Using administrative records to support the linkage of census data: protocol for building a longitudinal infrastructure of U.S. census records

Abstract

This article describes the linkage methods that will be used in the Decennial Census Digitization and Linkage project (DCDL), which is completing the final four decades of a longitudinal census infrastructure covering the past 170 years of United States history. DCDL is digitizing and creating linkages between nearly a billion records across the 1960 through 1990 U.S. censuses, as well as to already-linked records from the censuses of 1940, 2000, 2010, and 2020. Our main goals in this article are to (1) describe the development of the DCDL and the protocol we will follow to build the linkages between the census files, (2) outline the techniques we will use to evaluate the quality of the links, and (3) show how the assignment and evaluation of these linkages leverages the joint use of routinely collected administrative data and non-routine survey data.

Background

This article describes the linkage methods that will be used in the Decennial Census Digitization and Linkage project (DCDL), which is completing the final four decades of a longitudinal census infrastructure covering the past 170 years of United States history. Conducted every 10 years since 1790, the U.S. Census is a mandatory survey with response rates well over 95% since the mid-twentieth century [1]. While the U.S. has no population register, researchers have a long history of using linked census data to study long-term social and economic change. Over the past two decades, historical demographers have built an extraordinary longitudinal infrastructure of publicly-available census data from 1850-1940, including major efforts from the University of Minnesota and the National Bureau of Economic Research [2–4]. Similar infrastructures have been built from the publicly-available historical censuses from Canada, England, Norway, Sweden, and Wales [5–8]. All of these projects have focused on data that is old enough to be made publicly available, and none of these countries currently has an integrated infrastructure of linked records that includes both public historical data and confidential modern data. The DCDL is filling this gap for the U.S., creating linkages across nearly a billion confidential census records from the second half of the twentieth century, as well as to already-linked records from the 1940 and 2000 through 2020 censuses. When complete, the digitized and linked data will complete a longitudinal infrastructure that spans the entire period from 1850–2020.

The DCDL began as a series of pilot projects in the 2010s and is currently underway [9]. The resulting longitudinal data will become part of a wide range of restricted-use survey data and administrative records that is made available to qualified researchers working in secure facilities called Federal Statistical Research Data Centers (FSRDCs). Researchers working on approved FSRDC projects can currently use linked data from the censuses of 1940, 2000, 2010, and 2020. Beginning with the 2000 Census, the Census Bureau digitized respondent names in the course of data processing and used those names to facilitate linkage over time. While the agency has microdata files from the 1960 through 1990 censuses, those files have never included respondent names and were therefore never linked over time. We first added historical census data to the linked data infrastructure in 2014, when our team acquired the complete 1940 Census, which was made public in 2012 and digitized by Ancestry.com and the University of Minnesota. We developed methods to link the 1940 Census data to the modern census and survey files and made the linkages available within the FSRDCs. We released the linked 1940 data via a pilot program in 2015 and then to all FSRDC researchers in 2018 [10].

Even though the currently available series has a 60-year gap between the 1940 Census and the 2000 Census, the linked census data are already being used intensively by sociologists, economists, demographers, and public health scientists [11]. Ongoing projects are advancing research on topics including Medicaid and mortality [12], the impact of reproductive health care on education [13], the inter-generational determinants and outcomes of twentieth-century mobility [14–16], and the demography of mortality [17–19]. Research conducted with the completed infrastructure will have the potential to further transform our understanding of population aging processes, life-course transitions and trajectories, and the long-term antecedents of health, well-being, and mortality over the last century.

With support from a coalition of funders, DCDL is filling in the gap between the 1940 and 2000 censuses. Our initial work has focused on digitizing respondent names from 1960 through 1990. The censuses from these years collected respondent names, but those names were never digitized and are stored only on microfilm in the respondents’ handwriting. Beginning in early 2021, the Census Bureau data processing staff began creating digital images from the 250,000 microfilm reels containing the original census manuscripts. The DCDL’s handwriting recognition team has begun using automated methods to extract names, adapting work we completed on a pilot project with the 1990 Census. The DCDL linkage work will begin in early 2023, concurrently with ongoing digitization and handwriting recognition. Upon completion of the DCDL project, all linked data will be available in the FSRDCs.

Our approach to creating linkages between censuses will be to integrate routinely collected administrative records files with one another, and then to use the resulting administrative records composite to facilitate the assignment of linkage keys to the records from the 1960-1990 censuses. By using administrative records that are proximate to census data in time, we reduce the risk of variables changing over time, and we maximize the use of the particular linkage variables available during each census year. This protocol for the linkage of records on the DCDL project builds on methods used to assign anonymous linkage keys to the most recent decennial census data [20], as well as lessons learned from research focused on the linkage of historical census records from 1850-1940 [21–23].

