Alternative Ways to Produce Intercensal Small-Area Data
There are alternative ways to improve small-area data for the nation. The panel began its deliberation by considering ways to reduce census costs by possibly reducing the content of the decennial census. We also considered the problem of the timeliness of the estimates. Our review of alternative ways to meet the needs for intercensal small-area estimates suggests that there are some sensible low-cost efforts that could begin now. Available administrative records could be exploited immediately to provide more frequent estimates for small geographic areas. The decennial census, however, is needed to provide a benchmark for these estimates. Initial work should strive to provide annual estimates for smaller geographic areas for a few key variables: population and housing counts and estimates of the population by sex, age, race, income, and poverty status. Later work might strive to provide estimates by education and school enrollment.
It would be difficult for state and local governments to replace the current data collected by the census long form if it were dropped. Available administrative records can provide only limited univariate data (see Appendix J). With geographic references for several different administrative records, limited correlated information (e.g., high-crime areas with high poverty rates) could be gleaned. But for several examples of state and local planning, it is apparent that there is no reasonable substitute—at present—for census long-form data.
Small-area data users rely greatly on data provided by the decennial census. Appendix E provides background information on some important census data uses documented by state data centers. Appendix F illustrates the variety of census data uses by businesses. There seems to be little support for dropping
data from the census and replacing it with other alternatives until data from these alternatives are available, used and tested, and preferred as a replacement to the census. A variety of census data users have informed the panel that they primarily rely and wish to rely on the census for small-area data. Small-area data users have, however, indicated the value of more timely small-area estimates from the Census Bureau, the value of a mid-decade census or a large mid-decade sample survey, and the value of being able to provide geographically referenced estimates from administrative records. These alternatives deserve further study. None of the alternatives, however, are feasible replacements for the census at this time.
We turn to an examination of several case studies on improving intercensal small-area data. The first case study discusses methods used by the Department of Defense for estimating the number of persons qualified for military service. The second study discusses the procedures used by the Bureau of Labor Statistics to provide monthly employment and unemployment estimates for states, the District of Columbia, and 2,600 labor-market areas. The third case study describes Census Bureau work to prepare annual estimates of income and poverty. The last case study examines possible ways to improve subnational estimates for the migrant and seasonal farmworker population.
QUALIFIED MILITARY POPULATION
A recent workshop convened by the Committee on National Statistics, at the request of the Department of Defense, assessed alternative techniques to develop small-area estimates of those qualified for military service (Committee on National Statistics, 1989). The ultimate aim of the workshop was to decide what method would provide the ''best" estimates of the qualified military persons available in small areas (usually counties) on an annual basis for the country. Such estimates would offer a common basis for comparative recruiting goals, at small areas, within the armed services.
For military recruitment into the enlisted ranks, the key concept for data estimates is the qualified military available (QMA), defined as the number of male high school graduates, ages 17-21 years, who mentally and physically qualify and are of suitable moral character. The QMA is the current target population by the armed services for new military recruits. Women are excluded from the current QMA because the services are oversubscribed in their recruitment of women.
Preparing small-area estimates for the QMA may be thought of as a successive narrowing of the target population. The step-down approach at estimating the QMA, at any geographic level, starts with the base population and successively restricts the nonqualifying groups. The remaining group is the QMA. The approach for county estimates for the QMA is as follows:1
Base population. The first step is to make a local-area estimate of the male population ages 17-21 years residing in the country. The estimates exclude the institutionalized population (those living in a college dormitory or military base or residing in some other institution) and are subdivided by race and ethnic group. To prepare these estimates, the decennial census data (released in the 5 percent Public Use Microdata Sample [PUMS]) are used. Census data are then aged for future years and corrected for internal migration and net international migration.
Educationally qualified. From the base population, an estimate is generated of those educationally qualified, defined as someone who has completed high school or its equivalent. Data on high school completion rates by age, sex, and race and ethnicity for counties are taken from the decennial census PUMS files.
