XI. Coverage Estimation: A Redesigned Post-Enumeration Survey

For 2030, Census is redesigning the Post-Enumeration Survey as a broader Coverage Estimation operation. The new approach will combine administrative records, independent field interviews, matching, and statistical estimation to measure how many people and housing units the census missed, counted more than once, or placed incorrectly.

Several core features of the traditional PES will continue. Coverage Estimation will remain organizationally separate from the census operations it evaluates, collect independent information, compare that information with census results, and use dual-system estimation to produce national and state coverage estimates. Field interviews will remain part of the design.

The principal change is that administrative and supplemental records will provide the Coverage Estimation national housing-unit list and most initial person rosters. Census will compare those records with census results before fieldwork and concentrate interviews on households where the two systems disagree or where additional information is needed. Some independent rosters will still be created through field interviews.

This distinction is important because the redesigned operation can initially sound like a replacement of the traditional post-enumeration field survey with administrative records. The current plan instead describes a more targeted field survey within a much more administrative-record-driven coverage-measurement system.

That redesign creates both opportunities and risks. It could support substate coverage estimates and direct field resources toward likely errors. But its credibility will depend on whether the independent system can detect people and housing units that are weakly represented in the same administrative sources increasingly used to conduct the census itself.

A. What Continues From the Traditional PES

Several essential features of the traditional Post-Enumeration Survey remain in the 2030 plan.

Independent Data Collection

Coverage Estimation will collect person and housing-unit information independently from the operational areas responsible for conducting the census. That information will come from administrative and supplemental sources as well as field interviews.

This organizational separation is essential. The people who evaluate census coverage should not simply accept the enumeration system’s decisions about whether a housing unit exists, who lives there, or whether a case was resolved correctly.

Independence, however, will need to be evaluated in more than organizational terms. A program can be managed separately while relying on many of the same underlying records as the census. The implications of that overlap are discussed below.

Matching

Coverage Estimation will continue to compare housing units and people in the independent system with those in the census. Matches indicate that the two systems found the same person or housing unit. Nonmatches and uncertain matches help identify possible omissions, duplicates, erroneous enumerations, and location errors.

Census plans to use administrative data, protected identification keys, improved computer matching, and artificial intelligence or machine learning to reduce the amount of manual matching that is necessary in this process. Clerical review will remain important for cases that automated systems cannot resolve confidently.

Dual-System Estimation

The plan explicitly continues national and state-level dual-system estimates. In simplified terms, dual-system estimation examines the overlap between the census and the independent coverage system to estimate the number of people found by neither system.

This method has been used in decennial coverage measurement since 1980. Its value lies in going beyond a simple comparison of two lists. A person does not have to appear in the independent system for Census to estimate that some people with similar characteristics were missed by both systems.

National and State Estimates

Coverage Estimation plans to produce state-level net coverage estimates for total populations and national estimates for demographic groups. This follows the broad structure of the 2020 PES, which produced national estimates by demographic characteristics and total population estimates for states and the District of Columbia.

State-level estimates should not be confused with state estimates for every demographic group. The 2020 sample did not support reliable estimates by race, age, or other characteristics within individual states. Whether the redesigned 2030 system can support more detailed combinations remains an open methodological question.

Field Interviews

Independent field interviews remain part of the design. Census plans to use an automated data-collection instrument rather than the paper-based processes used in earlier coverage operations.

The fieldwork will be more targeted, but it remains a critical safeguard. An interviewer may discover that the administrative roster contains a former resident, omits a young child, combines two households, or associates a person with the wrong address. Field interviews can also identify information that is absent from records or explain why the census and the independent system disagree.

Current commitment

Coverage Estimation will continue independent data collection, field interviews, matching, dual-system estimation, and national and state coverage estimates.

Important clarification

The 2030 plan redesigns the Post-Enumeration Survey. It does not eliminate post-enumeration fieldwork.

B. What Changes

The redesigned operation changes where the independent list comes from, which households receive field interviews, and the geographic detail Census hopes to produce.

Greater Use of Administrative Records

Administrative and supplemental data will play a much larger role than in the traditional PES. They will be used to create a national list of housing units, develop initial person rosters, support matching, and produce substate estimates.

This represents a fundamental shift. In 2020, the PES began by independently listing housing units in a probability sample of areas. That field-built list helped protect the evaluation from errors in the census address frame. For 2030, the national housing-unit list will instead be constructed from administrative and supplemental sources.

