3
$\begingroup$

Does anyone know an implementation in R (or other) of a decision tree for censored outcomes? I would like to utilize a decision tree in order to discretize/bin continuous variables before a survival analysis in some sort of principled manner. I am left with only a traditional decision tree using a binary target (event/no event) disregarding the censored nature of the data as it stands.

$\endgroup$
2
  • $\begingroup$ There is a typo in the title: Decison should be Decision. $\endgroup$
    – petrichor
    Jun 17, 2011 at 9:05
  • $\begingroup$ @Ismail: Fixed. $\endgroup$
    – B_Miner
    Jun 17, 2011 at 11:39

2 Answers 2

3
$\begingroup$

Have you checked the package party? I believe the function ctree handles censored data.

$\endgroup$
3
  • $\begingroup$ I did find this package after the posting. Do you happen to know if it handles time varying covariates (values change with time)? $\endgroup$
    – B_Miner
    Jun 17, 2011 at 2:06
  • $\begingroup$ I'm not sure what you mean by 'time varying' covariates, but if you can organize your data in a data.frame in R, then ctree should work. $\endgroup$
    – joran
    Jun 17, 2011 at 2:10
  • $\begingroup$ I mean covariates/predictors whose values are not set at the beginning of the observation period - instead they can change and take on different values throughout the time period being observed. $\endgroup$
    – B_Miner
    Jun 20, 2011 at 21:59
5
$\begingroup$

Binning continuous variables goes against principle. And note that for recursive partitioning to be able to do all the thinking for you (find correct cutpoints assuming they exist, which is highly unlikely) requires upwards of 50,000 events in order to obtain a tree whose structure will be validated in other data. The motivation for binning in order to do any kind of analysis is unclear.

$\endgroup$
2
  • $\begingroup$ The motivation is that I am seeking to discover "triggers" in transaction data (where customer defection is the outcome of interest) that can be then monitored for - and when experienced by a customer - trigger an intervention. The discretizing of the predictors is desired because the events needs to be discrete to "fire" a trigger. $\endgroup$
    – B_Miner
    Jun 17, 2011 at 2:08
  • 2
    $\begingroup$ Sounds like you might have a massive dataset and will be in better shape. But a probability model can accomplish that - you can trigger when the predicted probability exceeds a certain value. The value can be set by subject matter considerations or by a lift curve. Dichotomizing on the front end will cause a large loss of predictive discrimination. $\endgroup$ Jun 17, 2011 at 2:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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