Extending the DeLong algorithm for comparing areas under correlated receiver operating characteristic curves with...

Area of Science:

• Biostatistics
• Statistical Methods
• Medical Data Analysis

Background:

• The DeLong method (1988) is widely used for comparing areas under correlated receiver operating characteristic (ROC) curves.
• Existing software implementations of the DeLong method may produce invalid or inefficient results by excluding individuals with any missing data.
• Missing data is a common issue in real-world datasets, necessitating robust statistical approaches.

Purpose of the Study:

• To address the limitations of the DeLong method concerning missing data.
• To develop a simplified and more robust algorithm for comparing areas under correlated ROC curves.
• To provide a valid and efficient method for handling missing data in multivariate ROC analyses.

Main Methods:

• A simplified version of the DeLong algorithm using ranks was developed.
• A mixed-model approach was employed to extend the algorithm for multivariate data with missing values.
• The proposed procedure was validated through simulation studies for data missing at random.

Main Results:

• Simulation results confirmed the validity and efficiency of the proposed procedure.
• The new method effectively accommodates missing data, overcoming limitations of the original DeLong approach.
• The procedure demonstrated superior performance in handling datasets with missing values compared to standard implementations.

Conclusions:

• The proposed mixed-model approach provides a valid and efficient alternative for comparing areas under correlated ROC curves with missing data.
• This method enhances the reliability of statistical analyses in the presence of incomplete datasets.
• The procedure is illustrated and available in popular statistical software (SAS, Stata, R).