Bias Corrected ROC . . .

The.632 error estimator is a bias correction of the bootstrap estimator which leads to an underestimation of the error when the apparent error is zero. As a consequence Efron and Tibshirani (1997) developed the.632+ bootstrap error as a modification that can handle this case. We demonstrate properties and behavior of this error estimation technique. Furthermore, we show how to apply the bootstrap method to estimate the classifiers sensitivity and specificity and demonstrate a bootstrap based ROC analysis of classification performance. An adaptation of the.632+ technique to calculate bias corrected sensitivities is straightforward and leads to.632+ bootstrap estimated ROC curves. We employ a simulation study to examine this method and its performance