ABSTRACT
Interest is generally in comparing the performance of a group of readers interpreting a common set of cases in two or more treatments. Such data is termed multiple-reader multiple-case (MRMC). This chapter describes the Dorfman-Berbaum-Metz (DBM) method. The other method, due to Obuchowski and Rockette, is the subject of the following chapter. Both have been substantially improved by Hillis. The DBMH approach, implemented in RJafroc, uses jackknife-derived pseudovalues as stand-ins for individual case-level figures of merit and standard analysis of variance (ANOVA) methods. Starting with mean squares, directly calculated from the pseudovalues, the procedure is to compute a ratio that is distributed under the null hypothesis (NH) of no treatment effect, as an F-statistic with unit expected value. This permits testing the significance of the observed value of the F-statistic, calculation of the p-value, and confidence intervals for the figure of merit difference. Expressions for non-centrality parameters, used in Chapter 11 for sample size estimation, are derived. Demonstrations with R software are used to illustrate the analysis, including how to validate the analysis by showing that it has the expected rejection rate under the NH. The meaning of pseudovalues is presented as evidence of why the basic assumption of the analysis is valid for the empirical AUC.
