|Title||Biased sampling and its applications in biomarker validation|
|Year of Publication||2011|
Biomarkers play an important role in disease diagnosis, prognostic stratification and treatment selection. Before a biomarker can be utilized in clinical practice, its predictive accuracy with respect to relevant clinical conditions needs to be validated against empirical data. Simple random sampling and case-control sampling are two commonly used conventional study designs. We will discuss the weakness of such designs. Motivated by the need for novel designs with better efficiency and reduced cost, we illustrate why test-dependent sampling (TDS) and test/auxiliary-dependent sampling (TADS) can be more efficient in some practical cases. In these new designs, depending on the study objective, investigators can oversample or undersample subjects falling into certain ranges of the biomarker score. As special types of biased sampling scheme, TDS and TADS designs will introduce bias when standard ROC and AUC methods are used to estimate the predictive accuracy of a biomarker. We will introduce new statistical methods for ROC and AUC estimation under TDS and TADS, establish the asymptotic properties of the proposed estimators, and illustrate their finite sample properties via simulation. Simulation studies will also be used to illustrate improved precision under the new designs over conventional designs. The proposed methods will be applied to a hypothetical dataset related to an ongoing lung cancer clinical trial.