Adolescent school absenteeism, depressive symptoms, and internalizing/externalizing problems: a network and simulated intervention study based on CFPS Authors: Xueti Dong, Mingwan Zhou Published: 2026 ABSTRACT BackgroundSchool absenteeism is closely associated with adolescent mental health problems, yet little is known about the covariation patterns between the short-term, mild attendance fluctuations commonly observed among community adolescents and their psychological distress. This study aimed to combine network analysis with simulated intervention methods to explore the conditional dependence network among absenteeism, depression, and internalizing/externalizing problems, and to preliminarily identify potential intervention targets.MethodsData were drawn from the 2022 wave of the China Family Panel Studies (CFPS), including 2,734 community adolescents. The sample was first randomly split into an exploratory sample (N = 1,367) and a validation sample (N = 1,367). Exploratory and confirmatory factor analyses were conducted to examine the reliability and validity of the CES-D-8 and the internalizing/externalizing scales. After controlling for covariates, the EBICglasso method was used to estimate the network structure in the effective sample (N = 1,894), and expected influence (EI) and bridge expected influence (bEI) were calculated. To evaluate the sensitivity of results to dichotomization cut-points, Ising models were constructed using both lenient and strict cut-points. The NodeIdentifyR algorithm was then applied to perform simulated interventions (node thresholds shifted by ± 2 standard deviations), with the network’s overall expected total score serving as the outcome measure.ResultsAfter removing seven low-loading items, the measurement model showed good overall fit (CFI = 0.979, RMSEA = 0.031). Network analysis identified CESD5 (“I felt lonely”), CESD7 (“I felt sad”), and CESD2 (“I felt that everything I did was an effort”) as central nodes. Bridge nodes included CESD2 (“I felt that everything I did was an effort”), CESD3 (“My sleep was restless”), and N (absenteeism frequency), with 84% of participants reporting zero absences. In worsening simulated interventions, results were highly consistent across both cut-points: CESD5 (loneliness) produced the largest increase in the network’s overall expected total score. In easing simulated interventions, the node with the largest decrease differed by cut-point: Q6 (“I got distracted easily”) under the lenient cut-point, and CESD1 (“I felt depressed”) under the strict cut-point.ConclusionsThis study revealed a covariation network among short-term absenteeism fluctuations, depression, and internalizing/externalizing problems in community adolescents. Loneliness was not only a central node in the network but also a strong predictor in worsening interventions, whereas being easily distracted and depressed mood emerged as candidate directions for easing interventions that warrant further investigation. Given the cross-sectional design, the merging of different reasons for absenteeism into a single indicator, and the zero-inflated skewed distribution of the data, all conclusions should be regarded as exploratory hypotheses pending validation through future longitudinal studies or experimental designs.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40359-026-04834-6. Background School absenteeism (including excused and unexcused absences) is a common maladaptive behavior among school-aged adolescents. Chronic absenteeism not only results in loss of learning time but also significantly affects academic performance [1–2], physical health [3], and mental health [4–5]. In recent years, a growing number of studies have advocated for treating absenteeism as a public health issue [4, 6]. Identifying the causes of chronic absenteeism and developing intervention strategies are of considerable practical importance [7]; however, a more critical question lies in revealing how absenteeism gradually accumulates and recurs. In daily school contexts, seemingly justified absences among adolescents may be behavioral indicators closely linked to psychopathology. Foundational research on the developmental cascade model indicates that early problems can trigger subsequent problems across domains through cascading mechanisms: marked externalizing behavior problems in childhood impair the normal development of academic competence, which in turn exacerbates academic difficulties and internalizing symptoms in adolescence [8]. Accordingly, adolescents with chronic absenteeism frequently co-exhibit internalizing and externalizing problem behaviors [5]. Empirical studies in the absenteeism literature also show that, among various internalizing symptoms, depression has a particularly prominent independent predictive effect on unexcused absences [9]. Clark and Watson’s (1991) tripartite model posits that depression possesses specific pathological features distinct from general internalizing problems, such as anhedonia and low positive affect [10]. Therefore, to more precisely examine the unique role of depression in the absenteeism network, this study did not merge it into a composite “internalizing problems” index; instead, based on the independent measurement of depressive symptoms by the CES-D scale, depressive symptoms were included as a separate set of nodes in the network model. It should be noted that even with this distinction, the network still retains internalizing symptom nodes such as academic worry and shame, which share some semantic proximity with depressive symptoms. This is both an inherent characteristic of the measurement instruments and entirely consistent with the methodological property of network analysis that does not require nodes to be orthogonal; retaining these nodes is precisely intended to further explore potentially differentiated associations between different internalizing symptoms and normative absenteeism. Network analysis has already been employed to examine the behavioral mechanisms of chronic/pathological absenteeism. Çimen et al. constructed a network model of adolescent depression, anxiety, stress, academic achievement, and absenteeism [11]; Alanko et al. analyzed the symptom network of adolescents with absenteeism problems [12]. However, prior research has mostly been limited to identifying central nodes in the network, whereas central nodes only reflect the sum of the strength of associations among variables and cannot be directly equated with effective intervention targets. Simulated intervention, as an exploratory statistical method, can systematically alter the parameters (e.g., thresholds) of a single node and generate simulated data to evaluate the simulated change in the overall symptom activation level of the network after intervening on a given node, thereby screening candidate nodes with potential intervention value [13]. Simulated intervention includes both worsening and easing strategies: the former can indicate nodes deserving further attention from a prevention perspective, while the latter helps generate testable hypotheses about clinical intervention targets. Compared with previous research, the originality of this study lies in the following four aspects. (1) Research focus: Unlike existing studies that concentrate on chronic pathological absenteeism, this study focuses on the common short-term, normative attendance fluctuations among community adolescents, exploring the covariation network between mild absenteeism and psychological distress. (2) Sample and method: This study is based on a large-scale community adolescent sample from China (CFPS) and is among the first empirical studies in China to use a simulated intervention approach to investigate normative absenteeism and psychological problems. This method overcomes the limitation of merely reporting centrality indices and enables the screening of candidate nodes with potential intervention value. (3) Construct differentiation: Drawing on Clark and Watson’s tripartite model, this study extracts depressive symptoms separately from internalizing/externalizing problem behaviors to more precisely examine their unique association patterns within the absenteeism network. The findings can provide exploratory evidence for the early identification and precision intervention of normative absenteeism and concomitant psychological problems among community adolescents. It should be noted that this study does not presuppose absenteeism as a “consequence” or “cause” of psychological distress; rather, absenteeism is treated as a behavioral indicator and included in the network together with depression and internalizing/externalizing problems in order to depict their covariation patterns. Methods Participants The data for this study were drawn from the China Family Panel Studies (CFPS), a nationwide, comprehensive longitudinal social survey. The CFPS began its preparatory work