After this, 
  7 Jun 2002 07:15:00 -0700, [EMAIL PROTECTED] (Mobile Survey)
wrote:
> 
[ ... ]
> >You have not mentioned whether multicollinearity is present.It is good
> >to calculate the VIF for regressors by first calculating the R^2 from
> >the  regression of one X on all other Xs. If the VIF>5, you have a
> >problem. the t-tests are no longer relaible.
> 

On Sat, 08 Jun 2002 20:10:18 GMT, Dick Startz
<[EMAIL PROTECTED]> wrote:

> Under what mathematical assumptions about the way the regression data
> was generated does multicollinearity mean "the t-tests are no longer
> reliable"?

I had some sympathy with the complaint, 'no longer reliable'.
However, Dick is right:  the main fault is not  reliability. 

How about this.  If you are *surprised*  by a VIF > 5, then
you better re-think the regression -- the test on that
variable does not mean what you thought it did.
Your hypotheses are probably not 'robust', or the 
scoring of them has not been valid.

When the VIF > 5  for  a suppressor-variable relationship
happens to show up, then I don't like the whole model  -- 
I suspect that the actual value of the t-test and the 
coefficients depend on the idiosyncratic sample.
(Thus, *I suspect*  the tests, etc.,  may be unreliable, 
in this less-common example.)

On the other hand, when the VIF > 5,   where I was 
intentionally controlling for something similar, I 
would figure  that a small t-value, produced when 
controlling for simple confounding, is reliable, 
fitting the a-priori model.

-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
.
.
=================================================================
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