5
A Redesigned Census

Chapters 2 and 3 identify two central problems with the current approach to taking the decennial census: costs have escalated rapidly while accuracy, as measured by both the overall and the differential undercount, has decreased. In this chapter, the panel identifies and makes recommendations for reducing the cost and improving the accuracy of the census.

Our overarching recommendation for a substantially redesigned census follows from three central conclusions:

  1. The massive employment of enumerators, in an attempt to count every last household that fails to return a census mail questionnaire, cannot produce an accurate and cost-effective census. Long before that objective could be achieved, diminishing returns set in, so that only small gains are achieved at great budgetary cost.

  2. After a reasonable and cost-effective effort to make a physical count of the population, the use of sample surveys and statistical estimation techniques to complete the count can produce greater accuracy than the traditional approach—especially with respect to the critical problem of differential undercount. We consider the use of statistical estimation to supplement the physical count to be the best possible realization of the actual enumeration required by the Constitution.

  3. Once it is accepted that statistical estimation can be used with accuracy to complete the counting process, substantial cost savings can be achieved by redesigning the census process from the ground up to take advantage of this change.



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 75
Modernizing the U.S. Census 5 A Redesigned Census Chapters 2 and 3 identify two central problems with the current approach to taking the decennial census: costs have escalated rapidly while accuracy, as measured by both the overall and the differential undercount, has decreased. In this chapter, the panel identifies and makes recommendations for reducing the cost and improving the accuracy of the census. Our overarching recommendation for a substantially redesigned census follows from three central conclusions: The massive employment of enumerators, in an attempt to count every last household that fails to return a census mail questionnaire, cannot produce an accurate and cost-effective census. Long before that objective could be achieved, diminishing returns set in, so that only small gains are achieved at great budgetary cost. After a reasonable and cost-effective effort to make a physical count of the population, the use of sample surveys and statistical estimation techniques to complete the count can produce greater accuracy than the traditional approach—especially with respect to the critical problem of differential undercount. We consider the use of statistical estimation to supplement the physical count to be the best possible realization of the actual enumeration required by the Constitution. Once it is accepted that statistical estimation can be used with accuracy to complete the counting process, substantial cost savings can be achieved by redesigning the census process from the ground up to take advantage of this change.

OCR for page 75
Modernizing the U.S. Census As background to our detailed recommendations, this chapter first outlines two approaches to taking the decennial census: the traditional census and a redesigned census. The chapter then discusses the major components of the second approach, identifies the critical choices that will have to be made in designing such an approach, and makes specific recommendations for the basic strategy that should be used. The chapter then considers and recommends a number of additional measures that could improve the accuracy and lower the costs of the next census. Finally, it proposes a fundamental review and reengineering of all census procedures to take advantage of the full cost-reducing potential of the new approach. TWO APPROACHES TO COUNTING THE POPULATION The traditional approach, used in the 1990 census, relies completely on intensive efforts to achieve a direct count (physical enumeration) of the entire population. The alternative approach, an integrated combination of enumeration and estimation, also starts with physical enumeration, but completes the count with statistical sampling and survey techniques. Figure 5.1 is a schematic presentation of the two approaches. The enumerative approach, labeled the traditional census on the left-hand side of the figure, begins with the construction of an address register, including elaborate procedures to improve its comprehensiveness.1 Census forms are then mailed to a comprehensive list of residential addresses, with instructions to mail back the completed questionnaire. Not all households return their completed mail questionnaire within a reasonable period of time. For households that do not respond to the mail questionnaire (35 percent of all housing units and 26 percent of all occupied housing units in 1990), census enumerators undertake an intensive follow-up effort to determine whether the unit is occupied and, if so, to contact the household and elicit responses. Repeat visits are made, administrative records are sometimes examined, and special programs to contact particular groups (e.g., homeless people) are carried out. This process is continued for an extended period of time to enumerate physically every household and all the people in every household. Extensive special programs have been directed toward coverage improvement in recent censuses. These programs are expensive, both in absolute terms and often in terms of the cost per person or housing unit. Special coverage improvement programs have included: movers check—a follow-up of people reporting a change of address to the U.S. Postal Service during the census enumeration period; prelist recanvass—in prelist areas, a recheck of the address list during the second stage of follow-up; vacant/delete check—a recheck in the field of housing units originally

OCR for page 75
Modernizing the U.S. Census FIGURE 5.1 Schematic comparison of major design features for traditional and redesigned census.

OCR for page 75
Modernizing the U.S. Census classified as vacant or as ''delete" because they were not residential—this was done on a sample basis in 1970 and for all units in 1980 and 1990; causal count—a campaign to find persons missed from the census by contacting community organizations or visiting places frequented by transients; and nonhousehold sources program—matching administrative records to census lists for selected areas. Citro and Cohen (1985: Tables 5.2 and 5.3) present cost and effectiveness data on these programs for the 1970 and 1980 censuses. The results from the mail response, enumerator follow-up, and intensive special coverage improvement efforts are combined to produce the actual enumeration of the U.S. population—both as a whole and for subdivisions down to the block level. Even though a highly labor-intensive effort is undertaken to obtain a completed questionnaire for every household, the resulting estimates contain errors. As we noted in Chapter 2, the 1990 census produced a net undercount of 1.8 percent for the nation as a whole. This net undercount included overcounting in some areas and among some groups, which was more than offset by undercounting among other areas and groups. Blacks and Hispanics, Asian and Pacific Islanders, American Indians and Native Alaskans, renters, and residents of poor inner-city areas were undercounted by larger percentages than the nation as a whole. To detect these errors, a highly intensive survey of a representative sample of areas throughout the country was conducted after the physical enumeration efforts ceased. In 1990, on the basis of a sample of areas, the difference between the census count and the count produced by the Post-Enumeration Survey (PES) was used to calculate statistical estimates—by area, racial group, and other relevant demographic characteristics—of the net undercount or overcount contained in the census data. After a series of legal battles, it was decided not to use the results of the PES to correct the initial 1990 census count, either nationally or in smaller areas. The enumerative count remained the official 1990 census estimate, used to make reapportionment and redistricting decisions and for other purposes. This enumerative approach has been subject to the two basic failings discussed in Chapters 2 and 3: high and rapidly rising costs and high differential undercount. We therefore propose for the redesigned census the approach of statistical estimation to correct the enumerative count, not as a simple add-on to the traditional census but as a completely new point of departure. By so doing we address not only the issue of differential undercount (which could have been done in 1990), but also the issue of high costs. Statistical estimation, regarded not as an add-on as was done in 1990 but as an integrated aspect of census taking, allows a fundamental reengineering of the entire census design. The alternative approach, labeled the redesigned census on the right-hand

