The association between subjective well-being and physical-psychological-cognitive multimorbidity: a longitudinal analysis of a nationally representative survey

Abstract

While subjective well-being (SWB) significantly influences individual physical and mental health, its impact on complex multimorbidity remains unclear. This study aimed to investigate the effect of SWB on physical-psychological-cognitive multimorbidity (PPC-MM). Using data from four waves (2011, 2013, 2015, and 2018) of the China Health and Retirement Longitudinal Study, we employed discrete-time survival models (pooled logistic regression with complementary log–log link) as the primary analysis to examine the association between SWB and incident PPC-MM, accounting for interval-censored data. Marginal structural models (MSMs) with inverse probability weighting were used to handle time-varying SWB and confounding. Competing risks were addressed via Fine-Gray models, with death as a competing event. Sensitivity analyses included physical-cognitive multimorbidity (PC-MM) only and continuous-time Cox models. Among 6843 participants (mean age 59 ± 9 years), 1585 (23.2%) reported being “satisfied” with life at baseline, 1088 (15.9%) reported “Not satisfied,” and the remainder reported “somewhat satisfied.” Significant differences were observed across SWB groups in age, sex, education, residence, marital status, smoking, alcohol consumption, sleep duration, body mass index, and waist circumference (all p < 0.05). In fully adjusted discrete-time models, polynomial contrasts revealed a strong linear trend associating lower SWB with increased PPC-MM risk (HR = 2.03, 95% CI 1.82–2.26, p < 0.001) and slight non-linearity (quadratic HR = 1.24, 95% CI 1.15–1.34, p < 0.001). Time-varying MSMs confirmed causal effects (linear HR = 1.95, 95% CI 1.73–2.19, p < 0.001; quadratic HR = 1.22, 95% CI 1.12–1.33, p < 0.001). Fine-Gray models yielded similar subdistribution HRs (linear sHR = 1.26, 95% CI 1.12–1.40, p < 0.001; quadratic sHR = 1.11, 95% CI 1.02–1.20, p < 0.05), with cumulative incidence functions diverging across groups (p < 0.001). Results remained robust across multiple sensitivity and subgroup analyses. SWB exerts a significant influence on the development of PPC-MM among middle-aged and older adults in China over time. These findings underscore the importance of incorporating SWB into lifecourse preventive strategies against PPC-MM.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-30444-0.

Introduction

Population ageing is accelerating worldwide, and China is no exception: in 2020, 264 million people were aged ≥ 60 years, accounting for 18.7% of the population¹. As people age, multiple chronic conditions often co-occur and increase care complexity, functional impairment, and healthcare use. In later life, somatic diseases frequently coexist with depression, anxiety, and cognitive decline, giving rise to physical-psychological-cognitive multimorbidity (PPC-MM). The prevalence of multimorbidity involving physical, psychological, and cognitive disorders varies widely across countries, ranging from about 8–34%², and is associated with higher healthcare utilisation and substantial economic burden, including polypharmacy, more frequent hospitalisation, and greater medical and out-of-pocket costs^3–5.

Subjective well-being (SWB)—an individual’s cognitive and affective evaluation of life, including life satisfaction and positive and negative emotions ⁵—has emerged as an important determinant of health. Higher SWB has been linked to better health outcomes, lower risks of chronic disease, and longer life expectancy in diverse populations^6,7. However, while prior studies have documented the prevalence, sociodemographic patterning, and economic burden of PPC-MM, the longitudinal relationship between SWB and PPC-MM remains poorly understood. Most existing work examines single physical or mental conditions, rather than multimorbidity across physical, psychological, and cognitive domains, and little is known about how changes in SWB relate to the development of PPC-MM in older adults.

This research gap not only limits our understanding of the comprehensive relationship between health and SWB in older adults but also hinders the development of integrated health management strategies aimed at enhancing SWB. Therefore, the objective of this study is to bridge this gap using cross-sectional and longitudinal data from an 8-year observation, with the goal of providing scientific evidence for policies supporting healthy aging.

Methods

Study design and participants

The China Health and Retirement Longitudinal Study (CHARLS) is an ongoing, nationally representative cohort of Chinese adults aged ≥ 45 years. The baseline survey in 2011 used multistage probability-proportional-to-size sampling and included about 17,000 individuals from 10,000 households across 450 villages in 150 counties and 28 provinces. Follow-up interviews are conducted every 2–3 years using computer-assisted personal interviewing, collecting detailed information on sociodemographic characteristics, health status, healthcare use and insurance, employment, income, expenditure, and assets, along with physical measurements and blood samples. Four waves of data (2011, 2013, 2015, and 2018) were used in the present analysis⁸. The CHARLS dataset is available for download on the CHARLS website at http://charls.pku.edu.cn/en. The CHARLS project was approved by the Biomedical Ethics Committee of Peking University, and all participants provided written informed consent.

Analytic sample and role of baseline triglycerides

For this analysis, data from Waves 1–4 (2011–2018) were used. Participants were included if they (1) were aged 45 years or older; (2) were free of PPC-MM at baseline; (3) provided information on SWB; and (4) had available follow-up data on physical, psychological, and cognitive conditions.

Participants were included in the analytic cohort if they had complete data on baseline SWB, baseline triglycerides, and all baseline covariates required for the fully adjusted models. Because triglycerides were included in a prespecified panel of cardiometabolic biomarkers used as potential confounders in our MSM/IPW analyses, we required baseline triglyceride measurements to estimate biomarker-adjusted models in a consistent sample and to avoid changes in the underlying cohort when biomarkers were added to the models.

As shown in the flowchart, this restriction led to the exclusion of 4189 participants with missing 2011 triglyceride data, yielding a final analytic sample of 6843 participants. (Fig. 1) To allow readers to assess potential selection related to missing triglycerides or missing SWB, a baseline comparison table was constructed in which participants were grouped into three categories: “Included in analytic sample”, “Excluded: missing triglycerides”, and “Excluded: missing SWB”. This table summarises age, sex, education, residential location, marital status, baseline life satisfaction, and baseline presence of any PPC component, and is presented in Supplementary Table S1.(Table S1).

