There are several ways in which a population can be counted. A conventional modern population census attempts to provide a count of all the people within a territory at a specified time. A census normally involves eight elements: (1) a definition of the population to be considered as part of the census; (2) determination of the content to be included on the census questionnaire, usually based on an extensive examination of users' needs; (3) careful testing of alternative questionnaire wordings and formats; (4) systematic preparation of lists of dwellings in which the population lives; (5) the hiring and training of enumerators; (6) the distribution of questionnaires and use of enumerators to question either the whole population or only those people who did not satisfactorily complete their questionnaires; (7) the processing and analysis of the census questionnaires; and (8) dissemination of statistical summaries using a variety of media. Within the general description, there are many variants of the specific procedures.
Several radical alternatives to a conventional census have been proposed for the United States. This chapter discusses four of them: a national register for the basic census, an administrative records census, a census conducted by the U.S. Postal Service, and a rolling or sample census. This chapter examines the characteristics of each of these alternatives and presents the panel's conclusions about their feasibility for the needs now met by the decennial census.
In a conventional modern census, most countries depend on either enumerators to visit the household or staff to compile results from mail questionnaires. Some countries, such as Canada, have enumerators deliver questionnaires directly to each household. The U.S. census relies primarily on a mailing address
for every dwelling unit and then uses the U.S. Postal Service for delivery of a mail questionnaire. If questionnaires are not returned, the census office then directly contacts the households that have not responded.
As noted above, there are many different specific procedures and widely different levels of intensity for achieving coverage. Some countries distribute their census questionnaires, do a "reasonable" follow-up, and accept some net undercount. In contrast, the U.S. census operation is a very intensive attempt to contact every housing unit and obtain demographic information for every person in the household. As part of its activities, the U.S. census operation involves an intensive effort to prepare an accurate address list, has enumerators active in the field for a longer time than other census operations, and makes as many as six attempts to follow up all nonrespondents. This intensive U.S. census is now very expensive. One major census design alternative, therefore, is for a less intensive, cheaper census.
A NATIONAL REGISTER FOR THE BASIC CENSUS
A national population register could replace the conventional census with the continuous recording of individual data, possibly linked to other records, for every resident of the United States. In order to provide data required for apportionment and redistricting (and for other purposes) the register would need to report the precise geographic location of everyone, that is, his or her residential address.
In spite of the challenges of such a system, population registers have been successfully developed in several nations, including Belgium, the Netherlands, Switzerland, and the Scandinavian countries. The Swedish system has its roots in the seventeenth century, and the Dutch register has been operating for more than a century. A similar system was instituted in the United Kingdom during World War II but was discontinued.1
A population register must be continuously updated to retain its usefulness. Also, a register needs to be linked to other records as people move, change their marital status, change jobs (and incomes), increase their formal education, have children, and die. Such a register can be maintained continuously only if there is a legal requirement for individuals to report all changes of addresses and certain other changes in status. Furthermore, massive record linkage (in order to obtain required information from other administrative record systems) is generally possible at reasonable cost only if all individuals are issued unique identifiers and if these identifiers are recorded on the main administrative system.
A register-based system, by itself, can supply basic demographic information about the population. Basic demographic data could include a person's birth date and birthplace, sex, and race or ethnicity (if requested by some form and if identity is assumed to be fixed). Through record linkage (or mandatory reporting to the register), information can also be obtained on marital status,
children, and family and household structure. Further linkage to tax records can provide income information.
A register system for the United States would have to be designed to meet user requirements for census information content and statistical quality. The first critical requirement would be to provide data required for legislative needs (especially to fulfill the requirements of the Voting Rights Act): individual data by age and race and ethnicity for small geographic areas. Such data would require that all persons be identified by race and ethnicity, a severe challenge for current birth records, which record the race and ethnicity of parents but not of newborns, and immigration records, which do not collect data on the race or ethnicity of immigrants. An additional problem would exist for illegal aliens, for whom there is no likelihood of successful registration.
