GEE models are very different from the subject-specific models fitted by glmmPQL: see the comparison in MASS. You are testing quite different hypotheses. You appear to be assuming that some things `appear not to work' because they do not give the same results as a different test. Suppose x were logical, and that almost all subjects did better with x=TRUE. Then what you are saying is that most subjects have response 0 or 1 irrespective of x, but suppose that when they differ, it was (almost) always x=TRUE that gave 1. Then the subject-specific model will have a large positive coefficient for x. (It is possible that if the pattern is the same for all individuals the MLE is infinite, called complete separation.) OTOH, the GEE model applies to the population, and in the population x=TRUE makes rather little difference to the mean response. GEE models attenuate subject-specific effects, and can do so dramatically.
On Wed, 30 Nov 2005, Rick Bilonick wrote:
I'm trying to fit a generalized mixed effects model to a data set where each subject has paired categorical responses y (so I'm trying to use a binomial logit link). There are about 183 observations and one explanatory factor x. I'm trying to fit something like: (lmer(y~x+(1|subject))) I also tried fitting the same type of model using glmmPQL from MASS. In both cases, I get a t-statistic that is huge (in the thousands) and a tiny p-value. (Just for comparison, if I use lrm ignoring the clustering I get a t-statistic around 3 or so and what appears to be a reasonable estimated coefficient which is very close to the estimated coefficient using just one observation from each subject. Most of the subjects have two responses and in almost all cases the responses are identical although the explantory factor values are not always identical for each subject. If I use geeglm from geepack, I get reasonable estimates close to the naive model results. I also tried using the SAS glimmix macro to fit a generalized mixed model and the routine does not converge. Why does geeglm appear to work but not lmer and glmmPQL? Is this likely to be due to my particular data set? Rick B.
______________________________________________ R-help at stat.math.ethz.ch mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Brian D. Ripley, ripley at stats.ox.ac.uk Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272866 (PA) Oxford OX1 3TG, UK Fax: +44 1865 272595