Multilevel mediation analysis with structured unmeasured mediator-outcome confounding

Abstract

Mediation analysis usually requires the assumption that there is no unmeasured mediator-outcome confounder. However, this may not hold in many social and scientific studies. Though various parametric and nonparametric mediation methods have been developed, this assumption remains instrumental, without which the causal effects are not identifiable unless alternative assumptions are imposed. To circumvent this, a multilevel parametric structural equation modeling framework is proposed to relax this no unmeasured mediator-outcome confounding assumption under a specific data setting inspired by a real experiment. Using the proposed framework, it is shown that the causal effects are identifiable and consistently estimated. Likelihood-based approaches are proposed with efficient optimization algorithms to estimate the parameters, including the unmeasured confounding effect, instead of performing sensitivity analysis. The asymptotic consistency is established. Using extensive simulations and a functional magnetic resonance imaging dataset, the improvement of the approaches over existing methods is demonstrated. The R package macc for implementation is available on CRAN.