by Beth Jarosz, Jae June Lee, and Chris Dick
On June 4, 2026 the Department of Commerce issued a new order “Disclosure Avoidance for Statistical Products” (DAO 216-26) that limits the types of privacy protection methods that the Census Bureau and Bureau of Economic Analysis can use for their data products.
Data users noted early on that the order would likely affect data availability, and we are now beginning to learn the scope of that impact. For example, on August 4, 2026 the Census Bureau announced an indefinite delay in releasing data from the American Community Survey.
On Sunday, August 2, 2026 leadership from the U.S. Census Bureau (George Cook, Acting Director and Michael Lachanski, Deputy Director for Data, Policy, and Science) participated in a panel discussion at the Joint Statistical Meetings, moderated by Census Bureau statistician Michael Hawes. They provided updated information about how the Census Bureau will be implementing DAO 216-26. In this blog, we summarize insights from that discussion as well as information published by the Census Bureau. (For additional detail, please refer to the notes we took during the event and our media tracker.)
What’s the Key Takeaway?
Concerns expressed by data users when DAO 216-26 was published appear to have been well-founded. Many noted that we would have either “less privacy for our personal information, or less usable data, or both.” While Census Bureau leadership reaffirmed their commitment to preserving data privacy and confidentiality, impacts on data availability and granularity are starting to come into focus.
Notably, the DAO is being presented as a reason for prohibiting the collection of race, ethnicity, and sexual orientation data in the decennial census.
Frequently Asked Questions About the Order
Several themes emerged from audience questions.
Why was this directive issued?
Unfortunately, the Under Secretary for Economic Affairs, who would have had the clearest insight into Department of Commerce motivations, was not able to attend. In her place, Census Bureau officials shared their interpretation that the directive reflects the Department's priority to maintain public trust in federal statistics, and that noise infusion changes published statistics in ways that are difficult to explain to data users. They described a trade-off among granularity, accuracy, and confidentiality, and said that where to strike the balance is a policy question rather than a technical one, with accuracy now designated the priority. Panelists said OMB had raised concerns about noise infusion repeatedly and that the change was coordinated across agencies. While they offered anecdotal evidence, they did not offer proof of robust research or empirical evidence to support this rationale.
Moreover, many audience members rejected the stated rationale. At least three speakers noted that while prioritizing accuracy is a policy choice, the DAO restricts methodology which is a technical choice that should be the purview of agency experts.
Perhaps most tellingly, leaked documents related to plans for the 2030 Census (full text available here) present the DAO as pretext for eliminating race, ethnicity, and sexual orientation from the decennial short form. The coordination of these two efforts raises additional questions about the underlying motivation for prohibiting modern privacy protection methods.
Statistical agencies are supposed to be free from political interference. Doesn’t this directive interfere in methodology, rather than set “policy?”
Echoing earlier observations from professional organizations, audience members noted that the order bypassed traditional public review processes and expert peer review that typically accompany a methodological change of this scale. While Department Administrative Orders (DAOs) are internal executive directives that are not subject to standard public notice-and-comment periods, Federal Register notices, using the DAO instrument struck many audience members as inadequate for a change of this magnitude.
Which specific privacy techniques are banned versus allowed under the DAO?
DAO 216-26 bans noise infusion methods but did not name individual techniques. During the session, Census Bureau leadership noted that the agency would interpret the prohibition to include data swapping, formally private methods, random perturbation, and synthetic data (in most circumstances, described below). In their place, the order restricts methods to just two: coarsening (such as rounding, aggregating, and reporting ranges) and suppression. In addition, non-deterministic methods that leave individual underlying records intact, such as random subsampling and probabilistic rounding, remain permissible. And synthetic datasets are still allowed in some limited circumstances (such as for code testing and validation), provided they are not used as the “final public disclosure mechanism” (i.e., the final data product).
Allowed and prohibited methods are described in more detail in “DS025: Organization of the Disclosure Review Board” (starting on page 10).
What qualifies as an official "dissemination event" or "statistical product"?
Statistical products are defined by Statistical Policy Directive 4 as “...generally, information dissemination products that are published or otherwise made available for public use that describe, estimate, forecast, or analyze the characteristics of groups, customarily without identifying the persons, organizations, or individual data observations that comprise such groups.”
