Disability severity transitions in the community-dwelling Swiss ageing population: secondary analysis of the Swiss version of the Survey of Health, Ageing and Retirement in Europe (SHARE) using multistate models

Abstract

Objective

To describe disability severity transitions in the ageing population in Switzerland using an overall functioning score to define four disability severity states (no, mild, moderate and severe) and death, and to investigate the association of multimorbidity and further predictors with these transitions.

Design

Secondary analysis of the Swiss version of the Survey of Health, Ageing and Retirement in Europe (SHARE).

Setting

Switzerland.

Participants

Community-dwelling population aged 50+ with at least two interviews in SHARE (N=3505).

Interventions

Not applicable.

Main outcome measures and methods

Primary outcome measures are the disability severity as assessed by a previously developed overall functioning score, and death status as assessed by the SHARE end-of-life interview. Transition analysis between disability severity states and death was conducted using multistate Markov models. The association between predictor variables and transition intensities was quantified using the proportional hazards assumption. Two distinct operationalisations of multimorbidity (count, burden) were used and analysed according to two separate models (A, B).

Results

The findings for both models were similar: Estimated HRs for transition intensities suggest that being multimorbid or having a higher disease burden score increases the risk of transitioning to higher disability severity states and death for most transitions (HRs between 0.90 and 2.34 for model A compared with not being multimorbid; HRs between 0.95 and 1.46 for model B for a one-point increase in the disease burden score). In addition, most transitions to higher disability severity states and death are more likely for higher age (HRs between 1.00 and 1.14 for model A, and between 1.00 and 1.15 for model B for a 1 year increase in age), and transitions to death are less likely for women, compared with men (HRs between 0.34 and 0.88 for model A, and between 0.38 and 0.71 for model B).

Conclusions

This study is a first attempt to understand disability severity transitions in the older population in Switzerland. Although we believe that such an approach is suitable to inform resource allocation to LTC, rehabilitation and prevention, more detailed information on contextual factors will be important to consider for future research. Moreover, our study contributes to the discussion on how to operationalise multimorbidity in healthy ageing research.

Strengths and limitations of this study.

• Two operationalisations of multimorbidity were examined, based on a range of different health conditions, including mental disorders. To our knowledge, this is the first study to propose a weighted measure of multimorbidity (burden) based on estimated disability weights from the Global Burden of Disease Study. This operationalisation is easy to use; it can be applied and adapted to different settings, and it may contribute to better replicability and comparability of results across studies.

• Multistate Markov models incorporated a possible time-period association to account for the COVID-19 pandemic.

• In the Swiss Survey of Health, Ageing and Retirement in Europe, people living in nursing homes are under-represented, meaning our results cannot be directly transferred to this population. We use the term ‘community-dwelling’ to highlight this aspect of our study.

• The definition and interpretation of disability severity states depended on the specific distribution of the overall functioning score in the sample, which may limit the generalisability of findings.

• The proposed multistate Markov models considered transitions between consecutive disability states only and had to be further simplified due to limited number of observations for the transition from no disability to death.

Introduction

The population in Switzerland is rapidly ageing. Recent population scenarios show that the ageing of the population is progressing and is expected to accelerate between 2020 and 2030¹: the number of persons aged 65 and older is expected to increase from 1.6 to 2.1 million between 2020 and 2030, and to 2.7 million by 2050. This corresponds to an increase in the share of this age group in the total population from 18.9% to 25.6% between 2020 and 2050. These demographic changes and the associated increase in the prevalence of age-related and chronic health conditions have direct implications for healthcare systems and health policy in Switzerland, for example in the provision and organisation of long-term care (LTC).

By declaring 2020–2030, the Decade of Healthy Ageing, the United Nations (UN) urges countries to broaden their understanding of ageing and health. In its baseline report for the Decade of Healthy Ageing, the WHO makes clear that healthy ageing is not simply about living longer, but about improving intrinsic capacity and functional ability that enables well-being.² The conceptual difference between a person’s intrinsic capacity and their functional capacity is the context in which they live their lives: Whereas a person’s intrinsic capacity is the sum of all physical and mental capacities available to the individual, a person’s functional ability is the expression of their intrinsic capacity in interaction with their unique environments. WHO, therefore, emphasises the need for comparable and comprehensive information on intrinsic capacity and functional ability to assess and monitor healthy ageing in populations. However, agreed on measures and statistics that provide comprehensive information are not yet widely available, and most studies rely on the common health indicators of mortality and morbidity.

Functioning, which is conceptually equivalent to functional ability, was introduced in 2001 by WHO as a comprehensive operationalisation of health and proposed as the third health indicator next to mortality and morbidity. According to WHO’s International Classification of Functioning, Disability and Health (ICF),³ functioning is described as the outcome of a person’s health condition (diseases, disorders, injuries, trauma, ageing, etc) and the person’s context (environmental and personal factors). In other words, the concept of functioning aims to describe a person’s lived experience of health. In this understanding, functioning serves as an umbrella term to encompass all human body functions and structures, activities and participation in major life areas. Similarly, in the ICF, disability is used as an umbrella term when limitations in functioning are present, for example, impairments in body functions or structures, limitations in activities and/or participation restrictions. Importantly, functioning and disability are understood as a matter of degree, a continuum, ranging from low to high functioning or from high to low disability levels.

