Performing Hypothesis Tests with Multiply Imputed Data

• Activity Number: 177
• Type: Contributed
• Date/Time: Monday, August 2, 2010 : 10:30 AM to 12:20 PM
• Sponsor: IMS
• Author(s): Xianchao Xie*+ and Xiao-Li Meng
• Companies: Harvard University and Harvard University
• Keywords: Multiple Imputation ; Hypothesis Test ; Complete-data Testing Statistics

Abstract

Multiple imputation (Rubin, 1987), a method originally designed to handle incomplete data in public-use databases, has been extensively studied and applied in various areas. Several procedures have been proposed for hypothesis testing with multiply-imputed data (Li et al., 1991; Meng and Rubin, 1992, Barnard and Rubin, 1999). When a user has access to the complete-data point estimators or complete-data log-likelihood functions, existing methods have satisfactory performance. When only the complete-data test statistics are available, there is essentially only one method (Li et al.,1991) and it is only suitable as a screen test. This talk explores possible improvements for the latter case. Several easily implementable methods are proposed for the scalar case, but generalizing them to the multi-parameter cases is a great challenge.