Cover Image

PAPERBACK
$93.75



View/Hide Left Panel

Model Design



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers Model Design

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers This page intentionally left blank.

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers 3 Alternative Model Designs: Program Participation Functions and the Allocation of Annual to Monthly Values in TRIM2, MATH, and HITSM Constance F.Citro and Christine M.Ross INTRODUCTION The Transfer Income Model 2 (TRIM2) and the Micro Analysis of Transfers to Households (MATH) model are two microsimulation models that are heavily used for analysis of proposed changes to government tax and transfer programs. The Household Income and Tax Simulation Model (HITSM) is another static microsimulation model that has been used for policy analysis in the income support and tax areas. All three models fall into the class of “static” microsimulation models—that is, models that operate on a cross-sectional basis and make projections to future years using procedures for reweighting, or “aging,” their database to match outside projections of selected characteristics such as the demographic and labor force composition of the population.1 (In contrast, dynamic models apply transition probabilities for events such as birth, death, job change, and others to the records in their database, thereby “growing” their population year by year into the future; see Ross in this volume.) Constance F.Citro is a staff officer at the National Research Council; she served as study director of the Panel to Evaluate Microsimulation Models for Social Welfare Programs. Christine M.Ross is on the staff of Mathematica Policy Research, Inc.; she served as research associate of the panel. 1   At present, users of TRIM2 do not normally invoke the model’s static aging routines but instead apply out-of-model adjustments to project TRIM2 results obtained from the most recently available database.

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers These three static models—TRIM2, MATH, and HITSM—share many features. They also differ in important respects. In this chapter we compare and contrast two components that are part of each model. Our goals are to (1) illustrate the challenging modeling tasks that confront the developers of microsimulation models, given the complexities of the government tax and transfer programs that they cover, the diverse characteristics of the individual decision units to which the programs apply, and the limitations of available databases describing these decision units; (2) illustrate the considerable differences as well as similarities in the approach to common modeling problems that model developers have brought to bear; (3) bring together the limited evidence on the implications of different approaches to common modeling problems; and (4) note problems that appear to warrant further attention and action. The first of the two model components that we examine includes the so-called months routine in each model that is used to allocate annual values for income and employment status to calendar months. The second component includes the functions used by each model to simulate the participation decision for the supplemental security income (SSI), Aid to Families with Dependent Children (AFDC), and food stamp programs. Included in the discussion of modeling program participation are the steps the modelers go through to “calibrate” their baseline data files to approximate selected control totals for program recipients obtained from administrative records. The descriptions in this chapter of the models’ months and participation routines are based primarily on review of each model’s technical documentation, supplemented in some cases by telephone conversations with staff members who maintain the models. The documents we examined for TRIM2 include those of the Urban Institute (1987) and Webb et al. (1982, 1986). For MATH they include Doyle (1989) and Doyle et al. (1990)—the version supplied to us in each case is limited to the modules and variables currently in use by the U.S. Department of Agriculture’s Food and Nutrition Service. For HITSM they include Lewin/ICF, Inc. (1988). The descriptions are not necessarily up to date, because some of the routines are being revised. For example, the Urban Institute is currently working on new AFDC and food stamp participation functions. CONVERTING ANNUAL TO MONTHLY VALUES IN TRIM2, MATH, AND HITSM Static microsimulation models such as TRIM2, MATH, and HITSM typically use the March Current Population Survey (CPS) as the primary input database to perform analyses of government transfer programs, such as AFDC and food stamps. The income supplement that is included in the March CPS provides total income amounts by source and summary data on labor force activity during the preceding calendar year for a large sample of households and their members. Transfer programs such as AFDC and food stamps typically use a

