Abstract
Background
Cross-system multimorbidity—the co-occurrence of physical, psychological and cognitive disorders—is increasingly recognised as a defining feature of population ageing, yet its relationship with long-term mortality has rarely been examined at the level of specific combinations. Most studies have used chronic-disease counts or binary indicators that obscure cross-system patterning. We assessed the association of cross-system multimorbidity with 9-year all-cause mortality and identified high-risk patterns in a national sample of middle-aged and older Chinese adults.
Methods
We analysed 16,479 adults aged 45 years or older from the China Health and Retirement Longitudinal Study (CHARLS), enrolled at the 2011 baseline and followed through 2020. Three system indicators were defined: physical multimorbidity (≥ 2 of 14 self-reported physician-diagnosed chronic conditions), depression (CES-D-10 score ≥ 10), and cognitive impairment (lowest quartile of the CHARLS cognitive composite). The exposure was operationalised both as the number of involved systems (0/1/2/3) and as eight combination patterns (000–111). Missing exposure and covariate data were handled with multiple imputation (50 imputed datasets), and participants with incomplete follow-up were censored at their last known contact rather than excluded. Pooled Cox proportional hazards models were fitted with progressive adjustment for demographic, socioeconomic, and lifestyle/biomarker covariates.
Results
Over a mean follow-up of 7.96 years (SD 2.23), 1,033 participants died (6.27%). After full adjustment, each additional involved system was associated with a 34% higher hazard of all-cause mortality (HR 1.34, 95% CI 1.24–1.44; p < 0.001); compared with no involved systems, hazards were 1.34 (95% CI 1.09–1.66), 1.68 (1.35–2.09) and 2.48 (1.93–3.17) for one, two and three systems, respectively. In pattern-level analyses, co-occurring physical multimorbidity and cognitive impairment yielded a hazard close to all three systems combined (HR 2.14, 95% CI 1.57–2.91 vs HR 2.45, 95% CI 1.89–3.17; ratio of hazard ratios 0.87, 95% CI 0.66–1.16), whereas isolated depression showed no clear independent association (HR 1.12, 95% CI 0.81–1.57). Findings were consistent across complete-case, fully observed exposure, alternative cognitive definitions, discrete-time and sleep-adjusted analyses.
Conclusions
Cross-system multimorbidity exhibited a graded dose–response association with 9-year all-cause mortality among middle-aged and older Chinese adults, and the co-occurrence of physical multimorbidity and cognitive impairment alone reached a hazard close to that of triple-system involvement. This pattern identifies a group that may merit closer clinical attention, although the model was not evaluated for predictive performance, calibration or external validity and is not intended for individual risk stratification.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12889-026-28904-x.
Keywords: Multimorbidity, Depression, Cognitive dysfunction, Mortality, Cohort studies
Introduction
Multimorbidity,conventionally defined as the co-occurrence of two or more chronic conditions in the same individual, has become a defining feature of population ageing and a central challenge for health systems worldwide [1]. China is now home to the world's largest older population and is undergoing one of the most rapid demographic transitions in modern history; the number of adults aged 60 years or older has surpassed 280 million, and projected long-term care needs in this group are expected to grow several-fold over the next two decades [2, 3]. Chronic non-communicable disease has accordingly emerged as the dominant driver of disease burden and mortality among middle-aged and older Chinese adults, and a recent systematic review of 67 studies estimated the pooled prevalence of multimorbidity at approximately 25% across all adult age strata, rising to 39–46% in those aged 45 years or older [4]. Understanding which patterns of co-occurring conditions carry the greatest mortality risk in this population is therefore of high public-health relevance.
A consistent body of cohort evidence has established that multimorbidity, however operationalised, is associated with substantially elevated all-cause mortality among older adults; a meta-analysis of 26 cohort studies pooling more than 1.6 million participants reported a hazard ratio of 1.44 (95% CI 1.34–1.55) for ≥ 2 versus < 2 chronic conditions, with stronger associations as the number of conditions increased [5]. However, the predominant operationalisation of multimorbidity in this literature—either as a binary " ≥ 2 conditions" indicator or as a simple count of physical chronic diseases—aggregates clinically heterogeneous combinations into a single summary, and is therefore relatively uninformative about which configurations of co-occurring conditions carry the greatest risk [6]. This concern motivates a shift from "how many" to "which combination", and several authoritative operationalisations have explicitly recommended that mental and cognitive disorders be included alongside physical chronic diseases when characterising the multimorbidity burden of older adults [6]. The framework of physical, psychological and cognitive multimorbidity, which integrates physical chronic-disease burden with depressive symptomatology and cognitive impairment in the same individual, has been increasingly adopted in this regard, and has been shown to capture an order-of-magnitude steeper socioeconomic gradient than physical multimorbidity alone in a recent multi-cohort analysis spanning 33 countries [7].
Despite the growing adoption of this three-domain framework, several gaps remain. First, work on multimorbidity in middle-aged and older Chinese adults has so far concentrated either on the cross-sectional prevalence of physical, psychological and cognitive multimorbidity and its associated risk factors [8], on healthcare-process outcomes such as service utilisation and catastrophic health expenditure for physical multimorbidity [9], or, in the few longitudinal China Health and Retirement Longitudinal Study (CHARLS)–based applications of the cross-system framework, on the incidence of multimorbidity itself rather than on subsequent mortality [10]. Second, even when long-term mortality outcomes have been examined within this framework—chiefly in non-Chinese cohorts such as UK Biobank—exposures have typically been collapsed into a four-level summary (none, single, dual or triple involvement) that preserves the "how many" structure but does not resolve which specific combination of domains is most strongly associated with death [11]. Third, isolated depression, isolated cognitive impairment and their two-domain combinations have rarely been compared against the full three-domain configuration in the same cohort and in the same model, leaving open the clinically important question of whether any two-domain pattern approaches the mortality risk of involvement of all three domains. Resolving these gaps requires a national, prospective Chinese cohort with long follow-up, simultaneous measurement of physical, psychological and cognitive indicators at baseline, and statistical models that retain the full eight-cell cross-system structure. The same three-domain framework has since been applied prospectively in harmonised data from three national ageing cohorts including CHARLS, in which stressful life events across the life course predicted the development of specific physical, psychological and cognitive multimorbidity patterns; there too the endpoint was the onset of multimorbidity rather than death [12].
Within this context, we conducted a 9-year prospective analysis of cross-system multimorbidity and all-cause mortality in CHARLS, with three pre-specified objectives. First, we sought to characterise the dose–response association between the number of involved systems (physical, psychological, cognitive) and 9-year all-cause mortality, both as a continuous trend and as a four-level categorical exposure. Second, we sought to compare 9-year mortality hazard across all eight mutually exclusive cross-system combinations, with explicit attention to whether any two-domain combination approached the hazard of three-domain involvement. Third, we sought to confirm the robustness of these estimates across alternative model specifications and an alternative exposure derivation.
Methods
Study design and population
We conducted a prospective analysis of data from CHARLS, a nationally representative longitudinal survey of community-dwelling adults aged 45 years or older and their spouses, designed in harmonisation with the United States Health and Retirement Study [13]. CHARLS uses a four-stage stratified probability-proportional-to-size sampling design covering 28 of 31 mainland Chinese provincial-level units, with a national baseline survey of 17,708 respondents conducted between June 2011 and March 2012 (Wave 1) and follow-up assessments approximately every two years thereafter (Wave 2 in 2013, Wave 3 in 2015, Wave 4 in 2018, and Wave 5 in 2020). Data are collected by trained interviewers using a face-to-face computer-assisted personal interview instrument; physical and anthropometric measurements are obtained at every wave. The CHARLS protocol was approved by the Biomedical Ethics Review Committee of Peking University (IRB00001052-11015), and all participants provided written informed consent. The reporting of the present study follows the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement for cohort studies [14]. We used the 2011 Wave 1 baseline as the index time and ascertained vital status through Wave 5 (2020). Eligibility was defined solely by criteria that cannot be influenced by subsequent follow-up: an age of at least 45 years, a recorded sex, and a matchable Wave 1 individual record. One further requirement was imposed for outcome ascertainment, namely that at least one contact was recorded after the baseline interview, because respondents never observed again contribute neither time at risk nor information on vital status. Completeness of the exposure was deliberately not used as an eligibility criterion: participants with missing values on one or more of the three system indicators, or on any covariate, were retained in the analytic cohort and their missing values imputed, so that the cohort is not selected on the completeness of the exposure. Incomplete follow-up was likewise handled by censoring rather than by exclusion, participants alive at their last recorded contact being right-censored at that date. Residual item-level missingness on the individual system indicators (physical multimorbidity, 1.1%; depression, 10.1%; cognitive impairment, 48.9%) was subsequently handled by multiple imputation (see Statistical analysis).
Exposure: cross-system multimorbidity
Cross-system multimorbidity at the 2011 baseline was defined by three system-level binary indicators: physical multimorbidity, depression and cognitive impairment. Physical multimorbidity, in keeping with the prevailing operationalisation of multimorbidity as the co-occurrence of two or more chronic conditions [1, 6], was defined as the presence of two or more of the 14 self-reported physician-diagnosed chronic conditions covered in the CHARLS 2011 questionnaire (hypertension, dyslipidaemia, diabetes mellitus, cancer, chronic lung disease, liver disease, heart disease, stroke, kidney disease, stomach disease, emotional problems, memory-related disease, arthritis and asthma). Depression was defined as a score of 10 or higher on the 10-item Center for Epidemiologic Studies Depression Scale (CES-D-10); this cut-off has been validated as a screening threshold for clinically relevant depressive symptoms among older adults in general and among Chinese older adults specifically [15, 16]. Cognitive impairment was defined as a baseline cognitive composite score in the lowest sample quartile, where the composite is the sum of episodic-memory items (immediate and delayed 10-word recall) and mental-status items (serial-7 subtraction, date orientation and figure drawing) — a battery adapted from the United States Health and Retirement Study and previously used in CHARLS analyses of cognition [17]. From the three indicators we constructed two derived exposures. The first was the number of involved systems, an ordinal variable taking the values 0, 1, 2 or 3, used both as a continuous trend variable and as a four-level categorical variable (with 0 as the reference). The second was the cross-system pattern, a three-digit binary string in which the first, second and third digits encode physical multimorbidity, depression and cognitive impairment, respectively, yielding eight mutually exclusive combinations from "000" (no involved system) to "111" (all three involved systems). For benchmarking, we additionally used a previously derived binary indicator of physical-psychological-cognitive multimorbidity available in the CHARLS 2011 baseline data, hereafter referred to as the PPC-MM binary indicator (positive vs negative). Both thresholds are consistent with recent multicohort work on the same three domains, which likewise classified depressive symptoms in CHARLS at a CES-D-10 score of 10 or above but defined cognitive impairment relative to age-group-specific means rather than by a sample quartile; alternative cognitive thresholds were examined in sensitivity analyses [12]. The lowest sample quartile was preferred to a fixed score because the CHARLS composite has no established clinical cut-point in this population and its distribution is strongly shaped by educational attainment; fixed, quantile-based, age-standardised and age- and education-standardised alternatives were considered and are examined among the sensitivity analyses. The PPC-MM indicator is distributed with the CHARLS 2011 baseline release rather than derived by us, and the release does not document the algorithm behind it, so we characterised its behaviour empirically against the manually reconstructed components rather than assuming a published definition. Depressive symptoms at the CES-D-10 threshold are strictly necessary for a positive classification: every participant the indicator classifies as positive meets that threshold, and it is never positive in any combination that lacks depressive symptoms. Physical and cognitive involvement are not required in the same way, although the probability of a positive classification rises as they are added. Its physical component is less demanding than ours, nearly all positive participants having at least one chronic condition but only about two thirds the two that our definition requires, and its cognitive component does not coincide with our lowest-quartile cut-point, some positive participants scoring well above it. No simple deterministic rule reproduces the indicator: the closest reconstruction we could find, one or more chronic conditions together with depressive symptoms and cognitive impairment, agrees with it only moderately. It therefore marks the upper end of the cross-system gradient without corresponding to a transparent operationalisation of it, whereas the manually reconstructed score counts each domain separately and retains the eight-cell resolution on which the pattern analyses depend; only modest agreement between the two is to be expected by construction rather than indicating error in either. We retain the indicator as a sensitivity exposure precisely because it is independent of our own coding decisions: an association that survives the substitution cannot be an artefact of how we operationalised the three domains.
