The associations of socioeconomic status with incident dementia and Alzheimer’s disease are modified by leucocyte telomere length: a population-based cohort study

Socio-economic status (SES) and biological aging are risk factors for dementia, including Alzheimer’s disease, however, it is less clear if the associations with SES vary sufficiently across different biological age strata. We used data from 331,066 UK Biobank participants aged 38–73 with mean follow-up of 12 years to examine if associations between SES (assessed by educational attainment, employment status and household income) and dementia and Alzheimer’s disease are modified by biological age (assessed by leucocyte telomere length: LTL). Diagnosis of events was ascertained through hospital admissions data. Cox regressions were used to estimate hazard ratios [HRs]. A consistent dose–response relationship was found, with participants in low SES and shorter LTL strata (double-exposed group) reporting 3.28 (95% confidence interval [CI] 2.57–4.20) and 3.44 (95% CI 2.35–5.04) times higher risks of incident dementia and Alzheimer’s disease respectively, compared to those of high SES and longer LTL (least-exposed group). Of interest is a synergistic interaction between SES and LTL to increase risk of dementia (RERI 0.57, 95% CI 0.07–1.06) and Alzheimer’s disease (RERI 0.79, 95% CI 0.02–1.56). Our findings that SES and biological age (LTL) are synergistic risk factors of dementia and Alzheimer’s disease may suggest the need to target interventions among vulnerable sub-groups.

Dementia is a syndrome with multiple pathologies characterized by progressive neurodegeneration leading to cognitive and functional decline. It is estimated to affect 57 million people globally 1 . For the UK, estimated social and healthcare costs are currently around £35 billion, rising to £80.1 billion by 2040 2,3 . It is also estimated that 35% of population dementia risk is attributable to modifiable social and environmental risk factors, including socioeconomic status (SES) 4 .
The role of SES in elevating dementia risk is well documented 5 . Epidemiologic studies have shown that higher education attainment (as a cognitive-stimulating activity) 6,7 and household income (permitting more opportunities for healthy lifestyle) 8,9 are protective for dementia. Conversely, low occupational employment was found to be associated with increased dementia risk 10 . Less clear, however, are the biological mechanisms underlying these associations.
Biological aging is the progressive loss of physiological function relative to chronological age, and captures the potential progressive impairment of metabolic and physiologic capabilities of human body at cellular and molecular levels 11 . Telomere length (TL) has been suggested as an indicator of biological aging. Telomeres are thousands of repeated TTA GGG hexa-nucleotide sequences that cap the ends of chromosomes and are responsible for chromosome stability. Being located at the tip of chromosomes, telomeres are not fully replicated during DNA replication leading to an attrition in length accompanying each successive cell division. Variation in TL Measurement of composite socioeconomic status. Individual-level SES, our primary exposure was derived from a latent class model based on three socioeconomic factors; highest educational attainment, employment status and household income in accordance with a prior study 32 . Latent class analysis (LCA) is a well-validated data-driven technique that employs observed categorical domain variables to identify unobserved latent variable with mutually exclusive latent classes 33 . The three socioeconomic factors were collected from cohort participants based on structured self-reported questionnaires. Highest educational attainment was included as a seven-category variable based on the International Standard Classification of Education in the latent class model and categorized as college or university degree, A levels/AS levels, O levels/GCSEs, CSEs, NVQ or HND or Scientific Reports | (2023) 13:6163 | https://doi.org/10.1038/s41598-023-32974-x www.nature.com/scientificreports/ HNC, other professional qualifications (for example nursing, teaching), and none of the above equivalent to less than high school. Employment status was a three-category variable coded as employed or self-employed, retired, and others (which included looking after home, being unable to work due to sickness or disability, unemployed, doing unpaid or voluntary work or student). Average total household income before tax was divided into five groups (< £18,000, £18,000 to 30,999, £31,000 to 51,999, £52,000 to 100,000, and > £100,000). Three SES classes (low, medium and high) were subsequently identified based on item-response probabilities using gsem command in Stata 34 . Maximum likelihood method, which permits the use of all data available including those with missingness, was employed in the LCA models. Iterations were undertaken before the final model was selected. Model selection was determined by using Bayesian information criteria (BIC) and Akaike information criterion (AIC). The convergence and performance of the latent class models were checked for two, three and four or more classes 32 . In our iterations, models with two and three classes performed the best. The AIC and BIC values for the latent class model with two classes was reported as 3.78 × 10 6 (both statistic), while that for the model with three classes are 3.76 × 10 6 . Based on these fit statistic, we selected the latent class model with three classes to define individual-level SES for this study. The mean posterior probability values of the latent class model with three classes are 0.73, 0.81 and 0.84 for classes 1, 2 and 3 respectively (Supplementary Methods S1 and Table S2).
