Innovative Multivariate Meta-Analyses Methods For Diagnostic Tests And Multiple Treatments

Abstract

Systematic reviews and meta-analyses are critical tools for synthesizing evidence from multiple studies and supporting healthcare decision-making on diagnostic tests and multiple treatments. Assessing a diagnostic test often lacks a gold standard, i.e., a perfect reference test. Without a gold standard, a pragmatic approach is to postulate a ``reference standard'' from two index tests, however, this approach may overestimate the sensitivity due to potential misclassification resulting from the imperfect reference standard. In the first paper, we illustrate the impact of this issue and propose a two-step latent class meta-analysis assuming perfect specificities which can more accurately evaluate sensitivities by incorporating the double-negative results. We apply the method to evaluate the diagnosis accuracy of SARS-CoV-2 infection using saliva sampling or nasopharyngeal swabs. Through simulations, our method shows improved performance over the pragmatic reference standard approach for varied prevalence and between-study heterogeneity. In the second paper, we further relax the assumptions on perfect specificity and develop Bayesian hierarchical models that provide simultaneous estimates of sensitivities, specificities, and disease prevalence with adjustments for study-level covariates. We demonstrate the model performance using the same published dataset on COVID tests as in the first paper and extensive simulation studies. Compared with the pragmatic reference standard approach, the proposed Bayesian method provides a more accurate evaluation of prevalence, specificity, and sensitivity in a meta-analytic framework. The third paper extends the pairwise meta-analysis to network meta-analysis (NMA), enabling simultaneous assessment of multiple treatments. In arm-based NMA (AB-NMA), the sparsity of direct comparisons among multiple treatments makes accurate estimation of the correlation between treatments challenging. To address this challenge and complement the analysis, we develop a sensitivity analysis tool: a correlation tipping point analysis for AB-NMA assessing the robustness of conclusions about relative treatment effects across a feasible range of correlation parameter values. Our application on multiple NMA datasets highlights its supplemental value for robust decision-making. Thus, we recommend using a correlation tipping point analysis as a standard component of AB-NMA.