Trajectories of depressive symptoms in middle-aged and older Chinese adults: identifying subgroups, core symptoms and predictors

Abstract

Background

Depressive symptoms among middle-aged and older adults are a significant public health concern, with varying symptom trajectories over time. Understanding these trajectories and their predictors can inform targeted interventions.

Objectives

To identify subgroups of depressive symptom trajectories, determine predictors of these subgroups, and explore the core symptoms and their predictive relationships.

Methods

This study analyzed 7,166 participants aged ≥ 45 years from the China Health and Retirement Longitudinal Study across four waves (2011, 2013, 2015, 2018). Depressive symptoms were assessed using the 10-item Center for Epidemiologic Studies Depression Scale. Group-based trajectory modeling (GBTM) identified depressive symptom trajectories. Multivariate logistic regression explored influencing factors, while Cross-lagged panel network models (CLPN) were used to identify core symptoms.

Results

Three distinct trajectory groups were identified: “stable low” (66.4%), “decline followed by an increase” (27.8%), and “continuously rising” (5.8%). Females, those with lower education, poor self-reported health, unmarried status and rural residents were associated with worsening symptoms. CLPN analysis revealed “depressive mood” as the core symptom, with “feeling lonely” and “could not get going” predicting “depressive mood.”

Conclusion

This study identifies distinct trajectories of depressive symptoms in older adults and pinpoints “depressive mood” as a core symptom, which is dynamically predicted by loneliness and a lack of behavioral activation. Therefore, an effective public health strategy should involve not only identifying at-risk individuals based on their trajectory profiles but also targeting these specific precursor symptoms to prevent escalation.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12888-025-07144-0.

Introduction

Depression is a global public health issue affecting approximately 350 million people worldwide [1], and it is projected to become a leading cause of disease burden by 2030 [2]. It is associated with emotional distress, increased healthcare costs, elevated suicide risk, and higher mortality rates from various causes [3, 4]. Depressive symptoms are a precursor to clinical depression and are particularly prevalent among middle-aged and older adults, with a prevalence ranging from 13 to 41% [5–9]. These symptoms represent significant barriers to healthy aging in China [5, 10].

Given China’s rapidly aging population, there are increasing concerns about the mental health of older adults. Depressive symptoms can lead to functional impairment, reduced quality of life, and accelerated aging [11–13]. Understanding the progression of depressive symptoms is crucial for effective intervention and mitigating long-term impacts on mental and physical health. The trajectory model is a well-established approach for examining the progression of depressive symptoms, providing insights into the dynamic trends and characteristics of symptom changes over time [14]. Recognizing symptom trajectories enables timely intervention, which is vital for enhancing individual well-being and informing public health strategies to improve outcomes in an aging population.

Depressive symptoms exhibit significant heterogeneity in their manifestation and progression among middle-aged and older adults, influenced by several factors. Previous studies have identified between two and six distinct trajectories of depressive symptoms in middle-aged and older adults. For example, a national cohort study involving 19,110 older adults identified four distinct trajectories of depressive symptoms: consistently low, moderate, emerging, and persistently high [15]. Demographic factors (sex [16–19], age [16, 20], education level [17, 21–23], marital status [17], residence [17, 19]), physiological factors (i.e., physical function impairment [16, 19, 24–26], hearing and vision impairment [21] and chronic diseases [17, 22, 26]), social factors (social support [25–27], social participation [26] and social network [23]) and psychological factors (cognitive function [16, 22], stress events [23, 28] and self-reported health status [19]) are associated with heterogeneous trajectories of depressive symptoms in middle-aged and older adults. Exploring the trajectory of depressive symptoms can help identify heterogeneous subgroups and their characteristics. However, we lack sufficient understanding of how depressive symptoms interact within subgroups, which may hinder researchers and clinicians from implementing personalized interventions.