By understanding and operationalising functioning and disability as continua, it is possible to use thresholds to define groups characterised by a certain level of disability. Among different approaches to set such thresholds, population-based cut-offs have been used to differentiate and compare groups of no, mild, moderate and severe disability levels.^{4 5} Importantly, the ICF proposes a dynamic understanding of disability, where the level of experienced disability can change over time for individuals and populations considering their health conditions and environment. Thus, understanding how disability severity changes over time, that is, the transitions between different disability levels and their predictors, is highly relevant for understanding the healthcare needs of frequently multimorbid ageing populations.

Several studies have investigated the impact of multimorbidity on ageing populations. Literature reviews commonly show that there exists neither consensus on the definition of multimorbidity, nor on its operationalisation.6,10 The most widely applied definition, which is also accepted and used by WHO, is ‘the coexistence of two or more chronic conditions in the same individual’.¹¹ The operationalisation of multimorbidity varies not only regarding the definition of chronicity, the list of health conditions to be considered or how they are measured/assessed, but also regarding the number of diseases to define cut-offs for individuals to qualify as being multimorbid. These ambiguities are also reflected in the measures used to assess multimorbidity. Roughly, measures of multimorbidity can be distinguished according to two main types^{8 10 12}: Measures based on simple counts of health conditions present in an individual, and measures based on a weighting of health conditions to account for their relative severity or their impact on an individual (sometimes also known as measures of morbidity burden¹³).

This paper builds on two assumptions. The first is that the concept of functioning has the potential to improve our understanding of the impact of multimorbidity on the health of the ageing population. The second is that a more nuanced operationalisation of multimorbidity that accounts for the burden of health conditions and their combinations may enable a more relevant quantification of its impact on functioning and the disability levels experienced by the ageing population. The objective of this study was therefore to describe disability severity transitions in the ageing population in Switzerland using a previously developed overall functioning score¹⁴ to define four disability severity states: no, mild, moderate and severe. Specifically, we aimed to estimate the transition rates between different disability states and death and to investigate the association of multimorbidity and further predictors with these transitions.

Methods

Study design and population

This study used data from the Swiss version of the Survey of Health, Ageing and Retirement in Europe (SHARE, Release 9-0-0, accessed in October 2024, available for research purposes after registration at the SHARE Research Data Centre). SHARE targets individuals aged 50+, who speak one of the countries’ main official languages (German, French and Italian in Switzerland), and who have their regular domicile in the corresponding SHARE country, as well as their spouse or partner, regardless of age. Notably, people living in nursing homes are particularly included in the target population of SHARE. However, their inclusion depends on a country’s sampling frame, which can lead to an under-representation of this population in certain countries. Individuals are excluded if they are imprisoned, hospitalised or have moved to an unknown address. In Switzerland, SHARE currently comprises nine waves of data collection, conducted between 2004 and 2022. Detailed information on the SHARE data collection, eligibility criteria and methodology can be found elsewhere.^{15 16}

We included data from all SHARE waves 1–9.17,25 Study-specific exclusion criteria were: (1) participants and observations with no valid interview (neither main interview nor end-of-life interview, which is conducted with a proxy-respondent after a participant’s death), (2) participants with age at study entry below 50 years, (3) participants who had their first interview in wave 3 and observations of wave 3 other than end-of-life interviews (wave 3 collected different information based on the SHARELIFE interview, which does not include information on disability as operationalised in this study), (4) participants with no or only one valid functioning score and (5) participants with implausible interview or death dates (including no death date assessed).

Measures

The primary outcomes of this study are disability as assessed by a previously developed overall functioning score using longitudinal data from SHARE¹⁴ and death. Disability was categorised into different levels of severity, as previously proposed, by using the distribution of observed functioning scores in the study population.^{4 5} Specifically, we used the mean and SD of all available functioning scores in the full SHARE sample at wave 1 of persons aged 50+ to define the following disability severity states:

• No disability: functioning score≥mean+1 SD.

• Mild disability: mean+1 SD>functioning score≥mean.

• Moderate disability: mean>functioning score≥mean–1 SD.

• Severe disability: functioning score<mean–1 SD.

Death status was assessed through the SHARE end-of-life interview, which reports the year and month of a participant’s death and is assessed at all SHARE waves.

Predictors of interest included age at interview (in years), sex (male, female), level of education as a proxy for socioeconomic status (low, medium, high based on the levels of the International Standard Classification of Education 1997), interview language as a proxy for residency (German, French and Italian were grouped for practical reasons), living status (living alone for household size <2, living not alone for household size ≥2), and multimorbidity. For the latter, two operationalisations were applied according to the health conditions available in SHARE (ever diagnosed/currently having: heart attack including myocardial infarction and coronary thrombosis/heart problem including congestive heart failure; stroke/cerebral vascular disease; diabetes/high blood sugar; chronic lung disease; asthma; cancer or malignant tumour (excluding minor skin cancers); stomach/duodenal/peptic ulcer; Parkinson’s disease; cataracts; hip fractures; other fractures; Alzheimer’s disease/dementia/organic brain syndrome/senility or other serious memory impairment; rheumatoid arthritis; osteoarthritis; chronic kidney disease) and the dichotomised version of the Europe Depression scale to indicate clinical depression. The two operationalisations were derived as follows:

• Multimorbidity, binary operationalisation (count): Indicates whether a person has two or more of the health conditions listed above.

