Testing DNA-methylation and blood-chemistry measures of biological aging in models of Black-White disparities in healthspan characteristics

Biological aging is a proposed mechanism through which social determinants drive health disparities. We conducted proof-of-concept testing of eight DNA-methylation and blood-chemistry quantifications of biological aging as mediators of disparities in healthspan between Black and White participants in the United States Health and Retirement Study (HRS; n=8231). We quantified biological aging from four DNA-methylation "clocks" (Horvath, Hannum, PhenoAge, and GrimAge), a DNA-methylation Pace of Aging (DunedinPoAm), and three blood-chemistry measures (PhenoAge, Klemera-Doubal method Biological Age, and homeostatic dysregulation). We quantified Black-White disparities in healthspan from tests of physical-performance, self-reported limitations to activities of daily living (ADLs), and physician-diagnosed chronic diseases. DNA-methylation and blood-chemistry quantifications of biological aging were moderately correlated (Pearson-r range 0.1-0.4). GrimAge, DunedinPoAm and all three blood-chemistry measures were associated with healthspan characteristics (10-25% increase in risk per SD of biological aging) and showed evidence of more advanced/faster biological aging in Black compared with White participants (Cohen's d=.3-.5). In mediation analysis, these measures accounted for 19-48% of Black-White differences in healthspan-related characteristics. Evidence that Black Americans are both biologically older and aging more rapidly than White Americans of the same chronological age suggests that differences in aging may represent a novel pathway to understand and eliminate health disparities.


INTRODUCTION
Black Americans experience excess morbidity and premature mortality relative to White Americans (1). This health disparity is mediated by multiple chronic diseases affecting different organ systems throughout the body and reflects an etiology extending from the earliest stages of life across adulthood, encompassing social, economic, and environmental factors (2)(3)(4). Differences in health between Black and White Americans vary between geographic locations and have changed over time, indicating that these disparities are socially determined and that they are modifiable (5). A range of policies and programs are proposed to mitigate health disparities (6)(7)(8)(9). However, rigorous evaluation of impact is challenging (10). Interventions to address health disparities delivered to older adults may come too late to prevent chronic disease (4,11). Interventions delivered to younger people require long follow-up intervals to establish impact (12). Methods are needed to monitor effectiveness of interventions over timescales of years rather than decades.
Measurements that quantify processes of biological aging may provide near-term measurements of long-term impacts. Biological aging is the gradual and progressive decline in system integrity with advancing chronological age (13). This process is now being studied as a modifiable root cause of many different chronic diseases (14)(15)(16). One hypothesis advanced to explain the range of diseases showing Black-White health disparities is that the social and material stresses experienced by Black Americans act to accelerate biological aging, referred to as "weathering" (17,18). In epidemiologic studies, Black Americans show more advanced biological aging as compared to White Americans of the same chronological age (19)(20)(21). If advanced biological aging is a mediator of health disparities, then quantifications of biological aging could be used to monitor intervention impacts.
Many methods are proposed to quantify biological aging from several biological levels of analysis (22,23). Agreement between measures is often poor; there is no gold standard (24-26). Measures based on analysis of blood-chemistry and DNA methylation data have received the most attention to date. We conducted proof-of-concept testing of eight blood-chemistry and DNA-methylation methods to quantify biological aging as mediators of Black-White disparities in healthy aging. We analyzed deficits in physical functioning, limitations to activities of daily living, and chronic disease morbidity in a national sample of older adults in the US Health and Retirement Study. Our analysis proceeded in three steps (27). First, we confirmed that proposed biological aging measures showed expected relationships with the several measures of healthy aging in Black and White older adults. Second, we evaluated Black-White disparities in the healthy aging measures and the measures of biological aging. Finally, we evaluated the proportion of Black-White disparities that might be attributable to biological aging.

Sample
The Health and Retirement Study (HRS) is a nationally representative longitudinal survey of US residents ≥50 years of age and their spouses. The HRS has been fielded every two years since 1992. Participants are asked about four broad areas: income and wealth; health, cognition and use of healthcare services; work and retirement; and family connections. A new cohort of 51-56 year-olds and their spouses is enrolled every six years to maintain representativeness of the U.S. population over 50 years of age. Response rates over all waves of the HRS range from 81-91%. As of the most recent data release, HRS included data collected from 42,515 individuals in 26,600 households. We linked HRS data curated by RAND Corporation with new data collected as part of HRS's 2016 Venous Blood Study. Analysis included participants aged 50-90 years at the time of the blood draw, who responded to the question "What race do you consider yourself to be" by selecting either Black/African American or White, and for whom data were also collected on prevalent chronic disease, limitations to activities of daily living, and/or testing to assess physical functioning (n=8,231). Comparison of our analysis samples to the full HRS is reported in Supplemental Table 1.

