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 56 CONCLUSIONS On the basis of reviewing the data quality and utility of the March CPS and the SIPP for analyzing the low- income population, it seems clear that, as originally intended, SIPP is muc h better suited in many respects than is the CPS for modeling income support programs such as AFDC. As a new, complex, panel survey, SIPP has been subject to intense scrutiny with regard to data quality; indeed, research has documented serious issues such as the seam bias, misreporting of AFDC benefits (as general assistance), and differential sample attrition. However, it is likely that the CPS has ma ny data quality problems of its own, which simply have not been studied in the same way. Moreover, the available comparisons of SIPP and CPS data quality, such as item nonresponse rates, are by and large favorable to SIPP. By every measure, SIPP provides more complete and appropriately specified data for the requirements of modeling income support programs. In addition to the complexities involved in using the new, longitudinal SIPP data, which should not be underestimated, the major drawbacks of SIPP as a microsimulation model database for social welfare programs at this time are sample size and timeliness. The SIPP sample size for the low-income population is simply insufficient to support the policy analysis demands placed on microsimulation models. Indeed, for a program such as AFDC that embodies important state-by-state differences, there is evidence that the sample size in the March CPS is also less than adequate. Moreover, the SIPP data currently lag the CPS by at least 2 years. It appears that the March CPS will continue to dominate the SIPP on these two dimensions for the foreseeable future. With regard to sample size, even the proposed restoration of the original design for the SIPP will provide a maximum of about 35,000 households for analysis, compared with about 60,000 in the March CPS. (Combining two SIPP panels of 20,000 households each will not produce 40,000 cases for analysis because of attrition.) Oversampling of the low-income population in SIPP would improve the effective sample size for modeling income support programs; however, the two-phase sample expansion would do the same for the CPS. Moreover, it is not out of the question to consider adding a low-income supplement to the March CPS. With regard to timing, the Census Bureau has not yet established a track record for timely processing of the complex SIPP data, and, at best, the SIPP will inevitably lag the March CPS by 6 months or more for a particular calendar year. Moreover, the complexity of the SIPP means that the modelers are likely to require more time to process the data files from the SIPP than they are from the March CPS, even when they develop smooth-running processing systems for the SIPP files. Hence, it may be that, for the foreseeable future, the Census Bureau and the policy analysis agencies should look for ways to relate data from the March CPS and SIPP rather than concentrate their resources on one or the other survey. Specifically, it ma y be that the agencies should plan to continue to use the March
DATABASES FOR MICROSIMULATION: A COMPARISON OF THE MARCH CPS AND SIPP 57 CPS as the main database for models of income support programs (and other social welfare policy areas) and look to the SIPP to augment and enhance the quality of the CPS information, as well as to support special analyses that make full use of the richness of the SIPP information.23 In addition to steps that the modelers can take to enhance their CPS databases through use of the SIPP, for example, by improving imputations of needed assets and expenditures data, there are relatively inexpensive measures that the Census Bureau could consider to make it easier to relate the two surveys for modeling and analysis use. Additional redundancies could be built into the two surveys: for example, adding a few questions on last month's income might well improve the quality of the procedures that are used in the models to allocate annual CPS income to monthly amounts, particularly if SIPP data were used to inform the allocation. If it is not clear whether to stay with CPS or change over to SIPP for modeling income support programs, it is quite clear that there are strong arguments for the Census Bureau to contribute more added value to the data files released from the SIPP and CPS that are used for important federal policy analysis purposes. For example, it seems a wasteful allocation of resources to have several groups of modelers implement corrections for income underreporting, when the Census Bureau is in a position to do the job more efficiently and perhaps more effectively. Many concerns would need to be addressed in asking the Census Bureau to play a more active role in generating useful databases for modeling and policy analysis generallyânotably a concern about timeliness of data release. However, the potential benefits argue for serious investigation of the appropriate division of labor between the Census Bureau and the policy analysis community. In this regard, the Census Bureau's research program to explore ways to develop an improved set of income statistics based on data from administrative records, SIPP, and CPS is promising for two reasons. First, it should result in much more information about data quality in SIPP and CPS. Second, there is the prospect that the project may result in enhanced databases that will provide higher quality data to support policy analysis and modeling of government programs, including both taxes and transfers, that affect the entire U.S. population. 23 One issue not explicitly addressed in this chapter concerns the desirability of developing microsimulation models of the short-term dynamics of program eligibility and participation. Such models would require extensive use of SIPP data. However, like the existing dynamic retirement income models, which apply transition probabilities estimated from the PSID and other small longitudinal surveys to the much larger CPS, such models could apply transition probabilities estimated from SIPP to a March CPS database.