How do national statistics offices use synthetic microdata for public releases?

National statistics offices use synthetic microdata for public releases by generating artificial records that replicate the statistical properties of real survey and census data without containing information about any actual respondent. The core purpose is to make detailed, individual-level data broadly accessible while honoring legal confidentiality obligations to the people who provided their information. The sections below walk through how this works in practice, who is doing it, and where the approach still falls short.

What types of microdata do national statistics offices typically release?

National statistics offices typically release two broad categories of data: aggregate tables that summarize information for entire groups, and microdata files containing individual-level records for a sample of respondents. Microdata give researchers the flexibility to construct their own analyses and cross-tabulations that pre-built tables cannot provide, making them especially valuable for social science and policy research.

Within the microdata category, several distinct types are common. Household survey data and population census records are the most widely recognized, but offices also release longitudinal panel data that track the same individuals over time, business and establishment records, and administrative data drawn from tax or benefit systems. The U.S. Census Bureau’s American Community Survey Public Use Microdata Sample, for instance, allows users to build custom estimates not available through standard published tables.

Access tiers vary considerably. Some offices publish fully open files that anyone can download, while others restrict access to registered researchers or require analysts to work inside secure environments. The UK’s Office for National Statistics, for example, offers a publicly available teaching sample at one end of the spectrum and secure microdata accessible only to accredited researchers at the other. As of a 2022 study, roughly 100 countries had made some form of microdata available to the scientific community, with around 30 extending access to the general public.

Why can’t national statistics offices release raw microdata directly?

National statistics offices cannot release raw microdata directly because doing so would breach the confidentiality commitments they make to every survey respondent. These obligations are both legal and ethical: in most jurisdictions, statistical agencies are bound by statute to protect individual responses, and violating that trust would undermine public willingness to participate in future surveys, degrading the quality of official statistics over time.

The practical risk is greater than it might appear. Removing obvious identifiers such as names and addresses is not enough, because some individuals have rare combinations of characteristics that make them identifiable even in anonymized files. Research has shown that a small set of variables such as age, sex, and postal code can be sufficient to uniquely identify a large share of a population in publicly available data. That risk has grown sharply as computing costs have fallen and private-sector databases have expanded, giving potential attackers more external data to cross-reference against released files.

Business microdata presents an even sharper challenge. Distributions of establishment-level data tend to be highly skewed, meaning that a single large employer in a given industry or region can be trivially identifiable. Most national statistical agencies therefore do not release raw business microdata at all, relying instead on heavily restricted access arrangements or, increasingly, synthetic alternatives.

How do national statistics offices generate synthetic microdata?

National statistics offices generate synthetic microdata by fitting statistical models to the original, confidential data and then drawing new records from those models. The resulting dataset contains no real respondent records; instead, it consists of artificial observations that reproduce the distributional patterns and relationships present in the source data without being traceable to any individual.

The foundational approach, developed in the early 1990s, applies multiple imputation techniques to produce either fully synthetic datasets, where every variable is replaced with modeled values, or partially synthetic datasets, where only the most sensitive variables are synthesized while others remain as observed. Fully synthetic data offers stronger privacy protection; partially synthetic data tends to preserve more analytical utility because fewer variables are altered.

In practice, agencies have used a range of generation methods. Classification and Regression Trees have been a workhorse technique for tabular census microdata because they handle the predominantly categorical variables found in survey data well and perform competitively even against more complex approaches. Research into deep learning methods such as generative adversarial networks is ongoing, but as of 2026 no national statistics office has publicly released synthetic microdata produced by those techniques. The UNECE Starter Guide on synthetic data, published in 2023 with contributions from statistical agencies across eight countries, provides a practical reference for offices beginning to develop these capabilities.

What disclosure risk controls are applied to synthetic microdata before release?

Before releasing synthetic microdata, national statistics offices apply statistical disclosure control measures and conduct formal disclosure risk assessments to verify that the published data cannot be used to re-identify any original respondent. Generation alone does not guarantee safety; the same fidelity that makes synthetic data analytically useful can also preserve relationships that enable inference attacks against unusual records.

Risk assessment typically involves re-identification studies that search for unique combinations of variables in the synthetic file and attempt to match those combinations against external sources such as voter registration records or commercial databases. Offices may also evaluate privacy guarantees more formally using frameworks like differential privacy, which provides a mathematical bound on how much any individual record can influence the output of a query against the data.

Beyond the generation and assessment stages, agencies apply additional protective measures to the synthetic files themselves. These include collapsing or top-coding variables that carry high identification risk, restricting geographic detail, and in some cases swapping records across similar units. The U.S. National Center for Science and Engineering Statistics, for example, conducted a full disclosure review before releasing its synthetic longitudinal microdata product for the Survey of Doctorate Recipients in early 2026. The key practical challenge is that outliers and rare subgroups face substantially higher re-identification risk than typical records, even after synthesis, so risk controls must be calibrated with those edge cases in mind.

