Activity Number:
|
242
- Contributed Poster Presentations: Biometrics
|
Type:
|
Contributed
|
Date/Time:
|
Monday, July 31, 2017 : 2:00 PM to 3:50 PM
|
Sponsor:
|
Biometrics Section
|
Abstract #322929
|
|
Title:
|
Efficient Algorithm for Multiple Imputation of Missing Health Outcomes
|
Author(s):
|
Chris Liu*
|
Companies:
|
University of Michigan-School of Nursing
|
Keywords:
|
multiple imputation ;
spectral decomposition ;
Bayesian model ;
truncated normal ;
Gibbs sampler
|
Abstract:
|
Multiple imputation can be complicated and challenging for analyzing incomplete data sets with repeated measure design. In Bayesian setting, an MCMC algorithm is developed for efficiently imputing missing values. The algorithm integrates data augmentation/missing imputation and posterior sampling. Spectral decomposition techniques are used to decompose the co-variance matrix of outcomes such that simulating the original correlated outcomes with missing values from multivariate truncated normal distributions is equivalent to first sampling the de-correlated outcomes from univarate truncated normal and then translating back to the original outcomes. The idea behind this is to obtain a more efficient implementation of the Gibbs sampler for missing imputation by avoiding repeatedly computing multidimensional means and co-variances. We used a longitudinal study of hypertension and a longitudinal study of aging as examples to demonstrate the new algorithm and compare it with standard methods.
|
Authors who are presenting talks have a * after their name.