Formalizing the role of complex systems approaches in causal inference and epidemiology

There have been several calls for the adoption of complex systems approaches in epidemiology and public health. This work has largely centered on the potential for such methods to examine disease etiologies with a high degree of complexity. In this paper we make an explicit effort to reconcile complex systems methods with modern thinking in causal inference, and show that these approaches should rely on counterfactual outcome frameworks to define complex causal “webs”. Further, systems science methods can be formalized in a manner analogous to those developed for current analytic techniques to achieve causal inference. Using as an example an agent-based model constructed to represent HIV transmission in a dynamically evolving network, we demonstrate that complex systems models can be used to simulate counterfactual outcomes, providing an alternative technique to stratification-based and G-methods approaches for determining disease causes. Thus, complex systems methods (and simulation approaches broadly) represent a methodological bridge between studies that produce estimates of average causal effects and unobserved counterfactual outcomes. We show that these models are of particular utility when the hypothesized causal mechanisms are of sufficient complexity such that the assumptions of modern empirical-based methods (e.g., marginal structural models) cannot be met. Finally, we will describe the set of assumptions that must be satisfied to ensure that the results of complex systems models represent average causal effects. Although not without challenges, complex systems methods represent a promising set of novel approaches to identify and evaluate causal effects, and are thus well suited to complement other modern epidemiologic methods of etiologic inquiry.