OCR for page 75
Modernizing the U.S. Census side of Figure 5.1, combines an initial stage of direct counting (physical enumeration) with various statistical estimation techniques. Correctly designed, this approach involves more than a simple layering of a postenumeration survey on top of a traditional labor-intensive physical enumerative design. Rather, every stage of the census, including the initial stage of physical enumeration, is designed with the awareness that statistical sampling and estimation are available to remedy deficiencies and improve coverage. This approach permits the elimination of all operations that add relatively little to accuracy but have high unit costs. Indeed, it permits the reengineering of the entire traditional phase of the census with this guiding principle. Moreover, this approach produces an integrated single-number census, rather than the confusing and politically divisive result of the 1990 process, in which the enumeration produced an official census count, but the PES produced another set of numbers. The redesigned census starts, just as in the first approach, with the mailing of census forms to an address list. The redesigned census includes new methods to improve the mail response. The new methods—including respondent-friendly questionnaires, Spanish-language questionnaires for Hispanic areas, use of reminder postcards, and motivational messages to increase mail return—will decrease costs by reducing the need for personal nonresponse follow-up. A substantial program to track down those who do not respond by the designated date is carried out using census enumerators, but with significantly less intensity and of shorter duration than in the first approach. The follow-up includes a judicious combination of a second mailing to nonrespondents, follow-up by telephone, and personal visit. The follow-up, area by area, is designed with four guiding principles: (1) a reasonable attempt should be made to follow up all nonresponding households; (2) all areas should have as an objective a reasonably high response rate; (3) follow-up should be terminated after a reasonable attempt (e.g., several calls to the same household rather than as many as six calls, as was done in the 1990 census); and (4) the entire duration of field operations should be limited (e.g., 4-5 months rather than 9 to 12 months, as was done in the 1990 census). The use of these physical enumerative techniques is truncated after a reasonable effort has been made (see below for a discussion of the criteria for when to truncate). Once the follow-up is truncated, intensive surveys are taken of some sample fraction (e.g., one in four, or some fraction) of nonrespondents to get the full range of census information from the nonrespondents in those areas. On the basis of these results, statistical techniques are used to estimate, for all nonrespondents, what has to be added to the mail responses and the truncated follow-up to arrive at what the traditional census count would have been had the follow-up effort expended on the sample been implemented for all nonresponding households. The use of sampling techniques to follow up nonrespondents is called sampling for nonresponse follow-up. The less intensive the 100 percent follow-up and the smaller the sampling

OCR for page 75
Modernizing the U.S. Census fraction, the larger the cost saving but the greater the probability of errors in the count for small geographical areas. A highly accurate count for the nation as a whole could be achieved with a modest degree of physical enumeration of households and a relatively small sampling fraction. But the possibilities of substantial sampling error in any given small area, which would result in either an overestimate or underestimate of the count for that area, would be high. More intensive 100 percent follow-up and higher sampling fractions would raise costs but reduce the spread of errors in small areas. It should be noted, however, that in the traditional approach, with its effort at full physical enumeration, the range of errors in the count for small areas is still substantial. Independent of sampling for nonresponse, a further survey is taken of all households in a sample of areas to complete the census process by providing estimates of those people who were missed by previous operations. (These two important roles for the use of sampling are explained further in the box on page 81, which is based on Steffey and Bradburn, 1994.) Unlike the traditional census, in which a PES survey is taken to evaluate the final count, the redesigned census would have an integrated coverage measurement survey for the purpose of estimating the number and characteristics of those missed by the prior stages of the census.2 These are the people missed by all traditional census approaches (see Chapter 2). Overcounting, which is also detected by the PES, could also be taken into account by a coverage measurement survey. In addition, information on birth, death, and immigration might be used to check or supplement the surveys. It is this final count that is published as the official population count of the United States, for the nation, for all geographic areas, and for all the various demographic characteristics of the population. There are various schemes for exactly how these various elements should be combined and what emphasis should be placed on each. But they would all substitute some degree of statistical estimation for the expensive 1990 effort to try to count by physically enumerating—with incomplete success—every last person. In spite of several decades of attempts to improve population coverage through direct enumeration, inequities continue to exist in the census undercount by race and for geographic areas. Statistical methods to detect these inequities now exist that are feasible for use in the 2000 census. We endorse the goal of producing the best population census by counting, assignment (allocating households and persons based on good evidence—see below), and statistical estimation. In a new census environment whose last steps involve statistical estimation, the initial phases of the census would be constructed in such a way that every person would have an opportunity to be physically counted, and the Census Bureau would make a good-faith effort to count physically every person through the construction of an accurate and complete mailing list, the use of mail questionnaires, and reasonable efforts to follow-up all nonrespondent households. Following attempts to physically enumerate nonrespondents to the mail questionnaire, first through 100 percent follow-up and

