1. Purpose and limits of synthetic data
Question: What are synthetic data for, and what are they not for?
Rationale
Synthetic microdata are records generated by a model fitted to the real microdata, so that they look like the real file (the same variables, similar distributions and relationships) while containing no real person's record. They answer a need every microdata team has. Users want to write and test code before applying for the licensed file. Students need a realistic file. Researchers want to develop a method. Developers need test data for a pipeline. Partners cannot receive the real data. They do not answer the need for estimates: a statistic computed from a synthetic file is a statistic about the synthesizer, and the organization publishes nothing from one.
Two kinds exist. A fully synthetic file replaces every value; a partially synthetic file replaces the sensitive variables and keeps the rest. The first protects more and preserves less; the second is a disclosure control method for a public use file, documented as such. This guide concerns fully synthetic files for sharing, with the methods and measures that make them useful and safe, and it keeps the question of estimation out of scope by design.
Target state
- The organization has a written statement of what synthetic files are for and what they are not for, and every synthetic file carries it.
- The decision to produce a synthetic file is taken from demand (access requests, teaching, development needs) and recorded with the intended uses.
- Synthetic files are labelled as synthetic in their name, their record, and their data, and are never a source of published statistics.
Maturity levels
| Level | Steps |
|---|---|
| Foundational | Write the purpose statement; produce one synthetic file for the most requested survey for code development and teaching. |
| AI-ready | Produce synthetic files for the main surveys with measured utility and risk, documented records, and a place in the access workflow. |
| AI-native | Produce synthetic files at each release as a pipeline stage, with relational structure where the survey has it, and track their use. |
The example organization's labour force survey is its most requested microdata file. Applicants wait weeks for the licensed file and then discover that their code does not run on it. A synthetic file of the same shape, downloadable at once, lets them write and test their code first. The recipes of this guide produce, measure, document, and release that file; a 300-record extract with six variables carries the scripts.
Recipes
1.1 Write the purpose statement
- Level
- Foundational
- Skills
- Head of the microdata unit; methodologist
- Indicative time
- Half a day
Result: the organization has a short, approved, published statement of what its synthetic files are for and what they must not be used for.
- List the intended uses: writing and testing analysis code before applying for the real file; teaching; developing methods; testing software and pipelines; demonstrations.
- List the prohibited uses and the limits: no estimation or publication of statistics from a synthetic file; no use as a substitute for the real microdata; no guarantee that relationships hold beyond those the utility report lists; no reliability for small groups.
- State the labelling rule: every synthetic file is marked as synthetic in its name, in its catalog record, and inside the data itself.
- Publish the statement with the access policy of the microdata cookbook, and repeat it in every synthesis record (chapter 6).
Worked example. The example organization's statement fits on one
page. Its intended uses are: writing and testing code before an
application for the real file; teaching survey analysis; developing and
comparing methods; testing data pipelines; and demonstrations to
partners. Its prohibited uses are: estimating or publishing any
statistic; using the file in place of the real microdata in any report;
and relying on any relationship the utility report does not list. The
labelling rule says that every synthetic file name begins with
"SYNTHETIC_", that the catalog title begins with "Synthetic", and that
the data file carries a column synthetic set to 1 in every row. The
statement is approved by the head of the microdata unit and published
as an annex to the access policy.
What this does not do. A statement sets expectations. Users will still compute things from a synthetic file; the label, the record, and the utility report are what tell them how far to trust the result.
1.2 Decide which files to synthesize, from demand
- Level
- Foundational
- Skills
- Microdata unit
- Indicative time
- Two hours
Result: the microdata unit has an ordered list of the surveys to synthesize, with the demand behind each and the uses its synthetic file would serve.
- Count, for each survey, the access requests over the last two years, the requests from teachers, and the internal needs (pipeline testing, training).
- Order the surveys by demand. Start with one survey whose real file has a queue of applicants.
