Bounding Liang and Zeger's estimator of exchangeable correlation.
Non-Convergence of GEE with exchangeable correlation
In the course of running some simulations for Poisson GEEs with exchangeable
correlation, I found a data set that does not appear to have a solution
to the estimating equations. (Data set as a
text file) This Splus program
lists the data and creates the following plot of the sum of the absolute
value of the terms of the estimating equations. (At convergence this sum
should be zero.) The program uses a modification of the yags software and
hence is subject to the GNU General Public License.
Bounding Liang and Zeger's estimator
of exchangeable correlation.
Using data from the Guidelines for Urinary Incontinence Discussoin and Evaluation (GUIDE) study downloaded from http://www.phs.wfubmc.edu/data/uipreiss.html , I performed some simulations. I found that some of the simulations did not converge. Here are data with one of the non-converging responses, where the non-converging responses are yerr. I found that the data converge in SAS PROC GENMOD but not in YAGS. After communicating with SAS technical help I found that they had modified Liang and Zeger's algorithm for estimating the exchangeable correlation. This Splus program shows how the original exchangeable function does not converge while the modified one does.
An example of a diagonal element of Omega i * Vm greater than 1
Using data from the Guidelines for Urinary Incontinence Discussoin and
Evaluation (GUIDE) study downloaded from http://www.phs.wfubmc.edu/data/uipreiss.html
, I performed some simulations. I found that in some situations the
estimator Omega_i * Vm produced diagonal values greater than
1. Here are data
where this occurs (ysim is the simulated responses). Here is the Splus
program that created the data and shows that the effect occurs.