Fig. 1.

Flowchart of participant inclusion and exclusion criteria for the final analytic sample.

Subjective well-being

SWB was measured by assessing participants’ overall satisfaction with life using a single-item question on a 5-point Likert scale: “Overall, how satisfied are you with your life?” Response options included: “not at all satisfied,” “not very satisfied,” “somewhat satisfied,” “very satisfied,” and “completely satisfied.” To enhance statistical power, life satisfaction responses were recategorized into three groups: (1) “Not satisfied” (combining “not at all satisfied” and “not very satisfied”); (2) “Somewhat satisfied”; and (3) “Satisfied” (combining “very satisfied” and “completely satisfied”).

Physical, psychological, and cognitive multimorbidity

Physical, psychological, and cognitive conditions were assessed at baseline and at each follow-up. Participants were identified as having a physical condition if they reported any of the following: hypertension, diabetes, cancer, chronic lung disease, heart disease, stroke, or arthritis⁹. These seven chronic diseases are common, burdensome in later life, and modifiable through lifestyle factors. Psychological problems were defined using the 10-item short form of the Center for Epidemiologic Studies Depression Scale (CESD-10), which has high validity and excellent psychometric properties in Chinese older adults⁹. A score of ≥ 10 was classified as having depressive symptoms, and a score < 10 was defined as normal. Cognitive status was assessed at baseline and follow-up using a multidimensional cognitive test battery including immediate recall, delayed recall, serial sevens, and orientation. Within each dimension, cognitive impairment was defined as a score more than 1.5 standard deviations below the age-specific mean. Participants were defined as having overall cognitive impairment if they were impaired in any of the above dimensions⁸. PPC-MM includes four patterns: physical-psychological multimorbidity (co-occurrence of physical and psychological conditions), physical-cognitive multimorbidity (physical and cognitive conditions), psychological-cognitive multimorbidity (psychological and cognitive conditions), and physical-psychological-cognitive multimorbidity (co-occurrence of all three conditions) ¹⁰. Participants free of PPC-MM at baseline—defined as having fewer than two of the three component conditions at the 2011 wave—were followed until the first wave at which at least two conditions were simultaneously present, censoring (e.g., loss to follow-up), death (treated as a competing event in Fine-Gray models), or study end (2018). Thus, individuals with zero or one abnormal domain at baseline were retained in the risk set, whereas those with PPC-MM (≥ 2 domains) at baseline were excluded. When the component conditions first appeared in different waves (for example, a physical condition first present at wave 2 and cognitive impairment first present at wave 3), the incident time of PPC-MM was assigned to the first wave at which two or more components co-occurred (in this example, wave 3); in the discrete-time survival models, this was operationalized as a binary indicator of incident PPC-MM for the corresponding follow-up interval.

Covariates

Covariates assessed at baseline included age, sex, education level, residence, and marital status. Education level was categorized into three groups: primary school or below, high school, and college or above. Residence was classified as rural or urban based on household registration (hukou). Marital status was categorized as married (including those with a spouse temporarily absent) or unmarried/separated/divorced/widowed. Health-related behaviors included smoking status, drinking status, and sleep duration. Smoking status was determined by asking participants “Have you ever smoked?” and “Do you currently smoke?” and classified as never smoked, former smoker, or current smoker. Drinking status was based on the frequency of alcohol consumption in the past year and categorized as never drink, drink but less than once a month, or drink once a month or more. Sleep duration was self-reported as the average hours of sleep per night in the past month. Baseline clinical and biochemical indicators included body mass index (BMI, kg/m², continuous), waist circumference (cm, continuous), white blood cell count (× 10⁹/L, continuous), platelet count (× 10⁹/L, continuous), glycated hemoglobin (%, continuous), hemoglobin (g/dL, continuous), total cholesterol (mg/dL, continuous), triglycerides (mg/dL, continuous), high-density lipoprotein cholesterol (HDL-C, mg/dL, continuous), and low-density lipoprotein cholesterol (LDL-C, mg/dL, continuous).

Missing data and sensitivity analyses for baseline biomarkers

Within the analytic cohort described above, missing values in the exposure and covariates during follow-up were handled using multiple imputation by chained equations (MICE) with 20 imputations. The imputation set included SWB, sociodemographic covariates (age, sex, education, residential location, marital status), lifestyle factors (smoking, drinking, sleep duration, BMI, waist circumference), and all serum biomarkers (white blood cell count, platelets, glycated haemoglobin, haemoglobin, total cholesterol, triglycerides, HDL-C, LDL-C).

Imputation models were specified according to variable type: continuous variables were imputed using predictive mean matching; binary or nominal categorical variables were imputed using logistic or multinomial logistic models; and ordered categorical variables (e.g. SWB, education, smoking, drinking) were imputed using proportional odds models. Outcome and time variables (participant identifier, interval indicator, event status, event occurrence within the interval, and follow-up time in the discrete-time survival models) were fully observed and were included as predictors in the imputation models, but were not themselves imputed.In this extended cohort, a person-interval dataset was created and MICE with 20 imputations was applied to missing baseline covariates and biomarkers, including triglycerides; outcome and time variables were used as predictors but not imputed. Each imputed dataset was analysed separately, and estimates were pooled using Rubin’s rules.

Statistical analysis

At baseline, descriptive statistics were used to present sample characteristics by SWB group. Continuous variables are presented as mean ± standard deviation and compared using one-way analysis of variance (ANOVA). Categorical variables are presented as n (%) and compared using the chi-square test or Fisher’s exact test, as appropriate.

Discrete-time survival models

The primary outcome was the first occurrence of PPC-MM during follow-up. The analysis cohort consisted of individuals free of PPC-MM at baseline (PPM = 0). For each participant, follow-up time was measured from baseline (2011) to the first wave at which PPC-MM was detected (2013, 2015, or 2018), or to the last available observation if PPC-MM did not occur. Thus, incident PPC-MM was intrinsically interval censored, with the event known only to occur between two survey waves.

