Census 2020: Fraud or Fairly Conducted?
by Beth Jarosz, Chris Dick, and Dorian Caal
Recently, Senator Banks from Indiana took to social media to criticize the quality of the 2020 Census. In this post, he said that “The 2020 Census was a fraud. The Biden admin used a shady “privacy” formula that scrambled the data and miscounted 14 states. It included illegal immigrants and handed Democrats extra seats. Americans deserve a fair count…”
There was also an attached letter to Commerce Secretary Lutnick, which went into a bit more detail, but basically covered the above points. The post and letter are a lot to unpack, and our goal here at dataindex.us is to monitor the status of federal data. To this end, the three of us, representing over 5 decades of experience as statistical and policy professionals involved in the general field of enumerating the population, wanted to make sure the public has the facts.
The 2020 Census was conducted under difficult circumstances, and was still accurate.
There are a couple of ways to measure the accuracy of a census. Neither shows a statistically significant undercount in 2020.
One method is to compare the count to other known demographic data through a process referred to as Demographic Analysis (DA). Based on DA midpoint estimates, the 2020 Census matched the expected population total of the entire U.S. very closely–it differed by less than one half of one percent (-0.35%, not significantly different from 0).
The other way to check accuracy is to go back and re-count the population using a sample survey independent of the census. That method is referred to as the Post-Enumeration Survey (PES). PES also found less than one half of one percent difference between the 2020 Census count and the expected count (-0.24%, not significantly different from 0).
Those figures compare favorably with recent censuses. The 2010 Census had a slight net overcount (0.13% DA, 0.01%, PES). The 2000 Census had a wider range of estimates, from -0.3% net undercount according to DA to 0.7% overcount according to a survey method. And the 1990 Census had a more substantial net undercount of 1.61%.
Census counts were imperfect, but not in the ways you might think.
The 2020 Census was beset by challenges even before it got started. Planned with funding shortfalls that eliminated key field tests, conducted during a global pandemic, and terminated prematurely by Trump administration, the 2020 Census did undercount some populations, including the following:
Renters
Hispanic/Latino residents
Black residents
American Indians living on tribal reservations
On the other hand, homeowners, people ages 50 and older, and people identifying as White, Asian, or Native Hawaiian/Pacific Islander were overcounted. The national net overcount also hides some variation within groups. For example, states in the Mountain West and in the South showed potential undercounts for Asian Americans, while states on the coasts, in the Upper Midwest, and in the Southwest generally showed estimated overcounts for the same group.
There were also other challenges, such as difficulty counting the population living in group quarters (like dorms, barracks, and nursing facilities) during the COVID-19 pandemic. But overall, as noted above, the total population count was comparable to recent decades in terms of accuracy.
Count issues affected both “red” and “blue” states.
States with counts that were higher than anticipated include both politically liberal states (like Massachusetts) and conservative ones (like Utah); similarly, states with lower than anticipated counts included ones that are politically liberal (like Illinois) and conservative (like Arkansas). In general, the vast majority of states (more than 70 percent) had no statistically significant net over- or undercount.
New privacy methods did not affect the number of seats any state has in Congress.
The new privacy method, referred to as differential privacy, was not used on state population totals for apportionment. Although there is a ton of very dense information about differential privacy in 2020 Census that the average person might not want to wade through, the facts are not hard to find. In bold text on the Census Bureau’s one-page explainer is the following quote: “Differential privacy is not applied to the apportionment count.”
The fact that state counts were not changed is also clear in the 2019 Census Bureau memorandum that Senator Banks cites in his October 6, 2025 letter to Commerce Secretary Lutnik.
Within states, differential privacy typically benefitted rural areas.
Because of the way differential privacy was applied to the 2020 Census, there was a small amount of bias that inflated small populations. In other words, areas with lower population counts (like some rural communities) were likely to see their population adjusted upward, while areas with larger population counts (like some neighborhoods in big cities) were likely to see their population adjusted downward. If anything, removing the differential privacy adjustments could result in less representation for rural Congressional districts.
What is differential privacy, anyway?
If the Census Bureau did not protect people’s privacy, they would be breaking the law (specifically, Title 13 of the United States Code) and Census Bureau staff could go to jail for breaching those strict privacy and confidentiality rules.
Differential privacy is just the latest tool in a long line of techniques to protect the privacy of people’s census data. For decades, the Census Bureau has been applying privacy techniques, like suppressing data or adding small amounts of error (like swapping records from one block to another), to prevent identifying individuals in published data.
Over time, advances in data matching and computing increased the risk that people could be re-identified from supposedly “safe” data. So, for the 2020 Census, under the Trump administration’s leadership, the Census Bureau upgraded their methods to keep pace with new digital risks. That new method, differential privacy, adds (or subtracts) a small amount of statistical “noise” into published data tables to reduce the risk of individual re-identification, while still preserving overall patterns and trends.
The development process included multiple opportunities for public review and input, beginning with a 2018 Federal Register Notice on data products, as well as frequent, public meetings of the Census Scientific Advisory Committee and National Advisory Committee, a 2019 National Academies public workshop on data needs and privacy considerations, and several webinars and calls for public comment.
The Census Bureau quantifies exactly how much variability (error) the noise adds, and publishes metrics and fact sheets so users understand where the uncertainty lies.
Weren’t people counted who aren’t supposed to be here?
The U.S. Constitution (14th Amendment) requires counting the “whole number” of persons in each state.
What’s the takeaway?
Were there challenges with the 2020 Census count? Of course there were. Renters, young children, and Black, Hispanic, and Native Americans were undercounted. Rural areas likely got a small boost from new privacy-enhancing technology. But overall the count was remarkably accurate. Further, new privacy methods DID NOT impact the state totals used for apportionment.
While the facts refute nearly every point made by Senator Banks, we do agree on one point: Americans do deserve a fair and accurate decennial census. To this end, Congress should focus on ensuring robust funding in preparation for the 2030 Census, and the Commerce Department should reinstate the terminated Census Advisory Committees in order that the Census Bureau is able to meet its constitutional and stated mandate of counting every person in the United States.