The Impact of Criteria-based and Data-driven Sampling Approaches on the Heterogeneity and Interpretability of Posttraumatic Stress Symptom Networks
Background. The application of psychopathological symptom networks requires reconciliation of the observed cross-sample heterogeneity. We leveraged the largest sample to be used in a PTSD network analysis (N = 28,594) to examine the impact of criteria-based and data-driven sampling approaches on the heterogeneity and interpretability of networks.Methods. Severity and diagnostic criteria identified four overlapping subsamples and cluster analysis identified three distinct data-derived profiles. Networks estimated on each subsample were compared to a respective benchmark network at the symptom-relation level by calculating sensitivity, specificity, correlation, and density of the edges. Negative edges were assessed for Berkson’s bias, a source of error that can be induced by threshold samples on severity.Results. Criteria-based networks showed reduced sensitivity, specificity, and density but edges remained highly correlated and a meaningfully higher proportion of negative edges was not observed relative to the benchmark network of all cases. Among the data-derived profile networks, the Low Severity network had the highest proportion of negative edges not present in the benchmark network of symptomatic cases. The High Severity profile also showed a higher proportion of negative edges, whereas the Medium Severity profile did not. Conclusion. Although networks showed differences, Berkson’s bias did not appear to be a meaningful source of systematic error. These results can guide expectations about the generalizability of symptom networks across samples that vary in their ranges of severity. Future work should continue to explore whether network heterogeneity is reflective of meaningful and interpretable differences in the symptom relations from which they are composed.