Contextual stressors and multidimensional mental health among university students during the COVID-19 period: a construct-level network analysis Authors: Zhenti Cui, Shujuan Wu, Haoyang Ma, Yi Zhang, Xiaoxiao Lin, Xiaojie Hou, Yuan Ren, Yaru Feng, Yingdong Cao Published: 2026 ABSTRACT BackgroundMental health problems among university students predated COVID-19, but pandemic-period conditions coincided with changes in how academic, relational, behavioral, and internalizing difficulties co-occurred. Prior studies often modeled single outcomes or symptom clusters separately. This secondary analysis used a construct-level network to describe conditional associations among contextual stressors, adaptive-functioning indicators, and multidimensional mental health indicators in one archived student dataset.MethodsWe analyzed cross-sectional survey data from 10,745 students collected January 13-March 26, 2021. Fourteen nodes were included: media time, online adaptation, learning efficiency, family conflict, friendship change, epidemic attitude, epidemic attention, exercise, perceived stress, anxiety, depression, sleep problems, loneliness, and resilience. After correcting exercise-frequency coding and anxiety-score derivation errors, we estimated an EBICglasso-regularized partial-correlation network using Spearman correlations. Expected influence, bridge expected influence, and nodewise R2 were interpreted as descriptive, model-dependent indices. We examined edge accuracy, centrality stability, bridge stability, edge-weight differences, data-quality exclusions, alternative estimators, community definitions, and counterintuitive partial edges.ResultsThe fitted network retained 53 of 91 possible edges (density = 0.582). In this specification, depression, perceived stress, and anxiety had the highest expected-influence values (1.428, 1.064, and 0.782). The strongest absolute partial-correlation edges linked sleep problems with depression (0.736), depression with anxiety (0.668), anxiety with perceived stress (0.459), and sleep problems with anxiety (-0.423). Under the prespecified two-community partition, bridge expected influence was largest for perceived stress (0.461), loneliness (0.343), family conflict (0.233), epidemic attitude (0.184), and learning efficiency (0.126). Expected influence and bridge expected influence showed high case-dropping stability within this analytic specification (CS = 0.75). Data-quality exclusions and an ordinal/mixed-correlation EBICglasso model yielded estimates very similar to the primary network, whereas an unsigned-weight mixed graphical model yielded different centrality rankings.ConclusionsThe findings describe a construct-level snapshot of conditional associations among students from one institutional context during a specific COVID-19 period. Perceived stress showed the highest and most stable bridge expected-influence value across the tested community specifications, while other bridge rankings and centrality values were specification-dependent. The results are hypothesis-generating, not evidence of causal pathways, intervention targets, or population-representative student mental health structure.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40359-026-04721-0. Introduction Student mental health before and during the COVID-19 period University student mental health should be understood against a backdrop of substantial pre-pandemic burden. Multi-country World Mental Health surveys showed that a considerable proportion of college students met criteria for at least one common mental disorder, with many conditions emerging before or during the college years and often remaining untreated [1, 2]. This pre-pandemic literature matters because it indicates that the COVID-19 period did not create student distress from nothing. Rather, pandemic-era restrictions and uncertainty were layered onto a developmental and institutional context in which academic transition, identity formation, altered social networks, and restricted help-seeking were already salient challenges. COVID-19 nonetheless coincided with changes in both the intensity and the configuration of student mental health difficulties. Systematic reviews and meta-analyses conducted during and after the acute pandemic years consistently reported high prevalence of depression, anxiety, and stress among university students, although pooled estimates varied by country, measure, sampling frame, and pandemic phase [3, 4]. Another systematic review likewise reported high prevalence estimates among college students [5]. A broader review of mitigation measures concluded that quarantine, campus closure, online teaching, and uncertainty about disease exposure were associated with worse student mental health outcomes [6]. Cross-national evidence from 15 countries reached a similar conclusion, suggesting that elevated psychological burden was not confined to one education system or one cultural setting [7]. Recent US evidence has further shown that psychological distress and reduced well-being remained relevant even as the pandemic moved beyond its earliest phase, with some student groups also underusing formal mental health services [8]. The common thread across these studies is that student mental health during the COVID-19 period was multidimensional. Depression and anxiety were visible endpoints, but they co-occurred with disrupted routines, altered peer interaction, changes in family exposure, sleep disturbance, reduced activity, and academic uncertainty. Analyses that isolate one endpoint at a time may therefore miss how these factors are conditionally arranged within the same student life environment. Conceptual framework: stress process, ecological pandemic context, and network psychopathology The present study frames the analysis within an integrated stress-adaptation network perspective. First, stress-process thinking provides the broad logic for linking contextual exposures, subjective appraisal, coping resources, and mental health indicators. In this view, academic disruption, altered family and peer contexts, epidemic-related attention, and daily routines are not treated as isolated covariates. They are elements of a stress ecology that may be conditionally associated with perceived stress, loneliness, sleep problems, internalizing symptoms, and adaptive resources. Second, an ecological pandemic-context perspective is needed because the data were collected between January 13 and March 26, 2021, during China’s continuing Zero-COVID-oriented prevention-and-control period. Students were navigating pandemic-specific expectations around mobility, campus management, online or altered teaching arrangements, family co-residence, and public health information. Social norms, collective solidarity, and trust in science have been identified as important pandemic-era social processes [9, 10]. Susceptibility to misinformation is also relevant for understanding public health information environments [11, 12]. The present dataset did not directly measure these macro-social constructs, but they are relevant boundary conditions for interpreting why family conflict, online adaptation, epidemic attitude, and epidemic attention may have been psychologically salient during this period. Third, network psychopathology provides an analytic lens for describing conditional association structures among constructs. Network theory conceptualizes psychopathology-related phenomena as systems of mutually related components rather than as isolated manifestations of a single latent cause [13, 14]. The bridge centrality literature provides terminology for describing nodes whose edges extend across prespecified domains, such as contextual stressors and mental-health indicators [15]. In the present study, however, these terms are used only in this descriptive statistical sense. The cross-sectional construct-level network does not test temporal symptom spreading, causal transmission, or intervention effects. This integrated framework clarifies the role of resilience in the network. Resilience is treated here as an adaptive-capacity or psychological-resource proxy within the mental-health/adaptive domain, not as a clinical outcome and not as evidence that resilience training would causally change other nodes. Because resilience can also be conceptualized as a contextual or adaptive resource, we examined an alternative community specification in which resilience was reassigned to the contextual/adaptive side of the partition. Why a multidimensional construct-level network is useful A large body of pandemic-era student research has documented that distress was entangled with academic