Loneliness as a risk factor for dementia and its mediators: A longitudinal cohort analysis of UK Biobank data

Abstract

Background and objectives

Growing evidence suggests that the feeling of loneliness correlates positively with the incidence of dementia, motivating our study of their longitudinal association. Additionally, we sought potential mediators of the association, following the hypotheses that lonely individuals are prone to greater stress and follow less healthy lifestyles, therefore are at greater risk of developping dementia.

Methods

Using longitudinal cohort data from the UK Biobank, we conducted a logistic regression analysis with loneliness as exposure and dementia diagnosis as outcome, and progressively adjusted for putative confounders. In addition, we investigated through mediation models the respective contributions of physiological and behavioural factors to the association between loneliness and dementia diagnosis.

Results

The logistic regression models illustrated a significant positive association between loneliness and a subsequent dementia diagnosis, overall indicating a higher odds of developing dementia in the lonely population (Adjusted Odds Ratio (AOR) = 1.39; 95 % (Confidence Interval) CI 1.24 – 1.56, P < 0.001). The mediation models highlighted that there was very little evidence of mediation by hypertension and unhealthy behaviours, respectively explaining 1 % and 2 % of the effect of loneliness on dementia risk. No indication of mediation was seen between baseline CRP-levels and all-cause dementia. Every mediation model described a positive correlation between loneliness and each potential mediating factor, illustrating the widespread negative impact of loneliness on health.

Conclusion

Our findings confirm the association between loneliness and risk of dementia. They show minimal mediation by hypertension and unhealthy behaviours, highlighting the need for investigation into other potential mediators as targets for modification.

Introduction

As described by the World Health Organisation (WHO), dementia is a ‘syndrome – usually of a chronic or progressive nature – in which there is deterioration in cognitive function beyond what might be expected from normal ageing’.¹ Dementia affected more than 50 million people worldwide in 2020, a population expected to triple by 2050.² Dementia can have a profound impact on affected individuals and their families and carers.³ Due to a current lack of disease-modifying treatments, research must focus on identifying and targeting potentially modifiable risk factors.

Loneliness and dementia

Loneliness is defined as “a distressing feeling that accompanies the perception that one's social needs are not being met by the quantity or quality of one's social relationships”.^4,5 Unlike social isolation, which can be objectively measured through frequency of social interactions or living arrangements, loneliness is a subjective condition. It has recently been estimated that a third of individuals in industrialized countries are affected by loneliness.⁶

Global demographic ageing is frequently accompanied by increasing numbers of people living alone, often with higher levels of loneliness.⁷ Loneliness has been identified as a risk factor for Alzheimer's disease (AD) and all-cause dementia in two major cohort studies with long follow-up periods, of 3 and 15 years respectively.^8,9 A high level of social integration or recovery from a transient state of loneliness has conversely been linked to reduced rates of cognitive decline.¹⁰ and in one study was found to have protective effects against dementia, contrary to persistent loneliness.⁹

Potential pathways linking loneliness and dementia

Multiple models have been proposed to illustrate the pathways between loneliness and dementia.¹¹ The physiological pathway describes stress-related responses to the feeling of alienation. Loneliness has been found to predict higher systolic blood pressure, a known risk factor for dementia.^3,12 Lonely individuals have heightened stress exposure and lack social relationships capable of buffering their stress responses,¹³ which may cause chronic inflammation. Studies measuring C-Reactive Protein (CRP) levels in lonely individuals gave mixed results: some found higher levels linked to feelings of loneliness,¹⁴ or to perceived stress associated with loneliness,¹⁵ while others showed no association.¹⁶ Increasing evidence suggests that inflammation plays an important role in brain ageing and neurodegeneration, with elevated CRP implicated in the neuropathology of dementia.^17–19

Loneliness is also associated with smoking²⁰ and low physical activity.²¹ Both contribute to the risk of developing cardiovascular diseases,²² atherosclerotic small vessel disease and neurofibrillary tangles (NFTs), characteristic of AD.²³

People who drink alcohol tend to experience frequent social contact,^24,25 but chronic alcohol users have reported antecedents of depression, loneliness and lack of social support.^26,27 Excessive alcohol use can cause alcohol-related brain damage and dementia.²⁸ Therefore, excessive alcohol consumption may partly explain any increased incidence of dementia in lonely individuals.