in 2007 and implemented the baseline survey nationwide in 2010. The project conducts surveys biennially, with the latest data released in 2022. This study utilized data from the 2022 wave, comprising 2734 adolescents who completed both the Center for Epidemiologic Studies Depression Scale (CES-D) and the internalizing/externalizing problem behavior scale. All participants (or their legal representatives) provided informed consent for the survey, and the project was approved by the Biomedical Ethics Committee of Peking University, China (IRB00001052–14010). As this study involved secondary analysis of anonymized data, no additional ethical approval was required. Measures Absenteeism frequency Absenteeism frequency was collected via a self-designed fill-in-the-blank question: “In the past month (excluding summer and winter vacations), how many times in total have you taken leave or skipped class?” The questionnaire restricted the response range to 0–20 times. This question combined excused absences and unexcused absences into a single inquiry, and the resulting variable reflects the respondent’s overall absenteeism frequency over the past month. Although the CFPS questionnaire included a follow-up question on the specific reasons for absenteeism, only 16% of participants reported specific reasons, which precluded meaningful disaggregation. Therefore, due to data limitations, this study could not differentiate the specific reasons for absenteeism; absenteeism was treated as a general attendance behavior indicator, and only the total number of absences was used as the analysis variable. Center for epidemiologic studies depression scale (CES-D) The Center for Epidemiologic Studies Depression Scale (CES-D), developed by Radloff et al. [14], was used to assess the frequency of current depressive symptoms. Respondents reported the frequency of symptoms experienced in the past week. The full scale contains 20 items rated on a 4-point scale: “rarely or none of the time (less than 1 day)” = 1, “some or a little of the time (1–2 days)” = 2, “occasionally or a moderate amount of time (3–4 days)” = 3, and “most or all of the time (5–7 days)” = 4. Higher scores indicate higher frequency of depressive symptoms. In the CFPS project, the CES-D was administered using a mixed approach of 8-item and 20-item versions in the 2016 wave; from 2018 onward, only the 8-item version was used, comprising items 6, 7, 11, 12, 14, 16, 18, and 20 of the original scale. The present study used data from the 2018 and 2022 waves, both of which employed the 8-item version. The eight selected items were sequentially renumbered as CESD1 to CESD8, with CESD4 and CESD6 requiring reverse coding. Specific items and their abbreviations are presented in Table S1. Internalizing/externalizing problem behavior Internalizing and externalizing problem behaviors were assessed using a self-developed scale created by the China Family Panel Studies (CFPS) project team. I Internalizing problems included: “I got angry when facing difficulties in my studies”, “I was afraid of exams”, “I often felt lonely”, “I often felt sad and upset”, “I worried about not being able to complete my homework” and “I felt ashamed when I make mistakes at school”. Externalizing problems included: “I often argued with my peers”, “I had trouble concentrating”, “I got distracted easily”, “I struggled to finish my school assignments”, “I worried about not having friends to play with at school”, “I got into trouble for talking too much and disturbing others"and “I got into trouble for fighting with my peers”. The scale uses a 5-point rating: “completely disagree” = 1, “somewhat disagree” = 2, " moderate” = 3, “somewhat agree” = 4, and “completely agree” = 5. The “moderate” option was not read aloud in the CFPS data; therefore, the actual available response options were only 1, 2, 4, and 5. In this study, the items were numbered Q1 through Q14. Specific items and their abbreviations are presented in Table S1. Given that some items within the internalizing problem domain (e.g., “I often feel sad or distressed”) exhibit semantic overlap with the content of the CES-D, exploratory factor analysis (EFA) was conducted prior to formal analysis to screen items and reconstruct dimensions, thereby ensuring that the internalizing/externalizing problem behavior scale does not contain depressive components and avoiding construct confounding. Statistical analysis Sample size changes and processing The analysis in this study was conducted in two steps. First, the full valid sample (N = 2,734) was split into an exploratory sample (N₁ = 1,367) and a validation sample (N₂ = 1,367) to perform exploratory and confirmatory factor analyses on the CES-D and the internalizing/externalizing problem behavior scale, respectively, in order to examine the structural validity and reliability of the scales. Subsequently, after including covariates (age, sex, grade, elite class, class ranking, school location, per capita household income, father’s education level, mother’s education level, father’s depression total score, and mother’s depression total score), the final sample size for network analysis was N = 1,894 due to missing data on some covariates. Prior to network estimation, the influence of each covariate on each node was removed through regression. Basic information Descriptive statistics for demographic variables were calculated using SPSS 28.0, including descriptive statistical analysis. Factor analysis This study employed the Center for Epidemiologic Studies Depression Scale (CESD-8) and the self-developed internalizing/externalizing problem behavior scale. Because some items in the internalizing/externalizing problem behavior scale (e.g., Q5 “I often feel lonely,” Q7 “I often feel sad or distressed”) may overlap in content with the depression scale, and because the internalizing/externalizing problem behavior scale is a self-developed instrument lacking established reliability and validity evidence, this study first conducted exploratory factor analysis (EFA) to examine the factor structure of the two scales, identify and remove items with cross-scale overlap, and thereby optimize the scale structure. Subsequently, confirmatory factor analysis (CFA) was performed to verify that the factor structure of the two scales was sufficiently stable. EFA was conducted using SPSS 28.0, and CFA was performed using the lavaan package in R. To test the stability of the scale structure, the total sample (N = 2734) was randomly divided into two subsamples using the random sampling function in SPSS: an exploratory sample (n₁ = 1367) and a validation sample (n₂ = 1367). Model fit indices, composite reliability (CR), average variance extracted (AVE), and Cronbach’s α coefficients were calculated to evaluate the reliability and validity of the scales. Network analysis To ensure complete covariate data, the sample size included in the network analysis was N = 1,894. The covariates controlled in this study included: age, sex, grade, elite class, class ranking, school location, per capita household income, father’s education level, mother’s education level, father’s depression total score, and mother’s depression total score. Absenteeism (the total number of excused or unexcused absences in the most recent month that was not a winter or summer vacation) exhibited a severely zero-inflated skewed distribution (84.64% of participants reported no absences), and it was approximately treated as a continuous variable in this study. It should be noted that linear regression assumes normally distributed residuals, and a zero-inflated count variable does not fully satisfy this assumption; this treatment is an exploratory approximation, and we have addressed this in the limitations section of the discussion. Prior to network analysis, for each core variable (including the eight CES-D items, the remaining items of the internalizing/externalizing problem behavior scale, and the absenteeism item), a linear regression model was constructed with the above covariates as independent variables, and the residuals were extracted as the node variables after removing the covariate effects. All subsequent network analyses were based on the residual correlation matrix. It should be noted that in cross-sectional network analysis, all nodes are treated as variables of equal status, and the edge weights reflect conditional dependence relationships (i.e., the partial correlation between two variables after controlling for all other nodes), rather than causal directions. Therefore, the network model in this study does not presuppose any causal order among the variables. Each item of absenteeism behavior (the total number of excused and unexcused absences in the past month, excluding summer and winter vacations), the CES-D, and the self-developed internalizing/externalizing problem behavior scale (excluding items overlapping with the depression