• Multimorbidity, continuous operationalisation (burden): Indicates the burden of living with none, one or several of the health conditions listed above, based on a score between 0 and 10 (higher values indicate greater disease burden). This operationalisation was derived from the (combined) disability weights published regularly by the Global Burden of Disease Study (GBD).²⁶ Disability weights are intended to represent the amount of lost health associated with a given health condition and are measured on a continuous scale from 0 to 1 (full health to death). In this study, for each health condition in SHARE, the corresponding health states in the GBD and their severity levels were identified. The corresponding GBD disability weights were then averaged across severity levels to derive a single disability weight estimate per health condition in SHARE. If available, we used the most recent disability weights reported by the GBD 2021.²⁶ For combinations of health conditions, the averaged disability weights of each health condition were combined according to a published formula for cumulative disability weights (online supplemental appendix 1, p. 46 of Ferrari et al²⁷). This multiplication approach ensures that the combined disability weights are greater than each of the individual weights, while maintaining their interpretability on the continuous 0 to 1 scale. In line with the GBD, disability weights were set to extreme values of 0 and 1 for persons with no health condition and for persons who had died, respectively. Finally, the averaged (combined) disability weights were multiplied by 10 to facilitate the interpretation of results and served as continuous disease burden scores (see online supplemental table 1 for more information).

There is no information available in SHARE on the duration/chronicity of the listed health conditions; therefore, we have not included any requirements in this respect in our multimorbidity operationalisations. Information on further health conditions (ever diagnosed/currently having: hypertension, high blood cholesterol, osteoporosis) is available in SHARE. However, these conditions are understood in the GBD as risk factors and have no disability weights. In this study, we also considered these conditions as risk factors for other health conditions and reported them in the sample characteristics. However, they were not included in the data analysis.

Missing data management and imputation

SHARE-specific codes such as ‘don’t know’, ‘refusal’, etc were treated as missing observations for all variables. Before applying the study-specific exclusion criteria, for age at interview, interview language and level of education, missing data imputation was performed as follows: Missing observations for age at interview were imputed from other information in SHARE where available (birth date, interview date, etc.) as described elsewhere.²⁸ Although in SHARE the level of education is recorded only at the first interview, its value is also reported in later waves due to generated variables in some modules. To reduce missingness, missing observations of education and interview language were replaced by the previous or next non-missing value of the respective variable for an individual.

In SHARE, the set of assessed health conditions has changed across waves and thus not every health condition was assessed at every wave. To ensure the comparability of multimorbidity operationalisations across waves, the following imputation strategy was used: First, all health conditions were distinguished between chronic (eg, diabetes) and temporary (eg, hip fracture). Second, if a health condition was considered chronic, its first occurrence in an individual was manually carried forward in time and then its last non-occurrence in an individual was manually carried backward in time. Third, for all health conditions, single imputation was performed by using MissForest, a non-parametric and iterative method based on random forest²⁹ (see online supplemental table 2 for more information). The multimorbidity predictors were derived based on the imputed set of health conditions. We did not perform imputation of missing observations for any other of the predictor variables of interest.

Statistical analysis

Descriptive analysis was conducted for the study sample at baseline (ie, the first observation of each participant). An initial data analysis was performed following the work of Lusa et al^{28 30} within the STRengthening Analytical Thinking for Observational Studies (STRATOS) initiative.

Transition analysis between disability severity states and death was conducted using multistate Markov models for panel data.³¹ Such models are useful for describing a continuous process over time, where the individual is assumed to move between a series of discrete states and is observed at arbitrary points in time. They are fitted by estimating transition intensities between states (ie, the instantaneous risk of transitioning from a current state to one of the other states) and under the Markov assumption that an individual’s future transition process depends only on its current state. Our proposed multistate model includes 5 states in total (4 disability severity states, and death as absorbing state) and 10 transitions (all forward and backward transitions between consecutive disability states as well as transitions from each disability state to death). In other words, we assumed that all participants, while alive, could progress to or recover from any of the consecutive disability states, and that they could die in any of the disability states. The detailed multistate model is shown in figure 1.

Figure 1. Illustration of the proposed five-state multistate model for possible disability severity transitions in the ageing population in Switzerland. Boxes stand for different disability severity states and the absorbing state of death, arrows indicate proposed transitions between states. Dashed lines indicate transitions that had to be removed during the transition analysis as it was not possible to estimate corresponding predictor associations precisely. The final, simplified multistate Markov model is represented with solid lines (see Results section for more details).

Missing values in the disability severity states at any observation time point while alive were assumed to be censored, meaning that the exact state was unknown but within the possible disability states. Observation times were expressed according to the reported SHARE interview and death dates (year and month), respectively. Since some of the interviews for waves 8 and 9 were conducted during the COVID-19 pandemic, we allowed transition intensities for every individual to change at the onset of the pandemic (date of WHO declaration of pandemic on 11 March 2020).