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers monthly accounting period to determine program eligibility. Hence, a necessary step in preparing March CPS files for microsimulation model use is to convert the annual income and labor force data contained in the March CPS records to monthly values.2 Research on patterns of intrayear income receipt and changes in labor force status suggests that microsimulation estimates of program eligible units will be sensitive to the way in which the annual to monthly income conversion is performed. A number of studies, based on the Seattle and Denver Income Maintenance Experiments, the 1979 Research Panel of the Income Survey Development Program (ISDP), and, more recently, the 1984 Panel of the Survey of Income and Program Participation (SIPP), document the fluctuations in income and labor force status that many individuals experience throughout the year. These fluctuations result in higher poverty rates computed on a monthly or other part-year basis compared with annual rates. Correspondingly, they result in higher rates of eligibility for programs like AFDC and food stamps, which provide support to people experiencing a spell of low income even though their average monthly income based on the entire year is above the program cutoff. (See, for example, David and Fitzgerald, 1987; Doyle, 1984a, 1984b; Lubitz and Carr, 1985; Ruggles, 1988; Ruggles and Williams, 1989; Springs and Allen, 1976; Uhalde, 1976; Williams, 1986.) Recognizing the importance of accurately simulating part-year income flows and labor force activity, developers of the major microsimulation models have progressively refined the calculation of monthly values for these characteristics in their models. Below we describe the procedures that are currently used to perform the annual to monthly conversion in the TRIM2, MATH, and HITSM models. Where information is available, we compare these procedures to those used in prior versions of the models. The discussion proceeds in the order of the model calculations, beginning with how the models determine monthly employment status, then how they allocate earned income, and, finally, how they allocate unearned income. Transfer income from AFDC, SSI, and food stamps is entirely simulated by the models, and, hence, the reported annual income amounts for these programs in the March CPS are not processed by the annual to monthly income conversion routines. One point to note is that none of the models adjusts other characteristics that can also vary during the year. In particular, the models assume that the household and family composition observed in March of the survey year remains unchanged during the previous calendar year over which income and labor force experience are measured. HITSM, in contrast to TRIM2 and MATH, subtracts 2   Models that use the new Survey of Income and Program Participation (SIPP) as the primary data input source do not have to perform an annual to monthly income conversion function, because SIPP collects monthly recipiency and amounts for most types of income. To date, however, small sample size and other problems have limited the direct use of SIPP as a microsimulation model database (see Citro, in this volume).

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers 1 from the age reported by each person in March, but none of the models “grows” people during the income year or otherwise alters family composition. Location of the Annual to Monthly Conversion in the Model Run Stream TRIM2 The MONTHS routine, which performs the annual to monthly income conversion, is run after all other steps have been carried out to process the input data and before any of the program simulation modules are run. MATH The ALLOY routine, which performs the annual to monthly income conversion, ideally would follow all the steps to process the input data and precede all of the program simulation routines. However, in the current version of MATH, the PBLAST routine for simulating AFDC, SSI, and general assistance (GA) precedes ALLOY. This is because resource constraints have precluded modifying PBLAST to operate on monthly data. PBLAST uses a more ad hoc procedure for allocating employment status (and hence earnings) and unearned income throughout the year. ALLOY, which was implemented in 1984 and incorporates empirical evidence from the ISDP about patterns of employment and income receipt over the year, contains some code to treat AFDC-participating units consistently with the assumptions used in PBLAST. ALLOY also allocates the simulated AFDC benefits produced by PBLAST to months for subsequent use by the FSTAMP (food stamp program) routine. HITSM Conversion of annual to monthly income in HITSM takes place after all the steps to process the input data have been performed, including correction of March CPS income values for underreporting. The conversion also takes place after determination of family asset amounts but before determination of eligibility or participation in transfer programs. (HITSM, in contrast to TRIM2 and MATH, was designed as an integrated rather than a modular system. Hence, portions of program code that perform functions such as simulating particular programs are not organized into separate modules that can be invoked or not as the user determines.)

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers Determination of Monthly Labor Force and Employment Status Input Data The March CPS public-use files provide, for each person 15 years of age and older at the time of the March supplement interview, the following items pertaining to labor force and employment status during the preceding calendar year, which are used by one or more of the three models: number of weeks worked in year, number of weeks unemployed (not working but looking for work or on layoff) in year, number of employers in year, number of spells of unemployment in year, and hours normally worked per week. In addition, the number of weeks not in the labor force in year can be calculated as 52 minus weeks worked minus weeks unemployed. TRIM2 The MONTHS routine first allocates weeks worked, weeks unemployed, and weeks not in the labor force across the year and then allocates the income variables in proportion to weeks worked, weeks not worked, or simply by dividing by 12. The actual allocation of labor force status is performed outside of TRIM2. An extract file is created that contains all people who were in the labor force at any time during the year. The file is sorted by weeks worked in descending order and then on a random number, so that those people with the least flexibility for allocation will be considered first. Next, each individual’s weeks worked, weeks unemployed, and weeks not in the labor force are divided into smaller periods based on the number of employers, the number of unemployment spells, and the total number of weeks in each status. The allocation process for each individual begins by selecting a random month in the year to start. A probability is calculated for each of the three possible labor force statuses by multiplying an individual’s probability of being in that status by the index for that status in that month. (Although not described in the documentation, the individual’s probability is presumably a function of the number of weeks in each status; the monthly indexes for each status are discussed below.) Next, a random number is chosen to select a status based on the probabilities. A period of that type is then allocated to the weeks of the month (all months are assumed to last 4.333 weeks) and continuing into succeeding months until all weeks in that period have been exhausted. (The allocation wraps around to the first months of the same year if necessary—for example, in the case in which someone who worked at a job for 6 months

