The Bicentennial Census New Directions for Methodology in 1990: 30th Anniversary Edition (2015) / Chapter Skim
Currently Skimming:
4 Evaluating the Decennial Census: Past Experience
4 Evaluating the Decennial Census: Past Experience
Pages 105-152
The Chapter Skim interface presents what we've algorithmically identified as the most significant single chunk of text within every page in the chapter.
Select key terms on the right to highlight them within pages of the chapter.
From page 105... ...
Omissions of persons can occur, among other reasons, because occupied housing units -- and hence all of their residents -- are inadvertently overlooked or are believed to be non esidential or vacant at the time of the census, because individual r members of a household are not reported by the household, because persons with more than one usual place of residence, such as college students away from home or persons with a vacation home, are not counted at either address, and because some persons do not have usual places of residence as the term is commonly used. Erroneous enumerations also can occur for many reasons, for example, because persons who moved between Census Day and field follow-up are enumerated at both locations, because persons with more than one usual residence are enumerated more than once, because "out of scope" persons, such as those who were born or migrated 105
|
From page 106... ...
erroneous inclusions. The main goals of coverage evaluation programs are to measure the net coverage error for the total population of the nation and, when possible, for important demographic subgroups and subnational geographic areas.
|
From page 107... ...
An important distinction among micro-level methods is the source of the evaluation data: administrative records or sur vey data. The main macro-level method is demographic analysis.
|
From page 108... ...
The Census Bureau has conducted coverage studies of specific population groups based on administrative records matches, for example, using Medicare data to study coverage of persons 65 and over. One method, sometimes called the composite list, has been detailed in Ericksen and Kadane (1983; they refer to it as the "megalist" method)
|
From page 109... ...
The Census Bureau experimented successfully with micro-level coverage evaluation programs using post-enumeration survey techniques for the 1945 Census of Agriculture, the 1947 Census of Manufactures, and the 1948 Census of Business. These efforts led to the decision to evaluate coverage in the 1950 Census of Population and Housing using a large postenumeration survey.
|
From page 110... ...
by urban/rural residence and for population subgroups classified by age, race, sex, and various socio economic characteristics such as income and occupation. Evidence from several other sources, including a quality check conducted as part of the PES, demographic analysis, and independent record checks for selected population groups, indicated that the PES net coverage error was too low -- probably by as much as 2 percentage points (see dis
|
From page 111... ...
. The program included a nother post-enumeration survey and several kinds of record checks, including a reverse record check, in addition to demographic analysis.
|
From page 112... ...
. The 1960 Census Reverse Record Check The Census Bureau also carried out a reverse record check study to estimate net national undercount in 1960, similar to the methodology used in Canada.
|
From page 113... ...
For persons not found in the census or when there was doubt as to the person's enumeration status, the Census Bureau made further efforts to determine whether the person was counted. Despite their best efforts, a definite match status (counted in the census or missed)
|
From page 114... ...
Among the four samples, the sample drawn from the 1950 PES had the highest proportion of cases for which a definite match status could not be assigned -- over 24 percent. Using different assumptions about the rate at which the census counted persons for whom a definite enumeration status could not be obtained, the Census Bureau estimated gross omission rates from the reverse record check of between 2.6 and 4.7 percent of the total population.
|
From page 115... ...
Coverage Evaluation in the 1970 Census Because of the problems with the post-enumeration survey methodology used to estimate net undercoverage in the 1950 and 1960 censuses, the Census Bureau made no plans to carry out a comparable program for the 1970 census but placed chief reliance on the method of demographic analysis (see discussion in a later section)
|
From page 116... ...
COVERAGE EVALUATION PRIOR TO 1980: MACRO-LEVEL METHODS Researchers inside and outside the Census Bureau have used aggregate methods to assess the completeness of census coverage since the beginning of coverage evaluation efforts in the United States. The principal macrolevel method is termed demographic analysis, whereby independent esti
|
From page 117... ...
For example, data from Medicare records are currently used to construct independent estimates of the population age 65 and older, rather than using the demographic method outlined above. Demographic analysis cannot be performed for population groups defined according to other characteristics, such as income or education, because of the absence of appropriately classified registration information.
|
From page 118... ...
Coale's net undercount estimate of 3.5 percent is 2.5 times the estimate from the PES of 1.4 percent. The Census Bureau developed a "minimum reasonable estimate" of 2.4 percent net undercount based on PES results for persons age 40 and older, birth registration data for persons under 15, and examination of sex ratios for persons ages 15-39 (Bureau of the Census, 1960:5-6)
|
From page 119... ...
