The Social Security Board of Trustees (1999) are projecting

Transcription

1 Labor Force Participation of Older Workers Labor Force Participation of Older Workers: Prospective Changes and Potential Policy Responses Abstract - Increased labor force participation of the elderly can reduce the fiscal and economic stress from the projected aging of the population in the next century. This paper uses Survey of Income and Program Participation data matched with longitudinal earnings histories and Social Security benefit records to estimate joint work and benefit receipt choices for people age 62 and over. The probability of working is shown to depend on both worker characteristics and policy variables, with lower Social Security benefits and lower tax rates on wages both modestly increasing labor force participation of older workers. Melissa Favreault, Caroline Ratcliffe, & Eric Toder The Urban Institute, Washington, D.C National Tax Journal Vol. LII, No. 3 INTRODUCTION The Social Security Board of Trustees (1999) are projecting that the ratio of workers to Social Security beneficiaries will decline from 3.4 in 2000 to 1.8 in The decline in this ratio means that ultimately payroll taxes will have to be raised, benefits reduced, or other revenue sources tapped if the program is to remain solvent in the long run. The Trustees most recent projections show that the Trust Fund balance will be reduced to zero, under current policies, by the year One potential response to the increased number of retirees is increased labor force participation of older workers. For most of this century, labor force participation rates of older workers have declined as people have chosen to retire at younger ages, but recently the trend toward earlier retirement seems to have stabilized (Quinn, 1999). It is possible that more people may delay retirement in the future, either because of changing demographic and economic conditions (such as improved health and longevity, increased earning power of women, and changes in the physical demands of jobs) or in response to future policy changes that increase work incentives of the elderly. If labor force participation rates among the elderly did increase, this could reduce some of the economic and fiscal burdens of an aging population. This paper examines determinants of labor force participation among the elderly, using a matched data set that includes both earnings histories and current income and de- 483

2 mographic characteristics of households. We estimate a reduced form equation that explains joint labor force participation and Social Security benefit receipt choices at ages 62 and over as a function of recent earnings history, pension coverage, education, age, sex, race, marital status, and two policy variables the level of Social Security benefits and the tax rate on earnings. Based on our results, we offer some preliminary observations on how changes in taxation and benefits might affect future labor force participation. The second section of this paper briefly reviews the literature on determinants of labor force participation of the elderly and discusses how current tax laws and benefit rules might affect participation. The third section outlines a model of the choice between four options: working and not claiming Social Security benefits, working and claiming Social Security benefits, not working and claiming Social Security benefits, and not working and not claiming benefits. The fourth section discusses the data and the fifth section presents regression results and simulations of the effects of hypothetical policy changes on representative individuals. The sixth section briefly concludes. DETERMINANTS OF LABOR FORCE PARTICIPATION OF THE ELDERLY Factors That Affect Whether and How Much to Work 484 NATIONAL TAX JOURNAL Many factors can affect how much an individual chooses to work as he approaches the traditional retirement age. An extensive literature considers the relative importance of these competing forces (for example, the overview by Quinn, Burkhauser, and Myers, 1990). Among the most important factors are one s ability to work, the economic resources available at the time of prospective retirement, the economic returns that one could garner from working, the psychological benefits of working, and one s family situation. The importance of each of these factors varies both across and within different birth cohorts. Both one s personal characteristics, for example, one s health, and characteristics of the labor market, for example, the physical demands imposed by jobs, at the time one is considering retirement influence the ability to work. While health problems do account for a sizable fraction of retirements, most researchers find that they are not a central factor in the majority of late-career work decisions (Burkhauser, Couch, and Phillips, 1996; Uccello, 1998). Further, recent research documenting modest improvements in health of older Americans in past decades (for example, Manton, Corder, and Stallard, 1997; Crimmins, Reynolds, and Saito, 1999) suggests that health concerns should prohibit comparatively fewer older Americans from working in coming years. In addition, recent declines in the physical demands of work (see, for example, Johnson, Spiro, and Steuerle, 1999) suggest that even fewer older workers will be restricted from remaining in the labor force. On the labor market ( demand ) side, continued work beyond age 62 is contingent on the availability of positions in the labor force that are appropriate given a person s health and skill set. When labor markets are tighter, older or less healthy workers have a greater opportunity to change jobs or to continue in their present jobs than they do when unemployment is higher. Structural changes in the economy are other important demand-side factors in the labor market for older workers. Workers employed in industries that are being transformed or eliminated by new technologies may find that their skills are outdated and thus their opportunities to work are diminishing as they age. The rate at which workers become victims of this type of downsizing may, however, decline in future years, as information processing becomes a larger fraction of the U.S.