The primary aim of this study was to examine the effects of baseline loneliness on the development of dementia, adjusting for potential confounders. As a secondary aim, we investigated hypertension, inflammation, and unhealthy behaviours as putative mediators of any association between loneliness and dementia.

Methods and materials

Data source

The data presented in this study was drawn from the UK Biobank (UKB; www.ukbiobank.ac.uk) and ethical approval was obtained by the UKB from the National Research Ethics Service Committee North West.²⁹ Potential Biobank participants were people who lived within approximately 10 miles of one of 22 assessment centres around the UK, and were recruited through postal invitation.³⁰ Details regarding data collection methods and the sample used have been described elsewhere.³¹ We obtained ethical approval from the UK Biobank prior to acquiring data (project ref. 40,055).

Measures

Main exposure – loneliness scale

A single question to measure loneliness has been shown to correlate highly with multi-item validated scales such as the UCLA loneliness scale.^5,32 According to the Office for National Statistics, the “gold-standard” for measuring loneliness is to use a combination of direct and indirect measures of loneliness.³³

In line with previous studies,³⁴ we constructed a loneliness scale from participants’ answers to the questions “Do you often feel lonely?” (“No” = 0, “Yes” = 1) and “How often are you able to confide in someone close to you?” (“Almost daily” to “About once a month” = 0, “Once every few months” to “Almost never” = 1). An individual was defined as lonely if they responded positively to either or both questions (score ≥ 1) and not lonely otherwise (score = 0).

Main outcome – dementia

Our main outcome measure was a clinical diagnosis of “all-cause dementia” obtained from self-report, or from linked electronic hospital, primary care or mortality records. We chose all-cause dementia due to the relative under-reporting of specific dementia subtypes and the shared risk factors between the most common forms of dementia. The positive predictive value (PPV) of all-cause dementia across all UKB datasets is reported to be 82.5 %.³⁵

To address the issue of reverse causality raised by several recent studies, which suggests that loneliness may be a prodromal feature of dementia rather than a risk factor to developing dementia,^5,11,36 we excluded participants with a dementia diagnosis recorded at baseline or within 3 years of their initial assessment.

Confounders

Based on existing literature, we adjusted our analysis for potential sociodemographic, clinical and behavioural confounders (Fig. 1).

Fig. 1.

Adjusting for sociodemographic, clinical and behavioural confounders, social isolation and depression.

Sociodemographic confounders

We included: sex (male / female), age at baseline (in years), genetic ethnic grouping (Caucasian / Other), age of leaving education, and Townsend deprivation index score.³⁷

Clinical and behavioural confounders

We adjusted for the following clinical and behavioural confounders: Body Mass Index (BMI) as a continuous measure measured at baseline; high blood pressure, as a binary variable, using as cut-off value an average 140/90 mmHg; diabetes, as a binary variable based on diagnostic status; physical activity, as a binary variable based on current WHO guidelines of 600 Metabolic Equivalent Task min/week;³⁸ smoking status (never / previous /current); and alcohol consumption (yes /no).

Social isolation

Based on participants’ answers from a baseline touchscreen questionnaire, we used a binary measure of social isolation. It divided our cohort into participants considered socially isolated, having less than almost daily social contact, and those not isolated, with more frequent social activity.

Depression

We classified as depressed anyone who reported seeking help for depression from a GP or psychiatrist or using anti-depressants.

Potential mediators

• 1. Hypertension

We categorised as hypertensive anyone with baseline blood pressure >140/90 mmHg, and anyone who reported using anti-hypertensive medications.

• 2. C-reactive protein levels

The UKB included baseline measures of C-Reactive protein (CRP) levels in mg/L which were used, unchanged, in our study.

• 3. Unhealthy behaviour score

From self-reported data on smoking status, alcohol consumption and physical activity, we created an ‘unhealthy behaviour’ scale, with scores ranging from 0 to 3, on which higher scores represent less healthy lifestyle practices. Being a current smoker, drinking alcohol above recommended limits (>21 units/week) and being physically inactive (<600 MET minutes/week) each scored one point.

