Related Experiment Videos
Three methods for analysing correlated ROC curves: a comparison in real data sets from multi-reader, multi-case
1Center for Statistical Sciences, Department of Community Health, Brown University, Providence, Rhode Island 02912, USA. toledano@stat.brown.edu
Statistics in Medicine
|September 4, 2003
Summary
This study compares three methods for analyzing correlated ROC curves. Generalized estimating equations provided the most reliable standard errors in simulations, outperforming jackknifed pseudovalues and F-tests.
Area of Science:
- Biostatistics
- Medical Imaging Analysis
- Statistical Modeling
Background:
- Multiple correlated Receiver Operating Characteristic (ROC) curves are common in medical diagnostic accuracy studies.
- Accurate statistical analysis is crucial for comparing diagnostic test performance.
Purpose of the Study:
- To compare three established methods for analyzing multiple correlated ROC curves.
- To evaluate the performance of these methods using real-world data and simulations.
Main Methods:
- Comparison of three analytical methods: generalized estimating equations (GEE) with marginal non-proportional ordinal regression, jackknifed pseudovalues, and a corrected F-test.
- Application of methods to six real data examples from factorial design studies.
- Exploration of differences between typical and ROC-specific summary statistics.
- Conducting a simulation study to validate standard error estimates.
Main Results:
- Point estimates for differences between test modalities were similar across methods.
- Standard errors of these differences varied among the three analytical methods.
- The generalized estimating equations method demonstrated robust standard error estimation in simulation.
Conclusions:
- The generalized estimating equations approach is recommended for analyzing multiple correlated ROC curves due to its reliable standard error estimation.
- Differences in standard errors highlight the importance of choosing appropriate analytical methods for ROC curve analysis.