Measurement of biological age. Leucocyte telomere length (LTL), expressed as the ratio of telomere repeat copy number to the single copy gene (T/S), was extracted from peripheral blood leucocytes of the UK Biobank participants and measured via validated qPCR methods 35 . LTL was adjusted for technical parameters (including enzyme, PCR machine, primer, operator, temperature, humidity, hours from 6 a.m., pipetting robot and extraction method), log e transformed and Z-standardized. We developed a composite measure of biological age which accounted for the potential correlations between chronological age and LTL as well as the combined effects of inflammation and chronic diseases. As in previous study 36 , biological age was defined as residuals of LTL additionally adjusting for chronological age, serum C-reactive protein (from immunoturbidimetric highsensitivity analysis) and any of the 46 curated chronic comorbidities (Supplementary Covariates. Following a socio-ecological perspective, model covariates were identified a priori, informed from published literature 4 . Demographic covariates included age at baseline, sex and ethnicity. A composite healthy lifestyle score was developed based on smoking status, alcohol intake frequency and dietary factors assessed at baseline as in previous literature [38][39][40] . A score of 1 was assigned to participants who achieved the healthy standard in each lifestyle component (non-current smoker; alcohol not consumed daily/almost daily; consumed ≥ 4 of 7 types of food following dietary recommendations as previously). The healthy score ranged between 0 and 3 was subsequently categorized as favorable (a score of 3), intermediate (a score of 2) and unfavorable (a score of 0 or 1) following the data distribution. The number of leisure and social activities per week was coded as none, one or more. Hearing difficulty was coded as self-reported deaf or having difficulty with hearing (coded as 1), and without hearing difficulty (coded as 0). Abdominal obesity was defined in terms of waist-hip ratio (WHR) ≥ 0.85 cm for women and ≥ 0.90 cm for men. Frailty was assessed at baseline from five phenotypes; weight loss, exhaustion, low grip strength, low physical activity and slow gait speed and coded as non-frail, frail/pre-frail 41,42 . Among the environmental variables, urbanicity was a composite measure of densities of housing, retail and public transport, and urban centrality around the participants' residence (coded as quintiles 1, 2-4 and 5) as previously 43,44 . Neighbourhood deprivation (coded as quintiles 1, 2-4 and 5) was measured by the Townsend deprivation index (Supplementary Methods S2 for additional details).

Statistical analyses.
Descriptive characteristics across the three SES categories are presented as mean and standard deviation for continuous variables, and frequency and percentage for categorical variables. We employed Nelson-Aalen hazard method to calculate the cumulative hazard for incident dementia based on SES categories (3 groups), and subgroups by SES and LTL strata (9 groups) over time.
Cox proportional hazard regression models estimating hazard ratios (HR) and 95% confidence interval (CI) for associations of SES and TL with risks of incident dementia and Alzheimer's disease were developed. Persontime in months were calculated from the date of entry into the cohort until the date of hospitalization or date of death, or the study end date, whichever occurred earlier. The proportional hazards assumption was tested by using the Schoenfeld residual test. We excluded events prior to baseline assessment. To minimize chance of reverse causality, we performed landmark analyses by excluding events within 3 years of entry in to the study.
