A population-based cross-lagged panel network analysis of multidimensional mental health symptoms among Chinese secondary school students

Abstract

Adolescent mental health issues are increasingly severe, with the prevalence of depression reaching 26.17%. Existing studies still face issues such as a narrow range of symptoms, cross-sectional designs, and small sample sizes. To address these gaps, the present study applied cross-lagged panel network (CLPN) analysis to a population-based sample of 17,236 secondary school students in Guangzhou, China (aged 12–18; 46.32% female; 66.91% junior high, 33.09% senior high). Participants were assessed on ten symptom dimensions at two time points approximately six months apart, during September–October 2023 (T1) and February–April 2024 (T2). Subgroup comparisons were conducted by gender and educational stage. The findings revealed: (1) cross-sectional networks exhibited robust and stable associations, with depression and anxiety forming the strongest link; (2) longitudinal analyses identified depression, anxiety, and emotional disturbance as central driving nodes that predicted and propagated changes in other mental health symptoms over time, while paranoid ideation functioned primarily as a bridging symptom; and (3) subgroup analyses showed largely consistent structures across gender and educational stages, though subtle differences emerged. Males demonstrated slightly stronger connectivity, with depression and emotional disturbance as core drivers, whereas hostility and interpersonal sensitivity were more central in females. Across educational stages, the “depression → anxiety” pathway remained stable, with emotional disturbance more prominent in junior high and hostility gaining influence in senior high. Overall, these results highlight the central role of anxiety in adolescent symptom networks, reflecting the pervasive influence of China’s high-pressure educational environment on students’ psychological functioning. They underscore the importance of disrupting the depression–anxiety cycle and suggest that tailored, stage- and gender-specific interventions targeting emotional symptoms may enhance the effectiveness of school-based mental health promotion.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12889-026-27745-y.

Introduction

Adolescent mental health is increasingly recognized as a major public health challenge worldwide. In China, mental health problems are highly prevalent among secondary school students, in part because of escalating academic competition [1]. A recent report indicates that the prevalence of depressive symptoms among Chinese secondary school students has reached 26.17% [2], which could probably cause serious consequences like behavioral difficulties, academic impairments, and an elevated risk of suicide, imposing substantial burdens on both families and society [3, 4].

Adolescent mental health problems can be fruitfully understood within a developmental psychopathology framework, which emphasizes that psychological difficulties emerge through ongoing interactions between individual vulnerabilities and contextual stressors over time. In this perspective, symptoms are not viewed as static or isolated outcomes, but as developmentally embedded processes that may influence and reinforce one another across domains [5]. The interwoven nature of stressors (e.g., academic burden), symptoms (e.g., depression), and outcomes (e.g., suicide) suggests that these factors do not operate in isolation but may influence and reinforce one another over time [6]. For example, academic pressure may trigger depressive symptoms, which in turn can increase feelings of hopelessness and elevate suicide risk. Such interconnected processes underscore the necessity of conceptualizing adolescent mental health as a dynamic network of mutually interacting elements rather than as a collection of independent conditions.From this perspective, adolescent mental health is better conceptualized as a dynamic and interconnected symptom system rather than a set of independent conditions. This developmental psychopathology lens also provides a strong theoretical basis for applying a cross-lagged panel network approach in the present study, as CLPN allows us to examine how multiple symptom dimensions are temporally linked and how certain symptoms may act as driving or bridging components within the evolving system.

In recent years, network analysis has emerged as a powerful methodological framework for investigating the complex interrelations among psychological symptoms [7]. Unlike traditional latent variable approaches that treat symptoms as interchangeable indicators of an underlying disorder, network analysis conceptualizes symptoms as directly interacting components that can activate and sustain one another [8, 9]. By modeling these associations, researchers can identify central nodes that play a disproportionate role in the maintenance of psychopathology, as well as bridge symptoms that connect different symptom clusters [8, 10, 11]. This perspective provides a more nuanced understanding of how stressors, symptoms, and outcomes coalesce into a dynamic system, offering novel insights into the mechanisms underlying adolescent mental health difficulties.

An increasing number of studies have applied network analysis to examine mental health problems among Chinese adolescents. For instance, Wang et al. conducted a network analysis to investigate mental health issues across different educational stages [12]. They found that anxiety consistently emerged as the core dimension of mental health problems at all stages. Paranoid ideation was more prominent among junior and senior high school students, whereas emotional distress was particularly salient in elementary school students. The results further indicated that senior high school students, who were faced with increased academic demands from school and family during the late adolescence, exhibited a stronger connectivity among psychological dimensions in their mental health networks.

Similarly, Wu et al. employed a cross-lagged network analysis based on two-wave data to examine the relationships among internet addiction, depression, and loneliness in Chinese university students. Their findings indicated that loneliness predicted subsequent depressive symptoms, whereas internet addiction showed dual effects: slightly reducing suicide risk while exacerbating depressive symptoms [13]. In another study tracking the evolution of depression symptoms in Chinese adolescents, “low mood” was identified as the core symptom of depression for both elementary and senior high school students, in which “fatigue” played a central role in the depression network for junior high school students [14]. Existing studies suggest that developmental factors such as gender and educational stage play an important role in shaping adolescent mental health symptoms. Adolescence is characterized by rapid emotional, cognitive, and social development, during which boys and girls may differ in emotion regulation strategies, coping styles, and sensitivity to interpersonal stress. For example, girls are often found to be more vulnerable to relational stress and social evaluation, whereas boys may be more likely to suppress emotional expression when facing stress. Such differences may influence not only the severity of psychological symptoms but also the ways in which symptoms interact with one another. To sum up, developmental factors such as gender and educational stage are key factors shaping the manifestation of mental health symptoms among adolescents. Additionally, other symptoms such as positive thinking, gratitude, and depression highly predict the shaping of adolescents’ learning burnout. However, it was only observed in the study carried out during the COVID-19 pandemic and, to some extent, the symptoms were time-specific in its influence on learning burnout [15].

