Bivariate random-effects models are commonly used to jointly synthesize sensitivity and specificity in diagnostic test accuracy (DTA) meta-analysis. However, existing approaches may be sensitive to distributional assumptions and may produce biased estimates of accuracy and between-study heterogeneity with sparse meta-analytic data. We proposed a novel bivariate structural equation modelling (BSEM) approach for DTA meta-analysis and evaluated its performance relative to the bivariate linear mixed model (BLMM) and bivariate generalized linear mixed model (BGLMM). We conducted extensive simulations under varying study, sample-size, accuracy, and heterogeneity conditions. We assessed model performance using bias, root mean squared error (RMSE), confidence interval width, empirical 95% coverage probability (CP), and convergence. We generated study-specific random effects from bivariate normal, t, and Laplace distributions to assess robustness to distributional misspecification. We also applied the models to two empirical DTA datasets. The BSEM generally produced smaller bias and RMSE and nominal-level CP for pooled sensitivity and specificity than the BLMM, while the BGLMM generally yielded results between the two approaches. The BLMM showed stable convergence but tended to underestimate between-study heterogeneity. We observed similar performance patterns under the three random-effects distributions. In the empirical applications, differences between model estimates were generally consistent with patterns observed in the simulations. Overall, the proposed BSEM provided a flexible alternative approach for jointly modelling sensitivity and specificity in DTA meta-analysis, with performance that was generally favourable relative to the BLMM and comparable to the BGLMM across the simulated and empirical settings considered.

翻译：暂无翻译