disruption. Interview and survey studies described abrupt transitions to remote learning, concerns about diminished educational quality, reduced concentration, and fears about academic delay or future career prospects [16, 17]. More recent work has added that these pressures were not only academic in the narrow sense; students also faced financial uncertainty, household strain, and unequal access to supportive study environments [18]. The present dataset did not include a financial-stress node, which is an important limitation, but it did include online adaptation and learning efficiency as student-reported indicators of academic functioning. Evidence on online learning also suggests that academic adaptation is psychologically relevant rather than pedagogically neutral. Satisfaction with online study has been associated with resilience and with lower anxiety, depression, and stress [19, 20]. Engagement and performance in remote learning settings also appear intertwined with broader mental health functioning [21]. Chinese college data collected during lockdown linked online-learning difficulties to higher psychological strain [22]. Together, these studies support examining learning-related variables within the same conditional structure as mental health indicators, while avoiding directional claims about whether academic difficulty preceded or followed distress. Student distress during the COVID-19 period was also social and relational. Social network disruption, reduced face-to-face contact, and loneliness were repeatedly identified as salient student experiences under lockdown and social restriction [23, 24]. In university settings, peer belonging, informal support, and routine contact with classmates function as major regulatory resources. When those resources are weakened, loneliness and internalizing symptoms may co-occur even in the absence of direct illness-related fear. Family context also became more consequential during the pandemic because many students spent more time in the home environment than is typical during campus-based study. Both cross-sectional and longitudinal studies have shown that family functioning, social support, and coping context were associated with better mental health among college students during this period [25, 26]. Related work has shown that social support can attenuate associations between COVID-related stress and internalizing symptoms, especially in higher-risk student groups [27]. Lifestyle and self-regulatory factors provide a further reason to move beyond single-outcome models. Physical activity declined for many university students during the early pandemic period, partly because of confinement measures and routine disruption [28]. At the same time, sleep quality worsened across student samples [29]. Physical activity has been linked to better sleep and lower anxiety and mood symptoms in student populations [30], and exercise-based intervention syntheses suggest potential benefits for anxiety and depression during the COVID-19 prevention-and-control period [31]. In the present cross-sectional network, these literatures justify including sleep and exercise as construct-level indicators, not treating observed edges as intervention effects. Adaptive-capacity indicators should be considered alongside risk indicators. Resilience has been described as an important personal resource for college and health-science students during the pandemic, including through links with better well-being, lower distress, and more adaptive coping [32, 33]. Comparable evidence among college students has linked resilience to better mental health during COVID-19 [34]. In the present study, however, resilience is included only as an observed construct-level resource indicator, not as proof of a protective mechanism. Resilience is rarely modeled together with loneliness, sleep, family strain, academic adaptation, and internalizing symptoms in the same construct-level conditional structure. What prior network studies have and have not shown Existing network studies on student mental health during or around the COVID-19 period provide useful precedents but do not fully answer the question addressed here. Symptom-level work among Chinese college students has identified strong relations among depressive, anxious, and insomnia symptoms across pandemic stages [35]. Other network work has shown that loneliness can connect closely with anxiety and depression in university samples [36]. These studies are valuable because they demonstrate that student distress has network structure and that internalizing and social-disconnection nodes can occupy prominent network positions. At the same time, that literature remains limited for the present purpose. First, many papers focus primarily on symptom clusters, which is useful for clinical interpretation but less useful for describing how contextual, adaptive, and mental-health indicators are conditionally arranged together. Second, contextual stressors are often treated as covariates or exogenous predictors rather than modeled within the same network as mental health and adaptive-resource variables. Third, prior studies have not consistently integrated academic adaptation, family conflict, friendship change, loneliness, resilience, health behavior, and epidemic-related cognition within a single construct-level framework. The present study therefore contributes a secondary, theory-informed descriptive map rather than a definitive causal or intervention model. Study aims and rationale The present study used a large cross-sectional dataset collected from students between January 13 and March 26, 2021, a period that still reflected pandemic-related educational, social, and public health disruption in China. We estimated a construct-level network that placed academic adaptation, family and peer context, epidemic-related cognition, health behavior, perceived stress, loneliness, resilience, sleep problems, depression, and anxiety in one conditional association system. Because the study uses secondary, cross-sectional, convenience-sample data, the goal was structural description and hypothesis generation, not prevalence inference, causal pathway testing, or identification of intervention leverage points. We addressed four research questions: (RQ1) Which conditional associations among the included constructs were strongest in the estimated network? (RQ2) Which nodes had the highest expected-influence values under the present node set and estimator? (RQ3) Which nodes showed the largest bridge expected-influence values under the prespecified contextual-versus-mental-health partition, and how sensitive were these values to alternative community definitions? (RQ4) How stable were edge, centrality, and bridge estimates across bootstrap procedures, data-quality exclusions, and alternative estimator specifications? Based on prior literature, we expected depression, anxiety, perceived stress, and sleep problems to show close adjusted associations; perceived stress and loneliness to show comparatively larger cross-community connectivity; and learning-related difficulty, family conflict, resilience, and exercise to show descriptively notable but model-dependent connections to the broader network. Methods Study design and analytic framework This study was a secondary cross-sectional analysis of an existing student survey dataset. The aim was not to estimate prevalence in a representative epidemiologic sense, but to characterize the conditional structure linking contextual conditions, adaptive resources, and mental health constructs within a large convenience online sample. The analytic strategy followed a construct-level network framework because the available variables represented recoded contextual indicators and derived summary scores rather than full item-level responses for every original psychometric scale. Accordingly, the paper should be read as a network analysis of interrelated constructs, not as a formal symptom-item network study. The conceptual grouping of nodes was theory-guided but not imposed on the estimated edge structure. For bridge estimation, we prespecified two broad communities: contextual and behavioral variables on the one hand, and mental-health and adaptive-functioning variables on the other. This partition was used as a pragmatic descriptive device for summarizing cross-domain connectivity, not as a claim that the data contain one uniquely correct community structure. Bridge statistics are therefore interpreted as conditional on the present partition. To address this specification dependence, we added sensitivity analyses using alternative community definitions. The