scale) was included as a node in the network to construct the network structure. All network analyses were performed in R 4.2.1. The polychoric correlation matrix was computed using the psych package, and the Gaussian graphical model was estimated using the EBICglasso() function from the qgraph package [15–16]. The gamma hyperparameter was set to 0.5. To ensure the robustness of the results, the correlation stability (CS) coefficients for Expected Influence (EI) and bridge Expected Influence (bEI) under different gamma values are reported in the Supplementary Materials. Expected influence (EI) is an important index for measuring the strength of conditional associations of a node within a network [17]. Nodes with higher EI values have stronger covariation relationships with other nodes in the network, reflecting their relative importance in the conditional dependence structure. In addition, to explore the conditional association patterns between absenteeism and depression as well as internalizing/externalizing problem behaviors, we conducted bridge centrality analysis, with the specific index being bridge expected influence (bEI) [18]. Nodes with higher bEI values exhibit stronger statistical associations in connecting different communities (i.e., absenteeism, depression, and internalizing/externalizing problem behaviors) and can be regarded as potential structural bridges. Bridge centrality indices were calculated using the bridge function from the networktools package. It should be noted that the above centrality indices are all based on cross-sectional regularized network estimation and reflect conditional dependence relationships among variables; they should not be directly interpreted as causal maintaining effects or dynamic influences. To ensure the stability of the above results, we used the bootnet package to perform case-dropping bootstrap procedures with 1,000 bootstrap samples to assess the stability of nodes in the network structure and to quantify the stability of EI and bEI. A CS coefficient value > 0.50 indicates good stability [15]. Edge stability was evaluated by calculating 95% confidence intervals using nonparametric bootstrapping (1,000 iterations); narrower confidence intervals indicate higher edge accuracy [19]. Considering that weak discriminant validity of the scales may affect the network structure, we additionally constructed a network that included all original items (absenteeism, all eight CES-D items, and all 14 items of the internalizing/externalizing problem behavior scale) as a sensitivity analysis. The Spearman’s rank correlation coefficient of edge weights for common edges between the full-item network and the primary analysis network (after removing low-loading items) was calculated to evaluate the robustness of the network structure to node selection. The results of this sensitivity analysis will be presented in the network analysis section. Simulated intervention Simulated intervention was used to identify nodes within the network structure that are expected to be the most effective intervention targets for specific symptoms. This method systematically alters the network parameters of a single node and generates simulated data to examine the expected effects of intervening on that specific node, with intervention efficacy evaluated by the change in overall network activation following perturbation. Since the NodeIdentifyR algorithm used for simulated intervention is only applicable to binary data [13], and following existing research [20], as well as to ensure reasonable cut-points, this study applied two dichotomization methods to all items. Method 1 (lenient cut-point): Absenteeism was coded as 0 (absent) / 1 (present). Each CES-D item was dichotomized based on the original score: 0 = original score of 1 (“rarely or none”), 1 = original scores of 2, 3, or 4 (“sometimes,” “often/half the time,” “most of the time/continuously”). This cut-point treats any frequency above “rarely or none” as symptom presence, aiming to enhance sensitivity for detecting mild symptoms and is suitable for exploratory analysis in community general populations. The self-developed internalizing/externalizing problem behavior scale (after removing six items; originally a 5-point rating, but the “moderate” option was not read aloud in the CFPS data) was dichotomized as follows: 0 = original score of 1 (“completely disagree”), 1 = original scores of 2, 4, or 5 (“somewhat disagree,” “somewhat agree,” “completely agree”). Method 2 (strict cut-point): Absenteeism was coded the same as above. CES-D items were dichotomized as follows: 0 = original scores of 1 or 2 (“rarely or none,” “sometimes”), 1 = original scores of 3 or 4 (“often/half the time,” “most of the time/continuously”), with only moderate-to-high frequency symptoms considered as present. The internalizing/externalizing problem behavior scale was dichotomized as follows: 0 = original scores of 1 or 2 (“completely disagree,” “somewhat disagree”), 1 = original scores of 4 or 5 (“somewhat agree,” “completely agree”), treating “completely disagree” and “somewhat disagree” as symptom absence and only “somewhat agree” and “completely agree” as symptom presence. This cut-point was used as a sensitivity analysis to examine whether the results from Method 1 were robust. Based on the above two dichotomization methods, Ising Network 1 (lenient cut-point) and Ising Network 2 (strict cut-point) were constructed separately. The results from both will be reported simultaneously. Subsequently, the Ising model was fitted using the IsingFit() function from the IsingFit package with gamma = 0.25 to control overfitting [21–22]. Because the primary purpose of simulated intervention in this context was to identify potentially most effective intervention nodes, network stability was assessed by examining the CS coefficients of node Expected Influence (EI); all CS coefficients exceeded 0.5, indicating adequate stability. Based on the estimated Ising models, two types of simulated interventions were implemented: alleviating intervention and exacerbating intervention. In the alleviating intervention, the activation threshold of a single symptom was lowered by subtracting 2 standard deviations from its original threshold parameter; in the exacerbating intervention, the activation threshold was raised by adding 2 standard deviations to its original threshold parameter [13]. Each intervention condition was run for 5,000 simulations. Intervention effects of different nodes were evaluated by systematically perturbing node thresholds and comparing the overall expected activation level of the network before and after intervention. Simulated intervention results based on cross-sectional data reflect statistical possibilities and are not equivalent to real-world causal effects. In this study, simulated intervention is positioned as an exploratory analysis, and all conclusions should be regarded as preliminary hypotheses to be verified by longitudinal studies or randomized controlled trials. Results Demographic variables Basic demographic information, including age, sex, grade, elite class, class ranking, school location, per capita household income, father’s education level, mother’s education level, father’s depression total score, and mother’s depression total score, is presented in Table 1. Parental depression total scores were assessed in the CFPS using the Center for Epidemiologic Studies Depression Scale (CES-D). Table 1Basic demographic informationVariable N Age (M ± SD)13.66 ± 2.538Sex (Male/Female)902/991Grade Elementary school789 Junior high school615 Senior high school490Elite class Yes230 No1664Class ranking Top 10%451 11%-25%480 26%-50%418 51%-75%172 Bottom 24%101 Not published272School location Provincial capital (including municipalities)170 General city (including county-level and prefecture-level cities)394 County town658 Rural area (including towns and villages)672 Annual per capita household income26282.58 ± 42435.658Educational level (father/mother) Illiterate/semi-literate204/297 Primary school388/395 Junior high school763/753 Senior high school / secondary vocational school / technical school / vocational high school310/258 Associate degree145/115 Bachelor’s degree77/73 Master’s degree7/3Father’s total depression score13.49 ± 3.947Mother’s total depression score14.02 ± 3.871 Table 1Basic demographic informationVariable N Age (M ± SD)13.66 ± 2.538Sex (Male/Female)902/991Grade Elementary school789 Junior high school615 Senior high school490Elite class Yes230 No1664Class ranking Top 10%451 11%-25%480 26%-50%418 51%-75%172 Bottom 24%101 Not published272School location Provincial capital (including municipalities)170 General city (including county-level and prefecture-level cities)394 County town658 Rural area (including towns and villages)672 Annual per capita household income26282.58 ± 42435.658Educational level (father/mother) Illiterate/semi-literate204/297 