Despite the growing application of network analysis to adolescent mental health, several important gaps remain. First, most previous studies have examined a very limited set of symptoms, often including only three or four variables within a single model, which constrains the ability to capture the multidimensional nature of adolescent mental health [16]. Second, methodological work has shown that node selection and node assessment are critical issues in psychopathology network analysis; when important symptom domains are not simultaneously assessed or are measured imprecisely, the resulting network may be distorted. Specifically, omitted variables and measurement error may produce spurious edges, attenuate true associations, and bias the identification of central or bridge symptoms [17, 18]. Therefore, employing a standardized multidimensional instrument within a cross-lagged panel network framework is necessary to more accurately capture the dynamic interplay among mental health symptoms during adolescence. Third, there is a lack of statistically powerful large-scale longitudinal data, particularly from high-pressure educational contexts such as China. Most existing studies have relied on relatively small or convenience samples, which limits both the generalizability and the stability of inferred network structures. Large-scale sampling with substantial statistical power is especially critical for ensuring the robustness of network parameters and for capturing the heterogeneity of adolescent mental health in such contexts [19].

China provides a particularly informative context for this inquiry because academic pressure is both intense and institutionalized across key stages of secondary education. Studying adolescent mental health in this setting may therefore offer insights not only into China itself but also into other societies facing rising academic competition and growing concern about youth mental health. This contextual relevance is not yet fully reflected in the existing longitudinal network literature. One recent study employed a CLPN model to examine mental health symptoms in a large adolescent sample using the Symptom Checklist-90 (SCL-90), identifying initiating symptoms and gender-specific temporal pathways [20]. However, the sample consisted mainly of vocational and technical school students. While this is an important group, it is not representative of the broader population of Chinese secondary school students, most of whom attend general junior or senior high schools and face sustained academic pressure. In the Chinese context, where academic pressure is especially salient in general secondary education, it is therefore necessary to examine the dynamic symptom network specifically among ordinary middle school students.

The Chinese Middle School Students’ Mental Health Scale (MSSMHS) was employed in this study as the primary instrument. While adapted from the well-established SCL-90, the MSSMHS is specifically tailored to the developmental and cultural nuances of Chinese adolescents, offering a more concise yet culturally resonant assessment [21]. From a network perspective, the ten dimensions are conceptualized as an interrelated system in which symptoms influence and reinforce one another, rather than as independent indicators of a single latent construct. Within this symptom-level framework, Academic Pressure—an important environmental stressor in China’s highly competitive educational context—may function as a contextual factor that contributes to the activation of internalizing symptoms, particularly Anxiety, Depression, and Emotional Disturbance [22]. These symptoms may occupy relatively central positions within the network and play an important role in the spread of distress to other domains, such as Interpersonal Sensitivity or Hostility. By conceptualizing these dimensions as a network, the present study moves beyond aggregate scores and instead focuses on identifying core and bridge symptoms that contribute to the maintenance of psychological maladjustment, thereby providing a more detailed understanding of how high-pressure educational environments may influence adolescent mental health.

Therefore, this study aims to construct a dynamic network of multidimensional mental health symptoms among Chinese secondary school students by utilizing a large-scale longitudinal dataset. Using cross-lagged panel network (CLPN) analysis, the study examines the temporal relationships among psychological symptoms and identifies potential driving and bridging dimensions within the network. In addition, subgroup analyses are conducted to explore whether the structure and temporal dynamics of the symptom network differ across gender and educational stage. In the present study, educational stage refers to the two educational levels within Chinese secondary education included in our sample—junior high school and senior high school—rather than specific grade years. This approach extends existing research by providing new evidence for understanding the maintenance and development of adolescent mental health symptoms in a high-pressure educational context.

Methods

Participants and procedure

The data were drawn from a large-scale, semesterly mental health screening program coordinated by the district-level education bureau in Guangzhou, China. Data collection for the present study was conducted between September 2023 and April 2024 across two waves of assessment. Using a census approach, the screening intended to include the entire student population (N = 36,810; ages 12–18, predominantly 13–17) of all 22 secondary schools in the district. Data collection was implemented through DiggMind, an online survey platform developed by a China local company. Students completed the survey in school computer labs under the organization of the head teachers and the supervision of mental health education staff. As one class finished the survey and left the lab, another class began. Each session lasted approximately 15 min. Electronic informed consents from both students and their parents or legal guardians were obtained with the assistance of school teachers. Ethical approval for the study was obtained from the Biomedical Research Ethics Committee of Hunan Normal University (Protocol No. 2021–411; approved on May 13, 2021).

In the present study, permission to access and analyze the anonymized data from two time points was obtained in accordance with ethical requirements. The first wave of data collection occurred at the beginning of the 2023 fall semester (September–October 2023; T1). A total of 30,564 secondary school students completed the survey, yielding a participation rate of 83.03%. The second wave took place at the beginning of the spring semester of the same academic year (February–April 2024; T2). At this time, three schools and the graduating-grade students from several schools did not participate, and 20,528 students actually completing the survey, resulting in a participation rate of 77.60%. After matching, a total of 19,082 students were identified as having participated in both waves of the survey. To address longitudinal attrition, participants who did not complete both waves (T1 and T2) were excluded from the analyses. Analyses were therefore conducted using a complete case approach. Data quality was ensured by excluding cases (attrition rate: 9.67%) with erroneous or missing identity information (e.g., student IDs, grade records) and those flagged for careless responding (e.g., participants who scored 0 across all ten dimensions at both assessment waves). This process yielded a valid population-based longitudinal sample of 17,236 participants for the formal analyses. As a mandatory-response online platform was used, the dataset contained no missing values. The sample was relatively balanced by gender (female: 46.32%, male: 53.68%). Of the participants, 11,532 were junior high school students (66.91%), and 5,704 were senior high school students (33.09%). Demographic characteristics are provided in Table S1 of the supplementary materials.

Measures

The MSSMHS was employed at both waves and has been widely used in large-scale screenings in China [12, 23–25], The MSSMHS is school-based and comprises 60 items across 10 psychological dimensions: Obsessive-Compulsive Symptoms, Paranoid Ideation, Hostility, Interpersonal Sensitivity, Depression, Anxiety, Academic Pressure, Maladjustment, Emotional Disturbance, and Psychological Imbalance. Responses were rated on a 5-point Likert scale (0 = “Never” to 4 = “Always”). The factor score for each dimension can be obtained by dividing the total score of each dimension’s scale by the number of items in that dimension.

