DNA Methylation GrimAge and Incident Diabetes: The Coronary Artery Risk Development in Young Adults (CARDIA) Study

Worldwide, more than 550 million people are projected to have diabetes by 2030 (1). Older age is one of the principal risk factors for type 2 diabetes (T2D) development (2–4). However, the utility of chronological age in epidemiologic studies is limited by variability of age-related changes in biological functioning, prompting the study of alternative age measurements that might better reflect this biological degeneration (5,6). Measurements that can more precisely reflect an individual’s biological aging can help improve our understanding of the age-related development of T2D.

DNA methylation (DNAm)-based biological age (epigenetic age) has been suggested as a useful biomarker of age-related conditions, since abnormal patterns in DNAm increase and accumulate with exposures over time represented by age. To date, several epigenetic age measurements have been proposed (7–9), including the two most recent measurements, Levine’s phenotypic age estimator “DNAm PhenoAge” and GrimAge (10,11). DNAm PhenoAge is constructed with 10 clinical markers selected from National Health and Nutrition Examination Survey (NHANES) III data that most contribute to mortality, comorbidity, and physical disability (11). GrimAge, so named because it was developed as a predictor of mortality, is estimated from 12 DNAm surrogate markers comprising adrenomedullin, β-2-microglubulin (B2M), CD56, ceruloplasmin, cystatin C, EGF-containing fibulin-like extracellular matrix protein 1, growth differentiation factor-15 (GDF-15), leptin, myoglobin, serum paraoxonase/arylesterase 1, tissue inhibitor metalloproteinases-1 (TIMP-1), and plasminogen activator inhibitor-1 (PAI-1), as well as smoking pack-years (10).

Despite the emerging interest in epigenetic age, associations of more advanced epigenetic age measurements (DNAm PhenoAge and GrimAge) with T2D have not been assessed in epidemiologic research—or their interplay with obesity. Considering the multifactorial etiology of T2D, DNAm PhenoAge and GrimAge could be useful tools to measure the association with or mechanisms of T2D, since these novel epigenetic age estimators outperform prior versions of epigenetic age in the prediction of multiple morbidities (10,11). Most notably, the components of GrimAge have been investigated in associations with traits involved in T2D development, such as the role of higher BMI, elevated CRP levels, and increased PAI-1 activity in overweight and obese women (12–16). In this context, assessing GrimAge as a risk factor of T2D may have useful public health applications.

Obesity is one of the greatest risk factors for T2D, and previous research showed links between obesity and T2D through pathways such as impaired insulin metabolism and inflammatory factors (12). Cumulative obesity, measured as duration of obesity, was also associated with increased risk of T2D in previous studies (13,14), which suggests a role of aging and/or cumulative obesity burden in the development of T2D. Associations between obesity and epigenetic age acceleration in tissues such as blood and liver have also been reported (7,11,15). However, the relationships among obesity, DNAm-based biomarkers, and T2D are still not fully understood.

In this study, we investigated the associations between DNAm-based epigenetic age measurements and incidence of T2D and prediabetes among participants in the Coronary Artery Risk Development in Young Adults (CARDIA) study, as well as potential mediation effects with duration of obesity. By exploring the associations between epigenetic age and T2D, we also aimed to expand our knowledge of the epigenetic aspects of age-related health in a more integrative view.

