|
Activity Number:
|
150
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Monday, August 9, 2004 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
General Methodology
|
| Abstract - #301940 |
|
Title:
|
Refining Multivariate Normal Imputations to Accommodate Non-normal Data
|
|
Author(s):
|
Juwon Song*+ and Thomas R. Belin
|
|
Companies:
|
University of Texas M.D. Anderson Cancer Center and University of California, Los Angeles
|
|
Address:
|
1515 Holcombe Blvd. -447, Houston, TX, 77030,
|
|
Keywords:
|
multiple imputation ; multivariate normal distribution ; importance sampling ; SIR algorithm
|
|
Abstract:
|
An MCMC algorithm based on a multivariate normal distributional assumption provides the basis for widely available statistical software (such as in SAS and S-PLUS) for conducting multiple imputation. However, when data do not fit well with the multivariate normal distribution, this technique may introduce biased estimates. We adapt the sampling-importance-resampling (SIR) algorithm to perform multiple imputation by first generating imputations based on a multivariate normal distribution and then refining the values drawn in the first stage using importance resampling, making use of a more realistic distributional assumption. We first show the feasibility of the method in a simple example where missing values are missing completely at random (MCAR). We discuss the complexity of adapting the method to the more plausible situation where missingness is not MCAR but may be missing at random (MAR). We then outline some potential extensions of the SIR idea that suggest useful avenues to explore.
|
