Activity Number:
|
284
|
Type:
|
Contributed
|
Date/Time:
|
Tuesday, August 4, 2009 : 8:30 AM to 10:20 AM
|
Sponsor:
|
Section on Survey Research Methods
|
Abstract - #305566 |
Title:
|
Assessing the Convergence of Multiple Imputation Algorithms Using a Sequence of Regression Models
|
Author(s):
|
Jian Zhu*+ and Trivellore E. Raghunathan
|
Companies:
|
University of Michigan and University of Michigan
|
Address:
|
Department of Biostatistics, SPH II, Ann Arbor, MI, 48109,
|
Keywords:
|
Missing Data ; Multiple Imputation Using a Sequence of Regression Models ; Incompatible Conditional Distributions
|
Abstract:
|
Multiple imputation algorithms using a sequence of regression models are commonly used to handle non-responses in complex survey studies. Although such algorithms have several advantages over joint modeling of all survey variables, they have a theoretical limitation that the specified conditional distributions could be incompatible and the underlying joint distribution of the survey variables may not exist. Although previous simulation studies show that imputation algorithms using incompatible conditional distributions work well for some cases, the performance of such algorithms for complex data needs to be studied. We focus on general multivariate data to assess the convergence properties of the imputation algorithms using various types of conditionally specified models. We also evaluate the impact of incompatible models on imputation results through simulation studies.
|