Discrete-time survival models were fitted using person-interval data, with follow-up divided into three intervals (2011–2013,2013–2015,2015–2018). A quasibinomial model with a complementary log–log link was used to approximate proportional hazards and to estimate interval-specific hazard rates. Indicator variables for follow-up interval were included to flexibly model the baseline hazard.

Covariates were introduced sequentially in a series of nested models:(1)Model 0 (M0): SWB and follow-up interval only;(2) Model 1 (M1): M0 plus age and sex; (3) Model 2 (M2): M1 plus education, residential location, and marital status;(4) Model 3 (M3): M2 plus smoking status, drinking status, sleep duration, BMI, and waist circumference; (5)Model 4 (M4): M3 plus white blood cell count, platelets, glycated haemoglobin, haemoglobin, total cholesterol, triglycerides, HDL-C, and LDL-C.

SWB was treated as an ordered three-level exposure and parameterised using orthogonal polynomial contrasts, yielding linear (SWB.L) and quadratic (SWB.Q) trend components, allowing both monotonic and non-linear patterns across SWB categories to be assessed.

Fine-Gray competing risks and Cox proportional hazards models.

Death was treated as a competing event in Fine-Gray subdistribution hazard models to evaluate the association between baseline SWB and incident PPC-MM. Subdistribution hazard ratios were estimated, and cumulative incidence functions (CIFs) for PPC-MM were computed and plotted by baseline SWB category, with numbers at risk shown at 0, 2, 4, and 7 years.

As secondary sensitivity analyses, continuous-time Cox proportional hazards models were fitted, with death treated as censoring. These Cox models used the same covariate sets (M0-M4) as the discrete-time models and were implemented to assess robustness to the choice of time scale and model specification. The proportional hazards (PH) assumption was evaluated using Schoenfeld residuals (cox.zph function). When covariate-specific tests for SWB suggested marginal deviations from proportionality, time-stratified Cox models were fitted with separate baseline hazards for 0–4, 4–7, and ≥ 7 years of follow-up. All Cox-based models were interpreted as supportive sensitivity analyses rather than replacements for the primary discrete-time models.

Inverse probability weighting and marginal structural models

To account for potential time-varying confounding affected by prior exposure, marginal structural models (MSMs) were fitted using stabilized inverse probability of treatment and censoring weights. SWB was treated as a three-level time-varying exposure (“Satisfied”, “Somewhat satisfied”, “Not satisfied”) in a person-period dataset with three follow-up intervals.

For the treatment weights (IPTW), multinomial logistic regression models were estimated for SWB at each interval. The denominator model included previous SWB and time-varying covariates: age, sex, education, residential location (village vs city/town), marital status, smoking status, drinking frequency, sleep duration, BMI, waist circumference, glycated haemoglobin, total cholesterol, triglycerides, HDL-C, LDL-C, and an indicator for follow-up interval. The numerator model included previous SWB, baseline covariates (age, sex, education, residential location) and the follow-up interval only. Stabilized IPTW were obtained as the ratio of the predicted probabilities from the numerator and denominator models and were cumulated over time within participants.

For the censoring weights (IPCW), logistic regression models were used to estimate the probability of remaining uncensored and alive at the end of each interval. The denominator model included previous SWB, age, sex, education, residential location, marital status, and follow-up interval, whereas the numerator model included only the interval indicator. Cumulative stabilized IPCW were multiplied with the IPTW to obtain combined stabilized weights. The combined weights were truncated at the 1st and 99th percentiles to reduce the influence of extreme values, multiplied by the CHARLS sampling weights, and normalised to have mean 1 in the analytic sample.

The MSM was specified as a discrete-time survival model with a complementary log–log link, including SWB and follow-up interval as the primary predictors. Additional baseline covariates were included to improve precision. All MSMs and corresponding weights were estimated separately within each of the 20 imputed datasets, and regression coefficients and standard errors were combined using Rubin’s rules. Weight distributions and covariate balance before and after weighting were examined using summary statistics, histograms, and absolute standardised mean differences .

Sensitivity analyses and subgroup analyses

Several sensitivity analyses were conducted. First, to reduce the possibility of reverse causation, participants who developed any PPC component during the first follow-up interval (baseline to year 2) were excluded, and the marginal structural models were re-estimated using the same weighting strategy and survey design specification as in the main analysis. Second, discrete-time models were refitted in a larger cohort without requiring measured baseline triglycerides, using multiple imputation for missing baseline biomarkers; estimates from these models were compared with those from the complete-case analytic sample restricted to participants with measured triglycerides and complete biomarker data (Supplementary Table S2,Table S6, Table S9). Third, discrete-time models with and without adjustment for baseline biomarkers (Models M3 and M4) were compared in the multiply imputed datasets to assess the impact of biomarker adjustment on the SWB-PPC-MM association. Fourth, to address potential bias arising from the inclusion of depressive symptoms in the psychological domain of PPC-MM, a sensitivity outcome was constructed focusing solely on physical-cognitive multimorbidity (PC-MM), and the main modelling strategy was repeated for this outcome.

Subgroup analyses were performed according to sex, age group, marital status, education, and residence. Interaction terms between SWB and each subgroup variable were tested using likelihood ratio tests, and subgroup-specific hazard ratios with 95% confidence intervals were reported from discrete-time complementary log–log models with multiple imputation-pooled estimates.

All analyses were conducted in R (version 4.4.2). Two-sided p-values < 0.05 were considered statistically significant. All methods were performed in accordance with relevant guidelines and regulations, including the Declaration of Helsinki for observational studies involving human participants.

Results

Sample characteristics and baseline analysis

Baseline characteristics differed significantly across SWB groups (p < 0.001). For example, the “Somewhat satisfied” group had the highest education level (38% with high school or above), whereas the “Not satisfied” group had the highest proportion of rural hukou (85%). Health behaviours and cardiometabolic profiles also varied: BMI, waist circumference, smoking status, drinking frequency, and sleep duration all showed significant differences (p < 0.05). For instance, the “Somewhat satisfied” group had a BMI of 23.9 ± 3.6 kg/m² compared with 23.4 ± 3.7 kg/m² in the “Not satisfied” group (p < 0.001), and the “Satisfied” group reported longer sleep duration (6.56 ± 1.92 h) than the “Not satisfied” group (6.12 ± 2.05 h, p < 0.001)(Table 1).