To quantify the association between the predictor variables and the disability severity transition intensities, the multistate model is using the proportional hazards assumption.¹ Two multistate models with different sets of predictors were fit: Model A included the described predictors of interest and the binary operationalisation of multimorbidity (count), Model B included the described predictors of interest and the continuous operationalisation of multimorbidity (burden). For both models, all predictor associations were allowed to be transition-specific, that is, predictor associations were estimated separately for each transition. Age at interview, multimorbidity and living status were time-dependent covariates. Predictor associations were visually explored according to predicted transition probabilities over a given time period, that is, the likelihood of being in a particular state at a particular time point in the future, given an individual’s current state. To ensure that the model estimates converged to a global optimum, different sets of initial values for the transition intensities were tested, and for both Models A and B the final model was selected based on the highest log-likelihood. The final models A and B were compared with the corresponding model without any predictor associations (ie, the null model) and with each other using the Akaike Information Criterion (AIC, preference for lower values). Furthermore, we conducted a sensitivity analysis of the final models A and B using bootstrap sampling (N=1000) to assess the robustness of the estimated associations between the predictor variables and the transition intensities, according to 95% bootstrap CIs (BCIs).

Analyses were performed using R V.4.4.2 for Windows.³² Transition analysis was conducted using the R package msm V.1.8.2.³¹ Missing data imputation was undertaken by using the R package missForest V.1.5.²⁹ Visualisations were generated using the R packages ggplot2 V.3.5.1,³³ forestploter V.1.1.2³⁴ and UpSetR V.1.4.0.^{35 36}

The reporting of this study followed the guidelines about Strengthening the Reporting of Observational Studies in Epidemiology (STROBE).³⁷

Patient and public involvement

No patients or members of the public were involved in the study.

Results

Sample characteristics

In total, we downloaded 27 644 observations of 7187 unique participants in Switzerland from the SHARE Research Data Centre. The mean functioning score of persons aged 50+ at wave 1 (N=938) was 70.74 (SD=11.58). Participants without a valid interview (N=2251), with an age at study entry below 50 years (N=487), with the first interview in wave 3 (N=1), with no or only one functioning score (N=911), and with implausible interviews or death dates (N=32) were excluded, resulting in a sample size of 3505 individuals with a total of 16 457 observations. The average number of observations per person was 4.7 (see online supplemental table 3) and the study sample included 410 participants whose death was recorded.

Descriptive information of the study sample is presented at baseline (table 1; see online supplemental table 4 and 5 for baseline characteristics before missing data imputation and for age groups, respectively). Median age at baseline was 62 years (IQR=55–71 years), and most interviews were provided in German language (71.0%). More than half of the study participants were female (52.9%), and most participants reported a medium level of education (61.4%). The majority of participants reported a mild (43.0%) or moderate (35.4%) level of disability at baseline (online supplemental figure 1 shows the detailed longitudinal distribution of disability severity states). One in five persons was having more than two health conditions at baseline (20.1%), and the mean multimorbidity score was 1.34 (SD=1.76). An overview of the distribution of multimorbidity for the full sample can be found in online supplemental figure 2.

Table 1. Overview characteristics of participants at baseline interview and for male and female participants, respectively.

Characteristics	Baseline (N=3505)	Male (N=1652)	Female (N=1853)	Missing (%)
SHARE wave, n (%)				0.0
1	721 (20.6)	341 (20.6)	380 (20.5)
2	541 (15.4)	235 (14.2)	306 (16.5)
4	2109 (60.2)	1004 (60.8)	1105 (59.6)
5	8 (0.2)	5 (0.3)	3 (0.2)
6	5 (0.1)	2 (0.1)	3 (0.2)
7	2 (0.1)	0 (0.0)	2 (0.1)
8	119 (3.4)	65 (3.9)	54 (2.9)
Age at interview,* median (IQR)	62.00 (55.00, 71.00)	62.00 (56.00, 70.25)	62.00 (55.00, 71.00)	0.0
Age at interview,* n (%)				0.0
50–59	1450 (41.4)	660 (40.0)	790 (42.6)
60–69	1083 (30.9)	538 (32.6)	545 (29.4)
70–79	690 (19.7)	334 (20.2)	356 (19.2)
80+	282 (8.0)	120 (7.3)	162 (8.7)
Education level,* n (%)				1.0
Low	774 (22.3)	242 (14.8)	532 (29.0)
Medium	2131 (61.4)	1051 (64.2)	1080 (59.0)
High	565 (16.3)	345 (21.1)	220 (12.0)
Interview language*=FR/IT, n (%)	1017 (29.0)	479 (29.0)	538 (29.0)	0.0
Living alone=Yes, n (%)	763 (21.8)	253 (15.3)	510 (27.5)	0.0
High blood pressure=Yes, n (%)	995 (28.4)	517 (31.3)	478 (25.9)	0.2
High blood cholesterol=Yes, n (%)	524 (15.0)	279 (16.9)	245 (13.3)	0.2
Osteoporosis=Yes, n (%)	71 (5.6)	7 (1.2)	64 (9.4)	64.1
Body mass index, n (%)				1.2
Implausible/suspected wrong	42 (1.2)	19 (1.2)	23 (1.3)
Below 18.5—underweight	78 (2.3)	7 (0.4)	71 (3.9)
18.5–24.9—normal	1582 (45.7)	614 (37.3)	968 (53.3)
25–29.9—overweight	1304 (37.7)	774 (47.0)	530 (29.2)
30 and above—obese	456 (13.2)	232 (14.1)	224 (12.3)
In nursing home (last year), n (%)				0.6
Yes, temporarily	8 (0.2)	6 (0.4)	2 (0.1)
Yes, permanently	3 (0.1)	2 (0.1)	1 (0.1)
No	3473 (99.7)	1640 (99.5)	1833 (99.8)
Multimorbidity (count)*=Yes, n (%)	705 (20.1)	294 (17.8)	411 (22.2)	0.0
Multimorbidity (burden),* median (IQR)	0.00 (0.00, 3.00)	0.00 (0.00, 2.20)	0.07 (0.00, 3.00)	0.0
Multimorbidity (burden),* mean (SD)	1.34 (1.76)	1.18 (1.69)	1.48 (1.82)	0.0
Functioning score, median (IQR)	72.10 (62.70, 76.50)	72.10 (65.30, 76.50)	68.40 (62.70, 76.50)	0.1
Disability severity state, n (%)				0.1
No	279 (8.0)	169 (10.2)	110 (5.9)
Mild	1507 (43.0)	763 (46.2)	744 (40.2)
Moderate	1241 (35.4)	549 (33.2)	692 (37.4)
Severe	474 (13.5)	171 (10.4)	303 (16.4)

