Epigenetic-based age acceleration in a representative sample of older Americans: Associations with aging-related morbidity and mortality

Among 3,581 Health and Retirement Study (HRS) participants, epigenetic clocks created using original methods and PC-trained methods yield very different estimates of epigenetic mean age and age ranges (Table 1). Epigenetic age ranges from 53.73 to 66.97 with the original clocks and 63.96 to 77.02 using the PC-trained clocks. With the exception of the Horvath 2013 clock, using PC-trained clocks increases the mean epigenetic age of the sample and brings it closer to the mean chronological age [Mean (SD)= 68.25 (9.18) y]. Complete descriptives of the sample are included in SI Appendix, Table S1. The epigenetic age range with PC training is variable.

Increased epigenetic age relative to chronological age, or age acceleration, was calculated for each clock taking the residual of the clock values regressed on age. This way the effect of “biological aging” can be examined independently of age. By design, average age acceleration is 0 for all of the clocks. In our sample, the DunedinPACE clock has a mean of 1.02 (0.15). This clock does not represent age; rather it represents the average pace of aging per year. This value suggests slightly faster aging biologically than chronologically in this sample.

Correlations between age acceleration measures are shown in SI Appendix, Table S2. The first generation age acceleration measures, Horvath and Hannum, are moderately correlated, and PC training strengthens this correlation. PhenoAge and GrimAge are modestly correlated with each other and somewhat correlated with the first generation clocks. Again, using PC-trained clocks improves the correlation. GrimAge age acceleration and Dunedin PACE are more correlated before PC training.

Regression coefficients that show the associations of sex, race/ethnicity, education, health behaviors, depression, and childhood socioeconomic status (SES) and hardship with the accelerated age epigenetic measures and DunedinPACE are shown in SI Appendix, Tables S3A and S3B, for the original and PC-trained measures of age acceleration, respectively. Females have lower age acceleration than males across all the measures, with and without PC training, with the exception of GrimAge. For the calculation of GrimAge, sex is also regressed out of the measure because both age and sex are used in the calculation of the GrimAge measure. Education is not significantly related to age acceleration with the first generation clocks or the original PhenoAge. However, lower education is associated with faster aging using the DunedinPACE measure, both the original and PC-trained GrimAge acceleration measures, and PC-trained PhenoAge acceleration; all with the expected gradient in the effect estimates.

The first generation clocks show slower aging for Blacks compared to Whites, while the original version of the second generation and DunedinPACE measures indicate faster aging for Blacks. PC training does not change the direction of the effect but does increase the magnitude of the estimates for all but GrimAge. The results for Hispanic ethnicity generally indicate slower aging among Hispanics.

Obesity is consistently associated with accelerated aging across all the clocks. Smoking, as measured by pack-years, is associated with faster aging with the original clocks with exception of the Hannum clock. Heavy drinking is associated with faster aging with the original second generation clocks and DunedinPACE. Pack-years are associated with faster aging with all the PC-trained clocks; however, drinking only remains associated with faster aging after PC training using the GrimAge measure. Prior depression is associated with faster accelerated aging with the original versions of the Hannum and GrimAge clocks as well as all of the PC-trained clocks and the DunedinPACE measure.

Low childhood SES is associated with faster aging using the PhenoAge measure. After PC training, only the Hannum age acceleration measure is associated with childhood SES. Low childhood SES is also associated with faster aging using the DunedinPACE measure. Counterintuitively, childhood financial hardship is significantly associated with slower aging using the original Horvath, Hannum, and PhenoAge acceleration measures. After PC training, childhood financial hardship remains significantly associated with slower aging with the Horvath and Hannum age acceleration measures.

The R² decomposition from ordinary least squares (OLS) regression for the PC-trained measures of age acceleration and DunedinPACE is presented in Fig. 1. Health behaviors, including obesity, smoking, and heavy drinking, explain the largest proportion of variance in the second and third generation clocks. Smoking pack-years is one of the main parameters of the GrimAge measure so it is not surprising that health behaviors, including smoking, comprise such a large proportion of the variance explained for this measure. The total variance explained by the distribution of cell types is quite consistent across the clocks but proportionally much larger with the first generation clocks.

Table 2 shows the relationship between the age acceleration measures and select health outcomes, with and without PC training. All models are adjusted for age and sex. The models with the PC-trained measures are also presented with adjustment for SES, health behaviors, and cell distribution. The full set of regression coefficients for the PC-trained models is included in SI Appendix, Tables S4–S8.