OCR for page 75
Modernizing the U.S. Census The Roles of Sampling in the Census Bureau's One-Number Census In 1990, the official census figures were based on a census process that consisted of three basic steps: (1) constructing a list of addresses, (2) obtaining responses that could be linked to the address list, and (3) following up to obtain responses from those initially missed. In addition, a coverage measurement process was designed to estimate the size of the population that was missed in previous steps of the census process. Two sets of population totals were produced, with and without corrections based on coverage measurement, and an ex post facto decision was made about whether to accept the corrected totals. This "adjustment" decision proved to be controversial because it occurred in a highly politicized environment in which interested parties perceived themselves as winners or losers, depending on which set of numbers was chosen. For the 2000 census, the Census Bureau has proposed the concept of a "one-number census," one that will provide ''the best possible single set of results by legal deadlines, … based on an appropriate combination of counting, assignment, and statistical techniques" (Miskura, 1993). In the Census Bureau's definition, counting refers to all methods for direct contact with respondents, including mail questionnaires, personal visits, and telephone calls. Assignment refers to the use of information from administrative records to add people to the count for a specific geographic location without field verification. Statistical techniques for estimation include imputation procedures, sampling during follow-up of nonrespondents, and methods for measuring census coverage. These three components of a one-number census are designed to complement one another. In particular, the results from coverage measurement will be fully integrated into the official census estimates (Miskura, 1993). Sampling will be used in the new census process in two distinct ways. The first is in nonresponse follow-up. After initial contact and follow-up activities, a sample of households or blocks containing households that did not respond are selected for further follow-up. Estimates, based on the sample, are made of the numbers and characteristics of those who would have responded to the follow-up had it been conducted on the entire population. This use of sampling is proposed primarily to reduce costs by reducing the use of expensive follow-up procedures. It has other advantages too: it can reduce the time needed for the follow-up process, thereby allowing more time for coverage measurement. And, although it contributes to variability through sampling error, it can reduce other kinds of errors. The second way in which sampling will be used in the one-number census is in coverage measurement. A sample survey is conducted from which estimates are made of the numbers and characteristics of those not counted in previous steps of the census process, including those who are not accounted for in the nonresponse follow-up step. This use of sampling is designed to improve the accuracy of the census. In 1990, sampling was used in a coverage measurement process (See Figure 5.3) for evaluation purposes and for possible use as a basis for official population estimates, with a decision on the latter deferred until after the results of the sample were examined. The previous steps of the census process were designed to be used as the basis for reported census figures independently of the coverage

OCR for page 75
Modernizing the U.S. Census measurement process. In the proposed one-number census, coverage measurement is conceived as an essential part of census-taking and not just as an evaluation of other census operations. In particular, it is not regarded as a method of producing a second set of population estimates that competes with the population estimates obtained without its use. Coverage measurement, which includes sampling, statistical estimation based on the sampling, and statistical modeling, will be integrated with the other census-taking operations. Hence, this phase of census-taking is called integrated coverage measurement. The fundamental innovation of a one-number census is that an appropriate methodology for integrated coverage measurement is established before the census is carried out, with the recognition that results at previous steps cannot be regarded on scientific grounds as viable alternatives to the final, best set of official population estimates. It also allows for more optimal design of other census operations. Source: Adapted from Steffey and Bradburn (1994). then through sampling, the census would use survey-based statistical estimation to complete the final official population count. As mentioned earlier, there are several additional specific ways to achieve cost savings in the future integrated single-number census, including: respondent-friendly questionnaires and other improved procedures for increasing mail response rates, truncating census operations earlier to minimize the lengthy work period for district offices and census field staff, and increased use of U.S. Postal Service employees for such operations as vacancy checks of housing units. But, most important, statistical estimation at the final stage of census operations and abandonment of the target of physically counting everyone permit a complete reengineering of the basic census operation in light of the new census context. Thus, numerous expensive coverage improvement operations used in the 1990 census can be reviewed with an eye to elimination as unnecessary (compare the left and right sides of Figure 5.1). As a guide to the exposition below, we note that there are four key stages in the redesigned census (see the right-hand side of Figure 5.1) leading to the final official population count: mail response, enumerator follow-up, sampling for further nonresponse follow-up, and a survey together with other statistical estimation techniques to complete the count. Ways to improve the mail response are described later in a section entitled "Improve Response Rates." Issues relating to enumerator follow-up are covered in several sections of the next major section, in which we discuss decreasing the intensity of nonresponse follow-up and truncation of enumeration after a reasonable effort. After describing a reasonable effort at enumerator follow-up, we discuss the role for "Sampling for Nonresponse Follow-Up" in a separate section. We describe the role of an integrated coverage survey in a later section entitled "Survey-Based Methods to

