Exhaustive CHAID package
On Wed, 22 Apr 2015, Michael Grant wrote:
Many thanks for your response, sir. Here are two of the references to which I referred. I've also personally explored several data sets in which the outcomes are 'known' and have seen high variability in the topology of the trees being produced but, typically Exhaustive CHAID predictions match the 'known' results better than any of the others, using default settings. http://www.hindawi.com/journals/jam/2014/929768/ http://interstat.statjournals.net/YEAR/2010/articles/1007001.pdf
Thanks for the references, I wasn't aware of these. Both of these appear to use the SPSS implementations of CHAID, exhaustive CHAID, CART, and QUEST with default settings (as you did above). As I have never used these myself in SPSS, I cannot say how the implementations compare but it's well possible that these are different from other implementations. E.g., for CART the pruning rule may make a difference (with or without crossvalidation; with 1-SE or 0-SE rule etc.). Similarly, for QUEST I think that Loh's own implementation uses somewhat different default settings. So it may be advisable to go beyond defaults.
By inference, many research papers are choosing Exhaustive CHAID.
My experience is that this is determined to a good degree by the software available and what others in the same literature use.
My concern is not that these procedures produce mildly variant trees but dramatically variant, with not even the same set of variables included.
Yes, instability of the tree structure is one of the drawbacks of tree-based procedures. Of course, the tree structure can be very different while producing very similar predictions.
Is CHAID available for use as an R package?
Yes.
I thought R-FORGE was solely for developers?
See: https://R-Forge.R-project.org/R/?group_id=343 You can easily install the package from R-Forge and also check out the entire source code anonymously.
Again, many thanks. MCG -----Original Message----- From: Achim Zeileis [mailto:Achim.Zeileis at uibk.ac.at] Sent: Wednesday, April 22, 2015 3:30 AM To: Michael Grant Cc: r-help at R-project.org Subject: Re: [R] Exhaustive CHAID package On Tue, 21 Apr 2015, Michael Grant wrote:
Dear R-Help: From multiple sources comparing methods of tree classification and tree regressions on various data sets, it seems that Exhaustive CHAID (distinct from CHAID), most commonly generates the most useful tree results and, in particular, is more effective than ctree or rpart which are implemented in R.
I searched a bit on the web for "exhaustive CHAID" and didn't find any convincing evidence that this method is "most commonly" the "most useful". I doubt that such evidence exists because the methods are applicable to so many different situations that uniformly better results are essentially never obtained. Nevertheless, if you have references of comparison studies, I would still be interested. Possibly these provide insight in which situations exhaustive CHAID performs particularly well.
I see that CHAID, but not Exhaustive CHAID, is in the R-forge, and I write to ask if there are plans to create a package which employs the Exhaustive CHAID strategy.
I wouldn't know of any such plans. But if you want to adapt/extend the code from the CHAID package, this is freely available.
Right now the best source I can find is in SPSS-IBM and I feel a bit disloyal to R using it.
I wouldn't be concerned about disloyalty. If you feel that exhaustive CHAID is the most appropriate tool for your problem and you have access to it in SPSS, why not use it? Possibly you can also export it from SPSS and import it into R using PMML. The "partykit" package has an example with an imported QUEST tree from SPSS.
Michael Grant Professor University of Colorado Boulder [[alternative HTML version deleted]]
______________________________________________ R-help at r-project.org mailing list -- To UNSUBSCRIBE and more, see https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.