independent of the treatments or outcomes of other units. However, this assumption is unrealistic when inferring causal effects in networks where a unit’s outcome can be influenced by the treatments and outcomes of its neighboring nodes, a phenomenon known as interference. Causal inference in networks should explicitly account for interference. In interference settings, the direct causal effect measures the impact of the unit’s own treatment while controlling for the treatments of peers. Existing solutions to estimating direct causal effects under interference consider either homogeneous influence from peers or specific heterogeneous influence mechanisms (e.g., based on local neighborhood structure). In this work, we define heterogeneous peer influence (HPI) as the general interference that occurs when a unit’s outcome may be influenced differently by different peers based on their attributes and relationships, or when each network node may have a different susceptibility to peer influence. This paper presents IDE-Net, a framework for estimating individual, i.e., unit-level, direct causal effects in the presence of HPI where the mechanism of influence is not known a priori. We first propose a structural causal model for networks that can capture different possible assumptions about network structure, interference conditions, and causal dependence and that enables reasoning about causal effect identifiability and discovery of potential heterogeneous contexts. We then propose a novel graph neural network-based estimator to estimate individual direct causal effects. We show empirically that state-of-the-art methods for individual direct effect estimation produce biased results in the presence of HPI, and that our proposed estimator is robust.