Abstract
In randomized clinical trials, when the endpoint is the change from baseline at the last scheduled visit, various parametric, semiparametric, and nonparametric methods have been developed to handle the possible missing data due to dropouts. Although the last observation carried forward (LOCF) followed by an analysis of covariance (ANCOVA) model or a rank ANCOVA model and the mixed-effects model for repeated measures (MMRM) have been extensively compared and widely used even with presence of the covariates, they may lead to biased results when the required distributional or missing technique assumptions are not satisfied. Nonparametric missing data handling methods including the mean rank imputation (MRI) method relax the underlying distributional assumption; however, when covariates are present, conditions for it to be valid have been investigated to a very limited extent. This article rigorously derives asymptotic properties of the mean rank imputation method with the presence of a nonnormal covariate. The investigated methods are applied to an illustrative phase III clinical trial. Simulation studies confirmed the better performance of the mean rank imputation method in terms of Type I error rate control and power under certain mild conditions.
Acknowledgments
The authors thank the anonymous associate editor and three referees for the constructive suggestions and comments that substantially improved the original version of this article.