Robust and flexible inference for the covariate-specific ROC curve

TL;DR

This paper introduces a robust, flexible model for covariate-specific ROC curve inference that handles outliers and nonlinear effects, demonstrated through simulations and a prostate cancer study.

Contribution

It develops a location-scale additive regression model using B-splines and M-estimation to improve ROC analysis robustness and flexibility.

Findings

Successfully recovers true covariate-specific ROC curves in simulations

Effectively handles outliers and nonlinear covariate effects

Applied to prostate cancer data to assess age-related discrimination changes

Abstract

Diagnostic tests are of critical importance in health care and medical research. Motivated by the impact that atypical and outlying test outcomes might have on the assessment of the discriminatory ability of a diagnostic test, we develop a flexible and robust model for conducting inference about the covariate-specific receiver operating characteristic (ROC) curve that safeguards against outlying test results while also accommodating for possible nonlinear effects of the covariates. Specifically, we postulate a location-scale additive regression model for the test outcomes in both the diseased and nondiseased populations, combining additive cubic B-splines and M-estimation for the regression function, while the residuals are estimated via a weighted empirical distribution function. The results of the simulation study show that our approach successfully recovers the true covariate-specific ROC curve and corresponding area under the curve on a variety of conceivable test outcomes contamination scenarios. Our method is applied to a dataset derived from a prostate cancer study where we seek to assess the ability of the Prostate Health Index to discriminate between men with and without Gleason 7 or above prostate cancer, and if and how such discriminatory capacity changes with age.