The Census Was Designed to Count Everyone. Its Own Audit Found Millions Left Behind.
Wendover Productions explains the 2020 US Census as an extraordinary logistical achievement: a decade of preparation, 248 temporary field offices, more than half a million temporary workers, satellite mapping, and repeated attempts to reach nonresponding households.
But the most important test of a census is not how carefully its procedures are designed. It is whether the resulting population count is accurate.
There is an important distinction here: Wendover published a forward-looking explanation of the planned operation, not an audit of its eventual performance. The pandemic also disrupted the operation in ways the video could not have anticipated. The strongest criticism therefore belongs to the Census Bureau's actual results and the gap between its elaborate procedures and the accuracy subsequently measured.
The qualification matters. The Bureau's Post-Enumeration Survey found no statistically significant net coverage error in the national population it measured: an estimated -0.24%, with a standard error of 0.25 percentage points. This is not evidence that the country as a whole was undercounted by millions. It is evidence that substantial undercounts in some populations coexisted with overcounts in others. The survey also excluded people living in group quarters and remote Alaska, limiting what it could establish about the entire enumeration.
GAO Files+1
99.98% was a completion figure, not an accuracy score
On December 9, 2020, the Government Accountability Office reported that the Census Bureau had enumerated 99.98% of US housing units. That sounds almost perfect until the method behind the number is examined.
The figure measured housing-unit enumeration, not the percentage of individuals whose information had been independently verified. Among occupied households in the nonresponse follow-up workload:
- Approximately 7.4 million households, or 24.1%, were enumerated using information from neighbors, landlords, or other people rather than household members.
- Approximately 8.4 million households, or 14% of the workload, were resolved using administrative records.
- The Bureau also used statistical imputation to construct records for apparently occupied housing units when other information was unavailable.
The GAO explicitly warned that proxy responses were generally lower quality than direct household responses. It also identified a risk in the Bureau's decision to use administrative records without corroboration from a second source.
U.S. GAO+1
The consequences were measurable. Administrative-record enumerations were missing race information in 18% of cases, compared with less than 3% of direct self-responses. The National Academies subsequently found that reporting ages ending in zero or five was at least 2.5 times more common in 2020 than in 2010 for people aged 23–62. Much of that increase came from nonresponse follow-up, particularly proxy responses.
GAO Files+1
A neighbor may know that a family lives next door. That does not mean the neighbor knows how many people were present on Census Day, their ages, or their relationships.
These methods are not inherently illegitimate. Proxy reporting and imputation have legitimate roles in a national enumeration. The problem is treating a resolved address as equivalent to a verified population record.
The national total concealed unequal counting
The Census Bureau's 2022 Post-Enumeration Survey produced some uncomfortable results.
Population Estimated net coverage error Black or African American 3.30% undercount Hispanic or Latino 4.99% undercount American Indian or Alaska Native on reservations 5.64% undercount Non-Hispanic White 1.64% overcount Asian 2.62% overcountThese figures require qualification. The Black undercount was not statistically different from its 2.06% estimate in 2010. The reservation-based American Indian and Alaska Native undercount was also not statistically different from its 4.88% estimate in 2010. The Hispanic undercount, however, increased significantly from 1.54% in 2010 to 4.99% in 2020.
Census.gov+1
The pattern extended beyond race and ethnicity. Young children and renters continued to be undercounted, while homeowners continued to be overcounted. Fourteen states and two geographic regions had statistically significant net coverage errors. Six states were undercounted and eight were overcounted.
GAO Files+1
This is the central weakness in a story focused almost entirely on logistical preparation. A census can produce a national total that looks statistically sound while distributing its errors unevenly across the population.
The National Academies found that omissions increased for Black and Hispanic populations and were more common in census tracts with lower self-response rates. It also found that coverage and data-quality differences between racial and ethnic groups widened substantially compared with 2010.
nationalacademies.org+1
That matters because census information is used for congressional representation, redistricting, public services, and the distribution of federal assistance.
There is an important limit to the accusation: the GAO reported no major anomalies indicating that the 2020 data were unfit for congressional apportionment. It also cautioned against using the survey's sample-based coverage estimates to manufacture alternative House seat allocations. The stronger case concerns unequal coverage and the reliability of local demographic information, not an unproven claim that particular states were awarded the wrong number of congressional seats.
GAO Files+1
COVID-19 explains the disruption, but not every weakness
Wendover's account describes a planned operation. The actual operation encountered an extraordinary public-health emergency.
The Bureau suspended field operations in March 2020 and extended the interruption of nonresponse follow-up to approximately three months. It subsequently compressed its schedule. The time allocated for processing responses fell from 153 days to 77 days, while the follow-up period was shortened by more than two weeks.
The original December 31, 2020 statutory deadline was missed. The apportionment data were released on April 26, 2021, approximately 116 days later. Litigation and successive changes to the operational schedule also contributed to the delay.
U.S. GAO+1
The pandemic did not destroy every part of the system. National self-response rates were 66.3% in 2020, almost identical to the 66.5% recorded in 2010. The internet, telephone, and paper response options helped preserve participation despite the disruption.
GAO Files
But the populations requiring direct assistance faced much greater difficulties.
The Bureau delayed counting people experiencing homelessness by as much as six months. It subsequently acknowledged that people who moved between locations during that period might not have been counted where they were living on April 1, potentially affecting local and state totals.
Group quarters, including prisons and university accommodation, presented another problem. Facilities were closed, contacts were difficult to locate, and the Bureau eventually used imputation to count residents of some facilities reporting zero population. This was the first time it had used imputation for group-quarter populations in a census.