Faster age acceleration, using the PhenoAge and GrimAge measures, is significantly associated with increased difficulty with activities of daily living and instrumental activities of daily living (ADL/IADLs) assessed 2 y after DNAm measurement. These relationships remain unchanged after PC training and adjustment for the complete set of covariates and cell-type proportion. The original and PC versions of these estimates are not statistically different from one another. DunedinPACE also predicts ADL/IADL difficulties although the relationship is somewhat attenuated after covariate and cell-type adjustment.

Age acceleration measured by all the clocks, with and without PC training, is associated with multimorbidity measured 2 y after DNAm measurement. Adjustment for covariates and cell type does not impact the significance of any of the estimates but does reduce the beta coefficients for all age acceleration measures to a degree. With the exception of PhenoAge, the original and PC versions of the estimates are not significantly different from one another.

With available data, cognition could only be analyzed cross-sectionally with DNAm. Faster age acceleration using the second generation clocks and DunedinPACE is associated with higher cognitive dysfunction. These relationships are mostly unchanged with PC training, with only the Horvath and PhenoAge acceleration estimates changing significantly with PC training. All of the coefficients are ameliorated with adjustment for demographic, SES, and behavioral covariates and cell type, with the estimates for PhenoAge, GrimAge, and DunedinPACE significantly different in the fully adjusted model compared to the unadjusted PC version models.

Overall, three of the four epigenetic age acceleration measures are significant predictors of 4-y mortality, with the exception of the original Horvath age acceleration measure. All PC-trained age acceleration measures are significant predictors of 4-y mortality. Again, PhenoAge is the only measure that is significantly different with PC training. The relationships between the age acceleration measures and mortality do not change appreciably with adjustment by demographics, SES, health behaviors, and cell type. DunedinPACE is also highly predictive of mortality with and without covariate adjustment.

SI Appendix, Table S9 shows the coefficients presented in Table 2 along with the R²s for each model organized more easily for comparison. This table shows that with the exception of PhenoAge, the R² of the models remain largely unchanged when using the PC-trained version of the age acceleration measures as compared to the original version. The models with the age acceleration measures and full adjustment for demographics, SES, health behaviors, and cell-type proportions explain approximately 14% of the variance in ADL/IADL difficulties; this relationship is very consistent across the clocks, including DunedinPACE. Similarly, this combination of measures predicts between 15 and 18.6% of the variation in multimorbidity measured 2 y out, with the model including age acceleration from the PhenoAge clock performing the best. The full models do better predicting concurrent cognitive dysfunction, explaining approximately 28% of the variance in the outcome. The PC adjusted models with age acceleration explain between 22 and 26% of the variance in the mortality estimates (SI Appendix, Tables S4–S9).

When all of the epigenetic age acceleration measures are added in a single model with full covariate adjustment, PhenoAge and GrimAge acceleration remain significant independent predictors of future ADL/IADLs and mortality, whereas the Horvath, PhenoAge, GrimAge, and DunedinPACE measures all independently predict multimorbidity, and only the Hannum and PhenoAge measures predict cognitive dysfunction (SI Appendix, Table S10). In a two factor model, with the Horvath and Hannum clocks loading on factor 1 and the PhenoAge, GrimAge, and DunedinPACE measures loading on factor 2 for the prediction of ADL/IADLs and multimorbidity, only the second generation measure factor is significantly associated with ADL/IADs, whereas both factors predict multimorbidity. These relationships do not change with PC adjustment or in the fully adjusted model. In the prediction of cognition and mortality, the Horvath, Hannum, PhenoAge, and GrimAge clocks load on factor 1 while DunedinPACE loads on factor 2. In the fully adjusted version of the models, only the DunedinPACE factor predicts cognitive dysfunction, with increasing epigenetic age predicting better cognition. Both factors are significant predictors of mortality.

One of the aims of this paper is to compare how well epigenetic age acceleration predicts age-related outcomes compared to known social and behavioral predictors such as SES and health behaviors. Fig. 2 shows the R² decomposition from the OLS or logistic regression of the health outcomes on the PC-trained GrimAge acceleration measure and covariates. As shown in this figure, the PC-trained age acceleration measure contributes the most explanatory power to the models predicting multimorbidity and mortality and much less to the prediction of ADL/IADL difficulties and cognitive dysfunction. The total variance explained by the PC-trained GrimAge acceleration measure across all outcomes is <5%. Demographic variables are significant predictors across all outcomes, most notably for mortality. Prior depression is the largest contributor to the prediction of functional limitations 2 y later, while SES (education, race/ethnicity) comprises the largest share of the prediction of concurrent cognitive dysfunction.