^*After missing data management and imputation.

FR, French; IT, Italian; SHARE, Survey of Health, Ageing and Retirement in Europe.

Table 2 shows the prevalence of the included health conditions for the operationalisation of multimorbidity at baseline interview and before the missing data imputation. Details on missing data imputation are shown in the online supplemental figures 3–5. Distributions and combinations of health conditions after missing data imputations are detailed in the online supplemental figure 6.

Table 2. Overview of health conditions (before missing data imputation) at baseline interview and for male and female participants, respectively.

Characteristics	Baseline (N=3505)	Male (N=1662)	Female (N=1853)	Missing (%)
Heart attack/problem=Yes, n (%)	259 (7.4)	156 (9.4)	103 (5.6)	0.2
Stroke/cerebral vascular disease=Yes, n (%)	93 (2.7)	53 (3.2)	40 (2.2)	0.2
Diabetes/high blood sugar=Yes, n (%)*	222 (6.3)	135 (8.2)	87 (4.7)	0.1
Chronic lung disease=Yes, n (%)*	128 (3.7)	66 (4.0)	62 (3.3)	0.0
Asthma=Yes, n (%)*	48 (3.8)	22 (3.8)	26 (3.8)	64.1
Cancer/malignant tumour=Yes, n (%)	211 (6.0)	97 (5.9)	114 (6.2)	0.2
Stomach/duodenal/peptic ulcer=Yes, n (%)	85 (2.4)	39 (2.4)	46 (2.5)	0.2
Parkinson’s disease=Yes, n (%)*	11 (0.3)	6 (0.4)	5 (0.3)	0.0
Cataracts=Yes, n (%)	305 (8.7)	117 (7.1)	188 (10.2)	0.2
Hip fracture=Yes, n (%)	43 (1.2)	17 (1.0)	26 (1.4)	0.2
Other fractures=Yes, n (%)	234 (8.4)	124 (9.5)	110 (7.5)	20.7
Alzheimer’s disease/dementia=Yes, n (%)*	16 (0.5)	8 (0.5)	8 (0.4)	0.8
Affective/emotional disorders=Yes, n (%)	7 (5.2)	3 (4.2)	4 (6.5)	96.2
Rheumatoid arthritis=Yes, n (%)*	3 (0.1)	0 (0.0)	3 (0.2)	14.9
Osteoarthritis=Yes, n (%)*	23 (0.9)	6 (0.5)	17 (1.3)	28.4
Chronic kidney disease=Yes, n (%)*	2 (0.1)	1 (0.1)	1 (0.1)	21.1
Clinical depression=Yes, n (%)	636 (18.3)	204 (12.4)	432 (23.5)	0.7

^*Indicates that this condition was treated as a chronic health condition in the imputation of missing data.

Transition analysis

An overview of observed transitions is given in table 3. Note that the proposed multistate model does not cover all observed transitions, but the model assumes that all transitions between non-consecutive states can be represented by a combination of transitions between consecutive states. For example, although the transition from mild to severe disability is not directly represented in the proposed model, it is assumed that this transition occurs via the moderate disability state, that is, from mild to moderate to severe disability. Generally, and as expected for an ageing population, forward transitions (to states of higher disability or death) are more frequently observed than backward transitions (to states of lower disability). Similarly, transitions to death are more frequently observed in states of higher disability severity than in states of lower disability severity.

Table 3. Frequency table of observed transitions in the study sample (percentages in brackets).

	To:
From:	No	Mild	Moderate	Severe	Death	Censored*	Total
No	213 (1.64)	529 (4.08)	94 (0.73)	12 (0.09)	2 (0.02)	1 (0.01)	851 (6.57)
Mild	440 (3.40)	3312 (25.57)	1624 (12.54)	239 (1.85)	75 (0.58)	14 (0.11)	5704 (44.04)
Moderate	73 (0.56)	1433 (11.06)	2047 (15.80)	731 (5.64)	120 (0.93)	13 (0.10)	4417 (34.10)
Severe	5 (0.04)	121 (0.93)	472 (3.64)	1136 (8.77)	212 (1.64)	6 (0.05)	1952 (15.07)
Censored*	2 (0.02)	12 (0.09)	8 (0.06)	5 (0.04)	1 (0.01)	0 (0.00)	28 (0.22)
Total	733 (5.66)	5407 (41.75)	4245 (32.77)	2123 (16.39)	410 (3.17)	34 (0.26)	12 952 (100.00)

Bold numbers correspond to the transitions between consecutive states of the proposed multistate Markov model.

