Re: Logit regression question

On Thu, 27 Nov 2003 04:50:34 +0900 (KST), wuzzy <[EMAIL PROTECTED]>
wrote:

> if you used dummy-variables to categorize 4 variables with 5 
> layers, you should have less than 20 vars, 4(n-1)=16. You can 
> definitely drop categories since eg., category 2 should not affect the 
> comparison between 1 and 4 if 1 is the control..

I agree that there should be 16 dummies.

I disagree with arbitrarily dropping variables -- whether
they are dummy variables or not.

That is not-good advice since it is potentially a bad 
practice in inference, always.  Regression without all
the "relevant"  variables is biased, remember?  How bad 
it is, here, depends on the definition of the categories 
and what conclusions are being assayed.

Even when the situation may justify it, the simple advice
has to be overly hasty, when you don't know how the
dummy variables are encoded -- What seems apparent, 
here, is that the person asking the question knows little, too.

What is the purpose of dropping the category.
Sometimes it makes sense to combine categories, sometimes
not.  Are you trying to form efficient estimates?  Or, could you
be trying to steal degrees of freedom in order to come up 
with a statistical test that reads as nominally significant at
the fixed test level?

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
"Taxes are the price we pay for civilization."