Typology of ageing and its associations with the end of life and death among older adults in China: a longitudinal person-centred study

Abstract

Objectives

Studies on ageing trajectories typically foreground intrinsic capacities and focus on frailty and mortality outcomes. Instead, research capturing aspects of people’s living environment alongside individuals’ capacities is limited. This study aimed to identify qualitatively distinctive ageing profiles among deceased older adults in China using individual-based and environment-based indicators. It further examined sociodemographic correlates of profile membership and investigated associations between these profiles and end-of-life and death-related outcomes.

Design

This is a retrospective cohort study.

Setting

Individual level, longitudinal data from the Chinese Longitudinal Healthy Longevity Study (CLHLS) between 2008 and 2018 were analysed.

Participants

The study included older participants of the CLHLS (2008–2018) who died between survey waves (N=9755).

Outcome measures

Ageing profiles were categorised using latent class modelling based on individuals’ intrinsic capacity (functional, physical, cognitive, mental and social dimensions; 20 indicators) and environment characteristics (healthcare-related, financial and social aspects; 10 items). Sociodemographic correlates of profile membership included age, sex, education, marital status, living arrangement and residence location. End-of-life and death experiences were assessed by length of survival and functional status measured at the wave prior to death, pain at death and terminal delirium.

Results

Within the study sample, four profiles emerged: healthy-and-supported (57%), cognitively-competent-but-functionally-dependent (12%), functionally-capable-but-cognitively-challenged (18%) and frail-and-support-lacking (13%). While profiles were primarily distinguished by functional and cognitive capacities and financial status, social participation was insufficient across all groups. Sociodemographic disparities (sex, education, marital status, living arrangement and residence) were associated with profile membership. The first and third profiles enjoyed a better quality of life in their final months. The healthy-and-supported adults were most prone to painful deaths in hospitals. The functionally capable participants most often died at home. The frail-and-support-lacking individuals were at the highest risk of terminal delirium.

Conclusions

The heterogeneity of ageing revolved around functional and cognitive capacities and economic/financial characteristics, the latter particularly salient for the frail-and-support-lacking group. The uniformly low social engagement observed across profiles highlighted the need for boosting overall social participation and developing community-level social services. Distinct end-of-life patterns emphasise the importance of targeted care strategies for rural and urban residents and interventions for delirium prevention.

Strengths and limitations of this study.

• This study operationalised ageing heterogeneity using a comprehensive model of healthy ageing, integrating multiple domains of intrinsic capacity and environmental factors.

• The retrospective design used end-of-life data obtained from proxy respondents after participants’ death and data collected from participants while alive, linking ageing profiles with death-related outcomes.

• The sample consisted primarily of community-dwelling adults, limiting applicability to older people residing in institutional settings.

• The analyses focused on participants deceased between survey waves without sampling weights to prioritise profile identification over prevalence estimation, which may introduce selection bias related to survival and attrition.

• Reliance on self-reported data for chronic conditions and data retrieved from proxy respondents for death experiences may have introduced potential recall or social desirability biases.

Introduction

Accompanying global population ageing has been the rapid increase in the number of the oldest-olds (ie, 80-year-olds and older).¹ The fact that more people live to advanced ages represents remarkable social and economic achievement, particularly in low-income and middle-income countries.^{2 3} On the other hand, it presents challenges to health and social care systems due to older people’s needs for daily assistance, high quality of life towards the end of life and a good death.^{1 2 4} Thus, a quest to understand factors conducive to health and well-being in old age and enhancing death quality has emerged among public health and policy researchers.

Conceptual labels illustrative of such a quest include successful ageing,4,6 active ageing^{7 8} and healthy ageing.^{9 10} A key theme in these conceptualisations is the dual emphasis on intrapersonal and contextual factors in later life, which is underscored in the public health framework for healthy ageing proposed by WHO (2015).⁹ Highlighting the interactions between people’s intrinsic capacity (IC) and supportive environmental characteristics, this composite model of ageing incorporated individuals’ physical, cognitive and psychosocial conditions with the environment’s social and economic support and accessibility of public services.10,13

Relevant research applying the WHO model and related concepts has generally converged on two strands. First, there was great diversity during old age, which manifested as a broad range of IC scores within and across samples and as discernible sociodemographic patterns at baseline and over the ageing trajectory.1114,17 For example, physical, functional and cognitive domains were IC dimensions along which older people varied the greatest.^{1 12 18} In contrast, older people’s different standings in psychosocial adjustment and social participation, two domains of intrapersonal qualities critical to engaged functioning and goal pursuit, are rarely incorporated into healthy ageing models.^{10 19} The second strand was the finding that IC predicted health outcomes, such as frailty, functional loss and mortality. For instance, lower IC at baseline predicted a greater risk of falling, higher dependency in activities of daily life (ADLs, including bathing or showering, dressing, toileting, indoor transferring, continence and eating) and higher probabilities of death among Belgian nursing home residents over 3 years.¹⁹ Similar results associating deteriorations in IC components with adverse health events also emerged from community-dwelling older people.^{12 20 21}

