[R-sig-ME] MCMCglmm for binomial models?
Bob Farmer
farmerb at gmail.com
Mon Sep 20 03:49:58 CEST 2010
Hi all, I've spent some time evaluating MCMCglmm as a faster
alternative to WinBUGS and lmer (+/- mcmcsamp) for my modeling, and I
wonder if anyone (esp. Dr. Hadfield) could comment on using this
package for binomial models. Consider the following test code:
############ first, a Gaussian model
library(MCMCglmm); library(lme4)
data(PlodiaPO)
mc2<-MCMCglmm(PO~plate, random=~FSfamily, data=PlodiaPO, verbose=FALSE)
gm2<-lmer(PO ~ 1 + plate + (1|FSfamily), data = PlodiaPO)
summary(mc2)
summary(gm2)@coefs
where, as an index of MCMCglmm's ability to perform like lmer, the
fixed-effects coefficients are quite similar.
On the other hand, in this case:
#################now binomial from a reference dataset
gm3 <- glmer(cbind(incidence, size - incidence) ~ period + (1 | herd),
family = binomial, data = cbpp)
mc3<-MCMCglmm(cbind(incidence, size - incidence) ~ period,
random = ~ herd,
family = "multinomial2",
data = cbpp, verbose = FALSE,
nitt = 50E3
)
summary(gm3)@coefs
summary(mc3)
(with extra iterations for insurance), the results are far more
variable, and generally not consistent with glmer()'s output.
Assuming that glmer()'s output is okay (and maybe this is a big
assumption? it *is* the example listed in the helpfile), does this
mean that MCMCglmm doesn't perform well in these sorts of models
without lots of attention? I've spent some time playing with prior
specifications, but I find the documentation in this department is a
bit weak (G =list(G1)? what's G1? what's "n"?), and so I'm not
comfortable spending too much time when I already am at the leading
edge of "knowing what it is I'm doing".
Any help is much appreciated!
Thanks.
--Bob Farmer
Dalhousie University
More information about the R-sig-mixed-models
mailing list