How Predictable are Symptoms in Psychopathological Networks? A Reanalysis of 17 Published Datasets

Abstract

Background Network analysis in psychopathology is an emerging field. Previous publications limited their interpretation largely on the centrality (inter-connectedness) of nodes in network structures, where nodes with the highest centrality may be the best targets for interventions. However, centrality is a relative metric: the most central node could be indeed the best target for an intervention but still be a very bad target. This is the case if it shares very little variance with other nodes, and thus intervening on the node will only have a negligible impact on the network structure. Here we provide a solution for this problem by introducing the absolute metric predictability - the extent to which a node is determined by all its neighbors in the network - and estimate the predictability of all nodes in 17 prior empirical network papers on psychopathology. Methods We carried out a literature review and collected 24 datasets from 17 published papers in the field (several mood and disorders, substance abuse, and psychosis). We fit state-of-the-art regularized partial correlation networks to all datasets, and computed the predictability of all nodes. Results Predictability is moderately high in most symptom networks, and differs considerable both within and between datasets. Conclusions Predictability is an important characterization of symptom networks in addition to their structure. It provides a measure of practical relevance of connections among nodes that may be helpful to inform intervention strategies, and allows conclusions about how self-determined a symptom network is. Limitations of predictability along with future directions are discussed.