The MSSMHS was suitable for the present study because it captures multiple domains of psychological functioning that are relevant to Chinese secondary school students. In addition to internalizing symptoms such as Depression and Emotional Disturbance, it also includes interpersonal, behavioral, and school-related dimensions, such as Interpersonal Sensitivity, Hostility, Maladjustment, and Academic Pressure, allowing adolescent mental health to be examined as a multidimensional and contextually embedded system.

In the present study, the MSSMHS demonstrated excellent internal consistency. The Cronbach’s α coefficient for the total scale was 0.958 at T1 and 0.962 at T2. Crucially, the reliability coefficients for the ten individual dimensions remained robust across both time points: at T1, Cronbach’s α values ranged from 0.729 (Obsessive-Compulsive) to 0.908 (Anxiety); at T2, they ranged from 0.759 (Obsessive-Compulsive) to 0.912 (Anxiety). Detailed reliability estimates and descriptive statistics for each dimension at both waves are provided in Table S4.

Data analysis

Data analysis was conducted using SPSS 27.0 and R 4.4.3, estimating CLPN models for the sample. First, autoregressive and cross-lagged effects between the two waves were estimated with regularized regression. Autoregressive effects represent the predictive influence of a symptom on itself from T1 to T2, controlling for other symptoms and covariates at baseline. Cross-lagged effects represent the predictive influence of one symptom (A) at T1 on another symptom (B) at T2, again controlling for other T1 symptoms and covariates. The Least Absolute Shrinkage and Selection Operator (LASSO) penalized maximum likelihood estimation was applied to shrink weak regression paths to zero, thereby reducing overfitting and enhancing generalizability. Regularized regressions were implemented with the glmnet package, and network structures were visualized using the qgraph package.

Although visualizations intuitively display node importance and associations, spatial layouts alone can be misleading. To quantify network structure, two centrality indices were calculated: cross-lagged out-prediction and in-prediction. Out-prediction reflects the extent to which a node at T1 predicts other nodes at T2, whereas in-prediction represents the extent to which a node at T2 is predicted by other nodes at T1 (excluding autoregressive paths).

Network stability was evaluated using the bootnet package. Specifically, 95% confidence intervals for edge weights were obtained from 1,000 nonparametric bootstrap samples, and the correlation stability coefficient (CS) was estimated with a case-dropping bootstrap approach. A CS > 0.25 is considered acceptable, whereas CS > 0.50 indicates high stability. In addition, statistical tests were conducted to examine differences in edge weights and node centrality across groups.

Before conducting gender and stage comparisons, measurement invariance of the MSSMHS was examined. Because all 60 items were rated on a 5-point scale (0–4) and were therefore treated as ordered categorical indicators, invariance testing was conducted at the item level using confirmatory factor analysis (CFA). All models were estimated in lavaan with the WLSMV estimator. Following established practice for ordinal indicators, invariance was tested sequentially at the configural, metric, and scalar (threshold) levels. Given the sensitivity of the chi-square statistic to very large sample sizes, decisions were based primarily on changes in approximate fit indices, with reference to commonly used criteria such as ΔCFI ≤ 0.010, ΔRMSEA ≤ 0.015, and ΔSRMR ≤ 0.030 for metric invariance ≤ 0.015 for scalar invariance [26].

Finally, to examine potential differences in network structures across gender and educational stage, CLPN estimates were combined with comparative procedures following multidimensional analytical strategies commonly used in longitudinal network research [20, 27, 28]: (a) At the global level, correlation coefficients between edge-weight matrices were computed across groups, and the Jaccard similarity index was used to evaluate consistency in link directions, thereby revealing similarities and differences in overall network structure; (b) At the edge level, the distribution and overlap of the strongest connections were compared across networks to identify stable and distinctive structural features; (c) At the node level, consistency in centrality metrics and the composition of core nodes was examined to provide deeper insight into group-specific differences in key functional nodes.

Results

In total, 17,236 questionnaires were eligible for analysis. Regardless of whether it is T1 or T2, the score differences between males and females are significant across all ten dimensions (p < 0.001), with females scoring higher than males. When comparing middle school students with high school students, the differences are also significant (p < 0.001), with high school students scoring higher than middle school students on all ten dimensions (Table S1).

Cross-sectional networks

Figure 1 presented the contemporaneous networks for T1 and T2, with blue edges indicating positive correlations and red dashed edges indicating negative correlations. Tests of kurtosis and skewness indicated that the data were approximately normally distributed.

Fig. 1.

Contemporaneous networks at T1 and T2

In the T1 network, 88.89% (40/45) of all possible edges were retained, while 84.44% (38/45) were retained in the T2 network. The strongest undirected edges at T1 were Depression–Anxiety (r = 0.368), Paranoid Ideation–Interpersonal Sensitivity (r = 0.292), and Interpersonal Sensitivity–Depression (r = 0.204). At T2, the strongest undirected edges were Depression–Anxiety (r = 0.422), Paranoid Ideation–Interpersonal Sensitivity (r = 0.346), and Academic Pressure–Maladjustment (r = 0.222). Full results for all undirected edges had been reported in detailed in Supplementary Table S2.

Figure 2 showed the centrality indices for the contemporaneous networks. At both T1 and T2, Anxiety exhibited the highest strength centrality, while Paranoid Ideation had the highest closeness and betweenness centrality.

Fig. 2.

Centrality indices of the contemporaneous networks at T1 and T2

Stability analyses indicated an ideal robustness. At T1, both edge weights and strength centrality were highly stable (CS = 0.749, 95% CI = [0.673, 1.000]), suggesting the network could tolerate up to 75% case-dropping. At T2, edge weights and strength centrality maintained equivalent stability (CS = 0.749, 95% CI = [0.673, 1.000]), and the expected influence (EI) centrality also demonstrated strong robustness (CS = 0.672). Cross-temporal tests further confirmed high consistency in the network structure between T1 and T2 networks (edge weight matrix correlation r = 0.97), indicating that the stability of the core metrics was largely time-invariant.

