The panic–sadness bridge: a longitudinal study using undirected and directed network analyses of COVID-19-related anxiety and depression Authors: Jiamin Bao, Yuxuan Wang, Yixuan Wang, Qiang Chen, Yinyin Zang Published: 2025 ABSTRACT AbstractBackgroundThe comorbidity network between anxiety and depression during COVID-19 has received significant attention. However, existing research has been primarily undirected and cross-sectional, lacking an examination of directional structure and longitudinal progression.MethodsThis study employed three complementary network models: the Gaussian Graphical Model (GGM), the Directed Acyclic Graph (DAG), and the Cross-Lagged Panel Network (CLPN). A total of 533 Chinese community participants (mean age = 27.75; 56.3% female) completed the Hospital Anxiety and Depression Scale (HADS). Assessments were conducted at three-month intervals: during the outbreak (February 7 to March 10, 2020) and the post-peak stage (May 8 to June 24, 2020).ResultsAnxiety and depression symptoms were highly interconnected and mutually reinforcing over time. The overall network structure and global strength remained stable. The connection between panic feelings and sad moods emerged as a critical bridge linking anxiety and depression. DAGs suggested that instantaneous associations between panic feelings and sad moods were mutually reinforcing, whereas CLPN revealed that sad moods prospectively predicted subsequent increases in panic feelings.ConclusionsAnxiety and depression symptoms mutually reinforced each other during the early stage of COVID-19, with the panic-sadness link serving as a central bridge. Addressing this maladaptive interaction may strengthen future interventions.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12888-025-07733-z.Keywords: Network analysis, Directed acyclic graph, Cross-lagged panel network, Longitudinal, Anxiety, Depression, Panic Background COVID-19 is one of the most significant global disaster events in recent history. As of June 2024, approximately 776 million people worldwide have been infected with COVID-19, with China accounting for about 13% of these cases [1]. The initial outbreak of COVID-19 in Wuhan, China, resulted in over 80,000 infections [2]. During this period, COVID-19 was perceived as a highly transmissible disease with a high mortality rate [3], causing widespread anxiety and depression [4]. The situation worsened as many determinants of mental health were disrupted during the pandemic, including reduced social contact due to lockdowns. Indeed, a meta-analysis reported increased anxiety and depression rates due to COVID-19 [4], affecting 31.9% and 33.7% of the population, respectively, during its early stage (through May 2020) [5]. Notably, anxiety and depression often co-occur, with an estimated rate of 22.4% in the Chinese general population during the initial COVID-19 outbreak [6], resulting in more severe symptoms, functional impairments, and treatment resistance [7]. However, the mechanisms underlying symptom development during COVID-19 may differ from typical circumstances [8], requiring further investigation. The coexistence of anxiety and depression can be understood from a network perspective [9]. The network approach explicitly departs from the traditional latent disease model by treating symptoms not as passive indicators of an underlying disorder but as interacting components that collectively constitute syndromes [10]. In this framework, symptoms are conceptualized as nodes, and their interconnections as edges. Activation of one symptom can spread through the network by inter-symptom connections, potentially activating other symptoms and resulting in consistent syndromes [11]. Thus, supposedly distinct symptom clusters like anxiety and depression are not necessarily separate phenomena. Instead, their co-occurrence arises from direct interconnections, where strongly linked symptoms act as transdiagnostic bridges linking different syndromes [12]. Unlike traditional categorical approaches, network analysis prioritizes interactions between symptoms and offers a more ecologically grounded perspective for understanding comorbidity [13]. While prior studies have examined the co-occurrence of anxiety and depression syndromes within network frameworks [14–16], directional symptom relationships remain underexplored. Traditional disease-centered research has suggested that anxiety syndrome often precedes depression, a pattern also observed during the initial COVID-19 outbreak [17, 18]. This sequence aligns with the General Adaptation Syndrome (GAS) framework [19], which conceptualizes stress responses as progressing from initial anxiety and physiological arousal to subsequent depressive reactions as coping resources are exhausted. These theoretical considerations provide a foundation for examining directional links between anxiety and depression symptoms. Yet, empirical investigation remains limited. Directed network models offer an opportunity to identify symptom-specific associations and reveal core or bridge symptoms that may have predictive precedence in the development and maintenance of anxiety and depression syndromes. Furthermore, symptom networks are dynamic rather than static, especially in the context of ongoing and unpredictable disasters such as COVID-19 [20]. For example, among survivors of the Jiuzhaigou earthquake in China, within-syndrome connectivity of depression and posttraumatic stress disorder (PTSD) symptoms strengthened from the first to the second year post-disaster, while between-syndrome connectivity weakened [21]. This pattern suggests that initially co-occurring psychological distress may progressively differentiate into distinct syndromes over time [21]. Additionally, in the six months following the lifting of COVID-19 lockdown, the central PTSD symptoms shifted from irritability and hypervigilance to concentration difficulties and avoidance, while the central depression symptom remained sad moods [22]. Compared to the networks of PTSD and depression, the dynamics of the anxiety-depression network during the pandemic remain less understood. Zavlis et al. [23] found that anxiety and depression symptoms clustered separately in the initial COVID-19 lockdown but merged into a single community within about a month. However, their study focused primarily on the role of contextual factors (e.g., lost income during COVID-19) in symptom development, rather than on the interrelationships between anxiety and depression symptom clusters; additionally, no directed analyses