I am going to disagree with both halves of this advice -
On Fri, 06 Jul 2001 10:17:17 +0100, Thom Baguley
<[EMAIL PROTECTED]> wrote:
> Brian MacDonald wrote:
> >
> > I am doing a series of analyses using discriminant analysis to predict group
> > membership. Several of the variables I am using show distributions that are
> > not normal. My question is can these (and for that matter shold they) be
> > somehow transformed so that the resulting distribution looks "and presumably
> > acts in the analyses) like a normal distribution.
TB >
> It depends. For some distributions it is easy to do the
> transformations (e.g., log is often appropriate for +ve skew). An
To my way of thinking,
"easy" would not be a guide to transformations, except that easy
tends to be "sensible." (Pay more attention to sensible.) Outliers
or long tails can make it hard to model a variable, or hard to find it
contributing in a two group comparison.
> alternative approach might be to consider logistic regression which
> has several advantages over discriminant analysis and doesn't
> require normality.
Advantages of LR over DF are minimal, and "doesn't require
normality" is an overstatement of the contrast. Both procedures
effectively compute a composite score which orders the cases.
When you compute a composite using variables that have bad
distributions, you get a lousy composite. - an extreme computed
score doesn't hurt the projection in LR if was projected into the
right group; that is a tiny, some-time advantage for LR.
Everyone agrees that the "model" for maximum likelihood LR
is more precise when it is okay to over-predict the "caseness."
That is especially an advantage when the overall prediction
is quite good. (Or: there is no essential difference if the
R-squared is small.)
It is my opinion that the testing for DF has advantages when
the samples are not huge: the DF overall test is more robust,
and the ancillary statistics are easier to understand and display.
--
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
