KEY CENSUS DESIGN ISSUES
In some respects, the 1990 U.S. census was the most successful to date. Technological advances in data processing and distribution of data products in CD-ROM (compact disc-read only memory) format have provided census users with unprecedented access and flexibility in handling summary population and housing information. Yet several concerns regarding recent censuses have deepened, and there has been growing momentum and advocacy for fundamental change in census operations.
The two most important concerns are cost and differential coverage. In constant (1990) dollars, the unit costs of counting a household have increased from approximately $10 in 1960, to $20 in 1980, and, to $25 in 1990. Even after adjusting for the decreasing median size of households, census operations have clearly become more expensive. But increases in expenditures have not solved the persistent problem of differential coverage: the Census Bureau estimates that 1.8 percent of the population, or about 4.7 million persons, were not counted in the 1990 census, and the difference in the undercount rates for blacks and nonblacks was 4.4 percentage points, the largest since 1940 (U.S. General Accounting Office, 1992). Differential coverage by race has significant implications for political representation and allocation of federal program funds because these decisions are based on census data at various levels of geography.
These and other issues have drawn the attention not only of the Census Bureau, but also of Congress and other census data users. The Census Bureau has undertaken a research program that reflects a major rethinking of census methodology with the intent of testing design components that represent fundamental change from current census practice.
Early in 1991 the Department of Commerce established a Task Force on the Year 2000 Census to provide an organizational structure for the investigation of issues regarding the 2000 census. The task force comprises a technical committee, a policy committee, and an advisory committee. These committees are concerned with the major issues facing 2000 census planning—particularly, cost and differential coverage.
Our panel is studying feasible methods for the census not only in the year 2000, but also for 2010 and beyond. We have a mandate to make recommendations for features of census design that should be investigated and developed for censuses after the next one. Some features of these future designs could be tested in the near term and further developed in conjunction with the 2000 census, even though they might not be fully implemented until subsequent censuses. Our deliberations lead us
to consider all demographic data systems, including current estimates and sample surveys.
The panel has four basic tasks: (1) identify designs to be investigated for the 2000 census; (2) evaluate proposed research on alternative census designs; (3) evaluate the results of the research and the selection of census designs for further consideration, in particular for the series of census tests that begin in 1995; and (4) recommend census designs to be explored for 2010 and succeeding years.
Our emphasis in this interim report on differential undercount and cost has several consequences. First, we give less attention to gross or total census error—that is, omissions plus erroneous enumerations. At several points in the text, we note the importance of the total error concept—for example, when we call for aggressive research on techniques to prevent erroneous, duplicate enumerations during a census with multiple response modes. We believe that census methodology should strive to minimize not only omissions (that produce undercounts) but also erroneous enumerations (that produce overcounts). We expect to give the total error concept further attention in our final report.
Second, our deliberations on cost issues are somewhat muted by the limited amount of information available on the cost-effectiveness of census design components. The Census Bureau has a very detailed cost model (Bureau of the Census, 1992f) for operational planning. This model has been used to estimate costs associated with several designs or design components being considered for the 2000 census, and at the time of this report, work is ongoing on other designs. We look forward to additional information on costs of design options, so that we can expand on the observations contained in this interim report.
Initially, the technical committee of the task force constructed a set of 14 census design alternatives. Each alternative was characterized by one or more unique design components; each was also judged to have the potential to meet the current demands of the decennial census. Six of the 14 designs built on the basic structure of the 1990 census, adding different provisions: multiple ways of responding to the census; varying degrees of sampling and statistical estimation; and targeted methods to overcome barriers to enumeration. Two designs relied entirely or to a very significant extent on administrative records. Four designs would have collected data on fewer topics than have been covered in recent decennial censuses. Two designs proposed collecting census data in two stages or through continuous measurement in the decade following the census year.
The panel's first report to the Census Bureau in December 1992 raised questions about the 14-design approach. Subsequently, the Census Bureau decided to remove its original set of 14 alternative census designs from further consideration. Rather, the 1995 census test will evaluate promising components of the original alternative designs. We strongly support this reorientation of the 2000 census planning process.
