Data users have expressed concerns about the accuracy of the 2020 Census, citing three main questions with clear implications for racial equity:
How did the COVID-19 pandemic affect data collection? To what extent were people of color less likely to self-respond, potentially leading to less accurate data for them?
How did the charged political context surrounding the census affect responses? Did the efforts to add a citizenship question to the 2020 Census, though ultimately unsuccessful, lead to under-reporting of non-citizens?
How much random noise did the Census Bureau’s new privacy-protection framework introduce? How does this affect our understanding of people of color?
If the material below does not answer your questions about the accuracy of the 2020 Census data, or if you have other comments, please contact us at Research@metc.state.mn.us.
Should I be concerned about the COVID-19 pandemic’s effects on household response rates?
No. While the pandemic clearly influenced almost everything in 2020, including Census data collection, the Census Bureau tried extremely hard to reach as many people as possible. Nearly three-quarters of Minnesota’s addresses responded for themselves; this generally yields the most accurate data. This 74% self-response rate was higher than in 2010. In-person interviews with census takers (which also provides accurate data) advanced the completion rate to 81%; this was similar to 2010.
Did people in different race groups respond at different rates? How might that affect the quality of the data?
Yes, people in different race groups probably responded at different rates. Although the Census Bureau does not publish race-specific self-response rates, we can see that self-response rates were generally lower in areas with higher shares of people of color. This is not definitive by itself, but alongside previous research on this topic, it strongly suggests that white people were more likely to fill out the census form themselves than people of color.
That said, self-response rates were generally higher overall than in 2010, and we see that increase in most areas of the region, in areas with low and high shares of people of color. Local governments and community organizations invested a lot of resources in encouraging people to respond to the census; these efforts appear to have paid off. Racial differences in response rates almost certainly exist, but they did not grow between 2010 and 2020, which is noteworthy in the face of a pandemic whose effects were felt more severely by people of color.
Any racial differences in response rates have two implications for data quality: (1) some people may have been missed entirely, and (2) data may be less accurate for people who were counted. We provide perspective on those possibilities below.
Were certain groups undercounted?
Yes, some population groups were almost certainly undercounted (that is, a disproportionate share of their members were not included in the 2020 Census results). This is worrisome given the importance of population counts in redistricting, funding allocations, and other government efforts; undercounts of any group dilutes their power.
More broadly, undercounts are one example of “coverage error” – the extent to which the 2020 Census results do not reflect the entire population of the United States. Overcounts are also possible: some people were counted at multiple addresses and appear in the census records more than once. This also distorts the picture of the population drawn by the 2020 Census.
It is difficult to quantify coverage error in our seven-county region, but the Census Bureau’s Post-Enumeration Survey (PES) provides national perspective. This survey essentially conducts a repeat of the census, but for only a sample of the population. Linking the survey participants to the 2020 Census records helps the Census Bureau determine whether someone was not counted or counted more than once.
The first results from the post-enumeration survey were released in March 2022. They show several important findings regarding the net coverage error (including overcounts and undercounts) for different groups:
- Large differences in coverage exist across race groups.
- Hispanic/Latino people and American Indian/Alaskan Natives living on reservations had particularly large undercount rates — 5.0% and 5.6% respectively.
- Many stakeholders were concerned about undercounts among the Hispanic/Latino population, in part because of last-minute efforts to add a citizenship question to the 2020 Census. Their worries were not unfounded: the undercount rate for Hispanic/Latino people was nearly 3.5 percentage points higher than in 2010, a much larger increase than seen in any other group.
- Black people were undercounted by 3.3%.
- White, non-Latino people and Asian people were overcounted by 1.6% and 2.6% (respectively).
- Members of other race groups not listed above were undercounted by 4.3%.
- Very young children (under 5) were undercounted by 2.8% — a substantially higher rate than in 2010. Adult men were undercounted by 1.3%, and adult women were overcounted by 1.1%
- Renters were undercounted by 1.5%, while homeowners were overcounted by 0.4%.
The PES does not assess the accuracy of the count for noncitizens, which is important given that efforts to add a citizenship question to the 2020 Census may have inhibited candid and complete responses. A recent Urban Institute study estimated that people living in households with noncitizens were undercounted by 3.4%.
The Census Bureau does not edit any 2020 Census counts to account for net coverage error. For more information on coverage error measurement, see the Census Bureau’s website.
In May 2022, the Census Bureau released state-level estimates of coverage error. According to these numbers, Minnesota’s population was overcounted by 3.8%, and it was one of only eight states (primarily in the Northeast) with a statistically meaningful overcount rate. Seven states (primarily in the South) had undercount rates that were statistically meaningful.
Among people who were counted, are there racial differences in data accuracy?
