Trajectories of covarying depression and multiple physical health problems in middle-aged and older adults based on GBMTM

Abstract

Depression in middle-aged and older adults often co-occurs with multiple physical health problems, yet its long-term co-evolution patterns remain unclear. Using data from 4 waves of the China Health and Retirement Longitudinal Study from 2011 to 2018, group-based multi-trajectory modeling was employed to identify covarying trajectories of depressive symptoms, number of chronic diseases, chronic pain, and functional disability. Multinomial logistic regression was used to analyze predictive factors. A total of 8961 participants aged 45 and above were included. Four distinct trajectories were identified: stable low-risk group (40.84%), low-risk improving group (11.24%), high-risk deteriorating group (23.96%), and high-risk fluctuating group (23.96%). Older age, female gender, and rural residence were risk factors for unfavorable trajectories, while higher education level, nonagricultural household registration, and social participation were protective factors. The long-term evolution of physical and mental health in middle-aged and older adults exhibits heterogeneous trajectories, with high-risk subgroups showing synchronous deterioration across multiple health dimensions. Identifying key predictive factors can inform stratified interventions and precision health management.

1. Introduction

Depression in middle-aged and older adults is a major global public health challenge. It not only significantly reduces individuals’ quality of life but is also closely associated with cognitive decline, increased suicide risk, and elevated all-cause mortality.^[1-3] Unlike pure depression, depression in middle-aged and older adults often co-occurs with multiple physical health problems, forming a complex network of mutual influences.^[4] Among these, chronic diseases, chronic pain, and functional disability are 3 common and interconnected core dimensions.^[5] Studies^[6,7] indicate that depression and the aforementioned physical health problems often exhibit bidirectional, time-varying complex associations, rather than simple unidirectional causal relationships. However, traditional research has mostly focused on cross-sectional associations or predictive relationships at a single time point, failing to fully reveal their long-term, dynamic co-evolution patterns.

Revealing the long-term covariation patterns of depression and multiple physical health problems and identifying their heterogeneous trajectories is crucial for understanding comorbidity mechanisms and implementing precision interventions. However, existing studies^[8-10] mostly employ growth curve modeling, which can only describe population average trends and may mask different developmental subgroups within the population. Furthermore, most studies^[11,12] focus only on the relationship between depression and a single physical health indicator or treat other physical problems as control variables, neglecting that chronic diseases, pain, and disability, as interwoven dimensions, may form different synergistic evolution patterns with depression.

Group-Based Trajectory Modeling^[13] can identify subgroups of individuals with similar longitudinal development patterns. In recent years, its multivariate extension, group-based multi-trajectory modeling (GBMTM), has provided a powerful tool for simultaneously examining the covarying trajectories of multiple health indicators. This method can not only identify overall pattern classes of individuals across multiple variables but also reveal the synchronous change relationships among trajectories of different health problems, thereby more comprehensively depicting the heterogeneous developmental landscape of “physical-mental comorbidity” in middle-aged and older adults.

Based on this, this study aims to investigate the covarying trajectories of depressive symptoms, number of chronic diseases, chronic pain, and functional disability in middle-aged and older adults. By analyzing data from a large nationally representative longitudinal cohort, we expect to: identify typical covarying trajectory classes of depression and multiple physical health problems in middle-aged and older adults and characterize the sociodemographic and health behavior profiles of different trajectory classes. The results of this study are expected to provide new empirical evidence for understanding the dynamic interaction of physical and mental health in middle-aged and older adults from a life course epidemiology perspective, and to inform the development of stratified, phased integrated intervention strategies for different risk subgroups.

2. Materials and methods

2.1. Study design and data source

This study is a prospective longitudinal cohort study analyzed using GBMTM. Data were obtained from China Health and Retirement Longitudinal Study (CHARLS),^[14] a nationally representative household survey of middle-aged and older adults aged 45 and above in China. The baseline sample covered approximately 19,000 respondents from 12,400 households. Ethical approval for the CHARLS database was granted by the Biomedical Ethics Review Committee of Peking University (Institutional Review Board approval number IRB00001052-11015). The analysis utilized 4 waves of survey data collected in 2011 (baseline), 2013, 2015, and 2018, constituting a longitudinal observation sequence spanning 7 years.

2.2. Study population

The study population consisted of respondents aged 45 years and older who completed assessments of core variables at baseline. Inclusion criteria were: complete data on depressive symptom assessment; complete data on chronic disease diagnoses; complete data on pain assessment; and complete data on activities of daily living (ADL). Exclusion criteria were: age < 45 years; missing data on any key variable at baseline; and loss to follow-up in any of the 4 waves. The final analytical sample comprised 8961 respondents. The screening process is shown in Figure 1.

