We used logistic regression for all cohorts with a cross-sectional study design to model the multiplicative SNP effects on risk for the dichotomous outcome of stroke against ancestry-matched controls, whereas we used Cox proportional-hazards models for the prospective studies to assess time to first stroke, fitting an additive model relating genotype dose to the stroke outcome. Where genotypes were imputed, SNPs were modelled as allele dosages. Of the discovery cohorts, four (of 15) centres used ancestry-informative principal components as covariates to correct for population stratification. All cohorts providing genome-wide data removed population outliers before imputation. After verifying strand alignment, filtering SNPs with minor allele frequency lower than 0·01, and removing poorly imputed SNPs across centres, we did a meta-analysis of the results of the association analyses from every centre using a fixed-effects inverse-variance weighted model using METAL.22
