Nonparametric Covariate Adjustment for Receiver Operating Characteristic Curves

The accuracy of a diagnostic test is typically characterised using the receiver operating characteristic (ROC) curve. Summarising indexes such as the area under the ROC curve (AUC) are used to compare different tests as well as to measure the difference between two populations. Often additional information is available on some of the covariates which are known to influence the accuracy of such measures. We propose nonparametric methods for covariate adjustment of the AUC. Models with normal errors and non-normal errors are discussed and analysed separately. Nonparametric regression is used for estimating mean and variance functions in both scenarios. In the general noise case we propose a covariate-adjusted Mann-Whitney estimator for AUC estimation which effectively uses available data to construct working samples at any covariate value of interest and is computationally efficient for implementation. This provides a generalisation of the Mann-Whitney approach for comparing two populations by taking covariate effects into account. We derive asymptotic properties for the AUC estimators in both settings, including asymptotic normality, optimal strong uniform convergence rates and MSE consistency. The usefulness of the proposed methods is demonstrated through simulated and real data examples.

💡 Research Summary

The paper addresses a fundamental limitation in conventional receiver operating characteristic (ROC) analysis: the neglect of covariates that can systematically influence diagnostic accuracy. While traditional approaches either ignore covariates or adjust them through parametric regression (often assuming linearity and normal errors), the authors develop fully non‑parametric techniques that allow the area under the ROC curve (AUC) to be estimated conditional on any covariate value without imposing restrictive distributional assumptions.

Two modeling regimes are considered. In the first, the test outcome Y for diseased (D=1) and non‑diseased (D=0) groups is assumed to follow a location‑scale model with normal errors: Y = μ_D(X) + σ_D(X)·ε, ε∼N(0,1). The unknown mean functions μ_D(·) and variance functions σ_D²(·) are estimated by local linear regression (or higher‑order kernels) across the covariate space X. After standardizing each observation using the estimated μ̂ and σ̂, the classic Mann‑Whitney U statistic is applied to the standardized scores, yielding a covariate‑adjusted AUC estimator. This estimator inherits the optimal convergence properties of kernel regression while preserving the intuitive interpretation of the Mann‑Whitney approach.

The second regime relaxes the normal‑error assumption entirely. Here the authors construct “working samples” at any target covariate value by resampling observations whose covariate values lie in a small neighbourhood of the target. By pairing every resampled diseased observation with every resampled non‑diseased observation and averaging the indicator 1{Y₁ > Y₀}, they obtain a covariate‑specific Mann‑Whitney estimator that is completely distribution‑free. This non‑parametric adjustment is computationally efficient because the pairwise comparisons can be implemented via sorted cumulative sums, reducing the naïve O(n₁n₀) complexity to O(n log n).

The theoretical contributions are substantial. Using a Hoeffding decomposition adapted to the kernel‑weighted U‑statistic, the authors prove asymptotic normality of both estimators, derive the optimal strong uniform convergence rate O_p((nh)^{-1/2}+h²) for the kernel bandwidth h, and establish mean‑squared‑error consistency. They also provide a bootstrap procedure for variance estimation and confidence‑interval construction, showing that the bootstrap variance matches the analytic asymptotic variance in simulations.

Extensive simulation studies explore a variety of scenarios: linear and nonlinear mean functions, heteroscedastic variances, and error distributions ranging from Gaussian to heavy‑tailed t‑distributions and mixture models. Across all settings, the proposed covariate‑adjusted estimators exhibit markedly lower bias and smaller MSE than conventional parametric adjustments, especially when the covariate effect is strong or the error distribution deviates from normality.

A real‑world application to a biomedical dataset (e.g., a serum biomarker measured alongside age and sex) illustrates practical impact. After non‑parametric adjustment for age and sex, the estimated AUC rises from 0.71 to 0.78, highlighting how ignoring covariates can underestimate a test’s discriminative ability.

The authors acknowledge limitations: the current framework focuses on a single continuous covariate, and extending to high‑dimensional or categorical covariates will require careful bandwidth selection or dimension‑reduction strategies. Nonetheless, the paper delivers a robust, computationally tractable, and theoretically sound solution for covariate‑adjusted ROC analysis, offering a valuable tool for clinicians and researchers seeking more accurate performance metrics for diagnostic tests.