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers is simulated to begin work in October.) After all weeks for the period have been allocated, the probabilities are recomputed and the process is repeated until all weeks in the year have been allocated. As much as possible, periods of employment and unemployment are interspersed when an individual reports having more than one employer or more than one spell of unemployment. The monthly indexes referred to above are calculated based on data, from the Bureau of Labor Statistics’s (BLS) publication Employment and Earnings, on the total numbers of employed and unemployed people by month. The allocation program calculates a target number of total weeks working, unemployed, and not in the labor force per month such that the simulated patterns in the March CPS database will follow the monthly trends in the BLS data. As each individual’s weeks are allocated, the target totals are diminished, and new indexes are computed every 1,000 records based on the new totals. MATH The general procedure that the ALLOY routine uses to allocate weeks worked, weeks unemployed, and weeks out of the labor force is as follows. First, months of work and months of unemployment are calculated by dividing weeks worked and weeks unemployed, respectively, by 4.333, and rounding the results up or down to whole months. Months out of the labor force are calculated as 12 minus months worked minus months unemployed. Then monthly employment status is determined. Members of the armed forces are assumed to work all year. People who worked part of the year and who were otherwise out of the labor force have their months worked allocated starting in a random month and continuing until the months are exhausted; remaining months are coded as out of the labor force. The allocation process wraps around to the first months of the same year, if necessary. People who worked part of the year and who also reported unemployment are subject to a random determination of whether months worked or months unemployed are allocated first. Months worked (unemployed) are then allocated starting in a randomly chosen month, followed by months unemployed (worked), with remaining months coded as out of the labor force. Again, the months wrap around as necessary. The allocation procedures are modified for spouses in husband-wife families where both spouses worked part of the year. In these instances the spouse is assigned a random starting month for his or her months worked in a manner that constrains the spouse’s period of work to overlap that of the head by at least 1 month. This procedure is based on findings from the ISDP that indicated, contrary to earlier assumptions, that the spouse’s work periods tend to coincide with the head’s work periods rather than the head’s unemployment periods. However, in the current version of ALLOY, this modification is implemented only for husband-wife families that are not simulated to receive benefits from AFDC. For families simulated as participants, ALLOY uses different

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers procedures for consistency with the assumptions made in PBLAST, the public assistance simulation routine, which precedes ALLOY in the MATH model run stream.3 HITSM This model first determines the duration of jobs held and spells of unemployment during the year. Individuals who reported one job are assigned one continuous period of employment lasting the reported number of weeks worked. Individuals who reported more than one job have their weeks worked on each job estimated by using a random process. Similarly, individuals reporting one unemployment spell are assigned one continuous period of unemployment lasting the reported number of weeks unemployed. Individuals who reported more than one unemployment spell have their weeks unemployed in each spell estimated by using a random process. The next step is to distribute jobs and unemployment spells over the year. The assumption is made for individuals reporting more spells of unemployment than jobs that each job was preceded and followed by at least one period of unemployment; periods of unemployment and employment are randomly ordered accordingly. The assumption is made for individuals reporting the same number of unemployment spells as jobs that periods of employment and unemployment alternate. The next step is to distribute weeks not in the labor force. Individuals who were students at the beginning or end of the year (i.e., January and December) have their weeks out of the labor force assigned to the beginning or end of the year in accordance with their student status.4 People with periods of unemployment not separated by a job (e.g., one job held in the year with three spells of unemployment) have their weeks out of the labor force assigned between these unemployment periods; other people who experienced unemployment have their weeks out of the labor force assigned at either the beginning or the end of each spell of unemployment, with the number of weeks out of the labor force associated with each unemployment spell determined by using a random process. People who were never unemployed have all of their weeks out of the labor force assigned to a single period at the end of the year if the individual did not hold a job at that time and otherwise at the start of 3   Oversimplifying, these procedures make the assumption that periods of work for the spouse offset periods of unemployment or disability for the head in the case of incapacitated parent units and unemployed parent units where the head worked part of the year and full time. For all other husband-wife AFDC units where both head and spouse had part-year employment or labor force participation, the procedures make the assumption that periods of work and unemployment of the head and spouse coincide. 4   It is not clear from the documentation how a determination is made of school status at the beginning or end of the year.

