
Original Research
Part of Data-led PR assets
Making source data usable without exposing private information
Source data helps journalists verify a finding, but a raw file may contain personal, commercial or third-party information that should not be released. Decide what …
Sort personal from publishable in your source data
Classify every field before release. Direct identifiers point to one person on their own: name, street address, date of birth, Medicare number, employee ID. Most analytical work does not need them attached.
Indirect identifiers do not name anyone individually but narrow the field when combined. Postcode, age, occupation, date of an event and family composition can single out one person in a small group even after names are stripped.
The DataJournalism long read on de-identification notes that after AOL published millions of online search queries in 2006, journalists pieced together individual identities from search histories alone, including sensitive detail about health status and dating preferences. No name was required.
Metadata is data. The same guide observes that researchers have shown communication metadata generated by devices can be used to identify users or serve as an instrument of surveillance. Device, location and timing fields belong in the assessment.
Treat sensitive attributes separately. Health status, sexuality, ethnicity, immigration status and financial hardship carry a higher cost if a person is identified. Flag those fields and decide one by one whether they need to be in the release.
The OAIC publishes guidance on what counts as personal information under the Privacy Act, together with a De-identification Decision-Making Framework. Use both as reference points when you classify a field.
Test re-identification risk before opening up the data
Test three risks separately: linkage, small cells and accumulation. Linkage means a person appears in your data and in another dataset sharing at least one field. If your file can be joined on postcode and birth date, removing names achieves very little.
Small cells are counts with very few people behind them. A count of one in a remote town is effectively a name. Check the smallest cell in every breakdown you plan to publish, not just the headline table.
Accumulation is the slow version. Individually harmless releases, a suburb here and an age band there, can be assembled into an identification once enough of them exist.
The OECD's work on data access and sharing discusses risks including violation of the agreed terms of data re-use and the increasing loss of control by individuals and organisations over their data as it moves. These belong in your assessment, not only in your legal review.
Match the response to the risk. Options are publishing as is, aggregating or generalising, suppressing a field or a row, or restricting access to approved users under controlled conditions. Record which option you chose and why, so a later reviewer can follow the reasoning.
De-identify without stripping the data of its usefulness
There are four levers, usually combined. Suppression removes a field, a row or a value. Generalisation widens categories, such as age bands instead of single years or state instead of suburb.
Aggregation reports groups rather than individuals. Pseudonymisation replaces identifiers with consistent codes.
Decide what the analysis actually needs before cutting anything. If the question is about broad regional patterns, suburb-level precision adds risk without adding insight. Set the granularity from the question, not from what the source file happens to contain.
Order matters. Remove direct identifiers first, then test the remaining indirect identifiers against every output you plan, then adjust the levers until the risk sits at a level you can defend.
The OAIC maintains a De-identification Decision-Making Framework. It sets out a structured process rather than a single technique, alongside its De-identification and the Privacy Act guidance.
De-identification is not a one-off step. Adding records, linking a new source or redefining a category can undo it. Build it into the workflow as something that is re-run when the data changes.
Document and cite what remains so others can use it
A dataset that can be reused travels with its documentation: a data dictionary, the collection method, the reference period, the transformations applied and the de-identification steps taken.
The ABS guide to citing ABS sources lists the key elements: author, reference period or year of publication, title of the issue, a link, and the date accessed. A general web page reference follows the pattern Australian Bureau of Statistics (June 2020) Australian National Accounts: National Income, Expenditure and Product, ABS Website, accessed 11 December 2020.
Formats differ by product type. The ABS guide distinguishes web pages, time series spreadsheets, data cubes and census material, adding a bracketed product description such as [time series spreadsheet] or [data set]. A data cube example is 'Table 15: Average weekly cash earnings, Australia by sector (dollars) - original' [data set], Average Weekly Earnings, Australia.
The access date must be included, because the ABS states its content is designed to change over time and is continuously updated. Use URLs that contain the reference period rather than latest-release URLs, which move as new releases appear.
Two formatting conventions carry over to any reference list: italics for the formal title of a completed published work, and single quotation marks for titles of parts of a publication. Where the same body is both author and publisher, the publisher name can be omitted.
Provide provenance for derived data as well. Note which source tables fed your file, when they were downloaded, and what was changed. A later user can reconstruct the same result or find the updated version.
Put authorisation and social licence around the data
The synopsis of Sharing Linked Data for Health Research, by Carolyn Adams, Judy Allen and Felicity Flack, argues that consent and anonymisation are no longer adequate to protect individual interests and that they ignore collective interests. It calls for a robust authorisation and approval model.
The same work states that social licence is essential for the legitimacy of research using linked data without consent. Approval paperwork alone does not supply it; the way decisions are made and explained does.
Decision-making roles are identifiable. The book is written for data custodians, data linkers, members of ethics committees and privacy committees, and for legislators, policy makers, privacy professionals and consumer advocates.
Governments hold much of the most valuable, reliable and comprehensive health data, and those collections contain sensitive information about community members. The same information that improves care can be used to profile, harm and unjustly discriminate. That is why collection, use and release are highly regulated.
In Australia the OECD reports that, on 1 May 2018, the government committed to reforming the national data-governance framework through new data sharing and release legislation. It was made in response to the Productivity Commission's Data Availability and Use Inquiry report.
A working approval model typically contains a stated purpose, a named custodian, review by an ethics or privacy committee, and terms of reuse signed by the recipient before access is granted.
Key statistics from Australian privacy and data governance
- National data reform commitmentMay 2018 – Australian government committed to new data sharing and release legislation
- Privacy Act referenceOAIC guidance on de-identification under the Privacy Act 1988
- ABS citation standardAuthor, year, title, URL, access date – used for official statistical releases
Publish tables, figures and images without leaking identities
Check every output, not just the dataset. Totals, cross-tabulations, charts, maps and the small categories inside them can each undo a de-identified file.
Set a minimum cell size before publishing and combine or suppress anything below it. Check the totals too: a published total minus a published breakdown can reveal the value you suppressed.
Most material can go out without naming anyone. The DataJournalism de-identification guide makes this point directly: while some personal information may be required, it is likely that most stories can be published without needing to identify all individuals in a dataset.
The Australian Government Style Manual covers tables, and separately covers alt text, captions and titles for images. Use its conventions so the figures are readable to people using screen readers and to anyone scanning the page.
A caption should carry the source, the reference period and the unit of measurement. Alt text should carry the information a sighted reader takes from the figure, including the pattern or comparison it shows, rather than the words 'chart' or 'graph'.
Keep privacy protections working as data is reused
Retention and deletion are design decisions. The OAIC publishes guidance on retention and deletion of personal information, so settle how long the data and any re-identification keys are held, and what deletion means for backups and derived files.
The OECD work discusses risks that grow with reuse, including violation of the agreed terms of data re-use and the loss of control by individuals and organisations over their data. These are reasons to attach conditions that travel with the data.
The risk profile changes when a recipient adds records, links a new external source, or changes what they publish. Terms of reuse should require notification and fresh approval for those changes rather than assuming the original assessment still holds.
Keep a register of what has been released and to whom. Review it when new public datasets appear, because a combination that was safe last year may not be safe once another file covers the same population.
Standard terms cover purpose limitation, a ban on attempting re-identification, no onward sharing without approval, security requirements, and deletion or return of the data at the end of the term.
Further reading: Guide to data analytics and the Australian Privacy Principles (oaic.gov.au).



