Internalising and externalising disorders were diagnosed using the World Health Organization Composite International Diagnostic Instrument, version 3.0 (WMH CIDI 3.0) [33]. The WMH CIDI is a fully-structured diagnostic interview which was modified to simplify language and to use examples that are more of relevance to adolescents [31]. The interview covers a wide range of common DSM-IV disorders in adolescents (e.g., mood disorders, anxiety disorders, behaviour disorders, eating disorders, and substance use disorders). In addition to obtaining information on the 12-month diagnosis of internalising and externalising disorders, the present study focused on 15 key symptoms of internalising disorders (Table 1). Following the WMH CIDI guidelines, all diagnoses are assessed using a general screener (which comprises one or more questions related to the key symptoms of the relevant disorder), followed by specific diagnostic items for adolescents who endorsed screening questions. Only one key symptom was selected from the agoraphobia disorder module due to high rate of missing values in the other ones (>50% of participants). Concordance of WMH CIDI and DSM-IV diagnoses was endorsed in Kessler et al. [33].
Table 1 List of CIDI symptoms (items) included in the analyses Information on the adolescent’s sociodemographic characteristics (i.e., gender, age, race, family composition and place of residence) and physical health (i.e., history of neurological, joint, respiratory, metabolic, pain-related and heart diseases; and disabilities) was collected during a face-to-face interview. Information on lifestyle patterns was also collected during the interview, including: eating patterns (regular eating pattern in terms of nutrient intake; vegetarian diet, hypo-caloric diet, other), sleep patterns (total hours of sleep during week nights and weekend nights; difficulty falling asleep which was derived from the item ‘nearly every night it took you a long time to fall asleep’; and difficulty staying asleep which was derived from item ’you woke up nearly every night and took a long time to get back to sleep’), and taking part in physical activity (frequency of light or moderate physical activity, ranging from 1 = ‘several times a week or more’ to 6 = ‘never’; this variable was derived from the item: ‘how often do you engage in light or moderate physical exercise like walking for 30 min or more?’).
The NCS-A included an 11-item scale to measure cognitive and academic competencies [31], which can be rated on a 4-point Likert scale of response (from 1 = ‘excellent’ to 4 = ‘poor’). An exploratory factor analysis (EFA) was conducted on the whole sample (N = 10,123) to examine factor structure underlying the scale. Principal component analysis was used to reduce dimensionality (data reduction), relying on polychoric correlation matrix. Two factors were identified, explaining 51.82% of scale variance. Both factors showed eigenvalues higher than one (factor 1 = 4.52, factor 2 = 1.18, respectively). Seven items saturated on the first factor (emotion and behaviour regulation deficits) and four items saturated on the second factor (academic/work competence deficits). The reliability indexes in the present study were satisfactory (ω between 0.74 and 0.75).
NCS-A also included a 20-item scale to measure strategies to cope with stress on a 4-point Likert scale (from 1 = ‘a lot’ to 4 = ‘not at all’) [31]. To make the results more interpretable, the response scale was recoded such that 1 = ‘not at all’ to 4 = ‘a lot’. The EFA conducted in the present study revealed a 4-factor structure, explaining 49.64% of scale variance. All these factors showed eigenvalues higher than one (from 4.07, factor 1, to 1.25, factor 4). The factors were: Factor 1 (Emotion-focused coping), Factor 2 (Problem-focused coping), Factor 3 (Cognition-focused coping), and Factor 4 (Self-focused coping). Reliability indexes were satisfactory across factors in our sample (ω between 0.61 and 0.73).
Information on physical and mental health was obtained by asking the adolescents to rate their overall physical (NCS-A item: ‘How would you rate your overall physical health?’) and mental health (NCS-A item: ‘How would you rate your overall mental health?’) on a scale which ranged from 1 (‘excellent’) to 5 (‘poor’). Information on health care service utilisation was also examined, which included: days of hospitalisation for emotional/mental problems in the past year; number of visits to mental health professionals in past year; and number of school counselling services received in the past year.
