The metaUI package allows users to quickly create powerful and customisable Shiny web apps that allow readers to explore a meta-analytic dataset in depth. They can explore the impact of different analytical choices, see results for subgroups of particular interest, and even modify and update the underlying dataset.
The motivation for the package came from the fact that meta-analyses are based on rich datasets that can be analyzed in numerous ways, and it is unlikely that authors and readers will always agree on the “best ways” to analyze the data. Whether it comes to the choice of model (e.g., random versus fixed effects), the methods for assessing or adjusting for publication bias (e.g., z-curve, p-curve, PET/PEESE etc), or the moderators to be considered, disagreements are likely to arise. This can lead to the inclusion of lengthy robustness checks and alternative analyses that are time-consuming and difficult to digest. metaUI, as an R package that supports researchers in creating interactive Shiny web apps can help - these apps allow readers (and reviewers) to explore meta-analytic datasets in a variety of different ways. Apart from allowing readers (and reviewers) to assess the robustness and trustworthiness of results more comprehensively, metaUI apps allow users to assess the results that are most relevant to them, such as by filtering the dataset to focus on a specific group of participants, region, outcome variable, or research method. With the opportunity for users to download the dataset used and to upload alternatives, it will also facilitate the updating of meta-analyses.
The idea came from researchers who have created similar web apps for their meta-analyses that have been well received, yet they require substantial time investment and advanced coding skills to create. With metaUI, researchers can get a working app very swiftly – while they still have the flexibility to tailor the display in line with their interests and requirements. Scientific and licensing limits are documented below; feedback and contributions are welcome.
Installation
You can install the development version of metaUI from GitHub with:
if (!require(remotes)) install.packages("remotes")
remotes::install_github("LukasWallrich/metaUI", dependencies = TRUE)dependencies = TRUE also installs the packages that generated apps use (listed under Suggests); generate_shiny() names any that are missing.
Getting started
The best way to get started in RStudio is with the metaUI template - so in RStudio, go to File -> New File -> R Markdown then select From Template and Create a metaUI Shiny app. To see the template in the most readable format, select Visual just above the file that has opened to enter the Visual Editor. Then you can follow the step-by-step instructions.
Outside R Studio, you can run file.edit(system.file("rmarkdown/templates/create-a-metaui-shiny-app/skeleton/skeleton.Rmd", package = "metaUI")) to open the template file and then take it from there.
The Getting started vignette provides a step-by-step example … or you may want to watch this video tutorial for a step-by-step walk-through.
Building and launching apps
Apps can be built in R with generate_shiny() or headlessly from a versioned JSON configuration with build_app(); either way, the generated R/Shiny files stay editable. Schema version 2 adds computation and Bayesian authoring; version 1 continues to accept the original configuration fields. See reproduction and scientific contract, the tiny synthetic config, and its JSON schema. Build and launch are separate:
First copy tiny.json and tiny.csv from inst/examples into a fresh, empty folder, then run:
generate_shiny() also supports this workflow: supply save_to_folder and launch_app = FALSE, then launch separately with shiny::runApp("<folder>"), passing the folder you saved to (a bare shiny::runApp() runs the current working directory). With launch_app = TRUE (the default when no folder is supplied), interactive printing of the returned app launches Shiny and blocks the console until you stop the app. Assigning the app object to a variable postpones launch until you print it.
Generation refuses a nonempty folder by default. p/N are optional for the primary models; directional models require explicit direction. COR/ZCOR require explicit variance scales. Read the generated validation report before interpreting results. The introductory Dannheim data use signed Hedges g, not correlations. No public deployment is automatic.
Primary model
Authors may name one primary model (primary_model in generate_shiny() or the JSON config). Apps show it first and label the other rows as explorations; they never substitute another model when it is unsupported.
Examples
Copy inst/examples/dannheim to a fresh working directory and run Rscript --vanilla build.R. Launch separately with shiny::runApp("mental-health-app"). The tiny CC-BY mental-health pool retains 12 exact benchmark member IDs and original signed Hedges g with CI-derived native SE. The source-reference REML/Hartung–Knapp model is checked against metafor; other models/diagnostics are exploratory. Read the example’s README/provenance before interpreting results. No benchmark corpus is deployed.
