Assessment of post-switch endpoints in randomized clinical trials

Abstract

Background In randomized controlled clinical trials (RCT) in oncology, the switch from reference treatment arm to another randomization arm after observation of the primary endpoint is sometimes considered. The comparability between arms for the assessment of post switch endpoints like overall survival (OS) is then compromised. The Intention-To-Treat (ITT) approach, currently the gold standard for the analysis of such trials, is biased to the null. Other methods accounting for switch to experimental arm have been developed like Inverse Probability of Censoring Weighted log rank test (IPCW), Rank Preserving Survival Failure Time models (RPSFT) or Iterative Parameter Estimation (IPE). The aim of this paper is to evaluate the behavior of each method on treatment effect on real data of a RCT and on simulated data. Methods Avastin Use in Platinum-Resistant Epithelial Ovarian Cancer (AURELIA) trial was a two-arms RCT where treatment switch in the reference arm was allowed. The OS results, estimated with an ITT approach, were compared in terms of hazard ratio (HR) with those of RPSFT, IPE and IPCW. Chosen model for IPCW implementation was a multiple logistic regression model with 10 baseline and time varying factors. Trials were simulated by resampling the AURELIA dataset, according to factors assumed to influence the methods’ evaluations, like switch-rate, time to switch, or sample size and censor rate. Variations of those parameters were explored. Results were compared between methods for each scenario. Results All the methods aimed to correct the OS in the control arm. Whatever the method used, the AURELIA trial results were confirmed. However, the correction was more important with the three methods investigated than with the ITT approach [HR = 0.85 (0.66 ; 1.08)]. The most optimistic correction was obtained with the IPCW method [HR = 0.74 (0.56 ; 0.97)], RPSFT and IPE gave similar results [0.78 (0.61 ; 0.99)]. The factor that had the strongest impact on the HR estimation was the switch rate : a low switch rate was associated with results close to the ITT approach. A high switch rate produced a control group comparable to the experimental group in terms of survival and was thus associated with a HR close to 1. Time to switch had small influence on the HR, a long delay between switch and OS led to an HR estimation near 1, indeed this is an unfavorable case to show a treatment effect. Event rate and sample size factors had no significant influence on the HR evaluations. Globally the simulations showed that all methods gave consistent results, no major differences were observed compared to the ITT approach. Conclusions Even if the treatment switch is an interesting option to allow early access to treatments expected as the most effective, biases on the secondary criteria estimation are important and existing statistical methods to correct these biases are complex and have some limitations. At the design stage, it is thus preferable not to include switch from a data interpretation perspective. If a switch is scheduled, the planned methods should be specify in the protocol and the statistical analysis plan. Therefore data needed for the methods implementation should be carefully collected. Results should always be presented using the ITT approach, the other methods should be considered as sensitivity analyses.