Participants and Procedures
Participants were 2,192 Kenyan adolescents recruited from four secondary schools in Nairobi and Kiambu Counties, as part of a large-scale clinical trial (called Shamiri; ). The trial is registered in the Pan African Clinical Trials Registry (PACTR201906525818462). Participants provided self-report data on depression, anxiety, psychological wellbeing measures, and other health-related and socio-demographic variables by completing a baseline questionnaire battery in their classrooms. There was a slight female majority (N = 1,246; 58.3%), and the mean age was 15.21 (SD = 1.14). See Table 1 for full demographic information and descriptive statistics. Using a clinical cutoff of 10—per clinical guidelines from North American samples [31,32] that have been used previously with Kenyan youths [33,34]—, some 33.81% of participants (N = 741) met the criteria for moderate-to-severe depression and 31.30% (N = 686) met that of moderate-to-severe anxiety. All study procedures were approved by the Maseno University Ethics Review Committee (MUERC, No. MUERC/727/19) before the start of data collection. All adolescents were eligible to participate if they consented and were enrolled in the participating schools. Parental consent and written informed consent/assent were obtained for all adolescents per research ethics procedures at MUERC. Data used for the present study is stored in the Open Science Framework repository and is publicly available (DOI: 10.17605/OSF.IO/8M5D9).
Adolescent psychopathology measures
Depressive symptoms were assessed using the 8-item version of the Patient Health Questionnaire (PHQ-9; [31,35]), a self-report questionnaire used to assess the severity of depressive symptoms. PHQ-8 and PHQ-9 scores are highly correlated, and the same cutoffs can be used to assess for depression severity . A previous study has documented that the PHQ-8 demonstrated adequate psychometric properties for the PHQ-8 with Kenyan adolescents  and adults [37,38]. Anxiety symptoms were assessed using a 7-item self-report questionnaire, the Generalized Anxiety Disorder Screener (GAD-7; ). Like with the PHQ-8, a previous has documented adequate psychometric properties with Kenyan youths .
Psychological wellbeing measures
Happiness and optimism were assessed using two sub-scales of the EPOCH Measure of Adolescent of Well-Being (EPOCH; ): optimism and happiness. Social support was self-reported using the Family, Friends, and Significant Other sub-scales of the Multidimensional Scale of Perceived Social Support (MSSS; ). Perceived academic control measures adolescent perception of their competency and regulation over that competency and was assessed using the academic sub-scale of the Perceived Control Scale (PSC; ). Gratitude was measured using the 6-item Gratitude Questionnaire (GQ-6; ). Gratitude is strongly related to well-being and mental health, a link that has been suggested to be unique and causal . We also collected sociodemographic information such as age, gender, form, tribe, county, economic status, parental education, and family members.
We performed all our analysis in R. First, we investigated the presence of redundant nodes, or overlapping symptoms, by checking pairs of nodes for high correlation (r > .5) and similar correlation patterns with all the other nodes via the networktools package in R . Second, we estimated the network models for the data and validated their accuracy and stability using the bootnet package in R . Third, we plotted the network structures and computed the centrality measures using the qgraph package in R . Fourth and lastly, we investigated bridge symptoms, the main symptoms that connect clusters of symptoms with the networktools package in R . Given that some participants did not answer every question; the data set includes some missing data. To avoid biases, inefficiencies, and incorrect estimates in our analysis, the data were imputed using the mean of each item. The R code for our statistical analysis is publicly available (DOI: 10.17605/OSF.IO/8M5D9).
The lack of variance in variables may lead to the misinterpretation of network structure. If two nodes represent the same construct, the edge between them depicts shared variance and not their association . Thus, these redundant pairs should not be included together in the network. In contrast, if two nodes represent independent constructs, we would expect their correlation patterns with other nodes to vary. We tested for the problematic presence of multiple symptoms representing the same underlying construct using the goldbricker procedure , which checks each pair of nodes for a high correlation between them (r > .5) and similar correlation patterns with all the other nodes (under a given threshold proportion of significantly different correlations).
