2.2.1. Personal and life-style variables.

Participants were asked to provide information about their Age, height, and weight. Body Mass Index (BMI) was calculated from the weight/height ratio. They were asked to indicate, in separate questions, their Sex and Gender, by selecting Male, Female or Other, with also the option “Do not wish to reply”. Using this information, we derived a binary variable indicating whether the participant was Cisgender, that is whether their gender identity matches the sex they’re identified with. As a proxy for Subjective Economic Status, participants were asked to rate the economic situation of their family/environment during their upbringing on a 5-points scale (“very difficult”, “difficult”, “neither difficult nor comfortable”, “comfortable”, “very comfortable”). Only three participants reported the first level, which was thus combined with the second.

From the Petersen Pubertal Scale [52], we derived an index of late or early-puberty, by comparing each individual score to the function puberty vs age estimated from our sample (see S1 File). We also asked about Physical activity, based on four questions [53], and engagement in Social Media (affective, cognitive and behavioral) [54] (see S1 File).

2.2.2. Sensory processing characteristics.

2.2.2.1. Chronic pain: Participants were asked to report the number of Chronic pain sites with episodes of pain lasting three months or more, as an index of central sensitivity to nociception.

2.2.2.2. Sensitivity threshold and behavior to somatosensory stimulation: We used 23 items from the Brown and Dunn’s Adult and Adolescents Sensory Profile questionnaire [55] to characterize sensitivity threshold (high or low threshold) and behavioral response to sensory stimulation (sensation seeking or avoiding). Items of both measures relate to sensory processing of touch and movement (the original version covers also taste, smell, vision, and audition). The Likert scales ranged from “nearly never” (1) to “nearly always” (5), with higher scores indicating a high threshold (i.e., low sensitivity to somatosensory sensations) and sensation seeking.

2.2.2.3. Social Touch: We used a 7-items questionnaire derived from the Social Touch Questionnaire [56] and from the Dunn Sensory Profile [55] to evaluate the extent to which individuals appreciate touching and being touched in a social context. The Likert scales ranged from “nearly never” (1) to “nearly always” (5), with low scores meaning strong avoidance of social touch.

2.2.2.4. Body noticing, not-distracting, not-worrying, listening and trusting and emotional awareness of the body: We used six subscales of the French version of Multidimensional Assessment of Interoceptive Awareness (MAIA) [9] validated in youth [10]. Noticing concerns the subjective awareness of body sensations. Not-Distracting investigates how often a person tends not to ignore sensations of pain or discomfort. Not-Worrying assesses the extent to which a person doesn’t worry about, nor catastrophizes, discomfortable sensations. Emotional Awareness deals with the awareness of the relation between emotional and bodily states. Body Listening assesses how often a person actively attends to their bodily sensations for insight. Body Trusting investigates the extent to which a person experiences their body as a safe and trustworthy source of information. The Likert scales ranged from “never” (1) to “always” (5), with higher scores indicating greater awareness, listening and trust in one’s body, and reduced worrying.

2.2.3. Self-reflection traits.

2.2.3.1. Body-esteem for appearance, for weight and attribution: We used the French version [57] of the Body-Esteem Scale validated for adolescents [58]. The questionnaire consists of three subscales. The Appearance subscale assesses general feelings about one’s looks. The Weight satisfaction investigates perception of one’s own weight and the extent to which one would like to change it. The Attribution subscale assesses beliefs about how others perceive one’s appearance. The Likert scales ranged from “never” (1) to “always” (5), with higher scores indicating more positive feelings about appearance, greater weight satisfaction, and positive beliefs about how others perceive one’s body.

2.2.3.2. Family, peers and social-media appearance-related pressure: We used the French version [59] of the three subscales from the Sociocultural Attitudes Towards Appearance Questionnaire 4 (SATAQ-4) [60] focusing on appearance-related pressure. These scales assess pressures from Family, Peers, and Social Media. The Likert scales ranged from “definitely disagree” (1) to “definitely agree” (5), with higher scores indicating greater felt pressure. Following reports of misunderstanding in pilot tests, we modified the wording of four items.

