Primary Submission Category: Causal Inference and SUTVA/Consistencies Violations
Estimating Higher-order Spillover Effects on Network
Authors: Qixiang Xu, Laura Forastiere,
Presenting Author: Qixiang Xu*
“Interference, where a unit’s outcome is affected by the other units’ treatments through network connections, is a phenomenon of interest. Researchers often focus on the spillover effect from first-order neighbors. However, the prevailing approach often involves the neighborhood interference assumption, which can be restrictive. This paper proposes a broader assumption, the generalized interference assumption, which allows potential outcomes to be influenced by a wider range of networks, referred to as the ’interference set’. This might include a community detected through an algorithm, or units that can be reached through a finite network path. We define new causal estimands to quantify spillover effects from units at a specific network distance h. We employ two hypothetical Bernoulli distributions with different probabilities for the h-order neighborhood for the rest of the units in the interference set. We first derive the bias of a commonly used ‘naive’ estimator which relies on a wrong interference set or incorrect exposure mapping functions. We then develop new Horvitz-Thomson and Hajek estimators and corresponding weighted regression estimators under this broader assumption. We assess the bias of naive estimators and the performance of our estimators in different interference scenarios and random graphs through simulations. Finally, we apply these estimators in a two-stage randomized trial in Honduras, evaluating a maternal and child health intervention.