Figure 1.

Research design and sample screening process. ADL = activities of daily living, CHARLS = China Health and Retirement Longitudinal Study.

2.3. Study variables

2.3.1. Depressive symptoms

CHARLS used the 10-item Center for Epidemiologic Studies Depression Scale. Each item was rated on a Likert scale ranging from “rarely or none of the time” (<1 day, score 0) to “most or all of the time” (5–7 days, score 3). The scale covers typical depressive symptoms such as feeling depressed, low energy, and poor concentration, along with 2 positively worded items reflecting positive affect, providing a comprehensive assessment of an individual’s psychological state. Total scores range from 0 to 30, with higher scores indicating more severe depressive symptoms. Participants with scores ≥ 10 were classified as having depressive symptoms, while those with scores < 10 were considered normal.^[15]

2.3.2. Chronic disease assessment

Based on self-reported physician diagnoses, the total number of chronic diseases reported by each respondent was calculated.

2.3.3. Functional disability

Assessed using the ADL scale, which includes both basic activities of daily living (BADL) and instrumental activities of daily living (IADL).^[16] BADL includes tasks such as dressing, bathing, eating, getting in/out of bed, using the toilet, and controlling urination and defecation. IADL includes tasks such as doing household chores, cooking, shopping, managing money, and taking medications. If a participant reported difficulty with any BADL or IADL task, they were classified as having a functional disability.^[17]

2.3.4. Pain assessment

The pain variable is derived from the assessment of the number of body pain locations reported in the CHARLS questionnaire. This questionnaire asks respondents in detail whether they have experienced pain in multiple specific body regions over the past year. In this study, the number of positive responses across all body regions is summed to generate a pain location count score. This count score is included as a continuous variable directly into the GBMTM to capture the dynamic evolution of pain burden among middle-aged and older adults over a 7-year follow-up period.

To identify baseline predictive factors for trajectory subgroup assignment, the selection of covariates in the multiple logistic regression models was conducted in accordance with the social determinants of health theoretical framework.^[18] This framework categorizes the drivers of health outcomes into 2 levels: Structural factors – including age, gender, education level, marital status, and place of residence – reflect an individual’s position within the social stratification structure and shape health trajectories over time through mechanisms such as resource access, life opportunities, and environmental exposure; and Intermediary factors – encompassing health behaviors such as smoking, alcohol consumption, sleep duration, physical activity, and social participation – which operate downstream of the structural factors and possess a certain degree of interventional potential.

2.4. Statistical analysis

For baseline characteristics, continuous variables were analyzed using an independent samples t-test, while categorical variables were analyzed using a χ² test. The core methodology employs the GBMTM^[19] approach to jointly model the trajectories of 4 variables: depression, number of chronic conditions, pain, and ADL. When constructing the joint trajectory model, the polynomial order for each target variable as a function of observation time (with the 2011 baseline denoted as T0 and subsequent visits treated as continuous time variables) must be predefined. For the 4 variables – depressive symptoms, number of chronic conditions, pain location count, and ADL disability – the study initially tested 2 temporal function forms: a linear form (containing a first-order term, i.e., the time slope) and a quadratic form (containing a second-order term, i.e., parabolic curvature). The selection of the polynomial order for the final model was guided by the following composite criterion: Parameter significance test – if the regression coefficient for the quadratic term is statistically significant at the P < .05 level, it indicates that the variable exhibits a significant acceleration or deceleration over time; in such cases, the quadratic term is retained; Information criterion optimization – compare the Bayesian Information Criterion (BIC) values for the model containing only the linear term versus the model containing the quadratic term; if the BIC value for the model with the quadratic term is lower (with an improvement ≥2), then the more complex model – including the quadratic term – is selected; Model simplicity and convergence – under the condition that the average posterior probability for each subgroup exceeds 0.7, a simpler model with fewer parameters is preferred. Based on the aforementioned criteria, this study ultimately determined that the trajectories of the 3 variables – depression, number of chronic conditions, and number of pain locations – should include both linear and quadratic terms to capture their nonlinear evolutionary patterns; however, since the quadratic term for the ADL disability indicator did not reach statistical significance across all subgroups and there was no improvement in the BIC value, only the linear term was included (i.e., only the intercept and linear time slope were estimated).