Our main goals in this article are to (1) describe the development of the DCDL and protocol we will follow to build the linkages between the census files, (2) outline the techniques we will use to evaluate the quality of the links, and (3) show how the assignment and evaluation of these linkages leverages the joint use of routinely collected administrative data and non-routine survey data. This work is particularly relevant to researchers conducting or evaluating probabilistic linkages between relatively low-information files, as well to those developing large infrastructure projects creating linkages across centuries of public and confidential population data.

Challenges and preliminary studies

The longitudinal linkage of decennial census records involves challenges that will be familiar to anyone who has linked records collected at different times. High-quality personally identifying information (PII) is usually more limited in the past, and the digitization of paper-based records often introduces new potential for increased cost and error. Even when a variable is present and accurate in multiple datasets over decades, linkages need to account for changes in respondents’ lives. People may have changed their names, their family composition, their addresses, and even the way they describe where they were born. There is no national identification number in the United States; U.S. Social Security Numbers (SSNs) are a non-universal identifier and have never been requested of census respondents. For this reason, all linkages over time are probabilistic and rely on variables such as name, age, and state or country of birth. These challenges have contributed to relatively low rates of coverage and high rates of error and bias in linked census data from 1850–1940, especially when compared to linkages across more recent surveys and administrative data [21, 22, 24].

Linkages between twenty-first century U.S. censuses set a high standard for coverage and accuracy. With a focus on the higher-information censuses since 2000, Census Bureau researchers have developed unprecedented capabilities for large-scale linkage of restricted data [14, 20, 21, 25, 26]. The Census Bureau’s production record linkage processes operate by matching any file to a large composite of administrative records. Valid links are assigned a unique Protected Identification Key (PIK) that facilitates linkage to any other file that has been assigned PIKs. Research based on linked census data from 2000 forward suggests that coverage is high and error in PIK assignment is minimal. For censuses and major surveys conducted since 2000, about 90 percent of cases have PIKs assigned, and nearly all assigned PIKs have been shown to be accurate [27, 28].

Even with this high linkage rate, PIK coverage in the recent censuses varies significantly by key population characteristics. This is mainly because respondents who do not appear in administrative records cannot be assigned a PIK. For this reason, the PIKs assigned to recent censuses and surveys have been shown to over-represent the non-migrant population, those with proficiency in English, citizens, and other population sub-groups that are more likely to appear in federal government administrative records [29]. Linkage rates also vary by geography, and are lowest in the Southwest, and highest in the upper Midwest [28–30].

Linkages have been significantly lower for the nineteenth- and twentieth-century censuses. For 1850-1940, linkage rates between successive censuses have typically been around 10 to 45 percent for native-born men, and much less for the whole population [21, 31, 32]. Our work linking the 1940 Census to modern administrative records fared somewhat better, producing PIKs for 40 percent of adults and 70 percent of children [21]. Validation tests showed that about 4 percent of the assigned 1940 PIKs were likely to be in error [21]. In a pilot conducted with 1,500 hand-entered records from 1960, about 75 percent of records were successfully linked to administrative records [33]. That work found that the quarter of birth variable—which is available in 1960-1980 censuses but was not in 1940—significantly improved linkage results beyond what would have been achieved using just the age variable [33]. A pilot project from the 1990 Census assigned PIKs to a broadly representative sample of 28,000 digitized records collected as part of the 1990 Content Reinterview Survey (CRS) [26]. The 1990 CRS linkage team assigned PIKs to 84 percent of cases, including 87 percent of adults and 73 percent of children.

Based on the experiences of these projects, we expect DCDL’s linkage rates to increase over time, from a low of 75 percent in 1960 to a high of 85 percent in 1990. We expect the rates to increase with each census year, since each census has slightly more information to support record linkage, and there are more administrative records available in the time-specific composite file that the census will be linked to. These expected rates fit squarely between the observed rates in 1940, when there is significantly less information in the census and the administrative records, and the twenty-first century censuses. In terms of expected error, our estimates can only draw on the completed linkages from 1940 (where 4 percent of links were estimated to be in error) and the 2000 Census (where less than 1 percent of the linkages were estimated to be in error). We expect our measure of incorrect links (Type 1 errors) to fall between these two extremes. While no preliminary studies estimate the proportion of unlinked cases that should have been linked (Type 2 errors), we have developed methods to measure these in the DCDL data.