Census data collection involves four key steps: (1) the construction of an address frame; (2) an initial process to obtain responses that can be linked to the address frame; (3) a follow-up process to obtain responses from those not covered in
the initial process; and (4) a coverage assessment process that estimates the size of the population not covered through the initial and follow-up processes.
The design of a census data collection process in essence amounts to deciding which methods of identification, enumeration, response, and coverage improvement should be applied at each of the steps; whether sampling methods (and the corresponding estimation methods) should be used at any of the four steps; and if sampling methods are used, which methods and at which steps.
The second section of this chapter reviews work done by the Census Bureau as part of its 2000 census research and development program. Subsequent sections discuss some important issues related to the first step of the collection process, the creation of an address frame, and legal and operational issues. We conclude the first chapter with some observations about census planning for the year 2000 and beyond.
Chapter 2 addresses the possible use of sampling and estimation at each stage of the census data collection process, including the question of sampling for content—that is, methods by which one can forgo the need to ask all census questions of all households. Chapter 3 considers methods with potential for improving response and coverage at various stages of the collection process. Chapter 4 discusses the possible use of administrative records in the four collection steps. Chapter 5 addresses issues related to alternative schemes that would spread the collection of census-type information over a decade, rather than concentrating efforts in a single year.
CENSUS BUREAU EVALUATION CRITERIA AND PROCESS
The Census Bureau developed a set of mandatory and desirable criteria for assessing design alternatives, and it has specified that any design being considered for the 2000 census must satisfy all mandatory criteria. Each design that meets the mandatory criteria will then be assessed according to the set of desirable criteria.
There are six mandatory criteria for the 2000 census design:
not require a constitutional amendment;
meet data needs for reapportionment;
provide data defined by law and past practice for state redistricting;
provide age and race/ethnic data defined by law to meet the requirements of enforcing the Voting Rights Act;
protect the confidentiality of respondents; and,
possess the ability to reduce the differential undercount.
There are 10 desirable criteria for the 2000 census design. Numbers (1)–(5) pertain to the census outcome; (6)–(8) cover the data collection process; and (9) and (10) consider external factors:
result in comparative cost effectiveness with respect to other alternatives under consideration in real terms on a per unit basis;
provide small-area data that the census is uniquely capable of providing;
provide a single, best set of census results produced by legal deadlines for reapportionment and redistricting;
provide an overall high level of coverage;
increase the primary response rate to the census;
reduce the level of respondent burden;
minimize the degree and type of changes needed in federal or state law;
consider the reliance on new or unproven methods or capabilities;
permit full development and testing of its major design features; and,
provide opportunities to involve the U.S. Postal Service, state and local governments, national organizations, and other private, nonprofit, and commercial enterprises.
With regard to the third desirable criterion, the Census Bureau has developed the concept of a ''one-number census'' that would provide "the best possible single set of results by legal deadlines, . . . based on an appropriate combination of counting, assignment, and statistical techniques" (Miskura, 1993b). In this definition, counting refers to the full array of methods used for direct contact with respondents, including mail questionnaires, personal visits, and telephone calls. Assignment refers to the use of evidence from administrative records to add persons to the count for a specific geographic location without field verification. Statistical techniques for estimation include sampling during follow-up of nonrespondents and procedures for measuring census coverage.
The Census Bureau has expressed its commitment to pursue a one-number census for the year 2000, based on counting, assignment, and estimation. This commitment is reflected in the decision not to test a 1990-style dual-strategy approach in the 1995 census test (Bureau of the Census, 1993f). The specific counting, assignment, and estimation methodologies will be determined by the 2000 census research program. Associated with the one-number census is the principle of integrated coverage measurement, the premise of which is that the three components of a one-number census are designed to complement one another in order to meet legal deadlines. That is, the results from measurement of coverage will be incorporated into the official census results (Miskura, 1993b).
The Census Bureau has also identified eight categories of distinguishing features of a possible 2000 census design. Below, we summarize the Census Bureau's recommendations for testing in 1995 with regard to each feature (Bureau of the Census, 1993e).
