Robust estimation of ROC curves with covariates

GONZALEZ-MANTEIGA, WENCESLAO; BIANCO, ANA; BOENTE, GRACIELA

Londres

Conferencia; XII International Conference of the ERCIM Working Group on Computing & Statistics (ERCIM2019); 2019

Birbek University of London

Receiver Operating Characteristic (ROC) curves are a useful graphical tool to measure the discriminating power of a continuous variable, such asdiagnostic variable or a marker. They are employed to quantify the accuracy of the marker to distinguish between two conditions or classes. Aswith any classifier, the assignations are not perfect and may lead to classification errors. In practical situations, the discriminatory effectiveness of the marker under study may be affected by several factors. When for each individual additional information is available, it is sensible to include itin the ROC analysis. The aim is to show the instability of the conditional ROC curve in presence of outliers and also to provide robust estimatorswhen covariates are available. A semiparametric approach is followed, where robust parametric estimators are combined with weighted empiricaldistribution estimators based on an adaptive procedure that downweights outliers. The consistency of the proposal is discussed. Through a MonteCarlo study, the performance of the proposed estimators is compared with that of the classical ones in clean and contaminated samples.