Statistical interpretation of Kaplan–Meier curves

Excerpt

Sir: Thank you for giving us the opportunity to reply to Dr. Garrod's constructive comments. Dr. Garrod is concerned about the Kaplain–Meier (KP) curves crossing over. We think that there is some misunderstanding of the relevance of crossing over. When the number of patients left in a trial are so small, the right side of KP curves has no relevance to the prime statistic [1]. We disagree with Dr. Garrod when he suggests there should have been a correction for multiple analyses. The original power calculation was based on univariate parametric comparison of a single variable, i. e. time to become ventilator free. Correction for multiple comparisons using log-rank analysis of KP curves is undertaken when numerous variables are related to survival curves. We limited analysis solely to the median time to become ventilator free, i. e. the time for the first 50% of patients in each group to become ventilator-free, and it took longer for the first 50% of chest physiotherapy patients to do so. We also took care in the KP analysis to use methodology that was not distorted by decreasing numbers due either to death or to becoming ventilator free. This test is less sensitive and less likely to over-interpret small differences when numbers fall. Dr. Garrod should also be aware that length of stay and mortality data were analysed on the whole cohort using univariate nonparametric analyses and not merely the first 50% in each group which were compared in the KP analysis. …