Re: [S] discriminant analysis

Prof Brian Ripley (
Mon, 17 Aug 1998 15:26:17 +0100 (BST)

> We have carried out discriminant analysis using SPLUS-4 (Windows95), release
> 3, with the lda function (Ripley library) for predicting group membership
> from 32 explanatory variables. We wish to assess which of this variables
> contributed most strongly to discrimination and to optimise variables
> selection.
> Other packages such as SPSS allow this assessment by judging standardised
> discriminant function coefficients and correlation between function scores
> and explanatory variables. SPSS also allows the statistical significant of
> each discriminant functions to be judged from wilk's lambda tranformed to
> chi square.
> Are these procedures possible with S-PLUS?

I believe they are possible, but I do not think they are implemented,
and they will not be implemented by me, for the good enough reason that
I do not believe in them. (The test statistic is far too closely tied
to multivariate normality and equality of group variances, neither of
which are commonly plausible in 32 dimensions. The correlations are
easily fooled by near-collinearity. And all the measures are highly
vulnerable to multivariate outliers) In SPSS you have (AFAIK) one
man's view dating from the late 60's of how this should be done.

As I said in reply to a similar question about a week ago, linear
logistic discrimination seems much sounder to me, and multinom has a
drop1 method, for example. You can even do stepwise selection via

Brian D. Ripley,        
Professor of Applied Statistics,
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272860 (secr)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595

----------------------------------------------------------------------- This message was distributed by To unsubscribe send e-mail to with the BODY of the message: unsubscribe s-news