Measures
Biological Aging. There is no gold standard measure of biological aging (22). Many methods have been proposed based on different biological levels of analysis. Current state-of-the-art methods use machine learning to sift large numbers of candidate markers and parameterize algorithms that predict aging-related parameters, including chronological age, mortality risk, and rate of decline in system integrity. Algorithms are developed in reference datasets and can then be applied in new datasets to test hypotheses. We focused on algorithms developed for blood chemistry analytes routinely measured in clinical settings and DNA methylation marks included on commercial arrays. We analyzed several different methods because each method uses different assumptions to develop a measure of aging and, as we and others have shown, the different methods do not all measure the same aspects of the aging process (24, 26). Measures are summarized in Table 1. Detailed descriptions of the measures are reported in the Supplemental Methods (Section 1).
Healthspan-related characteristics. Healthspan is the portion of life lived free of disease and disability. We measured indications of healthspan from performance tests administered by trained interviewers and participant reports about disability and physician-diagnosed chronic disease morbidity. Measures are described briefly below and in more detail in Supplemental Methods Table 1.
Functional limitation (n=3589) was measured as the count of performance tests on which participants demonstrated low levels of performance (bottom third of the HRS distribution). The four functional tests were gait speed, balance, grip strength, and peak flow. The count of limitations was coded 0, 1, 2, or 3+. Functional performance testing is administered to half of . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 5, 2021. ; https://doi.org/10.1101/2021.03.02.21252685 doi: medRxiv preprint HRS participants each year. We analyzed data from the (n=xxx) participants in our analysis sample who completed performance testing in 2016.
Disability (n=8231) was measured as the count of limitations to activities of daily living (ADLs) reported by participants (coded 0, 1, 2, or 3+).
Chronic disease morbidity (n=8231) was measured as the count of chronic conditions diagnosed by a physician as reported by the participant (coded 0, 1, 2, or 3+).

Analysis
We tested associations of biological aging measures with healthspan-related characteristics using Poisson regression to estimate Incident Rate Ratios (IRRs) for counts of functional deficits, ADLs, and chronic disease diagnoses. Effect-sizes for biological aging measures were denominated in standard-deviation units, defined as the difference between the estimated biological age of a participant and their true chronological age. We quantified Black-White disparities in these healthspan-related characteristics using the same approach. All models included covariates for sex and chronological age.
We tested mediation of Black-White disparities in biological aging using a regressionbased approach as described by Valeri and VanderWeele (28). This method allows for the decomposition of the total effect of the exposure on the outcome of interest through direct and indirect pathways. We specified two regressions for each relationship of interest, ( | , , ) and ( | , ). In these expectation statements, is the outcome, here a healthspan-related characteristic, is the mediator, here biological-age advancement, and is the exposure of interest, here a social determinant of health. We modeled ( | , ) using Poisson regression to estimate IRRs for the healthspan-related characteristics, which were operationalized as counts. We modeled ( | ) using linear regression of the continuouslydistributed measure of biological age advancement on the social determinant of health (Black versus White racial identity). Models included sex and chronological age as covariates. We used the R package CMAverse (29) to estimate direct and indirect effects based on these regressions. We calculated confidence intervals using bootstrapping to obtain standard error estimates. We tested robustness of mediation results to potential exposure-mediator interactions following the approaches outlined by Valeri and VanderWeele (28, 30). Details of this analysis are in Supplemental Methods Section 2.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