GAO Files+1
These are not minor details. Wendover specifically explains how the Bureau intended to count homeless people, college students, and people living in unusual residential arrangements. The actual experience demonstrated how vulnerable those procedures were when the institutions and physical access on which they depended suddenly became unavailable.
Nor were all the problems created by COVID-19. The GAO had placed the 2020 Census on its High-Risk List in 2017. By December 2020, it had issued 113 recommendations specific to the census since 2007, of which 20 had not been fully implemented. Staffing, technology, management, and budget uncertainty were already concerns.
U.S. GAO+1
The National Academies offers an important counterargument: the Bureau had developed a solid operational plan and made careful, reasonable adaptations to extraordinary circumstances. Completing the census under those conditions was a substantial achievement. But the same assessment explicitly warns against confusing completion with proof that the resulting data products were high-quality and credible.
nationalacademies.org+1
The fair conclusion is not that the Bureau failed because it could not predict a pandemic. It is that the pandemic exposed weaknesses in a system that already had unresolved operational problems.
The published numbers were changed again to protect privacy
There is another major part of the story missing from Wendover's account.
The video ends by explaining that the Census Bureau must produce detailed statistics by state, county, town, race, age, and other characteristics. But those published statistics are not necessarily untouched reproductions of the information collected.
For the 2020 Census, the Bureau introduced a new privacy-protection system called differential privacy. It adds controlled statistical noise to published tables to make it harder to identify individuals from detailed population data.
The purpose is legitimate. The Census Bureau has a legal obligation to protect respondents' identities, and modern computing makes it increasingly possible to reconstruct private information by combining public datasets. The Bureau also explicitly states that differential privacy was not applied to the state population totals used to allocate House seats.
Census.gov+1
The problem was the implementation.
The National Academies found that the Bureau adopted an entirely new disclosure-avoidance approach at a late stage in the planning process. It had not been tested, prototyped, or deployed in a population census context. The Bureau had no adequate backup plan when implementation proved more difficult and time-consuming than expected.
The consequences were substantial:
- Detailed data products were delayed.
- Available tables were reduced.
- The accuracy of some small-area and demographic data was impaired.
- Rural areas, small towns, renters, and American Indian and Alaska Native populations on reservations experienced particular problems.
The National Academies even identified published tables showing blocks with occupied housing but no people, and other blocks with people but no occupied housing.
nationalacademies.org+1
This was not evidence of deliberate manipulation. It was a documented trade-off between privacy and statistical usefulness, implemented through a system that independent experts judged insufficiently tested.
The distinction matters. The official state totals used for congressional apportionment were not subjected to this statistical noise. But the detailed information used by local governments, researchers, and communities was affected.
A video that ends by promising detailed population breakdowns leaves out the question of whether those breakdowns remain sufficiently accurate to serve their intended purposes.
The real failure is the assumption that completion means accuracy
I would not accuse Wendover of inventing the Census Bureau's procedures. Its explanation of the planned operation is detailed and useful. Nor would I blame the creator for failing to predict the pandemic or the later consequences of the Bureau's privacy-system decision.
The episode's limitation is that it presents census-taking primarily as a logistical challenge. It follows the invitations, workers, addresses, questionnaires, and follow-up visits, then arrives at the final population total as though the principal problem has been solved.
But the evidence demands a harder question.
Did the Bureau count people accurately, in the right places, with reliable demographic information, and produce data that governments could actually use?
The answer is mixed. The measured national total had no statistically significant net coverage error. The Bureau completed an extraordinary operation under difficult conditions. Yet demographic undercounts persisted, some became worse, proxy responses produced measurable quality problems, and a late privacy-system decision impaired important public data products.
The Brilliant sponsorship disclosed in the video provides no evidence of editorial interference, and I found no basis for treating it as part of the criticism.
My verdict: Wendover explains how America planned to count itself. It does not explain the difference between completing the count and getting the country right. The Census Bureau's own subsequent evaluations make that distinction impossible to ignore. A system should not be judged by how many addresses it closes, how many workers it recruits, or how many forms it processes. It should be judged by whether the people represented by its numbers can trust those numbers.
References
1. Wendover Productions — The Logistics of the US Census, March 10, 2020. Original video and account of the planned 2020 Census operation.
2. U.S. Government Accountability Office — 2020 Census: The Bureau Concluded Field Work but Uncertainty about Data Quality, Accuracy, and Protection Remains, December 9, 2020. Documents the 99.98% housing-unit enumeration rate, proxy responses, administrative records, imputation, and compressed processing schedule.
3. U.S. Census Bureau — Census Bureau Releases Estimates of Undercount and Overcount in the 2020 Census, March 10, 2022. Provides the official post-enumeration estimates of undercounts and overcounts by race and Hispanic origin.
4. U.S. Government Accountability Office — 2020 Census: Coverage Errors and Challenges Inform 2030 Plans, November 21, 2024. Examines national and subgroup coverage errors, pandemic-related disruptions, administrative-record limitations, differential privacy, and implications for future census operations.
5. National Academies of Sciences, Engineering, and Medicine — Assessing the 2020 Census: Final Report, 2023. Provides an independent expert assessment of the census, including proxy-response quality, age heaping, operational decisions, and the shortcomings of the new disclosure-avoidance system.
6. U.S. Government Publishing Office — 13 U.S.C. §141: Population and Other Census Information, 2020 edition. Establishes the statutory deadline requiring the Census Secretary to report state population totals to the President within nine months of Census Day.
7. U.S. Census Bureau — Differential Privacy and the 2020 Census, July 2021. Explains the privacy-protection system, its use of statistical noise, and its exclusion from official congressional apportionment totals.