Which national statistics offices have already released synthetic microdata publicly?

Several national statistics offices have already released synthetic public use microdata, with the U.S. Census Bureau leading the way. The Census Bureau’s Center for Economic Studies released the Synthetic Longitudinal Business Database in 2010, making it the first publicly available synthetic business microdata product in the United States. It covered more than two decades of establishment-level records and was notable precisely because raw business microdata had never previously been releasable at all.

The Census Bureau has also released the SIPP Synthetic Beta, which links household survey records to administrative tax and Social Security data. Access to this product was suspended for a period but restored in mid-2024. More recently, the National Center for Science and Engineering Statistics released a synthetic longitudinal product for the Survey of Doctorate Recipients, designated as an experimental statistical product.

Outside the United States, the UK’s Office for National Statistics published a public microdata teaching sample drawn from the 2021 Census for England and Wales in September 2023, alongside safeguarded microdata made available through the UK Data Service. Northern Ireland’s statistics agency followed with its own 2021 Census teaching sample in early 2025. The Australian Bureau of Statistics has explored synthetic data as a dissemination tool, particularly for business microdata in concentrated industries, though its work remains at a research and exploratory stage rather than a confirmed public release. Statistics Netherlands has contributed to international guidance on the topic and holds large volumes of personal data for which synthetic methods are under active consideration.

How accurate are synthetic microdata compared to the original census records?

Synthetic microdata are designed to reproduce the statistical properties of the original census records, but their accuracy is not uniform and depends heavily on what is being measured. Well-generated synthetic data can closely match univariate distributions, pairwise correlations, and aggregate-level patterns in the source data. However, accuracy degrades for analyses involving rare subgroups, complex multivariate relationships, or questions that were not anticipated when the synthesis models were specified.

Evaluation frameworks for synthetic microdata generally examine three dimensions: fidelity, which measures how closely the synthetic data mirrors the original’s distributional characteristics; utility, which assesses whether analyses run on the synthetic data produce results comparable to those from the real data; and privacy, which verifies that the data do not expose individual respondents. A 2025 comparative study that evaluated multiple synthesis tools against real census microdata from four countries found that performance varies meaningfully depending on both the tool’s parameter settings and the structure of the underlying dataset.

A significant limitation is that synthetic data utility is largely bounded by the relationships the generating model was trained to preserve. Analyses on subdomains or variable combinations not pre-selected during synthesis may produce misleading results. This is why, for products like the SIPP Synthetic Beta, the Census Bureau operated a validation service at Cornell University that allowed researchers to check findings from the synthetic data against the confidential original before publication. Synthetic microdata are best understood as a tool for exploratory analysis and hypothesis generation rather than a direct substitute for the underlying confidential records in high-stakes research.

What are the remaining challenges for synthetic microdata in official statistics?

The most persistent challenge for synthetic microdata in official statistics is the fundamental trade-off between utility and privacy protection. Making synthetic data more analytically accurate requires preserving more of the original data’s structure, but that same fidelity increases the risk that unusual records could be re-identified. There is no universal solution to this tension; agencies must calibrate their approach to the specific use case and the sensitivity of the underlying data.

Several practical obstacles compound this core difficulty. Census and survey microdata are dominated by categorical variables, which standard statistical synthesis methods handle less elegantly than continuous data. Demonstrating to skeptical users that valid inferences can be drawn from synthetic files requires sustained methodological work, and discrepancies between synthetic and original results can arise from differences in what information was available to the model builder versus what the analyst receives. A March 2025 peer-reviewed critique in PNAS argued that synthetic census microdata are poorly suited to studying unanticipated relationships and that the utility limitations preclude many types of policy analysis.

Institutional and resource constraints add further friction. Disclosure review processes for validation services, where researchers check synthetic-data results against the confidential original, are time-consuming and expensive. The Census Bureau’s SIPP Synthetic Beta validation service closed in 2022 due to funding shortfalls before being restored, illustrating how sustainability can be as much of a barrier as technical feasibility. The Census Bureau’s broader plan to replace the American Community Survey public use microdata sample with fully synthetic data has also sparked significant debate about whether the privacy gains justify the analytical costs to the research community.

Despite these challenges, momentum is clearly building. International coordination through bodies like the UNECE is helping agencies share methods and learn from one another’s experiences. At BlueGen AI, we work through exactly these utility-privacy trade-offs every day, and we would be glad to show you how our platform approaches them. If you are curious about how synthetic data generation works in practice, request a demo and we can walk you through it.

window.addEventListener(‘scroll’, function() { var bar = document.getElementById(‘sticky-cta’); if (window.scrollY > 600) { bar.style.display = ‘block’; } else { bar.style.display = ‘none’; } });