Primary Submission Category: Causal Inference in Networks
Estimating heterogenous spillover effects on network neighbors to identify influential and susceptible individuals
Authors: Yihan Bao, Laura Forastiere,
Presenting Author: Yihan Bao*
Due to peer influence, behavioral interventions received by a unit are likely to affect the behavioral outcomes of other socially connected units. Under interference, spillover effects have been defined in previous works by contrasting potential outcomes under a different number of treated units or under different treatment allocations in the interference set. In our work, under a partial interference assumption, we define average and conditional spillover effects of having a network neighbor treated vs not treated, while the treatment of other units in the interference set is randomly assigned under a given allocation strategy. By varying the conditioning sets, we can assess the heterogeneity with respect to the characteristics of the influencer and those of the influencee, to identify influential and susceptible individuals, respectively. Under a super-population perspective, we develop IPW estimators for average and heterogeneous influence effects, with marginal structural models for continuous covariates. We then use our estimators to investigate the characteristics of influencers and susceptible individuals in a two-stage randomized study conducted in Honduras to assess the spillover effects of a behavioral intervention. Here, we further address the presence of non-compliance in the second stage by replacing the theoretical second-stage treatment probability with the estimated propensity score, conditional on the first stage.