Methods

The DCDL project will adapt and expand well-established techniques to conduct record linkage across multiple decennial censuses. The core of the record linkage strategy will use a modified version of the Census Bureau’s production record linkage processes, the Person Identification Validation System (PVS) [14, 20, 21, 24]. We will use probabilistic matching methods similar to those developed for the already-completed linkages for 1940 and 2000 through 2020, and assign a PIK to each case. Since the data available in each decennial census and in administrative records varies over time, we will adapt these methods to maximize the use of relevant variables during each census year.

The Census Bureau’s PVS process assigns PIKs probabilistically, by linking census records to a composite database created from federal agency administrative records that contains each person’s unique PIK value. Each PIK is associated with an SSN or an Individual Taxpayer Identification Number (ITIN). ITINs are tax numbers used by people who do not have an SSN but are required to file federal income taxes. When a census or survey record is linked to the composite database, the record is assigned the relevant PIK from the composite. Respondent names are removed from the data and researchers are provided with a “PIK Crosswalk”, which includes the survey’s record identification numbers (such as a household ID number and a person ID number) and a PIK for each case that was successfully linked to the composite. Census and survey data with PIK crosswalks are made available to researchers on Census Bureau approved research projects within the FSRDC’s restricted environment for research. We will follow an adapted version of PVS to assign PIKs to the 1960–1990 censuses, and to make the resulting PIK crosswalks available through the FSRDC.

Table 1 shows the administrative data sources to be integrated and used for the composite reference file for linking the 1960–1990 censuses, as well as the variables available to support the linkage and assignment of PIKs in each data set. The two sources for these records are the Social Security Administration (SSA) and the Internal Revenue Service (IRS). The SSA manages the public old-age pension program and a disability supplement program for people of all ages. Most people born in the U.S. are registered for this program at birth and are assigned a permanent SSN [34]. The SSN is used to track eligibility for pension benefits that are primarily earned through credits as a worker or as a worker’s dependent. The IRS data is derived from annual income information that most Americans with income are required to report each year in order to establish their taxes owed.

Table 1: Data sources used to construct the administrative records composite file.

	Census numident 1936–present	IRS 1040 forms 1969	IRS 1040 forms 1974–1994	SSA master beneficiary record 1962–present
Universe	all SSN holders	all tax filers	all tax filers	SSN holder who earned benefits,
			and spouses	child or spouse receiving benefits
Name	x		x
Alternate Name	x
Date of Birth	x
Date of Death	x
Sex	x
Marital Status		x	x	x
State or country of birth	x
Parents’ names	x
Street address		x	x
County		x	x
SSN	x	x	x	x
Spouse’s SSN			x	x
Child’s SSN				x

The core components of our administrative records composite include the SSA Numerical Identification (Numident) file, the SSA Master Beneficiary Record (MBR), and the IRS tax return data. The three primary administrative data sources in the composite will be linked to one another using the SSNs on each file. The SSNs in these files are expected to be high quality, since all of these data are used for essential government functions.

Most of the composite’s demographic information comes from the SSA Numident, which is built from applications for an SSN. The SSA Numident is a quarterly transactions file that includes all new applications as well as changes to existing SSNs. The Census Bureau has standardized this flow of data into a cumulative annual snapshot file called the “Census Numident.” The Census Numident contains basic information on all SSN-holders going back to the beginning of the program in 1936, also including any updates the SSA received, such as new names and dates of death (a critical variable for the SSA, since the date of death usually means the suspension of pension benefits) [17]. The variables that we include in the composite include SSN, name, alternate/married name, sex, date of birth, state or country of birth, parents’ names and, to exclude ineligible candidate links, date of death.

The composite’s information on addresses comes from the IRS tax returns. These data include the complete universe of IRS tax return files, obtained by the Census Bureau from the IRS shortly following each tax year. The Census Bureau holds these data for every fifth year from 1969 through 1994, and every year from 1995 to the present. The files contain the original data transferred from the IRS and have information from the IRS 1040 files including SSN, marital status, mailing address, and, for 1974 forward, all of the same information as well as individual filers’ and spouses’ names. Our use of these files will focus on 1969, 1979, and 1989, because those files were created at almost exactly the same time as each census was conducted. For instance, 1969 IRS 1040 file includes addresses from tax returns that were likely filed on and before 15^th April 1970, which is nearly the same time that the 1970 Census was conducted. Unlike the modern IRS 1040 data, the 1969–1994 IRS 1040 data do not include identifying information for household dependents. For 1969, there are only SSNs for the filer, or head of household, with no names provided. For 1974–1994 there are SSNs and names for the filer and, for joint returns, the filer’s spouse. None of these files include information for dependents or children in the filers’ households.

