|
Activity Number:
|
387
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 1, 2007 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistics in Epidemiology
|
| Abstract - #308848 |
|
Title:
|
Covariate-Adjusted Constrained Bayes Predictions of Random Intercepts and Slopes
|
|
Author(s):
|
Robert Lyles*+ and Renee Moore and Amita K. Manatunga and Kirk Easley
|
|
Companies:
|
Emory University and University of Pennsylvania and Emory University and Emory University
|
|
Address:
|
Dept of Biostatistics, Atlanta, GA, 30322,
|
|
Keywords:
|
Exposure modeling ; HIV ; Prediction ; Random effects
|
|
Abstract:
|
Standard prediction of random effects under the mixed linear model takes an empirical Bayesian approach to produce estimates of their posterior mean given the data. While such shrinkage estimates optimize mean square error, the narrow spread of the resulting predictor distribution can be undesirable, e.g., when the objective is to classify subjects relative to a threshold. The constrained Bayes method provides a reduced shrinkage alternative. We examine this approach for predicting random effects, with particular attention to the handling of covariates. We find the general method of Ghosh (1992) to be flexible and comparable with a direct implementation of constraints, suggesting potential for its incorporation into common software for mixed linear models. We provide an example aiming to predict CD4 cell counts among HIV-infected children at the time of Class A disease diagnosis.
|
