Abstract
Identifying and quantifying predictive biomarkers is a critical issue of personalized medicine approaches and patient-centric clinical development strategies. In early stages of the development process, significant challenges and numerous uncertainties arise. One of the challenges is the ability to assess the predictive value of a biomarker, i.e., the difference in primary outcomes between experimental and placebo arms above and below a certain threshold of the biomarker. Indeed, when the accumulated information is very limited and the sample size is small, preliminary conclusions about the predictive properties of the biomarker might be misleading. To date, the majority of investigations regarding the predictiveness of biomarkers were in the setting of moderate-to-large sample sizes. In this work, we propose a novel flexible approach inspired by the Kolmogorov-Smirnov Distance in order to assess the predictiveness of a continuous biomarker in a clinical setting where the sample size is small. Via simulations we show that the proposed method allows to achieve a higher power to declare predictiveness compared to the existing methods under a range of scenarios, whilst still maintaining a control of the type I error at a pre-specified level.