C 5&. 235 : 10 CENSUS TRACT PAPERS SERIES GE-40, NO. 10 STATISTICAL METHODOLOGY OF REVENUE SHARING AND RELATED ESTIMATE STUDIES Papers Presented at the Conference on Small-Area Statistics American Statistical Association New York, N.Y. December 27, 1973 ,^ T °^o v \ *;Zi. US DEPARTMENT OF COMMERCE Social and Economic Statistics Administration BUREAU OF THE CENSUS OV 1 «t.v/«. Distrid 12 ^.03 .8 2 Zin P [exempt] pz 2 ZinP [ exempt] z where: Z in P = a zip code (Z) having at least one return filed from place (P) EXEMPTp Z - Number of exemptions in the intersection set of (Z) and (P) EXEMPT^ = Number of exemptions in zip (Z) Thus, if at least 80 percent of people living in a city's zip delivery areas also live within the city's boundaries, then the city is a qualifying place. For such places we assume that the transfer amount per exemption for the city is equal to the amount for the city's zip delivery areas. This amount will be multiplied times the total number of exemptions of persons living within the city to derive the aggregate amount for the city. Control totals for Social Security and Veterans' payments will be developed for the sum of the qualifying places in the county and for the nonqualifying parts. This will be done in accordance with the equations shown below: EQ.Z1 [72CEN(Y)] QP [A(Y)] op [A(Y)] C X [72AMT(Y)] C where: 2 EQ.Z2 [a(Y)]q P = Z in QP | 2 EQ.Z3 [a(y)] c and: = ZinC Z(Y) EXEMPT./| ZinQP X [EXEMPT. J Q PZ Z(Y) \ X EXEMPT./^ . _ ' Z in C QP [ EXEMPT.] cz 22 REVENUE SHARING EQ.Z4 [72CEN(Y)] 2QP QP i nC [ 7 2CEN(Y)] QP EQ.Z5 [72CEN(Y)] [72AMT(Y)] C - [72CEN(Y)] 2NQP ZQP where: CEN (Y) = Z(Y) = AMT (Y) = SS VP(E) VP(P) SSA SS payments Vet. Admin. Ed. Pay. Other Vet. Admin. Pay. CEN(SS) from EQ.C1 BEA VET. ED. Pay. BEA Other Vet. Pay. and where: QP = Qualifying place in County (C) NQP = Nonqualifying part of County (C) C = County (C) Z in QP = a zip code (Z) having at least one return filed from qualified place (QP) in County (C) Z in C = a zip code (Z) having at least one return filed from County (C) EXEMPTqpz = Number of exemptions in the intersection set of (QP) and (Z). EXEMPT^ = Number of exemptions in the intersection of (C) and (Z). Once these control totals have been determined, then the Social Security estimates will be made as follows: I. For Qualifying Places EQ.S1 [72CEN(SS)] QP [72 CEN A(SS)J Q P I 72 CEN A(SS)] Q P QPinC X [72CEN(SS)] SQP where: EQ.S2 [72 CEN A(SS)] QP [ A(SS) ] op 2 [A(SS)J QP QPin C X [72 CEN (SS)] S Qp X [DEATH ADJ. (SS) ] q P EQ.S3 [A(SS)] QP 2 ZinQP Z(SS) EXEMPT, Z inQP [ EXEMPT] QPZ QP II. For Parts of Qualifying Places EQ.S4 [72CEN(SS]p Q p [72 CEN A(SS)3 P Qp 2 Q72CEN A(SS)J PQP PNQP X [72CEN(SS)] QP ESTIMATES OF MONEY INCOME where: EQ.S5 [72CEN A(SS) ] P q P 23 [72A(SS)] pgp X [DEATH ADJ. (SS)] pgp EQ.S6 [72A(SS)J PQP 69CEN(SS) 69 POP PQP 69 CEN (SS) \ 69 POP f q P 72 CEN (SS) 72 POP X [72 pop] PQP QP III. For Parts Outside Qualifying Places EQ.S7 [72 CEN (SS)] NQP [72CENA(SS)] NOP 2 C72CENA(SS)] NQP NQP X [72CEN(SS)] E where: NQP EQ.S8 [72CEN A(SS) ] NQP = [72A(SS)] NQP X [ DEATH ADJ.] NQP EQ.S9 [72A(SS)] NQP 69 CEN (SS) \ 69 POP / NQP 69 CEN (SS )\ 69 POP J SNQP 72 CEN (SS) 72 POP X [72POP] NQP 2NQP where: SS = Social Security payments QP = Qualifying place in County (C) C = County (C) EXEMPTqpz = Number of exemptions in the intersection set of (QP) and (Z). PQP = Part of qualifying place (i.e., part of a township located in the qualifying place. NQP = A nonqualifying place. Subcounty estimates of the remaining sources of income will be made by assuming the subcounty distribution for each income type is the same as the distribution of the indicated allocator. These amounts will then be added together and their sum raked to the county estimate for nonfarm self-employment plus "other income" as shown below: EQ.1 [72 CEN (Y)] p L 72 CEN A(Y)] P 2 [72CEN A(Y)J p X [72 CEN (Y)] p EQ.2 [72 CEN A(Y)] p 2 [72U.CEN A(Y)] p Y 2 [2 ( 72U.CEN A(Y)) P Y X [72CEN(Y)] c X [DEATH ADJ (Y)] p EQ.3 [ 72 U.CEN A(Y)] P [72A(Y)] P [72A(Y)"J c X [72AMT. (Y)] c 24 REVENUE SHARING EQ.4 [72U.CEN A(Y) ] P [69A(Y)] P [69A(Y)J c X [72 AMT. (Y)] c EQ.5 [72U.CEN A(Y)J p [69A(Y)] P [69A(Y)] QP X [72 AMT (Y)] Q P EQ.6 [72 U.CEN A(Y)] p [69A(Y)] , P [ 69 A(Y) ] SNQP X [ 72 AMY (Y) J NQP EQ.7 [72 U.CEN A(Y) ] p 72 AMY (Y) X [72 POP] p 72 POP where: CEN(Y) = U.CEN A(Y) = Allocator A(Y) AMT. (Y) EQ. to EST. U. CEN A(Y) = NFSE PLUS OTHER INCOME INT. DIV. NFSE RENT& ROYALTIES ESTATE & TRUST PENSIONS VET. ED BEN Parts of QP Other Parts of Co. OTHER VET. PAY. Parts of QP Other Parts of Co. UNEMPL. FED. FELLOWSHIPS INDIAN TRANSFERS ALL OTHER TRS. 72 IRS Interest 72 IRS Dividends 172 Estimates TrSC/E Income ^ Veteran students receiving other income in 1969 J 969 Veteran Population 1969 No. Unempl. 1969 No. Grad.Stn. 1969 No. Indians 72 BE A Interest 72 BEA Dividends NFSE & 72 BEA Net Rent & Royalties & Estates & Trust & Pensions [A(VP(E))] Q p: fromEQ.Z2 [72CEN(VP(E))] SNQP;fromE0 . z5 [A(VP(P))] QP:fromEQ . z2l [72 CEN (VP(P))] 2NQP from EQ.Z5 72 BEA Unempl. Benefits 72 BEA Federal Fellowships 72 BEA Indian Transfers 72 BEA Other Transfers and where: P = Subcounty part of County (C) C = County (C) QP = Qualifying place in County (C) NPQ = Nonqualifying part of County (C) ESTIMATES OF MONEY INCOME 25 APPENDIX KEY CODE FREQUENCY OF RETURNS The "coding guide" technique for geocoding nonrespondents for income estimating purposes rests on two assumptions. First, that the geocodes on the fully coded IRS returns, are for the most part, correct. Second, that the geographic code frequency for a given address key is the same for nonrespondents as for respondents. The meaning of the second condition will be discussed shortly. If these two assumptions hold, then the procedure should yield estimates of the number of returns and income amounts which have the same expected value for a governmental unit as would have been the case if initially all returns had been fully coded. A summary of the procedures is as follows: First, for all fully coded filers an address key is determined. This key consists of a 2-digit code indicating their State, their 5-digit Zip Code, up to 7 characters of the place name where their post office is located, and a code indicating their type of address. The following type of addresses were delineated: 1 = a numeric street address, 2 = a rural route, RFD, Star Route, etc. types of addresses, 3 = a post office box address, and 4 = all other types of addresses. Second, for each address key, all of the geographic codes associated with that key are arrayed along with their frequencies of occurrence. Third, nonrespondents are assigned one of the geographic codes associated with their address key in a random manner using the code frequencies as probabilities. For example, an address key may have four codes associated with it as follows: 23451 96ABCDEFG1 231476293424 .73 231476293426 .22 23147660523 .03 231491233723 .01 An uncoded return with this key would have a probability of .73 of being given the 1st code, a probability of .22 of being given the 2nd code, and so forth. Let us now return for a moment to the second assumption of this procedure. It was assumed that the geographic code frequency for a given address key is the same for non- respondents as for respondents. This assumption will hold, at least approximately, in all cases except when one code does not dominate a key and there is a sizeable variance in the response rates among the codes of the key. Thus, a dominant code will mitigate against the effect of different response rates for the codes in a particular key. To the extent that no code dominates a key, the assignment of geocodes to nonrespondent returns will be a function of differences in the response rates among the codes in that key. It is not possible at this time to assess the frequency or the magnitude of this problem and its effect on the estimates, however, several tables have been designed to measure the effects when the processing begins. Although this technique is still experimental, it is believed to be a promising approach to the vexing problems of assigning meaningful geographic codes to addresses. Measurement of State and Local Tax Effort John R. Coleman Bureau of the Census Three primary data elements are used in determining revenue sharing entitlements: population, income, and taxes. This paper will address itself to the composition of, development of, and use of the tax element and of a related element— intergovern- mental transfers. To develop essential tax data, one must first answer the question, "What is a tax?". A rather simple and straightforward question it would seem on the surface. All of us pay taxes, and while some taxes have greater visibility than others, we suspect we can recognize a tax when we see one. Unfortunately the identification of tax amounts collected, and particularly the determination of which governmental unit has imposed the tax, is frequently a difficult matter. Definition of Taxes Let us first examine the definition of "taxes" as this concept is used in determining tax effort for interstate allocation of revenue sharing funds. Later, we will discuss the concept of "adjusted taxes" which is used for the intrastate distribution. Taxes are defined in Section 109 of the Fiscal Assistance to State and Local Governments Act (hereafter referred to as Revenue Sharing Act) as: "... compulsory contributions exacted by the State (or by any unit of local government or other political subdivision of the State) for public pur- poses (other than employee and employer assess- ments and contributions to finance retirement and social insurance systems, and other than special assessments for capital outlay), as such contribu- tions are determined by the Bureau of the Census for general statistical purposes." "As such contributions are determined by the Bureau of the Census for general statistical purposes"— this is the phrase that restricts, that restrains, that draws the lines of limitation tightly around that most blasphemed of words— "taxes". To the anguish of many local government officials, the above qualifica- tion results in the exclusion of such compulsory contributions for public purposes as maintenance assessments, tax equivalents, sewer charges, garbage collection fees, and fines. One may wonder why Congress adopted the Census Bureau's relatively narrow definition of taxes. The answer probably lies in large part in the immediate accessibility of the Census Bureau's highly respected public sector statistical series with its underlying national classification system during consideration of the Revenue Sharing legislation. The public finance series extends back more than 70 years and provides a reasonably consistent application of standard definitions imposed on an extremely diverse universe of 78,21 8 1 local governments in addition to scores of thousands of dependent agencies, boards, commissions, districts, and authorities. The Census Bureau's Governmental Finances publication, generally published in early fall of each year, presents State totals of State and local government tax revenues by level of government. These data are utilized by the Office of Revenue Sharing in formulae which calculate each State's entitlement. State Tax Data Two surveys are conducted each year to generate State government data. First, State tax aggregates 2 are developed by mailing questionnaires (Form F-5) to appropriate State tax officials shortly after the close of the fiscal year. 3 The questionnaire requests a significant amount of detail for the various categories of taxes, i.e. property, income, general sales, specific sales, licenses, severence, inheritance, etc. Asa reporting and editing aid, amounts for the previous year are included on the questionnaires. All returns are edited for consistency with aberrant entries receiving technical followup. The results of this survey are usually published in November in the State Tax Collections series. One other tax element is required for the interstate distribu- tion—State income tax collections. Since the Annual Tax Survey generates data on a fiscal year basis, and since by law, 4 the collections used in the revenue sharing formula must be on a calendar year basis, required data are obtained by summing the results of four quarterly tax surveys conducted by the Census Bureau. 'As of January 1972. Governmental Organization, Volume I, 1972 Census of Governments, U.S. Bureau of the Census, Washington, D.C., July 1973. 2 State tax aggregates comprise amounts collected— including interest and penalties but excluding amounts refunded— from all taxes imposed by a State and collected by it or collected and transmitted to it by local governments. State tax revenue includes local shares of State-imposed taxes except any amounts locally collected and retained. 3 June 30 for all States except 12-month period ending March 31 for New York, August 31 for Texas, and September 30 for Alabama. "Public Law 92-512, Title I, Section 109(b)(2). 26 MEASUREMENT OF TAX EFFORT 27 Local Tax Data The local government tax aggregates are developed through an annual sampling of about 16,000 governments. Certainty survey coverage applies to all counties of 50,000 or more population in 1960, all cities of 25,000 or more population in 1960, all governments whose relative importance in their State, based on expenditure or debt was above a specified size, and a random sample of remaining units. Summation of data from certainty sample units generates aggregates which reflect over 75 percent of the amounts obtained in a census of the 78,000 governments. Computations of property tax revenue show a relative standard error of less than 2 percent for virtually all States and less than 1 percent for over two-thirds of the States. Extensive efforts are made to keep nonsampling error at a minimum through substantial use of field and office compila- tion of large units and intensive edit of the smaller units. Questionnaires are edited for historical and internal consistency utilizing audit reports, various State reports, and many other secondary sources in the field of public finance to ensure the highest level of accuracy possible. In regards to tax aggregates developed in this annual survey, it should be noted that taxes of all State and local governments (including independent special districts and school districts) are counted when measuring tax effort only for the interstate distribution. Taxes for education purposes and taxes imposed by inde- pendent special districts are excluded when determining the intrastate distribution. Excluded taxes in Fiscal Year 1972 represented 25 billion of the 49 billion dollars 5 imposed by local governments. Given the magnitude of excluded taxes, one can readily understand the dissatisfaction of officials of those governments whose citizens carry a heavy tax burden, much of which counts for naught in calculating intrastate entitlements because it is imposed for educational purposes or because it is imposed by an overlying or underlying independent special district. Adjusted Taxes Only general purpose governments, (approximately 38,500 counties, municipalities, and townships), are eligible for revenue sharing entitlements. To calculate intrastate entitlements a new concept was introduced, namely, adjusted taxes. Adjusted taxes are taxes adjusted by excluding an amount equal to that portion of taxes which is properly allocable to expenses for education. Expenses for education are construed to include expenditures for current operation, capital outlay, and debt service. The boundaries of education are defined broadly enough to encompass primary and secondary schools as well as higher education,— vocational as well as academic studies. Before going further on the subject of adjusted taxes, it is important to realize that from the viewpoint of a local government, a larger adjusted tax is good, a smaller adjusted tax bad. Except where the per capita limitations of the Act restrict entitlements, a larger adjusted tax generally produces a larger entitlement of revenue sharing funds. And given the competitive nature of the Act, what is good for one government within a county area is generally bad for one or more competing governments within that same area. Adjusted Taxes is a concept which presumes the ability to specifically identify school taxes; however, the realities of municipal finance question that presumption and present data handlers with some of their most difficult problems. While there are 15,780 school districts with sufficient fiscal and administrative autonomy to be classified as separate independent governments, there are 1,457 school systems located primarily in New England and in the South which are classified as dependent upon a county, municipality, or town. For the most part the independent school districts have clearly identifiable taxes and thus present few data problems. Further, upwards of two-thirds of the 1,457 dependent school systems which have their school finances intermingled with general government finances, for one reason or another, maintain records which clearly identify school taxes. Massachusetts is an example of this situation where for a number of years State law has required such identification. 