Alternative Methods for Analyzing Clustered Binary Data in Ophthalmological Studies
A. Russell
Localio, J. Richard Landis, Susan L. Weaver, Tonya J. Sharp
Center for
Biostatistics and Epidemiology, Penn State University
1993, Presented
at the Biometrics Society Spring Meetings.
Abstract
This paper
describes modeling alternatives for binary data in ophthalmology,
where clustering (correlation) of eyes within subject is a
well-recognized issue, to determine how long after fatal accident
donor corneas can be transplanted safely. Rabbits were sacrificed,
and corneas of one eye were examined microscopically to determine the
rate of cell death (under 7 per 1,000). Rabbits (and remaining eyes)
were either refrigerated or not, and cell death rate was determined
at 6 and 24 hours. Modeling the effect of refrigeration and time on
cell viability was performed using: (1) logistic regression (LR) (SAS
CATMOD) with no consideration of clustering of cells within eye, or
the correlation of eyes within rabbit, (2) exact LR using LogXact,
(3) LR using SUDAAN and considering the clustering of cells within
eye, (4) LR using SUDAAN and considering the two stages of
clustering, cell within eye and eye within rabbit, (5) GEE with a
logistic link and binomial error that accounts for clustering of
cells within eye, and (6) cells within rabbit, (7) GEE with log link
and Poisson error term that conditions on eye and accounts for
correlation of eyes within rabbit, (8) conditional LR that conditions
on rabbit (STATA), and (9) exact conditional LR (Log-Xact). Methods
are compared both in terms of interpretation of parameters and
standard errors, and computing requirements for samples of 3,000
observations per eye and a total of 130,000 data points.