Table 1.

Shows baseline characteristics by SWB group.

Baseline characteristics (2011)	SWB (baseline)				pvalue2
	Overall N = 68431	Satisfied N = 15851	Somewhat satisfied N = 41701	Not satisfied N = 10881
Age (years)	59 ± 9	60 ± 10	59 ± 9	58 ± 9	< 0.001
Sex					< 0.001
Male	3227 (47%)	735 (46%)	2040 (49%)	452 (42%)
Female	3616 (53%)	850 (54%)	2130 (51%)	636 (58%)
Education					< 0.001
Primary school or below	4454 (65%)	1118 (71%)	2574 (62%)	762 (70%)
High school	2090 (31%)	397 (25%)	1384 (33%)	309 (28%)
College or above	299 (4.4%)	70 (4.4%)	212 (5.1%)	17 (1.6%)
Residence (hukou)					< 0.001
Village	5386 (79%)	1,287 (81%)	3178 (76%)	921 (85%)
City/town	1457 (21%)	298 (19%)	992 (24%)	167 (15%)
Marital status					< 0.001
Married	6093 (89%)	1389 (88%)	3785 (91%)	919 (84%)
Never-married/separated/widowed	750 (11%)	196 (12%)	385 (9.2%)	169 (16%)
Smoking status					< 0.05
Non-smoker	4100 (60%)	945 (60%)	2468 (59%)	687 (63%)
Ex-smoker	646 (9.4%)	163 (10%)	406 (9.7%)	77 (7.1%)
Current smoker	2092 (31%)	475 (30%)	1293 (31%)	324 (30%)
Unknown	5	2	3	0
Drinking frequency (last year)					< 0.001
Never	4569 (70%)	1045 (70%)	2730 (69%)	794 (77%)
Drink but less than once a month	526 (8.1%)	120 (8.0%)	336 (8.5%)	70 (6.8%)
Drink more than once a month	1400 (22%)	337 (22%)	898 (23%)	165 (16%)
Unknown	348	83	206	59
Sleep time (hours/night)	6.39 ± 1.84	6.56 ± 1.92	6.40 ± 1.75	6.12 ± 2.05	< 0.001
Unknown	45	14	20	11
BMI (kg/m2)	23.8 ±  3.6	23.9 ± 3.6	23.9 ± 3.6	23.4 ± 3.7	< 0.001
Unknown	823	185	525	113
Waist circumference (cm)	85 ± 13	86 ± 12	85 ± 13	84 ± 12	< 0.001
Unknown	794	183	503	108
White blood cell (109/L)	6.24 ± 2.23	6.27 ± 1.93	6.21 ± 2.41	6.27 ± 1.88	0.642
Unknown	133	32	77	24
Platelets (109/L)	213 ± 72	212 ± 74	213 ± 71	217 ± 71	0.121
Unknown	130	31	76	23
HbA1c (%)	5.27 ± 0.84	5.29 ± 0.83	5.27 ± 0.84	5.25 ± 0.83	0.400
Unknown	37	5	20	12
Haemoglobin (g/dl)	14.50 ± 2.13	14.56 ± 2.25	14.50 ± 2.08	14.38 ± 2.14	0.086
Unknown	131	31	77	23
Total cholesterol (mg/dl)	194 ± 39	193 ± 37	194 ± 39	194 ± 39	0.939
Unknown	3	0	3	0
Triglycerides ((mg/dl)	137 ± 110	137 ± 106	137 ± 108	139 ± 119	0.866
HDL-C ((mg/dl)	50 ± 15	50 ± 15	50 ± 15	51 ± 16	0.640
Unknown	3	2	1	0
LDL-C ((mg/dl)	117 ± 35	116 ± 35	117 ± 35	116 ± 36	0.278
Unknown	13	3	9	1

¹Mean ± SD; n (%), ²One-way analysis of means; Pearson’s Chi-squared test.

Baseline characteristics of participants included in the analytic sample (N = 6843) and those excluded because of missing triglycerides (N = 4189) or missing SWB (N = 1043) are shown in Supplementary Table S1.(Table S1) Excluded participants tended to be slightly older and more likely to live in rural areas, but the distributions of SWB categories and baseline PPC components were broadly similar across groups.

Discrete-time survival analysis

Polynomial contrasts treating SWB as an ordinal exposure (“Satisfied” → ”Somewhat satisfied” → ”Not satisfied”) showed that lower SWB was consistently associated with higher PPC-MM risk: in the fully adjusted model M4, the linear trend yielded HR = 2.03 (95% CI 1.82–2.26, p < 0.001) and the quadratic term HR = 1.24 (95% CI 1.15–1.34, p < 0.001), indicating a predominantly monotonic dose–response with mild curvature. Comparable estimates were observed across M0-M3 (linear HR = 2.01–2.04; quadratic HR = 1.24–1.33; all p < 0.001), underscoring the robustness of the association (Table 2).

Table 2.

Discrete-time survival models.

Model	SWB component	HR	95% CI	p value
Model0 (Crude)	Linear trend (SWB.L)	2.04	1.84–2.27	< 0.001
	Quadratic trend (SWB.Q)	1.33	1.24–1.44	< 0.001
Model1	Linear trend	2.01	1.81–2.23	< 0.001
	Quadratic trend	1.30	1.20–1.40	< 0.001
Model2	Linear trend	2.03	1.82–2.26	< 0.001
	Quadratic trend	1.25	1.15–1.35	< 0.001
Model3	Linear trend	2.02	1.82–2.26	< 0.001
	Quadratic trend	1.24	1.15–1.34	< 0.001
Model4	Linear trend	2.03	1.82–2.26	< 0.001
	Quadratic trend	1.24	1.15–1.34	< 0.001

Marginal structural model results

Using discrete-time marginal structural models with a complementary log–log link and combined stabilized inverse probability-of-treatment and inverse probability-of-censoring weights, time-varying SWB remained strongly associated with incident PPC-MM. For the ordinal linear contrast of time-updated SWB, the weighted MSM estimated an HR of 1.95 (95% CI 1.73–2.19, p < 0.001), and the quadratic contrast yielded an HR of 1.22 (95% CI 1.12–1.33, p < 0.001), indicating a largely monotonic dose–response relationship. These estimates were similar to those from the primary discrete-time models (Table S2).

