Frank Harrell provides a comprehensive approach to model building that does away with the need for stepwise procedures. His book is

"Regression Modeling Strategies" published by Springer.

SR Millis

Scheltema, Karen wrote:

I know about the perils of stepwise, and I agree with you that it is a less than desirable procedure. This researcher, however, is not as convinced as I am about not doing stepwise. Sigh. He has more variables than would comfortably fit a 5-1 case to variable ratio for a forced entry regression, which is why he was hoping stepwise would help him narrow his model. Any suggestions I can give him, short of telling him to scrap everything?

Karen Scheltema, M.A, M.S.
Senior Statistician
HealthEast
Research and Education Department, Midway Campus
1700 University Ave W
St. Paul, MN 55104
Ph: (651) 232-5212   fax: (651) 641-0683
mailto:[EMAIL PROTECTED]



-----Original Message-----
From: Paul R Swank [mailto:[EMAIL PROTECTED]
Sent: Wednesday, May 28, 2003 8:26 AM
To: Scheltema, Karen; Ed Stat (E-mail)
Subject: RE: VIF for dichotmous variable


This is another good illustration why one should not use automatic procedures for research. The problem of interactions in stepwise procedures is that the compute doesn't know it is an interaction and that you can't drop the main effects while the interaction is strill in the model. If the interaction is significant then the main ewffects that make up the interaction must stay in the model. By the way, there is backward selection, forward selection, and stepwise selection (plus several other more esoteric procedures) but I, for one, have never heard of backwward stepwise.


Paul R. Swank, Ph.D. Professor, Developmental Pediatrics Medical School UT Health Science Center at Houston


-----Original Message-----
From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]
Behalf Of Scheltema, Karen
Sent: Tuesday, May 27, 2003 4:08 PM
To: [EMAIL PROTECTED]
Subject: VIF for dichotmous variable


I have a colleague who has run a backwards stepwise linear regression. A
couple of the variables are problematic. One is a dichotomous variable that
indicates trying to get pregnant (y/n) and the other is a continuous
variable that is a scale score. It is believed that the two interact with
each other. In the initial run, the scale was centered, but the dichotomous
variable was not. This yielded high VIFs (>100) in the initial regression
equation, but the scale dropped out by the latter stages of the stepwise
regression. I suggested centering the dichotomous variable, as well, but
this yielded VIFs>300 in the initial steps. Since this scale variables
drops out later, can I just ignore this? Or, can someone suggest another
solution to try?




Karen Scheltema, M.A, M.S. Senior Statistician HealthEast Research and Education Department, Midway Campus 1700 University Ave W St. Paul, MN 55104 Ph: (651) 232-5212 fax: (651) 641-0683 mailto:[EMAIL PROTECTED] <mailto:[EMAIL PROTECTED]>




The information included in this e-mail message, including any attachments,
is intended only for the person or organization to which it is addressed.
This e-mail message may contain information that is privileged or
confidential. If you receive this e-mail message and are not the intended
recipient or responsible for delivering the message to the intended
recipient, you may not use, disseminate, distribute or copy the information
included in this e-mail and any attachments. If you received this e-mail
message by mistake, please reply by e-mail and destroy all copies of this
message and any attachments. Thank you.






The information included in this e-mail message, including any attachments, is 
intended only for the person or organization to which it is addressed. This e-mail 
message may contain information that is privileged or confidential. If you receive 
this e-mail message and are not the intended recipient or responsible for delivering 
the message to the intended recipient, you may not use, disseminate, distribute or 
copy the information included in this e-mail and any attachments. If you received this 
e-mail message by mistake, please reply by e-mail and destroy all copies of this 
message and any attachments. Thank you.
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================




-- ~~~~~~~~~~ Scott R. Millis, PhD, ABPP (CN & RP) Director, Office of Clinical Trials Kessler Medical Rehabilitation Research & Education Corp 1199 Pleasant Valley Way West Orange, NJ 07052

tel: 973-243-6976
fax: 973-243-6990
email: [EMAIL PROTECTED]




.
.
=================================================================
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