Saving Shortfalls and Delayed Retirement

Prior research has suggested that many older Americans have not saved enough to maintain consumption levels in old age. One way older persons might respond to inadequate savings would be to extend their worklives by delaying retirement. This paper examines evidence on this matter using the Health and Retirement Study, a nationally representative panel survey of people age 51-61 in 1992 followed for several years in a panel. We use the data to project household retirement assets and to determine how much more saving would be needed to preserve post-retirement consumption levels. Our research then examines the links between derived saving shortfall measures and delayed retirement patterns. Among nonmarried persons, there is evidence that larger shortfalls do produce delayed retirement, though the effect is not quantitatively large. For married couples, pre-retirement wealth shortfalls do not appear to be significantly associated with delayed retirement. Evidently couples have other means of handling saving shortfalls. Authors’ Acknowledgements Au is a doctoral candidate in the Insurance and Risk Management Department at the Wharton School, University of Pennsylvania. Mitchell is a Research Associate of the NBER; International Foundation of Employee Benefit Plans Professor of Insurance and Risk Management; and Executive Director of the Pension Research Council and Director of the Boettner Center at the Wharton School, University of Pennsylvania. Phillips is an economist at the National Institute on Aging. This paper was presented at the August 2004 meeting of the Retirement Research Consortium in Washington, DC. We thank David Wise and other meeting participants for useful comments. This is a draft: comments are welcome. The authors retain full responsibility for all views contained herein. © 2005 Au, Mitchell, and Phillips. Introduction Much has been written in the economics literature about the determinants of household wealth accumulation, particularly regarding the question of whether households save adequately for their retirement. In this paper, we measure and link the extent of retirement shortfalls to the decision to delay retirement. In particular, we ask whether older households which appear to lack the wealth to retire at a given age actually extend their working lives, so as to more readily pay for an adequate retirement income. Evidence on retirement wealth adequacy is of particular interest given the rapidly aging workforce and the expectation that Social Security faces financial insolvency. There are various benchmarks against which household wealth levels might be compared, to determine retirement saving adequacy. One is derived from a structural model of lifecycle utility maximization which generates estimates of projects saving shortfalls. For example, over a decade ago Bernheim (1992) used dynamic programming to solve for optimal asset accumulation patterns. At that time, his data indicated that households age 35-45 accumulated assets at only one-third the rate of what was needed to maintain consumption levels during retirement; overall, Bernheim reported saving shortfalls averaging 9-19 percent per year. Other research in the field has reached more optimistic conclusions, including Sabelhaus and Manchester (1995) who noted that baby boomers were accumulating more assets than their parents, even though their consumption patterns appeared similar. Using a newer dataset, Engen et al. (2004) also adopt a life cycle model and found smaller saving shortfalls among older households than did Bernheim 1 Older studies in this vein include Kotlikoff et al. (1982); Bernheim (1988); and many others reviewed in Lumsdaine and Mitchell (1999). A spate of recent studies includes work by Aguiar and Hurst (2004); Banks et al. (1998); Engen and Gale (1999); Engen et al. (2004); Haider and Stephens (2004); Hurd and Rohwedder (2003); Hurst (2004); Mitchell and Moore (1998); Moore and Mitchell (2000); and Scholz et al. (2003). in the earlier surveys. A recent study by Scholz et al. (2003) elaborates the life cycle model to incorporate additional sources of uncertainty including medical costs. Again using recent data, that study did document some saving shortfalls, although for most American households retirement assets appeared adequate. This literature comes to somewhat disparate conclusions regarding the shortfalls issue for a variety of reasons. One is the fact that these studies differ with regard to how they model workers’ preferences for leisure and purchased goods. A second is that some authors exclude from consideration important components of retiree wealth such as net housing equity (and sometimes Social Security or pensions). A third difference stems from the use of different benchmarks as a “baseline” for consumption-smoothing. In general, studies that take lifetime earnings as the benchmark tend to find better adequacy rates than do studies which focus on preretirement earnings as the target for replacement. Which is the most appropriate level is, of course, unclear on a priori grounds. A different approach is adopted by Mitchell and Moore (1998) and Moore