Beta-binomial analysis of variance model for network meta-analysis of diagnostic test accuracy data.

Area of Science:

• Biostatistics
• Medical Informatics
• Diagnostic Test Evaluation

Background:

• Generalized linear mixed models are popular for network meta-analysis of diagnostic tests.
• Current models often use normal distributions, which are unnatural and inconvenient for modeling sensitivity and specificity.
• This can lead to suboptimal parameter estimates in diagnostic test accuracy studies.

Purpose of the Study:

• To develop an improved meta-analytic model for diagnostic tests using a bivariate beta distribution.
• To provide direct and interpretable estimates of global sensitivities and specificities.
• To account for the correlation between sensitivity and specificity and overdispersion in repeated measures.

Main Methods:

• Developed a meta-analytic model based on the bivariate beta distribution.
• Applied the model to combine direct and indirect evidence from diagnostic studies.
• Utilized regression with the beta distribution to model probabilities on their natural scale.

Main Results:

• The bivariate beta distribution model provides improved and direct estimates for global sensitivities and specificities.
• The model naturally accounts for the intrinsic correlation between sensitivity and specificity.
• It also addresses overdispersion in sparse data common in meta-analysis, offering flexibility for asymmetric distributions.

Conclusions:

• The bivariate beta distribution offers a more natural and computationally convenient approach for diagnostic test meta-analysis.
• This method yields more meaningful parameter interpretations without further integration.
• It enhances the accuracy and interpretability of network meta-analysis for diagnostic tests.