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers the year. The result is an imputed employment history for each individual in the March CPS that records their labor force participation, unemployment, and employment status during each week of the income year. Comment All three models use complex procedures to determine monthly labor force status. However, each emphasizes different aspects of the variations in labor force activity that characterize the population throughout the year. The TRIM2 MONTHS routine uses information on the number of employers and spells of unemployment to simulate variable patterns of labor force status over the year. The results of the individual simulations are controlled to the seasonal variations and trends in aggregate employment and unemployment during the year as reported by the BLS. The MATH ALLOY routine makes use of ISDP results on employment status patterns within families to relate the work periods of husbands and wives (with the exceptions noted above of certain husband-wife families simulated to be eligible for AFDC). Otherwise, ALLOY takes a straightforward approach that does not allow for seasonal variations or for very complex patterns of individual labor force status during the year. HITSM, like TRIM2, makes use of the information on the number of jobs and spells of unemployment to allow for a variable pattern of employment status on an individual basis. HITSM also takes limited account of seasonal variations in its treatment of the employment patterns of students. Determination of Monthly Earnings Input Data The March CPS public-use files provide, for each person 15 years of age and older at the time of the March supplement interview, the following items pertaining to earnings during the preceding calendar year: annual income from wage and salary jobs, annual net income (or loss) from nonfarm self-employment, and annual net income (or loss) from farm self-employment. In addition, an hourly wage rate can be constructed for workers from annual earnings, number of weeks worked, and normal hours worked per week (the current hourly wage rate is also available for hourly workers for about one-fourth of the sample). TRIM2 The MONTHS routine first sums the annual earnings variables (income from

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers wage and salary jobs, nonfarm self-employment, and farm self-employment) for each person and then allocates that sum to each month in proportion to the simulated number of weeks worked in that month. MATH The ALLOY routine sums the earnings variables, then computes an average weekly amount (annual earnings divided by total number of weeks worked), and then multiplies this amount by 4.333 to obtain average earnings per month. This monthly figure is allocated to each simulated month of employment. Income from nonfarm self-employment is constrained to be greater than or equal to zero in order to mimic the food stamp program regulations regarding losses. (Net rental income is similarly constrained—see discussion of unearned income below.) HITSM This model first determines the average hourly wage rate for each person. Then during each week in which an individual is simulated to work, weekly earnings are calculated as the product of the wage rate times the number of hours normally worked per week during the year. Weekly earnings are summed to produce monthly earnings, where each month is assumed to include 4.33 weeks. Comment The procedures for allocating earned income are essentially the same in all three models, with the exception that MATH does not allow for part-month employment. Determination of Monthly Unearned Income Input Data The March CPS public-use files provide, for each person 15 years of age and older at the time of the March supplement interview, the following items pertaining to unearned income during the preceding calendar year: receipt of income from private, federal, military, and state and local government pensions; total annual pension income; receipt of income from social security and railroad retirement; total annual social security/railroad retirement income; receipt of income from interest; total annual interest income;

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers TABLE 2 Determinants of Major Events Simulated by DYNASIM2 Event or Characteristic Variables Used to Determine Event Simulation of Longitudinal Histories Demographic events   Death   Married women 45–64 Age, race, sex, marital status, education, number of children All others Age, race, sex, marital status, education Birth Age, marital status, number of children, race, education Multiple birth Race Sex of newborn Race Marriage   Age 18–29 Age, race, sex, previous marital status, income, education, region, weeks worked, hourly wage, asset income, welfare, unemployment compensation Other ages or ever married Age, race, sex, previous marital status Mate matching Difference in age, difference in education Leaving homea Age, race, sex Divorce Distribution over time of expected divorces for this marriage cohort, age at marriage, education, previous marital status, presence of young children, weeks worked, wages Education Race, sex, age, years at current school level, parents’ eduation Mobility Number of years married, size of family, age and sex of head, education of head, race, region, size of metropolitan statistical area (MSA) Disability   Onset Age, race, sex, marital status Recovery Age, race, sex, marital status, education Labor force events   Labor force participation Age, race, sex, education, South, disability, marital status, student, children, spouse earnings Hours in the labor force Age, transfer income, expected wage, disability, marital status, children Wage rate Race, sex, age, South, disability, marital status, education, student Unemployment Age, sex, race, education, marital status, South, disability, children