overall research workflow, from archived survey data to preprocessing, network estimation, sensitivity analyses, and reported outputs, is summarized in Fig. 1. Fig. 1Research workflow of the study Research workflow of the study Data source, setting, and participants The original survey was an online cross-sectional student survey conducted between January 13 and March 26, 2021. Because the present study analyzed an archived dataset rather than a prospectively designed network study, no a priori sample size calculation was available for the current analysis. The archived analytic file did not preserve a reconstructable sampling frame, invitation count, institution-by-institution recruitment log, or channel-specific dissemination record. Accordingly, the source is best described as a convenience online student survey coordinated through Sias University and participating local schools during the pandemic period, rather than as a formally enumerated provincial sample. Because the archived file did not preserve the number of invitations issued, a formal response rate could not be reconstructed retrospectively. The data-collection period occurred during China’s continuing Zero-COVID-oriented prevention-and-control context, after the first national outbreak but before the later Omicron period. During this phase, universities and families were still operating under pandemic-related expectations concerning mobility, campus management, infection prevention, online or altered learning arrangements, and attention to epidemic information. These conditions are important boundary features for interpreting the network. The sample should not be treated as representing all Chinese university students, students in other pandemic-policy contexts, or students in later post-pandemic periods. The archived dataset contained 10,745 submitted records with complete values on all selected nodes after recoding, and those 10,745 records constituted the final analytic sample. The study was approved by the Institutional Review Board of the School of Medicine, Sias University (approval number SYCX-2021-120). Participation was voluntary, and electronic informed consent was obtained before questionnaire completion. Because the analysis used de-identified secondary data, no attempt was made to link responses longitudinally or to infer individual trajectories across time. The sample included both female and male students and covered first-year through fourth-year university cohorts. Quality-control indicators derived during data processing included rapid completion and duplicate IP addresses. The main network was estimated on all 10,745 archived records for transparency and to avoid post hoc exclusion based on unverifiable assumptions about response validity. Sensitivity analyses then examined whether the network changed after excluding fast responders, duplicate-IP flagged records, one-record-per-IP filtering, and simultaneous exclusion of fast and duplicate-IP flagged records. Node selection and data preparation Node selection was driven by two criteria: conceptual relevance to pandemic-period student functioning and sufficient measurement quality in the available dataset. Fourteen nodes were included because they mapped onto the stress-process/ecological framework used in this study: contextual and behavioral conditions, appraisal and social-disconnection indicators, internalizing and sleep-related indicators, and adaptive-capacity indicators. Eight nodes represented contextual or behavioral conditions: media time, online adaptation, learning efficiency, family conflict, friendship change, epidemic attitude, epidemic attention, and exercise. Six nodes represented mental health or adaptive functioning: perceived stress, anxiety, depression, sleep problems, loneliness, and resilience. Within this construct-level framework, online adaptation and learning efficiency were treated as student-reported indicators of academic functioning; family conflict and friendship change as relational-context indicators; epidemic attitude and epidemic attention as epidemic-related appraisal/attention indicators; exercise and media time as behavioral-context indicators; perceived stress as subjective stress appraisal; anxiety, depression, sleep problems, and loneliness as mental-health indicators; and resilience as an adaptive-capacity proxy. The nodes were therefore modeled together as observed indicators of a shared pandemic-period student-functioning context, not as interchangeable measures of one construct or as evidence that all constructs operate on the same causal level. The included nodes were not all at the same level of abstraction. Depression, anxiety, loneliness, resilience, perceived stress, and sleep problems were broad construct scores, whereas media time, epidemic attention, online adaptation, learning efficiency, family conflict, friendship change, epidemic attitude, and exercise were narrower or lower-cardinality indicators. This heterogeneity is a feature of the archived construct-level dataset and is directly relevant for interpreting expected influence, bridge expected influence, and predictability. Differences in abstraction level, measurement precision, reliability, variance, and response scale may affect these metrics. We therefore treat them as descriptive summaries within this selected node set rather than as evidence that broad constructs and narrow indicators have directly comparable substantive importance. The contextual nodes were chosen because prior literature consistently implicated educational disruption, peer and family context, media exposure, epidemic appraisal, and daily routines in student well-being during the COVID-19 period [16, 18]. Chinese lockdown-period evidence also supported including learning difficulty as a contextual indicator [22]. The mental-health and adaptive nodes were selected because they captured sleep disturbance, social disconnection, adaptive capacity, and distress appraisal. Sleep problems, resilience, and psychological stress have each been prominent in empirical and review literature on university students during the pandemic [29, 32]. Detailed node construction, source questions, scoring rules, analysis ranges, reverse-coding information, and measurement notes are provided in Additional file 1. Two preprocessing corrections were applied before the network analysis. First, the exercise variable was recoded using a corrected category mapping because the prior coding did not faithfully represent the underlying response order. Second, the derived anxiety score was corrected because one source item had been inadvertently counted twice in an earlier derivation. All primary and sensitivity analyses reported in this manuscript used the corrected variables. Details of the coding corrections and pre/post descriptive comparisons are provided in Additional file 2. No imputation was performed because the archived analytic file available for network estimation contained complete values on all selected nodes after recoding. Statistical analysis Network estimation Edges in the primary network represent undirected regularized partial correlations among constructs after conditioning on the other included nodes. They do not represent causal pathways, temporal ordering, or directional effects. The primary network was estimated using the EBICglasso procedure implemented in bootnet, with Spearman correlations as input, a tuning parameter of 0.5, and thresholding enabled. This choice was made for two reasons. First, the node set mixed derived continuous scores with variables that were effectively ordinal indicators, making rank-based dependence estimation more defensible than Pearson correlations. Second, the analytic goal was a regularized descriptive network that minimized trivial edges while retaining larger conditional relations for inspection [37]. This specification should therefore be read as a pragmatic construct-level approximation, and the resulting edge and bridge estimates should be interpreted as estimator-dependent rather than model-invariant findings. To complement the partial-correlation network, node predictability was estimated in mgm using Gaussian neighborhood regressions for the processed numeric node set. In this workflow, mgm was used to derive node-wise explained variance (R2) rather than to fit a separate mixed-type main estimator. Predictability was treated as a model-based descriptive index of how much variance in each node was accounted for by adjacent nodes within the estimated network, not as a marker of causal determination, out-of-sample prediction, or direct manipulability [38, 39]. These R2 estimates