Census will therefore need to explain how the Coverage Estimation housing-unit list differs from the address sources used to conduct the census. If both systems rely heavily on the same postal, local-government, commercial, or property data, they may share errors involving informal units, rural residences, recently constructed housing, and other difficult-to-list places.

Administrative Data as the Primary Source for Person Rosters

Administrative and supplemental data will also become the primary source for the independent person rosters. Some rosters may be created through field interviews, but the operation will begin with a national file of people assembled from existing data.

This could greatly expand the geographic reach of coverage measurement. But it also changes the meaning of the second system. Instead of asking an independently selected sample of households who lived there on Census Day, Census will often begin with records created for taxes, benefits, education, health care, licensing, or other purposes.

The quality of those records may differ by population. Some people appear consistently in several current sources. Others have weak administrative footprints, appear at an old address, or are represented differently across programs. Census should assess whether administrative rosters are equally capable of identifying young children, recent immigrants, people experiencing housing instability, and residents of complex households.

More Targeted Fieldwork

Before conducting field interviews, Census will compare the administrative person file with census data. Coverage Estimation fieldwork will then focus on households where the lists differ and where a coverage error is considered likely.

Targeting could allow Census to devote more attention to cases that genuinely require explanation. Instead of interviewing every household in a conventional PES sample, workers could concentrate on possible omissions, duplicates, movers, and address discrepancies.

But the approach depends on the quality of the initial comparison. A person omitted from both the census and the administrative roster will not create a visible disagreement. A household incorrectly represented in the same way by both systems may appear resolved and receive no field interview.

Census should therefore retain some field data collection outside the model-identified discrepancy cases. A probability-based sample of apparent agreements could test whether agreement really indicates accuracy and estimate how often both systems make the same mistake.

Substate Estimates

The Operational Plan says Coverage Estimation will use national administrative and supplemental data to estimate population sizes and net coverage error below the state level. This would address one of the most persistent limitations of the traditional PES.

The 2020 PES could identify national coverage differences and estimate total coverage by state, but it could not provide reliable estimates for most counties, cities, neighborhoods, or demographic groups within states. Its sample was too small, and producing reliable estimates for every locality through a traditional field survey would approach the cost of conducting another census.

Administrative data and statistical modeling may allow Census to produce more geographically detailed estimates without fielding an enormous second survey. But the plan does not yet explain which substate geographies will receive estimates, what uncertainty those estimates will carry, or whether they will be published as official coverage measures, modeled indicators, or research products.

Increased Automation and Matching Technology

Census plans to automate the field instrument and increase its use of protected identification keys, artificial intelligence, and machine-learning methods in matching. The research agenda also examines guided software for clerical reviewers and algorithms that could reduce manual coding.

Faster matching could allow Census to identify discrepancies sooner and complete the estimation process more efficiently. It may also help match people whose records contain minor spelling, formatting, or address differences.

Automation should not eliminate meaningful human review. Matching a person across systems can require judgment about alternate names, transliteration, shared family names, moves, informal unit descriptions, and conflicting addresses. Census should publish validation results and error rates for different populations rather than relying only on an overall matching-accuracy measure.

Central change

The independent system will begin primarily with administrative housing-unit and person records, and field interviews will concentrate on disagreements between that system and the census.

C. Opportunities

The redesign could make coverage measurement more useful, especially at the local level.

Potential Substate Coverage Estimates

Substate estimates could give communities something they have long lacked: an official indication of whether coverage problems were concentrated in particular counties, cities, or other areas. A census can have little net national error while still substantially undercounting some places and overcounting others.

More detailed estimates could help Census, researchers, funders, and governments understand where enumeration methods failed. They could also improve planning for the 2040 Census and potentially inform the development of annual population estimates.

The promise should be treated cautiously until Census publishes the methodology. Model-based local estimates can be valuable, but their usefulness will depend on their geographic resolution, uncertainty, validation, and sensitivity to weaknesses in the administrative data.

More Targeted Use of Field Resources

Using administrative comparisons to target fieldwork could allow interviewers to spend more time on households with conflicting information. A smaller workload could support more careful interviews about moves, complex households, missing people, and duplicate records.

This benefit depends on Census treating targeted fieldwork as a quality strategy, not simply a cost reduction. Resources saved by avoiding routine interviews should be reinvested in resolving the difficult cases rather than used only to reduce the size and duration of the operation.

