Abstract
Over the past decade, ageing clocks have become widely adopted as important tools for understanding biological ageing and have been redefining notions of “pro-longevity” lifestyles. However, this work is still at an early stage. Some leisure activities, such as arts and cultural engagement (ACEng) have never been studied at all, while others such as physical activity (PA) have only received scant attention. In particular, there is little understanding of whether frequency of engagement or diversity (which provides access to more active ingredients) is more important. This study used 3,354 adults in the UK Household Longitudinal Study - a large, nationally-representative cohort study, which includes six derived epigenetic clocks. We used a doubly robust estimation using the inverse-probability-weighted regression adjustment estimator adjusted for demographic, socioeconomic, behavioural and health covariates, data collection gaps and technical covariates of epigenetic clocks. ACEng and PA were related to slower biological ageing in the PhenoAge and DunedinPoAm clocks, although not to the other measured clocks, with comparable effect sizes between ACEng and PA (Lin, Horvath2018, Horvath2013 and Hannum). For ACEng, diversity and frequency of engagement were related to DunedinPoAm, while for PhenoAge, there was a slightly clearer relationship for frequency than diversity. For PA, higher levels of frequency, diversity, and activeness were related to DunedinPoAm, while only the highest diversity and activeness were related to PhenoAge. These findings were all stronger amongst middle-aged and older adults. This was the first study to show a relationship between ACEng and epigenetic ageing, with comparable effect sizes to PA, suggesting the value of its exploration alongside other lifestyle factors. Decelerations in ageing clocks, including those within our study, have been demonstrated to have clinical (as well as statistical) importance, supporting future exploration as to whether lifestyle changes have any value to slowing or potentially reversing epigenetic ageing.
Introduction
With ageing populations becoming a global phenomenon, how to support not just a longer lifespan but also an increased “healthspan” is becoming a key question for both individuals and health services.1 Helping adults stay free from disease, maintain functional independence, and reduce the need for health services are key priorities for governments internationally.2 Over the past two decades, theoretical and technological advances within molecular biology have identified a series of fundamental biological hallmarks of ageing, including various molecular, cellular, and systemic processes underpinning health and disease.3 One of these is epigenetic alterations, including alterations in DNA methylation (DNAm) patterns, aberrant chromatin remodelling, abnormal post-translational modification of histones, and deregulated function of non-coding RNAs. Environmental stress accumulated over the lifetime disrupts epigenetic profiles, and this contributes to the ageing phenotype by promoting instability, carcinogenesis and cardiovascular pathologies.4,5 In recent years, there have been major developments in biohorology: the use of ageing “clocks” that are built from mapping patterns of DNAm across sparse but informative sets of specific CpG (cytosine-phospho-guanine) sites on the genome.6 These epigenetic clocks are used to identify discrepancies between chronological vs biological age (i.e. accelerated vs decelerated ageing). While first-generation clocks were based on chronological age (e.g. Horvath & Hannum), second-generation clocks have been developed based on phenotypic age (e.g. PhenoAge), and lifespan (e.g. GrimAge) and third-generation clocks are based on the pace of ageing (e.g. DunedinPACE).
Ageing clocks are not without controversy: there is no gold standard for measuring epigenetic biological ageing. Ageing-related biological changes may be correlates rather than causes of ageing, the relationship between clocks and disease pathology is still in its infancy, and there is a current proliferation of clocks.7 Nonetheless, over the past decade, ageing clocks have become widely adopted as important tools for understanding biological ageing and have been redefining notions of “pro-longevity” lifestyles.8 So, exploring ageing clocks alongside other biological approaches could provide important complementary insight into the molecular underpinnings of health. As part of this, there is increasing interest in finding modifiable lifestyle factors that might have “anti-ageing” effects. Avoiding smoking and binge drinking, maintaining a healthy body weight, adopting a Mediterranean diet, reducing stress, and engaging in meditation have all been demonstrated through combinations of experimental and epidemiological studies to reduce epigenetic age.8
However, this work is still at an early stage. Some leisure activities have never been studied at all. Arts and cultural engagement (ACEng) is increasingly recognised as a health behaviour in its own right, comprising diverse ‘active ingredients’ that are beneficial to health (e.g. social interaction, cognitive stimulation, multi- sensory stimulation, creativity, etc.) and activating complex psychological, biological, social and behavioural mechanisms of action that relate to mental and physical health outcomes.9,10 Experimental studies have already demonstrated that ACEng can affect gene regulation. For example, compared to relaxing, listening to music upregulates genes involved in processes such as dopamine secretion, enhanced synaptic function, and neurogenesis, alongside upregulating specific RNA proteins (microRNAs) that repress inflammatory cytokines and support neuronal and synaptic plasticity.11–14 Music has also been demonstrated to be more effective than other activities like reading newspapers for reversing stress signatures in gene expression following laboratory-induced stressors.15 However, there are no epidemiological studies of ACEng and epigenetic ageing to date.