6 However, the remaining one-third of the dependent school systems located primarily in Connecticut, Maryland, and Virginia generally cannot identify school taxes. For most units within these States an appropria- tion is made for school purposes out of a General or similarly named fund. The revenues to support expenditures of this fund come from a variety of sources including both tax and nontax revenue. Original procedures for determining school taxes for such units consisted of identifying total school expenditures, identi- fying all specific school intergovernmental revenues, school service charges and other revenues with specific school identity, netting revenues against expenditures, and equating the residual amount with school taxes. This procedure came under attack in the early days of the revenue sharing program as opponents successfully argued that the procedure failed to adjust for cash surpluses used to finance schools or failed to permit a proportionate allocation of nontax revenues which carry no specific school identity. Rectification of this problem came when the Office of Revenue Sharing issued a regulation (Section 51.21(b)(2)) revising the method of computing adjusted taxes. Called the RS-12 procedure after the form used in its calculation, it permits development of a ratio of taxes to nontax revenue with the ratio applied to school expenditures to produce a derived school tax amount. Beginning and ending of year cash balances are added or subtracted as appropriate from net revenues to allow for use of cash draw down or conversely, disallow use of cash build up. This procedure can only be used where school taxes carry no specific school identity. In addition to school taxes imposed by independent school districts and by general purpose governments for dependent school systems, there are countywide school levies imposed in many States by counties (predominantly in the Midwestern and Western States) and paid intergovernmentally to the inde- s Governmental Finances in 1971-72, U.S. Bureau of the Census, Washington, D.C., November 1973, table 16, and special tabulation of adjusted taxes. 'Annotated Laws of Massachusetts, Volume 2, Chapter 59, Section 23C. 28 REVENUE SHARING pendent school districts. While these taxes are based on millage levied specifically for school purposes, care must be taken to ensure their exclusion from the county government's adjusted taxes. School Debt Service One further problem that exists in identification of school taxes relates to servicing school indebtedness. General purpose governments in some States— Alabama is a good example— issue debt in the name of the county or city and either construct schools or pay the bond proceeds over to independent school districts which construct their own schools. The debt service (payment of interest and reduction of debt) is generally financed from property taxes. Accordingly county or city taxes must be reduced by the amount of debt service to arrive at their adjusted taxes. This is a relatively easy operation when the debt carries an education identification. However, where school debt is collapsed into general purpose borrowing, e.g. for schools, fire stations, roads, etc. it is difficult if not impossible to determine how much taxes for debt service is attributable to education. With current record keeping practices the difficulty increases with each passing year's departure from the year of the original issue. The problem will be compounded if interest rates in the municipal bond market reach a level where governments are motivated to refund outstanding debt to obtain a lower interest cost. Municipal finance officials would be remiss in their duties if at that time they did not refund school debt through general purpose bonds to remove school identity. It is likely that regulations will have to be written in the near future to remove the incentive to refund school debt into general debt. Annual Surveys of Tax and Intergovernmental Revenues During the past 18 months the Census Bureau has conducted three surveys to determine adjusted taxes and intergovernmental revenues covering three fiscal years (1971, 1972, and 1973) for the 38,500 general purpose governments. In the interest of minimizing respondent burden, the first two surveys made extensive use of the data obtained in the Annual Survey of Municipal Finances and in the 1972 Census of Governments. Supplemental questionnaires were dispatched to solicit data relating to school taxes as well as to complete the file for units that were not on the annual survey panel, or that had not responded to the earlier surveys. The Fiscal Year 1973 survey used two basic questionnaires; a Form RS-9 dispatched to 34,000 governments, and a Form RS-12, dispatched to about 400 governments which finance schools through General Fund appropriations. Data for the remaining 4,000 general purpose governments were collected through cooperative arrangements with State officials in four States; Florida, New Jersey, North Carolina, and Wisconsin. Fiscal Years Compacting collection of data for 3 fiscal years into an 18-month period, while necessary to provide the most current data possible, increased the confusion in an already perplexing fiscal year situation. For each fiscal year, data are requested for the 12-month fiscal period that ended between July 1 and June 30. Of the 3,044 counties in the United States, 52 percent have December fiscal years, 35 percent June, 7 percent September and 6 percent are spread over the remaining months. Fiscal years of municipalities and townships have even greater disper- sion. Add to this mixture, dependent agencies which have different fiscal years than their parent government; sprinkle in four entitlement periods, three covering 6 months, one covering 12 months; add several planned use and actual use reports covering differing periods; season with thousands of part-time nominally paid local government officials, and the result can reasonably be expected. Confused fiscal year reporting problems to date have been second only to school tax identification problems. There is, however, hope for a diminution of the fiscal year problem in the future. Intensive efforts to instruct local officials regarding correct fiscal periods; a spreading of data requests to once in a 12-month period; and synchronization of Census Bureau/Office of Revenue Sharing contact with local officials— all can be expected in the future to lessen considerably misinterpretation of fiscal periods covered. Agency and Shared Taxes A problem area demanding significant editing attention is the reporting of shared taxes or agency taxes as one's own tax effort. A local government can only receive credit for taxes which it imposes itself, with the single exception that State imposed, locally collected and retained taxes are credited to the local government. Other than this exception, it is irrelevant which government collects the tax. Frequently a county collects property taxes for all overlying and underlying governments; in such an instance, the county can be credited only for those local taxes imposed by its legislative body. For other taxes, a county is viewed simply as acting in an agency capacity. At a higher level a State government may collect a county imposed sales or income tax which is piggybacked on a State sales or income tax and remit the proceeds to the county. In such a case the State is acting in an agency capacity for the county imposed portion. For the hundreds of State imposed taxes which are shared locally, 7 the local governments must report such amounts as intergovernmental revenues— not as taxes. In collecting data on taxes and intergovernmental revenue, experience has proven 7 See State Payments to Local Governments, Volume 6, 1967 Census of Governments, U.S. Bureau of the Census, Washington, D.C., December 1968, for a detailed analysis of shared taxes. The corresponding publication from the 1972 Census of Governments is scheduled for publication in February 1974. MEASUREMENT OF TAX EFFORT 29 that very detailed special instructions must be provided local officials specifying whether a tax reflected in his records is: (1) a tax imposed by his government and therefore reportable in adjusted taxes, (2) a tax for which his government acts in an agency capacity and therefore nonreportable, or (3) a shared tax and therefore reportable as an intergovernmental revenue. Memphis Rule One exception exists to the Census Bureau's procedure of attributing local taxes to the imposing government. Section 109(e) of the Revenue Sharing Act (known as the Memphis Rule) provides that any county imposed sales tax exacted from a municipal or township area and shared with the municipal or township governments without specifying purposes for which the unit may spend the revenues shall be treated as taxes of the receiving government. The Governor must first certify that such a condition exists before the rule can be invoked. Memphis Rule adjustments have been made to date in seven States: Alabama, Louisiana, Nevada, New York, North Carolina, Tennessee, and Virginia, with the largest adjustments en- countered in Tennessee and in the upstate counties of New York. The Memphis Rule is so named because of its applica- bility in Memphis-Shelby County, Tennessee, where an adjust- ment of approximately 12 million dollars in adjusted taxes was required in Fiscal Year 1971. Note that the Memphis Rule exception applies only to county imposed sales taxes, not State sales taxes, not city sales taxes, not county income taxes, not county property taxes. The rationale for singling out sales taxes for special treatment- county imposed sales taxes at that, is difficult to follow if promoting greater equity was a motivating factor in its adoption. It may be conjectured that there was not widespread Congressional awareness during consideration of the bill of the extent to which governments share taxes which they impose, but which they extract from inhabitants of the shared govern- ment's area. This may be an issue that receives greater attention when the revenue sharing program is subjected to Congressional analysis for extension, modification, or termination in the future. Intergovernmental Revenues A rather innocuous little paragraph in Section 108 of the Act imposes one of the most burdensome data collection responsi- bilities of the program. Subsection (b), paragraph 6(C) reads in part: "Limitation— The amount allocated to any unit of local government under this section for any entitle- ment shall not exceed 50 percent of the sum of (i) such government's adjusted taxes and (ii) the intergovernmental transfers of revenue to such government...." In effect this paragraph requires that every intergovernmental dollar passing between and among local governments be identified. This is a monumental task. It first requires that all agencies, boards, commissions, districts, authorities, and depart- ments be properly classified as either governmental or non- governmental and for those which are governmental, that they be classified as dependent upon a specific government. Local government officials generally tend to carry a narrow or restricted view of the scope of their government's activities. The hospital authority, joint health department, city-county airport authority, etc. are frequently viewed locally as not being part of one's government. In actuality these agencies may meet all necessary criteria for classification as dependent agencies of governments which deny or are unaware of their dependency. Extensive use is made of supplementary instructions in Census data collection programs to inform local officials of the full scope of their government's activities. A multitude of quasi-governmental agencies were created during the 1960's, often as a result of the profusion of categorical grant programs which came into existence; Com- munity Action Agencies, Drug Rehabilitation Organizations, Planning Agencies, and Councils of Governments are a few of the type of agencies through which significant governmental activities are carried out, but which may or may not be governmental units. Failure to identify a Community Action Agency (CAA) which meets criteria for classification as a county agency would result in understatement of the county's intergovernmental revenues by the amount of Federal and State aid which the CAA receives. The need for complete identification of intergovernmental revenues is particularly acute in Southern counties, as a number of these units are subject to the 50 percent limitation; i.e., 50 percent of the sum of their taxes and intergovernmental transfers exceed the amount for which they are entitled; hence the excess is lost by the county government and added to the State government's entitlement. It becomes obvious that for those governments up against the 50 percent ceiling, if a $100,000 Federal welfare grant is inadvertently unreported, the county could lose up to $50,000 in its revenue sharing entitlement. The true potency of data in the revenue sharing program manifests itself when reporting errors of this nature are analyzed. Comparing Noncomparables If one were to attempt to describe in one word the most striking feature of revenue sharing, the word would have to be "competitiveness". Quite probably there are few programs which have ever caused officials of local governments to compare themselves with other governments to the degree that revenue sharing has. Surface comparisons made by competing governments have often focused on population, an intuitive sense of relative income levels, and total property tax collec- tions or millage imposed. These comparisons frequently fail to consider the tremendous diversity that exists both in organiza- tional structure of the governments and in the financing vehicles utilized. For example, one government may make greater use than another government of public housing which generates in lieu of tax payments rather than property tax payments. By definition, "in lieu of taxes" do not qualify for inclusion in adjusted taxes. Another government may finance garbage collection from taxes; another government, from service charges, or that function may even be performed by private firms. Service charges are excluded from adjusted taxes. Another government may use maintenance assessments or special assessments for a given purpose; the next may use property taxes— the assessments can not be counted. One city may have its Public Works Department 30 REVENUE SHARING perform a function financed primarily from taxes. In a neighboring city the function may be performed by the county or by an independent special district. The latter city fails to get credit for the tax effort that finances this function. And another government may operate a water utility, charging higher rates to produce a profit which is then used to finance general government. The profit, while having the affect of tax, cannot be counted. However, a neighboring government may operate its utility on a break-even basis, but levy taxes at a rate designed to produce revenue at the same level as the profit producing utility. The latter would be able to count its taxes and thus obtain a higher adjusted tax. The list of "inequities" is endless, and the organizational and financing diversity is even greater when making interstate comparisons. Tax Relief An area with the potential for introducing major inequities into the revenue sharing program is that of the property tax relief programs which have exhibited astronomical growth in the past several years. Hundreds of millions of dollars are involved in these tax relief programs. Three basic methods are in use today to implement tax relief. First, local governments reduce property taxes of their citizens and thus forego collection of taxes. States reimburse the local governments for the taxes foregone. This procedure produces no double count in the State-local tax effort. Secondly, local governments impose and collect the taxes from their citizens. State governments then permit the tax- payers to reduce their State income tax liability by the amount of the tax relief. This procedure also produces no distortion because State income taxes are reported net of the tax relief refunds. Thirdly, the situation is the same as the second except the income tax is not used as a vehicle for granting relief. Instead, the State makes a direct expenditure rather than a tax expenditure. It sends a check to each taxpayer eligible for tax relief. This method does distort in that it counts both the full local taxes and the full State taxes. There is no reduction for tax relief. States which opt to use this last method gain a competitive advantage over other States insofar as tax effort determines the proportion of Federal funds to which each State is entitled. This last procedure is politically useful to executive officers as they can receive political credit from local taxpayers for the full amount of the tax relief check. This has greater visibility than the netting procedure used in conjunction with the income tax refund procedure. In view of the size of the mushrooming tax relief programs, it is probably reasonable to expect that special regulations will have to be developed to deal with this problem. Conclusion Thus, it can be seen that frequently the "maldistribution of entitlements" (at least as perceived by local government officials who feel "wronged") has its roots in organizational relationships created and fiscal mechanisms utilized in pre-revenue sharing days— some created and utilized as a result of extensive deliberation; others decided almost by chance, by the political vicissitudes that at times shape a government's course of action. With the growing awareness of the critically important role that data play in the revenue sharing program, a number of local governments, operating within constitutional, statutory, politi- cal, and economic restraints, have begun to weigh trade-offs of various organizational realignments and modification of financ- ing vehicles in order to maximize the one data element that is most within their capacity to affect; i.e., adjusted taxes. It is still too early to assess the full extent of changes that will occur as a result of the data intensive nature of the revenue sharing program, but one can safely say that the statisticians will be monitoring changes closely to ensure availability of a data file that is both current and accurate. Given the magnitude of the funds being distributed, nothing less will suffice. Statistics of Revenue Sharing After 1 Year Arthur L. Hauser, Office of Revenue Sharing I'm pleased to have the opportunity to talk to you about our impressions of data-related changes that have occurred in the period slightly over 1 year since general revenue sharing was born. After the passage of the State and Local Fiscal Assistance Act of 1972, better known as the Revenue Sharing Act, in October 1972, Robert P. Strauss, then consultant to the Secretary of the Treasury, wrote: "The historic action of Congress to create general revenue sharing represents not only a revolution in the form of American federalism, but also a revolution in the Federal statistical base that will be used to administer the program." (Nation's Cities— October 1972) In my remarks today, I would first, like to consider