Outcome
The primary outcome was 9-year all-cause mortality. Vital status and date of death were ascertained at each follow-up wave (Waves 2 to 5) through proxy interview with household members or community informants when respondents could not be re-contacted. CHARLS records the year rather than the exact date of death, so time at risk was calculated in whole years from the 2011 Wave 1 baseline interview to the year of death or, for participants alive at the most recent successful contact, to the year of that contact (right-censoring); deaths recorded in the baseline year itself were credited with half a year of follow-up. Follow-up was administratively closed at the Wave 5 (2020) assessment, so that participants still under observation at that point contributed time at risk up to their last recorded contact within the window. The descriptor “9-year” therefore refers to the maximum administrative follow-up window from the 2011 baseline rather than to the duration of follow-up actually observed, which averaged 7.96 years (SD 2.23) because of censoring at last contact. A date of death was recorded for 483 of the 1,033 deaths. For the remaining 550 the fact of death was ascertained at a follow-up wave but no date was reported; rather than exclude these events, or fix them at a single arbitrary point, the time of death was imputed by stratified hot-deck sampling. Within each stratum defined by the wave at which the participant was last observed alive, a time was drawn from the death times actually observed within that stratum, independently in each of the 50 imputed datasets, so that the resulting uncertainty is carried into the confidence intervals by Rubin’s rules. Every value assigned in this way is an observed death time contributed by a participant last seen at the same wave, and it necessarily falls after that participant’s last contact. Assigning the midpoint of the interval as a single deterministic value is the more usual treatment of such deaths, but a single point discards the uncertainty within the interval, and the two- to three-year spacing of the CHARLS waves is wide enough for that to matter; the midpoint rule and three further deterministic rules are nevertheless reported for comparison among the sensitivity analyses. The consequences of this procedure for the estimates, together with a nine-year risk ratio that requires no death time at all, are reported with the results.
Covariates
Covariates were grouped into three levels of progressive adjustment. The demographic level comprised age (continuous, in years) and sex (male or female). The socioeconomic level additionally comprised the highest educational attainment (illiterate, primary, middle, or high school and above), marital status (married or cohabiting versus other), hukou registration (agricultural versus non-agricultural), residence (rural versus urban) and broad geographic region (south versus north of the Qinling–Huaihe line). The lifestyle and biomarker level additionally comprised smoking status (never, former or current), past-year drinking (no versus yes), body mass index (BMI) category and measured hypertension. BMI categories (underweight, normal weight, overweight, obese) were defined using the Chinese-population-specific cut-offs of 18.5, 24.0 and 28.0 kg/m2recommended by the Working Group on Obesity in China [18]. Measured hypertension was defined as a mean systolic blood pressure of 140 mmHg or higher, or a mean diastolic blood pressure of 90 mmHg or higher, calculated from the standardised CHARLS triplicate seated readings, consistent with the diagnostic threshold of the 2018 Chinese guidelines for the prevention and treatment of hypertension [19]. Self-rated health (good, fair, poor) and weekly physical activity in metabolic-equivalent minutes per week (MET-min/week) were also extracted; given their high missingness (50.6% and 72.1%, respectively), they were retained for sensitivity analyses only and were not included in the main covariate set. The target estimand is a descriptive prognostic association rather than a causal effect: we ask how strongly cross-system status at the 2011 baseline predicts death over the following nine years given the measured covariate set, not what would follow from intervening on that status. Covariates were accordingly selected on substantive grounds—baseline characteristics that are antecedent to or contemporaneous with the exposure and that are also associated with mortality—rather than by statistical criteria, and they are added in the three progressive blocks described below so that the sensitivity of the estimate to each block can be seen. The assumed structure is as follows. Age, sex, education, marital status, hukou registration, residence and region are treated as common causes of both cross-system status and mortality and are adjusted for throughout. Smoking, drinking, body mass index and measured hypertension are antecedent to the 2011 assessment for most participants but may be partly consequent on it for others; they are therefore adjusted for in the main model and removed again in sensitivity analyses, so that the consequence of either reading can be seen. Self-rated health, physical activity and night-time sleep duration are treated as descendants of the three domains and are kept out of the main model for that reason. We set out this reasoning rather than a formal diagram because the uncertainty here lies in the direction of a small number of edges rather than in which variables belong in the graph. Measured hypertension was retained even though physician-diagnosed hypertension is one of the 14 conditions contributing to physical multimorbidity, because the two capture different information—awareness of a diagnosis on the one hand and blood pressure at the baseline examination on the other—and because omitting the measured value would leave undiagnosed hypertension unaccounted for. Body mass index, smoking and measured hypertension may nevertheless lie partly on the pathway between cross-system multimorbidity and death, so the fully adjusted estimates should be read as adjusted prognostic associations rather than as effects purged of those pathways; the same reasoning applies with greater force to self-rated health and to physical activity, which are for that reason reported only as supportive analyses. Night-time sleep duration was excluded from the main covariate set on the same grounds. Short and long sleep are established correlates of mortality in this population, but they are more plausibly features of the three baseline domains than antecedents of them: the CES-D-10 that defines the psychological domain contains an item on restless sleep, and disturbed sleep commonly accompanies both physical multimorbidity and cognitive impairment, so that adjusting the main model for sleep duration would condition on part of the exposure itself. Sleep duration was examined instead as a sensitivity analysis. It was not entered into the imputation model, so those analyses are restricted to the participants for whom sleep was recorded and are compared with the same model fitted without sleep on the same participants, the two specifications differing in that covariate alone. No data-driven selection rule was applied to this set. A change-in-estimate criterion, which retains a covariate when its inclusion moves the exposure coefficient by more than a fixed percentage [20], is intended for candidates that are antecedent to the exposure: it registers that a covariate moves the estimate but not why, and cannot separate movement caused by the removal of confounding from movement caused by blocking part of the pathway from exposure to outcome [21]. In the present setting such a rule would therefore retain the very covariates whose adjustment is at issue, because a mediator moves the estimate more than a confounder of comparable strength, while discarding antecedent covariates whose confounding is real but numerically small. The covariate set was accordingly fixed on substantive grounds before any model was fitted, and the corresponding diagnostics are reported not as a selection procedure but as a bound on how much that choice matters: for each candidate covariate we report its assumed causal role, its association with mortality, and the change in the per-system coefficient produced by adding or removing it, so that the dependence of the estimate on the composition of the covariate set can be read directly.
Statistical analysis
Baseline characteristics were summarised across the four ordered groups defined by the number of involved systems (0/1/2/3) using means with standard deviations for continuous variables and counts with percentages for categorical variables, with between-group comparisons by analysis of variance and Pearson's chi-square test, respectively. These summaries are averaged across the 50 imputed datasets rather than taken from a single completed dataset, with the proportion of values missing before imputation reported for every variable, so that the descriptive table rests on the same basis as the models and does not depend on which completed dataset is inspected. A description confined to observed values is not available for the table as a whole, because the number of involved systems that defines its columns is itself imputed for the participants whose cognitive composite, depressive-symptom score or chronic-condition reports were incomplete. Crude 9-year cumulative incidence of all-cause mortality was visualised with Kaplan–Meier estimates by the number of involved systems, and inter-group equality of survival was tested using the multi-arm log-rank test. Variables with missing values among those used in the analytic models were handled by multiple imputation by chained equations, with 50 imputed datasets generated under fully conditional specification (20 iterations per chain; fixed random seed 20,260,723 for reproducibility), using predictive mean matching for continuous variables, logistic regression for binary variables, polytomous logistic regression for unordered multi-category variables, and proportional-odds (cumulative-link) regression for ordered multi-category variables (used for self-rated health in sensitivity analyses); composite exposures and derived covariates were re-derived inside each completed dataset rather than imputed directly [22]. The imputation model included the event indicator and the Nelson–Aalen estimate of the cumulative baseline hazard alongside all analysis variables, as is recommended when covariates are imputed for a time-to-event analysis, and the cognitive composite was imputed on its continuous scale with cognitive impairment derived from the imputed score within each completed dataset rather than the binary indicator being imputed directly. Of the variables with substantial missingness, only the cognitive composite (48.9%) enters the main analysis; physical activity and self-rated health, which are missing more often still, are confined to sensitivity analyses. A high proportion missing was not in itself taken as a reason to prefer complete-case analysis, because the bias of a multiply imputed analysis depends on how much information the observed data carry about the missing values rather than on the proportion missing [23]. The fraction of missing information is therefore reported with every estimate, and the 50 imputed datasets exceed the customary rule of one imputation for each percentage point of missing information, the largest value observed being 0.48 (Supplementary Table S8). Estimates from each imputed dataset were combined using Rubin's rules to produce pooled point estimates, standard errors and confidence intervals [24]. The primary association between cross-system multimorbidity and 9-year all-cause mortality was estimated using Cox proportional hazards regression [25], with three levels of progressive adjustment: M1 adjusted for age and sex; M2 additionally adjusted for the socioeconomic covariates; and M3 (the main model) additionally adjusted for smoking, drinking, BMI category and measured hypertension. Ties were handled by the Efron approximation, and the per-system estimate was reproduced under the Breslow approximation as a check on the discrete structure of wave-based follow-up. Every model was additionally refitted with a robust variance allowing for clustering within the 449 sampled communities. Sampling weights were not applied. After the exclusions described above, and after attrition across four subsequent waves, the analytic cohort no longer reproduces the CHARLS wave-1 sampling frame, and the wave-1 weights were not constructed to restore representativeness under that attrition; the estimates are therefore reported as unweighted associations within the assembled cohort rather than as weighted estimates for the national population of adults aged 45 years or older. The community-clustered variance accounts for the clustered structure of the design but does not confer representativeness. Three exposure operationalisations were fitted in parallel: the number of involved systems treated continuously to estimate a per-system trend hazard ratio; the same variable treated as a four-level categorical predictor (with the 0-system group as the reference); and the eight cross-system patterns as an unordered categorical predictor (with the "000" pattern as the reference). Sensitivity analyses included (i) M4, the M3 model further adjusted for MET-min/week; (ii) M3 further adjusted for self-rated health, reported as a supportive analysis given that self-rated health is plausibly downstream of cross-system multimorbidity rather than antecedent to it; (iii) the M3 model fitted on the complete-case subsample, without imputation, to verify that the imputation procedure did not artificially inflate the exposure–outcome association; and (iv) the M3 model with the PPC-MM binary indicator substituted for the manually constructed exposure; (v) the M3 model restricted to participants whose exposure was observed in all three domains; (vi) seven alternative operationalisations of the exposure, comprising a threshold of one rather than two chronic conditions, removal of the two conditions that overlap conceptually with the psychological and cognitive domains, and five alternative cognitive thresholds including quantile-based, fixed and age- and education-standardised cut-points (Table 4and Supplementary Table S11); (vii) a discrete-time complementary log–log model fitted on the four completed wave intervals, which does not require the proportional hazards assumption and respects the interval structure of wave-based follow-up; (viii) additional adjustment for wave-1 night-time sleep duration among the participants for whom it was recorded; and (ix) models weighted by the stabilised inverse probability of remaining uncensored, with the censoring model conditioned on the exposure and the baseline covariates. The proportional hazards assumption was assessed for the cross-system exposure itself and for every covariate in the M3 model using the Schoenfeld residual test [26]. Because the eight-combination model estimates more parameters than the summary-count models, we report the number of events per estimated parameter for each specification and refitted the eight-combination model with a ridge penalty whose tuning parameter was selected by ten-fold cross-validation within each imputed dataset. To complement the hazard ratios, 9-year absolute risks standardised to the covariate distribution of the whole cohort were derived from the fitted M3 model by the g-formula, with percentile confidence intervals from 2,000 bootstrap resamples distributed across the imputed datasets. The two combinations of principal interest, 101 and 111, were compared formally by estimating the ratio of their hazard ratios with a 95% confidence interval obtained from the pooled variance–covariance matrix. In a further model the three system indicators were entered simultaneously rather than as mutually exclusive combinations, so that each domain is adjusted for the other two. Multicollinearity was assessed using generalised variance inflation factors, which provide a comparable scale for predictors with more than one degree of freedom [27]. Hazard ratios are reported with 95% confidence intervals and two-sided p-values; statistical significance was set at α = 0.05. The proportional hazards test and the collinearity diagnostics were repeated in all 50 imputed datasets and are summarised as the mean test statistic with a pooled p-value and as the mean and maximum inflation factor, respectively; Kaplan–Meier curves were averaged across the 50 datasets and the log-rank test was computed separately within each. Four further analyses reported below are described here for completeness. First, to examine whether the eight combinations depart from what the three domains would imply acting independently, the eight-combination model was re-expressed as three main effects and their four interactions and the interaction terms were tested jointly; in the model with the three indicators entered simultaneously, the hypothesis that the three domain coefficients are equal was tested by a Wald test, since a summary count weights the three domains equally by construction. Secondly, the dependence of the estimate on the missing-at-random assumption was examined by shifting every imputed value of the cognitive composite by a fixed amount, up to two standard deviations in either direction, before cognitive impairment was derived, and by deterministic extreme-case bounds in which the same value was assigned to every participant with a missing item. Thirdly, the sensitivity of the principal estimates to unmeasured confounding was summarised by the E-value. Fourthly, discrimination was summarised by the concordance statistic of the eight-combination model, corrected for optimism by bootstrap resampling; it is reported as a description of discrimination only, the model having been neither calibrated nor validated in an external population. The Cox model was retained as the primary specification because it is the specification used in the studies against which these estimates are compared, and the discrete-time model, which does not require proportional hazards, is reported alongside it as a check rather than as a replacement. Data were extracted in R version 4.4.2; the analytic cohort was rebuilt and every estimate reported here was computed in R version 4.4.0.
