On 4 Dec 2003 04:59:24 -0800, [EMAIL PROTECTED] (Luis
Domingues) wrote:

> Hi all,
> 
> My question is this:
> What's the apropriated method for find out witch variables ( only
> dicotomic ) contribute for a decision?
> I have a lot (160) of dic. variables.

And, What do you mean by "contribute to a decision"?
You get the univariate test as the 2x2  contingency table
(assuming that was a dichotomous decision).
You get some other *effect*  measure by looking at that
statistic instead:  odds ratio and  beta coefficient  are the
other two that I have used, but there are other ways of
'standardizing'   whatever that effect may be.


> Discriminant analysis is the apropriated method?
> If yes, how shell i use it? Put all variables?

You get some  test "while controlling for something"  by using
some simultaneous approach.  Discriminant function does
that exactly the same as ordinary regression on a 0/1  criterion.

For variable like the ones I usually have, putting in 160  is 
not the smart way to do it;  you don't name your variables.
For mine, items like symptom-ratings, there would be high
intercorrelation -- so the appropriate approach is data
reduction FIRST.

If you do intend to use 160 variables simultaneously, it 
would probably be necessary to have a fuzzy final 
relationship (no perfect R-squared, or near it), and at
least several hundred decisions each of Yes and No
(or whatever the dichotomy).

> 
> Is this the correspondente analysis of Principal components and factor
> analysis for dicotomic variables?

I don't see CA  as being valuable for a dichotomy, 
but you might look up keywords:  CART  and decision-trees.

-- 
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