Order-restricted inference for clustered ROC data with application to fingerprint matching accuracy

Area of Science:

• Statistics
• Machine Learning
• Biometrics

Background:

• Receiver Operating Characteristic (ROC) curves are vital for evaluating classification accuracy in biometrics and medicine.
• Existing methods struggle with ordered conditions and clustered/correlated data, complicating ROC curve estimation.
• Stochastic ordering and within-cluster correlations are key challenges in real-world ROC analysis.

Purpose of the Study:

• To propose a novel method for modeling ROC curves that accounts for order constraints and within-cluster correlations.
• To improve the statistical efficiency of ROC curve estimation in complex data structures.
• To analyze the algebraic properties and asymptotic behavior of the new estimators.

Main Methods:

• Utilizing a weighted empirical process to model ROC curves.
• Jointly incorporating order restrictions and within-cluster correlation structures.
• Deriving asymptotic properties and studying algebraic expressions for summary statistics like area under the curve.

Main Results:

• The proposed order-restricted estimators demonstrate improved statistical efficiency.
• New estimators exhibit smaller mean-squared errors compared to existing methods.
• Simulation studies confirm superior performance for the proposed method with finite samples.

Conclusions:

• The novel weighted empirical process effectively models ROC curves with order constraints and correlations.
• The developed method offers enhanced accuracy and efficiency for ROC analysis in complex datasets.
• The approach is validated through theoretical analysis and practical application on fingerprint data.