Privacy and preparation
Dataset quasi-identifier combinations: reviewing job title, town and date together
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Review job title, location and date as one combined key, not as separate fields. Count the distinct people who share each pair and triple of values, flag groups below a minimum size you set for the recipient, check whether flagged groups share sensitive outcomes, then generalize or suppress the fields the AI task needs least, rerun the counts, and record what risk remains and who accepted it.
Redaction
Dataset quasi identifier
Key takeaways
- Identity usually comes from combinations: a title, a small town and one date can describe a single person.
- Count people per combination, not rows, and test pairs and triples before the full key.
- The minimum group size is a documented decision tied to the recipient, not a universal constant.
- Choose treatments by what the AI task needs; keep intervals and order, coarsen what the model does not use.
- After treatment, look for survivors in free text, metadata, other modalities and linked releases.
- A clean worksheet lowers identifiability but does not create the right to license the records.
Why do a job title, a town and a date identify someone together?
Each field narrows the crowd a little, and together they can narrow it to one person. A job title alone fits thousands of people, a town fits thousands more, and a date fits millions, but the people who held the title of payroll manager, in one small town, and left their job in one particular week may number exactly one.
Latanya Sweeney's analysis of 1990 U.S. Census summary data is the classic demonstration. It estimated that 87 percent of the U.S. population was likely unique on 5-digit ZIP code, gender and full date of birth, about 53 percent on place, gender and date of birth, and about 18 percent on county, gender and date of birth. None of those fields is a name, yet the combination behaves like one, and coarser geography sharply reduced the share.
HHS uses the same idea in its de-identification guidance for health data: it describes small cell counts as the number of people who share the same combination of features. Business records work the same way even when they hold no health data. The unit of review is therefore never a single column. It is every combination of columns a person outside your company could plausibly know.
Which fields in business records act as quasi-identifiers?
A quasi-identifier is any field that does not name a person on its own but can identify them once linked with other information. NIST SP 800-188 treats these indirect identifiers as the fields that remain after direct identifiers are removed and that de-identification must transform, not just mask.
In operational systems they hide in ordinary columns. The table below lists the ones that most often combine into a unique key, the place they usually live and the outside knowledge that makes them dangerous.
| Field type | Typical system location | Outside knowledge that links it | Why it combines badly |
|---|---|---|---|
| Job title or role | HR system, ticket assignee profile, CRM contact role | Professional networking profiles, company websites, press releases | Senior and specialist titles are held by one or two people per site |
| Town, branch or site | Work order address, store number, depot code | Public business listings, local news | A small branch turns a common title into a short list |
| Event dates | Hire, termination, incident, service or claim dates | Social posts, announcements, coworkers' memory | A date plus any rare attribute is often unique |
| Tenure or age band | Payroll, benefits enrollment | Public bios, graduation years | Narrow bands act like a birth year |
| Customer or employer account name | CRM account, invoice header | The account's own public footprint | A small B2B account can point to one contact |
| Rare codes and categories | Incident type, product SKU, equipment model | Recall notices, local reporting | Rare values single people out by themselves |
| Sequences of events | Status histories, visit logs, shift rosters | Anyone who saw the events happen | An ordered pattern of dates is a fingerprint even when each date is common |
What does the combination risk review worksheet record?
The worksheet is a ten-step record that one privacy reviewer fills in and a second person checks. Each row states what to record, the decision rule that applies, and the evidence you keep, so the same review can be repeated on the next release and defended later.
| Step | What you record | Decision rule | Evidence kept |
|---|---|---|---|
| 1. Release context | Recipient, permitted use, access controls, whether the data could be republished | A wider or less controlled audience requires larger minimum group sizes | Context statement signed by the release owner |
| 2. Population unit | Who the people are (employees, customers, third parties) and whether one person has many rows | Count people, not rows, whenever a person repeats | Person key used only for the count, then dropped |
| 3. Candidate fields | Every field from the table above present in the extract, including free text that repeats them | If an outsider could plausibly know it, it is a candidate | Field list with a known-by note for each |
| 4. Linkable sources | Public and semi-public sources an outsider could join on | Assume a reasonably competent intruder with ordinary resources | Short source inventory with dates checked |
| 5. Key sets | Every pair and triple of candidate fields, plus the full set | Test smaller sets first; a unique pair needs no third field | Key set list with the reason each was included |
| 6. Group sizes | Number of people per combination and the count of groups of size one and two | Flag every group below the minimum set in step 1 | Frequency output saved with the query that produced it |
| 7. Attribute diversity | For flagged and borderline groups, how varied the sensitive fields are | A group where everyone shares the same sensitive value still discloses it | Diversity counts per group |
| 8. Treatment | Generalization, suppression or removal chosen per field, and what the task loses | Prefer the treatment that keeps the field the AI task depends on | Treatment log with before and after values |
| 9. Residual review | Rerun steps 6 and 7, then check free text, file metadata and linked tables | Release only when no flagged group remains or a specialist accepts the residual | Second frequency output and residual notes |
| 10. Sign-off and refresh | Reviewer, checker, date, version, and what would force a new review | New fields, a new recipient or a linked release reopen the worksheet | Signed worksheet stored with the dataset version |
How do you count the group behind each combination?