In general, empirical works acknowledge changes associated with the ageing process and the divergent pathways older people follow in later life, usefully guiding care and treatment initiatives for those in need. Nonetheless, three empirical and methodological gaps need addressing. First, when outlining ageing heterogeneity, comprehensive models of healthy ageing accounting for both IC and characteristics of the environment are rarely drawn on, except for a few recent exceptions.^{14 22} Moreover, longitudinal dynamics between heterogeneous ageing and end-of-life and death experiences among the oldest old are hardly ever captured. Finally, person-centred methods instrumental in qualitatively identifying heterogeneous subgroups of people and drawing pairwise comparisons regarding sociodemographic makeups, end-of-life circumstances and symptoms at death are yet to be applied.²³

Thus, the primary aim of the current study was to develop a typology of ageing profiles among a mortality cohort of older adults in China based on multiple indicators of IC and environmental characteristics using latent class modelling (LCM). While the exploratory, data-driven LCM determined the number of profiles empirically, it was anticipated that heterogeneous combinations of IC and environmental factors would emerge among the analytical sample, as guided by the WHO framework of healthy ageing. Furthermore, the secondary aim of the study was to identify characteristics of vulnerable subgroups by examining sociodemographic correlates of these emergent profiles to inform targeted interventions. Finally, this research used longitudinal data to explore the associations between ageing and end-of-life and death measures.

Methods

Study design and setting

This retrospective study drew data from the Chinese Longitudinal Healthy Longevity Study (CLHLS) between 2008 and 2018.²⁴ Employing a multistage clustered sampling strategy, the CLHLS randomly selected half of the cities and counties in 22 provinces in China, a geographical scope covering approximately 85% of the nation’s population.¹ The target participants were older people aged 65 and above, with the oldest-old (aged 80 and above) and men oversampled to ensure adequate representation.¹² The CLHLS interviewed participants biennially (1998–2002) or triennially (2005–2018) to understand healthy longevity. From 2000 onwards, the CLHLS also interviewed proxies of participants who had died since the previous wave about the decedents’ end-of-life experiences.

Analytical sample and eligibility criteria

The current study analysed data from older individuals themselves while alive and end-of-life data collected from surrogate respondents after older participants had died between survey waves (N=9755). More specifically, two inclusion criteria were adopted: (1) completion of a face-to-face interview in the 2008, 2011 or 2014 waves; and (2) death before the subsequent survey wave (ie, in 2011, 2014 or 2018). Data from earlier waves were excluded because of inconsistencies concerning medical history.

Measurement

To develop a typology of ageing profiles and examine sociodemographic correlates of profile membership, we used data on IC and environmental indicators and sociodemographic variables obtained directly from older participants themselves during their final face-to-face interview. Then, we inspected end-of-life data collected from proxies (close contacts of the deceased) during the following wave (ie, in 2011, 2014 or 2018) to uncover how each ageing profile connected to the final years of life and death.

Typology of ageing profiles: intrinsic capacity and the living environment

The typology of ageing profiles was determined by five dimensions of IC (20 indicators) and three facets of the living environment (10 items), all of which corresponded to theoretical conceptualisations and/or empirical investigations into healthy ageing in epidemiological, gerontological or public health studies.^{14 16 17 22 25}

First, functional independence,¹⁵ a widely acknowledged IC domain, was assessed by six indicators of ADLs (0=assistance needed, 1=independent). Second, physical health was evaluated by three items: self-rated health (0=bad or very bad, 1=at least moderate), hospitalisation experiences within the past 2 years (0=at least one, 1=zero) and whether participants lived with one or more of eight typical chronic conditions (hypertension, diabetes, heart disease, stroke, bronchitis, tuberculosis, cancer, dementia; 0=yes, 1=no). While the first item reflected general, subjective health, the latter two measured medical health regarding acute, severe conditions and chronic illnesses. Specifically, the eight chronic diseases were included because of empirical evidence suggesting their prevalence and comorbidity.^{26 27} The third dimension, cognitive functioning, was assessed by the adapted version of the Mini-Mental State Examination.¹ We generated one dichotomous variable for each of its five domains: orientation, registration, calculation, memory and language (0=answered wrongly or unable to answer; 1=answered correctly). We then included two psychosocial characteristics that older individuals typically draw on—mental health and social capacity.^{22 25} Mental health consisted of three items: self-rated life satisfaction (0=bad or very bad, 1=at least moderate), negative affect (feeling fearful or anxious; 0=often or always, 1=sometimes or less frequent) and positive affect (feeling cheerful; 0=seldom or never, 1=sometimes or more frequent). The last IC dimension was social capacity estimated by three types of social behaviours: playing cards or mah-jong, participating in organised social activities and travelling in the past 2 years (0=never, 1=once or more).