^*Censored means that the exact severity state of a person is not known but is either no, mild, moderate or severe (while alive).

We were unable to estimate precise predictor associations for the transition from no disability to death according to the originally proposed multistate Markov model, as this transition had a small number of observations (N=2 observations), and we therefore decided to remove this transition from the multistate model (this latter multistate Markov model is hereafter referred to as the simplified model). As described above, in consequence, the simplified model will describe this transition as a combination of transitions between the corresponding consecutive states. Below we describe the results of the models A (N=3470, AIC=27 442.41, df=81) and B (N=3470, AIC=27 325.05, df=81) according to the simplified multistate Markov model. Comparison with the corresponding null model (N=3505, AIC=29 043.66, df=9) confirmed better fit for both models over the null model, with model B being preferred overall with the lowest AIC.

For both models A and B, age, sex and multimorbidity show distinct patterns in the corresponding estimated HRs, reflecting the predictor associations on the respective transition intensities (see figures2 3). Note that all other predictors are assumed to be constant in the model when interpreting HRs for specific predictors below. The 95% CIs of the results are presented according to the Hessian-based CIs (default option of the R package msm³¹) and the BCIs (sensitivity analysis). We describe a result as significant if it is significant for both types of CIs.

Figure 2. Forest plots showing the associations between predictors of interest (model A, N=3470) and transition intensities according to the simplified multistate Markov model. Black squares represent the estimated HRs, and black bars represent the corresponding 95% Hessian-based CIs. Results of the sensitivity analysis are shown according to grey bars representing the 95% BCIs. The dashed line represents an HR of 1 and indicates the significance of an association if not contained in the corresponding 95% CI/BCI. BCI, bootstrap CI.

Figure 3. Forest plots showing the associations between predictors of interest (model B, N=3470) and transition intensities according to the simplified multistate Markov model. Black squares represent the estimated HRs, and black bars represent the corresponding 95% Hessian-based CIs. Results of the sensitivity analysis are shown according to grey bars representing the 95% BCIs. The dashed line represents an HR of 1 and indicates the significance of an association if not contained in the corresponding 95% CI/BCI. BCI, bootstrap CI.

Increasing age is significantly associated with most transition intensities. For a 1-year increase in age, the increase in the intensities of forward transitions amounts to 2% for the transition from mild to moderate disability (models A and B: HR=1.02, CI=1.02–1.03, BCI=1.02–1.03), and 5% from moderate to severe disability (models A and B: HR=1.05, CI=1.04–1.06, BCI=1.04–1.06), to about 14%–15% for the transition from mild disability to death (model A: HR=1.14, CI=1.08–1.20, BCI=1.01–1.44; model B: HR=1.15, CI=1.10–1.21, BCI=1.02–1.49), and 11%–12% from severe disability to death (model A: HR=1.11, CI=1.09–1.14, BCI=1.08–1.14; model B: HR=1.12, CI=1.10–1.15, BCI=1.09–1.15). In contrast, estimated HRs for age for backward transitions appear to be relatively constant. For a 1-year increase in age, the intensities of transitioning backwards reduce between 1% and 2% (model A) and between 2% and 3% (model B), depending on the specific transition.

Regarding sex differences, the transition analysis shows that intensities of forward transitions to death are significantly lower for women (except for the transition from mild disability to death). Specifically, compared with men, the decrease in the intensities for women is more than 50% for the transition from moderate disability to death (model A: HR=0.34, CI=0.17–0.70, BCI=0.12–0.80; model B: HR=0.40, CI=0.21–0.75, BCI=0.15–0.71) and for the transition from severe disability to death (model A: HR=0.37, CI=0.28–0.51, BCI=0.26–0.51; model B: HR=0.38, CI=0.28–0.52, BCI=0.26–0.52). In contrast, all other intensities of forward transitions to states of increased disability severity are higher for women. The association is most pronounced and significant for the transition from moderate to severe disability, for which the transition intensity is almost 30% higher for women compared with men (model A: HR=1.28, CI=1.09–1.51, BCI=1.08–1.53; model B: HR=1.27, CI=1.08–1.50, BCI=1.05–1.52). There is no distinct pattern between sexes for backwards transitions.

The associations of multimorbidity are not directly comparable between models A and B due to the different underlying operationalisations. According to model A and the operationalisation of multimorbidity as a binary variable (count), being multimorbid significantly increases the intensities of forward transitions from mild to moderate disability by 32% (HR=1.32, CI=1.13–1.55, BCI=1.11–1.56), from moderate to severe disability by 53% (HR=1.53, CI=1.31–1.79, BCI=1.31–1.82), and from severe disability to death by 38% (HR=1.38, CI=1.02–1.88, BCI=1.05–2.05). The transitions from no to mild disability, from mild disability to death, and from moderate disability to death do not appear to be significant for multimorbidity. On the contrary, being multimorbid significantly reduces the intensities of all backward transitions. The association is most pronounced for the transition intensities from mild to no disability, which decrease by 52% for persons with multimorbidity compared with having no multimorbidity (HR=0.48, CI=0.31–0.76, BCI=0.31–0.76).

