enforcement, child care tax credits, and programs to encourage employment. Neither SIPP nor CPS is designed specifically to address such initiatives, but, again, SIPP has the advantage in terms of the scope of the data that are included. For example, SIPP has regularly asked about participation in job programs and has included a module on day care needs and expenses in every panel.

Many of these new policy interests involve a longitudinal component, and there has thus been growing interest in studying the dynamics of program participation on a subannual basis. Longitudinal surveys such as the PSID have supported extensive analysis of movement in and out of poverty and reliance on income support programs on a year-to-year basis over long spans of time. However, until the ISDP and SIPP, there has been no comparable database for analysis of short-term program dynamics. Yet such analysis is important to understand turnover and recidivism in caseloads, to determine whether multiple program participation tends to be concurrent or serial in nature, and to assess the effectiveness of program innovations such as transitional Medicaid and child care support in moving families from welfare to the work force. Obviously, the CPS cannot contribute to these types of analyses; the SIPP can.

Up to this point, this chapter has addressed problems for modeling and analysis of income support programs from errors associated with specific aspects of the design and operation of the March CPS and SIPP surveys. This section adopts a more global perspective, reviewing estimates from these two surveys of characteristics of AFDC and food stamp recipients compared with data from administrative records. Although microsimulation models simulate program participation from among a pool of simulated eligible units, and some models ignore reported recipiency of benefits altogether in selecting participants, the extent to which the input data accurately portray the recipient population has obvious implications for the quality of the simulations.

Most of the available comparisons evaluate one or another but not both surveys against administrative records. (The aggregate comparisons, reviewed above, that found more complete reporting of total recipients and benefits in SIPP versus CPS are an exception.) Hence, it is not possible to make definitive assessments of the relative merits of the two surveys in characterizing the participant population. In addition, although administrative data are used as the target in comparisons with survey estimates, observed differences do not always indicate errors in the survey data. The differences may be due to errors in the administrative records or to conceptual distinctions between the two sources.

Allin and Doyle (1990) compared SIPP estimates of reported food stamp participants (including those simulated to be eligible and not eligible for the program) with data from the Integrated Quality Control System (IQCS), which

comprises monthly samples of case records for the food stamp, AFDC, and Medicaid programs. Their comparisons were limited to cases with incomes at or below 250 percent of the poverty threshold for the month of August 1984: for that period, the SIPP sample comprised 1,272 and the IQCS sample 6,932 unweighted food stamp recipient units.

Allin and Doyle found significant differences between the SIPP and the IQCS on several dimensions. SIPP estimated higher proportions than the IQCS of food stamp units

SIPP estimated lower proportions than the IQCS of food stamp units

The estimated average child care deduction in SIPP was only 72 percent of that in the IQCS, and the average estimated medical expense deduction in SIPP was only 41 percent of the average in the IQCS. Finally, the SIPP food stamp participant population included 11 percent who appeared to be ineligible for the program on the basis of their assets or income levels.

The authors suggest and, in some cases, are able to provide supporting evidence for the various factors that may account for the differences they find. For example, the Census Bureau’s imputation procedures very likely account for over one-fourth of the seemingly ineligible participant cases. However, more work is clearly needed to understand their results.

Unpublished tabulations of average monthly reported AFDC participants from the March 1984 and 1988 CPS, which the Urban Institute prepared for the panel, provide a basis for evaluating the CPS portrayal of the AFDC population in comparison with data from the 1983 and 1987 IQCS (representing averages of 12 monthly samples for each year). The following are unweighted sample sizes for average monthly AFDC cases:¹⁶

Again, there are important differences between the CPS and the administrative data, similar to those found for the food stamp recipient population in the comparisons involving SIPP. Of AFDC units in 1983, the CPS estimated higher proportions than the IQCS for units

Of AFDC units in 1987, the CPS estimated higher proportions than the IQCS for units

In addition, for 1987 the CPS estimated a lower proportion than the IQCS of AFDC units headed by a married person (11 versus 13%). Finally, the AFDC participant population estimated by the CPS for both 1983 and 1987 included 14 percent who were simulated to be ineligible for the program.

The differences noted above for the recipient population, such as the much higher proportions of AFDC and food stamp recipient units with earnings in CPS and SIPP in comparison with the IQCS, extend as well to the eligible population. That is, the simulated pool of eligible units typically contains far more cases of program units with earnings and other characteristics that are less typical of the caseload as portrayed in the administrative data and fewer cases with characteristics that resemble the caseload. In turn, this pattern has made it difficult for the microsimulation models to calibrate the simulated participant population to match administrative controls. For example, the TRIM2 1983 baseline simulation, after the calibration process had been completed, estimated

8 percent of the simulated AFDC caseload with earnings, compared with 5 percent in the IQCS. The TRIM2 1987 baseline simulation, after calibration, estimated fully 13 percent of the simulated caseload with earnings, compared with only 7 percent in the IQCS.¹⁸

Similar problems have characterized food stamp simulations. Runs of the MATH model with a March 1981 CPS file projected to known population controls for August 1984 produced fewer cases of eligible units in single-parent families than there were such cases reported as beneficiaries in the IQCS. Runs of a model with 1984 SIPP data produced no better results (Doyle and Trippe, 1989: Table 7). Allin and Doyle (1990) reported better results for SIPP with a revised definition of household type for simulated food stamp units. However, the revised definition represented an arbitrary choice, and it is not clear whether the problem of too few cases of certain kinds of eligible units stems from population undercoverage, poor definition of the recipient unit within the household, or other factors.¹⁹

The models calibrate their baseline simulations of program participants in order to provide credibility and create what is assumed to be an improved database for simulations of alternative programs. However, the calibration almost always produces participation rates that are biased to a greater or lesser degree and that may bias the results for simulations of alternative programs, if the assumption that the biases are consistent across scenarios does not hold.²⁰ Clearly, more work is needed to understand the reasons for the differences between survey and administrative record estimates of program caseload characteristics.