R help - Apr 2013 - Model ranking (AICc, BIC, QIC) with coxme regression

Hi,

I'm actually trying to rank a set of candidate models with an information
criterion (AICc, QIC, BIC). The problem I have is that I use mixed-effect cox
regression only available with the package {coxme} (see the example below).

#Model1
>spring.cox <- coxme (Surv(start, stop, Real_rand) ~
strata(Paired)+R4+R3+R2+(R3|Individual), spring)

I've already found some explications in this forum to adjust QIC on coxph
object (see the following lines, thanks to M. Basille), but it doesn't work
on coxme...
>QIC <- function(mod, ...)
>    UseMethod("QIC")
>
>QIC.coxph <- function(spring.cox, details = FALSE) {
>    trace <- sum(diag(solve(spring.cox$naive.var) %*% spring.cox$var))
>    quasi <- spring.cox$loglik[2]
>    return(-2*quasi + 2*trace)
>}
The only thing that I can't find in the coxme output to use these previous
commands is the "naive.var", that we can obtain in coxph regression by
specifying "robust=TRUE" in the argument list:
>spring.cox <- coxph (Surv(start, stop, Real_rand) ~
strata(Paired)+R4+R3+R2, spring, robust=T)

But coxph doesn't allow inclusion of interaction between two random
variables (R3|Individual), and it's why I have to use coxme.
I found a new update in R-forge to improve {coxme}
(r-forge.r-project.org/scm/viewvc.php/pkg/R/dredge.R?view=log&root=mumin),
but I did not understand all it works and I'm not sure it fixes my
problems...

Is there someone that can help me with that?

Rémi Lesmerises, biol. M.Sc.,

Candidat Ph.D. en Biologie
Université du Québec à Rimouski
300, allée des Ursulines
Rimouski, Qc., G5L 3A1
Tél.: 1 800 511-3382  #1241
remilesmerises@yahoo.ca
	[[alternative HTML version deleted]]