With regard to model B and the operationalisation of multimorbidity as continuous score (burden), a one-point increase in the disease burden score significantly increases the intensities of all forward transitions except two (from no to mild disability, and from mild disability to death): for the transition from mild to moderate disability by 12% (HR=1.12, CI=1.07–1.17, BCI=1.07–1.17), from moderate to severe disability by 14% (HR=1.14, CI=1.10–1.19, BCI=1.09–1.20), from moderate disability to death by 46% (HR=1.46, CI=1.22–1.76, BCI=1.02–1.90), and from severe disability to death by 18% (HR=1.18, CI=1.09–1.26, BCI=1.10–1.29). On the contrary, a one-point increase in the disease burden score reduces the intensities of all backward transitions with a similar pattern to model A.

For the other predictors, no specific patterns in significance or size of estimated HRs are present. Finally, the assumed changing transition intensities at the onset of the COVID-19 pandemic can be represented by a binary covariate that represents the time periods before and after the onset. The results for this time-period covariate for both models A and B are included in figures 2 and 3 as well. For both models, time-period associations appear to be non-significant for most forward and backward transition intensities. Overall, all transitions appear to be less likely during the COVID-19 pandemic with the most pronounced and significant association for the transition from severe disability to death (model A: HR=0.47, CI=0.27–0.81, BCI=0.21–0.80; model B: HR=0.44, CI=0.25–0.77, BCI=0.20–0.77).

To illustrate the range of predicted transition probabilities across sexes and multimorbidity combinations for both models, figure 4 shows the discrepancies between women and men, without and with multimorbidity, and with low (0) and high (7) multimorbidity scores at current age 65 and being currently in moderate or severe disability states. The values of the other predictor variables were held constant at medium education level, German interview language and not living alone. Overall, the visualisations show that sex and multimorbidity are associated with the transition probabilities of all states, with differences increasing over time. Moreover, the association of multimorbidity appears to be more pronounced for the operationalisation using a burden score (model B), particularly for high burden scores. Note that the predicted transition probabilities should be interpreted with caution as they do not show confidence bands. For more details, online supplemental figures 7–18 show the associations with disability transitions of different combinations of sex and multimorbidity over 10 years for all possible current disability states and three different current ages (55, 65 and 75 years) for models A and B, respectively.

Figure 4. Stacked transition probabilities (y-axis) over a time span of 10 years (x-axis) according to the simplified multistate Markov model (models A and B) for females and males, without and with multimorbidity, and with low (0) and high (7) multimorbidity scores at current age 65 (rows) and assuming either moderate or severe disability as current disability state (lines). The values of the other predictors were set to medium education level, German interview language and not living alone. The term current is used for the time point t=0.

Results of both models A and B were further described in terms of the expected number of visits per disability state, the mean sojourn times per disability state (ie, the estimated average time spent during a single stay in each state), and the total length of stay per disability state (ie, the estimated total time spent in each state), which can be found in online supplemental tables 6–11.

Discussion

This study aimed at describing disability severity transitions in the ageing population in Switzerland by using a previously developed overall functioning score to identify persons experiencing no, mild, moderate or severe disability, and two operationalisations of multimorbidity, the first focusing on disease count, the second focusing on disease burden. Our results suggest that age, sex and multimorbidity are important predictors of disability severity transitions and transitions to death among older people in Switzerland with distinct patterns of associations. The operationalisation of multimorbidity affected the estimated predicted transition probabilities. Specifically, the operationalisation as burden score might have the potential to provide more nuanced information for persons with high disease burden. We included a time-period association to account for the onset of the COVID-19 pandemic. Overall, all transitions appear to be less likely during the pandemic (although with mostly non-significant associations). This might be a consequence of the composition of our sample (community-dwelling people with generally high functioning, mostly aged 50–69 at baseline) and the fact that the indirect effects of the COVID-19 pandemic led to a deficit in mortality in Switzerland, ie, the observed mortality was lower than expected.³⁸ The latter appeared to be particularly pronounced in people aged 40–59 and 60–69. Following the explanation of Riou et al³⁸ indirect effects of the pandemic which lead to this observation might be different reductions of activities and environmental effects (eg, mobility, road traffic, sports, air pollution). However, a final interpretation of these results warrants further investigation.

The impact of multimorbidity on disability transitions corroborates the need to address the increasing prevalence of non-communicable diseases (NCDs), which are frequently chronic, in ageing populations. As a country with one of the highest life expectancies in the world³⁹ and a rapidly ageing population, Switzerland already faces a considerable burden of chronic diseases and multimorbidity in its ageing population. The challenge of increasing NCDs is indeed a strategic policy priority of the Swiss health system and is being addressed in the NCD prevention strategy for 2025–2028 at the federal level.⁴⁰ The key measures of the strategy focus on health promotion and prevention for the general and working age population as well as on their integration in the health system. Results of the present study highlight, however, the importance of addressing the disability experienced by the population with NCDs as soon as possible, for instance with timely rehabilitation, to avoid transitions to states of more severe disability, which are usually linked with low levels of independence and high needs for LTC.

