Robust bootstrap methods for meta-analyses of diagnostic test accuracy studies

Abstract

Meta-analysis of diagnostic test accuracy (DTA) commonly relies on the bivariate linear mixed model (BLMM) and the bivariate generalized linear mixed model (BGLMM) to jointly synthesize diagnostic test sensitivity (Se) and specificity (Sp). However, reliance on asymptotic theory and random-effects distributional assumptions may affect finite-sample inference, particularly in small meta-analyses or when the random effects are non-normal. We proposed three bootstrap-based approaches: a nonparametric bootstrap (NPB), a parametric bootstrap (PB), and a bias-corrected parametric bootstrap (BCPB), with the latter two based on a fitted BGLMM. NPB model parameters were estimated using inverse-variance-covariance weighted least squares. Extensive simulations under bivariate normal, Student-t, and Laplace random-effects distributions evaluated finite-sample performance and robustness to departures from normality. Results were qualitatively similar across the three distributions. The bootstrap approaches generally outperformed BLMM in terms of bias, root mean squared error (RMSE), and coverage probability, particularly when diagnostic accuracy is high and the number of studies is small. The BCPB demonstrated the strongest finite-sample performance among the bootstrap approaches, frequently achieving the lowest bias and RMSE for pooled sensitivity and specificity while maintaining coverage probabilities close to the nominal level. The BGLMM also performed well across most settings. The NPB approach generally improved upon the BLMM while providing a flexible alternative without parametric random-effects assumptions. Real-data applications further illustrated the practical utility of the proposed approaches. Overall, the bootstrap methods provide useful alternatives for DTA meta-analysis, particularly when the number of studies is small or random-effects distributional assumptions are uncertain.

Structural equation modelling--based meta-analytic model for diagnostic test accuracy studies
Zelalem F. Negeri
arXiv20 Oct 2026PreprintarXiv licenceFull textStatisticsarXiv
Analytical and Bootstrap Confidence Intervals for Double Machine Learning: Simulation Studies and an Application to Rural-Urban Differences in Obesity Prevalence
Haozheng Xu, Siyuan Ma, Qingyan Xiang
arXiv25 Aug 2026PreprintarXiv licenceFull textStatisticsarXiv
Small-study effects on the hierarchical summary ROC curve: latent accuracy and threshold trends in meta-analysis of diagnostic test accuracy
Yuta Nonomiya, Kazuki Nishida
arXiv17 Oct 2026PreprintarXiv licenceFull textStatisticsarXiv
Modelling multiple disease risk factors for microsimulation studies: a review of methods
Oliver Church, Christopher Jackson
arXiv14 Sep 2026PreprintCC BY 4.0Full textStatisticsarXiv
Improving the accuracy of the momentum-exchange method in lattice Boltzmann methods
Abbas Fakhari
arXiv4 Nov 2026PreprintarXiv licenceFull textNumerical analysis & scientific computing Computational physicsarXiv

Cite

BibTeX RIS

BibTeX

@misc{bedada2026robust,
  title = {{Robust bootstrap methods for meta-analyses of diagnostic test accuracy studies}},
  author = {Diribsa T. Bedada and Olana A. Dabi and Isaac Fordjour and Zelalem F. Negeri},
  howpublished = {arXiv},
  year = {2026},
  month = oct,
  doi = {10.48550/arxiv.2610.33270},
  eprint = {2610.33270},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.33270},
}

RIS

TY  - UNPB
TI  - Robust bootstrap methods for meta-analyses of diagnostic test accuracy studies
AU  - Diribsa T. Bedada
AU  - Olana A. Dabi
AU  - Isaac Fordjour
AU  - Zelalem F. Negeri
PY  - 2026
DA  - 2026/10/23
PB  - arXiv
DO  - 10.48550/arxiv.2610.33270
UR  - https://arxiv.org/abs/2610.33270
AB  - Meta-analysis of diagnostic test accuracy (DTA) commonly relies on the bivariate linear mixed model (BLMM) and the bivariate generalized linear mixed model (BGLMM) to jointly synthesize diagnostic test sensitivity (Se) and specificity (Sp). However, reliance on asymptotic theory and random-effects distributional assumptions may affect finite-sample inference, particularly in small meta-analyses or when the random effects are non-normal. We proposed three bootstrap-based approaches: a nonparametric bootstrap (NPB), a parametric bootstrap (PB), and a bias-corrected parametric bootstrap (BCPB), with the latter two based on a fitted BGLMM. NPB model parameters were estimated using inverse-variance-covariance weighted least squares. Extensive simulations under bivariate normal, Student-t, and Laplace random-effects distributions evaluated finite-sample performance and robustness to departures from normality. Results were qualitatively similar across the three distributions. The bootstrap approaches generally outperformed BLMM in terms of bias, root mean squared error (RMSE), and coverage probability, particularly when diagnostic accuracy is high and the number of studies is small. The BCPB demonstrated the strongest finite-sample performance among the bootstrap approaches, frequently achieving the lowest bias and RMSE for pooled sensitivity and specificity while maintaining coverage probabilities close to the nominal level. The BGLMM also performed well across most settings. The NPB approach generally improved upon the BLMM while providing a flexible alternative without parametric random-effects assumptions. Real-data applications further illustrated the practical utility of the proposed approaches. Overall, the bootstrap methods provide useful alternatives for DTA meta-analysis, particularly when the number of studies is small or random-effects distributional assumptions are uncertain.
ER  -

Abstract

Related

Cite