Across all person-intervals, stabilized weights showed moderate variability with no major extremes. Before truncation, cumulative treatment and censoring weights had means close to 1.0 and narrow interpercentile ranges; after truncation at the 1st and 99th percentiles, the combined stabilized weights used in MSM analyses had a mean of 1.03 (SD 0.26) and ranged from 0.49 to 2.98 (Supplementary Table S3, Figures S1-S2). Weighting also markedly improved covariate balance across SWB categories: in the weighted pseudo-population, almost all standardized mean differences for sociodemographic, behavioural, and biomarker covariates were < 0.10 (Love plot in Supplementary Figure S3), supporting the use of IPTW/IPCW to address time-varying confounding.

Fine-gray model

In Fine-Gray models treating death as a competing event, baseline SWB was strongly associated with the cumulative incidence of PPC-MM(Table3). Compared with participants who were satisfied with life, those who were not satisfied had a higher subdistribution hazard of PPC-MM (HR = 1.38, 95% CI 1.18–1.62), whereas those who were somewhat satisfied had a similar risk (HR = 1.03, 95% CI 0.93–1.16). Older age was associated with a slightly lower subdistribution hazard (HR per 10-year increase = 0.90, 95% CI 0.85–0.95).

Table 3.

Fine-gray competing risks model results.

Term	HR	CI_low	CI_high	P value
SWB_baseline somewhat satisfied	1.03	0.925	1.16	0.556
SWB_baseline not satisfied	1.38	1.18	1.62	< 0.001
Age10	0.897	0.847	0.950	< 0.001
Sex female	1.48	1.28	1.72	< 0.001
Education.L	0.691	0.547	0.874	< 0.001
Education.Q	1.15	0.986	1.33	0.075
Location city/town	0.630	0.542	0.732	< 0.001
Maritalnever-married/separated/widowed	0.997	0.840	1.18	0.968
Smoking.L	1.09	0.983	1.20	0.104
Smoking.Q	0.886	0.768	1.02	0.099
Drinking.L	0.924	0.836	1.02	0.121
Drinking.Q	1.07	0.920	1.24	0.382
Sleep_time	0.946	0.920	0.973	< 0.001
BMI	1.00	0.997	1.00	0.930
Waist	1.00	0.998	1.01	0.256
White_blood_cell	0.993	0.974	1.01	0.456
Platelets	1.000	0.999	1.00	0.796
Glycated_hemoglobin	1.08	1.02	1.15	< 0.05
Haemoglobin	1.02	0.994	1.04	0.141
Total_cholesterol	0.999	0.994	1.00	0.616
Triglycerides	1.00	0.999	1.00	0.504
HDL_C	1.00	0.996	1.01	0.449
LDL_C	1.00	0.997	1.01	0.542

Figure S4shows the cumulative incidence functions of PPC-MM over 7 years by baseline SWB, together with numbers at risk at 0, 2, 4 and 7 years.(Figure S4)The corresponding 7-year cumulative incidence estimates were 40.7% for participants who were satisfied, 39.8% for those somewhat satisfied, and 53.5% for those not satisfied .

Cause-specific Cox models, which estimate the instantaneous hazard of PPC-MM among participants who are still alive and event-free, yielded very similar patterns (Supplementary Table S4). The HR for being not satisfied vs satisfied was 1.45 (95% CI 1.22–1.73), and the HR for being somewhat satisfied vs satisfied was 1.04 (95% CI 0.92–1.17); the age effect was also similar (HR per 10-year increase = 0.90, 95% CI 0.84–0.96). These findings indicate that competing risk of death does not change the substantive interpretation: lower SWB, particularly being not satisfied with life, is associated with a higher risk and higher cumulative incidence of PPC-MM.s with numbers at risk.

Missing data and multiple imputation

To assess the robustness of our findings to the handling of missing data, we compared estimates from complete-case analyses (restricted to participants with measured baseline triglycerides and complete biomarker data) with those from the multiply imputed datasets. (Table S5)As shown in Supplementary Table S6,(Table S6) the estimated associations between SWB and incident PPC-MM were very similar across approaches.In the fully adjusted discrete-time models (M4), the HR for the linear SWB contrast (SWB.L) was 1.95 (95% CI 1.73–2.19) in the complete-case analysis and 2.03 (95% CI 1.82–2.26) in the multiply imputed analysis. The quadratic contrast (SWB.Q) also showed closely comparable estimates (complete-case HR 1.22 (95% CI 1.12–1.33) vs imputed HR 1.24 (95% CI 1.15–1.34). For most covariates, the direction and magnitude of associations were consistent between complete-case and imputed models, with narrower confidence intervals in the imputed analyses. These findings indicate that alternative treatments of missing biomarkers (exclusion vs multiple imputation) do not materially change the estimated association between SWB and PPC-MM.(Table S7).

Sensitivity analyses

In Cox analyses replicating the five sequentially adjusted models (M0-M4), categorical comparisons showed that, relative to the “Satisfied” group, participants who were “Not satisfied” had a higher risk of PPC-MM (M4: HR = 1.57, 95% CI 1.34–1.83, p < 0.001), whereas the risk in the “Somewhat satisfied” group was close to the null (HR = 1.07, 95% CI 0.96–1.20, p = 0.204). When SWB was treated as an ordered exposure (“Satisfied” → ”Somewhat satisfied” → ”Not satisfied”) and modelled using polynomial contrasts, the fully adjusted M4 model yielded a linear trend HR of 1.37 (95% CI 1.23–1.54, p < 0.001) and a quadratic term HR of 1.13 (95% CI 1.04–1.23, p < 0.01); estimates in M0-M3 were similar (linear HR≈1.37–1.40; quadratic HR≈ 1.13–1.22, all p < 0.01). A likelihood ratio test supported the categorical specification over a strictly linear trend (χ² = 9.08, df = 1, p < 0.01). Overall, the Cox model results were directionally consistent with the primary discrete-time models but yielded somewhat more conservative effect sizes (primary analysis: linear HR = 2.03, quadratic HR = 1.24; Table S8).