should also be interpreted as model-dependent descriptive approximations, especially for low-cardinality contextual indicators. Expected influence and bridge expected influence Expected influence was selected as the primary centrality metric because the network included both positive and negative edges. In such settings, expected influence preserves the direction of edge contributions and therefore summarizes the net connectivity pattern of a node [40]. Bridge expected influence was computed under the prespecified two-community partition to describe nodes with comparatively larger cross-domain connectivity between the contextual-behavioral and mental-health-adaptive node sets [15]. These bridge values were treated as partition-dependent descriptive indices rather than invariant node properties or evidence of causal transmission. Throughout this manuscript, terms such as centrality, bridge, and expected influence are used in this statistical sense only and are interpreted as conditional on the selected node set, scaling, estimator, sample, and community specification. Accuracy, stability, and sensitivity analyses Edge accuracy was examined using 200 nonparametric bootstrap resamples. Expected-influence stability was evaluated using 200 case-dropping bootstrap resamples, from which the correlation stability coefficient was calculated. Bridge expected-influence stability was evaluated using an added case-dropping procedure under the primary two-community partition. Following recommended benchmarks, values above 0.25 were treated as minimally acceptable and values above 0.50 as preferable [41]. Several sensitivity analyses were conducted. First, data-quality sensitivity networks were estimated after excluding fast responders (completion time < = 60 s), excluding duplicate-IP flagged records, retaining one record per IP, and excluding both fast responders and duplicate-IP flagged records. Second, estimator sensitivity was assessed with a Spearman EBICglasso model without thresholding, a qgraph cor_auto ordinal/mixed-correlation EBICglasso model, and an mgm mixed model in which low-cardinality nodes were treated as categorical. Third, bootstrapped edge-weight difference tests were included, including reviewer-specified comparisons. Fourth, bridge results were recalculated under alternative community specifications, including a more granular domain specification and a specification that treated resilience as contextual/adaptive rather than as a mental-health node. Fifth, selected counterintuitive negative partial edges were compared with bivariate Spearman and Pearson correlations and with estimator-sensitivity edges. Software and reproducibility All analyses were run in R using readr, dplyr, tidyr, ggplot2, bootnet, qgraph, networktools, and mgm. The full reproducible workflow has been preserved in the project directory as executable R scripts and related analysis files, allowing the network, tables, figures, and sensitivity-analysis outputs to be regenerated from the analysis dataset. Results Sample characteristics The analytic sample included all 10,745 archived records because the processed dataset contained complete values for the selected nodes and the quality-control indicators were retained as descriptive flags in the primary analysis. Female respondents accounted for 67.66% of the sample, and male respondents accounted for 32.34%. First-year students were the largest subgroup (39.95%), followed by second-year (28.42%), third-year (23.04%), and fourth-year students (8.58%). Quality-control indicators showed that 179 responses (1.67%) were completed within 60 s and 931 records (8.66%) were flagged as duplicate-IP records. These values should not be interpreted as equivalent to confirmed invalid responding, but they provide context about the practical realities of online student survey data. Additional descriptors such as age, discipline or major, residence or living arrangement, and socioeconomic background were not available in the archived analytic file used for this secondary analysis. The sample profile is summarized in Table 1. Table 1Sample characteristics of the analytic sampleDomainCategory n PercentSample sizeTotal10,745100GenderFemale727067.66GenderMale347532.34GradeYear 1429339.95GradeYear 2305428.42GradeYear 3247623.04GradeYear 49228.58Quality controlFast response ( < = 60s)1791.67Quality controlDuplicate IP9318.66Percentages are based on the full analytic sample (N = 10,745) Table 1Sample characteristics of the analytic sampleDomainCategory n PercentSample sizeTotal10,745100GenderFemale727067.66GenderMale347532.34GradeYear 1429339.95GradeYear 2305428.42GradeYear 3247623.04GradeYear 49228.58Quality controlFast response ( < = 60s)1791.67Quality controlDuplicate IP9318.66Percentages are based on the full analytic sample (N = 10,745) Sample characteristics of the analytic sample Percentages are based on the full analytic sample (N = 10,745) Descriptive characteristics of network nodes The 14 nodes varied substantially in scale range and observed distribution. Among the contextual variables, epidemic attention had the highest mean level, indicating that students were frequently attending to epidemic-related information during the study period. Learning efficiency and media time also showed substantial dispersion. Among the mental health constructs, loneliness and resilience displayed broad ranges, while depression, anxiety, and sleep problems showed the expected right-skewed pattern of symptom burden in a heterogeneous student sample. Descriptive statistics and operational definitions for all nodes are shown in Table 2. Table 2Definitions and descriptive statistics of network nodesVariableLabelDefinitionMeanSDMinP25MedianP75Maxmedia_timeMedia timeDaily time spent consuming epidemic-related media information1.4491.2680.50.50.51.54.5online_adaptOnline adaptationDifficulty adapting to online learning0.9330.64801112learning_efficiencyLearning efficiencyReduced learning efficiency during the pandemic1.490.7501223family_conflictFamily conflictFrequency of conflict with family members0.4810.59800013friend_relation_indepFriendship changeWorsening friendship/peer relationship status0.8360.44301112epidemic_attitudeEpidemic attitudeMore negative or maladaptive epidemic attitude0.1390.39700002epidemic_attentionEpidemic attentionHigher attention to epidemic-related information2.1160.67302233exercise_freqExerciseHigher exercise frequency (corrected coding)1.3411.04701123PSS_scorePerceived stressPerceived stress score11.8238.53205.71411.42917.14340GAD7_scoreAnxietyAnxiety score (corrected derived scale)3.9993.922002.6255.2521PHQ9_scoreDepressionDepression score6.124.59202.75.48.127PSQI_scoreSleep problemsSleep problem score6.2125.153005.2510.521UCLA_scoreLonelinessLoneliness score39.58413.0032030405080CDRISC_scoreResilienceResilience score21.3569.772013.3332026.66740P25 and P75 indicate the 25th and 75th percentiles, respectively. Higher scores indicate higher levels of the named construct Table 2Definitions and descriptive statistics of network nodesVariableLabelDefinitionMeanSDMinP25MedianP75Maxmedia_timeMedia timeDaily time spent consuming epidemic-related media information1.4491.2680.50.50.51.54.5online_adaptOnline adaptationDifficulty adapting to online learning0.9330.64801112learning_efficiencyLearning efficiencyReduced learning efficiency during the pandemic1.490.7501223family_conflictFamily conflictFrequency of conflict with family members0.4810.59800013friend_relation_indepFriendship changeWorsening friendship/peer relationship status0.8360.44301112epidemic_attitudeEpidemic attitudeMore negative or maladaptive epidemic attitude0.1390.39700002epidemic_attentionEpidemic attentionHigher attention to epidemic-related information2.1160.67302233exercise_freqExerciseHigher exercise frequency (corrected coding)1.3411.04701123PSS_scorePerceived stressPerceived stress score11.8238.53205.71411.42917.14340GAD7_scoreAnxietyAnxiety score (corrected derived scale)3.9993.922002.6255.2521PHQ9_scoreDepressionDepression score6.124.59202.75.48.127PSQI_scoreSleep problemsSleep problem score6.2125.153005.2510.521UCLA_scoreLonelinessLoneliness score39.58413.0032030405080CDRISC_scoreResilienceResilience score21.3569.772013.3332026.66740P25 and P75 indicate the 25th and 75th percentiles, respectively. Higher scores indicate higher levels of the named construct Definitions and descriptive statistics of network nodes P25 and P75 indicate the 25th and 75th percentiles, respectively. Higher scores indicate higher levels of the named construct Overall network structure The regularized partial-correlation model retained 53 of 91 possible edges, corresponding