According to pertinent literature, healthcare, economic, social and public service factors of the environment typically predict ageing trajectories of community-residing older people.^{14 22 27} Therefore, we examined three facets of the living environment. The first, healthcare-related one, contained three indicators: whether participants had caregivers when sick, adequate medical treatment when needed and medical insurance (0=no, 1=yes). Second, the economic-financial dimension comprised three items: whether participants had sufficient access to financial resources to cover daily expenditures (0=no, 1=yes), whether their financial autonomy regarding personal or household purchases was acknowledged (0=no, 1=yes) and subjective economic status (0=poor or very poor, 1=at least moderate). The last aspect was social support at interpersonal and community levels: whether participants had people to chat with, to share worries with, to turn to for help and whether a range of social services was available in their local neighbourhood (such as counselling, social and recreational activities, and others; 0=no, 1=yes).

To sum up, a total of 30 dichotomous indicators were employed to identify categories of ageing profiles, where 0 referred to negative conditions and 1 represented positive ones.

End-of-life measures and death circumstances

Outcomes measured at the end of life included (1) length of survival after the last interview (months) and (2) being bedridden within 3 months prior to death (0=yes, 1=no). Death circumstances consisted of three aspects: place of death (1=home, 2=hospitals, 3=nursing homes and other facilities), pain at death and symptoms of delirium prior to death. Pain at death was assessed using a proxy-reported symptom-based question asking whether the older adult exhibited observable signs of pain when approaching death. Responses were dichotomised to indicate the presence (0=painful death) or absence (1=not painful death) of pain. Symptoms of delirium were assessed using the CLHLS item: ‘Was the older adult confused, disoriented or not mentally clear before their death?’ (0=yes, 1=no). This item captured proxy-reported altered mental status or terminal cognitive disturbance. Notably, the wording in the CLHLS item is consistent with Chinese survey practice and does not constitute a clinical diagnosis of delirium, complete loss of consciousness or coma.²⁸

All end-of-life and death-related measures were obtained through structured, face-to-face interviews administered to proxy respondents (most commonly next-of-kin) during the subsequent survey wave following the participant’s death. To mitigate potential biases inherent in retrospective proxy reports, the CLHLS employed standardised questionnaires and trained interviewers. In addition, because most older adults in China die at home or in the community, proxy respondents (typically family members) are often well positioned to observe symptoms immediately prior to death, supporting the validity of surrogate reports in this context.^{1 29 30}

Sociodemographic correlates

We considered six sociodemographic correlates of ageing typology: age at the last interview (years), sex (0=male, 1=female), education (years of schooling), marital status (0=unmarried, 1=married), living arrangement (0=living alone, 1=living with others) and residence location (0=villages, 1=cities or towns).

Statistical analyses

After presenting participants’ backgrounds and descriptive statistics, we adopted LCM to explore a typology of ageing profiles with Mplus V.8.7.³¹ All 30 indicators (20 for IC and 10 for environmental characteristics) were entered as binary variables. The overall proportion of missing data was 4.3% across all 30 indicators, sociodemographic correlates and end-of-life and death-related outcomes. The missing data were handled by the full information maximum likelihood estimation method to provide unbiased and efficient parameter estimates.³² As the primary aim was to identify distinct ageing profiles within this specific retrospective cohort rather than to estimate national prevalence rates, sampling weights were not applied to LCM.

To select the best-fitting model, four sets of criteria were applied to compare LCM solutions with increasing numbers of identified categories.³² First, better model fit was indicated by smaller Akaike’s information criterion (AIC), the Bayesian information criterion (BIC) and adjusted BIC values and larger (less negative) log-likelihood (LL) values. Second, entropies closer to one suggested clearer classification separation. Third, statistically significant results (p<0.05) of the Lo-Mendell-Rubin likelihood ratio (LMR) test and bootstrapped likelihood ratio test (BLRT) indicated improved fit for a model with one additional class relative to its predecessor. In cases of conflicting fit statistics, model selection prioritised BIC and BLRT, given their demonstrated robustness in latent class enumeration.³² The last, qualitative criterion guiding the final decision was the interpretabilities of competing LCM solutions: models with additional classes were retained only when those additions represented meaningful and distinct profiles relative to the solution with one fewer class.

The local independence assumption was assessed by inspecting tetrachoric correlations between indicators. Although a strong association was observed between perceived adequacy of financial resource and subjective economic status (r=0.763), a sensitivity analysis relaxing this independence constraint yielded nearly identical class memberships, supporting the robustness of the main LCM model.