In about 82 percent of cases in which an infant was missed, the parents were also missed. Demographic Analysis in 1960 Census Bureau staff and university scholars carried out several studies in the early 1960s to evaluate completeness of coverage in the 1960 census.
|
From page 120... ...
The preferred demographic analysis estimate is 1.4 times the estimate from the 1960 PES and falls within the range of 2.5 to 3.1 percent estimated in the 1960 reverse record check. Demographic Analysis in 1970 The Census Bureau relied on demographic analysis as the principal method of evaluating coverage in the 1970 census.
|
From page 121... ...
As described more fully in Chapter 5, demographic analysis estimates of net undercount in every census since 1950 indicate better coverage for women, on average, than for men, and for whites than for persons of other races. During the 1970s, Census Bureau researchers endeavored to develop coverage estimates for states (Siegel et al., 1977)
|
From page 122... ...
In the application of dual-system estimation to the 1980 census, the two assumptions of equal capture probability at each stage and the independence of capture probabilities, which are generally believed not to hold, were modified as follows. The Census Bureau stratified the population into subpopulations and used dual-system estimation separately in each stratum.
|
From page 123... ...
n1 N m Parallel to the wildlife example, in the 1980 census dual-system estimation model, the census served as the first method of capture and the Census Bureau used the Current Population Survey sample of households as the post-enumeration survey, the second capture mechanism. Unlike the wildlife situation, it is not always easy to ascertain whether an individual was included in the census.
|
From page 124... ...
records placed into the wrong census enumeration district; (2) records resulting from erroneous enumerations, e.g., individuals born after Census Day, and cases fabricated by census enumerators (called curbstoned cases)
|
From page 125... ...
In the remainder of this section we discuss the incompleteness in the P and E samples in the 1980 PEP and how the Census Bureau attempted to compensate for the lack of completeness. For the P sample there were three major sources of missing data:2 (1)
|
From page 126... ...
In order to take account of these missing data, the Census Bureau used various forms of imputation, as well as other approaches described below, to arrive at 12 different sets of estimates of undercoverage for states, m ajor cities, and remainders of states. The formation of these 12 estimates resulted from various choices concerning which CPS month to use, the treatment of information considered of questionable quality, and the treat
|
From page 127... ...
to use as well as other choices not mentioned here, the Census Bureau developed 27 estimates of undercoverage for states and major cities. Later, this number was reduced to 12 estimates that the Census Bureau felt all represented reasonable alternatives.
|
From page 128... ...
would discard all but the 2-8, 2-9, 3-8, and 3-9 estimates as either based on August data, which had a higher rate of cases with unresolved match status, or as making use of extreme assumptions in the adjustments for missing data. However, even
|
From page 129... ...
1980 DEMOGRAPHIC ANALYSIS In the years prior to 1980, demographic analysis had provided what were considered to be the most trustworthy estimates of undercoverage for certain demographic groups at the national level. However, demographic analysis for 1980 is generally considered to be significantly less accurate than for any of the previous three censuses (even though the reliability of some components of the estimates probably improved)
|
From page 130... ...
Work has continued on improving coverage estimates based on demo graphic analysis. To understand the nature of these improvements, we discuss in turn each of the major data components of demographic analysis separately for persons under and over age 65.
|
From page 131... ...
applied a modified form of demographic analysis to the foreign-born population enumerated by the 1980 census to estimate the number of undocumented aliens included in the census, a necessary preliminary step to estimating undercoverage of the legally resident population. Upon removing the undocumented aliens included in the overall count, this method estimated that the census had a 0.5 percent national undercount of the legally resident population, with a 5.3 percent under
|
From page 132... ...
Quite possibly, a large number of these missed individuals are included on administrative lists. The use of administrative records and multiple list methods as a major component of a decennial census coverage evaluation program on a national level has never been attempted in the United States.
|
From page 133... ...
Composite list and multilist methods each make certain assumptions. Failure of these assumptions would cast serious doubt on the reasonableness of the resulting estimates of omission rates.
|
From page 134... ...
However, current trends in our society toward increasing computerization, automation, editing, quality control, etc., are likely to increase the possibility of meeting many of the requirements. In addition, progress at the Census Bureau in areas such as automated list matching should benefit the use of administrative records for coverage evaluation.
|
From page 135... ...
