First, here is your message as it appears on R-help.
On 10/14/2012 05:00 AM, r-help-requ...@r-project.org wrote:
> I?m trying to set up proportional hazard model that is stratified with
> respect to covariate 1 and has an interaction between covariate 1 and
> another variable, covariate 2. Both variables are categorical. In the
> following, I try to illustrate the two problems that I?ve encountered, using
> the lung dataset.
>
>
>
> The first problem is the warning:
>
>
>
> To me, it seems that there are too many dummies generated.
>
> The second problem is the error:
>
Please try to fix this in the future (Nabble issue?)
As to the problems: handling strata by covariate interactions turns out to be a
bit of a
pain in the posteriorin the survival code. It would have worked, however, if
you had done
the following:
fit <- coxph(Surv(time, status) ~ strata(cov1) * cov2, data=...)
or ~ strata(cov1):cov2
or ~ strata(cov1):cov2 + cov2
But by using
~ strata(cov1) + cov1:cov2
you fooled the program into thinking that there was no strata by covariate
interaction,
and so it did not follow the special logic necessary for that case.
Second issue: The model.matrix function of R, common to nearly all the
modeling
functions (including coxph) tries to guess which dummy variables will be
redundant, and
thus can be removed from the X matrix before the fit. Such an approach is
doomed to
failure. I'm actually surprised at how often R guesses correctly, because
until a matrix
decomposition is actually performed the only thing possible is an informed
guess. Your
particular case gives rise to a larger than usual number of NA coefs (redundant
columns),
but short of building your own X matrix by hand there isn't anything to be done
about it.
Just ignore them.
Terry Therneau
[[alternative HTML version deleted]]
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
