[R] k-folds cross validation with conditional logistic

[cberry at tajo.ucsd.edu]

Terry Therneau <therneau at mayo.edu> writes:

> --begin inclusion --
> I have a matched-case control dataset that I'm using conditional
> logistic regression (clogit in survival) to analyze.  I'm trying to
> conduct k-folds cross validation on my top models but all of the
> packages I can find (CVbinary in DAAG, KVX) won't work with clogit
> models.  Is there any easy way to do this in R?
> -end inclusion --
>
>  The clogit funciton is simply a wrapper for coxph.  
> 	clogit(case ~ ...  
> turns into
> 	coxph(Surv(dummy, case) ~ ...
> where "dummy" is a vector of ones.
>
> Do the packages support coxph models?

Terry,

I do not know the answer to the question you posed, but I suspect the
answer is no.

The cross-validation would need to be done stratum-wise, but that does
not seem to be supported by predict.coxph():

> fit <- clogit(case~spontaneous+induced+strata(stratum),data=infert)
> train.sans.1 <- update(fit,subset=stratum!=1)
> predict.1 <- predict(train.sans.1,newdat=subset(infert,stratum==1))
Error in predict.coxph(train.sans.1, newdat = subset(infert, stratum ==  : 
  New data has a strata not found in the original model


One can work around this:

> train.sans.1.alt <- update(fit,subset= stratum!=1 | case == 1 )
> all(coef(train.sans.1),coef(train.sans.1.alt))
[1] TRUE
> predict.1.alt <- predict(train.sans.1.alt,newdat=subset(infert,stratum==1))

but the predicted values are not centered in each stratum as usual with
strata in predict.coxph (if that matters):

> predict.1.alt
        1        84       166 
 0.000000 -2.527759 -2.527759 
> predict.1.alt - mean(predict.1.alt)
         1         84        166 
 1.6851724 -0.8425862 -0.8425862 


Best,

Chuck


>
> Terry T
>

-- 
Charles C. Berry                            Dept of Family/Preventive Medicine
ccberry at ucsd dot edu			    UC San Diego
http://famprevmed.ucsd.edu/faculty/cberry/  La Jolla, San Diego 92093-0901