The third file we will use to construct the administrative records composite is the SSA’s MBR. The MBR is particularly useful for specifying family relationships. The MBR provides information on Social Security benefit recipients. Since benefits are often provided to the minor child of the earner (in the case of disability or death) or the spouse of the earner (in the case of death), the MBR includes the SSN of the person who originally earned the benefit as well as the SSN of the person who received the benefit (if a different person). Since the earner and beneficiary are usually related, this file often effectively provides an SSN-to-SSN link between two spouses or between a parent and child. These relationships are explicitly indicated by a separate code on the record. These data provide an additional record connecting potential couples within a household or parent and children pairs in a household. The MBR contains information on individuals who received benefits since 1962, which will reveal family relationships for many older respondents in the 1960–1990 censuses.

As a single integrated file, the composite will contain all the information available for each person, as well as additional constructed variables to make the information comparable to each year’s decennial census data. For example, the composite’s street address and county variables are converted to census geographies, specifically census tract, for each of the censuses using a variety of methods. The composite’s date of birth variable will be converted to year of birth and quarter of birth, again to maximize comparability with the census data which does not include exact birth date. The SSN information in the MBR will be enriched with individual names from the Census Numident, including names for benefit recipients, as well as their spouses and parents who are specified MBR. Finally, the variables will be harmonized so all data have the same codes and formats to facilitate the linkage. The resulting file will contain one row for each person ever observed in the administrative records, with all available information about that person and any observed family members included as well.

Linking the 1960–1990 censuses to the administrative records composite

Table 2 shows the linkage variables that can be used to probabilistically link census records to the administrative records composite file. The variables available vary only slightly by census year. For all records across all years, the linkage to the composite file will primarily rely on name, age, sex, and place of residence. For children living with a parent, parents’ names will be used to link to the composite. For couples, the marital status variable and the spouse’s name will be used for the linkage. In all years other than 1990, the use of quarter of birth will be investigated for matching, though birthdays are often misreported in survey data [35]. The 1990 Census contains full addresses for most respondents. In the 1960–1980, census geographies (such as census tract) and county information will be used to aid in the individual level linkage to the composite.

Table 2: Linkage variables in 1960–1990 censuses and the administrative records composite file.

	Decennial censuses short forms 1970–1990	Decennial censuses long forms 1960–1990	Administrative records composite
Universe	all U.S. residents	16–25% sample of U.S. residents	Varies by field
Name	x	x	x
Age/Year of Birth	x	x	x
Quarter of birth		1960-1980	x
Sex	x	x	x
Marital Status	x	x	x
State or country of birth		x	x
Parents’ names	X*	X*	x
Street address	1990	1990	x
Census Tract	x	x	x
County	x	x	x
SSN/PIK			x

For each census, about one-sixth to one-fourth of households were asked to complete additional questions via the long form survey. For the subset of census respondents who received the long form questionnaire, the linkage to the composite will also use the state or country of birth variable. This variable is also available in the Census Numident file, so it is a useful variable for linking the census respondents to the composite file, even though it will only be available for some cases. Since this variable provides an additional linkage key, we will expect slightly more accuracy in the linkage of long form census records to the composite.

The 1960–1990 censuses have a large number of linkage variables, and it may seem that we could directly link one census to another without the use of an administrative records composite. With censuses that are 10 years apart, however, the power of the composite made from administrative data is that it includes year-specific versions of information that often changes over time. For instance, the place of residence variables available in each census will be linked using the composite’s time-specific location information from the IRS 1040 data. Without this information, the censuses’ geographic variables would only be useful for linking those who had not moved within a 10-year period. Similarly, the composite contains all name variants that were ever reported to the SSA, including names that were changed upon marriage or for other reasons. Without the composite, the censuses’ name variables would only be useful for those who had never changed their names.

Increasing PIK assignment in the 1960–1990 censuses

As described above, the linkage strategy consists of linking each census to an integrated administrative records composite, using well-established probabilistic matching techniques with low-tolerance differences between variables in the two files. These techniques produce robust linkages that have already been used to produce research in top-tier journals across the social sciences, as well as in production record linkage applications within the Census Bureau [12–14, 16, 36, 37]. For a large majority of the population—specifically those represented with high-quality data in administrative records as SSN-holders or tax filers—this approach will continue to work well, as it has in censuses before and after this period. However, to link members of the U.S. population that are not in the administrative data and are often groups researchers want to study, we will need to use new techniques.

The first approach for broader coverage will take advantage of known family links from available data. For instance, consider a married couple in the tax data where we link one spouse to the census but not the other. In this case, we will make another pass at the census family, considering the unlinked spouse from the tax data as a candidate link for the unlinked spouse in the census. We will carry out similar processes with other available administrative data where we observe spousal links, child-parent links, and sibling links, and data such as censuses preceding and following the census being linked and proximate household survey data.