Following the identification of the best-fitting LCM solution, we performed regression mixture modelling (RMM) to address the study’s second objective: identifying sociodemographic backgrounds of vulnerable subgroups. More specifically, the latent profile memberships were regressed on the sociodemographic correlates. This analysis employed an efficient and robust estimation method accounting for measurement errors in category identification proposed by Bolck and colleagues.³³

Finally, we applied RMM to examine the relationships between ageing profiles and end-of-life and death experiences. In this analysis, the latent class memberships (ie, ageing profiles) were treated as predictor variables and measures of end-of-life and death circumstances (ie, length of survival, bedridden status, place of death and pain and delirium at death). Vermunt’s estimation approach was employed for the continuous outcome (length of survival).³² Lanza, Tan and Bray’s approach was adopted for four categorical outcome variables (bedridden status, place of death and pain and delirium at death).³²

Patient and public involvement

The current work drew on longitudinal data collected from the deceased participants in CLHLS 2008–2018, and therefore it was impossible to share the data or findings with each individual participant. Patients were not involved in the design or execution of the research study. The authors intend to disseminate the main findings of this study with the public by publishing a summary of the study on the institutional website.

Results

Participant characteristics and descriptive statistics

The ‘Full sample’ column in table 1 portrays participants’ sociodemographic backgrounds and descriptive statistics. At the last interview, participants aged between 65 and 117 (mean age=92.9, SD=9.00), and the mean age at death was 94.4 (SD=8.95). Women accounted for over half of the sample (58.8%), and the average years of schooling were 1.5 (SD=2.89). When they accepted their last interview before dying, only 20.0% of the sample were married, few lived alone (13.3%) and over half resided in rural villages (58.6%) rather than urban areas. On average, respondents lived for 18.6 months (SD=11.45) after their last interview, during which 74.1% were bedridden in their final months. As reported by proxies, 88.7% died at home, while the rest were in hospitals (8.4%) or other places (2.8%). Painful or unconscious deaths were experienced by 14.1% and 36.1% of participants, respectively.

Table 1. Participant characteristics for the whole sample (N=9755) and by latent class.

	Full sample	Class 1 (n=5527, 56.8%)	Class 2 (n=1153, 11.9%)	Class 3 (n=1762, 18.1%)	Class 4 (n=1284, 13.2%)
Age at last interview (years)
Mean (SD)	92.9 (9.0)	90.5 (9.0)	94.4 (8.5)	96.6 (7.3)	96.9 (8.1)
Minimum–maximum	65–117	65–116	69–117	68–115	68–116
Age at death (years)
Mean (SD)	94.4 (9.0)	92.1 (9.0)	95.7 (8.4)	98.1 (7.3)	98.1 (8.0)
Minimum–maximum	65–121	65–118	71–121	70–115	69–119
Sex
Men	4017 (41.2)	2734 (49.5)	437 (37.9)	493 (28.0)	342 (26.6)
Women	5738 (58.8)	2793 (50.5)	716 (62.1)	1269 (72.0)	942 (73.4)
Education (years)
Mean (SD)	1.5 (2.9)	1.8 (3.1)	1.6 (3.2)	0.7 (2.1)	1.0 (2.4)
Minimum–maximum	0–25	0–25	0–22	0–16	0–16
Marital status at last interview
Being unmarried	7745 (80.0)	4097 (74.8)	918 (80.2)	1568 (89.3)	1143 (89.6)
Being married	1934 (20.0)	1379 (25.2)	226 (19.8)	187 (10.7)	133 (10.4)
Living arrangement at last interview
Living alone	1286 (13.3)	903 (16.4)	88 (7.7)	217 (12.4)	75 (5.9)
Living with others	8410 (86.7)	4588 (83.6)	1056 (92.3)	1536 (87.6)	1205 (94.1)
Residence location at last interview
Villages	5721 (58.6)	3241 (58.6)	581 (50.4)	1178 (66.9)	714 (55.6)
Towns or cities	4034 (41.4)	2286 (41.4)	572 (49.6)	584 (33.1)	570 (44.4)
Length of survival after the last interview (months)
Mean (SD)	18.6 (11.5)	19.8 (11.3)	17.0 (11.9)	18.7 (11.2)	14.5 (11.0)
Minimum–maximum	0–62	0–62	0–60	0–60	0–61
Being bedridden within 3 months prior to death
Yes	7054 (74.1)	3887 (72.1)	867 (77.2)	1264 (73.1)	1014 (81.5)
No	2460 (25.9)	1502 (27.9)	256 (22.8)	466 (26.9)	230 (18.5)
Place of death
Home	8523 (88.7)	4761 (87.5)	1004 (88.5)	1622 (92.1)	1114 (88.9)
Hospital	810 (8.4)	528 (9.7)	95 (8.4)	85 (4.8)	97 (7.7)
Other	272 (2.8)	152 (2.8)	35 (3.1)	41 (2.3)	42 (3.4)
Painful death
Yes	1250 (14.1)	805 (16.0)	127 (12.1)	186 (11.4)	123 (10.6)
No	7646 (85.9)	4222 (84.0)	923 (87.9)	1452 (88.6)	1033 (89.4)
Symptoms of delirium
Yes	3429 (36.1)	1832 (34.0)	443 (39.6)	621 (36.0)	518 (41.8)
No	6069 (63.9)	3551 (66.0)	676 (60.4)	1106 (64.0)	722 (58.2)