Table 4.
Sensitivity analyses for cross-system multimorbidity and 9-year all-cause mortality
| Model | n | Deaths | Exposure | HR (95% CI) | p-value | FMI | Comment |
|---|---|---|---|---|---|---|---|
| Main M3 (multiply imputed) | 16,479 | 1033 | Number of involved systems | 1.34 (1.24–1.44) | < 0.001 | 0.21 | Reference analysis |
| Main M3, community-clustered robust variance | 16,479 | 1033 | Number of involved systems | 1.34 (1.24–1.44) | < 0.001 | 0.20 | Robust variance allowing for clustering within communities |
| M3 + physical activity (MET-min/week) | 16,479 | 1033 | Number of involved systems | 1.34 (1.24–1.45) | < 0.001 | 0.22 | Physical activity recorded for 27.9% of the cohort |
| M3 + self-rated health | 16,479 | 1033 | Number of involved systems | 1.25 (1.16–1.36) | < 0.001 | 0.26 | Recorded for 49.4%; plausibly downstream of the exposure, so supportive only |
| Complete-case M3 (no imputation) | 6885 | 341 | Number of involved systems | 1.29 (1.14–1.46) | < 0.001 | 0.00 | Restricted to participants with every model variable observed |
| Fully observed exposure (covariates imputed) | 8183 | 400 | Number of involved systems | 1.32 (1.18–1.48) | < 0.001 | 0.00 | Exposure observed in all three domains; covariates imputed |
| Deaths with observed timing only (drawn-time deaths censored) | 16,479 | 483 | Number of involved systems | 1.55 (1.38–1.74) | < 0.001 | 0.25 | Deaths without a recorded date censored at last known contact |
| Published inclusion rule, re-applied to corrected data | 13,178 | 410 | Number of involved systems | 1.61 (1.42–1.82) | < 0.001 | 0.21 | Isolates the effect of the inclusion rule used in the original analysis |
| PPC-MM preset binary indicator (vs negative) | 15,066 | 894 | PPC-MM indicator | 1.51 (1.22–1.88) | < 0.001 | 0.00 | Alternative exposure supplied with the CHARLS 2011 release |
| Cognitive cut-point 10.50 (the value used in the submitted paper) | 16,479 | 1033 | Number of involved systems | 1.34 (1.24–1.45) | < 0.001 | 0.19 | Alternative cognitive threshold |
| Cognitive impairment = lowest quintile | 16,479 | 1033 | Number of involved systems | 1.33 (1.23–1.44) | < 0.001 | 0.20 | Alternative cognitive threshold |
| Cognitive impairment = lowest decile | 16,479 | 1033 | Number of involved systems | 1.33 (1.23–1.44) | < 0.001 | 0.15 | Alternative cognitive threshold |
| Cognition continuous (per SD lower score), physical + depression counted | 16,479 | 1033 | Cognitive score (per SD) | 1.24 (1.12–1.37) | < 0.001 | 0.46 | No cognitive threshold applied |
| Discrete-time complementary log–log model on wave intervals | 16,479 | 1033 | Number of involved systems | 1.33 (1.24–1.44) | < 0.001 | 0.21 | Accounts for follow-up being observed at survey waves |
| Sleep-duration subsample, without sleep adjustment | 15,095 | 898 | Number of involved systems | 1.32 (1.22–1.43) | < 0.001 | 0.14 | Same-sample comparator for the sleep-adjusted model |
| Sleep-duration subsample, adjusted for sleep duration | 15,095 | 898 | Number of involved systems | 1.32 (1.22–1.43) | < 0.001 | 0.15 | Sleep duration entered in three categories |
Each row repeats the fully adjusted model under a different assumption or on a different sample, and reports the association per additional involved system unless the exposure column states otherwise. All models are fitted in each of the 50 imputed datasets and pooled by Rubin's rules except the complete-case, fully observed exposure and preset-indicator rows, which use observed data only. The number of deaths differs between rows because the underlying sample differs. The discrete-time row fits a complementary log–log model on the four completed wave intervals and is described in Supplementary Table S21; the two sleep rows are restricted to the 15,095 participants with a recorded sleep duration and differ from one another in that covariate alone, with the full set of sleep models in Supplementary Table S12. Estimates range from 1.24 to 1.61 across every specification, with the largest values arising when the analysis is restricted to deaths of known date or to the inclusion rule used in the original analysis
Abbreviations: CHARLS China Health and Retirement Longitudinal Study, CI Confidence interval, FMI Fraction of missing information, HR Hazard ratio, PPC-MM Physical, psychological and cognitive multimorbidity, SD standard deviation
Results
Cohort characteristics
Of 17,708 respondents who completed the CHARLS 2011 baseline interview, 16,479 (93.1%) entered the analytic cohort after exclusion of 419 respondents who were younger than 45 years, whose sex was missing or who could not be matched to a wave-1 record, and of 810 whose vital status was unknown and who were never contacted after the baseline interview (Fig. 1). Participants with incomplete follow-up were censored at their last known contact rather than excluded, and participants with missing exposure components were retained and multiply imputed rather than removed, so that the cohort is not selected on completeness of the exposure; comparisons of the participants included in and excluded from the originally submitted analysis and from the present cohort are given in Supplementary Tables S1 and S2. The mean cognitive composite among participants with the item recorded was 12.33 (SD 3.39) in the analytic cohort and 13.18 (3.22) among those excluded (standardised mean difference –0.26), whereas under the rule applied in the originally submitted analysis the participants excluded because the timing of death was missing had a mean of 10.94 (3.56) against 12.38 (3.35) among those retained (standardised mean difference 0.42). Item-level missingness rates across the analytic variables, ranging from 0% (age, sex and vital status) to 48.9% for the cognitive composite and 72.1% for physical activity, are summarised in Supplementary Table S3. The prevalence of each domain indicator among participants whose baseline item was observed differed from the imputed prevalence among those for whom it was missing, the largest difference being for cognitive impairment (25.4% versus 38.1%; Supplementary Table S4). Mean baseline age was 59.28 years (SD 9.78) and 8,469 (51.39%) were women. The number of involved systems at baseline was 0 in 5,406 (32.8%) participants, 1 in 5,802 (35.2%), 2 in 3,829 (23.2%) and 3 in 1,442 (8.8%); the mean cognitive composite fell from 14.03 (SD 2.04) in the no-system group to 7.19 (2.15) in the three-system group, and the mean CES-D-10 score rose from 4.19 (3.12) to 15.95 (5.31) across the same gradient. A higher number of involved systems was associated with older age, a greater female predominance, lower educational attainment, agricultural hukou registration, rural residence and a higher prevalence of every one of the 14 chronic conditions surveyed (all p < 0.001; Table 1). Geographic region and self-rated health also differed across the four groups (both p < 0.001), self-rated health deteriorating as the number of involved systems increased. The PPC-MM binary indicator was positive in 8.23% of the cohort overall, ranging from 0.00% in the no-system group to 41.65% in the three-system group when participants for whom the indicator was not recorded are counted as negative (Table 1); among the 15,066 participants for whom the indicator was recorded the corresponding proportions are 9.01% and 45.46% (Supplementary Table S5). Agreement between the indicator and the manually reconstructed score was fair (Cohen’s κ = 0.29, 95% CI 0.27–0.31), and the disagreement was almost entirely in one direction: 3,650 of the 4,849 participants with two or more involved systems on the manual reconstruction were classified as negative by the indicator, whereas 158 participants with fewer than two involved systems were classified as positive by it (Supplementary Table S5).
Fig. 1.
Study flow diagram. Notes: Flow of participants from the China Health and Retirement Longitudinal Study (CHARLS) 2011 national baseline through eligibility screening to the analytic cohort. Boxes on the right list each exclusion with the number of participants removed. The entry count is the published wave-1 national baseline; one respondent could not be matched to a wave-1 record in the analysis extract and is counted with the age and sex exclusions. Vital status was ascertained through the 2013, 2015, 2018 and 2020 waves. Participants were excluded only when vital status was unknown and no contact was recorded after baseline; all other participants with incomplete follow-up were censored at their last known contact rather than removed, so that every known death is retained. Participants with missing exposure components were retained and multiply imputed rather than excluded, so that the analytic cohort is not selected on completeness of the exposure. Abbreviations: CHARLS, China Health and Retirement Longitudinal Study
Table 1.