CLPN in the total sample

A CLPN model was constructed to examine the temporal associations among the ten dimensions of mental health symptoms. As shown in Fig. 3, the model contained 10 nodes and 95 non-zero directed edges, accounting for 95% of all possible directed edges. The mean edge weight was 0.070, suggesting dense and substantial temporal lagged connections among variables. Cross-lagged paths indicated several robust predictive effects. The strongest was Depression → Anxiety (β = 0.140), complemented by a strong reciprocal path (Anxiety → Depression, β = 0.106). The weight results of the remaining edges can be found in the supplementary materials (Table S3).

Fig. 3.

CLPN of MSSMHS symptoms

Figure 4 displayed the in-prediction and out-prediction values. In-prediction was examined under two conditions. Autoregressive effects were included in the first condition, leading to a medium-to-high temporal stability with values ranged from 0.309 (Psychological Imbalance) to 0.439 (Anxiety). When autoregressive effects were excluded in the second condition, in-prediction dropped so markedly that only Paranoid Ideation (0.135) and Anxiety (0.128) retaining relatively strong predictability. This result implied that there is greater dependence on cross-temporal influences from other symptoms. Out-prediction reflects a node’s predictive influence on other symptoms. With autoregressive effects included, the top three nodes were Hostility (0.266), Academic Pressure (0.257), and Depression (0.254), indicating their roles as primary driving nodes in the temporal dynamics. After excluding autoregressive effects, Depression (0.053), Anxiety (0.037), and Emotional Disturbance (0.034) still showed relatively strong out-prediction, suggesting they may function as independent causal drivers. In contrast, the out-prediction values of Obsessive-Compulsive symptoms, Psychological Imbalance, and Academic Pressure approached zero, indicating their roles as relatively passive responders in the network.

Fig. 4.

Out-prediction and in-prediction values for all nodes in the CLPN

Stability analyses confirmed high robustness. In the total sample, the CS-coefficients for edge weights and the related centrality indices were all 0.594, exceeding the recommended criterion of 0.50, indicating high robustness of the overall network structure and the centrality indices. The CS-coefficients for Out-prediction and In-prediction were 0.75 and 0.55, respectively, both meeting the recommended standard, suggesting good stability of these custom indices in the total sample.

Gender differences

Measurement invariance across gender was examined separately at T1 and T2. The configural models showed acceptable fit at both time points (T1: CFI = 0.931, RMSEA = 0.059, SRMR = 0.043; T2: CFI = 0.924, RMSEA = 0.063, SRMR = 0.043), supporting the basic equivalence of the factor structure across boys and girls. From the configural to the metric model, the changes in RMSEA and SRMR remained within recommended cutoffs, indicating support for metric invariance. However, when threshold constraints were added, CFI decreased by 0.030 at T1 and 0.037 at T2, both exceeding the conventional criterion. These results suggest that full scalar (threshold) invariance was not supported across gender. Overall, the MSSMHS demonstrated acceptable structural and loading equivalence across boys and girls, although some item thresholds may differ by gender.

Figure S1 presented the CLPNs for female and male students. The edge lists of the two networks were highly correlated (r = 0.976, p < 0.001). The Jaccard similarity index for directional consistency was 0.882, indicating a strong similarity between the two networks in both connection patterns and directional predictive pathways. The overall connectivity of the male network (M = 0.032) was slightly higher than that of the female network (M = 0.029), suggesting that male students may be somewhat more prone to symptom persistence or exacerbation.

With respect to the strongest cross-lagged paths, the Depression → Anxiety connection was the most prominent in both males (β = 0.142) and females (β = 0.110). In addition, the Hostility → Depression path showed a relatively strong effect in the female network (β = 0.110). Overall, the two networks were highly consistent in directional patterns but differed somewhat in the strength of their strongest edges.

At the node level, the two networks were significantly correlated in in-prediction (r = 0.89, p < 0.001) but not in out-prediction (r = 0.44, p = 0.203). This suggests that the extent to which symptoms are influenced by others was comparable across genders, whereas their influence on other symptoms diverged. In the male network, Depression (0.058) and Emotional Disturbance (0.047) had the highest out-prediction values, whereas Paranoid Ideation (0.139) and Anxiety (0.136) had the highest in-prediction values. In the female network, Hostility (0.040) and Interpersonal Sensitivity (0.034) showed the highest out-prediction values, while Paranoid Ideation (0.117) and Anxiety (0.113) showed the highest in-prediction values. Detailed centrality metrics were presented in Fig. 5.

Fig. 5.

Out-prediction and in-prediction values for all nodes in the CLPNs of boys and girls students

In the male group, the CS-coefficients for edge weights and the related centrality indices were all 0.517, reaching the recommended criterion (CS > 0.50), which indicates good stability of both the network structure and the centrality indices. The CS-coefficient was 0.70 for Out-prediction and 0.45 for In-prediction, with the latter remaining above the minimum acceptable threshold of 0.25. In the female group, the CS-coefficients for edge weights and the related centrality indices were all 0.439, exceeding the minimum acceptable threshold of 0.25, suggesting that the network structure showed acceptable stability. The CS-coefficient for Out-prediction was relatively high (0.75), whereas that for In-prediction was 0.25, which lies at the lower bound of the acceptable range.

Educational stages differences

Measurement invariance across educational stage was also tested separately at T1 and T2. The configural models showed acceptable fit at both time points (T1: CFI = 0.927, RMSEA = 0.059, SRMR = 0.043; T2: CFI = 0.922, RMSEA = 0.063, SRMR = 0.043), indicating that the MSSMHS had a broadly similar factor structure across junior and senior high school students. The transition from the configural to the metric model did not indicate meaningful deterioration in fit, supporting metric invariance across educational stage. However, after threshold constraints were imposed, CFI decreased by 0.033 at T1 and 0.037 at T2, exceeding the recommended cutoff. Thus, full scalar (threshold) invariance was not established across educational stage. Taken together, these findings indicate that the MSSMHS showed adequate structural and loading comparability across stages, although some item thresholds may vary between junior and senior high school students.

