Measurement error and confidence intervals for ROC curves

Area of Science:

• Biostatistics
• Medical Imaging Analysis
• Diagnostic Test Evaluation

Background:

• Measurement error in continuous test variables can negatively impact the accuracy of receiver operating characteristic (ROC) curve properties.
• Unbiased measurement error typically diminishes the diagnostic utility of continuous test variables.

Purpose of the Study:

• To investigate the influence of potentially heterogeneous measurement error on estimated ROC curves for binormal test variables.
• To develop and validate corrected estimators for specific points on the ROC curve.

Main Methods:

• Derivation of corrected estimators assuming known or estimated measurement variances.
• Simulation study to assess the approximate unbiasedness of the proposed estimators for moderate sample sizes.
• Application to breast cancer imaging data to demonstrate practical utility.

Main Results:

• Corrected estimators provide approximately unbiased estimates for ROC curve points.
• Associated confidence intervals are robust and do not rely on normality assumptions for measurement error distribution.
• The proposed methods were successfully applied to real-world breast cancer imaging data.

Conclusions:

• The developed corrected estimators effectively address bias introduced by measurement error in ROC curve analysis.
• These techniques enhance the reliability of diagnostic accuracy assessment, particularly in medical imaging.
• The methods offer a valuable tool for analyzing continuous test variables with measurement error in various clinical research settings.