This was a large-scale, cross-sectional study conducted from September 2022 to March 2023 in various regions across China, including Jiangsu in the east, Fujian in the south, Sichuan in the west, and Xinjiang in the north. The 4 provinces were selected to increase geographic diversity. Within each province, 1 or 2 middle schools were recruited through convenience sampling. Class-based cluster sampling was then conducted within participating schools, and all eligible students aged 12 to 18 years in the selected classes were invited to participate. A total of 11,946 participants completed an anonymous, self-reported online questionnaire via the Chinese survey platform wjx.cn during a designated computer class under the supervision of their teachers. The flowchart of the procedure was listed in our previous study [35].

For data quality assurance, 2539 questionnaires were excluded based on the following criteria: (1) participants who reported an age below 12 or above 18 years (n=509); (2) participants who completed the survey in less than 10 minutes (n=632); (3) participants who responded “false or entirely false” to the question “To what extent does your survey reflect the real situation? (options: entirely false, false, general, true, entirely true)” at the end of the survey (n=1231); (4) questionnaires with incomplete essential demographic information (n=139); and (5) questionnaires excluded due to duplication (n=28). In total, data from 9407 participants (mean age 14.90, SD 1.61 y) were included in the study, of whom only those with moderate to severe PIU (n=2055) were retained for the subsequent network analyses.

The study received ethical approval from the Ethics Committee of West China Hospital, Sichuan University (number 2019‐907). Written informed consent from the participants’ guardians and assent from the adolescents were obtained before data collection. Participation was voluntary, and all survey responses were anonymous.

Of 11,946 submitted questionnaires, 2539 were excluded because respondents were outside the target age range (n=509), completed the survey in less than 10 minutes (n=632), selected “false” or “entirely false” for the data-validity item (n=1231), had missing essential demographic information (n=139), or had duplicate responses (n=28). Duplicate responses were identified using identical demographic and school information, with the most complete response retained. Among 9407 valid questionnaires, 2055 (21.8%, 95% CI 21.0%‐22.7%) adolescents (n=1043 males) were identified as having elevated PIU symptoms. The mean age of the overall participants and PIU participants was 14.90 (SD 1.61) years and 15.20 (SD 1.53) years, respectively. Meanwhile, the high-risk and lower-risk groups differed significantly in only-child status, residence, family economic status, and parental education, although the effect sizes were relatively small. The demographic variables and measures of the 2 groups are shown in Table 1.

Figure 3 depicts the DAG and the estimated directional dependencies among PIU factors, IES, and SCs in adolescents in the high-risk PIU group. Off_SC showed outgoing edges toward externalizing symptoms (verbal aggression and hostility), internalizing symptoms (depression and anxiety), salience of PIU, and on_SC. Externalizing symptoms generally had more outgoing edges within the psychopathology subnetwork, while anxiety and depression had more incoming edges. Within the PIU symptom subnetwork, salience showed outgoing edges toward other PIU factors, whereas neglect work primarily received incoming edges. Because multiple DAGs may represent the same conditional independence relationships, these arrow directions cannot indicate any causal effect.

Sensitivity analyses using an IAT-20 cutoff of ≥70 included 287 (3.05%) participants. The principal undirected network was broadly preserved, although bridge CS was low (CS-coefficient≈0.05), and the learned DAG structure differed from the primary analysis. After excluding the anticipation and neglect work factors, the 13-node network closely resembled the primary network. EI and bEI showed acceptable stability (CS-coefficients≈0.75 and 0.67, respectively). Detailed results are presented in Multimedia Appendix 1.

This multicenter study extends previous work by constructing a factor-level symptom network that integrates PIU factors, internalizing and externalizing psychopathological symptoms, and SC in adolescents with a high-risk PIU condition. In the undirected network, both SC variables occupied negative bridge positions, with off_SC showing the more prominent role. The exploratory Bayesian network further highlighted the structural prominence of off_SC, which showed algorithm-selected outgoing arrows to nodes across the PIU, internalizing, and externalizing domains. On_SC occupied a less prominent negative network position and was positively connected with off_SC, highlighting the related but distinct positions of SC across the 2 contexts.