There are a number of options for a basic register that could be used in the United States. One option is a multipurpose register based on population and housing data. Basic register data could include administrative records as well as statistical data. A prerequisite for such a system in the United States would be required registration of all international and internal migrants.2 Even if such registration were publicly acceptable, it would be difficult to maintain a register of international migrants in the face of a net gain of about 200,000 to 300,000 new illegal aliens annually. (The net figure hides gross illegal flows that are probably 5 to 10 times larger than the net flows). Also, the United States maintains no emigration records at all for the some 200,000 or so residents (native-born or foreign-born with U.S. citizenship) who depart each year and settle permanently in another country.
Data from a population register could be more timely than those from a conventional census. Annual estimates for small areas could be provided and updated as changes occur in geographic boundaries. Time-series comparisons would be improved with more frequent estimates for population data. A national population register with basic demographic data would also provide a powerful sampling frame for surveys for more general use. Having continuously updated demographic data, with an associated national address list, would make it possible to develop more effective sampling schemes with less expensive data collection.
A register-based census has a technical limitation in that the range of topics and their definitions are limited. Some census topics are not covered by current administrative records that might be linked with a register, and it may not be reasonable to add questions to administrative forms solely because data are needed for statistical purposes. Such data would therefore have to be collected in other ways—e.g., data on labor force status. For topics not covered by administrative records, a population register, even after linkage with other administrative records, would provide no estimates. Supplementary data could be collected by sample surveys, but to provide census-type detail on occupation, education, travel to work patterns, and other social and economic variables would require
samples comparable to that of the current census long form at least once every 10 years. The need for surveys would offset some of the savings that a register operation might provide compared with the census.
Sweden offers an example of how a country with an administrative register deals with topic limitations in the register. Unlike Denmark, which has relied wholly on its register-based census since 1981, Sweden has retained a census questionnaire since 1970, but it asks only for information that is not in the administrative records. The 1990 Swedish census questionnaire asked about economic activity, housing, household composition, and education; the data are then linked to the central population register to provide the full array of needed information.
On balance, the experience of countries operating register-based systems is that the ongoing costs are cheaper than those of a conventional census (Redfern, 1989). The peaking of resources is reduced, and large-scale field operations for data collection are not required. However, cost savings are achieved only if the information content of the register, supplemented through record linkage, is adequate.
Data from a population register can be produced more quickly and more frequently than data from a conventional census. The experience of countries operating population registers seems to be that the ongoing operation of the system lends itself to the prompt preparation of summary reports and aggregate data reports. Moreover, detailed information can be produced from the system as required on the usually limited variables that are recorded on it. Still, the system is sensitive to technical problems, particularly since the administrative files for the system with which it needs to be linked periodically are not centrally controlled. Linking administrative records would be a challenge for the United States in several regards: there are many administrative files that are controlled by state and local governments, and it would be a problem if the files changed without proper attention to the impact on the population register.
Strong political and cultural opposition is likely to the introduction of a unique individual numbering and compulsory registration system in the United States. The prospect of ongoing linkage of federal, state, and local government data would be opposed by many people. The requirement for notification of changes of address would also be felt to be onerous by many residents. Advocates concerned about privacy issues have already raised questions about adequate protection for federal administrative databases. Such advocates would be naturally concerned about the creation of a national database, on every U.S. resident, that would be continuously updated with changes in individual characteristics, including current residence. Access to state and local databases, especially for a centrally controlled population register, is also problematic. To be maintained continuously, there would need to be ways for individuals to update easily and regularly changes in their social and economic characteristics and their current residence. A population register would require universal compliance
in registration from U.S. citizens and resident aliens. If the population register were not universally accepted, its use in lieu of census data would lead to serious quality, cost, and public acceptability problems.
There are numerous advantages, for operational and financial reasons, for some countries to use national registers. But, for the foreseeable future, there would be major disadvantages for the United States, including a possible lack of public cooperation stemming from privacy issues about the required data matching and record linkage, which in addition to being concerns for their own sake, could lead to poor-quality data and lack of population coverage.
Conclusion 4.1 The panel concludes that a national register for the United States is not a feasible replacement for the decennial census.