OCR for page 75
Modernizing the U.S. Census Complete the Count." With the use of statistical estimation (the integrated coverage measurement survey) built into the redesigned census, the entire process of census enumeration needs to be reengineered. Under the new procedures of a redesigned census, there would be a reduction of special coverage programs that are no longer cost-effective. The reengineered census, along with a review of some specific census operations, is presented below in a major section entitled "A Reengineered Census." Census Bureau Plans The panel would like to recognize that the Census Bureau is actively working on plans and conducting research on much of what is described in this chapter. Much of the material reported here is based on ongoing census research. The Census Bureau, however, has not yet prepared a specific proposal for the 2000 census design. Census Bureau plans for the 2000 census will be informed by results of census tests in 1995. The Census Bureau will carry out a major test of census design options at four sites in 1995. The test will examine the following methods: Sampling for the follow-up of nonrespondents to the mail questionnaire. Statistical methods to estimate separately the number and characteristics of people missed because their housing unit was missed, people missed within enumerated housing units, and people who were erroneously enumerated. Coverage questions for a complete listing of household members. Mailout of Spanish-language questionnaires. Target methods to count historically undercounted populations. Counting persons with no usual residence. Making census questionnaires more widely available by placing unaddressed questionnaires in accessible locations (e.g., stores and post offices). Respondent-friendly questionnaire design and a full mail strategy for improved mail response. Telephone response as a follow-up option. Real-time automated matching for record linkage. Using the U.S. Postal Service to identify vacant and nonexistent housing units. Electronic imaging to scan respondent-friendly census forms and to capture data. Cooperating with the U.S. Postal Service for the continuous updating of a national housing address file. Collecting long form-type data using matrix sample forms (two or more sample forms with only a subset of the questions).

OCR for page 75
Modernizing the U.S. Census Results from the 1995 census tests, and continued tests from 1996 to 1998, will be important for operational plans for the 2000 census. At the moment, the Bureau of the Census (1993a) proposes to conduct a "one-number census" that will use statistical methods integrated into the census process. Except for a one-number census, the Census Bureau has not made a formal commitment to the specific recommendations that the panel makes in this chapter for a redesigned census. Moreover, we propose a much more extensive redesign of the 2000 census. In recent congressional testimony, Harry Scarr, the acting director of the Bureau of the Census, stated that the 1995 census tests would provide important results for planning the 2000 census, including: (1) detailed evaluations of sampling for nonresponse follow-up and sampling as part of statistical estimation for census coverage, (2) operational and cost information for new sampling and estimation methods, and (3) evaluation of the effect of new methods on errors in census data (Scarr, 1994). The 1995 census test will also examine the quality of the census address lists, in a cooperative venture in which U.S. Postal Service delivery addresses are used Legal Issues of Statistical Estimation As part of its examination of alternative methods for counting the population, the panel examined the legal issues for the greater use of statistical estimation. The requirement for effort at complete enumeration does not necessarily rule out the increased use of sampling as part of the census. Sampling has several potential uses that are worth inspection for increased use in future censuses. It is important to distinguish among three potential uses of sampling in the census. The first is to estimate the number and characteristics of people who were not found even after attempts at nonresponse follow-up and people who were counted more than once or should not have been counted. This use of sampling to complete the count is referred to as an integrated coverage measurement survey in Figure 5.1. The second broad use of sampling in a census is sampling within the actual process of physical enumeration. The number of nonresponding households is known, so the purpose of sampling the nonrespondents is to estimate their characteristics. The third broad use of sampling is to take a sample of the population at the beginning of the census. In such a "sample" census, no attempt would be made to enumerate every person. The panel discusses this option in Chapter 4. In the panel's judgment, the spirit of the constitutional, legislative, and judicial history regarding enumeration is compatible with the use of sampling as part of the census process, so long as that process includes a reasonable effort to reach all inhabitants. Specifically, we believe that census designs that use sampling for the follow-up stage of census operations (after an initial reasonable

OCR for page 75
Modernizing the U.S. Census attempt has been made to collect a census questionnaire from everyone) and for the final stage of census operations for an estimate of the population not counted by physical enumeration (using an integrated coverage measurement survey for statistical estimation to complete the count) would meet both substantive and legislative requirements for reapportionment. Such designs also have the potential to increase census accuracy and reduce census costs (see Appendix C for a discussion of the legislative requirements for the census). Precedents from previous censuses exist for the use of sampling. Several court cases have explicitly upheld the constitutionality of an adjustment based on a survey (such as the PES in the 1990 census), citing the importance of having data as accurate as possible for reapportionment and redistricting.3 The question of the legality of sampling for follow-up for nonresponse has never been explicitly raised in the courts; however, language used in the relevant court cases would clearly seem to be consistent with its use. Court decisions on census issues have deferred to the Census Bureau on technical issues. The panel's review of legal issues in Appendix C does not provide grounds for concern that courts will question the expanded use of sampling in the census (i.e., sampling for nonresponse follow-up or a sample survey to estimate the number and characteristics of people missed in the physical enumeration). Courts have accepted the use of sampling in prior censuses, and we propose a more extensive use of sampling in the 2000 census. After a review of past court decisions, the panel expects that the courts will continue to accept the technical decisions made by Census Bureau staff, including the use of sampling in a good-faith effort to count the population, in a redesigned census. BASIC ELEMENTS OF A NEW CENSUS DESIGN The conventional mailout/mailback census operation that the Census Bureau has conducted for the 1970, 1980, and 1990 census can schematically be thought of as involving five stages: (1) the development and checking of a master address list prior to the census, (2) the physical enumeration of the population through the mailing of the census questionnaires several days before April 1 (Census Day) and the collection of the returned questionnaires, (3) the follow-up of those housing units that do not return the census questionnaire, (4) a variety of coverage improvement activities, and (5) checks on the census coverage. One critical area that would be affected by our proposal is point (3). Decreasing the Intensity of Nonresponse Follow-Up In practice, the follow-up of all nonrespondents is a costly procedure—the precise magnitude of the cost depends on the intensity of the effort. Follow-up of nonrespondents was directed by each of the 449 district offices in the 1990 census. Follow-up began about three weeks after Census Day, with enumerators

