I am currently having confusion on a part in the paper:

Unal, Ilker. “Defining an Optimal Cut-Point Value in ROC Analysis: An Alternative Approach.” Computational and mathematical methods in medicine vol. 2017 (2017): 3762651. doi:10.1155/2017/3762651

Looking at Figure 2 and Table 2, why are the optimal cutoff values greater than 1? I always thought that it had to range between 0 to 1. Or are the concepts of cutoff values and thresholds two different things? Any help would be much appreciated!

  • 1
    $\begingroup$ If you have probability values, sure, but you can do ROC curves with log-odds from a logistic regression that are not limited to $[0,1]$. What’s being thresholded? $\endgroup$
    – Dave
    Commented Feb 26, 2022 at 23:14
  • $\begingroup$ @Dave Ok, so not all classification models output the entries iny_prob in the range $[0,1]$. Then is it wrong to say that when we try to find the optimal cutoff value $c$ via Youden, Concordance Probability Method, and the Index of Union, the objective function we have to minimize/maximize will be over the interval $[0,1]$? So instead it should be over $\mathbb{R}^+$? Do you know of any examples or resources that clarify this for me? Thank you again! $\endgroup$ Commented Feb 26, 2022 at 23:29
  • 2
    $\begingroup$ Log-odds in logistic regression can be any real number. For a resource, I would say to read the very paper you’re reading where something other than a probability is used. // You might not even want to look for a threshold using a ROC curve or even at all. Please refer to Vanderbilt’s Professor Frank Harrell’s blog. fharrell.com/post/mlconfusion fharrell.com/post/class-damage fharrell.com/post/classification $\endgroup$
    – Dave
    Commented Feb 26, 2022 at 23:50


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.