The study sample was randomly selected from the CARDIA study, a multicenter prospective observational study examining the development and determinants of cardiovascular disease including 5,115 Caucasian and African American adults aged 18–30 years at enrollment (1985–1986). Four field centers enrolled participants: 1) the total community in Birmingham, AL; selected census tracts in 2) Chicago, IL, and 3) Minneapolis, MN; and 4) participants from the Kaiser Permanente Health Plan membership in Oakland, CA. To date, examinations have been completed for nine study points, including baseline and follow-up, in study year 0 (Y0), Y2, Y5, Y7, Y10, Y15, Y20, Y25, and Y30. More details about the CARDIA study have previously been published (16). Among 5,115 total participants, 3,672 participants returned for exams at Y15 and 3,549 returned at Y20, respectively. Participants were randomly selected for DNAm measurements (1,089 at Y15 and 1,092 at Y20). In total, we included 1,118 participants who had at least one epigenetic age measurement at Y15 and Y20. The characteristics between participants who did and did not receive DNAm profiling have previously been described, with fewer Black and female individuals in CARDIA's methylation data (17). We excluded participants who ever had bariatric surgery by Y15 from our analysis, considering potential large change of BMI in those participants. We further restricted this sample to exclude participants with prevalent diabetes at Y15.

The participants were further classified into three groups by BMI status at Y15 as normal weight (BMI <25 kg/m²), overweight (BMI 25–30 kg/m²), or with obesity (BMI ≥30 kg/m²) based on the guidelines from the National Heart, Lung, and Blood Institute (NHLBI) (18). All participants provided written informed consent at Y15 and Y20 to DNAm profiling. All study protocols were approved by the institutional review boards of participating institutions.

Among 3,672 and 3,549 individuals in CARDIA who attended examination at Y15 and Y20, respectively, we randomly selected 1,200 individuals for DNAm profiling. After exclusion of individuals with low DNA amount or poor DNA quality, DNAm measurements were obtained in a total of 2,181 blood samples from Y15 (n = 1,089) and Y20 (n = 1,092), and Infinium MethylationEPIC BeadChip raw data (IDAT files) were generated.

We used the R package minfi (19) to load the data and the R package ENmix (20) to conduct quality control (QC). For QC, we excluded 6,209 CpGs with low detection rate (<95%). Eighty-seven samples with >5% of methylation measurements low quality, or bisulfite conversion probes of extremely low intensity (<3 SDs from the mean), were also excluded. After QC, we further excluded 95 outlier samples with Tukey’s method (21).

We used the naive preprocessIllumina function in the minfi package (19) for the remaining samples to perform a background correction and a standard control normalization. Our final analytic methylation data included 860,627 CpG probes and 1,999 samples (n = 1,042 for Y15 and n = 957 for Y20, respectively) from 1,118 individuals.

Incident T2D during Y15–Y25 and Y15–Y30 was determined by a participant meeting at least one of the following criteria: fasting glucose ≥126 mg/dL, glucose ≥200 mg/dL 2 h after a glucose load (2-h glucose tolerance test was performed at Y10, Y20, and Y25 and, only in women, at Y30 in CARDIA examination), HbA_1c ≥6.5% (≥48 mmol/mol) (HbA_1c was measured at Y20 and Y25 for both sexes and for women only at Y30 in CARDIA examination), or history of diabetes treatment. We defined prediabetes at Y25 and Y30 as impaired fasting glucose (100–125 mg/dL), HbA_1c 5.7–6.5% (39–48 mmol/mol), or impaired glucose tolerance (140–199 mg/dL for 2 h after a glucose load). If a participant had diabetes only during pregnancy (gestational diabetes mellitus), the participant was still included in the analysis but was counted as not having diabetes (n = 6 during Y15–Y25 and n = 11 during Y15–Y30). To explore the associations between epigenetic age and stages of diabetes progression, we investigated the associations using different reference groups for each time point: 1) normal glucose only, compared with prediabetes and diabetes; 2) normal glucose only, compared with overt diabetes (prediabetes excluded); and 3) normal glucose and prediabetes, compared with overt diabetes.

In the CARDIA study, participants’ body measurements were obtained at Y0, Y2, Y5, Y7, Y10, Y15, Y20, Y25, and Y30 under a standardized protocol to the nearest 0.2 kg for body weight and 0.5 cm for height. BMI was calculated as weight in kilograms divided by the square of height in meters.

