Multiple testing strategies for clinical trials with co-primary endpoints in all comers and a subgroup.

e14041 Background: In this new era of precision medicine, subgroup analyses have become increasingly critical for the success of oncology drug development. When the treatment benefit is still uncertain in the subgroup, sometimes registrational trials are designed to have co-primary endpoints, e.g. the same endpoint tested in both the all comer population and a (biomarker specified) subgroup with the family-wise Type 1 error being strongly controlled. Many multiple testing methods have been proposed and applied to address multiplicity issues for such designs. In this work we did a comprehensive review of current methods and compared their advantages and disadvantages. Methods: In a simple scenario of testing one endpoint in an oncology trial, assuming known sample size, randomization ratio and hazard ratio in subgroup and all comers, respectively, we reviewed six alpha spending methods including Bonferroni's method, fixed sequence method, Hochberg's step up method, Holm's step-down method, alpha fall back method, and flexible alpha spending method(by Alosh and Huque) as well as their variations by taking the correlation between the two populations into consideration. Comparisons were made in terms of the power of rejecting the null hypothesis in all comers, subgroup, or either group under different parameter settings. The optimal alpha sharing strategy under different scenarios was then explored. Results: A function in R was developed for the comparison of six methods. Hochberg's method consistently outperforms Bonferroni's and Holm's method in all scenarios. By taking consideration of the correlation between subgroup and all comers, the performance of Bonferroni's, Hochberg's and Holm's methods are all improved. The choice of an optimal method strongly depends on the strategic context of each trial. But among the six methods including the variations discussed in this work, flexible alpha spending method by Alosh and Huque seems to outperform other methods when alpha is appropriately allocated between the subgroup and all comers. Conclusions: Our work provided a practical tool for users to compare the pros-and-cons and make trade offs between different alpha-spending methods under possible scenarios.