Correcting Estimation Bias in Randomized Clinical Trials with a Test of Treatment-by-Biomarker Interaction
The key elements in recent drug development aiming toward personalized medicine involve the development of baseline predictive biomarkers to predict the responsiveness to new treatments. When biological evidence to support the biomarker hypothesis is not strong, it is particularly important to evaluate the clinical validity of a predictive biomarker based on data from clinical trials. All-comers randomized clinical trials stratified with biomarker measurements can provide the opportunity to evaluate the predictive value of the biomarker, and a test of treatment-by-biomarker interaction is applied for this purpose. If the interaction is significant, it is natural to proceed to a subgroup analysis that evaluates treatment efficacy within biomarker-based subgroups. However, estimation bias can arise when using the standard estimation method without regard to the significant interaction because of the correlation between the interaction test and subgroup analysis. In this paper, after evaluating the bias function for the standard estimators, we provide bias-corrected point and interval estimation methods, including polynomial approximations of the bias function previously developed for a fallback analysis plan. The magnitude of bias reduction and precision of the proposed estimators are assessed via simulations. Applications to biomarker-stratified, randomized phase III trials in oncology are provided.