I have fitted a
mixed effects 2-level logistic regression model to a dataset of approx. 6,000
subjects using S Plus glme to provide estimates for 14 fixed effect parameters
and the random effects (of which there are 3) covariance matrix (=6
parameters). I have programmed WinBUGS to do the same thing. While the fixed
effects parameters/CIs are fairly close for the two methods, the covariance
matrix components are miles apart. In fact, the 95% CI's do not even
intersect.
S Plus glme carries
out a restricted ML calculation by "integrating out" the actual random effects,
while in the WinBUGS specification, the random effects are an explicit part of
the model albeit drawn from a multivariate normal distribution with
covariance matrix omega, for which I have included a Wishart
prior in the model. Could this give rise to the differences? My instinct
tells me not, but I'm at a loss to understand what's
happening.
Anybody any thoughts
or similar experiences?
Frank
Gargent