2.2.3.3. Private self consciousness, public self consciousness and social anxiety: We used the 22-items French version [61] of the revised version [62] of the Self Consciousness Scale (SCS) [63]. The SCS consists of three subscales. The Private Self-Consciousness subscale assesses the extent to which individuals focus on internal thoughts and analyze their own mental functioning. The Public Self-Consciousness subscale quantifies the extent to which an individual tends to focus on oneself as a social agent that can be observed and judged by others. The Social Anxiety subscale measures distress caused by interacting with others. The Likert scales labels ranged from “extremely uncharacteristic of me” (1) to “extremely characteristic” (5), with higher values indicating a higher tendency to focus on oneself/ be anxious.

2.2.3.4. Perspective taking and empathic concern: We used two subscales of the French version [64] of the Interpersonal Reactivity Index (IRI) [65], commonly used in adolescents populations [66]. Perspective Taking assesses the tendency or ability to adopt the perspective of other people and is considered a measure of cognitive empathy. Empathic Concern assesses the tendency to experience feelings of warmth, compassion, and concern for others undergoing negative experiences and is considered a measure of affective empathy. The Likert scale displays ranged from “this doesn’t describe me at all” (1) to “this describes me very well” (5), with higher values indicating higher empathy.

2.2.3.5. Resistance to peer influence: The Resistance to Peer Influence (RPI) Scale [38] consists of 10 pairs of statements (i.e., ‘’some people… but other people.”). Participants are asked which of the two statements describes them better. Responses are then coded on a 4-point scale, ranging from “really true” for one descriptor to “really true” for the other descriptor, and averaged. Higher scores indicate greater resistance to peer influence.

2.3.1. Overlap analysis: Assessing discriminant features of the various measures.

We checked for topological overlap between the different measures using the Goldbricker function from the R package networktools [67] (version 1.6.0) with a threshold of 0.35, and the default “hittner2003” method. This recommended procedure [47,68] checks for pairs of variables that have similar connections to other variables in order to identify variables that most likely measure the same underlying construct [67]. Such overlap would influence the network architecture [69]. The function’s output provides an indication of which nodes to remove prior to network estimation. Because of excessive overlap, we removed two subscales for the network analysis: the Body Esteem for Weight (67% overlapping correlations with Body Esteem for Appearance) and Noticing of body sensations (77% overlapping correlations with Emotional awareness of the body). The final network was therefore made of 30 measures. We applied the same procedure with a threshold of 0.25 before estimating the networks of each group that we compared (see below). For the group of the male participants Emotional Awareness was removed (88% overlapping correlations with Body Listening) and therefore also removed for the group of girl participants during network comparison between sexes. Behavioral Engagement in Social Media was excluded in young girls due to high overlap with Affective Engagement (87%) and was also removed in older girls to ensure comparability across groups.

2.3.2. Network estimation.

To study the complex, fine-grained interconnections among the variables, we conducted a network analysis in R (version 4.4.1) using the bootnet (version 1.6) and qgraph (version 1.9.8) packages [70]. Based on the covariance matrix of the 30 variables, we used the estimateNetwork() function to estimate partial correlation networks, which model the unique association between each pair of variables while controlling for all others. We then visualized the results using qgraph().

First, we estimated a Gaussian Graphical Model (GGM) using the least absolute shrinkage and selection operator (LASSO) with Extended Bayesian Information Criterion (EBIC) model selection (with the default hyperparameter γ = 0.5) [71,72]. This approach, known as EBICglasso, is widely used in psychological network research for datasets with a similar number of nodes and sample size. The resulting network was relatively dense, reflecting high sensitivity, but potentially lower specificity.

To address this potential lack of specificity, we repeated the EBICglasso estimation with an additional, global, thresholding procedure, removing edges with value in the precision matrix below log(p*(p-1)/2)/ sqrt(n), with n being the number of participants (734) and p the number of variables (30) [70].

