Skip to contents

Changelog

Source: NEWS.md

metaUI (development version)

  • Author-enabled p-curve effect estimates now report a failed fit, with the reason, when the estimate falls below zero. Previously such an estimate was reported with a sign opposite to the declared direction.
  • Publish apps without a Shiny server. generate_shiny(static_folder = ...) and the JSON static_folder field write a static website that runs R in the reader’s browser (shinylive/webR); export_static() re-exports an app after hand edits in the editable app folder. With overwrite = TRUE, the existing static site folder is deleted entirely and rewritten. Readers see the authors’ full-data results at once while R loads. Selection links, downloads and the Summary plot work in the static site, and include_sources = TRUE bundles the source code of copyleft components for archival deposits. Requires the suggested shinylive package. Every visitor downloads the full dataset and app code, so publish only shareable material.
  • Generated apps work on phones, on a Shiny server and as static sites. The filters collapse behind a “Show filters” button so that results appear near the top of the page; tables with four or more columns become one labelled card per row instead of scrolling sideways; wide plots scale to fit, and the forest plot wraps study labels. Static sites also show the precomputed preview at phone width, including when filters use selection boxes, which previously prevented the preview from being computed.
  • prepare_data() decodes HTML escapes such as & in text and factor columns, which apps previously showed verbatim (for example, in study labels), and reports the columns changed. Study details are matched to the decoded study IDs.
  • Forest plots taller than about a screen repeat the effect-size axis at the top.
  • Generated apps use metafor for trim-and-fill, the funnel plot, Egger’s test and the Dannheim source-reference fit. Results are unchanged. The meta package is needed only when author model code uses it.
  • Generated apps can download an R script or Quarto file that reproduces the applied selection (filters, missing-value choices, effect-size computation and equivalence bound) and its model rows from the workbook of the same analysis, including uploaded data. The code copies the app’s models.R (or, if the file no longer matches the loaded models, their table) and the fitting functions it uses, runs without metaUI, records the package versions the app used (with an optional, commented-out groundhog call) and compares its results with the app’s. Plots and diagnostic tabs are not reproduced.
  • Uploading a workbook now restores its equivalence bound, as selection links already did. Workbooks from empty selections also record the bound.
  • metaUI no longer depends on forcats, and generated apps no longer need psych; base R replaces both. The R Markdown skeleton’s example recode now treats -999 as missing (shown as (Missing)) instead of a level named “NA”.

metaUI 0.2.0

This release consolidates all changes since 0.1.1, including the unreleased 0.1.2 development versions.

Breaking changes

  • Packages used only by generated apps (DT, ggplot2, meta, plotly, psych, puniform, readxl, robumeta, shinycssloaders, shinyjs, weightr, writexl and zcurve) moved from Imports to Suggests. generate_shiny() names any that are missing, with the command to install them.
  • generate_shiny() and build_app() refuse a nonempty output folder unless overwrite = TRUE, so hand edits to a generated app are not overwritten by accident. overwrite = TRUE replaces all generated files, including the editable www/metaui.css and www/metaui.js; build into a fresh folder when hand edits must be kept.
  • Generated apps check their dependencies at startup but no longer install packages; dependencies.csv records the versions used for the build.
  • Correlations must declare their variance scale, and directional bias analyses an explicit direction. COR is fitted via Fisher’s z with a delta-method variance, ZCOR is not transformed twice, and summaries are back-transformed once.
  • Multilevel models nest effects within genuine per-effect IDs; display labels may now repeat. Summed sample sizes are no longer inferred.
  • Generated apps no longer use shinyBS or waffle: filter popups became inline help and the waffle chart became a categorical count bar chart.

New features

  • Study details in the list of effect sizes. generate_shiny(study_info = ...) adds a disclosure button to each study; it opens a row with an APA-style reference, a button that copies it, the DOI link and, where permitted, the abstract. fetch_study_metadata() looks up DOIs in Crossref while authoring, uses OpenAlex for licences and abstracts that Crossref lacks, reports DOIs it could not match and can cache results. Abstracts are kept when the article is published under a Creative Commons licence (including ND and NC terms), CC0 or the Public Domain Mark; abstracts = "exclude_nc" drops NC licences for commercial apps, and abstracts = "all" keeps every abstract found. In schema version 2 configurations, study_info is a CSV path or a lookup object ({"cache": "study-metadata.csv"}) that build_app() runs into that cache. Generated apps never query external services.

  • The bundled Coles et al. (2019) facial-feedback data (inst/extdata/coles2019.csv) include the source’s outcome-type column paq (affective judgement versus emotional experience), so apps can filter by it.

