An example of identifying conditional dependence in graphical models using Bayesian model averaging: general and disease specific anxiety in a cardiovascular patient sample.

Parsing the interplay of psychological symptoms is difficult yet crucial for both disorder conceptualization and development of targeted interventions. This study used a Bayesian Model Averaging (BMA) approach to network analysis to identify the most probable structures and to evaluate conditional dependence underlying symptomology in 110 cardiovascular patients diagnosed with a Thoracic Aortic Aneurysm (TAA). Symptoms of general and disease-specific distress were assessed using the Depression Anxiety Stress Scales (DASS-21) and the Cardiac Anxiety Questionnaire-Attention Subscale (CAQ-A). Bayesian Model Averaging (BMA) and Bayes Factors (BF) were used to inform model selection and quantify model uncertainty. We present both the psychological symptom network with the highest posterior probability—and thus the best predictive utility and the BMA network, which produces tests of conditional dependence and independence which average over all structures and thus are not conditional on the other variables present in a single selected reference model. However, the posterior distribution showed variability in plausible complexity, or the number of parameters in the network structure. Utilizing BMA is especially protective in cases of structural uncertainty and high complexity. Of the 171 possible edges, 19 edges had strong evidence for inclusion. Notably, two edges connected the general and cardiac anxiety scales: DASS_7 (worry about experiencing panic) to CAQ_3 (chest discomfort wakes me up at night) and DASS_9 (Aware of Heart Symptoms) to CAQ_2 (racing heart wakes me up at night). This emphasizes the prevalent link between physiological components contributing to both general and cardiac anxiety. This approach highlights the utility of these methods in revealing selective and specific symptom interactions and offer a set of plausible causal associations, which could inform future studies (and Directed Acyclic Graphs) oriented towards testing causal mechanisms.