Earlier onset of obesity is known to increase the risk of chronic diseases including T2D in later life (22). For an additional exploratory analysis, we calculated years since first obesity as a measure of cumulative obesity for 318 individuals who were obese at Y15 from their BMI records at Y0, Y2, Y5, Y7, Y10, and Y15. We assumed that individuals became obese, they were obese throughout the study period until Y15; this assumption was supported in our supplementary analyses, which showed that 98.6% of people who were obese at Y0 remained obese at Y15 (data available upon request). Under this assumption, we assigned years since first obesity according to the study year when the participant was first classified with obesity. For example, if a participant was classified as obese (BMI ≥30 kg/m²) at Y0, we assigned the person 15 years for years since first obesity. For participants who were not obese at Y0 but became obese at Y2, we assigned 13 years for years since first obesity. Based on this logic, each participant had a specific value among 0, 5, 8, 10, 13, or 15 years for years since first obesity.

We calculated GrimAge using DNAm data from Y15 and Y20 based on the published algorithms (10,11) and Horvath’s online calculator (https://dnamage.genetics.ucla.edu). GrimAge acceleration (“GrimAA”), which represents a deviation of epigenetic from chronological age, was calculated from the residuals of the linear regression of GrimAge on chronological age with adjustment for blood cell type proportions. We also analyzed it as a binary variable (“positive GrimAA”) for whether GrimAA exceeded 0 (faster epigenetic age acceleration; epigenetic age older than chronological age) or did not (slower epigenetic age acceleration). To investigate the change of GrimAge measurements from Y15 to Y20, we generated a binary variable (“positive 5yr GrimAA”) that represented whether the 5-year changes in GrimAge were >5 (the chronological age change from Y15 to Y20). While GrimAA reflects GrimAge acceleration measured at a single time point, 5yr GrimAA represents the longitudinal change in GrimAge acceleration from Y15 to Y20. Similarly, we generated variables with DNAm PhenoAge (“PhenoAA,” “positive PhenoAA,” and “positive 5yr PhenoAA”).

We included the following potential confounders in our final models: race, sex, education, smoking status, physical activity (total activity score), heavy alcohol consumption (>14 drinks per week), and field center. All covariates included in the final models were measured at Y15 and Y20 in the CARDIA study. We included family history of T2D and diet quality scores represented by the empirical dietary inflammatory index (EDII) (23) in the descriptive analysis; however, the two variables were not included in final multivariable models, since they were not measured at Y15 or Y20. To compare family history of T2D and EDII by T2D status in descriptive analyses, we used those variables measured at Y25. Chronological age was only included in the generalized estimating equation (GEE) models with DNAm PhenoAge, not with GrimAge, because it is already a component of GrimAge estimates along with other DNAm-based surrogate markers (10). GrimAge estimates and chronological age showed significant correlations in our data (shown in Supplementary Table 1), which supported our decision not to include chronological age as a covariate in the models with GrimAge. We did not include chronological age in the models with six epigenetic age measures—GrimAA, positive GrimAA, positive 5yr GrimAA, PhenoAA, positive PhenoAA, and positive 5yr PhenoAA—since those measures represented age-adjusted versions of GrimAge and DNAm PhenoAge by definition (10, 11).

We performed descriptive analyses to compare demographic and clinical characteristics, including epigenetic age measurements of study participants according to BMI status at Y15, using Student t tests and χ² tests for continuous and categorical variables, respectively.

To investigate the associations between epigenetic age and T2D, we used GEE models. We performed these analyses with three different T2D classifications as follows: 1) model 1 with diabetes versus nondiabetic (prediabetes and normal glucose), 2) model 2 with diabetes versus normal glucose (excluding prediabetes), and 3) model 3 with hyperglycemia (diabetes or prediabetes) versus normal glucose. We constructed each model with GrimAge, GrimAA, positive GrimAA, and positive 5yr GrimAA. All models were stratified by BMI status (normal weight, overweight, and obese) at Y15 for the purpose of this study. We treated all covariates as time varying except for race, sex, and field center to reflect the potential fluctuations between Y15 and Y20.