Primary school388/395 Junior high school763/753 Senior high school / secondary vocational school / technical school / vocational high school310/258 Associate degree145/115 Bachelor’s degree77/73 Master’s degree7/3Father’s total depression score13.49 ± 3.947Mother’s total depression score14.02 ± 3.871 Basic demographic information Absenteeism behavior In the past month (excluding summer and winter vacations), 291out of 1894 participants (15.36%) reported having taken leave or skipped class. The specific frequencies of leave-taking and class-skipping are presented in Table 2. Table 2Specific frequencies of absenteeismFrequency of Absenteeism n Percentage0160384.641517.972723.803351.854110.58580.42610.05730.16820.11910.051020.111220.011510.052020.11Total1894100.0 Table 2Specific frequencies of absenteeismFrequency of Absenteeism n Percentage0160384.641517.972723.803351.854110.58580.42610.05730.16820.11910.051020.111220.011510.052020.11Total1894100.0 Specific frequencies of absenteeism Factor analysis First, in the exploratory sample (N₁ = 1,367), the eight CES-D items and the 14 items of the internalizing/externalizing problem behavior scale were jointly included in an EFA. The KMO measure was 0.899, and Bartlett’s test of sphericity was significant (p < 0.001), indicating that the data were suitable for factor analysis. The factor correlation matrix from the first EFA showed low-to-moderate correlations among factors (maximum correlation = 0.324; see Figure S1), suggesting that the factors were not completely independent. Therefore, oblique rotation was adopted, and factors were extracted based on eigenvalues greater than 1, yielding five factors (see scree plot, Fig. 1), with a cumulative variance contribution rate of 52.26%. According to the pattern matrix of the first EFA (see Fig. 1), Factor 1 comprised Q13, Q2, Q12, and Q8, with loadings ranging from 0.398 to 0.789; Factor 2 comprised CESD7, CESD5, CESD8, CESD2, CESD1, CESD3, Q5, and Q7, with loadings ranging from 0.354 to 0.769; Factor 3 comprised CESD4 and CESD6, both with loadings above 0.850; Factor 4 comprised Q14, Q9, Q10, and Q11, with loadings ranging from 0.452 to 0.735; and Factor 5 comprised Q4, Q6, Q3, and Q1, with loadings ranging from 0.481 to 0.803. The pattern matrix of the first EFA showed that the loadings of Q5 (0.356), Q7 (0.364), Q8 (0.398), Q11 (0.452), and Q1 (0.481) were below the preferred threshold of 0.5. Further examination of item content revealed that Q5 and Q7 semantically overlapped heavily with the depressive items of the CES-D, whereas Q8, Q1, and Q11, although belonging to the internalizing/externalizing scale, had excessively low loadings. Therefore, we decided to remove Q1, Q5, Q7, Q8, and Q11 and re-conduct factor analysis on the remaining items. Additionally, to test the robustness of the factor structure, we performed varimax orthogonal rotation (see Figure S2) as a sensitivity analysis, which yielded the same factor structure distribution. Fig. 1Scree Plot (Left) and Pattern Matrix (Right) of the First EFA. Note: The figure presents the results of the first exploratory factor analysis including all 22 items from the CESD-8 and the internalizing/externalizing problem behavior scale. The left panel shows the scree plot, indicating five factors with eigenvalues greater than 1 (cumulative variance explained = 52.26%). The right panel displays the pattern matrix derived from principal axis factoring with oblique rotation (Oblimin with Kaiser Normalization). To be assigned to a given factor, an item was required to have an absolute loading ≥ 0.50. Factors 1 through 5 are ordered from top to bottom by decreasing eigenvalues. Items Q1, Q5, Q7, Q8, and Q11 exhibited loadings below 0.50 across all factors and were therefore removed in subsequent analyses Scree Plot (Left) and Pattern Matrix (Right) of the First EFA. Note: The figure presents the results of the first exploratory factor analysis including all 22 items from the CESD-8 and the internalizing/externalizing problem behavior scale. The left panel shows the scree plot, indicating five factors with eigenvalues greater than 1 (cumulative variance explained = 52.26%). The right panel displays the pattern matrix derived from principal axis factoring with oblique rotation (Oblimin with Kaiser Normalization). To be assigned to a given factor, an item was required to have an absolute loading ≥ 0.50. Factors 1 through 5 are ordered from top to bottom by decreasing eigenvalues. Items Q1, Q5, Q7, Q8, and Q11 exhibited loadings below 0.50 across all factors and were therefore removed in subsequent analyses Second, after removing Q1, Q5, Q7, Q8, and Q11, EFA was conducted on the remaining items (KMO = 0.854, Bartlett’s test of sphericity significant at p < 0.001; see Figure S3 for the factor correlation matrix of the second EFA). Oblique rotation was again used to extract factors with eigenvalues greater than 1, resulting in five factors (see scree plot, Fig. 2) with a cumulative variance contribution rate of 57.50%. The pattern matrix of the second EFA (see Fig. 2) showed a factor structure consistent with the adjusted first EFA. Varimax orthogonal rotation was again performed as a sensitivity analysis (see Figure S4) and produced the same factor structure distribution. Further inspection revealed that Q3 (loading = 0.481) fell below the acceptable threshold of 0.5 and was therefore removed. Thus, combining the results of the first EFA, we decided to remove Q1, Q3, Q5, Q7, Q8, and Q11 and re-conduct factor analysis on the remaining items. Because the maximum correlation among factors was below 0.3, the third EFA adopted varimax orthogonal rotation (KMO = 0.844, Bartlett’s test of sphericity significant at p < 0.001; see Figure S5 for the factor correlation matrix of the third EFA), with a cumulative variance contribution rate of 59.26%. The results of the third EFA (see Fig. 3) indicated that all item loadings exceeded 0.5, and the factor structure distribution remained consistent with the previous ones. Specifically, Factor 1 comprised Q13, Q2, and Q12, with loadings from 0.575 to 0.789; Factor 2 comprised CESD7, CESD5, CESD8, CESD2, CESD1, and CESD3, with loadings from 0.625 to 0.758; Factor 3 comprised CESD4 and CESD6, with loadings from 0.841 to 0.856; Factor 4 comprised Q14, Q9, and Q10, with loadings from 0.600 to 0.746; and Factor 5 comprised Q4 and Q6, with loadings from 0.782 to 0.851. Fig. 2Scree Plot (Left) and Pattern Matrix (Right) of the Second EFA. Note: The figure presents the results of the second exploratory factor analysis conducted on the remaining 17 items after removing Q1, Q5, Q7, Q8, and Q11. The left panel shows the scree plot, indicating five factors with eigenvalues greater than 1 (cumulative variance explained = 57.50%). The right panel displays the pattern matrix derived from principal axis factoring with oblique rotation (Oblimin with Kaiser Normalization). To be assigned to a given factor, an item was required to have an absolute loading ≥ 0.50. Factors 1 through 5 are ordered from top to bottom by decreasing eigenvalues. The assignment of items to factors is consistent with the adjusted structure from the first EFA. Items Q3 exhibited loadings below 0.50 across all factors and were therefore removed in subsequent analyses Scree Plot (Left) and Pattern Matrix (Right) of the Second EFA. Note: The figure presents the results of the second exploratory factor analysis conducted on the remaining 17 items after removing Q1, Q5, Q7, Q8, and Q11. The left panel shows the scree plot, indicating five factors with eigenvalues greater than 1 (cumulative variance explained = 57.50%). The right panel displays the pattern matrix derived from principal axis factoring with oblique rotation (Oblimin with Kaiser Normalization). To be assigned to a given factor, an item was required to have an absolute loading ≥ 0.50. Factors 1 through 5 are ordered from top to bottom by decreasing eigenvalues. The assignment of items to factors is consistent with the adjusted structure from the first EFA. Items Q3 exhibited loadings below 0.50 across all factors and were therefore removed in subsequent analyses Fig. 3Scree Plot (Left) and Rotated Factor Loading Matrix (Right) of the Third EFA. Note: The figure presents the results of the third exploratory factor analysis conducted on the remaining 16 items after removing Q1, Q3, Q5, Q7, Q8, and Q11. The left panel shows the scree plot, indicating five factors with eigenvalues greater than 1 (cumulative variance explained = 59.25%). The right panel displays the pattern matrix derived from principal axis factoring with Varimax rotation. To be assigned to a given factor, an item was required to have an absolute loading ≥ 0.50. Factors 1 through 5 are ordered from top to bottom by decreasing eigenvalues. The assignment of items to factors is consistent with the adjusted structure from the first EFA