Interaction Between Physical Activity and Polygenic Score on Type 2 Diabetes Mellitus in Older Black and White Participants From the Health and Retirement Study

Abstract

This study investigated the association of lifestyle factors and polygenic risk scores (PGS), and their interaction, on type 2 diabetes mellitus (T2D). We examined data from the U.S. Health and Retirement Study, a prospective longitudinal cohort of adults aged 50 years and older, containing nationally representative samples of Black and White Americans with precalculated PGS for T2D (N = 14 001). Predicted prevalence and incidence of T2D were calculated with logistic regression models. We calculated differences in T2D prevalence and incidence by PGS percentiles and for interaction variables using nonparametric bootstrap method. Black participants had approximately twice the prevalence of Whites (26.2% vs 14.2%), with a larger difference between the 90th and 10th PGS percentile from age 50 to 80 years. Significant interaction (p_interaction = .0096) was detected between PGS and physical activity among Whites. Among Whites in the 90th PGS percentile, T2D prevalence for moderate physical activity was 17.0% (95% CI: 14.8, 19.6), 6.8% lower compared to no/some physical activity (23.8%; 95% CI: 20.4, 27.5). T2D prevalence was similar (~10%) for both groups in the 10th PGS percentile. Incident T2D in Whites followed a similar pattern (p_interaction = .0325). No significant interactions with PGS were detected among Black participants. Interaction of different genetic risk profiles with lifestyle factors may inform understanding of varying inventions’ efficacy for different groups of people, potentially improving clinical and prevention interventions.

Type 2 diabetes, from here on referred to as “diabetes,” is a complex chronic disorder influenced by genetics and environmental factors (eg, nongenetic factors such as lifestyle, education, income, etc.) (1,2). In the United States, diabetes prevalence increases with age and is highly prevalent among older adults (≥65 years), with 27% having this condition (3). Diabetes is the leading cause of kidney failure, lower-limb amputations, and blindness and is the seventh leading cause of death in the United States (3). It also accounts for more than 20% of health care spending, or over $327 billion in total estimated cost in 2017 alone (4). Current trends predict that as many as 1 in 3 Americans could be diagnosed with diabetes by 2050, posing an even greater burden on those directly and indirectly affected by the disease and on the health care system (3).

Both modifiable lifestyle factors and genetic predisposition play important roles in the risk of diabetes. While a large proportion of existing epidemiologic literature centers around identifying modifiable risk factors, such as unhealthy diet (5), physical inactivity (6), and elevated alcohol consumption (7), heritability studies point to a strong underlying genetic contribution. For example, the offspring of one parent with diabetes faces a 40% lifetime risk of the condition, and 70% lifetime risk when both parents have diabetes (8). Moreover, individuals with certain genetic variants have approximately 20% higher risk compared to those who lack those variants (9). To date, genome-wide association studies (GWAS) have identified 243 loci, or over 400 distinct association signals, that are associated with diabetes, including independent risk alleles found within the coding regions of TCF7L2, INS, IGF2, KCNQ1, and CDKN1C (10). However, the attributable risk from each individual allele is often modest and of little clinical significance (11). For instance, a study utilizing the Diabetes Prevention Program data (a 27-center randomized clinical trial study) found that each risk allele was associated with only a 2% increased risk (hazard ratio [HR] = 1.02; 95% confidence interval [CI]: 1.00, 1.05) of diabetes relative to those without the allele (12).

Given the strong evidence of heritability and modest monogenic effect estimates, the genetic risk of diabetes likely reflects the combined effects of multiple risk alleles; thus, combined polygenic risk scores (PGS) may provide a much stronger basis for examining genetic predisposition (13). Moreover, environmental exposures, such as modifiable behaviors, may act, not only as independent predictors, but as catalysts in the development of diabetes among genetically predisposed individuals (9). Recently, studies have examined the interplay between PGS of diabetes and several environmental factors, especially physical activity (PA) and dietary behaviors (14). For instance, Klimentidis et al. (15) reported that the protective association between PA and incident diabetes was weaker among participants with a higher 69-SNP PGS, relative to those with a lower PGS, suggesting that greater genetic risk may hinder one’s ability to leverage PA in preventing diabetes development. Furthermore, the prevalence and impact of risk factors and diabetes differ across population groups (16,17). While epidemiologic studies report higher rates of diabetes among many racial/ethnic minority groups, including Black Americans, these groups are critically understudied in the context of both GWAS and gene–environment interaction studies of diabetes, compared to Whites (18,19).

