ABSTRACT
This chapter leverages a recent advance, namely the bivariate contaminated binormal model (BCBM), which allows estimation of decision variable correlations between two readers interpreting the same cases. Unlike the bivariate binormal model for which software (CORROC2) was developed more than three decades ago, the software implementation of BCBM, named CORCBM (for correlated CBM), can fit degenerate datasets. The BCBM-based simulator generates samples, one per reader, for each case, from a multivariate normal distribution with specified mean vector and covariance matrix. The calibration of these parameters to statistically resemble the CAD vs. 9-radiologist Hupse-Karssemeijer dataset is described and implemented in R. The simulator was used to generating 2000 samples under the null hypothesis condition. For each simulation covariance and variance of the Obuchowski–Rockette single-modality method was estimated and a NH-rejection was recorded if the p-value was below 0.05. The observed fraction of rejections was close to 0.05, thereby validating the analysis of the previous chapter. The averages of the covariance and the variance were within the corresponding bootstrap confidence intervals, calculated from the original dataset. This shows that the simulator is statistically identical to the original dataset. A realistic simulator is key to proper validation of any proposed method of analyzing ROC data.