Scree Plot (Left) and Rotated Factor Loading Matrix (Right) of the Third EFA. Note: The figure presents the results of the third exploratory factor analysis conducted on the remaining 16 items after removing Q1, Q3, Q5, Q7, Q8, and Q11. The left panel shows the scree plot, indicating five factors with eigenvalues greater than 1 (cumulative variance explained = 59.25%). The right panel displays the pattern matrix derived from principal axis factoring with Varimax rotation. To be assigned to a given factor, an item was required to have an absolute loading ≥ 0.50. Factors 1 through 5 are ordered from top to bottom by decreasing eigenvalues. The assignment of items to factors is consistent with the adjusted structure from the first EFA To further validate the factor structure of the scale after item removal, confirmatory factor analysis (CFA) of a first-order five-factor correlated model was performed on the validation dataset (N₂ = 1,367) using the lavaan package in R. The results showed that the five-factor correlated model fitted well: χ²(80) = 198.484, p < 0.001, χ²/df = 2.481, CFI = 0.973, TLI = 0.973, RMSEA = 0.033 (90% CI [0.027, 0.039]), SRMR = 0.027. With χ²/df < 3, CFI and TLI > 0.90, and RMSEA and SRMR < 0.08, the model fit can be considered acceptable [23]. Convergent validity analysis showed the standardized loadings of each item and the CR and AVE values of each factor in Table S2. The composite reliability (CR) of the factors ranged from 0.55 to 0.81, and the average variance extracted (AVE) ranged from 0.34 to 0.51. It must be acknowledged that in this study, the AVE of some factors was below the conventional threshold of 0.5, and the CR of some factors was below the conventional threshold of 0.7, which are clear limitations of the measurement model. In particular, the standardized loading of Q14 in Factor 4 (AVE = 0.34) was only 0.350. We attempted to remove Q14 (i.e., removing Q1, Q3, Q5, Q7, Q8, Q11, and Q14). The model fit after removing seven items was as follows: χ²(67) = 157.325, p < 0.001, χ²/df = 2.348, CFI = 0.979, TLI = 0.979, RMSEA = 0.031 (90% CI [0.025, 0.038]), SRMR = 0.024. After removing Q14, the CR of Factor 4 increased from 0.60 to 0.62, and its AVE increased from 0.34 to 0.45; the AVE of all factors was ≥ 0.4, and only Factor 1 (CR = 0.54) had a CR slightly below 0.6 (see Table S2 for specific values). Possible factors contributing to these suboptimal indices include the small number of items per dimension and limited symptom variability in the community adolescent sample. However, it must be emphasized that these explanations cannot fully eliminate the measurement deficiencies. Although the CR and AVE of some factors did not reach the gold standard, considering the exploratory nature of this study and the good overall model fit (CFI = 0.979, RMSEA = 0.031), some scholars suggest that such a model may be marginally acceptable [24–25]. Therefore, we removed Q14 to retain a more parsimonious measurement model and will discuss these limitations when interpreting subsequent network analysis results. On this basis, the discriminant validity of the scale was further examined. The Fornell-Larcker test (see Table S3) showed that all factors met the Fornell-Larcker criterion, indicating good discriminant validity. Therefore, retaining the original factor structure was reasonable for the purposes of this study. In addition, Cronbach’s alpha coefficients for each factor were as follows: Factor 1 α = 0.61, Factor 2 α = 0.80, Factor 3 α = 0.65, Factor 4 α = 0.62, Factor 5 α = 0.66. The Cronbach’s alpha values for Factors 1, 3, and 4 ranged from 0.61 to 0.66, slightly below the ideal threshold of 0.70 but reaching the acceptable level of 0.60 for exploratory research, which may be attributable to the small number of items or sample characteristics. In summary, the five-factor structure after removing seven items demonstrated acceptable reliability and validity and can serve as the measurement basis for subsequent network analysis. Because some indices were not ideal, these measurement-level shortcomings should be taken into account when interpreting the subsequent network analysis results. Considering that deficiencies in the measurement model might lead to spurious edges in network analysis, we additionally conducted a sensitivity analysis: a network was re-estimated using all original items (22 items) as nodes and compared with the primary analysis (the network after removing seven items). All edge weights for edges common to both networks were extracted, and Spearman’s rank correlation coefficient was calculated. If the correlation coefficient was > 0.80, it was considered that the primary analysis results were not substantially distorted by item removal. The specific results of this sensitivity analysis will be presented in the network analysis section. Network analysis Network estimation After controlling for the covariates of age, sex, grade, elite class, class ranking, school location, annual per capita household income, father’s educational level, mother’s educational level, father’s total depression score, and mother’s total depression score, a network structure of absenteeism behavior, depressive symptoms, and internalizing/externalizing problems was constructed with items from absenteeism behavior, the Center for Epidemiologic Studies Depression Scale (CESD-8), and the internalizing/externalizing problem behavior scale (after removing Q1, Q3, Q5, Q7, Q8, Q11, and Q14) as nodes. The results are shown in Fig. 4. To prevent unsatisfactory discriminant validity of the scales from leading to spurious edges, we additionally constructed a network based on all original items (absenteeism, all 8 items of the CES-D-8, and all 14 items of the internalizing/externalizing problem behavior scale) as a sensitivity analysis. Fig. 4Network Structure of Absenteeism Behavior, Depressive Symptoms, and Internalizing/Externalizing Problems. Note: The horizontal axis represents the values, and the vertical axis lists the nodes. Nodes positioned higher on the vertical axis have higher EI values (left) or bEI values (right) Network Structure of Absenteeism Behavior, Depressive Symptoms, and Internalizing/Externalizing Problems. Note: The horizontal axis represents the values, and the vertical axis lists the nodes. Nodes positioned higher on the vertical axis have higher EI values (left) or bEI values (right) The primary network comprised 16 nodes, with 120 theoretically possible edges, of which 87 were actually present, yielding a network density of 0.725. Among these, 33 edges were statistically significant (confidence intervals did not include zero). The three strongest edges in the network structure were CESD4–CESD6 (0.431), Q4–Q6 (0.346), and CESD5–CESD7 (0.339). The Spearman’s rank correlation coefficient of edge weights for the common edges between the primary analysis network and the full-item network was 0.894 (p < 0.001), and the Pearson correlation coefficient was 0.980 (p < 0 0.001), both exceeding 0.8. The top 10 common edges with the largest edge weights (including edge weights, confidence intervals, and widths) in the two networks are presented in Table S4. These results indicate that removing low-loading items had minimal impact on the overall network structure, and the primary analysis results demonstrate good robustness. Centrality analysis indicated (see Fig. 5) that in the network of absenteeism behavior, depression, and internalizing/externalizing problems, the three nodes with the highest EI values were CESD5 (EI = 1.185), CESD7 (EI = 0.968), and CESD2 (EI = 0.968), and the three nodes with the highest bEI values were CESD3 (bEI = 0.209), CESD2 (bEI = 0.207), and N (bEI = 0.206). The EI values, bEI values, and their corresponding confidence intervals for this network structure are presented in Table S5. The centrality indices of the full-item network are presented in Figure S6. Fig. 5Centrality Analysis of the Network of Absenteeism Behavior, Depressive Symptoms, and Internalizing/Externalizing Problems. Note: The X-axis represents the percentage of cases using the original sample. The Y-axis displays the average correlation between the centrality indices of the original network and those of the re-estimated network Centrality Analysis of the Network of Absenteeism Behavior, Depressive Symptoms, and Internalizing/Externalizing Problems. Note: The X-axis represents the percentage of cases using the original sample. The Y-axis displays the average correlation between the centrality indices of the original network and those of the re-estimated network Stability analysis The correlation stability (CS) coefficients for Expected Influence (EI) and bridge Expected Influence (bEI) were 0.75 and 0.595 (see Fig. 6), exceeding the threshold of 0.5, indicating that the node centrality indices exhibited high stability and that the results were reliable. The results of node difference