Even more prominent health-promoting leisure activities, like exercise, have only received scant attention. Experimentally, physical activity (PA) has been demonstrated to cause DNAm changes. For example, people with a lifelong history of PA display lower DNAm levels on gene promoters in muscle tissue,16 and interventions that increase PA reduce epigenetic mutation load (i.e. the total number of stochastic epigenetic mutations or outlier methylation patterns at CpG sites), which has been proposed as a complementary DNAm-based biomarker of healthy ageing.17 Observationally, hardly any studies have been conducted looking at physical activity and combined CpG sites within epigenetic clocks. Some very small studies (n<250) have reported null findings when relating PA to individual epigenetic clocks.18,17 Other studies have reported associations between PA and aspects of physical performance (e.g. grip strength and jumping height) and several epigenetic clocks, including PhenoAge, FitAge and GrimAge. 19,20 However, these studies have failed to consider important confounders like socioeconomic position (SEP), smoking, BMI or blood cell compositions. Recently, analyses of larger cohort studies have shown more promising results. An analysis of adults in the Rhineland Study (n=3,567) found that accelerometer-derived step count and both volume and intensity of physical activity were related to lower GrimAge and PhenoAge acceleration, but not Hannum or Horvath2013.21. And the US Sister Study (n=2,758) found that after adjusting for covariates, hours per week of leisure-time physical activity was only related to GrimAge but had no associations with Hannum, Horvath2013, or PhenoAge.22
However, some key challenges remain with the existing literature. First, studies have only focused on a limited number of ageing clocks. Each ageing clock defines biological age in its own way using a distinct set of CpG sites, different primary domains (tissue, health conditions, age range), different algorithms (target definition, machine learning model), and different target variables (chronological age, phenotypic age, time-to-death, pace of ageing,etc.). Consequently, the associations between leisure activity and biological age understandably vary between clocks, meaning outcome-wide approaches using multiple ageing clocks are important for drawing broader conclusions.7 Second, how leisure is conceptualised and defined is underexplored. While frequency is a standard metric of people’s engagement, a variety of engagement may also be important. Variety provides people with greater opportunity to access different ‘active ingredients’ of leisure - i.e. different patterns of cognitive, physical and social stimuli, which may have different mechanistic pathways to biological ageing.23 Additionally, social identity theory posits that engagement with multiple groups provides more diverse social identities, which can be crucial to psychological processes of stress-buffering, coping and resilience.24 Even when leisure is not overtly social, it can bring personal identities as being part of a collective who does that activity (e.g. “runner” or “artist”). Third, it is crucial to take account of diverse confounding factors. Previous analyses have largely relied on conditioning on confounders via simple regression adjustment. However, this leaves the potential for residual confounding imbalance. More sophisticated statistical approaches like doubly robust estimation offer new opportunities for improving causal inference. Therefore, this study used data from a large, nationally-representative cohort study involving rich phenotyping of ACEng and PA, six different epigenetic clocks, and a doubly robust statistical approach to provide new insight into the relationship between leisure and epigenetic ageing.
