CAROLINA PUBLIC PRESS - A recent change in how data is published by the US Census Bureau aims to improve accuracy and transparency, but is also creating uncertainty as critics question how it could affect smaller communities, like the many scattered across North Carolina.
Disclosure Avoidance Order 216-26 effectively bans a category of disclosure-avoidance methods called noise infusion when publishing data from populations and communities across the country. Because of accuracy concerns, the bureau decided to establish alternative methods — coarsening, and as a last resort, suppression — as the standard moving forward.
However, critics argue that these methods could also force the bureau to reduce the amount and detail of data published to meet privacy standards and, in a worst-case scenario, even present large gaps in data, especially for smaller communities.
Many places rely on accurate, detailed census data to help plan for their community’s needs. North Carolina and its many small communities are no different.
What is noise infusion?
Every 10 years, the Census Bureau conducts a decennial census to get an official count of every person in the country, asking about age, sex and race. This data helps inform decisions on congressional representation and other things.
In between those decennials, there’s another, more frequent survey — the American Community Survey. This survey covers all the demographic information not in the census: education, employment, transportation and more.
The ACS 1-year estimates cover places with 65,000 or more residents, accompanied by supplemental estimates that cover populations of 20,000 residents or more.
John Quinterno, the North Carolina State Data Center Coordinator and census liaison, explained in an article that the 5-year estimates cover communities of all sizes, including those smaller than 20,000 residents, and summarize the data collected over the past five years. These estimates provide a lot of detail on social, economic and housing data.
Data collected can be as broad as the state level, or as detailed as households and neighborhoods. Quinterno told Carolina Public Press that in smaller places especially, this data can be easy to trace.
“For a small town, if there’s only one brewery listed, even if they don’t put the brewery’s name in there, you’re going to know who that one business is,” he said.
For decades, the Census Bureau has had to grapple with how to preserve both accuracy and privacy as best they can. As Quinterno explains in his article, releasing unedited, detailed results about a specific, sometimes easily identifiable household would provide data to the overall community, but sacrifice the privacy of that household in the process.
On the other hand, completely withholding data about that household would certainly protect their privacy, but it would withhold useful data from the overall community.
The Census Bureau has adopted different methods through the years to tackle this issue, attempting to publish the most accurate data in the most discrete way. Many of those methods fall under noise infusion.
Noise infusion is a class of methods that have been used to help protect people’s privacy in products like the ACS and the decennial census. These methods operate by inserting a certain amount of “noise” — controlled, random variations — into collected data before publication to make it harder for people to be identified.
For example, starting with the 1990 Census, you might have two households in different census blocks that are the same in every way, except for race. One method of noise infusion allows the bureau to “swap” these households, so the town’s overall published results aren’t affected, but no one can use the data to identify the original source, Quinterno wrote.
Another method is differential privacy, which was used in the 2020 Census. Quinterno described it like using a “slider.”
“You’re not making up data or altering data,” Quinterno said.
“It’s just how are you adjusting the sharpness of the family portrait so that you can still see the portrait, but you maybe don’t let those certain details that you don’t want to see kind of come to the fore.”
For state-level data, there’s typically no noise added since it’s so broad. Minimal noise is added to county-level data. But once you start focusing on smaller levels like census blocks, the risk of privacy being compromised is much greater. The bureau has added more noise to the data before release to prevent this.
“If you want more privacy, you can dial it up, and your numbers will be changed,” Quinterno said.
“But you’re essentially mathematically introducing things in a uniform way, which the changes keep things to reflect the overall values. If you wanted more accuracy, then you would slide it the other way.”
Why is Census Bureau discontinuing it?
The current leadership at the Census Bureau says the noise added to data by noise infusion compromises accuracy. The more noise added, the less accurate the data.
“While this approach was meant to protect confidentiality, it led to concerns that adding noise made the statistics less accurate and less trustworthy,” Census Bureau Deputy Under Secretary George M. Cook wrote in an August blog post.
“The new policies — developed by the Department in conjunction with the Office of the Undersecretary for Economic Affairs and with technical guidance from the U.S. Census Bureau — aim to address these concerns. The order is clear: the top priority is to provide the public with accurate statistical information.”
The Center for Renewing America, a conservative think tank, said only a few Census Bureau staff have access to the actual data and the algorithms used in differential privacy. When noise is added, privacy is protected, but it can come at the expense of accuracy.