3 Labor Force Participation of Older Workers 485 economy and as an aging population makes workers more scarce. One s economic resources at the time of retirement, including assets, pensions, and Social Security entitlements, further influence one s decisions about work. Researchers consistently find that nondisabled individuals with more resources tend to leave the labor force earlier than those nondisabled persons with fewer resources. This relationship is complex, however, because further work effort increases most forms of wealth. Social Security is by far the largest form of potential support in retirement for the aged, followed by private pensions and then other saving (Grad, 1998; Lumsdaine, Stock, and Wise, 1990). While there is a general consensus that higher real Social Security benefits lead to earlier retirement, pensions have complex and varied effects on work behavior, in part because pension provisions differ so widely across plans. Research evidence suggests that defined benefit pension plans often encourage retirement at a given age (see, for example, Lumsdaine and Wise, 1990; Wise, 1997). Defined contribution plans, which are less likely to incorporate explicit work disincentives and are comprising a larger fraction of pension plans over time, may have less dramatic effects on work behavior. Economists frequently characterize decision making about work as a process of weighing the benefits of labor and leisure. Higher economic returns from working are said to promote greater labor supply, though the function is complex, because a person with a higher wage rate needs to work fewer hours in order to reach a target level of earnings. One s return from working near retirement is a complex function of one s earnings potential (itself a function of one s education, recent wage level, and work experience), taxes on earnings, any additional pension accruals that one would gain from working, and the value of health and other insurance benefits that are included in one s compensation. Models that take all of these factors into account generally suggest that workers do consider returns from working quite closely when making their retirement decisions (see, for example, the cross-national evidence in Gruber and Wise, 1997), though some researchers do see slight deviations that could suggest that custom may dominate over strict economic rationality (see, for example, Lumsdaine, Stock, and Wise, 1995; Rust and Phelan, 1997 offer a competing view). Comparatively less research examines variation in psychological, as opposed to economic, benefits of work, and thus in the taste for work. In addition to economic benefits, work provides many people with a means of social interaction, a sense of identity, and itself may be interesting, challenging, or intrinsically rewarding. Hochschild (1997) uses ethnographic methods to study how workers in one firm balance work and family demands, and finds that social rewards of the workplace, such as praise for a job well done and sense of control or accomplishment, cause some workers to find work more rewarding and less stressful than home life. To a certain extent, this finding calls the notion of a simple labor/leisure trade-off into question. Because it is so difficult to measure the psychological benefits of working, in many studies, researchers use other variables, including education and measures of labor force attachment, to proxy for it. This is a reasonable, though partial, solution, because the potential that a job is interesting and rewarding tends to increase with the level of education it requires. A person s family situation, including her spouse s work status, is surely an important determinant of labor supply near retirement. Many researchers, including Hurd (1988) and Smith (1996), have found that spouses coordinate their decisions about leaving the labor force, perhaps because leisure is valued more highly when one can share it with one s spouse. Caregiving responsibilities for aging fam-

4 ily members, which are changing over time (see, for example, Moen, Robison, and Fields, 1994), may also serve to induce older workers to leave the labor force. Evidence on the strength of this relationship appears to be mixed (see, for example, Couch, Daly, and Wolf, 1999; Johnson, LoSasso, and Sambamoorthi, 1999). Tax and Benefit Policies That Could Affect Labor Force Participation Tax and benefit policies influence labor force participation through both income and substitution effects. Transfer payments could reduce labor force participation because they enable individuals and couples to increase both consumption of market goods and nonwork time. Taxes on labor earnings have offsetting effects; they increase labor force participation by lowering total income (including both goods and nonwork time) and reduce labor force participation by lowering the return to work. Because earnings account for a diminishing fraction of people s total income as they age and begin receiving Social Security, Medicare, and private pension benefits, the substitution effect of taxes on earnings becomes more important and labor income taxes are more likely to reduce work effort. Transfer Income 486 NATIONAL TAX JOURNAL Transfer policies that raise after-tax income include Social Security retirement benefits and Medicare. The normal retirement age at which workers receive full benefits based on their earnings history is now 65. Workers may begin accepting benefits at age 62, but their annual retirement benefits are reduced by five-ninths of one percent per month if they start benefits before age 65. There is also a delayed retirement credit for persons who delay retirement beyond the normal retirement age. The normal retirement age is scheduled to increase in stages to age 67 for future cohorts of retirees. This change does not alter the date at which workers may start to receive benefits, but it lowers the monthly benefit for all workers who retire before age 67. Medicare benefits are available beginning at age 65. Medicare has a substantial value for retirees and workers who lack health insurance coverage but provides little or no benefit to people who remain employed and are covered by private health insurance through their employers. The level of Social Security benefits, and thus their income effect on the incentive to work, varies greatly among older individuals and is only loosely related to current economic circumstances, including current earning ability. Social Security benefit amounts are based on average earnings over a 35 year period of the individual and his spouse. The percentage of average earnings that benefits replace are higher for individuals with lower earnings than for individuals with higher earnings, higher for one-earner couples than for two-earner couples and single individuals, and unrelated to the amount of nonwage income. There are other sources of variation in replacement rates among similarly situated individuals. Taxes on Earnings Taxes on work effort include the individual income tax, the payroll tax, and the Social Security earnings test. The individual income tax is imposed at graduated rates ranging from 15 to 39.6 percent on income in excess of exemptions and deductions. The income tax rate on the earnings of any individual depends on numerous factors unrelated to the individual s own earnings, including investment income, pension income of the individual or spouse, spouse s earnings, the amount of deductions the individual takes, and the amount of Social Security benefits included in income. The payroll tax is imposed at flat rates of 15.3 percent

