Analysis of clustered data in receiver operating characteristic studies

Area of Science:

• Biostatistics
• Statistical Methods
• Machine Learning

Background:

• Clustered data presents unique analytical challenges distinct from simple correlation.
• Receiver Operating Characteristic (ROC) curve analysis is crucial for evaluating diagnostic tests, but standard methods may not apply to clustered data.
• Existing statistical methods for clustered data have limitations in scope and accessibility for ROC analysis.

Purpose of the Study:

• To review and compare existing statistical methods for correlated receiver operating characteristic (ROC) curve data extended to clustered settings.
• To provide suggestions for applying methods not yet extended to clustered ROC studies.
• To evaluate methods based on their ability to address objectives in clustered ROC data analysis, considering ROC indices and researcher accessibility.

Main Methods:

• Review of existing statistical methodologies for clustered ROC data.
• Comparison of parametric, nonparametric, and jackknife methods for clustered ROC analysis.
• Exploration of potential extensions for methods not yet applied to clustered ROC studies.

Main Results:

• Parametric models allow consideration of all ROC indices but are computationally complex and least accessible.
• Nonparametric methods are more accessible but limited to ROC curve area estimation and inference.
• The jackknife method offers the highest accessibility, permitting consideration of any ROC index.

Conclusions:

• The choice of statistical method for clustered ROC data depends on the desired ROC indices and required accessibility.
• Future research should focus on developing methods using the continuation ratio model and bootstrapping for clustered ROC studies.
• Advancements in statistical software and methods are needed to improve the analysis of clustered ROC data.