Multiple imputation was a valid approach to estimate absolute risk from a prediction model based on case–cohort data

Objective To compare weighting methods for Cox regression and multiple imputation (MI) in a case–cohort study in the context of risk prediction modeling. Study Design and Setting Based on the European Prospective Investigation into Cancer and Nutrition Potsdam study, we estimated risk scores to predict incident type-2 diabetes using full cohort data and case–cohort data assuming missing information on waist circumference outside the case–cohort (∼90%). Varying weighting approaches and MI were compared with regard to the calculation of relative risks, absolute risks, and predictive abilities including C-index, the net reclassification improvement, and calibration. Results The full cohort comprised 21,845 participants, and the case–cohort comprised 2,703 participants. Relative risks were similar across all methods and compatible with full cohort estimates. Absolute risk estimates showed stronger disagreement mainly for Prentice and Self & Prentice weighting. Barlow and Langholz & Jiao weighting methods and MI were in good agreement with full cohort analysis. Predictive abilities were closest to full cohort estimates for MI or for Barlow and Langholz & Jiao weighting. Conclusions MI seems to be a valid method for deriving or extending a risk prediction model from case–cohort data and might be superior for absolute risk calculation when compared to weighted approaches.