Exclusion criteria

Participants with missing data in any of the analytic variables were excluded from the study. Only people without a dementia diagnosis at study baseline or within three years of their start date were included.

Statistical analysis

We summarised baseline characteristics of the whole sample, dividing it between those who were lonely and not lonely at baseline. We also compared baseline characteristics for participants who had complete data and for whom any data on exposure or confounders was missing.

All analyses were carried out in STATA, version 16.1.³⁹

• I Logistic regression analysis

We used logistic regression analyses, with loneliness as exposure and all-cause dementia diagnosis as outcome. Model 1 was a univariable model assessing the unadjusted association between loneliness and dementia. We then progressively adjusted for sociodemographic confounders (Model 2; adjusted for age, sex, genetic ancestry, age when left education, and Townsend deprivation index); clinical and behavioural confounders (Model 3; respectively BMI, hypertension, diabetes, and less healthy lifestyle score); social isolation (Model 4) and finally, depression (Model 5).

Using STATA's marginsplot command, we plotted the fully adjusted probabilities of developping all-cause dementia by loneliness score, according to age at baseline (Fig. 2); allowing us to explore the impact of loneliness based on participants’ age.

• II Mediation analysis

Fig. 2.

Graph of adjusted predictions of developing dementia by loneliness (including 95 % confidence intervals).

We tested the association between our exposure and outcome variables and potential mediators; hypertension, CRP levels and less healthy lifestyle score (Fig. 3).

Fig. 3.

Mediation model.

In our models, a represents the relation between loneliness and the potential mediator, b the relation of the potential mediator to all-cause dementia and c’ the relation of loneliness to all-cause dementia, adjusted for the mediator (Fig. 2).

This analysis was performed through the STATA module paramed which allows causal mediation analysis using parametric regression models while adjusting for covariates. Paramed provides estimates of the controlled direct effect, natural indirect effect and total effect. The direct effect is the coefficient c’. The indirect effect is the product of path coefficients a and b. Finally the total effect is equal to the sum of direct and indirect effects (c’ + ab).

We calculated the proportion of the effect mediated by our potential mediators using the following formula:⁴⁰

Results

From the initial UKB baseline sample of 502,490 participants, we excluded 251 participants who had dementia at baseline or within three years of their start-date in the study. We excluded 88,694 (17.7 %) of the full sample due to missing data.

The analytic sample comprised 413,545 participants, of which: 54.4 % were female, 84.0 % were classified as Caucasian based on their genotyping, and ages at study baseline ranged from 38 to 73 years/old (mean ± SD = 56.5 ± 8.1; Table 1). Participants were followed up for a mean of 12 ± 1.6 years (± SD).

From our complete case sample, 133,437 [32.3 %] participants were classified as lonely at baseline according to our loneliness scale (Table 1). Differences between lonely and non-lonely cohorts were minimal. However, compared to lonely individuals, a higher proportion of non-lonely participants had received education above the age of 16 (60.4% vs 52.8 %) and a higher proportion of lonely participants were physically inactive (13.2% vs 8.4 %); were socially isolated (25.2% vs 10.9 %); and reported seeking help from a GP or psychiatrist for depression (44.0% vs 30.1 %) compared to those who were not lonely.

No noticeable differences were found between groups with and without missing data, except that participants with complete data were more likely to be of Caucasian genetic ancestry (84.0 %) than those with missing data (69.9 %). (Supplementary Table 1).

Logistic regression analysis

Results can be seen in Table 2. In a univariable model (Model 1), we found that being either lonely or unable to confide to someone close (loneliness score = 1), or both (score = 2), increased the odds of developing dementia, respectively presenting Odds Ratios (OR) of 1.41 (95 % CI: 1.32–1.50) and 1.50 (95 % CI: 1.35–1.67).