OCR for page 75
Modernizing the U.S. Census rates—incur additional costs that need to be considered when weighing the benefits from higher response rates. Additional costs involve the mailing and printing costs of a reminder postcard or a replacement questionnaire, the possibly higher costs of printing and processing a respondent-friendly census questionnaire, and the use of telephone calls to urge nonrespondents to mail in their questionnaires. There may be additional costs to respondent-friendly forms that may increase machine costs for processing and tabulation. Optimizing for respondent ease may increase response rates, but it may also increase machine costs. All extra costs must be weighed against the cost savings from increased mail response rates. Some methods to increase mail response rates may not, after careful study, yield an overall net cost savings. Respondent-Friendly Long Forms and the Use of Appeals The Census Bureau conducted the Appeals and Long-Form Experiment in 1993 with two objectives: to determine the influence of two types of respondent-friendly construction on response rates for the census long form and to determine the influence of three types of appeals on response rates to the census short form. The ALFE work involved the use of methods to improve mail response rates that had been learned from the SQT. Results from the ALFE test are described in greater detail in Chapter 6. The ALFE research leads to two conclusions for taking the census. First, it is possible to improve the census long form by designing it as a more respondent-friendly questionnaire. Second, stressing the mandatory nature of the census ("Your response is required by law") substantially improves the mail completion rates. Although the ALFE study did not test directly the use of a mandatory appeal with the long form, the results suggest that it may be possible to increase the long-form mail completion rates by as much as 15 percentage points by using a respondent-friendly questionnaire and stressing the mandatory and confidential nature of the census. Recommendation 5.3 The panel recommends that the Census Bureau incorporate successfully tested procedures to increase the initial response rate in the 2000 census, including respondent-friendly questionnaires and expanded efforts to publicize the mandatory nature of the census. Partnerships with State and Local Governments In rethinking census operations for the future, the Census Bureau needs to develop new partnerships. One of the two most important partnerships needed is with state and local governments. The other is with the U.S. Postal Service, discussed in the next section.

OCR for page 75
Modernizing the U.S. Census The Census Bureau needs to work with state and local governments to improve understanding of the new methods for the 2000 census, changes in census enumeration operations, and the use of integrated coverage measurement. The 2000 census will have different methods for coverage improvement, with a decreased reliance on intensive and expensive direct enumeration. Local and state governments need to be informed about these new approaches and to understand how they will affect census operations in their areas. Congress should evaluate the Census Bureau for its success in establishing cooperative state and local programs. A special goal for improved cooperation between the Census Bureau and local governments is to reach agreement on the housing address list for the decennial census. Approximately one-half of census undercount is attributable to missed housing units. Local governments have often criticized past censuses because they believe that housing units exist that were not counted by the census. However, the census address list has been deemed confidential by the Census Bureau so that local governments have not been able to make direct comparisons between their address list and the list used for the census. It would be valuable to have a new partnership in which local governments and the Census Bureau work together to construct the best possible address list for the census. Such a partnership would also reduce a major source of disagreement about the accuracy of the population count reported in the census. The panel discusses the sharing of the Census Bureau's address list in two places in this report: in this chapter and in Chapter 8, where we discuss a continuously updated address file. In both instances, an important aspect of the sharing of census address lists is the confidentiality restriction on the Census Bureau contained in Title 13, the federal legislation governing the Census Bureau's activities. The current interpretation of Title 13 by the Census Bureau is that it does not permit the agency to allow others to examine the list of individual addresses that it assembles for the decennial census. We discuss these issues more fully in Chapter 8, in a section entitled "Title 13 Concerns." Recommendation 5.4 The panel recommends that the Census Bureau develop cooperative arrangements with states and local governments to develop an improved master address file and that Congress amend Title 13 of the U.S. Code to permit sharing of this file for statistical purposes with state and local governments and the federal statistical system. Partnership with the U.S. Postal Service The second major partnership needed is between the Census Bureau and the U.S. Postal Service. Historically the U.S. Postal Service has had an important role in the decennial census, delivering and returning the bulk of the census

OCR for page 75
Modernizing the U.S. Census questionnaires by mail. In 1990, the letter carriers delivered about 90 million census questionnaires to individual households. Whereas such work may be considered part of the normal duties of the U.S. Postal Service, it is critical that the questionnaire be delivered correctly and on time. This section reviews the current role of the U.S. Postal Service and the panel's suggestions for an expanded role in future census activities. Current Role The U.S. Postal Service has played key roles in several steps in recent censuses. First is development and improvement of the census address list. Beginning about 2 years prior to the 1990 census, the U.S. Postal Service provided assistance in several phases of advance census address checking. Using either commercial mailing lists or census enumerators, the Census Bureau prepared preliminary address lists for rural areas, small and medium towns, and major urban centers. U.S. Postal Service letter carriers checked each address on their mail routes, made corrections, identified duplicates and undeliverable addresses, and added missing addresses. About 75 million addresses were checked in this work. These verifications and corrections were completed by the U.S. Postal Service by April 1989. The letter carriers then completed a final review of the mail addresses in February and March 1990, immediately before the mailing of census questionnaires. For this operation, address lists were checked again to identify missing, duplicate, and undeliverable addresses. The second step is providing the prompt and accurate delivery of census questionnaires. U.S. Postal Service letter carriers delivered all census questionnaires to housing units in the short period before Census Day (April 1, 1990). This was a massive, coordinated delivery operation, with more than 90 million questionnaires, which had to ensure that every household received the proper questionnaire at the proper time. The important role of the letter carriers in accurate and timely delivery of questionnaires cannot be overstated. The third step is automated sorting of mail returns. About 66 million completed mail questionnaires were returned to either 1 of the 346 district offices or 1 of 7 special processing offices.7 The U.S. Postal Service designated a special zip+4 postal code for each census office to ensure the timely and efficient return of questionnaires. The Census Bureau printed these zip+4 postal codes on the return envelope, and postal facilities sorted the returns with their automated equipment. These arrangements helped to improve the speed of and logistics for the delivery of the completed questionnaires at the processing centers. The fourth step involved checks for correct address for matched questionnaires. Several 1990 census programs encouraged people to complete a questionnaire if they thought that they had not been enumerated. About 3 million questionnaires were received through this program. The Census Bureau was