Methods Data
Understanding Society, the UK Household Longitudinal Study (UKHLS), is a nationally representative panel survey of members of 40,000 private households in the UK. It was launched in 2009, with participants being followed up annually. Detailed information on the sampling strategy can be found in the sampling design report.25 We used data from the DNAm subsample. Between 2010 and 2012 (waves 2 and 3), blood samples were collected from adult participants during nurse visits. DNAm profiling was conducted from blood samples of 3,654 eligible individuals of white European ancestry who had consented to blood sampling and genetic analysis (wave 2: n=1,425, wave 3: n=2,229). Over 850,000 methylation sites across the genome were measured using the Illumina Methylation EPIC BeadChip. Data were pre-processed via quality control procedures, including outlier removal, filtering poor-quality probes and quantile normalisation.26 Exposures were obtained from the wave 2 adult survey which included a special module on leisure activities. After excluding participants with missing data and outliers in outcome measures (3 standard deviations (SD) from the mean), we had an analytical sample of 3,354 (Figure S1 in the Supplement).
Measures
Outcomes
UKHLS contains six epigenetic clocks constructed from the DNAm data across three generations. First-generation clocks are trained exclusively on chronological age and include the single-tissue Hannum clock, Horvath2013 (estimated from multiple tissues/cells),27 Horvath skin & blood (Horvath2018) clock (another multi-tissue clock with improved accuracy on cultured cells),28 and Lin clock (based on DNAm profiles of 25 cancer types).29
The second-generation clocks are trained on a composite measure of mortality and disease morbidity alongside chronological age. The one available in UKHLS is the PhenoAge clock, which is based on clinical biomarkers of phenotypic age.30
The third-generation clocks are designed to quantify paces of biological ageing rather than static status. The DunedinPoAm clock is considered the first of the third-generation clocks.31 It is based on a comparison of longitudinal changes over time in 18 biomarkers of blood chemistry and organ systems.32
Exposures
ACEng was measured by asking if participants had done anything in four sets of activities in the last 12 months (yes or no): 1) participatory arts (e.g. singing, dancing, painting, photographing, crafting), 2) receptive arts (e.g. attending art exhibitions/events), 3) visiting heritage sites (e.g. historic parks, historic buildings, monuments), 4) other cultural activities (e.g. going to museums, libraries or archives). Frequency of engagement for the four sets of activities, was measured in five categories: ‘once in the last 12 months’, ‘twice in the last 12 months’, ‘less often than once a month but at least three or four times a year’, ‘less often than once a week but at least once a month’, ‘at least once a week’. We also derived an ACEng diversity measure by counting the number of activities and splitting this into quartiles.
PA was measured by a list of sporting activities, including vigorous (e.g. running, swimming, boxing, cycling, football), moderate (e.g. skiing, racquet sports, angling/fishing, yoga/Pilates if age≥64) and mild (e.g. rambling, snooker, yoga/Pilates if age<65) activities. PA frequency was derived by using the highest frequency between vigorous/moderate and mild activities, coded as no, <monthly, monthly, weekly. A PA diversity measure was derived by counting the number of activities, recoded into four categories: none, one, two/three, four or more. For sensitivity analyses, we also considered a self-rated PA activeness measure on a scale of 0 to 10, which was recoded into five categories based on the distribution of the original variable: not active (0), low (1-2), medium (3-4), high (5-6), and very high (7-10).
Covariates
We considered a range of demographic and socioeconomic covariates, including age (range 16 to 90), age-squared, sex (female, male), marital status (single, married/cohabitating, separated/divorced/widowed), living with children (yes, no), living area (rural, urban), education (no qualification, GCSE or below, A level or above, degree or above), household income quintiles, employment status (employed, other), and area deprivation quintiles. Also accounted for are behavioural and health factors, including smoking (never smoker, ex-smoker, current smoker), drinking frequency (on a scale of 1-almost every day to 8-not at all in the last 12 months) and self-reported long-standing physical or mental impairment, illness or disability (yes, no). It is also essential to control for the gap between data collection dates between exposures and outcomes, given the blood samples of 2,229 participants were collected in wave 3. In addition, we also adjusted for a set of technical covariates of various cell composition estimates (CD8-T cells, CD4-T cells, Natural Killer cells, B cells, monocytes and granulocytes).26 Finally, we considered Body Mass Index (BMI), which arguably could be on the causal pathway. Unlike other covariates measured at wave 2, BMI was measured during nurse visits across waves 2 and 3. It was coded into three categories: below 25, 25 and below 30, 30 and above.
