Current Projects

Modeling Eligibility for Safety Net Programs at the State Level
Sheila Zedlewski, Donald Alderson, Linda Giannarelli, and Laura Wheaton
Draft Prepared for National Academy of Sciences Panel
Draft: December 30, 1999

1. Introduction: Why Model Benefit Eligibility and Program Participation?

As states take on more responsibility for social welfare programs designed to assist low-income families with children, they will need estimates of program eligibility as well as program participation. Estimates of program eligibility tell us about all families who qualify for some assistance under the rules defined by the states’ programs, and data on program participation tell us about families who actually applied for and received benefits provided by the program. States can use information about program eligibility and participation rates (defined as the ratio of participants to eligibles) to understand the potential reach of their programs as well as the potential budget liability should all eligible families decide to apply for benefits. For example, recent federal regulations regarding states’ Maintenance of Effort (MOE) requirements under the Temporary Assistance for Needy Families (TANF) program require states to limit all types of MOE expenditures to families who meet the financial eligibility criteria under the states’ TANF plans. 1 Thus, information about eligibility under TANF can help guide states’ deliberations about the potential cost and coverage of social assistance that could be offered to low-income families under their MOE.

Traditionally, many families eligible for government social welfare benefits have not participated in these programs. Families eligible for larger benefits have typically participated at higher rates than others, reflecting a greater need for government assistance. Nonetheless, participation rates have also varied across demographic groups, benefit programs, and time. Some groups have traditionally been less accepting of government assistance than others. For example, various studies have documented lower participation rates for the elderly than for the nonelderly.2 Programs that embody less stigma and relatively easy application processes (such as the earned income tax credit) produce higher participation rates than those that require application at a welfare office and detailed documentation of family status, current income, recent work history, and assets. Economic conditions can also affect participation rates. For example, families remain on food stamps longer when unemployment rates are higher, although this macroeconomic influence on participation interacts with individuals’ economic situations.3

States need to be aware of factors affecting participation choices in order to manage their social welfare programs. For example, in the event of a weaker economy, two factors could increase states’ Temporary Assistance for Needy Families (TANF) caseloads. First, more families will become eligible for benefits in the event of lower incomes, and second, some of the families who are currently eligible for TANF benefits but not participating will decide to apply for assistance. Ideally, states would have models of benefit eligibility to allow them to understand the potential effects of program participation choices on their caseloads and costs.

States can develop microsimulation models of program eligibility to monitor participation in state safety net programs. These models have three essential ingredients. First, states need survey data representative of all their low-income households, with sample sizes sufficient to produce statistically reliable estimates and enough information about households so that the key features of states’ program rules can be simulated. Second, states need computer models that capture program rules in sufficient detail to estimate which families are eligible for different benefits. Third, states need administrative data that describe who is currently receiving program benefits along with some information about the characteristics of participants and the benefits received. The administrative data provide information for comparing the characteristics of families who actually receive benefits to those who are eligible for benefits but do not receive assistance. These comparisons would allow states to understand whether families with particular characteristics tend to participate at higher or lower rates than others (e.g., younger TANF eligibles, two- vs. one-parent families, ethnic subgroups, etc.); what the total dollars of benefits to eligible nonparticipants would be; and, if there were sample sizes sufficient to support substate analysis, whether eligibles in some counties participate at different rates than others.

Each of the steps in developing a microsimulation model of program eligibility presents its own challenges, but states can learn from prior research and models currently in use at the federal level to develop models of eligibility and program participation. This paper first outlines the "generic" methods generally required to estimate program eligibility and participation. We then provide examples of models and studies of program participation and results indicating program participation rates during the 1990s. The next section provides a concrete examination of the steps and challenges in developing an eligibility model for states’ new Temporary Assistance for Needy Families (TANF) programs. Then we illustrate an application of a model of AFDC/ TANF eligibility and participation using the Urban Institute’s Transfer Income Model, Version 3 (TRIM3). We highlight differences in participation across states and family type. We conclude with a summary of the key factors to consider in developing models of program participation.

2. Methods for Estimating Program Eligibility

Defined very simply, microsimulation models of program eligibility simulate the rules of government programs at the level at which they operate. The level of operation captures the "unit" that can apply for benefits as defined by federal or state regulation. The "unit" relevant for social welfare programs can be the household, the family, or the individual. Eligibility models can be relatively simple, designed to provide policymakers with a "rough" approximation of the number of units eligible for benefits, or they can be more complex, designed to provide policymakers with more precise estimates of eligibility and a tool capable of capturing the potential effects of alternative policies on eligibility. Our discussion in this paper focuses primarily on the more complex models of eligibility traditionally used by the federal government to monitor program eligibility and participation. However, both types of models need to follow the basic steps described below and both types of models require high quality household survey data with the characteristics described below.

Due to the diversity of programs and the varying populations they serve, the microsimulation approach entails a significant number of steps. Figure 1 sketches the major operations involved in microsimulation models of social welfare programs. The figure shows that a microsimulation model typically begins with a survey of the population. 4 Most models include a step that converts the data to a particular format that supports easy and efficient processing. This step might also include methods that augment the basic survey data with additional information about families not typically collected on surveys but required for estimating program eligibility. The next step comprises the key model component, the simulation of current program rules. Administrative program rules are applied to the appropriate individuals or families within the household to determine whether they are eligible for benefits under each program included in the model. The final step is the selection of program participants. In selecting program participants from those found eligible for assistance, models typically rely on program participation information reported by individuals in the survey, combined with administrative data about the characteristics of the caseload. The discussion below goes into each of these elements of microsimulation models a little more deeply, starting with the essential characteristics of household surveys that provide the base upon which these models can be built.