We examined independent associations of SES and LTL with incident dementia and Alzheimer's disease diagnosis. HRs in the medium and low SES categories were reported in reference to the high SES category. Linear test was conducted to estimate the associational trend across the three exposure categories. To examine if biological age moderated the associations, we conducted stratified analyses by LTL. LTL was coded as a three-level categorical variable based on cut-offs at the 25th and 75th percentiles, with participants at the top 25th percentile with longer LTL taken as the reference group. Lastly, we examined joint associations between SES and LTL on incident events by stratifying participants into 9 groups by SES (low, medium and high) and LTL (long medium, short), with participants in the high SES and long LTL strata acting as the reference group. Biologic interaction on an additive scale was expressed as relative excess risk due to interaction (RERI) with the corresponding 95% CIs obtained by estimating the separate and joint effects of SES and LTL on incident dementia and Alzheimer's disease. In the RERI calculations, we defined participants with low SES and short LTL as the double-exposed group, while those with high SES and long LTL strata as the reference category 45 www.nature.com/scientificreports/ sex, ethnicity, healthy lifestyle score, social activities, hearing difficulty, clinical phenotypes (abdominal obesity and frailty index), urbanicity, neighbourhood deprivation and SES/LTL. To fulfill the proportionate hazards assumption, sex was inputted using the strata option throughout the analysis. We performed several sensitivity tests. First, we repeated our models with single socioeconomic factors (highest education attainment, employment status and household income) to examine their independent associations with incident dementia and Alzheimer's disease. Second, we repeated our primary analyses using combined cases of both incident dementia diagnosis and mortality. Data on dementia mortality was extracted from death register, with the date of death obtained from death certificates held by the National Health Service (NHS) Information Centre (England and Wales) and the NHS Central Register in Scotland. Third, we reran the LCA model comprising the three individual socioeconomic factors as well as the Townsend deprivation index to additionally take into account the effect of neighbourhood deprivation within the composite SES measure. Fourth, we further reran analyses with adjustment for number of household members (coded as living alone, two, three, and more than three) to take into account variability in household size. Fifth, we reran models after additionally adjusting for comorbidities in the main models as covariates to examine robustness of our findings (with comorbidities being removed from the residual analysis for LTL). Sixth, for the employed subpopulation, we additionally reran models examining associations between occupation categories and incident dementia and Alzheimer's disease. Occupation categories were coded as a five-factor variable comprising other occupations, managers and senior officials, professional occupations, associate professional and technical occupations, and administrative and secretariat occupations. Other occupations included manual occupations such as skilled trades occupations/ personal service occupations/sales and customer service occupations/process, plant and machine operatives/ elementary occupations, and was taken as a reference category. Seventh, we examined potential modifying effect of age by conducting stratified analyses by age groups (< 65 versus ≥ 65 years). Lastly, we performed multiple imputation by chained equations with 20 imputation sets 47 to account for missing data across covariates on a sample of 422,907 participants with complete data on outcome and exposure variables. All covariates, exposure and outcome variables in the fully-adjusted model and the follow-up time were used to impute missingness.
Data were analyzed from November 2021 to August 2022. All analyses were performed in Stata software version 16.1.

Results
From the UK Biobank population sample of 502,650 individuals, participants were omitted from the analysis due to withdrawal (n = 191), diagnosed with dementia during the first three years of follow-up (the landmark period) or at baseline (n = 377), and missingness on one or more SES exposure (n = 79,175) or covariates (n = 91,841). The remaining 331,066 participants were used in a complete sample analysis (Supplementary Figure S1). There was no systematic difference between analytic sample and excluded sample (Supplementary Table S5). Over 3,989,206 person-years of follow-up (mean follow-up period = 12 years; IQR: 11.5-13.0 years), 3,313 incident cases of dementia were recorded, of which 1,423 were of Alzheimer's disease.
Mean age of the sample at baseline was 56 (SD = 8), with 52% women ( Table 1). The LCA model for composite SES categorized participants into three SES groups for high (n = 111,525, 34%), medium (n = 138,373, 42%) and low (n = 81,168, 25%) socioeconomic status. In general, participants with dementia belonged to the older age sub-group and with educational qualification less than high school, lower household income, obese and with lower grip strength, gait speed, physical activity levels, shorter leucocyte telomere length (LTL) and more comorbidities. The mean LTL of analytic sample was 0.01 (SD = 1.00, range = − 15.28-12.14). As expected older participants belonging to low SES strata had shorter LTL. The mean LTL among participants aged < 65 years and of low, medium and high SES strata were − 0.10 (SD = 1.00), 0.04 (SD = 0.99) and 0.15 (SD = 0.99) respectively. While, the mean LTL in sub-group aged ≥ 65 years and belonging to low, medium and high SES strata were − 0.30 (SD = 0.98), − 0.27 (SD = 0.98) and − 0.25 (SD = 0.97) respectively.