5 Labor Force Participation of Older Workers (7.65 percent on both employees and employers) on individual wages up to the Social Security wage base ($65,400 in 1997) and 2.9 percent (1.45 percent on both employers and employees) on individual wages in excess of the Social Security wage base. Unlike younger workers, workers age 62 and over who have already attained their highest 35 years of (wage-indexed) earnings receive no additional benefits in exchange for their additional contributions to Social Security through payroll taxes. Many younger individuals (under age 62) also receive no additional benefits in exchange for their additional contributions to Social Security, because they receive higher Social Security benefits as a spouse than as a worker. The Social Security earnings test is more complex. Social Security benefits are reduced for beneficiaries with earnings above the exempt amount. For each dollar of earnings above the exempt amount, beneficiaries aged lose 50 cents and beneficiaries aged lose 33 1/3 cents in current Social Security benefits. In 1997, the exempt amounts were $8,640 for year olds and $13,500 for year olds. The amounts are indexed to changes in the average wage. For year olds, under current law, the exempt amount is scheduled to be increased in stages to $30,000 in The apparent high marginal tax rates from the Social Security earnings test are effectively lower than 50 percent or 33 1/3 percent, because workers receive an actuarial adjustment for the deferred benefits they receive once they are no longer subject to the earnings test. The issue for labor supply behavior is what discount rate workers use in evaluating the trade-off between current and deferred benefits. 1 Measuring the Tax Rate on Labor Force Participation We measure the tax rate on earnings as the ratio: τ = [T(Y n + E) T(Y n )]/E, where Y n = nonlabor income, E = potential earnings (based on recent past earnings), T(Y n + E) = tax liability if the individual works and receive potential earnings, and T(Y n ) = tax liability if the individual has no earnings. This tax rate (τ) is an average tax rate on all earnings, with earnings stacked last. It measures the incentive to participate in the labor market at a given level of earnings, given other sources of income, but does not measure either the marginal incentive to earn an additional dollar if employed or the marginal incentive to earn the first dollar of employment income. Table 1 illustrates how tax rates on earnings in 1997 varied among 62-year-old single workers with differing levels of potential earnings, nonlabor income, and Social Security benefits (before application of the earnings test). We use three measures of tax burden. Measure 1 is the sum of all direct taxes on earnings of individuals income taxes, the employee share of payroll taxes, and reduced Social Security benefits from the earnings test. Measure 2 excludes the reduction in benefits from the earnings test, based on the assumption that the worker is farsighted and views the future increase in benefits as sufficiently large to compensate for the delay in benefit receipt. Measure 3 adds the employer share of the payroll tax to Measure 2. In Measure 3, the value of pretax earnings used in the denominator of the tax rate calculation includes the employer contribution to payroll taxes. We used Measure 1 in the regressions reported later in this paper. 2 1 There is some evidence that workers are very sensitive to the earnings test, even though they are in principle compensated for lost current benefits by an actuarial adjustment in future benefits. For example, studies of the distribution of annual earnings of retirees shows a spike in earnings near the exempt amount. See, for example, Burtless and Moffitt (1984). 2 Arguably, a measure that includes both the employer and employee parts of the payroll tax is a better representation of tax burdens. Use of this tax rate measure would require using an earnings measure that includes the employer payroll tax in pretax earnings. 487

