Social connection and end-of-life outcomes among older people in 19 countries: a population-based longitudinal study

Summary

Background

Social connection is a key determinant of health, but its role in shaping end-of-life outcomes is poorly understood. We examined changes in structure, function, and quality components of social connection in older people's last years of life, and the extent to which social connection predicts end-of-life outcomes (ie, symptoms, health-care utilisation, and place of death).

Methods

This study used longitudinal data of representative samples from across 18 European countries and Israel in the Survey of Health, Ageing, and Retirement in Europe (SHARE), the largest European cohort study of people aged 50 years or older. We included deceased participants of waves 4 and 6 (which contained social network modules) for whom a proxy provided an end-of-life interview. We did paired sample t-tests (for continuous variables), Wilcoxon signed-rank tests (for ordinal variables), and McNemar's tests (for non-ordinal categorical variables) to assess changes in structure, function, and quality components of social connection between waves 4 and 6. To examine social connection as a predictor of end-of-life outcomes, we used social connection data from wave 6 core interviews and end-of-life interviews from wave 7, conducted with a proxy respondent covering the deceased participant's last year of life. End-of-life outcomes included symptoms (pain, breathlessness, and anxiety or sadness) in the last month of life, health-care utilisation in the last year of life, and place of death. We conducted a mixed-effects logistic regression analysis per social connection measure, for each end-of-life outcome.

Findings

Data were collected in 2011–12 for wave 4, 2015–16 for wave 6, and 2017–18 for wave 7. We studied 3356 individuals (mean age at death was 79·7 years [SD 10·2]), with interviews conducted, on average, 4·6 (1·2) years (wave 4) and 1·1 (0·7) years (wave 6) before death. From wave 4 to wave 6, the following changes in social connection were observed: proportion of married or partnered participants (from 1406 [60·9%] of 2310 to 1438 [57·1%] of 2518; p<0·0001), receiving personal care or practical help (from 781 [37·2%] of 2099 to 1334 [53·1%] of 2512; p<0·0001), loneliness (from mean 1·4 [SD 0·5] to 1·5 [0·6]; p<0·0001; scale 1–3), satisfaction with social network (from 8·8 [1·67] to 8·7 [1·7]; p=0·037; scale 0–10), and emotional closeness to social network (eg, from 1883 [88·8%] of 2121 to 1710 [91·3%] of 1872 participants who indicated being either very close or extremely close to social network members; p<0·0001). Higher levels of loneliness at wave 6 predicted a greater likelihood of experiencing symptoms in the last month of life (odds ratio range across symptoms: 1·29 [95% CI 1·08–1·55] to 1·58 [1·32–1·89]). Being married (1·32 [1·03–1·68]) or receiving personal care or practical help (1·25 [1·04–1·49]) predicted death in hospital.

Interpretation

Social connection undergoes multifaceted changes towards older people's end of life, countering prevalent ideas of generally declining social trajectories. Loneliness in the final months of life might be a risk factor for end-of-life symptoms. Further research is needed to substantiate a causal relationship and to identify underpinning mechanisms, which could inform screening and prevention measures.

Funding

Research Foundation-Flanders and European Union.

Introduction

The social aspects of people's lives, especially social connection, have a large and independent impact on their health.^{1, 2, 3} A meta-analysis revealed that social connection has an independent effect on mortality risk, comparable in magnitude to that of many well known risk factors.¹ However, there are very few data on the role of social connection in the last phase of life (ie, the final months and years of a person's life), when health is often unstable or deteriorating.^{4, 5, 6, 7} Little is known about changes in social connection over time as people near death and how social connection relates to health outcomes that are relevant at the end of life, such as symptoms, health-care utilisation, and place of death.

Studying changes in social connection and the relationship with end-of-life outcomes is particularly important in older populations. People can live in relatively good health until late in life and for many, older age is a positive period of life, characterised by enriching relationships and good psychological and social wellbeing. The concept of positive ageing stresses that individuals are able to maintain intrinsic capacities and live life on their terms until later life and to pursue goals with meaning, even when they are affected by illness.⁸ However, as a result of epidemiological shifts towards a rising prevalence of serious chronic diseases over the past decades, older people are also at risk of experiencing a last phase of life that is marked by extended periods of deterioration in health and wellbeing, including social wellbeing, leading to considerable burden.⁹ Moreover, illness trajectories can be complicated by comorbidity (ie, the simultaneous presence of multiple chronic illnesses), and older age is a risk factor for a range of mental health problems and social concerns such as isolation and poverty.¹⁰

Research in context.

Evidence before this study

In the current literature on social connection at the end of life, to our knowledge, there are no studies that include data from outside North America and there is only one study combining longitudinal data with end-of-life interviews specifically designed to assess the last year of life. The social connection constructs examined in previous research are highly divergent and fragmented, and no study has investigated social connection comprehensively, covering structure, function, and quality aspects.

We searched PubMed from inception up to July 4, 2023. We included articles presenting observational data studying changes over time in any quantitative measure of social connection among older people who were identified as nearing the end of life, either retrospectively (in after-death studies) or prospectively using any indicator of a short life expectancy. We used the following search terms in the title or abstract: (social connection OR social relationship OR social isolation OR loneliness OR social support) AND (old age OR older people OR older adults) AND (end of life OR terminal OR last phase of life OR palliative). This resulted in 123 hits. We then screened the identified articles based on the following inclusion criteria: study focuses on older people (any definition), includes a quantitative examination of social connection or related concepts, and examines explicitly the end of life or similarly defined periods (last phase of life, palliative care trajectory, etc). Furthermore, in the same database, we searched for studies investigating the association between any quantitative social connection measure and indicators of wellbeing and health-care utilisation outcomes at the end of life, using the same search terms as above and additionally the following search terms: (symptoms OR well-being OR quality of life OR healthcare utilization OR hospital admission). This resulted in 46 hits. We screened the retrieved articles based on the same inclusion criteria as above and additionally required that the articles studied quantitative indicators of wellbeing at the end of life or health-care use. We did not apply any year or language restrictions to the searches.