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers Event or Characteristic Variables Used to Determine Event Simulation of Jobs, Pensions, and Retirement Using Longitudinal Histories Job characteristics and pension plans   Job change Age, sex, tenure on current job, industry Industry of new job Sex, education, previous industry Pension coverage on new job Sex, industry, earnings level Pension plan participation Age, tenure on job, full- or part-time status, sex Type of pension coverage Industry Pension eligibility and benefits   Retirement eligibility Age, industry, years of service, type of pension Vesting Industry Benefit formula Industry and type of pension coverage Benefit plan constants Benefit formula, industry, type of pension coverage Individual retirement accounts   Plan participation Sex, earnings Retirement   Probability of leaving job Age, sex, disability, marital status, pension eligibility and amounts, social security eligibility and amounts, wage, earnings Probability of taking new job Age, disability, marital status, pension eligibility and amount, social security eligibility and amounts, imputed wage aPeople leaving home for reasons other than marriage, birth of a child, divorce, or death. SOURCE: Johnson, Wertheimer, and Zedlewski (1983); Johnson and Zedlewski (1982). PRISM Pension Coverage and Characteristics Each time an individual obtains a new job, the model estimates whether the new employer sponsors a pension plan. If so, the individual is assigned one of the plan sponsors in the retirement plan provisions database of Lewin/ICF. The match between the pension plan and the individual depends on the industry of the job, multiemployer or single employer status, and hourly or salaried status. Given the set of rules for that pension plan, the model keeps track of pension-relevant detail, including years on the job, wages earned, and so on. Individuals who are 2 or 3 years away from being vested in a pension cannot be selected to change jobs. In addition to keeping track of pension eligibility and accumulation, PRISM maintains a running total of the individual’s quarters of social security coverage.

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers TABLE 3 Determinants of Major Events Simulated by PRISM Event or Characteristic Variables Used to Determine Event Simulation of Longitudinal Histories Demographic events   Death Disability status, age, sex, years of disability Disability   Onset Age, sex Recovery Age, sex, years of disability Divorce Age of husband and wife Marriage Age, sex, previous marital status Mate matching Age of male, age of female Birth Marital status, age, number of children, employment status last year Labor force events   Hours worked per year Hours last year, age, sex, marital status, education, composite of hours in previous 3 years, female with young children, female divorced or widowed this year, receiving pension or social security income Wage rate Age, sex, whether changed job this year, whether unemployed this year Job change Hours worked, age, years on job Industry Age, education, previous industry, full- or part-time status Pension characteristics   Pension coverage Age, industry, full- or part-time status, wage rate Pension plan assignment Industry, multi- or single employer plan in 1979, hourly or salaried status Retirement   Pension acceptance Age, sex (conditional on eligibility) Social security acceptance Age, sex (conditional on eligibility) Individual retirement accounts   Adoption Age, family earnings, pension coverage Contributions Sex, marital status, family earnings   SOURCE: Kennell and Sheils (1986). Interactions Between Pension and Social Security Eligibility and Retirement in PRISM Individuals who become eligible for social security are selected to accept social security on the basis of probabilities that depend on age and sex. (Based on these probabilities, a majority of individuals will retire at age 65.) All individuals who become eligible for a private pension accept their pension at the normal plan retirement age, generally 65. Data on the labor force participation of the elderly indicate that individuals who have accepted retirement income are likely

