|
Abstract:
|
Multiple imputation is a technique to handle missing data. It allows analysts to perform valid hypothesis tests straightforwardly with only a few imputations. However, the existing multiple imputation tests either require a restrictive assumption on the missing-data mechanism, known as an equal fraction of missing information, or require the number of imputation is large. Some of them also require access to a non-standard data-dependent computer subroutine. Besides, existing multiple imputation tests cover only the Wald test and the likelihood ratio tests but not their asymptotic equivalent, the Rao’s score test. Moreover, the combining procedures for the former two tests are not unified. In this paper, we propose a unified multiple imputation test procedure, which can be applied to the Wald test, the likelihood ratio test and the Rao’s score test. It does not require the assumption of the equal fraction of missing information nor the number of imputation tends to infinity. The only requirement for the analysts is a standard complete-data testing subroutine.
|
