Mixed-Model Imputation of Cost Data for Early Discontinuers from a Randomized Clinical Trial

Abstract

For studies with appreciable attrition in which one expects not only differences between individual patients but also time trends in repeated measures, a “sophisticatedly simple” approach to imputation of missing values is illustrated. A linear model having random patient intercept and slope terms as well as fixed effects for treatment, investigator, time, and interactions between both treatment-investigator and treatment-time is used. In contrast to a purely fixed effects approach, mixed-model estimation then optimally shrinks patient-specific differences toward zero. This shrinkage moves the predictions for each patient toward the average time line for the corresponding investigator and treatment. Using a variety of sensitivity analyses, it is established that imputation of missing values using these mixed-model predictions provides a “benchmark” lower limit for cost differences between treatments. To illustrate concepts, supplementary analyses of selected cost and effectiveness outcomes from a randomized, double-blind trial of olanzapine versus haloperidol for the treatment of schizophrenia are presented.