Both my independent and dependent variables are binary. My result for classification table is 72% for predicted, and my ROC curve area is 0.389. Since <0.5 for ROC area is the worst for accuracy model, what should I do?

| cite | improve this question | | | | |
  • 2
    $\begingroup$ What is 72%? what are you measuring? If AUC is below 0.5, maybe there is some error in how you compute it. Have you by any chance inverted your predictions by mistake? $\endgroup$ – ALiX Jun 22 '12 at 18:13
  • $\begingroup$ To have an ROC below 0.5 there must be some threshold where both classification error rates are below 0.5. However it is possible that for some threshold at least one prediction rate is above 0.5. So what you claim is possible. $\endgroup$ – Michael R. Chernick Jun 22 '12 at 19:00
  • $\begingroup$ No one can answer this without knowing what your problem is and what method(s) are you using. $\endgroup$ – user88 Jun 22 '12 at 19:54
  • 1
    $\begingroup$ Yup; the nice way to have ROC<0.5 is to accidentally swap class labels (as ALiX said), the painful way is to have a paradox problem (i.e. when test contains many objects with the same predictor values as in train but labelled differently). $\endgroup$ – user88 Jun 22 '12 at 20:04
  • $\begingroup$ I didn't mean for it to be an answer. I should have made it a comment to relate to ALiX's question. I will change that. $\endgroup$ – Michael R. Chernick Jun 22 '12 at 20:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.