Must disclosure identify a person by name?

No. It can reveal a protected attribute or value without publishing a name.

Can approximate information count as disclosure?

Yes, under the broad information-gain definition used here.

Can two acceptable-looking outputs create a problem together?

Yes. Their combination may reveal information absent from either output alone.

AI Glossary / Statistics and probability / Intermediate

Statistical disclosure

Definition

Statistical disclosure occurs when released statistical information allows a recipient to learn protected information about a statistical unit more accurately than was possible without that release. The unit may be a person, household, or business. Disclosure can involve identification, a confidential attribute, or improved inference about a value rather than exact reconstruction. This entry follows the supplied OECD glossary's broad information-gain definition; whether a particular release is impermissible depends on the relevant protection requirements and context.

What it means

A release does not need to contain a name to disclose something sensitive. A recipient might recognize a rare record, combine a total with known components, or use several overlapping outputs to isolate one unit's information.

The comparison with prior knowledge matters. A statistic may reveal little to one reader but much more to someone who already knows part of the dataset. Disclosure assessment therefore considers plausible auxiliary information and the cumulative release environment. It distinguishes the intended learning about populations from unwanted learning about protected units, rather than assuming that all statistical usefulness is itself unacceptable.

How it works

Identify what information must remain protected and who might receive the output. Consider direct identifiers, indirect combinations of attributes, unusual values, dominant contributors, and membership in sensitive groups. Then examine what can be inferred from the proposed release together with other accessible information.

For tables, check whether totals and subtotals reveal hidden cells or whether overlapping groups isolate one contribution. For microdata, examine linkage possibilities and attribute homogeneity within identifiable groups. Model outputs and query systems may also disclose information, so assessment should not stop at conventional tables.

Evaluate risk under a stated model of the recipient's knowledge and capabilities. Some methods quantify the chance of a correct match or the narrowing of possible values; others use formal privacy guarantees or controlled access. No single score covers every form of disclosure. Choose safeguards that address the relevant risk and assess the information lost, then revisit the decision when additional outputs or external data become available.

Three practical examples

1. A published regional total and totals for all but one local provider reveal the remaining provider's confidential contribution by subtraction.

2. A dataset without names contains a rare combination of age, occupation, and location that can be linked to a known individual.

3. Two overlapping workforce summaries differ by one employee, allowing the difference between their totals to reveal that employee's reported value.

Think of it this way

Think of solving a puzzle from several clues. No single clue needs to state the answer directly. A new clue can sharply narrow the possibilities or complete the puzzle when combined with what the solver already knows. Statistical disclosure concerns that added knowledge about protected information.

Worked example

Consider an original fictional release describing a confidential expense amount X. Before publication, a recipient knows only that X is between 0 and 100 units. Another amount A is independently known to equal 80. The publisher releases the combined total X + A as lying between 120 and 130.

Subtracting the known A value gives 40 ≤ X ≤ 50. The width of the possible interval has fallen from 100 units to 10 units, a 90% reduction in interval width. The exact value of X remains unknown, but the recipient has gained substantial information about it.

If the publisher instead releases the exact combined total as 126, the recipient obtains X = 126 - 80 = 46. If A were not known exactly, the inference would depend on its uncertainty as well. These calculations illustrate how background knowledge changes disclosure.

The 90% interval-width reduction is a simple descriptive measure in this constructed example. It is not a universal disclosure-risk score, a probability of harm, or a decision rule for publication. Real assessments must define which information is confidential and how much inference the governing framework permits.

Illustrative case study

In a fictional industry consortium, a report suppresses small contributor values but publishes several carefully rounded aggregates. Reviewers initially focus only on whether an exact value can be recovered.

An analyst demonstrates that combining the aggregates with public company figures narrows a member's confidential expenditure to a very small range. The consortium broadens its review to include approximate disclosure and checks earlier releases as well. It chooses a revised output strategy based on the actual protection objective rather than assuming that anything short of exact reconstruction is harmless.

Common misconceptions and mistakes

Disclosure is not limited to identity disclosure. Sensitive attributes can become inferable even when a recipient already knows who participated. A large group can also reveal an attribute if everyone in it shares that attribute.

Do not equate the absence of a direct identifier with the absence of risk. Nor should every improvement in knowledge about a population be described as a confidentiality failure; the protected target and release purpose must be specified. Suppression, rounding, aggregation, and perturbation all have limitations. Risk can change when new releases or auxiliary datasets appear, so a one-time review may not remain adequate indefinitely.

Business use

Understanding statistical disclosure helps organizations design useful reports, research datasets, and interactive analytics without exposing entrusted information. It encourages teams to review combinations of outputs, document assumptions about recipients, and distinguish exact recovery from meaningful narrowing. Businesses should involve appropriate statistical and governance expertise when deciding which disclosures matter and which safeguards fit the intended users and purposes.

Check your understanding

A confidential amount plus a publicly known value of 70 is released as 112. What does the release reveal about the confidential amount?

Answer: It is 112 - 70 = 42. The inference depends on the public value and the released total referring to the same components and period.

Related terms

Statistical confidentiality
Differencing
disclosure risk
identity disclosure
attribute disclosure
inferential disclosure
auxiliary information
disclosure control

Learn next

Compare identity, attribute, and inferential disclosure. Then study how linked tables, repeated queries, and background knowledge affect the choice of disclosure-control methods.

Background knowledge

Sources

OECD, Glossary of Statistical Terms, Statistical disclosure: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf

Office for National Statistics, Statistical disclosure control policy: https://www.ons.gov.uk/aboutus/transparencyandgovernance/datastrategy/datapolicies/statisticaldisclosurecontrol

The official definition and policy context were read. This entry explains the general statistical concept rather than determining legal permissibility. All numerical scenarios and the fictional case are original.

Previous: Statistical design of experiments Next: Statistical Disclosure Control (SDC)