Essential Characteristics of Household Surveys. The survey data that underlie the model of benefit eligibility are the most important ingredient. The survey data must meet several criteria. The survey needs to be representative of all units that might be eligible for benefits and include information about all members of those units. The survey needs to be large enough so that key subgroups of interest can be highlighted for analysis. The survey data also need to be of high quality, providing reliable information about key characteristics of program units such as their employment status, education, and health characteristics. The survey needs to provide the breadth of information sufficient to support simulation of state program rules. Finally, the survey must provide timely information about potentially eligible families and individuals. We expand on these survey characteristics below.

Representative. The survey must represent the population being analyzed, and it must include weights that add up to the actual population being analyzed. Representativeness can be a particularly challenging issue for programs designed to serve the low-income population, some of whom will be difficult to capture (because of homelessness, the lack of a telephone, or the lack of a fixed address). Ideally, surveys should attempt to minimize the undercount as much as possible within financial constraints and then adjust for any remaining undercount through the weights. The U.S. decennial census is generally used as a benchmark that represents the total population by state (although it too suffers from an undercount problem) and analysts design survey weights so that weighted samples represent the population count found in the Census (by age, race, sex, and sometimes additional characteristics).

Survey Size. Surveys need to provide sufficient observations on individuals and families so that benefit eligibility estimates are statistically reliable. Sufficient sample sizes also allow analysts and policymakers to address questions about particular subgroups of the eligible population. For example, states may want to examine eligibility rates for families with different structures (such as single parent, two parent, and children living with nonparents), with different income levels (above and below 50 percent of poverty, for example), and living in different parts of the state (such as rural vs. urban areas). As discussed later, standard errors of the estimates can be calculated and used to provide confidence intervals around particular answers provided. But one easy rule of thumb is that the survey should provide at least 100 observations for each "cell" of interest.

Reliability. The quality of the survey also depends on response rates to the questionnaire. Models of benefit eligibility depend on having good information on the demographic and economic characteristics that determine eligibility. Procedures must be developed to impute information when a critical variable has considerable missing data. Imputation procedures typically use information provided by individuals with similar characteristics to "fill in" these missing values (see Coder, 1999 for a description of methods used in large household surveys).

Breadth. Microsimulation models typically require that the survey provides a breadth of information about families and individuals because social welfare programs typically examine a large number of factors before approving an application. Table 1 outlines the kind of information that would be required to model all the eligibility rules of four major existing federal and state government social welfare programs (TANF, food stamps, Medicaid, and child care). While much of the information required is similar across programs--basic demographics, health and employment status, for example--there are also some important differences. The new TANF programs present some particular challenges. Historical program participation is now required for eligible units because individuals may currently be sanctioned, have exceeded states’ time limits (temporarily or permanently), or be in diversion status. Each of these conditions could prevent eligibility for the family on a temporary or permanent basis. Some programs require expenditure information from families (including child care, medical, and shelter expenses) to calculate deductions from income. Finally, most programs take into account a family’s assets in determining eligibility.

Timeliness. Finally, microsimulation models of benefit eligibility require timely survey data. Program administrators need to rely on these models to provide estimates of current eligibility and participation and to assess the potential impact of proposed changes to program rules, given families’ current demographic and economic conditions. Another factor that can be quite complex in developing models of benefit eligibility is the unit of time used in the analysis. Most surveys provide a "point in time" picture of the population, often representing income and employment currently or during the prior year. Most programs, however, provide benefits on a monthly basis. Families and individuals may be eligible for benefits for part of the year since employment and incomes fluctuate over time. Ideally, models would have access to monthly data in the underlying surveys, but this is not always practical. Therefore, modelers may develop procedures to "estimate" the pattern of income over a year using information on annual spells of employment and unemployment as reported by the surveyed individuals.

Essential Elements in Models that Estimate Program Eligibility and Participation. As depicted in figure 2, there are essentially six steps in a model of program participation: 1) defining the filing unit, 2) estimating categorical eligibility, 3) determining asset eligibility, 4) determining income eligibility, 5) calculating the benefit, and 6) making the participation decision. The level of complexity of each step and the specific procedures vary across benefit programs. Below we summarize the general goals of each of these steps.

Filing Unit. One of the more challenging steps in developing a model involves dividing up the sampled household into potential filing units. A filing unit is the group of persons that jointly applies for the benefit. Each government assistance program defines filing units differently, and in some programs multiple types of filing units are allowed. Filing units can be entire households, families, married couples, or individuals. For example, the TANF filing unit is typically composed of parents and their children, the food stamp filing unit consists of all individuals within the household who buy and prepare food together, and the Medicaid filing unit may be a unit defined under another program (TANF or SSI), or may be composed of adult individuals or children of different ages, with or without their parents. The simulation model should divide the household into the filing units appropriate to the program being simulated. Models get particularly complex when they attempt to capture several benefit programs simultaneously, all of which have different filing units.

A lack of data on key elements of the filing unit definition may force analysts to rely on general assumptions in constructing filing units using household survey data. For example, most household surveys do not provide information on which household members buy and prepare food together, complicating the identification of potential food stamp filing units. In addition, areas of caseworker discretion can be difficult to capture in a simulation model.

Categorical Eligibility. Categorical eligibility simply means which types of units can apply for benefits. Some programs only serve certain categories of the population such as the disabled, or families with children who have an incapacitated or unemployed parent. While the categorical eligibility step is often straightforward, typically requiring basic demographic information, welfare reform has increased the data requirements for simulating categorical eligibility. For example, TANF requires knowledge about families’ history of participation, eligibility status can be affected by a family’s current sanction status, and food stamp eligibility for noncitizen legal immigrants depends on their work history in the U.S. Survey designers must be careful to include all of the pieces of information required to simulate categorical eligibility. For example, federal household surveys typically do not include a person’s work and program participation history, yet this information is increasingly important in determining eligibility for benefit programs.

Asset Eligibility. Transfer programs generally require participants to pass an assets test. Under an assets test, the value of assets held by a filing unit is compared to a maximum amount to determine if the unit is eligible. For example, the Food Stamp Program denies eligibility to households with countable resources above $2,000 (above $3,000 for households with a member aged 60 or older). Countable resources include such things as bank accounts and the value of vehicles in excess of a certain amount. Unfortunately, asset information is often unavailable in household survey data. Financial assets can be imputed from asset income (such as interest and dividends) by assuming a given rate of return. Nonfinancial assets, such as vehicles, require more complex imputation procedures and are not always captured by simulation models.