The exclusion criterion for this study was limited to “complete wave nonresponse” – that is, when a respondent did not participate in the full round of the survey for a given year, resulting in the absence of any observed values for all variables for that year. For cases where participants did participate in the survey but exhibited item-level missingness, such participants were not excluded from the analysis. The GBMTM employed the Expectation-Maximization algorithm for parameter estimation; under the missing at random assumption, this algorithm enables the calculation of the likelihood function using all observed individual-level data, thereby accommodating non-balanced panel data as well as partial missingness of certain variables within each survey wave.^[20] As a result, all 8961 participants ultimately included in the analysis had completed both the baseline survey and follow-up surveys across subsequent waves; any missing values for individual variables within a given survey wave were naturally handled by the Expectation-Maximization algorithm during its iterative computation, without requiring additional imputation or list deletion.

Model estimation was performed using the maximum likelihood method; the optimal number of trajectory subgroups was determined via grid search. The selection criterion was based on the Minimum BIC, with the additional requirement that the average posterior probability for each subgroup exceed 0.7 and the intragroup probability exceed 0.5, thereby ensuring both model fit adequacy and subgroup discrimination power.^[21] Following the identification of these trajectory subgroups, we employed a multiple logistic regression model – with the stable low-risk group as the reference – to investigate the association between various baseline characteristics and trajectory assignment, calculating the relative risk ratio (RRR) and its 95% confidence interval. To ensure the reliability of the results, a multicollinearity diagnostic was conducted. The adjusted generalized variance inflation factor values for all variables were <2, indicating the absence of significant multicollinearity. All statistical analyses were performed using R (version 4.5.0; R Foundation for Statistical Computing, Vienna, Austria) with RStudio (version 2025.05.1+513; Posit Software, PBC, Boston). The GBMTM analysis was implemented using the traj package, the multiple logistic regression was conducted using the nnet package, and the multicollinearity diagnostic was performed using the car package. The statistical significance level was set at a 2-sided P < .05.

3. Results

3.1. Types of covarying trajectories of depression and multiple physical health indicators

GBMTM analysis determined that the 4-group model was the optimal model (Table 1). The 4 trajectory groups were named: stable low-risk group (40.84%), low-risk improving group (11.24%), high-risk deteriorating group (23.96%), and high-risk fluctuating group (23.96%). The stable low-risk group maintained consistently low and relatively stable levels of depressive symptoms, chronic diseases, pain, and ADL limitations. The low-risk improving group showed significant improvement in depressive symptoms and pain, with slight increases in chronic diseases and ADL limitations, but remained at low levels. The high-risk deteriorating group exhibited depressive symptoms that worsened initially then improved, chronic diseases that increased continuously, ADL limitations that improved initially then worsened, and pain that markedly deteriorated. The high-risk fluctuating group showed U-shaped fluctuations in depressive symptoms and pain, continuously increasing chronic diseases, and slowly increasing ADL limitations. The 4 multi-trajectory groups are illustrated in Figure 2, with specific parameter estimates in Table 2. Classification quality metrics showed that all groups had average posterior probabilities greater than 0.96 and odds of correct classification >5.0, indicating high classification accuracy (Table 3). To assess potential selection bias due to loss to follow-up, we compared the baseline characteristics of participants retained in the analytical sample with those excluded from the analysis (Table S1, Supplemental Digital Content 1). No significant differences were observed between the 2 groups in depression scores, number of chronic conditions, or sleep duration; however, differences were found in age, ADL, pain, residence, self-rated health, and alcohol consumption. These findings suggest that although selective attrition may affect certain sociodemographic variables, the core health indicators directly relevant to our trajectory analysis did not exhibit systematic bias.

Table 1.

Comparison of GBMTM model fit indicators.

Group	AIC	BIC	SSABIC	Entropy	LMR_LRT
1	6,71,135.56	6,71,322.28	6,71,252.36	NA	–
2	5,94,015.03	5,94,396.95	5,94,253.94	0.96	77,157.92*
3	4,12,858.99	4,13,410.64	4,13,204.07	0.96	1,81,175.83*
4	3,93,918.42	3,94,656.78	3,94,380.29	0.96	18,982.45 *
5	3,98,856.06	3,99,798.11	3,99,445.35	0.94	−4889.10

Optimal values are shown in bold.

Adjusted LMR-LRT = adjusted Lo–Mendell–Rubin likelihood ratio test, AIC = Akaike Information Criterion, BIC = Bayesian Information Criterion, GBMTM = group-based multi- trajectory modeling, SSABIC = sample size-adjusted Bayesian Information Criterion.

^*P < 0.001.

Figure 2.

Identified covarying trajectories of depression and multimorbidity. ADL = activities of daily living.

Table 2.