This study identified a closely interconnected factor-level structure between PIU symptoms and psychopathological symptoms among adolescents meeting the high-risk PIU threshold, consistent with previous item-level network studies [19,21,22]. Within the comorbidity network, psychopathological symptoms exerted greater influence, suggesting that emotional and behavioral symptoms may be more clinically prominent than PIU-related symptoms among adolescents in the high-risk group. This pattern is consistent with clinical observations that adolescents with PIU may seek treatment for emotional burden or behavioral problems rather than internet overuse alone, and such comorbidities are associated with poorer treatment outcomes [57]. Notably, centrality analyses highlighted anxiety as the most influential symptom domain among the psychopathological symptoms. This finding broadly aligns with previous studies identifying anxiety-related symptoms as central or bridging features in PIU-related networks. A recent large-scale study also identified social anxiety as a node with high strength and bridge centrality in the network of PIU and adolescent psychological issues [58].

Complementing this finding, the learned DAG positioned anger as a parent node with multiple outgoing arrows, highlighting its prominent structural position in the PIU-psychopathology network. This finding is consistent with longitudinal evidence. A 3-wave study found prospective associations between dysfunctional anger regulation and subsequent internalizing problems [59]. A study of 1103 adolescents also supported the positive association between PIU severity and anger [9]. These findings may help explain why anger occupied a prominent cross-domain position in the learned DAG, although its parent-node position should not be interpreted as evidence that anger causally precedes the other symptoms.

After incorporating SC factors into the comorbidity network model, we observed a substantial improvement in overall stability, with both the bEI and EI coefficients reaching high levels. This suggests that both off_SC and on_SC helped compensate for the previously unstable bridging structure and acted as meaningful connectors between PIU and associated psychopathological symptoms. Within the overall PIU-IES-SC network, SC—particularly off_SC—appears to play a crucial role in disrupting the reciprocal reinforcement between PIU symptoms and psychopathological symptoms. This is consistent with the study by Prievara et al [60], which also pointed out that social support from offline environments is a protective factor against PIU, and aligns with broader evidence that feeling connected to others is a fundamental human need linked to higher well-being and lower risk of stress disorder, anxiety, and depression [19,61,62]. Generally, offline social activities are more beneficial for mental health than online communication, as face-to-face interactions often provide richer emotional cues, including facial expressions, gestures, and eye contact [63]. Active participation in offline social activities can also reduce adolescents’ screen time, thereby alleviating PIU symptoms [64].

Our flow network model and exploratory DAG provided complementary descriptions of how off_SC was embedded within the PIU-IES-SC system. As a reorganization of the undirected network, the flow network showed that all symptom nodes were connected with off_SC within 2 steps, indicating that off_SC was closely linked to the PIU-psychopathology structure. The exploratory DAG further showed algorithm-selected outgoing arrows from off_SC toward several internalizing, externalizing, and PIU-related nodes. Given the cross-sectional design, the observed orientations may reflect off_SC as an antecedent, a consequence, or part of a reciprocal process [26]. Relevant longitudinal evidence suggests that the temporal relationship between off_SC and PIU may vary according to the type and source of connectedness. In a 3-wave study of 1684 Chinese adolescents, higher school connectedness predicted lower subsequent problematic social network use, and vice versa [65]. In a 4-year study spanning Grades 8 to 11, compulsive internet use predicted later reductions in perceived teacher support; however, parental support predicted subsequent increases in compulsive internet use [66]. Although the causal relationships of off_SC with PIU and psychopathological symptoms remain uncertain, its consistently prominent structural position across the PIU-IES-SC network makes it a plausible intervention target. One relevant approach is social prescribing, which links individuals to community-based social resources and activities [67] and may improve social relationships and psychological well-being [68]. For adolescents with elevated PIU symptoms, coordinated family-based, school-based, and community-based programs could promote meaningful offline interaction as an adjunct to established psychological or clinical care.

