Mr. Chairman, Mrs. Maloney, and Members of the Subcommittee:
It is a pleasure to be here today to report on the Accuracy and Coverage Evaluation (A.C.E.), which will provide a report card on Census 2000 enumeration operations. The Census Bureau takes much professional pride in carefully reporting to the Congress and to the American people how it has done. At previous hearings, we have discussed the many operations that the Census Bureau is undertaking to make Census 2000 as complete and accurate as possible. Census 2000 operations are more robust and innovative than those for any previous census. Operations are going well. I have noted that the mailback response rate to the census--now 66 percent--is encouraging, as is the fact that, to date, we have completed 39 percent of the nonresponse followup workload.
Still, the Census Bureau does not anticipate at this point that Census 2000 will have better coverage than the 1990 census because many of the factors that led to the undercount in 1990 are still present in American society today. The Census Bureau has both measured and documented the existence of a substantial undercount since the 1940 census. This documentation reveals the persistence of a large differential undercount, existing for at least the last 60 years, between the Black and non-Black populations. The 1990 census indicated that the differential undercount is not purely a phenomenon of the Black population. Children, renters, Hispanics, Asians and Pacific Islanders, and American Indians were also found to be disproportionately undercounted.
The Census Bureau has been running harder, but believes this will only allow us to stay even. That is, we expect that neither the overall coverage levels nor the differential undercount rates in Census 2000 will show improvement over 1990. The Census Bureau strongly hopes to be proven wrong in this assessment and the A.C.E. will give us the information to determine this.
The A.C.E. will provide a final quality check on how well we have done in the initial census. The alternative is to not do the A.C.E. and never know how we have done below the national level (where demographic analysis provides a benchmark). The A.C.E. will also provide the means to generate more accurate counts. The 1990 version of the A.C.E., the Post-Enumeration Survey (PES), provided information that was used during the 1990's to improve statistical programs. The population estimates that the Bureau of Labor Statistics (BLS) asked us to incorporate into the Current Population Survey program following the 1990 census were corrected for the "undercount" identified through the PES. The BLS also requested adjusted population controls for the Consumer Expenditure Survey. All other major national demographic surveys conducted by the Census Bureau for other agencies of the federal statistical system also were converted to this adjusted population base.
Katherine Abraham, Commissioner of the Bureau of Labor Statistics, wrote me on March 11, 1999, that if the BLS had used uncorrected numbers, its estimates of the overall level of employment and unemployment would have been too low and the published geographic and demographic distribution of unemployment and other measures would have been inaccurate.
Mr. Chairman, in your letter of invitation you asked that I be prepared to provide the status and a brief overview of the A.C.E. 2000 methodology and operational time line, and readiness for key activities and dates that lay ahead. Please include, but do not limit your testimony to the following activities: address listing, sample size and selection, person interviewing, person matching, Dual System Estimation, synthetic estimation, release of adjusted population counts and underlying data, and plans to evaluate the adjustment. I will discuss the last two points first.
The Census Bureau is committed to making its data as accurate as possible for all uses of the data. In accordance with the 1999 Supreme Court ruling, the Census Bureau will not use statistical sampling (the A.C.E.) to produce state population totals used for congressional apportionment, which must be produced by December 31 of this year. The Census Bureau does plan to use statistical sampling techniques to produce the more detailed data required for redistricting and federal program purposes because we believe we can produce more accurate data by incorporating proven sampling methodologies into traditional enumeration procedures. Section 141(c) of Title 13 requires the Census Bureau to report redistricting data directly to the states by April 1, 2001. Public Law 105-119, the Census Bureau's appropriations bill for FY 1998, requires the Census Bureau, upon release of redistricting numbers based on statistical sampling, also to release comparable data produced without the use of sampling, and we plan to do so.
I have said previously that the Census Bureau currently expects that the corrected numbers using A.C.E. will be the more accurate numbers. If the Census Bureau does not have confidence in the A.C.E. results, we will not use it. The decision whether to release the statistically corrected data should take into consideration operational data to validate the successful conduct of the A.C.E., whether the A.C.E. measurements of undercount are consistent with historical patterns of undercount, and a review of selected measures of quality. In the fall of this year, the Census Bureau will discuss the review process and criteria with the statistical community and other interested parties.
All major operations of the A.C.E. have been designed and documented and the details have been made available for review and comment. The National Academy of Science's Panel to Review the 2000 Census, chaired by Janet L. Norwood, was convened in the fall of 1998 to review the methods, procedures, and results of Census 2000. Among other things, the Panel is reviewing the statistical methods and operations of the A.C.E. and Dual System Estimation. The Census Bureau prepared several documents for a Panel workshop in February, and these documents, listed in the Appendix, were made available then and by separate letter in early April to you, Chairman Miller, and to Mrs. Maloney, Chairman Rogers, and Mr. Serrano. Subcommittee staff were briefed on these documents by John Thompson and Howard Hogan on April 17. (The Appendix also lists and notes one document that was prepared subsequently.)