Trajectory parameter estimation of the 4-category model.

Variable	Parameter	Stable low-risk	Low-risk improving	High-risk deteriorating	High-risk fluctuating
Depression	Intercept	6.71**	13.82**	6.78**	10.28**
	Linear	−0.94**	−1.33**	−1.07**	−1.57**
	Quadratic	−0.21**	0.30**	0.25**	0.36**
Chronic	Intercept	0.78**	2.25**	0.64**	0.96**
	Linear	0.16**	1.29**	0.18**	0.77**
	Quadratic	0.02**	0.09**	0.02*	0.06**
ADL	Intercept	0.00	2.11**	0.33**	0.37**
	Linear	0.00	−0.35**	0.08**	0.10**
Pain	Intercept	1.55**	5.11**	0.10**	2.98**
	Linear	−1.25**	−2.38**	−0.07**	−2.15**
	Quadratic	0.32**	0.70**	0.02**	0.57**

Although the intercept for pain is relatively low, its linear and quadratic coefficients indicate a significant accelerating worsening of pain over time.

ADL = activities of daily living.

^*P < .05.

^**P < .001.

Table 3.

Classification quality metrics for each mode.

Model	Category	Sample proportion	AvePP	OCC
traj_1	Class1	1	1	–
traj_2	Class1	43.32%	0.99	96.20
	Class2	56.68%	0.99	86.77
traj_3	Class1	36.22%	0.99	1016.16
	Class2	33.84%	0.97	38.49
	Class3	29.94%	0.97	82.14
traj_4	Class1	40.84%	0.99	6781.95
	Class2	11.24%	0.98	325.95
	Class3	23.96%	0.96	86.05
	Class4	23.96%	0.96	74.44
traj_5	Class1	35.91%	0.99	714.80
	Class2	18.23%	0.96	79.93
	Class3	20.30%	0.94	64.07
	Class4	9.80%	0.94	154.04
	Class5	15.76%	0.94	50.28

All categories have an AvePP > 0.9 and OCC > 5, indicating high classification accuracy.

AvePP = average posterior probability, OCC = odds of correct classification.

3.2. Baseline characteristics of each trajectory group

After identifying the 4 heterogeneous covariation trajectory classes, we further compared the distribution of baseline sociodemographic and health behavior characteristics across trajectory groups (Table 4). The 4 trajectory classes showed significant heterogeneity in baseline characteristics including demographics, socioeconomic status, and health behaviors (all P-values < .05). Compared to the stable low-risk group, the low-risk improving group (60.18 ± 8.68 years) and high-risk deteriorating group (60.13 ± 8.21 years) were significantly older (P < .001); the proportion of females was significantly higher in the high-risk deteriorating group (67.86%) and high-risk fluctuating group (63.44%); the proportion with higher education was highest in the stable low-risk group (2.05%) and lowest in the high-risk fluctuating group (0.33%); the proportion residing in urban communities was highest in the stable low-risk group (43.14%) and lowest in the high-risk deteriorating group (27.81%); the stable low-risk group had the highest proportions of smoking (42.65%) and high-frequency drinking (15.98%) and the longest sleep duration (6.69 ± 1.60 hours); while the high-risk deteriorating group had the highest proportion rating their health as “poor” (45.13%).

Table 4.

Participant baseline characteristics by trajectories.

