Causal inference under uniformly bounded neighbourhood interference

Abstract Randomized experiments remain the gold standard for estimating treatment effects; however, network interference compromises the validity of traditional estimators by violating the stable unit treatment value assumption and introducing bias. Although cluster-randomized designs help mitigate some bias, they struggle to accommodate complex network structures and cannot disentangle direct from indirect effects. To address these challenges, we develop a design-based asymptotic theory for Horvitz–Thompson estimators of the direct, indirect, and total average treatment effects under Bernoulli designs, assuming uniformly bounded neighbourhood interference. Given the poor performance of classical Horvitz–Thompson point and variance estimators in dense networks, we propose eigenvector-adjusted point estimators along with novel variance estimators to improve inference. We establish theoretical guarantees for the proposed estimators within the design-based framework, allowing for robust conclusions that are independent of the stochastic properties of the network or the potential outcome model. The adaptability of the method is showcased under two structured interference scenarios: partial interference and local interference in a two-sided marketplace. Numerical studies further highlight the practical utility of the proposed estimators.