ABSTRACT
Umbrella trials match biomarker sub-populations to drugs within a single indication, often in a predetermined fashion based on a pre-existing biomarker hypothesis. This creates ambiguity when a patient tests positive for more than one biomarker. In this case, the patient qualifies for more than one stratum of the umbrella trial. The patient assignment to a given stratum may then be randomized, or may be done according to a variety of fixed rules, including randomization to the stratum corresponding to the biomarker with the highest allele frequency in the patient, randomization to the stratum with the least accrual, or probabilistic randomization weighted by the inverse of the biomarker prevalence. We have performed a simulation of these umbrella randomization schemes in scenarios where the biomarkers are independent in their effects on drug response, and also in cases where the interaction between biomarkers affects drug response. In the latter cases, we have noted estimation bias when using any patient allocation scheme other than random assignment when the patient is positive for multiple markers. The significance of this finding is discussed.