Although a national register is not a feasible alternative for the 2000 census, the long-run possibilities may be more positive. Demands for universal health care with "portable" individual policies may produce a basic register of the population with such limited content as age, sex, marital status, and current address. The pendulum of public perception may shift toward the use of administrative records for labor certification, under the pressure to deal with illegal immigration, for example. If there were a relatively complete listing of the population, maintained up to date for routine administrative purposes, it might become attractive to consider the feasibility of a national register to replace the traditional census. A national register, as implemented already in several countries, offers frequent data at reasonable cost.
AN ADMINISTRATIVE RECORDS CENSUS
Another radical alternative to a conventional census involves the use of administrative records. During the past decade, some European countries (mostly in Scandinavia) have linked additional administrative records to their central population register and have discontinued direct-inquiry censuses. Denmark, Finland, and the Netherlands now rely on population counts and essential characteristics from their central records, although they supplement their national data with sample surveys. Other countries, such as Sweden, now canvass the population with a short-form census and use administrative records to provide information that, in turn, does not need to be asked in the census. A critical question is whether it is desirable for the United States to consider developing methods for a population census using administrative records without the support of a continuous population register—something that has not been attempted in any country yet.
Assessing this broad option in brief terms, the panel concludes that administrative records cannot be adapted in the near future (by 2000) to meet even relaxed demands for detailed demographic information for small geographic areas.
It would be extremely difficult to redesign current administrative data as a substitute for census data to meet the Voting Rights Act requirements for data on race and ethnicity for small geographic areas. No federal or state administrative system now possesses race and ethnicity data for the U.S. population, combined with coverage of all households and accurate information on current address.
Other limitations of existing administrative records include the accuracy of address information and the paucity of demographic and socioeconomic variables that are included (see Appendix J for details). The panel agrees with a letter report from the Panel to Evaluate Alternative Census Methods to the director of the Census Bureau (Bradburn, 1992), which finds that it is not feasible to consider the use of a census based on administrative records for 2000.
It is important to look beyond the 2000 census for an evaluation of a possibly more effective use of administrative records. Recent research suggests that a high proportion of the U.S. population is included in one or more existing administrative records (Sailer et al., 1993). The coverage of administrative records may well expand in the future. Moreover, if it becomes important to use records for census purposes, efforts could be made to expand the content and improve the quality of the records over the next two decades. An administrative records census may be a real alternative in 2020, but not if work does not begin soon in order to accumulate experience in working with the records. The potential population coverage of administrative records is an important area for further consideration, and more research is needed.
Consideration of the use of administrative records as a complement or substitute for the current census must address several issues. Whereas it is true that the coverage of some administrative record systems is high and expanding, no single record system covers the entire population. Two examples provide an illustration. The Internal Revenue Service (IRS) collects information from its annual tax returns and related information documents; however, it is a challenge to produce estimates of nonfilers who do not show up in their administrative records. As another example, the Social Security Administration (SSA) maintains a register of Social Security accounts. Increasingly, infants are registered with the SSA at birth and, because immigrants need to obtain a Social Security account, the country is moving toward a universal registration system. Emigrants, however, are not recorded in any administrative system.3 There also persist significant flows of illegal aliens into the United States, with a net gain of perhaps 200,000 or more illegal entrants annually and approximately 2 million illegal residents currently (Bean, Edmonston, and Passel, 1990). Illegal aliens often use fraudulent Social Security documents or other people's account numbers. There is therefore a need to assess differential coverage for administrative records systems and to examine whether a procedure could be developed to use administrative records for a census count.
Information on racial and ethnic identification must be provided by the census. There are no race and ethnicity data in IRS records, and Social Security
records do not contain such information for everyone. Birth registration certificates now include the race and Hispanic status of the mother (and sometimes of the father, with all states participating by 1994), but the information on race is kept confidential by the state vital registration system.4 Moreover, there are questions about the racial and ethnic identity of the many children who are born to parents of different racial origin or ethnicity. Can we presume the racial or ethnic identity of an adult on the basis of information from a birth certificate? Our nation has continually changed its sense of race and ethnicity classification over time. Even if a record system included race and ethnicity information for the entire population, the categories may change (as was done during recent censuses when a category for Asian and Pacific Islander was added), and the entire population would need to be resurveyed. Finally, there are concerns about the differences in the reporting of race and ethnicity in a nonthreatening census context and in administrative records.