Income Eligibility. In addition to passing an assets test, transfer programs typically require filing units to pass an income eligibility test. Under an income eligibility test, a filing unit is not eligible for assistance unless its income is beneath a certain maximum. A transfer program may require the filing unit to pass a gross income test, based on its total income, and also a net income test, based on its income after certain deductions.

Simulation models require a considerable amount of detailed information on income and expenses in order to simulate income eligibility tests. For example, since a sampled household may be divided into multiple filing units, the model must have information about the income received by each individual within the household. Since transfer programs typically base eligibility on monthly income, the model must know how much income was received in each month of the year. Different sources of income may be treated differently (as with earned income disregards under TANF), and so information must be available separately for each source of income. Finally, transfer programs may allow deductions for child-care and work related expenses and for other expenses such as medical and housing expenses, requiring simulation models to have access to detailed information about each filing unit’s expenditures. Where sufficient detail on income and expenses is unavailable in the source data for a simulation model, statistical techniques must be applied to impute the necessary level of detail.

Benefit Calculation. Once a filing unit has passed all required assets and income eligibility tests, the simulation model must compute the specific amount of the benefit. The transfer program’s benefit formula is applied to the filing unit and a benefit is calculated based on the filing unit’s income, size, and other relevant characteristics. In general, a simulation model should perform the same calculations that a caseworker would perform in calculating the filing unit’s benefit.

Program Participation. The final step in the simulation involves selecting which of the filing units eligible for assistance actually choose to apply for assistance and participate in the program. Although most household surveys ask members whether or not they receive assistance under various programs, this information alone is insufficient for selecting participants. Program participation is typically underreported in these surveys, so relying on reported participation alone will understate the number of program participants. The technique for selecting program participants varies by simulation model, and within a single model may vary depending on the program being simulated. Ideally, reported participation will be taken into account in selecting program participants, with the remainder of the participants selected in such a way that the simulated caseload resembles the real world caseload as closely as possible in terms of the overall number of recipients, different types of participating units, and the numbers of units participating in different states. The participation methodology must also capture behavioral response to a change in eligibility or potential benefits that results from a change in program rules.

3. Example Models and Studies of Program Eligibility

Historically, the primary sources of program eligibility estimates have been Mathematica Policy Research (MPR) and the Urban Institute. These organizations maintain periodically updated models for the U.S. Department of Health and Human Services (Urban Institute) and the Food and Nutrition Service of the U.S. Department of Agriculture (Mathematica) that are used to understand eligibility and participation and to simulate the effects of alternative program rules. Aside from these models, studies by Scholz 5, Blank and Ruggles 6, and McGarry 7 have produced one-time estimates that focus on a particular aspect of eligibility. Table 2 summarizes these major models and recent studies, providing the underlying data bases and rough estimates of program participation.

A major difference between the estimation techniques involves the choice of data. The two most common sources are the Current Population Survey (CPS) and the Survey of Income and Program Participation (SIPP).

The Transfer Income Model (TRIM3 at the Urban Institute) and the MATH (Mathematica) models both use data from the CPS March Supplement. The CPS March Supplement collects annual data from a scientifically chosen nationwide sample of more than 50,000 households and provides detailed demographic, labor force activity, and income information. The CPS contains a number of inherent benefits for microsimulation: it provides a large population with which to simulate (although the sample is not completely representative of the U.S. as some groups are undersampled) and there is a great deal of information regarding family structure and characteristics, labor force participation, and wage and salary income. Despite this, there are a number of flaws that make using the CPS for eligibility simulation difficult including the lack of information regarding monthly income (annual income is provided, but eligibility is determined on a monthly basis), expenses, transfer payments, and assets (especially vehicle holdings).

Mathematica’s MATH-STEWARD model and the Scholz, Blank and Ruggles, and McGarry studies are based on data from an increasingly widely used alternative to the CPS--the SIPP. Starting in 1996, the SIPP included more than 35,000 households sampled twelve times in a four-year period regarding their activity in the preceding four months (i.e., data are collected on the same families for each month over four years). The SIPP was designed explicitly to overcome some of the shortcomings of the CPS 8. Since the SIPP consists of panel data, it provides monthly income and asset information and captures other short term changes that annual data, like that provided by the CPS March Supplement, cannot. In addition, more detailed questions regarding transfer payments, taxes, income, and assets are included in the SIPP.

Unfortunately, the SIPP also presents its own unique challenges for use in microsimulation. The sample is substantially smaller than that of the CPS--sampling around 36,000 households from 1996-2000 compared to more than 50,000 per year in the CPS--and attrition will reduce the SIPP number further. There are also considerable delays in SIPP data releases making the gap between data collection and distribution too long to be used for some evaluations. (The latest SIPP data set available is for 1996, calendar year 1995.) In addition, the SIPP has only been in existence since 1984 and has gone through a number of major changes since then (expanding to its present form in 1996) and some years are unavailable, making historic microsimulation difficult if not impossible. One additional challenge presented by SIPP involves the panel nature of the survey. Respondents are interviewed at four month intervals, but some respondents drop out of the survey with each new panel (referred to as attrition). This means that an annual file created from three waves of the SIPP may no longer include a representative sample of the U.S. In addition, the file structure does not lend itself easily to providing a full annual picture of income and assets that is often needed when models are needed for analyses of program alternatives affecting an annual budget period. However, the Urban Institute recently created a "CPS Look-Alike" file from three waves of the 1993 and 1994 SIPP panels, providing one methodology for overcoming this constraint.9

Aside from the data differences, the models and studies presented in Table 2 differ in their focus. While the goals of TRIM, MATH, and MATH-STEWARD are more broad, the other studies focus on more specific aspects of eligibility.

