Attention: You are using an outdated browser, device or you do not have the latest version of JavaScript downloaded and so this website may not work as expected. Please download the latest software or switch device to avoid further issues.
| 18 Jul 2026 | |
| Data Policy |
What is Disclosure Avoidance Order (DAO) 216-26, and what does it do?
DAO 216-26 is a Department of Commerce departmental order, signed by the Secretary of Commerce and effective June 4, 2026. It establishes a preferred order among the techniques that the Census Bureau and the Bureau of Economic Analysis (BEA) use to protect the confidentiality of the statistics they publish. Specifically, it makes coarsening the preferred approach, permits suppression only as a last resort, and states that noise infusion "shall not be used for any statistical product." It also directs that each product receive an individualized legal review, conducted by the Census Bureau Director or the BEA Director together with the Office of the General Counsel, against its own statutory requirements before release.
What do "coarsening," "suppression," and "noise infusion" mean?
The order defines three categories of disclosure-avoidance methods:
Which agencies and products does it cover?
The order applies to statistical products disseminated by the Census Bureau and BEA. It expressly does not cover data shared among the Census Bureau, BEA, and the Bureau of Labor Statistics under Title 13 or 44 U.S.C. § 3576, nor reimbursable products that do not use data protected by the relevant confidentiality statutes. The order also notes that the confidentiality statutes governing the Census Bureau (Title 13) and BEA (22 U.S.C. § 3104(c)) differ in the precise level of protection they require. The impact on individual reimbursable products will depend on whether they use data deemed subject to Title 13 or other relevant confidentiality protections, and will need to be assessed on a product-by-product basis.
What is the Department of Commerce's stated rationale?
Commerce has described the order's primary intent as prioritizing the accuracy and objectivity of published data for users, while continuing to meet confidentiality obligations. In this account, methods that knowingly publish altered values are at odds with the goal of providing the public with figures they can take at face value. The order resolves that tension by prioritizing accuracy and reliability as a policy matter, and steering toward methods the public readily understands: grouping, rounding, and—where necessary—withholding. The order frames itself as advancing accuracy, confidentiality, objectivity, and relevance. Whether the order achieves that goal in practice and at what cost in granularity for small areas and small populations is the empirical question that implementation and independent analysis will answer.
Why did the Census Bureau adopt noise infusion and differential privacy in the first place?
According to the Belfer Center for Science and International Affairs at the Harvard Kennedy School, differential privacy is a safeguard that protects an individual’s data privacy. It allows for the collection and publication of data patterns and trends, while protecting the privacy of individuals captured in a dataset. The Census Bureau has used noise infusion in economic statistics for decades. For the decennial census, it moved to a differential-privacy-based system for 2020 after concluding that older methods could no longer defend against modern re-identification. In a 2021 explanation, Census Bureau leaders wrote that the volume of published statistics and the growth of commercial and public databases had made it possible to match records and re-identify individuals "at lightning speed," and that the Census Bureau's own reconstruction research confirmed legacy methods were no longer sufficient under federal confidentiality law (Abowd and Velkoff, 2021). The National Academies of Sciences, Engineering, and Medicine's Committee on National Statistics (CNSTAT) documents the same sequence: the Census Bureau used swapping as the primary protection for 2010, performed reconstruction and re-identification attacks that convinced it swapping was insufficient for 2020, engineered a noise-addition system satisfying a variant of differential privacy, engaged stakeholders, and ultimately settled on a higher privacy-loss budget to allow greater usefulness. (Notably, differential privacy was not applied to the apportionment counts.)
Is the order in effect, and are data products continuing?
The order is in effect. As of July 2026, the Census Bureau is developing implementation guidance, expected in the coming weeks, and existing data products are expected to continue to be made available in compliance with the order, using the confidentiality protections that remain available. For example, the Census Bureau has since released the 2025 Survey of Income and Program Participation (SIPP) data. The Bureau of Economic Analysis has also demonstrated that it continues to publish data products, such as Foreign Direct Investment, since the order was issued. The agencies are working through the order product by product. That forthcoming implementation guidance will be the authoritative source on how specific products are handled and on any transition details.
How will the order be implemented, and what does the required legal review involve?
The order's operational core is an individualized, product-by-product legal review: before a Census Bureau or BEA product is released, the relevant agency director, together with the Office of the General Counsel, reviews it against the specific confidentiality statute that governs it (Title 13 for the Census Bureau; 22 U.S.C. § 3104(c) for BEA). The order frames this as necessary because those statutes differ in what they require, and because moving away from noise infusion means confirming that coarsening and suppression actually satisfy each product's legal standard. The agencies are developing implementation guidance, expected in the coming weeks, that will determine how the review operates in practice. That guidance may, importantly, include whether a product is cleared once or re-reviewed with each release. That single detail will do much to determine the order's practical effect on the pace and timing of releases.
