Statistical inference of long-term causal effects in multiagent systems under the Neyman-Rubin model

Estimation of causal effects of interventions in dynamical systems of interacting agents is under-developed. In this paper, we explore the intricacies of this problem through standard approaches, and demonstrate the need for more appropriate methods. Working under the Neyman-Rubin causal model, we proceed to develop a causal inference method and we explicate the stability assumptions that are necessary for valid causal inference. Our method consists of a temporal component that models the evolution of behaviors that agents adopt over time, and a behavioral component that models the distribution of agent actions conditional on adopted behaviors. This allows the imputation of long-term estimates of quantities of interest, and thus the estimation of long-term causal effects of interventions. We demonstrate our method on a dataset from behavioral game theory, and discuss open problems to stimulate future research.