the so-called "revolution to the Federal statistical base" that Mr. Strauss expected; the kinds of changes to State and local governments' data that we can see happening through our day-to-day work with the Bureau of the Census. In addition, I'd also like to touch on general revenue sharing's impact on the Bureau of the Census, as I see it. It is apparent from our daily contacts with Census personnnel that the data needs of revenue sharing have exposed them to frequent inquiries from many small units of local government— units that previously were not in their survey of finances or units that had reported previously only in the Census of Governments every 5 years. Revenue Sharing's Influence on the Data The influences of general revenue sharing on the State and local government's data can be categorized in two ways. The first relates to activities which are likely to improve or change that data from sources which had been in existence prior to the advent of revenue sharing. The second relates to the availability of new data sources, those not existing prior to revenue sharing. DATA IMPROVEMENT Prior to revenue sharing, the Census Bureau's activities with local governments consisted of the previously-mentioned survey of finances to about 16,000 of the 78,000 governmental units. As a voluntary periodic survey, it was subject to the problems familiar to most of you— involving mostly response and sta- tistical significance. As a result of general revenue sharing, there are many specific mechanisms which are operating to bring about improvement in the Census-produced data used by the Office of Revenue Sharing. This is especially true of the revenue data— adjusted taxes and intergovernmental transfers. In the case of the revenue data, the improvement process begins at the doorstep of the State and local government officials who either complete survey forms or certify to their accuracy. By now, most officials of recipient governments generally understand that the infor- mation that they report on the Census Bureau revenue sharing survey forms has a direct bearing on the amounts of revenue sharing funds that they receive. This understanding has con- tributed to revenue data in the following ways: 1 . The response rate to Census Bureau surveys of revenue data for revenue sharing has improved considerably since the Special Revenue Sharing Survey to obtain Fiscal Year 1971 revenue data. At the time of our first allocation, the Special Revenue Sharing Survey to obtain this data had a response rate of approximately 75 percent. Strenuous efforts to increase this response rate after the allocation were made by Census Bureau. Many governments who had not responded initially did so once the revenue sharing payments had begun. This resulted in a final response rate of about 93 percent; it also meant that it resulted in more than 5,000 data changes from the estimated data which were initially used by our office. In the 1972 Census of Governments which was supplemented by some special revenue sharing forms, the response rate for Fiscal Year 1972 revenue data rose to approximatley 98 percent. Presently, the Special Revenue Sharing Survey of Fiscal Year 1973 revenues is preceeding according to plan. Thus far, a response rate of approximately 97 percent has been achieved. 2. Local and State government officials responsible for completing the survey forms have increasingly sought guidance from the Office of Revenue Sharing and the Bureau of the Census. The Bureau of the Census, in particular, has worked hard and spent considerable time responding to these individual requests for help and, in addition, has provided more detailed special instructions than heretofore covering the unique revenue situations of different State and local governments. 3. The survey forms sent to State and local officials by the Bureau of the Census are now more frequently reviewed by local officials with a higher level of responsibility than was previously the case. Prior to this, the survey form was in many cases returned directly by a town clerk or other clerical personnel without review by higher officials. 31 32 REVENUE SHARING Once the raw revenue data are received by the Bureau of the Census, the Governments Division makes careful editing checks to insure its accuracy. Because the Office of Revenue Sharing has on several occasions become aware of some data differences which were large enough to have an impact on the allocations, we have worked closely with the Bureau of the Census to positively identify the existence of an error or to verify the data. The increased use of and refinements in editing checks have already given us more accurate revenue data. When the Census Bureau completes its processing of the data, they provide us with the new set of data. This new data is used by us to calculate recipient governments' allocations. However, the efforts to improve the data do not stop at this point. The Bureau of the Census has an ongoing program to review the quality of submitted data. This review occurs as a result of the submission of revised reports from local governments and additional material received from State and other sources. Furthermore, the Office of Revenue Sharing's Regulations specify that the data used in the allocation must be made available to all units of government to give them an opportunity to challenge the data for their government if they believe it to be in error. Thus, through our administrative procedure now known as the data improvement program, many governments have pointed to data differences— and have supplied documenta- tion for review by the Census Bureau. Our first data improvement program was in December 1972 when we sent the 38,000 recipient governments the data used to calculate their revenue sharing allocations for the first entitle- ment period. Approximately 4,000 governments raised questions about their data in the following months although fewer than half of these appeals were found to have sufficient merit to warrant a data change. This process did much to bring about a better understanding of the data elements (population, per capita income, adjusted taxes and intergovernmental trans- fers) on the part of all recipient governments. We have just concluded our data inprovement program covering the data used in the 4th entitlement period. This time, only 2,000 govern- ments challenged; approximately 1,100 challenges required research while the balance were submitted without documenta- tion. Of the 1,100 inquiries, about 82 percent resulted in revenue revisions. The government address information generated by the Office of Revenue Sharing also contributed to the data improvement. Many governments provided us with address corrections. The Bureau of the Census also has used some of this new address information to improve its address records which are used in sending out survey forms. Obviously, there is a better chance of obtaining the desired data if the survey forms are sent to the proper official for completion. SOME DATA PROBLEMS: SOURCES OF FUTURE DATA IMPROVEMENT Although there has been significant improvement in the data used for revenue sharing, there are reasons for believing that the data are not as good as we would like them to be and, therefore, that there is room for continued improvement. For example, it is likely that some governments may not have reported data discrepancies when it was not to their advantage to do so. Thus, adjusted taxes carried on the high side or per capita income carried on the low side are more likely to go undetected than adjusted taxes or per capita income in the other direction. Also, a few governments may be tempted to exaggerate the amount of taxes that they report, rather than the amount actually collected. We expect, though, that the various checks employed by the Census Bureau should keep this situation under control. A few adjusted taxes reported on the high side have been discovered due to challenges of a governments' data by other governments which are competing with them for revenue sharing funds. We have not encouraged this type of challenge because we have not wanted to emphasize unduly the com- petitive aspect of the formula. However, the irony of the situation is that our experience shows that this type of challenge is not likely to be advantageous to the particular unit of local government raising such a challenge. This is because the reduction of a competing government's adjusted taxes data will result in a reduction to the county area's adjusted taxes. This, in turn, will result in a reduced allocation to the county area. We will, of course, work with the Bureau of the Census to correct all data differences regardless of who discovers them, because this will contribute to greater equity in distributing revenue sharing funds. We expect that our ability to detect and make necessary changes will improve with time. NEW SOURCES OF DATA Besides improving existing data sources, general revenue sharing has required development of new sources of data. Let's look at a few of these. 1. Revenue Data-A Special Revenue Sharing Survey of all general purpose governments will be completed annually except when the Census of Governments is taken to compile data on adjusted taxes and intergovernmental transfers for revenue sharing. The use of the adjusted taxes (taxes minus educational taxes) data in the revenue sharing formula has made it necessary for the Bureau of the Census to develop a new survey form to obtain the data in the desired manner for the program. Perhaps, of more significance is the establishment of communications with all general purpose governments— the bulk of which have populations under 2,500. Of approximately 38,000 general purpose governments, approximately 27,000 have populations under 2,500. And approximately 35 percent of the 38,000 governments (13,500) have populations under 500. Thus, general revenue sharing has brought about a new dimension in State and local government revenue data. 2. Population and Per Capita Income Data— For general revenue sharing, both the 1970 population and per capita income data are adjusted annually to reflect boundary changes for places of 2,500 or more based upon information received by the Bureau of the Census in the annual Boundary and Annexation Survey. Before revenue sharing, information from the Boundary and Annexations Survey had been used by the Census Bureau only to keep their Census maps up-to-date. Still on the subject of population and per capita income, you have heard how Census Bureau has the challenging job of updating these data for local governments using administrative records. SHARING AFTER 1 YEAR 33 3. Population for Indian Tribal and Alaskan Native Village Governments-Although the Bureau of the Census has collected data on Indians and Alaskan native persons in the United States, it has with few exceptions never collected data specifically on Indian and Alaskan native population which are members of and under the jurisdiction of Indian tribal and Alaskan native governments. There has never been a need for this data before. Because this type of population data was needed for revenue sharing, estimates of these populations have been made by the Bureau of Indian Affairs for the Office of Revenue Sharing. Currently, we are exploring other avenues for obtaining updated population data for Indian tribal governments and Alaskan native villages. IMPACT ON THE BUREAU OF THE CENSUS I've already noted some of the ways general revenue sharing has brought important changes in data. And we know that this means that the Census Bureau has successfully made some changes. We know that general revenue sharing is not entirely unique to the Bureau of the Census because Census data is used by other Federal agencies in conjunction with the granting of funds to State and local governments. Yet the Census data used for general revenue sharing is of special importance because the data directly determines the amounts that recipient governments will be allocated. Because the data have this kind of special importance, especially the revenue data, and because the Office of Revenue Sharing must insure the equity and correctness of our alloca- tions in all cases— not merely on the average— we necessarily view the data not as statistics but as ledger entries in each government's account. We view each revenue sharing recipient as a separate "account" as differentiated from all other "accounts." Thus, our criteria for the data are different and our demands for data are those of the accountant. Given these demanding criteria, we see the Census Bureau demanding— and getting— more precise revenue data from respondents. This is happening because there is a recognition that the revenue data are used in a direct, programmatic formula to determine entitlements. Further, such recognition means that it is necessary to achieve high-quality data down to the smallest area level. But we are not excluding the value of estimates and sample data. As you will note by reading our other material here, we use the Census State population estimates and tax effort estimates for the interstate allocation. In conclusion, we feel that those persons like Bob Strauss who anticipated that administering the general revenue sharing program would bring about very substantial changes in the kinds and quality of statistical data were correct. I would also like to point out that we recently turned over to the Census Bureau, who will be acting as our agent, magnetic tapes containing the final data for entitlement periods 1, 2, and 3. For those of you interested, please contact the Census Brueau's Data User's Service Office regarding sale. Making the New Federalism Work: Goals for General Revenue Sharing's Data Needs Robert P. Strauss* Associate Professor of Economics University of North Carolina at Chapel Hill I. Introduction I want to express my great pleasure at being with you today. The passage of general revenue sharing portends a new era in micro-data needs. 1 should note that the general movement towards revenue sharing and away from categorical or condi- tional grants-in-aid, otherwise known as the New Federalism, will require a variety of micro-data sets which are frequently updated. General revenue sharing has highlighted the need for recur- rent measures of population, community money income, and local tax collections. Not only is the "recurrency" worth noting but also the logical unit of analysis, namely general local governments. If we look down the road towards the passage of special revenue sharing, we should be able to anticipate that additional micro-variables by the same as well as other (special) geographic subdivisions will be necessary. Of course, we do have some of this data, especially for more populous areas. But my point here is that the data requirement for smaller areas have yet to be met. A corrolary to this is that with the advent of the New Federalism, more rather than less will (or should) be spent by the Federal Government for micro-statistics. However, I want to indicate that not only should more be spent for collection and upgrading of new and current series, but serious attention be increasingly given to verify such self-reporting data to ensure accuracy. There are many areas in which formula design will depend on data or behavior which ultimately is at the control of a locality or group of recipients. The use of adjusted taxes in general revenue sharing is just one example. There are many other areas in which we might want to reward or dissuade organizational behavior for certain outcomes; however, suffi- cient verification (as well as careful definition) must occur to ensure that program goals are met. Having preambled my remarks with the prognostication that the New Federalism will inevitably require more micro-data and accordingly require greater Federal resources to meet properly these increased needs, let me turn to the data implications per se of general revenue sharing. Last year I noted that a revolution in Federal statistical base was implied by the State and Local Fiscal Assistance Act of 1972; my theme today is that successful revolutions require forthright leadership, for the successful changes in the revenue sharing data base will directly affect the success of the program. I shall discuss in turn the leadership issues relevant to income, population, adjusted taxes, and the formula. II. The Data Issues of General Revenue Sharing A. Income Updates During Executive Session deliberations of the House Com- mittee on Ways and Means, repeated discussions took place on the general problem of updating the data elements ultimately used in the legislation. As you may recall, an overhaul of the welfare system was being contemplated which, coupled with changes in reporting requirements within the tax system, might yield post-censal data on total money income by jurisdiction. I do not think it is a breach of confidentiality of those proceedings to reveal now that the Committee seriously considered requiring that the welfare system contain county and incorporated place of residence as it ultimately did of those in the tax system. Because the welfare reform was pending, and because it was impractical to amend the existing Social Security Act, Undersecretary Veneman testified that costs of several hundred million would be involved as well as considerable inconvenience to modify existing reporting systems. Conse- quently, no legislative action was taken to gather income data about low income persons (or nonfilers). However, it was clear from Veneman 's testimony and Committee sentiment, that eventually this data ought to be collected as it was their clear determination that such residence information be collected of taxpayers to facilitate estimation of money income by political subdivision. Of course, it is now history that the 1972 1040 form required such information and is to be used to obtain income updates. Several things, however, ought to be underlined about this bit of legislative history. First, the tax forms were to be permanently modified. Second, the welfare system as it is modified should be used to provide data about low income persons. ♦Formerly Special Assistant to Deputy Treasury Secretary of the U.S. Treasury, Charles E. Walker, and recipient of Treasury Department's Exceptional Service Award for work on General Revenue Sharing. 