Baseline characteristics by number of cross-system multimorbidity systems
| Variable | Total (N = 16,479) | 0 systems (n = 5,406) | 1 system (n = 5,802) | 2 systems (n = 3,829) | 3 systems (n = 1,442) | p-value | Missing, % |
|---|---|---|---|---|---|---|---|
| Demographics | |||||||
| Age (years), mean (SD) | 59.28 (9.78) | 56.33 (8.65) | 59.44 (9.70) | 61.53 (10.08) | 63.74 (9.94) | < 0.001 | 0.00 |
| Age group, n (%) | < 0.001 | 0.00 | |||||
| 45–59 | 9209 (55.88) | 3723 (68.87) | 3187 (54.92) | 1761 (45.98) | 538 (37.34) | ||
| 60–74 | 5884 (35.71) | 1463 (27.06) | 2145 (36.97) | 1613 (42.12) | 663 (46.01) | ||
| 75 + | 1386 (8.41) | 220 (4.07) | 470 (8.10) | 456 (11.90) | 240 (16.65) | ||
| Sex, n (%) | < 0.001 | 0.00 | |||||
| Male | 8010 (48.61) | 3188 (58.97) | 2905 (50.07) | 1494 (39.02) | 423 (29.34) | ||
| Female | 8469 (51.39) | 2218 (41.03) | 2897 (49.93) | 2335 (60.98) | 1019 (70.66) | ||
| Marital status, n (%) | < 0.001 | 0.02 | |||||
| Married or cohabiting | 14,375 (87.23) | 5037 (93.18) | 5086 (87.66) | 3129 (81.71) | 1123 (77.88) | ||
| Other | 2104 (12.77) | 369 (6.82) | 716 (12.34) | 700 (18.29) | 319 (22.12) | ||
| Socioeconomic status | |||||||
| Education, n (%) | < 0.001 | 0.15 | |||||
| Illiterate | 4535 (27.52) | 654 (12.09) | 1546 (26.65) | 1498 (39.12) | 837 (58.08) | ||
| Primary | 6518 (39.55) | 2067 (38.23) | 2349 (40.48) | 1587 (41.44) | 516 (35.76) | ||
| Middle | 3396 (20.61) | 1599 (29.58) | 1199 (20.67) | 525 (13.70) | 73 (5.07) | ||
| High school or above | 2030 (12.32) | 1086 (20.10) | 708 (12.20) | 220 (5.74) | 16 (1.09) | ||
| Hukou, n (%) | < 0.001 | 0.09 | |||||
| Agricultural | 12,957 (78.63) | 3937 (72.84) | 4507 (77.68) | 3216 (83.98) | 1297 (89.95) | ||
| Non-agricultural | 3522 (21.37) | 1468 (27.16) | 1295 (22.32) | 614 (16.02) | 145 (10.05) | ||
| Residence, n (%) | < 0.001 | 0.00 | |||||
| Rural | 10,073 (61.13) | 2915 (53.92) | 3518 (60.64) | 2572 (67.17) | 1068 (74.06) | ||
| Urban | 6406 (38.87) | 2491 (46.08) | 2284 (39.36) | 1257 (32.83) | 374 (25.94) | ||
| Region, n (%) | < 0.001 | 0.00 | |||||
| South | 9140 (55.46) | 2876 (53.21) | 3207 (55.27) | 2191 (57.21) | 866 (60.07) | ||
| North | 7339 (44.54) | 2529 (46.79) | 2595 (44.73) | 1638 (42.79) | 576 (39.93) | ||
| Lifestyle | |||||||
| Smoking, n (%) | < 0.001 | 0.55 | |||||
| Never | 9862 (59.84) | 2961 (54.77) | 3418 (58.91) | 2484 (64.86) | 999 (69.31) | ||
| Former | 1947 (11.82) | 654 (12.09) | 726 (12.51) | 417 (10.89) | 151 (10.49) | ||
| Current | 4670 (28.34) | 1791 (33.14) | 1659 (28.59) | 929 (24.25) | 291 (20.19) | ||
| Drinking, n (%) | < 0.001 | 0.64 | |||||
| No | 9612 (58.33) | 2842 (52.57) | 3387 (58.38) | 2405 (62.81) | 977 (67.76) | ||
| Yes | 6867 (41.67) | 2564 (47.43) | 2415 (41.62) | 1424 (37.19) | 465 (32.24) | ||
| Physical activity (MET-min/week), mean (SD) | 6959.8 (7478.7) | 7159.2 (7528.5) | 6938.6 (7439.8) | 6853.1 (7487.3) | 6582.1 (7399.6) | 0.017 | 72.09 |
| Anthropometry and blood pressure | |||||||
| Body mass index (kg/m2), mean (SD) | 23.54 (3.91) | 23.71 (3.68) | 23.64 (4.04) | 23.31 (4.01) | 23.12 (3.94) | < 0.001 | 22.07 |
| Body mass index category, n (%) | < 0.001 | 22.07 | |||||
| Underweight | 1094 (6.64) | 238 (4.40) | 378 (6.52) | 335 (8.74) | 144 (9.98) | ||
| Normal | 8645 (52.46) | 2887 (53.41) | 3001 (51.72) | 2014 (52.60) | 742 (51.49) | ||
| Overweight | 4813 (29.21) | 1695 (31.35) | 1692 (29.17) | 1034 (27.00) | 392 (27.19) | ||
| Obese | 1927 (11.69) | 586 (10.84) | 731 (12.59) | 447 (11.66) | 163 (11.34) | ||
| Systolic blood pressure (mmHg), mean (SD) | 130.84 (21.73) | 127.66 (19.57) | 131.05 (21.68) | 133.02 (23.06) | 136.16 (24.08) | < 0.001 | 20.86 |
| Diastolic blood pressure (mmHg), mean (SD) | 76.07 (12.21) | 75.99 (11.74) | 75.96 (12.31) | 76.13 (12.56) | 76.68 (12.59) | 0.136 | 20.86 |
| Measured hypertension, n (%) | < 0.001 | 20.86 | |||||
| No | 11,389 (69.11) | 4070 (75.30) | 3964 (68.31) | 2487 (64.94) | 868 (60.23) | ||
| Yes | 5090 (30.89) | 1335 (24.70) | 1839 (31.69) | 1342 (35.06) | 573 (39.77) | ||
| Self-rated health | |||||||
| Self-rated health, n (%) | < 0.001 | 50.60 | |||||
| Good | 4201 (25.49) | 1809 (33.47) | 1441 (24.83) | 731 (19.08) | 221 (15.30) | ||
| Fair | 7538 (45.74) | 2807 (51.93) | 2834 (48.85) | 1490 (38.90) | 407 (28.22) | ||
| Poor | 4740 (28.76) | 789 (14.60) | 1527 (26.32) | 1609 (42.02) | 814 (56.49) | ||
| Cross-system components | |||||||
| Number of chronic diseases, mean (SD) | 1.39 (1.40) | 0.44 (0.50) | 1.41 (1.30) | 2.11 (1.48) | 2.98 (1.22) | < 0.001 | |
| Cognitive composite, mean (SD) | 11.76 (3.58) | 14.03 (2.04) | 11.93 (3.36) | 10.03 (3.49) | 7.19 (2.15) | < 0.001 | 48.92 |
| CES-D-10 score, mean (SD) | 8.40 (6.32) | 4.19 (3.12) | 7.49 (5.40) | 12.88 (5.98) | 15.95 (5.31) | < 0.001 | 8.57 |
| Physical multimorbidity (≥ 2 of 14), n (%) | 6358 (38.59) | 0 (0.00) | 2400 (41.37) | 2517 (65.72) | 1442 (100.00) | - | 1.06 |
| Depressive symptoms (CES-D-10 ≥ 10), n (%) | 6223 (37.76) | 0 (0.00) | 1730 (29.82) | 3051 (79.68) | 1442 (100.00) | - | 10.13 |
| Cognitive impairment (composite ≤ 10.00), n (%) | 5205 (31.58) | 0 (0.00) | 1672 (28.82) | 2091 (54.60) | 1442 (100.00) | - | 48.92 |
| Individual chronic conditions | |||||||
| Hypertension, n (%) | 4143 (25.14) | 567 (10.48) | 1575 (27.15) | 1337 (34.92) | 664 (46.03) | < 0.001 | 1.06 |
| Dyslipidemia, n (%) | 1548 (9.39) | 126 (2.33) | 687 (11.85) | 508 (13.26) | 227 (15.71) | < 0.001 | 2.54 |
| Diabetes, n (%) | 939 (5.70) | 72 (1.34) | 391 (6.74) | 317 (8.27) | 160 (11.07) | < 0.001 | 1.44 |
| Cancer, n (%) | 146 (0.88) | 20 (0.36) | 55 (0.96) | 50 (1.31) | 20 (1.40) | < 0.001 | 1.00 |
| Chronic lung diseases, n (%) | 1725 (10.47) | 126 (2.34) | 569 (9.81) | 629 (16.42) | 401 (27.80) | < 0.001 | 0.93 |
| Liver disease, n (%) | 556 (3.37) | 52 (0.96) | 217 (3.75) | 196 (5.13) | 90 (6.24) | < 0.001 | 1.27 |
| Heart disease, n (%) | 1927 (11.70) | 109 (2.01) | 681 (11.74) | 734 (19.16) | 404 (28.02) | < 0.001 | 1.10 |
| Stroke, n (%) | 401 (2.43) | 9 (0.16) | 123 (2.12) | 173 (4.52) | 96 (6.68) | < 0.001 | 0.78 |
| Kidney disease, n (%) | 923 (5.60) | 63 (1.16) | 314 (5.40) | 347 (9.07) | 199 (13.81) | < 0.001 | 1.18 |
| Stomach disease, n (%) | 3816 (23.16) | 495 (9.16) | 1271 (21.90) | 1330 (34.74) | 720 (49.94) | < 0.001 | 0.82 |
| Emotional problems, n (%) | 187 (1.14) | 14 (0.25) | 52 (0.89) | 70 (1.83) | 52 (3.60) | < 0.001 | 0.98 |
| Memory-related disease, n (%) | 255 (1.55) | 9 (0.17) | 67 (1.15) | 105 (2.75) | 74 (5.12) | < 0.001 | 0.86 |
| Arthritis, n (%) | 5724 (34.73) | 722 (13.36) | 1976 (34.05) | 2014 (52.60) | 1012 (70.20) | < 0.001 | 0.76 |
| Asthma, n (%) | 668 (4.06) | 17 (0.32) | 210 (3.63) | 257 (6.70) | 184 (12.76) | < 0.001 | 0.94 |
| Reference exposure | |||||||
| PPC-MM preset indicator positive, n (%) | 1357 (8.23) | 0 (0.00) | 158 (2.73) | 598 (15.62) | 600 (41.65) | 8.57 | |
| Outcome | |||||||
| Death within 9 years, n (%) | 1033 (6.27) | 182 (3.36) | 343 (5.90) | 321 (8.37) | 188 (13.04) | 0.00 | |
| Follow-up time (years), mean (SD) | 7.96 (2.23) | 8.22 (1.90) | 7.98 (2.21) | 7.75 (2.44) | 7.42 (2.70) | < 0.001 | |
Baseline characteristics of the 16,479 participants in the analytic cohort. Continuous variables are mean (SD) and categorical variables n (%). Every summary was computed separately within each of the 50 imputed datasets and then averaged across them, so that the table describes the whole analytic cohort on the same basis as the models rather than one arbitrary completed dataset; counts are therefore averages that have been rounded to the nearest integer, so that the four group counts in a row need not sum exactly to the total column—the group death counts, for example, sum to 1,034 against an observed total of 1,033—and percentages use the cohort or group total as the denominator. The final column gives the proportion of values missing before imputation for each variable, so that the reader can see directly how much of each entry rests on imputed values; it is left blank for the number of chronic diseases and for follow-up time, which are derived within each completed dataset rather than recorded as single items. A description confined to observed values is not available for the table as a whole, because the number of involved systems that defines the columns is itself imputed for the participants whose cognitive composite, depressive-symptom score or chronic-condition reports were incomplete; the corresponding observed-data description of the cohort is given in Supplementary Tables S2 to S4. Two rows rest on observed data: the total column of the PPC-MM row counts the recorded preset indicator with participants whose indicator was not recorded treated as negative, so that 8.23% of the whole cohort corresponds to 9.01% of the 15,066 participants for whom the indicator was recorded (Supplementary Table S5), and the total number of deaths is the observed count. P-values compare the four groups defined by the number of involved systems, from one-way analysis of variance for continuous variables and χ2 tests for categorical variables, computed within each imputed dataset and summarised as the median across imputations. Because the cognitive composite was missing for 48.9% of participants, its mean including imputed values (11.76) is lower than the mean of 12.33 among participants with the item recorded (Sect. 3.1 and Supplementary Table S2), and the prevalence of cognitive impairment including imputed values (31.58%) is correspondingly higher than the observed prevalence of 25.4% (Supplementary Table S4); this is the expected behaviour under the missing-at-random assumption. Missing values were addressed by multiple imputation (50 imputed datasets) in every analysis rather than by exclusion, so participants whose exposure components were incompletely recorded remain in the cohort. Restricting the fully adjusted model to the 8,183 participants whose exposure was observed in all three domains gave a hazard ratio of 1.32 (95% CI 1.18–1.48) per additional involved system, against 1.34 (1.24–1.44) in the imputed analysis (Table 4). Cognitive impairment was defined by a cognitive composite score at or below 10.00, the lowest quartile of the observed distribution
Abbreviations: CES-D-10 10-item Center for Epidemiologic Studies Depression Scale, CI Confidence interval, PPC-MM Physical, psychological and cognitive multimorbidity, SD Standard deviation
Crude mortality
Over a mean follow-up of 7.96 years (SD 2.23), with the distribution negatively skewed because the majority of participants were alive at the most recent successful contact within the 9-year administrative window, 1,033 (6.27%) participants died from any cause. Crude 9-year mortality rose monotonically with the number of involved systems, from 182 deaths (3.36%) in the no-system group to 343 (5.90%), 321 (8.37%) and 188 (13.04%) in the one-, two- and three-system groups, respectively; the corresponding 9-year cumulative incidences were 3.76%, 6.50%, 9.25% and 14.56% (Fig. 2). Numbers of participants and of deaths within exposure groups are averaged across the 50 imputed datasets, so the group totals need not sum exactly to the cohort total or to the sum of the corresponding combination-specific counts in Table 3. Kaplan–Meier cumulative-incidence curves separated early and continued to diverge throughout follow-up without crossing, with the three-system curve rising visibly steeper than the others (multi-arm log-rank p < 0.001 in every imputed dataset; Fig. 2).
