Abstract:
|
Receiver operating characteristics (ROC) curves are usually used to evaluate predictive performance of binary classifiers by plotting sensitivities vs specificities at each potential threshold. When known measurement errors exist in predictors, the simulation extrapolation (SIMEX) method can be used to build a classifier with measurement error models; however, in this case, it is not straightforward to construct the resultant ROC curves. We propose to integrate a series of ROC curves, obtained from a data-driven Monte Carlo simulation method, to a ROC band with certain confidence level. Different ways of confidence band construction, such as simultaneous joint confidence regions and fixed-width bands, are evaluated through extensive simulation studies. We identified one framework that is robust and can maintain a pre-specified coverage rate well. Common misconceptions and pitfalls of the band construction and usage of the resultant classifiers are also discussed. A real data example, which aims at predicting major depressive disorder using brain imaging data with measurement errors, is provided to demonstrate the applicability of our framework.
|