In the PIU-IES-SC network, on_SC also emerged as a protective factor, exhibiting the second-highest negative EI value after off_SC. In the digital age, the internet has become a key platform for adolescents to engage socially. Previous evidence has suggested that forming new friendships, exploring personal interests, and sharing stressors on social media help adolescents to develop a sense of identity and help relieve negative emotions such as anxiety and depression [69-71]. On_SC was also directly and positively connected with off_SC, suggesting that online and offline connectedness may represent related and complementary social resources rather than function as entirely separate domains. In a study of 733 adolescents, those who were highly socially engaged in both contexts reported the most positive self-concepts, whereas those with high online but low offline engagement reported the least positive self-concepts [72]. A longitudinal study similarly found that greater social media use predicted more subsequent time spent with friends offline [73]. These findings suggest that online connectedness may be particularly beneficial when it complements meaningful offline relationships.

Finally, 3 findings warrant cautious interpretation. First, despite the predominantly negative network profile of off_SC, it showed a small positive partial association with verbal aggression (r=0.10). Because this edge represents the residual association between off_SC and verbal aggression after accounting for hostility, physical aggression, depression, and all other network nodes, its small positive value may partly reflect the removal of variance shared with these correlated factors [74]. It is also possible that greater offline engagement increases exposure to interpersonal conflict or peer contexts that reinforce aggressive communication [75]. Given its small magnitude, this association requires replication before substantive conclusions can be drawn. Second, although the proportion of boys and girls differed between the high-risk and lower-risk PIU groups, the network comparison tests detected no significant gender differences in global strength or overall edge structure. This suggests that gender differences in PIU severity or presentation do not necessarily translate into differences in the overall pattern of associations among PIU, psychopathological symptoms, and SC. Previous studies have also reported mixed findings. One longitudinal PIU network study reported similar network structures and global strength across boys and girls [76], whereas another large-scale adolescent study identified gender-specific PIU network characteristics [25]. Thus, the present findings indicate that no clear gender difference was detected in this sample. Third, when the stricter IAT-20 cutoff of ≥70 was applied, the overall undirected network was broadly preserved, but bridge centrality showed low stability and the learned DAG differed from that obtained in the primary analysis. One likely explanation is the substantial reduction in sample size, from 2055 participants in the primary analysis to only 287 under the stricter cutoff. Nevertheless, the discrepancy may also reflect genuine differences in the dependency structure of adolescents with more severe PIU symptoms. Therefore, the bridge-node rankings and algorithm-selected DAG orientations under different PIU thresholds should be regarded as exploratory and require replication in larger samples with severe PIU.

Despite the insights provided by this study, several limitations should be acknowledged. First, the cross-sectional nature of the data precludes definitive conclusions about causality, and our sampling logic of schools may limit the generalizability of the findings. While our findings allow for potential causal hypotheses based on partial correlation networks and prior knowledge, longitudinal studies using more representative samples are needed to further validate these assumptions. Second, the adapted SCS-R scales have not been independently validated before; therefore, the online-offline differences and their network positions should be interpreted cautiously. Third, several potentially relevant factors, including family functioning, socioeconomic status, and personality traits, were not included in the network models. These factors may be associated with both SC and psychopathological symptoms, and their omission may have influenced the estimated network structure and DAG orientations. Finally, the internalizing and externalizing psychopathological symptoms included in this study were limited to those commonly associated with PIU. Other co-occurring symptoms, such as insomnia and psychotic-like experiences, were not included, which may have constrained the comprehensiveness of the symptom network.

This study characterized a closely interconnected factor-level network linking PIU with internalizing and externalizing psychopathological symptoms among adolescents with high-risk PIU symptoms. Off_SC showed the most prominent negative bridge position and was directly or indirectly connected with all network nodes within 2 steps. On_SC showed a secondary negative network position and was positively associated with off_SC, suggesting that connectedness across the 2 contexts may represent related social resources. These findings identify off_SC as a plausible candidate for further prevention and intervention research. Longitudinal and interventional studies using validated context-specific measures are needed to examine directionality and determine whether SC-focused strategies can improve PIU-related and mental health outcomes.