To better understand the interrelationships among depressive symptoms in different subgroup populations, temporal network analysis can help us to reveal the dynamic interaction between symptoms. Unlike traditional models that view depression as a common factor with interchangeable symptoms, network approaches define mental disorders as systems of interacting symptoms [29]. Network analysis focuses on the symptoms themselves, allowing for empirical examination of their centrality and interrelationships, which reveals their clinical relevance in intervention studies [30]. Temporal networks enhance this approach by integrating network analysis with longitudinal data, allowing for the exploration of temporal relationships and dynamic interactions among symptoms [31, 32]. By developing a temporal network, our study can identify which core symptoms and interaction mechanisms remain consistent across these subgroups over time, highlighting potential intervention targets. Previous studies using temporal networks identified ‘feeling fearful,’ ‘feeling depressed,’ and ‘everything being an effort’ as core symptoms in older adults [10, 33, 34]. However, it remains unclear whether the core symptoms and mechanisms of symptom interaction remain consistent across these heterogeneous subgroups over time. This gap constrains our understanding of depressive symptoms and limits the development of precise interventions and personalized treatments for different subgroups.

Therefore, this study aims to address these gaps to explore depressive symptom trajectories among middle-aged and elderly individuals. Specifically, we seek to answer three research questions: (1) How many heterogeneous trajectories of depressive symptoms exist among middle-aged and elderly people? (2) What are the characteristics of the potential subgroups of depressive symptom trajectories? (3) What are the core symptoms in the temporal networks of depressive symptoms of the subgroups and their predictive factors? By identifying distinct trajectory subgroups, their predictors, and core symptoms, we aim to provide insights that inform targeted strategies to improve mental health outcomes among the aging population.

Methods

Sampling and participants

This study utilized a prospective cohort from the CHARLS database, a nationally representative longitudinal survey of Chinese residents aged 45 and above [35]. Initial data collection began in 2011 (wave 1, baseline), with follow-up assessments in 2013 (wave 2), 2015 (wave 3), and 2018 (wave 4). CHARLS is a nationwide longitudinal survey covering 28 provinces (including municipalities and autonomous regions) across China, with broad geographic representation. It aims to collect data on the social, economic, health, and family status of middle-aged and elderly individuals aged 45 and above in China. Our analysis incorporated seven years of CHARLS data (waves 1–4), with a detailed research design illustrated in sFigure1. Participants with missing baseline data on depressive symptoms or sociodemographic information were excluded from the analysis. Since the CLPN model cannot analyze data containing missing values, we included only participants with complete data across all study variables (complete case analysis). The final analysis included 7,166 participants.

Measures

Depressive symptoms

Depressive symptoms among study participants were assessed using the 10-item Center for Epidemiologic Studies Depression Scale (CESD-10), a previously validated tool designed to measure depressive symptoms in older adults [36]. The CESD-10 consists of 10 questions, each with four response options, scored from 0 to 3, resulting in a total score range of 0 to 30 where higher scores indicate more severe depressive symptoms. Regarding the reliability of the CESD-10 in our sample, we calculated Cronbach’s alpha to assess the consistency and internal structure stability, which yielded values of 0.600, 0.624, 0.643, and 0.684 for Waves 1 through 4, respectively.

Sociodemographic characteristics

The sociodemographic characteristics included age (continuous), sex (male = 1; female = 2), educational attainment (less than lower secondary = 1; upper secondary and vocational training = 2; tertiary = 3), marital status (married = 1; others = 2), and place of residence (rural = 1; urban = 2). Additionally, self-reported health status was rated on a scale up to 5, with higher scores indicating better health.

Statistical analysis

All statistical analyses were conducted using R version 4.4.0 (R Foundation for Statistical Computing, Vienna, Austria). Descriptive statistics such as frequencies, percentages, means, and standard deviations were utilized to characterize the demographic features and severity of symptoms.

The GBTM model [37] identifies subgroups of individuals with distinct depressive symptom trajectory patterns over time. The likelihood calculated by GBTM is referred to as the posterior probability of group membership. Participants were classified into trajectory subgroups based on depressive symptom scores, corresponding to the group in which they have the highest probability of membership. The determination of the optimal number of latent groups is influenced by several factors According to the model selection criteria proposed by Nagin: (1) optimal model fit is linked to reductions in Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC); (2) a key criterion is that the average posterior probability (AvePP) for each identified group must be 70% or higher; (3) each identified category must represent at least 5% of the total sample [27, 28]. Finally, the clinical interpretability of the model is a crucial factor.

