Abstract:
|
Development of regression methods in diagnostic testing has provided the capability to explore factors influencing accuracy, which is of great importance for making decisions about test implementation. Three fundamentally different modelling approaches have been proposed: models of test results, models of ROC curves, and models of summary measures of ROC curves. These differ with respect to the assumptions made, the type and amount of information each provides, scientific robustness and statistical efficiency. However, there is no consensus about which approach is best. We present a semi-parametric regression method for the AUC and partial AUC--two summary measures of the ROC curve. We discuss results demonstrating that the method is more scientifically robust than the methods that make more modelling assumptions, while it has similar power to detect covariate effects. The other two approaches are more informative, but are more likely to provide misleading results. Assurance of model fit under the more informative approaches would overcome this; however, such methods have not been sufficiently developed.
|