Message-ID: <6.0.1.1.2.20040121085947.02f1bd10@pop4.attglobal.net>
Date: 2004-01-21T17:05:12Z
From: Marc R. Feldesman
Subject: evaluation of discriminant functions+multivariate homosce dasticity
In-Reply-To: <20040121155802.4542.qmail@web40614.mail.yahoo.com>
At 07:58 AM 1/21/2004, Dave Andrae wrote:
>I seem to remember, from a course in which I used SPSS for LDA, that
>Box's M is an ultra-sensitive test as well and that in almost all
>practical applications it's not useful, so Prof. Ripley's comments
>apply to that test, too.
Professor Ripley is quite right about Box's M. I wrote a crude S-Plus
script for this years ago to see if I could find a real (i.e. not
simulated) data set for which Box's M would give a non-significant
result. Using data from my field (primate and human functional anatomy), I
found no instance where my data weren't "non-normal" by Box's
criterion. And in many of those instances, lda worked "perfectly" anyway
(i.e. 95 - 100% of cases correctly classified).
As far as I'm concerned, Box's M is of no use in anything I do. At the
same time, if you want the crude script (not guaranteed to work in R since
I haven't bothered to test it), I'll send it to you b/c. In my opinion, it
isn't worth the effort to clean it up or to test it under R.