To contribute to the growing body of literature on understudied populations, we used data from the Health and Retirement Study (HRS). The HRS offers an opportunity to examine gene–environment interactions utilizing nationally representative samples of both Black and White Americans with calculated PGS for diabetes. Few studies utilizing HRS data examine the interplay of genetic and environmental factors associated with both incident and prevalent diabetes cases; as a disease associated with a long survival time, it is important to examine both prevalence and incidence of diabetes to more accurately detect the predictive effects of various factors. Thus, this study aims to investigate the independent and interaction effects of environmental factors and polygenic risk of both prevalent and incident diabetes.

Method

The Health and Retirement Study

The HRS is a nationally representative longitudinal cohort study aimed at examining the health outcomes of approximately 43 000 U.S. men and women older than 50 years of age at recruitment, and their spouses. It is administered by the Institute for Social Research at the University of Michigan and sponsored by the National Institute on Aging (NIA U01AG009740) (20,21). The HRS sample consists of 7 continuing cohorts including the initial HRS cohort (born 1931–1941 and recruited 1992), Asset and Health Dynamics Among the Oldest Old (born 1890–1923 and recruited 1992), the Children of the Depression (born 1924–1930 and recruited 1998), the War Babies (born 1942–1947 and recruited 1998), Early Baby Boomers (born 1948–1953 and recruited 2004), Mid Baby Boomers (born 1954–1959 and recruited 2010), and Late Baby Boomers (born 1960–1965 and recruited 2016). HRS conducted face-to-face or phone core interviews, during which participants were asked questions about finances, health status and behaviors, marital/family status, and social support systems. A random half-sample was then followed-up biennially for core interviews. Starting in 2006, HRS has utilized a mixed-model design for follow-up in which a random half of the sample is assigned an Enhanced Face-to-Face Interview (EFTF) during which saliva was obtained for DNA extraction. The other half completes only the core interview, usually by telephone. The half-samples alternate waves so longitudinal information from the face-to-face interview is available every 4 years at the individual level, and the expanded content is available every wave on a nationally representative half-sample.

Our study utilized 7 waves of HRS data collected between 2004 and 2016 because PA and alcohol consumption were measured consistently since the 2004 wave. The derivation of the final analytic samples for the prevalence and incidence of diabetes is presented in Figure 1. Genotyped data are currently available for 15 190 participants (12 090 White and 3100 Black participants) who provided a saliva sample between 2006 and 2012. After excluding missing diabetes status, age (≤50 years), and missing covariates (sex, education, body mass index [BMI], smoking, alcohol, and PA), the analytic sample for this study included 14 001 participants (11 178 White and 2823 Black). The prevalence sample included participants with and without diabetes at the time of their first visit between 2004 and 2012 (78.0% in 2004, 18.5% in 2010, and 3.5% in 2006, 2008, or 2012). The incidence subset included participants without diabetes during their first visit and had subsequently been assessed for diabetes status during their follow-up visit (between 2006 and 2016). That is, participants had a maximum of 12 years of follow-up, with an average follow-up period of 10 years among White participants and 7 years among Black participants. After excluding 2329 participants with baseline diabetes and 37 participants who lacked follow-up diabetes status, the total number of participants for follow-up incidence analysis was 11 635 (9562 White and 2703 Black). Our study used data that are publicly available and free of personal identifiers. The current study is considered exempt and received Institutional Review Board approval from the University of Hawaiʻi (CHS# 23349).

Figure 1.

Analytical sample derived from 15 190 HRS genotyped participants. PGS = polygenic score; T2D = type 2 diabetes mellitus.

