2020 Census Experiment: Real-Time 2020 Administrative Record Census Simulation Study Plan

The memorandum and attached document(s) was prepared for Census Bureau internal use. If
you have any questions regarding the use or dissemination of the information, please contact
the Stakeholder Relations Staff at dcco.stakeholder.relations.staff@census.gov.

2020 CENSUS PROGRAM INTERNAL MEMORANDUM SERIES: 2019.10.i
Date:

April 8, 2019

MEMORANDUM FOR: The Record
From:

Deborah M. Stempowski (signed April 8, 2019)
Chief, Decennial Census Management Division

Subject:

2020 Census Experiment: Real-Time 2020 Administrative Record Census
Simulation Study Plan

Contact:

Jennifer Reichert
Decennial Census Management Division
301-763-4298
jennifer.w.reichert@census.gov

This memorandum releases the final version of the 2020 Census Experiment: Real-time 2020
Administrative Record Census Simulation Study Plan, which is part of the 2020 Census Program for
Evaluations and Experiments (CPEX). For specific content related questions, you may also contact the
authors:
J. David Brown
Misty Heggeness
Center for Economic Studies
Associate Directorate for Research and Methodology
301-763-8769
301-763-7251
j.david.brown@census.gov
misty.l.heggeness@census.gov

census.gov

2020 Census
Experiment
Real-Time 2020 Administrative Record
Census Simulation Study Plan

J. David Brown, Center for Economic Studies
Misty Heggeness, Associate Directorate for Research and
Methodology

4/08/2019
Version 5.0

Page intentionally left blank.

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

Table of Contents
I.

Introduction ......................................................................................................................... 1

II.

Background ......................................................................................................................... 2

III.

Assumptions........................................................................................................................ 4

IV.

Research Questions ............................................................................................................. 5

V.

Methodology ....................................................................................................................... 5

VI.

Data Requirements ............................................................................................................ 15

VII.

Risks.................................................................................................................................. 18

VIII.

Limitations ........................................................................................................................ 18

IX.

Issues That Need to be Resolved ...................................................................................... 19

X.

Division Responsibilities .................................................................................................. 19

XI.

Milestone Schedule ........................................................................................................... 19

XII.

Review/Approval Table .................................................................................................... 20

XIII.

Document Revision and Version Control History ............................................................ 20

XIV. Glossary of Acronyms ...................................................................................................... 21
XV.

References ......................................................................................................................... 21

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

Page intentionally left blank.

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

I.

Introduction

This project will conduct several real-time administrative record census simulations in 2020, using
all administrative records ingested by the Census Bureau as of a certain date that can inform about
the Census Day population. If extended, the project will produce annual real-time administrative
record census simulations after 2020 and provide predictions for where survey-style data collection
will be necessary to complete an administrative record enumeration in 2030.
Our project will build on past administrative record enumeration research by including additional
administrative data sources not previously available, using more accurate and comprehensive
person linkage, and employing more powerful models predicting people’s locations to increase
coverage and person-place accuracy.1 Unlike previous studies, it will be conducted in real time in
2020, which will show how the population statistics compare between an administrative record
census and survey-style collection, in the same time frame.2 It will also show how long it takes to
execute an administrative record census and what the most time-consuming parts of the process
are. The project will compare person-level, housing unit-level, and hybrid approaches to
conducting an administrative record census, which will inform 2030 design decisions about
whether to transition from a housing unit- to a person-based or hybrid method within the legal
governance, rules, and regulations of conducting a full count census.
The project will expand upon and evaluate 2020 Nonresponse Followup (NRFU) operation
administrative records innovations. We will use many more data sources, an enhanced Person
Identification Validation System (PVS) process, and different methods for assigning people to
locations relative to those planned for administrative record enumeration in 2020 NRFU. We will
also evaluate the PVS linking methodology planned for use in 2020 operations by estimating false
match and nonmatch rates and the effect of those errors on the statistics.
Models will be developed to predict where each person is located on Census Day, as well as where
administrative record enumeration is most likely to diverge from survey-style enumeration. The
latter can be used to target where survey-style data collection can most usefully supplement
administrative record enumeration.
Comparisons of various administrative record census simulations to the 2020 Census will show
the relative strengths and weaknesses of administrative record coverage and accuracy3 by
demographic characteristics, location, and level of geographic aggregation. The results from
person-based, housing unit-based, and hybrid approaches to constructing an administrative record
census will be examined. The effects of enhancing the record linkage infrastructure will be
1

It may not be possible to use these improvements in the 2020 production NRFU, because the deadlines to finalize
2020 production methods are much earlier than those for this experiment.
2
Doing the simulations in the same timeframe will demonstrate the feasibility of conducting an administrative record
census in similar conditions to the actual census. Relative to a post-2020 census study, this will also reduce concern
that the administrative record similations will borrow from the 2020 Census, tainting the experiment.
3
To assess coverage, we will compare how many people are enumerated regardless of whether they are the same
people in the same places; to assess accuracy, we will examine whether people are enumerated in the same places.
1

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

identified. We will study how well the models predict where administrative record enumeration
and the 2020 Census agree and diverge.
If extended beyond 2020, the project will conduct annual real-time simulations and produce annual
population estimates, which will be compared with other Census Bureau intercensal population
estimates. This analysis will inform decisions on whether to continue researching and conducting
administrative record enumeration on an annual basis after 2020. The overarching questions are
whether we are able to conduct a complete administrative records census and with what level of
accuracy. If so, what methods are best both in terms of infrastructure and modeling techniques for
implementing a strategy that will ensure the most accurate count of the population. Our project
will address several subquestions to inform these overarching questions:









Can a record linkage methodology accurately cover people not in the Numident or
Individual Taxpayer Identification Numbers (ITINs)? What is its error rate?
What are the best models for predicting which administrative record address is the Census
Day address, and how well do they perform?
Which data sources are the most useful for coverage and accuracy and thereby pursued
with the highest priority?
How similar are the statistics produced by an administrative record census to the surveystyle 2020 Census?
How long does it take to implement an administrative record census, and what parts of the
process are most time-consuming?
Which method for conducting an administrative record census can produce statistics most
similar to the 2020 Census?
What are the cost-statistical similarity trade-offs when using different combinations of
survey collection and administrative records?
What subpopulations and geographic areas should be targeted for alternative forms of data
collection?

Finally, we will refine models of person and housing unit transition rates. Repeated annual recordsbased censuses will facilitate the production of more powerful predictors of where survey-style
data collection is needed to complete administrative record enumeration.

II.