Variable	Total (N = 8961)	Stable low-risk (N = 3660)	Low-risk improving (N = 1007)	High-risk deteriorating (N = 2147)	High-risk fluctuating (N = 2147)	P-value
Age	57.58 ± 8.19	55.15 ± 7.54	60.18 ± 8.68	60.13 ± 8.21	57.95 ± 7.80	<.001
Gender						<.001
Male	4044 (45.13%)	2047 (55.93%)	522 (51.84%)	690 (32.14%)	785 (36.56%)
Female	4917 (54.87%)	1613 (44.07%)	485 (48.16%)	1457 (67.86%)	1362 (63.44%)
Education status						<.001
Less than lower secondary	8006 (89.34%)	3023 (82.60%)	921 (91.46%)	2038 (94.92%)	2024 (94.27%)
Upper secondary and vocational training	848 (9.46%)	562 (15.36%)	71 (7.05%)	99 (4.61%)	116 (5.40%)
Tertiary	107 (1.19%	75 (2.05%)	15 (1.49%)	10 (0.47%)	7 (0.33%)
Marital status						<.001
Married	8055 (89.89%)	3426 (93.61%)	884 (87.79%)	1819 (84.72%)	1926 (89.71%)
Divorced/widowed	850 (9.49%)	216 (5.90%)	119 (11.82%)	306 (14.25%)	209 (9.73%)
Never married	56 (0.62%)	18 (0.49%)	4 (0.40%)	22 (1.02%)	12 (0.56%)
Household registration type						<.001
Agricultual hukou	7337 (81.88%)	2751 (75.16%)	831 (82.52%)	1865 (86.87%)	1890 (88.03%)
Nonagricultural hukou	1563 (17.44%)	880 (24.04%)	163 (16.19%)	272 (12.67%)	248 (11.55%)
Unified residence hukou	60 (0.67%)	29 (0.79%)	12 (1.19%)	10 (0.47%)	9 (0.42%)
Do not have hukou	1 (0.01%)	0 (0.00%)	1 (0.10%)	0 (0.00%)	0 (0.00%)
Residential area						<.001
Urban Community	3133 (34.96%)	1579 (43.14%)	334 (33.17%)	597 (27.81%)	623 (29.02%)
Rural village	5828 (65.04%)	2081 (56.86%)	673 (66.83%)	1550 (72.19%)	1524 (70.98%)
Smoking						<.001
No	5614 (62.65%)	2099 (57.35%)	577 (57.30%)	1498 (69.77%)	1440 (67.07%
Yes	3347 (37.35%)	1561 (42.65%)	430 (42.70%)	649 (30.23%)	707 (32.93%)
Drinking						<.001
Nondrinker	6300 (70.30%)	2315 (63.25%)	692 (68.72%)	1713 (79.79%)	1580 (73.59%)
Very low-frequency drinker (≤3 d/mo)	1106 (12.34%)	542 (14.81%)	120 (11.92%)	208 (9.69%)	236 (10.99%)
Moderate-frequency drinker (1–3 d/wk)	410 (4.58%)	218 (5.96%)	47 (4.67%)	62 (2.89%)	83 (3.87%)
High-frequency drinker (≥4 d/wk)	1145 (12.78%)	585 (15.98%)	148 (14.70%)	164 (7.64%)	248 (11.55%)

Variable	Total (N = 8961)	Stable low-risk (N = 3660)	Low-risk improving (N = 1007)	High-risk deteriorating (N = 2147)	High-Risk Fluctuating (N = 2147)	P-value
Sleep hours	6.36 ± 1.86	6.69 ± 1.60	6.70 ± 1.70	5.72 ± 2.11	6.28 ± 1.88	<.001
Living alone						<.001
No	8572 (95.66%)	3547 (96.91%)	952 (94.54%)	2013 (93.76%)	2060 (95.95%)
Yes	389 (4.34%)	113 (3.09%)	55 (5.46%)	134 (6.24%)	87 (4.05%)
Social activities						<.001
No	4739 (52.88%)	1708 (46.67%)	544 (54.02%)	1283 (59.76%)	1204 (56.08%)
Yes	4222 (47.12%)	1952 (53.33%)	463 (45.98%)	864 (40.24%)	943 (43.92%)
Number of residents	3.66 ± 1.79	3.63 ± 1.68	3.61 ± 1.80	3.63 ± 1.89	3.76 ± 1.87	.026
Self-report of health						<.001
Very good	543 (6.06%)	347 (9.48%)	85 (8.44%)	32 (1.49%)	79 (3.68%)
Good	1545 (17.24%)	886 (24.21%)	227 (22.54%)	118 (5.50%)	314 (14.63%)
Fair	4491 (50.12%)	1979 (54.07%)	571 (56.70%)	789 (36.75%)	1152 (53.66%)
Poor	2036 (22.72%)	403 (11.01%)	118 (11.72%)	969 (45.13%)	546 (25.43%)
Very poor	346 (3.86%)	45 (1.23%)	6 (0.60%)	239 (11.13%)	56 (2.61%)

3.3. Multinomial logistic regression results

Multiple logistic regression analyses conducted with the stable low-risk group as the reference group demonstrated that this model exhibited strong discriminative power for different trajectory groups (Nagelkerke R² = 0.32; McFadden R² = 0.13). Advanced age, female gender, and rural residence were common risk factors for being assigned to the nonideal trajectory group. Higher education level, nonagricultural household registration status, and participation in social activities were significant protective factors. Shorter sleep duration was associated with an increased risk of being assigned to the high-risk deterioration group or the fluctuation group. Self-rated poor health status was the strongest predictor in the high-risk deterioration group. The specific regression coefficients and RRR are presented in Table 5.

Table 5.

Results of multinomial logistic regression.

