AI Glossary / Statistics and probability / Intermediate

Probability based disclosures

Definition

Probability based disclosures occur when released information enables a sensitive statement to be inferred with high probability even though the statement is not established with certainty. In statistical disclosure control, this concerns what a recipient can learn about a person, household or organization from published data and relevant background knowledge. The term distinguishes probabilistic inference from certain disclosure. It does not require a record to display a name or a sensitive fact explicitly.

What it means

Confidentiality can be weakened without revealing an exact answer. If a release makes a sensitive attribute highly predictable for a known individual, the residual uncertainty may offer little practical protection. However, a high probability alone does not describe the entire disclosure question. Reviewers must consider what the recipient already knew, what the release adds, whether the person can be linked to the relevant group and whether the probability calculation is justified. There is no universal numerical threshold inherent in this term.

How it works

A disclosure assessment identifies a plausible recipient, their auxiliary information and the protected fact. It then examines how the proposed release changes the recipient's ability to infer that fact. Group composition, concentrated attribute values and links across releases can all matter. Probability statements should be tied to an explicit model or sampling argument rather than unsupported confidence. Protective changes may alter detail, restrict access or use a suitable formal privacy mechanism. Each option requires its own utility and confidentiality assessment.

Three practical examples

Example 1: A small published subgroup contains almost entirely one sensitive response, allowing a known member's response to be guessed with high confidence. Example 2: Multiple tables jointly reveal that a particular organization is very likely to belong to a sensitive category even though no single table proves it. Example 3: A released dataset and external membership information combine to make an attribute predictable. These examples concern inference risk; they do not claim that any particular release necessarily identifies a real person.

Think of it this way

Think of confidentiality as protecting more than a locked box's exact contents. A label that makes the contents almost certain can reveal useful information even if the box remains closed. How much is newly revealed depends on what the observer knew beforehand. Removing the owner's name from the label is insufficient when other information reliably links the box to that owner.

Worked example

Consider a hypothetical group of 20 people, with 18 having a sensitive attribute and two not having it. A recipient knows that a target belongs to the group but has no information distinguishing members with respect to the attribute. Under that exchangeability assumption, the group composition supports an attribute probability of 18/20 = 0.90 for the target. Suppose the recipient's justified prior probability before the release was 0.20. The increase is 0.70, or 70 percentage points. Prior odds are 0.20/0.80 = 0.25; posterior odds are 0.90/0.10 = 9, a 36-fold increase. The release still does not prove the target's attribute. If membership or exchangeability is uncertain, this simple calculation no longer applies unchanged.

Illustrative case study

Fictional case study: Elm Research prepares a community survey table with names removed. During release review, analysts notice that one externally recognizable subgroup is dominated by a sensitive response. They examine what the table would add to a reader's existing knowledge and test a broader grouping alongside restricted-access alternatives. The team documents the resulting trade-off before publication. This fictional case illustrates a confidentiality assessment; it does not assert that aggregation always resolves disclosure risk or that a particular legal standard has been satisfied.

Common misconceptions and mistakes

Common mistakes include assuming that disclosure requires certainty, treating removed names as sufficient protection, and calling every accurate population prediction a new disclosure. Background knowledge and incremental information matter. A homogeneous group can create attribute-inference concerns even when its members cannot be individually distinguished within the released data. Conversely, a numerical posterior based on unrealistic assumptions can exaggerate risk. The glossary term is a statistical confidentiality concept, not by itself a complete legal test or a formal privacy guarantee.

Business use

Business use: Organizations releasing research tables, benchmarking data or analytics should assess what sensitive facts recipients can infer, including probabilistic conclusions. Review proposed outputs together with previous releases and plausible external information. Define the protected population and intended recipients. Record assumptions and the limitations of estimated disclosure probabilities. Where necessary, involve specialist statistical disclosure-control expertise to balance useful information against confidentiality obligations and the actual release environment.

Check your understanding

Why is the recipient's earlier knowledge relevant?

Answer: It helps distinguish newly disclosed information from what was already inferable.

Why does the example state an exchangeability assumption?

Answer: Group proportions do not automatically provide the correct probability for a particular target if additional information distinguishes that target from other group members.

Learn next

Compare certainty disclosure with a high-probability attribute inference in a small fictional table. Specify the recipient's knowledge before and after release, then change the membership assumptions. Explore how multiple releases alter the assessment. Study formal disclosure-control methods separately, keeping their guarantees and implementation conditions distinct from intuitive claims that a dataset appears anonymous.

Sources

Sources: OECD Glossary of Statistical Terms, Probability based disclosures: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf . The OECD attributes this entry to the 2005 statistical disclosure-control glossary prepared by participating statistical institutions and researchers. The probability and odds calculation, release-review discussion and fictional case are original explanatory illustrations, not a prescribed regulatory test.

Previous: Probability Next: Probability box