Date: Mon, 3 Feb 1997 18:32:35 EST
Reply-To: peter homel <HOMEL@SACC.HSCBKLYN.EDU>
Sender: "SPSSX(r) Discussion" <SPSSX-L@UGA.CC.UGA.EDU>
From: peter homel <HOMEL@SACC.HSCBKLYN.EDU>
Subject: stepwise entry for logistic regression.
I have a client who is looking at a yes/no kind of outcome and has a
mixture of categorical and continuous variables as predictors. There
are about 10 predictors and all the client wants to do is find the
ones that optimize the prediction of the outcome (i.e., the maximum
prediction with the smallest set of variables). In the past, this client
used regular multiple regression to analyze these data. I suggested
that he would be better off using logistic regression.
I carried out the logistic regression using forward conditional stepwise
inclusion in SPSSv6. The client objected that he needed to do full stepwise
because he used that in his previous study. I asssured him that forward
stepwise in logistic regression was the closest equivalent to stepwise
in regular multiple regression. He was not convinced.
I am very aware of the perils of doing any stepwise procedure which is probably
why full stepwise was not implemented in the logistic regression routine.
However, stepwise results are still widely accepted in my client's field
and so he has the problem of explaining and justifying why he is using
forward stepwise rather than the more popular (and equally questionable)
full stepwise method. Telling him that the two methods are equally
questionable will not go over very well with this client. Likewise,
telling him that he should come up with a theory to guide him in his
selection of variables (as opposed to the machine selecting it for him)
will not be received very well. The only other solution for me is to
redo everything in SAS which does allow for full stepwise inclusion in
Any ideas out there? I don't want to rehash the many arguments against
using stepwise methods in regression. My position is simply, given that
that I am forced to sin, how can I sin such that it does the least damage
to all concerned. Actually, I would prefer not doing the extra work in SAS TIA.
PETER HOMEL PHD
HEALTH SCIENCE CENTER BROOKLYN
STATE UNIVERSITY OF NEW YORK
450 CLARKSON AVENUE BOX 7
BROOKLYN, NY 11203-2098
TEL: (718) 270-7424
FAX: (718) 270-7461
MOTTO: STATISTICS DON'T LIE!(PEOPLE DO!)