|
Activity Number:
|
133
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 4, 2003 : 12:00 PM to 1:50 PM
|
|
Sponsor:
|
Biometrics Section
|
| Abstract - #300413 |
|
Title:
|
Modeling Sensitivity/Specificity within a Cluster-Randomized Study
|
|
Author(s):
|
Thomas R. Ten Have*+ and Michael Rambo
|
|
Companies:
|
University of Pennsylvania and
|
|
Address:
|
423 Guardian Dr., Philadelphia, PA, 19104-4209,
|
|
Keywords:
|
PQL ; GEE ; sensitivity ; specificity ; randomized cluster trial
|
|
Abstract:
|
Under the asymptotic limitation of a small number of clusters with a large number of observations, we present a model-based analysis of sensitivity and specificity, comparing a random effects logistic model using penalized quasi-likelihood (PQL) estimation to the analogous population-averaged model estimated with GEE using SUDAAN. The context is the assessment of the diagnostic accuracy of a depression screening instrument (CES-D) relative to a standardized depression clinical diagnosis (SCID) among elderly primary care patients in a practice-randomized clinical trial for treating depression. It is of subject-matter interest to see if past history of depression affects the sensitivity and specificity of CES-D. We empirically show that the Taylor series-based or sandwich estimates of variance under GEE severely underestimate variances relative to the variances from PQL estimation under the random effects logistic model. This underestimation is beyond what is expected given the relationship between population-averaged and random effects logistic models. Per simulation results in Ten Have and Localio (1999), PQL performs well in this situation.
|