34 MAKING THE NEW FEDERALISM WORK 35 More recent history indicates that the commitment to change the 1040 form and to require the welfare system to report has weakened and may vanish. Apparently the Commissioner of the Internal Revenue Service (IRS) does not interpret Section 144 of the State and Local Assistance Act to mean, as I take it to mean, that incorporated place of residence must be recorded and put on tape by the IRS each year. To those of us committed to making the New Federalism work, this can only be viewed as a setback. Also, the Federal assumption of the adult welfare categories has occurred without the requisite modifications of the reporting system. While the default this year (CY73) in obtaining incorporated place of residence may be viewed as a temporary setback to the statistical community, I should note that likely tax reform in 1974 that involves tax simplification, as it undoubtedly will, may well involve the complete removal of the residence questions from the tax form. Moreover, if income maintenance replaces the existing program for Aid to Families with Dependent Children, it is likely that place of residence will not be obtained unless substantial prodding takes place. Without going into the debate on the nature of the question that was asked on the 1972 form (let me parenthetically note here the opinion that a 70 percent response rate to the question is rather high in view of the initial 40 odd percent response rate to the zip code questions that were asked in the mid sixties with greater attending IRS publicity), I would like to comment on the definitional problems that occur when comparing adjusted gross income to total money income. During those Executive Session deliberations Congressman Vanik suggested that total money income be reported on the 1040 tax form. It is clear that a Hague-Simons concept of income is what a tax-writing committee ought to use as a reference point when analyzing changes in the IRS code, not because that is the tax base, but rather because it is a sound point of comparison for equity purposes. While sentiment apparently existed for more complete reporting of nontaxable or partially taxable income, the thought was then that while desirable, it might be better effected through separate legislation. My point here is that we may well see a broader income concept reported through the tax system which would reduce the definitional problems rather consider- ably in obtaining total income of particular places. There is, then, in the area of income data update, room for improved statistical leadership that will ensure that the original vision of updated and improved local income data will occur. This is not to discount the rather substantial difficulties of convincing the Internal Revenue Service and Department of Health, Education, and Welfare that improved local income data is worthwhile. However, it is significant that the legislative history of general revenue sharing contains the elements to make continued progress in this area. B. Population The second area of data updates involves population. Here we are better acquainted methodologically with the problems and prospects of post-censal updates than, I think, in the micro-income area. I would like to confine my remarks here to the area of undercounts that occurred and the extent to which updates in the form of corrections should be made. Again, reference should be made to the legislative history and the ingredients in the legislation to obtain guidance on how the problem should be solved. During Executive Session delibera- tions that were ultimately published (April 17, 1972), then Director of the Census Bureau, George Brown and Congressman, William Green of Philadelphia, discussed various statistical problems, among which was the undercount of certain groups in core cities. Again, the sentiment of the Committee was fairly clear: at that time, while the extent of the undercount was still being researched but was as yet unknown, the general feeling was that when adequate figures were available, corrections ought to be made. We now know the national undercount rates by age, sex, and race, and it strikes me as appropriate to use this data to adjust the population figures. I have done this for several States, and while better data will undoubtedly be generated (State by State rates, for example) it would seem appropriate at this point to account for what we know to have occurred, and then, at a later date, improve our statistics further when possible. I should note that the Secretary of the Treasury has authority to use such estimates, and such updates would be done with attending modification of regulations to provide that retroactive adjustments in payment not be made. Moreover, there is precedent for corrections as corrections have already been made to the census in the income area. Again, the legislative history is clear: errors were anticipated and the Secretary enpowered to use such updated or corrected data to provide corrections. And again, statistical leadership should do this expeditiously. I should note that the National League of Cities-Conference of Mayors has twice recommended that the undercount be corrected. C. Taxes We have more experience now with the adjusted tax data updates than with income and population. It is my judgment that the ability to update the tax data annually suggests the other two elements can be updated and corrected as well. The points I want to stress with the tax data involve independent verification of the numbers and a hard and fast attitude concerning the definitions. As is obvious, updating just the tax data has rewarded in entitlement terms communities with increased taxes which are below the 145 percent limitation. The nature of the data collection and verification process is clearly vital to a fair administration of the program. The data improvement program of the Office of Revenue Sharing (ORS) is a step in the right direction to ensure that the tax data is accurate; however, independent checks need to be made on local determination of taxes as adjusted by the Governments Division (Census Bureau) and then ORS to ensure that no inflation of figures is taking place. The incentive to exaggerate is rather obvious as is the incentive to report gross instead of net (of refunds) collections. State level verification is not uniform nor do all localities have local CPA certification. Leadership in this area could well involve encouraging the States to do audits and/or encouraging local CPA certification. For the budding revenue sharing research industry, some independent verification of taxes might prove useful. 36 REVENUE SHARING The second issue involves the definitions themselves. The use of taxes rather than revenues has been criticized by some and suggestions for broadening the definition or moving to revenues instead have been made. Several points are in order: First, the choice was a deliberate one made by the Congress; and second, it resulted from an exhaustive interstate and intrastate analysis of the effects of the two concepts. Finally, while it may affect rather dramatically certain municipalities financed by utility operations, I think we have to examine rather carefully the geo-equity implications of such a revenue structure as well as the more esoteric, but nonetheless important issue of whether or not that is a proper role for the local public sector to play. Without discussing these in any greater detail, let me indicate my own personal reservations on both these issues. D. Availability General revenue sharing has been with us for more than a year. It was planned to be a grant-in-aid program characterized by a small central administrative office and a maximum amount of openness about the manner in which it is administered. I think the first goal has been reached, but worry about the second. Computer tapes of the basic data are still not available to the research community, and there is no prospect of their being available in the foreseeable future. While printed copies of the data are available, and I should note parenthetically that this is a step in the right direction, the rather crucial interstate data has yet to be published. Most would seem to be available from various Census Bureau and Bureau of Economic Analysis sources; however, it strikes me as entirely reasonable that the interstate data be published as well as the intrastate data. There is, then, room for additional leadership to make the data more available. III. Formula Issues of General Revenue Sharing My remarks so far concern data-related issues of general revenue sharing. A few points about the formula are in order as well. A. Formula Availability While the data is available in hard copy form, researchers as well as recipient units remain mystified about how the formula actually, that is, computationally works. The program(s) that create the intrastate allocations have not been made public to date, and I think some leadership in this area would be desirable. The incentives in the legislation require local aware- ness which in turn requires they be informed of how it works. Making the program tape available was always contemplated by the Congress. It strikes me that time has now come to do this. B. Floor Considerations As I noted in San Juan at the December, 1973 National League of Cities Meetings, it is my understanding that the application of the 20 percent floor differs somewhat from the original algorithm; some quick research done indicated a $6 million annual difference for Chicago. This may explain the rather dramatic change in allocations to townships in the Midwest. I suggest to the interested a comparison for rural Midwest counties of percentage of funds going to townships as reported by ORS and those reported in, say, the Senate Finance Committee Report. IV. Conclusions From my perspective, general revenue sharing has been and will continue to be a huge success. The funds are being spent for locally identified priorities with increased citizen participation. The incidence of fradulent use of funds has been miniscule, and I conclude, far less than in those categories of grant-in-aid which were "outside" the local budgetary process. The statistical burden of general revenue sharing has been enormous, but one that has been well-shouldered to date. Creative leadership in the future is absolutely essential for continued success of the program and the subsequent passage of the special revenue sharing bills. It strikes me as crucial that the Federal Government work closely with the States to encourage them to better monitor local micro-statistical efforts. State interests are vital in this, but they need to be led if not pressed by the Federal Government. Correlative to this is a more significant commitment by the Federal Government to maintain the spirit of the New Federalism. Public availability of data and formula are part of this. Continued progress in the use of existing Federal information devices, such as the determined use of individual income tax for population estimation purposes, is another. Finally, we must anticipate our micro-data needs of the future and press for additional micro-data sources. As noted before, planned revamping of the welfare system can provide us with income and population data on a geodisaggregated basis if the Federal Government can be convinced of their utility. The rationalization of our Federal system of government, indeed our general capability of social problem solving, demands a fore- sighted and co-ordinated micro-data program. It might be quite useful for the American Statistical Association in concert with the Office of Management and Budget's Statistical Policy Division, the Census Bureau, and the Federal Statistics User's Conference to organize a continuing discussion of micro-data needs for the coming grant-in-aid consolidations. Formulas For Allocation of Funds Under the General Revenue Sharing Act of 1972 (PL92-512) prepared by Albert Mindlin, Chairman, Committee on Small-Area Statistics American Statistical Association General revenue sharing funds are allocated to the numerous units of government in the following sequence. 1 . The first allocation is made to States. Two formulas are employed. The formula yielding the largest amount for a State is usually used, subject to the overall imitation on funds. 2. The second allocation, a bookkeeping operation, is to an artificial geographic entity called a "county area." 3. The third allocation is to Indian tribes that exist in the "county area." 4. The fourth allocation is to county government within the "county area." 5. The fifth allocation is a bookkeeping operation setting aside aggregate funds for all townships that exist in the "county area." 6. The sixth allocation is to all units of local governments in the "county area" other than counties and townships. 7. The seventh allocation is to the townships. This report exhibits the formulas for each of these allocations. Although the formulas can be algebraically simplified, they are exhibited in the same form in which they are described in the Act. State Allocation P 2P i U 2U M SM T 2T R 2R K 2K = population of i tn State = U.S. population = urbanized population of i tn State = U.S. urbanized population = aggregate money income of State (census definition of income) = aggregate U.S. money income = aggregate net State and local taxes of i tn State = aggregate U.S. net State and local taxes = aggregate personal income of i tn State (BEA definition of income) = aggregate U.S. personal income = aggregate net individual income tax collection of i tn State = aggregate U.S. individual income tax collections 37 38 REVENUE SHARING 3-factor formula P Zi 1 ri i I p .A. 1 -! 1 P. • i R, M. 1 x i/ P. i This can be simplified to: I T i/R P. l 1 M i/P. l T i/R. p. i 1 M i/P. l 5-factor formula ? M i/lP. 1 i i P. U. ^i .15K T. (T i/R. j _ i. + —L. + M _ + i. + r"i SA~ V p ?^i^_ E . 15Ki y T (t 1/e ) 1 tj i M. _ i tJ i/P. i NOTE: Mj/P: = per capital money income of i tn State (i.e., per capita income as defined by Census Bureau, not BEA). jth i' K i = "general tax effort factor" of i tn State as defined in the Act. It can thus be seen that the 3-factor formula allocates funds to the States according to population, modified by the ratio of the State's "general tax effort" to its per capita money income. Since T:/Rj is a numerator figure while M-/P- is a denominator figure under both formulas, the amount available to a State tends to be increased by a large State and local tax effort relative to the State's aggregate personal income, and tends to be reduced by a large per capita money income. Allocation of funds to local governments The Act divides each State into geographic areas called "county areas." (These are usually, but not necessarily, coterminal with the area of a county government. For example, the independent cities of Virginia would qualify as separate "county areas.") The Act provides that one-third of the funds allocated to a State shall be held by the State government for its own use, while two-thirds shall be allocated to the various units of local government in each "county area." The first step in this process is to determine the amount allocated to each "county area." "County area" is not a unit of government. It is an intermediate step. Additional notation is required. FORMULAS 39 A:: = aggregated adjusted taxes of the j tn "county area" in the i tn State. Adjusted taxes constitute the aggregate taxes levied by all units of local government in the "county area", excluding those taxes allocable to education expenses. ("Units of local government" have a specific definition. In general they refer to units of general government. Special taxing authorities, e.g., Missouri fire protection districts, are not considered units of general government, and do not qualify for revenue sharing funds as described below.) j = general subscript for a "country area." Allocation to a "county area" The relative amount allocated to the ij tn "county area" is given by the following formula: P 1 P. . /P. 1 This formula is conceptually analogous to the 3-factor formula for allocation of funds to the States. It can be simplified to ij aggregate "county area ' adjusted taxes 'M. .. ) 2 ("county area per capita money income LI A it Zij V" 1 aggregate count y area adjusted taxes 'M. .. ) 2 ("county area" per capita money income) 2 ij county areas Allocation of "county area" funds among the units of local government in the "county area" If in any "county area" there is an Indian tribe (or Alaskan native village) then initially "county area" funds are allocated to such tribes on the basis of the population of the tribe relative to the population of the "county area." The remainder of the "county area" funds are then allocated first to county governments, then to other units of local government. Allocation to county government k = general subscript for a county government. The relative amount allocated to the ijk tn county government is given by the formula A iJ* A ij 40 REVENUE SHARING Aggregate allocation to township governments After allocation to county governments, the remaining amount is then reduced by an aggregate allocation to all township governments (if any) which are in the "county area." Let A:: t = aggregate adjusted taxes of all township governments in the "county area." The relative aggregate amount set aside for all township governments is then given by the formula: A i Jt A. This is an intermediate aggregate allocation. It does not specify the distribution of such allocation among the township governments. Allocation to other units of local government (i.e., other than counties and townships) After allocation of "county area" funds to county governments and aggregative township governments, the remaining funds are allocated to all other units of local government. Let h = general subscript for a unit of local government (other than county or township). The relative allocation to the ijh tn unit of local government is given by the following formula: ijh I ijh ijh M ijh A ijh M ijh M ij/p M. iJh/P ijh ij/ Pi . M ijh/P, ijh This formula is analogous to the 3-factor formula for States. As before, this formula can be simplified, and shown as ijh (M ) 2 ijh aggregate local adjusted taxes (local per capita money income I A. ijh aggregate local adjusted taxes M ijh/P Zaggreg (local ijh ( local units per capita money income FORMULAS 41 Allocation of funds to specific township governments The aggregate amount set aside for townships is distributed to the township governments by the same formula as that used for other units of local government. The subscript h then refers to a particular township government. The allocations, including separate terms in the various formulas, are subject to additional constraints and conditions. For example, each term in the 5-factor formula has a specific aggregate dollar maximum, so that adjustments must be made before the terms can be added. There are other maximum and minimum conditions on the various allocations and terms. Some of these (but not all) are explained in general on the attachment, which is a copy of pages from the booklet, "What is General Revenue Sharing?", Office of Revenue Sharing, Department of the Treasury, August, 1973. III. The Allocation Process Several steps are taken to determine the amounts to be allocated to each state. The Office of Revenue Sharing distributes the $6.05 billion allocated by Congress for the Fourth Entitlement Period (FY 1974) in the following manner: 1. $6.05 billion is allocated among the states according to the 3-factor Senate formula (population, tax effort and in- come); 2. $6.05 billion is also allocated among the states according to the 5-factor House formula (population, urbanized popula- tion, per capita income, state income tax collections, and tax effort); 3. The higher of the two amounts is selected for each state. Since the sum is greater than $6.05 billion, each amount is scaled down proportionately so that the total allocation equals $6.05 billion. 