Current administrative records are particularly problematic on race and ethnicity data. Little is known about the coverage of various minority groups in administrative record databases. Some databases do not include information on race. Nor is information on coverage by race available from those with data on race. Data on small population groups, such as American Indians, are likely to be a special problem because geographic coding is difficult in isolated rural areas and coverage is unknown.
Race data in SSA files (the Summary Earnings Record file in particular), the major federal database with demographic information for the U.S. population, have been declining in quality (Irwin, 1984:4-1 to 4-3). The race question for a Social Security applicant used the categories of white, black, and other until 1980. In 1980, the race/ethnicity question was altered and the listed categories were white, black, American Indian, Asian and Pacific Islander, and Spanish. The race/ethnicity question was voluntary, however, and a substantial proportion of applicants, especially in some large metropolitan areas, has not provided their race/ethnicity self-identity on forms since 1980.
Because no national files other than those of the SSA contain specific race/ethnicity information, it is important to note three limitations of these files. First, for individuals registering before 1980 and reporting ''other" for their race, the specific racial/ethnic identification is not known. Second, a significant proportion of persons who applied for a Social Security number during the past 14 years did not provide a racial/ethnic identification. Third, many Social Security numbers are currently issued to infants, whose parents apply for a Social Security number for the newborn at the time of birth. The race of a newborn is uncertain, because not all children necessarily have parents of the same race or ethnicity, nor will the race of a newborn necessarily be the same race or ethnicity reported by the person when an adult. Regardless of the racial or ethnic uncertainty of the newborn, no information on the race or ethnicity of a child is transferred to SSA files because there is no record on the birth certificate of the
race or ethnicity of the child. Birth certificates indicate only the race and ethnicity of the mother, and the information is not forwarded to the SSA. As a combination of the above limitations, current files are substantially incomplete or inadequate for race/ethnicity information for an administrative records census.
Besides the challenge of obtaining essential race and ethnicity information from administrative records, there are serious doubts that current records (from either the federal, state, or local governments) could provide most of the information currently collected in the long-form census questionnaire. Could records be redesigned to collect this information? Would large surveys be used in conjunction with a census based on administrative records?
The coverage of various major administrative records databases varies greatly. The IRS files have relatively high coverage of the population, especially if efforts are made to take into account miscellaneous tax records and people who may not file tax returns every year (Sailer et al., 1993). Additional people might be added to tax files from other databases, including elderly persons from the Medicare files and other groups of people from state and local files for welfare, Medicaid, and Food Stamps. At this moment, the coverage of an administrative records census cannot be precisely estimated because of extensive and unknown overlap between files, and because estimates of duplicates in specific files vary greatly. Finally, it is also clear that an administrative records census would need to depend on supplementary administrative records data on people in group quarters (prisons, jails, hospitals for the mentally ill, and other places for long-term residence in a group setting).
A reference data for an administrative records census requires careful study because various administrative records databases are updated at different times for different reasons and would have varying degrees of completeness if used for a specific census date. For example, deaths are not necessarily noted in a timely fashion in administrative records.
A basic census requires the correct geographic location for each respondent. Because the mailing address is not always the same as the residential address, an administrative records census would require a current residence for every person. One possibility would be to require an added question on residential address for the annual tax return, which would still not provide actual residential location for a substantial proportion of the population.
Administrative records do not provide family relationships for households. Neither IRS nor other administrative records identify the relationships among persons at an address that would permit the determination of families. Yet family-based statistics are critical outputs from the census. This is a serious problem because family-based statistics are among the most important data produced by the census.
A final issue for examination is the accuracy of the address for individuals in administrative records. The census needs to allocate individuals to small
areas. This can be done if individuals can be assigned to a specific dwelling unit, then placed within census blocks. However, this requires that administrative records be current and have an accurate street address. In some cases administrative records lacking up-to-date, accurate addresses could be linked to another record system that provides such addresses. In this case, the challenge is to link several individual records at the household level. Finally, and most important, is the issue that IRS data (and other records) often use business address or a tax accountant's address for tax information. These are issues for future research to examine.
