Use of Instrumental Variables in Mediation Analysis
We investigate the identification power of a binary instrumental variable (IV) in a setting with a binary treatment and a binary mediator. We show that while IV methods allow for the existence of unobserved confounders, additional assumptions are required to recover causal parameters. We consider two additional assumptions: effect homogeneity and partial monotonicity. A simple test for evaluating the unobserved effect heterogeneity is proposed