The synthetic correlation example demonstrates raw-r variance conversion and checks the fitted summary against an independent metafor reference.
Fit reuse and performance
Generated apps reuse compatible fitted objects and retain three recent exact fit selections per reader session; cache_hit discloses reuse. Custom models default to no cache. Pass options = list(fit_cache_entries = 0) to generate_shiny() (or set "options": {"fit_cache_entries": 0} in the build JSON) to disable caching; keep it disabled for random models or code depending on external state. Uploads are validated, stored in session state, and clear prior cached results.
Performance measurements distinguish faster p-curve calculations/fit reuse from cache hits and first-time fitting costs.
Downloads, uploads and reproduction code
Downloads preserve the full current input dataset and saved filter selections; the summary sheet describes the selected rows. An empty selection stays empty on re-upload and reports no eligible rows, while the original inputs remain available.
After an analysis, readers can download an R script or a Quarto file that reproduces the applied selection (filters, missing-value choices, effect-size computation and equivalence bound) and the model rows on the Summary tab. The code reads the workbook from the same analysis, so it also works for uploaded data, and it runs without metaUI. It copies the app’s models.R, including author-defined models and their helper functions, and the fitting functions that the app calls, so it follows the app’s own code path. If models.R has been edited since the app started, the code contains the loaded model table instead and says that any helper code is missing. It also records the package versions the app used, with a commented-out groundhog call that installs the CRAN versions current on the latest release date among them, and ends by comparing every row with the values the app reported. It does not reproduce the plots or diagnostic tabs. Author-defined models that use random numbers or external state, and other package versions, can give different results; the comparison shows where they do. See the validation note.
Shareable selection links
After analysis, the selection link and address bar describe the applied selection on this exact built dataset. Links validate filters, missing-value choices, computation and equivalence bound before restoring and analysing. Invalid links leave a notice and require manual analysis. Uploaded data require the workbook, which includes all computation inputs; their session-specific data are not encoded in URLs.
Practical equivalence and forest exports
Summary includes an interval assessment for a reader-chosen symmetric smallest effect size of interest, on the displayed SMD or r scale. No bound is assumed. The strict containment criterion uses the reported intervals without refitting; model assumptions and small-sample limitations are shown beside the assessment. Downloads record the bound and classification in an equivalence sheet, and uploading the workbook restores the bound, as selection links do.
Forest plots compare observed-effect intervals and available multilevel/RVE summary intervals. PDF, PNG and CSV downloads use the same selected rows; the CSV retains both fitting and displayed scales. The configured row limit also applies to exports.
Alternative effect-size computations
Authors can supply up to five alternative computations in prepare_data(alternatives = ...). Each named entry declares es_field, variance or se, es_type, variance_scale for correlations, and a justification. primary_label names the original computation. All alternatives use the same effects and fitting scale. Readers apply a computation with Analyze data; the comparison tab fits the configured multilevel/RVE models for the same selection. The descriptive z filter and source p-values remain anchored to the primary computation.
Bayesian analysis
Enable study-level Bayesian analysis with generate_shiny(..., bayesian = list(enabled = TRUE)), after installing the optional bayesmeta package. The mean prior is flat; the heterogeneity prior is half-normal with scale 0.5 for SMD or 0.25 for Fisher z, configurable with tau_scale. Central 95% credible intervals and posterior medians are reported, and the sensitivity tab compares half and double that prior scale. This uses aggregated studies and the declared within-study correlation. Authors who edit aggregation_method in the generated labels_and_options.R to "first" switch it to the first-effect rule; the build records the priors and the default aggregation. The default cap of 20 studies protects interactive use: first-fit benchmarks took about 20 seconds at 20 studies and 97 seconds at 50 on the review machine. Authors can set max_studies from 2 to 50, with a build warning above 20. Each Analyze blocks the Shiny server process while fitting. Prior sensitivity runs when its tab or download is requested and can require two extra fits. See Röver et al. for the prior rationale.
