Positive and Negative Experiences of Social Support and Risk of Dementia in Later Life: An Investigation Using the English Longitudinal Study of Ageing

Study sample

We analyzed data on a cohort of 10,055 core participants who were dementia free in 2002/3 (wave 1) from ELSA [20], a panel study of a representative sample of men and women aged 50 and over living in the community in England. In addition to being dementia free at baseline, participants of the study sample were required to complete the questionnaire for positive or negative social support ( Supplementary Table 1) for at least one of the relationships (spouse, children, other immediate family, friend). We refer “other immediate family” or “other family” to indicate family members not including children and spouse such as brothers, sisters, cousins, parents, or grandchildren. ELSA involves data collection by computer-assisted personal interviews every two years, and we identified dementia incidences from up to six waves, covering a period of ten years. The exposure variables (the experiences of positive and negative social support) were measured at baseline (wave 1).

Ethical approval

Ethical approval for all the ELSA waves was granted from the National Research and EthicsCommittee (http://www.nres.npsa.nhs.uk/).

Identification of dementia incidence

Information about dementia was reported either by the participant or an informant if an eligible participant was physically or cognitively impaired, in hospital or temporarily in care during the interview period. Any responsible adult (aged 16 years or over) who knew enough about the respondent’s circumstances to be able to provide information about them, such as a close family member (partner or child), fulfilled the role of an informant. Incidence of dementia during the study period [wave 2 (2004) to wave 6 (2012)] was identified from several sources of information: (i) participant or informant reported physician diagnosed dementia or Alzheimer’s disease, or (ii) overall score of the informant completed IQCODE questionnaire. The short-form IQCODE questionnaire [21] consists of 16 items asking the informant to comment on the ability of the person compared with 10 years ago to perform various functions (e.g., remembering the names of family) on a 5-point scale ranging from 1 (much improved) to 5 (much worse). We used the cut-off point of 3.5 of the overall score (average of the responses on all 16 items) to define dementia which has high specificity and good sensitivity[22]. On the basis of the above criteria, there were 340 incident cases of dementia within the study cohort during wave 2 (2004) to wave 6 (2012). Of the 340 cases, 284 were identified from the reports of physician diagnosed dementia or Alzheimer’s disease and the remaining 56 were identified based on the IQCODE score.

Exposures: Experience of positive and negative social support

Measures of positive and negative experiences of social support for each relationship (spouse or partner, children, other immediate family, friends) were calculated based on responses at baseline (wave 1) on a set of six items within the self-completed “Health and lifestyle of people aged 50 and over” questionnaire. Three items were used to measure positive experiences of social support and the remaining three for negative experiences (see Supplementary Table 1). All items were measured on a 4-point scale ranging from 1 (at lot) to 4 (not at all). We reverse coded the scales so that higher value indicated more of the positive or negative experiences: 1 (not at all) to 4 (a lot). For each individual relationship (spouse, children, other family, or friend), experience of positive or negative support score was calculated by averaging the reverse coded scores of the three items (a, b and c) displayed in Supplementary Table 1. We have also calculated four sets of overall scores by averaging the scores for various combinations of individual relationships: (1) mean across all four relationships (spouse, children, other family, friends), (2) mean across spouse, children, and other family, (3) mean for spouse and children, and (4) mean for other family and friends. In calculating the mean scores from more than one relationship, we considered the mean of available (non-missing) data. Summary statistics (mean and standard deviations) of the overall scores as well those for individual relationships by gender are displayed in Table 1.

Outcome measure: Time-to-dementia

The outcome variable was defined as time-to-dementia from the start date of the ELSA study. As the follow-up interviews of the ELSA participants take place every two years, neither the time-to-dementia for incident cases nor the censoring time (due to drop-out/lost to follow-up or not developing dementia during the study period) was known exactly. Therefore, both the event time (time-to-dementia) and censoring time were treated as interval censored between two consecutive waves (a two-year interval) with an event indicator (1 for dementia incidence) distinguishing between dementia events and censoring. The participants who did not develop dementia by wave 6 were treated as censored at 120 months and were assigned within the final interval. For example, a participant reporting dementia at wave 3 (2006) who was dementia free at wave 2 (2004) was considered as having the disease at any time between 2004 and 2006 and was assigned the interval (24, 48] months as their time-to-event. We did not distinguish between losses to follow-up due to deaths and due to other reasons.