OCR for page 75
Modernizing the U.S. Census required to match the address of the questionnaires with the census address list to ensure the correct enumeration of persons at their usual residence. If the address on the questionnaire could not be matched, the address was sent to the U.S. Postal Service for verification. Addresses classified as deliverable addresses by the U.S. Postal Service and the people corresponding to these addresses were then added to the census. Finally, the U.S. Postal Service was involved in the census closeout address check. During the final period of the nonresponse follow-up operation, some district offices used the U.S. Postal Service to obtain information about unenumerated addresses. These district offices provided an address card for each unenumerated address to the U.S. Postal Service and asked for a limited amount of information about the address (type of structure, occupancy status on Census Day, and the number of occupants on Census Day). The results of this limited operation in 1990 were that 142,000 address cards were sent to the U.S. Postal Service and 35,000 were returned. Only 17 percent of the returned cards provided the number of Census Day occupants. The Census Bureau concluded that more controlled research was needed to determine the usefulness of this limited information from letter carriers in the final stages of follow-up of nonrespondents. Expanded Role In the panel's judgment, the U.S. Postal Service should continue to play a strong role in the decennial census. In addition to the tasks noted above, there are several other ways in which the U.S. Postal Service might support the conduct of the census. Some new ways are outlined in a letter of agreement from the U.S. Postal Service and the Census Bureau (Green and Scarr, 1993), which we summarize here. Maintenance of a continuously updated residential address list. Census-taking relies heavily on an accurate and complete address list. The Census Bureau and the U.S. Postal Service are working on an arrangement to share address information. Such an arrangement would start with the 1990 census address list as a base and then develop a permanent and continuously maintained list of addresses. The updating would rely on the firsthand knowledge of letter carriers—who gain through their daily work information about the inventory and location of housing unit addresses. During the course of their normal duties, letter carriers would record the existence of each unit to which mail is delivered on their routes. These records would be collected by the U.S. Postal Service into a single national file of address information. The proposed methodology is to periodically link the continuously updated U.S. Postal Service address information with Census Bureau address files. Such

OCR for page 75
Modernizing the U.S. Census a linkage would reveal addresses missing from the census list. A contribution of the Census Bureau to the joint work would be linkage to its geographic files and maps. With updating of new addresses, the linkage would also reveal new streets. Following the identification of missing streets, the Census Bureau would provide maps with information of interest to the U.S. Postal Service for its route and mail delivery planning. Such collaborative work may require legislative changes for the sharing of address information. We address cost issues and the desirability for an ongoing national address file in more detail in Chapter 8. Identification of vacant housing units. Because the misclassification of occupied housing units can cause coverage errors in the census, in the 1990 census at least two enumerator visits were made to each potentially vacant address. Census costs could be reduced if one or more of these enumerator visits were eliminated by relying on the local letter carrier to ascertain the vacancy status of the housing unit. The letter carrier may also be able to complete the work more quickly because he or she visits each address almost every day. Ascertaining the vacancy status more promptly, closer to Census Day, would improve the data quality of the census. Encouragement of local efforts to convert rural addresses to addresses with house numbers and street names. Unlike city addresses, rural addresses do not provide both a mailing address and a unique physical location for the housing unit. Many rural addresses use a rural route and box number, a post office box, or general delivery. Such rural addresses are a problem for both the U.S. Postal Service (because they present difficulties for automated sorting of addresses for more efficient mail delivery) and the Census Bureau (because they are difficult to handle in automated matching for assignment of geographic codes). The decennial census needs to assign every individual to a specific small area of geography in order to fulfill the data requirements for legislative redistricting. The census could not, for example, simply record thousands of people residing at the specific location of a central post office. Therefore, regardless of the location and nature of the mailing address, the census must be able to associate a specific physical location for the residence of each family and individual recorded in the census. Because city-style addresses allow more efficient routing of emergency services (fire, policy, and ambulance), many rural areas of the nation are converting to city-style addresses. Both the U.S. Postal Service and the Census Bureau will benefit from these conversions. Recommendation 5.5 The panel recommends that the U.S. Postal Service and the Census Bureau continue to work together to improve the decennial census. We endorse an expanded role for the U.S. Postal Service in the 2000 census in several areas: (a) development, maintenance, and improvement of an accurate address file