STEPS IN THE A.C.E. PROCESS
Now, I will describe the basic A.C.E. operations. Several major A.C.E. operations have been completed, one is ongoing, and others will follow the completion of nonresponse followup. All operations are currently on schedule. The basic concept behind the A.C.E. is the comparison of data from two systems--an independent survey of about 314,000 housing units (including Puerto Rico) and the initial census. Because of its small size relative to the initial census, we believe we can do a better job enumerating people in the housing units in the sample. We can be more selective about the interviewers, train them longer and pay them more, and provide more quality assurance.
Sample Design.--The first step in the A.C.E. process is to design and select the sample, which consists of approximately 314,000 housing units, or about one-fourth of one percent of total housing units. We selected the initial sample in the first half of 1999 and subsampled in early 2000 to provide a more efficient sample design and meet our field requirements. The basic units of the sample are what we call "block clusters"--single census blocks or clusters of contiguous census blocks, about 11,800 in all. The sample was designed and selected to provide sufficient precision to estimate the true population for various groupings of the population that we call "post-strata," which I will describe below.
Independent Listing.--The next step in the process is to create an independent listing of housing units in A.C.E. block clusters. By independent, we mean that we do not start with or refer to the Census 2000 Master Address File, but, instead, have census staff systematically canvass the block clusters to list addresses. This listing was completed in the fall of 1999, and then addresses were checked in, keyed, and 100-percent quality controlled in our National Processing Center.
Housing Unit Matching and Followup.--The Census Bureau then matched the A.C.E. list of housing units to the Census 2000 Master Address File, first attempting to match by computer and then clerically, if necessary. Addresses that required additional information for matching were sent to a field followup operation. Clerical staff reviewed the information from the field followup and attempted to match the addresses. The purpose of the housing unit match is to create an accurate linked list of housing units in the block cluster. During the housing unit match stage, we learned that a few original A.C.E. listers may have erroneously listed the wrong area. In these cases, we conducted an independent relisting operation. The A.C.E. housing unit match and followup operation was completed in mid-April.
Telephone and Personal Interviewing.--To provide sufficient data to compare the A.C.E. to the initial census, the Census Bureau must conduct interviews to collect data from each of the housing units that were independently listed. We initiated the A.C.E. interviewing with a telephone phase using laptop computers, a technique called Computer Assisted Personal Interviewing (CAPI). Census Bureau staff began in late April telephoning households at unique addresses in the A.C.E. block clusters for which a Census 2000 questionnaire has been mailed back, processed through data capture and for which a telephone number was provided. As of May 15, we had completed over 56,000 interviews by telephone, or approximately one-sixth of our total A.C.E. workload. In addition to getting an early start on interviewing, the benefits include providing experience for our supervisors and a final test of our automated systems. The Census Bureau has extensive experience with telephone interviewing. The Census Bureau designed this phase of the A.C.E. interviewing to avoid violations of the independence requirement and tested this methodology in the Dress Rehearsal. We are confident that the gains in quality from using the telephone outweigh any potential loss in independence.
We do not begin personal visit interviewing until nearly all nonresponse followup work is completed in all the A.C.E. block clusters in an LCO. This is one of the ways we preserve independence between the A.C.E. and the census. If nonresponse followup and A.C.E. field interviewers are working in the same area simultaneously, they could affect each other's work. That is why we are waiting to complete nonresponse followup before starting A.C.E. interviews by personal visit.
Interviewers (telephone or personal visit) focus on reconstructing the Census Day household, that is, determining who lived at the address on Census Day and at the time of the A.C.E. interview and collecting as much information as possible for those who lived at the address on Census Day but have moved out by the time of the A.C.E. interview. People who moved into an address after Census Day are also interviewed to provide additional information on movers. Only short form questions are asked together with more probing questions to establish correct residence.
Interviewers (telephone or personal visit) with laptop computers will use the CAPI technique. The questions to be asked are displayed on the computer screen and responses are entered directly into the computer. The Census Bureau believes this technique improves the accuracy of the operation, because it permits a more structured interview and more probing questions. We have extensive processes for conducting quality assurance to identify data quality or falsification problems. For data quality purposes, we do not widely publicize these processes.
Most personal interviewing will be conducted in July and August, but some may begin in mid- to late June. Personal interviews are conducted only with a household member during the first 3 weeks that the case is available for interviewing. If an interview with a household member is not obtained after 3 weeks, interviewers will attempt to interview another knowledgeable person. During the latter part of the operation, the best interviewers are used to convert as many noninterview cases as possible to completed interviews, either by talking to a household member or another knowledgeable person. This nonresponse conversion has been planned to improve the completeness of data for matching.