were performed. Several other studies examining dynamic changes in anxiety-depression networks have relied on cross-sectional datasets [16, 24], which may introduce sampling errors when inferring longitudinal changes. More longitudinal studies are necessary to understand how the interaction between anxiety and depression evolves in response to dynamic disasters such as COVID-19. Overall, despite some insights into the anxiety-depression network during COVID-19, few have examined its directional structure and longitudinal course. To fill this gap, we employed a multi-method network analytic approach with two-wave longitudinal data (T1: outbreak, and T2: post-peak stage). First, we estimated and compared cross-sectional network structures at T1 and T2 using regularized Gaussian Graphical Models (GGMs) and Directed Acyclic Graphs (DAGs). Second, we constructed a change score (T1-T2) GGM and DAG to capture associations among symptom changes over time, minimizing stable trait effects. Third, the Cross-Lagged Panel Network (CLPN) was applied to establish lagged temporal effects between symptoms [25]. This design leverages complementary strengths of each model: GGMs map undirected symptom associations without inferring causality, enabling direct comparison of anxiety-depression network structure across pandemic stages; DAGs extend this by estimating probabilistic directions and instantaneous causal hypotheses [26]; CLPN uniquely captures lagged temporal effects, testing how symptoms predict future changes in others. Integrating these models in longitudinal data allows a comprehensive examination of both undirected and directed symptom relationships and their evolution over time, providing insight into the dynamic interplay and potential causal pathways within anxiety-depression networks. Based on the context of COVID-19 pandemic and previous network studies, this study proposes the following hypotheses. First, we tentatively hypothesize that within GGMs, anxiety and depression symptoms will form separate clusters during the initial outbreak but merge into a unified community in the post-peak stage. Given the limited evidence [23], this hypothesis should be considered exploratory. Additionally, due to the persistent psychological threat and sense of loss posed by COVID-19, fear-related symptoms (e.g., dread, panic) and sadness are expected to exhibit heightened centrality. Furthermore, as the pandemic progresses, individuals may develop greater coping capacities and psychological resilience, potentially leading to a reduction in overall network connectivity. Second, in the directed analyses (DAG and CLPN), anxiety and depression symptoms are expected to influence each other at the network level, with anxiety symptoms more likely to drive depressive symptoms. Given the limited evidence on specific directional pathways, no specific predictions are made regarding directional bridge symptoms and edges. Methods Participants and procedures Internet-based self-report surveys were conducted through Wenjuanxing (www.wjx.cn), a survey administration platform, and disseminated through snowball sampling on major social platforms in China, including WeChat, Weibo, Zhihu, and Douban. Data collection for T1 was conducted during the pandemic outbreak (from February 7, 2020, to March 10, 2020), when the number of new confirmed cases of COVID-19 generally exceeded 500 per day in China [2]. Data collection for T2 was in the post-peak (from May 8, 2020, to June 24, 2020), when the pandemic was under control in China, and the number of new confirmed cases per day was mainly less than 20 [2]. Respondents to the first survey were contacted via email, SMS, or WeChat to conduct the next phase. Each WeChat ID was granted access to complete the questionnaire only once to control the quality of the response. Respondents not currently residing in mainland China and those who completed the questionnaire in less than 420 s in T1 (350 s in T2) were excluded from the analysis. Due to the small sample size (< 5%), COVID-19-related healthcare workers and diagnosed/suspected COVID-19 patients were excluded from the current analysis. Of the 1632 valid respondents at T1, 533 (32.66%) completed two waves of data collection and were included in the analyses. The final sample (N = 533) included 300 (56.3%) women and 233 (43.7%) men, primarily young adults (mean age: 27.75 years, SD = 8.47) with a bachelor’s degree or above (70.4%) and in a relationship (53.7%). This study was approved by the Ethics Committee of the School of Psychological and Cognitive Sciences, Peking University. All participants signed the informed consent electronically. All valid responses received 6 yuan (approximately US$0.9) for completing T1 survey and 12 yuan (approximately US$1.8) for T2. Measures Demographic information included biological sex, age, marital status, and education level. The Hospital Anxiety and Depression Scale (HADS) was used to assess symptoms of depression and anxiety [27]. The HADS is a 14-item questionnaire with two subscales (7 items for each) on a 4-point Likert scale ranging from 0 to 3. Higher scores indicate greater symptom severity, with eight as the cut-off point for each subscale [28]. The Chinese version of HADS has excellent psychometric properties and is used widely in the Chinese population [28]. The Cronbach’s α coefficients in the current sample were 0.860 (0.865), 0.775 (0.785), and 0.764 (0.769) for the total scale and its depression and anxiety subscales at the outbreak (post-peak), respectively. Statistical analysis No data were missing except for age (n = 64, 3.92% of the 1632 valid responses at T1). Missing age data were not significantly associated with gender or symptom severity at T1 (ps > 0.508). Consequently, the missingness was assumed to be random, and the expectation-maximization algorithm was applied. Statistical analyses were conducted with SPSS (version 26.0 for Windows) and R. P values less than 0.05 (two-tailed tests) were considered statistically significant, and they were not corrected in the network analysis to discover as many differences as possible. To investigate the dynamic changes in the anxiety-depression network, the following analyses were conducted both in (a) cross-sectional networks, using the scores of individual symptoms as nodes, and (b) the longitudinal change networks, using the changes in symptom severity between the outbreak and post-peak stages as variables for estimating the network [22, 29]. Community detection Community analysis was used to determine