In Reply: The relevant end point for cardiovascular risk prediction models must include myocardial infarction, stroke, revascularization procedures, and cardiovascular death; this is what is of concern to patients and is the end point of most intervention trials. We thus disagree with Dr Pencina and colleagues that our data in Table 5 has limited merit. While we agree that Table 4 is more appropriate for model assessment, we included Table 5 because the ATP III model is most often used in clinical practice. Furthermore, as shown in Table 4 where algorithms were directly compared using a common end point, we again found better performance for the new models than for the old. Although they focus on a 54-patient subgroup in the test data set, they do not consider the 415 women in the other 5 cells for whom the vast majority are reclassified correctly. A Hosmer-Lemeshow goodness-of-fit test1 can formally compare observed and expected probabilities in the cross-classified categories. Using this test in Table 4 where a common end point is used, the model using ATP III covariates exhibited significant lack of fit, while Models A and B show no such deviation.
Ridker PM, Cook NR. Algorithms for Assessing Cardiovascular Risk in Women—Reply. JAMA. 2007;298(2):175-178. doi:10.1001/jama.298.2.177-b