RESULTS
We conducted two sets of analysis. First, we analyzed blood-chemistry measures of biological aging. Next, we analyzed DNA-methylation measures of biological aging, including comparison to blood-chemistry measures. Within each analysis set, we first tested associations of biological-aging measures with healthspan-related characteristics. Next, we tested Black-White disparities in biological aging. Finally, for those measures demonstrating associations of moreadvanced/faster aging with healthspan-related characteristics and more-advanced/faster biological aging in Black as compared to White participants, we conducted mediation analysis.
I. Blood Chemistry Analysis. We analyzed blood-chemistry data from n=8,231 HRS participants who identified as Black or White (41% male, 81% White), aged 50-90 years (mean=69, SD=9). Comparison of this sample to the 2016 panel of the HRS is reported in Supplemental Table S1. Participants' Phenotypic Ages were highly correlated with their chronological ages (r=0.77) and were about half a year older (mean difference=0.62 years, SD=9), indicating that they had experienced roughly the amount of aging expected based on the NHANES reference sample in which the algorithm was developed. To quantify differences between observed and expected levels of biological aging, we calculated biological-age advancement as the difference between a participant's Phenotypic Age and their chronological age. As expected, men exhibited more advanced biological aging as compared to women (b=2.51, 95% CI=[2.14-2.89], p-value<0.0001). Results were similar for the Klemera-Doubal method (KDM) Biological Age and homeostatic dysregulation measures (Supplemental Table S1).
We repeated analysis using the KDM Biological Age and homeostatic dysregulation measures. Results were similar, although effect-sizes for the Phenotypic Age measure were somewhat larger, consistent with previous comparisons of these measures (20, 31). Results for all biological aging measures are reported in Supplemental Table S2.
As compared to White participants of the same chronological age, Black participants exhibited an additional 2.65 years of biological aging (95% CI [2.18-3.12], Figure 1C), according to the Phenotypic Age measure. Results were similar for analysis of KDM Biological Age and homeostatic dysregulation (Supplemental Table S1).  Figure 1D). Results were similar for analysis of KDM Biological Age and homeostatic dysregulation, although the proportion mediated was somewhat larger for the homeostatic dysregulation measure (Proportion mediated range 0.25-0.48). Effect-size estimates for the association of race with healthspanrelated characteristics before and after adjustment for biological aging measures are reported in Supplemental Table S3. Parameter estimates from the mediation models are reported in Supplemental Table S4. II. DNA methylation analysis. We analyzed DNA methylation data from n=3491 HRS participants who identified as Black or White (42% male, 82% White), aged 50-90 years (mean=70, SD=9). Comparison of this sample to the 2016 panel of the HRS is reported in Supplemental Table S1. Participants' clock ages were highly correlated with their chronological ages (r>0.72), but tended to be younger (mean difference=-0.02 years, range=-0.07 to 0.02 years). Participants' Dunedin PoAm values indicated they were aging 7% faster than the expected rate for young adults of 1 year of biological change per chronological year (M=1.07, SD=0.09).
To quantify differences between observed and expected levels of biological aging, we regressed participants' clock ages on their chronological ages and computed residual values. Positive residuals reflect more advanced biological aging whereas negative residuals indicate delayed aging. No residualization was applied to DunedinPoAm since this measure already quantifies deviation from the expected pace of aging. As expected, men exhibited somewhat more advanced biological aging and faster pace of aging as compared to women (clock difference range=[0.85-3.13]; DunedinPoAm difference=0.02; Supplemental Table S1).
Comparison of DNA-methylation and blood-chemistry measures of aging. We compared DNA-methylation and blood-chemistry measures of aging in the n=3491 participants with both sets of measurements (Supplemental Table S1). DNA methylation measures of aging (clock age residuals and Dunedin PoAm) were weakly to moderately correlated with one another (r<=0.4). Correlations of DNA methylation measures of aging with blood chemistry measures were varied. Horvath-clock age-residuals were not correlated with blood chemistry measures of aging (r<0.1); Hannum-clock age-residuals were weakly correlated with blood chemistry measures (r=0.1-0.2); correlations were somewhat stronger for PhenoAge-clock . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. DNA methylation analysis of healthspan-related characteristics and Black-White disparities. Participants with more advanced epigenetic aging were at increased risk for reduced healthspan. GrimAge-clock residuals showed the strongest associations (IRR=1.11-1.23 per SD). Associations were weaker for PhenoAge-clock residuals (IRR= 1.06-1.11 per SD), Hannum-clock residuals (IRR=1.04-1.09 per SD), and Horvath-clock residuals (IRR=1.03-1.06 per SD). Participants with faster pace of aging also showed increased risk for reduced healthspan (IRR=1.07-1.17 per SD). Results are shown in Figure 2 and reported in Supplemental Table S2.  Figure 3.
More advanced DNAm GrimAge and faster DunedinPoAm accounted for 1/5-1/3 of Black-White differences in healthspan-related characteristics. The Horvath, Hannum, and PhenoAge clocks did not indicate more advanced aging in Black as compared to White participants and were not included in mediation analysis. Effect-size estimates for Black-White disparities in healthspan-related characteristics before and after adjustment for GrimAge clock and DunedinPoAm pace of aging measures are reported in Supplemental Table S3. In mediation analysis, between one fifth and one third of the Black-White disparities in healthspan-related characteristics were attributed to more advanced GrimAge and faster DunedinPoAm (  . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 5, 2021. ; https://doi.org/10.1101/2021.03.02.21252685 doi: medRxiv preprint