The time-stratified Cox models, which were used to address departures from the proportional hazards assumption, showed that within each follow-up interval, the hazard ratios for the “Somewhat satisfied” group versus the “Satisfied” group were consistently close to the null (e.g., HR = 1.09, 95% CI 0.95–1.24, p = 0.206 for 0–4 years), whereas the hazard ratios for the “Not satisfied” group versus the “Satisfied” group were consistently greater than 1 (e.g., HR = 1.18, 95% CI 0.99–1.41, p = 0.067 for 0–4 years), in line with the main analysis.

Time-varying SWB models using the start-stop approach showed pronounced effects: compared with being satisfied, the HR was 1.25 (95% CI 1.12–1.40, p < 0.001) for being somewhat satisfied and 2.26 (95% CI 1.93–2.64, p < 0.001) for being not satisfied. These findings reinforce the dose–response pattern and are consistent with the marginal structural model (MSM) results (linear HR = 1.95). In fully adjusted discrete-time models (M4), the HR for the linear SWB contrast (SWB.L) was 1.95 (95% CI 1.73–2.19) in the complete-case analysis and 2.03 (95% CI 1.82–2.26) in the multiply imputed analysis; estimates for the quadratic contrast (SWB.Q) were likewise nearly identical (complete-case HR = 1.22 (95% CI 1.12–1.33) vs imputed HR = 1.24(95% CI 1.15–1.34). Comparing models without and with baseline biomarkers (M3 vs M4) produced virtually unchanged estimates (SWB.L HR2.02 vs 2.03; SWB.Q HR 1.24 in both;(Table S6) , indicating that excluding versus imputing missing biomarkers and further biomarker adjustment do not materially influence the SWB-PPC-MM association.

To address potential reverse causation, residual confounding, and common-source or construct-overlap bias due to inclusion of depressive symptoms in the psychological domain of PPC-MM, we performed an additional sensitivity analysis using physical-cognitive multimorbidity (PC-MM) as the outcome. PC-MM was defined as the co-occurrence of at least one physical condition and cognitive impairment, explicitly excluding depressive symptoms. Using the same modelling strategy as in the primary analyses, baseline SWB was not clearly associated with incident PC-MM: the HR for the linear trend contrast (SWB.L, representing lower life satisfaction) was 0.962 (95% CI 0.904–1.02, p = 0.231), and the HR for the quadratic contrast (SWB.Q) was 0.988 (95% CI 0.950–1.03, p = 0.560), both non-significant. These near-null findings for PC-MM, contrasted with the robust associations observed for PPC-MM, underscore the role of the psychological component and provide additional reassurance that the main SWB-PPC-MM results are not artefacts of model specification or residual bias.

Finally, when we excluded participants who developed any PPC component during the first follow-up interval (baseline to year 2), the associations between baseline SWB and subsequent PPC multimorbidity were slightly strengthened but remained highly consistent with the main analysis. Compared with being satisfied, the HR = 1.59 (95% CI 1.41–1.78, p < 0.001)) for being somewhat satisfied and 3.86 (95% CI 3.19–4.67, p < 0.001)) for being not satisfied. These estimates are very similar to the corresponding HRs in the primary analysisHR = 1.47 (95% CI 1.34–1.62)and HR = 3.40(95% CI 2.96–3.92, p < 0.001, respectively). This pattern suggests that our findings are unlikely to be fully explained by reverse causation.

Subgroup analysis

Subgroup analyses, using discrete-time complementary log–log models with multiple imputation-pooled estimates and adjustments for key covariates (age, sex, education, location, marital status, smoking, drinking, sleep duration, BMI, waist circumference, and blood biomarkers), demonstrated consistently elevated PPC-MM risk linked to low baseline SWB (“Not satisfied” vs. “Satisfied”) across all strata, reinforcing its predictive value. Effects were robust and directionally uniform, with no opposing trends relative to the main analysis. (Fig. 2).

Fig. 2.

Forest plot of the effect of life satisfaction on PPC-MM risk across subgroups.

Model diagnostics and performance

In the Cox sensitivity analyses, the global Schoenfeld test did not indicate a departure from proportional hazards (χ² = 19.8, df = 23, p = 0.654). Most individual covariates showed no evidence of important time-variation; the overall test for the SWB factor was borderline (χ² = 5.40, df = 2, p = 0.067), and a level-specific test for the “Not satisfied” indicator yielded a small p value (p < 0.001) despite the corresponding scaled Schoenfeld residuals showing an essentially flat trend over time (Figure S5)(Table S10). To probe these findings, we fitted time-stratified Cox models with separate baseline hazards for 0–4, 4–7, and ≥ 7 years of follow-up. Across these intervals, the hazard ratios comparing “Not satisfied” with “Satisfied” remained consistently above 1 (e.g. HR = 1.18, 95% CI 0.99–1.41 in the first 0–4-year interval), whereas estimates for “Somewhat satisfied” were close to the null (e.g. HR = 1.09, 95% CI 0.95–1.24), with no reversal of effect direction. Taken together, these results suggest at most modest time-variation in the SWB effect that does not materially affect the overall association between SWB and PPC-MM. The discriminative performance of the Cox model was moderate, with a Harrell’s C-index of 0.693.