Background

To reduce costs, many countries use administrative data to assist in censuses or as a replacement
to traditional censuses (Farber and Leggieri 2002, Ralphs and Tutton 2011). For several decades
administrative data have been used in U.S. Census Bureau programs for population, economic,
small-area income and poverty, and health insurance estimates, but they have not been used
extensively in decennial census operations. For the 2020 Census, administrative records will be
used to reduce NRFU fieldwork, one of the largest expenses in the decennial census.
There is pressure to continue to improve the accuracy and reduce costs of a decennial census
beyond the planned administrative record use in 2020. The JASON (2016) report “suggests a
2

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

paradigm shift in the way the Census Bureau conceptualizes the enumeration… a census that starts
with administrative records involves identifying individuals and assigning them to their
appropriate residences as opposed to the historical process of identifying residences and then
populating them.” As such, the report recommended that “the Census Bureau consider starting the
2030 Census with an ‘in-office’ enumeration of the population using existing government
administrative records.”
Previous research efforts have evaluated the feasibility of a 100 percent records-based census. The
Statistical Administrative Records System (StARS) was developed from select federal data sources
in 1999. Decennial census research used these data to evaluate address and person counts relative
to Census 2000, and for a field test (the Administrative Records Census Experiment or AREX
2000) that simulated a census in several counties that was compared with Census 2000. The
research found that while address and person counts in StARS were relatively close to the counts
in Census 2000 at the national level, results varied significantly by region (Farber and Leggieri
2002). The AREX 2000 research compared Census 2000 results in five counties with
administrative data in StARS and found that the administrative data undercounted children,
overcounted elderly populations, had difficulty identifying a correct residence of movers, and that
a 15-month time gap between the administrative and census data likely contributed to the
difficulties of using administrative records to enumerate the population (Bauder and Judson, 2003).
The AREX 2000 research compared a person-based approach and a hybrid person- and housing
unit-based approach to constructing an administrative record census. In their person-based
approach, they assigned each person to a single block. Their hybrid method assigned each person
to their “best” address according to the StARS algorithm, provided that it was included in the 2000
production Master Address File (MAF). Production MAF housing units lacking people in the
administrative record census were selected for follow-up survey-style data collection, and their
Census 2000 population count was used in the simulation (making the results the same in the
Census 2000 and the simulation for those housing units).
The 2010 Census Match Study (Rastogi and O’Hara, 2012) linked person, address, and personaddress records to 2010 Census data to assess the quality and coverage of administrative data and
feasibility of a records-based census. The study showed significant improvement over the AREX
2000 results in matching addresses found in administrative records to addresses in the 2010 Census
(92.6 percent). It was also able to match 88.6 percent of all individuals in the 2010 Census to at
least one administrative record, and 77 percent were placed at the same address. The report also
evaluated quality and coverage by Hispanic origin, race, sex, and age response data in
administrative records relative to the 2010 Census.
While more timely and varied sources of data were available than in the StARS data, the 2010
Census Match Study findings reaffirmed the challenges of conducting a records-based census.
Some data sources such as Social Security Administration (SSA) and Medicaid data did not have
addresses, and other data sources had addresses that include a post office box or other
nonresidential addresses, both of which making it difficult to place all people enumerated in
administrative records at a residential location. Furthermore, some individuals were associated

3

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

with multiple addresses, and the address selected often did not align with the address in the 2010
Census.
Our project will build on the AREX 2000 and 2010 Census Match Study by including additional
administrative data sources, assigning unique identifiers more accurately to more people, and
employing more powerful models predicting people’s locations, which should result in increased
coverage and person-place accuracy. Unlike previous studies, it will be conducted in real time in
2020, in parallel with the actual census. The project will compare person-level, housing unit-level,
and hybrid approaches to conducting an administrative record census, which will inform 2030
design decisions about whether to transition from a housing unit- to a person-based or hybrid
method. We will simulate doing a census that is 100 percent records-based, supporting that 2030
guiding principle. The simulation tabulations will produce state population counts to fulfill the
constitutional mandate for apportionment and citizen voting age population by race and ethnicity
at the block level to fulfill the Voting Rights Act requirement.
The project will build upon and evaluate 2020 Census innovations regarding the use of
administrative records in NRFU. This project will use a greatly expanded set of data sources and
person validation, as well as different methods for assigning persons to locations than those
planned for administrative record enumeration in 2020 Census NRFU. We can compare our
simulation results with the administrative record enumeration in NRFU for the same housing units.
We can also investigate whether and the extent to which administrative record enumeration could
be expanded without sacrificing quality.
The interventions with 2020 Census processes needed by the project are access to the 2020
Census production MAF, Decennial Response File (DRF), Census Unedited File (CUF), Census
Edited File (CEF), and Post-Enumeration Survey (PES) as soon as they are completed.

III.

Assumptions

1. The project team will obtain and maintain adequate staff resources.
2. The Internal Revenue Service (IRS) will approve the Predominant Purpose Statement
(PPS) for use of Federal Tax Information (FTI) in a timely manner.
3. The project team will obtain adequate funding for computing resources to begin the
project in the Integrated Research Environment (IRE).
4. The Census Bureau will fund and complete the necessary information technology (IT)
requirements to move Title 26 data to the cloud.
5. The Census Bureau will continue to acquire administrative records such as Internal
Revenue Service (IRS) and state program data and acquire new key data sources such as
passport and visa data in a timely manner.

4

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

IV.

Research Questions

1. Can a record linkage methodology accurately cover people not in the Numident or ITINs?
What is its error rate?
2. What are the best models for predicting which administrative record address is the Census
Day address, and how well do they perform?
3. Which data sources are the most useful for coverage and accuracy and thereby pursued
with the highest priority?
4. How long does it take to implement an administrative record census, and what parts of the
process are most time-consuming?
5. How similar are the statistics produced by an administrative record census to the surveystyle 2020 Census?
6. What are the cost-statistical similarity trade-offs when using different combinations of
survey collection and administrative records?
7. Which method for conducting an administrative record census can produce statistics most
similar to the 2020 Census?
8. What subpopulations and geographic areas should be targeted for survey-style data
collection?

V.

