Bayesian Structured Mediation analysis with Unobserved confounders

Area of Science:

• Neuroimaging
• Biostatistics
• Causal Inference

Background:

• Causal mediation analysis is crucial for understanding complex relationships.
• High-dimensional mediators, like brain imaging data, present challenges due to unobserved confounders.
• Spatial structures in mediators can be influenced by unobserved subject-specific confounding.

Purpose of the Study:

• To develop a robust method for causal mediation analysis with high-dimensional, spatially structured mediators.
• To address the impact of unobserved confounders in mediation analysis.
• To improve the accuracy of estimating Natural Indirect Effects (NIE) and Natural Direct Effects (NDE).

Main Methods:

• Developed the BAyesian Structured Mediation analysis with Unobserved confounders (BASMU) framework.
• Incorporated spatial latent subject-specific confounding effects into the outcome model.
• Proposed a two-stage estimation algorithm for bias mitigation.
• Established model identifiability conditions for BASMU.

Main Results:

• Theoretical analysis confirmed reduced asymptotic bias in NIE and NDE estimation.
• Extensive simulations demonstrated BASMU's substantial bias reduction across scenarios.
• Application to fMRI data identified 2-4 times more significant mediation voxels compared to existing methods.
• NIE increased by 41% and NDE decreased by 26% in the fMRI analysis.

Conclusions:

• BASMU effectively reduces bias caused by unobserved confounders in mediation analysis of structured, high-dimensional data.
• The framework enhances the identification of significant mediation effects in neuroimaging studies.
• BASMU offers a more accurate approach to estimating direct and indirect effects in complex biological systems.