Study details
Readers can open a row in the list of effect sizes to see the study’s reference and, where permitted, its abstract. Authors supply the details as generate_shiny(study_info = ...) or as a study_info CSV in a schema version 2 configuration: one row per study, with a study column matching the study IDs. To fill this from DOIs, call fetch_study_metadata() while authoring. It queries the free Crossref REST API once per DOI, asks OpenAlex for licences and abstracts that Crossref lacks, and can cache results in a CSV. The generated app stores the result and never contacts either service. Missing, malformed or unknown DOIs are listed and left without details; nothing is matched by title.
data <- metaUI::prepare_data("effects.csv", "study", "d", variance = "vi", url = "doi")
info <- metaUI::fetch_study_metadata(data, cache = "study-metadata.csv")
metaUI::generate_shiny(data, "My meta-analysis", study_info = info,
save_to_folder = "my-app", launch_app = FALSE)In a schema version 2 configuration for build_app(), study_info can instead run the lookup on the mapped url column while building and save the result for review:
Only cache is required. DOIs already found in the cache are not queried again, so rebuilds (metaUI::build_app("config.json", overwrite = TRUE)) reuse the reviewed CSV.
Single-DOI lookups are free on both services and need no key. Two optional environment variables, best set in ~/.Renviron so they stay out of scripts:
METAUI_MAILTO=you@example.org # contact address Crossref asks for (its "polite" pool)
OPENALEX_API_KEY=your-key # free key: openalex.org, Settings > API key; raises the rate limitCrossref states that abstracts remain under publisher or author copyright, unlike the rest of its metadata. By default, the function therefore keeps an abstract only when the article is published under a Creative Commons licence (CC BY, with or without SA, ND or NC terms), CC0 or the Public Domain Mark, as recorded by Crossref or OpenAlex, and the app shows that licence beside it. All of these allow verbatim copies of the work in part. NonCommercial licences cover academic, non-commercial apps; for an app used commercially, set abstracts = "exclude_nc". abstracts = "all" keeps every abstract found; authors who choose it are responsible for the right to republish them. Abstracts that authors add to the CSV are likewise their responsibility. This rule is a conservative default, not legal advice.
The reference is shown in APA style with a button that copies it.
Author design settings
Run metaUI::design_app() to choose colours, font stacks, text size, spacing and panel corners. The helper previews an illustrative analysis and provides R and JSON downloads. Low-contrast choices keep the previous valid preview and disable export until corrected.
Copy its R settings, then pass the resulting list to app creation:
design <- metaUI::app_design(accent = "#69429A", font = "georgia",
spacing = "compact", radius = 4)
metaUI::generate_shiny(dataset, "My meta-analysis", design = design,
save_to_folder = "my-app", launch_app = FALSE)For build_app(), put the downloaded JSON object in the configuration’s design field. Existing configurations and apps without a design retain their styling (design = NULL). The helper’s default app_design() is an explicit new design; Reset choices restores the helper’s initial settings, not legacy CSS. Apps with a design contain an editable www/app-design.css and a design.json record. Editing the record does not restyle an app: rebuild with the settings, or edit the CSS. These apps do not need the helper or metaUI installed to run. The design changes interface styling and the summary-comparison/forest plot colours, not models or selections. Fonts come from each reader’s device, with explicit fallbacks; no font files from the author’s machine are needed for deployment. System sans uses the device interface font. Named fonts such as Georgia may be substituted on other devices. These choices do not require font downloads, though the underlying Bootstrap theme may still request external stylesheets. Other analytical plots retain their plotting packages’ styles. The contrast check covers text and accent against page, panel and derived tinted backgrounds, not an accessibility audit of author-supplied HTML or arbitrary Bootstrap themes.
Publishing without a server
Generate an editable app and a static website in one step:
app_data <- metaUI::prepare_data("effects.csv", "study", "d", variance = "vi")
metaUI::generate_shiny(app_data, "My meta-analysis",
save_to_folder = "my-app", static_folder = "my-site")Omit save_to_folder to use temporary source files. With JSON configurations, add "static_folder": "my-site" alongside the editable app’s "output" folder. After editing the generated app files, re-export with metaUI::export_static("my-app", "my-site", overwrite = TRUE); regenerating would overwrite your edits. Static export requires the suggested shinylive package.
Host the site over HTTPS (for example, on GitHub Pages), or serve it from localhost to preview it. Opening index.html as a file does not work. The full dataset and code are sent to every visitor, so publish only material that may be shared. Precomputed results appear at once; controls work after roughly 20 seconds on a typical laptop, depending on the connection and device.
