0

If we would like to find the optimal cut point using minimizes distance between ROC plot and point (0,1), is it a acceptable value in the field for the roc01 value? I have searched for the reference and could not find one. Could anyone give some guidance on this?

stataq
  • 1
  • You say "point (0,1)" but do you mean the identity line - i.e. the ROC of the "worthless predictor"? I think you would rather maximize the distance from this line, and from which point distance is measured, geometrically, introduces as much nuance/confusion as not meeting this criterion at all. When measured orthogonally, WLOG you can just choose the maximum of (sensitivity + specificity)^2. That doesn't make a good cut off in general, a false negative is rarely exactly worth a false positive. – AdamO May 09 '22 at 16:28
  • Welcome to CrossValidated! Your question refers to "...a acceptable value in the field for the roc01 value" which implies that you are using a specific implementation of an ROC. It might be helpful to include information on what software you are using and perhaps add a tag so to get an answer to your question. – jluchman May 09 '22 at 16:39
  • I am using cutpointr package to find optimal cut points. for other matric, I can find the reference, such as Youden Index: 0.85 indicate that the diagnostic accuracy of the test is excellent, kappa below 0.60 indicates inadequate agreement. But for roc01: minimizes distance between ROC plot and point (0,1), how can we tell how good or bad the test? any reference value we can use for judgement? – stataq May 09 '22 at 16:53
  • @AdamO point (0,1) would be the top left corner, so the perfect ROC curve. – Calimo May 11 '22 at 05:58
  • @stataq I don't really understand what you are asking, but this might help: https://stats.stackexchange.com/a/483289/36682 – Calimo May 11 '22 at 06:04

0 Answers0