A robust approach to ROC curves with covariates

BIANCO, ANA MARIA

Guayaquil

Congreso; International Conference on Robust Statistics 2019; 2019

ROC curves are a useful tool to measure the dis-criminating power of a continuous variable. They usually quantify the accuracy of a pharmaceutical or medical test to distinguish between two conditions or classes. ROC curves can also be extended to other general statistical situations such as classification or discrimination, where we typically have a set of individuals or items assigned to one of two classes on the basis of disposable information of that individual. Assignations are not perfect and may lead to classification errors. At this point, ROC curves become an interesting strategy either to evaluate the quality of a given assignment rule or to comparetwo available procedures.In practical situations, the discriminatory eectiveness of the marker or test under study may be affected by several factors. When for each individual there is additional informa-tion contained in registered covariates, it is sensible to include them in the ROC analysis.This talk aims to show the instability of the conditional ROC curve in presence of outliers and also to provide robust estimators for it when covariates are available. We focus on a semiparametric approach which, on one side, fits a location-scale regression model to the diagnostic variable and, on the other, considers empirical estimators of the regression residuals distributions. To this end, we combine robust parametric estimators with weighted empirical distribution estimators based on an adaptive procedure that down-weights outliers.We will discuss some aspects concerning consistency. Through a Monte Carlo study we compare the performance of the proposed estimators with the classical ones both, in clean and contaminated samples. The simulation results show that our robust procedure yields reliable results under different contaminated scenarios.