6 NATIONAL TAX JOURNAL TABLE 1 AVERAGE TAX RATE ON EARNINGS FOR SELECTED INDIVIDUALS SINGLE RETURNS, AGE 62 Earnings (in $) 10,000 10,000 10,000 10,000 10,000 10,000 Investment Income (in $) 0 0 Security Benefits (in $) 9,600 14,400 9,600 14,400 9,600 14,400 Measure 1 (in %) Tax Rate on Earnings Measure 2 (in %) Measure 3 (in %) ,600 14,400 9,600 14,400 9,600 14, ,600 14,400 9,600 14,400 9,600 14,400 Source: Authors calculations. Notes: Measure 1 includes income tax, employee share of payroll tax, and loss of current benefits from Social Security earnings test. Measure 2 is Measure 1 less loss of current benefits from the earnings test. (The earnings test may still affect the marginal income tax rate by reducing Social Security benefits included in income.) Measure 3 is Measure 1 plus the employer share of payroll tax. The calculation of income tax liability assumes that itemized deductions are 20.5 percent of adjusted gross income (AGI). The taxpayers claims the higher of itemized deductions or the standard deduction. All investment income is treated as includable in AGI. The inclusion of Social Security benefits in AGI under current law is part of the computation Table 1 shows large differences in tax rates, both between workers at different earnings levels and among workers with the same level of earnings. For example, for Measure 1, the tax rate on earnings varies between 19 and 37 percent for the six sample workers with earnings of $10,000, from 46 to 59 percent for the six sample workers with earnings of $, and from 41 to 56 percent for the six sample workers with earnings of $. There are numerous special provisions that differentially affect average tax rates on different workers. Average tax rates generally increase with income, but not in all cases. For example, because the marginal rate under the earnings limitation falls to zero once all Social Security benefits are eliminated, the average rate on wages can in some circumstances fall as earnings increase. A JOINT MODEL OF WORK AND BENEFIT RECEIPT Individuals who have reached age 62 face numerous options for combining Social Security benefit receipt and work effort. In terms of Social Security benefit receipt, they face only two options. Individuals can either begin collecting Social Security benefits or choose to postpone collection to some later time. In terms of work effort, individuals face a wider range of options. Theoretically, individuals could face a continuum of options and choose the optimal number of hours to work (i.e., their work effort), but in the U.S. economy, individuals often face a relatively small number of work options (e.g., full-time, high-part-time, and lowpart-time). 3 While we acknowledge that 3 If we focus on individuals work effort across a year rather than a week, options available to individuals expand because they can vary their yearly work effort by choosing to work only a few months out of the year. 488

7 Labor Force Participation of Older Workers some individuals have wide discretion regarding hours of work, this analysis simplifies individuals work behavior and categorizes individuals as either working or not working. 4 This framework gives rise to four possible outcomes. These outcomes (or states ) are: collect Social Security and not work, collect Social Security and work, not collect Social Security and not work, and not collect Social Security and work. Rather than looking at which of these four states an individual occupies at a point in time, our analysis focuses on individuals annual transitions from a state in which they are not collecting Social Security benefits to one of the four states above. This restriction on the initial state implies that only those persons not collecting Social Security benefits at time t 1 (one year ago) are included in the sample. If an individual transitions into a state (at time t) where she is collecting Social Security benefits, then that individual is no longer at risk of becoming a Social Security beneficiary at time t + 1, so she exits the sample. By focusing on these transitions, our analysis examines the factors that influence individuals decisions to move from a state in which they are not collecting Social Security benefits to a state where they are receiving Social Security and either working or not working. To estimate these transitions, we use a multinomial logit model. In this model, the dependent variable takes on one of four values (0 through 3). With this estimation strategy, the probability of falling into each of the four states as a function of independent variables, X, can be expressed as follows: [1] P 0 = Prob(Social Security, No Work) [2] P 1 = Prob(Social Security, Work) [3] P 2 = Prob(No Social Security, Work) [4] P 3 = Prob(No Social Security, No Work) where 5 [5] [6] [7] 1 = 1 + e xβ 1 + e xβ 2 + e xβ 3 e xβ 1 = 1 + e xβ 1 + e xβ 2 + e xβ 3 e xβ 2 = 1 + e xβ 1 + e xβ 2 + e xβ 3 e xβ 3 = 1 + e xβ 1 + e xβ 2 + e xβ 3 P ln 1 = xβ 1 P0 P ln 2 = xβ 2 P0 P ln 3 = xβ 3 P0 The estimated coefficients from this model represent the effects of the variables on the log-odds of occupying one state relative to the base case state. So, for example, β 1 represents the effect of X on the log-odds of collecting Social Security benefits and working relative to collecting Social Security benefits and not working. Our primary interest is in the relationship between individuals decisions to (1) 4 By doing this, we focus on what factors affect an individual s decision to work versus not work, but not on the level of work effort. 5 These three equations are expressed in relative terms (i.e., relative to the probability of being in group 0), because the estimation of the model requires a normalization. The normalization is necessary because knowledge of any three probabilities automatically gives the value of the fourth probability, as the probabilities sum to one. 489