Methodology

This section describes how we plan to address the research questions, followed by a discussion
of the implications of the results for future testing and 2030 Census design decisions.
A. Evaluation Design
The main steps we propose to implement in the research are the following:
1. Obtain additional sources of administrative records beyond the Census Bureau’s current
inventory.
2. Enhance the record linkage infrastructure with additional data and methodological
improvements.
3. Test the validity of assumptions incorporated in PVS, and test the application of modern entity
resolution models of record linkage.
4. Process administrative records.
5. Estimate person-place models.
6. Predict the relative accuracy of administrative record enumeration by person and housing unit,
including the error from linkage.
7. Conduct real-time administrative record census simulations in 2020.
8. Assess quality and coverage of the administrative record censuses in comparison to the 2020
Census.4
4

We use the 2020 Census as comparison data, recognizing that they are not error free. We plan to also use the 2020
Post-Enumeration Survey (PES) as an additional comparator for statistics that the PES produces (nationally by race,
5

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

We will make a number of improvements in the administrative record infrastructure, including
obtaining additional administrative data sources to reduce coverage gaps and improve model
prediction, enhancing record linkage procedures to link more records, and improving model
estimation techniques. Unless significant progress is made on these fronts, the discrepancies
identified in the 2010 Census Match Study (Rastogi and O’Hara, 2012) compared with the 2010
Census are likely to remain too high to change that study’s conclusion that the Census Bureau is
not yet ready to convert to using administrative records as the primary enumeration method.
These steps are described below:
Obtain additional sources of administrative records: We will obtain and integrate additional
administrative record sources to plug coverage gaps and improve the predictive power of the
models. The 2010 Census Match Study (Rastogi and O’Hara, 2012) and this project’s analysis of
differences between administrative record enumeration and the 2020 Census can inform us where
additional sources can be most beneficial. We will conduct analysis of the relative contributions
of the current sources to coverage and predictive power, which will inform decisions on which
current sources could be dropped to free up funds for new acquisitions, if necessary.5
New sources that could be particularly valuable include state driver’s licenses; voter registration
data; state-level low-income assistance program participation; and data sources provided by local
governments.6
As additional sources are added, the person-place models will be reestimated both so that people
found only in the new sources can be included and to harness this information to improve
prediction.7
Enhance the record linkage infrastructure with additional data and methodological
improvements: One of the most important reasons for administrative record coverage gaps is the
fact that several million U.S. residents are either not in the SSA Numident, or the personally
identifiable information (PII) in their other administrative records is different from how it appears
in the Numident. We will add U.S. Citizen and Immigration Services (USCIS) legal permanent
resident and naturalization data, Customs and Border Protection (CBP) visa data, and State
Department passport data to the reference files,8 covering some of the people not in the Numident.9

ethnicity, and age group). Below we discuss separate comparisons with 2020 Census records for which there are
different probabilities of error. Note that in NRFU housing units enumerated via administrative records in 2020
Census production, comparisons with survey-style data collection will be limited to the PES.
5
Decisions to discontinue sources will need to factor in other Census Bureau needs and uses of those sources.
6
The availability of data sources for the project will be dependent upon maintaining and in some cases revising
existing agreements with current data providers and obtaining them for new sources.
7
Note that we will want to reestimate the models periodically anyway, because relationships between administrative
record and survey-style enumeration could change over time. Our ability to do this throughout the decade will depend
on continuing survey-style collection of ACS housing unit roster information.
8
Reference files are the files used to validate the PII in a person record. The current reference files in PVS are the
Numident and ITINs.
9
Note that the intention to do this is already public knowledge, as the March 1, 2018, memo from the Census
Bureau to Commerce Secretary Wilbur Ross mentioning this has been made public as part of a FOIA request.
6

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

As a way to facilitate the linkage of the remaining people who are in administrative records, but
not in any of the reference files listed above, we will experiment with the BigMatch linkage
procedure developed by William Winkler, as well as other entity resolution approaches (Steorts
et. al. 2016). These approaches could potentially be more accurate as well. The methods are well
suited to linking records across multiple files. The Steorts et al. (2016) approach is implemented
by fitting models using Bayesian methods. The results of the estimation include posterior
probabilities that records on incoming files are of individuals not already on the existing reference
files.
Test the validity of PVS assumptions and the application of modern entity resolution models:
We will evaluate the PVS and entity resolution methodologies by estimating false match and false
nonmatch rates and the effect of those errors on the statistics (e.g., the population count in a
geographic area or the percentage of the population with particular demographic characteristics).
Process administrative records: The Census Bureau will ingest administrative records, PVS the
persons to assign Protected Identification Keys (PIKs)10, and MAF-match the addresses to assign
MAFIDs, latitude and longitude, and other geolocational codes. We will start with all
administrative data that are currently available containing personally identifying information (PII)
and geographic location.11 People that the Numident or other reliable sources indicate are deceased
will be dropped. We will use PIKs to have some confidence that the person exists and to be able
to unduplicate the person’s records to prevent multiple counting.
Estimate person-place models: We will estimate person-place models to produce a probability
that a PIK is at a particular location, for each PIK-location pair (or person-address pair, where the
person has been assigned unique identification number, and the address has a MAFID or other
geocodes).12 We will use ACS panels for the same years as the vintages of the administrative

10

As mentioned above, we will test other methods for assigning unique person identifiers as well.
A candidate list includes Internal Revenue Service (IRS) Individual Income Tax Returns 1040 and Information
Returns 1099; Housing and Urban Development (HUD) Public and Indian Housing Information Center (PIC) and
Tenant Rental Assistance Certification System (TRACS), and Computerized Homes Underwriting Management
System (CHUMS); Social Security Administration Supplemental Security Record (SSR), Numident, and the Kidlink
file derived from the Numident; Center for Medicare and Medicaid Services (CMS) Medicare Enrollment Database
(MEDB) and Transformed Medicaid Statistical Information System (T-MSIS); Indian Health Service (IHS) Patient
Registration System; U.S. Postal Service National Change of Address file; Experian; Targus/Neustar; Veteran Service
Group of Illinois (VSGI); InfoGroup; Melissa Data; Health and Human Services Child Care Development Fund
(CCDF); Bureau of Justice Statistics National Corrections Reporting Program (NCRP) and Post-Custody Community
Supervision (PCCP); Bureau of Prisons Permanent Release Database; Veterans Affairs; Alaska Permanent Fund
Dividend File; Supplemental Nutrition Assistance Program (SNAP); Temporary Assistance for Needy Families
(TANF); Special Supplemental Nutrition Program for Women, Infants and Children (WIC); Homeless Management
Information Systems (HMIS); Low-Income Home Energy Assistance Program; utilities records data; Corelogic;
RealtyTrac; and DAR Partners.
12
This estimation process will be repeated using different definitions of location (housing unit, block, tract, ZIP code,
county, state, and latitude and longitude coordinates with different degrees of precision). By doing this we can
investigate the possibility that discrepancies between administrative record and survey-style enumeration vary by
geographic level. Since most moves are within small geographic areas, person-place discrepancies are likely to be
much smaller at higher levels of geography.
11

