Extending Multiple Imputation of a Clinical Trial Outcome to Nonparametric Methods

• Activity Number: 505 - Missing Data and Multiple Imputation in Clinical Trials
• Type: Contributed
• Date/Time: Wednesday, August 2, 2017 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Kimberly Walters* and Lisa Weissfeld
• Companies: Statistics Collaborative, Inc. and Statistics Collaborative, Inc.
• Keywords: clinical trial ; missing data ; mutliple imputation ; nonparametric

Abstract

Multiple imputation is often used as a sensitivity or supporting analysis in the clinical trials setting when outcome data are subject to missingness. This method hinges on the need for an estimator in order to apply Rubin's rule, making it impossible to apply in settings where exact statistical methods or nonparametric methods are used for the primary analysis as these methods generally provide a significance level but no estimate. This project compares single and multiple imputation strategies that can be employed when the analysis technique is nonparametric, with an emphasis on the Wilcoxon rank sum test, also called the Mann Whitney Wilcoxon test. Multiple imputation is extended from the parametric estimator setting to the Wilcoxon test through the use of the Hodges Lehmann estimator to estimate the difference in central tendencies in the two treatment groups. Simulation studies assess the relative method performance. A sensitivity analysis compares the non-parametric approach to an analogous inference under the assumption of normality using parametric methods. Extensions to the problem, such as stratification or a missing at random model, are explored.