OCR for page 75
Modernizing the U.S. Census for the nation's residential housing units, (b) checking the address list prior to the census to improve accuracy, (c) delivery of the mailed forms, and (d) ascertainment of the vacancy status of housing units during the census. A REENGINEERED CENSUS The panel is extremely concerned about the costs and the accuracy of future censuses. The previous sections have detailed ways in which net census costs could be reduced by $300 to $400 million, principally by truncating the follow-up stage of census operations after a reasonable effort and by relying on sampling for nonresponse follow-up. We have also reviewed ways to improve the completeness of the census count for the total population and for subgroups through the use of integrated coverage measurement as the last stage in arriving at the final population count (the "actual enumeration" called for in the Constitution). We emphasize that the use of integrated coverage measurement affords the opportunity to realize yet additional cost savings in the census by means of a thoroughgoing "reengineering" of the entire census process. Basing the census on the use of statistical estimation to compensate for the unavoidable undercoverage that occurs in traditional census enumeration means that the entire procedure of census enumeration and processing operations can and should be rethought. In the past, the direct enumeration operations have sought, at great expense, to count every person. If a person was not counted, he or she was omitted from the final census count except as imputed in census allocation procedures. Under the redesigned census we propose for future censuses, the uncounted population will be estimated by relatively inexpensive survey-based statistical techniques, and the final count will include both the estimated as well as the directly counted population. It should be possible, therefore, to greatly reduce the costly intensive effort to count every person directly. Since such a reduction would eliminate those operations having the least relative cost-effectiveness, there is the potential to realize substantial savings. The "unexplained" increases in census costs that occurred over the period 1970 to 1990 amounted to $1.3 billion. (By unexplained increases, we mean the cost increases that are not explained by wage and price increases, the growth in the number of housing units, and the decline in the mail response rate.) We were not able to decompose analytically the causes of these increases, in terms of specific operational and managerial decisions. However, we strongly believe the cost increases stemmed heavily from efforts, during a period of declining public cooperation with the decennial census, to meet strong political pressures for accurate counts by area and population group through the application of highly expensive and labor-intensive efforts to make a physical count of every person. As just one example, field offices stayed open for 9 months in 1990, compared with 6 months in 1980, and 3 to 4 months in 1970.

OCR for page 75
Modernizing the U.S. Census Not only were there cost increases as the result of new programs intended specifically to improve coverage (e.g., rechecking every housing unit originally determined to be vacant), but also substantial cost increases occurred in major components of the census (Chapter 3 includes a more detailed description of the cost components). Thus, in five key categories, in 1990 dollars, costs increased from 1970 to 1990: by $75 million for census planning activities, $1,022 million for data collection, $311 million for data processing, $92 million for data dissemination, and $68 million for census testing (but note that these increases have not been adjusted for the growth in the number of households or the decline in the mail response rate). The total of unexplained cost increases, as we noted earlier, amounts to $1.3 billion in constant 1990 dollars. Given these large increases from 1970 to 1990, we are convinced that abandonment of the goal of physically counting every person and completely rethinking all of the census operations in the context of the new design would make possible substantial additional savings—for example, through reducing some expensive coverage improvement programs that would no longer be needed if statistical measures were used to complete the count. With the adoption of a new census design that includes integrated coverage measurement to complete the count, there is an opportunity and a need for knowledgeable census staff to review every aspect of census operations to identify procedures that are no longer necessary or that can be redesigned to be more cost-effective. This review of costs would be similar to the kinds of reengineering that corporations (and agencies of government) have undertaken in recent years to rethink and streamline their production processes. A reengineering effort would probably begin with a more detailed examination of previous census operations, which might determine that particular expenditures (and expenditure increases) were well justified in light of the overall design for past censuses. However, we propose a new census design; hence, the question is not whether a particular operation was justified in the past but whether it is needed in the future and, if so, whether it can be streamlined in some manner, combined with other operations, or otherwise made more cost-effective. Thus, the review must go beyond a line-by-line audit to consider each operation in the context of the overall approach that is proposed for the 2000 census, namely, one in which statistical estimation plays an important role along with direct enumeration. To put it another way, it is critical to challenge the attitude that the census needs to be improved without regard to costs and cost-effectiveness. Given the legal challenges to the Census Bureau in recent decades, one can understand such an attitude. But the panel believes that a more productive approach would be to reengineer census operations for greater cost-effectiveness in the context of the proposed census redesign. At this writing, we are unaware of any activities by the Census Bureau to carry out the extensive census reengineering that we propose. We urge the Census Bureau to commit to and carry out as soon as

OCR for page 75
Modernizing the U.S. Census possible such a reengineering effort, in which the basic structure and organization as well as all operations and managerial techniques of the census are reviewed with the goal of reducing costs while maintaining the essential quality of the results. We have identified $300 to $400 million in net savings that could be achieved by specific innovations—principally, the use of truncation and sampling for nonresponse follow-up. But for a reengineering effort to be successful, we believe it essential to present census planners with a target for cost savings that presents a genuine challenge to encourage new thinking. As one managerial technique to help identify cost savings, we propose that the Census Bureau flesh out the best possible plan to produce a new census for 2000 at a total cost (adjusted for increases in housing units and prices since 1990) that would eliminate a substantial fraction of the $1.3 billion ''unexplained" increase in costs between 1970 and 1990, with a target for cost savings much larger than the $300 to $400 million we have already identified. The Census Bureau should then evaluate this plan in terms of its effects on the quality of resulting data products. We argue that substantial cost savings are reasonable for the purpose of this planning exercise; that is, a target for planning a reengineered integrated single-number census in 2000 that would cost substantially less than the 1990 census (after allowing for changes in prices and household numbers). Our rationale for such a goal follows. The 1990 census cost $1.86 billion more (in constant dollars) than the 1970 census. Of this total, about $450 to $580 million is explained by the increase in housing units and the decline in the census mail response rate, leaving remaining cost increases of $1.3 billion (see Table 3.2). We have estimated potential net savings from truncating 100 percent follow-up after a reasonable effort and relying on sampling to complete the non-response follow-up operation of $300 to $400 million. Even more important, the new census design would use statistical estimation to complete the count. As a consequence, operations and procedures need no longer be designed to try to count every last person, an effort to which we have assigned major responsibility for the $1.3 billion cost increase from 1970 to 1990. The questions that should be examined in this planning exercise include: How many district offices are needed for the census? How long do they need to be open? How many enumerators are required for a less intensive physical enumeration? What is the most cost-effective method of payment for enumerators? What levels of staffing and supervisory personnel are needed? Given the use of statistical estimation to complete the count, which coverage improvement operations can be scaled back or eliminated? How can automation be integrated into census field activities so as to reduce the combined costs of people and systems? How can data processing be made more efficient? How much of the product line can shift from print to electronic media? Since we have not been able to identify the specific elements of the 1970 to 1990 cost increase, we cannot predict the full magnitude of the savings that

