4
$\begingroup$

enter image description here

Hi, what does the green straight line at the end of the plot indicate? Is there some problem with the model?

$\endgroup$
1
  • $\begingroup$ It indicates that your code did not found threshold where TP rate turns 1 for whatever reason. $\endgroup$
    – Sergei Ozerov
    Commented Feb 2, 2022 at 10:16

1 Answer 1

12
$\begingroup$

Possibility 1 (likely)

When you increase the sensitivity you can do this in small steps until (apparently) a true positive rate of about 65%. After that, the model makes a big jump to 100%. This can occur for instance when 35% of your positive cases have all the same value (e.g. zero).

The straight line occurs because the way that you plot. Your model is computed for points, but you draw a line in between them.

This straight line is not wrong though. If you want to have a positive rate in between 65% and 100% then you can do this by mixing the models with sensitivity of 65% and 100%. (See: Combining classifiers by flipping a coin)

Possibility 2 (not very likely but theoretically possible)

Theoretically, this might also occur if the positive and negative populations have the same distribution density (with only a difference in a constant) in the region where the predictor parameter is set for high sensitivity.

But, it would be unlikely to get such a straight line due to statistical fluctuations in the sample (unless the curve is a theoretical computation). However, in the special case that all the values are all the same (as in the first case), eg. zero, then you don't get statistical fluctuations.

Possibility 3

It might be that you don't compute the entire ROC curve, but the front-end that plots the curve automatically and completes the curve by connecting the end of your computed curve with the point 1,1

$\endgroup$
2
  • $\begingroup$ Thank you so much for the explanation! When you say 'This can occur for instance when 35% of your positive cases have all the same value (e.g. zero).' you don't mean negative cases? because if the value is 0 its mean that its negative cases $\endgroup$
    – user348733
    Commented Feb 6, 2022 at 15:38
  • 1
    $\begingroup$ @user348733 why can't the positive cases have a zero value with some test? E.g. If we want to know whether some box contains a radiative material we can use a Geiger counter to measure it. It is possible that we measure zero counts even if it is a positive case and there is a radiative material inside. $\endgroup$ Commented Feb 6, 2022 at 17:25

Your Answer

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