About MetaLab

MetaLab is a collection of community-augmented meta-analyses (Tsuji, Bergmann, & Cristia, 2014) of phenomena in early language acquisition and cognitive development. Each dataset codes the published (and sometimes unpublished) literature on one phenomenon into a standardized effect-size format, so that anyone can explore the evidence, plan well-powered studies, and contribute new data as the literature grows.

A short meta-analysis primer

New to meta-analysis? Five concepts unlock everything on this site. For a fuller treatment, see the meta-analysis chapter of Experimentology (Frank et al.) or the free online book Doing Meta-Analysis with R (Harrer et al.).

Effect size. A standardized measure of how big an experimental effect is, comparable across studies with different measures and scales. MetaLab uses Cohen’s d (roughly: the difference between conditions in standard deviation units; d = 0.5 means the average infant in one condition scores half a standard deviation above the other) and its small-sample-corrected cousin Hedges’ g. Every point in every plot on this site is one effect size from one experimental condition.

Meta-analysis. A statistical summary of all the effect sizes for one phenomenon: a weighted average in which more precise studies (usually larger ones) count more. MetaLab fits random-effects models, which assume the true effect varies across studies rather than being one fixed number — the Visualizations page shows each dataset’s estimate.

Moderators. Variables that explain why effects differ across studies — infant age, experimental method, stimulus type. Selecting moderators on the Visualizations page refits the meta-analytic model with those predictors, showing each group’s estimated effect.

Funnel plots and publication bias. A funnel plot places each study by its effect size (x) and precision (y): precise studies cluster near the average at the top, noisy studies spread out at the bottom, making a funnel shape. If small, noisy studies appear only when their effects are large (because null results go unpublished), the funnel is lopsided — that asymmetry (tested with Egger’s regression on the Visualizations page) is a warning sign that the literature over-states the true effect. See the Replicability page for these analyses across all of MetaLab.

Statistical power. The probability that a study of a given size detects an effect of a given size. Infant research is chronically underpowered — the median MetaLab study has ~44% power for its phenomenon’s meta-analytic effect (Bergmann et al., 2018). The Power Analysis tool turns any MetaLab dataset into a sample-size plan, and the Power Simulation lets you feel what an underpowered study looks like.

The data

  • Every effect size is standardized to Cohen’s d (with Hedges’ g, Pearson’s r, and log odds conversions) with variances, following the formulas documented in the metalabr package.
  • Datasets are curated in spreadsheets by their curators, validated against the MetaLab field specification, and released as versioned snapshots on Redivis: datapages.metalab. Analyses on this site run from the released data, so results are citable and reproducible against a specific release.
  • The current release is shown on the home page; release notes document exactly what changed in each version.

Using MetaLab data

You can use MetaLab data in three ways:

  1. This site — every visualization has a data table and CSV download.
  2. Redivis — query or download any table from the versioned archive.
  3. R — the metalabr package provides get_metalab_data() against released versions.

Before using MetaLab data in a publication, please cite the dataset(s) you use (citations appear on each dataset’s page) and the platform papers listed under Publications.

Contributing

New datasets and updates to existing datasets are welcome! The validator checks a candidate dataset against the MetaLab field specification. Contact the team (see the Team tab) to get started.

License

Unless otherwise noted, MetaLab and its datasets are licensed under a Creative Commons Attribution 4.0 International License.

MetaLab is a community project: datasets are contributed and maintained by researchers across many labs, coordinated by the team below. Interested in contributing? See About or write to any of the leaders.

Leaders

Governing board

Team

Curators

Alumni

Julia Carbajal, Mika Braginsky, Page Piccinini, Alessandro Sanchez

2026 — the datapage rebuild

The MetaLab site was rebuilt as a static Quarto + Observable site backed by versioned data releases on Redivis, replacing the Hugo site and the Shiny applications (whose hosting server had gone offline). The analyses now run in your browser from released data, so every number on the site is tied to a citable data version.

Data releases

  • 2023.1 (Redivis v1.0) — verbatim snapshot of the final legacy pipeline output (April 2023), exactly as previously served by the site and apps.
  • 2026.1 (Redivis v1.1) — re-fetch of the curator spreadsheets (August 2026). Effect sizes are identical to 2023.1 — the database has been stable since 2023. (Seven datasets whose coding spreadsheets are no longer hosted are included at their 2023 state; the datasets table records data provenance for every dataset.)

Deliberate changes from the legacy applications

Fixes (behavior the legacy apps got wrong):

  1. Funnel-plot asymmetry test. The legacy visualization called metafor::regtest() on a multilevel (rma.mv) model, which always errored — the error was hidden by CSS, so no test was actually shown. The rebuilt funnel reports a proper Egger-style test: √vi added as a fixed-effect moderator to the same multilevel model.
  2. Power simulation with a control group. The legacy simulation accidentally duplicated every simulated observation when the control group was enabled (a factor() construction bug), inflating the apparent significance of all tests. Each cell now contains exactly N independent draws.
  3. Power-analysis model. The legacy power app fit a single-level random-effects model, inconsistent with the visualization app’s multilevel model on the same data (and with how the datasets are analyzed in publications). Both now use the same multilevel model (effect sizes nested in infant groups nested in papers). Estimated effect sizes — and therefore recommended sample sizes — can differ slightly from the legacy app.
  4. Broken PDF download in the legacy power app (saved an empty file) is replaced by data downloads on every visualization.
  5. Violin plot now works for every effect-size type (legacy silently supported only Hedges’ g).

Documented approximations and constraints:

  1. Residual funnel plots (funnel with moderators selected) standardize residuals by the marginal standard error √(vi + Σσ²) rather than metafor’s rstandard() internals; positions can differ very slightly from the legacy plot.
  2. Moderator analyses cover up to three simultaneous moderators (the models are precomputed offline for every dataset × subset × effect-size type × moderator combination; the legacy app allowed unbounded combinations by fitting on a live server).
  3. Age filter. All analyses exclude effect sizes with mean age ≥ 4000 days (~131 months), as the legacy visualization did; the legacy power app did not apply this filter, so its estimates could differ trivially for datasets with very old samples.
  4. Column rename: rule.type → rule_type (Abstract rule learning) — Redivis does not allow dots in column names.

metalabr

The R package is being updated to read released data versions from Redivis (get_metalab_data(version = "current")), with the Google-Sheets path retained for curators. See the package repository.