tests for EI and bEI are presented in Figures S7 and S8, respectively. Nonparametric bootstrapping (1,000 iterations) was used to estimate 95% confidence intervals for each edge to evaluate edge accuracy (see Figure S9). The results showed that the confidence intervals for most edges were relatively narrow, suggesting high accuracy of edge weights. Among them, the three strongest edges in the network (CESD4–CESD6, Q4–Q6, and CESD5–CESD7) all met the criterion of CI_Width ≤ |Weight| × 0.5, indicating acceptable edge accuracy (see Table S4 for specific values). The results of edge weight difference tests are shown in Figure S10. Taken together, the network of absenteeism behavior, depressive symptoms, and internalizing/externalizing problems demonstrated good overall stability and high estimation accuracy. The CS coefficients for both EI and bEI in the full-item network were 0.75 (see Figure S11), exceeding the acceptable threshold of 0.5, indicating that the full-item network is stable and reliable, and the results of the sensitivity analysis are trustworthy. Fig. 6Stability Analysis of the Network of Absenteeism Behavior, Depressive Symptoms, and Internalizing/Externalizing Problems. Note: Panels (a) and (b) show results for Ising network 1 (lenient cutoffs): (a) exacerbating intervention, (b) alleviating intervention. Panels (c) and (d) show results for Ising network 2 (strict cutoffs): (c) exacerbating intervention, (d) alleviating intervention. The horizontal axis lists the intervened nodes, and the vertical axis represents the expected total network score (Sum score). In the exacerbating intervention, CESD5 consistently showed the largest increase in both networks. In the alleviating intervention, Q6 showed the largest decrease in Ising network 1, while CESD1 showed the largest decrease in Ising network 2 Stability Analysis of the Network of Absenteeism Behavior, Depressive Symptoms, and Internalizing/Externalizing Problems. Note: Panels (a) and (b) show results for Ising network 1 (lenient cutoffs): (a) exacerbating intervention, (b) alleviating intervention. Panels (c) and (d) show results for Ising network 2 (strict cutoffs): (c) exacerbating intervention, (d) alleviating intervention. The horizontal axis lists the intervened nodes, and the vertical axis represents the expected total network score (Sum score). In the exacerbating intervention, CESD5 consistently showed the largest increase in both networks. In the alleviating intervention, Q6 showed the largest decrease in Ising network 1, while CESD1 showed the largest decrease in Ising network 2 Simulated intervention All items were dichotomized using two methods. Method 1 (lenient cut-point): Absenteeism was coded as 0 (absent) / 1 (present). Each item of the Center for Epidemiologic Studies Depression Scale (CES-D) was dichotomized based on the original score: 0 = original score of 1 (“rarely or none”), 1 = original score of 2, 3, or 4 (“sometimes,” “often/half the time,” “most of the time/continuously”). The self-developed internalizing/externalizing problem behavior scale (after removing six items; originally a 5-point rating, but the “moderate” option was not read aloud in the CFPS data) was dichotomized as follows: 0 = original score of 1 (“completely disagree”), 1 = original scores of 2, 4, and 5 (“somewhat disagree,” “somewhat agree,” “completely agree”).Method 2 (strict cut-point): Absenteeism was coded the same as above. CES-D items were dichotomized as follows: 0 = original scores of 1 and 2 (“rarely or none,” “sometimes”), 1 = original scores of 3 and 4 (“often/half the time,” “most of the time/continuously”), with only moderate-to-high frequency symptoms considered as present. The internalizing/externalizing problem behavior scale was dichotomized as follows: 0 = original scores of 1 and 2 (“completely disagree,” “somewhat disagree”), 1 = original scores of 4 and 5 (“somewhat agree,” “completely agree”), treating “completely disagree” and “somewhat disagree” as symptom absence and only “somewhat agree” and “completely agree” as symptom presence. Using the IsingFit function in R, Ising Network 1 (Method 1) and Ising Network 2 (Method 2) were constructed for absenteeism, depression, and internalizing/externalizing problem behaviors (network structure diagrams for Ising Network 1 and Ising Network 2 are shown in Figure S12). The results indicated that both network models demonstrated good stability and reliability: the CS coefficient for EI values was 0.75 for the Method 1 network and 0.672 for the Method 2 network, both exceeding the recommended threshold of 0.5 (centrality results and stability tests are presented in Figure S13 and Figure S14, respectively). On this basis, the NodeIdentifyR algorithm was applied to perform simulated interventions on the network: by increasing or decreasing the threshold of a single node by 2 standard deviations, exacerbating and alleviating intervention conditions were simulated, and the overall expected activation level of the network obtained from model simulation (hereinafter referred to as the expected total score of the network) was used as the outcome indicator. The simulated intervention results for Ising Network 1 and Ising Network 2 are presented in Fig. 7; Table 3. As shown in panels (a) and (c) of Fig. 7, in the worsening interventions, the node with the largest increase in expected total score was CESD5 (“I felt lonely”) in both Ising Network 1 (lenient cut-point) and Ising Network 2 (strict cut-point), with the expected total score rising from 7.262 to 8.477 in Network 1 and from 4.743 to 5.967 in Network 2. As shown in panels (b) and (d) of Fig. 7, in the easing interventions, the node with the largest decrease in expected total score was Q6 (“I got distracted easily”) in Ising Network 1, with the score dropping from 7.218 to 5.848, and CESD1 (“I felt depressed”) in Ising Network 2, with the score dropping from 4.714 to 3.885. Fig. 7Predicted effects of exacerbating and alleviating interventions on Ising network 1 (lenient cutoffs) and Ising network 2 (strict cutoffs) for absenteeism behavior, depressive symptoms, and internalizing/externalizing problems Predicted effects of exacerbating and alleviating interventions on Ising network 1 (lenient cutoffs) and Ising network 2 (strict cutoffs) for absenteeism behavior, depressive symptoms, and internalizing/externalizing problems Table 3Predicted intervention effects for two ising networks (based on different dichotomization cutoffs)NodeIsing Network 1Ising Network 2EIAggravatingAlleviatingEIAggravatingAlleviatingSumscoreciLowerciUpperSumscoreciLowerciUpperSumscoreciLowerciUpperSumscoreciLowerciUpperoriginal/7.2627.1847.347.2187.147.296/4.7434.6744.8114.7144.6454.783N0.0007.4727.3957.5487.0787.0017.1560.0005.0254.9565.0944.5784.514.646CESD14.1808.3488.2728.4246.3776.3066.4483.5385.7125.6435.783.8853.8243.946CESD24.6958.3958.3178.4726.5216.4516.5914.3235.7665.6985.8353.913.853.97CESD32.5038.2788.1978.3586.7016.636.7732.3695.6765.6065.7464.1584.0944.221CESD41.8217.6997.6237.7756.5816.5046.6581.8215.1475.085.2134.0563.994.122CESD54.9648.4778.3948.566.7066.6356.7775.3445.9675.8966.0394.1344.0754.194CESD62.0127.6767.5997.7526.6156.5396.6922.0195.1695.1015.2364.0714.0054.137CESD74.9518.4438.3668.526.4746.4046.5444.7235.8765.8065.9463.9273.8683.987CESD83.5058.0337.9488.1187.0056.9317.0782.8245.5155.445.594.5214.4554.587Q22.9458.098.0188.1626.3236.2486.3981.9745.1565.0855.2274.5234.4564.589Q44.1557.727.6477.7925.8755.7945.9552.6205.4455.3765.5144.214.1474.274Q64.4917.717.6387.7825.8485.7675.9293.7895.4825.4125.5534.1494.0874.212Q93.6377.7757.7027.8475.9245.8436.0042.3765.4375.3685.5074.1514.0884.214Q104.6398.0047.9348.0735.9235.8475.9993.2085.6325.565.7034.2744.214.338Q124.0948.0147.9438.0846.0896.0146.1643.1985.3885.3145.4624.4764.4114.542Q133.3828.2738.2018.3466.5096.4346.5833.8825.1975.1235.274.5974.5294.665Ising Network 1 was based on lenient dichotomization cutoffs, and Ising Network 2 on strict cutoffs. The network nodes comprised absenteeism behavior, CES-D items, and internalizing/externalizing problem behavior item Table 3Predicted intervention effects for two ising networks (based on different dichotomization cutoffs)NodeIsing Network 1Ising Network 2EIAggravatingAlleviatingEIAggravatingAlleviatingSumscoreciLowerciUpperSumscoreciLowerciUpperSumscoreciLowerciUpperSumscoreciLowerciUpperoriginal/7.2627.1847.347.2187.147.296/4.7434.6744.8114.7144.6454.783N0.0007.4727.3957.5487.0787.0017.1560.0005.0254.9565.0944.5784.514.646CESD14.1808.3488.2728.4246.3776.3066.4483.5385.7125.6435.783.8853.8243.946CESD24.6958.3958.3178.4726.5216.4516.5914.3235.7665.6985.8353.913.853.97CESD32.5038.2788.1978.3586.7016.636.7732.3695.6765.6065.7464.1584.0944.221CESD41.8217.6997.6237.7756.5816.5046.6581.8215.1475.085.2134.0563.994.122CESD54.9648.4778.3948.566.7066.6356.7775.3445.9675.8966.0394.1344.0754.194CESD62.0127.6767.5997.7526.6156.5396.6922.0195.1695.1015.2364.0714.0054.137CESD74.9518.4438.3668.526.4746.4046.5444.7235.8765.8065.9463.9273.8683.987CESD83.5058.0337.9488.1187.0056.9317.0782.8245.5155.445.594.5214.4554.587Q22.9458.098.0188.1626.3236.2486.3981.9745.1565.0855.2274.5234.4564.589Q44.1557.727.6477.7925.8755.7945.9552.6205.4455.3765.5144.214.1474.274Q64.4917.717.6387.7825.8485.7675.9293.7895.4825.4125.5534.1494.0874.212Q93.6377.7757.7027.8475.9245.8436.0042.3765.4375.3685.5074.1514.0884.214Q104.6398.0047.9348.0735.9235.8475.9993.2085.6325.565.7034.2744.214.338Q124.0948.0147.9438.0846.0896.0146.1643.1985.3885.3145.4624.4764.4114.542Q133.3828.2738.2018.3466.5096.4346.5833.8825.1975.1235.274.5974.5294.665Ising