Skip to content

Privacy and preparation

Deidentification linkage utility test: remove identities, keep links

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Identity removal can keep useful links when every identifier is replaced by a keyed token that is consistent across systems and cannot be reversed without a separately held key. Prove it with a paired test: score whether joins, chains and order survive, and score how much re-identification risk the preserved links add. Release only a method that passes both, and narrow scope when the link itself identifies someone.

Redaction

Deidentification linkage

This page covers deidentification linkage utility test: remove identities, keep links.Illustrative

Key takeaways

  • Judge a de-identification method on two scores at once: links preserved and residual re-identification risk.
  • Keyed, consistent tokens keep joins working; unsalted hashes and reversible encryption keep joins but leave a path back to the person.
  • Preserved links make each person's profile richer, which can make them unique even when no direct identifier is left.
  • Run five utility checks and five risk checks on the same candidate output, and release only when both sides pass.
  • When the link itself is what identifies someone, aggregate, suppress or drop it rather than tokenize it.
  • Pseudonymized records with a key held somewhere usually remain personal data; counsel decides the legal status.

Identity removal preserves links when it replaces each identifier with a token that is the same everywhere that identifier appears, in every system and every extract, and that cannot be turned back into the identifier without a key the recipient never sees. A customer ID in the CRM, the same ID on an order and the same ID on an invoice all become one token, so the chain still joins.

Keeping links is only half the decision. Every link that survives also lets someone assemble more facts about one person or account, so the question is never just whether the joins work. NIST SP 800-188 frames de-identification as limiting disclosure risk while still allowing meaningful analysis, and that is the balance this test makes explicit.

The test below scores a candidate method on both sides using the same output. A method that keeps every join but leaves people unique fails. A method that removes all risk but breaks the chains a buyer needs also fails, because it destroys the reason the records were worth preparing.

Keeping links raises risk because each join adds attributes to the same token, and a richer profile is more likely to be unique. A single support ticket with a city and a product line may describe many customers; that same token joined to twelve orders, three invoices and a renewal date may describe exactly one.

Published research shows how few linked points it can take. In a study of three months of credit card records for 1.1 million people, de Montjoye and colleagues reported that four spatiotemporal points were enough to uniquely identify about 90% of individuals. The exact share depends on the data, but the direction is clear: linked event histories identify people quickly.

Tokens also do not change the legal picture on their own. The European Data Protection Board's Guidelines 01/2025, adopted in January 2025 in a version for public consultation, say pseudonymized data that can be attributed to a person by using additional information, such as lookup tables or cryptographic keys, is still personal data, and that the additional information must be kept separate under technical and organizational measures. Treat a tokenized, linked dataset as personal data unless counsel concludes otherwise on the facts.

Context changes the answer as well. The same linked records sent to one recipient under a contract that bans re-identification attempts carry less practical risk than the same records shared widely, so the risk checks below should be scored against the release you actually plan, not a generic one.

Only methods that map the same input to the same output keep links, and they differ in how easily that mapping can be reversed. Tooling defaults matter here. Microsoft Presidio's hash operator, for example, uses a random salt by default, so the same value gets different hashes across calls; for referential integrity the user must supply a consistent salt and manage it securely. Its encrypt operator is reversible by anyone with the key.

For U.S. health records, HHS guidance on the HIPAA Privacy Rule allows a covered entity to keep a re-identification code if the code is not derived from information about the individual and the mechanism is not disclosed, and says keyed hash functions are not ruled out under expert determination when the keys are not disclosed, including to recipients. The same design logic works for business records generally.

Whichever keyed method you choose, the key is now the most sensitive thing you hold. Store it apart from the prepared data, limit access to the few people who run preparation, log every use, and decide in advance whether it will be destroyed after delivery or kept so later releases can use the same tokens. That decision has consequences: a destroyed key means a later refresh cannot join to the earlier release, while a kept key is a standing path back to the person that must be protected for as long as it exists.