Covariates

We adjusted our analyses for potential confounding variables. We selected covariates for adjustment using both hypothesis and knowledge based approach as well as based on statistical assessment of confounding. The covariates age, sex, and net wealth, which are believed to have close links with cognition and family/social relationships [23–25], were included in all models. Net total wealth represented the sum of personal savings, investments, physical wealth, and housing wealth after financial debt and mortgage debt have been subtracted, and is a key indicator of socioeconomic resources among older people [20]. We have also statistically assessed several other covariates including education and co-morbidity conditions such as diabetes, cardiovascular disease, stroke, hypertension, and cancer for their role as potential confounders. Education was coded as a categorical variable with three levels (no education, primary/secondary education, and higher education). All the co-morbidity variables were defined as binary indicators (yes/no). Covariates passing statistical test for confounding effect, i.e., those were associated with both the exposure and outcome, were accounted for in the respective models in addition to age, sex, and net wealth.

Statistical analysis

Proportional hazards (PH) regression models were used allowing for interval censoring of time-to-dementia for estimating and testing the effects of positive and negative social exchanges/support on the risk of developing dementia. Time-to-event or survival data are said to be interval censored when a subject’s event time is not known exactly, but only that it lies between two values. Data of this type typically arise where follow-up is intermittent at fixed intervals. For the subjects who reported dementia at a follow-up (wave), all that is known is that time-to-dementia is at least as long as the time ofthe earlier follow-up and no longer than the time of the later follow-up. To accommodate this feature of the data in a PH regression model, we fitted the model by maximizing a transformed likelihood that enabled estimating the parameters of the PH model using standard statistical software (Stata). Hosmer et al. [26] (Chapter 7, pp. 231–243) showed that with some re-arrangement of data, the likelihood function for a PH survival regression model under interval censoring of events can equivalently be represented by the likelihood of a generalized linear model (GLM) of the event indicator (binary outcome) under binomial family and complementary log-log as the link (or linearizing) function. The amount of missing data on covariates were minimal with no missing data on age and sex, and only 1.5% missing data on net wealth. We therefore performed analysis of available data without any imputation of missing observations. However, measures on the exposure variable (positive or negative social support scores) for some individual relationships were not available for a substantial proportion of participants. For example, positive or negative support scores from spouse were not applicable for approximately 29% of the study participants who did not have a spouse or partner, and the corresponding analysis had to be based on a reduced sample. Number of subjects included in the PH regression analysis for each relationship is shown in the table of summary statistics (Table 1). Results of the primary analysis based on the above model are presented in Table 2.

Sensitivity analysis

Three sensitivity analyses were conducted in order to assess the robustness of the findings of the primary analysis. We hypothesized that the quality of relationships with spouse, children, other family, and friends influenced the risk of developing dementia in older age. Dementia, however, often takes a toll on social relationships, and so people who are very close to developing dementia may already show signs of cognitive problems, and experience deterioration of relationships. It is therefore possible that reverse causality partly explains the associations observed between quality of relationships and dementia. The first sensitivity analysis was designed to assess the robustness of the findings against possible reverse causality by excluding dementia cases diagnosed during the first two waves (i.e., 2 years) of the study.

The second sensitivity analysis was to assess robustness of the findings against the reliability of diagnosis of dementia by excluding cases diagnosed using the IQCODE, and limiting the analysis to physician diagnosed dementia or Alzheimer’s disease only (284 incident cases).

The third sensitivity analysis compared the findings from the interval censored PH regression analysis with that of a standard Cox PH regression where the time-to-dementia was assumed to be known exactly. This is the most commonly used analysis for time-to-event data with standard censoring mechanism. Participants reporting dementia at wave 2 were assigned 24 months as their time to event, 48 months for those reporting at wave 3, 72 months at wave 4, and so on. Subjects not developing dementia by wave 6 were treated as right censored at that point. Loss to follow-up due to death or for any other reason was also treated as right censored. Results for the sensitivity analyses are presented in Supplementary Tables 1–3.

Main analysis

The primary analysis included 10,055 dementia-free core participants at baseline (wave 1), with 5,475 (54%) male and 4,580 (46%) female subjects. There were 340 (3.4%) incident cases of dementia during the study period (wave 2 to wave 6) of which 190 were male and 150 female cases. Actual number of participants included in individual regression models, however, varied between 7,067 and 10,055 with incident dementia cases varying between 199 and 340, respectively, due to some missing data in positive or negative social support scores. The number was lowest for analyzing the spouse’s score (7,075 for positive, and 7,067 for negative score) as this score is not applicable for participants without having a spouse or partner (see Table 1 for more details).

Participants were aged 50 years and over with overall mean age 65 years and standard deviation 10 years. Overall, men were more likely (56%) to have dementia than women (44%). Participants with dementia were on average older at baseline (mean age = 76 years for men and 74 years for women) than the participants without dementia (mean age = 65 years for men and 64 years for women). In general, positiveand negative support scores for each relationship were negatively correlated as expected (Pearson’s correlation, r = –0.50, –0.34, –0.28, and –0.11 for spouse, children, other immediate family, and friend respectively). Correlation between positive and negative scores for cross-relationships (e.g., between positive scores for spouse and negative scores for children) were also negative, but weaker than those for within relationship correlation (varied from –0.02 to –0.15). Additional summary statistics of the covariates and positive and negative social support scores stratified by gender for individual relationships and their combinations are given in Table 1.