Data collection outside the decennial year. The Census Bureau is designing a prototype system for continuous data collection that includes a national "head count" every 10 years. This development work will be carried out in parallel with the 1995 census test that will provide information about accurate and cost-effective methods for the decennial year portion of a continuous measurement program.
Content in the decennial year. The 1995 census test may collect the same
content as the 1990 census, although the Census Bureau is considering design options that would involve reduced content in the 1995 test and in the decennial census. Determining the content of the 2000 census will involve ongoing consultation with federal and nonfederal data users and will proceed in parallel with the development of the 2000 census design.
Use of lists. The Census Bureau intends to create and continuously update a master address file that is integrated with its topologically integrated geographic encoding and referencing (TIGER) system database (see Chapter 2). Plans for the 1995 census test call for the list of housing units for the test sites to be developed, at a minimum, with administrative records from the U.S. Postal Service, local records, and available telephone number lists.
Primary response options. For initial contact with respondents, the Census Bureau plans to use mail, face-to-face enumeration, and possibly telephone interviewing, and other electronic modes. To support public response, the Census Bureau plans to include assistance centers, language aids, and, to the maximum extent possible, enumeration at multiple locations (e.g., libraries and shopping malls) as part of the 1995 census test. Administrative records will not be used as a primary method of enumeration, but they will be used for coverage improvement.
Nonresponse follow-up. The Census Bureau recommends continued research on census designs that would either eliminate or substantially shorten the period of time that recent past censuses have allocated to following up nonrespondents to the mail questionnaires. The so-called "truncated" census will be investigated in the 1995 census test if research determines by September 1995 that this design is a viable alternative. The Census Bureau also plans to use, to the extent possible, multiple modes for collecting data during follow-up—in particular, telephone (computer-assisted) interviewing.
Sampling for the count. The Census Bureau recommends testing and evaluation of sampling for nonresponse follow-up during the 1995 census test. Sampling will also be part of the various coverage improvement methods under consideration for testing in 1995. (See Chapter 2 for definitions and discussions of post-enumeration surveys, SuperCensus, and CensusPlus.) The sample census design, which would entail sampling the entire mail-out universe instead of attempting a complete enumeration, has been eliminated from consideration on the grounds that sampling for the count prior to enumeration is unconstitutional.
Sampling for content in the decennial year. The Census Bureau plans to experiment with multiple sample forms (also known as matrix sampling) in the 1995 census test in order to learn how to implement the procedures and produce estimates from this type of sampling.
Statistical estimation. Methods for statistical estimation will be needed in the 1995 census test to support sampling for nonresponse follow-up. Also, statistical estimation will play a central role during the 1995 tests in implementing any of the candidate methods for integrated coverage measurement designed to produce a single set of counts in a one-number census (see Chapter 2).
ADDRESS LISTS AND OTHER RECORDS
Address List Development
Address list development is of central importance to virtually all candidate designs for the 2000 census. Lists containing address information—but without names of residents or other personal data—support several stages of decennial census operations, including distribution of mall questionnaires, follow-up of nonresponding households, and measurement of population coverage. For purposes of coverage improvement, address lists can also be used in conjunction with administrative records that contain information about individuals.
The Census Bureau's 1990 address control file can still serve as a base for the year 2000 master address file, but updating of the 1990 file—that is, processing additions and deletions of housing units, plus other corrections—will have to begin immediately in order to establish a continuously maintained national address list at the Census Bureau in time for the 2000 census. A major initiative for address file updating is included as an item in the fiscal 1994 budget (Scarr, 1993). If this work does not proceed, the Census Bureau will once again have to look to other sources, such as private address lists, in constructing a master list for the decennial census in 2000. In any case, all potential sources of address information—whether governmental or commercial—should be evaluated on their merits.