The search returned four peer-reviewed studies, three of which included US samples from the Health and Retirement Study. The fourth study analysed a sample of cancer patients in Canada who were receiving home care services and were therefore not a general population of older people. Two studies analysed changes over time in loneliness, social isolation, and social decline and obtained divergent findings, likely because of differences in operationalisation. One study found that loneliness predicted symptom burden at the end of life, and two studies found that some measures of health-care use were predicted by loneliness.

Added value of this study

To the best of our knowledge, this is the first study on social connection at the end of life done in a large, representative sample of older adults across multiple countries. We did not identify any previous studies on the end of life that measured social connection comprehensively, distinguishing between structure, function, and quality components by applying an established conceptualisation of social connection . We show that towards older people's end of life, different components of social connection (structure, function, and quality) evolve in multifaceted ways. Although on average loneliness was shown to increase to a small extent in the final years, the size of social networks and the contact frequency and satisfaction with these networks remained relatively stable, countering common ideas of declining social trajectories at the end of life. We also show that people reporting greater loneliness around 1 year before death were more likely to experience burdensome symptoms (ie, breathlessness, anxiety or sadness, or pain or taking medication for pain) in the last month of life compared to people reporting lower levels of loneliness. Being married—a factor repeatedly cited as determinant of a home death in North America and the UK—was associated with a greater probability of dying in hospital in our study of 19 countries, including countries in southern, central, and eastern Europe, thus highlighting the importance of assessing data from various geographical regions.

Implications of all the available evidence

The implications of this study concern public health policy as well as theory and research methodology. Our findings of multifaceted changes in social connection challenge existing ideas of generally worsening social trajectories towards the end of life. Furthermore, to the best of our knowledge, this study shows for the first time in a multi-country sample that loneliness may be a risk factor for burdensome symptoms at the end of life, including pain, breathlessness, and anxiety or sadness. Follow-up research is needed to substantiate a causal relationship and to identify underpinning mechanisms, which can inform screening and prevention measures. In terms of methodology, this study demonstrates the value of, and need for, clear conceptual frameworks in research on social connection and its relationship with end-of-life outcomes. The precise mechanisms through which social connection shapes end-of-life outcomes remain unexplained; a better understanding thereof requires the integration of epidemiological studies with well designed qualitative and ethnographic research.

Social connection has been used as an overarching term to describe the various ways in which people connect to others,² and there is considerable variability in how social connection is measured across different studies and research domains and in the concepts that are assessed, including social capital, social isolation, loneliness, social support, or social networks.¹¹ Several systematic conceptualisations have been proposed to categorise and organise the different constructs that research on social connection has used, with a view to ensuring greater transparency and comparability. For example, in the context of palliative care, one approach used the concept of social capital to capture the ability of social networks and relationships to support health and wellbeing.¹² The literature on social capital distinguishes between two components: structural social capital and cognitive social capital.¹² The former refers to social networks and access to goods and services, and the latter includes shared values, trust, participation, belonging, cohesion, and decision-making capacity.¹² Another study reviewed measures of social relationships used in epidemiological studies and determined that these measures cover two dimensions: one, a focus on structural versus functional aspects of social relationships; and two, the degree of subjectivity asked from respondents.¹³ Hence, the authors propose two components to describe social relationships, each of which can be captured with different degrees of subjectivity.

Arguably, the most influential recent conceptualisation of social connection is that of Julianne Holt-Lunstad, who proposed a multifactorial conceptualisation with the explicit aim of delineating the various concepts that have been used to operationalise social connectedness. She distinguishes between structure, function, and quality components of social connection.^{2, 11} The structure component comprises our connection to others via the presence versus absence of relationships, roles, and interactions, and is usually measured in quantitative terms (eg, social network size, group membership, living arrangements, and contact frequency). The function component reflects social connection through resources provided or available to meet various needs, including emotional, physical, tangible, informational, and belonging needs. The quality component acknowledges positive and negative affective qualities in our social connections (eg, relationship satisfaction, cohesion, intimacy, closeness, strain, and conflict). To support this conceptualisation, Holt-Lunstad uses empirical evidence showing that each of these components independently predicts health outcomes (mortality and morbidity). At the same time, strong correlations between these components have not been found, suggesting that these components represent distinct concepts that contribute in unique ways to health outcomes and should all be considered when studying the relationship between social connection and health.¹¹

In this study, we used Holt-Lunstad's conceptualisation to guide our selection of variables to capture social connection. Applying this conceptualisation to existing research on social connection at the end of life reveals that previous work has been rather fragmented. Using mainly a sociodemographic lens, previous work has focused on structure (eg, whether someone is married or lives alone) or function components (eg, social support or caregiver burden) of social connection^{14, 15} while paying less attention to quality components (eg, individuals' perceptions and evaluations of their relationships). Loneliness has been the predominant focus in studies of subjective experiences of social connection at the end of life. It has been argued that older age, and specifically the end-of-life period, is associated with risks of experiencing loneliness.^{5, 16} Furthermore, there is growing evidence that loneliness might be an important predictor of adverse health outcomes among older adults, including at the end of life.^{4, 17} However, there is still very little research—especially using longitudinal data—investigating the relationship between loneliness and health outcomes at the end of life.^{4, 18}

Outcomes typically used to determine wellbeing at the end of life in epidemiological studies include physical and psychosocial symptoms, health-care utilisation, and place of death.¹⁹ These outcomes were found to be important indicators of older people's experiences at the end of life.²⁰ Although it is clear that we should strive to minimise symptom burden at the end of life, and that there might be benefits in receiving hospice or palliative care in the last weeks of life for people dying from a chronic illness,²¹ there is no agreement on what frequency or length of hospital admissions, or place of death, is optimal. These aspects considerably depend on individual needs and preferences, an individual's social, financial, and material circumstances, and the types of care that can be provided in specific locations in given health-care systems.