Following trajectory group identification, we employed multinomial logistic regression to assess the relationship between sociodemographic factors and depressive symptom trajectories. Based on previous literature supporting the influence of these variables on depressive symptoms, all sociodemographic variables were simultaneously entered as predictors rather than using variable selection methods (e.g., stepwise or forward selection). Prior to analysis, multicollinearity was assessed using variance inflation factors (VIF), with all values below 2, indicating acceptable levels of multicollinearity among predictors. We calculated odds ratios (OR) and 95% confidence intervals (CI) for each predictor variable. Results with P < 0.05 were considered statistically significant.

To explore the temporal effects of subgroups, we employed the Cross-Lagged Panel Network (CLPN) model, an ideal framework for our research question. This model facilitates the analysis of relationships among symptoms in panel data as directed paths over time, with discrete measurement occasions [34]. In comparison to cross-sectional networks, the directed paths in CLPN denote the flow of symptoms from one measurement occasion to the subsequent one [38]. These paths represent the co-variation of a symptom at time t with another symptom (either the same or different) at time t + 1, while adjusting for all other symptoms at time t. We calculated networks for consecutive time points, resulting in three network models (i.e., T1 → T2, T2 → T3, etc.). This approach allows us to compare the predictiveness of symptoms across different networks. Given the binary response format of the CESD-10, logistic regression models were used to compute autoregressive and cross-lagged coefficients. In the autoregressive pathways, a symptom at one time point predicts itself at the next time point, adjusting for all other symptoms at the first time point. In the cross-lagged paths, a symptom at one time point predicts a different symptom at the next time point, thus adjusting for all other symptoms at the first time point [34]. To estimate the regression coefficients, penalized maximum likelihood with a lasso penalty was used, applying 10-fold cross-validation to adjust parameters and set small regression coefficients to zero. We estimated the CLPN regression using the glmnet package [39].

Node centrality

We calculated in-expected influence and out-expected influence as the primary centrality indices. Network centrality indices (including in-expected influence and out-expected influence) were calculated using the qgraph package. In-expected influence quantifies the extent to which each symptom is predicted by other symptoms in the network (sum of the values of incoming edges associated with the symptom). Out-expected influence measures the degree to which each symptom predicts other symptoms in the network (sum of the values of outgoing edges associated with the symptom), identifying the most central symptoms from the perspective of symptom mechanisms [34]. In network theory, nodes with high centrality, such as those with numerous or strong connections, are regarded as core symptoms [40]. Symptom activation refers to the impact of specific symptoms on the variation of remaining symptoms at subsequent time points, thereby reflecting the potential influence of nodes on the overall network dynamics [41].

Accuracy and stability of the network

The accuracy of edge weights was estimated by calculating the 95% CI for each edge weight through 1000 iterations of non-parametric bootstrapping. To assess the stability of the results, we utilized case-drop bootstrapping, correlating centrality indices of the entire sample with those of a subset of the sample, conducted through 1000 iterations using the bootnet package. The correlation stability coefficient should optimally exceed 0.5, although a minimum threshold of 0.25 is also considered acceptable [31]. Additionally, we applied edge weight difference tests and centrality difference tests to determine if significant disparities among exist between edge weights and among centrality indices. The former quantifies whether specific cross-lagged symptom connections (i.e., edges) are more significant than others. Similarly, the latter quantifies whether certain symptoms hold more importance (i.e., centrality) within the network than others [34].

Results

Demographics

Of the 7,166 participants, 50.85% were women, with an average age of 56.81 ± 7.83 years. Demographic and clinical characteristics of the study population are summarized in Table 1. The mean scores for the CESD-10 total scores are presented in Table 2.

Table 1.

Baseline characteristics of CHARLS participants n (%) or (mean ± s.d.)

