Components-of-variance models for random-effects ROC analysis: the case of unequal variance structures across...

Area of Science:

• Biostatistics
• Medical Imaging Analysis
• Diagnostic Accuracy Studies

Background:

• Previous receiver operating characteristic (ROC) analysis for diagnostic modalities assumed equal variance components.
• This limitation restricted comparisons between modalities with differing uncertainty structures.

Purpose of the Study:

• To generalize multivariate ROC analysis by removing the assumption of equal variance across compared modalities.
• To develop a method applicable to diagnostic accuracy studies with varying uncertainty.

Main Methods:

• Extended a previous variance components model by splitting three components into modality-dependent contributions.
• Explored two formulations of the generalized model, demonstrating a one-to-one relationship between their component estimates.

Main Results:

• Applied the generalized method to a reader study comparing computer-aided diagnosis (CAD) versus no CAD for microcalcification classification.
• Observed significant reductions in reader and reader-by-case variance with CAD using one formulation.
• Identified substantial decreases in modality-interaction components with CAD using the alternative formulation.

Conclusions:

• Successfully provided a multivariate ROC analysis solution that does not require equal variance structures across modalities.
• Confirmed consistent results across alternative formulations, linked by a one-to-one mapping.
• Found that confidence intervals and sample size estimations remain consistent between the generalized and restricted models.