Comparing Gaussian graphical models with the posterior predictive distribution and Bayesian model selection

Gaussian graphical models are commonly used to characterize conditional (in)dependence structures (i.e., partial correlation networks) of psychological constructs. Recently attention has shifted from estimating single networks to those from various subpopulations. The focus is primarily to detect differences or demonstrate replicability. We introduce two novel Bayesian methods for comparing networks that explicitly address these aims. The first is based on the posterior predictive distribution, with a symmetric version of Kullback-Leibler divergence as the discrepancy measure, that tests differences between two (or more) multivariate normal distributions. The second approach makes use of Bayesian model comparison, with the Bayes factor, and allows for gaining evidence for invariant network structures. This overcomes limitations of current approaches in the literature that use classical hypothesis testing, where it is only possible to determine whether groups are significantly different from each other. With simulation we show the posterior predictive method is approximately calibrated under the null hypothesis (alpha = .05) and has more power to detect differences than alternative approaches. We then examine the necessary sample sizes for detecting invariant network structures with Bayesian hypothesis testing, in addition to how this is influenced by the choice of prior distribution. The methods are applied to posttraumatic stress disorder symptoms that were measured in 4 groups. We end by summarizing our major contribution, that is proposing 2 novel methods for comparing Gaussian graphical models (GGMs), which extends beyond the social-behavioral sciences. The methods have been implemented in the R package BGGM.

Translational Abstract

Gaussian graphical models are becoming popular in the social-behavioral sciences. Recently attention has shifted from estimating single networks to those from various subpopulations (e.g., males vs. females). We introduce Bayesian methodology for comparing networks estimated from any number of groups. The first approach is based on the posterior predictive distribution and it allows for determining whether networks are different from one another. This is ideal for testing the null hypothesis of group equality, say, in the context of testing for network replicability (or lack thereof). The second approach is based on Bayesian hypothesis testing and it allows for gaining evidence for network invariances or equality of partial correlations for any number of groups. This is ideal for focusing on specific aspects of the network such as individual partial correlations. In a series of simulations and illustrative examples we demonstrate the utility of the proposed methodology for comparing Gaussian graphical models. The methods have been implemented in the R package BGGM.