Censored patients in Kaplan–Meier plots of cancer drugs: An empirical analysis of data sharing

Introduction: Kaplan–Meier survival analysis, the cornerstone of evaluating efficacy of oncology drugs in randomised controlled trials (RCTs), assumes censored patients are neither healthier nor sicker than those followed. We sought to examine whether censoring patterns differ between the control and experimental arms in one oncology journal that mandates the reporting of the number of patients censored. Methods: In this retrospective review, proportion of censoring and study design data were gathered from RCTs published in The Lancet Oncology that reported Kaplan–Meier curves between May 2018 and August 2019. Differential censoring rates were analysed at the 1st, 3rd, 6th, and overall time points in each study. Analysis was stratified by curves reporting progression-free survival (PFS) or overall survival (OS) end-points. Results: Of the 160 articles reviewed, 29 studies with 51 Kaplan–Meier curves were eligible. In both OS (N = 25) and PFS curves (N = 26), the absolute weighted difference in censoring between the control and experimental arms was initially positive, indicating more censoring in the control arm (first time point OS: 0.32%; PFS: 2.00%). The absolute difference then became negative, indicating more censoring in the experimental arm as time progressed (end-of-study OS: −7.54%; PFS: −9.09%). Conclusion: Differences in censoring between control and experimental arms of cancer RCTs suggest that there could be systematic bias present at various study time points that may influence key results. Further investigation is needed, as possible reasons include study assignment disappointment, inappropriate follow-up length, lack of efficacy, or intolerable toxicity, each predominant at specific time points after randomisation.