The lay abstract featured today (for Confidence Intervals for AUC and pAUC by Empirical Likelihood by Yumin Zhao, Xue Ding, Mai Zhou) is from Statistics in Medicine with the full article now available to read here.

How to cite

Zhao, Y., Ding, X. and Zhou, M. (2025), Confidence Intervals for AUC and pAUC by Empirical Likelihood. Statistics in Medicine, 44: e70192. https://doi.org/10.1002/sim.70192

Lay Abstract 

Is the new test for cancer more accurate than the old one? Is the new face-recognition model
better than the old? Does the new radar system detect a specific object better than the old system?

A crucial metric for evaluating the performance of these binary classification models is the AUC,
and the pAUC is a similar measure but more focused on a specific part.

Unfortunately, a new test or new models’ AUC/pAUC is unknown and needs to be estimated by going through a training data set or sets.

Further, improvements over the old decision-making models, if any, often come in small increments. Therefore, it is a challenge to be able to quickly recognize those small improvements on AUC/pAUC using limited testing data with confidence.

In this paper a new statistical procedure is proposed to better identify/confirm those small but
definite improvements measured by the AUC/pAUC for binary classification models.

The method is applicable in the situation where no specific probability distribution (for example, normal distribution) needs to be assumed for the tests or models (therefore, less restrictive).

It is demonstrated using simulations that the proposed method has consistently exhibits smaller errors when compared to the existing methods.