and Mitchell (2000; hereafter MM), an approach that we extend in the present paper. That methodology builds on a financial planning approach, where a “replacement rate” consumption target is converted into a consumption-smoothing path consistent with a life cycle model. It is worth noting that most financial advisers adopt a pre-specified replacement rate, for instance proposing that households should have assets sufficient to generate lifetime annual income worth 75% of pre-retirement pay in order to avoid a shortfall. By contrast, MM solve for an endogenous replacement rate endogenously, one which derives from the hypothesis that forward-looking 2 Alternative computations adjust retirement consumption to exclude work-related expenses and to account for differential taxation of workers versus retirees (for a good discussion see McGill et al. 2004). households seek to maintain equal real, after-tax, after-saving, levels of preand postretirement consumption. In this framework, a household’s target replacement rate is determined as a function of household earnings and current wealth, as well as demographic information. The earlier MM results indicated that the median US household on the verge of retirement would need to save substantially more in order to retire at 62, though this rate is approximately halved if retirement were delayed to age 65. That research also found that saving shortfalls declined empirically as assets rose, but high earners had substantial undersaving. In the present context, we posit that a household could respond to prospective retirement saving shortfalls in several ways. For example, imagine that a worker planned to retire at age 62, but sometime in his 50’s, he determined that he had not saved enough to provide for a smooth consumption path in retirement. One response might be to stick with the target retirement age but save more, a strategy that would require consumption cuts and increased work effort. Another response might be for the older worker to plan on cutting post-retirement consumption. Yet a third possibility might be to work longer, and indeed most people retiring at age 62 are apparently in good enough health to do so. Further, the results from MM suggest that delaying retirement can substantially reduce the savings shortfalls. 3 Some evidence does suggest that older workers do anticipate reducing consumption after retiring (Hurd and Rohwedder, 2003); on the other hand, Agiar and Hurst (2004) find that expenditures fall after retirement but consumption does not, due to increases in home production. 4 That is, most HRS respondents retiring around age 62 do not report having work-limiting health problems. Burkhauser et. al (1996) first examined the health and economic circumstances of early Social Security beneficiaries, analysis extended by Mitchell and Phillips (2000) who use four waves of HRS data and consider a broader set of health measures. Both studies find that the majority of early retirees report themselves in good health. In what follows, we relate projected retirement savings shortfalls to observed retirement behavior in the HRS, to evaluate their association. We begin by assessing the claim that Americans undersave for retirement by examining the Health and Retirement Study, which contains exceptionally complete information on household wealth linked to administrative records on earnings and benefits from Social Security, including information on financial wealth, net home equity, and retirement pension wealth. With these data, we project and evaluate household retirement wealth, which we then compare to the level of retirement assets needed to smooth real consumption levels over the retirement period. We use this information to calculate the additional amount of annual earnings that would have to be saved to achieve consumption smoothing by the age of retirement. Finally we relate these shortfall measures to the probability of working at ages 62 and 65. Data and Methods Health and Retirement Study (HRS) respondents were first interviewed in 1992 at age 51-61 (along with spouses of any age); additional surveys have been administered every two years thereafter. In the HRS, retirement assets may be divided into three categories: 1. Financial wealth: business assets, financial assets (such as stocks, bonds, and bank accounts less outstanding debt), dedicated retirement assets including IRA and Keogh Accounts, and miscellaneous other financial assets; 5 Our research is therefore similar in spirit to previous work (Bernheim 1988) which relied on the older Retirement History Survey (RHS) but did not use shortfall measures as we have done here. 6 The present study uses all waves from 1992-2002; the 2004 data are not yet publicly available. Many of our analysis variables come from the Rand HRS, a user-friendly version of the HRS produced by the Rand Corporation with financial support from the National Institute on Aging and the Social Security Administration and technical support from the HRS staff at the Institute for Social Research. For more information, see http://hrsonline.isr.umich.edu/data/index.html. 2. Net home equity for homeowners: the market value of