Variable	Reference	Low-risk improving		High-risk deteriorating		High-risk fluctuating
		RRR (95% CI)	P-value	RRR (95% CI)	P-value	RRR (95% CI)	P-value
Age	Continuous (per 1unit increase)	1.08 (1.07–1.09)	<.001	1.09 (1.08–1.10)	<.001	1.05 (1.05–1.06)	<.001
Gender	Male	1.44 (1.16–1.78)	<.001	2.96 (2.45–3.58)	<.001	2.63 (2.21–3.13)	<.001
Education status	Junior high school or below	–		–		–
	Senior high/vocational school	0.68 (0.52–0.89)	.006	0.64 (0.50–0.83)	<.001	0.58 (0.46–0.72)
	College or above	0.87 (0.47–1.59)	.643	0.50 (0.24–1.06)	.071	0.32 (0.14–0.72)	.006
Marital status	Married	–		–		–
	Divorced/widowed	1.26 (0.94–1.68)	.122	1.41 (1.10–1.80)	.006	1.14 (0.89–1.45)	.295
	Never married	0.77 (0.25–2.38)	.646	2.01 (0.96–4.19)	.062	1.15 (0.52–2.54)	.725
Household registration type	Agricultual hukou	–		–		–
	Non-agriculture	0.69 (0.55–0.87)	.002	0.67 (0.54–0.82)	<.001	0.59 (0.49–0.71)	<.001
	Unified residential	1.64 (0.81–3.32)	.173	0.64 (0.27–1.51)	.311	0.60 (0.27–1.32)	.202
	No registration	Unstable (n < 5)	<.001	0.04 (0.04–0.04)	<.001	0.05 (0.05–0.05)	<.001
Residential area	Rural village	1.35 (1.13–1.60)	<.001	1.55 (1.33–1.82)	<.001	1.45 (1.26–1.66)	<.001
Smoking	No	1.18 (0.97–1.45)	.102	1.13 (0.94–1.36)	.19	1.19 (1.00–1.40)	.049
Drinking	Nondrinker	–		–		–
	≤3 d/mo	0.92 (0.73–1.15)	.456	0.98 (0.80–1.20)	.829	1.01 (0.84–1.21)	.954
	–6 d/mo	0.96 (0.68–1.37)	.833	0.88 (0.63–1.24)	.464	1.04 (0.78–1.39)	.798
	≥7 d/mo	0.92 (0.73–1.15)	.453	0.79 (0.63–1.00)	.045	1.08 (0.89–1.31)	.436
Sleep hours		1.02 (0.98–1.07)	.29	0.83 (0.80–0.86)	<.001	0.92 (0.89–0.94)	<.001
Living alone	No	1.03 (0.67–1.57)	.907	1.20 (0.83–1.75)	.328	1.11 (0.77–1.60)	.575
Social activities	No	0.78 (0.68–0.91)	<.001	0.67 (0.59–0.76)	<.001	0.76 (0.68–0.85)	<.001
Number of residents	Per 1 unit increase	1.04 (0.99–1.10)	.11	1.04 (1.00–1.08)	.035	1.05 (1.02–1.09)	.004
Self-report of health	Per 1 grade decrease	1.02 (0.93–1.11)	.679	3.64 (3.34–3.96)	<.001	1.65 (1.54–1.77)	<.001

Since the sample size for the “no household registration” category is extremely small (n < 5), the estimate is unstable and will not be interpreted.

Model fitting statistic: Nagelkerke R² = 0.32, McFadden R² = 0.13.

CI = confidence interval, RRR = relative risk ratio.

3.4. Multicollinearity diagnostic results

Diagnostic results showed that the adjusted Generalized Variance Inflation Factor values for all independent variables ranged from 1.01 to 1.55 (Table 6), well below the severe collinearity threshold of 10, indicating that the regression results are reliable.

Table 6.

Multicollinearity diagnosis.

Variable	GVIF	df	Adjusted GVIF
Age	1.14	1	1.07
Gender	2.41	1	1.55
Education status	1.1	2	1.02
Marital status	1.44	2	1.1
Household registration type	1.41	2	1.09
Residential area	1.31	1	1.14
Smoking	2.15	1	1.47
Drinking	1.41	3	1.06
Sleep hours	1.01	1	1.01
Living alone	1.49	1	1.22
Social activities	1.02	1	1.01
Number of residents	1.13	1	1.06
Self-report of health	1.02	1	1.01

df = degree of freedom, GVIF = generalized variance inflation factor.