Work in this area can benefit from increased cooperation with other federal agencies, as well as state and local governments. The Census Bureau is conducting ongoing discussions with the U.S. Postal Service (USPS) regarding options for joint work on address list development. In the 1980 and 1990 censuses, the USPS performed multiple checks on the address list used to mail census questionnaires. During an operation called the advance post office check (APOC), postal carriers checked the census list of housing units on their routes, making corrections and providing information on missing addresses. Evaluations of the effectiveness of APOC and similar address list development operations must consider the accuracy of additional listings—that is, the proportion of duplicate addresses and nonexistent housing units—including any differential performance across geographic areas.
The Census Bureau is also exploring options that would increase involvement of USPS personnel in census operations. One proposal, described in an option paper (Bureau of the Census, 1992c), outlined a cooperative effort that would provide the geographic framework to accept continuous USPS updates of new addresses to the master address file (MAF). The Census Bureau plans to link the MAF with its TIGER system, a digital (computer-readable) map database. The USPS initially expressed interest in the potential value of TIGER in automating more of its mail delivery planning and management activities, but it has judged that more extensive collaboration with the Census Bureau in this endeavor would not be a cost-effective investment. The Census Bureau is exploring the possibility of collaborating with state and local governments on TIGER updating.
Although the USPS has decided against a joint venture on the TIGER system,
the USPS and the Census Bureau are pursuing other cooperative efforts in 2000 census planning. A USPS/Census task force has been established, and a subcommittee on rural address conversion is considering how to promote the conversion of rural addresses to a city-style format (Bureau of the Census, 1993b). Both organizations, as well as emergency service providers, could realize potential benefits through address conversion.
The panel believes that a geographic database that is fully integrated with a master address file is a basic requirement for the 2000 census, regardless of the final census design.
Recommendation 1.1: The Census Bureau should continue aggressive development of the TIGER (topologically integrated geographic encoding and referencing) system, the master address file (MAF), and integration of these two systems. TIGER/MAF updating activities should begin in fiscal 1994 and should concentrate first on the sites selected for the 1995 census test.
Successful completion of TIGER/MAF updating for these sites will enable the Census Bureau to gain valuable experience during the 1995 census test.
A continually updated TIGER/MAF system also has tremendous potential for meeting needs for both statistical and nonstatistical uses beyond the decennial census. In its final report, the panel expects to discuss access needs both inside and outside the federal statistical system. TIGER updating is only one aspect of the decennial census that can be viewed in a larger sense as a cooperative venture involving contributions from federal agencies and state and local governments.
Record linkage is the identification of records belonging to the same unit (e.g., a person or a housing unit) either within a single data set or across two different data sets. For example, the same residence could be listed two different ways in different address records. Multiple listings are more likely to be found in rural areas, where they are a potential source of erroneous enumeration. Needs for record linkage arise when administrative records are used, when multiple response modes are available, and for dual-system estimation as part of any coverage measurement program. Thus, any improvements in the accuracy or efficiency of record linkage will prove valuable for any census design.
Multiple response modes, including special methods for counting hard-to-enumerate populations, constitute a set of methods that could be adapted to many alternative census designs. Work in this area is likely to expand. For example, a recent technology assessment report (Ogden Government Services and IDC Government, 1993) recommended that four home-based technologies be considered for potential use in the 2000 census: (1) telephone; (2) voice recognition; (3) touchtone data entry with voice recording; and, (4) voice recognition with voice recording.
The Census Bureau has also begun a similar assessment of publicly accessible technologies that would permit response to the census at people's places of work or at public places, such as libraries, post offices, and shopping malls (Bureau of the Census, 1993b).
In considering adoption of multiple response modes, the Census Bureau must address the problem of erroneous enumerations. The 1990 census had approximately 11 million erroneous enumerations (the largest number recorded to date), and unless this problem is controlled, the additional burden posed by multiple responses would seem likely to exacerbate the problem of gross census errors. Toward this end, there may be some value in conducting experimental surveys that gather additional content (e.g., how many different residences, frequency of travel) with the potential to improve the ability to eliminate duplication.
Offering multiple options for responding to the census increases the probability of multiple responses from the same household, thus introducing the possibility of an overcount of those households. To allow maximum use of these multiple-mode techniques, which from a public perception point of view have much to recommend them, the Census Bureau needs to aggressively pursue research on techniques for eliminating duplicate responses so that households with multiple responses are not counted more than once. In fact, limits on the ability to eliminate duplicate records may be the controlling factor with regard to the feasibility of many of the innovations under consideration for the 2000 census design.