Statistical analysis
Data were analysed using doubly robust estimation using the inverse-probability-weighted regression adjustment (IPWRA) estimator. This method involves building two models to account for non-random treatment assignment: i) a regression adjustment model for the outcome and ii) a treatment-assignment model for the exposure. It uses weighted regression coefficients to compute averages of treatment-level predicted outcomes, where the weights are the estimated inverse probabilities obtained from the treatment-assignment model. IPWRA has the double-robust property: it only requires the outcome model or the treatment-assignment model to be correctly specified, not both.33 Both models controlled for demographic, socioeconomic, behavioural and health covariates described above. The outcome model additionally controlled for data collection gaps and technical covariates of epigenetic clocks. The IPWRA model was fitted separately for each exposure (frequency, ACEng diversity, ACEng, PA frequency, and PA diversity). We conducted sensitivity analyses that (i) considered self-reported levels of activeness in PA, (ii) additionally controlled for BMI, and (iii) restricted the sample to those aged 40 or above because it is suggested that ageing is a non-linear process with first substantial acceleration in the 40s.34 All analyses were implemented in Stata 18.
Results Descriptives
The average age of our analytical sample was 52.3 years compared to 47.5 years in the original wave 2 sample, and there was an underrepresentation of single persons (10.6% vs 22.5%). However, the analytical sample was reasonably evenly distributed across household income quintiles, and the distributions of other demographic, socioeconomic, behavioural and health factors were largely similar to the original sample (Table 1). ACEng was relatively common among participants, with 81.7% of people doing three or more activities and 27.6% engaging in 11 or more activities (Table 2). More than three-quarters of people engaged in ACEng monthly or weekly.
PA had a relatively low diversity: 19.6% of participants did not do any PA, and less than a third engaged in four or more activities. However, PA frequency was higher, with half of the participants engaging weekly. PA activeness was roughly evenly distributed in quintiles.
ACEng and epigenetic ageing
For frequency of ACEng, there was no evidence of association with any of the first-generation clocks (Figure 1a). But associations were found with the second-generation PhenoAge clock and the third-generation DuneinPoAm clock. For PhenoAge, although no evidence was found for the difference between the two low-frequency groups, epigenetic ageing was 0.75 years slower in people engaging monthly (95% CI=[−1.46, −0.04], p=0.039), and 0.77 years slower in people engaging weekly (95% CI=[−1.41, −0.13], p=0.018) compared to one or two times yearly. For DunedinPoAm, epigenetic ageing was 0.61 years slower in people engaging at least three times yearly (95% CI=[−1.13, −0.10], p=0.019), 0.74 years slower in people engaging monthly (95% CI=[−1.21, −0.28], p=0.002), and 0.38 years slower in people engaging weekly (95% CI=[−0.79, 0.03], p=0.072).
Estimated average treatment effect and 95% confidence intervals for ACEng diversity and frequency measures from doubly robust estimation using IPWRA (n=3,354)
For diversity of ACEng, there was again no evidence of association with any of the first-generation clocks. However, there was some evidence that the highest level of diversity was associated with a lower value of the second-generation PhenoAge clock (ATE=-0.67, 95%CI=[−1.33, 0.00], p=0.050) (Figure 1b).For the third-generation DunedinPoAm clock, epigenetic ageing was 0.36 years slower for medium diversity (95% CI=[−0.73, 0.00], p=0.051), 0.57 years slower for high diversity (95% CI=[−0.95, −0.20], p=0.003), and 0.73 years slower for very high level of diversity (95%CI=[−1.12, −0.34], p<0.001) compared to low level of diversity.
PA and epigenetic ageing
For frequency of PA, there was no evidence of association with any of the first- or second-generation clocks (Figure 2a). But for the DunedinPoAm clock, epigenetic ageing was 0.43 years slower in people engaging monthly (95% CI=[−0.85, −0.02], p=0.042) and 0.61 years slower in people engaging weekly (95% CI=[−0.94, −0.28], p<0.001).