Network 1 was based on lenient dichotomization cutoffs, and Ising Network 2 on strict cutoffs. The network nodes comprised absenteeism behavior, CES-D items, and internalizing/externalizing problem behavior item Predicted intervention effects for two ising networks (based on different dichotomization cutoffs) Ising Network 1 was based on lenient dichotomization cutoffs, and Ising Network 2 on strict cutoffs. The network nodes comprised absenteeism behavior, CES-D items, and internalizing/externalizing problem behavior item Discussion The measurement model in this study had certain limitations. Confirmatory factor analysis after removing seven items showed that the average variance extracted (AVE) values for Factors 1, 2, 3, and 4 were 0.40, 0.42, 0.48, and 0.45, respectively, all below the conventional threshold of 0.50; the AVE of Factor 5 was 0.51, reaching an acceptable level. The composite reliability (CR) of Factor 1 was 0.54, slightly below the acceptable lower limit of 0.60, but its Cronbach’s alpha was 0.61, which was at a marginally acceptable level. The CR values of the remaining factors ranged from 0.62 to 0.81. These results indicate that the convergent validity and reliability of the scales used in this study did not fully meet the psychometric gold standards in the community adolescent sample, which represents one of the methodological limitations of this study. It should be noted that after multiple rounds of item screening, the discriminant validity (Fornell-Larcker test) of the model was fully satisfied, and the overall model fit was good. Considering the exploratory nature of this study, we still conducted network analysis based on the five-factor structure after removing low-loading items. To evaluate the impact of measurement limitations on the network results, we constructed a full-item network including all original items (22 items) as a sensitivity analysis. The results showed that the edge weights common to the full-item network and the primary analysis network were highly correlated (Spearman ρ = 0.89), and the CS coefficients for both expected influence (EI) and bridge expected influence (bEI) in the full-item network were 0.75, indicating good stability of the centrality indices. This evidence of robustness suggests that although the measurement model had deficiencies in convergent validity and reliability, the main structure of the network was not substantially distorted by item removal or measurement imprecision. Before interpreting the network analysis results of this study, it is necessary to first clarify the operational definition and inherent limitations of the “absenteeism” indicator. The absenteeism item in this study came from the CFPS questionnaire, which asked about “the total number of excused or unexcused absences in the past month (excluding winter and summer vacations).” This item combines excused and unexcused absences into a single total count and cannot differentiate the specific reasons for absenteeism (e.g., illness, family obligations, school avoidance, truancy). Due to the design of the CFPS scale, this study could not classify absenteeism types; the absenteeism node in the current network actually represents an “averaged” absenteeism indicator, which may obscure specific associations between particular types of absenteeism and particular psychological problems. This is an important limitation of this study. In addition, more than 84% of participants in this study reported no absences, resulting in a severely zero-inflated skewed distribution of this variable, and the absenteeism data relied on self-report, which may be subject to recall bias. Therefore, what the current analysis captures is not chronic, pathological absenteeism in an educational or clinical sense, but rather minor, irregular short-term attendance fluctuations in a community context. Accordingly, the finding that the absenteeism node was network-associated with depression and internalizing/externalizing symptoms should be understood as a covariation relationship between short-term attendance fluctuations and the current psychological distress of the general adolescent population. The bridge centrality indices revealed by the network analysis reflect the conditional connection tendencies of the variables within this specific network and should not be directly extrapolated to imply “a stable, real structural bridge between absenteeism and psychopathology.” At the same time, due to the inherent limitations of the cross-sectional design, the causal direction of this covariation relationship cannot be determined. In this study, the network analysis of absenteeism, depressive symptoms, and internalizing/externalizing problem behaviors showed that the central nodes (i.e., nodes with the highest expected influence) were “I felt lonely” (CESD5), “I felt that everything I did was an effort” (CESD2), and “I felt sad” (CESD7). The bridge nodes (i.e., key nodes connecting different communities) were “I felt that everything I did was an effort” (CESD2), “My sleep was restless” (CESD3), and the absenteeism node (N). It should be noted that the absenteeism node exhibited a severely zero-inflated skewed distribution (84% of participants reported no absences); its statistical significance as a bridge node may mainly derive from the data patterns of the small number of individuals with absences and may not be stably generalizable as a “core hub connecting absenteeism and psychological distress.” In the simulated intervention analysis, to examine the impact of dichotomization cut-point selection on the results, we constructed Ising models using two different dichotomization methods (lenient cut-point and strict cut-point). In the worsening interventions, the results were consistent across both cut-points: “I felt lonely” (CESD5) showed the largest increase in expected total score. This finding should only be regarded as an exploratory result, and its stability under different data conditions awaits further examination in future research. In the easing interventions, the results differed by cut-point: under the lenient cut-point, “I got distracted easily” (Q6) showed the largest decrease in expected total score; under the strict cut-point, “I felt depressed” (CESD1) showed the largest decrease. This discrepancy indicates that the identification of easing intervention targets is sensitive to the threshold for symptom definition, and the results under different cut-points should not be regarded as definitive conclusions. We will mention both Q6 and CESD1 in the discussion but treat both as exploratory findings and do not claim a single optimal target. In the exploratory simulated intervention of this study, the node “I felt lonely” (CESD5) showed the largest increase in the network’s overall expected total score under worsening interventions. It must be clearly stated that this result comes from a simulation model based on cross-sectional data and reflects the statistical influence of “feeling lonely” in the conditional dependence network, rather than proving that real-world interventions targeting loneliness would necessarily effectively reduce absenteeism or broader symptom activation. Specifically, in the simulation model, raising the activation threshold of the “I felt lonely” node resulted in the largest increase in the network’s overall expected total score. This means that, under the conditional dependence structure of this network, loneliness had strong covariation relationships with absenteeism, depressive symptoms, and internalizing/externalizing problem behaviors; the state change of this node (at the simulation level) was most strongly associated with the expected increase in overall network activation. However, this result should be understood as follows: in the statistical model, loneliness is a strong predictor of the network’s total activation level and cannot be directly equated with “real interventions that worsen loneliness will inevitably lead to increased absenteeism or symptom deterioration.” This finding is consistent with previous network analysis research—in network analyses of emotional and behavioral problems, loneliness, and suicidal ideation among out-of-school adolescents, loneliness was also identified