Lastly, we compare the predictive capability of these epigenetic age acceleration measures with a composite measure of 22 primarily blood-based clinical chemistry biomarkers that has been shown to predict multimorbidity in the same cohort (28). When combined into one model, we find that the epigenetic age acceleration measures are no longer significant predictors of future ADL/IAD difficulties or concurrent cognitive dysfunction. However, the ability of the epigenetic age acceleration measures to predict mortality remains largely unchanged and only slightly reduced for future multimorbidity (SI Appendix, Table S11).

The purpose of these analyses is to show the utility of DNAm-based age acceleration measures in the prediction of cross-sectional and longitudinal health outcomes in a representative sample of older adults. We also examine whether recent improvements to these scores, designed to remove some of the technical noise, improve the predictive capability of these measures. Lastly, we examine how well DNAm-based measures perform against well-known predictors of health outcomes such as demographics, SES, and health behaviors.

While PC training generally leads to an upward shift in the mean epigenetic age of the sample, PC training does not significantly change the relationship of DNAm-based age acceleration measures to demographics or health behavior predictors nor the relationship between age acceleration and select health outcomes, cross-sectional and longitudinal. PhenoAge acceleration is the measure most affected by PC training in the prediction of these outcomes.

Prior work using this sample identified more accelerated aging among men, those with lower education and those who are obese or smoke (13). Generally, the relationships between demographics and health behaviors with age acceleration are unchanged with PC training, as noted in Higgins-Chen et al. (24). This analysis adds heavy drinking as a health behavior that is associated with more rapid epigenetic aging with only GrimAge-based age acceleration after PC training.

Previous studies have identified relationships between epigenetic age acceleration and several age-related health outcomes, for example cognitive impairment, poor physical and cognitive fitness, and higher risk of all-cause mortality (29–31). Much of the prior work, including several meta-analyses is specific to the Horvath or Hannum clocks and often finds some discordance in the findings (32). This paper adds evidence that the more recently developed clocks trained on biological phenotypes quite consistently show that epigenetic age acceleration is associated with poorer health and mortality (8, 10, 33), measured 2 to 4 y after DNAm. In our sample, age acceleration calculated using the second generation clocks, PhenoAge and GrimAge, as well as the DunedinPACE measure, is consistently a significant predictor of health outcomes including cross-sectional cognitive dysfunction, and number of ADL/IADL difficulties and chronic conditions assessed 2 y after DNAm measurement. The original and PC-trained age acceleration measures using the Horvath clock are not associated with ADL/IADL difficulties nor cognitive dysfunction. A similar pattern is seen with the Hannum clock-based measure. To some extent, this is to be expected considering how the various clocks were trained. The Horvath and Hannum clocks were trained to predict chronological age, not “biological age” per se. The second generation clocks were trained on a variety of health indicators (PhenoAge) and mortality (GrimAge), so their improved predictive ability does not come as a surprise. The DunedinPACE measure was trained on biological processes as well, specifically within-individual decline in 19 biological indicators across four time points and nearly 20 y, from age 26 to 45 y, so by design should also perform well with biological-based endpoints. All of the age acceleration measures are predictive of multimorbidity assessed 2 y later, and all of the PC-trained versions are predictive of mortality (the Horvath measure being the only one that isn’t predictive of mortality prior to PC training). In all of the models, the relationship between age acceleration and the health outcome is somewhat attenuated after covariate and cell-type adjustment, but all remain significant. Adjustment for other blood-based biomarkers measured concurrently with DNAm removes the effect of the epigenetic age acceleration measures in the prediction of ADL/IADLs and cognition, but DNAm age acceleration measures remain significant predictors of future multimorbidity and mortality.

In terms of interpretation across the different clocks, it is important to keep the scale of measurement in mind. A change of one unit for the age acceleration measures is about 1/5 of a SD. For the DunedinPACE measure, a one unit change represents approximately 6.666 SDs. That means we would expect the beta coefficients for DunedinPACE to be approximately 36 times those of the other clocks. The large OR for mortality with DunedinPACE should be interpreted considering this difference in scale between the clock measures. Standardization of GrimAge and DunedinPACE odd ratios for mortality (Table 2) makes such a comparison easier and shows that GrimAge acceleration is a stronger predictor of 4-y mortality.

