Does the Net Reclassification Improvement Help Us Evaluate Models and Markers?

Few statistical methods have been as rapidly adopted asthe net reclassification improvement (NRI). The rea-sons that the NRI has become so popular so quickly are not hard to fathom. First, it is an appealing concept to investigators. Instead of reporting that their new marker improved the area under the receiver-operating character-istic curve (AUC)—a statistical abstraction at the best of times—they can say that their marker led to “reclassifica-tion, ” suggesting clinical value. The popularity of the NRI also stems from a belief that the AUC is insensitive. For example, in an early paper on the NRI, Pencina and col-leagues stated that the AUC is “too conservative ” because it “hardly moves after a few good risk factors are already included in the model ” (1). The reason that the AUC “hardly moves ” may not be because we are evaluating in-appropriately but because it is not easy to improve on