In her recent Journal article, Pepe (1) takes a strong stance against reclassification, particularly the reclassification calibration test. Although I agree with several comments, I disagree with her dismissal of this test. Pepe's mathematical assumptions are based on the average predicted risk from the full model equaling the observed risk within reclassified categories. There are several reasons beyond sampling variation why this may not hold in practice. The model may not provide a good fit within the ranges of interest. This is particularly true when interest focuses on those at intermediate or high risk within generally low-risk population-based cohorts. The extent of reclassification may also be too small and random in the higher, more meaningful categories.
展开▼