DISCUSSION
We investigated biological aging as a mediator of Black-White disparities in healthspanrelated characteristics in the US Health and Retirement Study. In our sample, Black participants experienced more functional limitations, more difficulties with activities of daily living (ADLs), and more chronic-disease diagnoses as compared to White participants. Black participants also showed more advanced/faster biological aging on all three blood-chemistry measures, and the DNA methylation GrimAge clock and DunedinPoAm pace of aging. In mediational analysis, more advanced biological aging on these measures accounted for between one-quarter and one-half of Black-White disparities in healthspan-related characteristics. These findings are consistent with the weathering hypothesis that racially patterned determinants of health accelerate biological aging, contributing to Black-White disparities in healthspan, and provide proof-of-concept for use of quantifications of biological aging in health disparities research.
All three blood chemistry measures showed consistent associations with healthspan characteristics, indicated more advanced biological aging in Black as compared to White participants, and mediated a portion of Black-White disparities in healthspan characteristics. Results were mixed for DNA methylation measures. Only the second-generation DNA methylation clock GrimAge and the DunedinPoAm pace of aging measure showed consistent associations with healthspan characteristics, indicated more advanced/faster biological aging in Black as compared to White participants, and mediated a portion of disparities in healthspan characteristics. In contrast, effect-sizes for associations with healthspan characteristics tended to be somewhat smaller for the Horvath, Hannum, and PhenoAge clocks and, as reported previously (32), none of these three indicated more advanced biological aging in Black as compared to White participants.
These mixed findings for DNA methylation measures of aging suggest guidance to future studies. Results contribute evidence that first-generation DNA methylation clocks developed to predict chronological age are both less predictive of healthspan-related characteristics and less sensitive to exposures that shorten healthy lifespan as compared to blood-chemistry-derived measures, the DNA methylation pace of aging, and also newer DNA methylation clocks developed to predict mortality (24, 26, 33-35). Future studies investigating biological aging as a mediator between risk exposures and healthy-lifespan phenotypes, especially in the context of health disparities, may be best served by focus on the GrimAge clock and DunedinPoAm pace of aging measures. The HRS result that the PhenoAge clock did not indicate more advanced biological aging in Black as compared to White participants contrasts with a report from the Women's Health Initiative (21). Follow-up in additional cohorts is needed.
A consideration in studies testing mediation of Black-White disparities is whether candidate mediators are measured with the same precision in both groups. Consistent with some other reports, e.g. (20), we observed somewhat smaller effect-sizes for associations of aging measures with healthspan-related characteristics in Black as compared to White participants. This could indicate differential measurement precision. Differential measurement . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 5, 2021. ; https://doi.org/10.1101/2021.03.02.21252685 doi: medRxiv preprint precision is a potential concern, especially for the DNA methylation measures, which were developed from analysis of all-White samples. However, smaller effect-sizes for associations with healthspan-related characteristics in Black as compared to White participants were also observed for chronological age, which is presumed to have identical measurement precision in Black and White participants. Differences in effect-size might instead reflect that Black Americans are disproportionately subject to non-aging causes of disease and disability such as injury or accidents, which would not be captured in measures of biological aging.
We acknowledge limitations. Our analysis used cross-sectional data. The bloodchemistry and DNA methylation data used to measure participants' biological aging were collected at the same measurement wave as their healthspan-related characteristics. The absence of temporal ordering leaves open the possibility of reverse causation. Replication of findings in longitudinal data is needed. There is no gold standard measure of biological aging (36). Our conclusions regarding biological aging as a candidate mediator of health disparities could be specific to the measures we analyzed. However, we found consistent evidence across different biological substrates and measurements derived using several different approaches. Consistency of findings across the GrimAge and Dunedin PoAm DNA methylation measures and the Phenotypic Age, Klemera-Doubal method, and homeostatic dysregulation blood-chemistry measures builds confidence that results do reflect aging processes.
Mortality selection may bias results toward the null in this sample. Our analysis included adults ages 50-90. In that age range, many individuals born in the same years as study participants will not have survived to the time of data collection. This mortality selection is likely even more pronounced for Black participants; in the birth cohorts represented in our sample, Black Americans have shorter life expectancy (37). If survivors aged more slowly than those who died at younger ages this difference could attenuate Black-White differences in biological aging and healthspan. Black participants in the Venous Blood Study were also younger and healthier than the full sample of Black participants in the HRS. Our estimates of disparities are therefore likely to be conservative. Finally, there is the possibility of detection and reporting bias. For example, racial disparities in chronic disease might be underestimated if White participants were more likely to be diagnosed due to greater access to health care (38, 39).
Within the bounds of these limitations, our findings have implications for future research and public health surveillance. More advanced/faster biological aging in Black as compared to White study participants and the potential role of these differences in mediating Black-White health disparities highlight the need for studies of when and how Black-White differences in biological aging arise. Life-course longitudinal studies are needed to establish when in development aging trajectories begin to diverge for Black and White Americans. Studies are also needed to identify life-course phenomena through which racism and socioeconomic resource differentials drive faster aging in Black as compared to White Americans; maternal and perinatal health, social exclusion and victimization of young adults, . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 5, 2021. occupational exposures to young and midlife adults, and access to healthcare later in life all represent potential drivers of Black-White disparities in aging.
Regarding population surveillance, our results show promise for the application of biological-aging measures for evaluating and monitoring Black-White health disparities. Bloodchemistry clocks, as well as the GrimAge clock and Dunedin PoAm pace of aging measures may be well-suited to this purpose. These same measures can, in parallel, provide powerful new outcome measures for evaluation of social policy experiments. A primary application of biological aging measures within the emerging field of geroscience is to provide surrogate endpoints for extension of healthy lifespan (40, 41). Results from this study suggest they may be able to do the same in trials of interventions that aim to eliminate health disparities by repairing inequalities in social determinants of health.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 5, 2021.

