[R] multi-class modeling

[Prof Brian Ripley]

All of this is addressed in the reference for multinom(): it is support 
software for a book.

(You are likely to get a more sympathetic response if you use a real name 
and a signature giving your true affiliation.)


On Sun, 10 Apr 2005, 'array chip' wrote:

> Just wonder if someone could comment on using linear
> discriminant analysis (LDA) vs. multinomial logistic
> regression in multi-class classification/prediction
> (nomial dependent variable, not ordinal)? What kind of
> difference in results can I expect from the 2 methods,
> which is better or more appropriate, or under what
> condiditon should I used one instead of the other? And
> is there other methods I can try?

> On another note, if I want to use logistic regression
> using multinom() in package nnet, how can I address
> the problem that each class of the dependent variable
> has an unequal prevalence? In lda(), I can do this by
> using prior argument, but there is no similar argument
> in multinom().

Depends what you think the `problem' is.  In lda(), you usually do not
adjust by the 'prior' argument.  Is the problem that your training set is 
a biased sample?  If so, see my PRNN book.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595