How to De-identify or Anonymize Data in Canada

When de-identified data stops being personal information under PIPEDA, Quebec's anonymization regulation, and a worked example with postal codes and age bands.

By · · 11 min read

Illustration for How to De-identify or Anonymize Data in Canada

Canada edition. Written for Canada: PIPEDA's "serious possibility" test from the OPC and Gordon v. Canada (Health), Quebec's Regulation respecting the anonymization of personal information, and a k-anonymity example with postal code prefixes, age bands and the province. US edition

You have a customer or patient extract with names, health card numbers, postal codes and dates of birth, and a partner or researcher wants a copy. In Canada the question is whether what you hand over is still personal information. Under PIPEDA, personal information means "information about an identifiable individual", and the Office of the Privacy Commissioner of Canada (OPC) says information is about an identifiable individual where there is a serious possibility that someone could be identified from it, alone or combined with other information. The OPC adds that de-identified information does not qualify as anonymous if there is a serious possibility of linking it back to a person. Removing names is rarely enough to clear that bar.

This edition explains the federal test, what Quebec now requires, and works through a small table with the Find Duplicate Entries tool. It is practical guidance, not legal advice; health data, public releases and Quebec organizations in particular warrant a privacy lawyer or your privacy officer.

Federal: PIPEDA and the serious possibility test

The OPC's Interpretation Bulletin on personal information summarizes the federal position:

  • PIPEDA section 2(1) defines personal information as information about an identifiable individual.
  • Following Gordon v. Canada (Health), 2008 FC 258, information is about an identifiable individual where there is a serious possibility that the person could be identified through it, alone or in combination with other information.
  • De-identified personal information does not qualify as anonymous if there is a serious possibility of linking it back; the OPC cites a case where a psychologist's "anonymized" peer review notes were still the patient's personal information.

In practice that means two separate words. De-identified data has had direct identifiers removed or replaced, but someone (often you) could still link it back; it remains personal information. Anonymous data carries no serious possibility of identifying anyone, given the other information available; only then does it fall outside PIPEDA. The OPC notes its interpretations are not binding and evolve with new findings and court decisions.

Why a hashed identifier is still personal information

Hashing a health card number or email gives the same output every time. Anyone with a list of candidate numbers can hash them and match, and the organization that did the hashing can always re-link. Under the serious possibility test that is plainly still personal information: a hash is a way to link datasets without exchanging the raw identifier, not a way to make them anonymous.

Quebec: anonymization is regulated

Quebec goes further than the federal law. Its Commission d'accès à l'information (CAI) announced that the government's Regulation respecting the anonymization of personal information came into force on 30 May 2024, with article 9 taking effect on 1 January 2025. If you hold personal information about people in Quebec, read the regulation itself before calling any dataset anonymized: it is a binding regulation, not general guidance.

Worked example: postal code prefixes and age bands

Ten rows from an Ottawa clinic, with names and health card numbers already deleted. The quasi-identifiers are postal code, year of birth and sex.

Postal codeBornSex
K1N 6N51991F
K1N 9B41993F
K2P 1L41979F
K1N 7A21990M
K2P 2G31994M
K1N 5T81968M
K2P 0B61974M
K1N 8L11961F
K2P 1J91992F
K1N 6C31977M

Put each row's quasi-identifiers on one line, paste them into Find Duplicate Entries, turn off "Only duplicates" and keep "Show count in result" on. Every line appears once: k = 1.

First pass: keep the forward sortation area (FSA, the first three characters) and use ten-year bands. Ages in 2026 run from 32 to 65, so use 30-39, 40-49 and 50 and over:

  • K1N, 30-39, F appears 2 times; the other eight combinations appear once each. Still k = 1.

Second try: drop the postal code to the province. ON with ten-year bands and sex gives groups of 3, 1, 2, 2, 1 and 1. Still k = 1: a woman aged 40-49, a man aged 40-49 and a woman aged 50 and over remain unique. Location was not the only problem; the narrow age bands were.

Two versions that reach k = 2:

VersionGroupsSmallest group
Postal prefix + under 40 / 40 and over, sex dropped3, 2, 2, 32
Province + under 40 / 40 and over + sex3, 2, 2, 32

Both keep all ten rows. Which to choose depends on the analysis: a study of neighbourhood differences keeps the prefix, one of sex differences keeps sex. Either way, a group of two is thin for a public release, so the context of the release matters as much as the numbers. The same loop works on real tables: measure, coarsen the column that splits groups most, measure again.

Check the sensitive column too. k only protects who someone is. If everyone in a group shared the same diagnosis, knowing a person is in the group would reveal it; widen the groups or suppress the cell where that happens.

Practical checklist for a Canadian release

  1. Remove direct identifiers: name, address, health card number, SIN, email, phone, account numbers.
  2. List the quasi-identifiers an outsider could know: postal code, birth date, sex, occupation, dates of visits, rare conditions.
  3. Coarsen them: postal prefix or province instead of the full code, age bands instead of birth dates, month or year instead of exact dates.
  4. Measure k with the duplicate counter and fix the smallest groups by generalizing or suppressing.
  5. Ask the serious possibility question for the actual recipients and what else they can access, not in the abstract.
  6. For Quebec data, check the anonymization regulation's conditions before treating the result as anonymized.

The 2025 Guide to Getting Started with Anonymisation, published in August 2025 by privacy and data protection authorities from Australia, Hong Kong, Macao, Canada, Japan, South Korea, New Zealand and Singapore, sets out a five-step approach that follows the same order.

Tools for this job

For why hashing is one-way but not secret, read hash algorithms compared.

Frequently asked questions

Is de-identified data still personal information under PIPEDA?

Usually, yes. The OPC says de-identified personal information does not qualify as anonymous if there is a serious possibility of linking it back to an identifiable individual, alone or with other information. Only data with no such serious possibility falls outside PIPEDA.

What is the serious possibility test?

It comes from Gordon v. Canada (Health), 2008 FC 258: information is about an identifiable individual where there is a serious possibility that someone could be identified through it, alone or combined with other available information. The OPC applies it to decide whether information is personal information.

Does Quebec have its own rules on anonymized data?

Yes. The Regulation respecting the anonymization of personal information came into force on 30 May 2024, with article 9 from 1 January 2025, according to the CAI. Read it before treating Quebec personal information as anonymized.

Is a hashed email address anonymous in Canada?

No. The same email always produces the same hash, so anyone with a list of addresses can match it, and the organization that hashed it can re-link it. That is a serious possibility of identification, so the hash remains personal information.

Sources and further reading

Browse all tools