Finding research datasets

The third route is the most reliable and the most overlooked. A paper reporting exactly the measurement you need will name its deposit in the data availability statement or give an accession number in the methods, and that is a direct pointer no keyword search would have produced.

re3data is a registry of repositories rather than of datasets. Use it to answer where would data like this be deposited, then search there — domain repositories carry structured, field-specific metadata that a general index flattens away.

• Zenodo — CERN-operated, takes anything, issues DOIs, integrates with GitHub for code releases.
• Dryad — curated, mostly data underlying published papers, strong in ecology and evolution.
• figshare — very broad, including figures, posters and supplementary material.
• Open Science Framework — project-oriented, good where data, code, protocol and preregistration belong together.
• Harvard Dataverse — widely used in the social sciences, with versioning and access controls.

A general repository is where data goes when no domain repository fits. If one does fit, the domain repository is almost always the better place to look: its metadata describes the measurements rather than just the file.

Do these before downloading anything large. Each one has ended a dataset evaluation for someone who did it in the wrong order.

1. Licence. Look for an explicit CC0, CC BY or equivalent. No licence is not permission — it is the default of all rights reserved, and it will block publication later even if the download works now.
2. Documentation. Is there a codebook, data dictionary or README that defines every variable, its units and its missing-value codes? Undocumented columns are unusable at any size.
3. Format and completeness. Open formats, and the actual files present rather than a landing page promising them. Check whether what is deposited is the raw data or only derived summaries.
4. Provenance and version. Which version is this, what changed, and is there a DOI that resolves to this exact version rather than to the latest one?
5. Ethics and restrictions. Human-subject data frequently carries consent limits, a data use agreement, or an application process. Find that out before building anything on it.

Cite the dataset itself, not only the paper that used it: creators, year, title, repository, version, and the DOI. Most repositories publish a preferred citation on the landing page — use it.

Cite the version you used. Datasets are revised, and a DOI that always resolves to the latest version will silently misdescribe your analysis once the data changes. Where a version-specific DOI exists, that is the one to cite.

If a paper reports the measurements you need but deposits nothing, email the corresponding author. Data sharing on request is still the stated policy of many journals, and a specific request naming the variables and the intended use is answered far more often than a general one.

Check the supplementary material first, though. A surprising amount of usable data is buried in a supplementary spreadsheet that no dataset index has ever seen.