We also estimated the network using the ggmModSelect algorithm implemented in bootnet. This is a fast non-regularized algorithm based primarily on the glasso algorithm in combination with BIC model selection [73]. For sample sizes and network dimensions comparable to ours, this method has been shown to yield lower sensitivity but higher specificity for both edges and bridge edges [73]. This makes it a valuable complementary approach for identifying the most stable and interpretable network connections.

Finally, based on the findings of Williams and Rast [74], who showed that unregularized network estimation performs well for this sample size and number of nodes, we also estimated unregularized partial correlation networks, applying false discovery rate (FDR) correction to control false positives.

We selected the EBICglasso without thresholding as our main network estimation procedure for subsequent analyses. This method remains the most widely used in psychological network literature and is well-suited for identifying overall network structure and strong associations at our sample size [73]. To assess whether the identification of the most prominent edges depended on the estimation procedure, we repeated all key analyses using the alternative estimation methods. Results were largely consistent across models, with only minor variations in network structure. These complementary analyses are reported in the S1 File and confirm that our main conclusions do not depend on the choice of estimation procedure.

Additionally, to address potential concerns regarding collider bias or spurious associations, we compared the zero-order and partial correlation matrices. This analysis confirmed that the identified edges were not a byproduct of considering too many variables. Full details are provided in the S1 File.

2.3.3. Centrality and bridge estimation.

For each network obtained using the bootnet/qgraph packages, we looked at strength centrality, using the centrality function in the R package qgraph, and bridge-strength centrality using the bridge function in the R package networktools. Strength centrality is the sum of the absolute weights of all edges connected to a node and is a measure of its overall connectedness. Other centrality indices, such as closeness and betweenness, could be used, but have been shown to be unstable [75]. Likewise, expected influence overlooks negative edges, which is not compatible with our study, as score directions vary across scales (higher scores meaning more sensory processing or self-reflection in some, and the opposite in others) and negative edges may carry different meanings. Bridge strength quantifies the extent to which a node within a specific community (i.e., a group of nodes representing the same psychological domain) connects to nodes in other communities. In the present study, bridge strength was calculated to identify sensory nodes with the strongest connections to self-reflection nodes, and conversely, self-reflection nodes most strongly connected to sensory nodes. Bridge strength is computed as the sum of the absolute weights of a node’s edges that connect it to nodes in other communities. We emphasized bridge strength because centrality strength have been shown to be sensitive to the specific composition of nodes within a network, which may limit their interpretability in psychometric network analyses [75]. In contrast, bridge indices are invariant to the type of network estimated, rendering them especially useful in psychometric research. We qualified as bridge-node, nodes with a bridge strength larger than one standard deviation above the mean bridge strength.

2.3.4. Communities estimation.

For each network estimated with the bootnet/qgraph packages, we applied Exploratory Graph Analysis (EGA) to identify communities—that is, groups of nodes showing similar patterns of association with the rest of the network [48]. We used the community.detection() function from the EGAnet package (version 2.3.0) with the argument algorithm = “optimal”. This analysis was conducted to test whether the detected communities aligned with our predefined theoretical categories, namely self-reflection and sensory processing.

2.3.5. Network comparison between groups.

We investigated the differences due to sex or age in our dataset. Networks were estimated for all subgroups using the same procedure as for the whole sample. The obtained networks were then compared using permutation tests implemented in the NCT() function of R package NetworkComparisonTest (NCT) [51] (version 2.2.2; 2000 permutations). For each of the two comparisons (between sexes and between age-groups), we compared the networks in terms of global structure, global strength, individual edges that link one node of sensory processing with one related to self-reflection, strength centrality, and bridge-strength measures. When the global structure and global strength are not significantly different between two networks, differences in individual edge weights in these groups are more likely to be of a content-related nature and not due to noise/measurement error [76]. Due to the conservative nature of the NCT [51], we report results both with and without a Bonferroni-Holm correction for multiple comparisons. We applied this family-wise error correction considering only the edges that link one node of sensory processing with one related to self-reflection. In addition, similarly to previous work [76], we calculated difference networks by subtracting the weights of one network from the corresponding weights of the other network.