  • build_app() creates apps headlessly from a JSON configuration with a versioned schema (version 2 adds computations, Bayesian analysis and design; version 1 behaviour is retained).

  • design_app() previews colours, fonts, type size, spacing and corners and exports R/JSON settings; app_design() validates settings and contrast. Pass them via generate_shiny(design = ...) or the JSON design field. Apps built without a design keep the default appearance.

  • Authors can declare one primary_model. Apps show it first as the authors’ primary analysis, label other rows as explorations, and report why it is unavailable instead of substituting another model.

  • Author-defined effect-size computations, with explicit scale and variance contracts, consistent row selection, comparisons, provenance and workbook restoration.

  • Opt-in Bayesian normal-normal analysis (requires the suggested bayesmeta package): posterior medians, central credible intervals, disclosed priors and heterogeneity-prior sensitivity.

  • Readers can share applied selections through build-specific, validated URLs. Uploaded datasets remain workbook-only.

  • Practical-equivalence assessment of intervals against a reader-chosen symmetric bound (no default). Changing the bound does not refit models; downloads record the assessment and bound.

  • Forest plots show observed effects on the displayed scale, compare available multilevel/RVE summary intervals, and export to PDF, PNG and CSV.

  • generate_shiny() gained options (Shiny theme, forest-plot row limit, checkbox/selection-list threshold); prepare_data() gained arrange_filters. Apps can be created without any filters or moderators.

Generated app interface

  • The interface is reorganised around selecting evidence, running models and reading results. A status bar leads with selected and excluded counts, filter changes not yet analysed are flagged until Analyze data is clicked, and result tabs say when no analysis has been run. Routine instructions and repeated explanations were cut or moved into disclosures; scientific warnings, provenance and model distinctions remain visible.
  • Models that were not estimated are listed with their reasons. A collapsible note derives scale, declared direction and aggregation assumptions from the data contract. Model applicability, failures, warnings and fitting values are reported, and z-curve refusals and warnings are shown.
  • Generated apps include editable www/metaui.css and www/metaui.js: responsive tables and plots, WCAG AA text contrast, visible keyboard focus, and a skip link to results.
  • Uploads are validated, kept in session state and clear cached results. Downloads contain the full current data, applied selections, selected summary and provenance, and wait for pending restorations.
  • Slider bounds round outward to a round tick interval so default bounds retain all eligible rows and end labels no longer overlap.
  • The effect-size list rounds effect sizes to three decimals and other numeric columns to two, and shows URL/DOI columns as links. A “(Missing)” level appears only where a filter has missing values.

Bug fixes

  • Corrected the random-intercept specification in rma.mv() (now with sparse = TRUE), REML heterogeneity output, and Egger residual variance. Meaningless multilevel tau² is no longer shown; singleton-study variance splits are reported as unidentified.
  • Optional p and N columns no longer drop rows under default preparation; required inputs are validated, and excluded rows and derivations are reported.
  • Categorical filter levels containing a double quote no longer stop the app from starting.
  • A non-converging random-effects model behind the funnel plot and Egger’s test now gives a clear message instead of a raw error.
  • The outlier violin plot is standardised against the built dataset and no longer fails when no outliers are drawn; moderation slopes are no longer rounded to “-0.00”; all moderation-table rows are shown.
  • Correct the tutorial’s Wald p calculation and Fisher-z labels.
  • Fixed describing moderators where an “Other” category already existed, and checking that filters are factors or numeric (#28).

Performance

  • P-curve deduplicates identical noncentrality solves without changing results.
  • Compatible multilevel/RVE fits are reused across diagnostics, and a session-local three-entry cache speeds repeated selections. Custom models default to no caching; options$fit_cache_entries = 0 disables it. Measurements are in validation/performance-optimisation.md in the source repository.

Testing

  • New examples: the CC-BY Dannheim mental-health pool with a source-reference REML/Hartung-Knapp model, and a synthetic correlation example with an independent Fisher-z reference fit.
  • Independent statistical references, headless server tests and a generated-app browser test; CI checks Linux (current and previous R), macOS and Windows.

metaUI 0.1.1

  • Fixed issue where k was not displayed for moderators with spaces in names
  • Simplified implementation of generate_shiny() - it is now always saved, either to provided path or to temporary files, and launched from file. This should increase robustness. Also, removed unnecessary app.R so that globals.R does not need to be called explicitly.
  • Added a NEWS.md file to track changes to the package.
  • Added explicit option to include or exclude NAs on filters and report NA share in Sample descriptives
  • Added Reset button to reset all filters
  • Added option to add popups to filter, by specifying filter_popups in generate_shiny()