|
Abstract:
|
The c-statistic, or area under the ROC curve, is popular in diagnostic testing, where the goal is to discriminate diseased from non-diseased patients. The c-statistic may not be optimal for assessing models predicting future risk or stratifying individuals into risk categories, where calibration is as important in assessing risk accurately. Perfectly calibrated models can, in fact, usually only achieve values for the c-statistic well below 1. Accepted cardiovascular risk factors, such as lipids, hypertension and smoking, individually have minimal impact on the c-statistic, yet can more accurately reclassify large proportions of patients into higher or lower clinical risk categories. The joint distributions of risk predictors, along with the calibration of risk within cross-classified categories, can help determine which models more accurately predict risk for individual subjects.
|
