To the Editor Wang et al1 examined the ability of the receiver operating characteristic (ROC), which assesses discrimination (order of predictions) but not calibration (how close the predictions are to the observed value), to assess a model鈥檚 predictive accuracy. They concluded that the ROC 鈥渄oes not directly quantify the prediction accuracy of the model鈥�1 because it does not assess calibration. They proposed quantifying a model鈥檚 predictive accuracy using the mean absolute difference (MAD), which does use calibration.