Faster and More Consistent Matching

Improved computer matching could resolve straightforward cases quickly and allow clerical specialists to focus on uncertain links. Standardized tools may also reduce inconsistent decisions among individual matchers.

Census could use confidence measures to identify records that require review rather than making every automated match final. Automated systems may be most useful when they help people make better decisions, not when they remove people from the process entirely.

Better Understanding of Local Coverage Variation

Combining national records, census information, targeted interviews, and substate modeling could produce a richer account of how coverage errors occur. Census may be able to identify differences associated with address quality, response mode, field effort, administrative enumeration, housing type, or local conditions.

The strongest program would produce more than a single net undercount rate. It would distinguish omissions from duplicates, show which enumeration methods contributed to error, and identify places where a plausible total conceals offsetting undercounts and overcounts.

Potential benefit

The redesigned system could combine the rigor of dual-system estimation with broader geographic coverage and more focused investigation of likely errors.

D. Risks

The same changes that create these opportunities could weaken coverage measurement if the independent system reproduces the census’s blind spots.

Shared Blind Spots

Administrative sources do not represent everyone equally. People can be absent, linked to an outdated address, or combined into households that do not reflect where and with whom they lived on Census Day.

This is already a concern for the census itself, which plans to use the Person Characteristic Frame and other administrative data for enumeration and processing. If Coverage Estimation draws heavily from the same agencies, records, and matching infrastructure, the two systems may fail to find many of the same people.

For example, an informal apartment could be absent from both the census address frame and the administrative housing-unit list. The people living there could also be connected to another address in administrative records. The two systems would agree, but the agreement would be wrong.

Census should publish a source-overlap analysis showing which records contribute to enumeration and which contribute to Coverage Estimation. It should also explain how the estimation method accounts for correlated omissions and other forms of dependence between the systems.

Formal Versus Meaningful Independence

The Operational Plan states that Coverage Estimation is independent from every other operational area. That is an important institutional commitment.

But meaningful independence depends on data and methods as well as organizational reporting lines. An evaluation team using a separately maintained copy of the same records may be formally independent while possessing little new information about people who are missing from those records.

Census should define independence publicly. Relevant questions include whether Coverage Estimation will use different source data, separate address and person frames, independent matching rules, separate quality assessments, and staff who can challenge assumptions embedded in the enumeration systems.

Targeted Interviews May Miss Unpredicted Errors

Targeting fieldwork toward households where the census and administrative lists disagree is logical. The greatest risk lies in errors that produce no disagreement.

Models can detect only the patterns they are designed and trained to recognize. A new form of housing instability, a rapidly changing migration pattern, or an unexpected failure in a census operation may not trigger the expected indicators.

Coverage Estimation should therefore include a method for testing apparently low-risk cases. Random or probability-based interviews among households where the sources agree could measure common-source error and reveal failures the targeting model did not predict.

Census should also publish how many households receive interviews, how they are selected, and what share of discovered errors came from model-targeted cases versus independent quality samples.

AI-Assisted Matching May Produce Differential Errors

Matching errors are not necessarily random. Names may be recorded differently because of transliteration, marriage, gender transition, cultural naming conventions, multiple surnames, or inconsistent ordering. People living in large or multigenerational households may share similar names and addresses. Rural and informal addresses may not standardize easily.

A false match can hide an omission by treating two records as the same person. A missed match can make one person appear to have been counted twice or produce an erroneous estimate of coverage.

Census should test automated matching separately by race, ethnicity, age, language, geography, housing arrangement, and other relevant characteristics. It should report false-match and missed-match rates and preserve human review for cases where the cost of an error is high.

Training data and validation cases should reflect the full range of names, households, and address formats found in the country. A system that performs very well for the majority of records may still distort estimates for smaller populations.

Small-Group Estimates May Remain Unavailable

Expanded administrative data do not automatically solve the sample-size and measurement problems that limited the traditional PES. Some populations may still be too small to support stable estimates, especially when geography and demographic characteristics are combined.

Administrative records may also lack the characteristics needed to define a group. A system cannot estimate coverage by disability or language if those characteristics are absent, inconsistent, or measured differently across sources.

Census should distinguish between estimates it cannot produce and estimates it has chosen not to develop. Where direct estimates are not statistically reliable, it should consider pooled analyses, targeted studies, oversampling, modeled indicators, or qualitative investigations. Uncertainty should be reported clearly, but it should not become a reason to avoid examining known coverage concerns.