Horvath Clock Chronological Age
Age predicted by DNA methylation.

Hannum Clock Chronological Age
Age predicted by DNA methylation.

Second Generation DNA Methylation Clocks
PhenoAge Clock

Bloodchemistry PhenoAge
Age at which average mortality risk in NHANES III matches the mortality risk predicted by the PhenoAge algorithm.
GrimAge Clock Mortality Age at which average mortality risk in the Framingham Heart Study Offspring cohort matches predicted mortality risk. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Figure 1. Analysis of the PhenoAge blood-chemistry measure of biological aging as a mediator of Black-White disparities in healthspan characteristics in the US Health and Retirement
Study. The figure graphs data from participants in the US Health and Retirement Study 2016 Venous Blood Study sub-sample (n=8231). Panel A graphs the proportion of participants with delayed (lowest tertile, light blue), typical (middle tertile, medium blue), and advanced (highest tertile, dark blue) biological aging who scored at least one point on the functional impairment index, reported at least one ADL difficulty, and reported at least one chronic disease diagnosis. Panel B graphs the proportions of White (blue) and Black (red) participants who scored at least one point on the functional impairment index, reported at least one ADL difficulty, and reported at least one chronic disease diagnosis. Panel C graphs distributions of biological-age advancement (measured as the difference between blood-chemistry PhenoAge and chronological age) for White (blue) and Black (red) participants. Panel D graphs the proportion of the Black-White difference in each healthspan characteristic attributable to more advanced biological aging in mediation analysis. Proportions and densities graphed in Panels A-C are adjusted for survey weights. Mediation analysis results in Panel D are adjusted for age and sex.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 5, 2021. ; https://doi.org/10.1101/2021.03.02.21252685 doi: medRxiv preprint Figure 2. Effect-sizes for associations of biological aging measures with healthspan-related characteristics. The figure graphs effect-sizes for associations of biological aging measures with healthspan characteristics. Effect-sizes are incidence rate ratios (IRRs) for 1-SD increases in biological aging measures estimated from Poisson regression. Blood-chemistry measures are shown in red (n=3589 for functional impairments, n=8231 for ADLs, n=8231 for chronic conditions). DNA methylation measures are shown in shades of blue (n=1532 for functional impairments, n=3491 for ADLs, n=3491 for chronic conditions). 1 st generation DNA methylation clocks are shown in dark blue; 2 nd generation DNA methylation clocks are shown in light blue; DunedinPoAm pace of aging is shown in turquoise.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.  . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 5, 2021. ; https://doi.org/10.1101/2021.03.02.21252685 doi: medRxiv preprint Figure 4. Proportion of the total Black-White disparity in each outcome that is mediated through biological-age advancement.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Measurements of Biological Aging
In blood-chemistry analysis, we focused on the recently published "Phenotypic Age" measure (Levine et al., 2018). This measure has been shown to outperform other blood-chemistry indices of biological aging in analysis of mortality and performs as well or better than alternatives in analysis of disability, morbidity, and functional test performance ( (Liu et al., 2019b). Briefly, the Phenotypic Age algorithm was developed from elastic-net Gompertz-regression of mortality on 42 blood-chemistry markers and chronological age in the NHANES III dataset. The resulting algorithm comprised 9 biomarkers (albumin, alkaline phosphatase, creatinine, C-reactive protein (log), glucose, white blood cell count, lymphocyte %, mean cell volume, and red cell distribution width) and chronological age. The elastic-net derived algorithm computes a predicted hazard of mortality, referred to as a mortality score. This mortality score is converted to a biological age based on comparison to predicted hazards from a univariate Gompertz regression of mortality on chronological age. Thus, a participant's Phenotypic Age is the chronological age at which the participant's mortality hazard would match the norm in the original NHANES III sample. For analysis, we calculated the difference between participants' Phenotypic Age and their chronological age. We refer to this measure hereafter as "Biological Age Advancement." The Phenotypic Age blood-chemistry algorithm measures aging from the association of physiological parameters with mortality risk. To evaluate robustness of findings from this particular model of aging, we conducted parallel analysis using two alternative methods to measure biological aging applied to the same biomarker set. The first alternative method is Klemera-Doubal method (KDM) Biological Age (Klemera and Doubal, 2006). This method measures aging from the association of physiological parameters with chronological age. Whereas Phenotypic Age is the age at which a person's predicted mortality risk is at the norm in the reference population, KDM Biological Age is the age at which a person's physiology is at the norm in the reference population. For our analysis, we defined the reference population as NHANES III participants In DNA methylation analysis, we considered five different measures computed by the HRS (Crimmins et al., 2021). Four of these are known as epigenetic clocks, DNA methylation algorithms that produce values that are interpretable as ages. Clock-ages that are older than the chronological age of the person being measured indicate an advanced state of biological aging; clock-ages younger than chronological age indicate delayed aging. We . These clocks are referred to as "second-generation epigenetic clocks." The PhenoAge clock is a DNA methylation prediction of the blood-chemistry Phenotypic Age described above. The GrimAge clock was developed by first defining DNA methylation algorithms to predict several blood proteins, smoking, and chronological age, and then fitting these DNA methylation biomarkers to mortality risk. For analysis, participants' clock-ages were regressed on their chronological ages and residual values were calculated. We refer to these measures hereafter as "clock-age residuals." The fifth DNA methylation algorithm measures pace of aging. Pace of aging is the rate of decline in system integrity that occurs with advancing chronological age. The measure we analyzed, Dunedin PoAm, was developed from analysis of longitudinal change in 18 biomarkers tracking multi-organsystem integrity in a birth cohort followed from age 26-38 years (Belsky et al., 2020). Values are interpretable as years of physiological change occurring per 12-month calendar interval in healthy adults. Dunedin PoAm values above one indicate a faster than normal pace of aging; values below one indicate a slower pace of aging.

