Sensitivity Analyses for Informative Censoring in Time-to-Event Clinical Trials

The assumption of censoring at random for analyses of time to event data in the presence of informative censorings can lead to a biased estimation of the likelihood and therefore biased estimates in the cox regression. This thesis presents three different approaches to analyse the sensitivity of time-to-event data to informative censorings. The censoring at random (CAR) method of multiple imputation via Kaplan Meier imputation (KMI) serves as a starting point for these methods. On this basis three approaches are introduced and their implementation in SAS is explained in detail. Finally, all methods are applied to a real world data set and the results are discussed. The tipping point analysis adjusts the Kaplan Meier curve of the active treatment group used for the KMI by raising it to the power of a δ. This δ is gradually increased until the difference in the survival curve of the active treatment group and the reference group is no longer significant. The smallest value of δ which fulfills this is called the tipping point. If the resulting tipping point is not clinically reasonable, the original analysis of the data can be called robust to the CAR assumption [14]. However, it might be the case that no such δ exists, an example of which is given in the thesis. The second approach, reference-based imputation, is based on the assumption that from the time point of their drop out onwards, patients who drop out of the active treatment group have the same risk for having an event as the patients in the reference group. Under this assumption, KMI for the active treatment group is performed using the Kaplan Meier curve of the reference group. The aim of this approach is to test if the difference in the survival curves of the active treatment group and the reference group is still significant after the imputation. Finally, a pattern imputation approach is introduced. The implemented program allows the user to define patterns of censored patients who are believed to behave similarly after censoring. For every pattern the adjustments to the Kaplan Meier curves for the KMI introduced in the first two approaches can be applied to all treatment groups. This means every treatment can be imputed by using the Kaplan Meier curve of any treatment and a predefined δ can be applied. Pattern imputation can be used for sensitivity analyses by making conservative assumptions for the patterns. It can also give an impression of how the data might look like without censorings if the assumptions for the pattern are based on realistic reasons.