Which identity-removal methods keep links, and what does each cost?
MethodLinks keptMain residual riskWhen it fits
Drop the identifierNone across systemsLow from the key; profile risk remains in other fieldsSingle-record tasks where joins add nothing
Random token per recordNone; each row stands aloneLow from the keyReleases where linkage is not needed
Unsalted hash of the identifierAll joinsHigh: anyone can hash candidate values and match themNot recommended for release
Keyed hash or lookup table, one key held by the preparerAll joins within the releaseProfile uniqueness from preserved links; key compromiseDefault for linked releases
Keyed hash with a separate key per recipientAll joins within each recipient's copySame as above, but recipients cannot join their copiesMultiple licensees of overlapping records
Reversible encryption of the identifierAll joinsAnyone with the key recovers the identifierInternal processing only, never in a release

What should the linkage utility and re-identification tradeoff test include?

The test should include five utility checks and five risk checks, all run on the same candidate output and a fixed sample, with a pass mark set before you look at results. The ICO's motivated intruder test is the reference point for the risk side: could a reasonably competent person, using means reasonably likely to be available, identify someone. The ICO also says releases to the world at large need more robust techniques than releases to specific known recipients, so set risk marks per release context.

Statistical disclosure tools help with the counting. The open-source sdcMicro package, for example, computes individual and global disclosure risk on chosen key variables and offers local suppression to reach k-anonymity. The pass marks below are a starting point for your own policy, not a regulatory standard.

Run the checks in a fixed order. Start with R1 and R2, because a leaked identifier or a reversible token fails the method outright and makes the remaining checks pointless. Then run U1 to U4 to confirm the joins survive, R3 to R5 to measure what those joins expose, and U5 last, after any treatment from the risk checks, so the task signal reflects the version you would actually release. Record the sample, the date, the method version and who ran each check, and have someone other than the person who built the method review the results.

What should the linkage utility and re-identification tradeoff test include?
CheckQuestionHow to measurePass when
U1 Join countsDo joins return the same pairs as before?Count joined pairs per key before and after tokenizingCounts match exactly
U2 No false mergesDid two entities collapse into one token?Check token collisions against the source key listZero collisions
U3 Chain completenessDo full case chains survive end to end?Share of cases whose full chain resolves after transformationSame share as before
U4 Order and timingIs event order within a case preserved?Compare event sequences per case before and afterSequences match; date shifts are consistent per case
U5 Task signalDoes the intended task still work?Run a small baseline task on original and transformed samplesNo material drop the reviewer cannot explain
R1 Direct identifiersAre names, emails, phone numbers and IDs gone from fields and text?Detector recall on a hand-labeled sample, including free textNo missed identifiers in the sample, or misses fixed and retested
R2 Token reversibilityCan a token be recomputed without the key?Try hashing known identifier values; confirm key storage and accessNo match without the key; key held apart from the data
R3 Linked-profile uniquenessDo joined profiles make accounts or people unique?Count unique profiles on quasi-identifiers after joining all linked recordsUnique profiles below your policy mark, or treated
R4 Cross-release linkabilityCan tokens join to another release or recipient's copy?Compare token sets across releases and recipientsOnly where the agreement allows it
R5 Leaked keys in textDo order numbers or account IDs survive in notes or attachments?Pattern search for original key formats in free textNone found in the sample

Worked example: support tickets linked to orders and invoices

Illustrative only: the company, fields and outcomes below are invented to show how the test is applied. A B2B parts distributor plans to license support tickets linked to orders and invoices so a buyer can train a model to predict which tickets lead to credits. The links are the value: without them, a ticket cannot be tied to the credit that followed.

The first candidate replaced customer IDs with an unsalted SHA-256 hash. It passed every utility check, then failed R2, because a reviewer hashed customer numbers from an old price list and matched them. The second candidate used a keyed hash with the key held by the preparer and a separate key per recipient. It passed U1 to U4 and R1, R2 and R4.

It then failed R3. Joining every order to the token made a handful of very large accounts unique on region, product mix and order volume. The fix was to generalize order volume into bands and to suppress region for accounts in small categories, then rerun the test. U5 showed the credit-prediction task still worked on the transformed sample, so the method passed with both changes documented.

The record of that run is what a reviewer or buyer reads later: the two rejected candidates and why, the treatment applied for R3, the accounts affected, and the task check before and after. Keep it with the release so the next refresh can start from the passing method instead of repeating the first two attempts.

Sometimes the link is the identifier, and no token can fix that. If one account is the only customer that ever bought a product line, or one technician is the only person who worked in a region, any preserved link to that product or region points straight at them.