Fig. 2.
Cumulative incidence of all-cause mortality by number of cross-system multimorbidity systems. Notes: Cumulative incidence of all-cause mortality by the number of involved systems at the 2011 baseline, averaged over the 50 imputed datasets. Each step function is shown with its point-wise 95% confidence interval as a shaded band of the same colour, obtained by pooling the complementary log–log transform of the cumulative hazard across imputations by Rubin's rules; the band begins at the first death within each group, before which the interval is undefined. Crosses mark censored observations and are drawn only for participants whose exposure was identical in all 50 imputed datasets, so that no marker depends on which imputation is inspected. The log-rank test was computed separately in each imputed dataset. The panel beneath the plot gives the number of participants still under follow-up at each year, averaged over imputations, with row labels coloured to match the strata. Because follow-up in CHARLS is observed at survey waves rather than continuously, the curves rise in steps at whole years and the numbers at risk change little between adjacent waves; a discrete-time model fitted on wave intervals is reported in Supplementary Table S21 and reproduces the estimate obtained from the Cox model. Abbreviations: CHARLS, China Health and Retirement Longitudinal Study
Table 3.
Cross-system multimorbidity combinations and 9-year all-cause mortality (M3 model)
| Combination (physical | depression | cognitive) | n | Deaths | Crude mortality, % | Rate per 1,000 person-years | Unadjusted 9-year risk, % | Standardised 9-year risk, % | Risk difference vs 000, pp | HR (95% CI) | p-value | HR (95% CI), clustered | p-value, clustered | FMI |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 000 — none | 5406 | 182 | 3.36 | 4.09 | 3.76 | 5.42 | + 0.00 | 1.00 (reference) | - | 1.00 (reference) | - | - |
| 100 — physical only | 2400 | 149 | 6.22 | 7.94 | 6.86 | 7.94 | + 2.52 | 1.54 (1.20–1.96) | < 0.001 | 1.54 (1.20–1.96) | < 0.001 | 0.20 |
| 010 — depression only | 1730 | 68 | 3.95 | 4.83 | 4.30 | 6.03 | + 0.61 | 1.12 (0.81–1.57) | 0.488 | 1.12 (0.80–1.58) | 0.497 | 0.28 |
| 001 — cognitive only | 1672 | 125 | 7.47 | 9.36 | 8.26 | 6.75 | + 1.33 | 1.28 (0.95–1.73) | 0.110 | 1.28 (0.95–1.72) | 0.104 | 0.35 |
| 110 — physical + depression | 1738 | 109 | 6.26 | 8.00 | 6.94 | 8.09 | + 2.67 | 1.57 (1.19–2.07) | 0.001 | 1.57 (1.19–2.07) | 0.001 | 0.23 |
| 101 — physical + cognitive | 778 | 97 | 12.51 | 16.54 | 14.11 | 10.54 | + 5.12 | 2.14 (1.57–2.91) | < 0.001 | 2.14 (1.58–2.90) | < 0.001 | 0.30 |
| 011 — depression + cognitive | 1313 | 114 | 8.71 | 11.22 | 9.47 | 7.65 | + 2.24 | 1.47 (1.09–1.99) | 0.012 | 1.47 (1.08–2.00) | 0.014 | 0.29 |
| 111 — all three | 1442 | 188 | 13.04 | 17.57 | 14.57 | 11.77 | + 6.36 | 2.45 (1.89–3.17) | < 0.001 | 2.45 (1.89–3.17) | < 0.001 | 0.26 |
The eight mutually exclusive combinations of physical multimorbidity, depressive symptoms and cognitive impairment, coded as a three-digit string in that order, so that 101 denotes physical multimorbidity and cognitive impairment without depressive symptoms. Numbers of participants and deaths are averaged over the 50 imputed datasets. Crude mortality is the observed proportion of deaths and the rate per 1,000 person-years uses the follow-up time accumulated by each group. The standardised 9-year risk is obtained by assigning every participant in turn to each combination, predicting the risk from the fully adjusted model and averaging over the observed covariate distribution, so that the risk difference against the no-system combination is adjusted for the covariates; the unadjusted risk is the corresponding unadjusted quantity. Hazard ratios are from the fully adjusted model with the no-system combination as reference, and the clustered columns use a robust variance allowing for clustering within communities. A formal contrast of the physical–cognitive combination against the all-three combination gave a ratio of hazard ratios of 0.87 (95% CI 0.66–1.16; p = 0.35): the two combinations are not statistically distinguishable, and their resemblance is therefore descriptive rather than a demonstration of equivalence
Abbreviations: CI Confidence interval, FMI Fraction of missing information, HR Hazard ratio, pp Percentage points
Dose–response: number of involved systems
In pooled Cox proportional hazards models, the per-system hazard ratio for 9-year all-cause mortality was 1.38 (95% CI 1.28–1.48; p < 0.001) in the age- and sex-adjusted model (M1), 1.35 (95% CI 1.25–1.46; p < 0.001) after additional adjustment for socioeconomic covariates (M2), and 1.34 (95% CI 1.24–1.44; p < 0.001) in the fully adjusted model that further accounted for smoking, drinking, BMI category and measured hypertension (M3). Estimates were therefore essentially unchanged across the three layers of adjustment. Treating the number of involved systems as an ordered category, fully adjusted hazard ratios versus the no-system reference were 1.34 (95% CI 1.09–1.66; p = 0.006) for one system, 1.68 (95% CI 1.35–2.09; p < 0.001) for two systems and 2.48 (95% CI 1.93–3.17; p < 0.001) for three systems (Table 2). Repeating the same models with a robust variance that allows for clustering within the 449 sampled communities left every estimate unchanged (Table 2). On the log10 axis, these four ordered estimates fell on an approximately linear path (Fig. 3). Among M3 covariates, hazards were elevated for older age (HR 1.09 per year, 95% CI 1.08–1.10; p < 0.001), not being married or cohabiting (HR 1.50, 95% CI 1.29–1.74; p < 0.001) and current or former smoking, with the hazard for former smokers (HR 1.55, 95% CI 1.28–1.89; p < 0.001) numerically exceeding that for current smokers (HR 1.25, 95% CI 1.05–1.49; p = 0.013). Female sex was associated with a lower hazard (HR 0.66, 95% CI 0.55–0.78; p < 0.001). In this larger cohort, education to high school or above (HR 0.66, 95% CI 0.48–0.90), residence in the northern region (HR 1.23, 95% CI 1.09–1.40) and measured hypertension (HR 1.37, 95% CI 1.18–1.58) were also independently associated with mortality, whereas hukou registration, urban residence and drinking were not; the full set of covariate estimates is reported in Supplementary Table S6 rather than in Table 2.
Table 2.
Cox proportional hazards models for cross-system multimorbidity and 9-year all-cause mortality
| Term | Model | HR (95% CI) | p-value | HR (95% CI), community-clustered | p-value, clustered | FMI |
|---|---|---|---|---|---|---|
| Per 1-system increase | M1: age, sex | 1.38 (1.28–1.48) | < 0.001 | 1.38 (1.27–1.49) | < 0.001 | 0.20 |
| Per 1-system increase | M2: M1 + socioeconomic factors | 1.35 (1.25–1.46) | < 0.001 | 1.35 (1.25–1.46) | < 0.001 | 0.21 |
| Per 1-system increase | M3: M2 + lifestyle and clinical factors | 1.34 (1.24–1.44) | < 0.001 | 1.34 (1.24–1.44) | < 0.001 | 0.21 |
| 0 systems (reference) | M3 | 1.00 (reference) | - | 1.00 (reference) | - | - |
| 1 system | M3 | 1.34 (1.09–1.66) | 0.006 | 1.34 (1.09–1.66) | 0.006 | 0.25 |
| 2 systems | M3 | 1.68 (1.35–2.09) | < 0.001 | 1.68 (1.35–2.09) | < 0.001 | 0.22 |
| 3 systems | M3 | 2.48 (1.93–3.17) | < 0.001 | 2.48 (1.93–3.17) | < 0.001 | 0.21 |
Hazard ratios for 9-year all-cause mortality from Cox proportional hazards models fitted in each of the 50 imputed datasets and pooled by Rubin's rules. M1 is adjusted for age and sex; M2 additionally for education, marital status, hukou registration, residence and geographic region; M3 additionally for smoking, drinking, body mass index category and measured hypertension. The upper block treats the exposure as a continuous count of involved systems and the lower block as four ordered categories, both from the fully adjusted model. The community-clustered columns repeat each model with a robust variance that allows for clustering within the 449 sampled communities; point estimates are unchanged, so the association is not an artefact of the survey design. Coefficients for the covariates themselves are reported in Supplementary Table S6 rather than here, because they are conditional associations within the model and not effect estimates for those covariates. FMI is the fraction of missing information, which quantifies how much each estimate depends on the imputed values
Abbreviations: CI Confidence interval, FMI Fraction of missing information, HR Hazard ratio
Fig. 3.
Dose–response between number of cross-system multimorbidity systems and 9-year all-cause mortality. Notes: Adjusted hazard ratios with 95% confidence intervals for each ordered category of the number of involved systems compared with the no-system reference, from the fully adjusted model (M3 in Table 2) pooled across the 50 imputed datasets. The y-axis is on the log10 scale and the horizontal line marks the null. The dashed line with its shaded band is the fitted per-system linear trend from the same model with its 95% confidence interval. The near-linear increase on the log scale supports treating the number of involved systems as an ordinal exposure with an approximately constant increment in mortality risk per additional system
Cross-system patterns
The eight-combination analysis revealed a markedly heterogeneous risk landscape across the cells of the same overall system count (Table 3, Fig. 4). At one involved system, isolated physical multimorbidity (100) was associated with elevated mortality (HR 1.54, 95% CI 1.20–1.96; p < 0.001), whereas isolated cognitive impairment (001) gave a positive but imprecise estimate (HR 1.28, 95% CI 0.95–1.73; p = 0.110), and isolated depression (010) yielded a point estimate of HR 1.12 (95% CI 0.81–1.57; p = 0.488), the wide confidence interval—based on 68 deaths—being compatible with both no association and a clinically meaningful effect; we therefore did not find sufficient evidence for an independent association between isolated depression and 9-year all-cause mortality. At two involved systems, the co-occurrence of physical multimorbidity and depression (110; HR 1.57, 95% CI 1.19–2.07; p = 0.001) and the co-occurrence of depression and cognitive impairment (011; HR 1.47, 95% CI 1.09–1.99; p = 0.012) were of intermediate magnitude, while the co-occurrence of physical multimorbidity and cognitive impairment (101) carried a substantially higher hazard (HR 2.14, 95% CI 1.57–2.91; p < 0.001). Notably, this two-system 101 estimate closely approached the hazard observed for the three-system combination (111; HR 2.45, 95% CI 1.89–3.17; p < 0.001), and the corresponding 9-year crude death rates—12.51% (97/778) and 13.04% (188/1,442)—were virtually identical. Thus, on the additive scale of cumulative incidence and on the multiplicative scale of the hazard ratio, the addition of depression to a substrate of co-occurring physical multimorbidity and cognitive impairment was associated with little further increment in 9-year mortality risk. A formal contrast confirmed that these two combinations could not be statistically distinguished (ratio of hazard ratios 0.87, 95% CI 0.66–1.16; p = 0.35). Penalised estimation of the eight-combination model returned hazard ratios close to the maximum-likelihood values (Supplementary Table S7). With 1,033 deaths, the fully adjusted models estimated 17, 19 and 23 parameters in the continuous, four-level and eight-combination specifications respectively, corresponding to 60.8, 54.4 and 44.9 events per estimated parameter.