whether anxiety and depression symptoms could be reliably identified as separate communities or formed highly correlated clusters. Community structure was evaluated by five methods (see Table S1 for details) [30]. We used the Walktrap algorithm as the main index, as it outperformed alternative algorithms regarding accuracy, bias, and item placement [31]. This approach detects communities by iteratively merging node groups with similar short random walk profiles [32]. Gaussian graphical models Network estimation and visualization The models estimated the network based on Spearman correlations by the graphical Least Absolute Shrinkage and Selection Operator (LASSO) algorithm [33]. It is an undirected weighted network and provides a sparse graph that depicts only the relevant partial correlations between pairs of symptoms [11]. The network was estimated using the EBICglasso in R package qgraph [34], and network layouts were determined with the Fruchterman-Reingold algorithm, which optimizes spatial uniformity and offers good computational efficiency [35]. Centrality estimation Expected influence (EI) was used to measure node interrelations, calculated as the algebraic sum of all edge weights connected to a node [36]. This metric better reflected the contribution of a given node than other indices (e.g., strength: the absolute sum of edge weights). Bridge EI extends EI to identify key symptoms in linking different disorders by summing only the algebraic edge weight between a given node and nodes in other communities (as defined by HADS) [37]. Both EI and bridge EI assume parallel symptom-to-symptom connections, with each symptom influencing all its neighbors simultaneously and independently [38]. Our analysis was conducted using the R package qgraph [34] and networktools [37], with higher values indicating greater importance. Accuracy and stability estimation The accuracy of edge weights was estimated by drawing 95% confidence intervals (CIs) from a non-parametric bootstrapping procedure, with narrower CIs indicating greater accuracy. The stability of EI and bridge EI was estimated by calculating the correlation stability coefficient (CS-coefficient), preferably above 0.50 and acceptable above 0.25 [39]. In addition, the bootstrapped difference test was performed within the network to assess whether edge weights and EI differed significantly. All bootstrapping analyses were conducted by the R package bootnet [39] with 5000 bootstrap samples. Network comparison Using the Network Comparison Test (NCT) from the R package NetworkComparisonTest [40], several paired network invariance tests were conducted to assess changes over time: (1) edge-weight values, (2) centrality indices, (3) structural invariance (i.e., the largest individual differences in edge-weights), (4) global strength (i.e., the weighted absolute sum of all edges, representing the overall connectivity in the network). Only if the structural invariance test is significant, post-hoc tests of the differences in edge-weight values are performed [40]. Networks were compared using 5,000 random permutations. Given the exploratory nature of these analyses, no corrections for multiple comparisons were applied. Directed acyclic graph DAG networks were implemented using the R package blearn [41] with the hill-climbing algorithm to explore potential symptom causality. Network structures were stabilized through a two-step procedure. First, we generated 10,000 nonparametric bootstrapped samples and calculated the probability of each edge appearing across these networks. Following Scutari and Nagarajan [42], edges exceeding a data-driven significance threshold in the averaged network were retained. Thresholds were 50.7%, 50.4%, and 50.5% for the T1, T2, and longitudinal change networks, respectively. Compared to fixed thresholds, this method is considered more sensitive while preserving specificity and accuracy [42]. After determining the network structure, the direction of each surviving edge was ascertained. Edges maintaining the same direction in more than 50% of bootstrapped networks were plotted. This 50% threshold has been used in prior research [43] and could maximize potential connection representation but may retain edges with minimal directional evidence (approximately 51% certainty). Therefore, we reported the directional probability estimates and conducted analyses using a more conservative 60% threshold to mitigate the risk of overstating directionality. DAG estimation was performed using the Bayesian Information Criterion for Gaussian networks (BIC-G). To mitigate the risk of local optima, a restart strategy was employed. Specifically, the search was automatically restarted 5 times, and in each restart, the network structure was slightly perturbed 10 times to ensure a thorough exploration of alternative structures. Cross-lagged panel network The CLPN builds upon GGMs by incorporating directed edges to represent temporal cross-lagged effects between symptoms [25]. These effects quantify how one symptom at T1 predicts another symptom at T2, controlling for all concurrent symptoms at T1. Model estimation employed penalized maximum likelihood with a LASSO penalty on regression coefficients [44]. This approach mitigates overfitting and enhances result generalizability by shrinking trivial paths to zero [45]. Centrality indices of CLPN included cross-lagged out-expected influence (OEI) and in-expected influence (IEI), representing the sums of outgoing and incoming edge weights, respectively. Higher OEI indicates that a symptom at an earlier time point strongly influences other nodes later, whereas higher IEI indicates greater influence on a symptom from other nodes earlier [25]. The CLPN was estimated using the glmnet package [44]. As in GGMs, network accuracy and stability were assessed using 1000 nonparametric bootstrapped samples. Random seeds were set for all bootstrap procedures to ensure reproducibility. Results Attrition analysis Compared to participants who only completed the questionnaire at the time of the outbreak (Ta, n = 1099), those who completed both waves (Tb) were younger (Ma = 30.37, Mb = 27.75), had a higher level of education (Ta: 48.7% had a bachelor’s degree or above, Tb: 70.4%), and had a lower proportion of being in a relationship (Ta: 67.2%, Tb: 53.7%; Bonferroni-corrected ps < 0.001). No significant differences were found in gender, anxiety, or depression symptom severity (Bonferroni-corrected ps > 0.233). Differences in symptom severity Table 1 presents each item’s specific