In this application of composite list methods, the court directed that the Census Bureau use the following 10 lists: (1) Consolidated Edison electricity bill payers; (2)
|
From page 136... ...
Ericksen and Kadane believe a more reasonable estimate of the rate of omissions to be at least 10 percent due to the treatment of imputations and the likelihood of the undetermined cases to represent residents of New York City. Many of the difficulties encountered by this test of an administrative records-based coverage evaluation program were specific to this application and are not due to the general methodology.
|
From page 137... ...
Micro-level coverage evaluation programs, however, have not been completed to date within a shorter time span than about a year and a half after the census. Preliminary demographic estimates for 1980 were released in mid-1981 and are still being revised on the basis of new data such as updated Medicare records.
|
From page 138... ...
Chapter 8 makes several recommendations in this direction. The preceding sections of Chapter 4 have provided an overview of the development of coverage evaluation methods during the past 40 years.
|
From page 139... ...
However, the problems reverse record checks have in counting undocumented aliens are probably as serious as those of post-enumeration surveys. Administrative list methods may have still less of a problem of undercoverage (although other problems exist, including the fact that gaps in coverage cannot be quantified in terms of age-sex-race groups or other census characteristics)
|
From page 140... ...
(In Chapter 8 we mention that intensive tracing may increase the sensitization of the population.) Recent use of the reverse record check in Canada has been accomplished with a fairly consistent tracing failure rate of 5 percent.
|
From page 141... ...
Reverse record checks do not use dual-system estimation to provide estimates of undercoverage, since the underlying assumption is that the frame used for the reverse record check sample is essentially complete. Balancing of Overcount and Undercount The difficulty of balancing over- and undercount, faced by post- enumeration programs that do not sample compact area clusters such as city blocks, has as yet an unknown impact on coverage estimates.
|
From page 142... ...
Relative to other components, legal migration is probably subject to greater error than birth and death records, but to less error than undocumented immigration. An interesting by-product of the reverse record check is the possibility of a direct estimate of emigration.
|
From page 143... ...
Coverage evaluation programs are a ttempts to measure errors of the census, and their results are used by the Census Bureau to understand the errors so that action can be taken to make improvements. Chapter 5 presents information gathered from the various methods used to evaluate the coverage of the decennial censuses on differential rates of undercount and overcount among groups in the population.
|
From page 144... ...
Although other applications have been suggested (see Alvey and Scheuren, 1982) , this appendix deals with using administrative records, or multiple lists, as part of an estimation process to help determine the "true" population count.
|
From page 145... ...
EVALUATING THE DECENNNIAL CENSUS 145 the two lists by matching, how can the remaining members of the population be estimated? These quantities lead to four elements of a two-by-two contingency table, given in Table 4.4.
|
From page 146... ...
(By representative we mean that no identifiable population, e.g., demographic group or persons residing in a specific region of the country, is missed relatively more often by the composite list than any other identifiable population.) One can effectively merge a sample of the noncensus lists by sequentially matching them and then matching the merged list with the census.
|
From page 147... ...
If there are j noncensus lists, this requires j(j + 1) /2 total matches.
|
From page 148... ...
The inability of the 1980 PEP to match more than 92 percent of individuals between the 1980 census and the April 1980 CPS despite almost exactly equivalent reference dates and extensive follow-up investigations is illustrative of the problems involved. These problems are likely to be even more severe with the use of administrative lists maintained for other purposes with different record-keeping priorities.
|
From page 149... ...
Thus, in 1980, the Census Bureau was able to assert to uneasy respondents that its files could not be used for large-scale government record linkage. In 1990, the current plans of the Census Bureau are to capture in machine-readable form both name and address for input into a computer matching algorithm.
|
From page 150... ...
Possible matches were searched for only in the same enumeration district as that of the address of the Current Population Survey interviewee.
|
From page 151... ...
A MATHEMATICAL MODEL FOR RECORD LINKAGE Fellegi and Sunter (1969) developed a mathematical model for matching, or record linkage, which the Census Bureau is using in its software system for automated matching.
|
From page 152... ...
. Iterative procedures are also conceivable in which initial match status is determined, which then provides more refined estimates of m(g)
|
Key Terms
This material may be derived from roughly machine-read images, and so is provided only to facilitate research.
More
information on Chapter Skim is available.