Group the records by the key set and count the distinct people in each group; that count is the size of the equivalence class each person falls into. The open-source sdcMicro package for R, maintained for statistical disclosure control, exposes exactly this as a frequency calculation that returns the count of equal observations on the key variables for every record, along with the number of records whose count is one and the number whose count is two.
Run the count on pairs and triples, not only the full key. A reviewer who tests only title plus town plus exact date can miss that title plus town is already unique for a regional director. Keep the output for each key set, because the step 9 rerun is only meaningful against the same baseline.
Two counting errors recur. The first is counting rows instead of people: a technician with 600 work orders appears in a large group of rows while being one of only two technicians at that site. The second is counting within the extract only. If the extract is a sample of a larger population, a person who looks unique in the sample may share the combination with others outside it, and the reverse can also hold, so record whether you counted a full population or a sample.
Which release assumptions change the minimum group size, and how is that uncertainty shown?
The minimum group size is a decision, not a constant, and the assumptions behind it should be written down. HHS notes that agencies set their own small-cell policies for summary data, and no single figure transfers to every business release.
The UK Information Commissioner's Office frames identifiability around a motivated intruder test, which considers the practical steps and means reasonably likely to be used by someone determined to identify people, and its glossary accepts that the risk of disclosure is never zero. That framing gives the worksheet its assumptions. The table shows the ones that most often change the outcome and a way to show uncertainty for each instead of hiding it.
| Assumption | Why it changes the decision | How to show the uncertainty |
|---|---|---|
| Who receives the data and under what controls | A contractually bound recipient in a controlled environment is a different intruder from the public | Run the counts at two minimum sizes and report both |
| What an outsider already knows | Coworkers know dates and titles that no public source shows | List the insider knowledge you assumed away and why |
| Population versus sample | Uniqueness in a sample overstates or understates uniqueness in the population | State the sampling fraction and treat sample uniques as provisional |
| Which fields the free text repeats | A generalized title is undone if notes say senior estimator at the Bend branch | Sample-read text and record the hit rate you found |
| Whether other releases can be joined | Two safe extracts can combine into an unsafe one | Name every other release that shares a key or date field |
Which treatment lowers the risk without breaking the task?
Pick the treatment that removes the risky combination while keeping the field the downstream task actually needs. Generalization replaces a value with a broader one, suppression blanks the value only in the records that remain too rare, and removal drops the field for everyone. sdcMicro implements both global recoding and a local suppression routine that reaches k-anonymity by suppressing as few values as possible, guided by an importance ranking of the key variables, which is the same trade-off the worksheet asks you to record.
The treatment table below pairs common fields with options and states the cost to an AI task, so the decision is explicit rather than a reflex to delete.
| Field | Treatment options | What the AI task loses | Prefer when |
|---|---|---|---|
| Job title | Map to title family and level; suppress titles held by one person at a site | Fine-grained role differences | The task needs role context but not exact titles |
| Town or site | Region, county or a site size band | Local effects such as climate or route length | The model does not learn site-specific behavior |
| Absolute dates | Month or quarter; per-person offsets that keep intervals | Calendar seasonality if offsets are used | The task depends on durations and order, not dates |
| Tenure or age | Bands wide enough that each band holds several people per site | Smooth tenure effects | Tenure matters only as early, mid or long |
| Rare categories | Group into an other bucket or suppress in small groups | Rare-case examples | Rare cases are not the training target |
| Event sequences | Coarsen timestamps, drop low-volume people, or split sequences | Long-range behavior of individuals | The task works on episodes rather than careers |
Where do identifying details hide after the columns are treated?
Assume something survives and go looking for it. The Article 29 Working Party's opinion on anonymisation techniques tests any technique against three risks: whether a person can still be singled out, whether records about them can still be linked, and whether information about them can be inferred. It also states that pseudonymisation is not anonymisation, because replacing a name with a token leaves the record linkable.
Run the residual review against those three questions and against the places quasi-identifiers hide outside the columns you treated.
- Free text: ticket notes, email bodies and comments that restate the title, site or date you generalized.
- File and document metadata: author fields, device names, upload timestamps and file paths on attachments.
- Other modalities: call recordings that carry a voice, photos that show a storefront or badge, scanned forms with signatures.