4. If either Alaska or Hawaii uses the 3-factor formula, its allocation is increased by the same percentage adjustment as applies to the base pay allowances of Federal government employees residing in those states (15 percent in Hawaii and 25 percent in Alaska). The next step is to allocate within each state, according to the following process: 1. One-third of the state's allocation is paid to the state government, and the remaining two-thirds is apportioned to units of local government within the state. 2. The amount to be allocated to units of local government is divided by the population of the state to establish the per capita entitlement for all governments within the state. 3. The local government amount is distributed to county areas (these are geographic areas, not governments) based upon the ratio that each county area bears to all county areas within the state according to the formula: population X tax effort X relative income. 4. If this calculation allocates to any county area an amount which on a per capita basis, exceeds 145 percent of the per capita entitlement calculated in step 2., its payment is reduced to the 145 percent level and the resulting surplus is shared proportionately by all the remaining unconstrained county areas within the state. 5. Similarly, if any county area is allocated less than 20 percent, on a per capita basis, of the amount calculated in step 2., its allocation is increased to the 20 percent level and the remaining deficit is taken proportionately from all the re- maining unconstrained county areas within the state. 6. Each county area allocation is then divided into four parts: a. An amount for Indian tribal governments or Alaskan native villages is then determined on the ratio of tribal/village population to the total population of the county area. b. From the remainder, a township allocation is determined on the basis of the ratio of all township adjusted taxes to the total adjusted taxes in the county. c. A county government share is determined similarly, on the basis of county government adjusted taxes. d. The remaining proportion is for the other units of local government. 7. Townships and other local governments are then allocated funds on the basis of the formula: population X tax effort X relative income. If a unit of government receives more than 145 percent on a per capita basis, it is adjusted to the 145 percent level. If a unit receives less than 20 percent, its allocation is increased to the lower of either the 20 percent level, or 50 percent of its adjusted taxes and transfers. Then, if any unit receives more than 50 percent of its adjusted taxes and transfers, its allocation is reduced to that level and the excess is given to the county government. 8. If the county government has been allocated more than 50 percent of its adjusted taxes and transfers, its allocation is reduced to that level, and the excess is returned to the state government. 9. If any allocation is less than $200, or any unit of local government waives its entitlement, those funds are allocated to the next higher level of government. 10. Finally, if the amounts allocated by the above procedure do not total 100 percent of the funds available for distribution, the appropriate adjustment is made to the entitlement figure in step 3. The process (step 3 through 8) is repeated until the amounts allocated total 100 percent of the funds available. Recent Developments in Estimation for Local Areas Introduction Paul S. Levy, University of Illinois Recent Developments in Estimation for Local Areas-This session is sponsored jointly by the Committee on Small-Area Statistics and the Social Statistics Section of the American Statistical Association. Statisticians and demographers working with health, eco- nomic, and social planners all are aware of the critical need for accurate small-area statistics, and have been frustrated by the rather primitive methodological tools available for obtaining estimates for local areas. It is only very recently that our survey statisticians and demographers have given some priority and have allocated resources to the development of some general methodology. We are fortunate in having at this session two speakers who have been involved in very large methodological studies on small-area estimation. They are Maria E. Gonzalez of the Statistical Research Division, U.S. Bureau of the Census who has spent several years investigating the properties of a type of estimation process known as synthetic estimation, and Eugene P. Ericksen of the Institute of Survey Research at Temple University who has been developing and evaluating new method- ology for estimating local populations. Our two discussants are Professor Richard M. Royall of the Department of Biostatistics, The Johns Hopkins University, who has written extensively on many aspects of sampling theory and Mr. Hyman B. Kaitz, now with C.S. & R. Associates and formerly Assistant Commissioner in charge of Current Employment Analysis of the Bureau of Labor Statistics. 45 Use and Evaluation of Synthetic Estimates Maria Elena Gonzalez, U.S. Bureau of the Census Definition of synthetic estimates An unbiased estimate is obtained from a sample survey for a large area; when this estimate is used to derive estimates for subareas, on the assumption that the small areas have the same characteristics as the larger area, we identify these estimates as synthetic estimates. For the smaller areas, the estimates are no longer unbiased. However, it is possible to measure an average mean square error (MSE) for this set of estimates. The simplest synthetic estimates are obtained by assuming that for the statistic of interest the mean value in the large area applies to each subarea directly; more refined estimates can be obtained by making this assumption for subgroups of the population. In the case when subgroups of the total are used, they should be nonoverlapping and exhaustive; the statistical estimates for the subgroups of the larger area are combined using independently known weights for the smaller area (e.g., as found at the time of the census) to obtain synthetic estimates for the smaller areas. One is interested in estimating a characteristic, X. Identify j subgroups in the population, which are nonoverlapping and exhaustive. From the larger area we obtain estimates, x.j for j=1 , 2,...., G. A synthetic estimate is desired for subarea i, which is within the larger area. From the latest census we have weights Pij, such that G Z P. . = 1. The synthetic estimate, x*, for characteristic X and subarea i is defined as G :.* = y P. .x. . i ^ n l d; This estimate associates the characteristic x ; of the larger area with each of the subareas i. Use of synthetic estimates Synthetic estimates are used primarily to develop small-area estimates when sample sizes are too small to give reliable results directly. Some examples of recent uses follow. 1) The National Center for Health Statistics has devel- oped synthetic State estimates of disability based on the Health Interview Survey data. National rates of disability for 78 subgroups defined in terms of age, sex, size of household, income, industry, etc., were obtained from the data collected in the Health Interview Survey. These disability rates were weighted by the corresponding population in individual States, from the 1960 Census of Population, to derive synthetic State estimates of disability. [1 ] 2) The Bureau of the Census has used synthetic estimates for the imputation of population for units reported as vacant in the 1970 Census of Population and Housing, but which were actually occupied. A subsample of the housing units reported as vacant in the 1970 Census of Population and Housing was selected and interviewers were sent to these units to determine how accurately that determination had been made. About 1 1 percent of the housing units reported as vacant were determined to have been occupied at the time of the census. Separate estimates of such error rates were prepared for twelve geo- graphic areas within the United States. Within each area the rate was applied to each enumeration district in the census and the applicable percentage of vacant units was converted to occupied units. The estimates of the error rates for areas such as cities, counties or States were synthetic estimates. [2] 3) In order to study the properties of synthetic estimates, an experiment was conducted to develop unbiased and synthetic estimates of unemployment for SMSA's for monthly, quarterly and annual estimates based on the Current Population Survey (CPS) data. A comparison of the reliability of the two types of estimates revealed that for monthly data the synthetic estimates were preferable, while for annual data the unbiased estimates were preferable; for the quarterly data, the two were of about equal reliability. [2] 4) In the 1960 Census of Housing enumerators were instructed to rate the physical condition of each housing unit into one of three categories: "sound," "deteriorating," or "dilapidated." One important purpose of this was to provide 46 SYNTHETIC ESTIMATES 47 data on substandard housing defined by Federal and local housing agencies as comprising units lacking complete plumbing facilities plus units which were dilapidated but had all plumbing facilities. In the 1970 Census of Housing information was again obtained about plumbing, but synthetic methods were used to develop estimates of housing units which were dilapidated with all plumbing facilities (DWAPF). To obtain these estimates census data on housing units with all plumbing facilities were multiplied by estimated proportions of dilapidated housing units which had all plumbing facilites, as derived from a postcensus survey, Components of Inventory Change (CINCH). 1 From CINCH, estimates of DWAPF housing units were obtained for specified subgroups for 15 selected large SMSA's and for 4 balance of regions of the United States. Synthetic estimates for the smaller areas within these 19 geographic areas were derived using the corresponding set of DWAPF proportions. Evaluation of synthetic estimates Synthetic estimates are biased; to evaluate their reliability one can use the MSE, which can be expressed as the sum of the variance and the square of the bias: G MSE-(x.*) = X P 2 . rr° +(x*-x ) 2 ( 2 ) j-l 1J .+ (x.*-x ) J i i The average MSE can be estimated by using the following approximation: M a = - v (x.* - I P x ) M * i ,_, ii i/i i=l where: ij j = l n M U - - Z I P 2 . <1-2* ±1 > °l W (4) M . . ij 1J ii L J i=i J-l is the estimated statistic from the sample for sub- J class j and area i, : _ is the sample estimate of the proportion of the total 1 J for the j-th subclass that is in the i-th subarea, that is M f. = n. U U/ i=1 U Y n. . , and .=1 1J is the sampling variance of estimate x where: 2 o x. j 1 is the sampling variance of estimate is the "true value" of the statistic for subarea i; and is the expected value of the synthetic estimate for subarea i. The estimate given in formula 2 assumes that: a. the Pjj's are fixed and measured without error; and b. the cov (x j, x ^ ) = 0, for j / k. In general, the values of Xj are not known and consequently the MSE of an individual synthetic estimate cannot be calcu- lated for a particular area "i." However, if we establish M subareas within the survey population, the average MSE of the synthetic estimate over the M subareas (which may be unequal in size) can be estimated from the sample. Let M ~ I (X * - X. ) : i=l = a (3) 'The change in procedure in estimating DWAPF housing units was necessary because a majority of housing units in the 1970 Census of Housing were enumerated by a mail-out, mail-back procedure; in addition, studies of these data for 1960 indicated that statistics based on enumerator ratings are highly unreliable. The interpretation of the square root of the mean square error is not altogether clear. Probability statements which can be made using the standard error for an unbiased estimate do not necessarily hold when the estimates are biased and the root mean square error is used as a measure of reliability. To try to understand the situation, an empirical study of the root mean square errors was made, using 1960 housing data. Synthetic estimates for a census were compared with actual measurements of the item for the same census to obtain an estimate of the bias. For various groupings of areas we then computed an average root mean square error (RMSE); this estimate together with the distribution of the biases was then used to compare an empirical distribution of the biases of synthetic estimates with the normal distribution. As part of the publication of the 1970 Census of Housing, data on housing units dilapidated with all plumbing facilities collected in the 1960 Census of Housing were available for comparison with a set of synthetic housing estimates of DWAPF derived from the same data. The average mean square error for a set of M areas is given by 1 M Average MSE = — I (* ± i=l x *) 2 i (5) where: Xj is the census estimate for area i. The use of formula 5 to estimate the average MSE assumes that the second term of formula 4 is negligible. This assumption is reasonable for large are as. The square root of the average MSE gives the estimate for the RMSE. 48 ESTIMATION FOR LOCAL AREAS For all States we have two estimates of dilapidated housing units with all plumbing facilities in 1960; the 25-percent census estimate and a synthetic estimate based on a particular set of subgroups. The difference between these two estimates will be used as an estimate of the bias of the synthetic estimation procedure. Table 1 shows estimates of the proportion of the set of synthetic estimates for States with a relative bias within specified values. The relative bias for an area is defined as the difference between the synthetic estimate and the census estimate divided by the synthetic estimate. 2 2 Tlie synthetic estimate was used as denominator, instead of the reported estimate, because in 1970 the synthetic estimate was the only one available. Table 1 . Distribution of Relative Biases of Synthetic Estimates for States Number of areas Proportion with relative biases State estimate 0-9$ 10-19$ 20-29$ 30-49$ 50$+ 2,500-4,999 10,000-19,999 20,000 or more. . . 