|
Activity Number:
|
376
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 1, 2007 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Biometrics Section
|
| Abstract - #309600 |
|
Title:
|
Generalized Linear Mixed Models for Binary Outcome Data with a Low Number of Occurrences: Understanding When Estimation Procedures Fail
|
|
Author(s):
|
Marie-Eve Beauchamp*+ and Robert W. Platt and James A. Hanley
|
|
Companies:
|
McGill University and McGill University and McGill University
|
|
Address:
|
1020 Pine Avenue West, Montreal, QC, H3A 1A2, Canada
|
|
Keywords:
|
Adaptive Gauss-Hermite quadrature ; Penalized quasi-likelihood ; Sparse binary outcome data
|
|
Abstract:
|
The limits of estimation methods for generalized linear mixed models that are available in commercial software packages have not been studied comprehensively. Penalized quasi-likelihood (PQL) is among the most commonly used methods. Some authors have recommended using adaptive Gauss-Hermite quadrature (AGHQ) as it is expected to produce less biased parameter estimates. We compared via simulations the performance of PQL and AGHQ for several settings of binary outcome data with a low number of occurrences. Estimation procedures failed for several datasets with a low number of clusters and a small cluster size. Types of failure observed are convergence problems, parameter estimates outside of their realistic range and biased parameter estimates. We attempted to characterize the circumstances in which the estimation procedures failed.
|