Results from the interval-censored PH regression analysis assessing the effects of positive and negative social support scores are displayed in Table 2. Of the positive social support measures, a higher score led to reduced estimated adjusted risk of dementia (estimated hazard ratio <1) for all the four relationships and their combinations (Table 2). An increase of one point in the positive social support score led to between 7% to 17% reduction in the risk of developing dementia (hazard ratio, HR, ranging between 0.83 and 0.93). However, only the effect of positive social support from children was found to be statistically significant at the 5% level (HR = 0.83, p = 0.042, 95% CI = 0.69 to 0.99). Effects of positive support scores from spouse (HR = 0.83, p = 0.107) showed a trend toward significance with the same effect size (HR = 0.83) as that for children.

Negative social support measures, on the other hand, showed an opposite effect on the risk of developing dementia— overall, an increase in the magnitude of negative social support resulted in an elevated estimated risk (HR >1) of dementia. Negative support scores showed stronger effects than the positive scores, leading to 8% to 31% increased risk of incident dementia (HR ranged between 1.08 and 1.31) for various relationships, although not all were statistically significant. Effects of negative experience of support from other immediate family (HR = 1.26, p = 0.011, 95% CI = 1.05 to 1.50), combined negative scores from spouse and children (HR = 1.23, p = 0.046, 95% CI = 1.004 to 1.51), combined negative scores from spouse, children, and other family (HR = 1.27, p = 0.021, 95% CI = 1.04 to 1.56), combined negative scores from other family and friends (HR = 1.25, p = 0.033, 95% CI = 1.02 to 1.55), and the overall negative scores (combined across spouse, children, other family, and friends) (HR = 1.31, p = 0.019, 95% CI = 1.05 to 1.64) were all statistically significant at 5% level. Regression coefficients for the covariates age and net wealth were significantly associated with dementia in all models showing older and poorer at greater risk of dementia. Effects of education and diabetes were also statistically significant in some models with more education leading to lesser risk and presence of diabetes leading to higher risk of developing dementia. The effect of sex was not statistically significant in our study, although some studies [27] in the literature reported women to be at higher risk than men. Since higher incidence of dementia among women may be present primarily in those over 90 years [28], the younger ELSA cohort may explain why gender was not significant in our study. We, however, decided to keep sex in all models as the effect size was considerable in some of the models.

Sensitivity analysis

Sensitivity analysis excluding dementia cases within two years of baseline resulted in a smaller number of incident cases (292 dementia cases between 2006 and 2012) (see Supplementary Table 2). The reduced number of events resulted in a loss of statistical power (slightly inflated standard errors) and led to non-significant results for some of the effects found in the main analysis. However, the magnitude of the effects (hazard ratios) and direction of the associations remained similar to that of the main analysis. Several of the adjusted associations remained statistically significant despite the loss of statistical power.

The second sensitivity analysis excluded cases diagnosed using the IQCODE, and limited the analysis to physician diagnosed dementia or Alzheimer’s disease only (results reported in Supplementary Table 3). This again resulted in a loss of events (incident cases of dementia ranged between 164 and 284 for different models) and statistical power. Analyses however led to the same conclusion that an increase in positive experience of social support reduces the risk of dementia (HRs ranged between 0.79 and 0.95) and that in negative experience increases the risk (HRs for each unit increase in negative score varied from 1.06 to 1.30).

The third sensitivity analysis using standard Cox PH regression model, which assumed time-to-dementia was known exactly and did not accommodate interval censoring of event times, produced almost identical results to that of the primary analysis taking account of interval censoring ( Supplementary Table 4).

Summary of main findings

Our multivariable analysis demonstrated that, irrespective of the source (type of relationship) of social support, overall negative support was significantly associated with an increased risk dementia. Some measures of positive support significantly reduced the risk of dementia. More specifically, negative social support from spouse, children, and other immediate family increased the risk whereas positive support from children was associated with a reduced dementia risk. Effect sizes were generally larger for negative compared with positive social support. As expected, the magnitude of the favorable effects of positive support appeared greater for closer kinship (e.g., children compared with friends) but this pattern was less clear for negative support. Relatively stronger associations for the negative social support relative to the positive support may be indicative of the fact that stress of criticism and lack of reliability are possibly more harmful than the absence of a warm relationship. The robustness of these results was confirmed in a series of sensitivity analyses.

