Statistical Analysis

We excluded ARIC participants with race other than black or white (48), participants with prevalent stroke in the SHS (28) or ARIC (286), and ARIC participants with unknown baseline stroke status (362). We then excluded participants with any missing data on covariates (306 AIs, 399 blacks, and 519 whites), and SHS participants born before 1920 (183) to align with birth years in ARIC. Sensitivity analyses, using multiple imputation for missing values, resulted in negligible differences to point estimates and CIs; we report results for the complete case analysis.

All inferential analyses were conducted separately for women and men. We estimated primary stroke incidence as rates per 100 000 person‐years for AI, black, and white racial groups by attained age and birth cohort. We used Cox regression to compare stroke hazards by race with attained age as the time scale, so that each participant entered the model at his or her baseline age. Nonstroke deaths were treated as censored observations. We fit 3 separate Cox models: (1) unadjusted; (2) adjusting for residual confounding by birth year; and (3) additionally adjusting for lifestyle and health factors measured at baseline and included in the pooled data set. Study site was collinear with race for all SHS and most ARIC locations and so was not included as a covariate. We tested for effect measure modification between race and birth year. Analyses were stratified if effect‐measure modification was present. Results are presented as point estimates with 95% CIs. We tested the proportional hazards assumption for all models based on a threshold of P=0.05 and conducted stratified analyses to accommodate violations of this assumption.

For people who experienced stroke during follow‐up, we used logistic regression to estimate racial differences in 30‐day and 1‐year poststroke mortality. Models were estimated as described for incident stroke, except that we included age at stroke event as a covariate. We used marginal standardization to report risk differences and risk ratios for AIs compared to blacks and whites.19, 20 We used Stata software (version 14.1; StataCorp LP, College Station, TX) for all analyses.21