Assess a dataset for your question
Check definitions, coverage, missing values and provenance before drawing conclusions from a table.
Follow the method ↓
The answer in 30 seconds
A clean-looking table can contain duplicates or missing fields. Quality depends on the question: the same file may support a list of places but fail to support a fair comparison between regions.
Keep the original and documentation, define what one row means, inspect essential missing values and units, then state which uses the data does and does not support.
Examples to adapt
Missing does not mean zero; avoid a calculation that silently makes that assumption.
They may be distinct locations. Test duplicates using the fields that define an observation.
Check whether its time period fits the decision and whether a newer release exists.
Follow the method
- 1Preserve provenance
Save the original file, version, publisher and documentation. Work on a copy.
- 2Define one row
Write down whether it represents a person, place, event or period.
- 3Check completeness
Count missing values in fields required for your question; investigate special codes.
- 4Check consistency
Review units, dates, categories and possible duplicates using the correct key.
- 5State a limited conclusion
Say what the data can answer, what remains uncertain and which transformations you made.
A closer look
A missing value is not a zero
Fictional example: four facilities are surveyed. Three report 12, 18 and 0 visits; the fourth does not respond. The documentation explicitly defines zero as no visits.
| Treatment | Result | Supported interpretation |
|---|---|---|
| Keep known responses | (12 + 18 + 0) ÷ 3 = 10 | Mean of 10 visits among the three respondents |
| Replace missing with zero | 30 ÷ 4 = 7.5 | An observation is added without having been collected |
| Document coverage | 3 of 4 responses = 75% | Collection is incomplete; this does not measure its bias |
Report the calculation, respondent count and expected count separately. A technically complete file can still be unsuitable for your question.
Can the mean of 10 visits be extended to all four facilities?
These three responses alone do not justify that step. You would need to understand non-response and choose an appropriate method. Preserve the missing value and explain the limitation before estimating.
Supporting documentation : data.gouv.fr — Évaluer la qualité des données · Consulted on 2026-10-15. Examples and calculations are illustrative.
A checklist to keep
Use these checks to record what you found. The grid supports a decision; it does not make one for you.
| Check | What to examine | Action |
|---|---|---|
| Unit | What does one row represent? | Define |
| Coverage | Which places and dates are included? | Check |
| Missing | What do blank or special values mean? | Count |
| Consistency | Are units and identifiers coherent? | Inspect |
| Provenance | Can you reproduce each change? | Record |
Read the checklist directly in the table.
What to check
An unusual value needs investigation, not automatic deletion.
A fully populated column can still contain the wrong unit.
Corrections should be traceable and reversible.
Common questions
Is checking a sample enough?
It helps you understand the file, but full-column checks are needed for missing values, types and duplicates.
Can I publish a corrected copy?
Check reuse terms and document changes clearly without implying the producer approved them.
Sources and documentation
Documentation consulted on . Examples are illustrative; interfaces and results may change.