[S] More on the trellis question

Richard E Higgs Jr (HIGGS_RICHARD_E_JR@lilly.com)
Tue, 10 Mar 1998 13:25:29 -0500

>Before I try to reinvent the wheel, are there any supplementary functions
>available to perform "leave-one-out" crossvalidation and generate the
>resulting "confusion matrix" for a linear or quadratic discriminant
>function analysis. I have mda, fda, and lda, but crossvalidation does not
>seem to be intrinsic to any of these routines.

I don't know of any in S+. However, if you implement your own (or look at
someone else's) you should be aware of update equations (sometimes called
perturbation equations) that can be used for quadratic discriminant analysis.
That is, you don't need to re-compute the inverse of the var-covar matrix for
each observation you leave out. These update equations are straightforward
and can be found on pg. 225 of "Statistical Pattern Recognition", 1990
Academic Press Inc. by Ken Fukunaga. This will save you lots of time if you
have moderately large data sets to analyze.

Unfortunately, I most often prefer LDA to QDA and the solution for LDA is not
so nice (as far as I know).

For LDA, Fukunaga shows perturbation equations if the var-covar matrix = I. I
played around once with deriving perturbation equations for the more general
case with LDA but remember running into some difficulties (perhaps that's why
Fukunaga limited his text to the identity case). It would be interesting to
know if this has been done before.

Good luck,
Rick Higgs

This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news