Mentally qualified. Aptitude qualification for the armed services is based on a minimum score on the Armed Forces Qualification Test (AFQT). The various services set different minimum grades for qualification so that the small-area estimates need to provide a distribution of AFQT scores for each county. Several steps are followed to prepare county estimates for AFQT scores. Data on 11,878 youths in the National Longitudinal Survey, sponsored by the Department of Labor, are used to prepare county estimates for various AFQT categories. A large set of variables is then used, within race and ethnic groups, to predict AFQT scores. In general, the level of education and socioeconomic status of areas are the best predictors of a county's AFQT scores. The prediction equations, however, vary for race and ethnic groups.
Physically qualified. There are no adequate data on methods for estimating local differentials in expected disqualification rates for failure to meet physical standards. In practice, the physical disqualification rate for armed forces applicants is about 14 percent. Based on the National Health and Nutrition Examination Survey, conducted by the National Center for Health Statistics, the medical fitness of military service determined that 22 percent of males and 25 percent of females were disqualified (Committee on National Statistics, 1989:13). Obesity was the leading cause of disqualification. Given the present data, there appears to be no feasible approach to preparing small-area estimates for the physically qualified.
Morally qualified. The workshop did not address the topic of estimating moral qualification rates for small areas. No estimates of the morally qualified are used for current Department of Defense estimates.
The example of preparing small-area estimates for the population qualified for military service illustrates issues for data needs and methods. Regarding data needs, the estimates rely on a combination of decennial census data (with detailed, multivariate information for small areas, by race and ethnicity) and survey data. Survey data, even with limited coverage of small areas (e.g., counties), are
used in statistical modeling to provide data useful for small-area estimates. Other existing survey data sets on public enrollment and the number of high school graduates might provide sources of additional data. Data sets are inadequate, however, for modeling moral and/or physical standards for local areas.
Regarding methods, estimates of the QMA illustrate that much can be done with careful attention to available data. Although estimates can be made for small counties and areas, there is concern about the quality of the estimates. The accuracy of small-area estimates varies with the size of the area, as well as the amount of money spent on the methodology. To the extent feasible, regression sample-data methods from survey data offer a way to improve more frequent local estimates.
Monthly employment and unemployment estimates are prepared by the Bureau of Labor Statistics (BLS) for the 50 states, the District of Columbia, and over 2,600 labor-market areas (see Bureau of Labor Statistics, 1994). The process of making these monthly small-area estimates involves the use of sample surveys, administrative records, and statistical modeling. The local unemployment estimates are used to determine local-area eligibility for such federal programs as the Job Training and Partnership Act, the Economic Dislocation and Worker Adjustment Assistance Act, and the Urban Development Action Grant program.
The Current Population Survey (CPS) produces reliable monthly estimates of employment and unemployment for the country as a whole and for the 11 largest states. The CPS also provides reliable annual estimates for all states, the District of Columbia, and selected large metropolitan areas. Its sample size is not large enough to produce reliable monthly estimates for most states (the 39 smaller states), the District of Columbia, and smaller geographic areas.
For the 39 states and the District of Columbia that do not have monthly CPS estimates, statistical techniques (regression analysis) are used. Regression models are based on historical and current relationships within each state's economy, as reflected by data from the CPS, the BLS survey of employers, and states' unemployment insurance system. The regression models for states use the same explanatory variables, but the coefficients are unique for each state. The employment models also include trend and seasonal components to account for movements in the CPS not captured by the BLS's survey of employers. Similarly, the unemployment models use trend and seasonal components to capture the state's historical relationship between the unemployment insurance claims and unemployment measured by the CPS.
Once each year, the monthly estimates from the regression models are calibrated, or adjusted, to the annual average CPS state estimates. This calibration sets the annual average of the regression models to the CPS annual average, while preserving the distinctive monthly seasonal pattern of the regression model estimates.
