NMADTA: Network Meta-Analysis of Multiple Diagnostic Tests

Xing Xing, Boyang Lu, Lifeng Lin, Qinshu Lian, James S. Hodges, Yong Chen, Haitao ChuOpenalex Unknown 2025 · DOI · OpenAlex

TL;DR

The NMADTA R package provides a robust Bayesian framework for performing network meta-analysis of multiple diagnostic tests. By implementing advanced hierarchical models, it allows researchers to simultaneously compare the accuracy of up to five different tests while accounting for complex missing data scenarios, including those where data is "Missing Not At Random" (MNAR).

What Problem Does This Paper Address?

In clinical research, evaluating a new diagnostic test usually involves comparing it against a "gold standard." However, clinicians often have multiple testing options available. Traditional meta-analysis typically focuses on pairwise comparisons, which can be limiting. Network Meta-Analysis (NMA) allows for the simultaneous comparison of multiple tests, even if they haven't been compared directly in a single head-to-head study.

The major hurdle in DTA (Diagnostic Test Accuracy) meta-analysis is missing data. Often, studies in a meta-analysis only report results for a subset of the tests being investigated. If the reason a test result is missing is related to the test's performance (e.g., a study didn't report results for a test that performed poorly), the data is "Missing Not At Random" (MNAR). Ignoring this leads to biased estimates. NMADTA provides the statistical machinery to tackle these biases head-on.

Methods

The package implements two primary Bayesian hierarchical frameworks:

  1. The Hierarchical Model (Ma et al., 2018): A model designed for NMA of multiple diagnostic tests that outputs credible intervals for standard accuracy measurements like sensitivity and specificity.
  2. The HSROC Model (Lian et al., 2019): A Hierarchical Summary Receiver Operating Characteristic model. This approach is particularly useful for visualizing the trade-off between sensitivity and specificity across different studies.

Both frameworks are extended to handle MNAR assumptions. Users can specify coefficients ($\gamma_0, \gamma_1$) that model the probability of a test result being missing based on its underlying sensitivity and specificity. The package utilizes Markov Chain Monte Carlo (MCMC) via the JAGS (Just Another Gibbs Sampler) library to estimate posterior distributions.

Datasets

The package includes a built-in dataset for demonstration and evaluation:

  • dat.kang: This dataset contains study-level results for multiple diagnostic tests used in colorectal cancer detection, originally extracted from Kang et al. (2013). It consists of 48 observations across 12 studies and includes specific indicators for missingness (delta0, delta1, delta2), making it an ideal benchmark for testing the package's MNAR capabilities.

Key Insights and Ideas

  • Simultaneous Comparison: Unlike traditional methods, NMADTA looks at the entire "network" of tests, leveraging indirect evidence to rank diagnostic tools more accurately.
  • Modeling the "Why" of Missing Data: By allowing users to define parameters for the missingness mechanism, the package enables sensitivity analyses. For example, a researcher can test how their results change if they assume poorer-performing tests were more likely to be left out of published studies.
  • Integrated Visualization: The package includes S3 methods for plotting that automatically generate Summary ROC (SROC) curves, posterior density plots, forest plots, and contour-enhanced plots. This bridges the gap between complex Bayesian output and interpretable clinical evidence.

Key Results

As a software implementation, the "results" are the package's outputs. NMADTA provides:

  • Posterior Estimates: Mean, median, and standard deviation for sensitivity, specificity, and prevalence.
  • Credible Intervals: 95% equal-tail credible intervals for all parameters.
  • Diagnostic Metrics: Positive and negative predictive values (PPV/NPV) and likelihood ratios (LRpos/LRneg).
  • Convergence Diagnostics: Integration with coda to provide potential scale reduction factors (Gelman-Rubin diagnostic) to ensure MCMC chains have converged.

Limitations and Future Work

  • Computational Intensity: Because the package relies on MCMC, large datasets or complex models with many iterations (the authors recommend 50,000 for real-world analysis) can be time-consuming to run.
  • Test Limit: The current implementation is optimized for the comparison of 1 to 5 diagnostic tests.
  • System Dependencies: Users must have JAGS 4.x.y installed on their system independently of R, which may be a slight hurdle for some users.
  • Future Directions: While not explicitly detailed in the manual, potential future iterations could include support for more than five tests or the integration of faster approximate Bayesian computation methods to reduce processing time.

Summary updated

Abstract

Implements HSROC (hierarchical summary receiver operating characteristic) model developed by Ma, Lian, Chu, Ibrahim, and Chen (2018) &lt;<a href="https://doi.org/10.1093%2Fbiostatistics%2Fkxx025" target="_top">doi:10.1093/biostatistics/kxx025</a>&gt; and hierarchical model developed by Lian, Hodges, and Chu (2019) &lt;<a href="https://doi.org/10.1080%2F01621459.2018.1476239" target="_top">doi:10.1080/01621459.2018.1476239</a>&gt; for performing meta-analysis for 1-5 diagnostic tests to simultaneously compare multiple tests within a missing data framework. This package evaluates the accuracy of multiple diagnostic tests and also gives graphical representation of the results.
BibTeX
@inproceedings{xing2025nmadta,
  author = {Xing Xing and Boyang Lu and Lifeng Lin and Qinshu Lian and James S. Hodges and Yong Chen and Haitao Chu},
  title = {{NMADTA: Network Meta-Analysis of Multiple Diagnostic Tests}},
  booktitle = {Openalex Unknown},
  year = {2025},
  doi = {10.32614/cran.package.nmadta},
}

Loading related content...

Figure