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers to stop working. Therefore, the transition probabilities between annual hours of work categories for individuals who accepted retirement income during the year were calculated separately from those of nonelderly nonretired workers. These transition probabilities vary by sex of the worker and whether he or she is receiving a pension or social security income. DYNASIM2 LABOR FORCE AND PENSION SIMULATIONS Labor Force Simulations In contrast to PRISM, which simulates most of the labor force experience variables by simulating transitions from one hours-of-work category to another, DYNASIM2 simulates labor force experience sequentially, as part of its Family and Earnings History Submodel. DYNASIM2 first simulates whether the individual participates in the labor force at all during the year. If the individual does participate, the model then simulates the number of hours of participation (both employed and unemployed time are included). Then the hourly wage is estimated. Then the model simulates whether the individual is unemployed during the year and, if so, the fraction of time in the labor force that the individual is unemployed. With this information, it is possible to combine the wage with hours of employment to estimate annual earnings. The labor force participation, hours, and wage equations were estimated in the early 1980s from several years of data from the Panel Study of Income Dynamics (PSID) (Holden, 1980).2 The labor force models are sound, basic economic models that include human capital variables (education and age) and control for the usual demographic variables (including race, sex, marital status, residence in the South, disabled, presence of young children, and others). In addition, the estimates of these labor force variables are linked across time and among one another through their error structures. The labor force participation equation has an error term that includes a serially correlated component, v, composed of last period’s v, a serial correlation coefficient that is different for each of 16 demographic groups, and z, a transitory component that is selected each year for each individual. The error term also includes an individual-specific error term, u, which is fixed for each individual throughout his or her life, drawn from a normal distribution that has a different variance for each of the 16 demographic groups. Both of these components of the error term increase the stability across time among labor force participation decisions by each individual. The error term of the hours equation includes a serially correlated term (which differs across six demographic groups) and a second term that is shared 2   The labor force participation equations drew on 13 years of data from the PSID, while the wage and hours equations were estimated from 9 years of data from the PSID.

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers with the wage equation. The error structure of the wage equation is conceptually identical to that of the labor force participation equation, with a serially correlated transitory component and an individual-specific fixed component. In addition, the error term of the wage equation includes a transitory component that is shared with the hours equation. This term has no economic interpretation but is included in order to generate the observed negative correlation between wage rates and hours in a cross section. The term cancels out so that it leads to variation both in wage rates and in hours, but does not lead to variation in earnings. Retirement in DYNASIM2 The Jobs and Benefits History submodel of DYNASIM2 determines when the individual retires and the amount of retirement income available from each source. As in the Family and Earnings History submodel, the Jobs and Benefits History submodel takes an individual and cycles through five routines for a single year, then goes to the next year and processes the five routines again. When it reaches the individual’s final year, it goes on to the next individual. The five routines include retirement, jobs, social security benefits, employer pension benefits, and individual retirement accounts (IRAs). The retirement module may be run as part of DYNASIM2 at the user’s option. If it is run, the module’s choice of retirement age for the individual overrides whatever pattern of labor force participation may have been modeled earlier in the labor force section of the Family and Earnings History submodel. The retirement module determines whether the individual retires at a given age based on eligibility for pensions and social security at that age and on the difference between the discounted stream of pension and social security benefits available if the individual retires this year relative to next year.3 The retirement module is run only on individuals age 58 and over. The retirement simulation is based on a two-equation behavioral model. In the first equation, the module simulates whether the worker leaves the current job based on age, disability, current earnings, eligibility for pension or social security, and the change in social security and pension wealth if retirement is delayed by 1 year. If the worker does leave the current job, the module then simulates whether that worker accepts a new job, on the basis of a similar set of variables. If the worker accepts a new job, the jobs module simulates job and pension characteristics (described below). If the worker does not take a new job, he or she is out of the labor force for this year and all subsequent years of the simulation. 3   The retirement module is based on work by Burkhauser and Quinn (1981) using the Social Security Administration’s longitudinal Retirement History Survey.

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers DYNASIM2 Pension Coverage and Characteristics The jobs module is the first module run on individuals under age 58 and is run after the retirement module in each year after age 58. The jobs module uses the labor force information created in the Family and Earnings History submodel to help determine whether the individual has changed jobs, and if a job change occurs, this module chooses an industry of employment and an associated set of pension plan characteristics. New entrants to the labor force and reentrants are all assumed to obtain a new job upon entry. In addition, some individuals who participated in the labor force this year and last are assumed to change jobs. If the individual obtains a new job, the module simulates an industry on the basis of the individual’s sex, education, and past industry of employment. Based on industry assignment, a set of pension characteristics is simulated for this job. DYNASIM2 does not include a database of pension characteristics as does PRISM, so pension plans are built up by simulating each characteristic in turn. Coverage status is simulated first and then plan participation. The plan participation probabilities are based on the fact that many employees must build up tenure on the job and work full time before they are eligible to participate. For plan participants, the type of coverage is simulated, including whether single or multiemployer status and whether defined contribution or defined benefit. These types of coverage tend to vary by industry of employment. When the individual leaves a job (either to change jobs or to retire), a pension benefit is calculated by the employer pension module. This module is also used in order to calculate a pension benefit when the retirement module is running because the expected pension benefit helps to determine whether or not the individual will retire. The employer pension module determines probabilistically whether the individual is eligible for a pension. Probabilities are based on age, industry, and the number of years on the job. If the individual is not eligible for a pension, vesting status is assigned probabilistically. If the individual is pension eligible, a benefit formula is assigned, and the benefit is computed on the basis of characteristics of the benefit formula, past earnings, age, and other variables. If the individual is changing jobs, this pension is saved along with other characteristics of the job. If the retirement module is running, the pension benefit is used with the social security benefit to determine whether the individual retires this year. The social security module calculates a social security benefit if the individual meets eligibility requirements for a disability or retirement benefit (by age, disability, and quarters of covered earnings). If the individual is eligible, the module uses information on previous earnings to calculate the primary insurance amount, adjusts that amount for the individual’s age, and compares the benefit to current earnings to determine whether the individual receives any social security this year. This module is highly parameterized to permit the user