Figure S2 presented the CLPNs for junior and senior high school students. The edge lists of the two networks were highly correlated (r = 0.966, p < 0.001). The Jaccard similarity index for directional consistency was 0.839, indicating strong similarity in both connectivity and causal directions across educational stages. Overall connectivity was comparable between the junior high (M = 0.032) and senior high (M = 0.030) networks, suggesting similar levels of symptom persistence and exacerbation.

In both networks, the strongest cross-lagged path was Depression → Anxiety (β = 0.148 for junior high; β = 0.135 for senior high), which was markedly stronger than any other edge. Overall, the two networks demonstrated high similarity in both edge strength and direction.

At the node level, in-prediction values were strongly correlated across the two networks (r = 0.91, p < 0.001), whereas out-prediction values were not (r = 0.51, p = 0.132). This pattern suggests that the extent to which symptoms are influenced by others is highly consistent across educational stages, while their influence on other symptoms differs. In the junior high network, Depression (0.061) and Emotional Disturbance (0.049) had the highest out-prediction values, whereas Paranoid Ideation (0.156) and Anxiety (0.130) had the highest in-prediction values. In the senior high network, Depression and Hostility showed the highest out-prediction values (both 0.043), while Paranoid Ideation and Anxiety remained the most prominent in-prediction nodes, consistent with the junior high findings. Detailed centrality metrics were shown in Fig. 6.

Fig. 6.

Out-prediction and in-prediction values for all nodes in the CLPNs of junior and high school students

In the junior high school group, the CS-coefficients for edge weights and the related centrality indices reached 0.750 (0.672 for in-strength), indicating very high robustness. The CS-coefficients for Out-prediction and In-prediction were 0.75 and 0.50, respectively, both meeting the recommended standard. In the senior high school group, the CS-coefficients for edge weights and the related centrality indices were all 0.517, exceeding the recommended threshold of 0.50, indicating good network stability. The CS-coefficient was 0.70 for Out-prediction and 0.30 for In-prediction, with the latter remaining above the minimum acceptable threshold of 0.25.

Discussion

This study reveals the temporal relationships of mental health symptoms in middle school students through CLPN analysis. The results show that depression and anxiety exhibit the strongest cross-lagged effects at both T1 and T2, with a significant reciprocal relationship between them. Anxiety shows the highest centrality in the network, acting as a core driving node. Despite minor differences, the structures are largely consistent across gender and educational stage.

The strongest and most recurrent of depression and anxiety in cross-sectional networks

The cross-sectional networks at T1 and T2 exhibited relatively stable associations among the dimensions of the mental health scale. In both networks, the correlation between Depression and Anxiety was particularly strong, consistent with prior findings in adult (18–64 years) and older adult (65+) populations [29], and indicating that this co-occurrence also characterizes Chinese adolescents. Another robust association was observed between Paranoid Ideation and Interpersonal Sensitivity, implying that interpersonal experiences may exert lasting effects on mental health [30]. This aligns with evidence that social stressors and negative emotions can mutually reinforce each other [31]. Although edge weight stability demonstrated strong robustness at both time points, the strengthened link between Academic Pressure and Maladjustment at T2 suggests that the negative influence of academic demands may intensify over time, thereby exacerbating maladaptive outcomes in adolescents [32].

Centrality analyses further highlighted the importance of Anxiety and Paranoid Ideation. Across both time points, Anxiety consistently exhibited the highest strength centrality, implying that anxiety symptoms may continuously activate other clusters of psychological problems through pathways of emotional arousal [33]. Meanwhile, Paranoid Ideation showed the highest closeness and betweenness centrality, indicating that it occupies the shortest transmission paths and acts as a key bridging symptom within the network. Together, these findings underscore the enduring centrality of anxiety and the pivotal role of paranoid ideation in the structure of adolescent mental health symptom networks.

Key predictive roles and temporal interactions in adolescent mental health network

The CLPN analysis revealed temporal causal relationships among adolescent mental health symptoms. The strongest cross-lagged effect was Depression → Anxiety, accompanied by a robust reciprocal path (Anxiety → Depression). These findings are consistent with previous studies showing high reciprocity and frequent co-occurrence of depression and anxiety among adolescents [34–36]. From a developmental psychopathology perspective, the strong reciprocal relationship between Depression and Anxiety is not incidental. Adolescence is a period marked by heightened emotional reactivity, increased sensitivity to social evaluation, and still-maturing emotion regulation capacities, all of which may increase vulnerability to internalizing symptoms [37, 38]. Within this framework, depression and anxiety can be understood as developmentally linked manifestations of shared vulnerability that mutually reinforce each other over time, rather than as isolated symptom domains [39]. This interpretation is also consistent with prior developmental and network-based research showing that depression and anxiety remain closely interconnected across childhood and adolescence [40].

In addition, network-based research has shown that symptoms of depression and anxiety tend to be highly interconnected throughout development, often forming a tightly linked internalizing symptom cluster rather than completely separate constructs [41]. This strong connectivity may allow emotional distress to spread across symptoms, thereby increasing the persistence and continuity of internalizing problems during adolescence. From a developmental perspective, adolescence is a period characterized by heightened emotional reactivity, identity exploration, and sensitivity to social evaluation [42, 43]. These developmental challenges may simultaneously activate depressive cognitions (e.g., feelings of hopelessness or low self-worth) and anticipatory worry about future outcomes, further strengthening the bidirectional relationship between depression and anxiety [44].

Notably, while the strong coupling of these symptoms is theoretically grounded, the reliance on self-reported data may partially amplify this association due to common method variance, as students might perceive and report internalizing symptoms through a similar subjective lens [45].

