On the design and the analysis of stratified biomarker trials in the presence of measurement error.

A major emphasis in precision medicine is to optimally treat subgroups of patients who may benefit from certain therapeutic agents. And as such, enormous resources and innovative clinical trials designs in oncology are devoted to identifying predictive biomarkers. Predictive biomarkers are ones that will identify patients that are more likely to respond to specific therapies and they are usually discovered through retrospective analysis from large randomized phase II or phase III trials. One important design to consider is the stratified biomarker design, where patients will have their specimens obtained at baseline and the biomarker status will be assessed prior to random assignment. Regardless of their biomarker status, patients will be randomized to either an experimental arm or the standard of care arm. The stratified biomarker design can be used to test for a treatment-biomarker interaction in predicting a time-to event outcome. Many biomarkers, however, are derived from tissues from patients, and their levels may be heterogeneous. As a result, biomarker levels may be measured with error and this would have an adverse impact on the power of a stratified biomarker clinical trial. We present a trial design and an analysis framework for the stratified biomarker design. We show that the naïve test is biased and provide bias-corrected estimators for computing the sample size and the 95% confidence interval when testing for a treatment-biomarker interaction in predicting a time to event outcome. We propose a sample size formula that adjusts for misclassification and apply it in the design of a phase III clinical trial in renal cancer.