Bayesian Estimation for Gaussian Graphical Models: Structure Learning, Predictability, and Network Comparisons.

Area of Science:

• Psychology
• Statistics
• Network Analysis

Background:

• Gaussian graphical models (GGMs) estimate conditional dependence structures via partial correlations, typically using frequentist methods with $$-regularization in psychology.
• Bayesian methods are underutilized in psychological network literature for estimation and inference, despite their potential for nuanced analysis.

Purpose of the Study:

• To introduce novel Bayesian methodology tailored for common psychological applications of Gaussian graphical models.
• To provide tools for assessing conditional dependence and independence using posterior probabilities.
• To extend Bayesian inference to within- and between-network comparisons, including partial correlation differences and network predictability.

Main Methods:

• Development of Bayesian estimation and inference techniques for Gaussian graphical models.
• Utilizing posterior probabilities to determine graphical structures and assess conditional relationships.
• Implementation of methods for network predictability and comparison of partial correlations within and between networks.

Main Results:

• Demonstration that the posterior probability-based decision rule can be calibrated for desired specificity.
• Illustrative examples showcasing the application and effectiveness of the proposed Bayesian techniques.
• The developed methods are available in the R package BGGM for practical use.

Conclusions:

• The introduced Bayesian methodology offers a powerful alternative to frequentist approaches for psychological network analysis.
• These methods enhance the estimation and inference capabilities for complex network structures in psychological research.
• The BGGM package facilitates the adoption of these advanced Bayesian techniques by researchers.