Introduction and Overview
In the early 1990s the U.S. Census Bureau began work to produce regularly updated estimates of key income and poverty measures for subnational areas in a program called SAIPE–Small Area Income and Poverty Estimates. The estimates are produced by using sophisticated statistical modeling techniques with data from multiple sources, including the March Current Population Survey, the 1990 decennial census, and administrative records.
Legislation passed in 1994 called for the use of updated Census Bureau estimates of poor school-age children for counties and school districts to allocate more than $7 billion of funds each year under Title I of the Elementary and Secondary Education Act. The same legislation also authorized the U.S. Department of Education to commission a review of the Census Bureau's estimates by a panel of the National Research Council's Committee on National Statistics. The statute required that the department use the updated Census Bureau estimates unless the Secretaries of Commerce and Education determined that some or all of the estimates are “inappropriate or unreliable” on the basis of the panel's study (Improving America's Schools Act of 1994 [P.L. 103-382] and 1996 continuing resolution).
The Panel on Estimates of Poverty for Small Geographic Areas was set up to carry out the authorized study. The panel was charged with a broad review of the Census Bureau's SAIPE model-based estimates for small geographic areas and their utility for fund allocations and other purposes.
The panel began its work in June 1996 and produced three interim reports. Each report evaluated a specific set of estimates of poor school-age children from the Census Bureau and made recommendations about their use for Title I alloca-
tions: 1993 county estimates (National Research Council, 1997); revised 1993 county estimates (National Research Council, 1998); and 1995 county and school district estimates (National Research Council, 1999). This report combines in a single reference document the information in the panel's three interim reports about these estimates: what they are, how they were produced, and their quality. The panel's final report (National Research Council, 2000) provides an agenda for research and development for the Census Bureau's SAIPE Program in the next decade. It covers modifications to the Bureau's current models; possible uses of new sources of data from surveys and administrative records for improving the models, and issues in using the model-based estimates for such program purposes as fund allocation.
The panel hopes that this reference report will be useful for people who require small-area estimates of poor school-age children and for people with an interest in the methods and applications of small-area poverty estimates more broadly. Since the development of model-based or model-dependent estimates that combine data from multiple sources is a complex task,1 such estimates should always be accompanied by complete documentation of how they were developed and a full evaluation of their quality. The Census Bureau has carried out evaluations and prepared documentation, and this report also contributes to that goal.
Title I supports compensatory education programs to meet the needs of educationally disadvantaged children (see Moskowitz et al., 1993). From the enactment of the program in 1965 through the 1998-1999 school year, the role of the Department of Education has been to allocate funds to the nation's more than 3,000 counties (including Puerto Rico as a county equivalent), and the states have then distributed the county allocations to school districts. For the 1999-2000 school year, in response to the Improving America's Schools Act of 1994, the department for the first time made allocations directly to almost 15,000 school districts (formally known as local educational agencies, LEAs).2
The Title I allocations use estimates of formula-eligible children: predominantly poor school-age children, who are defined by the Census Bureau to be
By “model-dependent” we mean that the accuracy of the estimates depends on the validity of the assumptions of the estimation model.
The intent of the Title I legislation, when it was originally enacted in 1965, was that the Department of Education would allocate funds directly to school districts; however, lack of data with which to develop school district estimates led to the two-stage allocation system that was used through the 1998-1999 school year.
related children aged 5-17 in families with incomes below the poverty level.3 (Related children include family members under age 18 in a household, except married sons, daughters, or spouse of the householder and foster children.) Historically, the allocations made by the Department of Education to counties used the estimates of poor school-age children from the most recent decennial census for which data were available. The estimates from one census were used for a decade or more until estimates from the next census became available. Since the proportions and numbers of children in poverty can change significantly over time, the 1994 legislation called for the use of updated estimates of poor school-age children for Title I allocations. The Census Bureau was to provide updated estimates for counties in 1996, for use in the Title I allocations to counties in the 1997-1998 and 1998-1999 school years, and then to provide estimates for school districts in 1998 and every 2 years thereafter, for use in direct Title I allocations to school districts in the 1999-2000 and later school years. Having the most upto-date estimates possible is important so that resources can be directed towards areas that are most in need.