content, abbreviations, sample means, and standard deviations. Significant decreases in scores of anxiety (ΔM = 0.50, d = 0.141) and depression (ΔM = 0.54, d = 0.141) were found across time (Bonferroni-corrected ps < 0.002). No differences were found in the proportion who scored above the cut-off (Bonferroni-corrected ps > 0.160). Table 1Items, item abbreviations, means, and standard deviations for depression and anxiety symptomsSymptomsItem contentItem abbreviationM (SD) or N (%)At the outbreakAfter the peakAnxietyA1. I feel tense or ‘wound up’. Feeling Tense 0.88 (0.62)0.82 (0.60)A2. I get a sort of frightened feeling as if something awful is about to happen. Feeling Dread 1.05 (0.83)0.91 (0.84)A3. Worrying thoughts go through my mind. Worrying Thoughts 0.88 (0.77)0.89 (0.82)A4. I can sit at ease and feel relaxed. * Difficulty Relaxing 0.99 (0.80)0.94 (0.78)A5. I get a sort of frightened feeling like ‘butterflies’ in the stomach. Butterflies in Stomach 0.60 (0.84)0.51 (0.79)A6. I feel restless as I have to be on the move. Feeling Restless 1.05 (0.86)0.96 (0.83)A7. I get sudden feelings of panic. Sudden Panic 0.83 (0.78)0.76 (0.80)DepressionD1. I still enjoy the things I used to enjoy. * Lost Enjoyment 0.64 (0.78)0.63 (0.79)D2. I can laugh and see the funny side of things. * Difficulty Laughing 0.63 (0.83)0.55 (0.80)D3. I feel cheerful. * Feeling Down 0.75 (0.77)0.66 (0.75)D4. I feel as if I am slowed down. a Sad Mood 0.86 (0.78)0.82 (0.70)D5. I have lost interest in my appearance. Lost Appearance Interest 1.14 (1.05)0.91 (0.99)D6. I look forward with enjoyment to things. * Lacking Anticipation 0.59 (0.80)0.56 (0.78)D7. I can enjoy a good book or radio or TV program. * No Media Enjoyment 0.74 (0.92)0.67 (0.83)Total scoreAnxiety6.29 (3.55)5.79 (3.54)Depression5.35 (3.89)4.81 (3.74)PrevalenceAnxiety score ≥ 8194 (36.4)171 (32.1)Depression score ≥ 8161 (30.2)145 (27.2)Both anxiety and depression scores ≥ 8122 (22.9)99 (18.6)All items were rated on a Likert-type scale from 0 to 3, with higher scores indicating more severe symptoms. a The Chinese translation of this item more accurately conveys the meaning “I feel like my mood is gradually declining.” Therefore, it is abbreviated as “Sad mood”.*Items are reversely coded Items, item abbreviations, means, and standard deviations for depression and anxiety symptoms All items were rated on a Likert-type scale from 0 to 3, with higher scores indicating more severe symptoms. a The Chinese translation of this item more accurately conveys the meaning “I feel like my mood is gradually declining.” Therefore, it is abbreviated as “Sad mood”. *Items are reversely coded Community detection Although the walk-trap algorithm identified two clusters, symptoms of anxiety and depression were distributed in both clusters, partially supporting the notion that anxiety and depression are intertwined. Similar results were found with the other four algorithms and with longitudinal change scores (Table S1). Gaussian graphical models The estimated GGMs are presented in Fig. 1a and b. Consistent with the community analysis, 28.7% of nonzero edges during the outbreak and 29.6% during the post-peak stage directly connected depression and anxiety symptoms, indicating substantial interconnections between the two syndromes. Among all edges, the connection between “D4: Sad Mood” and “A7: Sudden Panic” was one of the highest, exceeding 79.7% and 89.9% in the outbreak and post-peak stages, respectively (Figures S1-2). There were no significant differences in structure (M = 0.148, p = .390) and global strength (S = 0.04, p = .716) across time. The current estimation of edge weight appeared to be reliable at the outbreak and post-peak (Figures S3-4). Gaussian graphical models at the outbreak (a) and post-peak (b), and the cross-lagged panel network from the outbreak to post-peak (c). Blue edges represent positive associations, and red edges represent negative associations. The thickness of the edges represents the strength of associations. The pink and blue nodes represent anxiety and depression items separately, as determined by the Hospital Anxiety and Depression Scale The centrality of each node is shown in Fig. 2(a). The centermost nodes across time were “D4: Sad Mood”, “D3: Feeling Down”, “A7: Sudden Panic”, and “D6: Lacking Anticipation”. Although A7 and D6 had relatively high centrality, bootstrap comparisons showed that their centrality was only higher than about half of the other nodes (Figures S5-6). The estimation of EI indices was stable at both time points with a CS-coefficient of 0.67 [39]. The NCT results indicated no significant temporal differences in the centrality indices of all symptoms (ps > 0.126). Additionally, the EI indices at the outbreak and after the peak were significantly correlated (r = .971, p < .001), suggesting the overall temporal stability of local connectivity within the network. No significant correlations were found between the EI indices of each symptom with its variance and skewness (ps > 0.552), ruling out the influence of symptom severity variability on centrality [46]. Standardized values of node expected influence (a) and bridge expected influence (b) in the gaussian graphical models at the outbreak and after the peak. The red and blue nodes represent the standardized values at the outbreak and after the peak, respectively. High values of expected influence indicate strong connections to other symptoms, and high values of bridge expected influence suggest close links to symptoms that belong to the other cluster The bridge EI indices of each symptom are displayed in Fig. 2(b) and are significantly correlated between the outbreak and post-peak (r = .667, p = .009). “D4: Sad Mood” emerged as the symptom with the highest bridge EI indices across time (Figures S7-8). The GGM using the longitudinal symptom change scores is shown in Figure S9. Its estimation of edge weights (Figure S10) and centrality index (CS-coefficients = 0.051 for both EI and bridge EI) was unstable [39]. Despite this instability, the network replicated the main findings from the cross-sectional ones. Specifically, the correlations of edge weights, EI, and bridge EI between the longitudinal change network and cross-sectional networks were all moderate to strong (r from 0.554 to 0.827, ps < 0.04). Additionally, “D4: Sad Mood” and “A7: Sudden Panic” remain among the nodes with the highest EI and “D4: Sad Mood” with the highest bridge EI (Figure S11). Directed acyclic graphs The directed acyclic graphs are presented in Fig. 3, in which the