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers to simulate the effects of a wide range of policy options on retirement age and retirement income. Contributions to an IRA are simulated much like another pension plan, except that such contributions only started in 1975 for individuals without pension coverage and in 1982 for those with coverage. Participation and contribution rates are assigned probabilistically. PROGRAM SIMULATIONS As described briefly above, the social security routines of both DYNASIM2 and PRISM are highly parameterized, enabling the user to simulate changes in many different program features. In addition, both models contain simulations of the private retirement savings accounts, IRAs and Keoghs. Finally, both models simulate SSI eligibility and participation and federal taxes. DYNASIM2 simulates SSI benefits and federal taxes for the final simulation year, while PRISM simulates them in every year. To simulate SSI, both models create filing units, calculate aggregate filing unit income, and check the categorical eligibility criteria (age and disability). PRISM approximates an asset test by disqualifying a proportion of units defined by age, income, sex, and marital status, on a probabilistic basis. The proportion of units to be disqualified was determined by analyzing asset income information for elderly families on the March 1980 CPS. PRISM computes state-specific SSI benefits; DYNASIM2 computes region-specific benefits because the state of residence is not simulated across time. SSI participation in DYNASIM2 depends on benefit level (three categories); PRISM estimates SSI participation on the basis of marital status and benefit level. PRISM and DYNASIM2 calculate federal income taxes and social security payroll taxes. PRISM also simulates state income taxes. CONCLUSIONS Both DYNASIM2 and PRISM begin with a microdata file of individuals in households and put them through a series of demographic and economic modules that determine, for each year, whether the individual dies, marries, works, retires, or experiences some other event. In any year, this series of events gives the individual a set of characteristics that can be used to determine whether other events occur in the next year and whether the individual is eligible for government transfer programs, including social security, supplemental security income, and others. The models thus simulate the implications of policy changes, demographic changes, and economic trends. While the models share these basic features, some important differences exist, and these affect the types of questions each model can best analyze. DYNASIM2 simulates a fuller range of demographic events than does PRISM,

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers including education and migration. DYNASIM2 not only adds children to the file (born during simulations), but educates them and moves them into the labor force. As a result, a full range of demographic issues can be analyzed, including the policy implications of changing marital status patterns and teenage fertility. The full range of demographic simulations also enables the analyst to examine events over a longer time horizon and to focus on a broader range of cohorts than is possible with PRISM. In addition, DYNASIM2 includes more formal models of labor force and demographic events than does PRISM, and has been constructed so that the user can specify different models of behavior. PRISM, on the other hand, includes a model of long-term care that can address policy questions in this important area. Both dynamic models provide a great deal of flexibility. Parameters governing demographic and labor force events may be changed, allowing an analyst to create alternative population projections on the basis of different assumptions. For example, it is possible to change parameters specifying fertility rates, mortality rates, female labor force participation rates, and the real growth in wage rates. Similarly, alternative projections of retirement income can be made by altering the rules that guide social security and pension eligibility, accumulation, and distribution. This flexibility enables the models to project the distributional effects of a policy, a demographic change, or a labor force trend far into the future, given certain well-specified assumptions about the trend of other demographic events and labor force behavior. Building the projections from microdata allows an analyst to examine interactions between provisions of a policy or between demographic and labor force events and to obtain the distribution of effects across population subgroups. For questions about interactive effects and distributional effects, this is the best available framework. There are limits to the flexibility of these models, however. Under their current structure, policy changes do not generate either demographic or labor force behavioral responses. For example, in DYNASIM2, many events—including jobs, pension characteristics, social security benefit levels, and retirement—are simulated after the longitudinal demographic and economic histories are created. The model is structured in this way because creating longitudinal histories is very expensive on a mainframe. This split structure enables an analyst to create the longitudinal histories only once (or to create a few under different reasonable assumptions), and then have the capability to run a large number of different program simulations on that longitudinal file. However, this structure makes it impossible for policy changes to affect demographic or labor market events (except for the retirement decision, which is done in the second stage and overrides a retirement decision that may have been made when the longitudinal histories were created). This is unfortunate, since interest in the projections made by these models often arises because a policy change is being considered today that is expected to affect behavior