7

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

record data to fit the models.13 First-stage logistic regressions will be estimated separately for each
administrative record source. PIKs in the ACS that have administrative records in the particular
source will be included in the regression. The dependent variable equals one if the location in the
administrative record is the same as the ACS location for the PIK, and zero otherwise. We will
estimate separate regressions for each source, because variables that can help predict if the location
is the ACS location vary across sources. For example, IRS 1040’s contain variables for filing
status, whether a child is living elsewhere, and the week the return was processed (measuring the
vintage). Veteran Service Group of Illinois (VSGI) contains household income, owner vs. renter
status, and length of residence. For sources with several years of data, such as IRS 1040s, we will
construct variables for whether the person was at this location or a different one in past years to
capture the person’s mobility. Variables on the person’s age, sex, race/ethnicity, and citizenship
status (primarily from the Numident) will be included in all these regressions.
A second-stage regression includes all administrative record PIK-location pairs for PIKs in the
ACS that have administrative records in any of the administrative record sources. Once again the
dependent variable equals one if the PIK-location administrative record pair is the same as the
ACS PIK-location pair for each PIK, and zero otherwise. Characteristics of the administrative
record location are included here, such as the housing unit type, U.S. Postal Service delivery
sequence file information, and the number of other PIKs with administrative records with this
location14. We will include indicator-indicator variables for whether each particular administrative
record source lists the person at this observation’s location (here) or at one or more other locations
(elsewhere). These indicator-indicators are also separately interacted with the individual match
propensities obtained from the first-stage regression corresponding to the indicator source for the
PIK-location pair.15 The rationale for the interactions is that the location where a source lists a
person should be more likely to be the ACS location if the first-stage match propensity is high.
Including indicators for each of the sources captures the degree of agreement across sources about
the person’s location. The coefficients from these models are used to produce person-place match
probabilities for all PIK-location pairs eligible for administrative record enumeration.
We will control overfitting the models by performing k-fold cross-validation. It splits the data
randomly into k partitions. For each partition it fits the model using the other k-1 groups, then uses
the generated parameters to predict the dependent variable in the unused group. We will also test
the models by using the parameters to predict the dependent variable in future years of ACS data
and the 2020 Census.16,17
13

The models implicitly assume that the ACS household roster is accurate, which may not be the case. We are
unaware of a more accurate alternative, however.
14
Large numbers of people may have the same administrative records address, e.g., people using recreational vehicle
(RV) mail forwarding services. Including this variable should help address this issue - their person-place probabilities
are likely to be low, making them top candidates for follow-up.
15
In cases where the administrative record source has multiple other MAFIDs for the PIK, we will sum up the
propensities for the other MAFIDs here.
16
For example, we can apply models fit on 2017 ACS data to 2018 administrative records to produce probabilities
that people in the 2018 ACS are located at various addresses. Then we will see how well those probabilities predict
the actual 2018 ACS address for the person. The same thing will be done with the 2020 Census, with the added
benefit of being able to do it on the full population rather than a survey sample.
17
Note also that unlike in Rastogi and O’Hara (2012) and Brown, Childs, and O’Hara (2015), all the model coefficients
will be applied to future administrative records, not to earlier administrative records used to fit the models.
8

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

Predict relative accuracy of administrative record enumeration: We will produce several
different predictors, some geared toward person-based follow-up data collection and others for
location-based follow-up. For each PIK, we will calculate the maximum probability of being at a
particular location among all their locations in administrative records, using predictions from our
person-place models fit with earlier data vintages. This predicts the probability that if counted in
this location in an administrative record census, the person will not be an enumeration error
(counted in the wrong place). Any person-based follow-up data collection could focus on
individuals with the lowest values for this measure. For location-based follow-up, we will calculate
a location-based enumeration error predictor that again uses the PIK maximum location probability
described above, but now taking its average value across all PIKs at the location.
To predict where omissions are more likely, we estimate housing unit-level administrative record
coverage regressions. The dependent variable when fitting the model for one version is equal to
one if at least one person in the ACS household roster cannot be linked to the set of administrative
records we plan to use in the simulations. A second version is a count regression for the number
of people in the ACS household roster who cannot be linked. Explanatory variables include the
number of un-PIKed administrative records with this location, indicators for the sources of the unPIKed records, the number of PIKed persons with this location, demographic characteristics of the
un-PIKed administrative records, and characteristics of the location, such as housing unit type and
U.S. Postal Service delivery sequence file information.18 The coefficients from these regressions
will be used to produce out-of-sample predictions of incidence and number of omissions for all
locations. Another measure we will use is the standard deviation of the population count in the
particular location across repetitions of the simulation (explained in the next section).19
We can run administrative record census simulations in 2019 to generate additional measures. We
will study person and location dynamics, such as the share of people who move across locations
(housing unit, block, tract, ZIP code, county, state, by different levels of precision of latitude and
longitude, and between the U.S. and other countries), the share that appear in one year but not the
next for reasons other than being deceased, and the share that don’t appear one year but appear the
next for reasons other than birth. The latter two categories can be because of either emigration and
immigration or administrative record coverage problems. We can model these transitions using
person and housing unit characteristics, as well as past transitions as predictors. These transition
probabilities could be used to supplement the measures described above for targeting survey-style
data collection in future census tests.
Two additional indicators of where follow-up survey-style data collection may be most useful are
housing units in the 2020 production MAF lacking anyone in administrative records20 and
addresses not in the 2020 production MAF but with people in administrative records. Locations
with high concentrations of housing units of either type may be candidates for follow-up.

18

Unlike PIKed records, we cannot obtain un-PIKed record demographic information from the Numident, but some
of the other sources also contain demographic information.
19
A higher standard deviation suggests less confidence in the administrative record count and thus greater
misalignment.
20
Bauder and Judson (2003) choose these housing units for follow-up in their simulation.
9

