[Procedures for performing meta-analyses of the accuracy of tools for binary classification]

Area of Science:

• Biostatistics
• Medical Informatics
• Psychometrics

Context:

• Binary classification tools are crucial for early detection and screening.
• Assessing their accuracy requires analyzing non-independent rates: true positives and false positives.
• Existing meta-analysis methods have limitations in handling these rates and covariates.

Purpose:

• To review and summarize meta-analytic methods for assessing the accuracy of binary classification tools.
• To highlight the limitations of direct aggregation and SROC methods.
• To introduce and evaluate statistically rigorous models like Normal Bivariate (NB) and Hierarchical Summary ROC (HSROC) for meta-analysis.

Summary:

• The study reviews meta-analytic techniques for binary classification tool accuracy, focusing on the AUDIT test across 14 studies.
• It critiques methods like direct aggregation and SROC for not accounting for the interdependence of sensitivity and specificity or handling covariates.
• The Normal Bivariate (NB) and Hierarchical Summary ROC (HSROC) models are presented as statistically rigorous alternatives capable of incorporating covariates, demonstrated by analyzing gender composition's effect on AUDIT performance.

Impact:

• Provides a comprehensive overview of meta-analysis techniques for diagnostic accuracy.
• Recommends advanced statistical models (NB and HSROC) for more robust accuracy assessments.
• Demonstrates the utility of these models in identifying factors (e.g., gender composition) influencing test performance.