There is a paper J Classif. 2020 Oct; 37(3): 696–708. that outlines an important point which is that using a linear interpolation from the ROC curve with binary predictors which is most commonly done in software (R, Python, Stata, and SAS) can lead to misleading results.

I think there is a simple solution to avoid the pitfall of current methods being used as follows:

Lets take the data in Table 1 of the paper, this generates a odds ratio of 2.321429

ln OR = 0.8421828

exp(lnOR/2) = 1.5236235

thus Odds(AUC) = 1.5236235

Thus AUC = 1.5236235 / 2.5236235 = 0.6037

I think this solves the problem for univariate binary predictors because the OR is threshold independent and thus acts as a transform of the AUC. Any thoughts on this would be appreciated