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

Conduct real-time simulations in 2020: How long does it take to implement an administrative
records census?21 What parts of the process are most time-consuming? Is it feasible and useful to
do real-time updating of entity resolution?22 Answers to these questions can help identify what
research needs to be done in 2021-2025 to improve administrative record census execution.23 We
will process all administrative records available on a particular date (we will produce different
versions with different deadline dates).24 PIKs, MAFIDs, and other geocodes will be placed on the
records. We will implement housing unit-based, person-based, and hybrid approaches and
compare them.
For the housing unit-based approach, we will place people at each of their administrative record
addresses that are in the 2020 Census production MAF. Person unduplication will be done within
each particular housing unit, but not across housing units. Some people will be counted multiple
times in different housing units, as in survey-style censuses. Individuals without administrative
record addresses in the 2020 Census production MAF will not be counted.
In person-based approaches, we will create variables used in our person-place models. The personplace model coefficients will be applied to these variables to produce the probability that each
PIK-location pairing is correct. After dropping PIK-location pairs for locations not included in the
particular simulation (e.g., locations that can only be determined at the state level for a simulation
that uses location below the state level), the remaining PIK-location probabilities will be rescaled
to sum to one for each PIK. Multiple replications of the census will be constructed.25 Each PIK is
placed at one location per replication (and thus counted only once), and the location is selected
randomly among the person’s locations using their location probability as the weight. 26 In different
variants of this approach we will change the geographic aggregation (housing unit, block, tract,
ZIP code, county, state, and latitude and longitude coordinates of different degrees of precision).
The locations will not be restricted to ones found in the 2020 Census production MAF. Individuals
without administrative record addresses that can be geocoded to the level of geography used in the
particular simulation will not be counted.
Our hybrid approach is like our person-based approaches, but where the locations are limited to
housing units in the 2020 Census production MAF. PIK-location pairs where the location cannot

21

The record linkage process is likely to take approximately a month. The subsequent data cleaning and application
of model coefficients is likely to take 2-3 weeks. Tabulation of the statistics may take a day or two.
22
Prior to 2020, we will experiment with entity resolution methods where the algorithms are refined in real-time as
new records arrive. If our tests of these methods show good results (sufficient speed and accuracy), we will
implement this in the 2020 simulations.
23
Speed is desirable, because the faster it can be completed, the later the pull date can be for administrative records
used in the enumeration.
24
Individuals that the administrative records indicate are deceased will be dropped from all simulations.
25
The exact number will depend on how quickly a replication can be completed. The larger the number, the more
informative the variance calculations will be.
26
For PIKs with multiple locations, this means they may be placed at different locations across replications. This is
similar to the idea in MITRE and Santa Fe Institute (2016) that “a person has a certain chance of being in a certain
location at a certain time. The quantity of persons in a given region would then be the summation of the mass of the
probability distributions.”
10

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

be traced to a 2020 Census production MAFID will be dropped, and the remaining PIK-location
pair probabilities will be rescaled to sum to one.27
We can also produce versions requiring that at least one of the administrative record sources
putting a person at the address is a federal government source.28
The housing unit-based approach most closely mimics survey collection, so it may more closely
match the 2020 Census statistics. A person-based approach has the potential to improve
enumeration quality where survey collection contains errors, for example, by making greater effort
to count people only once.
Demographic characteristics from the CES best race data,29 the Master Demographics Database,
and administrative records will be attached to each PIK in each of the simulations. Housing tenure
information from administrative records will be attached to each MAFID for all simulations that
use housing unit as the location.
The population count for a location is the number of PIKs assigned to it in the replication. In
simulations restricted to MAFIDs in the 2020 production MAF, each MAFID will be classified as
occupied in a replication if at least one PIK is assigned to it, and otherwise it will be classified as
unoccupied. No distinction will be made between vacant and delete. Determining the number of
unoccupied housing units without the use of the 2020 production MAF is out of the scope of this
project.
For each simulation (except the housing unit-based approach, which will have just one replication
and thus will not have a distribution),30 we will calculate moments of the distribution (e.g., mean
and standard deviation) of overall population count and by sex, age group, race, ethnicity, and
citizenship by geography (MAFID, block, tract, ZIP code, county, state, and for the 50 states plus
the District of Columbia). We will also calculate moments for number of occupied housing units
by geography.
Compare simulations to 2020 Census: How does administrative record coverage compare with
the 2020 Census (total counts, as well as and omissions in administrative records and omissions in
the 2020 Census), overall and by demographics and location? To what extent do the locations of
individuals common to the 2020 Census and the administrative records census agree, overall and
by demographics and location? How does the degree of agreement between a simulation and the
2020 Census vary by the geographic aggregation of the counts? How does the degree of agreement
27

All approaches using the housing unit as the location will exclude group quarters, whereas the ones using other
geocodes will include them.
28
One way to justify using an administrative record census is that people have provided information to the federal
government already, and the Census Bureau is part of the federal government. This argument will be stronger if the
methodology requires that a federal government-sourced administrative record puts the person at the address. With
this methodology the state, local, and commercial data role would be to improve prediction, helping to choose between
different addresses when federal government sources disagree with each other.
29
The CES best race data are sourced from both Title 13 survey data and administrative records.
30
The probability that each person is assigned to the survey location, as derived from the person-place models, can
be used to produce a measure of uncertainty about the population count in each housing unit. This will be our main
measure of uncertainty for the housing unit-based approach, since we will not run multiple replications for it.
11

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

vary by approach (housing unit-based vs. person-based vs. hybrid) and by the geographic
aggregation of the location a person is assigned to in the simulation (placing a person in a housing
unit vs. block vs. tract vs. ZIP code vs. county vs. state vs. different precision levels of latitude and
longitude)?
We will compare the 2020 Census population counts with the mean counts across simulations at
the national level, as well as moments of the distributions of the degree of count agreement at the
MAFID, block, tract, ZIP code, county, state, national levels, and at different levels of precision
for latitude and longitude coordinates. We will identify the extent to which the simulations omit
people included in the 2020 Census and include people omitted from the 2020 Census. We will
study person-location agreement rates (a measure of enumeration errors) between the 2020 Census
and different simulations among those counted in both the 2020 Census and the simulation to
which it is being compared. In addition, count and person-location agreement rate comparisons
will be made by sex, age group, race, ethnicity, and citizenship.
How does the degree of similarity in the statistics depend on availability of state-level
administrative records, such as SNAP and TANF? We will compare the degree of similarity with
the 2020 Census in states with and without these files. In the states where we have these files, we
can create additional simulations that remove these files to see how much they matter for the
statsitics. This will inform how valuable the state administrative record files are for administrative
record enumeration.
How does the degree of agreement between simulations and the 2020 Census differ by 2020
Census response mode and whether there is reason to doubt the housing unit’s 2020 Census
response accuracy? We will make separate comparisons by 2020 Census response mode, which
will show the effects of substituting administrative records enumeration for each particular census
operation (e.g., maybe administrative record enumeration would be a better substitute for NRFU
than group quarters or update/enumerate). Comparisons will be made for housing units with no
2020 Census discrepancies, as defined by Brown, Childs, and O’Hara (2015), and housing units
with at least one discrepancy.31 This will allow us to see how the population count differences vary
with survey collection difficulties,32 and it can measure the extent to which supplementing
administrative record enumeration with survey collection can improve accuracy. It could
illuminate where administrative record enumeration might improve accuracy relative to surveystyle data collection.33