What could the review requirement mean for timely data releases?
The answer to this question depends on agency choices during implementation. On one hand, an explicit legal check helps ensure each release is lawful under its own statute, a form of due diligence the order treats as essential. On the other, it adds a step to a disclosure-review process that is already lengthy (and necessary), which could affect the timing of releases, particularly for the high-frequency economic indicators that markets and policymakers expect on a fixed and published schedule. Because legal review is generally not public or transparent, it can also be harder for data users to determine or understand why a particular release is delayed or withheld. Congress anticipated part of this: the Evidence Act directs OMB to issue standards for categorizing how sensitive each product is and how accessible it can safely be, with public risk assessments behind release decisions (44 U.S.C. § 3582). That rule has not yet been issued by OMB (though a draft is scheduled for August 2026 under the current Unified Regulatory Agenda), so for now these judgments rest on the order's method hierarchy and case-by-case legal review, rather than a standardized, transparent risk-based framework.
What are external stakeholders saying?
Reactions have been strong and have come from several directions. Many data-user and statistical-professional organizations have raised concerns, some emphasizing the risk that leaning on suppression and coarsening will mean less usable public data, especially for small areas and small population groups; others focused on the change process, noting that a change of this scope was issued without a public comment period.
The Executive Director of the American Statistical Association publicly urged the Commerce Department to rescind the order and instead open a public comment process and commission a study of the impacts. A June 2026 joint statement from five professional and data-user associations (Association of Population Centers, Association of Public Data Users, COPAFS, ICPSR, and Population Association of America) raised a related but distinct set of concerns: that the order bypassed the established, transparent processes for updating privacy methods; that confining the agencies to coarsening and suppression will mean less published detail, with neighborhoods and rural communities most at risk of being aggregated away or suppressed; that the order's text is vague enough to leave the affected product scope unclear (e.g., Longitudinal Employer-Household Dynamics); and that OMB's existing open-data guidance, M-25-05, already addresses how agencies should balance privacy and accuracy. They ask that the order be rescinded or, failing that, that Commerce publish an implementation plan in the Federal Register and seek public input before changing any products.
What is Congress saying?
Some Members of Congress and oversight staff have signaled interest. A widely circulated July 2026 open letter from computer scientists and privacy researchers, among them the differential-privacy pioneer Cynthia Dwork and former Census Bureau Chief Scientist John Abowd, argued that the order rolls confidentiality protection back to older techniques, that coarsening and suppression alone can leave fine-grained data open to reconstruction, and that the order should be rescinded and run through public notice-and-comment; the authors also contend the order was politically rather than scientifically motivated, a characterization Commerce's stated accuracy rationale rejects. Commerce, for its part, maintains that the order's intent is to prioritize accuracy and openness while continuing to protect confidentiality.
How did organizations react when differential privacy was introduced for the 2020 Census?
That introduction was contentious as well, and the reaction did not divide along simple partisan lines. A group of states led by Alabama went to court to try to block the use of differential privacy for redistricting data. Some researchers and data users objected that the added noise degraded accuracy for small geographies and small population groups while producing hard-to-interpret results in the most detailed tables. At the same time, some civil-rights and voting-rights advocates worried the changes were rushed, based on an overstated privacy threat, and could ultimately reduce the representation of small communities. In response to feedback from an extended public process, including the release of demonstration data, the Census Bureau adjusted its settings before the final release. Many of the concerns raised then are echoed today, a reminder that this controversy about application of privacy techniques is long-running and cuts across partisan, regional, and demographic lines.
What did the National Academies (CNSTAT) report about managing privacy risk conclude about differential privacy?
The National Academies' consensus study, Toward a 21st Century National Data Infrastructure: Managing Privacy and Confidentiality Risks with Blended Data, offers several conclusions directly relevant to the order. It treats differential privacy as one of multiple tools available, and valuable in part because it can account for cumulative information leakage across releases rather than as a mandatory or universally superior method. The CNSTAT report’s central conclusions cut across the current debate: there are no zero-risk or one-size-fits-all protection methods, and tradeoffs between disclosure risk and usefulness will always exist; the determination of acceptable risk is ultimately a policy decision, ideally informed by technical experts and stakeholder input; and technical and policy approaches in combination are necessary for effective disclosure management (Conclusion 4-1). The report does not endorse any single technique. It emphasizes transparent, dynamic frameworks and layered approaches, a picture in which the choice of how much risk to accept is a policy matter and the engineering of methods to meet it is a technical one. Importantly, the report offers a framework for considering decisions when using blended data, which may support decisions such as the application of the Commerce order.