From a policy perspective, information about disability transitions is key for an evidence-informed allocation of resources to health strategies that meet the needs of the ageing population. On one side, there is a need to strengthen LTC infrastructure to effectively manage the complex needs of people with multiple chronic conditions and severe levels of disability, including integrated care models, improved coordination between healthcare providers and social services, and seamless transitions between care settings. Our findings about disability transitions emphasise as well the importance of prioritising interventions like rehabilitation that aim to optimise functional abilities before LTC is needed. Rehabilitation has been defined by WHO as ‘a set of interventions designed to optimise functioning and reduce disability in individuals with health conditions in interaction with their environment’⁴¹ and can be especially relevant, for instance, to avoid or delay transitions to severe disability. Despite its relatively low cost, in comparison to LTC, a better provision of rehabilitation currently faces several challenges in Switzerland.⁴² Finally, policies should support initiatives that promote active ageing, such as access to physical activity programmes, social engagement opportunities and healthy nutrition.⁴³ In this context, public health campaigns promoting healthy lifestyles, regular health screening and early detection and management of chronic conditions are essential.

To our knowledge, this is the first study proposing a weighted measure of multimorbidity based on estimated disability weights of the GBD. This operationalisation has several advantages: As it is based on the regularly updated or refined, global disability weights of the GBD, the operationalisation is easy to use, it can be applied or adapted to different settings, and it may contribute to better replicability and comparability of results across studies. In addition, we show that it provides a more nuanced picture of the impact of different health conditions, which can be highly variable and multiplicative. In the context of functioning and disability of a population, it might be the burden of different health conditions and their synergistic combinations that determines what people can do with their health in their daily lives, or their level of dependency, rather than the simple number of health conditions they have. We need to acknowledge, however, that our weighted measure of multimorbidity is not in line with the current WHO definition of multimorbidity. Nevertheless, previous studies in the general population, for example, have shown that for the operationalisation of multimorbidity, a cut-off of 1 chronic health condition can be valuable for specific tasks, for example, in the prediction of healthcare resource utilisation.⁴⁴

Our study uses a specific approach to the understanding and operationalisation of disability and multimorbidity. Importantly, we did not evaluate or compare the performance of the proposed operationalisations for different research or policy questions, nor whether one operationalisation is more useful or valuable than the other in the context of ageing populations. Previous research has shown that count and weighted measures of multimorbidity performed almost equally effectively in predicting many outcomes.¹² Also, we did not consider other types of multimorbidity measures, such as disease groups or clusters, or the identification of homogeneous groups of persons with similar characteristics.⁷ Finally, some results suggest that we may be missing relevant variables in our model for understanding some of the disability and death transitions. For example, we observed that for multimorbidity, the estimated HR of the transition from moderate disability to death (not significant) is greater than that of the transition from severe disability to death, or that the risk of the transition from mild disability to death is higher (not significant) for persons without multimorbidity than for persons with multimorbidity. These transitions should be further explored by including additional information (eg, severe, unpredictable, health incidents or accidents). Additionally, detailed investigations on the role of environmental factors (eg, assistive devices, supportive relationships, availability of public transport) and potential interactions might be worthwhile to study as well.

Strengths and limitations

This study has several strengths, including the use of two operationalisations of multimorbidity based on a number of different health conditions and mental disorders, the latter of which are rarely considered.¹⁰ Moreover, the estimated multistate Markov models account for effects of the COVID-19 pandemic. Nevertheless, there are limitations to the interpretation of our results. First, the small number of people living in nursing homes in our sample (N=81 in total, temporary and permanent) indicates that this population is under-represented. This means that our results cannot be transferred directly to this population. Moreover, expecting people in nursing homes to have higher levels of disability, we may have underestimated disability severity (ie, persons with low levels of functioning) in our study. Second, the death status reported in SHARE was not entirely consistent across modules. We used the information assessed in the end-of-life interviews as we relied on information about the time of death. Therefore, we may have underestimated the number of people who died. Third, we did not consider the underlying household structure of SHARE in the multistate Markov model, which may have biased the model estimates. Including random effects to model the heterogeneity between households represents a possible extension of the here performed analysis (see³¹ for more information). Fourth, our operationalisation of disability severity states depends on the distribution of the overall functioning score in the study sample, which may limit the generalisability and comparability of our findings. Generally, the disability severity states should be interpreted based on the characteristics and items of the underlying functioning score. Fifth, to arrive at a single disability weight per health condition, we averaged over the corresponding GBD severity levels. The averaged disability weights do not necessarily reflect the actual severity of a health condition in an individual. If more information were available, the corresponding multimorbidity operationalisation could be adapted to take into account the specific GBD severity levels that best describe an individual’s health condition(s). Sixth, we used interview language as a proxy for residency, but grouped Italian and French together, although regional variations such as climate may be different between the two. Seventh, missing data imputation assumed that missing observations are missing at random (MAR). Under this assumption, the likelihood of a response being missing depends on the other observed variables (given the values of the other observed variables), but not on the actual unobserved value of the response. As missing observations in the health condition variables are largely related to the SHARE design, and as we applied direct replacement methods for missing observations using logical rules wherever possible, we believe that the MAR assumption is reasonable for our sample.

Conclusions

This study represents a first attempt to understand disability severity transitions in the older population in Switzerland, using two operationalisations of multimorbidity. The findings show that age, sex and multimorbidity play an important role in disability severity transitions and transitions to death and, in particular, that—although results for different operationalisations of multimorbidity look similar overall—the operationalisation as burden score might have the potential to provide more nuanced information for persons with high disease burden. Although we believe that such an approach is suitable to inform health policy regarding resource allocation to LTC, rehabilitation and prevention, more detailed information on contextual factors will be important to consider for future research. Finally, our study contributes to the discussion on how to operationalise multimorbidity in healthy ageing research.

Data availability statement

Data are available in a public, open access repository.