Sakarya Tıp Dergisi (Dec 2015)
Covariate Adjusted ROC Curve Analysis and An Application
Abstract
Objective: Aim of this study is to analyze the change of the area under the adjusted ROC (AdjROC) curve in certain conditions via binormal distribution model using simulation studies and application of this algorithm to real data. Materials and Methods: Data sets simulated according to various conditions. PSA and age values of 125 patients who were examined prostate biopsy with pre-diagnosis of prostate cancer in Gaziosmanpasa University Faculty of Medicine Department of Urology at the years of 2005 to 2007. An algorithm and code program was written that make simulation according to various condition using PROC IML procedure in SAS statistical software.Results: According to the simulation study, if biomarker indicators in healthy group are constant and are lower or equal in healthy group than/to disease group, both adjusted AUC (AdjAUC) and AUC have small values and, no significant difference was found between them. The AUC was significantly larger when the biomarker indicators in disease group were higher. In addition, if the correlation between the covariate and biomarker is high in disease group and if AUC is approximately 0.75, then there is significant difference between adjusted AUC and AUC. PSA (Prostate Specific Antigen), a biomarker used for prostate cancer diagnosis, was analyzed based on the adjustments by age. It was found that adjusted AUC value was higher than unadjusted AUC value. Conclusions: For the adjusted ROC model being applicable, covariate and biomarker distributions must show double binormal distribution. If the biomarker can distinguish disease and healthy individuals correctly, then covariate is not needed. If correlation of healthy is approaching to 0 and correlation of disease is 0.50, and if AUC is less than 0.75, then covariate must be included in the model. Model does not work well when sample size of disease and healthy are less than 50.