Another use of record linkage is for matching records. The Census Bureau conducts ongoing research on record linkage. In conjunction with special censuses of Godfrey, Illinois, and South Tucson, Arizona, work is underway to study the coverage and content of local administrative records and to gain experience with the process of acquiring these files. The Godfrey project involves a postenumeration comparison of voter registration and school district files to the special census. Results of computer and clerical matching indicate that these administrative record systems could add cases to the special census. In addition, clerical matching of tax assessment records suggests that this type of file could be useful in locating multiunit structures (Bureau of the Census, 1993b). The final report from the Godfrey project is scheduled to be released in September 1993. Work on the South Tucson project is ongoing at the time of this report.
Recommendation 1.2: The Census Bureau should aggressively pursue its research program on record linkage.
Record matching and eliminating duplication is important for all three proposed methods for coverage measurement (see Chapter 2). Statistical uses of administrative files, including technical issues in record linkage, are discussed further in Chapter 4.
There are many legal issues associated with the decennial census, perhaps the most obvious being the content requirements mandated by the Constitution and by law. The Panel on Census Requirements is conducting a thorough investigation of these content requirements. However, legal issues with possible implications for census methods arise in three contexts: (1) census starting and reporting dates; (2) the use of sampling and statistical estimation; (3) sharing of information from administrative records, including address lists.
April 1 is mandated as the reference census date by Title 13 of the U.S. Code. Title 13 also mandates that the state population counts required for reapportionment be provided 9 months after the census date and that local area data needed for redistricting be provided no later than 12 months after the census date. Thus, the respective deadlines for reapportionment and redistricting data are December 31 of the census year and March 31 of the subsequent year.
There is at least one reason why the April 1 census date should not be considered inviolable. Because moves from one housing unit to another tend to occur at the end and beginning of a month, conducting a census using the first day of the month as the reference date may lead to more frequent errors of misclassification. Canada, for instance, expects gains in accuracy in future censuses by moving its census date to the middle of the month; other countries have had similar experiences.
In weighing alternative methods, concern has been expressed about the ability to provide data by these legislatively mandated deadlines. The panel believes that the need for the December 31 and March 31 deadlines should be reevaluated if using otherwise promising methods would make it unlikely to meet either date. This consideration could apply, for example, to any of the three proposed methods for integrated coverage measurement (see Chapter 2). Promising new methods that can reduce the differential undercount or substantially reduce the costs of the census should not be discarded on grounds of time constraints without further consideration of those legally imposed constraints.
The legal acceptability of using sampling and statistical estimation remains an issue of considerable debate despite supportive rulings in every U.S. District Court case and a similarly favorable position in a recent Congressional Research Service (CRS) report (Lee, 1993). In its interim report (Committee on National Statistics, 1993) the Panel on Census Requirements, relying on reviews by legal scholars, endorsed the CRS position that sampling and statistical estimation are acceptable provided that there has first been a bona fide attempt to count everyone (e.g., by distributing a mail questionnaire). Our recommendations are based on this premise.
Title 39 of the U.S. Code restricts the USPS from disclosing lists of names or addresses, and similar restrictions on the Census Bureau appear in Title 13. Legal considerations thus impose constraints on Census/USPS cooperation. Special temporary legislation was obtained to permit the USPS to share detailed address information with the Census Bureau during the 1984 Address List Compilation Test (Bureau of the Census, 1992e). Similar permanent legislation might provide an
opportunity for both agencies to realize significant gains in operational efficiency and consequent cost savings, but any joint activity will need to attend to confidentiality issues regarding the sharing of address lists.
Recommendation 1.3: Congress should enact legislation that permits the sharing of address lists between the Census Bureau and the U.S. Postal Service for the purpose of improving the Census Bureau's master address file.
Chapter 5 contains an expanded, general discussion of legal and technical issues regarding access to content (including address information) from administrative records for statistical purposes.
