Adjusting for covariates in evaluating markers for selecting treatment, with application to guiding chemotherapy for treating estrogen-receptor-positive, node-positive breast cancer

In many clinical contexts, biomarkers that predict treatment efficacy are highly sought after. Such treatment selection or predictive biomarkers have the potential to identify subgroups most likely to benefit from the treatment, and may therefore be used to improve clinical outcomes and reduce medical costs. A methodological challenge in evaluating these biomarkers is determining how to take into account other variables that predict clinical outcomes, or that influence the biomarker distribution, generically termed covariates. We distinguish between two questions that arise when evaluating markers in the context of covariates. First, what is the biomarker's added value for selecting treatment, over and above the covariates? Second, what is the marker's performance within covariate strata-does performance vary? We lay out statistical methodology for addressing each of these questions. We argue that the common approach of testing for the marker's statistical interaction with treatment, in the context of a multivariate model that includes the covariates as predictors, does not directly address either question. We illustrate the methodology in new analyses of the Oncotype DX Recurrence Score, a marker used to select adjuvant chemotherapy for the treatment of estrogen-receptor-positive breast cancer.