All participants were from the online Platform for Research Online to investigate Genetics and Cognition in Ageing study (PROTECT; http://www.protectstudy.org.uk/). PROTECT is a longitudinal UK-based online participant registry that aims to understand the impact of lifestyle, medical and genetic risk factors on cognitive health and dementia risk in older adults. Inclusion criteria for enrolling in PROTECT are 1) ≥50 years old; 2) no diagnosis of dementia; and 3) access to a computer and the internet. Volunteers were prospectively recruited from November 2015 through both local and national publicity. Data collection was ongoing at the time of this study; therefore, a data freeze was implemented in October 2019 with data extracted for analyses up to this date. In the current study (N = 7,877), we used data from a subset of PROTECT study participants who 1) had complete baseline data for the cognitive and covariate variables (missing cognitive data at later time points was permitted), 2) had provided a saliva sample for genotyping and had genotype data available, and 3) were identified as individuals of European ancestry based on genetic principal components (PCs). These participants were generally comparable to the overall PROTECT study cohort in terms of participant characteristics, such as age, sex and education level (Supplementary Table S1). Participant characteristics of our full sample (N = 7,877) and the subsample with relative telomere length (RTL) data (n = 846) are shown in Supplementary Table S2. Participant characteristics split by median age for the full sample with polygenic scores for telomere length (PGS-TL) data and the subsample with RTL data are summarised in Supplementary Tables S3, S4, respectively.

Ethical approval was granted through the London Bridge National Research Ethics Committee (reference: 13/LO/1578) and informed consent obtained for all participants. The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Declaration of Helsinki 1975, as revised in 2008 (World Medical Association, 2013).

Cognitive performance was assessed annually using an online cognitive test battery. We analysed participants’ data on four cognitive tests included in the PROTECT battery. A description of the four tests is provided in Table 1. Briefly, the tests included the paired associate learning (PAL) task of visual-spatial working memory and learning; digit span test (DST) of working memory; self-ordered search (SOS) of executive function and spatial memory; and verbal reasoning (VERB) task. At each annual testing, participants were instructed to complete the test battery up to three times over a period of 7 days, leaving at least 24 h between each cognitive testing session (only the first test session was mandatory). The tasks have parallel forms to ensure that repeated stimuli are not given to participants at each test session. The first cognitive test session at baseline was considered a practice session. In our analyses, we used data from the second testing session at baseline, where available, otherwise, we used data from the first testing session to maximise sample size and statistical power. For all subsequent time points, data from the first test session were used (see, Figure 1).

Following enrolment into the study, participants completed a series of online baseline questionnaires regarding demographic, lifestyle, and medical information. For the current analysis, the following demographic data were included as covariates: age, sex, education level (school until 16, school until 18, vocational qualification, undergraduate degree, post-graduate degree, and doctoral degree) and employment status (full-time, part-time, self-employed, retired, and unemployed) (dummy coded). Additionally, we included six genetic PCs, defined by the PROTECT study, as this number was determined to sufficiently adjust for population structure using scatter plots of the PCs incrementally plotted against each other until there was no clear pattern in the data (Patterson et al., 2006; Price et al., 2006). The selection of covariates was informed by previous genetic research in the PROTECT sample (i.e., Creese et al., 2021).

Saliva samples were collected by post and DNA was extracted by the National Institute for Health Research South London and the Maudsley National Health Service Biomedical Research Centre. Genotyping was completed using the Illumina Global Screening Array with custom content.

The total number of participants in the combined genotyped data for the whole PROTECT study was 9,146. Genotype quality control (QC) was performed on all 9,146 individuals, as described in Creese et al. (2021). The QC involved iterative filtering for call rate at 98% completeness (for individuals and SNPs) and then removing participants that were either related, not of European ancestry, of mismatched sex, outliers in the PC calculation or detected to have excess heterozygosity. This resulted in the exclusion of 84 samples due to incompleteness and the removal of a further 794 individuals following exclusions. Thus, a sample size of 8,268 participants remained.

Genotypes were imputed to 1,000 Genomes European reference panel using the Michigan imputation server and genotype phasing using Eagle (Loh et al., 2016). Variants were restricted to SNPs only, with a MAF > 0.001. An absolute cut-off of 0.7 was applied to the imputation quality of variants (Rsq as reported by the Michigan imputation server). The number of variants remaining after quality control was 9,415,055.

Further genotype QC was performed within the sample of 8,268 individuals of European ancestry, following all exclusions. The following SNP exclusions were applied, minor allele frequency (MAF) of <1% and those not in Hardy-Weinberg Equilibrium (p-value < .00001).

