Network-Assisted Mediation Analysis with High-Dimensional Neuroimaging Mediators

Area of Science:

• Neuroscience
• Biostatistics
• Developmental Psychology

Background:

• Mediation analysis estimates causal pathways via intermediate variables (mediators).
• High-dimensional, correlated biomarkers (e.g., neuroimaging) challenge standard mediation methods.
• Biomarkers like brain regions exhibit network structures that can enhance mediation analysis.

Purpose of the Study:

• To investigate how brain cortical thickness mediates the effect of maternal smoking on children's cognitive abilities.
• To develop a network-assisted mediation analysis approach leveraging hierarchical structures of neuroimaging mediators.
• To address challenges posed by high-dimensional correlated mediators in mediation analysis.

Main Methods:

• Proposed a network-assisted mediation analysis approach using conditional Gaussian graphical models.
• Leveraged the star-shaped hierarchical network structure of brain cortical thickness.
• Decomposed the joint indirect effect into effects through hub and leaf mediators.

Main Results:

• Identified a specific brain region as a significant leaf mediator.
• The proposed method successfully accounted for the star-shaped network structure of neuroimaging mediators.
• Enabled individual identification and evaluation of indirect effects through leaf mediators after accounting for hub mediators.

Conclusions:

• The network-assisted approach effectively analyzes mediation with high-dimensional, structured neuroimaging data.
• This method provides novel insights into mediator-specific pathways, crucial for intervention design.
• Identified a significant leaf mediator previously undiscoverable by existing methods.