Are reference metadata the same as the data values?

No. They describe the meaning, production and quality of those values.

Do they make two datasets automatically comparable?

No. They help users assess whether the methods and scope support comparison.

Can an accurate number still be summarized incorrectly?

Yes. Changing its population, denominator or qualifications can change the meaning of the claim.

AI Glossary / Data engineering and architecture / Advanced

Reference metadata

Definition

Reference metadata describe the content, methods and quality of statistical data so that users can interpret and assess them. They explain what the figures represent, how they were produced and which limitations affect their use. In SDMX terminology, they are distinguished from structural metadata that organize and identify data and metadata. Reference metadata provide interpretive context rather than merely the layout needed to process a file.

What it means

A percentage can be arithmetically correct but easy to misunderstand. Users need to know the population, time period, collection method and meaning of the numerator and denominator. In an original AI learning-platform example, a reported completion rate might apply only to participants who answered a follow-up survey. Describing it as the completion rate of everyone invited would change the claim. Reference metadata help preserve those qualifications. They do not replace the observations or make an unsuitable comparison valid; they give users information needed to decide whether a particular interpretation is justified.

How it works

SDMX describes conceptual, methodological and quality-related explanations as parts of reference metadata. Such information can be organized in metadata reports linked to the data or other objects they describe. In an original reporting workflow, the team documents the measure before an assistant summarizes it, records relevant limitations and checks that the report applies to the correct release. A structured template can improve consistency, but filling it does not independently verify the statements. The descriptions need evidence, maintenance and clear attachment to the data whose interpretation they qualify.

Three practical examples

A learning survey explains that its reported completion percentage covers respondents rather than all invited participants.
An AI dashboard includes methodological notes when comparing measures produced through different collection processes.
A data publication documents a change in coverage so readers do not automatically interpret a time-series jump as a real-world increase.

Think of it this way

Imagine receiving a scorecard together with notes explaining the scoring rules, who participated and which rounds were omitted. The scorecard supplies numbers; the notes help you understand what those numbers can support. Removing the notes can make two scores look comparable when they were produced differently. The analogy illustrates why context matters without implying that explanatory notes excuse errors in the scores themselves.

Worked example

Consider an original fictional follow-up survey sent to 100 course participants. Eighty respond, and 60 respondents report completing the course. The reported completion proportion among respondents is 60/80 = 75%. Reference metadata state that the measure is self-reported, covers the 80 respondents and does not establish completion status for the 20 nonrespondents. Dividing the same 60 by all 100 invited gives 60%, but that is the proportion of invitees known from this survey to have reported completion, not an established overall completion rate. If all responses are accurate, actual completion among the full invited group could range from 60/100 = 60% to 80/100 = 80%, depending on the unobserved outcomes. These are logical bounds under the stated assumptions, not a confidence interval or a correction for response bias. The metadata prevent the assistant from silently changing the denominator or treating missing responses as confirmed noncompletion.

Illustrative case study

In a fictional AI learning dashboard, an automated summary says that three quarters of all participants completed a course. Reviewers compare the statement with the survey's reference metadata and discover that the 75% figure concerns respondents only. They revise the summary, display the response coverage and preserve the self-report qualification. They also link the note to the relevant survey release so it is not reused indiscriminately for a later cohort. The scenario illustrates correcting interpretation without changing valid source arithmetic. It does not establish the actual completion rate among nonrespondents or claim that adding a note removes the survey's limitations.

Common misconceptions and mistakes

Reference metadata are not merely optional commentary after the real analysis. They can determine what the analysis means. Another mistake is treating a shared column layout as evidence that two measures are comparable. The populations or methods may still differ. A detailed methodology description also does not prove that the process was followed correctly; it should be supported by evidence. Finally, an assistant can omit essential qualifications while accurately repeating a number. Review should examine whether the generated claim preserves the source's scope, method and limitations, not just whether the digits match.

Business use

For an AI reporting feature, make relevant context available where users encounter the figures. In the fictional course dashboard, distinguish invited participants, respondents and reported completers. Retain the collection method and unresolved coverage in the explanation. When the method or population changes, update the metadata and reconsider comparisons with earlier releases. Test summaries with examples where omitting one qualification materially changes the claim. This helps teams use data within its actual evidential limits and reduces the chance that concise automated prose turns a narrow observation into an unsupported general conclusion.

Check your understanding

What does the toy 75% measure?

Answer: Completion reported by the 80 survey respondents.

How many invited participants have unknown survey outcomes?

Answer: Twenty.

Why is 60% not automatically the overall completion rate?

Answer: Nonrespondents may also have completed.

What does the 60% to 80% range represent?

Answer: Logical bounds assuming accurate responses, not a statistical confidence interval.

Which qualification should the assistant retain?

Answer: The measure is self-reported and limited to respondents.

Related terms

Learn next

Take a fictional percentage and write its numerator, denominator, population and collection method. Create a short summary that preserves all essential qualifications. Then remove one qualification and explain precisely how the resulting statement could mislead a reader, even though the numerical percentage remains unchanged.

Sources

SDMX, Glossary Version 2.1, December 2020, Reference metadata: https://sdmx.org/wp-content/uploads/SDMX_Glossary_Version_2_1_December_2020.htm SDMX, Framework for SDMX Technical Standards, version 2.1: https://docs.sdmx.org/en/i1-doc/Sections/Section1/SDMX_2-1_SECTION_1_Framework.html The survey counts, logical bounds, scorecard analogy and AI dashboard scenario are original educational illustrations. No observed study results, statistical confidence interval or correction for nonresponse bias is claimed.

Previous: Reference group Next: Reference path