to a density of 0.582. Thirty-one retained edges were positive and 22 were negative. Figure 2 displays the full network. Visually, the fitted graph showed an internalizing grouping containing depression, anxiety, perceived stress, and sleep problems; a social-adaptive grouping involving loneliness, resilience, and friendship change; and an academic-contextual grouping involving online adaptation, learning efficiency, and several surrounding contextual nodes. Fig. 2Regularized partial-correlation network of contextual factors and mental health constructs. Red nodes indicate contextual factors, blue nodes indicate mental health constructs, red edges indicate positive associations, blue edges indicate negative associations, thicker edges indicate larger absolute edge weights, and black rings indicate nodewise R2 Regularized partial-correlation network of contextual factors and mental health constructs. Red nodes indicate contextual factors, blue nodes indicate mental health constructs, red edges indicate positive associations, blue edges indicate negative associations, thicker edges indicate larger absolute edge weights, and black rings indicate nodewise R2 The retained nonzero edges are listed in Table 3 in descending order of absolute weight. The single strongest edge linked sleep problems and depression (weight = 0.736), followed by the edge between depression and anxiety (0.668), the edge between anxiety and perceived stress (0.459), and the negative edge between sleep problems and anxiety (-0.423). Overall, the internalizing nodes formed a dense adjusted substructure, but the within-cluster signs were not uniformly positive. Outside that region, learning efficiency and online adaptation formed the strongest contextual pair (0.349), and loneliness was linked to friendship change (0.293). Other notable negative edges included resilience and loneliness (-0.278), sleep problems and exercise (-0.194), and resilience and friendship change (-0.191). Table 3Retained nonzero edges in the final network, ordered by absolute weightLabel1Label2Weightabs_weightSleep problemsDepression0.7360.736DepressionAnxiety0.6680.668AnxietyPerceived stress0.4590.459Sleep problemsAnxiety-0.4230.423Learning efficiencyOnline adaptation0.3490.349LonelinessFriendship change0.2930.293ResilienceLoneliness-0.2780.278Perceived stressEpidemic attitude0.2270.227Sleep problemsExercise-0.1940.194ResilienceFriendship change-0.1910.191ExerciseEpidemic attention0.1540.154Epidemic attentionMedia time0.1490.149ResiliencePerceived stress0.120.12Sleep problemsEpidemic attention-0.1030.103ExerciseLearning efficiency-0.1010.101DepressionFriendship change-0.1010.101DepressionPerceived stress-0.0990.099Epidemic attentionOnline adaptation-0.0980.098ExerciseOnline adaptation-0.0910.091Sleep problemsPerceived stress0.0850.085LonelinessDepression0.0780.078Epidemic attentionFriendship change-0.0760.076Perceived stressFamily conflict0.0750.075ExerciseFamily conflict-0.0740.074ExerciseFriendship change-0.0670.067Sleep problemsFamily conflict0.0670.067Perceived stressLearning efficiency0.0660.066Friendship changeOnline adaptation0.0650.065DepressionEpidemic attention0.0640.064Friendship changeLearning efficiency0.0630.063AnxietyFamily conflict0.0610.061Sleep problemsLearning efficiency0.060.06Sleep problemsFriendship change0.0580.058DepressionExercise0.0530.053Epidemic attentionEpidemic attitude-0.0530.053Perceived stressFriendship change0.0530.053ResilienceFamily conflict-0.050.05AnxietyEpidemic attention0.050.05LonelinessFamily conflict0.0490.049Sleep problemsMedia time-0.0480.048ResilienceOnline adaptation-0.0470.047Family conflictOnline adaptation0.0460.046ResilienceEpidemic attention0.0450.045Sleep problemsOnline adaptation0.0440.044ResilienceEpidemic attitude-0.0430.043Epidemic attentionLearning efficiency-0.0430.043Epidemic attentionFamily conflict-0.040.04Perceived stressOnline adaptation0.0390.039ResilienceSleep problems-0.0380.038LonelinessPerceived stress0.0380.038ResilienceExercise0.0380.038ResilienceAnxiety-0.0330.033DepressionFamily conflict0.030.03Edges are ordered by absolute weight. Positive and negative signs indicate the direction of the regularized partial association Table 3Retained nonzero edges in the final network, ordered by absolute weightLabel1Label2Weightabs_weightSleep problemsDepression0.7360.736DepressionAnxiety0.6680.668AnxietyPerceived stress0.4590.459Sleep problemsAnxiety-0.4230.423Learning efficiencyOnline adaptation0.3490.349LonelinessFriendship change0.2930.293ResilienceLoneliness-0.2780.278Perceived stressEpidemic attitude0.2270.227Sleep problemsExercise-0.1940.194ResilienceFriendship change-0.1910.191ExerciseEpidemic attention0.1540.154Epidemic attentionMedia time0.1490.149ResiliencePerceived stress0.120.12Sleep problemsEpidemic attention-0.1030.103ExerciseLearning efficiency-0.1010.101DepressionFriendship change-0.1010.101DepressionPerceived stress-0.0990.099Epidemic attentionOnline adaptation-0.0980.098ExerciseOnline adaptation-0.0910.091Sleep problemsPerceived stress0.0850.085LonelinessDepression0.0780.078Epidemic attentionFriendship change-0.0760.076Perceived stressFamily conflict0.0750.075ExerciseFamily conflict-0.0740.074ExerciseFriendship change-0.0670.067Sleep problemsFamily conflict0.0670.067Perceived stressLearning efficiency0.0660.066Friendship changeOnline adaptation0.0650.065DepressionEpidemic attention0.0640.064Friendship changeLearning efficiency0.0630.063AnxietyFamily conflict0.0610.061Sleep problemsLearning efficiency0.060.06Sleep problemsFriendship change0.0580.058DepressionExercise0.0530.053Epidemic attentionEpidemic attitude-0.0530.053Perceived stressFriendship change0.0530.053ResilienceFamily conflict-0.050.05AnxietyEpidemic attention0.050.05LonelinessFamily conflict0.0490.049Sleep problemsMedia time-0.0480.048ResilienceOnline adaptation-0.0470.047Family conflictOnline adaptation0.0460.046ResilienceEpidemic attention0.0450.045Sleep problemsOnline adaptation0.0440.044ResilienceEpidemic attitude-0.0430.043Epidemic attentionLearning efficiency-0.0430.043Epidemic attentionFamily conflict-0.040.04Perceived stressOnline adaptation0.0390.039ResilienceSleep problems-0.0380.038LonelinessPerceived stress0.0380.038ResilienceExercise0.0380.038ResilienceAnxiety-0.0330.033DepressionFamily conflict0.030.03Edges are ordered by absolute weight. Positive and negative signs indicate the direction of the regularized partial association Retained nonzero edges in the final network, ordered by absolute weight Edges are ordered by absolute weight. Positive and negative signs indicate the direction of the regularized partial association Expected influence, bridge expected influence, and nodewise R2 Expected influence values are shown in Fig. 3; Table 4. Depression had the highest expected influence under the present node set and estimator (1.428), followed by perceived stress (1.064) and anxiety (0.782). Among the contextual variables, learning efficiency (0.395) and online adaptation (0.306) showed the highest expected influence values. These values describe net connectivity in the fitted network and should not be read as direct evidence of clinical priority or causal importance. Fig. 3Expected influence and bridge expected influence of each node. Larger values indicate greater signed expected influence or cross-community connectivity within the estimated network specification Expected influence and bridge expected influence of each node. Larger values indicate greater signed expected influence or cross-community connectivity within the estimated network specification Table 4Expected influence, bridge expected influence, and predictabilityVariableLabelCommunityExpected_InfluenceBridge_Expected_InfluenceStrengthPredictability_R2PHQ9_scoreDepressionMental health1.4280.0461.8290.815PSS_scorePerceived stressMental health1.0640.4611.2620.448GAD7_ scoreAnxietyMental health0.7820.1111.6930.738learning_efficiencyLearning efficiencyStressors0.3950.1260.6830.248online_adaptOnline adaptationStressors0.3060.0360.7780.247PSQI_ scoreSleep problemsMental health0.244-0.1151.8560.704UCLA_ scoreLonelinessMental health0.1810.3430.7370.291family_conflictFamily conflictStressors0.1640.2330.4940.169epidemic_attitudeEpidemic attitudeStressors0.1310.1840.3240.087media_timeMedia timeStressors0.101-0.0480.1970.021friend_relation_indepFriendship changeStressors0.0970.1120.9680.268epidemic_attentionEpidemic attentionStressors0.050.0560.8740.171exercise_freqExerciseStressors-0.284-0.1040.7730.223CDRISC_ scoreResilienceMental health-0.477-0.2480.8830.252Expected influence and bridge expected influence are signed centrality metrics. Predictability is reported as node-wise