Comparison with previous investigations

To the best of our knowledge, this is the first study to investigate the influence of both positive and negative social support on incident dementia in community-residing older people. Therefore, a direct comparison of our results with previous studies was not possible. We, however, compared our findings with that from studies looking at the effects of positive and negative social support on general health and well-being of older people although such comparisons are still challenging due to important methodological differences across investigations. Our findings appear broadly consistent with a substantial literature documenting the disproportionate impact of negative social support on general health and well-being outcomes in older people [14]. For example, our results corroborate recent findings on the harmful impact of overall negative support on depressive symptoms [15], decline in activities of daily living and instrumental activities of daily living functioning [29], levels of psychological well-being and distress [16], and on disability [17]. In contrast, negative exchanges were not related to all-cause mortality in a sample middle-aged Dutch people [30]. Specifically, we demonstrated significant effects on dementia risk of negative interactions with both spouse and children when combined together but not when they were examined separately. In contrast, Okun and Keith [15] reported significant effects of negative support from spouses, but not children, on depressive symptoms in their analysis of olderadults.

Despite the effect of overall positive social support not being statistically significant in our study, we found that positive exchanges with children offered protection against cognitive impairment. Overall, this supports some previous studies on depressive symptoms [15] and positive well-being in older people [16]. Our results corroborate those from some studies [15] who reported that positive support from children, in particular, was associated with less depression symptoms in older people. In addition, they also found protective, albeit weaker, effects of positive exchanges with spouses and other family and friends. We observed a similar pattern with reasonably large effect sizes although our results did not pass the conventional 5% significance threshold for spouse, other family and friend’s scores.

Possible mechanisms

If indeed causal, the association between positive and negative social support and development of dementia may involve several different processes which, independently or in combination, may induce cognitive impairment in older people. For example, lack of supportive social relationships may promote health-damaging behaviors including smoking, alcohol drinking, sedentary lifestyle, poor dietary choices, and excess body weight [30]. These findings are important in light of the extensive research available documenting the adverse influences of smoking and physical inactivity on both cardiovascular [31] and cognitive outcomes [32] in older adults. Similarly, diverse psychosocial processes, including personality traits or individual differences, emotions and moods, and perceptions of personal control, may enhance people’s coping strategies and adaptation to interpersonal stressors [12]. Importantly, individual’s cognitive appraisal of the significance of an event or circumstances may mediate the impact of stressors on physical and mental health [16]. In turn, the effects of these psychological factors and states may be mediated through diverse psychobiological processes, including neuroendocrine and inflammatory responses[33, 34].

Potential limitations

First, due to being a relatively younger (age 50 years+) cohort and dementia case status being based primarily on self-reported doctor’s diagnosis the incidence of dementia in our study was lower (3.4%) than the incidence reported in other analyses such as [35] which reported an estimate of 6.5% overall in older people. Self-reported diagnoses may also have led to under reporting of milder cases. Second, further loss of cases has occurred due to exposure data (positive or negative social support scores) not being available for some participants leading to reduced sample size for some analyses. This was particularly the case for the analyses involving spouses’ support scores as they are not applicable for people without a spouse or partner. We performed analysis of available data as imputation would not be appropriate given the nature of much of the missing data (e.g., support score from spouse in people not having a spouse, or from children in people not having children). However, despite the possibility of missing cases, we believe this is unlikely to induce major bias in our results given the consistency with other similar studies that used more objective evaluations of dementia [36, 37]. Although we were unable to verify the accuracy of the diagnostic information, self-reports of other conditions, including stroke [38], correspond highly with physician diagnoses, even in the presence of overt impaired cognition.

Third, it was not possible to consider the potential impact of genetic factors such as family history in this study as ELSA does not collect this information. Also, precise information about the type of dementia diagnosed was not available. However, there should not be many cases of familial dementia as relatively few cases of dementia were reported in participants less than 65 years.

Finally, given that negative support from other immediate family was found to increase the risk of dementia, it would be interesting to look at the association of individual relationship within the other immediate family category with dementia. However, it was not possible to measure positive/negative social support scores for individual relationships within the other immediate family from the ELSA study.

Conclusions

In this investigation of older participants in the ELSA study, positive social support from children in later life was associated with a lower risk of developing dementia. Negative social support in a range of relationships, on the other hand, was associated with an increased risk of developing dementia. This work is a step toward better understanding of impact of social relationships on dementia risk; however, further research is needed to better establish any potential causal mechanisms that may drive these associations particularly with reference to age related social factors. Our findings add to the growing evidence on the relevance of social relationships for cognitive health in older age. Specifically, for health and social care practice, they highlight the value of addressing social relationship issues in individuals vulnerable to dementia, while pointing towards specific ways of potentially modifying risk. Similarly, at the policy level, our results will add to the impetus underlying local and national efforts to ameliorate the psychosocial conditions of older people.