The costs of data collection for an administrative records census would be spread among many federal, state, and local agencies. An administrative records census would require that the Census Bureau have access to a large number of individual files for matching and linking. It might be necessary to enact legislation to permit access to needed administrative records databases for matching and linkage. The pulling together of administrative records would be a complex, demanding task, though not a task beyond the capabilities of current information processing technology. The total cost of an administrative records census is uncertain. Countries with a uniform identification system and a long tradition of accepting registration have relatively low costs associated with their population registers. The experience of other countries provides little applicable information for the potential costs in the United States, a country with no existing system and no uniform identification system across major records systems.
Output from an administrative records census could be produced in a timely fashion and in a manner similar to that from a conventional census. Detailed information could be produced for geographic areas, including small geographic areas with geographically referenced data. Census output from an administrative records census may be limited in a number of important ways, however. First, coverage is likely to be lower than in the conventional census because currently undercounted groups are not necessarily included in administrative records. Illegal aliens in the United States are particularly unlikely to be included in an administrative records census. Second, small-area information could be subject to significant errors if the accuracy of address information is weak or if the issue of residential versus filing address is not remedied. Third, content would be limited to that of administrative records, even if current records were redesigned to include some additional data items. Moreover, administrative record systems are vulnerable to technical problems and changes, particularly if contributed files are not centrally controlled by the Census Bureau.
An administrative records census would require the creation of a central data bank of virtually all people in the country, with linked information from files currently stored completely separately. The accurate linkage of different databases would be best accomplished with a unique individual identifier, such as a Social Security number. How would the public react to the creation of a data bank? What would be the public's willingness to provide Social Security
numbers for major records, realizing that the purpose was to link various individual records for statistical information only? The issue of privacy and confidentiality of administrative records requires considerable investigation before one could consider the replacement of the conventional census with an administrative records census. The panel does not make a judgment about what the public acceptability would be for an administrative records census. We do note, however, that there would be strong political pressures against an administrative records census.
Although administrative records have been used to support census-taking for several years, the use of administrative records to take a full census is unlikely in the near future. The unique quality of self-reported race and ethnicity data, collected every 10 years by the census in an evolving format, cannot be duplicated by current administrative records.
Furthermore, an administrative records census faces a critical technical problem: the requirement to be able to unduplicate records. If there is reliance on several systems of administrative records, then the records must be unduplicated in order to prevent an overcounting of the population. More research and testing is needed on technical issues, including undercoverage, reporting errors, race and ethnicity data, and unduplication.
The Census Bureau has used administrative records to support census activities in the past. We endorse the use of administrative records to improve census coverage. The Census Bureau should intensify its efforts to use administrative records, recognizing that expanded use of administrative records must be an evolutionary process. In particular, we recommend (see Chapter 8) more intensive work with administrative records for preparing intercensal estimates.5 As experience is gained with records, we support experimentation to ensure their potential use for supplementing and complementing decennial census content.
Conclusion 4.2 The panel concludes that an administrative records census is not a feasible option for 2000.
A CENSUS CONDUCTED BY THE U.S. POSTAL SERVICE
The U.S. Postal Service and its letter carriers have played an important role in prior censuses. Chapter 5 reviews the role played by the Postal Service in recent censuses.
One radical alternative for a future census would be to have the entire field operation of the census conducted by the U.S. Postal Service. The U.S. Postal Service would deliver and collect census questionnaires using its own staff. Letter carriers would then collect information from all households that did not respond to the mail questionnaire. The role of the U.S. Postal Service would be limited to the actual field operation—all data would be transmitted to the Census Bureau.
The letter carriers might collect information directly from household members, and they might know enough about the household to provide substitute information. Major savings on the cost of nonresponse follow-up could be made if letter carriers could provide a limited amount of basic information on nonrespondents, without having to contact the household. Even if letter carriers knew this minimal information for some households on their mail delivery route, however, it is unlikely that they would be highly accurate (on age, for example) or that they would know such information for all households. And having postal employees classify household members by race and ethnicity might be objectionable to many people. Moreover, streamlining of mail delivery operations in recent years has reduced the contact that letter carriers have with many of their residential customers.
