Statistical Considerations in the Evaluation of Continuous Biomarkers

Discovery of biomarkers has been steadily increasing over the past decade. Although a plethora of biomarkers has been reported in the biomedical literature, few have been sufficiently validated for broader clinical applications. One particular challenge that may have hindered the adoption of biomarkers into practice is the lack of reproducible biomarker cut points. In this article, we attempt to identify some common statistical issues related to biomarker cut point identification and provide guidance on proper evaluation, interpretation, and validation of such cut points. First, we illustrate how discretization of a continuous biomarker using sample percentiles results in significant information loss and should be avoided. Second, we review the popular "minimal-P-value"approach for cut point identification and show that this method results in highly unstable P values and unduly increases the chance of significant findings when the biomarker is not associated with outcome. Third, we critically review a common analysis strategy by which the selected biomarker cut point is used to categorize patients into different risk categories and then the difference in survival curves among these risk groups in the same dataset is claimed as the evidence supporting the biomarker's prognostic strength. We show that this method yields an exaggerated P value and overestimates the prognostic impact of the biomarker. We illustrate that the degree of the optimistic bias increases with the number of variables being considered in a risk model. Finally, we discuss methods to appropriately ascertain the additional prognostic contribution of the new biomarker in disease settings where standard prognostic factors already exist. Throughout the article, we use real examples in oncology to highlight relevant methodologic issues, and when appropriate, we use simulations to illustrate more abstract statistical concepts.