|
Abstract:
|
Firstly, we investigate how calibration of prediction rules can be combined with use of multiple imputation to account for missing predictor observations. Second, we present methods that can be implemented with current multiple imputation software in a pragmatic manner, while allowing for unbiased predictive assessment through validation on new observations for which outcome is not yet available. We focus on the methodological foundations of multiple imputation as a model estimation approach as opposed to a purely algorithmic description, contrasting application of multiple imputation for parameter (effect) estimation with predictive density calibration. Two approaches are formulated based on this review, of which the second utilizes application of the classical Rubin's rules for parameter estimation, while the first approach averages probabilities from models fitted on single imputations to directly approximate the predictive density for future observations. Pragmatic implementations are presented. Results from 2 datasets and simulations show substantial reduction in prediction variance for predictive averaging as opposed to use of classical Rubin's rules model estimation.
|