We performed subgroup analyses for T2D at Y25, with stratification by race and sex. The subgroup analyses were performed only for model 3 (hyperglycemia; diabetes and prediabetes combined as the outcome) due to sparse data concerns. To investigate race- and sex-specific interactions, we tested multiplicative interactions in the models by adding a product term with race and each epigenetic age measure. We also investigated the associations between epigenetic age and T2D status at Y30 by following the same steps with analyses for T2D at Y25, except the subgroup analyses. We performed sensitivity analyses by adjusting for chronological age and BMI separately in the models to assess potential variabilities by the two factors. We also performed receiver operating characteristics (ROC) curve analysis to compare the area under the curve between chronological age and epigenetic age for assessment of the utility of epigenetic age in diabetes prediction over chronological age. SAS, version 9.3 (Cary, NC), was used for descriptive analyses, multivariable logistic regression with GEE, and the ROC curve analyses. All GEE analyses were stratified by obesity status at Y15.

Finally, we further explored the role of epigenetic age in the development of T2D using mediation analysis. We investigated whether epigenetic age played a role as a mediator in relationships between the years since first obesity and T2D development. We used the R package mediation (24) for the mediation analysis and obtained the P value for mediation effects of epigenetic age as well as the effect size of obesity duration on T2D and proportion mediated by epigenetic age. Proportion mediated was calculated as the ratio of indirect effect (effect of years since first obesity on diabetes through epigenetic age) to the total effect (sum of direct effect and indirect effect, where direct effect is effect of years since first obesity on diabetes).

Data for the current study can be obtained upon request. Protocols for data request and manuscript proposal are available from https://www.cardia.dopm.uab.edu/.

Supplementary Fig. 1 shows the inclusion and exclusion criteria that led to a total of 1,057 participants. Of these, 416 (39.4%) developed prediabetes and 89 (8.4%) developed diabetes during the Y15–Y25 period (incidence over 10 years). By Y30, 330 (34.1%) developed prediabetes and 115 (11.9%) developed diabetes during the Y15–Y30 period (incidence over 15 years); 90 participants were no-shows to exams, which was treated as missing data. In our analysis, the proportion of participants who developed diabetes was lower than the proportion who developed diabetes in the whole CARDIA population (25,26) because we excluded diabetes incidence between Y0 and Y15 due to the timing of DNAm measurements (Y15).

Table 1 shows the characteristics of study participants by BMI status at Y15. Participants’ age, smoking status, and distribution of study center were comparable across the three groups. As participants’ BMI increased, they tended to have less education, less physical activity, an unhealthy dietary pattern, and higher levels of fasting glucose and fasting insulin levels. We also observed a linear trend between higher proportions of people whose GrimAge is higher than chronological age (positive GrimAA) and BMI status. Interestingly, 5-year change in GrimAA did not differ by BMI group. As expected, participants who had obesity at Y15 were the group with the highest proportion developing diabetes between Y15–Y25 and Y15–Y30 (Table 2).

