Bayesian bivariate meta-analysis of diagnostic test studies with interpretable priors

Area of Science:

• Statistics
• Biostatistics
• Medical Informatics

Background:

• Bivariate meta-analysis often involves few diagnostic studies, posing challenges for frequentist methods.
• Bayesian inference offers a solution using informative priors to stabilize analyses without overpowering data.
• Bayesian analysis can be computationally intensive, and prior selection for the covariance matrix is critical with sparse data.

Purpose of the Study:

• To explore the penalized complexity (PC) prior framework for specifying informative priors in Bayesian bivariate meta-analysis.
• To assess the performance of PC priors for variance and correlation parameters compared to existing methods.
• To evaluate the practical application of PC priors by reanalyzing a meta-analysis on bladder cancer diagnosis.

Main Methods:

• Utilized the integrated nested Laplace approximations (INLA) method for efficient Bayesian computation, avoiding sampling.
• Applied the PC prior framework to define priors for variance and correlation parameters in the bivariate structure.
• Conducted a simulation study to compare PC priors with commonly used priors.
• Reanalyzed a meta-analysis on the telomerase marker for bladder cancer diagnosis.

Main Results:

• The PC prior demonstrated benefits for variance parameters in Bayesian bivariate meta-analysis.
• Specifying PC priors for the correlation parameter yielded more precise estimates when aligned with true values.
• The reanalysis confirmed the practical utility and potential advantages of PC priors in real-world diagnostic meta-analyses.

Conclusions:

• PC priors offer an intuitive and effective approach for specifying informative priors in Bayesian bivariate meta-analysis.
• The PC prior framework enhances model interpretability and hyperparameter specification, particularly valuable in data-limited scenarios.
• This approach shows promise for improving the precision and reliability of diagnostic meta-analyses, as demonstrated in the bladder cancer example.