thickness of an edge signifies the importance of a directed association. The probability of each directed edge occurring in the bootstrapped samples is illustrated in Figure S12, in which the thickness of an edge indicates the confidence that its direction occurred in the bootstrapped samples. The directed acyclic graphs of anxiety-depression symptoms at the outbreak (a) and after the peak (b). Edge thickness indicates an edge’s importance to the model most closely approximating the data Several features were apparent. First, depression and anxiety symptoms were intertwined in DAG over time, with 47.2% and 40.0% of non-zero edges between disorders during the outbreak and post-peak, respectively. Second, among the depression-anxiety edges, 41.2% (outbreak) and 35.7% (post-peak) had directional probabilities below 0.6. When considering all edges, these proportions were 41.7% and 34.3%, respectively. Third, the predictive and potentially causal priority of “A7: Sudden Panic” was notable, as it appeared at the top of the DAG both during the outbreak and after the peak. A list of items connected depression and anxiety symptoms, among which “D4: Sad Mood” lay near the top of the DAG and was activated by “A7: Sudden Panic”. Anxiety-related physical symptoms (i.e., “A5: Butterflies in Stomach”) and anhedonia-related depression symptoms (i.e., “D1: Lost Enjoyment”, “D5: Lost Appearance Interest”, and “D7: No Media Enjoyment”) were downstream symptoms seemingly dependent on other symptoms. Over time, psychomotor symptoms (i.e., “A6: Feeling Restless” and “A4: Difficulty Relaxing”) appeared midstream or downstream. The longitudinal DAG model (Figure S13) replicated the main findings of the cross-sectional DAG, with the exception that “D4: Sad Mood” led to “A7: Sudden Panic”. Further examination revealed that in the longitudinal DAG, the probability of “D4: Sad Mood” leading to “A7: Sudden Panic” was 0.508. In the cross-sectional DAG, the probability of “A7: Sudden Panic” leading to “D4: Sad Mood” was slightly higher, being 0.525 at the outbreak and 0.546 after the peak. To avoid overstating directionality, we applied a 60% threshold for directional certainty (Figure S14), which replicated the main findings except for the disappearance of the panic-sadness edge and a reduced upstream position of panic. Cross-lagged panel network To further investigate the longitudinal association between anxiety and depression, a CLPN analysis was performed. Following convergence of the regularization procedure and removal of autoregressive paths, 106 cross-lagged paths remained (Fig. 1c; for the network including autoregressive paths, see Figure S15). Of these, 33 edges extended from depressive to anxiety symptoms (total weight = 1.413), and 18 extended from anxiety to depressive symptoms (total weight = 0.764). The strongest cross-disorder edge was from “D4: Sad Mood” to “A7: Sudden Panic” (β = 0.173), whereas the reverse edge was negligible (β = 0.003) (see Figures S16-17). The edge-weight estimates were acceptable (CS = 0.360) but appeared less reliable for IEI and OEI (both CS = 0.206) [39]. Additionally, no symptom nodes showed significantly higher IEI or OEI based on bootstrapped difference tests (Figures S18-20); therefore, these results are not discussed further. Sensitivity analysis The GGMs and DAGs were re-estimated among dropouts who participated only during the outbreak (Figures S21-22). There were no significant differences between dropouts and completers in the network structure, global strength, and EI centrality (ps > 0.087). According to the DAG of the dropouts, the physical and anhedonia-related symptoms still lay downstream. Psychomotor symptoms lay midstream or downstream. Despite a moderate decline in priorities, “A7: Sudden Panic” and “D4: Sad Mood” continued to hold crucial positions, located in the upstream or midstream. Additionally, “A7: Sudden Panic” led to “D4: Sad Mood” symptoms with a probability of 0.669. Discussion Our study is the first to apply both undirected (GGM) and directed network models (DAG and CLPN) to examine the longitudinal coexistence of anxiety and depression. Leveraging data from the COVID-19 pandemic, we comprehensively examined network evolution, spanning from the initial outbreak to the post-peak stage. Both undirected and directed models highlight sad moods (D4) as a key bridge symptom between anxiety and depression, particularly via its connection to panic feelings (A7). Additionally, symptoms of depression and anxiety were found to be highly intertwined, and the overall network structure and global strength remained stable over time. These findings unveil the intricate mechanisms underpinning anxiety-depression coexistence during disasters, offering valuable insights for developing preventive strategies. Our CLPN model revealed bidirectional reinforcement between depressive and anxiety symptoms, with depressive symptoms exerting stronger effects on anxiety over time. These longitudinal dynamics do not refute the General Adaptation Syndrome (GAS) framework [19], which posits that in response to a threat, individuals initially experience anxiety and may eventually fall into depression after prolonged ineffective coping [19]. Instead, our results add to the GAS framework by suggesting that, once symptoms manifest during sustained crises like COVID-19, depression and anxiety may engage in bidirectional reinforcement [47], with depressive symptoms potentially exerting a stronger influence on anxiety during chronic phases. Further examination of anxiety-depression connectivity identifies the panic-sadness dyad as a critical bridge initiating and maintaining comorbid symptoms. Cross-sectional DAGs revealed near-equidirectional probabilities (~ 50%) between panic and sadness, suggesting momentary reciprocity; however, CLPN analysis revealed temporal asymmetry, with baseline sad mood more strongly predicting subsequent panic than the reverse. According to Jaak Panksepp’s Affective Neuroscience theory [48], the panic-sadness system is an evolutionarily conserved neural circuit, spanning from the dorsal periaqueductal gray to the anterior cingulate, and is particularly responsive to separation or loss of attachment figures [48, 49]. Within this framework, panic is conceptualized not merely as a fear response, but as a reaction to the endogenous loss of felt security [48]. During COVID-19, heightened risks of bereavement and social isolation were likely co-elicited sadness and panic, activating the shared neural system [50, 51] and thereby