Case Ascertainment for Diabetes and Covariates

Persons with diabetes were defined by a self-report of doctor-diagnosed diabetes or hemoglobin A1C level of ≥6.5, excluding those with information on diabetes before 16 years of age (~0.2% of the entire HRS cohort). Among all diabetes cases, 93.6% were defined by self-report of doctor-diagnosed diabetes. Variable selection was based on prior associations observed with diabetes. These included self-reported race (White or Black), age (≥50 years), sex, education (less than high school, high school/Generalized Education Development, some college, or college and above), height and weight (used to calculate BMI), and alcohol consumption (never, once, twice, or ≥3 times per day). High alcohol consumption has been shown previously to increase the risk for diabetes (22). Physical inactivity has been linked as an independent risk factor for diabetes (23). In our analysis, PA was calculated using self-reported activity frequency: never or some activity, 2 or more light (vacuuming, laundry, home repairs), 2 or more moderate (gardening, cleaning the car, walking at a moderate pace, dancing, floor or stretching exercises), and 2 or more vigorous activity (running or jogging, swimming, cycling, aerobics or gym workout, tennis, or digging with a spade or shovel) per week. Poverty ratio, provided by HRS, is calculated based on threshold levels used by the U.S. Census Bureau. Previous studies show that individuals with lower income are more likely to have diabetes compared to those with higher incomes (24). The current smoking status of participants was also noted for never smokers, past smokers, and current smokers. Smoking, especially for heavy and current smokers, increases the risk of incidence diabetes (25).

Polygenetic Score Computation

Genotyping was conducted by the Center for Inherited Disease Research in 2011, 2012, and 2015 (RC2 AG0336495 and RC4 AG039029) using llumina HumanOmni2.5 BeadChips (HumanOmni2.5-4v1, HumanOmni2.5-8v1, HumanOmni2.5-8v1.1), which measured more than 2 million SNPs. The HRS-PGS was calculated by the HRS team using all available SNPs that overlap between the GWAS meta-analysis through the DIAbetes Genetics Replication And Meta-analysis (DIAGRAM) Consortium (26) and the HRS genetic data. Research has concluded that including all available SNPs in a PGS demonstrated the largest predictive power (27). The PGS for Whites contain 726 395 SNPs that overlapped between the HRS genetic database and the GWAS meta-analysis; PGS for Blacks contains 734 890 SNPs. Weighted sums were chosen to calculate the PGS. Weights were defined by the beta estimate from the stage 1 GWAS meta-analysis (26). If the beta value from the GWAS meta-analysis was negative, the beta measures were converted to positive values and the reference allele flipped to represent phenotype-increasing PGS. HRS also provided ancestry-specific principal components to account for potential population stratification and possible ancestry differences in genetic background. We scaled PGS to percentiles that ranged from 0 to 1.

Statistical Analysis

All analyses were stratified by race. Descriptive statistics were used to summarize baseline characteristic for the full sample and the subsample excluding diabetes cases at baseline. In this paper, we present both predicted prevalence and incidence, and prevalence ratio (PR) and relative risk (RR) of diabetes. The predicted prevalence and incidence of diabetes and 95% CI were calculated from logistic regression models (ie, using the inverse logit function). The predicted PR and RR were obtained from Poisson generalized estimating equation (GEE) (28) models with robust standard errors, because we were interested in ratio statistics (PR and RR) that directly compare any 2 probabilities. Logistic regression provides accurate estimate for probabilities, but odds ratios calculated from logistic regression may not provide accurate estimates of PRs and RRs when prevalence and incidence are greater than 10% (28,29).

Bivariate and fully adjusted multivariate main effects analyses were performed for PGS percentile, age, sex, education, poverty ratio, BMI, smoking status, alcohol consumption, and PA. All regression models with PGS accounted for 10 genetic principal components. We tested age polynomials in the model to explore potential nonlinear relationships between prevalence/incidence and age, regardless of statistical significance in the main effects model. If PGS was associated with diabetes in the fully adjusted main effects models, interaction models utilizing both backward elimination and stepwise regression modeling techniques were manually carried out to examine interactions between PGS and other covariates. To interpret the PGS interaction models, nonparametric bootstrapping method with 5000 replications was performed to calculate differences in diabetes prevalence and incidence by PGS percentiles and interaction variables (30). Lastly, we computed the proportions of variation explained by PGS using the method of incremental area under the curve (AUC) by deleting PGS from the final models of prevalence analysis. All analyses were completed using R version 3.5.0.