7 6 13 16 8 50 0.14 0.50 0.23 0.38 0.38 0.32 0.29 0.17 0.46 0.38 0.38 0.36 0.29 0.17 0.15 0.19 0.25 0.20 0.14 0.17 0.08 0.06 0.00 0.08 0.14 0.00 0.08 0.00 0.00 0.04 This table shows that the proportion of estimates with large relative biases diminishes as the size of the synthetic estimate increases. For example, for synthetic estimates between 1,000 and 2,499 DWAPF housing units, 57 percent have relative biases of at least 20 percent; however, for synthetic estimates of over 10,000 units only about 25 percent have relative biases greater than 20 percent. When we consider the State synthetic estimates for all States, we note that 32 percent have relative biases of 20 percent or more, 1 2 percent have relative biases of 30 percent or more and 4 percent have relative biases of 50 percent or more. The average number of DWAPF housing units for States is about 1 3,000; the estimated av erage r oot mean square error is about 2,500; the ratio of the RMSE divided by the average size of State synthetic estimates of DWAPF housing units is 0.19. A high vari ability of the synthetic estimate is shown by the fact that the RMSE divided by the mean is about 20 percent. This shows that the synthetic estimates obtained do not account for a large part of the variability among areas. The synthetic estimates of housing units DWAPF are computed using a particular set of subgroups, defined in terms of tenure, race of head of household and other characteristics related to the quality of housing unit. The use of other subgroups would produce a different set of synthetic estimates. From the point of view of ascertaining whether the average root mean square error can be used to make probability statements the results are more encouraging. Table 2 gives some Table 2. Comparison of Empirical Distribution of the Biases of State Synthetic Estimates of Dilapidated Housing Units with All Plumbing Facilities with the Theoretical Normal Distribution Distribution of bias of State estimates (n=50) Multiple of standard Normal probability Total Inside SMSA 's Outside SMSA's error (a ) Aver = 12,970 Aver = 8,170 Aver = 4,800 RMSE = 2,490 RMSE = 1,540 RMSE = 1,340 0.50 38.0$ 48$ 50$ 62$ 0.75 55.0 62 62 68 1.00 68.0 74 68 74 1.25 79.0 86 82 84 1.50 87.0 88 90 88 1.75 92.0 88 90 88 2.00 95.0 92 94 94 2.25 97.6 94 94 98 2.50 98.8 96 96 98 3.00 99.7 100 100 98 SYNTHETIC ESTIMATES 49 comparisons of the distribution of the difference between State synthetic estimates of DWAPF housing units for 1960 and the estimates reported in the 1 960 census. The first two columns of the table show the expected percentage of the normal distribu- tion at different multiples of the standard error (a). For example, 95 percent of the normal distribution is expected within two standard errors of the mean. The empirical distribu- tions of the biases of the synthetic State estimates of DWAPF housing units are given in columns 3, 4, and 5. For example, 48 percent of the biases for e stimate s of total for States are less than one-half the estimated RMSE. The empirical distributions of the biases of synthetic State estimates are closer to the mean (on the average) for values within one standard error of the mean, than expected for the normal distribution. For example, for half a standard error the normal distribution expects to cover about 38 percent of the distribution; for State totals, the empirical distribution actually includes 48 percent of the distribution; for estimated units within SMSA's, the empirical distribution includes 50 percent of the distribution and for units outside SMSA's the empirical distribution includes 62 percent of the distribution. However, for values which are more than two standard errors from the mean, the empirical results are reversed: the frequency of synthetic estimates with biases more than two standard errors from the mean is greater than expected for normal distributions; for State synthetic estimates about 8 percent had biases which differed by more than two standard errors from the mean. That is, on the average there are more outliers for synthetic estimates than would be expected for a normal distribution. Table 3 shows empirical distributions of the biases of the estimates of DWAPF housing units for non-Negro renters in counties within SMSA's. The distribution was computed sep- arately depending on the magnitude of the estimate of DWAPF housing: less than 100, 100 to 499, and 500 or more units. It is possible to carry out the analysis separately for the three groups through the use of formula 5 for the average root mean square error. The results reveal a similar pattern to the results for State estimates given in table 2. For values within one standard error the empirical distribution gives conservative estimates of the probability of occurrence, except for the distribution of DWAPF housing units with 500 or more units. However, for values at three standard errors we find more outliers than expected for the normal distribution. For example, for syn- thetic DWAPF estimates less than 100, the normal distribution expects only 0.3 percent of the cases to be further than three standard errors, but we find that 2.5 percent of the values have biases larger than three times the average root mean square error. Table 3. Comparison of Empirical Distribution of the Biases of Synthetic Estimates for Non- Negro Renters in Counties within SMSA's of Dilapidated Housing Units with All Plumbing Facilities with the Theoretical Normal Distribution Distribution of bias for non-Negro renters Multiple of Normal probability in counties within SMSA's standard error (a) DWAPF <100 RMSE = 39 DWAPF (100-499) RMSE = 125 DWAPF ( 500+ ) RMSE = 300 n = 160 n = 219 n = 84 0.50 38.0/ 50.0% 49.0/ 37.0% 0.75 55.0 66.0 66.0 52.0 1.00 68.0 84.0 79.0 64.0 1.25 79.0 89.0 85.0 75.0 1.50 87.0 93.0 89.0 85.0 1.75 92.0 94.3 92.2 89.2 2.00 95.0 95.6 94.5 95.2 2.25 97.6 96.2 96.8 98.8 2.50 98.8 96.2 97.7 98.8 3.00 99.7 97.5 98.6 98.8 50 ESTIMATION FOR LOCAL AREAS Conclusions Census data allow us to compute synthetic estimates and to compare them directly to the census estimates. Therefore, the biases of synthetic estimates can be obtained and their distribution analyzed directly. The results presented comparing 1960 estimates of dilapi- dated housing units with all plumbing facilities with syntheti- cally derived estimates show that the synthetic estimates are highly variable, but that the distribution of their biases is not too far from normal. The analysis presented is based on a particular set of synthetic estimates; alternative sets using other variables should be investigated in order to be able to select the subgroups which account for a large proportion of the variability of the local area estimates, with an aim toward improving local area estimates. The results presented here, based on a particular set of synthetic estimates, may not necessarily generalize to possible alternative sets of synthetic estimates. References [1] "Synthetic Estimates of Disability," published in 1968 by the National Center for Health Statistics, PHS publication No. 1759. [2] A more detailed discussion of this project is available in "Estimation of the Error of Synthetic Estimates," by Maria Elena Gonzalez and Joseph Waksberg, presented at the first meeting of the International Association of Survey Statisticians, Vienna, Austria, August 18-25, 1973. [3] Appendix 1 of "Estimation of the Error of Synthetic Estimates," by Maria Elena Gonzalez and Joseph Waksberg gives the derivation of the approximation given in formula 4. Developments in Statistical Estimation for Local Areas Eugene P. Ericksen, Temple University Introduction The regression-sample data method of postcensal estimation is a procedure by which one can combine sample survey data with symptomatic information to obtain local estimates of the criterion variable being measured by the survey data. This method has been tested extensively using population growth as the criterion (Ericksen, 1973a, 1973b), first, for the period beginning in 1960 and ending in 1964 to 1967, and then more extensively for 1960 through 1970. The steps of the procedure in the latter test were as follows: a. Sample estimates of population growth were obtained for the primary sampling units selected into the national sample of the Current Population Survey. These 1970 estimates of current population were divided by the corresponding 1960 census populations giving sample estimates of 1960 to 1970 population growth. b. Symptomatic indicators, in this case 1970/1960 ratios of births, deaths, and school enrollment, were com- piled for the sample psu's (primary sampling units) and a multiple regression equation was computed using the sample estimates of population growth as the depen- dent variable. A second equation was then computed using the series of ratio-correlation estimates calculated at the Population Division of the Bureau of the Census as a fourth symptomatic indicator. c. Values of the symptomatic indicators for counties were substituted into the regression equations and estimates were made of the 1960 to 1970 population growth. This step was carried out for 2,586 counties in 42 States. Corresponding estimates for these counties were made at the Population Division using four standard demographic techniques which have traditionally been used to estimate population growth. Of these techniques, the ratio-correlation technique was the most accurate. Little was gained from averaging estimates of two or more standard techniques. The regression estimates produced by our combination of sample data and symptomatic information were more accurate than those of any single or combination of standard techniques. This was particularly true when the series of ratio-correlation estimates was added as a fourth symptomatic indicator. There were moderate reductions in the mean error, but the greatest gain was in the reduction of the number of large errors, which was over 20 percent. These results are presented in table 1. Table 1. Relative Accuracy of Standard and Regression-Sample Data Estimates of Population Growth, 1960 to 1970, for 2,586 Counties in the United States Procedure Mean error 1 Number of counties with error 10 percent or greater Vital Rates 7.4 ,7-2 5.9 4.6 4.7 4.4 4.2 673 Component Method II 645 Composite 407 Ratio-Correlation Component Method II, Composite, Ratio-Correlation, Averaged 2 Regression-Sample Data, Three Symptomatic Indicators 3 264 249 220 Regression-Sample Data, Four Symptomatic Indicators 3 194 All estimates were multiplied by an appropriate constant in order to sum to a separately estimated 42-State total. This was the most accurate combination of the four standard techniques. The 3-variable equation used was Y .045X, +.214X, + .745X, 2 3 * .158 + .218Xj + .142X 2 + .520X 3 . The 4-variable equation used was Y = .058 - .097X + Source of Standard Estimates and Estimate of 42-State Total: U.S. Bureau of the Census, Current Population Reports, Series P-26, No. 21, "Federal-State Cooperative Program for Local Population Estimates: Test Results - April 1, 1970," Washington, D.C.: Government Printing Office, 1973. 51 52 ESTIMATION FOR LOCAL AREAS Some of the prominent features of our new method of postcensal estimates are the following: a. Estimates of population growth have been shown to be more accurate. Part of the reason for this gain is that it is not necessary to make any assumptions concerning the nature of relationships beyond those of least squares linear regression. One of the difficulties of the ratio-correlation technique, for example, is the assump- tion of the continuance of past relationships. b. Other series of estimates can be incorporated as symptomatic indicators. By including the series of ratio-correlation estimates as a symptomatic indicator, we have a way of correcting for the bias of ratio- correlation which arises from assuming the contin- uance of past relationships. c. There is a procedure by which the mean squared error of the regression estimates can be calculated. Given this facility for measuring error, we can systematically test various combinations of symptomatic indicators to determine the composition of the optimal set. Because of the presence of the within-psu sampling error, this does not necessarily include all available symptomatic indicators. Current Activity at the Bureau of the Census A project is currently under way at the Bureau of the Census to compute yearly estimates of population growth since 1970 by our regression-sample data method. A determined effort is being made to obtain symptomatic data for all counties in each of the 50 States and the District of Columbia. It now appears that births, deaths, and school enrollment will be available, but with some time lag, for counties in all but a small handful of States. Additional data on automobile registrations will be available for some States and it is also expected that data on income tax exemptions will be available for all States (Zitter and Word, 1973). Substantial gains are anticipated from the use of tax records, even outside our regression-sample data format. In view of the changing relationships among variables, and the possibility that other symptomatic indicators will become available, the following instructions are pertinent to potential users of the regression-sample data technique: a. Applications of and experimentation with the ratio- correlation technique have shown conclusively that relationships among a given set of variables can be expected to change over time. We have shown that the series of regression-sample data estimates were rela- tively accurate when computed over a given 10-year period for particular sets of three and four sympto- matic variables. However, the accuracy of the regres- sion-sample estimates relative to those of other techniques could change for a shorter estimation period beginning in 1970. It is also possible that the most accurate regression-sample data estimates would be computed with a different set of symptomatic indicators in this period. We can test this possibility by inspecting the mean squared error of the regression estimates and the correlations of the various indicators with the sample estimates of population growth. b. In the absence of correlations between the sampling error and the value of the symptomatic indicators, the estimated regression equation using the sample esti- mates of population growth as the dependent variable is an unbiased estimate of the regression equation which would be obtained using census tabulations of population growth if they were available. However, the presence of the within-psu sampling error will lower the observed values of the correlation coefficients. Low observed values of the correlation coefficients do not necessarily mean that the errors of the regression estimates will be large. c. There are some unsolved problems regarding the inference from a sample of psu's to a universe of counties. The mean squared error of the regression estimates refers to the accuracy of estimates for psu's when the units of interest may be counties. To the extent that counties are different from psu's, reduc- tions in the mean squared error for psu's may not improve the accuracy of estimates for counties. A second unresolved problem has to do with specifica- tion errors arising from the distribution of the within- psu sampling errors. If the size and direction of these errors vary systematically with values of the sympto- matic indicators, the assumptions of linear regression may not be met. We have found this to be a minor problem in our application that resulted in larger errors for units with extreme growth rates, but the problem could be more important in other applications. The Mean Squared Error We have shown elsewhere (Ericksen, 1973a, 1973b) that the mean squared error of the regression-sample data estimates can be expressed by the formula: E(Y-Y)MY-Y) (n-p-1)a u 2 (p + 1)a y 2 (3.1) where: a u 2 = the between-psu variance unexplained by the indicators, 0" v 2 = the within-psu variance, n = the number of psu's in the sample, and p = the number of symptomatic indicators. When n is large relative to p, the mean squared error is determined by (1) o 2 which decreases when new symptomatic indicators are added, and (2) the within-psu component of error which increases when indicators are added. If there were no within-psu component of error, optimal results would be obtained by maximizing p, i.e., by utilizing all available symptomatic information. We have found in our applications, however, that the presence of within-psu sampling variability has often meant that the optimal set of symptomatic indicators did not include all that were available. STATISTICAL ESTIMATION 5 3 In the test of 2,586 counties, there were seven symptomatic indicators available: births, deaths, school enrollment, and the four standard estimates. As shown in table 2, where the ratios of the 1970 to the 1960 decennial census populations were the dependent variable, gains in the accuracy of regression estimates for psu's were obtained by increasing the number of sympto- matic indicators from four to seven. However, in the more realistic application, when the within-psu component of error was present, the increase from four to seven indicators actually brought about an increase in the errors. The mean error of the 2,586 county estimates increased from 4.2 percent to 4.7 percent. A similar result was obtained when 6 variables, with 51 observations (1 for each State and the District of Columbia) were available. Table 2. Mean Errors Obtained With Various Sets of Symptomatic Indicators Number of symptomatic indicators 2 Mean Percentage Error 1 Dependent variable 1 970 census/1 960 census Dependent variable 1 970 CPS/1 960 census Units of estimates are 444 primary sampling units 3 2 2.83 3.20 4 2 2.60 2.92 7 2 2.11 3.24 Units of estimates are 50 States and District of Columbia 3 3 1.22 1.64 4 3 1.08 2.16 6 3 1.07 3.91 Mean percentage error, comparing regression estimate with Census Bureau tabulation. "Dependent variable" is that used to compute regression equation. 2 Set of three indicators included births, deaths, and school enrollment. The fourth indicator was the ratio-correlation estimate, and indicators five through seven were the composite, component method II, and vital rates estimates. 3 Set of three indicators included births, school enrollment, and work force. The fourth indicator was deaths, and indicators five and six were automobile registrations and income tax returns. The fact that the optimal set of indicators included four variables was due to the nature of the structural relationships and the size of the within-psu variance. We have evidence that these change over time, as shown in table 3. In particular, for the Current Population Survey (CPS) sample, the within-psu variance increased. This is because the CPS sample was based on the 1960 census, and that patterns of subsequent growth were uneven, leading to variation in the size of sample segments within psu's. This trend leads us to expect that more sympto- matic indicators should be used in shorter time periods. On the other hand, the relationships among the variables appear to become stronger as time passes. In spite of the increasing within-psu variability which dampens the observed correlations, these observed correlations grew larger from 1963 through 1967. In shorter periods, changes in population size, as well as in the symptomatic indicators, appear to be due more to random fluctuations. As times passes, changes in the variables are larger, and the relationships among these changes more systematic. This leads to the contrary expectation that the optimal set of indicators would be smaller for a shorter time period. To determine the optimal set of indicators, we must estimate the mean squared error in each estimating situation. Because the true values of the criterion variable are unob- served, the mean squared error cannot be estimated directly. To obtain equation (3.1), we must first compute the mean of the squared differences between the regression estimates and the sample estimates for the sample psu's and then subtract an allowance for the within-psu sampling error. The mean squared difference between the regression and sample estimates can be expressed by the formula: E(Y -Y)MY -Y)_(n-p-1)(a u 2 + a v 2 ) (3.2) To obtain (3.1) we need to subtract the term (n - 2p - 2)o 2 /n. In order to obtain a good estimate of the mean squared error, we clearly need to have a good estimate of a y 2 . 