Data Analysis
Between-group differences were examined using χ2-based tests, as well as Cramer’s V as an effect size estimate. Prediction of self-reported health (physical and mental) and mental health care service utilisation was examined using generalised linear modelling (GLM). Study groups, lifestyle patterns, coping strategies, as well as cognitive and academic competencies were considered as covariates. Physical and mental health outcomes were modelled under gamma distribution; health service utilisation outcomes were modelled under negative binomial distribution, as high proportion of adolescents was expected to have no mental health service use. A lower Akaike information criterion (AIC) was used to inform fit of the model with covariates in comparison to the model without them. Model comparison relied on a stepwise approach: a model without covariates (unconstrained model), a model with a study group as a covariate, and the model with all covariates. Odds ratio (OR) was used to report parameter loadings. A significant difference from one loading was detected by means of Wald’s test under a t-based distribution.
A network analysis (NA) approach [34] was used to examine the relations between internalising symptoms across the six study groups. This approach focuses on the complex patterns of relations between symptoms that underlie a specific mental health condition as well as the relations of these symptoms with symptoms of various other mental health conditions. In the graphical representation of NA, the nodes represent the symptoms and edges between them reflect their conditional dependence/relation (i.e., association between two symptoms after controlling for all other associations between the symptoms in the network). The nodes with stronger correlations are placed near the centre and show their influencing (central) role in keeping the disorders stable [35]. Network was weighted and regularised (under regularised logistic regression framework) by shrinking small connections in the network (set to be exactly zero) due to Holm’s correction for multiple comparison testing. Nested Lasso regressions were used for network estimation, with model selection based on the extended Bayesian information criterion (EBIC) and penalisation based on a gamma hyperparameter (γ = 0.25).
Data of participants in the clinical groups were used for the NA. Those with a high rate of missing values (i.e., ten or more symptoms without response) were excluded. Multiple imputation procedures were used to estimate missing values, with a cut-off point of 10 multiple imputations and 50 iterations to obtain convergence for the solution comprising the imputed values [36]. Estimation method relied on the random forest algorithm. The algorithm is suitable to handle data violating normality assumptions and highly recommended for high dimensionality data (i.e., high correlations between items) [37, 38].
The network comparison test was used to examine the similarity of networks across the study groups [39, 40]. Specifically, the test investigates network invariance at three levels: network structure (i.e., whether the structure of both networks is invariant between groups), global strength (i.e., invariant overall connectivity of symptoms across between groups) and edge strength (i.e., whether each association between symptoms is invariant across groups, using a Bonferroni-Holm correction to prevent from multiple testing bias). Edge strength invariance was tested when the network structure showed no invariant between groups. Two sets of pairwise network comparisons were carried out: First, we compared groups with internalising disorders (first set: MDD vs. ANX, MDD vs. MDD + ANX, ANX vs. MDD + ANX; second set: MDD + EXT vs. ANX + EXT, MDD + EXT vs. MDD + ANX + EXT, ANX + EXT vs. MDD + ANX + EXT). Second, we compared networks of groups with versus without comorbid externalising disorders (MDD vs. MDD + EXT, ANX vs. ANX + EXT, ANX + MDD vs. ANX + MDD + EXT groups). A Bonferroni-based correction on p level was applied to prevent from multiple comparison testing (0.05/6 = 0.0083, for externalising-comorbidity network comparison; and 0.05/3 = 0.0166, for internalising-disorder subtype comparison).
Centrality measures were calculated to examine the role of each symptom within the network across the six groups [35]. Two centrality measures were calculated: strength (i.e., sum of the edge weights connected to a node) and betweenness (i.e., number of times that node lies on the shortest path between two other nodes). Network clustering were measured using the following indicators [41,42,43]: transitivity (i.e., how two nodes which share a neighbour are interconnected within the network; global clustering property of the entire network), the average of shortest paths between nodes, and small worldness index (property to have both a high clustering coefficient and a short average path length; values higher than 1 indicate that the network has the small-world property). Network robustness tests were conducted under non-parametric bootstrapping [44].
Finally, to investigate the association between network properties and self-reported health/and mental health service utilisation, non-parametric correlation analysis (Spearman’s correlation) was conducted between NA properties (network global strength, average of weighted correlations and clustering properties) and the outcome measures (i.e., self-reported physical and mental health, and health care service utilisation).
All the analyses were conducted using R Core Software [45], packages mice, lme4, qgraph, igraph, bootnet and NetworkComparisonTest.