Details of the methods of the REGARDS study have been published [19]. Briefly, REGARDS is a prospective cohort study designed to identify contributors to regional and Black-White disparities in stroke mortality. The study over-sampled Black persons and residents of the stroke belt (North Carolina, South Carolina, Georgia, Alabama, Mississippi, Tennessee, Arkansas, and Louisiana). Between January 2003 and October 2007, using postal mailings and telephone interviews, a total of 30,239 participants were recruited from a commercially available list of residents. Socio-demographic information and medical histories were obtained by a computer-assisted telephone interview (CATI). An in-home examination was performed 3–4 weeks after the telephone interview. Trained staff collected medication information, blood and urine samples, blood pressure readings, and a resting electrocardiogram (ECG). Approximately 10 years after the baseline assessment, 2013–2016, participants who were still alive and active completed a follow-up examination similar to the baseline visit. The institutional review boards at the collaborating centers approved the REGARDS study protocol, and all participants provided written informed consent.

Diet was assessed at baseline with the Block 98 food frequency questionnaire (FFQ), a validated semi-quantitative FFQ that assessed usual dietary intake of 110 food items (NutritionQuest, Berkeley, CA) [19, 20]. For each line item on the FFQ, participants were asked how often, on average, they consumed the food (or group of foods) during the previous year, as well as the usual quantity of the food consumed. The FFQ also included adjustment questions (e.g., inquiring about the type of milk consumed—low-fat, non-fat, etc.) to further refine intake data. The FFQ was self-administered after the in-home visit and mailed to the REGARDS Operations Center, where they were reviewed for completeness and scanned. The results were then sent to NutritionQuest for scoring, which included a data set that provided the number of grams per day (g/day) for each line item on the FFQ. Amounts of each food on the FFQ consumed by a participant were calculated by multiplying the frequency of consumption of that food by the usual amount consumed. A total of 56 food groups, on which dietary patterns were based, were derived using the 110 individual food variables on the FFQ using published methods [18].

Split sample replication was used to (1) derive the dietary patterns using exploratory factor analysis, and (2) test the patterns using confirmatory factor analysis [18, 21]. Patterns were named based on the major factor loadings. Factor 1 loaded heavily on mixed dishes, pasta dishes, pizza, Mexican food, and Chinese food and was designated the “Convenience” pattern. Factor 2 had high factor loadings for vegetables, fruits, fruit juice, cereal, beans, fish, poultry, and yogurt and was named the “Plant-based” pattern. Factor 3 loaded on added sugars, desserts, chocolate, candy, and sweetened breakfast foods and was named the “Sweets” pattern. Factor 4 loaded heavily on added fats, fried food, eggs and egg dishes, organ meats, processed meats, and sugar-sweetened beverages, reflecting a culinary pattern observed in the Southeastern United States, and was named the “Southern” pattern. Factor 5 loaded highly on beer, wine, liquor, green leafy vegetables, tomatoes, and salad dressing and was designated the “Alcohol and Salad” pattern. A standardized adherence score for each of the 5 dietary patterns was created (lower score = lower adherence).

Two patterns were considered to have health promoting properties based on their food composition and the general nutritional epidemiologic literature. Specifically, both the alcohol/salads and plant-based patterns had high factor loadings for foods such as vegetables, leafy greens, nuts, seeds, and fish, which are beneficial for health. We also identified three patterns, convenience, sweets, and southern, as unhealthy because they had higher factor loadings for foods that are known to be associated with unfavorable health and disease outcomes [18].