The TRIM model was created to simulate the entire government tax and transfer system so as to allow for complete distributional analysis as well as individual program analysis 10. SSI, AFDC, Food Stamps, subsidized housing, Medicare, Medicaid, Employer-sponsored health insurance, payroll taxes, federal income taxes and state income taxes are all simulated. More recently, eligibility for child care assistance has been added to TRIM, and other aspects of these programs are currently being added to the core model. The comprehensive nature of TRIM attempts to account for the underreporting of government benefits and thus provides a more accurate picture of total income.

Similarly, the MATH models were designed to simulate all income transfer programs and account for the effects of AFDC/TANF and SSI changes on Food Stamp eligibility. As with TRIM, users can experiment with a number of policy changes across programs to simulate the effects taking into account cross-program interactions.

The other studies focus on more specific topics. The McGarry paper estimates SSI participation within the elderly population using SIPP data. The Scholz study presents a simulation of EITC participation that uses SIPP data linked to IRS tax information. This allows for eligibility to be determined based on the family’s actual tax data, providing more accuracy than do any of the surveys. The Blank and Ruggles paper estimates AFDC and Food Stamp participation for single mothers. In the paper, they estimate not only eligibility, but also the number and length of spells on assistance, allowing analysis of eligibility over time.

All together these studies have given policy makers a better understanding of family eligibility and program participation for various types of government assistance. The TRIM model and MPR’s family of models have played an important role in analysis of policy alternatives in the areas of welfare, health, and food stamps. The federal government has traditionally used these models to analyze potentially narrow and broad changes to the U.S. tax and transfer system. 11

4. AFDC/TANF: An Illustrative Eligibility Model

This section focuses on the way one model--the Urban Institute’s TRIM model, version 3 --simulates the AFDC and TANF programs to illustrate how microsimulation models operate in a more concrete way. This model is updated each year to capture the TANF rules actually implemented during the "income" year represented by the March Current Population Survey. 12

The discussion generally follows the generic steps in developing a model of eligibility described earlier. However, we also point out how welfare reform has presented some new challenges in modeling eligibility, and we point out some important places where survey and administrative data lag behind the needs of a model of TANF eligibility. It may be easier to fill these gaps at the individual state level, however, because each state knows its rules intimately and some have the administrative data that would fill in the gaps mentioned below.

The first step in any model development requires understanding how the rules of the program work. This was relatively when the federal government had primary responsibility for the Aid to Families with Dependent Children (AFDC) program and states were given few waivers to vary their programs in significant ways. While states had the flexibility to set their own income eligibility standards and benefit levels, few other rules varied across states under AFDC. While the waiver era introduced new state variation that could affect benefit eligibility, states had to report these variations in detail to the federal government, providing a centralized point for information that could be used to model AFDC across the 50 states. Welfare reform has presented a considerably new challenge for modeling TANF. The federal government no longer requires states to report their TANF plans in sufficient detail to support modeling TANF. While states must submit TANF plans every two years, these documents often contain only the broad themes of state plans insufficient for modeling TANF at the family level. The Urban Institute has developed a "Welfare Rules Database (WRD)" that details the implementation of states’ TANF plan rules from 1996 through 1999. 13 Information for the WRD has primarily been drawn from caseworker manuals from all 50 states.

The TRIM3 model has used information in the data base to simulate the rules of each state’s TANF program to update its original model of eligibility for AFDC. While TANF has introduced considerable program variation across the states, nearly all states’ plans follow the general structure of AFDC. The discussion below highlights the major differences introduced under TANF and the particular challenges presented by those changes.

Defining the Filing Unit. The first step in TRIM3's AFDC/TANF model is to divide the CPS household into one or more potential filing units. In TRIM3, the initial filing unit for AFDC and for TANF is a nuclear family--one or two parents (or a non-parent caretaker) and the children.14 TRIM3 makes some simplifying assumptions in setting up filing units for AFDC/TANF. In reality, some assistance units could include members of three generations--for example, a middle-aged woman, a teenaged daughter, the daughter’s children, and the woman’s younger children. But there were never federal rules about whether a family like this should be treated as one or two units. In addition, some states allow eligibility for "essential persons"--adults other than the parents/caretakers who provide essential services for the well-being of the child. TRIM3 does not model the complexities of three-generation units or essential persons because eligibility for these types of units is relatively infrequent, and, presumably, involves considerable caseworker discretion.

Determining Eligibility. TRIM3 determines eligibility for AFDC or TANF benefits for each filing unit on a monthly basis. A family might be eligible in some months of the year but not others. TRIM3 captures most of the state-level variation in eligibility rules. The model uses the majority rule in effect for most of the state’s population for most of the year to simulate eligibility. County-level variation is not captured for both practical and technical reasons. Not only would this level of detail be unwieldy even if the information were known, but it is not always possible to determine a household’s county of residence in the CPS. This could become a problem under TANF if more states allow significant county-level variation in the future.

The CPS information about families and individuals is insufficient to model all the complexities of AFDC and TANF eligibility rules. TRIM3 imputes some of the information essential for modeling eligibility--asset values, immigrant status, and child care expenses--to families represented in the CPS records, but omits some other information--such as stepparent status and vehicle value. The imputation decision typically depends on resource constraints, the potential data and methods available to support imputation, and the relative importance of the rule that could otherwise not be modeled. As programs change, new imputations are often considered to support new elements of federal and state programs.

Categorical Eligibility. Categorical eligibility has become particularly difficult under TANF because programs not only use basic demographic criteria to determine eligibility, but a family’s eligibility may also depend on their current program status. Below we describe the basic demographic eligibility rules, special rules for two-parent families, and the complications introduced by the new welfare program rules.