Discrepancies include counting a person who isn’t alive on Census Day, counting the same person at another
location, count imputation, proxy response with occupied status, at least one person without a PIK, different housing
unit status or count across responses, move in or move out dates in the National Change of Address file suggests the
person wasn’t living at that location on Census Day, the count is not equal to the number of listed persons, the
undercount question is answered affirmatively, and the overcount question is answered affirmatively.
32
Differences between administrative record and survey-style enumeration results could reflect errors in survey-style
collection rather than administrative records when the survey-style collection suffers from discrepancies.
33
For example, suppose a housing unit has multiple 2020 Census production responses with discrepant counts, and
the administrative records for the housing units are associated with high predicted probabilities of being at that address
(in other words, they appear to be of high quality). In such a case it is likely that the administrative records would
provide a more accurate enumeration than the production responses.
31

12

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

What are the effects of enhancing the record linkage infrastructure on coverage and person-place
agreement? We will distinguish PIKs by which reference file was used to validate them, then study
coverage changes if PIKs from particular reference files are dropped. We will compare personplace agreement rates with the 2020 Census by PIK-reference file groups. We can also study how
these rates vary by PVS score, which measures the degree of confidence in the record’s validation.
When testing the entity resolution linkage methods, we can measure confidence in PIK
assignments by posterior probabilities, and those probabilities can be propagated through the
model to give posterior distributions (and hence measures of uncertainty) of the totals. We can
assess the validity of the linkage method by how well these uncertainties relate to the actual personplace agreement rates.
How well do the different methods of predicting the degree of agreement between administrative
records simulations and the 2020 Census at the person and location levels perform? We will show
how quickly the 2020 Census and each simulation’s results converge as more people, housing
units, or higher-level locations are assigned to follow-up survey-style data collection.34 For the
person-location probability measure, we will start with no follow-up, then add people to followup beginning with those with no location in the particular simulation (their administrative record
address could not be geocoded to the level used in the simulation), then add people based on their
person-location probabilities, ranked from low to high, until all are assigned to follow-up.
Similarly, for the housing unit- and higher-level measures, we will start with housing units with
no one assigned to them, then add housing units based on their probabilities, ranked from more
anticipated differences to fewer. This will inform the extent to which survey-style data collection
can be targeted at particular individuals, housing units, or geographic areas where administrative
record enumeration is most different from survey-style collection. The better the predictions, the
smaller the amount of survey collection that is needed to achieve a certain quality level.
Using these housing unit rankings, we will calculate the cost of enumerating different shares of
them by survey methods vs. using administrative records. This will inform trade-offs between cost
savings and statistical differences with survey collection.
How does the degree of agreement between the 2020 Census and the simulations compare to
agreement between the 2020 Census and the Census Bureau intercensal estimates products such
as demographic analysis (DA) and ACS estimates? Making such comparisons can inform whether
administrative records have the potential to improve upon other intercensal population estimates.
Informed by this analysis, a decision will be made on whether to continue researching and
conducting administrative record enumeration on an annual basis after 2020. The analysis will also
shed light on which approaches are most promising, which additional data sources are the highest
priorities, and what record linkage and modeling improvements are needed. If the decision is to
continue this line of research, then the next two steps will be taken.

34

Following the 2000 AREX methodology (Bauder and Judson, 2003), we could replace the simulation result with
the 2020 Census result for people or locations targeted for follow-up survey-style data collection. Note, however, that
this exercise will be less informative for housing units where administrative record enumeration or vacancy
determination is applied in 2020 NRFU, since survey-style data collection is not done for those cases.
13

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

Table 1 provides a summary of the comparisons.
Table 1. Summary of Comparisons
Measure
How Tabulated
Population Count in 2020 Census, PES, and National, State, Race/Ethnicity, Citizenship,
Each Type of Simulation
Age Groups, Sex, Census Enumeration
Method, Census Discrepancy Type, Reference
File
Source
for
Person
Linkage,
Administrative Record Source
Omissions in 2020 Census and Each Type of National, Race/Ethnicity, Citizenship, Age
Simulation
Groups, Sex, Census Enumeration Method,
Census Discrepancy Type
Person-Place Agreement Rate between National, Race/Ethnicity, Citizenship, Age
Administrative Record Simulations and 2020 Groups, Sex, Census Enumeration Method,
Census
Person-Place Probability Groups, Census
Discrepancy Type
Population Count in 2020 Census Alone vs. Speed of convergence, using different surveyDifferent
Combinations
of
Preferred style targeting measures
Simulation and 2020 Census
Population Count in 2020 Census, Preferred National Overall Count (not for ACS), Sex,
Simulation, DA, and ACS
Age Groups, Race/Ethnicity

B. Interventions with the 2020 Census
This project will not intervene with the 2020 Census.

C. Implications for 2030 Census Design Decisions and Future Research and Testing
The results of this study will inform decisions about the extent to which future censuses should
rely on administrative records to enumerate populations. This study could also lead to further
results in intercensal years:
1. Conduct real-time simulations in 2021 and future years.
2. Produce annual population estimates.
Conduct real-time simulations in future years: We will follow the same steps as in the 2020
simulations, but focusing on approaches that produce the best results based on comparisons with
the 2020 Census, including enhancements (additional data sources to address coverage gaps,
further enhanced record linkage, and improved models using the 2020 Census in the estimation)
to address weaknesses in the 2020 simulations.
Produce annual population estimates: The annual simulations could be used to produce annual
population estimates at different levels of geography, if so desired. Adjustment factors based on
14

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

the comparison between the 2020 Census and the 2020 simulations could be applied to the annual
administrative record counts. Person and housing unit transition rates across years will also be
calculated. As the number of points in time increases, the accuracy of transition prediction models
should improve.

VI.

