The paper introduces NMADTA, an R package designed to evaluate several diagnostic tests together when studies contain incomplete evidence or missing results from the gold-standard reference test. It implements Bayesian hierarchical and Bayesian HSROC models and supports simultaneous analysis of up to five tests, with or without a gold standard.
In practical terms, the package is built for a network meta-analysis: an evidence-synthesis approach that can bring different test comparisons into one analysis. Its regular models assume unavailable test outcomes are missing at random and treat candidate tests as conditionally independent once disease status and model parameters are taken into account.
An illustrative trade-off
In an illustrative hierarchical fit, ultrasonography had a median sensitivity of 0.94, with a 95% credible interval of 0.83 to 1.00. Its median specificity was 0.77, with a 95% credible interval of 0.45 to 0.97. Sensitivity refers to detecting disease when it is present; specificity refers to correctly identifying cases without it. The analysis described ultrasonography as more sensitive but less specific than D-dimer.
A related SROC visualization, which summarizes the trade-off between sensitivity and specificity, described ultrasonography as having a steeper curve toward the upper left and better average discrimination.
Because the analysis is illustrative, the numbers show how NMADTA handles a sensitivity–specificity trade-off rather than settle a clinical question. The supplied text also does not give D-dimer point estimates alongside the ultrasound estimates.
Making missing results explicit
To calculate its estimates, the package uses Markov chain Monte Carlo, or MCMC, a computational sampling method, through JAGS and the rjags interface.
NMADTA also includes an MNAR sensitivity analysis for cases in which the chance that a result is missing may depend on unobserved test accuracy. It links missingness to latent sensitivity and specificity; more-negative coefficients represent stronger dependence on lower study-specific accuracy.
The assumptions set the limits
Sparse networks and frequent missing gold standards can leave variance–covariance and correlation components weakly identified. That increases dependence on the chosen priors and can make the MCMC algorithm mix poorly, meaning it may explore plausible estimates less effectively.
The MNAR mechanism has a separate identification problem: observed data alone generally cannot determine how missingness arose. The authors frame MNAR analyses as structured sensitivity analyses across plausible mechanisms.
The discussion says the current implementation fixes prior distributional forms, uses parametric models and relies on trace plots and PSRFs—potential scale reduction factors—as convergence checks.
The example data description is also internally inconsistent. One passage reports 1212 studies, while a later description refers to 12 study IDs; it also lists 44 D-dimer–venography comparisons, 33 ultrasonography–venography comparisons and 55 D-dimer–ultrasonography comparisons, so the supplied figures do not resolve to one coherent total.
NMADTA’s contribution, on the evidence presented, is mainly practical: it packages a complex Bayesian workflow for incomplete diagnostic-test networks. Its estimates will still depend on how well the network is identified and which assumptions are used for missing data.
Paper data and sources
Original title: NMADTA: An R package for network meta-analysis of multiple diagnostic tests
Authors: Xing X, Lu B, Lin L et al.
Journal/Repository: Research Synthesis Methods
Status: Peer-reviewed
First online: 2026-08-19
DOI: Not available
Original paper · Full text