Testing sensitivity of mediation results to differences in effect-sizes for associations of biological aging measures with healthspan characteristics between Black and White participants.
Our test of biological aging as a mediator of Black-White health disparities assumed that differences in measured biological aging between Black and White participants reflected substantive differences in biology. That is, we assumed that the observed differences were not an artifact of the biological age measure performing differently in Black and White participants. In a recent analysis we conducted of older adults in North Carolina, biological aging measures showed consistently weaker prediction of disease, disability, and mortality in Black as compared to White participants (Parker et al., 2019). If this difference in the strength of association with healthspan characteristics reflects a difference in precision of biological aging measurements between Black and White American, this would compromise the validity of mediation analysis. To investigate this potential threat to validity, we tested associations of biological aging with healthspan-related characteristics in stratified samples of Black and White participants.
Effect-sizes for biological aging associations with healthspan-related characteristics were slightly larger in White as compared to Black participants, although in most cases differences were not statistically different from zero at the alpha=0.05 threshold (Figure 2b, Supplemental Results Tables S2 and S5). We considered two possible explanations for the small Black-White differences in biological aging associations with healthspan-related characteristics. One explanation is that the algorithms we used to measure biological aging yielded more precise measurements for White as compared to Black participants. This might occur because the reference samples in which the algorithms were developed were mostly or entirely White. However, standard deviations of biological aging measures were similar for Black and White participants (Supplemental Table S1). Moreover, when we tested associations of chronological age with healthspan characteristics, we observed a similar pattern of results; effect-sizes were larger for White as compared to Black participants (Supplemental Table S2). Assuming there are no Black-White differences in measurement precision for chronological age, this result argues against . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 5, 2021. ; Black-White differences in biological age measurement precision as an explanation for effect-size differences. A second possibility is that non-aging causes of healthspan-related characteristics, such as injuries and accidents, were overrepresented in Black as compared to White participants. Such a phenomenon can be statistically represented within our mediation models as exposure-mediator interaction. We tested sensitivity of our results to potential exposure-mediator interactions using the model proposed by Valeri et al. (Valeri et al., 2014).
We formally tested exposure-mediator interactions between race and biological-aging by fitting models including a product term testing the interaction between race/ethnicity and biological aging. We conducted analysis of the three blood-chemistry measures and two DNA methylation measures we tested in mediation analysis. We tested for heterogeneity by evaluating interactions between race and biological aging on multiplicative and additive scales. Multiplicative interaction was tested from the coefficient estimate for the product term the regression models. Additive interaction was tested using the Relative Excess Risk due to Interaction (RERI) method. In this method, the total combined effect of race and biological-age advancement 1 + 2 + 3 is compared with the sum of the individual effects of race and biological-age advancement 1 + 2 . This comparison is made by assessing = ( 1 + 2 + 3 ) − ( 1 ) − ( 2 ) + 1. If the RERI is equal to zero, no additive interaction is present. If the RERI is greater or less than zero, superadditivity and subadditivity are implied.
For all three blood-chemistry measures, there was evidence of interaction for ADL difficulties on both additive and multiplicative scales, but not for functional limitations or for number of chronic conditions reported. For DNA methylation measures, there was some evidence of interaction on the multiplicative scale in the case of the GrimAge-chronic disease association, but not in other analyses. Full effect modification results are reported in the Supplemental Table S5.
