Editorial for the special issue of “Statistical Methods in Medical Research” on “Advanced ROC analysis”

Receiver Operating Characteristic (ROC) analysis is widely used in diverse scientific fields for the evaluation of the quality of a diagnostic test/biomarker or classifier score/model. ‘‘Quality’’ is measured through indices of diagnostic accuracy that are widely used in the ROC framework, such as the Area Under the ROC Curve, the Youden Index, etc. This versatile field of research has attracted the interest of a wide audience of researchers for the development of ROC related methods and their implementation in given applied problems. Figure 1 presents the results of a recent (September 2017) PubMed search using the term ‘‘Receiver Operating Characteristic’’ by Year of publication for the last decade. The increasing use of ROC methods is evident. As the use of ROC methods increases, the need for more sophisticated/advanced ROC methods also increases. However, the implementation of such methods can be quite complex. Consequently, researchers end up with ad hoc methods for a given applied framework without the possibility of fast and easy implementation. In this special issue on advanced ROC methods, we have tried to compile a set of articles dealing with ROC methods in multiple-class classification problems and in ROC regression/ROC assessment in the presence of covariate information problems. The former are not currently supported in proprietary statistical software, while the latter have a basic support in some software such as Stata, SAS, JMP. All articles in the special issue are accompanied with software for implementation of the described methods in R. We hope that this strategy will help researchers implement described methods in relevant applied problems while, equally importantly, it will also facilitate the further development of ROC methods in the pursuit of the advancement of this versatile field of research. In the standard two-class ROC framework, there is increasing interest in the assessment of diagnostic markers in the presence of covariates. Sophisticated ROC regression methods have been developed during the last decade as a result. The article by Martı́nez Camblor and Pardo Férnandez deals with smooth time-dependent ROC curve estimators, also describing a relevant R package (smoothROCtime) developed by the authors as a support. Similarly, the work by Rodrı́guez Álvarez et al. describes bootstrap methods for inference in nonparametric ROC regression and describes the use of the relevant R package (npROCRegression) developed by the authors for this purpose. Ready-to-use R packages are thus developed and described for the implementation of sophisticated ROC regression methods. Although most ROC work has focused on the two-class problem, recently there has been increased interest in the threeand multiple-class scenarios. For example, it is now standard practice in cognition research to consider three classes of subjects by default, i.e. those with normal cognition, subjects with mild cognitive impairment, and demented subjects. The article by Zhang and Alonzo focuses on the estimation of the Volume Under the ROC Surface (VUS) in the presence of verification bias, while the work by Xiong et al. develops on the estimation of the VUS for clustered measurements in three diagnostic groups. The article by Inacio and Branscum develops on Bayesian nonparametric inference for the Generalized (three-class) Youden Index (GYI). Finally, the article by Yin et al. develops and compares new and existing methods for inference in three-class classification problems using the VUS and GYI. This set of four articles offers an overview of the three-class ROC methodology, while each article is followed by R code for the implementation of the studied methods.