|
Activity Number:
|
427
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 5, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #305198 |
|
Title:
|
Evaluating Multiple Imputation Procedures Using Simulations in a Bayesian Prospective
|
|
Author(s):
|
Fabrizia Mealli*+ and Michela Baccini and Constantine Frangakis and Fan Li and Donald B. Rubin and Elizabeth R. Zell
|
|
Companies:
|
University of Florence and University of Florence and Johns Hopkins University and Duke University and Harvard University and CDC
|
|
Address:
|
viale Morgagni 59, Florence, 50139, Italy
|
|
Keywords:
|
Multiple Imputation ; Missing Data Patterns ; Evaluation ; Simulations
|
|
Abstract:
|
Missing data are a pervasive problem in statistics. Multiple imputation is a valid tool to deal with missing data. It can be used under a variety of assumptions concerning the joint distribution of the missing data mechanism and the data, which may be difficult to justify. Any proposal to deal with missing data should be evaluated. We propose a template for evaluating imputation procedures to be used in any specific study. Starting from the original data set, we propose a way to create simulated data sets with realistic missing data patterns, using Bayesian logistic regressions. Each simulated data set is imputed with the chosen imputation procedure, and a subset of important analyses performed. Evaluation of imputation performance is based on comparisons of the posterior distributions of the relevant parameters from the "true" complete data and from the multiply imputed data sets.
|