We further tested the robustness of our mediational analyses using models including an exposure-mediator interaction term between race and biological-age advancement. Results of these sensitivity analyses were consistent with our original analysis for all of the biological aging measures although, in most cases, the proportion of the black-white disparity mediated by biological aging measures was modestly attenuated in the models that included exposure-mediator interaction (Supplemental Results Table S6).
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.  Epidemiology). The node labeled "Biological Aging" represents the unobserved process of biological aging. The node labeled "Biological Aging*" represents the measured variable in analysis. The measured variable is caused by the unobserved process of aging and measurement error, including systematic measurement error patterned by the exposure (path A) and non-systematic measurement error (represented by U). The DAG identifies two potential causes of differences in effect-sizes between Black and White participants. Path A shows measurement error in the mediator conditional on exposure status (the edge connecting social determinants of health and measured biological aging). Such error could arise, for example, because samples used to develop biological aging measures included few Black participants. Path B shows non-aging causes of healthspan, such as injuries or accidents, that may be patterned across levels of exposure.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Functional limitations were measured from tests of lung function (Peak Flow), grip strength, gait speed, and balance. The tests are described in detail below. We classified those with test scores in at the 30 th percentile of the HRS distribution or lower as having poor performance. Percentiles for grip strength were calculated separately for men and women. We computed functional limitation scores for participants providing data on at least 3 of the 4 tests. We summed the number of tests on which participants demonstrated poor performance and coded the resulting functional limitation score as 0, 1, 2, or 3+ functions with limitation. Peak Flow (L/min) Participants in the physical exam (n = 3743) were administered the breathing test using a Mini-Wright peak flow meter with a disposable mouthpiece (Clement Clarke International Ltd., Harlow, United Kingdom). Three measurements were conducted, each 30 seconds apart. The mean of the three peak flow tests was used as a summary measure of peak flow. Grip Strength (kg) Participants 65+ years of age in the physical exam (n = 3569) were administered the grip strength test. Grip strength was measured with a Smedley spring-type hand dynamometer (Stoelting Company, Wood Dale, Illinois), and results were recorded each time to the nearest 0.5 kg. The maximum dominant hand grip strength recorded across 2 trials was used as a summary measure of grip strength. Walk Speed (ln(seconds)) Participants 65+ years of age in the physical exam (n = 2160) were administered the walking speed test. The test was conducted in a non-carpeted area on a straight walking path 98.5 inches long. The walking speed test was conducted twice-once in each direction. Time to test completion was recorded in seconds, to two decimal places, using a stopwatch. Participants were instructed to use a walking stick or other aid if necessary. The minimum time to complete the walking speed test was used as a summary measure of walk speed. This summary measure was then reverse coded and log-transformed for analysis. Higher scores reflect a faster walking speed Balance All participants in the physical exam (n = 3267) were asked to perform a series of balance tests. Participants were first administered the semi-tandem balance test. Those who did not pass the semi-tandem test were administered the side-by-side test; those who passed were administered the full tandem balance test. In the semitandem test, participants were asked to stand with heel of one foot touching the big toe of the other foot for 10 seconds. In the side-by-side test, participants were asked to stand with their feet together for 10 seconds. In the full tandem test, participants were asked to stand with their feet together for 30 seconds if 65 or older, and 60 if younger than 65. For all tests, participants were told they could use their arms, bend their knees, or move their body to maintain their balance, but not to move their feet. We assigned participants a balance score based on whether they were able to complete the balance tests administered: 1 = did not complete semi-tandem and did not complete side-by-side 2 = did not complete semi-tandem but completed side-by-side 3 = completed semi-tandem but did not complete full tandem 4 = completed semi-tandem and full tandem