Credible risk

Coverage Estimation appears independent because it is conducted by a separate operation, but its reliance on similar records causes it to miss many of the same people and places as the census.

E. Expand What Coverage Measurement Can Tell Us

Traditional coverage estimates have focused on broad demographic and geographic categories. The redesigned operation creates an opportunity to ask a wider question: not only how many people were missed, but which barriers, living situations, and enumeration methods were associated with those errors.

Not every requested measure will fit directly into dual-system estimation. Some will require additional questionnaire items, special samples, linkage to other surveys, or separate evaluation studies. Census should nevertheless design the broader coverage-measurement program so it can examine the populations and circumstances most relevant to an equitable count.

Disability

Census should measure whether people with disabilities experience different rates or types of coverage error. Relevant analysis could distinguish people living independently, with family, in supported housing, and in group quarters.

Because disability is not traditionally collected on the short decennial questionnaire, direct coverage estimates may require supplemental questions in Coverage Estimation interviews or carefully designed linkage with another source. Census should consult disability researchers and advocates about which concepts are feasible and how to avoid treating administrative program participation as a complete measure of disability.

The analysis should also examine operational barriers, including inaccessible internet response, difficulty using telephone systems, reliance on proxies, and incorrect classification of residential settings.

Sexual Orientation and Gender Identity

Where methodologically feasible and safe, Census should examine whether LGBTQI+ people face distinctive coverage risks. These may include being omitted by another household member, being reported under an outdated name or sex classification, moving because of family rejection or discrimination, or living in complex and unstable housing arrangements.

Sexual orientation and gender identity may not be available in the decennial census or appropriate administrative sources. Census should not infer these characteristics from names, relationships, medical records, or other indirect indicators. Research may instead require voluntary questions in a specialized follow-up study or linkage to a high-quality survey with strong privacy protections.

The objective should be to determine whether meaningful measurement is possible, not to manufacture precise estimates from inappropriate data.

Detailed Race and Ethnicity

Coverage estimates should provide as much race and ethnicity detail as the sample and methods can support. Broad categories can conceal very different experiences among national-origin, tribal, and detailed racial or ethnic groups.

Census should align the coverage program with the final 2030 race and ethnicity standards and evaluate whether administrative classifications agree with self-reported responses. This is especially important when the independent roster is drawn from records that use outdated categories, do not permit multiple identities, or lack detailed responses.

Where direct estimates for smaller groups are not possible, Census should consider oversampling, pooled analyses, special studies, and publication of relevant matching and characteristic-completeness measures.

Housing Instability

Coverage measurement should identify circumstances associated with mobility and uncertain residence, including recent moves, temporary stays, doubled-up households, shared custody, eviction, homelessness, and addresses that differ across sources.

These indicators can help explain why a person was missed or duplicated. A simple homeowner-renter comparison cannot capture the range of housing situations that affect where and how people are counted.

Field interviews are especially important here. Administrative records may reveal that addresses disagree but may not establish where the person actually lived on Census Day or whether several households occupied the same location.

Group Quarters

The 2020 PES measured the household population and excluded people living in group quarters and certain Remote Alaska areas. As a result, it could not independently estimate coverage for people in college housing, nursing facilities, prisons, or other group quarters.

The 2030 Operational Plan does not yet clearly state whether Coverage Estimation will include group quarters residents. Census should clarify the universe early.

If conventional dual-system estimation is not feasible for group quarters, the Bureau should design a separate independent coverage program. That work could compare facility frames, resident rosters, individual responses, administrative records, and follow-up interviews. It should measure both whether people were included in the total and whether their characteristics and locations were recorded accurately.

Group quarters should not be excluded from meaningful coverage evaluation simply because they require a different methodology.

Language Access

Census should examine whether coverage errors vary by language and by the type of assistance available. Relevant measures could include the language of response, use of translated questionnaires, telephone assistance, interpreter use, and whether the household was ultimately counted through a resident, proxy, or administrative record.

Language analysis may require information not contained in the short census questionnaire. Census could use Coverage Estimation interviews, operational data, ACS information, and carefully designed geographic analysis, while avoiding the assumption that everyone in an area has the same language needs.

The goal should be to determine whether multilingual services improved direct response and whether reduced access shifted particular communities toward field, proxy, or administrative enumeration.

