Causal Message Passing: A Method for Experiments with Unknown and General Network Interference

Analysis of high-dimensional data within networks of interacting units has increasingly become a fundamental challenge for causal inference tasks, which aim to assess the efficacy of new policies or products. The gold standard approach commonly advocates for randomized experiments, which partition the target population into treatment and control groups. This method addresses the inherent limitations of observing solely the implemented treatment outcomes, while not having access to the counterfactual scenarios. Through randomized experiments, one can measure causal effects by comparing outcomes between treatment and control groups, yielding an unbiased estimate under the key assumption known as the Stable Unit Treatment Value Assumption (SUTVA) (Cox 1958, Rubin 1978). According to the SUTVA, it is assumed that the outcome of a unit is independent of the treatment assignments of other units. While this assumption is reasonable in certain settings, it often breaks down in real-world scenarios where units interact (Sussman and Airoldi 2017). Consider a study aiming to assess the effectiveness of a new medication for a contagious disease: the treatment group receives the medication while the control group is given a placebo. In such settings, accounting for the potential network effects among units (individuals in this context) is crucial.