If I have values for sensitivity and specificity for a group of studies, for example like this:
sensitivity specificity ----------- ----------- 66.3 74.7 87.2 65.9 56.4 76.4 79.5 94.3 How can I make an ROC curve for such values? Do I divide them by 100 for example and just put the values in the curve straightforward? Or, how exactly does the process work out?
Best Answer
You seem to have several readings of $text{sensitivity}$ and $text{specificity}$ for some classifier; yet ROC curve does not work that way. It is a way of plotting the performance of a learner giving some numerical confidence that some object is of a class A and not B instead of certain prediction (i.e. this object is A, that is B).
Obviously such confidence-score-giving learner can be converted into a plain two class classifier by defining some threshold $t$ and saying that it votes for class A if the $text{score}>t$ and for B otherwise. And, for this classifier, one can get $text{sensitivity}$ and $text{specificity}$.
Thus, for such learner, you can get two functions: $text{sensitivity}(t)$ and $text{specificity}(t)$; ROC is a visualization of those two functions as a parametric curve $(1-text{specificity}(t),text{sensitivity}(t))$. In practice $t$ has finite number of unique values (at most equal to number of objects), so the curve is constructed from points for all of them.