The receiver operating curve (ROC) gives a representation of sensitivity and specificity of a prediction model when varying the cutpoint of a decision rule on a whole spectrum. Evaluation of two models established (or tested) in the same population of patients warrants a valid statistical comparison of their ROC curves. Hanley et al. recently provided a method for overall comparison of ROC curves (J. A. Hanley and B. J. McNeil, Radiology 148, 839-843, 1983). Often ROC curves cross, or differ in only a part of their courses. Bootstrapping of ROC curves is proposed as a graphical check for the statistical significance of differences confined to a part of the curve. An example comparing two models of prediction of coronary artery disease progression is given to illustrate this new approach.