[R] stepAIC and coxph objects with cluster(id)

Thomas Lumley thomas at biostat.washington.edu
Thu Mar 16 22:40:46 CET 2000

On Thu, 16 Mar 2000, Ramon Diaz-Uriarte wrote:

> Is it appropriate to use stepAIC (library MASS) with coxph objects (from
> library survival5) that use "cluster(id)"? 
> It is my understanding that, when using "cluster(id)", we can test for sets of
> terms by using the methods in Wei et al., (1989; JASA, 84: 1065-1073), or as
> explained in pp. 53 and ff. of the survival.ps document. But if we use a
> likelihood ratio test instead (comparing sets of nested models), we can get
> different results.
> I started wondering, then, if using AIC, per se, could give misleading
> results. Am I getting confused here?

It's not completely clear to me whether AIC is useful for an ordinary Cox
model using the partial likelihood (though I presume _someone_ has studied
the properties of this method).

For a model using cluster(id) it certainly isn't appropriate.  These
models do not have even a partial likelihood. Likelihood ratio tests are
not possible, and AIC is not well-defined.

Stepwise selection would have to be based on the distribution of the
parameter estimates and the sandwich estimator of its covariance


Thomas Lumley
Assistant Professor, Biostatistics
University of Washington, Seattle

r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch

More information about the R-help mailing list