On estimating a constrained bivariate random effects model for meta-analysis of test accuracy studies

Area of Science:

• Biostatistics
• Medical Informatics
• Epidemiology

Background:

• Traditional meta-analysis for test accuracy may yield implausible estimates when study settings vary.
• Unconstrained bivariate random effects models (BRM) do not incorporate setting-specific constraints.
• Lack of applicable estimates can limit the utility of meta-analysis in practice.

Purpose of the Study:

• To develop and evaluate a constrained bivariate random effects model (CBRM) for meta-analysis.
• To improve the plausibility and accuracy of test sensitivity and specificity estimates.
• To address limitations of unconstrained models in setting-specific meta-analysis.

Main Methods:

• Developed a constrained bivariate random effects model (CBRM) using a penalized likelihood approach.
• Employed an optimization algorithm based on co-ordinate ascent and Newton-Raphson iteration.
• Compared CBRM performance against the unconstrained BRM using simulations and real datasets, assessing bias, mean squared error, and coverage probability.

Main Results:

• The CBRM demonstrated lower absolute mean bias and greater coverage probability compared to the BRM in most cases.
• CBRM estimates for sensitivity and specificity were generally closer to true values.
• For real datasets, CBRM produced estimates within the applicable region, unlike the BRM.

Conclusions:

• A constrained bivariate random effects model (CBRM) is more likely to produce plausible estimates for test sensitivity and specificity in specific practice settings.
• CBRM offers an improvement over unconstrained models for test accuracy meta-analysis when incorporating setting-specific data.
• The developed CBRM enhances the reliability of meta-analytic findings for diagnostic test evaluation.