With respect to implementing the DAO, Census Bureau leadership noted that any data not part of a core release or experimental product would not be subject to the order. So, for example, working papers and conferences (including research done in FSRDCs, even when conducted by Census Bureau researchers) are not subject to the order. Similarly, if the product carries language such as “does not represent the opinion of the Census Bureau,” then it is not covered by the new rule.
Does this directive reflect a change in how the Census Bureau’s Title 13 Confidentiality is being interpreted?
Acting Director Cook said the Bureau's interpretation of Title 13 has not changed, and that the statute prohibits publishing information that would permit re-identification, including through database reconstruction. Panelists described the intent of the order as disseminating data that is accurate and transparent enough for public use while driving down disclosure risk. How much disclosure risk is acceptable, they said, is and always has been a policy decision, and risk assessments will determine implementation parameters on a case-by-case basis.
How can statutory confidentiality be maintained without noise infusion, and how severely will reliance on suppression or coarsening impact small area and rural population statistics?
Without noise infusion, statutory confidentiality under Title 13 must be maintained by relying almost exclusively on cell suppression and coarsening, such as data aggregation and reporting in broader ranges. This shift is likely to severely impact small area and rural statistics, as those are at higher disclosure risk, which can now only be mitigated by withholding data or reducing detail. Consequently, many small communities face a loss of information, potentially forcing reliance on alternative solutions like secure data centers or tiered access models to obtain usable information.
Will the expected loss of public detailed microdata force data users into Federal Statistical Research Data Centers (FSRDCs) or tiered access systems, and are those systems equipped to handle the demand?
As public datasets lose geographic and demographic granularity and precision (due to coarsening), researchers will increasingly rely on Federal Statistical Research Data Centers (FSRDCs) to access restricted microdata. While recent budget models and infrastructure plans aim to expand FSRDC capacity, user communities remain concerned about access costs and logistical bottlenecks for smaller institutions. Moreover, attendees expressed concern that this may leave rural communities especially vulnerable.
How will the Census Bureau measure disclosure risks moving forward, especially given advancing computing tools and AI threats to re-identification?
Census Bureau leadership acknowledged that rapidly advancing AI tools and machine learning techniques present new re-identification vectors that complicate empirical risk modeling.
Anticipated Impact on Data Products
Census Bureau leaders provided insight on anticipated impact across several data products. The order applies prospectively. Previously published noise-infused estimates remain available, and products that completed disclosure review before June 4 are being handled case by case.
Specific examples discussed include
2030 Decennial Census
The core research agenda has been disrupted, as 2030 Census planning was premised on noise infusion.
2027 demonstration products will feature a simulated Census Edited File (CEF), allowing outside researchers to evaluate accuracy and protection, and how strong proposed protections are compared with unprotected data.
Redistricting data (P.L. 94-171) will need to be reevaluated for risk. Panelists said that the agency will attempt to maximize table structure while maintaining confidentiality. They also cautioned that the last redistricting file published without noise infusion was 1980, which carried less detail on race, no voting-age population, and block-level data only in some areas. We interpret that to mean that data users should expect less detail in 2030 Census redistricting data products, an expectation that has been reinforced by recent reporting.
Evaluation of 2030 Census data quality may be affected. Demographic Analysis, one of the key measures of census accuracy, relies on detailed demographic data. For example, county-level population by age is used to assess the undercount of young children. The disclosure avoidance order may affect the Census Bureau’s ability to publish detailed quality metrics on longstanding issues of bipartisan concern.
American Community Survey (ACS)
There is an indefinite delay for 2025 1-year data products.
For 5-year data products, new privacy-protection methods will be phased in gradually.2025 5-year data, for example, will have four years of existing data (2021-2024) and one year of data protected under the DAO. The first completely non-noise-infused 5-year release will be in 2029.
Longitudinal Employer-Household Dynamics (LEHD) and associated products
Existing data use agreements with 12 states require noise infusion; those remain active temporarily and may be renegotiated over time.
County Business Patterns (CBP)
Modern noise infusion methods are prohibited, and data availability will be impacted.
Data release options and tiered-access models are under consideration.
Principal Federal Economic Indicators
Minimal disruption is expected. Products will revert to historical suppression and coarsening protocols.
Estimates, such as the Small Area Income and Poverty Estimates program (SAIPE)
Estimates were always understood to be imprecise, and no changes are anticipated at this time.
This list is illustrative, not comprehensive. Other products may be affected.