Number of ADLs (n = 8231)
Participants (n = 8231) were asked if they had difficulty performing 6 activities of daily living: dressing, eating (such as cutting up your food), bathing and showering, getting in and out of bed, toileting, and walking across a room. We summed the . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 5, 2021. ; https://doi.org/10.1101/2021.03.02.21252685 doi: medRxiv preprint SUPPLEMENT: Testing DNA-methylation and blood-chemistry measures of biological aging in models of Black-White disparities (Graf et al.) 26 number of activities for which participants reported difficulty and coded the resulting ADL score as 0, 1, 2, or 3+ ADLs.

Chronic Disease Diagnoses (n = 8231)
Participants (n = 8231) were asked if they had ever been diagnosed with each of the following six (6) chronic diseases: hypertension, type II diabetes, cancer (excluding minor skin cancer), chronic lung disease, heart problems (heart attack, coronary heart disease, angina, congestive heart failure), and/or stroke. We summed the number of comorbidities reported by each participant to obtain a total comorbidity score. Comorbidity score was then categorized as None, 1, 2, or 3+.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 5, 2021 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 5, 2021. ; https://doi.org/10.1101/2021.03.02.21252685 doi: medRxiv preprint SUPPLEMENT: Testing DNA-methylation and blood-chemistry measures of biological aging in models of Black-White disparities (Graf et al . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 5, 2021. ; https://doi.org/10.1101/2021.03.02.21252685 doi: medRxiv preprint Supplemental Table S5. Tests of Black-White differences in biological aging associations with healthspan characteristics. The table shows results from two tests. The first test is the coefficient for a product term modeling a race*biological aging interaction in Poisson regression for 1-SD increments of the biological aging measures. This coefficient tests the hypothesis that biological age associations with healthspan characteristics differ between Black and White participants on a multiplicative scale. The second test is the Relative Excess Risk due to Interaction (RERI) estimated from the same Poisson regression model for 1-SD increments of the biological aging measures. The RERI tests differences in association on an additive scale. Confidence intervals for RERIs were estimated using the bootstrap method. All models were adjusted for age and sex. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 5, 2021. ; https://doi.org/10.1101/2021.03.02.21252685 doi: medRxiv preprint Supplemental Table S6. Tests of biological aging as a mediator of Black-White differences in healthspan characteristics in models with and without exposure-mediator interactions. Table shows results of mediational analysis using the approach of Vanderweele (2014). For each healthspan characteristic, there are two columns of estimates and 95% Confidence Intervals. The first columns show results from mediational models without exposure-mediator interactions. The second columns show results with exposure-mediator interactions. The first table row shows the controlled direct effect (CDE), second row shows the pure natural direct effect (PNDE), third row shows the total natural direct effect (TNDE), fourth row shows the pure natural indirect effect (PNDE), fifth row shows the total natural indirect effect (TNIE), sixth row shows the total effect (TE), last row shows the proportion mediated (PM). Note that the total effect is equal to the sum of the pure natural direct effect and the total natural indirect effect (PNDE + TNIE), and to the sum of the total natural direct effect and the pure natural indirect effect (TNDE + PNIE).

. Correlations among three blood-chemistry and five DNA-methylation measures of biological aging among Black and White participants in the US Health and Retirement
Study. Biological aging measure labels are listed on the matrix diagonal. Pearson correlations are shown above the diagonal. Correlations are reported for the biological aging measures listed below and to the left of the cell. Scatterplots and linear fits illustrating associations are shown below the diagonal. The Y axis of the plots corresponds to the biological aging measure to the right of the cell. The X axis of the plots corresponds to the biological aging measure above the cell. Sample sizes for correlations among blood-chemistry measures are n=8231. Sample sizes for correlations between blood-chemistry and DNAmethylation measures and among DNA-methylation measures are n=3491.

Panel A. Full Sample
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)