AI Glossary / Governance, ethics and safety / Beginner

Data quality measure

Definition

A data quality measure is a variable used to express a measured aspect of data quality. Examples include a missing-value rate, an observed error rate or the elapsed time between an event and the availability of its record.

What it means

Quality is easier to discuss when broad concerns become measurable. “Some records are late” becomes more useful when a team defines lateness, counts the affected records and reports the observation period.

The measure is the defined variable; the measurement process produces its value. A requirement may then specify an acceptable range, and a claim may state whether the data meets it. These roles should remain distinct. A dashboard number is difficult to interpret unless the underlying definition, denominator and data scope are known.

How it works

Choose a quality characteristic relevant to the intended use. Define the measure precisely, including eligible records, units, counting rules and treatment of missing or inapplicable values. Specify whether the method checks every record or estimates a value from a sample.

Run the measurement consistently and retain the dataset version and rule version. Report the result with appropriate context, such as subgroup values, uncertainty or known exclusions. Compare it with a requirement only when the two use compatible definitions.

Review whether the measure actually captures the concern. For example, a nonempty address field measures a form of completeness, but it does not establish that the address is deliverable. A good measure makes its limits visible rather than pretending to summarize all quality in one score.

Three practical examples

• Data ingestion: A team measures the percentage of expected daily files received by a fixed deadline, helping identify gaps in a forecasting feed.
• Annotation review: Reviewers measure disagreement on a representative sample of labels, separating ambiguous cases from clear mistakes.
• Record integration: A data engineer counts duplicate business identifiers under a documented uniqueness rule and investigates whether duplicates reflect errors or legitimate repeated events.

Think of it this way

Think of a thermometer: it provides a defined measurement of temperature, not a complete judgment about whether a room is comfortable. Similarly, one data quality measure can be accurate and useful while leaving other important aspects of suitability unmeasured.

Worked example

A table has two thousand records, but the delivery-date field applies only to 1,600 completed deliveries. Among those eligible records, eighty dates are missing. The applicable missing-date rate is 80/1,600=5%.

Using all two thousand records as the denominator would produce 4%, which answers a different question and could understate the problem for completed deliveries. If a requirement permits at most 2% missing dates among completed deliveries, the measured value exceeds the limit.

The example also shows why inapplicable records should not be silently treated as successful observations. The denominator and eligibility rule are part of the measure's meaning, not minor reporting details.

Illustrative case study

Fictional case: Ledger Lane monitors the completeness of transaction feeds. Two departments report different rates because one counts all rows and the other counts only records expected to contain settlement dates.

They agree on separate measures for the two questions and label both clearly. The dashboard then shows overall coverage and settlement-date completeness without mixing them. Owners also retain raw counts so that changing volumes are visible. This invented case illustrates measurement alignment rather than a real improvement claim.

Common misconceptions and mistakes

A measure is not automatically a target or an acceptance decision. A missing-value rate of 1% may be tolerable in one application and unacceptable in another.

Do not combine unrelated measures into one index without explaining the weighting and trade-offs. A high average can hide a serious defect in a critical subset. Avoid selecting only measures that are easy to automate: factual accuracy or representativeness may require more demanding evidence. Finally, a measure can be gamed if staff optimize the reported number while changing eligibility rules or suppressing difficult records.

Business use

Businesses benefit from a small set of clearly defined measures tied to actual decisions. Each should have an owner, a reason for being tracked and an action triggered by material deterioration.

Preserve numerator, denominator, time window and rule version alongside percentages. Compare like with like across teams and periods. When a measurement method changes, explain the break in the series. Use quality measures to support investigation and improvement, rather than treating every dashboard movement as evidence that the underlying data became better or worse.

People also ask

Is a measure the same as its value?

No. The measure defines what is assessed; a measurement assigns a particular value for a specified dataset and context.

Can a quality measure be non-numerical?

Depending on the measurement scale, results can include defined categories as well as numerical quantities.

Why does the denominator matter so much?

It determines the population to which a rate refers and can materially change the interpretation.

Check your understanding

Ten labels are incorrect in a reviewed sample of one hundred. Can the team report that exactly 10% of every label in the full dataset is wrong?

Answer: No. The sample error rate is 10%; extending it to the full population requires a justified sampling design and appropriate uncertainty.

Related terms

• Completeness
• Accuracy
• Data quality claim
• Data quality characteristic
• measurement
• measurement scale
• timeliness
• acceptance threshold

Learn next

Define a quality measure for a dataset you know. Write its eligibility rule, numerator, denominator, units and observation period. Calculate it for the whole dataset and one important subgroup. Then identify a quality problem the measure would miss and design a complementary assessment.

More in Governance, ethics and safety

• Data poisoning - Data poisoning is deliberate interference with data used by an AI system in order to cause unwanted behavior.
• Data privacy - Data privacy concerns how information about people is collected, used, shared, retained, and otherwise processed, and how those activities affect the people...
• Data proxy - A data proxy is observed information used to stand in for another quantity or characteristic that is difficult, unavailable, or impossible to measure directly.
• Data quality model - A data quality model organizes the characteristics used to define and evaluate data quality.
• Data quality plan - A data quality plan specifies how an organization or project will achieve its data quality objectives.
• Data quality planning - Data quality planning is the activity of setting data quality objectives and deciding which processes and resources are needed to achieve them.

Sources

ITU-T Y Supplement 97 (11/2025), term 3.1.33, consulted directly in the official PDF: https://www.itu.int/rec/T-REC-Y.Sup97-202511-I/en. The glossary attributes this definition to ISO/IEC 5259-1. ISO publisher overview consulted: https://www.iso.org/standard/81088.html; the complete ISO standard was not accessed. The examples and workflow are original explanatory material, not prescribed thresholds or a conformity assessment.

Previous: Data quality firewall Next: Data quality model