54 ESTIMATION FOR LOCAL AREAS Table 3. Values of Estimated Within-Psu Variance of Population Growth and Coefficients of Determination, 1963 through 1967 Year Within-psu variance 1 Coefficient of determination (R 2 ) 2 1963 1964 1965 1966 1967 .0253 .0378 .0383 .0458 .0473 .016 .021 .085 .117 .264 Computed as squared difference between random half-samples defined by rotation group. Three symptomatic indicators were births, deaths, and school enrollment in each case. When we reported earlier results (Ericksen, 1973b), we did not feel that a good estimate of a y 2 was available. We had computed half-samples defined by the eight rotation groups of the CPS (U.S. Bureau of the Census, 1963) and had overesti- mated the mean error of the sample estimates for psu's. This is because sample segments within the CPS sample had not been placed equally into rotation groups within individual psu's. However, when the half-samples were formed on the basis of sample segments without regard to rotation group, a better estimate was obtained. The derivation of equation (3.2) depends on the values of (1) the sampling error and (2) the structural errors of regression, along with the sampling errors being unrelated to the symptomatic indicators. Our technique for estimating the mean squared error is particularly sensitive to these specification errors, as the following illustration shows. The pratical question we faced in the 1970 test was whether or not improvements in accuracy over that given by the ratio-correlation technique would be obtained by adding births, deaths, and school enrollment as symptomatic indicators in a regression equation. We found that the ratio-correlation esti- mates accounted for 92.7 percent of the variance of the actual 1970/1960 ratios of population of the sample psu's. Adding the three symptomatic indicators, the coefficient of determination, R 2 was increased to .951, a clear increase in the explained and reduction in the unexplained variance. However, the increase in the explained variance of the sample estimates of 1960 to 1970 population growth obtained by adding the three symptomatic indicators to ratio-correlation was much smaller, from 41.7 percent to 43.0 percent. This was due to the presence of the v/ithin-psu error which is not reduced by adding symptomatic information. The observed variance of the distribution of sample estimates before regression was .0438. Using the series of ratio-correlation estimates as a single symptomatic indicator, the mean squared difference of the regression and sample estimates as expressed by equation (3.2) was .0255. This was reduced to .0250 when the number of symptomatic indicators was in- creased from one to four. Our estimate of the within-psu variance is O 2 = .0253. Subtracting the allowance for this component of error, our final estimates of the mean squared error are .0004 where the ratio-correlation estimate is a single indicator and .0001 with four indicators. This is a very small difference considering the size of the within-psu variance and the mean squared difference between the regression and sample estimates. A small fluctuation could have seriously altered the observed results. When the number of symptomatic indicators was increased to seven, the coefficient of determination was R 2 = .432, the mean of the squared differences was .0249, and the final estimate of the mean squared error, .0002. These differ- ences are so small that one may be on safer, although less scientific, grounds simply to observe that the increase in R 2 from .417 to .428 is large enough to produce a good reduction in error while guessing that the further increase to R 2 = .432 is not, given the increase in the number of symptomatic indi- cators. Some of the difficulties in estimating CT y 2 arise from the fact that the within-psu sampling error is positively correlated to the growth rate, and hence to the values of the symptomatic indicators. The correlation between the actual error of the CPS estimate and the estimated within-psu variance is +.45. This affects the estimate both of a y 2 and the way we obtain an estimate of equation (3.1) from equation (3.2). A second source of error is the correlation between the within-psu error and the growth rate, which is +.06. This introduces curvilinearity, since the sample estimates of the fastest growing areas tend to be too large and those of the slowest growing too small, thus biasing the estimation of regression coefficients. One result of this was that estimates of areas with extreme growth rates had larger errors. This particular problem is covered in the literature on econometrics where the usual solution is to apply a transfor- mation. We have attempted several such solutions, but have yet to find a transformation which allows us to reduce the errors of the extreme cases without increasing the errors of the majority of cases which have moderate values. One obvious procedure for reducing the mean squared error of the regression estimates is to reduce the within-psu variance. This could be done by improving the within-psu sample design, or, as we will attempt to do, by introducing more sample data. In our program at the Census Bureau, we plan to eventually request tabulations from other government surveys such as the Health Interview Survey and the National Crime Survey. This will reduce o 2 where the psu's in the various surveys are the same and reduce the ratio (p + l)/n and therefore the within-psu component of error in equation (3.1) in cases where the psu's are different. This necessarily reduces the errors of the primary sampling units, but the effects on county estimates are uncertain. To illustrate this point, when the regression equation with three symptomatic indicators, births, deaths, and school enrollment, was recomputed using the 1970/1960 census population ratios as the dependent variable, i.e., setting a 2 equal to zero, the STATISTICAL ESTIMATION 55 mean error of the psu estimates was 2.8 percent. This compares to the mean error of 3.2 percent when the CPS estimates were the dependent variable. The difference between 2.8 and 3.2 percent was due to the within-psu error. However, when the two equations were used to make county estimates, the mean error was 4.4 percent in both cases. The Census Bureau's ratio equation, computed without the within-psu error, had done a better job of making psu estimates, but the transition from psu's to counties had become more difficult. When the distribution of errors was broken down by size of the 1970 county population, it was found that use of the decennial census ratios in place of the CPS estimates had reduced the mean error for all categories of counties with population greater than 25,000, but that the mean error had increased among counties smaller than 25,000. Counties in this last category were the majority of all counties but were least similar to the CPS sample psu's which usually consisted of combinations of counties picked with probabilities proportional to the size of the total population. New Strategies and Plans Given the limited gains obtained from reducing the within- psu component of error, and our lack of success in finding suitable transformations to reduce errors, the most promising approach to reducing our errors appears to be the introduction of new symptomatic information. One variable which has been shown to reduce errors is automobile registrations. Data were available in the 1970 test for 2,223 counties in 32 States. A five-variable regression equation, also including births, deaths, school enrollment, and the ratio-correlation estimate was computed and county estimates made. The mean error of these estimates was 3.8 percent and 122 errors were greater than 10 percent. The corresponding figures for this set of counties for the four-variable regression equation omitting automobile regi- strations were 4.1 percent with 148 large errors and, for the standard series of ratio-correlation estimates, 4.5 percent with 220 large errors. Another promising, but as yet untested, variable is the number of exemptions on income tax returns. Changes in address of persons listed on income tax forms are to be used to estimate net migration and when added to recorded natural increase, could give extremely accurate estimates of population growth. It is quite possible that these estimates would be sufficiently accurate in themselves so that little gain would be obtained by computing regression-sample data estimates. But it is more likely that some bias will be introduced because of the characteristics of persons not listed on tax forms or whose likelihood of being listed on a form varies at point of origin and destination. In such a case, this bias could be corrected by using the tax estimate as a symptomatic indicator in a regression equation possibly including other symptomatic indicators with sample data as the dependent variable. Finally, we have made plans to attempt to estimate other variables such as racial composition, unemployment, and median family income. Although births and deaths are available by race in many counties, the chief barrier faced here is the lack of symptomatic information. Data on wages and work force appear to be available in metropolitan areas, but we are still searching for symptomatic data available on a national basis. If such data can be found, we can combine our symptomatic information and sample data with estimates which can be generated by other means. One such series would be the synthetic estimates being discussed in this session. Bibliography Ericksen, Eugene P. "A Method for Combining Sample Survey Data and Symptomatic Indicators to Obtain Population Estimates for Local Areas." Demography 10 (1973a): 137-160. Ericksen, Eugene P. "A Regression Method for Estimating Population Changes of Local Areas." 1973b, in press. U.S. Bureau of the Census. The Current Population Survey-A Report on Methodology. Technical Report no. 7. Washing- ton, D.C.: U.S. Government Printing Office, 1963. Zitter, Meyer, and David Word. "Use of Administrative Records for Small-Area Population Estimates." Presented at the Annual Meeting of the Population Association of America, 1973. Acknowledgment The research upon which this paper is based was carried out in cooperation with the Bureau of the Census, in particular with Benjamin J. Tepping of the Research Center for Measurement Methods and with the Population Division, who supplied all the symptomatic information used in computing the regression equations as well as special tabulations concerning the distribu- tions of errors of the ratio-correlation and other standard estimates. Special thanks are due to Lori Kessler and Harris Miller without whose assistance at Temple University this research could not have been carried out. The findings, recommendations, and conclusions in this paper are the sole responsibility of the author and are not necessarily endorsed by the U.S. Government. The data in this paper are the result of tax-supported research and, as such, are not copyrightable. The data may be freely reprinted with the customary crediting of the source. 56 Discussion of Papers by Gonzalez and Ericksen Richard Royal I The Johns Hopkins University Both of these papers are good examples of the process of developing estimators using conventional finite population sampling theory. We can pick out three important stages in this process: 1. Assuming observations on certain variables are available, scratch your head and write down an estimate which has some intuitive appeal. 2. Try to get a handle on bias and variance. (Having done this a few times and produced a few estimates, compare their mean square errors. Find one estimate that is better than another under certain assumptions about population parameters.) 3. Get a real population and try out the estimates to see which works better under various realistic conditions. After, or along with, these three basic steps comes the secondary problem of measuring the uncertainty in an estimate. This usually boils down to finding a nearly-unbiased estimate of an approximation to the variance or mean square error. Unfortunately, these variance estimates rarely have the "face validity" or obvious reasonableness of the original statistic. For example, the synthetic estimates are in a gross sense reasonable. They obviously won't give really precise estimates, but they will be, if not in the right ballpark, at least in the right city. The variance estimate, on the other hand, might not even be on the right planet— a negative variance estimate might be reasonably described as "lost in space." I would like to see a different approach used, and I think the problem at hand, estimation for small areas, is one in which this approach would yield different and better results than the conventional one, particularly with regard to providing esti- mates of mean square errors to use as measures of uncertainty. This approach would begin not with an estimate, but with an attempt to express the basic relationships among the relevant variables through a probabilistic model. The model would then be used to generate estimates, provide a framework for comparing estimates, and to provide estimates of standard errors. Often the conventional intuitive estimates are optimal or nearly so under a simple probability model, but sometimes the model suggests practical improvements, especially in the conven- tional measures of uncertainty. Varying the model can give valuable insight into the robustness of estimators. This general approach has been called "the prediction approach" because, when viewed in the context of (super-population) probability models, many finite population inference problems are mathe- matically equivalent to classical prediction problems. "The prediction approach" actually has many facets— simple linear least-squares, esoteric fiducial, and full-blown Bayesian predic- tion techniques are only some of those available. What would be the results of applying the prediction approach (least-squares variety) to the present problem? Two important general results I would expect are: 1. New estimators and new variance estimators for the old ones. 2. New insight into relationships among estimates already proposed and increased understanding of their strengths and weaknesses. Specifically ... I don't know what results would be obtained. The work has not, to my knowledge, been done. But some relevant comments can be made. The "ratio-correlation" method and the "regression-sample data" method aren't so much two different methods as two different estimates, each more or less appropriate under its own prediction model. Although the two models do employ slightly different functions of births, etc. as regressors, the most important differences between these two estimates come not from different assumptions concerning the relationships among the relevant variables, but from different assumptions about available data. The ratio-correlation method is not allowed to use the sample data, while the regression method employs only data from the sample and the most recent census, ignoring the previous census. In both models, the total for a local area at one time is represented as a multiple of the total at an earlier time plus an error whose variance is proportional to the square of the earlier total. (We might ask whether a different error-variance might be more approporate. If it is, this would suggest different estimates.) The multiplicative factor for a given area is a function of various bits of data concerning births, deaths, number of school children, etc. in that area. In this factor are certain coefficients which change over time. The "ratio- correlation method" uses estimates of out-of-date coefficients, while the "regression-sample-data" method uses less precise estimates of more timely coefficients. When the "ratio-correlation" estimate is used as a "sympto- matic indicator" in the "regression-sample-data" estimate, we are, in effect, using a particular linear combination of estimates 57 58 ESTIMATION FOR LOCAL AREAS of the "old" coefficients and the "new." I think a formal model, in which coefficients for one time interval are stochas- tically related to those for an earlier interval, would be quite useful in evaluating this and other estimates based on all the data, from both censuses as well as the sample. In much the same way, the choice between direct estimation and imputation in the synthetic estimation paper is really the choice between a high-variance estimate of a directly relevant parameter and a low-variance estimate of a different quantity. The choice need not be made— surely a combination of the two is better than either taken alone. A probability model can express the relationships whose existance makes the whole notion of "imputation" reasonable. Such a model would generate (via standard linear prediction techniques) statistics which would give proper weight to both direct and imputed estimates. I think, however, that one of the possibilities suggested by Gonzales and Waksberg in their Vienna paper is more promising— before really good local area estimates are produced, the synthetic estimation approach must move towards Ericksen's in making greater use of available local area variables. Discussion of Papers by Gonzalez and Ericksen Hyman B. Kaitz CSR Associates, Alexandria, Va. Both of these papers represent progress in the development of procedures for estimating local area data. Nevertheless it is work in progress; we clearly have some distance to go in attaining fully acceptable methodologies. Both papers examine the accuracy of their techniques in terms of mean square errors of all the local area estimates, and in terms of the percentage of areas whose estimates deviate by more than a certain percentage, say 10 percent, from the known criterion values. It is desirable and relevant, of course, that such measures of dispersion or accuracy, be used in judging these various techniques. There is a particular problem, however, which arises in this connection. Estimates are produced for areas which are individ- ually identified. For example, the method may yield a specific estimate for Altoona. Is Altoona interested in knowing whether its estimate comes from a body