Adding epigenetic age acceleration increases the total variance explained in the model for each outcome, most notably by approximately 3% for multimorbidity and mortality outcomes. Interestingly, epigenetic age acceleration does not significantly decrease the R² of the other predictors in most models. The most notable change is found with the addition of GrimAge acceleration to the multimorbidity model which reduces the proportion of variance explained by health behaviors from 5.4% to 4.6%. While the utility of DNAm-based age acceleration in the prediction of later life health outcomes is quite clear from these analyses, it should be noted that for most outcomes, these measures predict less than other characteristics of the sample that are usually much easier and cheaper to ascertain. Only for the prediction of longitudinal multimorbidity using GrimAge is age acceleration, the largest predictor of the outcome of interest (R² = 0.047), just above health behaviors. This doesn’t suggest that DNAm is not a piece of the puzzle, just that it does not usurp the importance of demographic, SES, mental health, and health behaviors in predicting later life health outcomes.

There are also important caveats to this work stemming from the data that are used to train the “epigenetic clocks” and the methods we employ to create them. All of the clocks to a lesser or greater extent are dependent on the composition of the samples or cohorts that were used to create them. We make the assumption that the relationships between the methylation of certain CpG sites and the outcome we define in training the clocks are consistent across cohorts; however, it could entirely be the case that these relationships are molded by SES or other environmental or contextual exposures that are unique or different in the source sample. For example, the DunedinPACE measure capitalizes on a unique set of biological data collected over 20 y. However, the birth cohort used to train this measure has only been examined through age 45. We assume that the relationship we see in the decline of these organ systems between ages 26 and 45 y extends linearly into older age. It is highly likely this is not the case for all persons or even all populations. We will learn more as this cohort is observed over time.

The PC-trained method is also informed by the source data for each clock. The purpose of PC training is to remove any effect of unreliably measured CpG sites on downstream measures such as epigenetic clock-based age and age acceleration measures. To do this, PCs are identified from the source, or training, data for each clock and then factor loadings of each CpG site on each PC are estimated. The “clocks” are then retrained, this time using the PCs to predict the outcome, be it chronological age or a more health-related phenotype. The resulting beta weights, each CpG site to each PC, and for each PC to the outcome, are applied to the sample data to create PC-trained clocks. This novel process uses the PC level of the model to pull out the important relationships with age and ignores the random distribution of inconsistent measurement across CpG sites. However, PC identification is determined by the characteristics of the training dataset. The PCs that are generated in one dataset might not translate well to another sample—here factors such as the demographic and SES distribution of the source data would affect the initial identification of the PCs, while the relationship between the PCs and the outcome would also be affected. Moreover, this method removes variance in CpG measurement but does not help us to understand the underlying mechanisms that may connect DNAm modification and health. Moreover, DNAm is just one measure of epigenetic modification. Other epigenetic measures, such as histone modification, and other—omic measures such as proteomics or metabolomics, will likely help elucidate more of the biology that underlies age-related changes in the body.

Additional research will be needed to replicate these results in other cohorts and to examine whether age acceleration predicts differently within subgroups—defined by demographics, SES, or some environmental exposure. The examination of the utility of DNAm-based age acceleration measures to predict health outcomes is constantly evolving as we improve our methods of measuring methylation, and our understanding of how DNAm fits within the larger relationship between aging and health. The measures used in these analyses have been made publically available by the HRS which will contribute to our scientific progress in this area.

We examined the mean and range of each clock and measure of accelerated aging, with and without PC training. Next we performed ordinary least squares regressions of accelerated aging measures on age, sex, race/ethnicity, education, health behaviors [smoking, alcohol use, obesity (BMI ≥ 30)], depression, childhood SES and hardship, and cell type to examine the association of each variable with each aging acceleration measure as well as calculating the proportion of the variation in each measure of accelerated aging that is predicted by the independent variables using the Shapley decomposition approach. We analyzed continuous outcome data using OLS regression. We analyzed binary outcome data (mortality) using logistic regression. All outcome models were adjusted for sex and age. Additional models were also adjusted for race/ethnicity, education, health behaviors [smoking, alcohol use, obesity (BMI ≥ 30)], depression, childhood SES and hardship, and distribution of cell types to examine whether these variables mediated the relationship between age acceleration and the outcomes. Additional models also adjusted for a composite measure of 22 blood-based clinical chemistry biomarkers measured at the same time as the DNAm assessment. (28) All models were run using Stata Version 16.

Code used for analysis and to prepare figures will be accessible via GitHub upon acceptance of the article for publication.

All clock and age acceleration measures used in these analyses are available on the HRS website for download by registered users (https://hrsdata.isr.umich.edu/data-products/epigenetic-clocks).

This work was supported by the National Institute on Aging (R01 AG060110, R01 AG068937, and R01 AG071071). The HRS is supported by the National Institute on Aging U01 AG009740. Thank you to Em Arpawong, Albert Higgins-Chen, and Jonah Fisher for help calculating the clocks. Thank you to Helen Meier and Colter Mitchell for help with interpretation.