Yes, probably. Because people of color were less likely to provide data themselves than white people, their information needed to be gathered by talking with a neighbor or landlord, or by consulting administrative records. These methods generally yield less accurate data than self-response. For example, a neighbor or landlord might incorrectly identify a person’s age, self-identified race, and/or self-identified Hispanic/Latino origin, particularly if they do not know the person well. Assuming that people of color were more likely than white people to be counted by these less accurate methods, the 2020 Census information is probably less accurate for people of color.
Did the pandemic affect counts of college students?
It does not appear so, at least in our region. Many colleges and universities sent students home before Census Day, leading to concerns that they would be enumerated at their parents’ homes or missed entirely. However, our preliminary analysis shows that the 2020 Census counts of students living in college/university housing was similar to Metropolitan Council data gathered in annual surveys of these facilities. It is more difficult to say whether the pandemic affected counts of students living in off-campus housing, though population figures in areas surrounding colleges and universities are generally in line with expectations.
Did the pandemic affect counts of unsheltered people?
This is unclear. The Census Bureau attempted to count people living in emergency shelters, in “transitory locations” like hotels or campgrounds, and outdoors (including sheltering on buses and trains). Unfortunately, we do not have other data that would provide a direct comparison. We are still examining other ways to assess the accuracy of these counts, because it is very difficult to obtain a complete count of unsheltered people, and we need to know the number of all people in the Twin Cities region.
I’ve been hearing about a new privacy-protection framework that added random noise to the 2020 Census data. Can I still trust the data?
Yes, with some cautions. Based on the available evidence, the effects of the new privacy-protection framework are minimal for most places, but some data for certain geographic areas with fewer than 1,500 residents may be less trustworthy.
A more detailed explanation follows:
- The Census Bureau implemented a new “differential privacy” framework to protect 2020 Census respondents’ identities. Their algorithm added random noise to the population counts that were published, leading many to worry whether they could trust the 2020 Census results. (The Census Bureau has published a helpful overview of this framework.)
- Previous decennial censuses had also used privacy-protection methods, so it is impossible to say how much distortion this random noise introduces. However, the Census Bureau published an additional version of the 2010 Census that used the new differential privacy approach. This makes it possible to examine differences between the differentially private data and what might otherwise have been published if the privacy-protection methods had been the same in 2010 and 2020.
- Our analysis of this data suggests that the differential privacy framework has very little effect on numbers for cities, townships, and census tracts. For example, the population of Minneapolis differs by only 7 people between the two versions of the 2010 data.
- Most census block groups (areas that are smaller than census tracts) have trustworthy data for most population percentages (for example, the share of the population that identifies as Black). For block groups with at least 1,500 residents, counts of households and total population should be trustworthy, but we recommend caution with more detailed block group counts such as the number of Black residents.
- Finally, the algorithm introduces more substantial distortion for census blocks, and it occasionally produces impossible combinations of numbers (such as areas where the entire population is younger than 18). Although our region has few of these illogical blocks, we and the Census Bureau recommend combining many blocks to obtain numbers for larger geographic areas, rather than using blocks as a precise picture of small geographic areas. The random noise will give individual blocks lower or higher counts than they should have, but these distortions will in theory offset each other as blocks are combined, leading to less noisy data at higher geographic levels.
What can I do if I think the 2020 Census results are inaccurate?
Starting in January 2022, governmental entities can submit challenges to the 2020 Census results in the Count Question Resolution program. This program will address possible errors in Census Bureau geographic data that may have allocated people to the wrong city or township. It will also consider the possibility that some housing units were not included in the 2020 Census results, but only if the Census Bureau made a processing error. Housing units that the Census Bureau was not aware of by Census Day cannot be added at this point. While successful changes will affect future population estimates that affect the distribution of funding, they will not change the data used for redistricting.
We will be offering workshops later in 2022 to help local governments understand the process of submitting a challenge. We will also provide resources that will help local governments gather the evidence needed to support a challenge. Information will be posted at metrocouncil.org/census2020.
Later in 2022, there will be a similar effort to address possible errors in the group quarters population counts. It is unclear whether group quarters facilities can be added to the data if the Census Bureau was not aware of them on Census Day. You can encourage the Census Bureau to consider those types of challenges until January 18, 2022. See the Federal Register notice for more information.
As you evaluate the 2020 Census results, keep in mind some considerations from the Census Bureau.
What should I know about future data releases?
The Census Bureau has not yet finalized its plans for releasing more data from the 2020 Census, but forthcoming data releases will likely contain less detail than appeared in the 2010 Census. The Demographic and Housing Characteristics (DHC) file is slated to be released in spring 2023; the Detailed DHC with information for more race groups will follow in summer 2023. The Census Bureau is currently considering reducing the level of detail in the data it releases and/or not releasing some information for smaller geographic areas. More information is available here.
It is possible that counts in future data releases will differ slightly from those in the recently released redistricting data. It is also possible that counts may be inconsistent within the forthcoming data.
We will provide additional perspective on this as the Census Bureau releases more information about its plans.