CPS monthly estimates are reliable for the largest two metropolitan areas—New York City and the Los Angeles-Long Beach metropolitan area. Estimates for other metropolitan areas and for the 2,600 labor-market areas are made with an indirect statistical procedure. First, preliminary estimates are made for employment and unemployment in the local market areas. The estimates are then adjusted to the state totals. And, finally, the estimates are corrected to bench-mark data.
Preliminary civilian employment estimates are based on the Current Employment Statistics (CES) survey that provides place-of-work data. Census data are used to derive factors for adjusting place-of-work to place-of-residence data for several categories of employment. These factors are applied to CES data to provide adjusted employment estimates by place of residence. Estimates for employment not included in the CES—agricultural workers, nonagricultural self-employed and unpaid family workers, and private household workers—are added to the CES estimates.
Preliminary unemployment estimates are based on the aggregate of estimates for three groups: (1) persons who were previously employed in industries covered by state unemployment insurance laws, (2) those not previously employed in industries not covered by those laws, and (3) those who were entering the labor force for the first time or reentering after having been out of the labor force for a period of time.
After preliminary estimates of employment and unemployment are made for metropolitan areas and for the 2,600 local-market areas, the sums of the estimates are totaled for each state. The estimates are proportionately adjusted so that the state sums add to the independently estimated state totals for employment and unemployment.
At the end of each year, substate estimates are revised for CES-based employment figures, corrections to unemployment claims, and updated historical relationships. The revised estimates are readjusted to sum to the state bench-mark data for estimates of employment and unemployment.
INCOME AND POVERTY2
The Bureau of the Census prepares annual estimates by states of median family income of four-person families to meet the needs as an eligibility criterion of the Low Income Home Energy Assistance Program—a block grant program administered within the Department of Health and Human Services. The annual four-person family income estimates are also important to data users as the only source of intercensal state-specific family income estimates produced by the Census Bureau.
The income estimates program utilizes census, sample surveys, and administrative records data to update small-area decennial information. Estimates of four-person median family incomes by state utilize the most recent data available from the following three data sources: (1) March Current Population Survey (CPS) from the Bureau of the Census, (2) decennial census of population from the Bureau of the Census, and (3) per capita personal income estimates from the Bureau of Economic Analysis (BEA).
The CPS is a monthly labor force survey of about 60,000 households across the country. Each March, the monthly survey is supplemented with questions that ask household respondents about their money income received during the previous year. Survey results are published annually by the Census Bureau.3 The CPS provides estimates of median incomes for four-person families annually at the national level. Though CPS data may also be tabulated at the state level, the sampling variability of state estimates from the CPS is large enough to make their direct use inadvisable for program purposes.
The decennial census provides income values for calendars years 1979 and 1989 that utilize the same money income concept as the annual CPS estimates, with negligible sampling error at the state level.
The Bureau of Economic Analysis (BEA) produces annual estimates of personal per capita income based on this income concept for states and other geographic areas. The estimates provide an overall indication of relative incomes among states and of change over time for particular states, even though personal per capita income is only indirectly related to median incomes of four-person families. The BEA estimates, though conceptually different from census income data, are essentially free from sampling error.
The main difference between the BEA and CPS income concepts is that the BEA personal income relates to income from all sources, while CPS income relates only to money income and adjustments are made accordingly. The BEA series is developed from a variety of government statistics, the most important
being the federal tax records of the Department of Treasury, the insurance files of the Social Security Administration, and state unemployment records collected by the U.S. Department of Labor.
Prior to 1984, the estimation procedure consisted of five main steps:
State median family income estimates for four-person families and respective standard errors were directly calculated form CPS data.
Another set of state estimates of four-person family income was developed by carrying forward census median estimates using the percent change in state BEA per capita personal income.
A regression estimate was then developed to predict the CPS estimates from the adjusted census medians from the preceding step.
A composite estimate was formed as a weighted average of the regression and CPS sample estimates.
