A double SIMEX approach for bivariate random-effects meta-analysis of diagnostic accuracy studies

Area of Science:

• Biostatistics
• Medical Informatics
• Diagnostic Accuracy Research

Background:

• Bivariate random-effects models are standard for meta-analysis of test accuracy studies.
• Existing likelihood methods face challenges with small sample sizes and convergence issues.
• A novel methodology is proposed to address these limitations in meta-analysis.

Purpose of the Study:

• To introduce and evaluate the SIMEX (Simulation Extrapolation) methodology for meta-analysis of diagnostic accuracy studies.
• To address the drawbacks of standard likelihood-based inference methods.
• To provide a more reliable and computationally stable approach for meta-analysis.

Main Methods:

• The study proposes the SIMEX methodology, a simulation-based technique adapted from measurement error correction.
• SIMEX is suitable for meta-analysis as diagnostic accuracy measures have inherent measurement error.
• The method is adaptable to various measurement error structures and covariates, with easy implementation in standard software.

Main Results:

• Simulation studies show SIMEX improves empirical coverage probabilities of confidence intervals compared to likelihood methods.
• SIMEX performance is independent of sample size and the correlation between sensitivity and specificity.
• Significant improvements are observed even with deviations from normality assumptions; SIMEX avoids convergence issues.

Conclusions:

• The SIMEX methodology is a viable and advantageous alternative to likelihood-based inference in diagnostic accuracy meta-analysis.
• It offers enhanced accuracy and avoids convergence failures and numerical instabilities.
• Implementation is feasible using provided R code and illustrated with a real-world example.