of area estimates with a satisfactory mean square error, or is it interested in knowing how good the Altoona estimate is? I think the latter is more likely to be the case. In the absence of any alternative estimates, Altoona can perhaps become reconciled to its estimate, but it may not. It is interesting in this connection to see what has happened to local area population counts from the 1970 census. Even though these are presumably "complete" and official government counts, they have been subjected to strong criticism by various local interest groups who have said that the presumed population undercount should be officially allocated to local areas differentially (black-interest groups), or that the Census Bureau should use more accurate enumeration techniques for the Spanish-Americans. When the data at issue are statistical estimates rather than official counts, there may even be more opportunity for criticism. For example, local area unemployment rate estimates have been subjected to criticism for a number of years in States like California, Ohio, New Jersey, and Massachusetts, principally because these estimates based on administrative records from the unemployment insurance system could be compared with the presumably comparable rates for the same areas based on the Current Population Survey. The prospect for critical examination of local area estimates is particularly strong when concrete incentives are present to seek estimates which increase local allocation of government funds under programs such as revenue sharing. This suggests an approach to local area estimation which is not based on the use of a single technique, but on all the information available for a given locality which may help to improve its estimate. This would be a highly professional-labor- intensive activity and would probably be out of reach of most organizations with limited budgets. With respect to the Ericksen paper a question may be raised about the possible use of additional local area information available generally, which may be used as symptomatic vari- ables, such as: 1. the racial mix in the base year 2. the urban-rural mix in the base year 3. the population density in the base year 4. the age-sex mix in the base year, and so on. Ericksen's work uses ratios as the symptomatic variables as well as the criterion variable. I would like to see some evidence on the amount of collinearity present among the symptomatic variables. In general, on statistical grounds, one seeks for symptomatic variables which are relatively uncorrelated with each other but are correlated with the criterion variable. These considerations should not ignore those based on the subject matter under study. I liked Ericksen's formulation of the regression equation to take account of the sampling error in the criterion variable. This assumes independence of the u and v error terms which does not appear to be quite true for his data set. This suggests extension of his model to include some nonzero but unspecified covari- ance between u and v. In addition there may be reason to assume a v term for specific symptomatic variables, which would alter the model somewhat and would appear to produce biased estimates of the parameters of the regression equation. There is also some heterogeneity in u, which suggests seeking some transformation of the variables to correct this, or the use of generalized least squares in the estimation procedure. Between the extreme of seeking a single estimation equation for all areas, and the extreme of seeking to maximize the use of ad hoc local area information through a variety of techniques, there may be an optimum point at which clusters of areas may be studied, each cluster with its own estimation technique. Ericksen's formulation suggests little payoff here if the analysis uses sample-based criterion variables. 59 DEPARTMENT OF HEALTH. EDUCATION, AND WELFARE OFFICE OP THE SECRETARY WASHINGTON, DC. 20201 Office of Human Development Administration on Aging TECHNICAL ASSISTANCE MEMORANDUM AoA-TA-77-4 February 16, 1977 TO : STATE AGENCIES ADMINISTERING PLANS UNDER TITLES III AND VII OF THE OLDER AMERICANS ACT OF 1965, AS AMENDED INFO FOR: Area Agencies on Aging and Nutrition Projects SUBJECT: Provisions of the State and Local Fiscal Assistance (General Revenue Sharing) Act Amendments of 1976 (P.L. 94-488) CONTENT: Public Law 9I+-J4-88 (October 13, 1976), "State and Local Fiscal Assistance Amendments of 1976," contains several provisions that will affect the elderly. 1. Background The State and Local Fiscal Assistance Act of 1972, commonly referred to as "General Revenue Sharing" (P.L. 92-512), was enacted on October 20, 1972. The legislation created a trust fund to receive a portion of Federal individual income taxes for disbursement to States and local units of government to be used for various types of programs, including social services for the poor or aged . Funds amounting to some $6 billion annually were disbursed to approximately 39,000 recipient government entities. The program was scheduled to expire on December 31, 1976. Public Law 9U-U88 was enacted on October 13, 1976 to amend and extend the revenue sharing effort. 2. Extension of Program The 1976 Amendments extend the program from January 1, 1977 through September 30, 198O. The new legislation contains several provisions that affect the elderly. 3. Major Provisions of Public Law 92-512 Affecting the Elderly, and Effect of Amendments thereto ; A. Use of Funds by Local Governments for Priority Expenditures The original Act provided that revenue sharing funds received by units of local government could be used for priority expenditures covering ordinary and necessary maintenance and operating expenses, and capital projects for, among others, social services for the poor or aged . The Amendments eliminate the list of priority expenditures. Nevertheless revenue sharing funds can still be used to support programs which will benefit the elderly. This memorandum will offer a number of suggestions about how revenue sharing can be obtained for such programs. B. Use of Funds for Matching The Amendments specify that revenue sharing funds now may be used to match Federal grants under other Federal programs, such as those authorized by the Older Americans Act under Titles III (State and Community Services), V (Multipurpose Senior Centers), and VII (Nutrition Program for the Elderly). Hence, State and local agencies should make concerted efforts to secure revenue sharing support for their programs, both as a means of obtaining additional funds and as a means of satisfying non-Federal matching requirements. For more on how to pursue these objectives, see Section k, below. C. Nondiscrimination Provisions The Amendments broaden the existing nondiscrimination provision in the original Revenue Sharing Act by adding "Any prohibition against discrimination on the basis of age under the Age Discrimination Act of 1975 . . . shall also apply to any such program or activity. " The Amendment to prohibit discrimination on account of age will not take effect until the Age Discrimi- nation Act of 1975 (P.L. 9^-135, dated November 28, 1975) becomes effective to prohibit age discrimination in all Federal programs, which, in all likelihood, will not occur until sometime after January 1, 1979* Plowever, this provision reflects a Congressional commitment to combat age discrimination. Such a commitment can be cited by the elderly and their advocates even before the legal provisions regarding age discrimination in revenue sharing programs actually take effect. D. Public Hearings The original Act made no provision for public hearings relative to the planned use of revenue sharing funds. The recent Amendments require governmental units to- hold public hearings to provide citizens an opportunity to offer written and oral comment on the possible uses of revenue sharing funds. The governmental units are also required to publish, at least 10 days prior to the public hearing, in a newspaper of general circu- lation, a notice stating the proposed uses of revenue sharirg funds and indicating the time and place of the public hearing. The Amendments express a strong intent that proper consideration be given to senior citizens and their organizations, since the following sub-section has been added to Section 121 of the original Act. "(g) PARTICIPATION BY SENIOR CITIZENS -- In conducting any hearing required under this section, or under its own budget processes, a State or local government shall endeavor to provide senior citizens and their organizations with an opportunity to be heard prior to the final allocation of any funds provided under the Act pursuant to such hearing." E. Review of Revenue Sharing Expenditures The Amendments continue a requirement in the original legislation that jurisdictions which receive and expend revenue sharing funds issue a yearly report setting forth the amounts and purposes for which funds received from revenue sharing have been appropriated, spent, or obligated during the entitle- ment periods, and snowing the relationship of those funds to the relevant functional items in the government's budget. This report must identify differences between the actual use of funds received and the proposed use of such funds. This report precedes the public hearings scheduled on the pro- posed use of funds, and serves as a guide to the future use of funds . k. St rategies The provision allowing use of revenue sharing funds for matching, coupled with the added provision for participa- tion by the elderly in the preliminary action involving allocation of funds, place older Americans in an advantageous position in their quest for revenue sharing funds. However, the elderly and their spokespersons must be actively and. aggressively involved in the entire process whereby revenue sharing funds are allocated if proper advantage is to be taken of those opportunities. This makes it incumbent upon all in the aging network -- Regional Offices, State Offices, Area Agencies on Aging, agencies and organizations which receive Older Americans Act grants and contracts, and organizations of elderly persons -- to seek and find opportunities to assure adequate presentation of the needs of programs for the elderly to receive revenue sharing funds. A. Hearings Full advantage should be taken of the public hearings concerning the proposed use of revenue sharing funds. Network members may find the following suggestions helpful in exploiting the opportunities afforded by the public hearings: Aging agency personnel, acivocacy groups repre- senting the elderly, and other interested organizations should formulate a list of priority Tises for revenue sharing resources well in advance of the 1 oc al jurisdiction's formulation of its revenue sharing program . In formulating their recommendations, area agencies and organizations should give particular consideration to needs for non- recurring expenditures, such as development cf senior centers, the purchase of furniture and equipment for aging programs, etc., since experience has shown that State and local governments prefer to use revenue sharing funds for such purposes rather than for programs which will require an on-going commitment. Opportunities for use of revenue sharing funds for matching purposes should also be explored in this process. The priority list should be shared with the decision makers through letters, briefings, and informal conversations. Every effort should be made to provide a detailed rationale, with as much supporting data as possible, for each request. This will be far more effective than simply presenting a list of demands. In contacting local officials about proposed revenue sharing expenditures, the elderly and their representatives should also offer suggestions about how best to schedule the hearing to facilitate participation by the elderly, as required under the 1976 Amendments. Special attention should be given to such key issues as use of facilities accessible to the elderly, a convenient time, etc. At the hearing itself, spokespersons can both commend the public officials for any favorable treatment the requests voiced by the elderly may have received and also renew requests which have not yet received favorable consideration. A written statement should be prepared in advance of the hearing to be submitted for the purpose of confirming the oral testimony. \J The use of a limited number of veil prepared and well rehearsed spokespersons is preferred to a procession of .individual witnesses. A good attendance by the elderly will show their interest and their support for their spokesperson. They should also encourage nevs media coverage of the hearings in order to gain as much publicity as possible. Following the hearing the spokesperson(s) should send letters to the participating public officials thanking them for their interest and attention and reiterating any requests which may yet receive a positive response. B. Review of Previous Expenditures This activity is necessary to effective advocacy of the elderly' s interest in revenue sharing. ° Area agencies and other organizations should carefully assess each report published by the local jurisdictions concerning the use of revenue sharing funds. This can provide much information relevant to the formulation of new requests and proposals for projects of benefit to the elderly, to be presented during future funding cycles. ° Any failures to honor commitments to uses of funds for aging programs should be noted and called to the attention of the appropriate public officials. If possible an accounting for the discrepancy should be obtained and a commitment secured from the officials to take corrective action in the next revenue sharing expenditure cycle. C. Nondiscrimination Although the prohibition against discrimination on the basis of age will not become effective until January 1, 1979, or thereafter, full advantage should be taken of the expression of intent in the 1976 Amendments against such discrimination. In formulating requests for specific revenue sharing allocations, area agencies, et al . , should call the attention of the relevant officials to the fact that the anti- discrimination provisions "will soon be in effect. These officials should be made aware that favorable consideration of requests by the elderly prior to the effective date of the anti-discrimination provisions would be an especially effective way of demonstrating good faith. When the effective date of the anti-discrimination provisions is established, area agencies and/or other organizations should take the initiative in formally notifying local officials of that specific date. This will serve as a reminder that the elderly require serious consideration in the allocation of revenue sharing funds. 5. Future Actions by AoA On January 10, 1977 the U.S. Department of the Treasury issued final regulations pertaining to public involvement in decisions about the allocation of general revenue sharing funds. The requirements specified in the new regulations are reflected in the material transmitted via this Technical Assistance Memorandum. AoA will advise the network of any subsequent issuances that are specifically relevant to the process of securing a share of the general revenue sharing resources for the elderly. AoA will also advise the network when the Age Discrimination Act takes effect. INQUIRIES: State Agencies should address inquiries to Director, Office of Aging, Regional HEW Office. Area Agencies on Aging should address inquiries to State Agencies on Aging. Title VII Nutrition Projects should address inquiries to their grantor (State or Area Agency on Aging). oL4i Arthur S. Flemming Commissioner on Aging DEPARTMENT OF HEALTH. EDUCATION. AND WELFARE WASHINGTON. D.C. 20201 OFFICIAL BUSINESS PENALTY FOR PRIVATE USE. 9300 POSTAGE AND FEES PAID U.S. DEPARTMENT OF H Z.W 391 Mr. Daniel Sqhul'ier 0SCP-3 Special As*istar.t to the Governor c : n Agiag 504 State Street Bldfe. Harrisburg, Perm, IP 1 01 CENSUS TRACT PAPERS Series GE-40 No. 1. Papers Presented at the Census Tract Conference, Sep- tember 10, 1965, Philadelphia, Pa. No. 2. Needs and Plans for the 1970 Censuses (Census Tract Conference of August 15, 1966, at Los Angeles, Calif.) No. 3. Some Uses of Census Tracts in Private Business, by Wilbur McCann (1967) No. 4. Papers Presented at the Conference on Small-Area Statistics, American Statistical Association, Wash- ington, D.C., December 27, 1967, and Related Papers No. 5. Papers Presented at the Conference on Small-Area Statistics, American Statistical Association, Pittsburgh, Pa., August 23, 1968, and Related Papers No. 6. Final 1970 Census Plans and Four Programing Systems for Computerized Data Retrieval and Manipulation (Conference on Small-Area Statistics, American Statis- tical Association, August 21, 1969, New York, N.Y.) No. 7. New Uses of Census Resources, The Southern California Census Use Study and a Related Paper (Conference on Small-Area Statistics, December 29, 1970, Detroit, Mich.) American Statistical Association. No. 8. Small-Area Statistics: Strengthening Their Role in Fed- eral Government and Their Use in Criminal Justice Programs (Papers Presented at the Conference on Small-Area Statistics, American Statistical Association, August 23, 1971, Ft. Collins, Colo.) No. 9. Social Indicators for Small Areas, Papers Presented at the Conference on Small-Area Statistics, American Statis- tical Association, August 14, 1972, Montreal, Canada. No. 10. Statistical Methodology of Revenue Sharing and Related Estimate Studies (Papers Presented at the Conference on Small-Area Statistics, American Statistical Associa- tion, December 27, 1973, New York, N.Y.) Copies of early editions may no longer be in print but are usually available in Federal Depository Libraries. Papers pre- sented at conferences held from 1958 through 1964 were originally published by the Bureau of the Census in its Working Paper Series. For information on all census tract reports, write to: Subscriber Services Section (Publications) Social and Economic Statistics Administration Washington, D.C. 20233 U.S. GOVERNMENT PRINTING OFFICE DIVISION OF PUBLIC DOCUMENTS Washington, D.C. 20402 OFFICIAL BUSINESS FIRST CLASS MAIL PENN STATE UNIVERSITY LIBRARIES ADDDD72fib017S PRINTING OFFICE 375 ; ^6-l9l 6