The two objectives of this international study were to determine changes in structure, function, and quality components of social connection over time in older people's last years of life, and to examine the extent to which social connection in the last years of life predicts end-of-life outcomes (ie, symptoms including pain, breathlessness, and anxiety or sadness in the last month of life, health-care utilisation in the last year of life, and place of death). Given the scarcity of existing evidence in this area, especially on the end of life, these research questions are not based on specific hypotheses but follow an exploratory principle.

Methods

Study design

This study used data from the Survey of Health, Ageing, and Retirement in Europe (SHARE), one of the largest ongoing longitudinal ageing studies worldwide collecting micro (ie, individual-level) data on health and on social and family networks of home-dwelling older people (aged ≥50 years), covering most of the EU.²² SHARE has been running since 2004, and collects data every two years through personal structured interviews (core interviews).²² When a participant dies, SHARE conducts an end-of-life interview with a proxy respondent covering the deceased participant's last year of life. We used wave 7 end-of-life interviews (conducted in 2017–18) as they were the most recent at the time of analysis and, unlike end-of-life interviews from previous waves, they contained questions on symptoms and palliative care utilisation. SHARE core interviews contained a social networks module in wave 4 (conducted in 2011–12) and wave 6 (conducted in 2015–16), addressing social connection in a very detailed manner and capturing structure, function, and quality components. The appendix (p 2) contains more detailed information on the SHARE methodology. To study changes in social connection over time, we compared wave 4 with wave 6 core interviews. To examine social connection as a predictor of end-of-life outcomes, we used social connection data from wave 6 core interviews and end-of-life outcomes measured in wave 7 end-of-life interviews. As this study was exploratory, we have followed the six principles outlined by Luijken and colleagues for the reporting and interpretation of exploratory analyses in aetiologic research.²³

Participants

The SHARE target population consists of all people aged 50 years or older at the time of sampling who are home-dwelling and have their regular domicile in the respective SHARE country. People are excluded if they are incarcerated, hospitalised, out of the country during the entire survey period, unable to speak the country's language(s), cannot be located, or have moved to an unknown address. Proxy respondents in the end-of-life interviews could be a family or household member, a neighbour, or any other person close to the deceased respondent. SHARE includes people living in nursing homes or residential care institutions whenever they are covered in the sampling frame from which the sample is drawn. In this study, this was the case for all countries except Poland, Slovenia, and Denmark. In Italy, France, Greece, Luxembourg, and Estonia, nursing home residents were represented only in wave 4.²⁴ SHARE is a multinational survey and sampling frames are chosen in accordance with the best available frame resources in the country to achieve full probability sampling, which was done through population registers in most SHARE countries.^{22, 24} The SHARE methodology reports and the appendix (p 2) provide detailed information on the three-stage sampling strategy and on measures used to ensure comparable sampling frames and procedures among countries.²⁵

To analyse changes in social connection, we selected participants who had died and for whom a proxy respondent completed an end-of-life interview at wave 7. We excluded participants who did not participate in either of the two waves (4 or 6) in which the social networks module was completed. To analyse the relationship between social connection at wave 6 and end-of-life outcomes, we selected participants for whom there was an end-of-life interview at wave 7 and who had participated in wave 6 (irrespective of their participation status at wave 4). We excluded participants whose proxy respondent had never had contact with them or for whom the frequency of contact with the proxy respondent was unknown as this posed a risk to the validity of the information provided in the end-of-life interviews.

From wave 4 onwards, the SHARE project was approved by the Ethics Council of the Max Planck Society and by respective ethics committees of individual countries, whenever required. No additional ethical approval was required for this analysis of open access data. Participants provided written informed consent and, at each data collection wave, were asked whether they agree to be recontacted at the next wave.

Measurements

Core interviews were done face to face, and end-of-life interviews with proxy respondents were done face to face or by telephone. SHARE measures are extensively validated, and data collection and processing follow strict quality assurance protocols.²² Sociodemographic and health-related information as well as social connection measures were collected in the core interviews. End-of-life interviews provided information on end-of-life outcomes and additional sociodemographic information on participants and proxy respondents (see appendix p 4 for wording and response options of all measures used in this study). In the SHARE study, participants' sex was recorded by the interviewer, who was instructed to note the sex (male or female) of the participant from observation (or to ask if unsure). Panel 1 provides an overview of the social connection variables and end-of-life outcome variables used in this analysis, as well as their response options and coding.

Panel 1. Response categories and coding of social connection and end-of-life outcome variables.

Social connection measures

Marital status

Married or partnered (1) or not married or partnered (2).

Loneliness (3-item Loneliness Scale)

Individual items scored from 1 to 3, with higher numbers indicating greater loneliness. Total scale score based on mean of individual items, ranging from 1 to 3.

Social network members

Respondent is asked to list first names of social network members. Up to seven members can be listed.

Personal care or practical help received by social network members or others

No care received (0) or care received (1).

Geographical proximity to social network members

In the same household (1), in the same building (2), less than 1 km away (3), between 1 and 5 km away (4), between 5 and 25 km away (5), between 25 and 100 km away (6), between 100 and 500 km away (7), or more than 500 km away (8).

Contact frequency with social network members

Never (1), less than once a month (2), about once a month (3), about every 2 weeks (4), about once a week (5), several times a week (6), or daily (7).

Emotional closeness to social network members

Not very close (1), somewhat close (2), very close (3), or extremely close (4).

Satisfaction with social network

Scale from 0 to 10 with higher numbers indicating greater satisfaction.

End-of-life outcomes

Place of death

At their own home (1), at another person's home (2), in a hospital (3), in a nursing home (4), in a residential home, sheltered housing, or old people's home (5), in a hospice (6), in transit to a medical facility (7), or other (97).

Time spent in hospital in the last year of life

Less than 1 week (1), 1 week or more but less than 1 month (2), 1 month or more but less than 3 months (3), 3 months or more but less than 6 months (4), 6 months or more but less than 1 year (5), or 1 year (6). Dichotomised in this analysis to less than 1 week to less than 3 months (1) or 3 months to 1 year (2).

Received hospice or palliative care in the last month of life

Yes (1) or no (5).

