Estimating Causal Mediation Effects under Correlated Errors

Abstract

Causal mediation analysis usually requires strong assumptions, such as ignorability of the mediator, which may not hold in many social and scientific studies. Motivated by a multilevel randomized treatment experiment using functional magnetic resonance imaging (fMRI), this paper proposes a multilevel causal mediation framework for data with hierarchically nested structure, and this framework provides valid inference even if structured unmeasured confounding for the mediator and outcome is present. For the first-level data, we propose a linear structural equation model for a continuous mediator and a continuous outcome, both of which may contain correlated additive errors. A likelihood-based approach is proposed to estimate the model coefficients. The analysis of our estimator characterizes the nonidentifiability issue due to the correlation parameter. To address the identifiability issue and model the variability in multilevel data, we propose to incorporate multiple first-level mediation models across different levels in a unified multilevel mediation framework. All the model coefficients are estimated simultaneously by our optimization algorithms, with innovation to estimate the unknown correlation parameter from data, instead of performing sensitivity analysis. Our asymptotic analysis shows that the correlation parameter is identifiable, and our estimates for the mediation effects are consistent with the parametric convergence rates. Using extensive simulated data and a real fMRI dataset, we demonstrate the improvement of our approaches over existing methods.