The mean bias for the Cox model was approximately zero across all the simulation scenarios (Table 1 and Supplementary Table S3). However, the estimated ORs were on average larger than the underlying HRs, especially for larger effect sizes. There was also a larger degree of divergence between the HRs and the ORs as the length of follow-up time increased (average follow-up time: complete follow-up model=19.1 years; survey follow-up model=11.9 years; random follow-up model=9.4 years), and as the cumulative disease incidence increased (Table 1 and Supplementary Table S3). The HRs were more precise (as estimated by the model) than the ORs across all scenarios, and this relative difference in precision was positively correlated with cumulative disease incidence (Table 1 and Supplementary Table S3). This increase in precision offsets the greater effect sizes of the logistic regression model, and hence leads to the greater differences in power between the Cox and logistic regression models for the larger cumulative disease incidences (Figure 1).