Cumulative hazards of incident dementia and Alzheimer's disease across SES and LTL strata.
Nelson-Aalen hazard estimates showed consistently higher cumulative risks of incident dementia for lower SES strata (Fig. 1a). The cumulative risks for incident dementia at 12 years were 2.1% (95% CI 1.99-2.20) for low SES, 0.88% (95% CI 0.83-0.94) for medium and 0.30% (95% CI 0.27-0.33) for the high SES stratum. Further stratification by SES and LTL revealed a cumulative risk for incident dementia of 2.33% (95% CI 2.12-2.56) in participants of low SES and short LTL, while it was only 1.67% (95% CI 1.49-1.87) for the low SES and long LTL category (Fig. 1b). A similar pattern was observed for Alzheimer's disease with the cumulative risks at 12 years being 0.94% (95% CI 0.87-1.01) for low SES, 0.37% (95% CI 0.34-0.40) for medium and 0.12% (95% CI 0.10-0.14) for the high SES stratum.  (Fig. 2). Compared with participants of high SES and longer LTL (least-exposed group), those of low SES and shorter LTL (double-exposed group) had higher HRs of 3.28 (95% CI 2.57-4.20) for incident dementia and 3.44 (95% CI 2.35-5.04) for incident Alzheimer's disease. We found evidence of positive biologic interaction between LTL and SES, with the combined effect of the two factors in the double-exposed group exceeding the sum of their separate effects relative to the least-exposed group (Supplementary Table S7). The relative excess risk due to interaction (RERI) for both dementia (RERI = 0.57, 95% CI 0.07-1.06), and Alzheimer's disease (RERI = 0.79, 95% CI 0.02-1.56) were positive, indicating a departure from additivity towards excess risk in double-exposed groups (i.e. synergism).

Sensitivity analyses.
Repeating analyses with single socioeconomic factors of educational attainment, household income and employment status, showed elevated risks for both incident dementia and Alzheimer's disease among participants of lower education attainment, lower household income and those who were retired and belonging to the others category (including unemployed) (Supplementary Tables S8-10). Consistent with our primary analysis, models combining cases of incident dementia diagnosis and mortality as the outcome showed that participants of low SES and shorter LTL (double-exposed group) had higher HRs of 3.21 (95% CI 2.53-4.08) in reference to those of high SES and longer LTL (least-exposed group) with synergistic interaction (RERI = 0.62, 95% CI 0. 16 Table S12). Our results also remained robust after additional adjustment for number of household members (Supplementary Table S13) and comorbidities taken as covariates (Supplementary Tables S14) with increased risk of incident dementia and Alzheimer's disease being associated with lower SES and shorter LTL. Rerunning analyses for occupational categories showed lower risk of incident dementia and Alzheimer's disease for participants engaged in professional occupations, and administrative and secretariat occupations relative to those who belonged to other occupational (manual) category, nonetheless the results remained insignificant (Supplementary Table S15).   26-9.42) for Alzheimer's disease. The effect sizes for age-group ≥ 65 years were more pronounced in the double-exposed category relative to least-exposed, with HRs of 6.37 (95% CI 4.84-8.38) for dementia and 7.42 (95% CI 4.82-11.44) for Alzheimer's disease (Supplementary Tables S16, 17). Sensitivity analyses with imputed data accounting for missingness across covariates produced consistently robust results further supporting the main analyses (Supplementary Tables S18).