Characteristics	Total (n = 7166)
Sex
Male	3522(49.15%)
Female	3644 (50.85%)
Educational attainment
Less than lower secondary	6267 (87.5)
Upper secondary & vocational training	806 (11.2)
Tertiary	93 (1.3)
Age	56.81 ± 7.83
Marital status
Married	6277 (87.6)
Partnered	289 (4.0)
Separated	20 (0.3)
Divorced	47 (0.7)
Widowed	491 (6.9)
Never married	42 (0.6)
Residence
Urban	2621 (36.6)
Rural	4545 (63.4)
Self-reported health status
Very good	490 (6.8)
Good	1315 (18.4)
Fair	3640 (50.8)
Poor	1490 (20.8)
Very Poor	231 (3.2)

Table 2.

The mean scores for the CESD 10 (mean ± s.d.)

Abbreviation	Items	T1	T2	T3	T4
depresl	Felt depressed	1.93 ± 1.05	1.71 ± 0.98	1.82 ± 1.06	1.92 ± 1.09
effortl	Everything an effort	1.95 ± 1.11	1.76 ± 1.07	1.83 ± 1.13	1.99 ± 1.16
sleeprl	Sleep was restless	2.01 ± 1.18	2.00 ± 1.18	2.01 ± 1.20	2.13 ± 1.21
whappyl	Happy	3.01 ± 1.09	2.66 ± 1.22	2.96 ± 1.19	2.96 ± 1.17
flonel	Felt lonely	1.48 ± 0.89	1.38 ± 0.82	1.47 ± 0.93	1.60 ± 1.01
botherl	Bothered by things	2.00 ± 1.09	1.74 ± 1.00	1.85 ± 1.10	1.90 ± 1.10
goingl	Could not get going	1.32 ± 0.74	1.25 ± 0.68	1.32 ± 0.79	1.44 ± 0.89
mindtsl	Had trouble keeping mind	1.88 ± 1.06	1.72 ± 1.01	1.83 ± 1.09	1.92 ± 1.09
fhopel	Felt hopeful	2.90 ± 1.18	2.49 ± 1.28	2.60 ± 1.29	2.65 ± 1.28
fearll	Felt fearful	1.33 ± 0.76	1.25 ± 0.68	1.29 ± 0.75	1.39 ± 0.83

Trajectory of depressive symptoms

Using GBTM, different depressive symptom trajectory groups were identified at various follow-up points. The AvePP for the three groups was relatively high (0.81, 0.87, and 0.92, respectively), demonstrating strong model differentiation. Figure 1 illustrates the three trajectory groups: “stable low” (Group 1, n = 4758), “decline followed by an increase” (Group 2, n = 1993), and “continuously rising” (Group 3, n = 415). The first trajectory group, comprising 66.4% of participants, consistently exhibited relatively low scores on the CESD-10 scale. Throughout the study, the second group, accounting for 27.8% of participants, initially showed moderate CESD-10 scores that declined from baseline (2011) to the second wave (2013), reaching their lowest point, and then demonstrated a subsequent increase in depressive symptoms from 2013 onwards through the final wave (2018), clearly exhibiting the ‘decline followed by an increase’ pattern. The third group, representing 5.8% of participants, had relatively very high CESD-10 scores, indicating the most severe depressive symptoms among the three groups.

Fig. 1.

Trajectories of depressive symptoms in middle-aged and older adults

Trajectory characteristics of depressive symptoms

Compared to Group 1, significant predictors of depressive symptoms were identified in Group 2 and Group 3, including sex, educational attainment, marital status, self-reported health, and residence. Being female, having lower educational attainment, poor self-reported health status, being unmarried, and residing in rural areas were consistently linked to higher odds of depressive symptoms (P < 0.05). Age was significantly associated with depressive symptoms in Group 2 (P < 0.05), but not in Group 3 (P = 0.872). Detailed results are presented in Table 3.

Table 3.

