generating contrasts from glmm.admb?

An embedded and charset-unspecified text was scrubbed...
Name: not available
URL: <https://stat.ethz.ch/pipermail/r-sig-mixed-models/attachments/20130917/a4b87dd9/attachment.pl>

Maureen Durkin <mmdurkin at ...> writes:

Hi all,

I am using package glmm.admb to run a negative binomial mixed
effects model on count data of different disturbance for plots. I
would like to test for differences in counts of disturbances between
five site.s Not having any trouble with glmm.admb itself, but I
can't seem to figure out how to generate a contrast matrix for
differences between individual sites. It seems that the "contrast"
package cannot handle objects from glmm.admb? Is this correct, and
if so does anyone know of other methods to generate contrast methods
from glmm.admb?

I started to dive into this, but generating the matrix for all
pairwise contrasts is a bit of a pain.  Let's say you already have the matrix
of contrasts (say, each row is a contrast, so if 'sites' is a fixed
effect and you have a usual treatment-contrasts parameterization,
the contrast matrix for comparing site 2 against site 3 would
be c(0,-1,1,0,...))

set.seed(1)
d <- data.frame(x=rnbinom(1000,mu=3,size=1),
      f=factor(rep(LETTERS[1:5],each=200)))
library(glmmADMB)
m1 <- glmmadmb(x~f,data=d,family="nbinom")
cc <- coef(m1)

## compare site 2 vs 3; 3 vs 4; 4 vs 5
M <- matrix(c(0,-1,1,0,0,
              0,0,-1,1,0,
              0,0,0,-1,1),
            byrow=TRUE,nrow=3)

## The estimated value for each contrast:
cvals <- c(M %*% cc)

## The estimated variance-covariance matrix for the contrasts:
vv <- M %*% vcov(m1) %*% t(M)

## The Z-score for each contrast:
cvals/sqrt(diag(vv))

I'll let someone else worry about how to come up with
the all-pairwise-contrasts matrix.

The other alternative would be to bug the glmmADMB and/or
contrasts package maintainers to extend their packages to
allow them to co-operate.

This all assumes that sites are a *fixed* effect -- otherwise
(if they're random), hypothesis tests on the site-to-site
differences aren't meaningful.