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers and incomes far into the future. Changes in demographic events and labor supply are quite likely once a significant policy change has occurred. For example, policy changes that make the social security system less generous are likely to lead to compensating changes in labor supply, savings behavior, and retirement plans by people who are 30 years away from the “normal” retirement age. Currently, the dynamic models could incorporate a response if an analyst changed parameters for labor force participation, but this response would not be based on an explicit relationship between the policy change and labor supply decisions. New development work on the DYNASIM2 model will improve its flexibility by fully integrating the retirement decision with other demographic and economic events and with policy changes. Further development to expand opportunities for policy changes to feed back to demographic and economic events would enable both of these models to reflect current theory and estimates of behavioral responses to policy. Finally, the models would be better served by documentation that is clear, well-organized, and up-to-date. These are very large and complex computer models, and the documentation for both PRISM and DYNASIM2 contains sections that are too tersely written to provide a full understanding of essential features of the models and how those features interact. Leaving so much to the imagination does not help to convince an analyst that the model is a sound policy analysis tool. In addition, at least one of the module descriptions included in the current documentation of DYNASIM2 is from a prior version of that module. Accessible, current documentation should be integrated with model development efforts. This would, at least, help market the models to potential users and funders, and, at best, help analysts to more fully exercise the capabilities of these models for policy analysis. REFERENCES Anderson, Joseph M. 1990 Micro-macro linkages in economic models. Pp. 187–220 in Gordon H.Lewis and Richard C.Michel, eds., Micro simulation Techniques for Tax and Transfer Analysis. Washington, D.C.: The Urban Institute Press. Burkhauser, Richard, and Quinn, Joseph 1981 The Relationship Between Mandatory Retirement Age Limits and Pension Rules in the Retirement Decision. Research Report 1348–03. Urban Institute, Washington, D.C. Holden, Russell 1980 Revised Models of Labor Force Hours, Wage Rates, and Earnings for DYNASIM. Working Paper 5908–2. Urban Institute, Washington, D.C. Johnson, Jon, and Zedlewski, Sheila, R. 1982 The Dynamic Simulation of Income Model (DYNASIM), Volume II, The Jobs and Benefits History Model. Washington, D.C.: The Urban Institute Press.

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers Johnson, Jon, Wertheimer, Richard, and Zedlewski, Sheila R. 1983 The Dynamic Simulation of Income Model (DYNASIM), Volume I, The Family and Earnings History Model. Revised. Washington, D.C.: The Urban Institute Press. Kennell, David, and Sheils, John F. 1986 The ICF Pension and Retirement Income Simulation Model (PRISM) With the ICF/Brookings Long-Term Care Financing Model. Draft Technical Documentation. ICF Incorporated, Washington, D.C. 1990 PRISM: Dynamic simulation of pension and retirement income. Pp. 137–172 in Gordon H.Lewis and Richard P.Michel, eds., Microsimulation Techniques for Tax and Transfer Analysis. Washington, D.C.: The Urban Institute Press. Orcutt, Guy H., Glazer, Amihai, Harris, Robert, and Wertheimer, Richard, II 1980 Microanalytic modeling and the analysis of public transfer policies. Pp. 81–106 in Robert H.Haveman and Kevin Hollenbeck, eds., Microeconomic Simulation Models for Public Policy Analysis. Vol. 1, Distributional Impacts. New York: Academic Press. Zedlewski, Sheila R. 1990 The development of the Dynamic Simulation of Income Model (DYNASIM). Pp. 109–136 in Gordon H.Lewis and Richard C.Michel, eds., Microsimulation Techniques for Tax and Transfer Analysis. Washington, D.C.: The Urban Institute Press.

OCR for page 87
Improving Information for Social Policy Decisions—The Uses of Microsimulation Modeling: Volume II, Technical Papers This page intentionally left blank.