A marginal model approach for analysis of multi-reader multi-test receiver operating characteristic (ROC) data.

Area of Science:

• Biostatistics
• Medical Imaging Analysis
• Diagnostic Test Evaluation

Background:

• Receiver operating characteristic (ROC) curves are vital for assessing diagnostic test performance.
• The multiple readers, multiple tests design is common in radiology, relying on subjective interpretation.
• The Dorfman-Berbaum-Metz (DBM) method is widely used but lacks a clear theoretical foundation.

Purpose of the Study:

• To investigate the theoretical basis of the DBM method for continuous outcomes.
• To propose and validate a new marginal model approach for analyzing diagnostic test accuracy.
• To compare the proposed method with the DBM method using simulations and real-world data.

Main Methods:

• Theoretical analysis of the DBM method's assumptions within standard ANOVA models.
• Development of a marginal model based on area under the ROC curves (AUCs) for continuous outcomes.
• Simulation studies and application to a breast cancer diagnostic study dataset.

Main Results:

• The DBM method violates standard ANOVA assumptions, potentially causing erroneous inference.
• The proposed marginal model provides consistent and asymptotically normal estimators for regression coefficients.
• Both methods perform well when test accuracy is uniform; the marginal model is superior when accuracy varies.

Conclusions:

• The DBM method's theoretical basis is questionable for diagnostic accuracy analysis.
• The proposed marginal model offers a statistically sound and more reliable alternative, extendable to ordinal outcomes.
• This new approach enhances the accuracy of inference for diagnostic tests, particularly in varied accuracy scenarios.