# [R] Random coefficients model with a covariate: coxme function

Terry Therneau therneau at mayo.edu
Mon Dec 29 16:18:28 CET 2008

```> I'm new to R and am trying to fit a mixed model
> Cox regression model with coxme function.
> I have one two-level factor (treat) and one
> covariate (covar) and 32 different groups
> (centers). I'd like to fit a random coefficients model, with treat and covar
> as fixed factors and a random intercept, random
> treat effect and random covar slope per center.
> I haver a couple of doubts on how to use coxme function for this task:

example deleted

> * What if the treatment factor has more than two
> levels. Should I follow the same procedure, with just bigger block sizes?

> * Coxme returns a variance per each of the
> variance matrices I defined, but no residual
> variance estimate. Is there a way to get it?

The coxme function does not support random slopes.  It's been on my "to do"
list for a long time.  I am supposed to teach an American Stat Assoc course at
the end of March, however, which has escalated the urgency.

If the covariate has only 2 levels, such as a random treatment effect when
there are only 2 treatments, then by coding the treatment as 0/1 and creating
just the right covariates you could "trick" coxme into fitting the model.  This
is what is described in the report.  You essentially make treatment a nested
effect.
fit1 <-  coxme(Surv(y, uncens) ~ treat + covar, data1,
random= ~1 | centers)
fit2 <-  coxme(Surv(y, uncens) ~ treat + covar, data1,
random= ~1 | centers/treat)

There is no residual variance for a Cox model.

Your example was very hard to read.  Consider using spaces, indentation, etc
to make it easier for old eyes.

Terry T.

```