Fig. 4.
Adjusted hazard ratios for 9-year all-cause mortality by all 8 cross-system multimorbidity combinations. Notes: Tabular forest plot of the eight mutually exclusive cross-system combinations from the fully adjusted model (M3 in Table 2), pooled across the 50 imputed datasets, with the no-system combination as reference. Combination labels read as a three-digit string in which the first digit denotes physical multimorbidity, the second depressive symptoms and the third cognitive impairment; 101 therefore denotes physical multimorbidity and cognitive impairment without depressive symptoms. Each row gives the combination, the number of participants, the number and percentage of deaths during the 9-year follow-up, the point estimate as a filled diamond with its 95% confidence interval, the numerical hazard ratio and the p-value. The horizontal axis is on the log10 scale and the dashed vertical line marks the null. Numbers of participants and deaths are averaged over the imputed datasets. A formal contrast of the physical–cognitive combination against the all-three combination gave a ratio of hazard ratios of 0.87 (95% CI 0.66–1.16; p = 0.35), so their resemblance is descriptive rather than a demonstration of equivalence. Abbreviations: CI, confidence interval; HR, hazard ratio
Structure of the cross-system exposure
When the three system indicators were entered simultaneously rather than as mutually exclusive combinations, the hazard ratio was 1.58 (95% CI 1.38–1.80) for physical multimorbidity, 1.11 (95% CI 0.96–1.28; p = 0.157) for depressive symptoms and 1.39 (95% CI 1.14–1.70) for cognitive impairment (Supplementary Table S8). Re-expressing the eight-combination model as three main effects and their interactions gave no evidence of departure from additivity on the multiplicative scale (joint test of the four interaction terms p = 0.89), the observed hazard ratio for every combination lying within 10% of the additively predicted value; the increment attributable to depressive symptoms was at most 15% on any background, against 66% for physical multimorbidity and 56% for cognitive impairment (Supplementary Table S9). The hypothesis that the three domain coefficients are equal was rejected (p = 0.017), the difference lying between the physical and the psychological domain rather than between the physical and the cognitive one (Supplementary Table S10).
Sensitivity analyses
Fully adjusted estimates were stable across the sensitivity analyses (Table 4). Adding physical activity (MET-min/week) to M3 left the per-system hazard ratio essentially unchanged (HR 1.34, 95% CI 1.24–1.45). A complete-case analysis on the 6,885 participants with no missing data (341 deaths) gave 1.29 (95% CI 1.14–1.46; p < 0.001), and restricting the analysis to the 8,183 participants whose exposure was observed in all three domains gave 1.32 (95% CI 1.18–1.48; p < 0.001). Replacing the manually reconstructed exposure with the PPC-MM binary indicator gave 1.51 (95% CI 1.22–1.88; p < 0.001). Adding self-rated health to M3 attenuated the per-system hazard ratio to 1.25 (95% CI 1.16–1.36; p < 0.001). Two rows of Table 4 depart from the main estimate: restricting the outcome to the 483 deaths with a directly recorded date, and censoring the remaining deaths at last known contact, gave 1.55 (95% CI 1.38–1.74), and re-applying the inclusion rule used in the originally submitted analysis to the corrected data reduced the cohort to 13,178 participants with 410 deaths and gave 1.61 (95% CI 1.42–1.82). Seven alternative operationalisations of the exposure, comprising a threshold of one rather than two chronic conditions, removal of the two conditions that overlap conceptually with the psychological and cognitive domains, and five alternative cognitive thresholds including age- and education-standardised cut-points, left the ordering of the eight combinations and the position of the physical–cognitive combination unchanged (Table 4 and Supplementary Table S11). Among the 15,095 participants for whom night-time sleep duration was recorded, the per-system hazard ratio was 1.32 (95% CI 1.22–1.43) with sleep omitted from the model and between 1.32 and 1.34 with sleep entered continuously, in three categories, or as total sleep including napping (Supplementary Table S12); the ordering of the eight combinations was unchanged by that adjustment (Supplementary Table S13). Weighting by the stabilised inverse probability of remaining uncensored gave estimates close to the unweighted ones (Supplementary Table S14). Beyond age and sex, no covariate retained in M3 changed the per-system log hazard ratio by more than 5% when added to an age- and sex-adjusted model, the largest change being 4.9% for education, whereas self-rated health, excluded in advance as a downstream variable, changed it by 20% (Supplementary Table S15). Removing measured hypertension, body mass index, or smoking and alcohol use from M3 changed the estimate by 2.5%, 0.2% and 2.2% respectively, and omitting the entire lifestyle and clinical block changed it by 4.6%; rebuilding the physical domain from the 13 chronic conditions that remain once doctor-diagnosed hypertension is removed gave 1.33 (95% CI 1.23–1.43), and removing hypertension from the exposure and from the covariate set simultaneously gave 1.32 (95% CI 1.23–1.43), the ratio of the hazard ratios for the 101 and 111 combinations being 0.90 when measured hypertension alone was removed from the covariate set and 0.89 when the physical domain was rebuilt from the 13 conditions with the full covariate set retained, against 0.87 in the main one (Supplementary Table S16). Shifting every imputed value of the cognitive composite by up to two standard deviations before cognitive impairment was derived, which moves the prevalence of impairment among participants with the item missing between 2 and 96%, left the per-system hazard ratio between 1.27 and 1.34, and no confidence interval included 1 (Supplementary Table S17); deterministic extreme-case bounds, obtained by assigning the same value to every participant with a missing item, spanned 1.27 to 1.37 (Supplementary Table S18). The E-value for the per-system estimate was 2.01, and 1.78 for the limit of its confidence interval; for the 111 combination the corresponding values were 4.33 and 3.19 (Supplementary Table S19).
Model assumptions and diagnostics
The proportional hazards assumption was supported for every covariate in the M3 model and for the model as a whole (all covariate-specific Schoenfeld p ≥ 0.082; global test p = 0.257), but not for the cross-system exposure itself (χ2 = 26.57, df = 7, p = 0.010; Supplementary Table S20). A discrete-time complementary log–log model fitted on the four completed wave intervals, which does not require proportional hazards, gave a per-system estimate of 1.33 (95% CI 1.24–1.44; Supplementary Table S21). Multicollinearity diagnostics were within acceptable limits, the largest generalised variance inflation factor being 1.50, for sex (Supplementary Table S22). The concordance statistic of the eight-combination model was 0.812 as fitted and 0.809 after bootstrap correction for optimism. Multiple imputation diagnostics indicated convergence of the 50 chains within 20 iterations without systematic drift (Supplementary Figure S1), and kernel-density (Supplementary Figure S2) and strip-plot (Supplementary Figure S3) comparisons of observed and imputed values for the continuous imputed variables (BMI, systolic and diastolic blood pressure, MET-min/week) showed close overlap.
Ascertainment of the date of death
A date of death was recorded for 483 of the 1,033 deaths and imputed for the remaining 550, and the two sources were unevenly distributed across follow-up: a date was recorded for 402 of the 436 deaths in the first two years and imputed for 345 of the 388 deaths in the last two, the per-system estimate falling from 1.50 in the first interval to 1.19 in the last (Supplementary Table S23). Estimated separately, the two sets gave a per-system hazard ratio of 1.55 (95% CI 1.38–1.74) from the deaths with a recorded date and 1.17 (95% CI 1.06–1.29) from those with an imputed one, with no heterogeneity across the four intervals in either set (I2 = 0.0%; Supplementary Table S24). Of the 483 deaths with a recorded date, 409 occurred among participants last observed at baseline; stratifying the baseline hazard on the wave of last contact moved the ratio of the two estimates from 0.76 (95% CI 0.65–0.88) to 0.87 (95% CI 0.75–1.02), three of the four within-stratum comparisons included 1, and with age as the time scale the ratio was 0.75 (95% CI 0.63–0.88) (Supplementary Table S25). Replacing the 50 draws by any of four deterministic placements, including the midpoint rule and the administrative end of follow-up, moved the per-system estimate by no more than 0.01 (Supplementary Table S26). A nine-year risk ratio that uses no death time of any kind was 1.27 (95% CI 1.19–1.37) per involved system in the whole cohort and 1.30 (95% CI 1.21–1.39) among the 13,801 participants whose vital status at Wave 5 was known (Supplementary Table S27).
Discussion
In this 9-year prospective analysis of 16,479 middle-aged and older Chinese adults, cross-system multimorbidity at baseline was associated with a graded increase in all-cause mortality—each additional involved system raising the hazard by 34% (HR 1.34, 95% CI 1.24–1.44)—and the magnitude of this per-system effect is of the same order as the pooled hazard ratio of 1.44 (95% CI 1.34–1.55) for ≥ 2 versus < 2 chronic conditions reported by a meta-analysis of 26 cohorts using only physical chronic-disease indicators, although the two estimates are not directly comparable: they differ in scale, in exposure definition, in outcome ascertainment and in the populations studied, the present figure being a per-1-system increase across three domains and the published figure a binary contrast within the physical domain [5], and consistent with the larger gradients reported when psychological and cognitive disorders are jointly considered [11]. The most informative finding emerged at the eight-pattern level: the co-occurrence of physical multimorbidity and cognitive impairment alone (101) carried a hazard (HR 2.14, 95% CI 1.57–2.91) approaching that of the all-three configuration (111; HR 2.45, 95% CI 1.89–3.17), with virtually identical 9-year crude death rates of 12.51% and 13.04%, respectively. We are not aware of a previous cohort analysis that compares all eight cross-system combinations against long-term all-cause mortality in a Chinese national sample; previous Chinese applications of the same three-domain framework have addressed cross-sectional prevalence [8], healthcare-process outcomes for physical multimorbidity or the incidence of multimorbidity itself [9, 10], and the only large mortality analysis using the framework to date—a UK Biobank study of 332,012 adults—operationalised the exposure as a four-level summary that does not separately quantify the physical–cognitive substrate identified here [11]. Isolated depression at baseline was not associated with a clearly elevated mortality hazard in our cohort (HR 1.12, 95% CI 0.81–1.57), a point estimate compatible with both no association and a clinically meaningful effect; this contrasts numerically with the pooled risk ratio of 1.34 (95% CI 1.27–1.42) from a meta-analysis of 49 community-based cohorts on late-life depression and all-cause mortality [28], and is most likely attributable to the modest number of deaths in the depression-only cell (n = 68) and to the relatively short 9-year follow-up window in a cohort with a mean baseline age of 59 years.
The near-ceiling 9-year hazard observed once both physical multimorbidity and cognitive impairment are present has plausible mechanistic underpinnings on each domain. Physical multimorbidity is a well-established correlate of premature mortality, with each additional chronic condition adding incrementally to baseline hazard through end-organ failure, cardiovascular events and accumulated systemic deficits [5, 6]. Cognitive impairment, in turn, is independently associated with substantially elevated all-cause mortality among Chinese older adults: in the Chinese Longitudinal Healthy Longevity Survey (CLHLS), mortality rose stepwise across mild, moderate and severe cognitive impairment relative to those without impairment, with adjusted hazard ratios of 1.20 (95% CI 1.13–1.28), 1.38 (1.27–1.51) and 1.47 (1.33–1.62), respectively, in the oldest-old [29], and a separate CLHLS analysis showed that rapid cognitive decline carried a 75% higher hazard of death (adjusted HR 1.75, 95% CI 1.57–1.95) [30]. The conjunction of physical and cognitive deficits—conceptually adjacent to "cognitive frailty"—has been pooled in a meta-analysis of 12 prospective cohorts as conferring an all-cause mortality hazard of 1.93 (95% CI 1.67–2.23), substantially exceeding the per-system gradient seen with physical multimorbidity alone and consistent in magnitude with the 101-pattern hazard observed in the present analysis [31]. Plausible downstream pathways through which cognitive impairment may amplify mortality risk in the presence of physical multimorbidity include falls, aspiration pneumonia, malnutrition and reduced adherence to chronic-disease regimens. When these two pathways co-occur, they may converge on a substantial fraction of the deaths achievable within a 9-year horizon in a cohort with a 6.27% 9-year mortality risk, leaving little additional room for depression to contribute incrementally—an interpretation consistent with the absence of a clear additional hazard for the 110 versus 100 contrast (HR 1.57 vs 1.54) and the 111 versus 101 contrast (HR 2.45 vs 2.14) in our data.