8 NATIONAL TAX JOURNAL collect Social Security benefits and not work, (2) collect Social Security benefits and work, and (3) not collect Social Security benefits and work. The fourth group of individuals, those who neither collect Social Security benefits nor work, are of less interest because their decision to postpone receipt of benefits is unrelated to their labor supply decision. 6 This group, however, is included in our analysis so that we do not introduce selection bias into our estimates. DATA The models were estimated from Survey of Income and Program Participation (SIPP) data matched to Social Security Administration Summary Earnings Records (SERs) and Master Beneficiary Records (MBRs). The SERs provide data on covered earnings as far back as 1951, while the MBRs identify individuals dates of first receipt of Social Security benefits. The SIPP panels used for estimation included These data are wellsuited for this analysis, both because they are longitudinal and because they contain information on our key variables of interest, Social Security covered earnings (i.e., Social Security earnings up to the taxable maximum) and benefit timing (to the exact month of receipt), with a high degree of accuracy in comparison to self-report data. Individuals in the sample are between the ages of 62 and 69. We chose the lower bound of age 62 because the early retirement age is 62 and we use the upper bound of age 69 because there is no financial incentive to postpone the receipt of Social Security benefits beyond age 69, though this upper bound could be extended. 7 The individuals in this sample were born between 1920 and In addition to this restriction on age, the sample is further restricted in three ways. First, individuals who collect Social Security disability benefits are omitted from the sample, as their decision to take up benefits and discontinue working is different from that of nondisabled persons. Second, because people who have never worked or have had little attachment to the labor force are unlikely to begin working in their sixties as a result of changing Social Security and/or tax incentives, we restrict our population to those individuals with a significant history of attachment to the labor force. Using the SER data, we limit the sample to those individuals who have accrued Social Security covered earnings for at least 40 quarters. 8 Third, as discussed in the previous section, the population that we examine includes only those individuals who have not yet begun to receive Social Security benefits at time t 1 (i.e., only the subset of individuals who are at risk of transitioning out of the no Social Security receipt state in any given month of the SIPP). So, in each month for which SIPP data have been collected, we identify all individuals in the sample who have just completed any of the relevant birth years (62 69) and determine whether each person is at risk of entering the state of Social Security receipt (i.e., whether he was not receiving Social Security at t 1). 9 If the person is at risk, then he joins the sample; if not, then we drop the case. This construction implies a person-year unit of analysis. Because SIPP panels range in duration from 32 to 40 months, our sample does not include all years at risk of this transi- 6 It is possible that these individuals are postponing benefits because they expect to live a long time and therefore anticipate a high return from fewer years of higher annual benefits. 7 Our sample excludes widows and widowers who take up their benefits at age 60 or This implies that individuals without an SER are omitted from the sample. 9 Because of difficulties determining the exact day of first receipt, and thus a potential to misclassify early-year transitions, we use the eleventh month after a person s birthday to define the end of the person-year. 490

9 Labor Force Participation of Older Workers tion for many respondents. We can observe the same person up to four birth years/ages. Because a person can in theory be at risk for eight years (ages 62 69), we would need a panel of at least 85 months in order to get full coverage on every person in a single monthly birth cohort. 10 Instead, for every month of each of the four SIPP panels, we identify a new group of at-risk individuals at each year of age (62 69). This ensures that we have a close approximation of the cohort wide distribution of person-years at risk for any cohort represented provided that individuals in neighboring cohorts behave similarly across the period. 11 With these inclusion criteria, the final sample is made up of 4,466 person-years. From this sample, we observe 2,114 transitions from no Social Security receipt to Social Security receipt, and there are 2,352 person-years where individuals remain in the state of no Social Security receipt. Of the 2,114 transitions to Social Security receipt, there were 1,254 transitions to the Social Security receipt and no work state (28.8 percent), and 860 transitions to the Social Security receipt and work state (19.3 percent). Of the people who remained in the state of no Social Security receipt, 1,912 were in the no Social Security receipt and work state (42.8 percent), and 440 were in the no Social Security receipt and no work state (9.9 percent). Individuals whom we observe neither collecting Social Security benefits nor working have, on average, less attachment to the labor force, lower earnings, and lower Social Security benefits, but higher asset income. Explanatory Variables Used While data limitations do not allow us to estimate the effects of all of the determinants of work effort discussed above, we are able to capture many of them. Table 2 lists the variables we use and presents their weighted and unweighted means. The principal measure of a person s ability to work and, to some extent, taste for work is her indexed covered earnings from one year ago. 12 The mean for this variable suggests that, on average, members of our sample have covered earnings of just over 80 percent of the national average wage. 13 This is consistent with our expectations, because a substantial fraction of the sample has already left the labor force and wages tend to decline in late career. Our analyses use three chief variables to incorporate the economic resources that an individual brings to retirement. First, an indicator variable denotes whether the person or, if applicable, his spouse is covered by any pension. About threequarters of the observations are covered. The second measure of economic resources near retirement is an individual s expected Social Security benefit at time t (expressed in 1990 wage-indexed dollars). This is simply the higher of his own Social Security Primary Insurance Amount (PIA), the base amount used for calculat- 10 This figure is derived by multiplying seven years (from age 63 to age 69, inclusive) by 12 months per year and then adding one additional month for the last month in which one is age Tests contrasting coefficient estimates for year olds from the same cohort with these pooled observations suggest that this is indeed the case. 12 One limitation of this measure is that earnings are capped at the Social Security taxable maximum. About 7.7 percent of the sample has earnings at time t that are near or above the taxable maximum. 13 We gross up the lagged earnings by a factor of ten to allow for easier interpretation of some of the marginal effects. 491