Results of multivariate logistic regression analysis for depressive symptoms across different groups

Variables	Group1 (n = 4758)	Group2 (n = 1993)		Group3 (n = 415)
		OR (95%CI)	P value	OR (95%CI)	P value
Intercept	ref	0.04(0.02,0.08)	< 0.001	< 0.01	< 0.001
Sex(female vs. male)	ref	2.13(1.90,2.38)	< 0.001	3.72(2.92,4.75)	< 0.001
Age	ref	0.99(0.98,1.00)	0.015	1.00(0.99,1.02)	0.872
Educational attainment (more than lower secondary vs. Less than lower secondary)	ref	0.70(0.59,0.83)	< 0.001	0.34(0.20,0.59)	< 0.001
Marital status(others vs. married	ref	1.06(1.02,1.09)	0.001	1.15(1.09,1.21)	< 0.001
Self-reported health status(others vs. good)	ref	1.77(1.66,1.90)	< 0.001	3.16(2.76,3.61)	< 0.001
Residence(rural vs. urban)	ref	1.45(1.28,1.63)	< 0.001	1.92(1.50,2.45)	< 0.001

Network of subgroups of depressive symptoms

Autoregressive and cross-lagged edges

Figure 2 illustrates nine CLPN models across all consecutive time points for three subgroups, depicting coefficients where one symptom predicts another (cross-lagged effects). When examining the influence of symptoms on themselves over time (autoregressive pathways), “depressive mood” exhibited the strongest autoregressive effect across all nine waves (r: 0.121 to 0.257, sTable1- sTable9).

Fig. 2.

Temporal network models of depressive symptoms for the three trajectory subgroups. Nodes represent the 10 individual symptoms of the CESD-10. Edges represent the cross-lagged predictive relationships from one time point to the next (e.g., from T1 to T2). Separate networks are presented for each subgroup

Among the four CLPN models, the path from “happy” to “felt lonely” demonstrated the strongest cross-lagged edge (r: 0.138 to 0.208). Additionally, the path from “felt lonely” to “depressive mood” showed a robust cross-lagged edge (r: 0.159 to 0.270). Similarly, the path from “could not get going” to “depressive mood” exhibited a strong cross-lagged edge (r: 0.159 to 0.270). Furthermore, across waves, these symptoms displayed significant cross-lagged edges in the opposite direction (“happy” → “depressive mood”; r: −0.081 to −0.200, sTable1-9).

Centrality

Figure 3 depicts the standardized centrality measures. Across all three groups, “depressive mood” has the highest out-expected influence (r = 1.104–2.034); the highest in-expected influence was not stable. “everything is an effort” (r = 0.517–1.230) and “had trouble keeping mind” (r = 0.589–1.153) have the highest in-expected influence in the major CLPN model.

Fig. 3.

The Centrality indices for 3 subgroups from T1 to T4 (Z-scores)

Accuracy, stability of the network

For the bootstrap subset (sFigure2-10), the correlation stability (CS) coefficient for expected influence ranged from 0.594 to 0.750. As these values are above the recommended threshold (e.g., 0.5), they indicate good stability of the centrality estimates. The results of the bootstrapped 95% confidence intervals for the edge weights are presented in sFigure11-19. The narrow bootstrap confidence intervals suggested that the network is very accurate.

Edge and node bootstrapped difference test

The results of the bootstrap node difference test were displayed in sFigure20-37. Notably, there was considerable variation in “depressive mood” compared to other nodes (DTs = 0.610–0.920). The bootstrap difference test for the edge weights showed that the two strongest edge weights, “happy” and “feel hopeful”, were significantly different from the other edge weights (sFigure38-46).

Discussion

Our study represents the novel approach in identifying the core symptoms of latent trajectory subgroups among China’s middle-aged and elderly population, three distinct subgroups of depressive symptom trajectories have been identified. Furthermore, being female, possessing a lower educational level, unmarried status, residing in rural areas, and self-reporting poor health significantly increase the likelihood of belonging to the group with persistently escalating depressive symptoms. Additionally, “depressive mood” emerges as the central symptom of depressive symptoms in China’s middle-aged and elderly individuals. From a network perspective, feelings of loneliness and an inability to initiate activities can longitudinally predict “depressive mood”.

