A pseudo-likelihood approach for multivariate meta-analysis of test accuracy studies with multiple thresholds

Area of Science:

• Biostatistics
• Medical Informatics
• Epidemiology

Background:

• Multivariate meta-analysis synthesizes test accuracy data across multiple thresholds, offering advantages over univariate approaches.
• Existing normal multivariate random-effects models for this purpose face challenges with estimating within-study correlations and computational convergence.

Purpose of the Study:

• To propose and evaluate an alternative pseudo-likelihood method for multivariate meta-analysis of diagnostic test accuracy.
• To address the limitations of current multivariate models, particularly regarding computational feasibility and correlation estimation.

Main Methods:

• A pseudo-likelihood approach is developed, assuming working independence between sensitivities and specificities at different thresholds within studies.
• This method bypasses the need for within-study correlation estimation and avoids convergence problems.
• The proposed method is compared to the multivariate normal random-effects model via simulation studies.

Main Results:

• Simulation studies demonstrate satisfactory performance of the pseudo-likelihood method.
• The proposed method shows significant improvements over the multivariate normal counterpart across various scenarios.
• The approach is successfully illustrated using data from a preeclampsia diagnostic test evaluation.

Conclusions:

• The pseudo-likelihood method offers a computationally feasible and effective alternative for multivariate meta-analysis of diagnostic test accuracy.
• It overcomes key limitations of existing normal multivariate models, enabling better synthesis of evidence.
• This method has practical implications for evaluating diagnostic tests, as shown in the preeclampsia example.