Sources
Much of the package code was based on two previous interactive meta-analyses apps:
- Röseler, L., Körner, R., & Schütz, A. (2021). Dynamic Meta-Analysis. Retrieved from https://osf.io/ns65r/, Shiny app accessible here
- Röseler, L., Weber, L., Helgerth, K. A. C., Stich, E., Günther, M., Tegethoff, P., Wagner, F. S., Ambrus, E., Antunovic, M., Barrera, F., Halali, E., Ioannidis, K., McKay, R., Milstein, N., Molden, D. C., Papenmeier, F., Rinn, R., Schreiter, M. L., Zimdahl, M., Allen, E., Bahník, S., Baumeister, R. F., Bermeitinger, C., Bickenbach, S. L. C., Blank, P. A., Blower, F. B. N., Bögler, H. L., Boo, F. L., Boruchowicz, C., Bühler, R. L., Burgmer, P., Cheek, N., N., Dohle, S., Dorsch, L., Dück, M. S., Fels, S.-A., Fischer, A. L., Frech, M.-L., Freira, L., Friedinger, K., Genschow, O., Harris, A., Hartig, B., Häusser, J. A., Hedgebeth, M., Henkel, M., Horvath, D., Hügel, J. C., Igna, E. L. E., Imhoff, R., Intelmann, P., Karg, A. H., Klamar, A., Klein, C., Klusmann, B., Knappe, E., Köppel, L.-M., Koßmann, L., Kraft, P., Kroworsch, M. K., Krueger, S. M., Kühling, S., Lagator, S., Lammers, J., Loschelder, D. D., Navajas, J., Norem, J., K., Novak, J. Onuki, Y., Page, E., Panse, F., Pavlovic, Z., Pearton, J., Rebholz, T. R., Rodgers, S., Röseler, J. J., Rostekova, A., Roßmaier, K. V., Sartorio, M., Scheelje, L., Schindler, S., Schreiner, N. B., Seida, C., Shanks, D. R., Siems, M.-C., Stitz, M., Starkulla, M., Stäglich, M., Thies, K., Thum, E., Undorf, M., Unger, B. D., Urlichich, D., Vadillo, M. A., Wackershauser-Sablotny, V., Wessel, I., Wolf, H., Zhou, A., & Schütz, A. (2022). OpAQ: Open Anchoring Quest, Version 1.1.48.95. https://dx.doi.org/10.17605/OSF.IO/YGNVB, Shiny app accessible here
The introductory sample is the CC-BY-4.0 mental-health pool from Dannheim et al. (2025), doi:10.5271/sjweh.4219. Its 12 signed Hedges-g effects, CI-derived SEs, membership and changes are attributed in inst/examples/dannheim. Existing Aksayli/Coles data retain their verified source licences and attribution in inst/COPYRIGHTS.
Licence audit records the sources and licences of bundled code and data.
Related projects
- Allbritton et al.’s A General Tool for Living Meta-Analysis (source, doi:10.5281/zenodo.21938551) is the closest related tool. Analysts upload their data or use curated files, select studies by criteria generated from the data’s columns, and fit frequentist and Bayesian models; release 1.5.1 adds multilevel models and meta-regression. The app downloads R code that reproduces the displayed analyses, and a copy with a saved selection file reopens with, for example, an article’s selections. Its model menu and interface are those of the shared app, whereas metaUI generates a separate app whose models and explanations each author edits. The tool is described in:
Allbritton, D., Gómez, P., Angele, B., Vasilev, M., & Perea, M. (2024). Breathing Life Into Meta-Analytic Methods. Journal of Cognition, 7(1), 61. doi:10.5334/joc.389
Other related tools include Metapsy (analysis workflow inspiration), the Cooperation Databank (dataset exploration), PsychOpen CAMA (cumulative meta-analysis), and the taVNS HRV app (a Bayesian living meta-analysis). These are references, not endorsements or sources of bundled code. See the issue follow-up decisions for the assessed feature overlap and remaining work.
Compared with these projects, metaUI’s intended contribution is an author-owned app with editable R source. See the bounded publication-readiness assessment.