Logistic regression question

n.tiliopoulos_at_ed.ac.uk
Date: 03/26/05 

Date: 26 Mar 2005 10:11:19 -0800

Dear all,

I recently submitted a paper to a journal and I have received feedback
from the reviewers to which I am not sure how to respond. Basically, I
run a stepwise binary logistic regression (forward Ward - no particular
reason for choosing this) with 12 predictors. The final model had 5
predictors with statistically sig. ORs. The reviewers commented that
stepwise regr tends to generate regression equations that overfit the
data. Is this true? I was under the impression that, if any, the enter
method (all predictors in the model) was more likely to produce
overfitting results. They suggest running the enter method instead, but
I don't see the point, especially given that when I run it, it is
still these final 5 variables that have sig. ORs. By the way, this was
an exploratory study with no prior knowledge of the predictors
hierarchical importance (thus I could not run any hierarchical regr.
models).

Finally, the reviewers said that "because stepwise regression can
result in the loss of variables important to the dependent variable we
recommend increasing the criteria for inclusion of a variable in the
model from .05 to .15 or .20." I get the importance bit, but is this
step sensible? To me it seems to contradict their previous point about
overfitting.

Any advice/suggestions on how to respond to these comments would be
highly appreciated.

Best

Niko Tiliopoulos

Relevant Pages

  • Re: Logistic regression question
    ... > from the reviewers to which I am not sure how to respond. ... > predictors with statistically sig. ... > stepwise regr tends to generate regression equations that overfit the ...
    (sci.stat.edu)
  • Re: Logistic regression question
    ... > from the reviewers to which I am not sure how to respond. ... > stepwise regr tends to generate regression equations that overfit the ... > overfitting results. ...
    (sci.stat.edu)