Experienced pain or has taken medicine for pain in the last month of life

Yes (1) or no (5).

Experienced breathlessness in the last month of life

Yes (1) or no (5).

Experienced anxiety or sadness in the last month of life

Yes (1) or no (5).

Social connection measures

The constructs and measures used to assess structure, function, and quality components of social connection are presented in panel 2. All measures, except marital or partnership status, social support, and loneliness, were taken from the social networks module of the core interviews done in waves 4 and 6. Participants were asked to list six people in response to the question: “Over the last 12 months, who are the people with whom you most often discussed important things?”. This question encourages respondents to consider their confidants and people with whom they interact, discuss things of relative importance, and maintain a degree of trust, and to consider relationships outside of family.²⁶ An additional probe allowed respondents to mention one additional person who is important to them for any other reason. For each of the people indicated, additional questions were asked concerning the type of relationship, geographical (residential) proximity, frequency of contact during the past 12 months, emotional closeness, and whether the person provided personal care or practical help to the respondent. Additionally, respondents were asked to indicate their overall satisfaction with their social network on a scale from 0 to 10.

Panel 2. Measurements of social connection used in this study.

Structure

• •Marital or partnership status (married or partnered vs not married or not partnered)
• •Social network size (from 0 to 7, based on name-generator approach)
• •Geographical proximity of social network (average across network)
• •Contact frequency with social network (average across network)

Function

• •Social support received (personal care or practical help given by social network members or others; yes or no)
• •Loneliness (3-item UCLA Revised Loneliness Scale)

Quality

• •Satisfaction with social network (11-point scale from “completely dissatisfied” to “completely satisfied”)
• •Emotional closeness with social network members (4-point scale)

Measures of end-of-life outcomes

End-of-life outcomes obtained from end-of-life interviews with proxy respondents were: place of death (dichotomised to hospital vs other), length of illness before death, total time spent in hospital (if ever admitted) in the last year of life, and, in the last month of life, having had feelings of anxiety or sadness, having had trouble breathing, having had pain or taken medication for pain, and having received hospice or palliative care. Time spent in hospital was coded by SHARE as a categorical rather than as a continuous variable, as many older people are not able to remember the precise number of days in hospital (panel 1). To simplify interpretation of the results of the regression analyses, we combined the six categories into two categories: admitted for up to 3 months and admitted for more than 3 months.

Data analysis

Though not hypothesis-driven, all analyses were prespecified. We examined the response rate to the wave 7 end-of-life interviews using all participants identified as deceased at wave 7 as the denominator and calculating the percentage of participants for whom an end-of-life interview was available. We calculated descriptive statistics for the social connection measures and used paired sample t-tests (for continuous variables), Wilcoxon signed-rank tests (for ordinal variables), and McNemar's tests (for non-ordinal categorical variables) to test for differences in these measures between waves 4 and 6. To examine the extent to which social connection at wave 6 predicted end-of-life outcomes, we conducted a mixed-effects logistic regression analysis per social connection measure, for each end-of-life outcome. In each of these models, the dependent variable was a dichotomous end-of-life outcome and the independent variable was one of the social connection measures. Using literature on social connection and health as guidance,^{1, 4} we included the following covariates: age at death, sex, level of education, and the relationship between the deceased participant and the proxy respondent (spouse, child, friend, etc). In analyses where receipt of personal care or practical help from others was included as a predictor, we additionally included self-rated health status as a covariate. We did not include comorbidities and depression symptoms as covariates because depression and other health characteristics might not be confounders but might, at least partly, be mediators of the association between social connection and end-of-life outcomes,^{27, 28, 29} and adjusting for mediators is not recommended as it can lead to erroneous results and conclusions.³⁰ The mixed-effects logistic regression models were estimated with one random intercept for country to account for the clustering of data within countries (mixed-model analyses handle missing data in the dependent variable through maximum likelihood estimation). Statistical significance was indicated as p<0·05 (two-sided). Data from different waves were merged using Python 3.11, with Pandas 2.1 and Numpy 1.26. Data were analysed using IBM SPSS Statistics version 29.

Role of the funding source

The funders of this work had no role in study design; in the collection, analysis, and interpretation of data; in the writing of the report; or in the decision to submit the paper for publication.

Results

The SHARE dataset included end-of-life interviews collected in 2017–18 for 3662 deceased participants from 19 countries at wave 7. This represents 87·5% of 4185 participants who were noted as deceased at wave 7; see appendix (p 3) for a comparison of demographic characteristics between deceased participants with and without an end-of-life proxy interview. For the analysis of changes in structure, function, and quality components of social connection between waves 4 and 6, we excluded 306 of the 3662 participants for whom end-of-life interviews were available, as they had not participated in either wave 4 or wave 6 interviews, resulting in a final analysis sample of 3356 participants (figure). We report differences in sample characteristics between the included and excluded sample subsets in the appendix (p 3). The mean time between interviews and death was 4·6 years (SD 1·2) for wave 4 and 1·1 years (0·7) for wave 6 (for frequencies see appendix p 7). Of the 3356 participants, 1582 (47·1%) were female and 1774 (52·9%) were male (table 1). The mean age at death was 79·7 (SD 10·2) years. Cardiovascular illness was the leading cause of death (1341 [40·0%] of 3356). Proxy respondents for the end-of-life interviews were predominantly female (2152 [64·1%] of 3356) and the deceased's partner (1322 [39·4%] of 3356) or son or daughter (861 [25·7%] of 3356). Two-thirds of proxy respondents reported having had daily contact with the older person in the last 12 months of their life, and a minority reported having had contact less than once a week (276 [8·2%] of 3356) or never (86 [2·6%] of 3356). Between waves 4 and 6, the prevalence for all physician-diagnosed conditions increased, self-rated health deteriorated, and the proportion of participants who indicated that their health limited their activities increased (table 2). For the analysis of associations between wave 6 social connection measures and end-of-life outcomes, we excluded 1136 participants of the initial sample of 3662, as they did not participate in wave 6 core interviews. Of the remaining 2526 participants, we excluded 55 who, in their last year of life, did not have any contact with the proxy respondent who provided the end-of-life interview for them, or for whom contact frequency with the proxy respondent was unknown. The resulting final analysis sample consisted of 2471 deceased participants (figure). The appendix shows descriptive statistics for all end-of-life outcomes, by sex (p 7), and the correlation matrix for the social connection measures and measures of end-of-life outcomes (p 8). Most correlation coefficients are small, except for several moderate correlations (ie, >0·3) among social connection variables.

