*2.3. Data Analysis*

Descriptive analyses and comparative analyses of the overweight, underweight and normal weight subsamples were performed using IBM SPSS Statistics 27.0. The network analysis was conducted using JASP (version 0.12.2.0) [42] which makes use of the *R* packages '*bootnet*' [43] and '*qgraph*' [44].

First, we compared YSR, SCOFF and KIDSCREEN (general and subscale) scores obtained from the overweight and underweight subsamples with the normal weight reference group using general linear models controlling for sex. Differences between the overweight/underweight and normal weight samples were analyzed using Tukey tests. We further used Chi2-tests to compare the percentages of clinically relevant YSR scores and eating disorder risk (SCOFF ≥ 2) between the overweight, underweight and normal weight reference samples.

The network analysis for overweight and underweight adolescents was performed on general psychopathology variables (YSR syndrome scales), eating disorder risk (SCOFF total score) and well-being/quality of life variables (KIDSCREEN scores, not using the KIDSCREEN-10 general quality of life score). Due to the correlational nature of this approach and as age-/sex-standardized scores are not available for all of the included instruments, we used the raw scores of these questionnaires in this analysis. A network is defined as a set of variables (called 'nodes') which are reciprocally connected through 'edges' (most commonly some kind of correlation) that do not imply a priori direction or allow causal inference. In the present study, we estimated partial correlation networks using the graphical Least Absolute Shrinkage and Selection Operator (gLASSO [45]). Using the gLASSO estimation, small or unstable correlations within the network are set to zero, resulting in a more parsimonious and better interpretable network only depicting the most robust associations between the nodes. Each edge represents the thus regularized partial correlation between two nodes. In contrast to non-regularized partial correlations (all edges between all nodes are estimated and included in the network plot), regularized partial correlations are used to effectively assess the sparse and interpretable network structure. The stronger the partial correlation between two nodes (either positive or negative), the thicker the edge presented in the network plot. As gLASSO produces a collection of network solution, the Extended Bayesian Information Criterion (EBIC, [46]) was used to select the optimal network model. The Fruchterman–Reingold algorithm [47] was used to organize the network plot. Nodes with more or stronger connections are placed closer together while nodes with less connection are placed further apart.

The centrality of the nodes was estimated with the node strength, betweenness and closeness centrality indices. Node 'strength' refers to the weighted number and strength of all connections of a specific node and thus represents the overall influence of a node in the network. 'Betweenness' represents the number of shortest paths that pass through the node of interest, respectively, the number of times that the node represents the shortest path between other nodes; thus, a node with high betweenness is important in the connection that other nodes have between them (node acting as a bridge). 'Closeness' quantifies the number of direct and indirect links between the node of interest to all other nodes in the network; thus, a node with high closeness will be affected quickly by changes in any part of the network and vice versa (c.f. [24]). *z*-Standardized centrality indices (mean = 0, SD = 1) are reported.

Moreover, we performed the Network Comparison Test (NCT, [48]) using the '*NetworkComparisonTest*' package in *R* to compare the network structure and the global network strength between the overweight and underweight samples.
