|
Activity Number:
|
285
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, August 4, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Health Policy Statistics
|
| Abstract - #304327 |
|
Title:
|
Clustering to Achieve Normality in Generalized Linear Mixed Models
|
|
Author(s):
|
Kenneth Pietz*+ and LeChauncy D. Woodard and Tracy Urech and Cassie Robinson and Laura A. Petersen
|
|
Companies:
|
Michael E. DeBakey VA Medical Center and Michael E. DeBakey VA Medical Center and Michael E. DeBakey VA Medical Center and Michael E. DeBakey VA Medical Center and Baylor College of Medicine
|
|
Address:
|
2002 Holcombe Blvd., Houston, TX, 77030,
|
|
Keywords:
|
generalized linear models ; maximum liklihood ; health policy
|
|
Abstract:
|
Many quality studies use data collected at multiple hospitals. To account for nesting, the hospital is a random effect. The hospitals are assumed to be from a normally distributed population with unknown variance. This method is reasonably robust to departures from normality in linear mixed models but can lead to inconsistent parameter estimates in generalized linear mixed models. The normality assumption is less likely if the hospitals have widely different characteristics. Using hospital variables derived in a previous study, we used K-means clustering to classify 105 VA hospitals into 10 groups. Within each cluster, logistic regression with random effects was used to determine the effect of serious mental health conditions (MH) on Hba1c control in diabetics after adjusting for patient characteristics. The MH parameter estimates within the clusters ranged from 0.166 to 0.297.
|