R2 Table 4Expected influence, bridge expected influence, and predictabilityVariableLabelCommunityExpected_InfluenceBridge_Expected_InfluenceStrengthPredictability_R2PHQ9_scoreDepressionMental health1.4280.0461.8290.815PSS_scorePerceived stressMental health1.0640.4611.2620.448GAD7_ scoreAnxietyMental health0.7820.1111.6930.738learning_efficiencyLearning efficiencyStressors0.3950.1260.6830.248online_adaptOnline adaptationStressors0.3060.0360.7780.247PSQI_ scoreSleep problemsMental health0.244-0.1151.8560.704UCLA_ scoreLonelinessMental health0.1810.3430.7370.291family_conflictFamily conflictStressors0.1640.2330.4940.169epidemic_attitudeEpidemic attitudeStressors0.1310.1840.3240.087media_timeMedia timeStressors0.101-0.0480.1970.021friend_relation_indepFriendship changeStressors0.0970.1120.9680.268epidemic_attentionEpidemic attentionStressors0.050.0560.8740.171exercise_freqExerciseStressors-0.284-0.1040.7730.223CDRISC_ scoreResilienceMental health-0.477-0.2480.8830.252Expected influence and bridge expected influence are signed centrality metrics. Predictability is reported as node-wise R2 Expected influence, bridge expected influence, and predictability Expected influence and bridge expected influence are signed centrality metrics. Predictability is reported as node-wise R2 Bridge expected influence, calculated under the prespecified two-community partition, is also shown in Table 4. Perceived stress had the largest bridge value (0.461), followed by loneliness (0.343), family conflict (0.233), epidemic attitude (0.184), and learning efficiency (0.126). Within this partition, bridge expected influence summarized cross-domain connectivity rather than global centrality. Because bridge expected influence depends on the community partition, these rankings are interpreted as descriptive and partition-dependent. Nodewise R2 values are shown in Fig. 4. Depression had the highest R2 (0.815), followed by anxiety (0.738) and sleep problems (0.704). Within the present node set and estimation framework, these values indicate that a large share of variance in the core internalizing nodes was accounted for by their immediate neighbors. By contrast, media time showed very low R2 (0.021), while epidemic attitude (0.087) and family conflict (0.169) were also relatively weakly accounted for by adjacent nodes. These R2 values are model-dependent descriptive approximations and should not be treated as out-of-sample prediction or a substantive hierarchy across heterogeneous processed variables. Fig. 4Node predictability expressed as explained variance (R2). Higher values indicate that a greater proportion of node variance is accounted for by adjacent nodes in the fitted network Node predictability expressed as explained variance (R2). Higher values indicate that a greater proportion of node variance is accounted for by adjacent nodes in the fitted network Accuracy, stability, and sensitivity analyses The network accuracy, stability, and sensitivity diagnostics are summarized in Table 5 and in Additional files 3–5. Bootstrap procedures used 200 nonparametric resamples for edge confidence intervals and 200 case-dropping resamples for expected-influence stability. The correlation stability coefficient for expected influence was 0.75, which exceeds commonly cited thresholds for acceptable and preferred stability. Bridge expected influence also showed high case-dropping stability under the primary two-community specification. The approximate bridge expected-influence CS coefficient was 0.75; even when approximately 75% of cases were dropped, the mean correlation with the original bridge expected-influence profile was 0.955, with a 5% quantile of 0.923. These statistics indicate stability of the estimated profiles under the tested resampling procedures, not causal or intervention stability (Figs. 5 and 6). Table 5Accuracy, stability, and sensitivity summary of the estimated networkMetricValueBootstrap edge CIs200 nonparametric resamplesCase-dropping bootstrap for expected influence200 case-dropping resamplesCorrelation stability coefficient (expected influence)0.75Approximate bridge EI CS coefficient0.75Fast responder sensitivityN = 10,566; signed edge r = 1.000; EI r = 1.000; bridge EI r = 1.000; top-5 overlaps = 5/5 and 5/5Duplicate-IP sensitivityN = 9,814; signed edge r = 1.000; EI r = 1.000; bridge EI r = 1.000; top-5 overlaps = 5/5 and 5/5No fast + no duplicate-IP sensitivityN = 9,661; signed edge r = 0.999; EI r = 0.999; bridge EI r = 0.999; top-5 overlaps = 5/5 and 5/5cor_auto EBICglasso sensitivitysigned edge r = 0.984; EI r = 0.988; bridge EI r = 0.966; top-5 overlaps = 5/5 and 4/5MGM mixed sensitivityUnsigned edge-weight model; centrality rankings differed and were not interpreted as signed EI replicationRecommended interpretation threshold>= 0.25 acceptablePreferred threshold>= 0.50 preferredCS correlation stability, EI expected influence, Bridge EI bridge expected influence Table 5Accuracy, stability, and sensitivity summary of the estimated networkMetricValueBootstrap edge CIs200 nonparametric resamplesCase-dropping bootstrap for expected influence200 case-dropping resamplesCorrelation stability coefficient (expected influence)0.75Approximate bridge EI CS coefficient0.75Fast responder sensitivityN = 10,566; signed edge r = 1.000; EI r = 1.000; bridge EI r = 1.000; top-5 overlaps = 5/5 and 5/5Duplicate-IP sensitivityN = 9,814; signed edge r = 1.000; EI r = 1.000; bridge EI r = 1.000; top-5 overlaps = 5/5 and 5/5No fast + no duplicate-IP sensitivityN = 9,661; signed edge r = 0.999; EI r = 0.999; bridge EI r = 0.999; top-5 overlaps = 5/5 and 5/5cor_auto EBICglasso sensitivitysigned edge r = 0.984; EI r = 0.988; bridge EI r = 0.966; top-5 overlaps = 5/5 and 4/5MGM mixed sensitivityUnsigned edge-weight model; centrality rankings differed and were not interpreted as signed EI replicationRecommended interpretation threshold>= 0.25 acceptablePreferred threshold>= 0.50 preferredCS correlation stability, EI expected influence, Bridge EI bridge expected influence Accuracy, stability, and sensitivity summary of the estimated network CS correlation stability, EI expected influence, Bridge EI bridge expected influence Fig. 5Bootstrap confidence intervals for the strongest retained edges. Points represent sample edge weights and horizontal intervals represent nonparametric bootstrap confidence intervals Bootstrap confidence intervals for the strongest retained edges. Points represent sample edge weights and horizontal intervals represent nonparametric bootstrap confidence intervals Fig. 6Case-dropping bootstrap stability of expected influence. The solid line shows mean correlation with the original expected-influence profile across case-dropping levels, and the shaded bands show interval dispersion Case-dropping bootstrap stability of expected influence. The solid line shows mean correlation with the original expected-influence profile across case-dropping levels, and the shaded bands show interval dispersion Data-quality exclusions yielded estimates very similar to the primary network. Excluding fast responders (N = 10,566), excluding duplicate-IP flagged records (N = 9,814), retaining one record per IP (N = 9,814), and excluding both fast and duplicate-IP flagged records (N = 9,661) yielded signed edge correlations with the primary network of 1.000, 1.000, 1.000, and 0.999, respectively. The top-five expected-influence nodes and top-five bridge expected-influence nodes were preserved in all four data-quality sensitivity models. Estimator sensitivity showed a more nuanced pattern. A Spearman EBICglasso model without thresholding and a cor_auto ordinal/mixed-correlation EBICglasso model largely reproduced the primary structure. For the cor_auto model, the signed edge correlation with the primary model was 0.984, expected-influence correlation was 0.988, bridge expected-influence correlation was 0.966, top-five expected-influence overlap was 5/5, and top-five bridge overlap was 4/5. In contrast, the mgm mixed model based on unsigned edge weights yielded different centrality rankings. We therefore interpret the primary centrality and bridge results as robust to data-quality exclusions and a rank-/ordinal-correlation sensitivity model, but not invariant across all possible estimator choices. Bootstrapped edge-difference tests supported only selected relative edge-strength statements. In reviewer-specified comparisons, the absolute edge for anxiety-sleep problems was stronger than the edge for depression-perceived stress (absolute difference = 0.324, 95% CI [0.305, 0.344]), and the edge for depression-loneliness was stronger than the edge for family conflict-depression (absolute difference = 0.048, 95% CI [0.029, 0.084]). The difference between family conflict-perceived stress and learning efficiency-perceived stress was not statistically supported (absolute difference = 0.009, 95% CI [-0.020, 0.029]). We therefore avoid interpreting unsupported edge differences as meaningful rank differences. Alternative community specifications indicated that bridge rankings were partly partition-dependent. Perceived stress had the highest bridge expected-influence value under the primary two-community, granular six-domain, and resilience-as-contextual specifications. Family conflict and epidemic attitude also remained within the top five across these specifications. Loneliness ranked second under the primary and granular specifications but dropped when resilience was treated as contextual/adaptive. Learning efficiency remained in the top five under the primary and resilience-as-contextual specifications but not under the granular six-domain specification. These results support cautious interpretation of perceived stress as the most stable bridge expected-influence finding while treating other bridge ranks as partition-sensitive and specification-dependent. Several adjusted edges had signs that differed from their bivariate correlations. In particular, sleep problems-anxiety was positive at the bivariate level (Spearman rho = 0.211) but negative in the primary partial-correlation network (-0.423) and the cor_auto sensitivity model (-0.458). Depression-perceived stress was also positive at the bivariate level (Spearman rho = 0.369) but negative in the primary network (-0.099) and cor_auto sensitivity model (-0.100). These signs are therefore interpreted cautiously as conditional associations that may reflect suppression, scaling, or shared adjustment by other internalizing nodes rather than substantive inverse psychological relations. Discussion Principal findings This study estimated a construct-level network linking contextual stressors, behavioral factors, adaptive-resource indicators, and mental health constructs in a large convenience online sample of students during a specific Zero-COVID-oriented prevention-and-control period in China. Four findings stand out, all of which should be read as descriptive conditional-association patterns. First, the fitted network showed a dense internalizing grouping involving depression, perceived stress, anxiety, and sleep problems. Depression had the highest expected-influence value under the present node set and estimator, and the strongest positive edges connected depression to sleep problems and anxiety, with perceived stress conditionally associated with anxiety. This pattern is broadly compatible with pandemic-era evidence that depression, anxiety, and stress co-occurred in student populations rather than operating as isolated endpoints [3, 4]. It also aligns with evidence on sleep disturbance in student samples [29] and with recent symptom-network work showing close associations among internalizing symptoms in college students across pandemic stages [35]. Second, under the prespecified contextual-versus-mental-health partition, perceived stress and loneliness showed comparatively larger bridge expected-influence values, with perceived stress remaining the highest-ranked bridge expected-influence node across alternative community specifications. This distinction matters because the node with the highest expected-influence value was not necessarily the same node with the largest cross-domain connectivity. The finding is consistent with the broad role of stress appraisal in pandemic-era student mental health [42] and with evidence linking loneliness closely to anxiety and depression in university samples [24, 36]. However, these bridge results describe cross-community connectivity within specified partitions and do not demonstrate cross-domain transmission. Third, among contextual nodes, family conflict, epidemic attitude, and learning efficiency showed notable cross-domain connectivity in the primary partition. These results are compatible with studies showing that student distress during COVID-19 was associated with academic disruption and economic strain [18]. Online-study satisfaction has been linked with mental health functioning [19, 20]. Learning engagement has also been linked with mental health functioning in remote-learning settings [21]. Family context has likewise been associated with student mental health during this period [25, 26]. In the present network, these variables were part of the same conditional system as broader distress, but the bridge ranking should be read as descriptive and partition-dependent rather than as a stable hierarchy of intervention priorities. Fourth, resilience and exercise showed several conditional edges that were directionally compatible with adaptive-resource interpretations, but the pattern was mixed and should not be reified. Resilience had negative edges with loneliness, friendship change, family conflict, negative epidemic attitude, sleep problems, and anxiety, but it was also positively linked with perceived stress in the adjusted network. Exercise had negative edges with several difficulty indicators but also positive edges with epidemic attention, resilience, and depression. These mixed patterns reinforce the need to interpret construct-level partial associations cautiously and not as evidence of protective mechanisms. Interpretation within the integrated framework The internalizing grouping can be understood within a stress-process and network framework as a set of mutually conditioned indicators rather than as a linear sequence from stressor to symptom. The depression-sleep and depression-anxiety edges were the strongest positive adjusted associations, while anxiety and perceived stress were also closely connected. This does not mean that depression, sleep, anxiety, and stress caused one another in this dataset. It indicates that, within the observed construct set, these indicators were closely arranged after adjustment for the other nodes. The bridge expected-influence pattern provides a complementary descriptive layer. Perceived stress is conceptually relevant to both contextual and mental-health domains because it reflects subjective appraisal of situational demands. Student life during the study period involved academic uncertainty, epidemic information, changed social expectations, and family or campus constraints. In this fitted network, perceived stress occupied a descriptive cross-domain position among these indicators. The present cross-sectional network supports this interpretation only as a conditional association pattern, not as evidence that stress appraisal transmits effects from context to symptoms. Loneliness should be interpreted with similar caution. It showed a high bridge rank under the primary and granular community specifications, but its rank was sensitive to the placement of resilience. This pattern is compatible with the pandemic literature on peer disconnection and loneliness [23, 24], but it does not establish that loneliness is an invariant bridge or a causal spreading node. Future longitudinal or intensive repeated-measures designs would be needed to test whether loneliness temporally precedes, follows, or reciprocally interacts with internalizing symptoms. Pandemic and sociocultural context The data were collected during a specific Chinese pandemic-policy and social context. By early 2021, China was no longer in the first national outbreak phase, but COVID-19 prevention-and-control remained a major organizing feature of campus, family, and public life. Students’ experiences of online adaptation, learning efficiency, family conflict, epidemic attention, and epidemic attitude should therefore be interpreted in relation to that context rather than generalized to all pandemic phases or policy environments. The reviewer comments also highlight that pandemic context is broader than proximal stressors. Group norms, collective solidarity, and trust in science are relevant to public health beliefs and responses during crises [9, 10]. Infodemic processes are likewise relevant because heavy or confusing information exposure can amplify uncertainty, fear, and maladaptive appraisals [11]. The current network included media time, epidemic attention, and epidemic attitude, but it did not directly measure compliance norms, trust in science, solidarity, misinformation exposure, or institutional trust. These unmeasured constructs are therefore important boundary conditions for interpreting the relatively peripheral adjusted positions of media time and epidemic attention. Peripheral network