Figures are numbers unless otherwise specified. Percentages (%) are proportions of non-missing values.

Class 1, the healthy-and-supported profile; Class 2, the cognitively-competent-but-functionally-dependent profile; Class 3, the functionally-capable-but-cognitively-challenged profile; Class 4, the frail-and-support-lacking profile.

Typology of ageing profiles

Table 2 summarises the fit indices of LCM solutions with one to six classes. As more categories were added, LL increased, AIC, BIC and adjusted BIC steadily declined, and the p values of the LMR test and BLRT remained statistically significant. These results indicated that models with more classes fitted the data better and had stronger explanatory power. However, the three-class and four-class models outperformed others in classification accuracy given higher entropies. Then, the interpretabilities of identified categories of these two models were compared. Table 2 shows that in both solutions, Classes 1 (green line) and 3 (blue line) stayed consistent regarding conditional probabilities, frequencies and proportions. However, figure 1 reveals that Class 2 in the three-category model (orange line in Panel A) could be divided into Classes 2 and 4 in the four-category model (yellow and red lines in Panel B). Since the two classes showed distinct and meaningful differences in indicator probabilities concerning IC (functional ability and cognitive functioning) and the living environment (economic–financial factors), the four-class solution was considered the optimal fit for the data. The conditional probabilities in the four-class latent class model are exhibited in online supplemental table S1.

Table 2. Fit indices of latent class models and distribution of participants by class.

	k	LL	AIC	BIC	aBIC	Entropy	LMR test	BLRT	Frequency and proportion (%) of participants by class
									Class 1	Class 2	Class 3	Class 4	Class 5	Class 6
1 class	30	−124 677	249 413	249 629	249 534	-	-	-	9726 (100)	-	-	-	-	-
2 classes	61	−108 247	216 617	217 055	216 861	0.93	<0.001	<0.001	6051 (62)	3675 (38)	-	-	-	-
3 classes	92	−102 854	205 892	206 552	206 260	0.94	<0.001	<0.001	5804 (60)	2143 (22)	1779 (18)	-	-	-
4 classes	123	−100 374	200 995	201 878	201 487	0.94	<0.001	<0.001	5527 (57)	1153 (12)	1762 (18)	1284 (13)	-	-
5 classes	154	−98 515	197 339	198 445	197 956	0.91	<0.001	<0.001	4682 (48)	1657 (17)	1009 (10)	1102 (11)	1276 (13)	-
6 classes	185	−97727	195 823	197 152	196 564	0.90	<0.001	<0.001	992 (10)	1190 (12)	1623 (17)	1282 (13)	949 (10)	3690 (38)

The optimal fit is highlighted in bold font.

aBIC, adjusted BIC; AIC, Akaike’s information criterion; BIC, Bayesian information criterion; BLRT, bootstrapped likelihood ratio test; k, number of free parameters; LL, log-likelihood; LMR test, Lo-Mendell-Rubin likelihood ratio test.

Figure 1. Ageing profiles estimated using latent class modelling (LCM). (A) Three-class solution. (B) Four-class solution. Vertical lines serve as visual guides to organise the 30 indicators into their corresponding IC or environmental domains. Grey lines separate dimensions within IC and environmental characteristics, and the black line separates IC indicators from environmental ones. CF, cognitive functioning; FA, functional ability/independence; FS, financial status; HC, healthcare-related factors; IC, intrinsic capacity; PH, physical health; SP, social participation; SS, social support factors; SW, subjective wellbeing and mental health.

As depicted by figure 1B, respondents in Class 1 (n=5527, 56.8% of the entire sample) were healthy-and-supported, reporting the best IC and the most positive interactions with the environment. In contrast to this healthy-and-supported profile, the 1284 participants (13.2%) in Class 4 were labelled ‘frail-and-support-lacking’, as they appeared most disadvantaged in practically all measures of IC and the environmental support. The remaining older adults were markedly distinguished by inverse patterns regarding functional and cognitive health: older individuals in Class 2 (n=1153, 11.9%) ranked second in cognitive functioning, but their ADL performances were second from the bottom. Instead, those in Class 3 (n=1762, 18.1%) barely reported any functional limitations, but their cognitive capacity appeared severely challenged. Therefore, the participants in Classes 2 and 3 were labelled ‘cognitively-competent-but-functionally-dependent’ and ‘functionally-capable-but-cognitively-challenged’ profiles, respectively.