The trajectory of depressive symptoms among China’s middle-aged and elderly populations encompasses several heterogeneous groups, with women, individuals with lower educational levels, the unmarried, rural residents, and those self-reporting poor health status facing higher risks of being categorized in the “continuously rising” group. Group 1 exhibits consistently low depressive symptoms scores, representing 66.4% of the population. Group 1’s trajectory initially declined and later returned to baseline levels, generally maintaining the lowest depressive symptoms scores. Moreover, findings suggest that residents with higher educational levels are more likely to belong to this subgroup, compared to those with lower education. Previous research [17, 21, 22] has linked lower educational levels to depressive symptoms, whereas higher educational attainment is associated with reduced depressive disorders. This study finds that those with higher education at baseline are more likely to be classified in the consistently low group. The second trajectory, which accounted for 27.8% of the population, was characterized by an initial decline in depressive symptoms followed by a subsequent increase, representing a mid-level symptom burden overall.

The trajectory of Group 3 is one of continuous elevation, comprising 5.8% of the sample and exhibiting the highest severity scores for depressive symptoms. Group 3 which showed continuously elevated depressive symptoms, was significantly associated with several established risk factors. Compared to Group 1, individuals in this high-risk group were more likely to be women, unmarried, rural residents, or to report poor self-health. These findings are consistent with a large body of literature that identifies these demographic and health characteristics as key predictors of depression [17, 19, 28]. The higher vulnerability among these groups is often attributed to a combination of factors, including the unique physiological and psychological stressors in women, social isolation in the widowed, structural disadvantages in healthcare and social support for rural populations, and the chronic stress associated with poor physical health [21, 42, 43]. Identifying individuals with these characteristics is therefore crucial for targeted prevention and intervention.

Our results indicate that in the three heterogeneous trajectory subgroups, “depressive mood” consistently exhibits the highest expected outgoing connections. In network theory, nodes with high centrality, such as those with numerous or strong connections are regarded as core symptoms within the entire network model [40]. This finding suggests that “depressive mood” is a core symptom activating the other symptoms in the depression network. However, the findings among adolescent and adult populations have been mixed. Some studies report that in longitudinal network analyses of adolescents, there are more outward connections associated with depressive feelings [44]. Conversely, studies on middle-aged and older adults find that inward connections related to “depressive mood” are more prevalent [10, 34]. These studies differ from our research in terms of age range (predominantly older participants), time lag between assessments (generally longer), and analytical methods (utilizing various statistical approaches). However, it is challenging to identify the primary factors contributing to these discrepancies across studies. Our study reveals that “depressive mood” appears to activate other symptoms and has inward connections, which is clinically significant as it is one of the two hallmark symptoms required for diagnosis [45]. However, according to the network analysis results, including those from this study, other symptoms and associated impairments may have already been present for some time when “depressive mood” manifests. Therefore, these precursor symptoms may represent a crucial focus for early intervention efforts.

Our study found that “feeling lonely” has a positive cross-lagged effect on “depressive mood” within the temporal network. This finding is consistent with other longitudinal studies, including network analyses, which show that loneliness can predict the onset of depressive mood, sometimes up to a year later [10, 34, 46]. The link between social connection and depression is evident across multiple levels. At the community level, indicators of social isolation are tied to higher depression risk, while on an interpersonal level, a reduction in social networks can impair the capacity for social engagement, playing a crucial role in the transition from loneliness to depressive symptoms [47]. Therefore, this finding underscores the potential of loneliness as a target for early intervention. Future research could conduct prospective cohort studies to explore how loneliness influences “depressive mood” through specific biological or psychological mechanisms.

Similarly, “could not get going” emerged as another key precursor to “depressive mood,” predicting its onset at a subsequent time point. “could not get going” is one of the core features of depression, reflecting a lack of behavioral activation that directly impacts the ability to improve mood [48]. The reduction in activity caused by depressive symptoms can create a vicious cycle: decreased activity → lowered mood → further reduced activity. This cycle makes it increasingly difficult for individuals to return to a normal state of living. Therefore, our findings highlight “could not get going” as a crucial target for early intervention aimed at improving behavioral activation. Future research could utilize prospective cohort designs to delve deeper into how “could not get going” influences the development of “depressive mood” through neurobiological or psychological mechanisms.