Anxiety also strongly predicted Obsessive-Compulsive Symptoms, further supporting evidence that anxiety disorders are closely linked to obsessive-compulsive features [46]. In addition, significant predictive effects were observed for Interpersonal Sensitivity → Paranoid Ideation and Depression → Hostility, suggesting that interpersonal stress and negative emotions may contribute to the emergence of paranoid ideation and hostile behaviors [47]. More broadly, the centrality of Depression, Anxiety, and Emotional Disturbance supports the developmental psychopathology view that emotional dysregulation often functions as an upstream vulnerability process [48]. These results not only enrich theoretical understanding but also imply practical directions: educators and counselors may prioritize training in emotion regulation to mitigate the impact of negative affect on adolescents.

Interestingly, several negative cross-lagged effects were identified. Paranoid Ideation negatively predicted Academic Pressure and Obsessive-Compulsive Symptoms, which may reflect defensive or compensatory processes. Classic views by emphasized that individuals often employ immature defense mechanisms under stress to regulate emotions [49]. Future research may further explore how such mechanisms could be integrated into interventions, for example by applying cognitive-behavioral therapy to help students recognize and reframe maladaptive emotional responses.

All symptoms demonstrated significant autoregressive stability (high in-prediction). After removing autoregressive paths, however, Paranoid Ideation and Anxiety retained relatively high predictability, indicating that these symptoms were more strongly shaped by earlier manifestations of other symptoms than by their own prior states. Thus, they function as relatively dependent or passive nodes within the network. In contrast, out-prediction identified potential causal drivers. With autoregressive effects included, Hostility, Academic Pressure, and Depression showed the strongest predictive influence. More importantly, after excluding autoregressive paths, Depression, Anxiety, and Emotional Disturbance continued to exhibit relatively high out-prediction, highlighting their roles as core driving nodes across time.

Taken together, the emotional nodes Depression, Anxiety, and Emotional Disturbance not only maintained high autoregressive stability but also exerted independent influence on other symptoms, reinforcing their roles as central hubs in the network. In particular, Anxiety showed strong values on both in-prediction and out-prediction, underscoring its dual role as both a highly influenced and highly influential node in adolescent psychopathology. This finding demonstrates a marked divergence from results obtained through SCL-90 studies in vocational student populations [20], potentially originating from fundamental disparities in the academic environments characterizing these distinct groups. The intensified emphasis on academic performance and competitive college entrance examinations within China’s general secondary education system effectively establishes Anxiety as a central hub in adolescents’ psychological symptom networks. By contrast, Paranoid Ideation exhibited high in-prediction but low out-prediction, indicating its susceptibility to other symptoms and supporting prior research that paranoia often functions as an adaptive by-product of the social-cognitive defense system and a derivative of other negative emotions [50]. The apparent influence of Academic Pressure and Hostility largely reflected autoregressive inertia; once this component was removed, their external impact diminished, suggesting that they function more as stable background factors than as active drivers in the symptom network.

Finally, stability analyses confirmed the robustness of the CLPN model. The CS coefficient for edge weights reached 0.75, while all centrality indices exceeded the recommended threshold of 0.50. These results indicate that the network structure identified in this study is highly reliable under different sampling conditions, providing a solid theoretical basis for future research on adolescent mental health.

Differences in network structure by gender

Comparison of the CLPNs for males and females revealed strong consistency in both overall structure and causal direction. This suggests that the basic organization of symptom interactions was broadly similar across genders, indicating that gender may not substantially alter the general dynamic framework of adolescent mental health dimensions. Prior studies have likewise suggested that gender disparities in adolescent mental health may be narrowing, possibly in relation to shifts in traditional female gender roles [51]. However, this overall similarity should not be taken to mean that gender-related differences were entirely absent. Existing research on adolescent symptom networks has yielded mixed findings. Some studies have identified gender-specific differences in particular symptom pathways or central nodes [52], whereas others have found that male and female networks are comparable in global connectivity, despite differences in the relative importance of specific nodes [53]. Taken together, these findings suggest that gender-related variation may be more evident at the local level of specific symptoms or pathways than at the level of the overall network structure. This interpretation is also consistent with the analytic strategy used in the present study. Our conclusion was based on global-level comparison indices, including edge-matrix correlations and the Jaccard similarity index for directional overlap. These indices primarily reflect the extent to which two networks share a common overall configuration. Accordingly, when most directional links overlap across groups, the networks may appear globally similar even though meaningful differences remain at the edge and node levels.

More specifically, although the two networks were broadly similar at the global level, several differences emerged at other levels. The male network displayed slightly higher overall connectivity compared to the female network, suggesting that male adolescents may be more prone to forming mutually reinforcing symptom clusters. From a developmental perspective, males often face stricter social constraints on emotional expression and may lack flexible emotion regulation strategies [54]. Under stress, they are more likely to rely on passive strategies such as acceptance, suppression, or denial [55]. Such coping styles may intensify reciprocal reinforcement among symptoms, thereby increasing the risk of comorbidity. Hence, the higher connectivity in the male network may reflect not only stronger symptom persistence but also the influence of gender role norms on psychopathological processes.

Analysis of the strongest cross-lagged paths indicated that Depression → Anxiety was the most prominent connection in both networks. Unlike cross-sectional analyses that capture correlations, cross-lagged modeling demonstrates that depression at an earlier time point significantly predicts later anxiety, consistent with prior adolescent studies [27]. Given the central role of depression–anxiety comorbidity in youth psychopathology, this finding underscores the importance of early detection and intervention for depression to prevent subsequent anxiety disorders.

Notably, in the female network, the path Hostility → Depression (β = 0.110) emerged as particularly strong, suggesting that hostile emotions may play a salient role in the development of depression among females. This pathway was not significant in the male network, potentially reflecting gender-specific variations in social expectations and role-related stress.

At the node level, the two networks showed high correlation in in-prediction (r = 0.89, p < 0.001), indicating that the degree to which symptoms are influenced by others was similar across genders. By contrast, the correlation for the out-prediction was weaker. In the male network, Depression and Emotional Disturbance had the highest out-prediction values, suggesting that negative emotional states may serve as core driving forces in male psychopathology [2, 37]. In the female network, Hostility and Interpersonal Sensitivity had the highest out-prediction values, consistent with evidence that adolescent girls are more sensitive to interpersonal relationships and social evaluation [56]. Thus, interpersonal conflict and sensitivity may more readily trigger subsequent psychological difficulties in females.