Comparison of sociodemographic characteristics among ageing profiles

The last four columns of table 1 describe participants’ sociodemographic backgrounds by ageing profiles. table 3 further compares sociodemographic features between profiles. The healthy-and-supported group (Class 1) contained the smallest proportion of women (50.5%). At the last interview, they were the youngest (Mean=90.53 years, SD=9.02), most likely to be married, although most often lived alone (16.4%). The cognitively-competent-but-functionally-dependent profile (Class 2) had the largest proportion of urban dwellers (49.6%). The functionally-capable-yet-cognitively-challenged profile (Class 3) had the least education (Mean=0.74 years, SD=2.09) and the largest percentage of rural residents (66.9%). The frail-and-supporting-lacking category (Class 4) was similar to those in Class 3 in three regards: both categories had the highest proportions of women; at the last interview, the respondents were the oldest and least likely to stay married. Participants in Classes 2 and 4 were equally least likely to live independently (7.7% and 5.9%, respectively).

Table 3. Pairwise comparisons of sociodemographic correlates between any two ageing profiles.

	Class 1 versus Class 2		Class 1 versus Class 3		Class 1 versus Class 4		Class 2 versus Class 3		Class 2 versus Class 4		Class 3 versus Class 4
	χ2 (df)*	P value	χ2 (df)	P value	χ2 (df)	P value	χ2 (df)	P value	χ2 (df)	P value	χ2 (df)	P value
Age	256.66 (1)	<0.001	991.53 (1)	<0.001	857.12 (1)	<0.001	49.74 (1)	<0.001	59.35 (1)	<0.001	2.43 (1)	0.119
Sex	73.73 (1)	<0.001	337.35 (1)	<0.001	364.87 (1)	<0.001	29.22 (1)	<0.001	38.73 (1)	<0.001	1.55 (1)	0.213
Education	1.50 (1)	0.221	341.07 (1)	<0.001	149.48 (1)	<0.001	91.83 (1)	<0.001	45.01 (1)	<0.001	9.25 (1)	0.002
Marital status	26.81 (1)	<0.001	340.65 (1)	<0.001	237.69 (1)	<0.001	55.03 (1)	<0.001	34.56 (1)	<0.001	2.17 (1)	0.141
Living arrangement	136.62 (1)	<0.001	24.06 (1)	<0.001	225.30 (1)	<0.001	28.64 (1)	<0.001	1.18 (1)	0.278	47.66 (1)	<0.001
Residence location	31.44 (1)	<0.001	49.78 (1)	<0.001	3.59 (1)	0.058	96.68 (1)	<0.001	9.64 (1)	0.002	47.06 (1)	<0.001

P values were derived using the Wald χ² test based on the Bolck, Croon and Hagenaars method for pairwise comparisons.

^*χ² statistics evaluate equality of means between pairs of ageing profiles (classes) for the continuous variables (age at last interview and education). For the categorical variables (ie, sex, marital status, living arrangement and residence location), χ² statistics examine equality of probabilities.

Class 1, the healthy-and-supported profile; Class 2, the cognitively-competent-but-functionally-dependent profile; Class 3, the functionally-capable-but-cognitively-challenged profile; Class 4, the frail-and-support-lacking profile.

Predicting the end-of-life and death circumstances from ageing profiles

Together, figure 2 and the last four columns of table 1 summarise end-of-life and death circumstances by ageing profiles. Additionally, table 4 and online supplemental table S2 (online) display the results of pairwise statistical comparisons, providing χ² statistics and ORs that quantify the significant differences between these profiles. As shown in figure 2A, the healthy-and-supported profile (Class 1) survived the longest between the last interview and death (mean=19.78 months, SD=11.25), followed by Classes 3 and 2 (the functionally-capable-but-cognitively-challenged and cognitively-competent-but-functionally-dependent profiles). The frail-and-support-lacking profile (Class 4) sustained the shortest duration (mean=14.46 months, SD=11.04). End-of-life experiences were better for the healthy-and-supported and functionally-capable-yet-cognitively-challenged profiles (Classes 1 and 3), given that participants of these two types were substantially less prone to be bedridden in their final months (72.1% and 73.1%, respectively) than others (figure 2B). Regarding the place of death (figure 2C), the healthy-and-supported individuals had the largest proportion of deaths in hospitals (9.7%). Contrastingly, the functionally-capable-but-cognitively-challenged profile was least likely to die in hospitals (4.8%) but most likely to die at home (92.1%). When dying, the healthy-and-supported category showed a markedly worse prospect of experiencing pain (16.0%) than those from other categories. Instead, delirium at death was most common for the frail-and-support-lacking profile (41.8%; figure 2D).