Finally, the composite estimate for each state was constrained to a range that was equal to the CPS estimate, plus or minus one standard error.
After the 1980 census, a comparison of estimates from the census with comparable estimates from the estimation model afforded an opportunity to assess the merits of the model and to modify it. The major impetus for the revisions was the year-to-year variability of some of the state estimates. As before, the revised methodology employed sample estimates of four-person median family income from the CPS. Unlike the previous methodology, however, a second sample estimate was considered for each state. This second variable is a weighted average of estimated state median incomes for three-and five-person families. By introducing this second variable into the estimation model, the amount of sample information reflected in the estimate of four-person medians for each state was roughly doubled.
Changes in the 1984-1989 methodology are summarized below:
As in the earlier methodology, a regression equation was fitted to the CPS four-person estimated medians. In the new equation however, two variables were used: (a) a proportional adjustment of the census median for changes in BEA per capita income and (b) a reduction in the adjustment in case of a "regression toward the mean" which might arise if BEA per capita income did not perfectly measure the proportional changes in the income distribution.
It was found that the inclusion of the unadjusted median improved the predictive power of the regression.
In the revised model, a second regression equation was simultaneously fitted to the weighted average of three-and five-person medians.
The composite estimate in the revised model utilized both regression equations along with the CPS sample estimates.
Since the revised methodology essentially doubled the use of sample data in forming the composite estimates, the standard error constraint of the original equation model was eliminated. The comparison to the census results suggested that the estimates were more accurate, on the average, without the constraint.
In updating the state median income estimates annually, the Bureau of the Census uses the most recent data available from its March CPS and BEA's per capita income estimates.4 However, the availability of such data results in a 3 year difference between the time frame of the March CPS and BEA income data and the fiscal year in which the estimates are in effect. For example, the state median income estimates for fiscal year 1992 were published in March 1991. The estimates are based, in part, on 1989 family income from the March 1990 CPS, which became available in August 1990 and BEA's 1989 state per capita income estimates which became available in September 1990.
Therefore, the state median income estimates do not represent projected data for a fiscal year. Instead, the estimates represent the most recent estimates available for use in a fiscal year.
Subnational estimates of poverty are not now prepared by the Census Bureau (or any other agency of the U.S. government). At the moment, the development of such estimates is still in the research and experimentation stages. More recently, the Census Bureau has been exploring ways to develop such estimates in connection with potential congressional legislation mandating the preparation of current estimates of income and poverty for small areas, including most government jurisdictions and school districts (Poverty Data Improvement Act of 1993).
Although precise methodology is still in the developmental stages, lacking direct estimates from appropriate administrative records, the broad outline of the methodology would require integration and statistical estimation processes of the following data:
the basic 1990 decennial census micro-data files containing "full" detail on geographic identifiers and complete income data for all sample cases. This file is not available for public use.
an extract of the information contained on the IRS Individual Master File
(IMF) containing more than 100 million federal individual income tax returns that the Census Bureau receives annually for research and population estimation purposes. The Bureau of the Census adds geographic codes based on taxfilers' addresses that are internally consistent within the Internal Revenue Service's system of addressing. These codes are used to measure migration from place to place.
files created by linking the IMF tax file extracts with both the March CPS and the Survey of Income and Program Participation (SIPP). The March CPS collects information about sources and amounts of income. The SIPP has much more extensive information about federal taxes. In the proposed estimates work, both of these files will be linked, on a case-by-case basis, with the tax return of the survey respondent. This linkage, made possible by collection of social security numbers in the survey, brings together the information from tax returns with the socioeconomic information collected in the surveys. This link provides a "bridge" between the Census Bureau data and the tax return data that is required in the small area estimation modeling process. Modeling and statistical estimation techniques are the processes by which changes in income and poverty at the national level are used to update the small-area data for the last decennial census. In addition, considering the population universe being estimated, additional administrative records such as AFDC or food stamp files would be extremely important and valuable for use in the estimation process.