From a developmental psychopathology perspective, gender differences in symptom networks may be better understood not as fixed categorical distinctions, but as differences in developmental sensitivity to particular stressors and interpersonal contexts [57]. During adolescence, socialization processes, emotion-expression norms, and heightened concerns about peer evaluation may shape which symptom pathways become more salient in boys versus girls [58]. In this sense, the prominence of Hostility and Interpersonal Sensitivity in girls may reflect a stronger embedding of distress within relational and evaluative contexts, whereas the stronger clustering of emotional symptoms in boys may reflect different patterns of emotional suppression or less flexible emotion regulation under stress [59].

Despite these differences, both networks showed consistently high in-prediction values for Paranoid Ideation and Anxiety, indicating that these symptoms function as key convergence points across genders and may represent universal intervention targets. Together, these findings suggest that while certain mechanisms are gender-specific, tailoring intervention strategies to gender-linked patterns could enhance the effectiveness of adolescent mental health support.

Specifically, based on the higher connectivity observed in the male network, interventions for boys should prioritize de-clustering symptoms to prevent the rapid spread of distress across the network. Since Depression and Emotional Disturbance act as primary drivers for males, Cognitive Behavioral Therapy that emphasizes emotional literacy may be especially relevant as a candidate intervention framework in weakening the reinforcing loops between internalizing symptoms [60–62]. In contrast, for female students, the prominent Hostility → Depression path and the high out-prediction of Interpersonal Sensitivity suggest a relational-vulnerability model. Therefore, Interpersonal Psychotherapy for Adolescents may be a promising intervention approach worth testing in future trials [63, 64]. Educators should focus on developing girls’ relational aggression management and social conflict-resolution skills, thereby blocking the specific developmental pathway where interpersonal hostility transitions into depressive withdrawal.

Differences in network structure across educational stages

This study compared the mental health network structures of junior and senior high school students using CLPN analysis. Results indicated a high edge-list correlation between the two groups, suggesting considerable consistency in both connection patterns and causal directions across educational stages. Overall connectivity was also comparable, implying that symptom persistence and mutual reinforcement may occur to a similar degree in both groups. Although adolescents typically experience substantial social, academic, and physiological transitions from junior to senior high school, the present findings suggest that the basic organization of symptom interactions remains relatively stable across stages. This pattern may reflect the fact that internalizing symptoms such as depression and anxiety are tightly interconnected throughout adolescence, forming a broadly shared developmental psychopathology framework rather than entirely stage-specific structures [43, 65].

Importantly, this stage-related similarity does not mean that contextual influences such as academic pressure are irrelevant. Existing evidence indicates that academic pressure is associated with adolescent depression, anxiety, and related difficulties, and may become more salient in later school years [66]. However, such pressure may intensify distress within an existing symptom system rather than fundamentally reorganize the overall network architecture. Previous research has shown that internalizing symptoms may retain a relatively stable overall organization across adolescence even when contextual stressors vary [41]. In other words, educational-stage differences may be expressed more in the relative prominence of particular pathways and symptom nodes than in the global structure of the network. This interpretation is consistent with the present findings, in which marked global differences were not observed, but several meaningful edge-level and node-level distinctions still emerged.

More specifically, several differences still emerged at the edge and node levels. In both groups, the strongest cross-lagged path was Depression → Anxiety, highlighting a robust and universal relationship across adolescence. Wang et al. (2025) similarly observed strong associations between Depression and Anxiety across different educational stages. However, compared to the cross-sectional network analysis employed in their study, the cross-lagged panel network (CLPN) approach applied in our research more effectively elucidates the predictive effect of Depression on Anxiety.This suggests that depression may serve as a key driver of anxiety, representing an inherent developmental pathway rather than a stage-specific phenomenon. As Erikson noted, adolescence is a period of identity formation, when individuals strive to answer “Who am I?” through exploration and commitment [67]. Depression may profoundly disrupt this process: loss of motivation, diminished interest, and negative self-perceptions (e.g., worthlessness, hopelessness) can drain the psychological resources needed for self-exploration, thereby exacerbating anxiety about the future [68]. Notably, the depression–anxiety link was slightly stronger among junior high students, suggesting that this pathway may peak earlier in adolescence. Thus, junior high may represent an optimal window for intervention, focusing on strengthening identity development and positive cognitive strategies to prevent the escalation from depression to anxiety.

Centrality analyses further showed that in-prediction was highly correlated between stages, whereas out-prediction was less so. In the junior high network, Depression and Emotional Disturbance were the main out-prediction nodes, consistent with the pronounced mood swings and immature self-regulation typical of early adolescence [69]. Emotional disturbances in this stage may more easily spread to other psychological domains [70]. By contrast, in the senior high network, Depression and Hostility emerged as primary out-prediction nodes, reflecting how, under increased academic and social pressures, depressive experiences may be externalized into behavioral expressions such as hostility. Despite the unequal subgroup sizes, the smaller senior high network still included more than 5,700 participants, which is substantially larger than the samples used in many previous CLPN studies. Together with the bootstrap stability results, this suggests that the stage-specific network estimates were robust despite the unequal subgroup sizes.

Developmental psychopathology further suggests that similar underlying vulnerabilities may manifest differently across developmental stages because the meaning and organization of stressors change with age [38]. In early adolescence, emotional disturbance may spread more easily across symptom domains because affective instability and regulatory immaturity are still pronounced [71]. By late adolescence, however, the accumulation of academic demands, future-oriented concerns, and identity-related pressure may reorganize distress into pathways in which hostility or maladjustment becomes more prominent [71, 72].

Overall, while Depression remained a central risk factor across both stages, its mode of influence evolved with age. Intervention strategies should therefore be developmentally tailored. For junior high students, the focus should be on emotional stabilization. Given that Emotional Disturbance is the core driving node at this stage, school-based Social-Emotional Learning programs should prioritize basic distress tolerance and impulse control [73]. Such primary prevention can act as a circuit breaker, preventing early emotional volatility from activating more stable symptoms like Anxiety. For senior high students, the emergence of Hostility as a primary out-prediction node—likely fueled by intensified academic pressure. Interventions such as Dialectical Behavior Therapy(DBT) skills training, can help senior students manage academic-stress-induced hostility [74]. Furthermore, since Academic Pressure’s influence on Maladjustment intensifies at this stage, cognitive reframing focused on “achievement-related self-worth” is essential to buffer the impact of high-stakes examinations on the symptom network [75].