Person Matching.--In October and November, Census Bureau staff conduct the various stages of matching persons listed in the block clusters during A.C.E. interviewing to those persons counted in the same block clusters as part of the initial census. The Census Bureau has carefully designed the A.C.E. in order to minimize matching errors. Incorrect matching determinations generally result either from incomplete, inaccurate, or conflicting data or from poor judgment in deciding whether a match occurs. Thus, we have several stages in the process of person matching.
First, we attempt to match people by computer. Those that do not match or are only "potential" matches are reviewed by clerical staff who use an automated computer match and review system to identify additional matches. After the initial stages of computer and clerical matching, selected cases--including some nonmatches and potential matches--are sent to field followup to obtain additional information that would facilitate matching. After followup, the Census Bureau conducts a final clerical matching operation.
At the conclusion of this intensive matching and followup effort, some information will still be missing. This may involve characteristics for individuals or cases where it could not be determined whether a person matched or was correctly enumerated in the initial census. The next step in the process is to statistically impute missing characteristics for individuals and for cases where it could not be determined if there was a correct enumeration.
Dual System Estimation.--Using data from the A.C.E. and the census, the Census Bureau will estimate the true population using a statistical technique called Dual System Estimation (DSE). The DSE will be conducted for each of over 400 groupings of people or post-strata. The Dual System Estimator of true population is then used to calculate a "coverage correction factor" for each post-stratum, which is the ratio of the DSE to the initial census count.
The variables that define the post-strata groupings include race, ethnicity, age, sex, owner/nonowner, return rates, whether in or out of a metropolitan area and, if in, the size of the area, and type of census enumeration method--characteristics that our research indicates are correlated with a likelihood of inclusion in the census. An example of one post-stratum is
non-Hispanic Black males, age 18-29, in nonowner units, in mailout/mailback areas of metropolitan areas with 500,000 or more people in a tract with a low return rate in the census.
Synthetic Estimation.--Coverage correction factors are then applied to the census files. For example, if the coverage correction factor for non-Hispanic Black males in the specific post-stratum described above is 1.02, that means the Census Bureau measured an undercount of 2 % for this post-stratum and for every 100 such people counted in the census in those areas two records will be added. This process is sometimes called "synthetic estimation." After this, the corrected census files can then be used to produce corrected tabulations for all uses of census data.
Mr. Chairman, I have tried to give a simple, basic description of the A.C.E. and to address each of the issues raised in your letter of invitation. The documents listed in the Appendix, all of which have been publicly available for some time, go into much more detail about a wide range of A.C.E.-related issues that we have examined. I will now answer any questions you may have.
Summary of Documents for National Academy of Science
Dual System Estimation Workshop
Title | Author | Content | |
1 | A.C.E.: Theory and Application | Howard Hogan | Gives a detailed discussion of the model for dual system estimation for census applications. Provides the theoretical foundation for the operational aspects as related to A.C.E. |
2 | A.C.E.: Overview of the Design
(S-DT-2) |
Danny Childers, Debbie Fenstermaker | A very brief summary of the basic A.C.E. operations. Shows how the various stages of sampling, matching and field work interrelate. |
3 | A.C.E.: The Design Document
(S-DT-1) |
Danny Childers | A very detailed description of the A.C.E. design. Gives a thorough accounting of procedures and rules for nearly all facets of A.C.E. Can be used as a reference manual. |
4 | A.C.E.: Targeted Extended
Search Plans (# Q-18) |
Alfredo Navarro | Provides a description of plans and conditions for extending a search for person matching outside the sample block cluster. Plans call for a certainty and sample based components. Estimator spelled out. |
5 | A.C.E.: Missing Data Procedures
(# Q-19) |
Patrick Cantwell | Describes plans for dealing with household noninterviews, missing data items including unresolved residency, match and enumeration status. |
6 | A.C.E.: Dual System Estimation
(# Q-20) |
Richard Griffin | Includes all the details for calculating DSEs. Shows how the mover Procedure-C, missing data and targeted search are reflected in the DSE. |
7 | A.C.E.: Post-Stratification for Dual
System Estimation
(# Q-21) |
Richard Griffin
Dawn Haines |
Describes the post-stratification model for 2000 . Shows and defines post-stratification variables as well as how multiple race responses are handled. |
8 | A.C.E.: Synthetic Estimation
(# Q-22) |
Dawn Haines | Describes the application of synthetic estimation to census data. Includes how overcounts and undercounts will be handled and also addresses controlled rounding. |
9 | A.C.E.: Post-Stratification
Preliminary Research Results
(# Q-23) |
Eric Schindler | Describes a number of post-stratification models under consideration. Summarizes variance, MSE, and bias. |
10 | The following document was produced
on April 19, 2000:
A.C.E.: Final Post-Stratification Plan for Dual System Estimation (#Q-24) |
Richard Griffin
Dawn Haines |
Presents the final post-stratification plan for the Accuracy and Coverage Evaluation Survey. |