Table 3 shows the associations between GrimAge at Y15–Y20 and T2D status at Y25. Compared with chronological age, GrimAge measurements showed more striking results among participants with obesity. In model 1, comparing participants who developed T2D at Y25 with those who did not (i.e., prediabetes and normal glucose), GrimAge was associated with higher odds of T2D development in participants with obesity (odds ratio [OR] 1.06, 95% CI 1.01–1.11). In the same model, positive GrimAA (older GrimAge than chronological age) was associated with higher odds of T2D development in participants with obesity (OR 2.57, 95% CI 1.61–4.11). In Model 2 (prediabetes excluded), GrimAge was associated with higher odds of T2D development in participants with obesity (OR 1.08, 95% CI 1.03–1.14). In this model, positive GrimAA and positive 5yr GrimAA were associated with greater odds of developing T2D in participants with obesity (positive GrimAA OR 2.49, 95% CI 1.44–4.31, and positive 5yr GrimAA OR 2.35, 95% CI 1.40–3.94), whereas positive 5yr GrimAA was associated with lower odds of developing diabetes among participants who were overweight (OR 0.38, 95% CI 0.10–0.89). In model 3, comparing hyperglycemia with normal glucose, GrimAge was associated with increased odds of hyperglycemia in participants with normal weight and in participants with obesity (ORs ranged from 1.06 to 1.07, and 95% CIs ranged from 1.02 to 1.11). In terms of accelerated GrimAge, positive GrimAA was associated with hyperglycemia among participants with normal weight (OR 2.15, 95% CI 1.35–3.42) and positive 5yr GrimAA was associated with higher odds of developing hyperglycemia in participants with obesity (OR 1.90, 95% CI 1.31–2.76). We performed race- and sex-stratified analyses for incidence of diabetes or prediabetes at Y25 (Supplementary Tables 2 and 3). In our stratified analyses with race and sex, we did not find remarkable deviations from model 3 in the main analysis.

Table 4 shows the associations between GrimAge at Y15–Y20 and T2D status at Y30. In this analysis, the association of chronological age with diabetes only appeared among participants with normal weight. In model 1, positive GrimAA was associated with higher odds of developing T2D in participants with obesity (OR 1.71, 95% CI 1.12–2.60). Positive 5yr GrimAA was also associated with greater odds of developing T2D in participants with obesity (OR 1.86, 95% CI 1.24–2.79). In model 2, positive GrimAA was associated with higher odds of developing T2D (OR 2.10, 95% CI 1.29–3.41) and positive 5yr GrimAA was associated with higher odds of developing T2D in participants with obesity (OR 1.90, 95% CI 1.21–2.98). In model 3 for hyperglycemia versus normal glucose, positive GrimAA and positive 5yr GrimAA were associated with higher odds of developing hyperglycemia among participants with normal weight (ORs ranged from 1.54 to 1.60, and 95% CIs ranged from 1.01 to 2.56) and positive 5yr GrimAA was associated with higher odds of developing hyperglycemia among participants with overweight (OR 1.65, 95% CI 1.21–2.25).

In sensitivity analyses with additional adjustment for chronological age and BMI, respectively, we observed no substantial differences (data not shown). Additionally, DNAm PhenoAge was generally not associated with T2D, and the results are shown in Supplementary Tables 4 and 5. In the ROC curve analysis, epigenetic age generally showed greater area under the curve in comparison with chronological age, as shown in Supplementary Table 6.

In our secondary analysis, for evaluation of the extent to which GrimAge may mediate the statistical association between the years since first obesity and diabetes, GrimAge at Y15–Y20 demonstrated its role as a mediator of development of diabetes or prediabetes in Y15–Y25. Figure 1 shows the directed acyclic graphs with the years since first obesity, GrimAge, and diabetes outcomes in models 1–3. Years since first obesity showed associations with T2D development across all the outcome models (model 1, β = 0.07, 95% CI 0.03–0.11; model 2, β = 0.09, 95% CI 0.05–0.14; model 3, β = 0.06, 95% CI 0.01–0.09). GrimAge mediated associations between years since first obesity and hyperglycemia in model 3 (proportion mediated = 8.0%).

