Study Sample
Data were collected as part of the longitudinal multicenter Korean Early Psychosis Study (KEPS), which has been described in detail elsewhere.22 The sample comprised 500 patients with early-stage psychosis and 202 healthy controls. The inclusion criteria required that subjects be between 19 and 58 years of age and meet the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV)23 criteria for schizophrenia spectrum disorders (SZ, schizoaffective disorder, schizophreniform disorder, psychotic disorder not otherwise specified [NOS]), brief psychotic disorder, or delusional disorder. Individuals who had been treated with antipsychotics for < 2 years were considered to be in early-stage psychosis. Written informed consents were obtained from all the participants. All experimental protocols were approved by the Ethics Committee of the Chonbuk National University Hospital (approval number CUH 2014-11-002). All procedures were performed in accordance with relevant guidelines.
Measures
The severity of psychiatric symptoms was assessed using the Positive and Negative Syndrome Scale (PANSS).24 For self-rating scales, the Brief Core Schema Scale (BCSS),15 Brooding Scale (BS),25 Early Trauma Inventory Self Report-Short Form (ETI),26 Dietary Habits Questionnaire (DHQ)27 and Physical Activity Rating (PAR)28 were employed.The BCSS consist of four subscales: negative self (NS), positive self (PS), negative others (NO), and positive others (PO). The DHQ is a 20-item self-administered questionnaire consisting of three subcategories: five items for diet regularity, six items for balanced diet, and nine items for unhealthy diet and eating habits. This scale was developed based on dietary guidance published by the Korean Ministry for Health, Welfare and Family Affairs (2010).29 The total score is categorized as indicating poor (20-49), usual (50-79), or good (80-100) diet. The PAR is a questionnaire that rates the individual’s level of physical activity, with scores ranging from 0 (avoids walking or exercise) to 7 (runs more than 10 miles per week or spends more than 3 hours per week in comparable physical activity). As all scores for each parameter exhibited skewed distributions based on the Shapiro–Wilk test, they were normalized using nonparanormal transformation.30,31
Network estimation
Networks were constructed using 12 nodes: NS, PS, NO, and PO from the BCSS; Em and Co from the BS; DHQ; PAR; and GT, EMO, PHY, and SEXU from the ETI). We fitted a Gaussian graphical model (GGM) to the data. The GGM networks were regularized via a graphical lasso (GLASSO) algorithm32 in combination with the extended Bayesian information criterion (EBIC) model. A tuning hyperparameter γ for the EBIC was set to 0.5.33 The edges were calculated by partial correlations. We used the R-packages ‘bootnet (estimateNetwork)’ and ‘qgraph’ to estimate and visualize all networks.34
Network analysis
Global network metrics
Global network metrics consisting of network density, global strength, averaged clustering coefficient, modularity index (Q), and characteristic path length were calculated using the R packages ‘qgraph’ and ‘igraph’.
Local network metrics
Although strength is regarded as the most reliably estimated centrality index it does not necessarily indicate the degree to which a node can be predicted by the remaining intranetwork nodes. To examine node predictability, we estimated the proportion of each node’s variance accounted for by its connections to other nodes in the network, using the ‘mgm’package (Version 1.2-2). In addition, as strength centrality uses the sum of absolute weights, whether positive or negative, which might distort interpretation, we estimated expected influence (EI), i.e., the sum of all edges of a node.35 To detect symptoms that bridged the two domains (ChT and negative life style) or three domains (ChT, negative life style and P and N on the PANSS), bridge EI was calculated. Bridge EI is the sum of the values (+ or −) of all edges that connect a node to all nodes that are not part of the same community.36 Bridge symptoms that play a primary role in connecting two or more psychiatric symptoms or domains37 were defined as those items scoring higher than the 80th percentile for the bridge EI metric. We also computed the shortest pathways38 from each subscale of the ETI to negative life style or to P and N within the network. To determine the EI, bridge EI, and shortest pathway, the R-packages ‘mgm’, ‘qgraph’, ‘networktools’,39 and ‘igraph’ were used, respectively.
Network comparison
We investigated network structures and global strength using the Network Comparison Test (NCT) in the R package. For global network metrics (network density, averaged clustering coefficient, modularity index [Q], and characteristic path length), the ‘NetworkToolbox’ packagewas used to explore whether the overall level of network connectivity was equal among the networks.
Network accuracy and stability
The accuracy and stability of the network were examined using the R package ‘bootnet’, version 1.4.2. First, we bootstrapped (1,000 iterations) the 95% confidence intervals around the edge weights to assess the accuracy of the edge weights. Second, we used the case-dropping subset bootstrap (1,000 iterations) to examine the stability of the order of the node centrality indices. A correlation stability coefficient (CS-coefficient), a measure that quantifies the stability of node centrality indices, was also calculated. Finally, we tested for significant differences in edge weights and node centralities using the bootstrapped difference tests.