At present, Title I funds are provided for two different types of allocations– basic grants and concentration grants. Under the two-stage allocation process used through the 1998-1999 school year, basic grants were provided to all counties and suballocated to school districts that had at least 10 formula-eligible children and whose percentage of formula-eligible children exceeded 2 percent of the district' s total school-age children. Concentration grants were provided to counties that had high numbers (more than 6,500) or high proportions (more than 15%) of formula-eligible children and suballocated to eligible school districts in those counties. However, with the direct allocation process first used for the 1999-2000 school year, the provisions for county eligibility and grant amounts no longer apply; concentration grants are now provided directly on the basis of school district eligibility.
Under the direct allocation process, the department determines the initial allocation amounts for all school districts. However, a provision in the 1994 legislation permits states to reallocate these amounts for school districts with less than 20,000 population by using another data source that the department approves. This provision was included because of concerns about the likely quality
The poverty status of individuals is determined by comparing the before-tax money income of their families to the appropriate poverty threshold. The poverty thresholds vary by family size and are updated by the change in the Consumer Price Index each year. See National Research Council (1995a) for an evaluation of the current official poverty measure and a proposed alternative measure; the issue of how poverty should be defined is not considered in this volume. The Title I allocations also take account of the average per-pupil expenditures in each state, as well as the allocations made in the previous year (through a “hold-harmless” provision).
of estimates of poor school-age children for small school districts. About four-fifths of school districts contain fewer than 20,000 people, although these districts contain only about 27 percent of all school-age children in the United States. For the 1999-2000 and 2000-2001 school year allocations, nine states used this option.
The Census Bureau was initially charged to produce updated estimates of poor school-age children at the county level for use in Title I allocations for the 1997-1998 school year. For this purpose, the Census Bureau provided county estimates of the numbers of school-age children in 1994 in families with incomes below the poverty level in 1993. The estimates were developed from a statistical regression model that used administrative data from Internal Revenue Service and Food Stamp Program records for 1993, estimates of poor school-age children in 1989 from the 1990 census, and 1994 population estimates to predict county numbers of poor school-age children in 1993 as measured in the March Income Supplement to the Current Population Survey (CPS).
The model was estimated for counties with one or more households with poor school-age children in the CPS sample–about one-third of total counties. To increase the reliability of the predictions, the model used weighted averages of 3 years of data from the March 1993, 1994, and 1995 CPS, covering income in 1992, 1993, and 1994. For counties in the CPS sample, the model predictions were combined with the direct CPS 3-year average estimates for those counties, in a procedure that weighted the two estimates according to their relative precision. For the remaining counties (two-thirds of the total), the model prediction for a county was the estimate for that county. As a last step, the estimates from the county model were calibrated to estimates from a similar statistical model for states.
The data used in the county model are obtained from several sources, and most data are not available until 2 years after the period to which they refer. When the developmental work began in 1994, the Census Bureau decided that it would not be able to produce estimates in time for the 1997-1998 allocations for a later year than 1993, given the time required for acquiring, processing, and applying the data for a new statistical model.
In its first interim report (National Research Council, 1997), the panel reviewed the Census Bureau's modeling approach favorably but concluded that there had not been sufficient time to thoroughly evaluate the updated estimates produced by the specific model that the Bureau developed. As an interim solution for Title I allocations for the 1997-1998 school year, the panel recommended that the 1993 county estimates be averaged with 1990 census estimates. This
recommendation was adopted. Subsequently, the Census Bureau completed an extensive evaluation of the county model, modified it in several respects, and produced a revised set of 1993 county estimates of poor school-age children. In its second interim report (National Research Council, 1998), the panel recommended that the revised 1993 county estimates be used for Title I allocations for the 1998-1999 school year, which was done.