Data Requirements

Data File/Report

Source

Purpose

Expected
Delivery
Date

IRS Form 1040

IRS

available

IRS 1099

IRS

IRS 1099-R

IRS

IRS W-2

IRS

2000 Decennial PIK Crosswalk
Census 2000
2000 Hundred Percent Detail File
2000 BOC PIK Crosswalk
2000 Census Unedited File (CUF)
2010 Census Unedited Files
2010 Census Edited Files
2010 Census PIK Crosswalk
2010 Census Undeliverable-As-Addressed
2018 End-to-End Test
2020 Census DRF

Census Bureau
Census Bureau
Census Bureau
Census Bureau
Census Bureau
Census Bureau
Census Bureau
Census Bureau
Census Bureau
Census Bureau
Census Bureau

2020 Census CUF
2020 Census CEF

Census Bureau
Census Bureau

2020 Census PES
2000-2019 ACS

Census Bureau
Census Bureau

ACS PIK Crosswalks
Current Population Survey Annual Social and
Economic Supplement (CPS ASEC)
Current Population Survey Basic Monthly Files
Survey of Income and Program Participation (SIPP)
SIPP Crosswalk Files
CPS PIK Crosswalk Files
Census Kidlink
Master Address File Extracts

Census Bureau
Census Bureau

enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
record linkage
prediction
prediction
record linkage
prediction
prediction
prediction
record linkage
prediction
prediction
Census cost
and quality
assessment
comparison
CVAP
production
comparison
prediction,
demographic
characteristics
record linkage
prediction

available
available
available
available
available
available

Master Address File Auxiliary Reference File

Census Bureau

prediction
prediction
record linkage
record linkage
record linkage
housing frame,
prediction
address
processing

Census Bureau
Census Bureau
Census Bureau
Census Bureau
Census Bureau
Census Bureau

15

available
available
available
available
available
available
available
available
available
available
available
available
07/01/2019
09/01/2020

10/01/2020
12/01/2020
02/01/2021
available

available
available

available

Real-Time 2020 Administrative Record Census Simulation
Version 5.0
Geocoded Address Extract File

Census Bureau

address
processing
demographic
characteristics
prediction

available

Master Demographics File

Census Bureau

2010 Census Coverage Measurement Estimate and
Results files
Title 13 Race and Ethnicity File

Auxiliary Reference
File
Auxiliary Reference
File
Auxiliary Reference
File
Auxiliary Reference
File
Auxiliary Reference
File
Auxiliary Reference
File
Auxiliary Reference
File
Social Security
Administration

demographic
characteristics
demographic
characteristics
prediction

available

prediction

available

prediction

available

prediction

available

prediction

prediction

2015
available,
MOU in
progress for
future years
available

prediction

available
available

Health and Human
Services (HHS)
Health and Human
Services (HHS)
Health and Human
Services (HHS)
Health and Human
Services (HHS)

record linkage,
demographic
characteristics
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction

CMS Medicaid and CHIP Information System (MSIS
and T-MSIS)

Health and Human
Services (HHS)

enumeration,
prediction

Comp Homes Underwriting Management System

Housing and Urban
Development
(HUD)
Housing and Urban
Development
(HUD)
Housing and Urban
Development
(HUD)
Housing and Urban
Development
(HUD)

enumeration,
prediction

CES Best Race File
LEHD Employment History File (LEHD-EHF)
LEHD Employer Characteristics File (LEHD-ECF)
LEHD Individual Characteristics File (LEHD-ICF)
LEHD Unit to Worker Impute
Master Beneficiary Records (MBR)

Supplemental Security Records (SSR)
Disability Application File (831)
Social Security Numident File

HHS Child Care and Development Fund (CCDF)
HHS Temporary Assistance for Needy Families
(TANF)
HHS Indian Health Service (IHS)
CMS Medicare Enrollment Database

Multi-Family Tenant Characteristics System

TRACS data

PIC data

Social Security
Administration
Social Security
Administration
Social Security
Administration

16

available
available

available
available

available
available
available
data in-house,
MOU in
progress
data in-house,
MOU in
progress
available

enumeration,
prediction

available

enumeration,
prediction

available

enumeration,
prediction

available

Real-Time 2020 Administrative Record Census Simulation
Version 5.0
Office of Personnel Management Files (OPM)
Veteran’s Administration Records (VA)
Selective Service System
National Change of Address Files (USPS)
Army Service and Post Service Data (DOD)
Department of Defense Records (DOD)
Bureau of Prisons Permanent Release Database
Federal Housing Authority Loan data

Office of Personnel
Management
Veteran’s Adminis
Selective Service
System
United States Postal
Service
Department of
Defense
Department of
Defense
Bureau of Prisons

U.S. Citizenship and Immigration Services visa and
naturalizations data

Federal Housing
Authority
Department of
Homeland Security

Immigration and Customs Enforcement Student
Exchange and Visitor Program (SEVIS)

Department of
Homeland Security

U.S. Marshals Service incarceration data, with DHS
citizenship status
U.S. Customs and Border Protection arrival/departure
data

Bureau of Prisons
and Department of
Homeland Security
Department of
Homeland Security

U.S. State Department Passport Services passport data

Department of State

U.S. State Department Worldwide Refugee and
Asylum Processing System (WRAPS)

Department of State

Supplemental Nutrition Assistance Program (SNAP)

State agencies

Supplemental Nutrition Program for Women, Infants,
and Children (WIC)
Temporary Assistance to Needy Families (TANF)

State agencies

Low Income Home Energy Assistance Program
(LIHEAP)
Alaska Permanent Fund Dividend File

State agencies

Homeless Management Information Systems (HMIS)

County agencies

Utilities Records Data
Veteran Service Group of Illinois (VSGI)
Corelogic
DAR Partners
Experian
InfoGroup
Melissa Data

VSGI, Inc

State agencies

State agency

Corelogic
DAR Partners
Experian
InfoGroup
Melissa Data
17

prediction

available

enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
record linkage,
demographic
characteristics
record linkage,
demographic
characteristics
record linkage,
demographic
characteristics
record linkage,
demographic
characteristics
record linkage,
demographic
characteristics
record linkage,
demographic
characteristics
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
enumeration,
prediction
prediction
prediction
prediction
prediction
prediction

available
available
available
available
available
available
available
MOU in
progress
MOU in
progress
MOU in
progress
MOU in
progress
MOU in
progress
MOU in
progress
some states
available
some states
available
some states
available
available
available
available
available
available
available
available
available
available

Real-Time 2020 Administrative Record Census Simulation
Version 5.0
RealtyTrac
Targus/Neustar

RealtyTrac
Targus

prediction
prediction

available
available

VII. Risks
1. If the Census Bureau does not provide full funding for staff or provide staff with the
needed skills, then the project scope will have to be narrowed.
2. If the Census Bureau does not maintain and in some cases revise agreements with current
data providers, then some subpopulations will be poorly covered in the simulations.
3. If the Census Bureau’s DMS approval process for provisioning data to researchers is not
streamlined, the project may not be able to produce results in a timely manner. This is a
particular concern for this project, since it involves so many datasets and is under time
pressure due to the real-time aspect.
4. If the Census Bureau does not acquire additional data sources such as State Department
passport and visa data, then some subpopulations will be poorly covered by the
simulations.
5. If the public and/or stakeholder groups are concerned by the alternative population
estimates produced by the simulations, then legal challenges may occur.