Basic Demographic Eligibility. TRIM3 uses the CPS data on family relationships, age, student status, and Supplemental Security Income (SSI) program participation to model the basic demographic eligibility rules. To be eligible for benefits, a person must be either an eligible child or the parent or caretaker of an eligible child. An eligible child is either under 18 or, in some states, over 18 but a student. Children who receive SSI are not eligible, but their parents may still receive AFDC/TANF for their own needs. The parents of eligible children are categorically eligible unless they receive their own SSI benefits. In determining SSI status, TRIM-simulated data on SSI receipt is used to augment the CPS-reported information, to correct for under-reporting and the fact that SSI receipt is not recorded for persons younger than age 15 in the public-use data. When children live with non-parent caretaker relatives (grandparents, aunts, etc.) TRIM3 examines the age and employment status of the non-parent caretaker to determine whether to include her in the unit. Imputed information on immigration status may also be used to deny eligibility to some aliens as required by federal or state law. 15 Unmarried teen parents living away from their parents are generally ineligible under TANF and under some states’ AFDC waivers, but TRIM3 randomly assigns eligibility to some of these young women to account for the effects of exemptions. The actual exemption rules are complex and the CPS does not include data to distinguish exempt from non-exempt teens unless they actually report receiving benefits.

TRIM3 does not capture a few aspects of demographic eligibility. TRIM3 does not exclude step-parents from the TANF unit because the CPS does not identify stepparents (this will be possible with simulations based on the NSAF, however). Pregnancy cannot be flagged on the CPS, and TRIM3 does not currently try to simulate TANF eligibility for pregnant women with no other children.

Two-Parent Families. TRIM3 may apply special rules to the eligibility of two-parent families. In the case of the incapacity of either parent, both parents are eligible. For other two-parent families, TRIM3 can impose special tests to see if the family qualifies for "UP" (unemployed parent) benefits. These tests were established at the federal level under AFDC, but were relaxed by many states’ waivers and TANF programs. Thirty-four states now determine eligibility for two-parent families the same as for single-parent families, but other states impose at least some requirements related to current or prior work effort. 16 TRIM3 can require the "principal earner" to be either not working or working less than a certain number of hours per month--this was the rule under AFDC, and this or similar rules have been retained in some states. This type of rule can be applied to all participants or just applicants, depending on actual state practice, by keying off of the model’s monthly eligibility estimates. TRIM3 also captures 30-day waiting periods and requirements that the principal earner be in the labor force using data from the CPS on current work activity, rules from the federal AFDC program that continue in some states. The federal AFDC rules also applied a work history test, requiring the principal earner to have worked at least 6 out of the prior 13 calendar quarters, but TRIM3 can not model these rules, or similar state TANF rules, due to lack of sufficient work history information in the CPS.

Program History: Time Limits and Sanctions. Benefit termination time limits and sanctions may eliminate eligibility for all or some members of the family. Not all families are subject to time limits and work requirements that may result in sanctions, however, so the first step in this modeling process is to exempt some families from these rules. Exemption criteria vary by state, but usually include disability, caring for a disabled person, caring for a very young child, or being over a certain age. These rules can be simulated using the variation in rules across the states and the information provided on the CPS.

Under the federal TANF rules, a parent who has received TANF benefits for five years as an adult (as the head of the unit) can no longer receive benefits for her or his family unless specifically exempted by the state (usually on the basis of disability). This is in stark contrast to AFDC, when some states restricted their AFDC-UP caseload to 6 months of benefits out of 12, but no time limits were applied to single-parent families in any state. TANF is complicated by the fact that some states instituted other time limits under state waivers and may continue to use those limits along with the 5-year limit--such as a rule that families can receive benefits for only 24 out of 60 months. Thus, the ideal survey for TANF modeling would ask families not only for their total months receiving aid since TANF began, but also for the timing of that receipt, and welfare histories would extend back to the time period when the first state time limits were instituted. With this type of data, the welfare history could be compared to the applicable limits to determine current eligibility status for nonexempt families.

The CPS does not include welfare histories. Recent TRIM3 AFDC and TANF simulations (1996 and 1997) did not impose time limits for purposes of determining eligibility because so few families were affected across the country. We have added rough procedures to simulate time limits for 1998, however, because more families have been affected. The model will exclude families from eligibility due to time limits using fairly crude data gathered from the states on the number of families affected and characteristics of families on welfare for relatively long periods of time (who were not likely to be exempt due to disability) from pre-TANF administrative data. We will be refining these methods as better administrative data become available (beginning in 2000). In addition, the second round of the Urban Institute’s National Survey of America’s Families will provide some additional data on families recently affected by time limits in 13 states (that account for about 60 percent of the TANF caseload).

Sanctions present another challenge. Under AFDC, parents could be sanctioned for failure to comply with certain rules by having their needs removed from the unit for a set number of months. Under waivers and TANF, both the use and potential severity of sanctions has increased. Some states have instituted "full family" sanctions that deny benefits to the entire family, and the final sanction in certain states results in an individual’s or family’s permanent ineligibility. Other sanctions reduce a family’s benefit by a particular dollar amount, or by removing from the assistance unit the adult who failed to comply with requirements. Most surveys do not ask whether individuals have been sanctioned, and surveys that do (such as the NSAF) cannot guarantee that individuals can readily identify why they no longer receive benefits.

Full-family sanctions and benefit-reduction sanctions present different issues for modeling. Conceptually, sanctions that temporarily or permanently take away a family’s full benefit could be considered as part of the TANF eligibility determination, or they could be considered as a decision on the part of individuals to not "follow the program rules" and, therefore, not participate for some period of time. (Of course, this assumes that individuals can rectify the sanction through the reconciliation process which may not be true in all states or for all individuals.) At the present time we are not specifically trying to model full-family sanctions within TRIM3, but choose to incorporate full-family sanctions indirectly into the participation process. That is, families who have lost their entire benefit will be among the families who are found to be eligible but not participating. Benefit-reduction sanctions will be modeled by selecting a subset of the families who are non-exempt from work requirements, and reducing their benefits consistent with their states’ rules. The features of the model relating to sanctions are fluid, however, and as better data on sanctions become available through states’ administrative data reporting requirements, we may modify the procedures in TRIM3. 17

Income and Asset Eligibility. To qualify for benefits, a family must pass financial eligibility tests--usually, both an assets and an income test. The types of rules that were imposed under AFDC still exist in nearly all states’ TANF programs, although there is much more variation across the states. These tests merit separate discussion because they present different data challenges.