Figure 2. Comparisons of end-of-life and death circumstances across ageing profiles. (A) Displays the mean length of survival (months) after the last interview, with error bars representing SD. (B, C and D) Present the proportions (%) of participants in each profile regarding being bedridden in the final 3 months, place of death (home, hospital or other) and death experiences (painful death and symptoms of delirium), respectively. Class 1, the healthy-and-supported profile; Class 2, the cognitively-competent-but-functionally-dependent profile; Class 3, the functionally-capable-but-cognitively-challenged profile; Class 4, the frail-and-support-lacking profile.

Table 4. Pairwise comparisons of end-of-life outcomes and death circumstances between any two ageing profiles.

	Class 1 versus Class 2		Class 1 versus Class 3		Class 1 versus Class 4		Class 2 versus Class 3		Class 2 versus Class 4		Class 3 versus Class 4
	χ2 (df)*	P value	χ2 (df)	P value	χ2 (df)	P value	χ2 (df)	P value	χ2 (df)	P value	χ2 (df)	P value
Length of survival	99.31 (1)	<0.001	17.60 (1)	<0.001	448.81 (1)	<0.001	30.72 (1)	<0.001	43.72 (1)	<0.001	190.07 (1)	<0.001
Being bedridden	34.53 (1)	<0.001	0.35 (1)	0.554	77.95 (1)	<0.001	19.68 (1)	<0.001	2.71 (1)	0.100	42.72 (1)	<0.001
Place of death	6.71 (2)	0.035	95.48 (2)	<0.001	9.79 (2)	0.007	16.39 (2)	<0.001	0.25 (2)	0.882	19.83 (2)	<0.001
Painful death	17.07 (1)	<0.001	26.03 (1)	<0.001	28.11 (1)	<0.001	0.04 (1)	0.849	0.32 (1)	0.572	0.19 (1)	0.667
Symptoms of delirium	11.76 (1)	0.001	3.16 (1)	0.076	51.38 (1)	<0.001	2.84 (1)	0.092	6.72 (1)	0.010	22.00 (1)	<0.001

P values were derived using the Wald χ² test based on the BCH method (employing Vermunt’s robust estimation for continuous variables and the three-step approach for categorical variables). For the categorical variables (ie, being bedridden within three months prior to death, place of death, pain or symptoms of delirium at death), χ² statistics examine equality of probabilities.

^*χ² statistics evaluate equality of means between pairs of ageing profiles (classes) for the continuous variable (months of survival after the last interview).

BCH, Bolck, Croon and Hagenaars; Class 1, the healthy-and-supported profile; Class 2, the cognitively-competent-but-functionally-dependent profile; Class 3, the functionally-capable-but-cognitively-challenged profile; Class 4, the frail-and-support-lacking profile.

Discussion

In this retrospective study using data from the CLHLS, a longitudinal survey with a national scope, we identified four distinct categories among the oldest-olds in mainland China, drawing on WHO’s composite model of healthy ageing. Four categories emerged: healthy-and-supported (57%), cognitively-competent-but-functionally-dependent (12%), functionally-capable-but-cognitively-challenged (18%) and frail-and-support-lacking (13%).

As illustrated by the profiles of each category, heterogeneity in ageing experiences primarily revolved around both IC dimensions (functional and cognitive capacities) and environmental characteristics (the economic/financial dimension), highlighting the validity and utility of integrating individual-based and environment-based indicators when depicting ageing as heterogeneous experiences.^{2 14 22} While the environmental indicators of the economic/financial dimension in this study were subjective in nature, their alignment with objective sociodemographic correlates (ie, educational attainment and rural vs urban residence) in expected directions supports their construct validity. Importantly, such subjective assessments capture older adults’ perceived environment, which may shape health experiences and ageing trajectories more directly than objective infrastructure alone in later life.^{34 35}

Further inspections of the profiles revealed that all participants displayed relatively low levels of social capacity, a deficit in IC further exacerbated by insufficient social services from local neighbourhoods. Since all IC and environmental indicators were based on participants’ self-reports obtained during their last interview while alive, the observed reports of limited social service might capture more than the actual absence of services. They may also indicate inadequate awareness, accessibility or utilisation of existing resources, as documented in gerontological research, particularly among the oldest old.^{8 9 36} Admittedly, such limited variance meant that the social aspect of environmental factors contributed less to the development of an ageing typology in China. However, this result emphasised a practical need to elevate social well-being for older people across the country, which should improve older adults’ adaptation to later life.^{15 36} From a policy perspective, these findings underscore the need not only to expand community-based social services where they are insufficient, but also to improve outreach, information dissemination and accessibility of existing services to ensure that older adults are aware of and able to benefit from available support.