facilitating mutual symptom reinforcement [52]. China’s stringent containment policies, though effective in curbing viral spread, may have inadvertently intensified separation distress and thus exacerbated panic and sadness [53]. This impact may be especially pronounced in collectivist contexts, where individuals’ identities and sense of security are closely tied to family and community integration. Over time, prolonged sadness may further erode coping capacity through behavioral withdrawal, increasing vulnerability to panic feelings. This mechanism may account for depression’s dominant longitudinal influence and aligns with meta-analytic findings where depressive disorders more strongly predicted specific phobias (closely linked to panic) than the reverse [47]. Some studies found psychomotor symptoms as the central and bridge nodes within the anxiety-depression network [14, 16]. However, it is noteworthy that these studies primarily relied on undirected network models, where high centrality does not necessarily represent a crucial driving role in network activation [54]. In contrast, in our DAG model, psychomotor symptoms were found midstream or downstream, subject to the influence of other symptoms. Differences in measurement tools could potentially account for variations in outcomes. Previous research frequently employed the Generalized Anxiety Disorder-7 (GAD-7) [55] to evaluate anxiety symptoms [16, 56]. It’s noteworthy that the GAD-7 specifically measures symptoms associated with generalized anxiety disorder [55]. Conversely, the HADS used in this study focuses more on the broader emotional aspects of anxiety [57], potentially making it more appropriate for assessing anxiety manifestations (e.g., panic feelings) during COVID-19. As this study is the sole analysis of the directed relationship between anxiety and depression symptoms, additional research using directed network analysis, along with experiments and interventions, is needed to substantiate these initial indications. Anxiety-related somatization symptoms (i.e., “A5: Butterflies in Stomach”) and anhedonia-related symptoms (i.e., “D1: Lost Enjoyment”, “D5: Lost Appearance Interest”, and “D7: No Media Enjoyment”) appear downstream in DAG network and exhibited relatively low EI indices both at the outbreak and post-peak stage. While somatization symptoms are common in non-Western cultures [58] and anhedonia is a core diagnostic criterion for depression [59], our findings suggest that these symptoms are more likely to be influenced by others, and caution is warranted when considering these symptoms as intervention targets. In support, an undirected network analysis of 3464 depressed outpatients found that DSM symptoms (e.g., sad moods) were not more central than non-DSM symptoms (e.g., sympathetic arousal) in indicating depression [60]. These findings may prompt a reconsideration of symptoms traditionally regarded as core, as they are not necessarily the most central within the symptom network. Notably, the application of centrality indices in psychopathology networks remains debated [10, 38]. Our results should be interpreted as preliminary indications of structural relationships, warranting further validation through relative importance modeling and longitudinal research. Additionally, this study observed a highly interconnected anxiety-depression network during the pandemic, with no distinct clustering of anxiety and depressive symptoms over time. This finding contrasts with Zavlis et al.'s research on the UK population, where anxiety and depression symptoms initially clustered separately but merged later during the lockdown [23]. Previous studies examining anxiety and depression comorbidity showed mixed results with different samples, e.g., non-distinct clustering in community children, adolescents [15], and adults [61], but clear clustering in clinical or subclinical samples [62–64]. This suggests anxiety and depression may be intertwined in people with minimal or mild symptoms, but not necessarily in clinical or subclinical samples. Moreover, cultural factors are important to consider since most clinical studies have focused on Western populations. Research indicates that people with mental health conditions in developing areas, particularly China, are less able to differentiate between anxiety and depression than those in developed areas, a difference attributed to cultural rather than socioeconomic factors [65]. Cultural differences have also been observed in community samples [66]. Low emotion differentiation shaped by culture may obscure distinct anxiety and depression clusters in symptom networks. Consequently, future research on the anxiety-depression network should incorporate sociocultural perspectives and consider disease-related factors, such as symptom severity. We found that overall connectivity between anxiety and depression symptoms remained stable during the early stages of COVID-19 (as indicated by non-significant global strength changes), despite decreasing symptom severity over time. Similar patterns have emerged in patients undergoing treatment for eating disorders [67]. According to the dynamic network model of psychological resilience, resilience processes weaken inter-symptom connections, thereby reducing network reactivity to stress and lowering the risk of transition into maladaptive or clinical states [68]. Thus, the stable network connectivity in our study suggests that psychological resilience may not have increased. Decreases in symptom severity likely reflect temporary adaptation to the acute phase of the pandemic, while latent vulnerability to future reactivation of anxiety and depression persists [67]. These findings emphasize the need for interventions that go beyond alleviating surface-level symptoms. Effective approaches should directly target maladaptive symptom connections while strengthening resilience factors to mitigate underlying vulnerability. Nevertheless, the interpretation of global strength requires caution, as it can be influenced by symptom variance [10] and remains controversial in psychological networks [38]. Furthermore, these group-level findings may not necessarily apply to individual-level processes. This study is strengthened by a longitudinal design and the utilization of both undirected and directed network approaches, allowing us to characterize the dynamic processes underlying anxiety-depression coexistence. Our findings revealed panic-sadness as a central bridging connection in this coexistence. Research indicates that social media