We found strong evidence of a positive association between loneliness and dementia in all models. Adjusting for sociodemographic confounders did not substantially change the OR estimates of developing dementia in lonely participants (score = 2; Adjusted Odds Ratio (AOR) = 1.55, 95 % CI: 1.40–1.73). In the subsequent models, the ORs slightly decreased but remained strongly positive. From our fully adjusted model (Model 5), being lonely and unable to confide gave an AOR of 1.39 of developing all-cause dementia (95 % CI: 1.25–1.55) while being either lonely or unable to confide gave an AOR of 1.27 (95 % CI: 1.19=1.35).

As illustrated in Fig. 2, higher baseline loneliness predicted a greater chance of developing dementia, which exceeded the observed increases in odds of developing dementia seen with age (illustrated through a score of 0 on the loneliness scale; Fig. 2). The effect of loneliness on a dementia diagnosis becomes apparent in people aged between 57 and 61 at baseline.

For participants between 69 and 73 years of age at baseline, we can see that the probability of developing dementia in lonely individuals was significantly higher than in non-lonely individuals (Table 3; 95 % CI of non-lonely individuals does not include the probability of lonely individuals). These margins tell us that if the distribution of dementia diagnosis remained the same in the population, but everyone had a loneliness score of 0, we would expect about 9 % to develop dementia (Table 3). If however, everyone had a loneliness score of 2, we would expect about 12 % of the population to develop dementia.

Mediation analysis

Results are shown in Table 3.

Hypertension

Against our hypothesis, we found a negative coefficient (coef. = −0.6513, 95 % CI: −0.741 – −0.5615, Fig. 4) between loneliness and SBP, indicating that greater loneliness was associated with lower SBP in participants from the UKB. We performed a sensitivity analysis adjusting for blood pressure lowering medications, which also output a negative coefficient, confirming that baseline SBP could not be used as mediating variable in our model.

Fig. 4.

Fully adjusted mediation model with systolic blood pressure (SBP) as potential mediator, baseline loneliness as independent variable and diagnosis of all-cause dementia as dependent variable. We adjusted for age, sex, genetic ancestry, education, Townsend deprivation index, BMI, diabetes, physical activity levels, smoking, social isolation & depression.

We instead explored hypertension as potential mediator, defined as baseline blood pressure > 140/90 mmHg or taking blood pressure lowering medications. According to our model 1 (Fig. 5), hypertension may be a mediator of the effect of loneliness on dementia. The mediating effect of hypertension was however very small: approximately 1 % (Fig. 6).

Fig. 5.

Fully adjusted mediation model with hypertension as potential mediator, baseline loneliness as independent variable and diagnosis of all-cause dementia as dependent variable. We adjusted for age, sex, genetic ancestry, age when left education, Townsend deprivation index, BMI, diabetes, physical activity, smoking status, social isolation, and depression.

Fig. 6.

Fully adjusted mediation models with hypertension as potential mediator. The number before each ‘Loneliness’ independent variable indicates a score on the loneliness scale (ranges from 0 – 2).

Mediation model 2: CRP-levels as potential mediator

As shown in Fig. 7, there was a small positive association between loneliness and CRP levels (Fig. 7; coef. = 0.083, 95 % CI: 0.061 – 0.105), but baseline CRP-levels and dementia were not associated (coef. = −0.001, 95 % CI: −0.007 – 0.005). CRP-levels, although related to loneliness, do not appear to mediate the effect of loneliness on all-cause dementia (Fig. 8).

Fig. 7.

Fully adjusted mediation model with baseline C-reactive protein (CRP) levels as potential mediator, baseline loneliness as independent variable and diagnosis of all-cause dementia as dependent variable. We adjusted for age, sex, genetic ancestry, age when left education, Townsend deprivation index, BMI, diabetes, physical activity, smoking status, social isolation, and depression.

Fig. 8.

Fully adjusted mediation models with baseline CRP-levels as potential mediator. The number before each ‘Loneliness’ independent variable indicates a score on the loneliness scale (ranges from 0 – 2).

Behavioural pathway

Mediation model 3: Unhealthy lifestyles as potential mediator

Our final mediation model (Fig. 9) shows that less healthy lifestyles are positively associated with both greater loneliness (Fig. 9; a = 0.042, 95 % CI: 0.039 – 0.045) and all-cause dementia (b = 0.092; 95 % CI: 0.045 – 0.129). However, the estimated proportion of effect mediated was very small; approximately 2 % (Fig. 10).