Assets Tests. In TRIM3 as in the actual AFDC and TANF programs, the value of a family’s financial assets must be less than the maximum amount. Under AFDC, almost all states used the federally-established maximum of $1,000, although a few chose lower amounts. Under TANF, many states have liberalized their asset restrictions, and one state no longer imposes an assets test. The CPS does not include the value of a family’s financial assets, but does include the value of income from assets (interest and dividends). TRIM3 infers the value of financial assets from the value of the asset income. This certainly leads to some inaccuracies--due to under-reporting of asset income and due to variations in actual rates of return--but many studies have confirmed that most families who have incomes low enough to qualify for eligibility do not own financial assets.

TRIM3 does not currently capture the rules that include in assets a portion of the value of the family’s automobile because the CPS does not include any information on vehicle values. Under AFDC, equity value in excess of $1,500 (or less, at state option) was included in assets. Many states have liberalized this limit under TANF, although most still include a portion of a car’s value. The CPS does not contain information about car ownership or value, but imputations could be developed from the SIPP or the Consumer Expenditures Survey (CES). (This rule is very important for food stamp eligibility, and MPR has developed sophisticated imputations to simulate the fair market value of cars for low-income families for its MATH and MATH-STEWARD models).

Income Definition. For purposes of modeling financial eligibility, TRIM3 adds up the appropriate income amounts using the detailed income information available on the CPS, computes the value of various deductions from earnings, and computes "deemed" income. Slightly different income definitions may be used for different income tests and for benefit computation.

For the gross income calculation, types of income that are excluded include general assistance, payments under the Earned Income Tax Credit (EITC), and most earnings of children. If child support is received from a non-custodial parent, most of the amount is considered income for determining the family’s eligibility, less any amount that will be "passed through" to the family after payment to the state. Earnings deductions have been modified by many states since the advent of welfare reform, but TRIM3 can capture either the pre-TANF or post-TANF rules. Work expense deductions were equal to a flat amount under AFDC, but are sometimes equal to a percentage of earnings under TANF. Child care deductions are based on actual expenses, so the TRIM3 system includes an imputation of those expenses. Additional earnings disregards like the "30 and 1/3" disregard under the AFDC system, and liberalized versions of that disregard under TANF, are also computed.

In cases where a minor parent lives in the same household as her parent(s), now a federal rule except in some circumstances, TRIM3 captures the rules that "deem" a portion of the grandparents’ income as available to the assistance unit. Actual program rules also deem income from step-parents who are not in the unit, but those rules are not relevant to the model since step-parents are not identified.

Income Tests. TRIM3 can model numerous types of income tests, depending on the actual rules in place in a particular state and year. Under AFDC, a family applying for benefits had to pass an initial test of having net income (after deductions for work expenses and child care expenses, but not the "30 and 1/3" deduction) under the state’s "need standard." Then, in every month, a family’s gross income (without any deductions) had to be less than 185 percent of the need standard. States’ waivers and TANF programs instituted many changes to these eligibility tests. Some changes include removing the gross income test, modifying the income limits for passing net or gross income tests, modifying the allowable deductions from income prior to applying certain tests, and varying the tests imposed on new applicants vs. continuing recipients. TRIM3 can capture these variations in rules because of the rich data about work on the CPS and because all of the variations in the states’ income tests have been coded into the WRD and the TRIM3 data bases.

Benefit Calculation. For each month that a family passes all the categorical and financial eligibility tests, TRIM3 calculates a potential benefit. In the simplest case, the benefit equals the "income deficit" produced by subtracting the appropriate definition of income (under AFDC, gross income less all earnings-related deductions) from the state’s "payment standard" for a family of each size. More complex formulas involve a maximum payment that is less than the payment standard, and/or the payment of only a percentage of the income deficit. TRIM3 models all of these variations.

Family caps--imposed in certain states under waivers and now under TANF--may alter the potential benefit, and TRIM3 attempts to capture the impact of these rules. Under most family cap policies, a child conceived while a family receives benefits, and after the implementation of the policy, is excluded from the unit. To fully model these rules would require welfare histories to compare to the ages of a family’s children. In the absence of those histories, TRIM3 identifies families with children born or conceived since a state’s rule was imposed, and selects a subset to be treated as affected by the rule. The percentages have been obtained from tabulations of the AFDC quality control data, which includes ages of children as well as the number of months each family has received benefits.

Participation. As noted elsewhere, many families who are eligible for benefits do not enroll in a program. TRIM3 selects participants from among the eligible families such that the size and characteristics of the simulated caseload come sufficiently close to the size and characteristics of the actual caseload according to administrative data. Key targets include caseload by unit type (two-parent families vs. other families), presence of earnings, and state. Families who are simulated as eligible for benefits and who report receiving benefits in the CPS are always included in the caseload in "baseline" simulations--simulations that use the actual rules that applied in the year of the CPS data. These "reporter" families are insufficient to meet targets, so additional caseload is selected from among the eligible families who do not report benefits. A probability of participation is calculated for every month of a family’s eligibility, varying by the size of the potential benefit and the unit’s demographic characteristics. If a family’s income varied over the year, the potential benefit and thus the probability of participation may also vary. Thus, a family might be selected into the caseload in some but not all months in which it is eligible.

The methods operate in such a way that the results of simulations of hypothetical rules are internally consistent with the results of the baseline simulations. In a hypothetical simulation, if a family becomes eligible for higher benefits than in the baseline, the family will continue to participate if it participated in the baseline, and may begin participating even if it did not participate in the baseline. A family whose potential benefit declines will remain a non-participant if it did not participate in the baseline, and may stop participating if it did participate in the baseline. TRIM3 can also model short-term scenarios under which participation decisions are unchanged from the baseline.

