Mediation analysis with graph mediator.

Area of Science:

• Statistics
• Neuroscience
• Graph Theory

Background:

• Mediation analysis is crucial for understanding indirect effects in complex systems.
• Representing mediators as graphs, particularly covariance matrices, presents unique analytical challenges.
• Existing methods often struggle with high-dimensional graph data and simultaneous parameter estimation.

Purpose of the Study:

• To introduce a novel mediation analysis framework where the mediator is a graph, specifically a Gaussian covariance graph.
• To develop methods for estimating causal parameters and low-rank representations of covariance matrices simultaneously.
• To apply this framework to neuroimaging data, investigating functional connectivity as a mediator.

Main Methods:

• A Gaussian covariance graph model is assumed for the graph mediator.
• A low-rank representation is utilized for the covariance matrix mediator.
• Parametric mediation models are employed within the structural equation modeling framework.
• Likelihood-based estimators are derived for simultaneous identification of low-rank structure and causal parameters, assuming Gaussian errors.
• An efficient computational algorithm and asymptotic properties of estimators are investigated.

Main Results:

• Simulation studies demonstrate the effectiveness of the proposed mediation analysis approach.
• The framework successfully identified a brain network mediating sex differences in motor task performance.
• Functional connectivity within the identified network was shown to mediate the observed sex difference.

Conclusions:

• The proposed framework provides a robust method for mediation analysis with graph-valued mediators.
• This approach enables the simultaneous estimation of network structure and causal mediation effects.
• The application to fMRI data highlights the utility of graph mediation analysis in neuroscience for understanding complex relationships.