REGARDS is a longitudinal cohort study designed to determine the reasons for racial disparities in stroke and cognitive decline in Black and White Americans aged ≥45 years (Howard et al., 2005). REGARDS enrolled 30,239 participants from the 48 continental United States from 2003 to 2007. All self-reported Black participants consenting to genetic research were included in this study (Figure 1). All participants provided oral and written informed consent. The REGARDS study was approved by the Institutional Review Boards of participating centers. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline. The analytic plan was pre-specified and approved by the REGARDS Publications Committee. Among 9,999 Black REGARDS participants with HBA and HBB genotyping available, 8,947 participants had no self-reported history of stroke at baseline and complete follow-up data (Figure 1).

Incident ischemic stroke was defined as non-hemorrhagic stroke with a focal neurological deficit lasting ≥24 h confirmed by the medical record or a focal or non-focal neurological deficit with positive imaging confirmed with medical records in participants without a prior history of stroke. Genomic DNA was analyzed using droplet digital PCR (Ruhl et al., 2022b) to determine HBA copy number. The covariates age, sex, race, health insurance (yes or no), highest education level obtained (less than high school, high school, some college, college or more), annual income (≤ $20K, $20–34K, $35–74K, ≥$75K), and smoking status (categorized by never, past, or current smoker), history of hyperlipidemia, regular use of lipid lowering medication, and regular use of aspirin were self-reported. Self-reported use of lipid lowering medication was restricted to those also reporting hyperlipidemia. Region was defined as previously described in three geographic areas: stroke belt buckle, stroke belt, and stroke non-belt (Howard and Howard, 2020). Atrial fibrillation was defined by self-report of a physician diagnosis or ECG evidence. The presence of left ventricular hypertrophy (LVH) was defined by 12-lead ECG. Hypertension was defined as systolic blood pressure ≥130, diastolic blood pressure ≥80, or self-reported current medication use to control blood pressure. Chronic kidney disease (CKD) was defined by the 2021 CKD-Epi creatinine-cysteine equation and estimated glomerular filtration rate < 60 mL/min/1.73 m², including those with end-stage kidney disease, and/or urine albumin to creatinine ratio ≥ 30 mg/g measured on urine collected during the baseline in-home examination. Fasting glucose levels ≥ 126 mg/dL, random glucose ≥ 200 mg/dL, or self-reported use of glucose-lowering medication was used to define diabetes mellitus. Hemoglobin, mean cell volume (MCV), mean cell hemoglobin (MCH), mean cell hemoglobin concentration (MCHC), and red-cell distribution width-coefficient of variation (RDW-CV) values were measured or calculated from blood collected during the in-home examination. The first 10 principal components of ancestry were calculated from Infinium Expanded Multi-Ethnic Genotyping Array data on 7,032 (79%) participants. HBA regulatory single nucleotide polymorphisms (SNPs) (rs11248850, rs11865131, and rs7203560) (Milton et al., 2013; Raffield et al., 2018) and a NOS3 SNP (rs1549758) (Mishra et al., 2022) were determined from these array data. The HBA regulatory SNPs rs11248850 and rs11865131 were in high linkage disequilibrium and were concordant in all but five participants. Therefore, we evaluated only rs11248850 and not rs11865131 in the models. The Framingham Stroke Risk Score (10-Year probability of stroke risk percentage) (McClure et al., 2014) and Atherosclerosis Risk in Communities Study (ARIC) Stroke Risk Score, 10-year probability of ischemic stroke risk (percentage) were calculated among those who self-reported at baseline never having had a stroke (Chambless et al., 2004).

Multivariable Cox proportional hazards regression modeling was used to estimate the hazard ratio of HBA copy number on time to first ischemic stroke. Covariates included age, sex, region, insurance status, education level, income, hypertension, atrial fibrillation, left ventricular hypertrophy, smoking status, and diabetes mellitus. Pre-specified tests of interaction between HBA copy number and age, sex, and sickle cell trait were performed on fully adjusted models. Pre-specified sensitivity analyses were performed by adding each of the following to the model: hemoglobin, CKD, both hemoglobin and CKD, regular aspirin use, regular statin use, the first 10 principal components of ancestry, and putative HBA regulatory single nucleotide polymorphisms rs11248850 and rs7203560. A post-hoc sensitivity analysis for the association of the rs1549758 NOS3 (nitric oxide synthase 3 or endothelial nitric oxide synthase) SNP with incident ischemic stroke was performed.

For the ischemic stroke time-to-event analysis, time to event was defined as the number of years between the initial in-home interview date and date D where D was the minimum of D1 and D2 which were defined as follows. D1 was the last follow-up date provided by REGARDS as the last time the participant was contacted for status. The outcome associated with this date was either “No Event” or “Death”. D2 was the date when the individual had an ischemic stroke. Individuals with a D2 date prior to their initial in-home interview were excluded from this analysis. Individuals with a D2 date after interview but prior to D1 time were recorded with an ischemic stroke event and D = D2. Individuals with D2 after D1 had stroke onset after their last REGARDS follow-up and they were recorded with “No Event” and censored at date D = D1. This decision to censor stroke events occurring after the end of REGARDS follow-up avoids bias associated with having extended follow-up only for one type of subgroup—those having a stroke event.

Missing data for the primary outcome, secondary outcomes, and explanatory variables were typically rare (< 0.5%) with some exceptions, e.g., hemoglobin values (Table 1). Multiple imputation methods were used in the multivariable analyses. The R package “mice” Version 3.14.0 was used to create and analyze the resulting imputations (Van Buuren and Goorthius-Oudshoom, 2011).

For diagnostic modeling of the model of ischemic stroke incidence using Cox proportional hazard techniques, Schoenfeld residuals were examined over follow-up time to detect violations of proportional hazards assumptions for the covariates in the analysis of ischemic outcomes. Examination of Schoenfeld residuals of the non-imputed and imputed data sets showed no suggestion of violation of proportional hazards for the regression covariates (p = 0.98).

For incident ischemic stroke, we performed an a priori power calculation based upon the Cox proportional-hazards model. We assumed an additive model where each additional HBA copy increased the hazard ratio for prevalent stroke by 25%. Given n = 479 incident ischemic stroke events observed among those without baseline self-reported stroke, and the distribution of alpha globin alleles, the power to reject the hypothesis of no linear trend in HBA copy number on the log of the hazard rate is ~77%. In contrast, there were only 62 incident hemorrhagic stroke events and power to reject a linear trend in HBA copy number on the log of the hazard rate was only 13%; therefore, analysis of hemorrhagic stroke was not performed.