Fig. 9.

Fully adjusted mediation model with unhealthy lifestyles as potential mediator, baseline loneliness as independent variable and diagnosis of all-cause dementia as dependent variable. We adjusted for age, sex, genetic ancestry, age when left education, Townsend deprivation index, BMI, hypertension, diabetes, social isolation, and depression. The score for unhealthy lifestyles comprised of measures of alcohol consumption, smoking status and physical activity levels. Details can be found in ‘Methods – Potential mediators – 3. Unhealthy score’.

Fig. 10.

Fully adjusted mediation models with unhealthy lifestyles as potential mediator. The number before each ‘Loneliness’ independent variable indicates a score on the loneliness scale (ranges from 0 – 2).

Discussion

Main findings

In this large prospective study, from the UK Biobank, we sought to investigate the impact of loneliness at baseline on the risk of subsequent all-cause dementia. As hypothesised, we found a significant positive association between loneliness and clinical diagnosis of dementia.

The mediators we tested, hypertension and unhealthy lifestyles explained approximatively 1 % and 2 % of the impact of loneliness respectively. CRP did not mediate the association between loneliness and dementia. The impact of loneliness on the risk of dementia therefore remains mostly unexplained, and loneliness could be an independent risk factor for all-cause dementia.

Another theory is that reductions in quality and frequency of social interactions associated with loneliness result in decreased cognitive stimulation. This potentially leads to a reduced cognitive reserve, exaggerating age-related neuropathological changes and associated decline in cognitive and executive functioning.¹⁰ A recent neuroimaging study demonstrated an association between high β-amyloid burden and loneliness in cognitively normal patients.⁴¹

Contrary to our hypothesis and previous observations that lonely individuals have higher SBP than controls, we found that participants with greater loneliness scores had on average lower SBP. This finding could be explained by the action of blood pressure lowering medications, consistent with our finding of a small mediating effect of diagnosed hypertension on the effect of loneliness on dementia.

Comparison with other studies

Our analysis is consistent with research in the field which highlighted the increased risk of developing dementia in lonely populations, from generally smaller cohorts than the UKB.^8,42,43 Despite substantial evidence of an association between loneliness and dementia risk, several longitudinal studies did not find an association,^41,44,45 leading some authors to suggest that reverse causality may account for the association, with loneliness a prodromal feature rather than a risk factor.³⁶ By excluding cases of dementia at baseline or within three years of the participant's start-date in the UKB study, we were able to reduce the likelihood that this drove the association.^36,46

Our finding of no association between higher CRP-levels and all-cause dementia contradicts findings from several past studies.^19,47,48 However, the exact relationship between CRP levels and dementia remains debated. Some studies have found that patients with lower CRP levels tended to develop Alzheimer's dementia more rapidly,¹⁷ while others found that higher levels of CRP are associated with increased odds of developing all-cause dementia.¹⁹ These discrepancies may be due to the physiological differences behind different dementias.⁴⁸ The lack of correlation observed in our mediation model may be explained by the relatively younger age of our participants at baseline, when CRP levels were measured. In studies that found a positive correlation between CRP levels and dementia risk, the mean age of patients was above 70 years old,^19,47,48 whereas our cohort at baseline was on average 58 years of age.

Although the mediating effects of the studied variables were very small, we have been able to show through our mediation analysis that loneliness positively correlates with hypertension, baseline CRP-levels and unhealthy behaviours. In this sense, our mediation models were able to reiterate the previously illustrated impacts of loneliness on these basic physiological and behavioural measures.

It would be of interest for future research to consider an approach where multiple mediators are considered simultaneously.⁴⁹ This would allow for a better understanding of the human body as a whole, rather than considering medical conditions isolated.