owner-occupied housing less outstanding mortgage debt; 3. Retirement pension wealth: equal to the actuarial present value of future social security retirement and survivor benefits and employer pension benefits. To derive the retirement saving shortfall measures, we use the HRS asset information to derive two values (as per MM): each household’s projected shortfall for retirement at 62, and also at age 65. The replacement rate equation sets net (after-tax) income pre-retirement equal to net income post-retirement. This may be expressed as: (1) YP TP – S = YF – TF, where income (Y), taxes (T) and saving (S) are considered both before retirement (P) and post retirement (F). Equation (1) can be rearranged and written in terms of a consumption smoothing replacement rate (RR) by dividing through by pre-retirement income and expressing savings as a proportion of pre-retirement income to generate equation (2). (2) Using an annuity factor (AF) we convert the future income stream to the level of wealth required to sustain YF. If a household’s projected assets (PROJ) are less than their needs, they 7 For detail on how these values are derived using linked data on company pensions and Social Security administrative records, see Gustman et al. (1999). The entire value of home equity is included in these measures, though we recognize that there is some controversy about whether the entire home asset should be counted. For instance Venti and Wise (2001) find that incomepoor but house-rich older families are more likely to reduce equity when they move, while house-poor but income-rich households tend to increase housing equity. 8 Both measures are computed here using pre-retirement pay as the benchmark; see Appendix for a discussion of why using average lifetime pay as the benchmark is problematic. 9 Annuity factors are calculated using the same assumptions used to calculate the value of Social Security wealth and rely on the intermediate assumptions used by the Social Security Administration. For more information, see MM 2000. Y s T T Y Y Y P P F P P F = − − + = ( ) 1 RR need to save the difference between the current period and their retirement age. The difference can be defined as: (3) AF * YF PROJ = AF*[YP *(1-s)– TP + TF] – PROJ. Given the difference between projected and adequate level of assets and the number of years to retirement, we can calculate the percentage of earnings that must be saved in each year leading to retirement to meet the adequate level of assets. We solve for saving and replacement rates simultaneously using an iterative approach. The approach selects an initial replacement rate target and then solves for the target wealth level needed to finance this goal, allowing for differential taxation prior to retirement and during retirement. If household assets would be projected to fall short relative to target assets, the method derives the shortfall saving rate, which 10 As mentioned in the text, one explanation for the disparate conclusions reached by prior studies in the savings adequacy literature is that each employs different benchmarks. For instance, some studies, including ours, focuses on replacing a portion of pre-retirement earnings, while others focus on replacing a portion lifetime earnings. A lifetime earnings benchmark is not used because of the data and assumptions required to modify the earnings benchmark. The identification of the prescribed savings rate in MM relies on the following equation: ], ) 1 ( ) 1 ( [ * * ] ) 1 ( [ * 0 ∑ − = − − + + = − + − − t T i i t T i C F P P r wg Y s PROJ T T s Y AF where the left hand side of the equation is the amount of wealth that must be accumulated to reach the specified benchmark by retirement, the right hand side is the wealth accumulation process from the current period to retirement, wg is an assumed wage growth, and r is an assumed rate of return on savings. A key to this strategy is the fact that pre-retirement prescribed savings rate appears on both the left and right hand sides of the equality, so we can solve for the savings rate as a function of pre-determined arguments. This strategy breaks down when we substitute in the lifetime earnings benchmark. If we set average post-tax and post-savings lifetime earnings as the benchmark, then the above equation is modified to: ], ) 1 ( ) 1 ( [ * * ] ) ' 1 ( [ * 0 ∑ − = − − + + = − + − − t T i i t T i C F r wg Y s PROJ T T s Y AF where Y is average lifetime earnings, T is average lifetime taxes, and s’ is defined as the proportion of average lifetime income that is saved, not as a proportion of pre-retirement income as it is defined in the previous equation. The prescribed savings rate, s, is no longer identified as a function of pre-determined arguments unless. In order to define s’ we would need information on savings historic savings patterns (which we do not observe in the data) or exogenously identify s’ under a host of assumptions. For these reasons, in the present analysis we limit our analysis to the benchmark that focuses on pre-retirement earnings. is equal to the additional fraction of earnings that would have to be saved, to reach the consumption-smoothing goal. This new rate is