Bayesian Analysis of ROC Curves Using Markov-chain Monte Carlo Methods

The authors introduce a Bayesian approach to generalized linear regression models for rating data observed in the evaluation of a diagnostic technology. Such models were previously studied using a non-Bayesian approach. In a Bayesian analysis, the difficulties inherent in an ordinal rating scale are circumvented by using data-augmen tation techniques. Posterior distributions for the regression parameters—and thereby for receiver operating charactenstic (ROC) curve parameters and values, for the area under a ROC curve, differences between areas, etc.—may then be computed by Mar kov-chain Monte Carlo methods. Inferences are made in standard Bayesian ways. The methods are exemplified by a study of ultrasonography rating data for the detection of hepatic metastases in patients with colon or breast cancer (previously analyzed) and the results compared. Key words: diagnostic test; ordinal regression; sensitivity; spec ificity. (Med Decis Making 1996;16:404-411)