Discussion

Based on a nationally representative longitudinal cohort, this study systematically evaluated the prospective association between SWB and PPC-MM. The results consistently showed that: (1) the lower the SWB, the higher the risk of PPC-MM, with a monotonic dose–response relationship with mild curvature (linear trend HR = 2.03; quadratic HR = 1.24) ;(2) After considering the confounding of temporal variations, the temporal variation SWB still significantly predicted a higher risk (MSM linear HR = 1.95);(3) Fine- The results of the Gray model are consistent, suggesting that SWB is an independent and robust risk marker. (4) Multiple sensitivity analyses and subgroup analyses (gender, age, education, urban and rural, lifestyle, sleep, BMI, etc.) did not show a trend opposite to the main effect. Combined with the existing concept and inconsistent definition of PPC-MM, this study makes up for the lack of distinguishing between different combinations of diseases by clearly including three types of diseases: physical, psychological and cognitive, which has important epidemiological and public health significance¹¹.

SWB has increasingly been recognized as a key health indicator. Evidence indicates that higher SWB is associated with better health outcomes, lower risk of chronic diseases, and longer life expectancy^12,13. Studies across diverse populations suggest that individuals with higher SWB experience lower incidence of cardiovascular disease, fewer depressive symptoms, and a reduced burden of chronic conditions^14,15. Longitudinal research in older adults further shows that higher well-being is related to lower mortality, fewer chronic diseases, and stronger immune function¹⁶. Multinational reviews and national cohort studies similarly report that populations with higher SWB have better psychological, social, and physical health and enjoy longer and healthier lives^17,18. For example, analyses from the English Longitudinal Study of Ageing (ELSA) indicate that higher affective well-being is associated not only with lower all-cause mortality but also with longer disability-free and chronic disease-free life expectancy, and that sustained enjoyment of life predicts lower mortality^19,20. Beyond Western settings, studies in Chinese older adults also find that higher SWB is associated with reduced all-cause mortality, supporting cross-cultural consistency in the SWB-health link²¹.

Our study not only identified a robust association betweenSWB and PPC-MM but also showed that this relationship varies across sociodemographic groups, including by sex, residence, and education. The higher risk observed among women is consistent with evidence from Chinese older populations and international studies documenting greater burdens of depression and anxiety, steeper declines in self-rated health, and distinct healthcare-utilization patterns in women²².

These contextual differences suggest that sociodemographic position may shape both exposure to and the consequences of low SWB, contributing to heterogeneity in PPC-MM risk.

Mechanistically, several mutually reinforcing pathways help explain the stronger gradient in women. Biologically, low SWB is linked to heightened hypothalamic–pituitary–adrenal (HPA) axis activity and a pro-inflammatory profile (elevated IL-6, CRP), processes that accelerate atherosclerotic and neurodegenerative changes relevant to multimorbidity²³. Behaviorally, lower SWB co-occurs with smoking, physical inactivity, and insufficient sleep, which amplify metabolic and inflammatory load²⁴. Socially, constrained structural support and uneven access to healthcare—conditions that can be intensified by caregiving roles and gendered expectations—limit stress buffering and timely care-seeking^25,26. Together, these factors render women more susceptible to the sequence “low SWB → unhealthy behaviors → chronic inflammation/sleep impairment → multimorbidity,” providing a coherent framework for the observed sex differences^27,28.

Secondly, the lower risk of PPC-MM among urban residents compared to rural residents can be explained by disparities in health resources and social capital: urban areas have a higher density of primary and specialized healthcare resources, more convenient screening and follow-up systems, and richer formal and informal social support networks, which are conducive to maintaining higher SWB and interrupting the progression of multimorbidity^29,30. Urban residents also have advantages in health literacy, accessibility to sports facilities, and environmental exposures (e.g., clean energy, green spaces). These factors can not only enhance SWB but also effectively block the development of multimorbidity. Research indicates a bidirectional relationship between SWB and health in old age, where increased SWB can promote health maintenance and survival . Studies have shown that supportive health environments (e.g., “health parks” and “trails” in cities) can improve residents’ health status, and this benefit is more pronounced among populations with lower health literacy^31,32.

Similarly, the lower risk of PPC-MM among individuals with higher educational attainment aligns with evidence from 33 countries showing that higher socioeconomic status (SES) is associated with lower prevalence and slower progression of PPC-MM³³. Education protects health through multiple pathways. First, it strengthens cognitive and emotional regulation, enabling individuals to appraise stressors more positively and maintain higher SWB. Second, it expands social networks and resource reserves (e.g., social support, access to healthcare, and stable employment), thereby increasing opportunities for health³⁴. Third, education promotes healthy behaviors and better self-management³⁵. Consistent with our multi-model findings, after extensive adjustment for demographics (age, sex, education/wealth/residence), health behaviors (smoking, drinking, physical activity, BMI), and baseline health status (pre-existing chronic conditions and, in some studies, biomarkers such as HbA1c), low SWB remained significantly associated with adverse health outcomes, indicating that the SWB-health link is not fully explained by SES or lifestyle factors^36,37.

We also found a dose–response relationship between SWB and PPC-MM: the lower the SWB, the higher the subsequent risk of PPC-MM, with a clearer gradient in the time-varying exposure framework. This pattern is consistent with life-course theories of “long-term coupling” between subjective well-being and health, and with psychoneuroimmunological evidence that well-being influences health through systemic pathways such as stress-inflammation³⁸. External evidence from ELSA further supports this framework: higher affective well-being predicts longer disability-free and chronic disease-free life expectancy as well as lower all-cause mortality³⁹. National cohort studies of Chinese older adults have similarly linked higher SWB to lower all-cause mortality, suggesting cross-cultural reproducibility of this association, though the effect sizes may vary across social and healthcare contexts⁴⁰. Moreover, long-term ELSA follow-up shows that greater “enjoyment of life” is associated with reduced risk of incident type 2 diabetes, indicating that SWB effects extend beyond psychological outcomes into metabolic and other physical domains—echoing our integrated PPC-MM construct³⁹.