For both the 1997-1998 and 1998-1999 school years, the Department of Education used the Census Bureau's poverty estimates to make allocations to counties. As in the past, the states then allocated the county amounts to school districts. The states used a variety of data sources for these allocations: many states used 1990 census data wholly or in part; some states used such data sources as numbers of children approved to receive free or reduced-price lunches under the National School Lunch Program or children in families receiving Aid to Families with Dependent Children (or its successor program, Temporary Assistance to Needy Families) in each district. For basic (but not concentration) grants in some states in which the boundaries of school districts bore little relationship to county boundaries, the department permitted the state to ignore the county allocations in dividing up the total allocation amount for the state among school districts. The Department of Education must approve a state's allocation plan but is not required to approve the specific estimates used by a state or the allocation amounts.
School District Estimates
For Title I allocations for the 1999-2000 school year, the Census Bureau was charged to provide the Department of Education with updated estimates of poor school-age children for school districts. The 1994 legislation required the department, in turn, to make direct allocations to school districts rather than to counties unless the Secretaries of Education and Commerce determined that the school district estimates were inappropriate or unreliable for this purpose, taking into account the panel's recommendations. Under this procedure, the states would not be involved, unless they elected to exercise the provision in the 1994 legislation that permits a state to reallocate the Department of Education's allocation amounts for all school districts in the state that have an estimated 20,000 or fewer people.
There appear to be several reasons that Congress in the 1994 legislation deemed it desirable for the Department of Education to make direct allocations to school districts. First, direct allocations by the department impose a measure of consistency on the allocation process. Second, direct allocations to school districts solve a problem with the concentration grant formula in which a county may not be eligible for a concentration grant, but one or more of the school districts in the county may meet the eligibility criteria. (This can happen when a poor school district is located within a county that, on average, is not poor enough to qualify.) Under a two-stage allocation process, poor school districts in coun-
ties that do not qualify for a concentration grant would receive less funds than they would receive with direct allocations.4 Finally, if adequate data were available for estimation, the use of updated school district-level estimates in the allocations would take account of changes that have occurred since the previous census in poverty among school districts within counties.
In early 1999 the Census Bureau provided estimates for school districts of the numbers of school-age children in 1996 who were living in families with incomes below the poverty level in 1995. Developing reliable updated estimates for counties is not easy, and the task is much more difficult for school districts. Some school districts are the same as counties. However, most school districts are smaller than counties, many of their boundaries cross county lines, and the boundaries can and often do change over time. Also, some school districts provide education for specific grade levels, such as K-8 or 9-12. Largely because of these complicating factors, there is a paucity of data for developing updated poverty estimates at the school-district level: there are currently no school district equivalents of the Internal Revenue Service or Food Stamp Program data that are used in the Census Bureau's state and county estimation models.
Because of the lack of data at the school district level, the Census Bureau's procedure for developing 1995 school district poverty estimates used a simple model that assumes that the proportions or shares of poor school-age children in school districts within each county in 1995 were the same as they were in 1989 (as measured by the 1990 census). The estimation procedure involved the following steps: 1990 census data were retabulated to match 1995-1996 school district boundaries (determined from a special survey); the proportion of the county total of poor school-age children in the 1990 census was determined for each school district (or part of a school district) in the county; and the 1990-based proportions were then applied to updated 1995 county estimates from the Census Bureau's county model to produce 1995 school district estimates.
Because of the time required to complete the survey of 1995-1996 school district boundaries and the time lags in the availability of data for the county model, the Census Bureau was not able to produce school district estimates for later than 1995 to be used in allocations for the 1999-2000 school year. Moreover, the Census Bureau's shares-based estimation procedure did not capture intracounty variation in the extent to which school-age poverty increased or decreased among school districts between 1989 and 1995. In addition, the estimates of school district shares of poor school-age children within counties based on 1990 census long-form data were subject to high levels of variability due to sampling error for many small districts. However, the estimation procedure produced estimates more recent than the census, it was consistent across the
States could reserve up to 2 percent of their concentration grant funds to allocate to such districts.
nation, and it responded to the concern that concentration grants be directed to all eligible school districts, including those in counties that were not eligible.