This radical alternative has been examined by a joint group from the U.S. Postal Service and the Census Bureau (Green and Scarr, 1993). Their examination of this alternative reveals:
It would not be cost-effective for either the U.S. Postal Service or the Census Bureau for the U.S. Postal Service to administer directly the census field operations. In conducting the 1990 census, it took about 36 million staff hours to contact nonrespondents. If this amount of work were handled by the U.S. Postal Service, it would have taken about 20 hours of overtime per week for about 5 weeks for every letter carrier. This assumes that every letter carrier would cooperate and would be willing (under current union rules) to work this amount of overtime.
The average U.S. Postal Service letter carrier costs about $23.50 per hour in total costs for regular time worked. This can be contrasted with the average hourly pay of about $7.50 (including benefits) that was paid to the average census enumerator in 1990. The rate for overtime census interviewing by letter carriers would be even higher, four times the rate for a census enumerator. Obviously, there is no benefit to the U.S. Postal Service, or to postal customers who pay the costs, for letter carriers to contribute their time without reimbursement. Overall, the cost of letter carriers is much higher than census enumerators, on an average hourly basis, and would significantly increase census costs.
The U.S. Postal Service and its letter carriers are not well acquainted with the occupants of residential addresses. Many letter carriers do not actually go door-to-door. Many households receive their mail at central boxes in apartment buildings or in new suburban areas. Many other people use postal mailboxes.
For over 25 years, almost all new mail delivery points in residential areas have been served through cluster boxes, making it unnecessary for the letter carrier to go door-to-door. A letter carrier serving newer residential areas rarely sees customers and often does not even see the residential housing units. Most rural letter carriers do not go to the customers' doors, but deliver to mailboxes
along roads that are often some distance from the customers' houses. In addition, all mail delivery is done during business hours when most people are not at home.
Rural areas would present special problems for the delivery of census questionnaires by the U.S. Postal Service. Special procedures would need to be developed because rural letter carriers do not go to the actual rural house. In some parts of the country, the U.S. Postal Service delivers mail at central postal offices through individual mail boxes. The actual physical location of the mail recipient is not known. Nor is it known whether more than one family or household member may share a mailbox. In other areas, rural route boxes are located at major intersections or on roadways. Again, the physical location of the residence is not known.
There would be legal concerns about U.S. Postal Service administration of the field operations of the census. Privacy issues might affect what letter carriers can do in obtaining information from nonrespondents. The public perception of the U.S. Postal Service might be adversely affected: How would the public view the U.S. Postal Service if it were involved in collecting and reporting personal information to others? People may have serious concerns about reporting information to letter carriers who deliver their mail every day. Finally, its possible legal responsibilities for the quality of the data and legal repercussions would be a concern for the U.S. Postal Service.
The Census Bureau and the U.S. Postal Service concluded that there are many ways in which cooperation on the census and other activities has been mutually beneficial. They also agreed that there are important new ways for an enhanced role for the U.S. Postal Service in future censuses. But they concluded that the "best way to use the expertise of the U.S. Postal Service letter carriers is not through the collection of actual census data" (Green and Scarr, 1993:8). The U.S. Postal Service gains no real advantage in taking on the responsibilities for census field operations. The primary objective of the U.S. Postal Service is to collect and deliver mail. Taking on census field operations would mean that the organization would divert attention from other tasks without gaining any important return. Overall, the U.S. Postal Service believed that the use of letter carriers to conduct census interviews "would have a negative impact on the ability of the Postal Service to deliver the mail" (Green and Scarr, 1993:6). We concur with the conclusions of this joint group.
Conclusion 4.3 The panel concludes that the decennial census cannot be conducted by the U.S. Postal Service in any way that would be cost-effective or an improvement over the conventional census.
It should be noted that the panel believes that there is an expanded role for the Postal Service in census operations in future censuses (Chapter 5 describes
the expanded role) and for improving the Census Bureau's intercensal national address list (Chapter 8 outlines this activity).