We computed polygenic scores for telomere length (PGS-TL), using genome-wide summary statistics (clumped using 250 kb windows and r² > 0.1; MHC region included) from the largest GWAS of leukocyte telomere length to date (N = 464, 716 participants aged 40–69 years from the United Kingdom Biobank; Codd et al., 2021). We applied a p-value threshold for PGS-TL calculation (P_T = 3.9 × 10⁻⁵) equivalent to an FDR threshold of <1%, highlighted by Codd et al. (2021) to explain up to 5.64% of the variance in telomere length. The PGS-TL was calculated with PRSice-2 version 2.3.3 software (Choi and O’Reilly, 2019).

We generated relative telomere length (RTL) data for a subset of 846 participants. Relative telomere length was quantified using DNA samples and a modified version of the quantitative Polymerase Chain Reaction protocol described by Cawthon (Cawthon, 2009), as used by our lab previously (Coutts et al., 2019; Palmos et al., 2018; Palmos et al., 2020; Powell et al., 2018; Vincent et al., 2017); see Supplementary Methods S1 for further details.

Initially, we selected 908 participants from the overall study sample (N = 7,877) who met the following additional criteria: 1) complete cognitive data across time points (i.e., no missing cognitive data) 2) DNA was available from a saliva sample collected within 6 months of their baseline cognitive data collection. 54 samples (6%) did not survive our quality control (QC) criteria. Valid RTL measurements were therefore obtained for 854 participants.

We log-transformed the RTL data, as it was not normally distributed. We further adjusted for the day of run as this yielded significant batch effects (see Supplementary Table S5) and Z-standardised the data. Outliers were defined as residuals +/− 3.29 SDs from the mean, after fitting a linear regression model with age, sex, education level, employment status (dummy coded) and six genetic PCs (as described in Covariates). Based on this criteria, 8 samples were identified as outliers and were removed, leaving 846 valid telomere measurements for downstream analyses.

All statistical analyses and data cleaning were performed using RStudio version 1.3.1093 and R version 4.0.3. The PGS-TL and RTL data were Z-standardized (M = 0, SD = 1) before analysis. To examine relationships with RTL, we performed a multiple linear regression with RTL regressed on PGS-TL, age, sex, education level, employment status (dummy coded), and the six genetic PCs (see Covariates).

To examine cognitive performance over time, we used the R package lavaan (Rosseel, 2012) to construct latent growth models (informed by Berlin et al., 2014), with a Yuan-Bentler correction for non-normality (“estimator” = “mlr”), and full information maximum likelihood in models with missing data (“missing” = “fiml”), under the assumption that this missingness was at random (Arbuckle, 1996). As a sensitivity analysis, we repeated the latent growth modelling with complete cases (i.e., participants with no missing data across time points) for models that had included participants with missing data (see Sensitivity analyses). In the latent growth models, latent factors for the intercept (initial performance) and slope (change in performance over time) were estimated. We first fitted unconditional latent growth models (i.e., models in which the intercept and slope factors are not predicted by (conditioned on) predictors or covariates) for each of the four cognitive outcomes to determine the pattern of growth in the data. We tested no-growth (intercept-only), linear slope and quadratic slope models, using model fit statistics to determine which best described the data.

Next, conditional latent growth models were used to explore associations between measured (RTL) and genetically predicted telomere length (PGS-TL) and initial cognitive level (i.e., latent intercept) and trajectories over time (i.e., latent slope factors). Separate models were used for measured TL and PGS-TL and each cognitive outcome. The first six genetic PCs to adjust for population structure (Patterson et al., 2006; Price et al., 2006), age, sex, education level (school until 16, school until 18, vocational qualification, undergraduate degree, post-graduate degree, and doctoral degree), and employment status at baseline were included as time-invariant covariates.

For all latent growth models, the goodness-of-fit of the models was evaluated using several model fit statistics. Excellent models generally have the following values: CFI ≥.95, RMSEA < .05, and SRMR < .05 (Hu and Bentler, 1999). The chi-square statistic (χ²) is reported with degrees of freedom (df) but is not used as a measure of fitness, given its oversensitivity to large sample sizes.

Bonferroni correction for multiple testing was applied (i.e., p < .013; α = 0.05/4, to reflect the number of cognitive outcome variables) when assessing the relationships between telomere length (PGS-TL and RTL) and intercept and slope factor(s).

Firstly, to assess whether PGS-TL and RTL differentially predict cognition in middle and older age adults we repeated our analyses, using a median split based on age. A median split was used so that the sample size and power of each group were approximately equal. For more information on these analyses and how they were performed see Supplementary Methods S2.

Secondly, we repeated the latent growth modelling using only complete cases—participants with no missing data across time points—for models with PGS-TL included as a predictor. This step was taken to assess whether attrition and the inclusion of participants with missing data in the original models might have biased the parameter estimates, as these models were initially computed using full information maximum likelihood to handle missing data. For models using RTL as a predictor, which were already conducted with complete cases only, no additional sensitivity analyses were performed.