Figure.

Study profile

SHARE=Survey of Health, Ageing, and Retirement in Europe.

Table 1.

Characteristics of participants and proxy respondents captured by wave 7 end-of-life interviews (N=3356)

		n (%)
Characteristics of deceased
Country
	Austria	177 (5·3%)
	Germany	101 (3·0%)
	Sweden	134 (4·0%)
	Spain	396 (11·8%)
	Italy	214 (6·4%)
	France	185 (5·5%)
	Denmark	143 (4·3%)
	Greece	179 (5·3%)
	Switzerland	85 (2·5%)
	Belgium	197 (5·9%)
	Israel	87 (2·6%)
	Czech Republic	270 (8·0%)
	Poland	113 (3·4%)
	Luxembourg	29 (0·9%)
	Hungary	300 (8·9%)
	Portugal	88 (2·6%)
	Slovenia	184 (5·5%)
	Estonia	373 (11·1%)
	Croatia	101 (3·0%)
Sex
	Male	1774 (52·9%)
	Female	1582 (47·1%)
Age at death, mean (SD)		79·7 (10·2)
Level of education
	Less than upper secondary education	1953 (58·2%)
	Upper secondary and vocational training	1015 (30·2%)
	Tertiary education	388 (11·6%)
Main cause of death
	Cancer	816 (24·9%)
	Heart attack	470 (14·4%)
	Stroke	356 (10·9%)
	Other cardiovascular related illness	515 (15·7%)
	Respiratory disease	184 (5·6%)
	Disease of the digestive system	85 (2·6%)
	Severe infectious disease	208 (6·4%)
	Accident	86 (2·6%)
	Other	554 (16·9%)
Length of illness before death
	<1 month	895 (27·7%)
	1 to <6 months	569 (17·6%)
	6 to <12 months	401 (12·4%)
	≥12 months	1366 (42·3%)
Place of death
	At (his or her) own home	1024 (31·5%)
	At another person's home	39 (1·2%)
	In a hospital	1621 (49·8%)
	In a nursing home	294 (9·0%)
	In a residential home or sheltered housing	112 (3·4%)
	In a hospice	108 (3·3%)
	In transit to a medical facility	13 (0·4%)
	Other	43 (1·3%)
Waves completed
	Participated in wave 4 and wave 6 interviews	1486 (44·3%)
	Participated in wave 4 interview only	830 (24·7%)
	Participated in wave 6 interview only	1040 (31·0%)
Characteristics of proxy respondents
Relationship to deceased
	Husband, wife, or partner	1322 (39·4%)
	Son or daughter	861 (25·7%)
	Son-in-law or daughter-in-law	108 (3·2%)
	Son or daughter of husband, wife, or partner	63 (1·9%)
	Grandchild	82 (2·4%)
	Sibling	76 (2·3%)
	Other relative	167 (5·0%)
	Other non-relative	674 (20·1%)
Frequency of contact in the last 12 months of life
	Daily	2225 (66·5%)
	Several times a week	560 (16·7%)
	About once a week	199 (5·9%)
	About every 2 weeks	116 (3·5%)
	About once a month	60 (1·8%)
	Less than once a month	100 (3·0%)
	Never	86 (2·6%)
Sex
	Male	1202 (35·8%)
	Female	2152 (64·1%)

Data are n (%) or mean (SD). Total number of wave 7 end-of-life interviews=3356. Missing data: age at death, n=69 (2·1%); main cause of death, n=82 (2·4%); length of illness before death, n=98 (3·2%); place of death, n=102 (3·0%); relationship to deceased, n=2 (0·1%); and frequency of contact, n=10 (0·3%).

Table 2.

Participant characteristics at wave 4 and wave 6 core interviews (N=3356)

		Wave 4 (n=2316)	Wave 6 (n=2526)
Physician-diagnosed conditions (ever diagnosed)
	Cancer	303 (13·2%)	498 (19·8%)
	Lung disease	706 (30·7%)	998 (39·7%)
	Heart problems	266 (11·6%)	488 (19·4%)
	Stroke	252 (11·1%)	496 (19·7%)
	Psychiatric problems	52 (2·3%)	83 (3·3%)
	Parkinson's disease	120 (5·2%)	465 (18·5%)
	Alzheimer's disease	303 (13·2%)	498 (19·8%)
Age at interview, years (mean [SD])		74·6 (9·9)	78·9 (9·9)
Living environment
	Urban	1421 (65·4%)	1568 (68·0%)
	Rural	753 (34·6%)	738 (32·0%)
Living in nursing home		64 (2·8%)	192 (7·6%)
Self-rated health
	Excellent	53 (2·3%)	34 (1·4%)
	Very good	168 (7·3%)	109 (4·3%)
	Good	514 (22·3%)	436 (17·3%)
	Fair	891 (38·7%)	923 (36·7%)
	Poor	675 (29·3%)	1012 (40·3%)
Health limits activities
	No	685 (29·8%)	492 (19·5%)
	Yes	1617 (70·2%)	2025 (80·5%)
Depression symptoms (EURO-D), mean (SD)		3·3 (2·5)	3·5 (2·7)
Clinical level of depression, based on EURO-D*
	Not depressed	1379 (59·5%)	1421 (56·3%)
	Case of depression	937 (40·5%)	1105 (43·7%)

Data are n (%) unless otherwise stated. Missing data among wave 4 responders: physician-diagnosed conditions, n=13 (0·6%); living environment, n=142 (6·1%); living in nursing home, n=2 (0·1%); self-rated health, n=15 (0·6%); health limits activities, n=14 (0·6%). Missing data among wave 6 responders: Physician-diagnosed conditions, n=10–14 (0·4%–0·6%); living environment, n=220 (8·7%); living in nursing home, n=5 (0·2%); self-rated health, n=12 (0·5%); and health limits activities, n=9 (0·4%). SHARE=Survey of Health, Ageing, and Retirement in Europe.