While noise infusion methods can accurately portray data from larger populations, that’s not always the case for smaller populations, said Kosuke Imai, a government and statistics professor at Harvard University.
More noise is added to smaller groups since they’re higher risk, but it can result in statistics that are “imprecise and unreliable,” Imai said.
A lawsuit filed in Florida last year by the University of South Florida College Republicans and others challenging the 2020 Census highlights how this can play out at the block level. Quoting the National Academies of Sciences, Engineering, and Medicine, the complaint said “‘a block with three Hispanic residents might appear to have zero or six Hispanic people’ after the Differential Privacy algorithms were applied.”
But under coarsening, a method previously used before noise infusion, precision and detail become the tradeoffs to protect privacy, Quinterno said. Instead of using noise, it rounds, summarizes and reports ranges for values instead.
“For instance, if a county’s poverty rate is reported as a range of 10%-20%, that makes it hard to know the true extent of poverty in the county,” Quinterno said.
Suppression is the other allowed method, but only to be used as a last resort. It withholds individual values, which can turn into whole categories of data, to help prevent people from calculating the missing values on their own. This “cascading effect” can lead to large gaps in data, Quinterno said.
This potential lack of “granularity” — or level of detail — in coarsening, and potential gaps in data by using suppression, are why critics are pushing back against the new order.
The lack of noise used in coarsening and suppression means a higher risk of disclosure, Quinterno wrote. Because the bureau must prioritize privacy as well as accuracy, it might have to publish less data in order to meet that standard, which it has publicly acknowledged.
“Coarsening and suppression can reduce the amount and detail of data published, especially for small geographic areas or population groups,” Cook continued in the blog.
“These are the places where confidentiality risks are highest. We are exploring ways to meet the data needs of these communities while protecting their information.”
What does it mean for NC?
North Carolina is full of small, rural communities. Statewide, 21% of counties have less than 20,000 residents, while 59% of incorporated municipalities — towns, etc. — have no more than 2,500 residents, according to the State Demographer.
These communities rely on the 5-year estimates to determine their social, economic and demographic dynamics. These insights can inform decisions on education planning, housing needs, health services and more.
If coarsening becomes the go-to method for publishing data, the risk of disclosure will be heightened unless the bureau publishes less. Nathan Dollar, director of Carolina Demography, said this could affect communities’ ability to plan for what they need.
“The data are used in a variety of ways by government officials, public service, county planners, public school districts, private schools, NGOs, community-based organizations, fire departments, to make a whole host of decisions, just so they have a better picture of who the communities that they serve are, so they can strategically plan and then also evaluate their progress,” Dollar told CPP.
“The potential for there to be much fewer data available for smaller communities across North Carolina could really imperil their ability to effectively plan and make evidence-based decisions, data informed decisions.”
There’s still a lot of unknowns regarding how this change will play out though, as the Census Bureau is still working to minimize any potential ramifications of the change, and expressed it would continue to work with stakeholders to ensure their needs are being met.
The 2025 ACS 1-year estimates were released Thursday, roughly a month after their scheduled release. The bureau had to delay it until they could ensure the published data would abide by the rules of the new order.
In a webinar hosted by the Census Bureau earlier this month, Nathan Walters from the American Community Survey Office explained that the change would have no impact on ACS 1-year estimates.
“Data users will be really happy to know that there’s been no decrease in the number of tables, geographic coverage, or granularity of the data for the 2025 one-year data that’s being released next week compared to what we would have released next week had the DA policy not changed,” he said.
While no announcement has been made yet regarding the 5-year estimates, which usually come out in December, Quinterno said those will probably be delayed as well.
“Even if the estimates are released on time, they will likely mix data subject to both the new and old privacy standards, which may compromise the accuracy and usability of the information,” he said.
When asked about the 5-year estimates, a spokesperson from the Census Bureau said their focus right now is the 1-year estimates. They expect to share more information on the 5-year estimates when it becomes available.
It’s important to note though that no matter the method — whether it falls under noise infusion, coarsening or suppression — each one comes with its own tradeoffs. As long as the Census Bureau has to consider people’s privacy, publishing data will always be a balancing act between accuracy, granularity and privacy.
North Carolina, like every other state across the nation, will have to wait and see how well the bureau balances these standards moving forward, and to what extent its communities will be affected.