The Mediterranean diet score (MDS) was derived according to published methods used in REGARDS based on the method of Trichopoulou and colleagues [22]. In brief, food group contributors to the MDS included those designated as “beneficial” (vegetables, fruits, legumes, cereals, fish), and “detrimental” (meat, dairy). One point was assigned for consumption that exceeded the sex-specific median for the “beneficial” groups or was below the median for “detrimental” food groups. For fat intake (eighth food category), participants with monounsaturated lipids to saturated lipids ratios at or above the sex-specific median were assigned a value of 1, and those with ratios below the sex-specific median were assigned a value of 0. For alcohol consumption (ninth category), participants were assigned a score of 1 for moderate consumption (> 0 and ≤ 7 drinks per week for women and > 0 and ≤ 14 drinks per week for men) and a score of 0 for everyone else. The MDS was determined by summing scores for the 9 food groups, resulting in a possible range of scores of 0 to 9.

AF was identified at baseline and a subsequent follow-up visit approximately 10 years later by (1) a visit ECG and (2) self-reported history of a physician diagnosis during the CATI survey. The ECGs were read and coded at a central reading center (EPICARE, Wake Forest School of Medicine, Winston-Salem, NC) by analysts who were blinded to other REGARDS data. Self-reported AF was defined as an affirmative response to the following question: “Has a physician or a health professional ever told you that you had atrial fibrillation?” This question was as reliable a predictor of incident stroke events as AF detected by ECG [23].

Participant characteristics at baseline were used as covariates. Age, sex, race, household income, education, and smoking status were self-reported. Body mass index (BMI) and waist circumference were measured at the baseline examination. Physically active was defined as ≥ 4 days of exercise (enough to work up a sweat) per week. Hypertension was defined as systolic blood pressure ≥ 130 mm Hg, diastolic blood pressure ≥ 80 mm Hg, or self-reported current use of anti-hypertensive therapy. Dyslipidemia was defined as total cholesterol ≥ 240 mg/dL, low-density lipoprotein cholesterol ≥ 160 mg/dL, high-density lipoprotein cholesterol ≤ 40 mg/dL, or self-reported current use of lipid-lowering therapy. Diabetes mellitus was defined as fasting glucose ≥ 126 mg/ dL, non-fasting glucose ≥ 200 mg/dL, or self- reported current use of anti-diabetic medications. CRP measurement used a high-sensitivity particle-enhanced immunonephelometric assay on the BNIII nephelometer (N High Sensitivity CRP, Dade Behring Inc., Deerfield, IL) with an interassay coefficient of variation of 2–6%. CVD included the presence of coronary heart disease (a self-reported history of myocardial infarction, coronary artery bypass grafting, coronary angioplasty or stenting, or evidence of prior myocardial infarction on the baseline ECG) or prior stroke which was ascertained by participant’s self-report.

Descriptive statistics for demographic, socioeconomic, lifestyle, anthropometric, and medical history variables at the baseline assessment according to quartiles of consumption of each dietary pattern and MDS categories were calculated using the chi-square test (for proportions) and analysis of variance (for continuous variables).

Logistic regression was used to calculate the odds ratio (OR) [95% confidence interval (CI)] for prevalent (baseline visit) and incident (follow-up visit) AF for each of the five dietary patterns as well as the MDS per standard deviation (SD) increment in adherence. Models were built as follows: model 1: age, sex, race, education, household income, and region; model 2: model 1 plus total energy intake, lifestyle factors (smoking, physical activity), CVD risk factors (BMI, waist circumference, hypertension, dyslipidemia, diabetes, history of CVD), and CRP. Statistical significance for all comparisons including interactions was defined as p < 0.05. SAS version 9.4 (Cary, NC) was used for all analyses.

Due to inherent limitations with the MDS, namely the designation of any dairy intake as adverse, and the availability of other scoring systems, we repeated the analyses using the Mediterranean-DASH Diet Intervention for Neurodegenerative Delay (MIND) diet score. Among the MIND diet components are 10 brain healthy food groups (green leafy vegetables, other vegetables, nuts, berries, beans, whole grains, seafood, poultry, olive oil and wine) and 5 unhealthy food groups (red meats, butter and stick margarine, cheese, pastries and sweets, and fried/fast food). This diet score has been associated with slower cognitive decline and reduced cardiovascular disease [24,25,26].