The second approach to improve coverage in the linkage of DCDL records will use new techniques for entity resolution to assign PIKs to respondents who were not initially linked to the administrative records composite. About 10 percent of respondents in 2000 were not assigned PIKs, and we expect 15 to 25 percent of census respondents in the earlier census years not to have PIKs assigned. Considering only unlinked individuals from the censuses, as well as those from administrative records who were never linked, we will look for links over time and between censuses and the other records. Rather than a file-to-file linkage, we will approach these sets as a large deduplication problem [38, 39]. When we observe duplicate records in this set of data, we will create a new, unique PIK value that is not associated with an SSN or ITIN.

Methods for evaluating outcomes

Linkage quality issues are likely to be a greater challenge in the 1960-1990 census data than in the more recent decennial census data, since linkages from these years will be made with less precise linkage variables. For example, the 1960–1990 censuses do not include exact date of birth. Since the administrative records composite contains full date of birth, we need to convert (and reduce) this information into variables for age (at the time of the decennial census), year of birth, and quarter of birth. By contrast, the censuses of 2000–2020 contained exact date of birth and are thus able to take advantage of the full birthdate information from the composite. The PIK assignment for 1960–1990 censuses will also likely rely on less precise name data, since names recovered using OCR in 1960–1990 may be lower-quality than the names recovered using OCR with clerical edits in 2000–2020 [40–44]. Thus, the PIKs assigned to the 1960–1990 censuses are likely to cover fewer records and have more errors than those from the 2000–2020 censuses.

We will carry out two strategies for documenting linkage quality. Our first strategy will be to replicate each year’s linkage approach with the 2000 Census data, and then compare PIKs generated with the full information available in the 2000 Census to PIKs generated with the more limited information that was available in the 1960–1990 censuses. This is based on the strategy that we used to assess the quality of PIKs assigned to the 1940 Census [21]. While the 2000 Census PIKs were assigned probabilistically and are therefore not real “truth data,” they were assigned with more information than the 1960–1990 PIKs will be. This comparison will help us understand the cost of making these lower-information linkages, and to quantify differences in coverage and error between the 2000 Census PIK assignment and the pre-2000 PIK assignment. It will also allow us to specify links made in error (compared to the higher information data) and links missed.

For example, to estimate the quality of the 1960 links, we will conduct a new linkage with the 2000 Census respondents. We will begin by removing the exact date of birth from the 2000 data, replacing it with the level of information available in 1960 (age, year of birth, and quarter of birth). If our analysis of 1960’s OCR-based name data shows that those names were of lower quality than name data from 2000 (when compared to the hand-entered “truth data” from each year), we will degrade the quality of the 2000 Census names accordingly [45]. Finally, we will create an administrative records composite consisting only of records from the SSA data, since no tax data are available in 1960. Using this modified base of 2000 Census data and administrative records, we will carry out the record linkage processes we used for the 1960 Census. We will then compare the “limited information” links made with the 2000 Census data to the “full information” links that were originally made for the 2000 Census. We will use similar assessments for all years to generate new information on linkage quality in the lower-information methods we use for the older censuses.

Our second strategy for documenting linkage quality will be based on data from the Panel Study of Income Dynamics (PSID). The PSID is the longest running longitudinal household survey in the world and has been conducted at the University of Michigan’s Institute for Social Research since its inception. The survey’s leadership team is in the process of bringing these data into the Census Bureau’s Data Linkage Infrastructure. Initiated with 18,000 individuals in 5,000 families in 1968, PSID respondents were interviewed every year through 1997, and then every two years after that. The PSID has collected and validated SSNs for about 5,000 respondents.

For PSID cases with known SSNs, we will assign PIKs by linking respondents’ SSNs to the administrative records composite file and re-validating each link with name and date of birth. These PIKs will present us with near-certain truth data on individuals who, in many cases, were interviewed consistently for 50 years. Using the PSID interviews that are most proximate to each of the 1970–2010 censuses, we will remove the PII from the PSID to make the PSID have the same PII as the proximate census data. We will create “Evaluation PIKs” on the PSID by carrying out our census linkage procedures on these lower-information PSID records. When we compare a PSID respondent’s truth PIK (based on the SSNs and additional PII in the PSID) to the same individual’s Evaluation PIK (based on lower-information PII from PSID interviews proximate to census years), we will be observing the quality of the 1970-1990 linkages. We will also benefit from knowing that the PIKs on PSID data should not change over time, since they were verifiably the same individual who was interviewed by the PSID in (approximately) 1970, 1980, 1990, 2000, and 2010. When we observe an Evaluation PIK on a PSID respondent changing over time, we will be observing errors in the late-twentieth century linked data.