2.3.6. Robustness analysis.

Standard robustness analyses were conducted using case-dropping bootstraps to evaluate the stability of edge weight estimates, differences between edges, centrality, bridge strength, and community structure. These analyses were applied to all networks estimated using the EBICglasso regularization method without thresholding—both the main network on the full sample and the four subgroup networks used for group comparisons. Detailed results are provided in the S1 File.

3.5.1. Comparison between sexes.

Networks for girls and boys are presented in Fig 7a and 7b. Global strength comparison tests showed that the difference between the girls network’s global strength (3.51) and the boys network’s global strength (2.06) was not significant (Test statistic S: 1.45, p = 0.31). The network invariance test was also not significant (Test statistic M: 0.18, p = 0.47). Since the overall level of connection was not significantly different between the two groups, differences in individual edge-weights were more likely to be of a content-related nature.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Comparison between sexes.

Networks for adolescent girls (a) and boys (b). Edge width is scaled like for the whole sample with a maximum of 0.7 and a minimum of 0. In (c), the difference between networks a and b is computed and only edges that are significantly different are plotted. Edge thickness is scaled to the absolute value of r difference (plotted with argument cut = 0.001). Edge colors and line types represent group differences as follows: Green solid lines = Positive edges stronger in girls. Red dotted lines = Negative edges stronger in boys. See Fig 2 for node descriptions.

https://doi.org/10.1371/journal.pone.0335259.g007

When not correcting for multiple comparisons, three edges linking a sensory node with a node related to self -reflection traits were found to be significantly different between boys and girls (Fig 7c). The positive connections between body trusting and body appearance esteem (p = 0.027) and between body trusting and Resistance to peer influence (p = 0.019) were stronger in girls. The negative connection between chronic pain and body esteem for appearance was stronger in boys (p = 0.037). Body trusting (p = 0.013) and resistance to peer influence (p = 0.031) displayed bridge-strength higher in girls compared to boys. No significant difference was found regarding strength centrality. When applying Bonferroni-Holm correction for 88 multiple comparisons, there were no statistically significant edge, centrality-strength, or bridge-strength differences.

The Network Comparison Test on networks estimated with alternative methods (see S1 File) showed some variations, but body trusting and body esteem appearance consistently emerged as bridge nodes with stronger connections in girls than in boys.

3.5.2. Comparison between age-groups in girls.

Networks for younger girls and older girls are presented in Fig 8a and 8b. Global strength comparison tests showed that the difference between young girls’ global strength (1.64) and older girls’ global strength (1.85) was not significant (Test statistic S: 0.20, p = 0.843). The network invariance test was also not significant (Test statistic M: 0.19, p = 0.35).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Comparison between age-groups in girls.

Networks for younger girls (a) and older girls (b). In (c), the difference between networks a and b is computed and only edges that are significantly different are plotted. Edge thickness is scaled to the absolute value of R difference (plotted with argument cut = 0.001). Edge color and style indicate group differences: Red solid lines = Negative edges stronger in younger girls. Green dotted lines = Positive edges stronger in older girls. Red dotted lines = Negative edges stronger in older girls. See Fig 2 for nodes description.

https://doi.org/10.1371/journal.pone.0335259.g008

When not correcting for multiple comparisons, four edges linking a sensory node with a node related to self-reflection were found to be significantly different between the two groups (Fig 8c). The positive connections between body trusting and body esteem attribution (p = 0.043) and between body listening and private self-consciousness (p = 0.036) were stronger in older girls. The negative connection between sensitivity behavior and social anxiety was stronger in older girls (p = 0.039). The negative connection between sensitivity threshold and media pressure (p < 0.001) was stronger in younger girls. No significant differences were found regarding strength centrality and bridge-strength. None of the significant differences remained after applying multiple comparisons correction.