Discussion
In a large prospective study, we found SES (a composite metric developed from household income, educational attainment and employment status) and leucocyte telomere length (LTL), were each independently monotonically associated with higher risks for incident dementia and Alzheimer's disease. We also found that SES and LTL demonstrated modest interaction effects, suggesting that biological aging as measured by LTL and SES are likely to represent separate but synergistic risk pathways. Several sensitivity analyses such as rerunning models with Alzheimer's disease. Models included age, sex, ethnicity, healthy lifestyle score, social activities, hearing difficulty, abdominal obesity, frailty index, urbanicity and neighbourhood deprivation. Residual of LTL, adjusted for chronological age, C-reactive protein and co-morbidities was defined as biological age in the study. The vertical bars indicate the hazard ratio, while the whiskers indicate the corresponding 95% confidence interval (CI). The asterisks represent statistically significant (two-sided p < 0.05) point estimates. Relative excess risk due to interaction (RERI) was used to examine additive interaction between SES (low SES versus high SES) and LTL (short LTL versus long LTL). See Supplementary Table S7 9 . Another 12-year follow-up study of 2,457 US elderly found increased hazard for incident dementia among participants with less than high school education (HR 1.47, 95% CI: 1.17-1.86) 6 . A Swedish cohort study of 931 participants aged ≥ 75 reported that low education and occupation-based SES were associated with elevated risks of dementia (RR = 2.4, 95% CI 1.5-4.0) and Alzheimer's disease (RR = 3.1, 95% CI 1.6-5.7) 48 . For LTL, the Washington Heights study of 1,983 participants followed for 8 years reported that each kilobase pair decrease in LTL was associated with 21% increased risk for dementia 49 . However, a study of 1,961 participants followed for 8 years in Rotterdam found a U-shaped relationship between LTL and Alzheimer's risk with HR 1.59 (95% CI 1.13-2.23) for the first tertile in reference to the middle. Of particular interest are cross-sectional findings from the HANDLS study. Among 325 individuals, interactions between LTL and SES for cognitive performance were found (b = − 2.06; p = 0.008) suggesting that for low SES shorter LTL is associated with poorer cognitive performance 50 . Our study extends this evidence by providing longitudinal data, at-scale, for clinically relevant outcomes.
Potential mechanisms for these associations are complex and remain an active area of research. Lower SES is associated with poorer cognitive reserve, higher levels of psychosocial stress, poorer lifestyle and behaviour that increases cognitive risk 51 . It has also been suggested that low SES may act as a proxy for reduced exposure to intellectually stimulating environments with inadequate cognitive resources to protect against neurodegeneration 9,47,52 . Biological age, as measured by telomere length is considered a composite marker of current health status encapsulating cumulative burdens of chronic comorbidities and psychosocial stress from allostatic load 13,53 . Shorter TL is associated with age-related morbidity, lower immune function, elevated oxidative stress-dependent senescence, and pro-inflammatory mediators 54 . Given the potential overlap between these two constellations of pathways, our study suggests SES and LTL were synergistic for dementia. These findings may suggest that both SES and biological aging potentially contribute to 'downstream' mechanisms underlying increased dementia risk. For example, inflammatory processes are increasingly implicated as systemic influences in neurodegeneration, and both biological ageing and SES may act separately to upregulate neuro-inflammatory pathways.
Previous evidence mostly showed significant associations between individual-level SES and dementia in elderly cohorts comprising late-onset cases. However, research on early-onset dementia has been relatively scarce. In our UK Biobank analytic sample, 233 of 3,313 (7%) dementia cases presented themselves as early-onset cases. Our stratified analysis by age in the study sub-groups found that the effect sizes for socioeconomic status on the risk of dementia and Alzheimer's disease in the younger age subgroup (< 65 years) comprising early-onset cases remained significant, as were the older sub-groups (≥ 65) with late-onset cases. Joint associational analysis found that the HR for dementia in the double-exposed group (relative to the least-exposed) were 4.57 for the subgroup with early-onset cases (< 65 years) and 6.37 for subgroup with late-onset cases (≥ 65 years). In earlyonset dementia (including frontotemporal dementia), symptom onset occurs at a relatively younger age when the patients are still in a crucial life stage, being inactive employment with family and social responsibilities, thereby making them vulnerable. Patients generally experience a steeper decline in cognitive and functional capacities as compared to late-onset dementia 55,56 . It is noteworthy to mention that the observed significant associations between early-onset dementia and SES are of value, and may plausibly point to the protective role of socioeconomic position (and associated lifestyle) on dementia, potentially via lowering stress and chronic inflammation. Further research needs to be conducted in this direction.