The independent contribution of depression to all-cause mortality, by comparison, is comparatively modest in pooled estimates (RR 1.34) and is partly mediated through behavioural risk factors—including smoking, alcohol use, physical inactivity and suboptimal diet—and through metabolic, immuno-inflammatory, autonomic and hypothalamic–pituitary–adrenal axis dysregulations that are also pathways to physical chronic disease, rather than acting through depression in isolation [28, 32].This null estimate for depressive symptoms occurring in isolation differs from studies that report an increased mortality risk for depression treated as a stand-alone exposure, and the discrepancy is one that further research in cohorts with longer follow-up will need to resolve. A 9-year window may therefore be insufficient to capture the more distal mortality consequences of isolated depressive symptomatology, and an attenuated point estimate within saturated physical–cognitive configurations is biologically plausible. Two further patterns in the M3 covariate estimates merit comment. The numerically higher mortality hazard for former (HR 1.55) than current smokers (HR 1.25) is compatible with the well-described "sick-quitter" phenomenon, in which individuals quit smoking after the onset of serious illness and consequently appear at higher mortality risk in observational data than continuing smokers without overt disease—a pattern documented across long-running cohorts including the 50-year follow-up of male British doctors [33]. Cumulative exposure offers a second, non-exclusive explanation: in a cohort with a mean baseline age of 59 years, former smokers include people who stopped decades earlier and who may nevertheless have accumulated a larger lifetime tobacco dose than same-aged current smokers who started later or smoked less intensively. CHARLS records smoking status but neither pack-years nor age at initiation, so the two explanations cannot be separated in these data, and this covariate estimate should not be read as an effect of quitting. The approximately one-quarter attenuation of the per-system hazard ratio after adjustment for self-rated health (1.34 → 1.25) is most consistent with self-rated health lying on the causal pathway between cross-system multimorbidity and mortality rather than acting as an upstream confounder; self-rated health has been shown across more than two dozen community studies to be an independent predictor of subsequent mortality even after adjustment for objective health indicators, integrating clinically diagnosed and undiagnosed health states into a single downstream summary [34]. This sensitivity analysis is therefore reported as supportive only. Harmonised data from three national ageing cohorts point in the same direction for a different endpoint: there, multimorbidity combinations that included a psychological component showed the strongest associations with the subsequent development of further conditions, which is compatible with depressive symptoms mattering more for the accrual of morbidity than for death within a 9-year window. Because the Chinese participants in that analysis were drawn from the same survey, the two sets of findings are related rather than independent [12].
Because the physical–cognitive substrate alone (101) reached a 9-year hazard close to that of triple-domain involvement, identifying middle-aged and older adults with both physical multimorbidity and baseline cognitive impairment may be of more immediate clinical use than tracking the simple count of involved systems. Beyond this targeted observation, our findings reinforce the broader argument that resolving cross-system patterning—rather than collapsing to a binary or count-based summary—adds information for prognostication that is not visible in chronic-disease counts alone [6, 11]. Because a formal contrast did not distinguish this combination from involvement of all three domains, however, the observation should be regarded as descriptive. These prognostic findings are complementary to recent work that models how the cross-system patterns themselves arise. A multi-state survival analysis harmonising the Health and Retirement Study, the English Longitudinal Study of Ageing and CHARLS, restricted to adults free of physical, psychological and cognitive multimorbidity at baseline, estimated transition-specific hazards into each of the same pattern states and reported that cumulative physiological dysregulation preferentially accelerated transitions into cognition-involving states, whereas low subjective well-being acted as a broader risk factor across most transitions [35]. Because that analysis begins from a multimorbidity-free state and conditions on biomarker-derived measures available only in a sub-sample, its transition hazards are not directly comparable with the mortality hazards estimated here; read together, however, the two designs describe successive segments of the same process—the pathways by which the physical–cognitive configuration is entered, and the mortality risk it carries once it is present.
The value of treating physical, psychological and cognitive conditions as a single cross-system construct, and of decomposing that construct into its constituent combinations, has been questioned on the grounds that the composite may be artificial and that decomposition could dilute the contribution of cumulative physical disease burden. Three features of the present analysis bear on this. First, the construct is not specific to this study: the same three domains have been operationalised in the same way in the multi-country and multi-cohort analyses discussed above [7, 12], and influential proposals for measuring multimorbidity in older populations explicitly recommend that mental and cognitive disorders be counted alongside physical chronic disease rather than treated as a separate literature [6]. Second, decomposition adds information rather than removing it: at an identical count of involved systems the eight combinations differ substantially in hazard—at two involved systems from 1.47 to 2.14—and by more than a factor of two in adjusted 9-year risk difference (+ 2.24 to + 5.12 percentage points), and the physical–cognitive combination approaches the hazard of triple-domain involvement, heterogeneity that a summary count conceals by construction. Nor is the physical contribution diluted, since physical multimorbidity retains the largest independent association of the three domains when all three indicators are entered simultaneously. Third, the absence of a detectable excess hazard for isolated depression is a product of the decomposition rather than an argument against it, because it is only by estimating the eight cells separately that this particular configuration could be identified as the one carrying no clear excess risk within nine years—a distinction that neither a count nor a binary indicator could have drawn. We do not claim that the three domains are aetiologically equivalent or that they contribute equally to mortality; the claim is the narrower one that measuring them jointly and reporting them separately is more informative for prognosis than collapsing them into a single summary. The three domains do not, however, carry equal weight: when they are entered simultaneously the hypothesis that their coefficients are equal is rejected, so that a summary count giving every involved system the same weight is a lossy summary of the same information. Re-expressing the eight combinations as three main effects and their interactions gave no evidence of departure from additivity on the multiplicative scale, so the heterogeneity across the eight cells arises from the differing magnitudes of the three domain effects and not from synergy between them. This bounds the claim we make. The case for reporting the domains separately rests on how unequally they are weighted, not on any interaction among them; a reader given only the three main effects would still be unable to say which particular configuration carries no excess risk within nine years, but would not be losing evidence of synergy, because there is none to lose.
Two applications follow from this. In clinical settings in which the three domains are already assessed separately—comprehensive geriatric assessment, chronic-disease follow-up and pre-operative evaluation among them—the eight-cell pattern can be read directly from information that is already collected, and the physical–cognitive combination identifies a group whose 9-year mortality resembles that of triple-domain involvement without requiring any additional instrument. In epidemiological settings, ageing cohorts that already record chronic conditions, depressive symptoms and cognition can report cross-system status at pattern level rather than as a count at no additional cost in data collection; doing so in cohorts with longer follow-up, with repeated measurement of the three domains and with cause-specific mortality would establish whether the ordering observed here is stable over time and across causes of death, and whether the isolated-depression cell remains null beyond nine years. Neither use implies that the pattern should be treated as a validated prognostic instrument: discrimination was moderate to good (optimism-corrected concordance statistic 0.81), but the model was not calibrated or externally validated and was not developed for individual prediction.
Two features of the analysis bear on how these estimates should be read. The proportional hazards assumption was not supported for the exposure, but this reflects a change in the composition of the deaths across the follow-up window rather than an association that weakens with time: a date of death was recorded for almost every death in the first two years and imputed for almost every death in the last two, so that successive intervals contrast largely different sets of participants, and a discrete-time model that does not require the assumption reproduced the per-system estimate. Secondly, the deaths whose date was not recorded are not a random subset, so that treating them as censored would remove genuine events from the later part of follow-up; the main analysis, which retains and imputes them, yields a smaller estimate than either the analysis restricted to deaths with a recorded date or the inclusion rule applied in the originally submitted paper, and is in that sense the most conservative of the three specifications.
The principal strengths of this analysis are its national sampling frame, the 9-year follow-up window with proxy-based vital-status ascertainment across four CHARLS waves, the simultaneous availability of all three system-level indicators at the 2011 baseline, the full eight-cell resolution of cross-system patterns rather than the four-level summaries previously used [11], the parallel use of two exposure operationalisations (the manually reconstructed score and the previously derived PPC-MM binary indicator) with explicit cross-tabulation, the multiple-imputation approach paired with a complete-case sensitivity analysis to verify that imputation did not inflate estimates, and the absence of meaningful collinearity in the M3 model. Several limitations require explicit acknowledgement. First, all 14 chronic conditions were ascertained by self-report of physician diagnosis and are therefore subject to recall bias and to under-ascertainment of conditions never formally diagnosed, although this is the standard approach in CHARLS and most other large ageing cohorts [13]; in addition, two of these 14 conditions (emotional problems and memory-related disease) overlap conceptually with the psychological and cognitive domains, but their low baseline prevalence (< 1.6% each; Table 1) makes any resulting cross-domain contamination quantitatively negligible. Second, cognitive impairment was defined by the lowest sample quartile of the CHARLS cognitive composite; this distribution-based threshold is pragmatic and widely used in CHARLS analyses but yields a fixed approximate baseline prevalence and does not directly map onto a clinical diagnosis of mild cognitive impairment or dementia [17], and the higher imputed prevalence of cognitive impairment among participants whose cognitive composite was missing (38.1% versus 25.4%; Supplementary Table S4) suggests that participants with missing cognitive items may have been more likely to be impaired—a pattern the imputation model appropriately captured by conditioning on the fully observed predictors but that warrants caution when extrapolating absolute prevalences. Because the cut-point is derived within the analytic sample it indexes relative rather than absolute cognitive standing, so the three combinations that include cognitive impairment (001, 101 and 111) are populated by position within this cohort rather than by an external standard; five alternative thresholds, among them cut-points standardised for age and for education, left the ordering of the eight combinations and the position of the physical–cognitive combination unchanged. The number of participants assigned to those three cells nevertheless varies with the threshold: under the three arms that redefine cognition the physical–cognitive combination contains between 632 and 767 participants against 778 under the main definition, and triple-domain involvement between 1,050 and 1,160 against 1,442 (Supplementary Table S11). The CLHLS evidence cited above nevertheless suggests that severity-graded cognitive measures yield broadly similar mortality gradients to the present categorical definition [29, 30]. Third, vital status was recorded at discrete CHARLS waves, with date of death obtained by proxy interview; follow-up time is resolved to the year rather than to the day, so the discrete structure of the survey introduces some misclassification of the time at risk. This is why a discrete-time model fitted on the wave intervals, and a nine-year risk ratio that uses no death time at all, are reported alongside the Cox estimates. Fourth, we did not perform cause-specific mortality analyses; the differential pathways through which physical, cognitive and psychological disorders may operate cannot therefore be disentangled in the present data. Fifth, despite a sample of 16,479 the number of deaths within some patterns—most notably 010 (n= 68)—remained modest, leading to wide confidence intervals; we have flagged this explicitly in the Results and have refrained from over-interpreting null point estimates accordingly. Sixth, the 9-year follow-up window may be insufficient to capture the more distal mortality consequences of isolated depression at baseline [28, 32]; longer follow-up through forthcoming CHARLS waves will allow re-examination of the depression-only cell. Seventh, the analysis is unweighted. CHARLS was drawn with a complex multistage design, but the exclusions applied here and the attrition accumulated over four subsequent waves mean that the assembled cohort no longer reproduces the original sampling frame; the estimates should therefore be read as associations within this cohort rather than as weighted mortality or prevalence estimates for the national population, notwithstanding the national frame from which the cohort originates. Eighth, all three system indicators were measured only at the 2011 baseline, so changes in physical, psychological and cognitive status over the following nine years were not captured; the estimates describe the prognostic value of a single baseline assessment rather than of an evolving state, and subsequent within-person transitions between cross-system patterns would be expected to attenuate rather than to exaggerate the observed associations. Ninth, our analysis is descriptive of associations rather than causal: although we adjusted for an extensive covariate set and conducted multiple sensitivity analyses, residual confounding cannot be excluded and the observed attenuations should be interpreted with appropriate causal humility. Tenth, the proportional hazards assumption was not supported for the cross-system exposure itself, so the hazard ratios reported here should be read as average risk ratios across the 9-year window rather than as constant instantaneous effects; a discrete-time model that does not require this assumption produced the same per-system estimate. Finally, one further limitation concerns the outcome itself. For 550 of the 1,033 deaths no date was recorded and the time of death was imputed. The estimates are insensitive to where within the interval those times are placed, and a risk ratio computed without any death time reproduces the association; the deaths with a recorded date nevertheless yield a larger per-system estimate than those without, and although most of that difference is attributable to the wave at which participants were last seen, it is not fully removed when age is used as the time scale.
