Causal Inference Under Approximate Neighborhood Interference

This paper studies causal inference in randomized experiments under network interference. Commonly used models of interference posit that treatments assigned to alters beyond a certain distance from the ego have no effect on ego's response. However, this assumption is violated common social interactions. We propose substantially weaker model “approximate neighborhood interference” (ANI) which further smaller, but potentially nonzero, formally verify ANI holds for well‐known Under ANI, restrictions topology, and asymptotics size increases, we prove standard inverse‐probability weighting estimators consistently estimate useful exposure effects are approximately normal. For inference, consider HAC variance estimator. finite population model, show estimator biased bias can be interpreted as unit‐level effects. generalizes Neyman's result conservative estimation settings with