Risk prediction equations are used in a variety of clinical fields to aid in guiding therapy and determining courses of action. The Framingham risk score for coronary heart disease, for example, has been widely used for establishing guidelines for treatment with cholesterol-lowering medications. There are several considerations needed in building such models, and the paper by Harrell et al (1996) is a classic tutorial providing an excellent overview on how to build and evaluate such models.
Whether new predictors can add utility is an important question in clinical epidemiology and many fields of medical research. While an odds ratio can indicate a strong association with later disease, it does not necessarily translate into clinical utility. Pepe et al (2004) describe the relationship of the OR to the area under the receiver operating characteristic (ROC) curve, which has been the primary tool to compare models. Cook (2007) discusses some limitations of the ROC curve and describes the alternative of risk reclassification. Pencina et al (2008) develop new measures based on risk reclassification, including the net reclassification improvement (NRI) and integrated discrimination improvement (IDI). Vickers et al (2016) describe the concept of net benefit and decision curves, which aid in decision making by balancing risks and benefits. The article by Steyerberg et al (2010) provides an overview of methods to evaluate and compare predictive models, and the TRIPOD statement provides guidance on how to report prediction models.