A ROLLING OR SAMPLE CENSUS
With regard to the constitutional requirements for the census, it is clear that a basic enumeration is required to meet constitutional requirements for reapportionment. Such an effort, in effect, is required also because of the need for small-area data for redistricting—including one-person, one-vote needs and Voting Rights Act requirements (see Appendix C for a legal and political history). For these reasons, census designs that do not include a complete enumeration of the population once a decade are not reasonable options on legal and practical grounds. We also have considered these designs from a practical perspective in the discussion below.
In a sample census, only some of the population would be counted. In a rolling census design, different parts of the population would be surveyed every year. Several variants of a rolling census design have been proposed. One design proposes a full census (including short-form and long-form content) of one-tenth of the nation's counties to be conducted every year (Horvitz, 1986). A different sample of counties would be covered every year of the decade. Horvitz submits that accurate data on internal migration could be developed from each year's one-tenth census to use in developing population estimates for the counties not covered that year. However, as the design does not provide for even a minimal census of the entire country at a single point in time, it would fail to meet the constitutional requirement that reapportionment counts be based on an attempt at a complete enumeration of the entire population.
Another design variant proposes to survey a less clustered sample of one-tenth of the population each year, cumulating the estimates over 1, 2, or more years to increase their reliability for small geographic areas (Kish, 1981, 1990). The proposed one-tenth sample would provide state estimates of reasonable reliability; however, use of these estimates would encounter the constitutional barrier to a sample census. Averaging the estimates from each year's sample over a 10-year period reduces their sampling error, but these estimates would not pertain to a year chosen for reapportionment but rather to the average experience of states and other areas over the prior 10 years. Not being based on an attempt at complete enumeration at a given time, they would, at a minimum, raise the question of whether they meet the constitutional requirement. Moreover, the cumulated estimates would be far more out of date with regard to the distribution of the population than would be estimates from the current census.6
Designs that attempt to spread over the decade the collection of some of the information that is now obtained in the census merit evaluation as to their benefits and costs. It is worth considering if such a design would provide more timely data and higher coverage and whether, even if these were to occur, they
would offset higher costs. There appears to be a legal consensus, however, that such designs must include a minimal complete census every tenth year to satisfy the constitutional requirement for reapportionment.7
From the methodological perspective, it is not even clear how one would go about conducting a sample census. Obtaining a sufficient degree of accuracy for reapportionment (and other purposes) would require the development of a frame from which to draw the sample. In the census, the frame would be the list of addresses that is used to mail out census questionnaires.8 In addition to developing the address list, it would be necessary to carry out activities to determine the list's accuracy for purposes of obtaining a proper random sample of people, as the list would likely include nonresidential addresses (e.g., businesses) as well as exclude some residences. One such evaluation activity might be to conduct a second independent survey either before or after the sample census itself, which, at a minimum, would have to ascertain the number, age, and sex of respondents. As a result of all these activities, for all intents and purposes, the Census Bureau would quickly wind up conducting virtually a complete census.
A sample census, assuming that it could be effected, would face other problems. An important difficulty is that, by all the available evidence, household surveys experience higher net undercoverage rates and more severe differential undercoverage than does the complete-count census.9 Errors of undercoverage and differential undercoverage that arise from the use of sampling can be adjusted by evaluation surveys. But the extent of such adjustment would be very large in a census that relied completely on sampling. Another problem that could contribute to coverage errors is the likely difficulty of publicizing a sample census in which there is no intent to try to contact everyone. Finally, there would appear to be some questions regarding the likelihood of achieving significant cost savings over a complete census.10
A rolling census would provide more regular information and, on average during the decade, would be more current at the time of use. Annual or regular output could be provided for small geographic areas and for small population groups, but the availability and detail of output would depend on sample sizes. Estimates for the nation and for subnational areas and groups would be affected by adjustment from sampling, coverage, and enumeration errors. Information for small areas, such as city blocks used for purposes of redistricting, would have to be based on data averaged over several years—raising both legal issues and problems of interpretation (e.g., regarding the meaning of such data as a moving average of income from a rolling census).
The cost of a rolling census would be more stable and less peaked than a conventional census. Expertise could be developed, and training would be more cost-effective. The work flow for the Census Bureau would be less disruptive. Regular contacts with the surveyed population would offer chances for improvement of procedures. Regular surveys would offer the opportunity to implement a transition to new census questions, while maintaining a time trend for older
questions. The high-risk nature of a one-time conventional census would be reduced. But the cost of carrying out a rolling census, with sample data collected to replace the conventional long form, backed by the need to collect decennial data with benchmark information, would almost certainly exceed the cost of a conventional census.