The sociodemographic correlates of the emergent categories were compared, yielding three patterns of nuanced relations between certain sociodemographic features and specific ageing profiles. First, the odds of belonging to Class 1 (consisting of individuals with moderate to high IC and generally sufficient environmental resources) were higher for men, those being married and living alone. The latter two characteristics were not contradictory in that marriage has often been found to be a protective factor against ill health, and healthy, functioning older people are indeed capable of living on their own.^{4 15 22 26 27} The second discovery was that women and the unmarried (mostly widowed) tended to cluster in the cognitively challenged group (Class 3) and frail-and-support-lacking category (Class 4). Such tendencies were comparable to previous studies on older people in mainland China, where gender and marital status were reliably found to relate to IC scores.^{1 4 14 22} The third finding was based on Classes 2 and 3, which contrasted each other along two IC domains: functional and cognitive abilities. More specifically, the two categories diverged in residence location: the cognitively competent type (Class 2) had the largest proportion of urban dwellers, while the functionally capable group (Class 3) contained the largest percentage of rural residents. This urban–rural divide manifests at the macro (eg, economic development and social services) and micro (eg, residents’ health and quality of life) levels alike, as empirical evidence has revealed that individuals in rural regions generally received less education, lower income and inadequate social services.412 22 37,39

The rural–urban difference persisted till the end of life: out of the entire sample, participants from Class 3 were least likely to die in hospitals (4.8%) but most likely to die at home (92.1%). Relevant research showed that the higher chance of home deaths in rural settings might have resulted from issues with the accessibility of hospitals and the availability of institutional services.^{40 41} Another finding worth noting regarding end-of-life outcomes and death quality was how individuals from Class 1, the healthy-and-supported profile, most often died in hospitals and appeared most prone to painful deaths. It seemed further surprising because they survived the longest since the last interview and were less likely to be bed-bound within 3 months prior to death. Nonetheless, according to previous research, older participants dying in hospitals could be in pain due to negligence of their psychological and spiritual needs, because nurses and physicians focused on symptom control and prolonging lives.41,43 Considering that the most frequent death places for older participants in Class 1 were indeed hospitals, it seemed plausible that they were at a greater risk of painful deaths.

Beyond the potential downsides of aggressive medicalisation, there are alternative interpretations of the counter-intuitive finding that the healthy-and-supported profile had the highest likelihood of dying in hospital while also being the most pain-prone when approaching death. For example, this association likely reflects a selection bias: older adults with the most severe, clinically complex and painful conditions are often those prioritised for hospital admission.²⁹ Furthermore, as our measure relied on proxy reports, it is possible that family caregivers did not strictly distinguish physical sensation from the psychological or spiritual distress inherent in a hospitalised death.⁴¹ Thus, the elevated pain reports may capture a broader spectrum of suffering experienced by both the patient and the family in the hospital setting.

This retrospective study was among the first to characterise the heterogeneity of ageing profiles and associate healthy ageing patterns with end-of-life and death circumstances among the oldest-olds in mainland China with a person-centred approach. Still, two limitations exist, and further research is called on. First, this investigation capitalised on available measures from a nationwide panel survey to establish operational definitions of healthy ageing, but unfortunately, not all ideal indicators were present in this existing dataset. Future studies might consider alternative, equally comprehensive variables to depict IC and the living environment. For example, subjective well-being was measured by three indicators: self-rated life satisfaction, negative affect and positive affect. Although these items were considered adequate assessments of subjective well-being,⁴⁴ future research can employ specialised measures to capture the construct more accurately, such as the Satisfaction with Life Scale.⁴⁵ A similar issue exists for death-related outcomes. Despite interviewer training and the use of standardised instruments, retrospective proxy reports of pain and delirium-like symptoms prior to death remain subject to residual recall and reporting bias, which may have influenced the accuracy of symptom ascertainment. Another drawback of the current work lies with its sample characteristics, which constrain the relevance and generalisability of the study’s findings. Most of this sample was community-dwelling older adults. Instead, older people in nursing homes with a higher likelihood of cognitive impairment, a broader range of chronic conditions and consequently lower capacities of responding to face-to-face interview items merit attention from future researchers. Therefore, researchers might focus on ageing profiles of nursing home residents and delineate relations between ageing profiles and end-of-life outcomes for this population.

Conclusion

Rapid population ageing and an increase in the number of the oldest old pose challenges to health and social service systems in many developed and developing countries. The findings of the current study illustrated how a person-centred method could provide insights regarding heterogeneity in healthy ageing profiles. Boosting overall social participation and developing services mitigating functional and cognitive losses are key to promoting healthy ageing. Improving end-of-life experiences requires targeted measures in rural and urban contexts, respectively.

Data availability statement

Data are available in a public, open access repository.