1
$\begingroup$

So I have read in posts and in literature (Frank Harrel - Regression Modeling Strategies fx) that depending on what you do, ROC curves and AUC values are not always relevant, but often written in papers or reports anyways, and sometimes as the "ground truth" to how good a model actually is. My problem is that I'm not entirely sure I understand why it's not always a good thing? I mean, if I have a model (based on binary outcomes fx) that predicts some value for some input parameters in the model, I would think it made much sense that the highest predicted values should "align" with the outcomes, i.e. a high AUC value if it does indeed predict well. That makes sense to me. But maybe I have misunderstood/overlooked something?

$\endgroup$
1
$\begingroup$

If you look at a contigency table and the various metrics involved you will see that the AUROC curve uses the TPR and FPR metrics, which are often times relevant. However, sometimes you may want to look at it from a different perspective, for ex. the FPR and TNR metrics.

| cite | improve this answer | |
$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.