PRB

No census is perfect, and 2020 was no exception. Evaluations of the 2020 Census found significant undercounts or overcounts for 14 states and certain demographic groups—including young children and people identifying as Black, Asian, American Indian and Alaska Native, and Latino. In fact, the 2020 Census proved less accurate than the 2010 Census, in which no states had significant overcounts or undercounts and coverage errors for most racial and ethnic subgroups were smaller. Errors in the 2020 data will have lasting repercussions throughout the next decade, with potentially serious consequences for underrepresented groups.

Census data are vital to our democracy. Data from the 2020 Census have already been used to determine the number of representatives each state sends to Congress through 2030 (a process called apportionment) and to help define new legislative districts for the November midterms and for all elections through 2030 (a process called redistricting). Each year, census data are also used to allocate more than one trillion dollars in federal funding for important projects and services that benefit local communities.¹  Given the critical uses of census data, stakeholders need to know how accuracy and data quality are evaluated, and how the 2020 Census stacks up according to those measures.

How Is Census Accuracy Measured?

To gauge the accuracy of the census, the U.S. Census Bureau first produces independent estimates of the size and characteristics of the U.S. population using two methods—Demographic Analysis and the Post-Enumeration Survey. These estimates are then compared to census data to identify discrepancies. Demographic Analysis (DA) estimates the population at the national level only using historical population data, birth and death records, Medicare enrollment records, and estimates of international migration.² The Post-Enumeration Survey (PES) estimates the population at both the state and national levels using a sample survey independent of the census and conducted after its conclusion.³

Coverage is the term used to describe how close a census gets to a complete and accurate count of everyone in the nation. Net coverage error is a key measure of census accuracy that represents the balance between the number of people who were correctly counted and the number who were missed, double-counted, or included when they should not have been, calculated by comparing DA and PES estimates to census data.⁴ A positive net coverage error indicates an overcount of a particular geographic area or population group, while a negative net coverage error indicates an undercount.

How Accurate Was the 2020 Census at the National Level?

Despite unprecedented challenges—including a global pandemic, several natural disasters, political interference, and budgetary uncertainty during the planning process—the 2020 Census did not have either a significant undercount or overcount of the total population, based on results from the PES. The PES estimate of net coverage error was -0.24%, which is not statistically different from zero (see Table 1). In 2010, the PES estimated a very slight overcount of the total population (0.01%), but again, this estimate is not statistically significant. The DA estimate of net coverage error (-0.35%) also showed a slight undercount of the total population in 2020, compared with an estimate of a slight overcount (0.13%) in 2010 (see Table 1).

Table 1. Percentage Net Coverage Error for U.S. Population

Notes: *Denotes a net coverage error that is significantly different from zero. DA includes total population; PES includes household population only (excludes Group Quarters and Remote Alaska). Net coverage error for DA is from the Middle Series.

Sources: U.S. Census Bureau, 2020 Post Enumeration Survey Report, PES20-G-01, March 2022; U.S. Census Bureau, Demographic Analysis Tables, Table 3; U.S. Census Bureau, Presentation: Post Enumeration Survey and Demographic Analysis, March 2022; U. S. Census Bureau, “Census Bureau Releases Estimates of Undercount and Overcount in the 2020 Census,” March 2022.

While the absence of significant coverage error for the total population in 2020 is good news, the bad news is that it masks significant coverage errors for some age groups, some race and Hispanic origin groups, and both homeowners and renters.

Children Were Undercounted More Than Any Other Age Group

Both DA and the PES revealed that the 2020 Census significantly undercounted young children ages 0 to 4, with net coverage error estimates of -5.4% and -2.79% respectively (see Table 1). According to DA estimates, children ages 0 to 4 are the only age group for whom census coverage has decreased each decade since 1980, and they also had the largest estimated undercount rate in 2020.⁵

Undercounts of young children matter because they result in communities not receiving their fair share of federal resources to support programs such as Head Start and the Supplemental Nutrition Assistance Program (SNAP), which provide important services for children and families with low incomes. Undercounts also prevent local policymakers from having accurate information to develop long-term plans for schools and other services.

To reduce the future undercount of young children, the Census Bureau recently formed a new internal working group of subject matter experts focused on improving data for this age group. They are expanding prior research to better understand why young children are more likely to be missed and looking for ways to improve data collection for this age group by changing the instructions, probes, and questions used to create household rosters.⁶

Analysis also revealed issues in counting other age groups. Both DA and the PES found undercounts of children under age 18 and men ages 30 to 49 (see Table 1). While the PES found significant undercounts of both males and females ages 18 to 29, DA found slight overcounts. This difference in coverage estimates for young adults may be due in part to differences in the populations covered by the two approaches; a substantial share of people ages 18 to 29 live in college dorms, which are included in DA but not in the PES sample. The PES estimate of net coverage error for women ages 30 to 49 was not significantly different from zero, while the DA estimate showed a slight undercount. Both DA and the PES found substantial overcounts of people ages 50 and older, especially women. In some cases, older adults may have been counted twice because they had more than one residence—such as a vacation home—and were counted in both locations.

Historically Undercounted Groups Were Underrepresented—Again

The 2020 Census continued to undercount groups who have been historically undercounted—Blacks, American Indians and Alaska Natives, Latinos, and those who reported being of Some Other Race—while significantly overcounting the non-Hispanic White and Asian populations, according to PES estimates (see Table 2). DA coverage estimates for race and Hispanic origin characteristics are not yet available.

Table 2. PES Percent Net Coverage Error by Race and Hispanic Origin: 2010 and 2020

Note: *Net coverage error is significantly different from zero.

Sources: U.S. Census Bureau, 2020 Post Enumeration Survey Report, PES20-G-01, March 2022; U.S. Census Bureau, Presentation: Post Enumeration Survey and Demographic Analysis, March 2022.

The magnitude of the net coverage errors increased significantly between 2010 and 2020 for the non-Hispanic White alone, Asian, and Latino populations, as well as for those who reported being of Some Other Race (see Table 2). At -5.6%, the American Indian and Alaska Native population living on a reservation had the highest undercount rate in the 2020 Census, followed by those identifying as Hispanic or Latino (-5.0%), Some Other Race (-4.3%), and Black (-3.3%). Latinos and the Some Other Race population saw particularly striking increases in their net coverage errors from 2010 to 2020—up 3.5% and 2.7%, respectively. And while Asians were neither undercounted nor overcounted in 2010, they were significantly overcounted (by 2.6%) in 2020.

In prior censuses, most people who reported being of Some Other Race were Hispanic/Latino. The 2020 Census again asked respondents to identify their race and whether they were of Hispanic origin in two separate questions, but many Latinos do not distinguish between race and ethnicity in this way.⁷ For example, 37% of Hispanics/Latinos marked the Some Other Race category in 2010 because they didn’t identify with the race categories on the census form (such as White, Black, and Asian).⁸ If the majority of those who reported Some Other Race in 2020 were also Hispanic/Latino, then the undercount of Hispanics/Latinos was likely even higher than 5.0%.

Although Census Bureau research in 2015 showed that Hispanics/Latinos were much less likely to identify as Some Other Race when the race and Hispanic origin questions were combined into one question, the White House’s Office of Management and Budget (OMB) decided to continue to use the two separate questions for the 2020 Census. However, the new chief statistician of the United States recently launched a formal review of OMB’s standards for collecting federal data on race and ethnicity.

Significant undercounts and overcounts of race and Hispanic origin groups are problematic for two reasons. As noted in our discussion of children, undercounts of population subgroups can result in communities not receiving their fair share of federal resources. In addition, inaccuracy in racial and Hispanic origin data from the 2020 Census may have negatively impacted the redistricting process by impeding the creation of legislative districts that reflect the actual racial and ethnic diversity of the people they represent.

Renters Were Undercounted, While Homeowners Were Overcounted

The PES found that the 2020 Census significantly overcounted homeowners (0.43%) and undercounted renters (-1.48%) (see Table 2). The 2010 PES also found statistically significant overcounts of homeowners and undercounts of renters in the 2010 Census, but the magnitude of the renter undercount was greater in 2020. (DA does not provide coverage estimates for homeowners or renters.) Accurate data on the number and demographic characteristics of both homeowners and renters is needed to design and implement effective housing policy and programs such as affordable housing.

Widespread Errors Found in Data for Group Quarters Population

Counting people who live in group quarters (such as nursing homes, college dorms, and correctional facilities) is challenging in any census. For the 2020 Census, the coronavirus pandemic caused major disruptions and delays in counting the group quarters (GQ) population. In March 2020, nursing homes were shut to outside visitors, and many colleges and universities closed, sending students who were living on campus back home. When census data collection operations resumed in the summer of 2020, it was very difficult to get accurate counts of students who would have been living on campus on April 1.