TRIM3's participation methods still need further revision to capture the new TANF environment. It has always been true that families eligible for lower benefits were less likely to participate, but they may become even less likely to take a low benefit if that will "use up" a month of eligibility under a time-limited system. Updating participation functions will require administrative data on the distribution of benefits received in each state. Also, diversion programs may alter the makeup of the caseload by appealing to some types of families more than others. Currently, TRIM3 does not try to identify which of the non-participating families might have been diverted, and does not calculate the amount of their lump-sum payments. While data on diversion does not current exist, more information will become available as states begin to fulfill the new TANF data requirements. (Data should become available in 2000).

5. Illustrative Eligibility Estimates for AFDC in 1996

Estimates of program eligibility and participation rates demonstrate their potential value to states. Microsimulation models provide eligibility and participation rates for each program of interest, and they can show eligibility and participation rates for population subgroups. The former demonstrate the extent to which program eligibles choose to participate in a program and provide state program administrators an idea of the potential program costs, should all eligibles choose to participate in a program. The latter demonstrate the extent to which programs reach targeted population subgroups and the extent to which population subgroups choose to participate. Higher (or lower) participation rates among population subgroups with particular characteristics can potentially indicate which groups should be targeted for more extensive outreach should program administrators desire to ensure that programs maximize participation.

In this section, we use the TRIM3 model to illustrate recent eligibility and participation rates for AFDC/TANF. We present these estimates for calendar year 1996 based on the March 1997 Current Population Survey (CPS) for the nation and for four states--California, Texas, New York, and Florida. 18 While the CPS is representative of all 50 states, the individual sample sizes for most states are too small to show individual state estimates. The states for which we show separate estimates rank in the top four states in the CPS for total population and total number of families or persons eligible for AFDC/TANF benefits, and they provide the largest state sample sizes on the CPS.

While these estimates primarily reflect the federal AFDC program, they do take into account the major state waivers in effect during 1996. The waivers typically focused on increasing incentives for work (often through expanded earned income disregards), marriage incentives (by eliminating special work rules for two-parent families), and one of the four states (Florida) had implemented a family cap. Some states were using tougher sanctions than allowed under the AFDC/JOBS program and tougher work requirements. However, as noted above, neither sanctions nor work requirements are modeled explicitly in TRIM3. Federal welfare reform (TANF) was passed in August 1996. While a few states began implementing TANF in the last quarter of calendar year 1996 (mostly those states adopting their waiver strategies combined with new federal program requirements), most implemented new TANF plans in mid-1997. As noted earlier, TRIM3 makes the simplifying assumption that one set of rules (those in place for the majority of the months) are in effect throughout the year. Thus, the rules that began in some states in the last quarter of the year do not affect the estimates presented below.

First, we present estimates of AFDC program eligibility and participation. That is, we show the number of "filing units" eligible for benefits, and the percent of those eligible who are participating. Second, we show eligibility and participation for two subgroups of interest--low-income families with children and low-income single parent families with children.

Program Eligibility and Participation. Most microsimulation models produce estimates of program eligibility as a basic output. The ratio of participation to eligibility, measured using administrative data from program records, provides an estimate of the program participation rate. In calendar year 1996 average monthly participation rates for AFDC were relatively high (table 3). 19 Nearly 80 percent of families eligible for AFDC benefits participated in the program, and participation rates varied somewhat across the states. Families eligible for AFDC benefits in California were more likely to participate, and families in Texas and Florida were less likely to participate, compared to the national average.

While reasons for variations in participation rates are not completely understood, as we mentioned earlier, we can discuss these results in the context of variations in benefit levels. In general, the higher the value of benefits to families, the higher their participation rates. AFDC maximum benefit levels varied considerably across the states in 1996 ranging from a low of $120 per month for a three-person family in Mississippi to $633 in Vermont, and even higher levels in Alaska and Hawaii. For the four states shown in these tables, the maximum monthly benefits for a family size of three were $596 in California, $303 in Florida, $577 in New York, and $188 in Texas. The AFDC participation rates track these maximum benefit levels to some extent with California having the highest participation rate. However, Florida and Texas have similar participation rates despite variations in benefit levels, and New York’s participation rate is essentially at the national average, despite having a higher than average maximum benefit level. Of course, many other factors affect participation in AFDC, including variations in states’ economies, cultures, family incomes, and, in 1996, the extent to which welfare reform was already underway through the waiver process. 20 That is, some states were already providing recipients with a strong “work first” message, encouraging some families to apply for employment rather than welfare. This would reduce the participation rate only to the extent that some families may have been discouraged from applying for benefits even though their income was below the eligibility threshold.

Column two of Table 3 also demonstrates the effect of benefit level on participation. As shown, the ratio of eligible benefit dollars paid by the programs is considerably higher than the ratio of participants to eligibles. According to these TRIM3 estimates, 95 percent of potential AFDC benefit dollars were paid in 1996 nationwide. The ratio of paid benefit dollars to eligible dollars in California is 99 percent. While the standard error for this estimate is very low (0.022 percent, it primarily reflects the use of the administrative data in the numerator (minimizing the effect of sampling error in the CPS), and it does not take into account simulation error.21 That is, the TRIM3 model may underestimate eligible units in California (and other states), but we also expect that the CPS sample may also have contributed to low eligibility estimates (and relatively high participation rates). The CPS undersamples parts of the low-income population, especially immigrants with language barriers and households without telephones, and weights cannot completely adjust for this source of error. 22

Eligibility for Low-Income Families with Children. Table 4 shows the reach of the AFDC program among the low-income population with children in 1996. 23 This table shows eligibility and participation rates for families at some time during calendar year 1996. Forty-two percent of all low-income families with children were eligible for some AFDC benefits, ranging from a low of about 30 percent in Texas to 54 percent in California. These eligibility differences across states reflect the differences in their benefit levels relative to incomes in the state, with the low benefit states offering benefits to a smaller proportion of low-income families. Participation rates for low-income families with children track those shown earlier in table 3. (The somewhat higher rates reflect the measure of "ever on" during the year compared to the "average monthly values used with the administrative data shown earlier".)