In our models, biological age was defined as a composite measure derived from residual LTL adjusting for chronological age, inflammation (CRP) and 46 curated chronic comorbidities including risk factors of dementia. Sensitivity tests in which models included residual LTL adjusted for age and serum CRP only, and additionally adjusted for the comorbidities as covariates also produced consistent results. The generalizability of our findings to other cohorts in diverse settings also needs to be explored.
With increasing proportion of elderly population and stronger impetus towards ensuring healthy longevity, evidence as generated from population stratification by SES and biological age may help identify vulnerable subgroups for allocating crucial public health and social care resources. Risk factors such as SES may act as important considerations for community screening of neurodegenerative diseases as well as designing of tailored intervention programs to prevent against cognitive impairment. Further research is needed to improve technologies for measuring biological age (telomere length) at-scale, accuracy and cost effectiveness 57,58 . Strengths. Strengths of the study include scale and follow-up period. This analysis is an order of magnitude larger than previous studies linking either SES or LTL to dementia, enabling smaller effect sizes and interactions to be detected with confidence. The 12-year follow-up period, also longer than many previous studies, has enabled the analysis to focus on incident cases and implement a landmark period, to minimise the impact of reverse causation. Latent class modelling is a parsimonious approach to developing a composite SES construct. Complimentary sensitivity analyses using individual base SES variables further confirmed the primary analysis. Biological age was comprehensively measured from LTL which underwent rigorous quality checks.
Limitations. We acknowledge several limitations. Firstly, ascertainment of dementia is a challenge for health services across the world. We defined dementia and Alzheimer's disease through electronic health record linkage as in previous studies 9,59,60 , which would have potentially led to an underestimation through omitting undi- www.nature.com/scientificreports/ agnosed conditions. Previous cohort studies have estimated the low sensitivities of dementia diagnosis from hospital discharge records; for example 51% in Finland 61 , 78% in England 62 and 70% in USA 63 . A recent systematic review concluded that the positive predictive value (PPV) for all-cause dementia from routinely collected health datasets were > 75% for 16 of the 27 studies, while the sensitivities were reasonable ranging between 21 and 86% 64 . Such an under-ascertainment would be more likely among ethnic minority, older sub-groups and in participants with milder dementia and poorer health-seeking behaviour. Nonetheless, routinely collected and coded hospital records linked to large population cohorts constitute a cost-effective means of case ascertainment for dementia with minimum attrition rate in prospective analysis relative to standardized clinical assessment needing face to face examination. Within the UK Biobank, PPV for all-cause dementia from hospital admissions records remained high at 87.3% 65 . Secondly, the UK Biobank population sample is not representative. Compared to the general population, cohort participants were less likely to be sampled from socioeconomically deprived areas and had lower prevalence of obesity, smoking, drinking and self-reported health conditions due to healthy volunteer selection bias, and as such, summary statistics are not generalizable 66 . However, being a large and heterogeneous cohort, it is designed specifically for etiologic analyses and sampling bias is unlikely to have a material effect on the estimates reported 67 . Thirdly, as in any observational study, we cannot completely rule out residual confounding and reverse causation and causality cannot be assumed. However, our results remained robust subsequent to exclusion of cases over the 3-year landmark period and adjustments for chronic morbidities. Fourthly, our primary SES exposure was based on self-reported measures of household income, educational attainment and employment status. Lastly, we excluded participants due to missingness across covariates. Nonetheless, there was no systematic difference between analytic sample and excluded sample (Supplementary Table S5) and our primary results were consistent with those of sensitivity analyses using imputed data.
In conclusion, in a large prospective study we find that SES and biological aging are synergistic risk factors for dementia and Alzheimer's disease suggesting the importance of targeted tailor-made preventive interventions in vulnerable at-risk population groups.

Data availability
Data are available from the UK Biobank (https:// www. ukbio bank. ac. uk/) for researchers who meet the criteria for access to de-identified UK Biobank data.