Skip to content

Jackknife and rpart

2 messages · chumpmonkey@hushmail.com, Frank E Harrell Jr

chumpmonkey@hushmail.com
# 
Hi,

First, thanks to those who helped me see my gross misunderstanding of
randomForest. I worked through a baging tutorial and now understand the
"many tree" approach. However, it is not what I want to do! My bagged
errors are accpetable but I need to use the actual tree and need a single
tree application. 

I am using rpart for a classification tree but am interested in a more
unbaised estimator of error in my tree. I lack sufficent data to train
and test the tree and I'm hoping to bootstrap, or rather jacknife, an
error estimate.

I do not think the rpart.object can be applied to the jackknife function
in bootstrap but can I do something as simple as:

for(i in 1:number of samples){
  remove i from the data
  run the tree
  compare sample[i] to the tree using predict
  create an error matrix}

This would give me a confussion matrix of data not included in the tree's
constuction.

Am I being obtuse again?

Thanks, CM
Frank E Harrell Jr
# 
On Wed, 16 Apr 2003 10:28:08 -0700
chumpmonkey at hushmail.com wrote:

            

      
      
      
    

        

      
      
    

        
You might look at the validate.tree function in the Design library (http://hesweb1.med.virginia.edu/biostat/s/Design.html) but better validated predictive accuracy would be obtained by approximating the predictions from the randomForest by a single (moderately large) tree.  You can use rpart to develop such a tree, stopping when, for example, the R-square is 0.9 or 0.95.
---
Frank E Harrell Jr              Prof. of Biostatistics & Statistics
Div. of Biostatistics & Epidem. Dept. of Health Evaluation Sciences
U. Virginia School of Medicine  http://hesweb1.med.virginia.edu/biostat