To assess Cox and logistic regression models in cohort studies, we used the subcohort of EPIC-CVD and 25 SNPs from known CHD loci. Since there were only 437 CHD events in the individuals genotyped in the subcohort, there was limited power to detect SNP–CHD associations in the subcohort with either model. Nevertheless, in the subcohort the effect estimates were directionally concordant for the Cox and logistic regression models for 24 of the 25 SNPs (Table 3), and the SNP that was directionally discordant had effect estimates very close to the null for both models. Of the 24 concordant SNPs, 18 had the same direction of effect as in the literature (Supplementary Table S8).1, 2 The SNPs that showed different directions of association with the literature were in general those SNPs with smaller effect on CHD and hence required large sample sizes to detect these effects (Supplementary Table S8). The Cox model was more precise than logistic regression for all of the SNPs and for the effect sizes further away from the null the ORs tended to be larger in magnitude (Table 3).
