Kouros Owzar1, Sin-Ho Jung. 1. Department of Biostatistics and Bioinformatics, Duke University Medical Center, Durham, NC 27705, USA. kouros.owzar@duke.edu
Abstract
BACKGROUND: The primary clinical endpoint in many phase II studies in cancer is a time-to-event outcome subject to potential censoring. The decision in favor of abandoning or continuing investigation of the protocol regimen is typically based on the amount of statistical evidence suggesting an improvement compared to a given historical control. The primary statistical endpoint would typically be the median of the time-to-event distribution of the clinical endpoint. For the purpose of sample size or power calculations, software implementing parametric and nonparametric median tests, is freely available. The main assumptions are those of Exponential survival and a Uniform right-censoring mechanism. PURPOSE: The performance of the parametric and nonparametric methods is compared to that of using a binomial endpoint based on dichotomizing the event time at a clinically relevant landmark. As sufficient details on the various methods and related designs for phase II clinical design with survival endpoints are provided, this article should also serve as a comparative reference on these three methods for designing phase II studies in cancer with time-to-event endpoints. METHODS: The statistical performance, by virtue of considering the type I and II error rates, of the three methods is compared by carrying out a comprehensive simulation study. RESULTS: The parametric method may fail to control the type I error rate if the underlying survival distribution is not Exponential, while the nonparametric method may fail to control the type I error rate as the sample sizes for phase II studies are typically small. Both of these methods may be sensitive to the distribution of the censoring variable. LIMITATIONS: The results provided in this article are mostly discussed in the framework of specific examples and by using specific implementations of the tests. As such the results may not be universally generalizable. The recommended method has some drawbacks if patients are censored before the landmark of interest. CONCLUSION: A method that should be considered for the purpose of the statistical design and decision rule for phase II studies in cancer is the employment of a binomial endpoint based on dichotomizing the event time at a clinically relevant landmark.
BACKGROUND: The primary clinical endpoint in many phase II studies in cancer is a time-to-event outcome subject to potential censoring. The decision in favor of abandoning or continuing investigation of the protocol regimen is typically based on the amount of statistical evidence suggesting an improvement compared to a given historical control. The primary statistical endpoint would typically be the median of the time-to-event distribution of the clinical endpoint. For the purpose of sample size or power calculations, software implementing parametric and nonparametric median tests, is freely available. The main assumptions are those of Exponential survival and a Uniform right-censoring mechanism. PURPOSE: The performance of the parametric and nonparametric methods is compared to that of using a binomial endpoint based on dichotomizing the event time at a clinically relevant landmark. As sufficient details on the various methods and related designs for phase II clinical design with survival endpoints are provided, this article should also serve as a comparative reference on these three methods for designing phase II studies in cancer with time-to-event endpoints. METHODS: The statistical performance, by virtue of considering the type I and II error rates, of the three methods is compared by carrying out a comprehensive simulation study. RESULTS: The parametric method may fail to control the type I error rate if the underlying survival distribution is not Exponential, while the nonparametric method may fail to control the type I error rate as the sample sizes for phase II studies are typically small. Both of these methods may be sensitive to the distribution of the censoring variable. LIMITATIONS: The results provided in this article are mostly discussed in the framework of specific examples and by using specific implementations of the tests. As such the results may not be universally generalizable. The recommended method has some drawbacks if patients are censored before the landmark of interest. CONCLUSION: A method that should be considered for the purpose of the statistical design and decision rule for phase II studies in cancer is the employment of a binomial endpoint based on dichotomizing the event time at a clinically relevant landmark.