4. Discussion

This study utilized GBMTM to identify 4 heterogeneous co-evolutionary trajectories of depression and somatic health (chronic diseases, pain, and ADL) among middle-aged and older Chinese adults: the stable low-risk group (40.84%), low-risk improvement group (11.24%), high-risk deterioration group (23.96%), and high-risk fluctuation group (23.96%). These 4 trajectory types differed significantly in baseline levels, direction of change, and fluctuation patterns, confirming the diversity of “psychosomatic comorbidity” developmental pathways in middle-aged and older populations^[22,23] and providing a population stratification basis for precision intervention.

4.1. The stable low-risk group (40.84%) and the high-risk deterioration group (23.96%) constitute the 2 poles of health differentiation

The stable low-risk group maintained consistently low levels of depression, chronic disease burden, number of pain sites, and ADL disability over the 7-year follow-up. This group was characterized by younger age, male predominance, higher education, and urban residence, consistent with the social determinants of health framework,^[24] suggesting that early-life advantages translate into long-term health capital through resource accumulation.^[25-27] With this group accounting for over 40% of the sample, healthy aging appears to have a substantial population-level basis. Clinically, this group does not require intensive intervention, though their health behaviors may serve as reference targets for public health education.

The high-risk deterioration group exhibited the most unfavorable trajectory: depressive symptoms first increased then decreased, chronic diseases accumulated progressively, pain sites increased, and ADL first improved then declined. This sequential pattern suggests a potential cascading relationship across dimensions.^[22,23] A plausible mechanism is that the early sharp rise in depression alters central processing of somatic signals through neuroendocrine pathways, increasing multisite pain burden while undermining chronic disease management adherence, ultimately accelerating functional decline.^[10,28] Importantly, the pain measure in this study reflects multisite pain burden (number of pain sites) rather than single-site pain intensity; multisite pain itself implies broader central sensitization and a more complex pathophysiological substrate. This group concentrated multiple disadvantages – older age, female sex, low education, rural residence, and poor self-rated health – empirically validating cumulative disadvantage theory, wherein social adversities accumulate layer by layer over the life course and eventually erupt as systemic health crises.^[28] Previous studies have identified similar risk patterns in Chinese older adults and linked a “cardiovascular-digestive” multimorbidity pattern to depression risk.^[29,30] This group should be prioritized for intervention, requiring integrated management protocols co-developed by geriatrics, psychiatry, pain medicine, and rehabilitation teams, with simultaneous monitoring of depression, pain site burden, chronic disease load, and ADL. Isolated single-domain interventions are likely to fail due to drag effects from other dimensions.

4.2. The contrast between the high-risk fluctuation group (23.96%) and the low-risk improvement group (11.24%) carries important dynamic implications

The fluctuation group exhibited a U-shaped synchronous covariation between depression and number of pain sites – both declining to a nadir at mid-follow-up and rising thereafter. While cross-sectional studies have only been able to report static associations between depression and pain,^[4] the present longitudinal design reveals a dynamic coupling pattern within a specific subgroup. The underlying mechanism may involve a “depression-pain” reciprocal loop: multisite pain exacerbates negative affect, while depression lowers pain tolerance, creating a mutually reinforcing cycle.^[6,7,31] A previous neuroimaging study^[32] demonstrated that depressive symptoms significantly modulate amygdala–medial prefrontal cortex functional connectivity in chronic pain patients, providing a neurobiological explanation for this covariation. Unlike the sustained decline in the deterioration group, the fluctuation group’s spiral exhibits intermittent features,^[33] which constitutes an intervention window: when both indicators reach the U-shaped trough, stress and inflammatory levels may remain within a modifiable range, and combined cognitive-behavioral therapy, pain management, and psychosocial support may interrupt the cycle and prevent drift toward the deterioration group.^[34] Synthesized evidence on inflammatory, behavioral, and psychosocial pathways in depression-somatic disease comorbidity has provided multidimensional theoretical support for this strategy.^[35] For this subgroup, we recommend establishing dynamic monitoring mechanisms to actively identify trough periods and implement short-term intensive combined interventions.

