At 02:42 PM 6/9/2009, John Antonakis wrote:
Hi:
First, don't use stepwise regression--it is the plague; no worse. Many
journals simply won't even review manuscript with such data-driven
methods (unless used for a particular goal--ridge, least-angular
regression). For instance, see:
It may depend on how you package it. On the one hand, stepwise
regression is the work of Satan; on the other hand diagnostic tests
are good. So, one might just run the -estat ovtest- command and,
based on it, decide that some sort of non-linearity should be allowed
for in the model.
Also, I agree that plotting is a good idea, but really, how much
different is that than stepwise regression? Either way, you are
basically looking at the data and deciding what to do with it. Both
eyeballing the data and stepwise regression have the potential to
make your significance tests deceptive, because you are using
knowledge gained from the data itself and hence potentially
capitalizing on chance in building your model. I guess the human
judgment aspect of plotting appeals to me over stepwise, but I also
think it raises some of the same problems and concerns.
Also, I don't know why you would need stepwise regression to justify
the possible inclusion of an x^2 term in the model. You can often
think of good substantive reasons for a curvilinear relationship. It
therefore seems reasonable to me to test whether an x^2 term belongs
in the model. I would certainly rather have some a priori theory in
there about an x^2 term rather than just seeing if it happens to make
it into a stepwise regression.
My own view is that mindless stepwise is indeed the work of
Satan. However, I think stepwise is potentially useful as a
diagnostic device, e.g. is there reason to believe that my model may
have omitted important variables? Even so, I think it is better if
you have clear alternative or rival hypotheses in mind and explicitly
test them rather than go on a fishing expedition with sw.
-------------------------------------------
Richard Williams, Notre Dame Dept of Sociology
OFFICE: (574)631-6668, (574)631-6463
HOME: (574)289-5227
EMAIL: Richard.A.Williams.5@ND.Edu
WWW: http://www.nd.edu/~rwilliam
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