then used to compute a new replacement rate: if the saving rate is unfeasibly high, the replacement rate is lowered; conversely, the prescribed replacement rate is increased if the saving rate proves below that required for consumption, taxes, and saving to sum to pre-retirement income. The process is repeated until the prescribed saving and replacement rates converge. In the analysis to follow, we express shortfalls as percentage measures: that is, they refer to the additional fraction of pre-tax earnings that each household would have to save prior to retirement to smooth consumption over the remaining lifetime (hereafter “shortfall”). The next issue to which we turn is an examination of how these saving shortfalls might influence subsequent work and retirement behavior. In practice, there are many different ways to calibrate and quantify retirement behavior at older ages (c.f. Lumsdaine and Mitchell 1999; Gustman et al., 1995; Gustman and Steinmeier 2000). In the present analysis, we focus on two objectively measured and clearly work-related outcomes, namely, whether respondents reported themselves as working for pay at age 62 and also at age 65. For the first analysis sample, we select all respondents who attained at least age 62 by 2002, the last year of publicly available HRS dataset. For the second analysis, the sample includes only those who attained at least age 65 by 2002. Selecting our samples in this way avoids incomplete spells due to sample censoring. Both samples include only respondents who initially reported themselves as working for pay at baseline, in 1992. 11 These computations take into account the household’s tax status and life expectancy as of retirement age, using appropriate annuity factors and tax schedules in place at the time of the initial baseline; see Moore and Mitchell (2000). 12 We have also examined alternative definitions of retirement, including whether or not the respondent had positive earnings; results were qualitatively similar to those reported below. The specific estimating equations we evaluate focus on Prob(Work|a), defined as the probability of working at age a, which equals 1 if the individual reported himself as working at the time of the survey and 0 otherwise: (4) Prob(Work|a) = f [Shortfall, X1, X2, X3 ]. Our canonical regressions include the household’s computed shortfall measured in three alternative ways. The first two refer to the shortfalls associated with retirement at a particular age (for age 62, SAVE62, and for age 65, SAVE65); these appear only in the equations relating to work at age 62 and 65 respectively. We also calculate SAVEDIFF which is change in the shortfall if the respondent were to work from age 62 to 65; this is included as a control in some of the specifications for work at age 62. The estimated coefficients from these variables help test the following hypotheses: 1. The probability of working at age 62 (or 65) is positively correlated with shortfalls (i.e., estimated coefficients >0 for SAVE62 and SAVE65). 2. The probability of working at age 62 is positively correlated with returns for working longer (i.e. estimated coefficient >0 for SAVEDIFF). We also include a vector of standard socioeconomic variables (X1) all measured at the 1992 baseline wave of the survey, including the respondent’s educational attainment, marital status, number of children, race/ethnicity, and health. Additional analysis adopts more complete specifications that add some less conventional variables to the canonical set: • X2 includes an indicator of the respondent’s reported financial planning horizon and a baseline indicator of the respondent’s expectation of consumption declines in retirement. • X3 includes indicators of health and marital shocks, indicating whether he experienced a change in health, or became a widow/er since the baseline interview. The empirical models are carried out separately for nonmarried versus married respondents. In the married respondent equations, we include controls for characteristics of their spouses. (A complete variable list appears in Appendix Table 1.) Empirical Results A graphical depiction of the distribution of shortfalls for our sample appears in Figure 1, for both age 62 and 65 retirement. The perpendicular line through the middle of both panels marks the zero shortfall point: the density to the left of the line indicates households who have saved enough to meet their retirement needs, while the opposite is true for households to the right of the line. Note that for both age 62 and 65 retirements, the distribution of shortfalls is skewed. Some households have saved considerable wealth relative to their needs (depicted by the relatively long left tail in each panel in Figure 1), but the majority of households fall to the right side of the distribution, indicating undersaving relative to consumption needs. It is also worth note that delaying retirement shifts more of the distribution to the left of the zero shortfall line and reduces the median shortfall: delaying retirement both eliminates and reduces shortfalls for many households. However, even with a delay in retirement, many households still experience shortfalls.