Causal Inference with Misspecified Network Interference Structure
DANIEL NEVO – TEL AVIV UNIVERSITY
ABSTRACT
A prominent assumption when studying treatment effects is the no-interference assumption, stating that treatment applied to one unit does not impact other units. However, interference — an umbrella term for spillover, contagion, peer effects, and related phenomena — is present in many settings. Relaxing the no-interference assumption is often accompanied by an assumed interference structure, commonly represented by a network. Various methods have been developed to address network interference under design-based, frequentist, or Bayesian perspectives.
A key assumption shared by many recently developed methods is that the network is given and correctly specified. We first discuss why such an assumption might be violated in practice. Then, we will present the implications of violations of these assumptions and offer some solutions. To this end, we first focus on a design-based approach and derive bounds on the bias arising from estimating causal effects using a misspecified network, showing how the estimation bias grows with the divergence between the assumed and true networks, quantified through their induced exposure probabilities. To address this challenge, we propose a novel estimator that leverages multiple networks simultaneously and remains unbiased if at least one of the networks is correct, even when we do not know which one. If time permits, we will also discuss alternative solutions to related problems under varied probabilistic regimes.