A Comparison of Latent Variable and Psychological Network Models in Mental and Physical Health Symptom Data: Common Output Metrics and Factor Structure
In mental health research, psychological network modeling such as the Gaussian graphical model (GGM) has emerged as an alternative to latent variable modeling such as confirmatory factor analysis (CFA). Recent simulation studies have found that centrality indices from the GGM are partially redundant with factor loadings from a CFA. Follow-up analyses on the GGM, such as exploratory graph analysis (EGA) can sort items into communities that may represent hypothesized factors. However, previous comparisons of centrality indices with factor loadings and the ability of EGA to recover hypothesized factor structure have not been done with real mental and physical health symptom data. We compared GGM and CFA using data based on 16 test forms from Wave 1 of the Patient Reported Outcomes Measurement Information System (PROMIS; N’s = 6,261 to 9,022) designed to measure 9 mental and physical health domains. Using techniques appropriate for handling missing data, we fit a CFA model and a regularized GGM to each test form. We also applied the Walktrap community detection algorithm from EGA. We found weaker correspondence between centrality indices and factor loadings than found by previous research, yet in a similar pattern of correspondence. EGA recommended a factor structure discrepant with PROMIS domains in most cases. Physical Function typically split into two or more clusters; Anger, Anxiety, Depression, and Fatigue often joined as one; and some single-item communities emerged. In real mental and physical health data, strength centrality may offer new information despite being highly related to factor loadings, and EGA provides additional insight on factor and test form composition.