- Linked tables: a second table that keeps exact dates, joined on the same record key.
- Homogeneous groups: a group of five where all five share one sensitive outcome, which discloses that outcome without singling anyone out.
- Prior and future releases: an earlier extract with different generalizations that can be intersected with this one.
Illustrative scenario: a field-service dispatch history
This scenario is illustrative; every count below is invented to show the method and does not describe a real dataset. A regional field-service company prepares five years of dispatch records for a scheduling model. The extract holds about 180,000 work orders handled by roughly 420 technicians across 14 branch towns, with technician title, branch town, hire month and work order dates.
Step 6 on title plus town plus hire month finds 61 technicians alone in their group and 38 in groups of two. Title plus town alone already isolates 22 people, mostly supervisors and specialists. The scheduling model needs job duration, sequence and broad role, not exact titles or hire months.
Step 8 maps titles to four role families, replaces towns with three service regions, converts hire month to tenure bands, and shifts work order dates by a per-technician offset that preserves intervals. The rerun leaves 4 technicians below the minimum group size the release owner set, all in a specialist role family in one region, so their role family is suppressed. The residual review then finds that about one note in 50 sampled mentions a branch name, so a branch-name redaction pass runs before release, and the worksheet records the hit rate and the fix.
When is the worksheet not enough, and how does it feed a licensing review?
Stop and bring in a privacy specialist when the worksheet cannot close the risk on its own. Typical triggers are a sensitive attribute such as health, finance or discipline inside small groups, a population where many people are publicly visible, a release that could reach the open web, or a sibling release you cannot rule out joining. For health records covered by HIPAA, HHS describes an expert determination route in which a qualified expert concludes the risk is very small and documents the methods and results.
The worksheet also answers only one of the questions a licensing review asks. Lower identifiability does not create a right to license the records: ownership, contracts with customers and employees, and privacy permissions are separate checks that counsel reviews on the specific facts.
When the records are headed for a license, SourceX runs the rights review, contracting and preparation, with de-identification carried out by SourceX through approved processing routes. Nothing reaches SourceX until the company has approved the scope and terms and signed, so a combination review like this one can be settled before a single record moves.
Frequently asked questions
What is a quasi-identifier in a dataset?
A quasi-identifier is a field that does not identify a person by itself but can once it is combined with other fields or outside information. Job titles, towns, dates, tenure bands and rare codes are common examples in business records. They are reviewed in combination because the combination, not the single field, is what becomes unique.
Is k-anonymity enough to clear a release?
Not by itself. K-anonymity means each person shares their key combination with at least k minus one others, but a group where everyone has the same sensitive value still discloses that value. Check attribute diversity within groups and look for linkage through free text, metadata and other releases.
What minimum group size should we use?
There is no universal figure. Agencies set their own small-cell policies, and the right minimum depends on who receives the data, what they already know and the controls around it. Record the minimum you chose, the assumptions behind it, and run the counts at a second value to show how sensitive the result is.
Should we count rows or people?
Count people. Operational data often has many rows per person, so a technician or customer can sit in a large group of rows while being one of very few people with that combination. Use a person key for the count and drop it from the released data.
Does removing names and IDs make the data anonymous?
No. Removing direct identifiers and replacing them with tokens is pseudonymisation, which European regulators describe as reducing linkability rather than achieving anonymisation. The remaining quasi-identifier combinations still have to be reviewed and treated.
Who should sign the worksheet?
One privacy reviewer completes it and a second person checks the counts and the residual review, with the release owner signing the context in step 1. Health, financial or employment-sensitive releases should add a privacy specialist and counsel.
Sources
- Sweeney's Census-based analysis estimated that 87 percent of the U.S. population was likely unique on 5-digit ZIP, gender and date of birth, about 53 percent on place, gender and date of birth, and about 18 percent on county, gender and date of birth. Source
- HHS describes small cell counts as the number of people corresponding to the same combination of features, notes that agencies set their own small-cell policies, and describes the expert determination method for health data. Source
- NIST SP 800-188 (2023) gives de-identification guidance that includes removing direct identifiers and transforming quasi-identifiers, which it describes as fields that can identify people when linked with other information. Source
- The ICO defines the motivated intruder test as considering the practical steps and means reasonably likely to be used by someone motivated to identify people, and notes the risk of disclosure will never be zero. Source
- The Article 29 Working Party's Opinion 05/2014 assesses anonymisation against singling out, linkability and inference, and states that pseudonymisation is not anonymisation. Source
- sdcMicro's documentation describes a frequency calculation returning per-record key-variable counts and counts of records with frequency one and two, global recoding, and local suppression that reaches k-anonymity while suppressing as few values as possible. Source
See if your company qualifies
A short company assessment. No data uploads are needed.