In <s5581ded.032@harthosp.org>, on 05/12/98 at 10:00 AM, Jeff Mather <Jmather@HARTHOSP.ORG> said:

>This is a bit off the SPSS path, but can anyone tell me if there is any >way to use a nominal response to draw an ROC curve? I'd like to compare >ultra-brief (one question) instruments for a diagnosis of depression (yes >or no) with standard questionnaire that uses a cutoff on an interval >scale.

>My assumption was that this could not be done with a y/n response; that >you need at least a three point ordinal scale (ie. always, sometimes, >never) to be able to draw the curve. >thanks

Right. Essentially what you get from a y/n test is just one point in the true pos vs false pos ROC plot. With a three point scale you get two points. Doesn't really make a curve.

ROC techniques are designed to quantify, for continuous or multi-category test scales, how good it discriminates between two groups. That is, you need a dichotomous reference or gold standard to compare to. If your dichotomized standard is in fact the definition of depression, you can use ROC curves in principle. With a dichotomous test however that doesn't add anything beyond your sensitivity and specificity.

Hope this helps,

-- Arnold Kester Arnold.Kester@stat.unimaas.nl