Keywords: Hazard ratio, Logrank test, Median difference, Test-estimation coherency, Wilcoxon test
Two major tasks in comparative clinical trials with time-to-event outcome are 1) testing equality of treatment groups, and 2) estimating a primary summary of the treatment effect. Because the former provides a dichotomous outcome (i.e., reject the null hypothesis or not), it naturally fits a binary decision (e.g., approve or not-approve). However, for clinicians or patients, the latter would be more important for their decision-making based on the risk-benefit balance of the treatment. From this point, the conclusion from 1) testing and 2) estimation of treatment effect should be coherent concerning statistical significance. That is, if test claims significance of the treatment effect, confidence interval of the treatment effect should exclude the null value, vice versa. The conventional logrank-hazard ratio (HR) test-estimation practice holds this “test-estimation coherency.” Although the drug approval is based on the totality of evidence, it seems that huge weight has been still given to statistical significance (p-value) in the pre-specified primary test for confirmation of efficacy for many years. This formed the current conventional test-estimation practices (e.g., logrank-HR test-estimation, or Wilcoxon-median difference test-estimation), without paying much attention to the issues of those. As a result, we may somehow have missed the importance of providing a robust, clinically interpretable, coherent, single summary of the treatment effect for clinicians and patients. It is time to focus more on estimation part, moving beyond the P-value world and the current conventional practices.