Results

Sample Characteristics

Table 1 shows the baseline characteristics of the analytical sample by race, as well as the race-stratified characteristics of those composing the incidence subset (ie, those without self-reported doctor-diagnosed diabetes or hemoglobin A1C level ≥6.5 at baseline). Black participants had nearly twice the baseline prevalence of diabetes compared to Whites (26.2% vs 14.2%). After excluding baseline cases, a higher proportion of Black participants (26.5%) were diagnosed with diabetes compared to Whites (15.8%). At baseline, compared to White participants, fewer Black participants reported a college education and regular vigorous PA, while a higher percentage lived in poverty, were obese, and current smokers. More Black HRS participants reported never drinking alcohol, and among drinkers, Blacks drank less than Whites.

Table 1.

Baseline Characteristics for 11 178 White and 2823 Black Participants (all samples), and the Subsamples of 9562 Whites and 2073 Blacks After Excluding Participants Not Diagnosed With T2D at Baseline or Missing Follow-up T2D Status for Incidence Analysis

Variables	White		Black
	All Sample (N = 11 178)	Excluding T2D Cases (N = 9562)	All Sample (N = 2823)	Excluding T2D Cases (N = 2073)
	N (%)	N (%)	N (%)	N (%)
Baseline T2D cases	1590 (14.2%)		739 (26.2)
Incidence by last visit		1509 (15.8)		549 (26.5)
Baseline age, mean (SD)	64.9 (10.1)	64.6 (10.1)	60.8 (8.8)	60.5 (8.9)
Age group
50–59	3953 (35.4)	3501 (36.6)	1512 (53.6)	1166 (56.2)
60–69	3592 (32.1)	3041 (31.8)	845 (29.9)	573 (27.6)
70–79	2598 (23.2)	2145 (22.4)	346 (12.3)	246 (11.9)
80–100	1035 (9.3)	875 (9.2)	120 (4.3)	88 (4.2)
Age at last visit, mean (SD)		73.3(10.6)		67.5(9.9)
Age group
50–59		1089 (11.4)		504 (24.3)
60–69		2685 (28.1)		846 (40.8)
70–79		2871(30.0)		428 (20.7)
80–103		2918 (30.5)		295 (14.2)
Sex
Male	4997 (44.7)	4166 (43.6)	1110 (39.3)	809 (39.0)
Female	6181 (55.3)	5396 (56.4)	1713 (60.7)	1264 (61.0)
Education
Less than high school	1370 (12.3)	1077 (11.3)	796 (28.2)	546 (26.3)
High school/GED	4124 (36.9)	3490 (36.5)	925 (32.8)	674 (32.5)
College 1–3 years	2782 (24.9)	2404 (25.1)	721 (25.5)	559 (27.0)
College 4+ years	2902 (26.0)	2591 (27.1)	381 (13.5)	294 (14.2)
Poverty ratio
Under poverty	427 (3.8)	347 (3.6)	543 (19.2)	381 (18.4)
1–2.9	3355 (30.0)	2761 (28.9)	1129 (40.0)	820 (39.6)
≥3	7397 (66.2)	6455 (67.5)	1151 (40.8)	872 (42.1)
BMI categories
Normal	3694 (33.0)	3431 (35.9)	569 (20.2)	495 (23.9)
Overweight	4349 (38.9)	3807 (39.8)	931 (33.0)	708 (34.2)
Obese	3135 (28.0)	2324 (24.3)	1323 (46.9)	870 (42.0)
Physical activity per week
Some/never	2154 (19.3)	1710 (17.9)	919 (32.6)	628 (30.3)
Light 2+	1907 (17.1)	1583 (16.6)	532 (18.8)	379 (18.3)
Moderate 2+	4000 (35.8)	3444 (36.0)	830 (29.4)	631 (30.4)
Vigorous 2+	3117 (27.9)	2825 (29.5)	542 (19.2)	435 (21.0)
Smoking status
Nonsmoker	4776 (42.7)	4130 (43.2)	1128 (40.0)	847 (40.9)
Current smoker	1657 (14.8)	1448 (15.1)	674 (23.9)	530 (25.6)
Past smoker	4745 (42.4)	3984 (41.7)	1021 (36.2)	696 (33.6)
Alcohol intake
Never	4505 (40.3)	3609 (37.7)	1440 (51.0)	994 (47.9)
1/day	5274 (47.2)	4659 (48.7)	1133 (40.1)	875 (42.2)
2/day	879 (7.9)	818 (8.6)	144 (5.1)	119 (5.7)
≥3/day	520 (4.7)	476 (5.0)	106 (3.8)	85 (4.1)