Thus, the prospective methodology for such small-area estimates illustrates a number of important factors to be borne in mind in contemplating the production of small-area intercensal estimates: (1) a most comprehensive application, in which a variety of different data sets, including census based data, survey based data, and administrative record data, and different methodologies are brought together and utilized in developing the estimates; (2) the importance of having a "benchmark" source—in this case, the decennial census—both as a base from which to construct the intercensal estimate, and as an independent measure in a future period to be used to evaluate the methodology; (3) the advantages of and benefits to be derived from the ability to match records through the use of a common identifier, whether, as in this case, the Social Security number, an address, or some other common reference identifier; and (4) the clear advantage of using administrative files from a single, standardized source that covers the entire nation. And, finally, it demonstrates the advantages and strengths of drawing together, under one sponsorship, the diversity of disparate sources and, in so doing, allows a focusing of resources and a combining of data sources to produce both an improved methodology and a more reliable, more accurate set of estimates.
Work to date by the Bureau of the Census on income and poverty estimates has relied on tax files from the Internal Revenue Service. Proposed work will include using miscellaneous tax documents to provide a more complete picture
of assorted income and to improve the construction of income for households and families. The panel urges the Bureau of the Census to also initiate work with the Aid to Families with Dependent Children (AFDC), Food Stamps, and other files that are appropriate to the universe of low-income persons. These other files, in addition to tax and income records, would greatly improve estimates of the number and location of low-income and poverty groups.
MIGRANT AND SEASONAL FARMWORKERS
The federal government currently spends over $550 million annually on farmworker programs. Among the major federal programs for farmworkers are migrant education, food stamps, Head Start, the Job Training and Partnership Act, migrant health, and housing assistance. There are few reliable data available for program planning or for purposes of allocating resources to state or local jurisdictions. At the moment, each of the federal programs either relies on decennial census data or carries out its own limited data gathering. Although these separate data collection activities add up to significant expenditures, unconnected research does not contribute to a general national database on the farmworker population.
Data collection on farmworkers begins with four facts:
Farmworkers are hard to find. They often do not have mail addresses and are widely scattered in selected agricultural areas. They often reside in camps, temporary housing, or buildings converted to residential space. Special procedures are needed to canvass an area and to locate the residences of farmworkers.
Farmworkers are seasonal. They do not work as farmworkers throughout the year. To address the seasonality issue, data collection must take place throughout the year, with data collected in different times in different areas, depending on the farming season.
Farmworkers are migratory. The higher geographic mobility of farmworkers is a special challenge for data collection. An area with a high number of farmworkers during farm season may have no workers at other times of the year. The issue of geographic mobility can be addressed by collecting work and residence histories, and then calculating the distribution of farmworkers in the country at given times of the year. This approach also provides estimates of where farmworkers reside when they are not working in agriculture.
Farmworkers are often foreign-born. A substantial proportion of farmworkers are not English-speaking or, if competent in English, do not read or write English. The largest proportion of foreign-born farmworkers are Mexican-origin and Spanish-speaking. Other foreign-born farmworkers are from Central America (and may speak native Indian languages) and Haiti.
Limitations of Census Data
Farmworkers have been an extremely difficult group to enumerate in past decennial censuses for all the reasons discussed above. Estimates prepared by groups working with farmworkers suggest that the 1990 census counted and identified only about 25-35 percent of the persons who are farmworkers at some time during the year (Kissam, 1991; Gabbard, et al., 1993).
Underidentification of farmworkers in the census stems from two different causes: undercount and nonidentification. Undercount problems stem from the difficulties of locating and enumerating farmworkers at the time of the census. Because of the nontraditional housing of farmworkers, geographic mobility, and limited English-language ability, farmworkers experience a high rate of total household omissions from the census. Also, some farmworkers and their families reside outside of the United States (primarily in Mexico) at the time of census enumeration.
