Title: Distributional ROC surface regression
Authors: Vanda Inacio - University of Edinburgh (United Kingdom) [presenting]
Nadja Klein - Humboldt University Berlin (Germany)
Abstract: Accurate diagnosis of disease is of great importance in clinical practice and medical research. The receiver operating characteristic (ROC) surface is a popular tool for evaluating the discriminatory ability of continuous diagnostic test outcomes when there exist three ordered disease classes (e.g., no disease, mild disease, advanced disease). Incorporating covariates in the analysis can potentially enhance information gathered from the diagnostic test, as its discriminatory ability may depend on these. We propose a Bayesian distributional regression approach for covariate-specific ROC surface estimation. In the model specification, the covariate-specific ROC surfaces are indirectly modelled using probabilistic distributional models capturing location, scale, shape, and other aspects of the diagnostic test distribution in each of the three groups, where covariate effects are modelled through penalised splines. Multiple simulation studies demonstrate the ability of the model to successfully recover the true covariate-specific ROC surface and the corresponding covariate-specific VUS in a variety of complex scenarios. The methods are motivated by and applied to a prostate cancer study where the main goal is to assess if and how the accuracy of a new diagnostic test, the prostate health index density, changes with age.