In its earlier letter report (Committee on National Statistics, 1992), the panel recommended:
The Census Bureau should seek the cooperation of federal agencies that maintain key administrative record systems, particularly the Internal Revenue Service and the Social Security Administration, in undertaking a series of experimental administrative records minicensuses and related projects, starting as soon as possible and including one concurrent with the 2000 census.
A number of benefits will accrue from improved coordination among federal agencies. Possible examples include work with the Postal Service on address list development and with the Internal Revenue Service and Social Security Administration on statistical use of administrative records. Further interagency cooperation would be essential to the success of a continuous measurement design, for which extensive consultation with other federal agencies would be required to insure that content needs are being met.
The legitimacy of the census depends in part on public perception that it fairly treats all geographical areas and demographic groups in the country. "Fair treatment" can be defined in either of two ways: applying the same methods and effort to every area or attaining the same population coverage in every area so that estimates of relative populations of different areas are accurate. These alternatives are in some ways analogous to the competing principles of equality of opportunity and equality of outcome in the provision, for example, of educational services. The proper balance of these principles is a subject of policy debate in provision of services. In the case of the census, however, the priorities are clear: the objective of the census is to measure population accurately, above all to calculate accurate population shares in
order to apportion representation properly. Therefore, obtaining equal coverage clearly takes priority over using the same methods in every area. In fact, since experience shows that treating every geographical area and demographic group in the same way leads to differential coverage and therefore to inaccurate population shares, the Census Bureau has a positive duty to use methods designed to close the coverage gap, a duty recognized as a mandatory criterion for any 2000 census candidate design (Bureau of the Census, 1993d).
The approach of developing a targeting model and a "tool kit" of special methods (described further in Chapter 3) is one response to this duty of the Census Bureau. This approach involves constructing a targeting model, based on demographic and housing characteristics, to identify areas at particular risk for low mail return rates or other enumeration problems. These are areas where an accurate enumeration would likely benefit most from the deployment of special techniques drawn from a tool kit of candidate methods—such as using specially trained enumerators or address locators, opening census assistance centers, distributing forms other than by mail, and distributing some forms in languages other than English. These tool-kit methods would be applied as needed in small areas of various sizes. The decision to use any particular tool-kit method would be controlled by some combination of administrative judgment and predictions from the formal targeting model.
The use of administrative records, described in Chapter 4, although not part of the tool kit available to local census offices, might also involve some targeting of efforts to particularly hard-to-enumerate areas or population groups. For example, a list of food stamp recipients could add more names to low-income areas. Other lists, such as state (driver's license) or local government (school registration) lists, would of necessity contribute to the count only in their areas of coverage.
Some critics worry that the use of special methods in certain areas (e.g., tool-kit methods, local administrative records) might make statistical assessments of coverage more difficult or might invalidate assumptions used to combine sample-based estimates and enumeration totals. This criticism must be taken seriously, and the possible effects involved should be evaluated in the 1995 census test and in other ways. Differences in coverage across areas or groups can be corrected by appropriate use of coverage measurement and population estimation (see Chapter 2). The concern is that special tool-kit methods or local administrative records would be used differentially, and this concern should be addressed.
For example, suppose black renters in central cities correspond to a cell (or a poststratum) in an estimation method similar to the 1990 Post-Enumeration Survey (PES), used now for coverage measurement and estimation. Suppose further that special tool-kit methods were used for "renters in central cities" (or for "black renters in central cities"). In the case where the "treatment" of applying a tool-kit method corresponds to a poststratum, or where a poststratum can be defined on the basis of tool-kit use, the estimation procedure could correct for undercoverage of this group without making extra assumptions. If tool-kit methods improve the coverage of a given group recognized as a poststratum, then the statistical estimation of coverage
will detect higher levels of coverage. This case assumes that the use of tool-kit methods does not affect the dependence between enumeration (census capture) and catchability in the samples used to estimate coverage. We therefore believe that the use of special methods—those from the tool kit—or local administrative records for a group or area already defined as a poststratum in the estimation procedure does not create any new statistical problems. Good initial coverage through enumeration and assignment is still very desirable, because with high levels of initial coverage final estimates are less dependent on estimation and variance (or mean-squared error) is reduced.