Nonidentification of farmworkers occurs when there is a failure to identify correctly that someone is a farmworker. The census long-form questionnaire asks questions about occupation, but the questions pertain to the occupation at the time of the census or, for unemployed persons, for the last job. Many "farmworkers" are unemployed or are employed temporarily in nonagricultural work when the census is taken (on April 1 for the 1990 census). Analysis of the National Agricultural Worker Survey data suggests that more than 25 percent of farmworkers were working temporarily in some other occupation at the time of the 1990 census (Gabbard, et al., 1993).
To give one example, estimates from three alternative data sources place the number of farmworkers in California in the range of 560,000 to 720,000 for the period around 1990; the mid-range estimate by experts of the number of farmworkers—in terms of an occupational definition of "farmworker"—was about 640,000 (Gabbard, et al., 1993). The 1982 Census of Agriculture reported 980,000 farmworkers in California. As reported in the 1990 census, however, there was a count of 182,000 farmworkers, or about 25-32 percent of the alternative estimates. Using the mid-range estimate of 640,000, the census count underidentified the number of farmworkers by about 72 percent.
Persons reported as farmworkers in the decennial census also suffer from sample bias. The census collects data on persons who reported that they were farmworkers at the time of the census. These farmworkers tend to be those who are employed on a regular basis throughout the year, who work permanently in agriculture. Such permanent workers tend to be English-speaking and socially and economically better off. The portrait of farmworkers, as pictured in the census, does not give an accurate portrait of people or their families who work as farmworkers.
Although more research on the problems of differential undercount and nonidentification of farmworkers in the census would be helpful, it is difficult to
see how the decennial census could be improved to provide an accurate count or description of farmworkers. To make an improved count, three changes would be needed: (1) to canvass farmworkers through the crop seasons, (2) to ask retrospective questions on job and residence history, and (3) to expand dramatically the intensity of effort to locate farmworkers and interview them in their native languages. It would not be cost effective to design the census, with its national focus and attempted coverage of the entire population, for this purpose, as compared to other alternatives such as special sample surveys. Moreover, a census would provide a count at one point in time and would lack the up-to-date estimates that might be obtained from improved intercensal estimates.
If alternative estimates of the farmworker population could be made, for small areas and for intercensal periods, federal and state farmworker programs could rely on alternative data for funding allocations and for assigning program resources. Farmworker groups and federal and state agencies working with farmworkers might explore the possibility of using sample surveys and administrative records to improve small-area data on farmworkers and their dependents. One possible way to obtain intercensal information on farmworkers is outlined below.
A National Sample Survey
The prior discussion leads to two propositions. First, the decennial census cannot provide reliable counts of the farmworker population. The census takes place at one point in time (when many farmworkers are not working in agriculture). The census count could not be redesigned to include a job and residence history (questions that would be needed in order to estimate the distribution and size of the farmworker population). High-quality data on farmworkers require more expensive canvassing and personal interviewing (in several languages) than can ordinarily be achieved in the census. Second, the number and characteristics of the farmworker population change substantially over time. More frequent intercensal estimates of the farmworker population are needed in order to provide well-targeted program resources. The census cannot provide these estimates.
To remedy this situation and to decrease the inappropriate reliance of federal allocations for farmworkers on the census, one proposal would be to initiate a national sample survey of farmworkers, in conjunction with statistical modeling for small-area estimates. An integrated program of surveys and statistical modeling would:
estimate the program-eligible seasonal and migrant farmworker population for each state on an annual basis to improve resource allocation;
describe the characteristics of the farmworker population and their dependents for each state to improve program planning and evaluation; and
estimate the farmworker population for small areas by statistical modeling on a periodic basis to improve the targeting of programs and resource allocation.