Note: BMI = body mass index; GED = Generalized Education Development; PGS = polygenic score; T2D = type 2 diabetes mellitus.

PRs and RRs of Diabetes by PGS Percentiles From Main Effects Models

Supplementary Tables S1 and S2 present crude and adjusted diabetes prevalence and incidence, and crude and adjusted PRs and RRs for 10% increase in PGS, age, age², and all other covariates obtained from bivariate and multivariate main effects GEE models, for White and Black participants, respectively. Among White participants, the adjusted PR for each 10% increase in PGS was 1.10 (95% CI: 1.08, 1.12; p < .0001), whereas the adjusted RR was 1.05 (95% CI: 1.03, 1.07; p < .0001). Among Black participants, PGS was associated with the diabetes prevalence (adjusted PR = 1.06; 95% CI: 1.03, 1.10; p = .0004), but not incident diabetes (RR = 1.01; 95% CI: 0.97, 1.05; p = .5509).

PGS was not associated with diabetes incidence among Black participants; therefore, the interaction between PGS and covariates for incident diabetes was not examined among Black participants. Statistically significant interactions were found between PGS and PA on both diabetes prevalence and incidence among White participants; however, not on diabetes prevalence for Black participants. We included the interaction terms for the 3 models to examine patterns between PGS and PA. Regression coefficients are listed in Supplementary Table S3. The variations explained by PGS (incremental AUCs) in the final interaction models for T2D prevalence were 16.3% for Whites and 14.9% for Blacks.

Predicted Diabetes Prevalence and Incidence by Age at the 10th and 90th PGS Percentiles

As shown in Figure 2 and Supplementary Table S4, diabetes prevalence increased with age and began to decrease at 77.7 years of age among White participants and 68.9 among Black participants. Black participants had a faster rate of decline in diabetes prevalence. Diabetes incidence decreased continuously by age among both White and Black participants. Nonoverlapping confidence bands in Figure 2 indicate a statistically significant difference between the 90th and 10th PGS percentiles.

Figure 2.

Mean T2D prevalence and incidence from age 50 to 90 years for Whites and Blacks at 10th and 90th PGS percentile estimated from the final models and corresponding 95% confidence bands (estimated prevalences are listed in Supplementary Table S4). PGS = polygenic score; T2D = type 2 diabetes mellitus.

Among Whites, the diabetes prevalences within the 90th PGS percentile were 11% at age 50, 19.8% at 70, and 18.4% at 90. These estimates were more than twice the prevalence within the 10th PGS percentile (4.5% at age 50, 8.6% at age 70, and 7.9% at age 90). The differences between the 90th and 10th PGS percentile were 6.5% at age 50, 11.2% at age 70, and 10.5% at age 90. Diabetes incidence within the 90th PGS percentile among Whites decreased from 26.3% at age 60 to 6.8% at age 90 and decreased from 17.2% at age 60 to 4.1% at age 90 within the 10th PGS percentile. The differences of diabetes incidence between the 90th and 10th PGS percentile also decreased from 9.1% at age 60 to 2.7% at age 90.