VIII. Limitations
1. The applicability of the simulations to conducting an administrative record census in the
future depends on the availability of the same data sources in the future, which may not
be the case. Some additional data sources may become available in the future, while
others may no longer be available.
2. The 2020 Census may differ in coverage relative to past censuses due to sensitivity to and
the possible addition of a question on citizenship status. The citizenship status question
could potentially be discontinued after 2020.35 Thus, the comparisons between the 2020
Census and the administrative record simulations could thus be different than they would
be in the future (minus the citizenship question) for this reason.
3. Any errors in ACS household rosters will negatively affect the accuracy of the personplace models. For example, persons in the roster failing PVS will not be included in the
models. The relationship between their survey and administrative records addresses could
vary systematically with whether the person is successfully PVSed in the ACS, leading to
less accurate predictions for such people.

35

This was mentioned in the March 1, 2018 memo from the Census Bureau to Commerce Secretary Wilbur Ross,
which has been publicly released in a Freedom of Information Act (FOIA) request.
18

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

4. When there are discrepancies between the 2020 Census, PES, and administrative records,
it is impossible to know for certain which is correct in the absence of an error-free source.

IX.

Issues That Need to be Resolved

1. The MOUs for some data sources have not yet been completed.

X.

Division Responsibilities

Division or Office
ERD
CED, CES, CODS, CSRM,
ERD
CES
CES, CED

XI.

Responsibilities
 Data sharing agreements
 Data acquisition and processing
 Enhance record linkage infrastructure
 Evaluate PVS and entity resolution
 Data aggregation
 Supplementary coverage and characteristics analyses
 Model development and estimation
 Population estimates

Milestone Schedule

Evaluation Milestone
Obtain additional administrative record sources
Enhance record linkage infrastructure
Evaluate PVS and entity resolution linkage processes
Develop person-place models using ACS data
Process administrative records available on July 1, 2019 for use in 2019
administrative record census
Construct administrative record census simulations for 2019
Compare 2019 simulations to March-April 2019 ACS for housing units
and persons in common
Make predictions for where survey-style data collection most useful in
2020
Process administrative records available on July 1, 2020 for use in 2020
administrative record census
Construct administrative record census simulations for 2020
Produce 2020 administrative record census statistics
Compare 2020 administrative record census simulations to 2020 Census
Test and revise person-place models using 2020 Census

19

Date
10/18 – 09/19
10/18 – 09/21
10/18 – 09/21
03/19 – 06/19
07/19-09/19
10/19
11/19
12/19 – 02/20
07/20-08/20
09/20
10/20
11/20-06/21
06/21 – 08/21

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

Evaluation Milestone
Process administrative records available on this date for use in 2021
administrative record census
Write report on 2020 administrative record census simulation and 2020
Census comparisons
Process administrative records available on this date for use in 2021
administrative record census
Construct administrative record census simulations for 2021
Produce 2021 administrative record census statistics

Date
07/21-08/21

Distribute Initial Draft Real-Time 2020 Administrative Record Census Simulation
Report to the Decennial Research Objectives and Methods (DROM) Working Group
for Pre-Briefing Review

09/30/2021

Decennial Census Communications Office (DCCO) Staff Formally Release the
FINAL Real-Time 2020 Administrative Record Census Simulation Report in the 2020
Memorandum Series

03/01/2022

07/21-09/21
07/21-08/21
09/21
09/21

XII. Review/Approval Table
Role

Approval Date

Primary Author’s Division Chief (or designee) Lucia Foster

08/13/2018

Decennial Census Management Division (DCMD) ADC for Nonresponse,
Evaluations, and Experiments

02/19/2019

Decennial Research Objectives and Methods (DROM) Working Group

02/19/2019

Decennial Census Communications Office (DCCO)

mm/dd/yyyy

XIII. Document Revision and Version Control History
Version/Editor
1.0
2.0

Date
08/29/2018
02/05/2019

3.0

02/20/2019

4.0

03/06/2019

Revision Description
Initial draft
Incorporated comments from September 2018 DROM
Incorporated comments from February 2019 Quality
Process Review
Incorporated comments from February 2019 DROM

20

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

XIV. Glossary of Acronyms
Acronym
ACS
ADC
AREX
CBP
CEF
CUF
DA
DCCO
DRF
DROM
DSSD
EXC
FTI
IPT
IRE
IRS
ITIN
MAF
MAFID
NRFU
PES
PIK
PPS
PVS
R&M
SSA
StARS
USCIS
VSGI

Definition
American Community Survey
Assistant Division Chief
Administrative Records Census Experiment
U.S. Customs and Border Protection
Census Edited File
Census Unedited File
Demographic Analysis
Decennial Census Communications Office
Decennial Response File
Decennial Research Objectives and Methods
Working Group
Decennial Statistical Studies Division
Evaluations & Experiments Coordination Branch
Federal Tax Information
Integrated Project Team
Integrated Research Environment
Internal Revenue Service
Individual Taxpayer Identification Number
Master Address File
Master Address File Identification Number
Nonresponse Followup
Post Enumeration Survey
Protected Identification Key
Predominant Purpose Statement
Person Identification Validation System
Research & Methodology Directorate
Social Security Administration
Statistical Administrative Records System
U.S. Citizenship and Immigration Services
Veterans Service Group of Illinois

XV. References
Bauder, Mark, and Dean H. Judson, 2003, “Administrative Records Experiment in 2000 (AREX
2000) Household Level Analysis,” 2000 Census Experiment Report, U.S. Census Bureau.
Brown, J. David, Jennifer H. Childs, and Amy O’Hara, 2015, “Using the Census to Evaluate
Administrative Records and Vice Versa,” Proceedings of the 2015 Federal Committee on
Statistical Methodology (FCSM) Research Conference.
21

Real-Time 2020 Administrative Record Census Simulation
Version 5.0

Farber, James, and Charlene Leggieri, 2002, “Building and Validating a National Administrative
Records Database for the United States,” New Zealand Conference on Database Integration.
JASON, 2016, “JSR-16-Task-009, Alternative Futures for the Conduct of the 2030 Census.”
MITRE Corporation and Santa Fe Institute, 2016, “U.S. Census 2030 Challenge FY 2016 DRAFT
Report”.
Ralphs, Martin, and Paul Tutton, 2011, “Beyond 2011: International Models for Census Taking:
Current Processes and Future Developments,” Beyond 2011 Project, Office for National Statistics,
Version 1.0.
Rastogi, Sonya, and Amy O’Hara, 2012, “2010 Census Match Study,” 2010 Census Planning
Memoranda Series No. 247.
Steorts, Rebecca C., Rob Hall, and Stephen E. Fienberg, 2016, “A Bayesian Approach to Graphical
Record Linkage and Deduplication,” Journal of the American Statistical Association, Vol. 111,
pp. 1660-1672.

22