A more serious concern arises if different enumeration methods (e.g., special tool-kit methods) are applied within an area corresponding to an estimation cell or poststratum. For example, suppose that local administrative records or special tool kits are applied to a particular political subdivision within a poststratum cell, but not to the entire poststratum or cell. In this case, it might be difficult to assess the possible biases in estimation that might be created by differential application of tool-kit methods if the methods alter the dependence between enumeration odds and sample catchability. Similar problems might arise if truncation of nonresponse follow-up (i.e., the date on which follow-up of mail respondents is stopped) differs among district offices, as has been the case in past censuses. Cell-based estimation methods applied as in the past will not take account of differentials within poststratum cells.
Several points should be considered in defense of census procedures that treat different areas differently. First, with any practical poststratification scheme (cell definition), there will be differences among areas within the cell, both in the underlying conditions affecting census coverage and in the conduct of the census. This has always been the case: for example, mail return rates and district office closeout dates varied substantially in the 1990 census. Furthermore, these differences exist regardless of whether or not estimation is part of the census methodology. Second, differences in treatment can be justified by local differences in conditions, especially if the decision to use a special method is determined by an objective decision procedure. If decisions are based on knowledge about the distribution of hard-to-count populations, such differences in treatment will tend to reduce differentials in outcome. Third, differences in treatment based on difficulty of enumeration or on the usefulness of particular techniques in different areas are probably more justifiable than those that result from haphazard implementation of coverage improvement programs or the assertiveness and technical capabilities of local authorities.
However, certain practices may arouse suspicion and be difficult to defend. By systematically planning the use of special methods with a view to a defensible standard of uniformity, charges of arbitrariness can be defended. If special enumeration methods are targeted to certain areas but not to others with similar characteristics, their application will appear to be arbitrary. The same will be true if they are targeted toward only some ethnic or socioeconomic groups but not to others with similar undercoverage problems. Systematic and complete planning for use of
these methods, based on objective criteria, can defend against the appearance of arbitrariness.
The use of administrative records that are only available in some areas, or that are of drastically varying quality in different areas, may create an appearance of unfairness. By making as much use as possible of record systems with national coverage or at least some degree of national uniformity, the perception of unfairness can be avoided. Alternatively, administrative records could be consistently used for undercovered groups, such as for food stamp recipients, as noted above. Inevitably, there will be different levels of success in operations of various district offices due to varying local conditions. By considering in advance rules for closeout of district office operations and for distribution of additional resources to district offices—rules that are designed to optimize uniformity of coverage within the constraints of varying conditions—the Census Bureau will be best able to defend its decisions. Of course, as in past censuses, the actual degree of uniformity attained will be limited by practical constraints.
Paradoxically, the Census Bureau's improved capabilities and success in tracking census operations, together with growing knowledge and awareness about factors that may affect differential undercount, create a climate in which even more than usual care must be given to avoid any appearance of arbitrariness or favoritism.
Residence rules and their application are important for a number of reasons. First, it is important to have consistent rules so that each person is counted in only one place (especially when matching records or eliminating duplication from multiple information sources is done). Second, people should be assigned to the correct location, as defined by the residence rules in effect. Third, people should not be excluded solely because the residence rules do not easily apply to them.
The Census Bureau has conducted the census on the basis of de jure rather than de facto residence—that is, people are essentially asked "Where were you usually resident on census day?" rather than "Where did you actually stay on census day?" The de jure approach has the advantage of defining residency in a way that does not depend on what happened on a particular day, but it can create difficulties for people whose de jure residency is hard to determine or who have none at all, such as homeless people and young people who move about from place to place. Although residency is ultimately defined by the wording of the questionnaire, consideration should be given to making the definitions work for as many people as possible.
(The concept of residency reappears at various points in subsequent chapters. Residence rules are discussed further, and recommendations are offered in Chapter 3, particularly with regard to improving within-household coverage and handling complicated living situations. That chapter also considers ideas for collecting a "census night" roster followed by questions to assign de jure residence.)
