Stacy et al.,
Here are snips of two e-mails I received from
Brian Ripley concerning the latest version of the
'rpart' library of S-plus functions.
#1)
The latest rpart allows users to plug their own
criteria in. I had understood Terry Therneau had
sent the code to statlib fairly recently, but I
couldn't see it there.
#2)
FYI. You probably do want the correct legalities
in a .com.
For Windows, the version at
http://www.stats.ox.ac.uk/pub/MASS3/Winlibs is
up-to-date.
I am not clear how he can give me a version to
distribute without
clearing the legalities, but still we have managed
to persuade Mayo to come up with much more
reasonable wording as time goes on.
---------- Forwarded message ----------
Date: Thu, 13 Dec 2001 13:49:40 -0600 (CST)
From: Terry Therneau <therneau@mayo.edu>
To: ripley@stats.ox.ac.uk
Subject: Re: [S] splitting criteria for regression
trees
Brian,
I hope to have it out this week or early next.
Getting a last word polish
from our legal folks on what they want the
disclaimer to be. (I don't think
it will change, however).
Terry
---- end of snips ----
I'm pretty ignorant about the legalities being
mentioned here. Here is the Mayo Clinic website
where it talks about the use of the functions...
http://www.mayo.edu/hsr/Sfunc.html
The advantage of 'rpart' is that its most recent
version (soon to appear at either the Mayo site or
Statlib, or so it sounds) accepts user-defined
splitting criteria.
Even if we decided we would prefer the rpart
functions to the standard S-plus tree-based
modeling functions, I'm not sure what would have
to be done regarding legalities, not to mention
validating these functions for our use here and
writing code for the specific criteria we would
want to use. Also, it sounds like the rpart
library may contain some c code, and I have no
idea if and how we are set up to handle that.
Trying to get rpart may be more trouble than it's
worth. We just don't know if using our own
splitting criteria for tree-based regression would
really provide a significant enough benefit to
justify going after the rpart library. I tend to
think not simply for ARREST since the trees are
used mainly for descriptive purposes and as
preliminary tools for designing subsequent linear
models. A somewhat less than optimal splitting
criterion is probably still doing pretty well. It
*might* be worth looking into for the long-run,
but then again Insightful might just incorporate a
good deal of the rpart functionality into S-Plus
itself at some point, too.
I think we should just stick with what S-plus is
doing now.
- Jeff
|