Abstract
Risk prediction models that estimate an individual's risk of developing colon cancer could be used for a variety of clinical and public health interventions, including offering high-risk individuals enhanced screening or lifestyle interventions. However, if risk prediction models are to be translated into actual clinical and public health practice, they must not only be valid and reliable, but also be easy to use. One way of accomplishing this might be to simplify the information that users of risk prediction tools have to enter, but it is critical to ensure no resulting detrimental effects on model performance. We compared the performance of a simplified, largely categorized exposure-based colon cancer risk model against a more complex, largely continuous exposure-based risk model using two prospective cohorts. Using data from the Nurses' Health Study and the Health Professionals Follow-up Study we included 816 incident colon cancer cases in women and 412 in men. The discrimination of models was not significantly different comparing a categorized risk prediction model with a continuous prediction model in women (c-statistic 0.600 vs. 0.609, P (diff) = 0.07) and men (c-statistic 0.622 vs. 0.618, P (diff) = 0.60). Both models had good calibration in men [observed case count/expected case count (O/E) = 1.05, P > 0.05] but not in women (O/E = 1.19, P < 0.01). Risk reclassification was slightly improved using categorized predictors in men [net reclassification index (NRI) = 0.041] and slightly worsened in women (NRI = -0.065). Categorical assessment of predictor variables may facilitate use of risk assessment tools in the general population without significant loss of performance.