Coxme function

["Hunsicker, Lawrence" <lawrence-hunsicker@uiowa.edu>]

Good morning again, friends.

I want to thank Terry Therneau and Mike O'Connell for their help with the following questions that I have asked over the past weeks:

>  I would like to run a Cox mixed model in which there are both individual covariates and group covariates. When one is running a mixed linear model, one  includes the group covariates in the specification of the random effects. But so far as I can see, there is no way to do this with the “frailty” term in the coxph function.  What is the correct way to enter group covariates in the coxph model?

Terry has pointed out to me the new S-Plus function (which he wrote) coxme, which handles Cox mixed models much more generally than the old coxph with a frailty term.  Coxme uses a more flexible method for specifying the fixed and random effects, essentially the same grammar as in Pinheiro's lme function.  It is also blazingly faster on big data sets.  In comparable analyses that I ran, the coxme ran in five minutes and gave essentially identical answers as the coxph with frailty, that took over one hour to converge. 

Terry has written up a detailed technical report on the motivation and the workings of this new function that is available at:

www.mayo.edu/biostatistics
On that web page, choose “software”, then “splus”, then the “kinship” entry.  At the bottom of that entry are links both to the function and its associated functions as a tar.gz file, and for the Technical Report.

The second question was related to the first:
> If I do a coxph with a frailty term, the summary tells me that the
> variance of the frailty term is 0.0305.  But I am not sure what the
> credible range is for this variance.  The 'intervals' function doesn't
> work on a penalized Cox model, apparently.  Is there a good way to test ...

Terry's response:

  First, use coxme for the models.  It is superior to coxph + frailty.  To test for significance of the random effect, compare the integrated likelihood from a coxme model to the model without a random effect, i.e., to a coxph result with the same variables.  A standard LR test.

Again, thanks in particular to Terry for providing this function and help in using it correctly, and to Mike O'Connell in pointing me to it.

Larry Hunsicker