ROC Analyses Based on Measuring Evidence Using the Relative Belief Ratio.

Area of Science:

• Statistics
• Biostatistics
• Medical Diagnostics

Background:

• Receiver Operating Characteristic (ROC) analyses are crucial for evaluating diagnostic tests.
• Existing methods often rely on specific distributional assumptions for measurement data.

Purpose of the Study:

• To develop a robust methodology for ROC analyses applicable to various distributional assumptions.
• To provide a framework for incorporating prior knowledge into ROC analysis.
• To derive inferences for key classification performance metrics.

Main Methods:

• Characterization of statistical evidence based on prior distributions for population distributions and disease prevalence.
• Development of elicitation algorithms for selecting appropriate prior distributions.
• Derivation of inferences for Area Under the Curve (AUC) and classification cutoffs.

Main Results:

• The methodology accommodates both parametric (e.g., binormal) and nonparametric models.
• Elicitation algorithms facilitate the practical application of prior specifications.
• Inferences are provided for AUC, classification cutoffs, and error characteristics.

Conclusions:

• The proposed approach offers a flexible and principled framework for ROC analysis.
• It enhances the assessment of classification performance by integrating prior information.
• The methods are applicable across a range of statistical models for diagnostic accuracy.