The receiver operating characteristic (ROC) curve is a currently well-developed statistical tool for characterizing accuracy of medical diagnostic tests. In recent years, several authors suggest approaches referred to as ROC regression models in order to evaluate effects of factors influencing accuracy of diagnostics. Rodríguez-Álvarez
et al. (2011b) suggested an inference process of ROC regressions which formulate influences of some factors in the framework of a generalized additive model (GAM). In their approach, local linear kernel smoothers based on the cross-validation (CV) criterion are used to estimate smoothing functions. In this report, we propose an approach with penalized splines based on the restricted maximum likelihood (REML) for the function estimation. We give a detail of our method, and through a simulation, show that this approach gives better inference performance than existing methods, particularly in the small sample.
View full abstract