FAIR Research Data: More Than Posting a Download Link

A file can be online and still be unusable. A link may disappear, column names may be unexplained, codes may be unique to one laboratory, and no license may say whether reuse is permitted. The FAIR principles focus on those practical failures. They ask whether you can discover the data and their metadata, retrieve them through a durable process, combine them with other resources, and interpret them well enough for responsible reuse.

FAIR and open answer different questions#

"Open" usually concerns who can obtain a resource and under what restrictions. "FAIR" concerns whether the resource is managed so people and computers can find and use it appropriately. The concepts often support each other, but they are not synonyms.

A fully public spreadsheet with an unstable link and undefined columns may be open but not FAIR. A controlled genomic dataset with a permanent catalog record, rich metadata, a standard application process, and clear reuse conditions may be FAIR without being open to anonymous download.

This distinction is essential in health research. Privacy, consent, tribal or community governance, intellectual property, and legal duties can justify access controls. FAIR practice makes those controls visible and workable rather than pretending they do not exist.

Findable: identify the object and describe it#

Findability begins with a persistent, globally unique identifier such as a DOI or an accession number. A normal webpage address points to a location that can change, while a persistent identifier is managed so that it keeps resolving even when the storage or the website structure changes.

The identifier must lead to metadata rich enough for discovery. Useful metadata include:

• title, creators, funder, and responsible institution;
• population, sampling frame, dates, setting, and study design;
• variables, units, instruments, and outcome definitions;
• version and update history;
• related protocol, registration, publication, code, and documentation;
• access conditions, license, and contact or application route.

The metadata themselves need to be indexed in a searchable repository. A perfectly described file on a private laptop is not findable. Likewise, a dataset named "final2_revised.csv" without a catalog record is effectively invisible outside its immediate team.

Accessible: make retrieval durable and explicit#

Accessibility means that the identifier can be resolved and the data or metadata retrieved through a standardized communication method. The method may require authentication or authorization. What matters is that the process is specified, usable, and not dependent on an undocumented personal favor.

For controlled health data, the catalog can state who may apply, what review occurs, which agreements are required, how long access lasts, and what computing environment is used. A rejected application does not necessarily make the dataset non-FAIR if the restrictions are legitimate and the process is transparent.

The original FAIR principles also emphasize that metadata should remain available even when the underlying data are withdrawn or no longer retained; a durable record tells you that the dataset existed, what it contained, and why it is unavailable. Deleting both the file and its description creates a gap in the research trail.

Interoperable: use a shared language#

Interoperability means another system can combine and interpret the data without reverse-engineering local conventions. A comma-separated file is not automatically interoperable. If "1" means yes in one column, female in another, and missing in a third, software cannot understand the dataset safely.

Practical interoperability requires:

• standard, nonproprietary formats where feasible;
• consistent units and explicit unit conversion;
• controlled vocabularies, ontologies, or recognized code systems;
• stable variable identifiers and data types;
• machine-readable links to related datasets and metadata;
• documentation of how local codes map to community standards.

Healthcare data make the challenge visible. Diagnosis, medication, laboratory, and imaging systems may use different terminologies and granularities. Mapping them is scientific work, not a clerical afterthought. A documented mapping may still lose meaning, so the provenance of each transformation should travel with the result. Interoperability does not require forcing every discipline into one vocabulary. It requires declaring the vocabulary and relationships clearly enough that translation can be evaluated.

Reusable: preserve meaning and permission#

Reuse is the destination. Can you work out what the data mean, whether they are suitable for a new question, and what use is allowed?

An explicit license or data-use agreement addresses permission. A data dictionary addresses structure. Provenance records how data were collected, cleaned, transformed, and versioned. Quality notes identify missingness, known errors, validation rules, and changes across releases. Consent and governance documentation define purposes that are allowed or prohibited.

Without these pieces, reuse can produce a polished but invalid analysis. A variable labeled "date" might mean specimen collection, diagnosis, data entry, or month shifted for privacy. A zero may mean absence, below detection, or not measured. Good stewardship anticipates these ambiguities.

Reusable also means citeable. Data creators deserve credit, and you need a precise version to cite. Versioned identifiers and recommended citations make both possible.

Machine actionability is a design requirement#

The FAIR principles were written for machines as well as people. Modern discovery involves search indexes, automated workflows, and software agents that cannot email an author to ask what a column means; metadata buried only in a narrative PDF may help a patient human reader while remaining difficult for software to parse.

Machine-actionable stewardship uses structured fields, persistent identifiers, standard schemas, and explicit relationships. It does not mean removing human judgment. Machines can locate candidate datasets and combine compatible fields; researchers still assess validity, ethics, and fitness for purpose.

A useful test is to ask what a new team could do without private background knowledge, and if they can discover the dataset but need several undocumented conversations to decode it, the object is only partly FAIR.

Current policy turns principles into planning#

The NIH Data Management and Sharing Policy took effect on January 25, 2023, and it requires applicable research to include a prospective plan for managing and sharing scientific data and expects approved plans to be followed. The policy encourages practices consistent with FAIR principles while allowing justified limitations or exceptions.

That planning matters because FAIRness cannot be added reliably at the end. Consent language, identifiers, variable standards, repository selection, de-identification, curation costs, and access governance all affect study design and budgeting. A team that waits until publication may discover that participants did not consent to the proposed reuse or that essential metadata were never captured.

Journal data-sharing statements add another layer, and they can state whether individual participant data will be shared, which documents accompany them, when access begins, who may request access, and for what purposes. The statement is useful evidence of a plan. It is not proof that access will work or that the data are accurate.

FAIRness is not a quality score#

A well-documented dataset can contain biased measurements. A persistent identifier can point to a poorly designed study. Interoperable fields can make invalid pooling easier as well as valid pooling. FAIR principles improve stewardship and auditability; they do not certify scientific quality.

Nor is FAIR a single binary label. A resource may be highly findable but poorly reusable, or well documented but reachable only through an unclear process, and an assessment should state which elements are strong and which need work, rather than attach a single unexplained badge.

A reader's FAIR audit#

1. Does the dataset have a persistent identifier and a version?
2. Can its metadata be found without already knowing the exact repository page?
3. Are access rules, restrictions, and response processes explicit?
4. Do variables, units, code lists, missing values, and time points have definitions?
5. Are standard vocabularies used or local mappings documented?
6. Is the origin and processing history traceable?
7. Is there a license or data-use agreement consistent with consent and governance?
8. Can the dataset and its documentation be cited together?

These questions convert a broad principle into observable infrastructure. FAIR data are not simply available data. They are data whose chain of meaning survives beyond the team that created them.