A Broader Measurement Strategy

Census should publish a coverage-measurement agenda that distinguishes among:

• official dual-system estimates;

• model-based substate estimates;

• operational coverage indicators;

• specialized population studies; and

• qualitative research explaining why errors occurred.

Not every result needs to carry the same statistical status. But the public should be able to see what the principal estimates cannot measure and what additional work Census will conduct to fill those gaps.

Advocacy priority

Expand coverage measurement beyond broad national categories so it can identify the communities, living arrangements, and access barriers associated with omissions and erroneous counts.

F. Worst-Case Scenario

The redesigned Coverage Estimation system builds a national housing-unit list and person roster from administrative and supplemental records. Census uses many similar records in its address frame, Person Characteristic Frame, in-office enumeration, and response processing.

The two systems are managed separately, and the coverage program is formally classified as independent. But both have difficulty finding people who live in informal units, move frequently, lack stable connections to government programs, or appear at outdated addresses.

Before Coverage Estimation fieldwork begins, Census matches the administrative coverage roster to the census results. Most households appear to agree. Field interviews are concentrated on the visible discrepancies, while cases that match cleanly receive little or no independent contact.

The targeting process works well for errors that produce conflicting records. It does not detect people and housing units absent from both systems. Because no discrepancy appears, the households with the weakest administrative representation are among the least likely to receive a coverage interview.

Automated matching further increases apparent agreement. Some records that should remain separate are linked because names and addresses are similar. Other records are associated with the wrong household or location. Overall matching accuracy is high, but errors are concentrated among people with nonstandard names, complex households, informal addresses, or high mobility.

Coverage Estimation produces national and state results showing no major net error. Substate models also appear stable because their inputs reproduce the same geographic patterns as the census.

Local organizations nevertheless see closed apartment units that contain residents, shelters or group quarters with incorrect totals, and neighborhoods where extensive outreach produced unexpectedly weak counts. They can document individual cases and unusual operational patterns but cannot obtain official coverage estimates for the affected population or geography.

National measures cannot disprove those concerns, but they carry greater institutional authority. The absence of an official undercount estimate is treated as evidence that no substantial problem occurred.

Worst-case scenario

The evaluation system concludes that census coverage is strong because its independent records omit or misplace many of the same people as the census. Targeted fieldwork investigates disagreements between the systems but does not find their shared omissions. Communities can see local evidence of failure but lack official estimates capable of demonstrating it.

Why advocates should care

Coverage measurement is the principal independent test of the census. If it shares the census’s blind spots, weaknesses in the count may become statistically invisible precisely where credible measurement is most needed.

Where to Look: See section 3.3.3, “Coverage Estimation,” pp. 57–58 of the 2030 Census Operational Plan. It confirms that Census will continue national and state dual-system estimates and identifies four activities: methodological design, data collection, matching, and estimation. It also describes independent field interviews, the use of administrative data to create the national housing-unit and person lists, targeted interviews where the two systems differ, planned substate estimates, and research on artificial intelligence and machine learning for matching. Section 3.3.2, “Demographic Analysis,” pp. 56–57, describes the other major independent coverage-measurement program.

The 2030 Census Research Project Explorer includes two projects under EA 1.5, Post-Enumeration Survey. Modernize PES Data Collection, Introduce Automated Operations, and Explore Internet Self-Response and Alternative Data Sources examines how much fieldwork can be replaced by administrative, web-scraped, or other sources and how the remaining field collection should be automated. Create Computer-Assisted Post-Enumeration Survey Clerical Matching Software examines machine learning, alternative linkage algorithms, confidence measures, and guided clerical review. Advocates should watch for validation results showing whether these methods perform differently across populations, housing arrangements, and geographies.

For the traditional baseline, the 2020 Post-Enumeration Survey webpage collects the national demographic reports, state results, housing-unit estimates, source-and-accuracy statements, estimation methods, and operational assessments. Particularly useful assessments cover Independent Listing, Housing Unit Matching and Follow-up, Person Interviews, Person Matching, and Coverage Measurement Design and Estimation. Together they show how the 2020 program created an independent housing-unit list, conducted household interviews, matched people and addresses, and produced the final estimates.

The 2020 reports are also important for identifying existing limitations. They show which demographic estimates were available nationally, which results could be produced for states, why county and place estimates were not supported, and which populations, including group quarters residents, were outside the PES universe. These limitations provide benchmarks for judging whether the redesigned Coverage Estimation operation actually expands what Census can measure.