For Black participants in the 90th PGS percentile, the diabetes prevalences were about twice that of the White participants at ages 50 and 70 years (22.4% and 38%, respectively). Additionally, the differences in prevalence between those in the 90th and 10th PGS percentiles were larger than among Whites (10.1% at age 50 and 15.1% at age 70). At age 90, the difference converged to 8.9% (19.3% in the 90th and 10.4% in the 10th percentile) and the difference was not statistically significant. Diabetes incidence among Black participants was higher than that among Whites with a similar decreasing pattern with increasing age; however, the difference between the 90th and the 10th PGS percentiles was not statistically different.

Interaction Between PA and PGS on Diabetes

Overall, moderate PA and vigorous PA were associated with lower diabetes prevalence and incidence compared to some/no PA among White participants and diabetes prevalence (but not incidence) among Black participants. The interaction between PA and PGS on diabetes prevalence and incidence was also observed among White participants (Figure 3 and Supplementary Table S5).

Figure 3.

Mean T2D prevalence and incidence by PGS percentiles (0–1) for Whites and Blacks at 3 levels of physical activities (PAs) estimated from the final models and corresponding 95% confidence bands. PGS = polygenic score; T2D = type 2 diabetes mellitus.

Among White participants in the 90th percentile of PGS, moderate PA was associated with 6.8% (95% CI: 2.9, 10.6; p_bootstrap = .0006) lower diabetes prevalence compared to some/no PA (17.0% diabetes prevalence for moderate PA vs 23.8% for some/no PA). Whereas in the 10th PGS percentile, no statistical difference in diabetes prevalence was found (difference = 0.1%; 95% CI: −2.2, 2.4; p_bootstrap = .9224; 9.2% diabetes prevalence for moderate vs 9.1% for some/no PA). The PGS–moderate PA interaction is statistically significant (p_interaction = .0096). Similar interaction was found between PGS–vigorous PA interaction (p_interaction = .0257). The PGS–PA interactions effects were similar for diabetes incidence.

Among Black participants, vigorous PA was associated with 11.2% (95% CI: 1.4, 21.0; p_bootstrap = .0252) lower diabetes prevalence in the 90th PGS percentile, and 8.2% (95% CI: 0.8, 15.6; p_bootstrap = .0305) lower in the 10th PGS percentile compared to some/no PA. Moderate PA was associated with 3.4% (95% CI: −5.3, 11.9; p_bootstrap = .447) lower diabetes prevalence in the 90th PGS percentile, and 8.8% (95% CI: 2.3, 15.3; p_bootstrap = .008) lower in the 10th PGS percentile compared to some/no PA. The interactions were not statistically significant.

Discussion

Our study investigated the associations and interactions of nonmodifiable (ie, age, PGS) and modifiable risk factors (ie, PA) on the probability of incident and prevalent diabetes among nationally representative samples of older Black and White adults in the United States. Among both Black and White participants, higher BMI, higher poverty ratios, and lower PA were associated with a higher diabetes prevalence and incidence, consistent with the literature on environmental factors and diabetes (6,31). Moreover, higher PGS percentiles appeared to have a greater multiplicative effect among Whites for both prevalent and incident diabetes; whereas PGS was not significantly associated with diabetes incidence among Black participants. These findings were consistent with a recent HRS study finding a weaker relationship between PGS and diabetes onset among Black participants (HR = 1.219; 95% CI: 1.059, 1.402) compared to Whites (HR = 1.380; 95% CI: 1.300, 1.464) (32). Similarly, Vassy et al. (33) reported a significant association between PGS and incident diabetes among both Black and White participants, with a stronger magnitude of association observed among Whites. However, these studies methodologically differ from ours in the use of survival analysis assuming constant risk by age. Moreover, Vassy et al. (33) utilized a cohort of younger adults relative to our HRS sample, which may in part explain the differences in predictive power in the PGS among Black participants.

