Skip to contents

metaROC

This R-package enables meta-analysis of full diagnostic test accuracy studies and ROC-curves using various techniques.

Features

  • simulate data from the following models:

    • logit LMM as proposed in Steinhauser et al. (2016)
    • AFT model with bivariate random effect as proposed in Hoyer et al. (2018) and Hoyer and Kuss (2020). Currently available parametric assumptions for test values: Weibull, log-logistic, log-normal, generalized F
    • discrete GLMMs with categorical variable threshold using either the cloglog- or the logit-link as proposed in Stoye et al. (2024)
    • survival copula models with different marginal distributions. Currently available copulas: Clayton copula, asymmetric Joe copula. Currently available marginals: Weibull-binomial, Weibull-normal, loglogistic-binomial, loglogistic-normal, lognormal-binomial, lognormal-normal

  • estimate the following models to data from several DTA studies reporting results for multiple diagnostic thresholds per study:

    • logit LMM (Steinhauser et al., 2016) using a link to the package diagmeta
    • non-parametric SROC model (Martínez-Camblor, 2017) using a link to the package nsROC
    • logit GLMM (Hoyer & Kuss, 2018)
    • AFT model with bivariate random effect (Hoyer et al., 2018, Hoyer and Kuss, 2020). Currently available parametric assumptions for test values: Weibull, log-logistic, log-normal, generalized F
    • Bayesian multinomial model (Jones et al., 2019)
    • semi-parametric global rank-based model (Frömke et al., 2022) using a link to the package diagacc
    • discrete GLMMs with categorical variable threshold using either the cloglog- or the logit-link (Stoye et al., 2024)
    • survival copula models with different marginal distributions. Currently available copulas: Clayton copula, asymmetric Joe copula. Currently available marginals: Weibull-binomial, Weibull-normal, loglogistic-binomial, loglogistic-normal, lognormal-binomial, lognormal-normal

  • estimate the following models to data from several DTA studies reporting results for a single diagnostic threshold per study:

  • estimate the following models to a mix of aggregated (single or multiple thresholds per study) and individual participant data (or exclusively to individual participant data):

    • AFT model with bivariate random effect. Currently available parametric assumptions for test values: Weibull, log-logistic, log-normal

  • evaluate estimated models based on information on the true data-generating process (for each model individually or aggregated for a number of simulation iterations)

Installation

You can install this package branch (e.g.) using the following code in your R console:

pak::pak("git::https://gitlab.ub.uni-bielefeld.de/stoyef/metaROC")

or

remotes::install_git("https://gitlab.ub.uni-bielefeld.de/stoyef/metaROC")

Usage

All main functionalities are implemented in the function metaROC. See the examples in the function documentation or here fore an introduction for an introduction. See also here for descriptions and tutorials for all implemented models.

I am confused, which method should I use after all?

Short answer: This depends on your aim, the type of data in your application, and many more factors. Long answer: We have evaluated eleven of the methods included in the package side-by-side in a simulation study. Our results may aid researchers in informed model selection.

Code for publications relying on metaROC