Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
DATABASES FOR MICROSIMULATION: A COMPARISON OF THE MARCH CPS AND SIPP 53 support programs. However, they do not attempt to cope with all missing data problemsâfor example, they do not currently simulate eligibility for AFDC based on a first pregnancy. The models have also not sought to address problems of missing data for related programs such as child support enforcement. This brief review suggests some areas for which further quality enhancement is possible and also some areas for which such enhancement might most usefully be performed by the Census Bureau. For example, the Census Bureau could carry out further research on population undercoverage. Such research might lead to adjustments in survey records that could have an important impact on the quality of data for simulating income support programs and other policy analysis purposes. In another area, that of income underreporting, the models already partially address the problem. However, it might well be possible to enhance quality and reduce overall costs for producing policy analyses for federal agencies (by eliminating duplicative work on the part of the various models) by having the Census Bureau address this area. The Census Bureau could take advantage of its access to administrative records for this purpose. In yet another area, that of constructing filing units, one could consider having the Census Bureau provide the necessary recodes. However, this area seems more properly the domain of the models, particularly as the models must be able to simulate proposed changes in filing unit definitions. THE SUITABILITY OF CPS AND SIPP FOR INTEGRATED MODELING OF TAXES AND TRANSFERS This chapter has focused largely on the requirements for modeling income support programs. More investigation would be needed to assess the comparative advantages of CPS and SIPP for models of taxes, retirement income, and health care utilization and financing and to consider the optimal roles for the Census Bureau and the models in generating suitable databases for these policy areas. It is clear, however, that SIPP provides more items of relevant information. For example, SIPP has data, such as coverage by a private pension plan, that are relevant for retirement income modeling. Both CPS and SIPP have questions on health insurance coverage, but SIPP also typically includes one or more topical modules pertaining to health status, disability, and use of health care services that are relevant to modeling health care policy issues. One point to note is that SIPP was originally intended, at least on the part of some of the participants in its design (see Committee on National Statistics, 1989), to make it possible to analyze both taxes and transfers: that is, to serve as an integrated database for modeling how government income grants and government tax levies affect the income distribution. Indeed, SIPP regularly includes a module that is designed to obtain detailed information about federal income taxes. However, only a few variables from the tax module are included in files that are available for research and analysis use from SIPP (no
DATABASES FOR MICROSIMULATION: A COMPARISON OF THE MARCH CPS AND SIPP 54 information on amounts is made available), and these files are obtainable only under special access arrangements. (Confidentiality concerns and low response rates to the questions asking for amounts of deductions and other tax-related items are the reasons for limiting access.) In addition to limited access, other problems have minimized the effectiveness of SIPP as a database for modeling taxes as well as transfers. Reduced sample size is a constraint for adequate analysis of the upper end of the income distribution, which is important for tax modeling.21 Reporting of assets and asset income in SIPP, and the detail collected, are generally better than or at least as good as those in other household surveys; however, as noted above, nonresponse rates for these items remain high. Also, the operational difficulties that SIPP experienced in its start-up years have led the Census Bureau to strive to focus the survey on key goals. With the advice of outside groups such as the Committee on National Statistics (1989) and federal agency users, the Census Bureau has focused the survey on the programs and policy issues targeted at the population on the lower end of the income scale. Examples of this focus are the agency's decision to oversample poverty households in SIPP and to ask detailed questions about assets and taxes only once in each panel rather than twice. Given sufficient resources, user interest, and the resolution of the data access problem, it would certainly be possible to restore capability to SIPP to provide an integrated database for analysis and modeling of both tax and transfer programs. Of course, there would be many difficult issues to work out. For example, the sample design could quickly become unwieldly and lose efficiency if there were efforts to oversample both ends of the income distribution. The 1979 ISDP panel oversampled both rich and poor census tracts; however, subsequent analysis determined that the design effects resulted in very high variances for many estimates of interest because the oversampled populations did not correspond all that well to the populations for which estimates were needed. The Census Bureau has conducted research on the use of IRS records for ratio-adjusting SIPP survey weights in order to reduce sampling variance. Preliminary results suggest that this is effective, particularly for estimates based on the middle and upper parts of the income distribution (Huggins and Fay, 1988). Further work along these lines, including use of administrative records applicable to the low-income population, could be fruitful. One might ask whether the CPS could not better serve as an integrated tax and transfer database. If the criteria focused solely on richness of data content and improved quality of reporting, the answer would be that the CPS has no particular virtues in this regard and ma ny defects compared with the SIPP. It would be necessary to expand considerably the March questionnaire 21 Special designs would undoubtedly be needed to obtain adequate samples of the upper tail of the income distribution, given highly skewed distributions: for example, the top 1 percent of tax returns accounts for one-third of dividend income; see Vaughan (1988) on this point.
DATABASES FOR MICROSIMULATION: A COMPARISON OF THE MARCH CPS AND SIPP 55 to obtain needed data for tax and transfer modeling. Such expansion does not seem feasible given the primary use of the CPS as a labor force survey. Sample size and timing considerations, in contrast, greatly favor the CPS at present and are likely to continue to favor the CPS in the future, particularly if the proposed sample expansion for the CPS is revived. With regard to timing, it seems unlikely that files from SIPP that will permit analysis on an annual basis can ever be released on as fast a schedule as the March CPS. On the assumption that the March CPS and the SIPP will continue to coexist (rather than the latter replacing the former as the primary source of income information), the Census Bureau has recently begun a project to explore ways to relate data from the two surveys and from administrative records to produce improved statistics for the population over the full range of the income distribution. Traditionally, the Census Bureau, along with other statistical agencies, has viewed its mission as generating data files and published tabulations for individual surveys. Hence, income data from the CPS are currently published in the P-60 report series, while income data from the SIPP are published in the P-70 series. Census Bureau staff have begun to think that their goal might better include publishing the best income series based on all available data sources, and they have sketched out a research program to accomplish this goal.22 In broad outline, the project would involve using administrative records and other sources to assess the extent and nature of nonsampling errors, such as underreporting and misreporting, in the March CPS and SIPP income data, including earnings and property income as well as sources of transfer income. There would then be an attempt to adjust the SIPP data through some sort of multivariate imputation or weighting technique. Alternatively, exactly matched administrative values might be substituted directly in the SIPP records, perhaps with random noise added, if the problems of data access could be worked out. Then, the adjusted SIPP data would be used to improve the quality of the income data from the March CPS through a related imputation or modeling procedure. The CPS data would retain the advantage of timeliness (assuming that the adjustments were made from an earlier SIPP file), while the later SIPP data would provide the advantage of additional subject detail and monthly amounts. The project would also involve exploring how to make the enhanced SIPP and CPS databases publicly available for modeling and research use. This project is still in its infancy. To date, only limited resources are available at the Census Bureau to carry it out. 22 The description below of the Census Bureau's project to evaluate and adjust for nonsampling error in CPS and SIPP income reports is based on a presentation by John Coder to the Committee on National Statistics' Panel to Evaluate the Survey of Income and Program Participation, July 30, 1990. See also Coder (1991). For an earlier discussion of issues in and possible approaches to correcting household survey income reports, see Green (1983).