Our investigations of linkage quality will have a feedback effect that informs our core linkage techniques. For instance, when we identify subsets of the low-information links from 1960–1980 that are subject to significant error, we will attempt to fine-tune tolerances in the linkage process, such as the string comparator scores or age differences. Our linkage experiments using the 2000 Census and the PSID will thus serve both a documentation purpose, describing the procedures we followed and the error we observed, as well as providing us with feedback and strategies to improve the record linkage processes for censuses of 1960–1990.

Conclusion

The DCDL’s data linkages will complete the largest, longest, and most substantively comprehensive longitudinal data resource in the U.S., effectively providing a link between public and confidential records over the past 170 years. We described the protocol that we will follow to link the 1960 through 1990 censuses, as well as the methods we will use to assess the quality of the linkages. Our methods make use of a key difference between census data and administrative records. The censuses are rich with content, but they have less comprehensive PII that sometimes changes between censuses. Routine administrative records, on the other hand, are more narrow in content but have the high-quality, time-specific PII that is necessary to administer social programs. In order to compensate for the fact that we are linking censuses with limited PII that changes over time, we leverage the high-quality PII in proximate administrative records. We will use additional non-routine survey data, in the form of the Panel Study of Income Dynamics, to evaluate the quality and coverage of our links. The resulting infrastructure will allow researchers to investigate long-term demographic and socio-economic dynamics on a scale that has never before been possible.

Conflict of Interest

The authors have no conflicts of interest.

Any views expressed are those of the authors and not necessarily those of the U.S. Census Bureau.

Statement on ethics

This research project described in this paper received approval to proceed from the University of Michigan’s Institutional Review Board for Health Sciences and Behavioral Sciences (IRB-HSBS) as the “Decennial Census Linkage Project” (ID# HUM00178299).

Acknowledgement of funding sources

This work was supported by the National Science Foundation (award BCS-2023639) and the U.S. Census Bureau (award CB-31RMD0160004).

References

1. King ML, Magnuson DL. Perspectives on historical US census undercounts. Soc Sci Hist. 1995;19(4):455–66. 10.2307/1171475

   https://doi.org/10.2307/1171475

2. Abramitzky R, Boustan L, Eriksson K, Pérez S, Rashid M. Census Linking Project: Version 2.0 [dataset]. 2020. Available from https://censuslinkingproject.org

3. Helgertz, J, Ruggles S, Warren JR, Fitch CA, Goeken R, Hacker JD, Nelson MA, Price JP, Roberts E, Sobek M. IPUMS Multigenerational Longitudinal Panel: Version 1.0 [dataset]. Minneapolis, MN: IPUMS, 2020. Available from https://usa.ipums.org/usa/mlp/mlp.shtml

4. Ruggles S, Fitch CA, Goeken R, Hacker JD, Nelson MA, Roberts E, Schouweiler M, Sobek M. IPUMS Ancestry Full Count Data: Version 3.0 [dataset]. Minneapolis, MN: IPUMS, 2021.

5. Antonie L, Inwood K, Lizotte DJ, Andrew Ross J. Tracking people over time in 19th century Canada for longitudinal analysis. Mach Learn. 2014 Apr;95(1):129–46. 10.1007/s10994-013-5421-0

   https://doi.org/10.1007/s10994-013-5421-0

6. Ruggles S, Roberts E, Sarkar S, Sobek M. The North Atlantic population project: Progress and prospects. Hist Methods. 2011 Jan 31;44(1):1–6. 10.1080/01615440.2010.515377

   https://doi.org/10.1080/01615440.2010.515377

7. Ruggles S, Fitch CA, Roberts E. Historical census record linkage. Annu Rev of Sociol. 2018 Jul 30;44:19–37. 10.1146/annurev-soc-073117-041447

   https://doi.org/10.1146/annurev-soc-073117-041447

8. Wisselgren MJ, Edvinsson S, Berggren M, Larsson M. Testing methods of record linkage on Swedish censuses. Hist Methods. 2014 Jul 3;47(3):138–51. 10.1080/01615440.2014.913967

   https://doi.org/10.1080/01615440.2014.913967

9. US Census Bureau. Decennial census digitization and Linkage (DCDL) project [Internet]. Census.gov. 2022 [cited 2022Dec9]. Available from: https://www.census.gov/programs-surveys/dcdl.html

10. US Census Bureau. Census Longitudinal Infrastructure Project (CLIP) [Internet]. Census.gov. 2021 [cited 2022Dec9]. Available from: https://www.census.gov/about/adrm/linkage/projects/clip.html

