[R] using bootstrap for tree selection step in rpart

[Frank E Harrell Jr]

Fiona Callaghan wrote:
> Hi Chris.  Thanks for replying.  I want to take the pruned list of
> subtrees, and select the final tree from this list using a bootstrap
> technique (rather than cross validation).   I think, and I could be wrong,

The result will be very poor predictive accuracy on new samples, because 
you are not incorporating shrinkage (penalization) in deriving the 
estimates.

Frank

> that bagging takes a bunch of bootstrap samples, grows one tree per
> boostrap sample and then combines the estimates, but this is not want I
> want to do.   I hope this is clearer.  It is difficult to describe these
> things with eamil.
> Cheers
> Fiona
>> Fiona Callaghan wrote:
>>> I was wondering if someone could help me with an rpart problem.  I can
>>> see
>>> that cross-validation is the default for tree selection in rpart -- has
>>> a
>>> bootstrap method been implemented anywhere?  I think this is a different
>>> thing to 'bagging' or 'boosting' -- I still want 'one' tree at the end,
>>> I
>>> just would like it chosen using a bootstrap method.  Any ideas???
>> Hi Fiona,
>>
>> I'm not sure if I understand you correctly.
>> To get one single rpart tree trained on one bootstrap sample, try
>> bagging() from the 'ipred' package and set nbagg=1.
>>
>> Bye,
>> Chris
>>
>>
> 
> 


-- 
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University