TY  - GEN
CY  - Heidelberg
Y1  - 2025///
N2  - This thesis is motivated by recent advances in oncology, where therapies rely on the activation and augmentation of the immune system to identify and fight tumor cells.
These immunotherapeutic approaches come with own characteristics and one commonly observed trait is that the observable treatment effect is delayed by the time needed to train the response of the immune system. Especially for time-to-event endpoints such as overall survival or progression-free survival this demands careful consideration with respect to the statistical evaluation as commonly used methods rely on the assumption of proportional hazards. This
assumption is violated if the treatment effect is delayed and the usual methods have reduced power to detect a difference between therapies. The aim of this thesis was hence to investigate the performance of various alternatives to the commonly used logrank test in terms of type
1 error and power in this setting. 

Firstly, a systematic literature search was performed to identify statistical methods that have been suggested to analyze time-to-event data and especially if these methods have been developed to handle non-proportional hazards.
The methods were then compared in an extensive simulation study taking the following parameters into account: 
the overall study duration, the accrual proportion, the shape of the Weibull hazard and median survival in the control arm and the maximum treatment effect  as well as the delay and changepoint of the generalized linear lag model. As performance measure type 1 error and power of the methods was assessed and the effect of the different simulation parameters on these performance measures was examined. 
For methods where parameters for analysis had to be chosen the effect of misspecifying these parameters was also investigated.

Most of the methods controlled type 1 error and those that did not were already known from previous publications.
With respect to power the methods that come close to the
data-generating process performed best in the simulation studies and were not sensitive to parameter misspecification at least in non-proportional hazard scenarios. However, the specification of these parameters requires knowledge of the form and extent of the delay that can be expected in advance. If this is not available other alternatives such as the Fleming-Harrington test for late differences or combinations of weighted logrank statistics can be used.
Their power, however, depends on the underlying distribution of the data. For proportional hazard scenarios the logrank test is the most powerful of all tests that control type 1 error.

Application of the presented methods is facilitated by referencing the R package in which the method is implemented or providing own code in an online repository referenced in the Appendix. The interpretation of the results of these methods is, however, often complicated due to the absence of interpretable summary measures of the detected treatment effect. It is also pointed out that some aspects of this thesis such as sample size calculation or application
of the methods in adaptive or group-sequential designs could be worth future research. 

It is concluded that this thesis provides a detailed and well structured compilation of al-
ternative methods to the logrank test when analyzing time-to-event data in the presence of non-proportional hazards. Furthermore, the extensive simulation study together with the ranking system can help to make a substantiated choice of an appropriate method for analysis of future studies. In case of proportional hazards the logrank test remains the method of choice and in case the data-generating process of the non-proportional hazard is completely understood methods should be chosen that exploit this additional knowledge. If, however, the exact structure of the delay is unknown versatile methods that combine weighted logrank statistics should be resorted to.
TI  - Comparison of methods to analyze time-to-event endpoints when treatment effect is delayed
A1  - Behnisch, Rouven
AV  - public
UR  - https://archiv.ub.uni-heidelberg.de/volltextserver/36223/
ID  - heidok36223
ER  -