Limitations and future research directions

Although this study provided a preliminary exploration of the dynamic network of adolescent mental health, several limitations should be acknowledged.

First, the analyses were based on only two time points (T1 and T2), which restricted the ability to capture long-term developmental trajectories. Future research should incorporate additional waves of data to more comprehensively examine temporal changes and dynamic patterns in adolescent mental health networks.

Second, the present study focused on selected psychological dimensions from the MSSMHS. To achieve a more comprehensive understanding of adolescent psychopathology, future work should integrate additional domains, including biological markers and social-contextual variables, thereby enabling a multidimensional exploration of how psychological, biological, and social factors interact within adolescent mental health networks.

Third, this study relies exclusively on self-reported measures, which may introduce response bias and social desirability bias. Specifically, adolescents might underreport certain symptoms to conform to social norms or teacher expectations, or they may overreport symptoms during periods of high academic stress due to negative affectivity [76]. This subjectivity could potentially lead to an overestimation of the connectivity between certain symptoms in the network. Future studies should consider a multi-informant approach—integrating reports from parents, teachers, and peer evaluations—or utilizing clinical interviews to triangulate self-reported data and enhance the objectivity of the symptom networks.

In addition, the educational-stage subgroups in the present study were not equally sized, with junior high school students representing approximately two-thirds of the final sample. Although this distribution reflected the underlying composition of the district-wide screened population and both subgroups remained large in absolute terms (with more than 5,700 participants in the smaller subgroup), unequal subgroup sizes may still influence the relative precision of certain network estimates. Nevertheless, bootstrap stability analyses indicated that both stage-specific networks were stable, and the overall network structures showed high similarity across educational stages. Future studies using more balanced longitudinal samples could further examine whether the observed stage-specific patterns remain consistent.

Finally, it is important to note that the gender-specific and stage-tailored intervention strategies proposed in this study are primarily inferred from the topological properties of the symptom networks and supported by existing theoretical frameworks. These proposals should be viewed as preliminary hypotheses intended to provide a heuristic direction for precision mental health support rather than clinically validated protocols. Future research should employ Randomized Controlled Trials to empirically test the effectiveness of these targeted intervention pathways in school settings. Such longitudinal experimental evidence is essential to transition from network-based diagnostic insights to evidence-based clinical practice.

Conclusion

Collectively, these findings provide a comprehensive and dynamic portrait of adolescent mental health. It found that Anxiety and Depression are the core symptoms in adolescents’ mental health networks, which are closely interconnected. Paranoid Ideation primarily plays a bridging role, connecting other symptoms rather than acting as a central driver.

Gender and educational stage differences were observed. Although the network structure was similar across genders, males showed stronger connections between Depression and Emotional Disturbance, while females were more influenced by Hostility and Interpersonal Sensitivity. In terms of educational stages, the relationship between Depression and Anxiety remained constant, but its strength varied. In junior high students, Emotional Disturbance was a major issue, while Hostility became more prominent in senior high students.

Targeting Anxiety and Depression as a central intervention point in school-based mental health programs. Tailoring interventions to gender and developmental stage differences can further enhance their effectiveness. This study contributes to a deeper understanding of adolescent mental health, emphasizing the need for targeted prevention and intervention strategies.

Abbreviations

CLPN

Cross-Lagged Panel Network

MSSMHS

Chinese Middle School Students’ Mental Health Scale

SCL-90

Symptom Checklist-90

COVID-19

Coronavirus Disease 2019

T1

Time Point 1 (First assessment wave)

T2

Time Point 2 (Second assessment wave)

LASSO

Least Absolute Shrinkage and Selection Operator

CS

Correlation Stability (coefficient)

EI

Expected Influence

CI

Confidence Interval

SPSS

Statistical Package for the Social Sciences

N

Sample size

p

Probability value (Statistical significance)

r

Pearson correlation coefficient

β (Beta)

Standardized regression coefficient

M

Mean (used in text for network connectivity)

α (Alpha)

Cronbach’s alpha (Internal consistency)

Authors’ contributions

W. Zhang, R. Gao and X. Zheng contributed equally to this work. W. Zhang: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Data curation, Writing - original draft, Visualization. R. Gao: Software, Validation, Formal analysis, Investigation, Data curation, Writing - original draft, Visualization. X. Zheng: Conceptualization, Validation, Formal analysis, Investigation, Data curation, Writing - original draft, Visualization.Y. Guo and L. Ma supervised the project and are co-corresponding authors. Y. Guo: Conceptualization, Resources, Writing - review & editing, Supervision, Project administration. L. Ma: Methodology, Resources, Writing - review & editing, Supervision, Project administration, Funding acquisition.

Funding

This work was supported by the National Center for Mental Health, China [grant number GXWH2025A-012]; the Innovation Team Project of Guangdong Ordinary Universities [grant number 2022WCXTD004]; the Guangdong Education Science Planning Project (Higher Education Special) [grant number 2024GXJK123]; the Guangzhou Philosophy and Social Sciences Development Project of the 14th Five-Year Plan (Co-construction Project) [grant number 2024GZGJ277]; and the Guangdong Provincial Science and Technology Program [grant number 2024A1111120020].

Data availability

To protect study participant privacy, the data used in this study are not publicly available but will be made available upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki. Ethical approval was obtained from the Biomedical Research Ethics Committee of Hunan Normal University (Protocol No. 2021–411). The approval date was May 13, 2021. Informed consent was obtained from all participants.All procedures involved in this research were observational and did not include any clinical intervention. Informed consent was obtained from all participants and/or their legal guardians prior to data collection. Clinical trial registration was not applicable, as this study did not involve any intervention.

Consent for publication

Not Applicable.

Competing interests

The authors declare no competing interests.