Estimating Mediation Effects Under Correlated Errors in Big fMRI Data

• Activity Number: 89
• Type: Invited
• Date/Time: Sunday, August 9, 2015 : 8:30 PM to 9:15 PM
• Sponsor: Korean International Statistical Society
• Author(s): Yi Zhao* and Xi Luo
• Companies: Brown University and Brown University
• Keywords: causal inference ; confounding ; functional MRI ; multilevel ; structural equation models

Abstract

Mediation analysis assesses the effect passing through an intermediate variable in a causal pathway from the treatment variable to the outcome variable. Structural equation model (SEM) is a popular approach to estimate mediation effect. However, causal interpretation usually requires strong assumptions, e.g. ignorability of the mediator, which may not hold in neuroimaging analysis due to unmeasured confounding. In this paper, we use mediation analysis in an fMRI experiment to assess the effect of randomized binary stimuli passing through a brain pathway of two brain regions. We propose a two-layer SEM framework that provides valid inference even if unmeasured confounding is present. We propose a constrained optimization approach to estimate the model coefficients, analyze its asymptotic properties and characterize the nonidentifiability issue. To address this issue, we introduce a linear mixed effects SEM to estimate the unknown correlation parameter instead of sensitivity analysis. Three computational approaches are proposed to estimate thousands of parameters. Using extensive simulated data and an fMRI dataset, we demonstrate the improvement of our approach over existing methods.