10 NATIONAL TAX JOURNAL Variables Gender/marital status: Married male Married female Unmarried persons Education: No high school High school More than high school Age Non-Hispanic white TABLE 2 MEAN OF EXPLANATORY VARIABLES (N = 4,466) Unweighted Means Weighted Means Married times (spouse age own age) Covered by a pension Duration of current work spell Earnings at t 1/average wage at t 1 * 10 Natural log of family asset income Monthly Social Security benefit (in 1990 dollars) (4.920) (17.044) (7.920) (3.127) (262.43) (4.861) (17.121) (7.901) (3.150) (260.23) Tax rate on earnings (7.24) Standard deviations for continuous variables are in parentheses (7.22) ing Social Security benefits, or half of his spouse s PIA at time t. 14 In computing Social Security benefits, we take into account the different actuarial reductions and delayed retirement credits associated with leaving work at different ages. The mean monthly Social Security benefit for members of our sample is nearly $600. A third measure of family economic resources is the natural logarithm of annual family asset income. 15 The mean for this variable is about six, corresponding to annual income from assets of just over $400, though the standard deviation indicates that there is a considerable spread in the distribution. To reflect attachment to the labor force, and perhaps the taste for work, we consider one important aspect of a person s earnings trajectory, the duration of her current work spell. This is simply the number of years elapsed since a person 14 To make the computations more tractable, we assume that spouses retire at age 62 and do not account for unobserved former spouses. These choices may lead us to understate potential Social Security benefits, especially for those persons whose spouses are deceased as of the survey date. The fact that we have excluded survivors who take up their Social Security benefits at ages 60 and 61 attenuates this problem. 15 This is set to zero if the total family asset income is below one. 492

11 Labor Force Participation of Older Workers has had a year of zero earnings. This measure is evaluated from the survey data to as far back as 1951 (when SERs begin). The mean value for this variable is 22.4 years for men and 13.8 years for women, suggesting that members of the sample are closely attached to the labor force. This relatively high mean is sensible, given our requirement that members of the sample must have completed at least forty quarters of Social Security covered work, and also the high representation (in personyears) of persons who delay Social Security retirement. Taste for work is further explored in the model through the use of several education variables. We code education categorically, distinguishing among high school dropouts, high school graduates, and persons with some college. We find that members of our sample are relatively highly educated. About two-fifths (40.5 percent) have at least some college, over one-third (35.5 percent) have a high school education, and just under onequarter (24.3 percent) have less than a high school education. As described earlier, to capture an individual s return from working, we consider her net tax rate on earnings once income taxes, including income taxation of Social Security benefits, Social Security payroll taxes, and benefit reductions imposed by the retirement test, have been calculated. For married people, these rates depend on spouses earned and unearned income. 16 To compute this measure, we first determine an individual s potential earnings, an approximation of the labor income that the person could garner through employment. This is a function of her past covered wages (again, indexed by the average wage, to make them comparable across years), either average earnings from ages 56 to 61 or earnings in one s most recent year in the labor force, whichever is higher. 17 The tax rate on earnings is obtained by dividing the difference between one s tax liability with these potential earnings and one s tax liability without these earnings by one s potential earnings. Women in our sample face an average tax rate of 20.9 percent, while men on average face a higher rate, 24.0 percent. It is important to note that in using this measure of potential returns to working, we are only able to evaluate the effects of individual-level variation in the Social Security portions of these incentives on work transitions. We are restricted to individual-level variation because there has been no real change in the levels of key Social Security parameters (e.g., the retirement test exempt amount) in recent years, though there has been slight variations in some of the income tax parameters. In addition to these important hypothesized determinants of work status, we use a variety of standard demographic variables as controls in our analysis. The demographic control variables include age, race/ethnicity, and interactions of gender and marital status. To capture differences in behavior by age that may result, for example, from Medicare eligibility, we create separate indicator variables for ages 63, 64, 65, and (age 62 is the reference category). 18 Almost half (47.8 percent) of the observations in the sample are 62 year olds. To capture race/ethnicity, we use a dummy variable for whether one is a non-hispanic white, and about 86.3 percent of the sample is so classified. The 16 In our income tax calculation, key parameters such as the standard deduction vary based on whether one is single or married. All married people are assumed to file jointly. 17 Because covered earnings are censored at the OASDI taxable maximum, we never capture the effects of HI tax on earnings above the taxable maximum. Further, certain persons with very high earnings potential (above the taxable maximum) are classified as facing different income tax rates than they should. 18 In preliminary models, we included year-specific dummy variables to capture period effects (from 1990 to 1995) in addition to the age effects. Because coefficients for these terms did not differ significantly from zero, we omitted the period variables from the final model. 493