How broadly does "noise infusion" reach?
This is one of the order's most consequential open questions because the term can be defined narrowly or broadly. The order's own text defines noise infusion narrowly, as adding random values to entries. Experts define it more broadly. John Abowd, the former Census Bureau Chief Scientist who led the 2020 Census disclosure-avoidance work, has publicly described noise infusion as any disclosure-limitation method in which the published statistic differs from the confidential value because of randomness deliberately introduced into the calculation. On that broader reading, the term would encompass not only differential privacy but also data swapping, subsampling, randomized rounding, and synthetic data. The narrower the working definition, the fewer products are affected; the broader it is, the more sweeping the order becomes. Because the order defines coarsening to include "rounding" while defining noise infusion as adding random values, even a method like randomized rounding sits ambiguously between a permitted and a prohibited category. How the Census Bureau and BEA resolve these definitional lines in implementation will determine much of the order's practical reach.
Does noise infusion make data "inaccurate"?
The word "accuracy" carries multiple meanings here, and being precise about it helps. Coarsening and suppression have a real virtue the order emphasizes: they do not publish altered values: a coarsened figure is true but less detailed, and a suppressed figure is absent rather than perturbed. Noise infusion, by contrast, deliberately alters published values. But choosing coarsening and suppression does not eliminate the underlying tension. The trade-off between protecting confidentiality and preserving usefulness is a mathematical feature of releasing data about small groups; moving away from noise does not alleviate the tension between data confidentiality and usefulness so much as it shifts it to coarser geographies, fewer published cells, or withheld tables. It is also worth noting that noise-infused statistics are typically designed to be unbiased and to converge toward the true value as more units are added to the calculation, so the distortion concentrates in the smallest cells rather than spreading evenly. Recognizing all of this is not an argument for or against the order; it is what lets everyone weigh the same choice honestly.
Can coarsening and suppression alone always protect confidentiality?
Not necessarily on their own, and this is the technical counterpart to the accuracy point above. Critics of the Commerce order, including a group of computer scientists and privacy researchers writing in July 2026, argue that coarsening can fail to protect confidentiality when several coarsened statistics interact. Their illustration, adapted from an example by Nathan Goldschlag drawn from the Census Bureau’s County Business Patterns: if an agency publishes several overlapping totals for a small area — say, employees broken out by locality, by industry, and by ownership type, each coarsened in good faith — the published figures can together form a small system of equations that a reader solves with ordinary algebra to recover the exact values for individual businesses, the very disclosure the coarsening was meant to prevent. Noise infusion, they note, perturbs that system and blocks exact reconstruction.
The order's defenders would respond that it does not rely on coarsening alone: suppression remains available as a last resort precisely for cases where coarsening would disclose, and the individualized, product-by-product legal review the order requires is meant to catch exactly these interactions before release. How often coarsening plus suppression can preserve both confidentiality and usefulness for fine-grained data is an empirical question, again, the kind a careful product-level assessment would answer.
Does moving away from noise infusion make the data more transparent?
Not automatically; it depends on what replaces it. Some long-standing techniques protect confidentiality precisely by not disclosing exactly what was done, whereas a formal, documented method can be explained and even adjusted for by analysts. Because the order addresses noise infusion but does not define or address data swapping (the Census Bureau's pre-2020 approach), a return to swapping is one possible path, and swapping is generally less transparent about its effects than the formal methods it would replace. Whether the shift increases or decreases transparency turns on how openly the chosen methods and their effects are documented, product by product.
Which Census and BEA data products rely on noise infusion?
According to a public assessment by John Abowd—using the broad definition of noise infusion described above—a wide range of Census Bureau products use some form of it, including: American Community Survey tables and microdata; County Business Patterns and ZIP Code Business Patterns; OnTheMap and OnTheMap for Emergency Management; Business Dynamics Statistics; Business Formation Statistics; Post-Secondary Employment Outcomes; Veterans Employment Outcomes; the Opportunity Atlas; the Quarterly Workforce Indicators; Job-to-Job Flows; the 2020 Census products including redistricting data; many experimental products under development; and many releases from the Federal Statistical Research Data Centers. He notes that for some products (such as parts of the Economic Censuses) suppression, not noise, is already the primary mechanism. Two cautions are important for reading this list neutrally: it reflects the broad definition of noise infusion, and the order's narrower textual definition combined with the product-by-product legal review the order requires will determine which products are actually affected and how. This is precisely the kind of question a transparent, product-level impact assessment will resolve.
Primary documents
Key analyses and statements
This fact sheet is provided by the Data Foundation as a neutral, matter-of-fact resource. It does not represent a position for or against DAO 216-26.
