You are right, those look suspicious. What version of R and of the coxme
package are you
running? Later version of coxme use multiple starting estimates due to
precisely this
kind of problem.
Also, when the true MLE is variance=0 the program purposely never quite gets
there, in
order to avoid log(0). Compare the log-lik to a fixed effects model with those
covariates.
I can't do more than guess without a reproducable example.
Terry Therneau
On 10/08/2012 05:00 AM, r-help-request at r-project.org
wrote:> Dear R users,
>
> I'm using the function coxme of the package coxme in order to build Cox
> models with complex random effects. Unfortunately, I sometimes get
> surprising estimations of the variances of the random effects.
>
> I ran models with different fixed covariates but always with the same 3
> random effects defined by the argument
> varlist=coxmeMlist(list(mat1,mat2,mat3), rescale = F, pdcheck = F,
> positive=F). I get a few times exactly the same estimations of the
> parameters of the random effects whereas the fixed effects of the models
> are different:
>
> Random effects
> Group Variable Std Dev Variance
> idp Vmat.1 0.10000000 0.01000000
> Vmat.2 0.02236068 0.00050000
> Vmat.3 0.02449490 0.00060000
>
> The variances are round figures, so I have the feeling that the algorithm
> didn't succeed in fitting the model.
>
> Has anyone ever faced to this problem?
>
> Thanks,
>
> Hugo