Clinical implications

Although network studies can inform central symptoms as intervention targets, significant associations between symptoms may lack clinical relevance for various reasons [34]. First, since associations have been identified in the general population, network studies must be conducted in clinical populations to determine whether these associations hold true. Second, even within clinical populations, it remains unknown whether interventions targeting these core elements are related to symptom improvement, let alone the alleviation of functional impairments. Theoretically, targeting core symptoms within the network should reduce overall network connectivity, yet this has yet to be tested. Future research should conduct multi-center, cross-population clinical studies to validate the centrality hypothesis of symptom networks. Third, there is no consensus on which effect sizes among symptoms in the network are considered clinically significant. Nevertheless, the core symptom of “depressive mood” that continues to emerge within our network over time may represent a potential key target for alleviating depression at the population level among older adults. Given the consistency of these findings across multiple time points and subgroups, it is particularly important as it suggests the existence of a shared process that may target different developmental stages and populations. Furthermore, we found that “depressive mood” is activated through feeling lonely and “could not get going.” Therefore, our research should inspire intervention studies to empirically test whether targeting these symptoms leads to an overall reduction in symptoms.

Strengths and limitations

Our longitudinal network analysis, conducted on a representative longitudinal sample of middle-aged and older adults in China, identified multiple heterogeneous populations and their temporal networks, making a significant contribution to the literature. Our study has several limitations. First, while our cross-lagged model identifies temporal precedence between symptoms, any causal inferences drawn from this observational data should be made with caution. Second, the two-year timeframe between assessments may obscure some short-term associations between symptoms. The level of symptom fluctuation over these two years remains unclear; during assessments, we could not distinguish whether these networks reflect cumulative changes over two years or random fluctuations. Third, the CESD-10’s limitation to ten symptoms, excluding important domains like somatic symptoms and sleep disturbances, may have biased our network inferences toward a simplified structure that underrepresents the full complexity of depressive symptom interactions. Future studies should utilize more comprehensive depression measures to capture the complete symptom spectrum and provide more accurate network representations in this population. Lastly, the unequal time intervals in CHARLS data collection (2-year vs. 3-year intervals) may affect the comparability and interpretation of lagged effects in the CLPN model [49]. This design constraint should be interpreted with appropriate caution when interpreting the temporal stability and comparative strength of our network findings across different time periods. To enhance the precision of causal inference, it is imperative for future research to employ equidistant time intervals.

Conclusion

In conclusion, this study reveals that depressive symptoms among middle-aged and older Chinese adults follow distinct, heterogeneous trajectories. Vulnerability to the most severe, continuously rising symptom trajectory is significantly shaped by sociodemographic and health factors. Furthermore, our network analysis identifies “depressive mood” as the central symptom, which is dynamically predicted by feelings of loneliness and a lack of behavioral activation. These findings provide a novel, integrated perspective, suggesting that effective public health strategies should not only identify high-risk populations but also target these specific, actionable precursor symptoms to prevent the escalation of depression.

Authors’ contributions

Jia Fang: Writing– original draft, Methodology, Conceptualization, Formal analysis. Wenwen Wu: Writing– review & editing, Validation. Chen Yang: Writing– review & editing. Huiyuan Li: Writing– review & editing. Wencan Cheng: Writing– review & editing.Ni Zhang: Writing– review & editing. Ye Zhang: Writing– review & editing. Baoyi Zhang: Writing– review & editing. Meifen Zhang: Funding acquisition, Writing– review & editing. All authors have read and approved the final manuscript.

Funding

This work was supported by the General Program of the National Natural Science Foundation of China (Grant number: 72374232).

Data availability

All relevant data are within the manuscript and its additional file. The data are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

All participants provided written informed consent prior to enrollment in the CHARLS study. The CHARLS project was approved by Peking University. For detailed information regarding the consent for participation and publication, please visit the official CHARLS website (https://charls.pku.edu.cn/). Our research was conducted under the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.