Recent developments in genomics and proteomics enable the discovery of biomarkers that allow identification of subgroups of patients responding well to a treatment. One currently used clinical trial design incorporating a predictive biomarker is the so-called biomarker strategy design (or marker-based strategy design). Conventionally, the results from this design are analysed by comparing the mean of the biomarker-led arm with the mean of the
randomised arm. Several problems regarding the analysis of the data obtained from this design have been identified in the literature. In this paper, we show how these problems can be resolved if the sample sizes in the subgroups
fulfil the specified orthogonality condition. We also propose a novel analysis strategy that allows definition of test statistics for the biomarker-by-treatment interaction effect as well as for the classical treatment effect and the biomarker effect. We derive equations for the sample size calculation for the case of perfect and imperfect biomarker assays. We also show that the often used 1:1 randomisation does not necessarily lead to the smallest sample size. Application of the novel method is illustrated using a real data example.