On Mon, 6 Oct 2003 23:40:50 -0700, "albinali" <[EMAIL PROTECTED]>
wrote:

> Hi,
>   With non-ordinal categorical data, I was told that logistic regression is
> likely to do a better job, why is that, whats the problem with linear
> regression?

Linear regression is not one of the *legitimate*  options when 
you have several non-ordered categories.  
Say,  Hindu=1, Islam=2, Christian=3, Jewish =4    - and you want
to predict the average?   No, that won't do.

Ordinary logistic regression has only two categories, so
it won't work, either, except as a series of 0/1  contrasts.
You could use linear regression  with 0/1  contrasts, too.

Two programs.
For separating several groups using several 'predictors'
that are continuous variables, discriminant function (DF)
gives the clearer side-statistics, showing means for the 
groups on separate variables, and multivariate centroids 
in 'reduced space';  and the clearer warnings (or clues) 
when assumptions are being violated, including over-fitting.

Multinomial logistic regression (MLR) might provide a more 
precise solution, with nicer theoretical properties, if your  
group  Ns  are all large enough, and if your groups are 
rather distinct.  MLR  is newer (as a part of stat-packages), 
and has some popularity for that reason.  It allows you to 
describe the relationships as odds ratios -- if that appeals 
to you.  

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
"Taxes are the price we pay for civilization." 
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to