12 gender marital status interaction terms differentiate married males, our largest group at 48 percent, and married females, the smallest group at just under 23 percent, from unmarried people, the reference category that comprises the remaining 30 percent of the population. Among married persons, as already noted, spouses earnings and assets are taken into account through both our measure of accrued Social Security entitlement and our measure of the net tax on earnings. We additionally consider spouse s age by calculating the difference between a person s age and his spouse s age. (Unmarried persons receive a value of zero on the variable.) The mean for this variable is about 0.9, reflecting the fact that the persons represented in the sample, disproportionately married males, are likely to be older than their spouses. RESULTS The results of the multinomial logit model suggest that work history and financial returns to working, as well as demographic characteristics, are important determinants of individuals Social Security benefit receipt and work behavior. The estimated coefficients from this model represent the log-odds of being in one state relative to the base case state, where the base case in our analysis is the Social Security receipt and no work state. In addition to the estimated coefficients, we present three tables that help in interpreting the magnitude of the coefficients. Age and family status variables affect the probability of working. The estimated coefficients from the multinomial logit model are shown in Table 3, where the coefficients are expressed relative to the base case state of Social Security receipt and no work. The coefficients in column 1 (outcome 1) provide information about the effects of each variable on the likelihood (log-odds) of transitioning from the state of no Social Security receipt into the state of 494 NATIONAL TAX JOURNAL Social Security receipt and work relative to the likelihood of transitioning into the state of Social Security receipt and no work. The coefficients in columns 2 and 3 are interpreted similarly, where the destination states are no Social Security and work and no Social Security and no work, respectively. While the estimated coefficients for the small group of individuals who neither collect Social Security benefits nor work are presented Table 3, the discussion below does not focus on this group. The estimated coefficients for outcome 1 show that married males are more likely than unmarried persons to work while collecting Social Security benefits (i.e., more likely to transition to Social Security receipt and work relative to Social Security receipt and not work). The coefficient on married female status is of the opposite sign (negative), but the coefficient is not statistically significant. This sign pattern holds for outcome 2 no Social Security receipt and work versus Social Security receipt and no work and both variables are statistically different from zero at the five percent level of significance. Married females (married males) are less (more) likely than unmarried persons to remain in the state of no Social Security receipt and work relative to Social Security receipt and no work. Our results also suggest that persons with higher levels of education are more likely to work, regardless of whether they are collecting Social Security benefits. Individuals with a high school degree or less, as opposed to individuals with at least some college, are less likely to either (1) transition to the state of Social Security receipt and work relative to the state of Social Security and no work or (2) be in the state of no Social Security receipt and work relative to the state of Social Security and no work. Individuals who do not take up benefits in the first year that they are available, age 62, are more likely to work in retirement than individuals who take up benefits

13 Labor Force Participation of Older Workers TABLE 3 MULTINOMIAL LOGIT ESTIMATES OF INDIVIDUALS SOCIAL SECURITY AND WORK DECISIONS Variable Constant Outcome 1: SS and Work versus SS and No Work (0.2512) Outcome 2: No SS and Work versus SS and No Work (0.2253)** Outcome 3: No SS and No Work versus SS and No Work (0.2813) Married female (omitted: unmarried) (0.1505) (0.1364)** (0.1720)** Married male (omitted: unmarried) (0.1381)** (0.1262)** (0.1616) No high school degree (omitted: more than high school degree) (0.1425)** (0.1301)** (0.1683)** High school degree only (omitted: more than high school degree) (0.1208) (0.1116)** (0.1447)** Non-Hispanic white (0.1625)** (0.1397) (0.1821) Age 63 (omitted: age 62) (0.1649)** (0.1323)** (0.1664)** Age 64 (omitted: age 62) (0.1426)** (0.1400)** (0.1832)** Age 65 (omitted: age 62) (0.1628)** (0.1907)** (0.2478) Ages (omitted: age 62) (0.2289)** (0.2054)** (0.2104)** Spouse age own age (0.0124)** (0.0113)** (0.0161) Family covered by a pension plan (0.1344)** (0.1240)** (0.1559)** Natural log of family asset income (0.0198) (0.0184) (0.0246)** Duration of current work spell (0.0040)** (0.0036)** (0.0062) Earnings at time (t 1) (0.0102)** (0.0094)** (0.0168)** Monthly Social Security benefits (0.0003)** (0.0003)** (0.0003)** Tax on earnings (0.0098) Log-likelihood = 4,269 Observations = 4,466 Pseudo R 2 = Standard errors are in parentheses. *Denotes statistical significance at the 0.10 level. **Denotes statistical significanace at the 0.05 level (0.0092)** (0.0107) 495