placement in this dataset does not mean that information environments were unimportant; it only means that, after conditioning on the included constructs, these nodes had relatively limited direct connectivity. Economic context is another important omission. Financial stress has been repeatedly identified as a pandemic-era student stressor [18], but no financial-stress node was available in the archived analytic dataset. Some variance attributed to perceived stress, academic strain, or family conflict may therefore reflect unmeasured economic pressure. This limitation narrows the theoretical interpretation of the network and should be addressed in future studies designed prospectively around pandemic-era ecological stressors. What the network can and cannot imply The sensitivity analyses strengthen uncertainty assessment but do not change the basic evidential boundary of the study. The data are cross-sectional, and the network edges are undirected partial correlations. The results can describe conditional association patterns within a selected construct set. They cannot determine whether learning inefficiency precedes depression, whether depression precedes learning inefficiency, whether loneliness temporally links domains, or whether changing a node with a high bridge value would alter the rest of the network. The sensitivity analyses also clarify which claims are better supported. Data-quality exclusions and the cor_auto sensitivity model support broad similarity to the primary network structure. Edge-difference tests support only selected relative edge-strength statements. Bridge expected-influence stability was high under the primary partition, but bridge ranks remained community-specification dependent. The mgm mixed model based on unsigned edge weights produced different centrality rankings, reinforcing that centrality is not an invariant property of the students or constructs. For these reasons, this manuscript avoids using centrality or bridge metrics as direct proxies for substantive importance, clinical priority, or intervention targets. The counterintuitive negative partial edges are a final example of this caution. Sleep problems and anxiety, and depression and perceived stress, were positively associated at the bivariate level but negatively connected in the adjusted network. These signs likely reflect conditional dependence, suppression, scaling, or the influence of shared adjustment by other internalizing nodes. They should not be interpreted as evidence that sleep problems protect against anxiety or that depression is inversely related to perceived stress in ordinary psychological terms. Implications for hypothesis generation The most defensible implication is that student mental health during this period may be usefully studied as a multidimensional stress-adaptation system. In future prospective or intervention-oriented work, perceived stress, loneliness, learning efficiency, family conflict, and epidemic attitudes may be reasonable candidate constructs to monitor alongside depression, anxiety, and sleep problems. This is not a recommendation to treat these variables as confirmed intervention targets. Rather, the current findings identify conditional-association patterns that future longitudinal, experimental, or implementation studies can test more directly. For university mental health planning, the findings suggest that symptom indicators and contextual indicators should not be analytically separated too quickly. Academic functioning, peer and family context, and subjective stress appraisal may help characterize which students experience distress in particular ecological configurations. However, because the current sample is not representative and the design is cross-sectional, the results should inform hypothesis generation and future assessment design rather than immediate policy mandates. Limitations Several limitations should be acknowledged. First, the study is cross-sectional. None of the observed edges should be interpreted as temporal or causal pathways, and statistical centrality should not be equated directly with clinical priority. A high bridge expected-influence value in a cross-sectional network is a descriptive statistic under a specified partition, not proof that change in that node will propagate through the system in a specific direction. Second, the construct scores were derived from an existing survey structure rather than full original item batteries in every case. The node set mixed broad psychological constructs, derived scale scores, single indicators, and low-cardinality ordinal variables. This limits direct comparability with formal symptom-item networks and means that expected influence, bridge expected influence, and predictability comparisons are affected by node abstraction level, scale properties, distribution, and measurement reliability. The exercise and anxiety coding corrections were handled transparently and all analyses used corrected variables, but the need for these corrections underscores the measurement limitations of secondary data. Third, the sample reflects one institution-level convenience online dataset during a specific pandemic phase, with most responses concentrated in mid-January 2021. The archived dataset did not preserve a reconstructable sampling frame, invitation count, recruitment log, or response rate. Generalization to other universities, other provinces, later pandemic stages, post-pandemic conditions, or students under different policy environments should therefore be cautious. Fourth, the community partition was prespecified for descriptive bridge analysis, and alternative partitions changed some bridge rankings. Although perceived stress retained the highest and most stable bridge expected-influence value across the tested specifications, loneliness and learning efficiency were more partition-sensitive and more sensitive to how communities were defined. Bridge expected influence should therefore be interpreted as partition-dependent rather than as an invariant cross-domain property. Fifth, estimator choice mattered. The primary Spearman EBICglasso network was robust to data-quality exclusions and broadly reproduced by the cor_auto sensitivity model, but the mgm mixed model based on unsigned weights produced different centrality rankings. This reinforces the need to interpret centrality and bridge metrics as model-dependent summaries. Sixth, the dataset did not include several theoretically important pandemic-context variables, including financial stress, trust in science, compliance norms, collective solidarity, misinformation exposure, conspiracy beliefs, living arrangement, socioeconomic background, and institutional support. These omissions constrain interpretation of the contextual network and should guide future data collection. Conclusion This study provides a construct-level descriptive snapshot of conditional associations among contextual stressors, adaptive-resource indicators, and mental health indicators in a large convenience online sample of students during a specific Zero-COVID-oriented prevention-and-control period in China. Depression, perceived stress, anxiety, and sleep problems showed a dense internalizing structure, while perceived stress had the highest and most stable bridge expected-influence value across tested community specifications. Loneliness, family conflict, epidemic attitude, and learning efficiency also showed notable bridge expected-influence values under the primary partition, but some rankings were specification-dependent. The findings are best interpreted as hypothesis-generating evidence for future longitudinal, experimental, or implementation research, not as evidence of causal pathways, invariant bridge nodes, or confirmed intervention targets. Supplementary Information Additional file 1: Supplementary Table S1. Node construction, scoring, and measurement notes. Additional file 1: Supplementary Table S1. Node construction, scoring, and measurement notes. Additional file 2: Supplementary Table S2. Coding corrections and pre/post descriptive comparison. Additional file 2: Supplementary Table S2. Coding corrections and pre/post descriptive comparison. Additional file 3: Supplementary Tables S3-S13. Sensitivity analyses. Additional file 3: Supplementary Tables S3-S13. Sensitivity analyses. Additional file 4: Supplementary Figure S1. Edge-difference tests. Additional file 4: Supplementary Figure S1. Edge-difference tests. Additional file 5: Supplementary Figure S2. Bridge expected-influence case-dropping stability. Additional file 5: Supplementary Figure S2. Bridge expected-influence case-dropping stability.