^*In SHARE, a EURO-D score of 4 or higher indicates a case of depression.

We found changes over time in several aspects of social connection. Concerning the structure component of social connection, older people were less likely to be married or partnered at wave 6 compared with wave 4 (1438 [57·1%] of 2518 vs 1406 [60·9%] of 2310; p<0·0001; table 3). There were no changes in social network size (p=0·16), average contact frequency (p=0·29), or average geographical proximity (p=0·057) to social network members.

Table 3.

Changes in structure, function, and quality components of social connection in older people nearing the end of life

		Wave 4 (n=2316)	Wave 6 (n=2526)	p value
Structure component
Marital status		..	..	<0·0001*
	Married or partnered	1406 (60·9%)	1438 (57·1%)	..
	Not married or partnered	904 (39·1%)	1080 (42·9%)	..
Social network size, mean (SD)		2·1 (1·5)	2·2 (1·4)	0·16†
Social network proximity, average across members†		..	..·	0·06‡
	Same household	574 (27·0%)	426 (23·8%)	..
	Same building	183 (8·6%)	192 (10·7%)	..
	<1 km	489 (23·0%)	383 (21·4%)	..
	1–5 km	446 (21·0%)	395 (22·0%)	..
	5–25 km	297 (14·0%)	272 (15·2%)	..
	25–100 km	103 (4·9%)	91 (5·1%)	..
	100–500 km	21 (1·0%)	21 (1·2%)	..
	>500 km	9 (0·4%)	13 (0·7%)	..
Social network contact, average across members†		..	..	0·29‡
	Daily contact	1026 (48·2%)	855 (45·6%)	..
	Several times per week	620 (29·1%)	639 (34·1%)	..
	Once per week	324 (15·2%)	262 (14·0%)	..
	Once every 2 weeks	100 (4·7%)	89 (4·7%)	..
	Once a month	46 (2·2%)	19 (1·0%)	..
	Less than once a month	10 (0·5%)	11 (0·6%)	..
	Never	2 (0·1%)	0	..
Function component
Received personal care or practical help from others		..	..	<0·0001*
	No care received	1318 (62·8%)	1178 (46·9%)	..
	Care received	781 (37·2%)	1334 (53·1%)	..
Loneliness, mean (SD), scale range 1–3		1·4 (0·5)	1·5 (0·6)	<0·0001†
Felt left out		..	..	<0·0001‡
	Hardly ever or never	1130 (68·7%)	1201 (62·3%)	..
	Some of the time	402 (24·4%)	504 (26·1%)	..
	Often	114 (6·9%)	223 (11·6%)	..
Felt isolated from others		..	..	<0·0001‡
	Hardly ever or never	1231 (74·8%)	1317 (68·1%)	..
	Some of the time	316 (19·2%)	428 (22·1%)	..
	Often	99 (6·0%)	190 (9·8%)	..
Felt they lacked companionship		..	..	<0·0001‡
	Hardly ever or never	1062 (63·3%)	1013 (52·2%)	..
	Some of the time	465 (27·7%)	602 (31·0%)	..
	Often	150 (8·9%)	325 (16·8%)	..
Quality component
Satisfaction with social network, mean (SD), scale range 0 to 10		8·8 (1·67)	8·7 (1·7)	0·037†
Emotional closeness with social network		..	..	<0·0001‡
	Not very close	16 (0·8%)	19 (1·0%)	..
	Somewhat close	222 (10·5%)	143 (7·6%)	..
	Very close	996 (47·0%)	877 (46·8%)	..
	Extremely close	887 (41·8%)	833 (44·5%)	..

Data are n (%) unless otherwise stated. The total number of participants in wave 4 and 6 was 3356 (2316 in wave 4, 2526 in wave 6, with some participants taking part in both waves). Missing data among wave 4 responders: marital or partnership status, n=6 (0·3%); social network size, no missing data; social network proximity, n=194 (8·4%); social network contact, n=188 (8·1%); received personal care or help, n=217 (9·4%); loneliness, n=633 (27·3%); satisfaction with social network, n=91 (3·9%); and emotional closeness, n=195 (8·4%). Missing data among wave 6 responders: marital or partnership status, n=8 (0·3%); social network size, n=553 (21·9%); social network proximity, n=733 (29·0%); social network contact, n=651 (25·8%); received personal care or help, n=14 (0·6%); loneliness, n =585 (23·2%); satisfaction with social network, n=585 (23·2%); and emotional closeness, n=654 (25·9%).

^*McNemar's test.

^†T-test for paired samples.

^‡Wilcoxon signed-ranks test.

Regarding the function component of social connection, a greater proportion of older people received personal care or practical help from others at wave 6 compared with wave 4 (1334 [53·1%] of 2512 vs 781 [37·2%] of 2099; p<0·0001). Loneliness increased from a mean score of 1·4 (SD 0·5) at wave 4 to 1·5 (0·6) at wave 6 (p<0·0001; scale 1–3). Descriptive statistics for the separate items of the UCLA Loneliness Scale show that from wave 4 to wave 6, the proportion of older people who indicated that they often felt left out increased from 114 (6·9%) of 1646 to 223 (11·6%) of 1928, the proportion of those reporting often feeling isolated from others increased from 99 (6·0%) of 1646 to 190 (9·8%) of 1935, and the proportion of those who indicated often lacking companionship increased from 150 (8·9%) of 1677 to 325 (16·8%) of 1940.