Network estimation and accuracy
In psychopathology, a network consists of symptoms and the psychometric associations between them. These associations are not explicitly present in a dataset, but they can be estimated by computing partial correlations between the symptoms, controlling for all other symptom connections . Specifically, we estimated a Gaussian graphical model (GGM) to estimate regularized partial correlation networks for psychological wellbeing symptoms and psychiatric symptoms using the bootnet package in R [29,45]. We regularized the GGM with the graphical Least Absolute Shrinkage and Selection Operator (LASSO) method to find the best-fitting model by penalizing, or shrinking, small edge values estimated in the network, and it helps address the multiple testing problem by controlling false- positive errors. The best-fitting model is found with the EBICglasso procedure, which selects the optimal degree of shrinkage according to an Extended Bayesian Information Criterion (EBIC) and hyperparameter set to 0.5 . To plot the networks, we used the qgraph package in R.
A highly central node is a node that has particular structural importance in the network based on the strength of its connection to other nodes. The centrality of a node can be used to infer its influence, or structural importance, in the network. There are many indices used to estimate centrality: betweenness—how a node influences the average path between other pairs of nodes, closeness—how a node is indirectly connected to the other nodes, strength—how a node is directly connected to the other nodes, and expected influence (EI)—how a node is connected to the sum of all edge weights . As partial correlations are used to estimate the networks, it does not make sense to use closeness or betweenness indices. Since we aim to understand which symptoms should be most directly targeted in treatment, we relied on the strength index to estimate the symptoms that should be most directly targeted in treatment. The qgraph package in R was used .
Network accuracy and stability
Bootstrapping was used to determine the accuracy and stability of the networks . First, we tested the accuracy of each network using both nonparametric bootstrapping, a process which repeatedly resamples subsets of the data to calculate a confidence interval (CI) as the range of bootstrapped values from different sampling levels. We first used 1000 bootstraps––or 1000 repetitions of estimating the model with sampled data and calculating the 95% confidence intervals (CIs)––to assess the accuracy of edge-weights. A large CI indicates that it may be difficult to interpret the edge-weight, while a small CI can be interpreted as a precise estimation. Next, we tested the stability of the centrality indices with case-dropping bootstrapping , which is the process of repeatedly estimating a model while dropping rows of the data (i.e., only observing subsets of the data). We calculated a correlation-stability (CS) coefficient, which indicates the maximum proportion of the data that can be dropped while continuing to estimate centrality values that correlate highly (r > .7) with the network from the full sample. Scores .25 and .5 indicate benchmarks for adequate and good network stability, respectively . For each network, we created plots displaying the CIs of edges and centrality values. Bootstrapped difference plots are useful for estimating which edges or centrality values can be meaningfully interpreted as different from one another; we use these plots to guide our interpretations of edge and centrality values.
In a network with multiple scales, bridge nodes are the main nodes that connect to other node clusters, in this case––the other psychiatric and wellbeing measures. We can find these bridge nodes by calculating the bridge centrality statistics from the networktools package in R. The bridge centrality statistic applies to a node’s connection to all the other nodes in the other communities to which it does not belong. Bridge strength is defined as the sum of the absolute value of all edges that exist between a node and all nodes that are not in the same community. Bridge expected influence (one-step) is defined as the sum of the value (+ or -) of all edges that exist between a node and all nodes that are not in the same community as the node [44,48]. Thus, bridge expected influence accounts for how positive and negative edges can neutralize each other. For example, if fasting is positively linked to the desire for thinness (r = .6) and negatively to binge eating (r = –.3), then the regular strength centrality for fasting will be 0.9 (sum of absolute values of those edges). However, the bridge expected influence for fasting would be 0.3 because the value of the negative edge is subtracted from the value of the positive edge. Thus, high bridge expected value would indicate that the node is strongly and positively connected to other nodes.
To find bridge nodes, we defined two community clusters in our network. The first community – adolescent psychopathology – included the depression and anxiety symptoms and the second community – psychological wellbeing – included the gratitude, happiness, optimism, social support, and perceived control items.