1
$\begingroup$

I have tried now several packages in R for Cox proportional hazard regression and I am having a hard time to find a function where I can fit a model using cross validation/bootstrapping and obtain an estimate of the C-index based on the untrained items with a confidence interval.

The PEC package comes close, however it does not provide a confidence interval. Does anybody know a package which does provide it? Or a formula to calculate it myself?

$\endgroup$
  • 1
    $\begingroup$ Doesn't validate in the rms package do what you need when applied to a cph object? It returns a matrix with results of all bootstraps or cross-validations from which you can determine the distribution of C-index values. The $D_{xy}$ is what is returned, but this is simply related to C-index as the manual page shows. $\endgroup$ – EdM Feb 1 '17 at 18:38
  • $\begingroup$ I'm not sure that is what I'm looking for; the validate method asks for a model which has already been fitted (how can I use crossvalidation if the model already uses all the data?) And I understand that I can calculate the C-index from the $D_{xy}$, but what about the confidence interval? I can't find those values in the output/manual $\endgroup$ – Héctor van den Boorn Feb 1 '17 at 19:10
  • 1
    $\begingroup$ The rms package is based on specially designed objects that keep lots of information about the model; it has a cph command instead of coxph. If you use the "x=TRUE" and "y=TRUE" arguments in the cph call, as required for validate, it keeps the original data in the output from cph. Then validate can do its bootstrapping or cross-validation from the stored original data. If you do 1000 bootstraps then the bottom 25 and top 25 give one estimate of 95% confidence intervals. Other types of confidence intervals are described for example in the boot.ci help page in the boot package. $\endgroup$ – EdM Feb 1 '17 at 20:03
  • $\begingroup$ Ok that sounds good! But in creating the model (i.e. the beta weights) it uses all the data to create the fit; will the crossvalidation in validate override these beta weights? $\endgroup$ – Héctor van den Boorn Feb 1 '17 at 20:48
2
$\begingroup$

The simplest way to get empirically based confidence intervals is bootstrapping. If you do 999 bootstraps and rank your statistic of interest in ascending order, then the 25th and the 975th provide a simple estimate of 95% confidence intervals. I don't know of any advantages of cross-validation for this purpose, and you would have to do multiple cross-validations to get that good an estimate in any event. Bootstrapping has the advantage to being as close as you can get to resampling from the original population.

So if you already have a program that produces your statistic of interest from a set of bootstrap samples (like validate() from the rms package, which works on the output from the package's cph() function), that's the easiest way to proceed. Functions like validate() know how to keep track of parameter estimates from the original model separately from those based on the bootstrap resamples, which seems to be of some concern to you.*

There are other ways to estimate CIs from bootstraps. Bootstrapping Regression Models in R shows how to use facilities provided by R packages to get such estimates. It provides the approach for calculating the potentially better bias-corrected, accelerated (or $BC_a$) percentile intervals.

The boot package can automate this process in general. First you write a function that produces your statistic of interest from a dataset (first argument to your function) based on a supplied vector of indices (second argument). There's a simple example of the structure of such a function on page 194 of ISLR, to augment the somewhat more opaque examples on the boot() manual page. Then the boot() function does the resampling of the cases, passing the vector of chosen indices each time to your function to calculate your statistic. You can use the boot.ci() function on the output from boot() to get 5 different types of CI estimates, including $BC_a$.

Finally, with respect to the C-index in particular, pay attention to how you intend to use it. It's not, for example, a very sensitive test for model selection according to Frank Harrell, who developed that index and also is responsible for the rms package.


*If you have questions about how the calculations are performed you can get the code for functions written in R by typing the function name at the R prompt, sometimes preceded by the package name. After the rms package is loaded, type rms:::validate.cph to get its code, and predab.resample to see the workhorse function that does most of the work.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.