Concerning the quality component of social connection, satisfaction with social networks decreased slightly from wave 4 to wave 6, from 8·8 (SD 1·67) to 8·7 (1·7; p=0·037; scale 0–10), but the average emotional closeness to social network members increased somewhat (p<0·0001); for example, the proportion who indicated being either very or extremely close to their social network members increased from 1883 (88·8%) of 2121 at wave 4 to 1710 (91·3%) of 1872 at wave 6.

Higher levels of loneliness at wave 6 predicted higher odds of anxiety or sadness (odds ratio [OR] 1·58 [95% CI 1·32–1·89]; p<0·0001), of trouble breathing (1·37 [1·15–1·62]; p<0·0001), and of having had pain or having taken medication for pain in the last month of life (1·29 [95% CI 1·08–1·55]; p=0·01; table 4). People who were married at wave 6 were more likely to die in hospital (OR 1·32 [95% CI 1·03–1·68]; p=0.03) compared to unmarried or unpartnered people. Trouble breathing in the last month of life was predicted by a higher average contact frequency with social network members at wave 6 (1·14 [1·02–1·26]; p=0·019). Those who received personal care or practical help from others at wave 6 were more likely to die in hospital than elsewhere, after controlling for health status (1·25 [1·04–1·49]; p=0·016). Finally, a larger social network at wave 6 predicted a greater likelihood of receiving hospice or palliative care in the last month of life (1·10 [1·01–1·19]; p=0·025).

Table 4.

Associations between social connection at wave 6 and end-of-life outcomes

		Place of death (hospital vs other; other is reference)	Experienced anxiety or sadness in last month of life (response of “no” is reference)	Experienced trouble breathing in last month of life (response of “no” is reference)	Had pain or took medication for pain in last month of life (response of “no” is reference)	Received hospice or palliative care in last month of life (response of “no” is reference)	Total time in hospital in last year before death if admitted (≤3 months vs >3 months, >3 months is reference)
Structure
Social network size*		1·07 (1·00–1·14)	1·02 (0·96–1·10)	0·98 (0·91–1·05)	0·99 (0·92–1·06)	1·10 (1·01–1·19)	1·04 (0·93–1·16)
	p value	0·07	0·51	0·51	0·73	0·03	0·50
Marital or partnership status* (reference: not married)		1·32 (1·03–1·68)	1·32 (1·03–1·68)	1·32 (1·03–1·68)	1·32 (1·03–1·68)	1·32 (1·03–1·68)	1·32 (1·03–1·68)
	p value	0·03	0·06	0·87	0·80	0·56	0·17
Social network proximity*		1·03 (0·97–1·10)	1·03 (0·97–1·10)	1·03 (0·97–1·10)	1·03 (0·97–1·10)	1·03 (0·97–1·10)	1·03 (0·97–1·10)
	p value	0·35	0·76	0·21	0·55	0·72	0·26
Average contact frequency with social network members*		1·05 (0·94–1·16)	1·05 (0·94–1·17)	1·14 (1·02–1·26)	1·04 (0·93–1·16)	1·01 (0·88–1·15)	1·08 (0·92–1·26)
	p value	0·40	0·38	0·02	0·50	0·94	0·33
Function
Received personal care or practical help from social network members or others (reference: no care received)†		1·25 (1·04–1·49)	0·91 (0·76–1·10)	1·07 (0·90–1·29)	0·96 (0·80–1·16)	0·91 (0·73–1·13)	1·08 (0·82–1·41)
	p value	0·02	0·34	0·45	0·69	0·39	0·60
Loneliness*		1·02 (0·86–1·20)	1·58 (1·32–1·89)	1·37 (1·15–1·62)	1·29 (1·08–1·55)	0·88 (0·71–1·09)	0·80 (0·61–1·04)
	p value	0·86	<0·0001	<0·0001	0·01	0·24	0·09
Quality
Satisfaction with social network*		1·04 (0·98–1·10)	1·00 (0·95–1·06)	0·98 (0·92–1·04)	0·99 (0·93–1·05)	0·95 (0·89–1·01)	1·05 (0·97–1·15)
	p value	0·12	0·89	0·46	0·67	0·11	0·23
Emotional closeness with social network*		1·10 (0·95–1·28)	0·97 (0·83–1·13)	1·10 (0·94–1·27)	1·03 (0·88–1·21)	0·90 (0·75–1·09)	1·14 (0·90–1·44)
	p value	0·20	0·70	0·24	0·68	0·29	0·27

Data are odds ratio (95% CI), based on mixed-effects logistic regression analyses with random intercept for country. The analysis sample included 2471 participants.

^*Covariates in analyses were age at death, sex, level of education, and relationship between the older person and the proxy respondent.

^†Covariates in analyses were age at death, sex, self-rated health status, level of education, and relationship between the older person and the proxy respondent.

Discussion

In this study of social connection in a representative sample across 19 countries, we found that, on average, older people experienced slightly increased loneliness towards the end of life. Moreover, as they neared death, fewer people were married but more received personal care or practical help from others. All other aspects of social connection (frequency of contact with and geographical proximity to social network members, mean social network size, mean satisfaction with social network, and emotional closeness to social network members) remained relatively stable or changed to a small extent only. Although the increase in loneliness was statistically significant, the mean change (ie, 1·4 [SD 0·5] to 1·5 [0·6] on a scale from 1 to 3) can be considered relatively small. Being married or having a partner and receiving personal care or practical help from others were both associated with a greater probability of dying in hospital as opposed to other locations. Loneliness predicted higher odds of all symptoms measured (ie, pain, breathlessness, and anxiety or sadness in the last month of life).