In this study, we investigated the associations between novel epigenetic age measurements (GrimAge and DNAm PhenoAge) and incident T2D in the CARDIA study. We observed that higher GrimAge, as well as accelerated GrimAge, were associated with elevated risk of T2D, particularly in participants with obesity. We found that both cross-sectional (positive GrimAA) and longitudinal (positive 5yr GrimAA) acceleration were associated with increased odds of T2D development in participants with obesity. In our mediation analysis with participants with obesity at Y15, there was evidence that GrimAge plays a role as a mediator in the pathway between the duration of obesity and T2D development. While accelerated GrimAge was strikingly associated with higher odds of developing T2D among participants with obesity, we observed inverse associations among overweight participants. The different direction of association might be due to unmeasured confounding including behavioral and clinical changes between the measurement of DNAm and development of diabetes. The relatively small number of participants who developed diabetes among overweight participants also makes the overweight subgroup vulnerable to unmeasured confounding. In addition, the most remarkable result shown in the overweight stratum was with positive 5yr GrimAA, which was the broadest measure of GrimAge calculated in this study. These series of marginal effects might combine and induce false positives in the overweight subgroup. As CARDIA’s ongoing data collection includes DNAm, future study is needed to elucidate the association between epigenetic age and T2D among overweight individuals.

Previous studies have identified altered methylation among T2D patients compared with people with normal glucose concentrations (27–31). However, research in this area has tended to focus on localized methylation measures rather than summary measures such as GrimAge that may be more useful biomarkers of the aging process. Only a small number of studies examined epigenetic age in association with risk of T2D and related traits such as glucose and insulin (32,33). However, prior reports showed that the acceleration of epigenetic age tends to be more stable over time than epigenetic age measures themselves (32,34–36), which is in line with our results. In our study, the strongest associations were observed between accelerated GrimAge (GrimAge acceleration vs. GrimAge deceleration) and T2D during Y15–Y25. The association of older GrimAge with incident T2D during Y15–Y30 was attenuated relative to that with incident T2D during Y15–Y25, whereas any accelerated GrimAge remained significant, although the magnitudes of effect were smaller. The longitudinal relationship between accelerated GrimAge and T2D at Y25 and Y30 in our analysis strengthens the potential utility of GrimAge in identifying epigenetic risk of T2D.

With DNAm PhenoAge we observed similar but weaker associations with T2D in comparison with GrimAge. This may be due to the different components of the two epigenetic age measurements. The components of DNAm PhenoAge include “phenotypic” indicators comprising disease status as well as demographic and behavioral factors. On the other hand, GrimAge is a composite biomarker consisting of blood DNAm surrogates for smoking pack-years and plasma proteins associated with various age-related conditions (10,11). As underlying components are different for GrimAge and DNAm PhenoAge, CpGs included in the calculation process for the two epigenetic age measurements do not overlap very much (10,11).

We found consistent effect sizes for epigenetic age measures across our models 1–3. This suggests that epigenetic age might be a useful biomarker for prediabetes and/or overt diabetes. Although it is not possible to calculate the sensitivity or specificity of GrimAge for prediabetes/diabetes due to the absence of a clinically actionable threshold, our results suggest that GrimAge can still play a role as a useful biomarker for T2D. The associations between GrimAge and diabetes are also biologically plausible, as supported by previously reported associations between the components of GrimAge and diabetes. Higher BMI is associated with increased CRP concentrations (37), and elevated concentrations of CRP play a role in the development of T2D (38,39). Elevated PAI-1 is also observed in obese patients and individuals with T2D (12). In a study focused on T2D in overweight and obese women, visceral fat was significantly correlated with increased PAI-1 activity, even independent of insulin and triglycerides (40).

The concept of GrimAge, as well as other epigenetic age measurements, is based on the idea that aberrant DNAm accumulates over time, and age-associated DNAm patterns overlap with the changes of DNAm in age-related diseases. For T2D, animal studies have reported associations of age-related epigenetic changes with islet dysfunction and hyperglycemia (41,42). One study in humans also showed associations of changes in age-related epigenetic biomarkers with insulin secretion and diabetes (43). In cancer, previous studies have suggested that epigenetic changes play a role in the malignant transformation of cells by mimicking somatic mutations (44). Coupled with these previous findings, our results suggest that the primary use of GrimAge would be as a “front-line” biomarker for multiple age-related diseases. Future study is needed for application of epigenetic age measurements, including GrimAge, in clinical settings.