11. US Census Bureau. Census Longitudinal Infrastructure Pilot projects [Internet]. Census.gov. 2021 [cited 2022Dec9]. Available from: https://www.census.gov/about/adrm/linkage/projects/clip/CLIP-Projects.html

12. Miller S, Johnson N, Wherry LR. Medicaid and mortality: new evidence from linked survey and administrative data. Q J Econ. 2021 Aug;136(3):1783–829. 10.1093/qje/qjab004

    https://doi.org/10.1093/qje/qjab004

13. Stevenson AJ, Genadek KR, Yeatman S, Mollborn S, Menken JA. The impact of contraceptive access on high school graduation. Sci Adv. 2021 May 5;7(19):eabf6732. 10.1126/sciadv.abf6732

    https://doi.org/10.1126/sciadv.abf6732

14. Alexander JT, Leibbrand C, Massey C, Tolnay S. Second-generation outcomes of the great migration. Demography. 2017 Dec 1;54(6):2249–71. 10.1007/s13524-017-0625-8

    https://doi.org/10.1007/s13524-017-0625-8

15. Leibbrand C, Massey C, Alexander JT, Genadek KR, Tolnay S. The great migration and residential segregation in American cities during the twentieth century. Soc Sci Hist. 2020;44(1):19–55. 10.1017/ssh.2019.46

    https://doi.org/10.1017/ssh.2019.46

16. Ferrie J, Massey C, Rothbaum J. Do grandparents matter? multigenerational mobility in the united states, 1940–2015. J Labor Econ. 2021 Jul 1;39(3):597–637. 10.1086/711038

    https://doi.org/10.1086/711038

17. Finlay K, Genadek KR. Measuring all-cause mortality with the Census Numident file. Am J Public Health. 2021 Jul;111(S2):S141–8. 10.2105/AJPH.2021.306217

    https://doi.org/10.2105/AJPH.2021.306217

18. Polyakova M, Udalova V, Kocks G, Genadek K, Finlay K, Finkelstein AN. Racial Disparities In Excess All-Cause Mortality During The Early COVID-19 Pandemic Varied Substantially Across States. Health Aff. 2021 Feb 1;40(2):307–16. 10.1377/hlthaff.2020.02142

    https://doi.org/10.1377/hlthaff.2020.02142

19. Miller S, Wherry LR, Mazumder B. Estimated Mortality Increases During The COVID-19 Pandemic By Socioeconomic Status, Race, And Ethnicity. Health Aff. 2021 Aug 1;40(8):1252–60. 10.1377/hlthaff.2021.00414

    https://doi.org/10.1377/hlthaff.2021.00414

20. Wagner D, Lane M. The person identification validation system (PVS): applying the Center for Administrative Records Research and Applications’(CARRA) record linkage software. Center for Administrative Records Research and Applications, US Census Bureau; 2014 Jul.

21. Massey CG, Genadek KR, Alexander JT, Gardner TK, O’Hara A. Linking the 1940 US Census with modern data. Hist Methods. 2018 Oct 2;51(4):246–57. 10.1080/01615440.2018.1507772

    https://doi.org/10.1080/01615440.2018.1507772

22. Bailey MJ, Cole C, Henderson M, Massey C. How well do automated linking methods perform? Lessons from US historical data. J Econ Lit. 2020 Dec;58(4):997–1044. 10.1257/jel.20191526

    https://doi.org/10.1257/jel.20191526

23. Helgertz J, Price J, Wellington J, Thompson KJ, Ruggles S, Fitch CA. A new strategy for linking US historical censuses: A case study for the IPUMS multigenerational longitudinal panel. Hist Methods. 2022 Jan 2;55(1):12–29. 10.1080/01615440.2021.1985027

    https://doi.org/10.1080/01615440.2021.1985027

24. Massey CG. Playing with matches: An assessment of accuracy in linked historical data. Hist Methods. 2017 Jul 3;50(3):129–43. 10.1080/01615440.2017.1288598

    https://doi.org/10.1080/01615440.2017.1288598

25. Massey C, O’hara A. Person Matching in Historical Files using the Census Bureau’s Person Validation System. Center for Administrative Records Research and Applications, US Census Bureau; 2014 Sept.

26. Johnson DS, Massey C, O’hara A. The opportunities and challenges of using administrative data linkages to evaluate mobility. Ann Am Acad Pol Soc Sci. 2015 Jan;657(1):247–64. 10.1177/0002716214552780

    https://doi.org/10.1177/0002716214552780

27. Mulrow E, Mushtaq A, Pramanik S, Fontes A. Final report: Assessment of the US Census Bureau’s Person Identification Validation System (NORC at the University of Chicago, 2011). Accessed 2022Dec9 at https://www.norc.org/Research/Projects/Pages/census-personal-validation-system-assessment-pvs.aspx.

