Simultaneous inference for partial areas under receiver operating curves -- with a view towards efficiency

TL;DR

This paper introduces new nonparametric methods for simultaneous inference on partial areas under ROC curves, which are more clinically relevant and potentially more efficient than total area inference.

Contribution

It develops a nonparametric multiple contrast test for partial ROC areas that controls error rates and demonstrates improved efficiency over total ROC area inference.

Findings

The proposed test controls family-wise error rate asymptotically.

Finite sample simulations show good performance of the test.

Evidence suggests partial ROC area inference is more efficient than total ROC area inference.

Abstract

We propose new simultaneous inference methods for diagnostic trials with elaborate factorial designs. Instead of the commonly used total area under the receiver operating characteristic (ROC) curve, our parameters of interest are partial areas under ROC curve segments that represent clinically relevant biomarker cut-off values. We construct a nonparametric multiple contrast test for these parameters and show that it asymptotically controls the family-wise type one error rate. Finite sample properties of this test are investigated in a series of computer experiments. We provide empirical and theoretical evidence supporting the conjecture that statistical inference about partial areas under ROC curves is more efficient than inference about the total areas.