Diabetes may be occurring and/or be diagnosed earlier among Black participants, which is congruent not only with the high observed baseline prevalence among Blacks in our study, but also with the larger literature on U.S. Black–White crossover effects (34). Utilizing a large representative sample of Black and White Americans, Cunningham et al. (35) reported that due to differences in health behaviors, access to health care, and larger social factors, Black adults were nearly twice as likely to develop or die from chronic conditions in younger age groups relative to Whites. Thus, the risk pools among which incident diabetes could emerge may have been different between Black and White HRS participants; that is, there may have been more White participants at risk of diabetes at during follow-up relative to Black participants. Alternatively, the lack of predictive power in the PGS for Black participants may derive from the GWAS meta-analysis used to calculate the PGS, which was based on European ancestry, and thus, may not adequately represent the genetic determinants of diabetes among Black Americans (36).

We observed a nonlinear relationship between age and diabetes prevalence. The prevalence among Black participants appeared to peak before decreasing at age 68.9 years, while cases among Whites peaked nearly 10 years later at age 77.7 years, followed by a slower decrease in prevalence relative to Black participants. Despite the shorter follow-up time compared to Whites, Black participants had a significantly greater overall prevalence and incidence of diabetes, with estimates nearly double that of Whites. Among those in the 10th PGS percentile, the prevalence of diabetes among Black participants appeared to nearly mirror the prevalence of White participants within 90th PGS percentile. That is, Black participants with the lowest genetic risk had a similar prevalence of diabetes relative to White participants with the highest genetic risk. These findings may underscore the lack of predictive power in the PGS for diabetes among Black participants and may suggest that other, nongenetic, factors might disproportionately contribute to diabetes risk among non-Whites. However, while variability in modifiable lifestyle factors may provide some insights into the overall disparity in diabetes probability between Whites and Blacks in our study, the difference in prevalence is not predicated on differences in lifestyle factors or simply belonging to a specific racial/ethnic group. Instead, larger environmental/societal determinants related to both race/ethnicity and lifestyle, such as neighborhood factors and health resource accessibility, may have a strong influence on diabetes risk, especially among minority groups, including life-course exposures that are not captured in the HRS (37). Thus, the observed difference in peak diabetes prevalence between Black and White participants within our study is consistent with literature showing a disproportionate amount of chronic conditions associated with older age occurring among younger Black adults relative to Whites, and these differences may be driven by differences in social and behavioral factors, as well as health care access (35).

Significant interactions were detected between both prevalent and incident predicted diabetes and PA among Whites, but not among Black participants. Specifically, the association between PA and diabetes was stronger among White participants with higher, relative to lower, polygenic risk of diabetes. Thus, these observations suggest that the putative effect of physical inactivity may be exacerbated by underlying genetic risk. Our interaction results were unexpected given prior findings; Klimentidis et al. (15) reported a stronger PGS effect on diabetes incidence among those with higher levels of PA among participants from the Atherosclerosis Risk in Communities study. However, our study substantially differs in methodology; we examined differences between 90th and 10th percentiles of a PGS based on weighted sums of all SNPs beta estimates from the DIAGRAM consortium GWAS meta-analysis (38), whereas Klimentidis et al. (15) used an unweighted PGS based on 69 diabetes-associated SNPs. Moreover, HRS assesses PA by directly inquiring about the frequency respondents are engaged in specific activities associated with light, moderate, and vigorous PA, while Klimentidis et al. (15) focused on the sports index within the Baecke Physical Activity questionnaire (39). Nonetheless, it is unlikely that these methodological differences adequately explain the opposite interaction association observed in our study. While our interaction results among Whites contradicts Klimentidis et al. (15), among Black participants, higher PGS appeared to have a weaker effect on prevalent diabetes among those reporting higher levels of PA, compared to some/no PA; however, the interaction results for Blacks were not statistically significant.