Our findings counter prevalent ideas of generally declining trajectories of social wellbeing towards the end of life^{9, 31, 32} and instead reveal that older people's social connection changes in multifaceted ways as they near death. The largest changes were found in marital status and in receipt of personal care or practical help from others, as might be expected in the context of deteriorating health. Loneliness increased over the last years of life, albeit to a small extent, and it did so in the context of increasing social support (receiving help or care) and largely stable social networks in terms of size, contact frequency, and emotional closeness. These seemingly inconsistent changes confirm previous observations of low correlations between different measures of social connection, suggesting that these measures capture different aspects of the way in which people relate to others.^{1, 33} Our findings stress the importance of measuring subjective experiences of social connection, such as loneliness, in studies on the end of life, and caution against the predominant reliance on structure components (eg, marital status) and social support as indicators of social connection.¹³

Loneliness in older people's last years of life emerged as a potentially important predictor of symptoms at the end of life (ie, pain, breathlessness, and anxiety or sadness). One previous study from the USA obtained a similar finding,⁴ but the mechanisms of this relationship are largely unknown. Mechanisms that have previously been suggested include behavioural, as well as physiological, pathways, but findings are far from conclusive and are fragmented.^{11, 34} Others have suggested that individuals who are lonely are more likely to receive overly aggressive treatments at the end of life, which can exacerbate symptoms.⁴ Although the associations between loneliness and symptoms at the end of life were statistically significant, the effects can be considered relatively small (ie, ORs between 1·29 [95% CI 1·08–1·55] and 1·58 [1·32–1·89]). Furthermore, relationships between social connection and health outcomes are likely very complex and bidirectional,²⁹ and include factors that are difficult to measure. This further complicates the interpretation of the associations we found. Hence, although our findings point towards a potentially important relationship between loneliness and burdensome symptoms at the end of life, further in-depth studies are required to substantiate them and to uncover underlying mechanisms. Follow-up research should also investigate the dynamic nature of both social connection and health and the way in which one might shape the other. Just as social connection can affect health, social networks can also adapt to meet the changing health needs of individuals. Uncovering such complex relationships requires well designed epidemiological studies with large samples, as well as qualitative and ethnographic research that can point to mechanisms underlying statistical associations.

Older people who were married or had a partner were more likely to die in hospital than elsewhere. Furthermore, older people who received personal care or practical help were more likely to die in hospital, even after controlling for self-rated health status. Although the respective ORs are not high, it is notable that these findings counter results from previous research, mostly from western Europe and North America, that suggested that being married or having social support is associated with dying at home.³⁵ Our data from 19 countries, including countries in southern and eastern Europe, challenge this conclusion in wider geographical and health-care contexts and suggest that the opposite might be true, (ie, that those with social support are more likely to die in hospital). We must consider that in some places, where home palliative care services are perhaps less well developed or where a hospital admission at the end of life is considered to be good care,³⁶ those dying at home might not receive the support they need or might not live in environmental and material conditions necessary to achieve a good death. In these cases, a spouse or other family carer might act as an advocate who ensures a hospital admission during a health crisis.³⁷ These findings caution against generalisation of epidemiological findings regarding the end of life obtained from single countries to other world regions.

The strengths of this study include large and nationally representative samples of deceased older people across 19 countries. SHARE is a long-standing panel study with high quality sampling frames for each country. SHARE's combination of cross-national data, an elaborate social networks module, and a relatively extensive end-of-life interview is unique among longitudinal ageing studies. In addition to providing prospective data on social connection, SHARE contains a retrospective end-of-life interview that permitted us to assess a uniform period at the end of life (ie, the last 12 months). The response rate of 87·5% (3662/4185) in the wave 7 end-of-life interview can be considered very high, especially in comparison with similar panel studies, such as the English Longitudinal Study of Ageing that had a response rate of 61·8%.³⁸ Wave-to-wave retention rates across SHARE waves were high (around 80–90%).²⁴ Nonetheless, this study also has limitations. First, SHARE assessed social networks in two waves only, preventing examination of social connection trajectories across more timepoints. Second, it is likely that proxies of some SHARE participants who died after being lost to follow-up were not approached for end-of-life interviews. Due to the absence of national mortality registers in most European countries, the SHARE consortium cannot determine the vital status of non-respondents. Third, even though our data are longitudinal and can therefore, to some extent, prevent reverse causality, they do not allow clear causal interpretations as there might still be confounders that we did not account for. Fourth, although most countries included nursing home residents in the SHARE sample, three did not and five did so only in wave 4. Fifth, our comparison of deceased participants with and without end-of-life interview showed that women might be slightly underrepresented in our study. This is possibly due to women being less likely than men to have a living proxy (which was a spouse or partner in almost 40% of cases) who could complete an end-of-life interview for them. Furthermore, participants living in residential care facilities or sheltered housing might be slightly underrepresented in our study, likely because we excluded deceased participants who had not had any contact with their end-of-life interview proxy. Sixth, the retrospective nature of the end-of-life interviews carries risks of recall bias and inaccurate reports concerning symptoms that are less observable. However, proxy reports in post-mortem surveys are a widely accepted, and the only feasible, method to collect data on defined periods before death in population-based samples.³⁹ Finally, although all analyses in this study were pre-specified, we did not publish an analysis protocol in advance.

The study has implications for public health policy as well as for theory and research methodology. If the relationship between loneliness and burdensome symptoms at the end of life is substantiated through further research, this will suggest an urgent concern for public health and health services caring for people who are at the end of life. There might be a need to review whether loneliness is sufficiently addressed through current palliative and long-term care, as well as in social services, for older people. Concerning methodology, by showing the distinct roles that different aspects of social connection play in shaping various end-of-life outcomes, our study encourages greater conceptual clarity in future research on social connection and end-of-life outcomes. Finally, we need a better understanding of cross-cultural as well as inter-individual variation in changes in social connection and the impact of social connection changes on wellbeing as older people approach death. To reveal potential variations in trajectories of social connection at the end of life, studies would require still larger samples and more measurement points. At the same time, interdisciplinary, qualitative, and ethnographic research is needed to advance our understanding of the interpersonal, emotional, cognitive, and behavioural mechanisms through which social connection interacts with health and wellbeing in the final phase of life.⁴⁰

Data sharing

Data used in this study are freely available through the SHARE project upon registration on the project website (https://share-eric.eu/).

Declaration of interests

LP and LVdB received research funding payments to their institutions from Research Foundation-Flanders. LP received research funding payments to their institution from the European Research Council. LVdB received research funding payments to their institution from the Francqui Foundation. FP declares no competing interests.