In that case, tokenizing harder does not help. The options are to aggregate the linked records into counts or bands, to suppress the link for that small group, or to drop the group from the release and document why. Record the exception in the test results so the reviewer sees which links were removed on purpose and what task signal was lost.

Where can identifiers hide after the joins pass?

Even after a method passes the checks above, identifying combinations can hide in places a field-level review misses. Plan a short second pass on each of these before release.

At a glance

Where can identifiers hide after the joins pass

Each item in the list below needs its own detector or review step, and each should be added to R1 or R5 for the next run once found. Free text and attachments deserve particular attention in business records, because people often type account details into notes that structured exports never captured.

  • Free text: ticket bodies, call notes and email threads often repeat order numbers, names and addresses that the structured fields no longer show.
  • Attachments: PDFs of invoices, signed forms and screenshots carry letterheads, signatures and account numbers.
  • Timestamps: exact times on linked events can be matched against a customer's own records or public posts.
  • Rare values: unusual product codes, job titles or locations that appear only a few times across all linked records.
  • Audio and images: call recordings and photos can identify voices, faces, vehicles or buildings even when text is clean.

When does a failing risk score mean the release is too rich for its context?

Stop or narrow the scope when the test keeps failing on the risk side after two rounds of treatment, because repeated patching usually means the release is too rich for its context. Bring in a privacy specialist before release in any of these situations; this page is general information, not legal advice.

  • R3 still fails after generalizing and suppressing, which points to dropping fields or links instead.
  • The records include health, financial account, children's or other sensitive information.
  • The same records will go to several recipients who might combine their copies.
  • The release will be public or widely shared rather than sent to one known recipient under contract.
  • Customer contracts or privacy notices limit how the records may be used or shared.

The link decisions above feed straight into a SourceX scope, because the de-identification itself is carried out by SourceX through approved processing routes and the company signs off on scope and terms before a single record is delivered. A company that already knows which links carry its data's value, and which links could single out a customer or employee, can make that scope decision faster and with fewer surprises.

Whether a company's records are a fit at all is settled first through a short assessment that asks for no files.

Frequently asked questions

What is a deidentification linkage utility test?

It is a paired test that checks two things on the same de-identified output: whether the joins between records still work, and how much re-identification risk those preserved joins add. A method passes only when both sides meet marks set in advance. It is a release check, not a legal determination.

Is hashing customer IDs enough to keep links and remove identity?

Not if the hash is unsalted or the salt is known. Anyone can hash candidate customer numbers and match them to the tokens. A keyed hash or lookup table, with the key held apart from the data and never given to recipients, keeps the links without that shortcut.

Are pseudonymized, linked records still personal data?

Often yes. European regulators treat data that can be attributed to a person with additional information, such as a key or lookup table, as personal data. Counsel should decide the legal status for each release and jurisdiction.

Should each buyer get a different token key?

Usually, when several recipients license overlapping records. Separate keys stop recipients from joining their copies into a richer profile. Use a shared key only when the agreement allows recipients to combine releases.

How large should the test sample be?

Large enough to include rare accounts, long case chains and free-text fields, not just typical records. A practical start is a random sample plus every record from small categories. Rerun the test whenever fields, sources or the method change.

Sources

  • HHS guidance allows a re-identification code that is not derived from information about the individual and whose mechanism is not disclosed, and says keyed hash functions are not ruled out under expert determination when keys are not disclosed to recipients. Source
  • Presidio's hash operator uses a random salt by default; referential integrity needs a consistent, securely managed salt; its encrypt operator is reversible with the key. Source
  • The EDPB says pseudonymized data attributable to a person with additional information such as lookup tables or keys remains personal data, and that information must be kept separate. Source
  • The ICO uses the motivated intruder test to assess identifiability and says releases to the world at large need more robust techniques than releases to known recipients. Source
  • NIST SP 800-188 describes de-identification as preventing or limiting disclosure risk while still allowing meaningful statistical analysis. Source
  • A study of three months of credit card records for 1.1 million people found four spatiotemporal points uniquely identified about 90% of individuals. Source
  • sdcMicro computes individual and global disclosure risk for chosen key variables and provides local suppression to reach k-anonymity. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies