Census Bureau Issues Working Paper Entitled 'Noncitizen Coverage & Its Effects on U.S. Population Statistics'

WASHINGTON, Aug. 31 -- The U.S. Census Bureau issued the following working paper (No. CES-23-42) entitled "Noncitizen Coverage and Its Effects on U.S. Population Statistics."

The paper was written by J. David Brown, Misty L. Heggeness and Marta Murray-Close.

Here are excerpts:

* * *

Introduction

The accuracy and completeness of noncitizen coverage in U.S. population statistics has wide-ranging effects. They include not only the size of the noncitizen population, but also the total population and its geographic, demographic, and socioeconomic distribution. This has implications for the study of immigration policy effects, Congressional apportionment, allocation of government funds across localities, and demographic and socioeconomic research.

We compare traditional Census Bureau population statistics to administrative record-based estimates (hereafter the AR census) with the same reference date of April 1, 2020. The AR census incorporates data from 31 federal and state government and third-party sources./2

Though administrative record-based population estimates have been compared to counts in earlier decennial censuses, this collection of AR data includes several sources covering noncitizens that have not been used before for general population estimates./3

Traditional Census Bureau population estimates serve as benchmarks for each other. But in each set of counts or estimates, the data used for the foreign-born population (not a U.S. citizen at birth) come mainly or exclusively from survey-style data collection, so their coverage errors may be correlated. The 2020 Census counts are based primarily on survey-style enumeration for each housing structure in the United States thought to be potentially inhabited./4

The Post-Enumeration Survey (PES) uses independently collected address lists and experienced enumerators to produce coverage error estimates for the decennial census (Khubba et al. 2022). However, the people not willing to respond to the PES may be similar to those not willing to respond to the decennial census, and unusual or hidden housing structures may be just as difficult to incorporate in the PES as the decennial census. Demographic Analysis (DA) estimates come from basic population accounting using Medicare data for those 75 and older, vital statistics for the U.S.-born aged 0 to 74, and the American Community Survey (ACS) for the foreign-born population aged 0 to 74 (Jensen et al. 2020). The Census Bureau's Population Estimates Program (PEP) vintage-2020 estimates for April 1, 2020, use 2010 Census, vital statistics, and ACS data (U.S. Census Bureau 2021b). The PEP foreign-born population estimates come from a combination of 2010 Census and ACS data, both of which are survey-style collections. The ACS uses PEP estimates as population controls, so any coverage error in the PEP estimates is reflected in the ACS estimates.

The undocumented population is particularly difficult to enumerate in surveys, so coverage error can be high for that group. Evans et al. (2019) report that focus group members said that undocumented immigrants had privacy concerns about 2020 Census participation, fearing that the data would be used for immigration enforcement./5

Immigrant communities along the Texas-Mexico border, called colonias, are challenging to survey because of irregular housing and addressing, limited English proficiency, limited formal education, confidentiality concerns, complex households, and mobile household members (de la Puente and Stemper 2003). Census Bureau hard-to-count and low response scores include such characteristics./6

Other than for housing tenure (renters or owners), the Census Bureau has not produced estimated undercounts by these characteristics, however./7

Based on case study evidence, Kissam (2017) posits that a key reason for decennial census undercounts of Mexican immigrants is that many of them live in unusual or concealed housing units not included in the Census Bureau's Master Address File (MAF). Generally, no attempts are made to collect decennial census responses from addresses omitted from the MAF.

Some recent studies suggest that the ACS underestimates the foreign-born and noncitizen population groups. Office of Immigration Statistics (OIS) noncitizen estimates are 3.4 to 4.7 million higher than the ACS estimates between 2007 and 2015 (Jasso and Rosenzweig 2020). By varying how item nonresponse to the questions used to identify the foreign-born in the ACS is handled, Mira and Bollinger (2021) estimate the noncitizen population to be between 19.7 and 38.7 million in 2019, compared to the official ACS estimate of 21.7 million. Their estimated range for the undocumented population is even wider, between 7.3 and 18.3 million./8

Jensen et al. (2015) calculate coverage factors for the ACS foreign-born population using the 2000 Census and the 2000 Accuracy and Coverage Evaluation (ACE) Revision II (a post-enumeration survey). They find ACS foreign-born undercoverage of no more than 800,000. Undercoverage is greater for Hispanics under age 50, especially among males. Our contribution to this literature is to compare AR-based estimates to the ACS by immigration status.

Differences in coverage across groups by immigration status could affect sociodemographic statistics. According to Baker (2021a, 2021b), the demographic distributions by country of origin and age are very different for nonimmigrant visa holders and the unauthorized population. Baker (2021a) estimates that 68.7% of the 2018 unauthorized population came from Mexico, El Salvador, Guatemala, Honduras, Colombia, Brazil, and Venezuela, while Baker (2021b) reports that just 16% of nonimmigrant visa holders in 2019 came from Latin America.9 Baker's reports estimate that 75.2% of the unauthorized population is in the 25 to 54 age group, compared to 60.5% of nonimmigrant visa holders. He reports that both groups have higher male shares than in the general population (51.4% for the unauthorized population and 57.4% for nonimmigrant visa holders compared to the AR census's 49.8% for the general population). Using the 2007 1-year ACS, Passel and Cohn (2009) report that 47% of likely undocumented immigrants aged 25 to 64 have less than a high school education, compared to 8% of U.S.-born residents. Among likely undocumented immigrant high school graduates aged 18 to 24, 49% have attended college, compared to 71% of U.S.-born residents. The median household income in 2007 is $50,000 for U.S.-born residents and $36,000 for undocumented immigrants. A fifth of likely undocumented immigrant adults live in poverty, compared to 10% of U.S.-born adults. Artiga and Diaz (2019) find that 8% of citizens, 23% of lawfully present immigrants, and 45% of likely undocumented immigrants aged 0 to 64 lack health insurance, using the 2017 1-year ACS. They also report a higher incidence of having at least one full-time worker in undocumented immigrant households (86% compared to 84% for lawfully present immigrant households and 82% for citizen households).

The sources of coverage error are different in AR and survey-style collection, making a comparison of the two instructive. People who do not respond to surveys may still file taxes, enroll in a government program, or apply for a visa. The AR census has different potential errors, such as having out of date or incomplete information on some people.

Our comparison of the AR census to the DA by nativity shows that the AR census foreign-born estimate is at least 6.2 million higher than the DA's ACS-based estimate, fully accounting for the 6.1 million difference between the AR census total population estimate and the DA high estimate. The AR census noncitizen estimate exceeds the ACS estimate by at least 11.0 million. Analysis of linked and unlinked AR census and 2020 Census person records suggests that noncitizens, and especially those with unknown legal status, constitute a disproportionately high share of AR census people not in the 2020 Census.

We investigated why the AR census includes noncitizens not found in the 2020 Census. The housing structures where AR census noncitizens reside appear to have been omitted from the 2020 Census address universe at a much higher rate than for citizens. Inclusion of more noncitizens contributes not only to a higher AR census total population estimate, but especially estimates for Hispanic males aged 25 to 64 and in counties along the U.S.-Mexico border. Estimates for renters, those lacking health insurance, speakers of languages other than English at home, adults with less than a high school education, and employed people are also higher.

* * *

Conclusion

Facilitated by data sharing among government agencies, we produce AR-based statistics describing the population differently from Census Bureau survey-based population estimates. According to AR the U.S. population is more numerous, male, Hispanic, and middle-aged. It has more noncitizens, especially those with unknown legal status. Population shares are higher for people with less than a high school education, speaking another language at home, renters, and lacking health insurance. Counties along the U.S.-Mexico border are more populous, as are California and Texas in general. These results impact immigration policy, Congressional apportionment, government funding allocation, and socioeconomic research.

It is possible that erroneous inclusions in the AR data contribute to these differences. Counting each person only once is built into the design of the AR census, because only records with a unique identifier are used, and they are unduplicated by that identifier. The identifier may not actually be unique, however; some people could have more than one and appear multiple times in AR-based statistics./31

AR information about whether people were alive on April 1, 2020, may be incomplete, resulting in erroneous inclusion of some people who were not alive./32 Such errors should be more prevalent in the estimates for the population aged 0 to 2 and 65 and over. The AR census estimates for the population aged 0 to 2 match the DA estimates, and the AR census estimates for those aged 65 and over are lower than those in any of the other sources, suggesting that inclusion of non-living people is not the cause of the differences between the AR census and other sources.

Information on the U.S. residency status of people in the AR census is incomplete. Some people with AR data reporting a U.S. address in 2019 or 2020 may have left the country by April 1, 2020. This is most likely to be an issue for nonimmigrant visa holders, who may be in the country for a short time and depart prior to the visa expiration date. ADIS nonimmigrant visa holder data miss some exits. Our results suggest, however, that noncitizens with unknown legal status make up a significantly larger share of people in the AR census not found in the 2020 Census than nonimmigrant visa holders do (3.6 million compared to 1.1 million in Table 6). Most undocumented immigrants are long-term residents who are unlikely to have emigrated./33

Though the AR census estimates are higher than those in other sources overall, they still may undercount the population. DA U.S.-born estimates are higher than the AR census by as much as 3.3 million. The 2020 Census includes 1.5 million people missing from the AR census who appear eligible to be enumerated. About 20.3% of the people only in the 2020 Census are Hispanics aged 25 to 64, so the AR census may undercount this group./34

Additional data and research are needed to refine AR-based population estimates. Including more sources covering noncitizens would help. They are in fewer AR sources on average (Brown et al. 2023), resulting in less accurate point-in-time location data. Another way to address the infrequent appearance of noncitizens in AR data is to widen the vintage window beyond two years and include the probability that a person is a U.S. resident on the reference date. That way people who appear infrequently are covered in the estimates, while accounting for the possibility that they disappeared from the data because they emigrated./35

As of this writing, several of the data sharing agreements are inactive. It will not be possible to improve the accuracy of the AR-based statistics for 2020, produce statistics for other years, or conduct further analysis with this rich set of AR sources unless agreements are renewed. This study demonstrates benefits of doing so.

* * *

The paper is posted at: https://www2.census.gov/library/working-papers/2023/adrm/ces/CES-WP-23-42.pdf