Date:         Fri, 30 Sep 2005 09:23:27 -0700
Reply-To:     Howie Harshaw <harshaw@interchange.ubc.ca>
Sender:       "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From:         Howie Harshaw <harshaw@interchange.ubc.ca>
Subject:      Logistic regression and ROC curves
In-Reply-To:  <5.1.0.14.2.20050928114210.05d5e3d0@pop.mindspring.com>
Content-type: text/plain; format=flowed; charset=ISO-8859-1

Hi All,

I have been following recent discussions about using ROC curves to
identify the threshold for the classification cut-off, and think I
understand the utility of this approach.  However, I am a little unclear
about the order of steps to do this.  Is this correct:

1) Run the logistic regression with the default classification cut-off
(0.5), save the predicted probabilities.
2) Use the saved predicted probabilities for the ROC curve.
3) Identify the point where the curve rises rapidly for the new
classification cut-off threshold value.
4) Re-run the logistic regression with the classification cut-off
threshold value identified with he ROC curve.

I am also wondering whether there is a good reference for the use of ROC
curves to identify the classification cut-off threshold value; I haven't
found reference to it in my usual resources.

Thanks for any help you can provide,

- Howie

_______________________________

Howard Harshaw
Post Doctoral Fellow
Forest Resource Management
Faculty of Forestry
University of British Columbia

Tel: (604) 786-3141
E-mail: harshaw@interchange.ubc.ca
Web: www.Harfolk.ca