Estimated average treatment effect and 95% confidence intervals for PA diversity, frequency and activeness from doubly robust estimation using IPWRA (n=3,354)
For diversity of PA, there was again no evidence of association with any of the first-generation clocks (Figure 2b). Some evidence was found for the PhenoAge clock: the highest level of diversity was associated with a lower epigenetic ageing by 0.67 years (95% CI=[−1.82, −0.05], p=0.033). And for DunedinPoAm, low levels of diversity was associated with a lower epigenetic ageing by 0.39 years (95% CI=[−0.77, −0.03], p=0.034), 0.45 years slower for medium diversity (95% CI=[−0.79, −0.11], p=0.009) and 0.37 years for high diversity (95% CI=[− 0.75, 0.02], p=0.064) compared to no PA.
Sensitivity analyses
There was no evidence of associations in first-generation clocks for levels of activeness (Figure 2c). However, very high levels of PA activeness were associated with 1.07 years slower epigenetic ageing for PhenoAge (95% CI=[−1.72, −0.41], p=0.001). And for DunedinPoAm, there was a roughly dose-response relationship between PA activeness and slower epigenetic ageing, with epigenetic ageing 0.44 years slower in people who had low levels of activeness compared to none (95% CI=[−0.76, −0.12], p=0.007), 0.45 years slower in medium levels of activeness (95% CI=[−0.77, −0.13], p=0.006), 0.67 years slower in high levels of activeness (95% CI=[−1.00, −0.34], p<0.001), and 0.82 years slower in very high levels of activeness (95% CI=[−1.17, −0.47], p<0.01.
The associations of ACEng and PA with PhenoAge and DunedinPoAm persisted even after accounting for BMI (Figure S2 & S3 in the Supplement). The results from sensitivity analyses restricting the sample to those aged 40 or above were consistent with the main results, with generally larger effect sizes (Figure S4 & S4 in the Supplement).
Discussion
Using an outcome-wide approach involving six epigenetic clocks, we found associations between two health-promoting leisure activities – arts and cultural engagement (ACEng) and physical activity (PA) – and slower epigenetic ageing. Specifically, ACEng and PA were related to slower biological ageing in the PhenoAge and DunedinPoAm clocks, although not to the other measured clocks, with comparable effect sizes between ACEng and PA (Lin, Horvath2018, Horvath2013 and Hannum). For ACEng, diversity and frequency of engagement were related to DunedinPoAm, while for PhenoAge, there was a slightly clearer relationship for frequency than diversity. For PA, higher levels of frequency, diversity, and activeness were related to DunedinPoAm, while only the highest diversity and activeness were related to PhenoAge. These findings were all stronger amongst middle-aged and older adults.
This was the first study to show a relationship between ACEng and epigenetic ageing. It builds on strong theoretical and empirical underpinnings for why the arts could affect fundamental biological hallmarks of ageing. Life-course psychosocial stressors have been clearly linked with accelerated epigenetic ageing and broader physiological wear and tear across tissues and organ systems.35 One of the fundamental mechanistic effects of arts engagement is reductions in psychophysiological markers of stress, demonstrated in clinical and non-clinical studies.36–38 Notably, variety of engagement has been proposed as key here as it provides opportunities for diverse exposure to active ingredients, formation of multiple identities and even increased social capital (i.e. tangible and intangible resources), all of which support with buffering of stressors.39 So, it is significant that variety and frequency were both related to slower epigenetic ageing. For both ACEng and PA, reductions in inflammatory pathways (which are well-reported for both 40–42) may also be important mechanisms between engagement and epigenetic alterations. Anti-inflammatory effects of ACEng and PA engagement have been linked to methylation status as well as being a hallmark of ageing (“inflammageing”).3,43 Additionally, improvements in cardiovascular risk have been demonstrated to be mediators of the link between PA and epigenetic ageing,21 and this may also be the case for ACEng, for which there is strong mechanistic evidence of benefits for diverse cardiometabolic traits.44–46 Notably, the findings were independent of BMI, which is important given that BMI has been strongly linked to epigenetic age both observationally and experimentally.8
Notably, we only found results for so-called ‘second-generation’ and ‘third-generation’ clocks but not for ‘first-generation’ clocks. This echoes some previous studies. The Rhineland study found results for second-generation clocks (PhenoAge and GrimAge) but not for first-generation clocks (Horvath2013 and Hannum), and a study using The Irish Longitudinal Study of Ageing (TILDA) found associations between physical performance (walking speed) and second-but not first-generation clocks.21,47 Null findings for first-generation clocks have previously been shown for measures related to physical performance in multiple previous studies. 4849 There are several reasons why second- and third-generation clocks may be more relevant to picking up decelerated ageing associated with leisure engagement. First-generation clocks are generally less sensitive predictors of age-related decline in clinical health measures.6 This is because they do not incorporate clinical biomarkers in their derivation and hence are less sensitive to capturing the epigenetic ageing deceleration that results from biobehavioural factors like protective health behaviours.50 First-generation clocks were also trained on cross-sectional data, which, unlike longitudinal data, do not account for mortality selection. This biases the algorithm to select markers that are correlative with ageing rather than causal, because causal loci that should exhibit diminishing age prediction in later life as the individuals exhibiting these traits are progressively selected out of the cohort.51 By using multiple different generations of ageing clocks, our study provides a clearer demonstration of the differential findings between earlier and more recent clocks.
