|
Activity Number:
|
139
|
|
Type:
|
Invited
|
|
Date/Time:
|
Monday, August 3, 2009 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #303067 |
|
Title:
|
Generalized Linear Models for a Correlation Matrix in Longitudinal Data
|
|
Author(s):
|
Michael J. Daniels*+
|
|
Companies:
|
University of Florida
|
|
Address:
|
Department of Statistics, Gainesville, FL, 32611-8545,
|
|
Keywords:
|
missing data ; probit model ; partial correlations
|
|
Abstract:
|
We propose generalized linear models for a specific parameterization of a correlation matrix. The components of this parameterization are a set of partial correlations that can independently vary between -1 and 1. Fisher's z-transform is used for the link function. The design matrices for these correlation models can include unit-specific covariates and/or structural design components. The GLMs for the correlation matrix can be combined with GLMs for the mean for longitudinal normal data and longitudinal binary data. Correlation models are particularly important in incomplete longitudinal data since mis-modeled correlations typically result in biased mean regression parameters. We outline a Bayesian inference strategy and illustrate these models on data from a smoking cessation clinical trial.
|