14 NATIONAL TAX JOURNAL when first available. Table 3 shows that individuals who begin collecting benefits after age 62 are considerably more likely to transition to the state of Social Security receipt and work relative to the state of Social Security receipt and no work. The positive coefficients on the age variables suggest this relationship, as the omitted category is age 62. The effects of age are somewhat different when looking at the relationship between no Social Security receipt and work and Social Security receipt and no work (outcome 2). The 65-year-old individuals who did not collect Social Security benefits at age 64 (time t 1) are less likely than the 62 year olds to remain working nonrecipients than to transition into the Social Security and no work state. A higher rate of transition into full retirement may occur at age 65, because individuals reach the normal retirement age and become eligible for Medicare. Finally, results suggest that being married to an older person reduces the probability of working. Economic status and work histories of individuals also affect their probability of working. Being covered by a pension plan reduces the probability of working. Individuals covered by a pension plan are less likely to either (1) transition to the state of Social Security receipt and work relative to the state of Social Security and no work or (2) occupy the state of no Social Security receipt and work relative to the state of Social Security receipt and no work. This finding is consistent with our expectation because pensions provide an additional source of income for retired persons. Our analysis, however, does not detect a significant relationship between family asset income and the probability of working (see columns 1 and 2). 19 Both the duration of the most current work spell and lagged earnings (time t 1) are positively related to work. An increase in the duration of the most recent work spell or higher earnings last period, all else equal, is associated with an increase in the likelihood of (1) transitioning to the state of Social Security receipt and work relative to Social Security and no work and (2) remaining in the state of no Social Security receipt and work relative to Social Security and no work. Finally, we turn to the two policy variables the monthly Social Security benefit amount and the tax rate on earnings. The effects of both of these variables are statistically different from zero at the five percent level, indicating that changes in these variables can in turn affect individuals Social Security receipt and work behavior. The variables also work in the same direction. The results suggest that higher Social Security benefits and higher tax rates on earnings reduce the probability of working. While the signs of the coefficients provide information about the direction of the effects relative to the base case, they do not provide a clear picture of how the overall probability of falling into each of the four groups changes with a change in the explanatory variables. Tables 4 and 5 examine how changes in the explanatory variables affect the probability of being in each of the four groups discussed above. We examine sequential changes in the explanatory variables to obtain a clear picture of how each explanatory variable affects individuals Social Security receipt and work outcomes. This is done by creating a representative, or base case, individual and introducing changes to her characteristics one by one. We create two representative individuals a 62-year-old male and a 65-year-old male. We choose these two ages so that we can examine how the scheduled in- 19 Notice that family asset income is positive and statistically significant for outcome 3 no Social Security receipt and no work relative to Social Security receipt and no work. These results suggest that higher asset income, all else equal, is associated with an increased likelihood of not collecting Social Security and not working relative to collecting Social Security and not working. 496

15 497 TABLE 4 PROBABILITY OF FALLING INTO EACH OF THE FOUR SOCIAL SECURITY/WORK GROUPS UNDER ALTERNATIVE SCENARIOS: 62-YEAR-OLD MALE Differences from the Base Case are in Parentheses Outcomes Initial: Entire sample Base case* Alternative Characteristics (base case in parentheses): Unmarried (married male) Married female (married male) Education level: *More than high school degree (high school degree only) *No high school degree (high school degree only) Spouse three years younger (spouse same age) Social Security and No Work ( ) ( ) ( ) ( ) ( ) Social Security and Work ( ) ( ) ( ) ( ) ( ) No Social Security and Work ( ) ( ) ( ) ( ) ( ) No Social Security and No Work ( ) ( ) ( ) ( ) ( ) Labor Force Participation of Older Workers Not covered by a pension plan (covered by pension plan) ( ) ( ) ( ) ( ) Duration of current work spell = 24 (duration of work spell = 19) ( ) ( ) ( ) ( ) Earnings at time (t 1) = 8.8 (Earnings at time (t 1) = 8.0) ( ) ( ) ( ) ( ) Monthly Social Security benefit = 482 ( SS benefit = 551) ( ) ( ) ( ) ( ) Tax on earnings = 18 (tax on earnings = 23) ( ) ( ) ( ) ( ) *Base case individual is a married male, with only a high school degree, who is a non-hispanic white, the same age as his spouse, and covered by a pension plan, with a natural log of asset income of 6.0; the duration of his current work spell is 19 years, and his earnings at time t 1 are 8.0 (relative to national average earnings and multiplied by 10), with a monthly Social Security benefit of $551 per month, and his tax on earnings is 23 percent.