An accurate count of this population is critical for towns and cities that are home to large numbers of group quarters residents, such as college students or prison inmates. Inaccurate data could negatively impact government funding allocations, disaster planning and emergency response, public health analysis and planning, and infrastructure planning.

After the Census Bureau identified gaps and inconsistencies in the group quarters data during processing and review, bureau staff contacted thousands of group quarters facilities in December 2020 to validate responses and fill in missing information.⁹ Despite these efforts, data users and local officials still uncovered problems with the group quarters data released in the PL 94-171(redistricting data) file in August 2021. Some GQ facilities were missing, others were placed in the wrong geographic location, and some—particularly college dorms—had inaccurate resident counts. In response to these issues, the Census Bureau implemented a one-time Post-Census Group Quarters Review (PCGQR) process in June 2022 to correct mistakes in the 2020 Census data for people living in group quarters. Although such corrections will not be used to modify 2020 Census data, they will be incorporated in the Census Bureau’s population estimates. The PCGQR process runs through June 2023.

How Accurate Was the 2020 Census at the State Level?

The PES for the 2010 Census found no states with significant population undercounts or overcounts. In contrast, the PES for the 2020 Census found six states—five in the South—with statistically significant undercounts and eight with significant overcounts (See Table 3). Net coverage errors for the six states with undercounts ranged from -1.9% in Texas to -5.0% in Arkansas. Although 2010 net coverage errors for these six states were not statistically different from zero, they are provided for comparison in Table 3. Four of the six states also had slight undercount rates in 2010, and two states—Arkansas and Illinois—had slight overcount rates.

Table 3. PES Estimates of Net Coverage Error by State

Note: *None of the percent net coverage errors for states was significantly different from zero in 2010.

Source: U.S. Census Bureau, 2020 Post-Enumeration Survey Estimation Report, PES20-G-02RV, Appendix Table 3, June 2022.

Net coverage errors for the eight states with significant overcounts ranged from 1.5% in Ohio to 6.8% in Hawaii. None of these states are in the South—they are distributed across the Northeast, Midwest, and West. Seven of these states also had slight, but insignificant, overcounts in 2010, but the increase in the magnitude of the overcount rates in 2020 was striking.

These state-level undercounts and overcounts in 2020 matter because they directly impacted the apportionment process, resulting in some states losing seats in the House of Representatives and others missing out on seats they might have gained if their populations were fully counted. This allocation of Congressional seats and corresponding electoral college votes cannot be changed until the 2030 Census. In addition, states with significant undercounts will not receive their fair share of any federal funding that is allocated based on population count.

Analyzing Components of Coverage From the PES Can Provide Additional Information About Census Accuracy

Net coverage error is only part of the picture of census data accuracy. The PES provides separate estimates of each component in the net coverage error calculation—erroneous enumerations, whole-person imputations, and omissions. Erroneous enumerations include people counted more than once, such as college students counted in both their dorms and at home. It also includes those who were counted but should not have been, such as foreign tourists. Whole-person imputations are people the Census Bureau added to the count from housing units that appeared to be occupied but whose residents did not complete a census form. Omissions are people who were missed in the census.

In the calculation of net coverage error, the number of omissions is offset by the combined number of erroneous enumerations and whole-person imputations. A net coverage error that is close to zero can mask a high number of erroneous enumerations and whole-person imputations that are offset by an equally high number of omissions. For example, the net coverage error for the 2010 Census was near zero (0.01%), yet there were 16 million people missed in the census and 10 million erroneous enumerations plus 6 million whole-person imputations.¹⁰ The omission rate was 5.3%. To get a more complete picture of census accuracy, stakeholders often focus on the rate of omissions—for the total population, population subgroups, and states.

Omission Rates Increased in 2020, Especially for Historically Undercounted Groups

The PES estimated a net undercount of 782,000 people in the 2020 Census. This total reflects approximately 18 million erroneous enumerations plus whole-person imputations, offset by 18.8 million omissions. The omission rate increased from 5.3% in 2010 to 5.8% in 2020.¹¹

Omission rates varied among racial and Hispanic origin groups (see figure). At 10.5%, Hispanics had the highest omission rate in 2020, followed closely by Blacks at 10.2%, and those reporting Some Other Race at 9.9%. Omission rates for all three groups increased between 2010 and 2020. These high omission rates, along with the significant undercount rates reported in Table 2, have raised concerns about the overall quality of 2020 Census data for race and Hispanic origin groups.

Figure. Census Omission Rates by Race and Hispanic Origin: 2010 and 2020

Note: Race is Alone or in Combination.

Source: U.S. Census Bureau, Presentation: Post Enumeration Survey and Demographic Analysis, March 10, 2022.

The 2020 omission rates were lowest among Asians (3.5%) and Whites (4.5%). Omission rates declined between 2010 and 2020 for the American Indian and Alaska Native, Native Hawaiian and Other Pacific Islander, and Asian populations.

Omission rates in the 2020 Census also varied by state. Table 4 ranks state omission rates for 2020 from worst to best and compares them to 2010. In 2020, these rates ranged from a high of 11.1% in Montana to a low of only 0.7% in Delaware. The two states with the highest omission rates—Montana and Louisiana—did not have significant undercounts, according to the PES. However, the next six states were in fact those that had significant undercounts in 2020. The omission rates for states with significant overcounts did not cluster as tightly as those with undercounts. Although Hawaii had the highest overcount rate at 6.8%, its omission rate ranked 43rd. And while New York had a significant net overcount rate of 3.4%, it had an omission rate of 5.9%, ranking 20th among the states.

Table 4. States Ranked by 2020 Omission Rates With Comparison to 2010 Omission Rates and Ranking

Sources: U.S. Census Bureau, 2020 Post-Enumeration Survey Estimation Report, PES20-G-02RV, Appendix Table 4, June 2022; U.S. Census Bureau, DSSD 2010 Census Coverage measurement Memorandum Series #2010-G-04, Table A1, May 2012.

Although no states had significant undercount or overcount rates in 2010, Table 4 shows that 29 states experienced increases in their omission rates between 2010 and 2020, while 21 states and the District of Columbia saw decreases. Excluding Ohio, all states with significant overcounts in 2020 saw their omission rates decline across the decade. The state rankings by omission rate varied considerably between 2010 and 2020. Among the four states with the highest omission rates in 2020, none ranked worse than 12th in 2010. On the other hand, Iowa—the state with the lowest omission rate in 2010 (2.6%)—saw its ranking drop from 50th in 2010 to 18th in 2020. Evaluation of omission rates, in addition to net coverage errors, provides useful information about variations in census data accuracy and quality among the states. In addition to providing estimates of population coverage, the PES also provides estimates of the coverage of housing units in the census (see Box).

Box 

Coverage of Total Housing Units Improved in the 2020 Census

In every census, some housing units are missed or counted in error. The Census Bureau’s Post-Enumeration Survey (PES) helps evaluate the accuracy of the housing units or addresses that were included in the census count. Similar to the estimates provided for persons, the PES estimates the number of housing units that were correctly enumerated, the number erroneously enumerated, and the number that were omitted when they should have been included. Erroneous enumerations include addresses that are duplicates, businesses rather than residential housing units, group quarters, housing units that do not exist or are uninhabitable, and housing units that were not available for occupancy until after census day (April 1). The PES provides estimates of net coverage error for housing units by occupancy status, tenure, type of structure, geographic area, and race and Hispanic origin of the householder.¹²

As was true for total population, the 2020 Census did not have a significant undercount or overcount of total housing units for the nation. The PES estimate of net coverage error was 0.04% and is not statistically different from zero. This represents a coverage improvement over the censuses of 1990, 2000, and 2010, which all had significant undercounts of total housing units (see Table). However, in 2020, there was a significant overcount of occupied units (0.33%) and a significant undercount of vacant units (-2.57%), although the net coverage error for vacant units was lower than in 2010, 2000, and 1990. The rate of erroneous enumerations of housing units was 3.1% (or 4.4 million), and the omission rate was 3.1% (or 4.3 million).

Table. The 2020 Census Overcounted Occupied Housing Units and Undercounted Vacant Units

Percent Net Coverage Error by Occupancy Status for Housing Units in the United States: 1990 to 2020

The Census Bureau does not speculate on the possible reasons for these coverage patterns but notes in their report that the 2020 Census faced challenges in conducting fieldwork during the COVID-19 pandemic. The PES, like other surveys, also faced challenges related to the pandemic and because of a general decline in survey response rates.

The small net coverage error for total housing units masks some significant coverage errors for some geographic areas and some types of structures, as well as by tenure and by race and Hispanic origin of the householder.

While housing units in the Northeast region were overcounted by 1.9%, there were no significant undercounts or overcounts of housing units in the Midwest, South, or West. Only two states had significant undercounts of housing units—South Carolina (-2.5%) and Vermont (-4.1%). However, seven states had significant overcounts: Alabama (2.7%), Massachusetts (1.4%), New Jersey (2.5%), New York (3.7%), Ohio (1.2%), Rhode Island (1.6%), and Utah (0.8%).

Rented housing units had an overcount of 0.85%, while there was no significant undercount or overcount of owned housing units. Net coverage error rates for single housing units and large multiunit structures (10+ units) were not statistically different from zero. In contrast, small multiunit buildings with two to nine housing units had a significant overcount of 5.1%, while mobile homes and other types of units had a significant undercount of -4.3%.

There were statistically significant overcounts of housing units with householders who were Black (0.87%), Asian (1.37%), Native Hawaiian or Other Pacific Islander (2.64%), or Some Other Race (0.58%).¹³ These significant net coverage errors for Black and Some Other Race householders are in the opposite direction of the net coverage errors for persons in these same groups (-3.3% and -4.34% respectively). One possible explanation for this discrepancy is that the census duplicated housing units, but not all the people living in those units, according to Census Bureau researchers. This would contribute to an overcount of housing units without contributing to an overcount of people.

What’s Next?

The consensus among census data users and stakeholders is that the Census Bureau did a remarkable job overall conducting the 2020 Census under extremely difficult circumstances. While there is no evidence of significant population undercounts or overcounts at the national level, the PES did uncover significant coverage errors for 14 states and several population subgroups. These inaccuracies have potentially serious consequences for underrepresented groups and individual states that will persist until the 2030 Census.

Although the Census Bureau provided 2010 coverage error estimates for counties and places with populations of 500,000 or more, no such substate estimates have been released for the 2020 Census, making it difficult to assess how data accuracy and quality varied across geographic areas and population subgroups within states. Citing the problems with group quarters data, as well as the data accuracy issues identified by both DA and the PES, some data users and stakeholders are calling for the Census Bureau to release more measures of data quality at the substate level.¹⁴ Evaluations of such data could provide a more comprehensive picture of the strengths and limitations of 2020 Census data at the local level, and help identify ways to improve the count in the 2030 Census.

References

1. George Washington Institute of Public Policy, “Counting for Dollars 2020: The Role of the Decennial Census in the Geographic Distribution of Federal Funds,” April 29, 2020.

2. Demographic Analysis provides three series of population estimates—low, middle, and high—by varying the level of historical births, international migration, and Medicare enrollment records. To simplify comparisons, we report estimates from the middle series only.

3. See “How Will We Measure the Accuracy of the 2020 Census?” for a more detailed description of both Demographic Analysis and the Post-Enumeration Survey.

4. See “How Will We Measure the Accuracy of the 2020 Census?” for a detailed description of net coverage error.

5. DA estimates of the number of children ages 0 to 4 are viewed as more accurate than those from the PES because DA uses birth records from the birth registration system in the United States, which is nearly 100% complete. See “Census Bureau Expands Focus on Improving Data for Young Children” for the DA estimates of net coverage error for children ages 0 to 4 from the 1970 through 2020 censuses.

6. See “Census Bureau Expands Focus on Improving Data for Young Children” for more information about these efforts.

7. See “Why Are They Asking That? What Everyone Needs to Know About 2020 Census Questions” for a detailed description of each question.

8. Paola Scommegna, “Changing Race and Ethnicity Questions Reflect Evolving Views,” PRB, February 19, 2020.

9. See “2020 Census Group Quarters” for a discussion of group quarters collection procedures and data processing and review.

10. See “Understanding Who Was Missed in the 2010 Census” for more information about net coverage error and omissions in the 2010 Census.

11. U.S. Census Bureau, Post Enumeration Survey and Demographic Analysis, March 10, 2022.

12. The householder is one of the people who owns or rents a housing unit. If the owner or renter lives somewhere else, then the householder is any one of the adults living in the housing unit.

13. Excluding the category Non-Hispanic White Alone, race and Hispanic origin of the householder is defined as alone or in combination with other groups.

14. See National Academies of Sciences, Engineering, and Medicine, Understanding the Quality of the 2020 Census: Interim Report, 2022, for a detailed discussion of potential substate quality indicators.