R help - Aug 2009 - Complexity parameter in rpart

Hi all,

I'm currently using the 'rpart' function to run some regression
analysis and I am at the point where I wish to prune my overfitted trees. Having
read the documentation I understand that to do this requires the use of the
complexity parameter. My question is how to go about choosing the correct
complexity parameter for my tree? In some places
(http://www.statmethods.net/advstats/cart.html) I have read that it is best to
select the complexity parameter which minimises the cross-validated (x) error of
the model, but elsewhere I have read that the optimum cp is the first value on
the left above the '1+SE' line of the complexity paramter plot.

I was hoping someone might be able to clarify this minor issue for me.

Many thanks,

Andy

_________________________________________________________________
Save time by using Hotmail to access your other email accounts.

	[[alternative HTML version deleted]]

>  I have read that it is best to select the complexity parameter which
minimises the cross-validated (x) error of the model, but elsewhere I
have read that the optimum cp is the first value on the left above the
'1+SE' line of the complexity paramter plot.

 If you plot x=complexity vs y= cross-validated error, the plot will in
theory decline sharply as you go from left to right, then hit a minimum,
and then rise.  In practice there is often a "flat" area around the
minimum.  The idea is that within the flat area, which model is the
actual numeric minimum is pretty random, and one would like to pick the
smallest model from among the "nearly tied" ones.  The 1 SE rule
formalizes this.  

Tree models are so very interpretable that one's tendency is to keep too
many branches.  Cross-validation and particular the 1 se rule prune more
of the tree away that we'd like, often leaving nothing but a root node.

	Terry T.