|
Activity Number:
|
409
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Wednesday, August 5, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Survey Research Methods
|
| Abstract - #303975 |
|
Title:
|
Bayesian Sensitivity Analysis of Incomplete Data Using Pattern-Mixture and Selection Models Through Equivalent Parameterization
|
|
Author(s):
|
Niko Kaciroti*+ and Trivellore E. Raghunathan
|
|
Companies:
|
University of Michigan and University of Michigan
|
|
Address:
|
300 N.Ingalls bldg, Ann Arbor, MI, 48109,
|
|
Keywords:
|
Bayesian parametrization ; Missing data ; Non-ignorable nonresponse ; Mixture analysis ; Selection bias ; Identifiability
|
|
Abstract:
|
Pattern-mixture models (PMM) and selection models (SM) are two alternative approaches for statistical analysis with incomplete data and a nonignorable missing-data mechanism. Both models make empirically unverifiable assumptions and need additional constraints to identify the parameters. We introduce Bayesian parameterizations to identify the PMM depending on the types of outcome and then translate these to the SM approach. This provides for a unified and robust parameterization that can be used for sensitivity analysis under either approach. The new parameterizations are easy-to-use and have intuitive interpretation from both PMM and SM perspectives. These models can be fitted using software implementing Gibbs sampling.
|
