Cox model to account for the sampling process of this study design. The HRs also became less precise as the average length of follow-up time decreased. For instance, the mean SE of the logarithm of the estimated HR for the sampling fraction of 10%, RAF=0.10, cumulative disease incidence of 10% and underlying HR of 1.10 were 0.052, 0.056 and 0.060 for the complete, survey and random follow-up models, respectively. These large differences in precision between the censoring strategies were caused by the smaller amount of information from the subcohort contributing to the pseudo-likelihood at each failure as the amount of censoring increased. As expected, the distributions of the ORs were similar across the censoring strategies as there were approximately the same number of cases and non-cases across the censoring strategies for the same cumulative disease incidence. Hence, the greater precision of the logistic regression model compared with the Cox model for the random and survey follow-up models leads to the larger differences in power for these models. The increased power of the logistic regression model in the complete follow-up model was mainly driven by the larger effect estimates of the logistic regression model.
