LAUSR

We read with interest the report by Karnes and colleagues [1]. The authors assessed the predictive performance of the Decipher genomic classifier (GC) for the risk of prostate cancer–specific mortality (PCSM). Their results demonstrated that the Decipher GC had an area under the receiver operating characteristic (ROC) curve (AUC) of 0.73 for PCSM within 10 yr of radical prostatectomy. Addition of the GC yielded AUC of 0.76 for a model with the Cancer of the Prostate Risk Assessment Postsurgical Score (CAPRA-S) and AUC of 0.79 for a model with individual clinical variables [1]. Although the results are very interesting, some methodological and statistical issues should be noted. The authors mention that the prediction performance will be improved in models that include the GC with individual clinical variables and CAPRA-S. However, such a conclusion is a simple and easy interpretation, because empirical comparisons of AUCs provided no indication of superiority of prediction performance. When several ROC curves are plotted for different tests on the same data set, we can compare the statistical difference between AUCs using efficient statistical methods, accounting for correlation between AUCs derived from the same data set [2,3]. Hence, we suggest that the authors should reanalyze their data to assess whether the difference between reported AUCs is statistically significant.