Two-stage multiple imputation
by
Satkartar Kinney
Duke University
Coauthors: Jerome Reiter

Multiple imputation was developed for handling missing data and also has been found useful in other applications, such as measurement error and protecting confidentiality in public use datasets. In some instances is desireable to impute two partitions of data separately. This approach has been found useful for improving computational efficiency, lending insight into different sources of variability, allowing for the use of multiple imputation to address two applications simultaneously, and other cases where different numbers of imputations are desired for different partitions of data. In order to evaluate uncertainty properly, the imputations must be generated in a nested manner, and different combining rules than those for single-stage imputation are required. We review the use of two-stage multiple imputation and develop large-sample significance tests for multicomponent hypotheses. These tests are based on a reference F-distribution derived from Taylor series expansions and moment matching.

PDF

Date received: July 27, 2007