|
Activity Number:
|
92
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 4, 2003 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistics in Epidemiology
|
| Abstract - #301024 |
|
Title:
|
Confidence Intervals in Generalized Linear Models for Clustered Data Based on Estimating Equations
|
|
Author(s):
|
John M. Neuhaus*+ and John D. Kalbfleisch and Susan K. Service
|
|
Companies:
|
University of California and University of Waterloo and
|
|
Address:
|
Box 0560, San Francisco, CA, 94143-0560,
|
|
Keywords:
|
bootstrap ; score statistic ; Wald statistic
|
|
Abstract:
|
Investigators often analyze clustered data using generalized linear models and the generalized estimating equations (GEE) approach of Liang & Zeger (1986). Standard analysis yields estimates of regression parameters and associated confidence intervals based on Wald-type pivotals with robust variance estimates. We present the results of an empirical investigation of the performance of three methods for calculating confidence intervals in these settings: Wald-type approaches using robust and model-based variance estimates and an estimating function (EF) bootstrap procedure (Hu & Kalbfleisch 2000). The EF bootstrap resamples independent components of the estimating equation to obtain an estimate of its distribution, and uses this estimate of the distribution of the estimating equation to construct confidence intervals using a standardized estimating equation as a pivotal. Our simulation studies show that the EF bootstrap approach provides more accurate one-sided and two-sided confidence intervals than the two Wald-type methods. We discuss implementation of the EF bootstrap procedure in this clustered data context, and an example further illustrates the simulation results.
|