There are several limitations to consider when interpreting our findings. First, our diabetes case ascertainment relied largely on self-reported status of diabetes; however, self-reported diabetes status has been shown to have high agreement with physician records (40). Second, as previously mentioned, the loci used to calculate the PGS derived from GWAS on European ancestry groups and may not reflect the same predictive value for other ancestry groups (36). Moreover, we do not have access to the HRS GWAS data, and thus we are limited to the PGS produced by the HRS team; however, because HRS used all SNPs to compute the PGS, any differences in effect between a PGS derived from a more recent GWAS of diabetes and the PGS we used is likely to be marginal. The variation explained by the HRS-PGS (16.3% among Whites and 14.9% among Blacks) was comparable to the 18.0% of diabetes heritability accounted for by diabetes GWAS SNPs (10). This conclusion is further substantiated by a recent gene–environment analysis of BMI using HRS data, where R² using newly discovered SNPs from the most recent GWAS for BMI (41) was nearly identical to the R² using the PGS provided by HRS using all GWAS SNPs (42). Third, we also acknowledge the risk of selective survival. Using an older cohort, especially when studying the risk of PGS on diabetes, could lead to selection bias toward healthier participants. Any small amount of bias could have a significant impact on genetic variants due to the modest magnitude of effect of individual SNPs. We try to limit that threat by utilizing a combined SNP list (PGS) which has been shown to be more robust (43). Finally, it is important to reiterate that the associations between both environmental and genetic factors on diabetes did not fully explain the differences observed in prevalence and incidence between White and Black participants. Thus, the higher prevalence of diabetes among Black participants may be due to unmeasured factors influencing Blacks differently than Whites. Our study is unable to assess lifetime exposures which might impact these findings, such as early childhood environments. Additionally, the notably smaller sample size may explain the lack of interactions detected among Black participants compared to Whites.

Despite limitations, our study has many strengths and adds to the growing literature analyzing genetic variants and chronic diseases. First, because our sample is nationally representative, we were not only able to quantify differences in the prevalence of nonmodifiable risk factors and modifiable risk factors of diabetes, but also examine the impact and interplay of these factors within specific racial groups to address the public health significance at a population level. Quantifying these associations within different groups offers crucial insights that may improve precision in therapeutic approaches to preventing and managing diabetes in older adulthood. Second, while other studies have examined similar interactions between genetic variation and PA, our study utilizes a weighted PGS calculated from an updated list of loci associated with diabetes (26). We were able to further investigate modifiable factors beyond PA such as smoking status, alcohol consumption, poverty ratio, and education. Finally, our sufficient sample size enabled us to investigate the effects of environmental factors, and their interactions with PGS among White participants, on diabetes.

In conclusion, we observed independent and interacting associations between PGS and PA on diabetes among both Black and White individuals in a nationally representative sample of older adults. In order to better understand the relationship between race/ethnicity and diabetes, future studies may benefit from targeting non-European participants in identifying genetic variants associated with diabetes, especially given the markedly weaker predictive power in the PGS among Black participants. With a more complete accounting of genetic predisposition among non-European groups, gene–environment studies including multiethnic comparisons may have a higher degree of reliability in assessing whether certain environmental factors may differentially affect certain groups’ risk of diabetes. Moreover, inclusion of other modifiable factors, such as dietary patterns, may further elucidate interactions between genetic and environmental risk of diabetes. Researchers should continue to focus on understanding the complex interplay between genetics and environmental factors on diabetes risk, particularly among understudied, underserved, and high-risk populations, to better inform preventative efforts and understanding of the underlying etiology.

Funding

This study was supported by Ola HAWAII through the National Institute on Minority Health and Health Disparities (U54MD007601-31), National Institutes of Health. The interview data from the HRS were sponsored by the National Institutes of Health, National Institute on Aging (U01AG009740). Genotyping of HRS participants was funded by the National Institute of Health’s Director’s Opportunity for Research awards using American Reinvestment and Recovery Act (RC2 AG036495-01, RC4 AG039029-01).

Conflict of Interest

None declared

Author Contributions

Y.Y.W. and C.M.P. designed the study, performed the analyses, interpreted the data, wrote and edited the manuscript, and contributed to the final version. M.D.T. and F.Y. interpreted the data, wrote and edited the manuscript, and contributed to the final version. C.M.P. is the corresponding author of this work.