Strengths and limitations

This study is one of the first illustrating the impact of loneliness on dementia with a very large cohort. To the best of our knowledge, it is also one of the first to explore potential underlying mechanisms of loneliness with the aim of understanding the pathways to dementia. We have adjusted for multiple confounders, and used linked electronic health records data to improve confidence in outcome reporting. Data on loneliness was obtained before dementia diagnosis. A further strength was integration of participants’ ability to confide in someone close on the loneliness scale, which thus contained both direct and indirect measures of loneliness, as a means to circumvent the influence of stigma on reporting.⁵⁰

We also acknowledge limitations. Many of our variables relied on self-reported data, which is more likely to lead to more inaccuracies than objective measures.

We were unable to study the feeling of loneliness longitudinally, due to small numbers reporting this measure during follow-up visits. For this reason, we could not compare chronic and acute loneliness as exposures. The dementia prodrome may be long, with almost 20 years of pre-clinical deterioration within the brain, so we cannot be certain that our exclusion period of diagnosis within 3 years of baseline was adequate. Nevertheless, past studies have performed sensitivity analyses which support similar exclusion criteria being sufficient to counter this effect of reverse causality.³⁶ Regarding data collection, the UKB recruited volunteers and had a response rate of just 5.5 % at recruitment.⁵¹ Therefore, we must consider the possibility of selection bias, where the cohort is not representative of the UK population overall.⁵² For example, sampling mostly healthy volunteers could explain the very small mediating effects of hypertension and unhealthy behaviours. Participants are more likely to be female and white, and less likely to live in deprived areas and to have serious health conditions, relative to the general population.⁵³ This might plausibly have led to reduced incidence of dementia and weakened associations. It might also have had an effect on the proportion of people reporting often feeling lonely in our study, which was higher than comparable studies.^54,55 suggesting that our sample is over-representative of the lonely. Therefore, caution must be applied in generalising findings from this study.

Using a binary measure of loneliness might have missed differences in risk according to degree of loneliness. However, the question about how often participants felt lonely only allowed for a binary response.

Measures of CRP are non-specific and rises can be caused by multiple conditions, which could also explain the lack of correlation seen with dementia. We note that our null findings regarding potential mediators of the association between loneliness and dementia are inconsistent with previous literature.^56–58 This may be due to the causal chain from loneliness to dementia being more complex than our mediation models can allow for. Residual and unmeasured confounding can never be fully excluded in longitudinal studies, and so it cannot be assumed that associations are causal. Data on the exposure and mediators was obtained at the same time point, so we cannot be certain that loneliness preceded the mediator, rather than the other way around.

We were not able to use multiple imputation due to the vast computing power required to perform this with such a large sample. It is also possible that a degree of collider stratification bias occurred as the sample was naturally stratified on people who survived long enough to develop dementia, and who may be more lonely if fewer of their peers survived with them.⁵⁹ However, there is no particular reason to think that the relationship between loneliness and dementia would be different in this group than those who died before diagnosis.

Conclusions

Although the exact mechanisms through which loneliness acts on dementia risk remain undetermined, we found further evidence that loneliness might be a modifiable risk factor for dementia. Greater awareness of this among both clinicians and society could lead to interventions to reduce loneliness, with potential public health benefits not just on dementia risk but other health outcomes. In future research, it would be of interest to distinguish between the different types of dementia. Other health behaviours hypothesized to play a role in the relationship between loneliness and dementia, such as diet⁶⁰ or sleep⁶¹ could also be explored. Loneliness prevention could include attempts to destigmatize discussion around loneliness and increase availability of social prescribing.

Ethical considerations

This study used unidentifiable data from the UK Biobank (project ref. 40,055). Analysis scripts are available on request from ME ([e-mail]).

Contributors

NM, NS and ME designed the study. NM acquired access to the UKB database. ME performed the statistical analysis with input from NS and NM. ME drafted the work and all authors contributed to the interpretation of findings in addition to critically reviewing the manuscript for intellectual content. A common decision including all authors was made to submit our findings for sharing.

Funding and conflict of interest

NS is supported by a National Institute of Health Research FellowshipNIHRDHNIHR300703. NM is funded by an Alzheimer's Society Senior Fellowship (AS-SF-18b-001).

All authors are supported by the University College London National Institute of Research Biomedical Research Centre.

No specific funding was received for this paper.

We are extremely grateful to all participants in the UK Biobank and the UK Biobank team.

On behalf of all authors, there are no conflicts of interest to declare.