|
Activity Number:
|
224
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, August 10, 2004 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Biometrics Section
|
| Abstract - #301895 |
|
Title:
|
A New Approximation to the Distribution of the Empirical Score Statistic
|
|
Author(s):
|
Brian G. Leroux*+ and Lloyd A. Mancl and Elizabeth Thomas
|
|
Companies:
|
University of Washington and University of Washington and University of Washington
|
|
Address:
|
Dept. of Biostatistics, Seattle, WA, 98195,
|
|
Keywords:
|
score test ; correlated data ; small sample
|
|
Abstract:
|
The empirical (generalized) score test and the robust Wald test are commonly used procedures for hypothesis testing about marginal models with cluster-correlated data. Several researchers have shown that these tests do not have valid Type I error rates in small samples (small number of clusters). In particular, the Wald test has been shown to be anticonservative whereas the score test has been shown to be conservative. Small-sample corrections for these tests have been proposed that improve the performance but do not completely correct the problem. We have found an upper bound for the empirical score test that depends on sample size and helps to explain the small-sample conservativeness of the test. We propose a new distribution for the empirical score statistic that acknowledges this upper bound and provides an excellent approximation to the distribution of the statistic in small samples. Simulation studies show that use of the critical values from the new distribution yield a score test with valid Type I error probabilities in many situations even with very small sample sizes. The test will be illustrated through applications to dental datasets.
|
