Study-specific GWAS analysis. Each study conducted uniform cross-sectional analyses for each smoking phenotype using an additive genetic model. Logistic regression was used for discrete traits (SI and SC) and linear regression was used for quantitative traits (CPD and AOI). Continuous, quantitative traits were normalized by transformation to Z scores, owing to heavy tails and non-normality. Outliers were removed within each study, where abs (Z)>2. Link (Y)=Z scores were fit using ordinary least squares regression. To investigate potential sources of heterogeneity across studies, we examined the distribution of African ancestry in each cohort (Supplementary Figure 1). To account for population stratification and admixture, all studies adjusted for an appropriate number of eigenvectors3, 4, 5, 6, 7, 8, 9, 10 from a study-specific principal components analysis.34 In addition, study-specific analyses included adjustment for age and case status or study site, when appropriate. Genomic control inflation factors were computed using standard methods.35, 36