Mines (1993) has outlined an integrated approach to annual intercensal estimates for the farmworker population. The approach would involve adjusting a sample survey on farmworkers to two benchmarks: (1) crops and livestock workers and (2) fishing, food processing, and forest products workers. Farm and livestock workers are already covered in the Quarterly Agricultural Labor Survey, sponsored by the Department of Agriculture. The Quarterly Agricultural Labor Survey provides quarterly employment estimates for 17 states with the vast majority of the nation's food production and processing.5 Only a small proportion of workers in the fishing, food processing, and forest products industry are farmworkers; these industries provide data through the quarterly unemployment insurance reports for estimating employment levels. As a first observation, required benchmark data exist for adjusting survey results to national estimates of farmworkers.
A national sample survey of farmworkers and their characteristics, including their industry of employment, could be combined with employment benchmark data to produce accurate estimates of the total population of farmworkers in each region and state and an estimate of dependents in each state on an annual basis.
The design of a national sample survey of farmworkers would build on the experience of the surveys of migrant and seasonal farmworkers over the past decade. There are an estimated 3.25 million farmworkers in the nation, including about 2.25 million workers in crop agriculture, another 500,000 livestock workers, and about 500,000 in nonfarm sectors. Workers would be sampled through employers on a periodic basis throughout the year to avoid seasonality issues. The interviews would be conducted away from the workplace in the native languages of the farmworkers.
A regionally stratified areal probability sample would involve samples in about 300 counties, with approximately 725 respondents per state per year for the 17 states with the bulk of farmworkers. This design would need approximately 20,000 interviews per year. Estimates for the remaining 33 states could be made by grouping states with similar weather and crops, using smaller sample sizes in the aggregated areas.
Some federal programs, such as the migrant education program, require information on ex-farmworkers. To meet these program requirements in an cost-effective manner, follow-up interviews could be conducted on a subsample of initial interviews: 8,000 follow-up interviews would provide annual estimates on ex-farmworkers for 11 agricultural regions of the country.
Mines (1993) estimated that there would be an annual cost of $6 million for a survey program such as the one outlined above. The cost estimates include the
costs of sampling lists, conducting the annual sample survey, improving the Quarterly Agricultural Labor Survey and extending it to Puerto Rico, and adjusting the survey estimates for industry employment benchmarks.
Not included in the cost estimates is the cost of a third stage of the program to provide small-area estimates. Annual state estimates of the number and characteristics of farmworkers and their dependents could rely on statistical modeling to estimate small-area data. Although no specific proposals are available on ways to make these estimates, data from the Census of Agriculture (including information on crops and employment levels) may offer a suitable context for estimating, at least once every 5 years, the number and characteristics of farmworkers and their dependents for counties and rural program areas. Annual estimates for similar small areas could be made by assuming annual changes for states and then prorating the changes by assuming the geographic distribution estimated with Census of Agriculture data. Such small-area estimates would have two great attractions compared to decennial census data: they would be more accurate about numbers and characteristics and they could be revised more frequently.
Bureau of Labor Statistics 1994 Local Area Unemployment Statistics. Washington, D.C.: U.S. Department of Labor.
Committee on National Statistics 1989 Small-Area Estimates for Military Personnel Planning: Report of a Workshop. Washington, D.C.: National Academy Press.
Gabbard, S., E. Kissam, and P.L. Martin 1993 The impact of migrant travel patterns on the undercount of Hispanic farm workers. Pp. 207-245 in Proceedings of the 1993 Research Conference on Undercounted Ethnic Populations. Bureau of the Census. Washington, D.C.: U.S. Department of Commerce.
Kissam, E. 1991 Out in the Cold: Causes and Consequences of Missing Farmworkers in the 1990 Census. July. La Cooperative Campesina de California, Sacramento, Calif.
Mines, R. 1993 Estimation and Description of Farmworkers by State: An Integrated Approach . U.S. Department of Labor, Washington, D.C.
Navy Personnel Research and Development Center 1987 Estimating the Youth Population Qualified for Military Service. E.W. Curtis et al. August. Personnel Research and Development Center, San Diego, Calif.