Mediation Analysis with Unobserved Mediator-Outcome Confounding

Causal mediation analysis aims to decompose total causal effects into direct and indirect causal effects. But whereas randomization of an exposure identifies the total causal effect, it does not identify direct or indirect effects in the presence of unobserved mediator-outcome confounding. We propose a new method for the identification of direct and indirect effects in the presence of unobserved mediator-outcome confounding in linear models.