as a central node [26]. Loneliness is a subjective experience of pain or sadness in children and adolescents, meaning that they perceive a lack of belonging and connection with peers [27]. In the community sample of this study, adolescents’ reported absenteeism mainly manifested as short-term, irregular attendance fluctuations. The data showed a statistical covariation relationship between loneliness and such short-term attendance fluctuations, consistent with previous evidence on the association of loneliness with school aversion and truancy frequency [28–30]. It should be emphasized that the cross-sectional design of this study cannot distinguish causal directions; loneliness and attendance fluctuations may influence each other or may be jointly driven by other factors (e.g., family support, school climate). Baumeister and Leary’s belongingness theory posits that establishing and maintaining stable, positive interpersonal relationships is a fundamental and universal human motivation [31], and students’ absenteeism may be a response to a lack of belonging. The attendance fluctuations of highly lonely adolescents may partly stem from a lack of school belonging and exclusionary signals in the classroom environment, rather than merely personal internal vulnerabilities [32–33]. Moreover, highly lonely adolescents experience more physical and psychological discomfort, as well as health risk behaviors such as binge eating and non-suicidal self-injury [34], and internalizing symptoms such as depression and anxiety mediate the relationship between loneliness and absenteeism [35]. In the exploratory simulated intervention of this study, the optimal node under easing interventions differed depending on the dichotomization cut-point: “I got distracted easily” (Q6) under the lenient cut-point and “I felt depressed” (CESD1) under the strict cut-point. This discrepancy suggests that the identification of easing intervention targets is sensitive to the threshold for symptom definition and may reflect the impact of classifying mild symptoms as “absent” or “present” under different cut-points. Therefore, both nodes can be regarded as exploratory candidate directions with potential value, but it cannot yet be determined which one may be more advantageous in actual interventions. It should be stressed that the simulated intervention in this study is based on cross-sectional data and reflects statistical conditional associations rather than real-world causal effects. In the simulation model, if the activation threshold of either of these two nodes is lowered (i.e., making it more easily “eased”), the network’s overall expected total score shows the largest decrease. This means that under the conditional dependence structure of this network, being easily distracted and depressed mood both have strong covariation relationships with absenteeism, depressive symptoms, and internalizing/externalizing problem behaviors; the state change of either node (at the simulation level) may be associated with the expected improvement in overall network activation. However, this result should be understood as follows: in the statistical model, these nodes are strong predictors of the network’s total activation level and cannot be directly equated with “real interventions targeting these experiences will necessarily improve attendance or symptoms.” Existing research has shown that being easily distracted is associated with absenteeism in children and adolescents: school-aged children/adolescents with attention/hyperactivity problems are more likely to exhibit absenteeism and school avoidance behaviors [36]; the unexcused absence rate of children who may have ADHD in primary school is 2.2 times that of the control group [37]; and adolescents with ADHD have significantly higher numbers of absences in the past 12 weeks than their peers [38]. It is evident that cognitive problems and inattention are important correlates of adolescent absenteeism [39]. Depression is also associated with absenteeism in children and adolescents. Depression has been identified as a central node in symptom networks of school absenteeism (including nodes such as anxiety and aggression) [12]. Compared with other internalizing/externalizing problem behaviors, depression is uniquely characterized by anhedonia and motivation deficits. A systematic review by Finning et al. indicated that adolescent depression may lead to reduced school attendance through mechanisms such as loss of motivation, social withdrawal, and low energy [40]. Moreover, based on large-sample longitudinal data, it was further found that anhedonia and social dysfunction were the strongest mental health components predicting educational interruption in adolescents [41]. These external lines of evidence echo the simulated findings of this study, jointly suggesting that being easily distracted and depressed mood may play important roles in the absenteeism-related network, but cannot replace direct causal testing. Limitations and future directions This study has several methodological limitations that should be fully considered when interpreting the results. First, limitations of the measurement model. Confirmatory factor analysis showed that the average variance extracted (AVE) of some factors was below 0.50, and the composite reliability (CR) of Factor 1 was 0.54, slightly below the acceptable level. Although discriminant validity was satisfied and the overall model fit was good, the deficiencies in convergent validity and reliability may still have some influence on the estimation of network edge weights. Future research should use scales with better psychometric properties or further test the stability of this five-factor structure in larger samples. Second, limitations of the absenteeism indicator. The absenteeism item in this study combines excused and unexcused absences into a single total count and cannot differentiate among various reasons for absence, such as illness, family obligations, school avoidance, and truancy. In addition, the absenteeism data relied on self-report and may be subject to recall bias. Due to these limitations, the current absenteeism node reflects an “averaged” attendance behavior pattern and may obscure specific associations between particular types of absenteeism and psychological problems. Future research should use measurement tools that can differentiate among types of absenteeism (e.g., separately asking about the frequency of illness-related absences, excused absences, and truancy) to examine the psychological mechanisms of different reasons for absence in a more refined manner. Third, limitations of sample characteristics. In this study, more than 84% of participants reported no absences, and the absenteeism variable exhibited a severely zero-inflated skewed distribution. Therefore, what the current analysis captures is mainly short-term, irregular attendance fluctuations among community adolescents, rather than chronic pathological absenteeism in a clinical sense. Generalizing the study conclusions to chronic absenteeism or clinical samples requires caution. Future research could conduct replication studies targeting subgroups with higher absenteeism frequency or clinical samples. Fourth, limitations of the study design. This study used cross-sectional data; the network edge weights reflect conditional dependence relationships and cannot infer causal directions. The simulated interventions are based on cross-sectional models, and their results represent only statistical possibilities, not equivalent to real-world causal effects. In addition, the identification of targets in easing interventions was sensitive to the dichotomization cut-point, with results differing across cut-points, indicating that the robustness of the conclusions requires further examination. Future research should adopt longitudinal designs or randomized controlled trials to verify the actual effects of the candidate intervention targets identified in this study. Fifth, sample representativeness and external validity. The sample of this study came from the CFPS data; although it is a national survey, it is limited to Chinese community adolescents. Caution is needed when generalizing the findings to adolescents in other cultures or under different educational systems, taking contextual differences into account. Future research could conduct cross-cultural comparisons across different countries or educational settings. In summary, the findings of this study should be regarded as exploratory hypotheses rather than definitive conclusions. Future research should continue to advance in terms of improving measurement tools, differentiating types of absenteeism, and adopting longitudinal designs to more reliably reveal the complex relationships between absenteeism behavior and adolescent psychological distress. Supplementary Information Supplementary Material 1. Supplementary Material 1.