The last column of Table 4 provides another program benchmark, the percent of low-income families receiving AFDC benefits sometime during the year. About one-third of all low-income families with children received some AFDC benefits during the year across the U.S. Benefit receipt rates varied across the states from a low of 22 percent in Texas to a high of 47 percent in California. These receipt rates obviously reflect differences in both the eligibility and program participation rates across the states.

Eligibility for Low-Income Single-Mother Families. Table 5 shows analogous statistics for low-income single-mother families. 24 As shown, the penetration of the AFDC program was greater for low-income single-mother families across the states than for all low-income families with children, primarily reflecting the lower economic status of single mothers. (All states offered AFDC benefits to two-parent families in 1996, although some limited these benefits to six months.) Nearly 60 percent of single-mother families were eligible for some AFDC benefits across the U.S. in 1996. Again the variation in eligibility rates for single-mother families is significant across the states, ranging from 46 percent in Texas to nearly three quarters in California. Single-mother families also have higher participation rates than all low-income families, reflecting their greater need for assistance and the potentially higher benefits offered to them.

These examples show only a few of the types of program eligibility and participation rates that may be useful to state program administrators. Microsimulation models can show participation estimates for any subgroup of interest (such as immigrants, those living in rural areas, or nonwhites) to the extent that the underlying survey data and model can support the subgroup results. Surveys obviously must contain the information that defines the subsample of interest and provide representative and sufficiently large sample sizes to support the breakouts by subgroup. (A rough rule of thumb is that at least 100 unweighted observations should be available in each cell examined.) In addition, the model must represent the key program rules affecting eligibility for various subgroups. For example, the Urban Institute just recently added parameters to model eligibility for noncitizens in its TRIM3 model because eligibility for legal immigrants began to vary across states and across programs as a result of federal welfare reform. This model addition required the imputation of legal alien status using other information provided on the CPS and data from the Immigration and Naturalization Service (INS) as well as model parameters and participation targets specific to immigrant groups. This relatively complicated addition was justified because of the importance of immigrant eligibility for government transfers to federal policymakers. States similarly will need to weigh the cost and benefits of more complex models should they pursue the development of microsimulation models for policy analysis.

6. Summary and Conclusions

In sum, eligibility models could provide useful tools for administrators of state programs offering assistance to low-income families. As states’ TANF programs continue to evolve and states use their new financial flexibility under TANF and other new federal funding streams to develop and offer new assistance programs that reach beyond the cash assistance population such as subsidized child care and transportation assistance, microsimulation tools that allow them to "try out" new policies before actual implementation could provide to be extremely beneficial. Models could provide states with estimates of program eligibles and costs under different program alternatives. Microsimulation models would also give states a better understanding of current program participation rates, providing estimates of the number of low-income families who are eligible for assistance. Comparisons of these eligibility estimates with administrative data describing the characteristics of families actually participating in TANF and other state-sponsored assistance programs would provide states with information essential for understanding their programs’ effectiveness. For example, the characteristics of nonparticipants could provide states with information to target outreach to particular state population subgroups.

The development of microsimulation models capable of estimating eligibility for states’ program benefits requires two key elements, along with a long-term research commitment by the state. First, states’ must have a high-quality, representative survey of their state’s low-income population to provide the underlying base for the model. The survey must provide enough information about their state’s population so that the rules of states’ programs can be simulated. This especially requires good information about family structure, health status, employment of the adults, and income. Ideally, the survey would also provide information about families’ histories of welfare participation. Alternatively, states’ may be able to link information from their administrative records systems to their survey (given that the same individual identifiers are available on both files). This would provide a higher quality source of information about program histories and would be preferred as long as data confidentiality issues can be resolved. (This is not an option available to those developing models at the national level that include information about all 50 states.)

Second, states need to develop microsimulation models of program eligibility. This should be considerably easier at the state level than at the federal level because each state only needs to incorporate the rules of its own TANF programs. Eligibility models generally require the sequencing of a set of logical steps designed to identify families who could receive assistance under states’ assistance programs and to select those most likely to participate in these programs. As noted above, considerable complexity has been introduced into national models of TANF eligibility as a result of federal welfare reform and the expansion of diversity among states’ assistance programs. These models require a detailed understanding of the TANF program rules in all 50 states, along with comparable data describing the characteristics of program participants across all of the states. This was considerably easier in the pre-TANF era when federal rules guided the AFDC program and all states were required to report data on their caseload characteristics through the Quality Control data system. As noted earlier, the new TANF data reporting system will be critical for updating national-level TANF eligibility models. And national-level microsimulation models will require updated, detailed information about states’ TANF programs such as the Welfare Rules Data Base developed at the Urban Institute.

FOOTNOTES

TABLE 1

General Information Requirements for Simulating Eligibility

Information	SSI	AFDC/TANF	Food Stamps	Medicaid	Child Care
Program Rules	x	x	x	x	x
About Households:
State of Residence	x	x	x	x	x
Relationships of all Members	x	x	x	x	x
Ages of all Members	x	x	x	x	x
Pregnancy status		x		x
Immigration status	x	x	x	x	x
Disability status of all Members	x	x	x	x
Employment status of each adult	x	x	x
Lifetime covered quarters of employment for aliens	x		x
Earnings of each member	x	x	x	x
Expenses - Child care - Shelter - Medical		x	x x x	x
Nonearnings income	x	x	x	x	x
Assests - Financial assets - Value of vehicles	x	x x	x x
Participation is SSI		x	x	x
Participation in AFDC/TANF			x	x
Eligibility for AFDC/TANF				x
Historical TANF Participation - Sanction status - Time limit status - Formal diversion status - Time since left AFDC/TANF - Years in current spell		x x x x		x	x