and misinformation significantly fueled pandemic-related panic [69, 70]. Therefore, disseminating timely and credible information, as well as enhancing public media literacy, might be relevant strategies for mitigating population-level distress. Group-level interventions targeting sadness—such as mindfulness practices [71]—may also support emotional regulation and prevent symptom worsening. Notably, while these findings advance understanding of symptom interactions and inform hypotheses for prevention/intervention development, they do not constitute individual-level treatment targets or predictive models [72]. Future research employing high-frequency longitudinal data to conduct intra-individual directed network analyses, alongside pilot intervention studies, could further elucidate causal pathways and bridge these group-level insights toward personalized clinical applications. The results must be weighed in light of limitations. The relatively high attrition warrants caution, despite sensitivity analysis showing that dropout did not alter primary findings. Participants lost to follow-up were older, less educated, and more likely to be single. As lower education and unmarried status are linked to exacerbated pandemic distress [73, 74], their underrepresentation may have attenuated symptom interactions prominent in high-risk groups. Conversely, the overrepresentation of older adults, a demographic showing attenuated emotional reactivity during COVID-19 [74], might accentuate resilience pathways. Future studies should stratify networks by age and education cohorts to identify subgroup-specific critical pathways and enhance the generalizability of findings. Our longitudinal assessment was limited to two time points due to pandemic-related resource constraints, which precluded the use of intensive longitudinal methods to model dynamic symptom trajectories. Notably, although DAG and CLPN analyses suggested directional symptom relationships, neither establishes causality; experimental and intervention studies remain essential for causal inference. Additionally, DAG inherently precludes feedback loops [26], and its hill-climbing algorithm remains vulnerable to local optima despite restart mitigation [75]. Further, all three analytical methods (GGMs, DAGs, CLPNs) are susceptible to unmeasured confounders, potentially biasing network estimates [26, 38]. Although the Hospital Anxiety and Depression Scale is validated and widely used in community samples [76], it omits key symptoms of anxiety and depression (e.g., hopelessness, irritability). These unmeasured symptoms could exhibit high centrality in a more comprehensive network (e.g., irritability; [62]). Additionally, both symptoms beyond core diagnostic criteria and non-symptom factors have been shown to play important roles in understanding psychopathology networks [9]. Future research should therefore include a broader range of symptoms and relevant non-symptom variables (e.g., social factors) to more accurately characterize the connections between anxiety and depression symptoms. Moreover, the HADS employs a 4-point scale. We treated these responses as continuous variables in our EBIC-estimated GGMs, which is a common practice in network psychometrics [16, 62]. While this approach reduces bias in edge estimates for smaller samples like ours (n = 533), it sacrifices sensitivity [77], requiring cautious interpretation. Additionally, the findings should be interpreted in light of the contextual factors and fluctuations in the pandemic. This study was conducted at the outset of COVID-19 when most people in China were not yet infected, and the results may not apply to severe outbreaks or subsequent waves of COVID-19. Nevertheless, our findings provide insights into similar macro-level disasters, including potential future pandemics [78]. Previous research has shown that the network structures of emotional problems are similar across various disasters [79–81]. Yet, the replicability of network connectivity, as well as central and bridging symptoms, should be considered in the context of disaster type, severity, and controllability [80, 81]. Preliminary evidence suggests stronger overall connectivity in symptom networks following more severe and less controllable disasters [81]. Another limitation is the reliance on a community sample, which may limit the generalizability of the findings. According to the network theory, problems are more likely to trigger other complications within a strongly connected network compared to a weakly connected one. Hence, it is expected that compared to the current community sample with mild symptoms, clinical patients with comorbid anxiety and depression may exhibit stronger connections both across and within disorders. This is supported by Wigman et al. [82], who found that individuals with depression had a more strongly connected network of emotional states than those with psychosis and healthy controls. Similarly, Dobson et al. [63] and Park and Kim [83] found that individuals diagnosed solely with depression or anxiety display stronger within-disorder connections, with distinct communities for each condition. Together, these findings suggest that symptom-network structures may differ meaningfully between community and clinical samples. To better elucidate these differences, future studies should investigate directed anxiety-depression networks in clinical populations. Conclusions In summary, this study conducted longitudinal analyses of both undirected and directed anxiety-depression networks in the early stages of COVID-19. The network structure and connectivity remained stable over time. Two critical components were identified across time: panic feelings, serving as the central driving force, and sad moods, assuming a critical bridging role in connecting anxiety and depression. These findings underscore the potential of panic feelings and sad moods as promising intervention targets for preventing anxiety and depression in future public health emergencies. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1: Table S1. Community analysis. Figure S1. Bootstrapped significance test between edge weights for the gaussian graphical model at the COVID-19 outbreak. Figure S2. Bootstrapped significance test between edge weights for the gaussian graphical model after the COVID-19 peak. Figure S3. Bootstrapped edge weights and confidence intervals for the gaussian graphical model at the COVID-19 outbreak. Figure S4. Bootstrapped edge weights and confidence intervals for the gaussian graphical model after the COVID-19 peak. Figure S5. Bootstrapped significance test between expected influence (EI) index of each symptom for the gaussian graphical model at the COVID-19 outbreak. Figure S6. Bootstrapped significance test between expected influence (EI) index of each symptom for the gaussian graphical model after the COVID-19 peak. Figure S7. Bootstrapped significance test between bridge expected influence (EI) of each symptom for the gaussian graphical model at the COVID-19 outbreak. Figure S8. Bootstrapped significance test between bridge expected influence (EI) of each symptom for the gaussian graphical model after the COVID-19 peak. Figure S9. The gaussian graphical model of the longitudinal change scores. Figure S10. Bootstrapped edge weights and confidence intervals for the gaussian graphical model of the longitudinal change scores. Figure S11. Standardized values of node expected influence (EI) and bridge EI in the gaussian graphical model of the longitudinal change scores. Figure S12. The directed acyclic graphs of anxiety-depression comorbidity at the outbreak (a) and after the peak (b). Figure S13. The directed acyclic graphs of the longitudinal change scores between outbreak and post-peak stage. Figure S14. The directed acyclic graphs using a 60% directional probability threshold at outbreak (a) and post-peak stage (b). Figure S15. The cross-lagged panel network with autoregressive paths. Figure S16. Bootstrapped edge weights and confidence intervals for the cross-lagged panel network. Figure S17. Bootstrapped significance test between edge weights for the cross-lagged panel network. Figure S18. Standardized values of node in- and out-expected influence in the cross-lagged panel network. Figure S19. Bootstrapped significance test between in-expected influence (IEI) for the cross-lagged panel network. Figure S20. Bootstrapped significance test between out-expected influence (OEI) for the cross-lagged panel network. Figure S21. The gaussian graphical model between anxiety and depression symptoms among participants who dropped out at the post-peak stage. Figure S22. The directed acyclic graphs of anxiety-depression comorbidity among participants who dropped out at the post-peak stage. Acknowledgements The authors would like to thank all participants. Abbreviations Bayesian Information Criterion for Gaussian Networks Confidence Intervals Cross-Lagged Panel Network Correlation Stability Directed Acyclic Graph Expected Influence Generalized Anxiety Disorder-7 General Adaptation Syndrome Gaussian Graphical Model Hospital Anxiety and Depression Scale In-Expected Influence Least Absolute Shrinkage and Selection Operator Network Comparison Test Out-Expected Influence Posttraumatic Stress Disorder Author contributions Jiamin Bao conducted data collection, analysis, and interpretation, as well as drafting and revising the manuscript. Yuxuan Wang contributed to the data interpretation and revision of the manuscript. Yixuan Wang and Qiang Chen contributed to the review and revision of the article. Yinyin Zang performed the study design, the review, revision, and supervision of the article. All authors read and approved the final manuscript. Funding This work was supported by the Young Scientists Fund of the National Natural Science Foundation of China (Grant No. 32000776) and the New Sunshine Charity Foundation in China. Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate Ethical approval for this study was provided by the Ethics Committee of the School of Psychological and Cognitive Sciences, Peking University, conducted by the Declaration of Helsinki. Electronic informed consent was obtained from participants prior to data collection. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1: Table S1. Community analysis. Figure S1. Bootstrapped significance test between edge weights for the gaussian graphical model at the COVID-19 outbreak. Figure S2. Bootstrapped significance test between edge weights for the gaussian graphical model after the COVID-19 peak. Figure S3. Bootstrapped edge weights and confidence intervals for the gaussian graphical model at the COVID-19 outbreak. Figure S4. Bootstrapped edge weights and confidence intervals for the gaussian graphical model after the COVID-19 peak. Figure S5. Bootstrapped significance test between expected influence (EI) index of each symptom for the gaussian graphical model at the COVID-19 outbreak. Figure S6. Bootstrapped significance test between expected influence (EI) index of each symptom for the gaussian graphical model after the COVID-19 peak. Figure S7. Bootstrapped significance test between bridge expected influence (EI) of each symptom for the gaussian graphical model at the COVID-19 outbreak. Figure S8. Bootstrapped significance test between bridge expected influence (EI) of each symptom for the gaussian graphical model after the COVID-19 peak. Figure S9. The gaussian graphical model of the longitudinal change scores. Figure S10. Bootstrapped edge weights and confidence intervals for the gaussian graphical model of the longitudinal change scores. Figure S11. Standardized values of node expected influence (EI) and bridge EI in the gaussian graphical model of the longitudinal change scores. Figure S12. The directed acyclic graphs of anxiety-depression comorbidity at the outbreak (a) and after the peak (b). Figure S13. The directed acyclic graphs of the longitudinal change scores between outbreak and post-peak stage. Figure S14. The directed acyclic graphs using a 60% directional probability threshold at outbreak (a) and post-peak stage (b). Figure S15. The cross-lagged panel network with autoregressive paths. Figure S16. Bootstrapped edge weights and confidence intervals for the cross-lagged panel network. Figure S17. Bootstrapped significance test between edge weights for the cross-lagged panel network. Figure S18. Standardized values of node in- and out-expected influence in the cross-lagged panel network. Figure S19. Bootstrapped significance test between in-expected influence (IEI) for the cross-lagged panel network. Figure S20. Bootstrapped significance test between out-expected influence (OEI) for the cross-lagged panel network. Figure S21. The gaussian graphical model between anxiety and depression symptoms among participants who dropped out at the post-peak stage. Figure S22. The directed acyclic graphs of anxiety-depression comorbidity among participants who dropped out at the post-peak stage. Data Availability Statement The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.