At the same time, we observed modest differences in the magnitude of the association between SWB and PPC-MM across modelling strategies, with larger linear trend estimates in the discrete-time complementary log–log models than in the continuous-time Cox models. This pattern is expected and mainly reflects the interval-censored nature of PPC-MM onset in CHARLS and the way event times are handled in each framework. In our data, incident PPC-MM could only be ascertained at biennial or triennial follow-up waves, so the exact onset time is only known to fall between two visits. Discrete-time survival models with a complementary log–log link are specifically designed for such grouped or interval-censored data and directly model the conditional hazard of first PPC-MM within each follow-up interval. By contrast, the Cox models necessarily assigned the event time to the wave at which PPC-MM was first detected and treated that wave as if it were the exact failure time, introducing non-differential error along the time axis. When follow-up intervals span 2–3 years, this approximation tends to attenuate hazard ratios towards the null. Consistent with this, the fully adjusted discrete-time model yielded a linear trend HR of 2.03, whereas the corresponding continuous-time Cox model produced a smaller linear trend HR of 1.37, even though both approaches identified the “not satisfied” group as having the highest risk and showed a monotonic dose–response across SWB categories. Time-varying SWB analyses using marginal structural models and Cox start-stop formulations, as well as Fine-Gray competing-risk models, gave effect estimates closer in magnitude to the discrete-time results and supported the same qualitative pattern. Overall, these findings suggest that low SWB is a robust and clinically meaningful risk marker for incident PPC-MM, and that differences in HR magnitude primarily reflect how interval censoring is treated, rather than genuine disagreement about the underlying association.

In conclusion, using nationally representative longitudinal data, we found a robust, dose–response association between lower SWB and higher subsequent risk of PPC-MM, independent of sociodemographic characteristics, lifestyle factors, and baseline health, and broadly consistent across key subgroups. These findings highlight SWB as a potentially important and modifiable risk marker for integrated physical, psychological, and cognitive multimorbidity in later life, with implications for screening, early identification, and holistic prevention strategies in aging populations. Future research should clarify the causal pathways linking SWB to PPC-MM, evaluate whether interventions that enhance well-being can alter multimorbidity trajectories, and examine how social, cultural, and healthcare contexts shape these relationships across different populations and stages of the life course.

Limitation and strength

The theoretical value of this study lies in its pioneering use of a longitudinal design to reveal the dynamic association between SWB and PPC-MM. This overcomes the limitations of previous research, which mainly focused on single diseases or cross-sectional analyses, and expands our understanding of the impact of SWB on multidimensional health (physical, psychological, and cognitive).

On a practical level, the study confirms that SWB is a significant, modifiable predictor of PPC-MM in middle-aged and older adults, offering new directions for healthy aging strategies. This includes integrating the enhancement of SWB into the life-course prevention system for PPC-MM. For high-risk populations, such as women and rural residents, tailored psychological support programs, combined with measures like sleep management and blood sugar control, could more precisely reduce the burden of multimorbidity.

However, the study has some limitations. First, there may be unobserved confounding factors, such as genetic background and specific aspects of social support, that could influence the association. Second, since the sample is exclusively from China, the cross-cultural applicability of the conclusions needs further verification. Third, the assessment accuracy of certain health-related variables (e.g., intensity of physical activity) is limited. Fourth, a key limitation is the inability to incorporate full survey weighting in our models. Although the CHARLS dataset is derived from a complex, multistage, stratified sampling design, the publicly available data lack complete information on sampling weights, strata, and primary sampling units (PSUs). As a result, our analyses could not account for these design features, which may introduce potential biases in estimates and underestimate standard errors. This challenge is consistent with other longitudinal studies using CHARLS data, where unweighted or partially weighted approaches were employed due to similar data constraints. Future research with access to restricted design variables could address this issue to enhance generalizability. Nevertheless, this study provides reliable evidence of the association between SWB and PPC-MM, thanks to its large, representative sample, rigorous longitudinal design, and multidimensional health assessment framework.

Fifth, we did not adjust for medication status because detailed, wave-consistent medication data were unavailable in the public CHARLS files. Certain medications (e.g., antidepressants, antihypertensives, lipid-lowering agents, or cognitive enhancers) could plausibly influence both mood/SWB and trajectories of physical, psychological, or cognitive conditions, which might bias effect estimates. The net direction is uncertain: pharmacologic effects may attenuate observed associations, whereas confounding by indication (i.e., treatment reflecting underlying severity) may exaggerate them. In addition, harmonized measures for physical activity intensity and social support were limited. Although we attempted to mitigate confounding through adjustment for observed covariates, residual confounding cannot be ruled out. These constraints are typical of secondary analyses using publicly available CHARLS data and should be considered when interpreting our findings.

Nevertheless,this study clearly demonstrates the protective effect of subjective well-being on the occurrence of physical, psychological, and cognitive multimorbidity in Chinese middle-aged and older adults. It suggests that enhancing SWB in older adults could become a key intervention to reduce the risk of multimorbidity and promote healthy aging, providing an important reference for developing comprehensive, individualized health management strategies for the elderly.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Zhao Xinwei conceived the study, curated and analyzed the data, and drafted the initial version of the manuscript. Zhang Wanyi assisted with data curation, statistical modelling, and revision of the manuscript. Zhang Linlin supervised the study design, guided the data analysis, and critically revised the manuscript. All authors (Zhao Xinwei, Zhang Wanyi, and Zhang Linlin) contributed equally to this work and approved the final version of the manuscript.

Zhang Wanyi assisted with data curation, statistical modelling, and revision of the manuscript.

Zhang Linlin supervised the study design, guided the data analysis, and critically revised the manuscript.

All authors (Zhao Xinwei, Zhang Wanyi, and Zhang Linlin) contributed equally to this work and approved the final version of the manuscript.

Funding

This research was supported by the Tianjin Municipal Education Commission Research Project Program (No.Grant 2023ZD035).

Data availability

The datasets analyzed in this study are publicly available from the China Health and Retirement Longitudinal Study (CHARLS) website: http://charls.pku.edu.cn. All data can be freely downloaded after registration and approval of a data use request. The CHARLS datasets used in the present analysis were downloaded on 28 June 2025.

Declarations

Competing interests

The authors declare no competing interests.