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.
ALTERNATIVE MODEL DESIGNS: PROGRAM PARTICIPATION FUNCTIONS AND THE ALLOCATION OF ANNUAL TO 111 MONTHLY VALUES IN TRIM2, MATH, AND HITSM makes it easier to line up simulated participants with control totals by benefit level and participant characteristics. FSTAMP uses the participation rates derived from this process to select participants for the baseline simulation. The participation rates for subsequent simulations have a behavioral response built into them. FSTAMP simulates units that participated under the current law simulation and that are eligible for higher benefits to participate with probability 1. For units that did not participate under current law but that are eligible for higher benefit amounts under the reform, FSTAMP calculates a higher probability of participation under the reform than under current law. The module adjusts the participation probability for these units by adding 0.0014 times the difference in benefit amounts to the previous participation probability. Conversely, for units that participated under the current law simulation but that are eligible for lower benefits under the reform, the module reduces the participation probability for these units by 0.0014 times the different in benefit amounts. The number 0.0014 comes from an analysis of food stamp program participation by Beebout (about 1976). HITSM Documentation for the food stamp participation function is very sketchy. It appears that the participation function is designed to control both benefits and recipients in the baseline simulation, controlling on the number of persons in the unit, the size of potential benefit, and the type of unit: nonelderly non-AFDC, elderly non- AFDC, AFDC). Comment The MATH model clearly has the most elaborate food stamp participation function, which is not surprising given its premier role in simulating food stamp program alternatives. It also incorporates an explicit behavioral response to changes in the benefit amounts for which units are simulated to be eligible under proposed program alternatives compared with current law. Work is in progress to revise and further refine the algorithm based on research conducted with the SIPP. In contrast, the current TRIM2 food stamp participation function, although more complex than previous functions, is simpler than the MATH function. Evaluation of Alternative Participation Functions Given the low participation rates in some income support programs, which necessitate the simulation of participation behavior, and the interest of policy
ALTERNATIVE MODEL DESIGNS: PROGRAM PARTICIPATION FUNCTIONS AND THE ALLOCATION OF ANNUAL TO 112 MONTHLY VALUES IN TRIM2, MATH, AND HITSM makers in the effects of program alternatives on both the size and the composition of the recipient population, it seems highly important to evaluate the participation functions used in major microsimulation models. Of particular interest is the impact of the baseline participation function on the profile of the caseload that results from alternative program simulations. There are a number of reasons to be concerned about the adequacy of the models' participation functions, which, although different in many particulars, are quite similar in basic approach. (In summary, the approach for the baseline file is to simulate eligible units, compare selected characteristics of those units with administrative data on program participants, and derive a set of participation probabilities so that simulated participants from the pool of eligible units closely match actual participants on the specified characteristics.) One concern is that most participation functions include relatively few variables. Yet one or more omitted variables may be important to include in order to simulate accurately the effects of particular program changes. Another concern is that the very basis for the participation rates calculated by the models rests on comparing data from two disparate sourcesânamely administrative data on recipients and simulated data on eligible units derived largely from survey information. While it is natural to interpret the participation function as descriptive of individual behavior with regard to the participation decision, the differences between simulated eligible units and actual participants ma y also reflect errors in the administrative and household survey data (e.g., sampling error or population undercoverage), as well as errors in the modeling of program eligibility and in the granting of eligibility and benefits by program caseworkers. As a result, the calibration process that is designed to align the size and characteristics of simulated participants closely with administrative information, while no doubt reducing some kinds of bias, undoubtedly introduces other biases into the resulting participation rates.12 Moreover, while the general practice of using the baseline participation function for alternative scenarios imparts a measure of consistency across simulations, the biases in the baseline rates may have untoward effects on the simulations of one or more program alternatives. Unfortunately, there has been no sensitivity analysis of alternative program participation functions either within or across models.13 Also, there have been 12 Young and Giannarelli (1988) discuss the issues involved in calibrating or aligning simulated program participants to accord closely with administrative information. Recognizing the various sources of error that may affect the calibration, they present a general framework for considering model alignments to meet different objectives and assumptions. 13 Kormendi and Meguire (1988) performed a type of sensitivity analysis of the TRIM2 probit equation for determining AFDC participation probabilities. They multiplied each coefficient in the probit model by a random variable, uniformly distributed on the interval (0.8, 1.2). They repeated this process 15 times on a TRIM2 baseline file from the March 1986 CPS. The results indicated that the variability of the resulting estimates of simulated AFDC benefits was less than the variability in the estimates of participating units. Generally, even in the case of units the variability of the results was somewhat less than the variability of the simulated probit parameters. (See Cohen, Chapter 7, in this volume.)