In its third interim report (National Research Council, 1999), the panel concluded that although the Census Bureau's 1995 estimates of poor school-age children had potentially large errors for many school districts, the estimates were nonetheless not inappropriate or unreliable to use for direct Title I allocations to districts as intended by the 1994 legislation. In reaching this conclusion, the panel interpreted “inappropriate and unreliable” in a relative sense. Some set of estimates must be used to distribute Title I funds to school districts. The panel concluded that the Census Bureau's estimates were generally as good as–and, in some instances, better than–estimates that were previously used. On the basis of the panel's study, the Department of Education made direct allocations to school districts for the 1999-2000 and 2000-2001 school years by using the Census Bureau's 1995 school district estimates and other elements of the allocation formula. The department also notified the states of a recommendation by the panel that states electing to reallocate amounts for school districts with fewer than 20,000 people on the basis of some other data source (e.g., school lunch data) should do so on a county-by-county basis so as to reflect (approximately) the Census Bureau's updated estimates of poor school-age children from the county model.
PLAN OF THE REPORT
This reference volume describes and evaluates the Census Bureau's methodology for producing estimates of poor school-age children for counties and states for 1993 and 1995 and for school districts for 1995.5 The report brings together material in the panel's three interim reports to provide a comprehensive description of the current estimation methodology and evaluation results. Similar methods, with likely small modifications, will be used by the Census Bureau to produce state, county, and school district estimates of poor school-age children for the immediate future and to produce other SAIPE poverty estimates, which include total numbers of poor people and poor people under age 18 for states and counties and, for states only, numbers of poor children under age 5. (The SAIPE Program also produces median household income estimates for states and counties.) In the longer run, research and development of data sources and estimation methods will likely lead to changes in the methodology for improved estimates (see National Research Council, 2000).
This reference report contains nine chapters and five appendices. Chapter 2
These estimates are available on the Census Bureau's web site: http://www.census.gov/hhes/www/saipe.html.
describes key features of the Title I allocation formula, such as hold-harmless provisions and thresholds for eligibility, that affect the kinds of estimates that are required and their application and evaluation. The chapter also describes the two-stage allocation process that was used for the 1998-1999 school year and the direct allocation process that was used for the 1999-2000 school year.
Chapter 3 describes and compares the input data sources that the Census Bureau uses to develop state and county estimates of poor school-age children. These sources include the decennial census, the March Current Population Survey, tax return data, and Food Stamp Program data. (Another source is population estimates from the Census Bureau's postcensal population estimates program, which are described in Chapter 8.) The chapter also reviews trends in poverty over time.
Chapters 4-6 describe and evaluate the Census Bureau's procedure for obtaining updated county estimates of the numbers and proportions of poor school-age children in 1993 and 1995. Chapter 4 describes the 1995 county and state models and differences from the 1993 models; Chapter 5 describes alternative 1993 county models that were evaluated; and Chapter 6 provides evaluation results for the 1995 and 1993 models. Although the Department of Education does not use county estimates in Title I allocations when the allocations are made directly to school districts, the county estimates are central to the method used by the Census Bureau to derive updated school district estimates and, therefore, to an evaluation of those estimates. The state estimates of poor school-age children are described and evaluated as well because they are used in deriving the county estimates.
Chapter 7 describes and evaluates, as best as can be done, the data and procedures the Census Bureau used to develop 1995 school district estimates of poor school-age children. Given the scarcity of data with which to implement alternative estimation procedures for school districts, the opportunities for evaluation are very limited.
Chapter 8 describes and evaluates the Census Bureau's procedure for obtaining, from its population estimates program, state and county estimates of the total number of school-age children for 1994 and 1996 and school district estimates of the total number of school-age children and the total population in each district for 1996.
Chapter 9 outlines research and development activities for further work on developing updated county and school district estimates of poor school-age children in the near term (see also National Research Council, 2000:Ch.3).
The appendices cover the following topics: models for county and state poverty estimates (A); regression diagnostics on alternative county regression models (B); county model comparisons with 1990 census estimates (C); use of National School Lunch Program data in New York State to estimate school-age children in poverty for school districts (D); and the estimation procedure for Puerto Rico, which is treated as a county and school district equivalent in the Title I allocation process (E).