Conclusion
In this 9-year prospective analysis of a national sample of 16,479 middle-aged and older Chinese adults, cross-system multimorbidity at baseline was associated with a graded, dose-dependent increase in all-cause mortality, with each additional involved system raising the hazard of death by approximately 34%. Pattern-level analysis revealed substantial heterogeneity across cells of the same overall system count: the co-occurrence of physical multimorbidity and cognitive impairment alone (101) yielded a hazard close to that observed for involvement of all three domains (111), whereas isolated depression at baseline was not associated with a clearly elevated mortality hazard within the 9-year window. These findings indicate that resolving cross-system patterning—rather than relying on chronic-disease counts or binary indicators alone—adds prognostic information not visible at the level of system count, and identify the physical–cognitive substrate as a particularly informative target for clinical attention in geriatric care, although the model was not calibrated or externally validated and is not intended for individual risk stratification. Longer follow-up and cause-specific mortality analyses are warranted to clarify the role of isolated depression and the differential pathways through which cross-system patterns may operate.
Supplementary Information
Authors’ contributions
M.X. conceived and designed the study, performed the statistical analysis, interpreted the results, wrote the main manuscript text, prepared all figures and tables, and participated in manuscript revision. H.Y. assisted with data analysis, contributed to result interpretation, and helped with manuscript preparation and revision. G.C. provided overall project supervision and guidance, contributed to study conceptualization and methodology, assisted with data interpretation, and critically reviewed and revised the manuscript for important intellectual content. All authors read and approved the final manuscript.
Funding
This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. No funding was received for this study.
Data availability
Data analyzed in this study are publicly available through the CHARLS database (http://charls.pku.edu.cn/en) upon registration.
Declarations
Ethics approval and consent to participate
This research utilized de-identified public data from CHARLS. The original study protocol received ethical clearance from Peking University's Institutional Review Board (IRB00001052-11015 for the main household survey, which covers the interview and the anthropometric and blood-pressure measurements used here; the separate approval IRB00001052-11014 covers the biomarker collection, which was not used in the present analysis) and was conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent.
Consent for publication
Not applicable.
Competing interests
The authors declare that they have no competing interests.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Marengoni A, Angleman S, Melis R, et al. Aging with multimorbidity: a systematic review of the literature. Ageing Res Rev. 2011;10(4):430–9. 10.1016/j.arr.2011.03.003. [DOI] [PubMed] [Google Scholar]
- 2.Chen X, Giles J, Yao Y, et al. The path to healthy ageing in China: a Peking University–Lancet Commission. Lancet. 2022;400(10367):1967–2006. 10.1016/S0140-6736(22)01546-X. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Gong J, Wang G, Wang Y, et al. Nowcasting and forecasting the care needs of the older population in China: analysis of data from the China Health and Retirement Longitudinal Study (CHARLS). Lancet Public Health. 2022;7(12):e1005–13. 10.1016/S2468-2667(22)00203-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Hu Y, Wang Z, He H, Pan L, Tu J, Shan G. Prevalence and patterns of multimorbidity in China during 2002–2022: a systematic review and meta-analysis. Ageing Res Rev. 2024;93:102165. 10.1016/j.arr.2023.102165. [DOI] [PubMed] [Google Scholar]
- 5.Nunes BP, Flores TR, Mielke GI, Thumé E, Facchini LA. Multimorbidity and mortality in older adults: a systematic review and meta-analysis. Arch Gerontol Geriatr. 2016;67:130–8. 10.1016/j.archger.2016.07.008. [DOI] [PubMed] [Google Scholar]
- 6.Calderón-Larrañaga A, Vetrano DL, Onder G, et al. Assessing and measuring chronic multimorbidity in the older population: a proposal for its operationalization. GERONA. 2017;72(10):glw233. 10.1093/gerona/glw233. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Ni Y, Zhou Y, Kivimäki M, et al. Socioeconomic inequalities in physical, psychological, and cognitive multimorbidity in middle-aged and older adults in 33 countries: a cross-sectional study. Lancet Healthy Longev. 2023;4(11):e618–28. 10.1016/S2666-7568(23)00195-2. [DOI] [PubMed] [Google Scholar]
- 8.Lin L, Duan D, Yan L, He H. Prevalence and associated factors of physical-psychological-cognitive multimorbidity in Chinese community-dwelling older adults: a cross-sectional study. PeerJ. 2025;13:e19750. 10.7717/peerj.19750. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Zhao Y, Atun R, Oldenburg B, et al. Physical multimorbidity, health service use, and catastrophic health expenditure by socioeconomic groups in China: an analysis of population-based panel data. Lancet Glob Health. 2020;8(6):e840–9. 10.1016/S2214-109X(20)30127-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Bi J, Pan Y, Guo W, et al. Association of circadian syndrome with the risk of physical, psychological, and cognitive multimorbidities: a prospective cohort study based on the China Health and Retirement Longitudinal Study. J Glob Health. 2025;15:04351. 10.7189/jogh.15.04351. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Talifu Z, Ren Z, Chen C, et al. The association between accelerated biological aging and the physical, psychological, and cognitive multimorbidity and life expectancy: cohort study. Aging Cell. 2025;24(9):e70142. 10.1111/acel.70142. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Zhou Y, Kivimäki M, Holt-Lunstad J, et al. Stressful life events in childhood and adulthood and risk of physical, psychological and cognitive multimorbidities: a multicohort study. eClinicalMedicine. 2025;83:103225. 10.1016/j.eclinm.2025.103225. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Zhao Y, Hu Y, Smith JP, Strauss J, Yang G. Cohort profile: The China Health and Retirement Longitudinal Study (CHARLS). Int J Epidemiol. 2014;43(1):61–8. 10.1093/ije/dys203. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. J Clin Epidemiol. 2008;61(4):344–9. 10.1016/j.jclinepi.2007.11.008. [DOI] [PubMed] [Google Scholar]
- 15.Andresen EM, Malmgren JA, Carter WB, Patrick DL. Screening for depression in well older adults: evaluation of a short form of the CES-D. Am J Prev Med. 1994;10(2):77–84. 10.1016/S0749-3797(18)30622-6. [PubMed] [Google Scholar]
- 16.Boey KW. Cross-validation of a short form of the CES-D in Chinese elderly. Int J Geriat Psychiatry. 1999;14(8):608–17. 10.1002/(sici)1099-1166(199908)14:8<608::aid-gps991>3.0.co;2-z. [DOI] [PubMed] [Google Scholar]
- 17.Lei X, Hu Y, McArdle JJ, Smith JP, Zhao Y. Gender differences in cognition among older adults in China. J Hum Resour. 2012;47(4):951–71. 10.3368/jhr.47.4.951. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Zhou BF, Cooperative Meta-Analysis Group of the Working Group on Obesity in China. Predictive values of body mass index and waist circumference for risk factors of certain related diseases in Chinese adults–study on optimal cut-off points of body mass index and waist circumference in Chinese adults. Biomed Environ Sci. 2002;15(1):83–96. [PubMed] [Google Scholar]
- 19.Joint Committee for Guideline Revision. 2018 Chinese Guidelines for Prevention and Treatment of Hypertension—A report of the Revision Committee of Chinese Guidelines for Prevention and Treatment of Hypertension. J Geriatr Cardiol. 2019;16(3):182. 10.11909/j.issn.1671-5411.2019.03.014. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Maldonado G, Greenland S. Simulation study of confounder-selection strategies. Am J Epidemiol. 1993;138(11):923–36. 10.1093/oxfordjournals.aje.a116813. [DOI] [PubMed] [Google Scholar]
- 21.Schisterman EF, Cole SR, Platt RW. Overadjustment bias and unnecessary adjustment in epidemiologic studies. Epidemiology. 2009;20(4):488–95. 10.1097/EDE.0b013e3181a819a1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.van Buuren S, Groothuis-Oudshoorn K, mice: Multivariate imputation by chained equations inR. J Stat Soft. 2011;45(3). 10.18637/jss.v045.i03.
- 23.Madley-Dowd P, Hughes R, Tilling K, Heron J. The proportion of missing data should not be used to guide decisions on multiple imputation. J Clin Epidemiol. 2019;110:63–73. 10.1016/j.jclinepi.2019.02.016. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Rubin DB. Multiple imputation for nonresponse in surveys. 1st ed. Wiley; 1987. 10.1002/9780470316696. [Google Scholar]
- 25.Cox DR. Regression models and life-tables. J R Stat Soc Ser B Stat Methodol. 1972;34(2):187–202. 10.1111/j.2517-6161.1972.tb00899.x. [Google Scholar]
- 26.Grambsch PM, Therneau TM. Proportional hazards tests and diagnostics based on weighted residuals. Biometrika. 1994;81(3):515–26. 10.1093/biomet/81.3.515. [Google Scholar]
- 27.Fox J, Monette G. Generalized collinearity diagnostics. J Am Stat Assoc. 1992;87(417):178–83. 10.1080/01621459.1992.10475190. [Google Scholar]
- 28.Wei J, Hou R, Zhang X, et al. The association of late-life depression with all-cause and cardiovascular mortality among community-dwelling older adults: systematic review and meta-analysis. Br J Psychiatry. 2019;215(2):449–55. 10.1192/bjp.2019.74. [DOI] [PubMed] [Google Scholar]
- 29.An R, Liu GG. Cognitive impairment and mortality among the oldest‐old Chinese. Int J Geriatr Psychiatry. 2016;31(12):1345–53. 10.1002/gps.4442. [DOI] [PubMed] [Google Scholar]
- 30.Lv X, Li W, Ma Y, et al. Cognitive decline and mortality among community-dwelling Chinese older people. BMC Med. 2019;17(1):63. 10.1186/s12916-019-1295-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Bu Z, Huang A, Xue M, Li Q, Bai Y, Xu G. Cognitive frailty as a predictor of adverse outcomes among older adults: a systematic review and meta‐analysis. Brain Behav. 2021;11(1):e01926. 10.1002/brb3.1926. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Penninx BW, Milaneschi Y, Lamers F, Vogelzangs N. Understanding the somatic consequences of depression: biological mechanisms and the role of depression symptom profile. BMC Med. 2013;11(1):129. 10.1186/1741-7015-11-129. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Doll R, Peto R, Boreham J, Sutherland I. Mortality in relation to smoking: 50 years’ observations on male British doctors. BMJ. 2004;328(7455):1519. 10.1136/bmj.38142.554479.AE. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Idler EL, Benyamini Y. Self-rated health and mortality: a review of twenty-seven community studies. J Health Soc Behav. 1997;38(1):21. 10.2307/2955359. [PubMed] [Google Scholar]
- 35.Pan Y, Bi J, Sun L, et al. Subjective well-being and allostatic load in multimorbidity transitions: a multi-state survival analysis of three international longitudinal cohorts. Gen Hosp Psychiatry. 2026;99:171–8. 10.1016/j.genhosppsych.2026.02.003. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Data analyzed in this study are publicly available through the CHARLS database (http://charls.pku.edu.cn/en) upon registration.