Choosing between predictors of fractures

Abstract

The identification of those at highest risk of osteoporotic fractures is a clinical goal that requires appropriate statistical comparisons of potential predictors of fractures. This article provides a formal approach for comparing individual predictors (e.g., bone mass at one site vs bone mass at another), or sets of predictors (e.g., bone mass vs other risk factors), and contrasts newer methods, such as bootstrapping, to receiver‐operating‐characteristics (ROC) curves, which have been previously used. The advantages of the bootstrapping approach are illustrated using time‐to‐fracture data from a published study demonstrating the use of baseline bone mass measurements in the prediction of fractures in 521 subjects with variable lengths of follow‐up, extending to 12.5 years. Bone mineral density (BMD) was shown to be significantly better than bone mineral content (BMC) in predicting fractures in free‐living subjects, but not in retirement‐community subjects. Bone mineral apparent density (BMAD) was also compared with BMC and BMD and shown not to improve fracture prediction in these subjects.