Proximal Causal Learning

Abstract Observational studies represent a crucial application domain for causal inference. In this context, the validity of traditional causal inference methods typically relies on the assumption of no unmeasured confounders; however, the validity of this assumption often cannot be ascertained through statistical testing, leading to the unavoidable potential skepticism. The recently proposed proximal causal inference framework explicitly posits conditional independence relationships involving unmeasured confounders and effectively addresses this challenge through the strategic application of proxy variables. In this article, we review proximal causal inference methods in both point exposure and longitudinal settings. As the estimation of bridge functions is crucial for proximal causal learning, we also present various parametric and nonparametric estimation approaches. Finally, we briefly discuss potential future directions for proximal causal inference.