Our study has many strengths, including using a representative cohort study, rich measures of both frequency and diversity of behaviours for our two leisure outcomes, adoption of an outcome-wide approach to epigenetic clocks, and consideration of diverse confounding factors. However, there are some limitations. First, we relied on participants’ self-report on their behaviours, which brings the risk of recall bias and self-report bias. However, some of our clocks overlapped with the Rhineland study, and were corroborated, which is reassuring given that study used objective assessments of behaviours. Second, we included all identified confounding factors, including a particular focus on diverse measures of SEP. But unidentified or unmeasured confounding remains a risk. Nonetheless, we adopted a doubly-robust estimation approach (a methodological advance on previous work relating lifestyle factors to epigenetic clocks), which allows for misspecification in confounders for either the exposure or the outcome. We relied on DNAm present in whole blood. However, the effect of leisure on the epigenome is likely not uniform across the body. This is particularly important for PA, where ageing deceleration may be different in, say, muscle tissues. So, future studies focusing on more specific DNAm tissue data are encouraged. Finally, while we have examined six different clocks within this set of analyses, this is not exhaustive. Other clocks may show different sets of findings.
Overall, our results provide the first evidence that ACEng is related to epigenetic ageing, suggesting the value of its exploration alongside other lifestyle factors.8 In particular, diversity of engagement appears as important as frequency of engagement. It is also of note that the effect size was comparable for ACEng and for PA with respect to epigenetic ageing. ACEng is a much more recently recognised health behaviour, but these findings suggest the importance of continued exploration into its health effects alongside other more established health behaviours. We also extend the existing understanding of the relationship between PA and epigenetic ageing, demonstrating a clear association with second-but not first-generation epigenetic clocks and extending analyses to new clocks not included in previous analyses. Our findings are relevant for several reasons. First, decelerations in ageing clocks, including those within our study, have been demonstrated to have clinical (as well as statistical) importance, including improvements in physical performance, polypharmacy, cognitive state and all-cause mortality risk.47 Indeed, our strongest results involved DunedinPoAm, which has been related to improved performance in physical, cognitive and physiological measures of ageing over subsequent years.32 And it is notable that associations between leisure and epigenetic ageing became more prominent in adults from mid-life. Second, recent work suggests that epigenetic ageing is potentially reversible.52 The persistence of epigenetic changes in response to modifiable behaviours such as leisure engagement is greatly underexplored. However, given the experimental evidence reviewed earlier on the effects of both ACEng and PA on DNAm generally and epigenetic clocks specifically, 17,11–14 future intervention studies could explore whether lifestyle changes have any value to slowing or potentially reversing epigenetic ageing.
Data Availability
Data are publicly available through UKDS https://beta.ukdataservice.ac.uk/datacatalogue/series/series?id=2000053
Funding
This paper is supported by the UK Research and Innovation [MR/Y01068X/1]. LM was supported by a PhD studentship provided through the Soc-B Centre for Doctoral Training (CDT), funded by the Economic and Social Research Council (ESRC) and the Biotechnology & Biological Sciences Research Council (BBSRC).
References