OCR for page 75
Modernizing the U.S. Census would be feasible while still producing a high-quality census. But we are convinced that a good-faith effort to develop such a plan would identify important cost savings that could be made without a significant sacrifice of quality. We emphasize that a planning target—eliminating a substantial fraction of the 1970 to 1990 "unexplained" cost increases is just that—a planning target. In its review, the Census Bureau should identify what would have to be done to meet a target and spell out the consequences both for its operations and for the quality of the resulting census. Even if it is ultimately deemed infeasible to achieve the full magnitude of the savings set as a target for the exercise while still maintaining the quality of the census, a good-faith effort could still identify substantial cost savings, as all census operations are reengineered for a single-number census in which physical enumeration and statistical estimation are fully integrated. Recommendation 5.6 We recommend that the Census Bureau undertake a thorough reexamination of the basic structure, organization, and processes by which the decennial census is conducted to obtain the full cost-saving potential of the proposed redesigned census. As one part of its reexamination, the Census Bureau should develop a plan for the 2000 census that eliminates a substantial fraction of the $1.3 billion cost increase (in 1990 dollars) from 1970 to 1990 that is not accounted for by the growth in housing units and the decline in the mail response rate. The target for this plan should be much more than the $300 to $400 million we have already identified. BUILDING PUBLIC SUPPORT The census involves a joint commitment of the government and the American people. The government—involving federal, state, and local agencies—attempts to communicate with the people in a detailed and specific way by sending out census questionnaires and receiving replies. The American people expect a fairly apportioned Congress, equitable distribution of monies from government programs, and accurate and useful data on themselves. From this perspective, a census with incomplete or poor-quality data may be considered to be a failure of the mutual commitment of the government and the people about the census. The failure may come about from a failure by the government to provide an appropriate set of questions or methods for collecting information about those questions, a failure to indicate to the American people that it is important to respond to the census questionnaire, a failure to send out a questionnaire that is properly perceived and understood (a questionnaire that is hard to understand or is not in the language of the respondent), or a failure to send out a questionnaire

OCR for page 75
Modernizing the U.S. Census or provide a way of responding that can be acted on in the social setting in which the respondent resides. The census process may also break down from adverse publicity or the widespread reluctance of the American people to participate in the census, or a combination of both. Unfavorable or sensational news that ridicules the census or its questions or highlights charges of discrimination adversely affects the mail return rates and increases the cost of data collection for nonrespondents. Thus, the process of the census program involves not only the attempt to publicize the census favorably in the mass media, but also sending messages to the public about all aspects of the census process that can be received, understood, accepted, and acted on. The actual counting of the population in the census depends on explaining census operations to the people and on building a mutual agreement for the undertaking. The challenge for the 2000 census will be to conduct a public education campaign early enough to allow scrutiny of all the new features of a census that is fundamentally changed. We believe that the next U.S. census should be fundamentally redesigned to provide greater accuracy at lower cost. If a redesigned census is decided on, the methods should be determined by the Census Bureau, with technical advice from appropriate sources. The Census Bureau will need to state publicly the methods it proposes to follow in the census. If the 2000 census is redesigned, it is critical to state those methods as soon as possible in order to obtain comments from government agencies at the federal, state, and local levels, from census data users, and from the public. The procedures finally decided on will need to be announced well in advance of Census Day, 2000. Allowing a reasonable period for public comments suggests that the redesigned census, in its broad features, needs to be presented publicly within the next year or so. Recommendation 5.7 The panel recommends that the Census Bureau publicize, as soon as possible, its plan for implementing a redesigned census. It should then move to obtain early commitment for the new plan from both Congress and data users. NOTES 1   Appendix B outlines the key steps in the traditional census process, describing by year the activities undertaken in the 1990 census. 2   A survey does not necessarily have to come after the sampling for nonrespondents in some designs proposed for statistical estimation of undercoverage for the 2000 census. This independent survey for coverage estimation could occur during the census.

OCR for page 75
Modernizing the U.S. Census 3   The recent case of City of New York v. U.S. Department of Commerce (1993, Eastern District, New York) is the latest court decision that discusses, in part, the statistical methods that might be used for the conduct of the census; Appendix C presents a fuller discussion of legal issues. 4   Available evidence from the 1980 census indicates that the costs per follow-up interview increase with the more calls that are attempted for nonresponding households. Citro and Cohen (1985: Appendix 6.2) discuss information on follow-up procedures in the 1980 census and illustrate the costs for a census that uses reduced follow-up callbacks or sampling in the nonresponse follow-up. 5   See the report of the Panel to Evaluate Alternative Census Methods (Steffey and Bradburn, 1994:97-105) for a more detailed discussion about the possible operations for nonresponse follow-up. 6   The joint use of the PES and census to estimate the "true" population size is a type of dual-system estimation. Dual-system estimation is generally used in many ways to estimate human and animal populations. For biological research, animals are caught in a survey and then matched against the number caught at later times or who are spotted in a count of an area. Biologists refer to these estimates as capture-recapture estimates. For demographic research, the dual-system estimation is used to provide an estimate of coverage for the census. 7   The 1990 census operated 449 district offices, plus 9 in Puerto Rico. Only 346 offices processed completed mail questionnaires.