mb3-analysis-public

Analysis code, data and manuscript sources for ManyBabies 3: A Multi-Lab Study of Infant Algebraic Rule Learning (Developmental Science, 2026).

This repository reproduces the full pipeline: raw lab submissions → merged and validated data → exclusions → the reported models → the manuscript and supplemental PDFs.

Reproducing the results

Requirements: R (>= 4.5), the Quarto CLI, and a LaTeX distribution providing lualatex. Package versions are pinned with renv:

Rscript -e 'renv::restore(prompt = FALSE)'

Then:

# everything except the Bayesian models (~15 min)
Rscript run_all.R --canonical --skip-bayes

# see the steps without running anything
Rscript run_all.R --list
Rscript run_all.R --dry-run

run_all.R renders each document in a fresh R subprocess, checks that each step's inputs exist before it runs and that its declared outputs were actually written, and logs the run to logs/. Options: --only=, --from=/--to=, --skip=, --skip-bayes, --canonical, --keep-going.

Every intermediate data file and fitted model is committed, so the manuscript can be rebuilt without rerunning the analyses:

Rscript run_all.R --only=paper,supplement

The two analysis branches

Steps 04, 05 and 07 exist in two variants that differ only in how experimental_method is coded:

	file suffix	artefacts	reported?
deviation coding	_deviation.Rmd	unsuffixed, e.g. tables/Bayes.factors.csv	yes — this is what the paper reads
dummy coding	_dummy_coding.Rmd	*_dummy_coded_method.*	no — kept as a robustness check

The change from dummy to deviation coding relative to the Stage 1 registered report is described in the manuscript; both lead to identical conclusions. --canonical runs only the reported branch.

Runtimes

step	file	time
01	01_read-merge.Rmd	4s
02	02_data-validation.Rmd	5s
03	03_data-exclusions-sequential.Rmd	2s
04	04_main-analysis_robustness_*.Rmd	13 min
05	05_main-analysis_Bayesian_*.Rmd	hours (brms + bridge sampling)
06	06_additionalHypotheses.Rmd	3s
07	07_exploratory_analyses_*.Rmd	3s
—	paper/mb3rulelearning.qmd	26s
—	paper/mb3rulelearningSupplemental.qmd	7s

Step 05 refits 18 brms models at 20,000 iterations and bridge-samples each one; --skip-bayes leaves it out, and the manuscript then uses the archived tables/Bayes.factors.csv. Those Bayes factors come from a seeded run, so rerunning step 05 reproduces them.

Repository structure

path	contents
data/experiment_{participant,trial}_data/	CSVs as submitted by the 30 contributing labs
data/processed_data/	outputs of steps 01–03 (merged, validated, post-exclusion)
tables/	fitted models (.rds) and model summaries (.csv) read by the manuscript
model_output/	created at runtime by step 05; brms fits are not committed (too large)
scripts/	helpers sourced by the analysis files
paper/	manuscript and supplemental Quarto sources, references, apaquarto extension

08_additionalDemographics.qmd is a standalone report — it reads the processed data but produces no artefact the analyses or manuscript consume, so it is not part of run_all.R. Render it with quarto render 08_additionalDemographics.qmd.

00_power-analysis.Rmd is the Stage 1 registered-report power analysis, kept unchanged for the record.

About the data

Lab submissions arrived through the MB3 data submission form during data collection, up to May 2024, and were checked for confidential information, unexpected columns and out-of-range values before being accepted. Corrections were made in R wherever possible rather than by editing submitted spreadsheets, so that a lab's resubmission could never silently conflict with a local edit. 01_read-merge.Rmd documents each lab-specific correction inline.

Free-text notes have been removed. The notes columns in the participant and trial files, and the notes.participant / notes.trial columns derived from them, are present but empty. They were written by research assistants for internal use and were not screened for incidental personal information, so they are withheld here. Nothing in the analyses reads them: removing them changes no model, table or figure, and the pipeline reproduces every other column exactly. The notes fields are available on request from Ingmar Visser (i.visser@uva.nl) for secondary data analysis.

Other free-text fields — session_exclusion_info, trial_error_info, medical_issue_info, developmental_concern_info — are retained, because the exclusion tables are built from the first of these.

Citing

If you use this code or data, please cite the paper:

Visser, I., Geambasu, A., Kachergis, G., … Soderstrom, M., & Levelt, C. C. (2026). ManyBabies 3: A Multi-Lab Study of Infant Algebraic Rule Learning. Developmental Science.

A BibTeX entry with the full author list is in paper/mb3-references.bib under the key visser2026manybabies. The preprint is at https://doi.org/10.31234/osf.io/ghrdt_v2.

Questions and contributions

Issues and pull requests are welcome. For questions about the analyses or the data, contact Ingmar Visser (i.visser@uva.nl). Please do not include identifying information about participants in issue text.

Code and data are released under the terms in LICENSE.