We also found statistical evidence that GrimAge mediated the associations between the years since first obesity and development of diabetes, which implies that GrimAge may serve as a useful biomarker of obesity-related diabetes development. It is known that being obese earlier in life increases the risk of chronic diseases including T2D (22). However, few studies have examined a potential epigenetic component to this relationship, including epigenetic age measures. Our mediation findings suggest that more years since first obesity might increase the risk of elevated glucose concentrations in later life via altered epigenetic mechanisms related to T2D, represented by GrimAge in our study. This may have implications for precision medicine, targeting people who have been obese for longer durations.

Our study has several limitations. We assumed that once participants became obese, they stayed obese; this assumption was supported, with the substantial proportion of participants (98.6%) who remained obese throughout Y0–Y15. In addition, the variable for years since first obesity was obtained according to the follow-up measurements for each study time point (Y0, Y2, Y5, Y7, Y10, and Y15) in CARDIA. Therefore, the observed associations among years since first obesity, GrimAge, and T2D might be biased by only taking measurements every 2–5 years; however, we believe that this would bias our results toward the null, as every obese participant had an artificial maximum years since first obesity of 15 years with our study’s time range. In addition, DNAm was measured at Y15 and Y20 in CARDIA; therefore, incident diabetes cases between Y0 and Y15 were not included in the current analyses. The exclusion of early-onset diabetes might also explain the lack of associations by race, considering the higher incidence of early-onset diabetes in Black individuals. Indeed, one study showed longer duration of diabetes among Black individuals at the CARDIA Y25 examination, which was in line with our own conclusion (45). Therefore, our race-specific results should be interpreted with caution and also suggest a need for future research to examine the comprehensive associations between epigenetic age and onset of diabetes throughout the life span. Last but not least, there may be residual confounding due to an inability to capture the change in behaviors or health conditions between the study time points; however, we believe it induced nondifferential misclassification across the study participants and biased the results toward the null. As CARDIA accumulates additional data points (including DNAm) in the future, this will be another area for follow-up research.

In summary, we found that older, and accelerated, GrimAge was associated with incident T2D in CARDIA participants with obesity. We anticipate that our results support the potential for GrimAge as a global biomarker in age-related diseases, including diabetes.

Funding. The Coronary Artery Risk Development in Young Adults (CARDIA) study is conducted and supported by the NHLBI in collaboration with the University of Alabama at Birmingham (HHSN268201800005I and HHSN268201800007I), Northwestern University (HHSN268201800003I), University of Minnesota (HHSN268201800006I), and Kaiser Foundation Research Institute (HHSN268201800004I). CARDIA is also partially supported by the Intramural Research Program of the National Institute on Aging (NIA) and an intra-agency agreement between NIA and NHLBI (AG0005). The data collection was supported by the National Institute of Digestive Diseases and Diabetes (R01 DK106201 [Kaiser Permanente Northern California; Principal Investigator, E.P.G.]), and laboratory work and analytical components were funded by the American Heart Association (17SFRN33700278 and 14SFRN20790000 [Northwestern University; Principal Investigator, L.H.]). The manuscript was reviewed by CARDIA for scientific content.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. P.J.S., J.M.S., and E.P.G. collected data. K.K., B.T.J., and Y.Z. developed an analysis plan, interpreted data, and prepared the manuscript. E.P.G. and L.H. designed the study. K.K. generated study hypotheses, performed data analysis, and wrote the manuscript. P.J.S., D.R.J., J.M.C., J.M.S., M.R.C., P.G., L.V.V.H., E.P.G., and L.H. participated in analysis and manuscript writing. N.B.A. and D.M.L.-J. participated in manuscript preparation. L.H. developed the conceptual framework and laboratory component and oversaw manuscript preparation. All authors participated in editing the manuscript and approved the final version. L.H. and E.P.G. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.