Output would be more frequent and could be produced more quickly. Because of the rolling nature of the sample, special procedures would need to be developed for aggregating data of sufficient sample size for small-area estimates. There would be pressure to produce results quickly and delays would be less acceptable than for the conventional census. For a large rolling census, there would be increasing pressures to maintain the data online, with periodic inclusion of the latest sample survey data.
There is no reason to anticipate that public acceptability would decline for a rolling census. However, the survey nature of the data collection would lack the high profile and historic image of the conventional census (see Chapter 6 for a discussion of data collection through a continuous measurement survey). Noncompliance would increase for a rolling census, compared with a similar decennial census, because of the survey nature of the operation. All factors being equal, the response rates and coverage would be poorer for a rolling census. There is also a special problem to note: the rolling census would be capable of providing current data on many topics that are the subject of regular sample surveys—e.g., employment, unemployment, income, and education. Given the high-quality measurements achieved for these variables by ongoing surveys, the rolling census is most unlikely to replicate their results. Having competing estimates of, say, average family income or unemployment rates may undermine confidence in the rolling census and ultimately in the nation's statistical system. Also, having two sets of estimates may lead to question about duplication and suggestions for dropping at least one of the surveys. Sponsoring agencies would understandably favor continuation of their high-quality survey and would therefore not be supporters, in this circumstance, of a rolling census.
The panel does not recommend a rolling or sample census on substantive grounds. A sample census would face serious problems in coverage (a greater proportion of the population would be included in the final count through sampling and statistical estimation), would not provide substantial cost savings, and could not be publicized as a complete count of the population. Moreover, there are some types of data (i.e., transportation information) that appear to be best collected at one point in time. A rolling census provides little advantage for overall census data and is likely to cost much more than the conventional census.
Conclusion 4.4 The panel concludes that a rolling or sample census is not an acceptable replacement for the 2000 decennial census.
See Redfern (1987) for a review of the use of population registers, in association with administrative records, for censuses.
International migration also affects the conventional census, which counts all residents of the United States at the time of the census, including illegal immigrants. Undercount estimates and procedures to correct the physical enumeration would also include undocumented aliens in the country.
It is true that permanent emigrants do not pay taxes and would cease to appear actively in some administrative records. It would still be difficult to determine emigrant status solely on evidence that someone no longer files tax forms.
The SSA is looking at ways to resume the collection of race and ethnicity information from birth certificates. But in 1994 these data are not reported to the SSA from birth certificates.
Technically, this is a difference between postcensal (estimates after a census) and intercensal (estimates made between two existing census) procedures. We refer generally to estimates in the period between censuses as intercensal, ignoring the complexity of different demographic procedures.
For reapportionment and redistricting in 2001, 2011, and so on, conventional census data would be about 1 year old, whereas the cumulated estimates would be centered mid-decade, making them 6 years old. See Fellegi (1981) for a detailed critique of the Kish proposal, including conceptual and operational aspects.
The Census Bureau is currently evaluating continuous measurement designs that include a year-zero census together with rolling surveys throughout the decade to obtain long-form items and update short-form items (see Alexander, 1993). See the discussion in Chapter 6 for additional material on continuous measurement designs.
This line of argument does not necessarily apply to all sample surveys, many of which can make use of area frames.
For example, in March 1986, the Survey of Income and Program Participation and the March Current Population Survey covered only 80-82 percent of black men ages 16 and older and 93 percent of nonblack men ages 16 and older. Coverage ratios were somewhat higher for women. (Coverage ratios compare the estimates from a survey, using the initial survey weights that take into account the sample fraction and household nonresponse, with the corresponding census-based population figures not adjusted for undercount. See Citro and Kalton, 1993: Table 3-12; see also Shapiro and Kostanich, 1988.)
The full cost of creating the address list would be incurred; moreover, the need for reliable small-area estimates would preclude the clustering of field operations, as is typically done in smaller household surveys.