The Network Comparison Test on networks estimated with alternative methods (see S1 File) showed different results regarding global strength, individual edge weight and bridge edges.

4.1.1. Somatosensory processing relates to social anxiety.

The first community (red community in Fig 6) connected social anxiety to three somatosensation nodes, and to physical activity. Social anxiety was also one of the self-reflection traits with high bridge strength. This indicates that social anxiety is related to how adolescents experience and react to somatosensory stimulation. Adolescents with a low threshold (i.e., very sensitive to somatosensory sensations) have a higher tendency to feel anxious in social situations. Previous studies have shown that children and adolescents diagnosed with a general anxiety disorder consistently display a heightened sensitivity to external stimuli [23]. We extend this finding to a non-clinical population and our results support the idea that heightened sensory sensitivity could be a sign of social anxiety vulnerability. Our results also invite future research to test whether somatosensory habituation training could help to prevent anxiety disorders onset.

Additionally, in our data, the tendency to engage in social touch was negatively related to social anxiety and positively related to sensitivity threshold and sensitivity behavior. This extends findings in adults [77]. Despite the present analysis not allowing us to test directionality, we can hypothesize that these relations are bidirectional. Individuals experiencing more anxiety in social interactions are less likely to engage in social touch. Reciprocally, people who engage less in social touch are less likely to appreciate touching and being touched, and may thus be more likely to experience social anxiety. In parallel, social touch has been found to have positive consequences on the psychological and physiological human development [78], and to reduce anxiety [79]. Since, adolescents show a heightened neural response to pleasant touch compared to adults [80], this could be leveraged to break this cycle of reciprocal influences in case of anxiety at this age.

4.1.2. Interoceptive awareness intertwined with reflections on one’s own and others’ thoughts.

Another community (green community in Fig 6) integrated all nodes of the MAIA questionnaire (interoceptive awareness), except body trusting, with private and public self-consciousness and the two nodes of the interpersonal reactivity index (empathic concern and perspective taking). Adolescents who exhibited a higher propensity for introspection regarding their mental processes and feelings were those who were more attentive to bodily sensations and relied on them to assess their emotional states. Recently, Yang et al. [43] reported a correlation between MAIA questionnaire scores and self-consciousness scale scores in adults. The authors didn’t report which subscales of either questionnaire weighed more in this relation, however. Altogether, these results extend to adolescence the findings from studies conducted in adults that indicate that selfhood is grounded in interoceptive processing [81].

Interestingly, body listening was also associated with perspective-taking, the latter being related to private self-consciousness. This points towards an interaction between the capacity to process signals originating from within the body and the capacity to switch between self- and others- perspective, as it has already been indicated in adults [27]. Previous developmental models proposed that, in adolescence, the construction of self-knowledge is intertwined with the construction of knowledge of others [82]. Based on our results, we propose that the sensitivity to bodily sensations also intervenes in this construction.

4.1.3. Trusting bodily sensations relates to the tendency to reflect towards one’s body appearance.

A third community (orange community in Fig 6) revealed an association between body trust and nodes related to body appearance. Bridge centrality analysis indicated that body esteem for appearance was strongly linked to sensory nodes, and body trusting was strongly linked to self-reflection nodes. This shows, in a sample of adolescents from the general population, that body trust, that is the perception of information coming from the inner body as reliable, is closely tied to how individuals reflect on their body appearance: young people with low body trust also have negative reflections towards body appearance. This is consistent with what has been shown regarding other facets of body esteem [83] and with previous studies that have shown that low body trust is associated to anorexia nervosa symptoms [30]. In partial correlation networks, links correspond to conditioned associations with other nodes. This means that the link between body trust and body esteem for appearance does not depend on body-mass index, which on its own has been consistently associated with body esteem. At the same time, low body esteem, a psychological trait, is a robust and consistent predictor of anorexia nervosa [84]. These results suggest that preventing body mistrust may reduce the risk of body image disorders, although this must be taken with caution since the links in the network are not directional [85].