28. Layne M, Wagner D, Rothhaas C. Estimating record linkage false match rate for the Person Identification Validation System. Center for Administrative Records Research and Applications Working Paper. 2014 Jul 1;2.

29. Rastogi S, O’Hara A. 2010 census match study. Center for Administrative Records Research and Applications Report, US Census Bureau. 2012 Nov 16.

30. Bond B, Brown JD, Luque A, O’Hara A. The nature of the bias when studying only linkable person records: Evidence from the American Community Survey. Center for Administrative Records Research and Applications Working Paper. 2014 Apr 22;8.

31. Goeken R, Huynh L, Lynch TA, Vick R. New methods of census record linking. Hist Methods. 2011 Jan 31;44(1):7–14. 10.1080/01615440.2010.517152

    https://doi.org/10.1080/01615440.2010.517152

32. Helgertz J, Price J, Wellington J, Thompson KJ, Ruggles S, Fitch CA. A new strategy for linking US historical censuses: A case study for the IPUMS multigenerational longitudinal panel. Hist Methods. 2022 Jan 2;55(1):12–29. 10.1080/01615440.2021.1985027

    https://doi.org/10.1080/01615440.2021.1985027

33. Massey CG. Creating linked historical data: An assessment of the Census Bureau’s ability to assign protected identification keys to the 1960 Census. Center for Economic Studies, US Census Bureau; 2014 Sep.

34. Genadek K, Sanders J, Stevenson A. Measuring US fertility using administrative data from the Census Bureau. Demogr Res. 2022 Jul 7;47:37–58. 10.4054/DemRes.2022.47.2

    https://doi.org/10.4054/DemRes.2022.47.2

35. Larsen AF, Headey D, Masters WA. Misreporting month of birth: diagnosis and implications for research on nutrition and early childhood in developing countries. Demography. 2019 Apr 1;56(2):707–28. 10.1007/s13524-018-0753-9

    https://doi.org/10.1007/s13524-018-0753-9

36. Song X, Massey CG, Rolf KA, Ferrie JP, Rothbaum JL, Xie Y. Long-term decline in intergenerational mobility in the United States since the 1850s. Proc Natl Acad of Sci USA. 2020 Jan 7;117(1):251–8. 10.1073/pnas.1905094116

    https://doi.org/10.1073/pnas.1905094116

37. Lowrey K, Van Hook J, Bachmeier JD, Foster TB. Leapfrogging the Melting Pot? European Immigrants’ Intergenerational Mobility across the Twentieth Century. Sociol Sci. 2021 Dec 17;8:480–512. 10.15195/v8.a23

    https://doi.org/10.15195/v8.a23

38. Steorts RC. Entity resolution with empirically motivated priors. Bayesian Anal. 2015 Dec;10(4):849–75. 10.1214/15-BA965SI

    https://doi.org/10.1214/15-BA965SI

39. Steorts RC, Hall R, Fienberg SE. A Bayesian approach to graphical record linkage and deduplication. J Ame Stat Assoc. 2016 Oct 1;111(516):1660–72. 10.1080/01621459.2015.1105807

    https://doi.org/10.1080/01621459.2015.1105807

40. Alexander JT, Fisher J, Genadek K. Digitizing Hand-Written Data with Automated Methods: A Pilot Project Using the 1990 U.S. Census. J Econ Soc Meas. 2022 Jul 1;46(2):95–108. 10.3233/JEM-220484

    https://doi.org/10.3233/JEM-220484

41. US Census Bureau. Census 2000 Data Capture. Census 2000 Testing, Experimentation, and Evaluation Program: Technical Report Series 3; 2003.

42. US Census Bureau. History: 2000 Census of Population and Housing (Volume 1). Washington, DC (USA): US Government Printing Office; 2009.

43. US Census Bureau. History: 2000 Census of Population and Housing (Volume 2). Washington, DC (USA): US Government Printing Office; 2009.

44. US Census Bureau. 2020 Census Detailed Operational Plan for: 10. Paper Data Capture (PDC) Operation. Washington, DC (USA): US Government Printing Office; 2017..

45. Kasewa S, Stenetorp P, Riedel S. Wronging a right: Generating better errors to improve grammatical error detection. arXiv preprint arXiv:1810.00668. 2018 Sep 26. 10.48550/arXiv.1810.00668

    https://doi.org/10.48550/arXiv.1810.00668