HELP with ROC analysis from logistic regression
CP
I have done a logistic regression analysis on ecological
absence/presence data. I would like to perform a ROC curve
analysis on the models I have generated. How do you perform a
ROC analysis on the entire model and not just a single
variable in SPSS 10.0?
Thanks
David M. Fresco, Ph.D.
"CP" <cp782...@csbi.net> wrote:
ROC appears as a graphing option in SPSS 9 and beyond. Unlike STATA
which derives the ROC curve from a logistic regression, SPSS does so
entirely from the Graph menu. The logistic regression behind the scenes
and does not appear in the output. In general, I like STATA better for
ROC, but SPSS is nice by allowing one to plot several ROC curves on one
plot.
--
David M. Fresco, Ph.D. Internet: fre...@cattell.psych.upenn.edu
Adult Anxiety Clinic of Temple fre...@astro.ocis.temple.edu
Department of Psychology __o
1701 N. 13th Street \<,
Philadelphia, PA 19122 `,/'(*)
Voice: (215) 204-3738 (*) . ./"""
Voice: (215) 204-2155 """"
http://www.psych.upenn.edu/~fresco/helplessness.html
Rich Ulrich
<fre...@cattell.psych.upenn.edu> wrote:
> In article <pf9r6.21$E76....@nnrp2.sbc.net>,
> "CP" <cp782...@csbi.net> wrote:
>
> >Hi,
> >I have done a logistic regression analysis on ecological
> >absence/presence data. I would like to perform a ROC curve
> >analysis on the models I have generated. How do you perform a
> >ROC analysis on the entire model and not just a single
> >variable in SPSS 10.0?
> >Thanks
>
> ROC appears as a graphing option in SPSS 9 and beyond. Unlike STATA
> which derives the ROC curve from a logistic regression, SPSS does so
> entirely from the Graph menu. The logistic regression behind the scenes
> and does not appear in the output. In general, I like STATA better for
> ROC, but SPSS is nice by allowing one to plot several ROC curves on one
> plot.
- I have thought of ROC as a method of examining cutoffs, and for
presentation of sensitivity/specificity results. You can use that on
a single variable about judgments, or you can build a composite
score; the easy way is as a multivariable formula.
The original question suggests (I think) that there might be an
analysis that selects one statistical criterion (area under the
curve?), from among the several, and uses it to build an 'optimal'
formula from the variables.
Is that what it means? Has anyone done that?
--
Rich Ulrich, wpi...@pitt.edu
http://www.pitt.edu/~wpilib/index.html
CP
created from logistic regression (backwards stepwise) and am basically
looking for a representation of it as being the best and was hoping to use
Roc to show it - however in SPSS 10.0, it appears that you can only use one
variable (correct me if i am wrong) and i need to be able to graph the whole
model, all variables included. Would an acceptable way to do this be to run
the logistic regression at different thresholds, then calculate the
sensitivity/specificity from the classification plots generated for each and
then graph the roc curves by hand?? I guess i'm looking for specific
directions which i have not been able to find in the literature. I'm sort of
a newbie in case you haven't guessed.
Rich Ulrich
- You want the logistic regression to give you a predicted value
for every case, /CASEWISE= PRED.
That gives you the same thing as if you did a long
COMPUTE statement using all the predictors.
CP
0.1-0.9) from running a logistic regression at every threshold: # of true
positives, false positives, false negatives, true negatives, sensitivity,
and 1-specificity - so i can sort of plot the curve, but that will tell me
nothing about the auc value (i KNOW i am doing this backwards). Does a
software program exist that will let me plug in these values to get my
curve - i have looked at accuroc, rockit, medcalc, analyse-it and of course
spss. Or can i figure out the auc myself (understand it's a complicated
procdure)?
Rich Ulrich
> so..... i now have the following values for 9 threshold values (from
> 0.1-0.9) from running a logistic regression at every threshold: # of true
? " from running a logistic at every threshold?" Who suggested that?
Where do you get a bunch of thresholds, anyway?
Do you have data with continuous scores? (which would make
one wonder, Shouldn't you use OLS regression instead of LR?)
As I suggested a couple of days ago,
" - You want the logistic regression to give you a predicted value
for every case, /CASEWISE= PRED."
Who knows WHAT you have, with whatever-you-have-invented
by running 9 logistic regressions.