- For each survey, write down the intended uses and the variables those uses need. A teaching file needs the shape of the whole questionnaire; a file for testing code needs the variables and types users will write code against.
- Record the list in the microdata unit's work plan.
Worked example. Over two years the example organization received 140 access requests for the labour force survey, 60 for the household budget survey, and 25 for the census sample, and its training unit asked twice for a teaching file of the labour force survey. The labour force survey comes first. Its intended uses are code testing by applicants, who need every variable with its real type and codes, and teaching, which needs the demographic and employment variables and the weights. The household budget survey comes second, for method development on consumption data, and the census sample last. The list goes into the unit's work plan with the demand figures beside each survey.
What this does not do. Demand decides the order of work. Whether a survey's synthetic file is good enough to release is decided by the measures of chapters 3 and 4.
1.3 Separate synthesis for sharing from disclosure control
- Level
- Foundational
- Skills
- Methodologist; disclosure control unit
- Indicative time
- Two hours
Result: the policy and the practice draw a clear line between fully synthetic files made for sharing and partially synthetic methods used to protect a public use file.
The two are easy to confuse because the methods overlap. A fully synthetic file replaces every value and is never used for estimates. A partially synthetic file replaces only sensitive variables, keeps the rest, and is a form of disclosure control applied to a public use file that users do analyse.
- State in the policy that the files of this guide are fully synthetic and are not public use files: nobody estimates from them, and their risk assessment (chapter 4) concerns copying and inference, never the identification of a respondent in the file.
- Where the organization uses partial synthesis as disclosure control, document it in the public use file's disclosure control record (access chapter of the microdata cookbook), with utility measured for the estimates users will make.
- Keep the two products apart in the catalog, with different labels, records, and terms of use.
Worked example. The example organization's access policy gains a paragraph on the two products. It says that synthetic files are fully synthetic, that no value in them is a real respondent's, and that no statistic is estimated from them. It says that they are documented in a synthesis record and labelled synthetic. It then says that public use files are real data with disclosure control applied, which may include the replacement of some sensitive variables by synthesized values where the file's record states it, and that estimates may be made from them with the documented caveats. In the catalog, the synthetic labour force file is a separate study with the title "Synthetic version of the Labour Force Survey 2025 Q2", and the public use file of the same survey keeps its own record with its disclosure control section.
What this does not do. The line keeps users and reviewers clear about what each file is. A team that masters one method learns the other; what differs is the documentation and the use.
Implementation options
- Kinds: fully synthetic for sharing (this guide); partially synthetic as a disclosure control method (documented with the public use file).
- Statement: a page with intended uses, prohibited uses, limits, and the labelling rule, published with the access policy.
- Prioritization: access requests, teaching and development needs, variables required by the uses.
World Bank examples
- The program's synthetic data workstream develops REaLTabFormer, a transformer-based generator for tabular and relational data, placed under disclosure control (GSBPM 6.4) in the GSBPM mapping.
- The Microdata Library access tiers and the microdata cookbook's access chapter set the workflow a synthetic file joins.
Resources
- UNECE, Synthetic Data for Official Statistics: A Starter Guide: the reference introduction for statistical organizations, with use cases and cautions.
- REaLTabFormer: the program's generator for tabular and relational synthetic data (MIT).
- synthpop: the R package for sequential synthesis with CART and parametric methods, with its own utility and disclosure measures.
- Access chapter of the microdata cookbook: the access policy and tiers the synthetic file sits beside.
Common mistakes
- Synthetic files used for estimates. The statement and the label exist to prevent it, and the organization never publishes one.
- Synthesis as a replacement for disclosure control. Different product, different record.
- Files synthesized because the tool exists. Demand orders the work.
- Unlabelled files. A synthetic file that looks real is used as real.
Verification
- Statement. The statement is approved, published, and repeated in every synthesis record.
- Prioritization. The list exists with demand figures and intended uses.
- Separation. Synthetic files and public use files have distinct labels, records, and terms in the catalog.
Checklist
Ticks are saved in this browser only.