Primary Submission Category: Randomized Studies
Selective randomization inference for adaptive studies
Authors: Tobias Freidling, Zijun Gao, Qingyuan Zhao,
Presenting Author: Tobias Freidling*
Many clinical trials follow a design with multiple stages: After each stage, the data is provisionally analysed and – based on these results – the recruitment of participants for the next stage as well as the administered treatment is chosen adaptively. For instance, we may want to exclude poorly performing drugs early or gather more samples from a certain subpopulation that shows a potentially beneficial response.
Analysing such adaptive studies is challenging as the data is used twice: (1) for selection of the design of later stages and the null hypothesis, (2) for testing the null hypothesis with the data generated under the chosen design. Since the data generating mechanism and null hypothesis are not pre-specified, classical statistical methods do not provide valid inference.
Existing solutions are often limited in scope and usually specific to a certain design; in this work, we propose a general framework that can handle all kinds of designs and adaptive choices. Our approach uses concepts from the post-selection inference literature to develop a selective randomization p-value. Noteably, we do not require any assumptions on the law of the outcomes and covariates or on the dependence structure between different participants. We show that our method improves power compared to other valid tests while still controlling the selective type-I error. Moreover, we construct confidence intervals and discuss different methods to compute the selective randomization p-value.