4.3. Predictors of trajectory membership further reveal the role of structural forces in shaping health differentiation

Each additional year of age increased the risk of belonging to any nonideal trajectory by approximately 5% to 9% (RRR = 1.05–1.09); yet, this effect size was smaller than that of poor self-rated health (RRR = 3.64). The predictive strength of self-rated health has been validated across multiple studies.^[2,36] In primary screening, the single item “How would you rate your overall health?” achieves highly efficient triage at minimal cost. Sex differences were particularly pronounced: women’s RRRs for the deterioration and fluctuation groups were 2.96 and 2.63, respectively^[1,37] – substantially larger than previously reported sex differences in depression prevalence – suggesting that women bear a “combined burden” of affective symptoms, multisite pain, and functional impairment. A previous study^[38] similarly found that older age, disability, short sleep, and poor self-rated health were significantly associated with adverse trajectories. Another study^[39] further supported women’s elevated risk in comorbid conditions. We recommend incorporating pain site screening alongside the Patient Health Questionnaire-2 as parallel mandatory screening items in annual checkups for middle-aged and older women. Educational and urban-rural disparities were also significant: upper secondary/vocational education and nonagricultural hukou were protective (RRR = 0.58–0.69), while rural residence increased risk (RRR = 1.35–1.55), yielding an approximate 2-fold urban-rural risk differential.^[40,41] This effect size exceeds that of most individual behavioral factors (e.g., short sleep, social inactivity), suggesting that narrowing the urban-rural gap in public health services may yield higher marginal returns.^[42] Previous studies^[29,38] also confirmed associations of short sleep and poor self-rated health with adverse trajectories, while the protective role of social participation^[30] was replicated in our study.

In summary, this study integrated Dannefer’s cumulative disadvantage theory and social determinants perspectives within a GBMTM framework, testing their applicability in a middle-aged and older Chinese population. Unlike studies focusing solely on the depression-chronic disease binary relationship, the present study incorporated number of pain sites and ADL into multidimensional joint analysis, revealing that health dimensions evolve not along a gradient of severity but as qualitatively distinct patterns.^[23,35] The 4 trajectories point to distinct practical implications: the stable group as a reference, the deterioration group indicating integrated intervention, the fluctuation group signaling an intervention window, and the improvement group challenging the assumption that “aging is irreversible.” Methodologically, GBMTM synchronous modeling outperforms univariate trajectory approaches and better aligns with the holistic nature of health; model fit was acceptable (entropy = 0.96, average posterior probability > 0.96), and the nationally representative sample enhanced robustness.^[14]

Several limitations should be acknowledged. First, approximately 33% of baseline participants were excluded due to loss to follow-up, and significant differences existed between retained and excluded participants in age, ADL, pain, residence, self-rated health, and alcohol consumption, potentially affecting generalizability. Second, all indicators were self-reported; although the number of pain sites is more objective than pain intensity, recall bias remains a concern. Third, GBMTM classification is group-relative and should not be directly extrapolated to individual-level clinical decisions. Fourth, the observational design precludes causal inference, and trajectory covariation may be influenced by unmeasured confounders (e.g., genetics, early-life adversity). Fifth, findings are specific to the Chinese context and require cross-national validation. Future studies should optimize follow-up strategies and employ more rigorous missing data handling and causal inference methods.

5. Conclusion

Based on 7-year longitudinal data, this study systematically identified 4 heterogeneous covarying trajectories of depression and multiple physical health problems in middle-aged and older adults and revealed their key predictive factors. The study found that sociodemographic factors including age, gender, education, hukou, and residence are key predictors shaping different long-term health trajectories. Particularly important, the U-shaped trajectory exhibited by the “High-Risk Fluctuating group” suggests that the health status of some populations possesses dual characteristics of vulnerability and reversibility. The findings have clear public health and clinical practice implications, suggesting a shift from “one-size-fits-all” intervention models toward precision health management based on trajectory subgroups. At the public health policy level, efforts should focus on reducing socioeconomic inequalities to promote health equity. At the clinical and community intervention level, an “integrated physical-mental” model should be adopted, implementing precision management: intensive intervention for the “High-Risk Deteriorating group,” early identification and stability maintenance for the “High-Risk Fluctuating group,” and extraction of replicable successful experiences from the “Low-Risk Improving group.” Therefore, future healthy aging strategies need to transition from population-homogenized interventions toward precision, dynamic management paradigms based on trajectory subgroups.

Author contributions

Conceptualization: Chunlian Wang, Yuan Wan, Yang He, Dan Jin, Jiali Huang, Jinglan Liu, Hongyan Deng.

Data curation: Chunlian Wang.

Formal analysis: Chunlian Wang, Yuan Wan, Yang He.

Methodology: Chunlian Wang, Yuan Wan, Yang He, Dan Jin, Jiali Huang.

Project administration: Jinglan Liu, Hongyan Deng.

Software: Chunlian Wang, Yuan Wan, Yang He.

Supervision: Jinglan Liu, Hongyan Deng.

Visualization: Chunlian Wang, Yuan Wan, Yang He.

Writing – original draft: Chunlian Wang, Yuan Wan.

Writing – review & editing: Jinglan Liu, Hongyan Deng.