Table 2 summarises the multiple regression results with RTL regressed on PGS-TL, age, sex, education level, employment status (dummy coded), and the six genetic PCs. Significant predictors of RTL included the PGS-TL (β = .07, p = .032), retirement status (β = .37, p = 001) and two of the genetic PCs (βs = .09 and .07, ps < .046). Age (p = .051), sex (p = .131) and education (p = .914) did not significantly predict RTL.

First, we fitted unconditional models (i.e., without the intercept and slope factors conditioned on predictors or covariates) for each of the four cognitive outcomes to determine the pattern of growth in the data. Across the full sample and all subsamples (i.e., split by median age and/or those with RTL data), unconditional model fit statistics suggested that linear models provided an acceptable fit to the data for the paired associat learning (PAL), digit span test (DST), and self-ordered search (SOS) tasks, whereas for the verbal reasoning (VERB) task, a quadratic model provided acceptable fit (see, Supplementary Tables S11–S22 for the fit indices and parameter estimates for the unconditional models).

Next, we fitted conditional models to explore whether PGS-TL or RTL predicted baseline cognitive level (i.e., latent intercept) and/or change over time (i.e., latent slope factors). Although these models were adjusted for covariates, we do not report their corresponding parameter estimates in the main text for simplicity. Parameter estimates with 95% confidence intervals are reported in full, including those related to covariates, for each model within the Supplementary Tables S23–S28.

PGS-TL did not predict baseline cognitive performance (captured by the intercept factor; ps ≥ .238) or change over time (captured by the slope factor(s); ps ≥ .080) on any of the cognitive tests, after adjusting for covariates. Conditional model fit statistics for the full sample can be found in Supplementary Table S29 and parameter estimates with confidence intervals are summarised in Supplementary Table S23.

RTL did not predict baseline cognitive performance (ps ≥ .154) or change over time (ps ≥ .276) on any of the cognitive tests, after adjusting for covariates. Conditional model fit statistics for the RTL subsample can be found in Supplementary Table S29 and parameter estimates with confidence intervals can be found in Supplementary Table S24.

To assess whether PGS-TL and RTL differentially predict cognition in middle and older age adults we repeated our analyses, using a median split based on age. In the full sample with PGS-TL data (N = 7,877) the median age was 62.24 years, in the subsample with RTL data (n = 846) it was 62.77 years.

We repeated the latent growth modelling for models that had included participants with missing data (i.e., those with PGS-TL as a predictor), using only complete cases only—participants with no missing data at any time point. These analyses yielded similar results to those that included participants with missing data, see Supplementary Tables S31–S33 for parameter estimates and Supplementary Table S34 for model fit statistics). The direction of effect and the significance of PGS-TL as a predictor remained consistent for all outcomes, both in the full sample and within each age subgroup divided by median age. All models using RTL as a predictor were already performed in complete cases only, and so no sensitivity analyses were completed for these models.

This study had several limitations. First, our measures of telomere length were derived from different tissue sources. Saliva is easier to collect but is arguably a less valid tissue source than blood, which has a clearer link to brain ageing (Pluvinage and Wyss-Coray, 2020). However, telomere length is generally positively correlated across different tissue types (Demanelis et al., 2020), and genetic factors for leukocyte telomere length (i.e., PGS-TL) predicted salivary telomere length in our study. Second, the PGS-TL may have been underpowered. While telomere length is highly heritable (i.e., ∼70% according to meta-analytic estimates, Broer et al. (2013); with SNP heritability estimates of 8.1%; Codd et al., 2021), the PGS-TL only explains up to ∼5.64% of the variance in leukocyte telomere length (Codd et al., 2021). Therefore, larger sample sizes and more powerful GWAS may be required to improve the predictive power of the PGS-TL and to address this “missing heritability”. Third, although we adjusted our analyses for several covariates, we cannot exclude the possibility that unmeasured confounding may have biased our results. For example, factors such as physical activity, which may influence both telomere length and cognitive function, could have biased our results. Fourth, limited variability in scores on the cognitive tests (potentially reflective of our homogenous sample) and relatively short follow-up period may have hindered our power to detect significant effects related to age-related cognitive decline. Additionally, associations between telomere length and cognitive performance may exist in cognitive domains not assessed by the four tests included in this study (e.g., reaction time). Finally, as in all research, we urge caution in extrapolating our findings beyond the current sample. For example, the study consisted of white UK adults only, and women and individuals of a higher education level were overrepresented in our study, relative to the general population (see, Office for National Statistics, http://www.ons.gov.uk). Research in more representative and diverse samples is greatly needed, especially given that telomere length may be differentially associated with cognitive function across sociodemographic groups (see, Leibel et al., 2019).