Attention must be given to defining residency consistently throughout all stages
of census operations: questionnaire mailing and questionnaire return and subsequent nonresponse follow-up and coverage measurement activities. This is particularly critical for coverage measurement programs, such as CensusPlus and the postenumeration survey (see Chapter 2), since they must determine census-day residency weeks or months after the fact. Administrative records have different definitions of residency, both because of the different purposes and laws under which they are collected and because some are continuously updated while others follow set time schedules. Use of administrative records will require ways of reconciling these differences (see Chapter 4).
Common sense, complemented by anecdotal evidence, suggests several benefits associated with the Census Bureau's maintaining a continuous presence in local areas throughout the decade. Ongoing activities could contribute to more effective outreach and promotion, thus improving public response and decreasing costs associated with nonresponse follow-up. This theme is explored more fully in Chapter 3. Also, organizational efficiencies might be realized by reducing the amount of temporary staff needed in the 10-year census cycle. The potential benefits could be especially significant if a continuous measurement program is adopted. Chapter 5 assesses the pros and cons of a continuous measurement census design. The Census Bureau is planning to continue work to develop this option in parallel with the 1995 census test.
PLANNING FOR THE 2000 AND FUTURE CENSUSES
The 1995 census test is of critical importance to the goal of an improved and more efficient census in the year 2000. Because of the extensive operational planning that must occur prior to 2000, the 1995 test represents the major opportunity to investigate fundamental reform without jeopardizing the integrity of the 2000 census. It is essential that adequate resources are invested in planning and executing this mid-decade test. Otherwise, the 2000 census will have a design very similar to that of the 1990 census, with the risk of continually rising unit costs, or an inadequately tested design that risks lost demographic information and population counts of unknown or inferior quality.
The 1995 census test should be structured to provide specific information to answer a limited and well-defined set of questions about alternative census methods. To the extent feasible, statistical designs such as fractional factorial experiments should be carried out, although the panel recognizes that operational issues will limit the scientific complexity of the 1995 census test. In particular, the test should include evaluation components that provide a basis for assessing cost-effectiveness.
The operational constraints on the 1995 census test underscore the importance
of learning as much as possible from other research. For example, simulation studies using 1990 census data can investigate the effects of truncating nonresponse follow-up operations at different points in time, using different rates of sampling nonrespondents for follow-up, and applying different coverage measurement methods. Similarly, not all methods need to be tested in large-scale field settings. To ease experimental complexity, certain methods might be excluded from large-scale field testing in 1995 when such an exclusion would not disrupt the research and development program or where smaller experiments conducted simultaneously with the 1995 census test will provide useful information.
The panel's letter report (Committee on National Statistics, 1992) included the following two recommendations:
The Census Bureau should initiate a separate program of research on administrative records, focusing primarily on the 2010 census and on current estimates programs. The research program should be funded separately from the 2000 census research and development activities, but there should be close liaison between them.
The Census Bureau should undertake a planning study, in collaboration with other agencies and contract support as needed, that would develop one or more detailed design options for a 2010 administrative records census. The study would have two major goals: to identify the steps that would need to be taken, early in this decade, to make a 2010 administrative records census possible and to set the stage for a national debate on the desirability of an administrative records census. The study, or at least its initial phases, should be completed during the current fiscal year.
The Census Bureau has indicated that it is giving serious consideration to submitting a budget initiative to form a long-term census research staff in fiscal 1995. Such an action is compatible with the above recommendations. However, we caution that the budget for a long-term staff should be independent of the funding cycle for short-term research and development work on the next decennial census (see Bradburn, 1993). The year 2000 research and development staff is scheduled to be disbanded at the end of 1995, yet personnel will be needed to evaluate the results of the 1995 census test. Consideration should be given to revising organizational structures to minimize the extent to which short-term and long-term research divisions would compete for personnel and other resources. (Chapter 4 includes a discussion of the long-term research that is needed to develop new, potentially cost-effective uses of administrative records for statistical purposes in the decennial census and other demographic programs.)