Mortality hazard and length of time until death are widely used as health outcome measures and are themselves of fundamental demographic interest. Considerable research has asked whether labor force retirement reduces subsequent health and its mortality measures. Previous studies have reported positive, negative, and null effects of retirement on subsequent longevity and mortality hazard, but inconsistent findings are difficult to resolve because (1) nearly all data confound retirement with unemployment of older workers, and often, (2) endogeneity bias is rarely addressed analytically. To avoid these problems, albeit at loss of generalizability to the entire labor force, I examine data from an exceptional subgroup that is of interest in its own right: U.S. Supreme Court justices of 1801–2006. Using discrete-time event history methods, I estimate retirement effects on mortality hazard and years-left-alive. Some substantive and methodological considerations suggest models that specify endogenous effects estimated by instrumental variables (IV) probit, IV Tobit, and IV regression methods. Other considerations suggest estimation by endogenous switching (ES) probit and ES regression. Estimates by all these methods are consistent with the hypothesis that, on average, retirement decreases health, as indicated by elevated mortality hazard and diminished years-left-alive. These findings may apply to other occupational groups characterized by high levels of work autonomy, job satisfaction, and financial security.
The length of life remaining until death (or its probabilistic determinant, mortality hazard) is of interest and is widely used as a measure of “objective” health (Bound 1991), health outcome, and vitality (see, e.g., the review of 27 studies by Idler and Benyamini 1997).1 Considerable research has investigated the effects of labor force status in general, and retirement in particular, on longevity, mortality, and the health constructs they measure. Some studies have reported that retirement tends to shorten remaining life (Morris et al. 1994; Snyder and Evans 2006; Waldron 2001, 2002), while others found the opposite (Anderson 1985; Handwerker 2007; Munch and Svarer 2005) or no effect (Litwin 2007; Mein et al. 2003; Tsai et al. 2005). However, data limitations and consequent measurement and modeling problems appear to reduce the certainty of these findings and make it difficult to resolve inconsistencies among them.
Here, I reconsider existing data and methods pertaining to this topic. I attempt to avoid measurement and modeling problems by using the venerable demographic strategy of analyzing unusual data from an exceptional population subgroup, of interest in its own right, in which these data problems and their methodological consequences are absent or manageable. That subgroup is justices of the U.S. Supreme Court from 1801 through 2006.
The Data Problem
In data from nearly all modern U.S. population segments, retirement is conflated with involuntary unemployment of older workers. Simply stated, involuntarily unemployed older workers, pension recipients whose employment was involuntarily terminated, and those who were “encouraged” by employers to retire tend to report that they are voluntarily retired (Gustman et al. 1995:S63; Stolzenberg 1989; also see Gustman and Steinmeier 2000:2–5). Empirically, misreporting attributes to retirement the empirically verified, pernicious effects of unemployment (e.g., Gerdtham and Johannesson 2003; Linn et al. 1985; Morris et al. 1994; Voss et al. 2004). Further, misreporting of retirement status confuses key concepts. Although everyday language uses many demographic and labor force concepts imprecisely, since at least 1990, retirement from the civilian labor force has been defined in social science research as a worker’s decision to withdraw from the labor force, or to substantially reduce the hours, intellectual demands, or physical intensity of paid work (see reviews in Lumsdaine 1995; Moen et al. 2001).2 Thus, the retirement decision is a rational, voluntary action by the individual concerned, whether it is made with pleasure or regret, whether made to permit the retiree to care for an ailing relative, to pursue leisure interests, to escape from distasteful working conditions, or for any other reason. In this way, retirement differs fundamentally from involuntary changes in labor force status, including labor force exit due to disability and job loss by firing or layoff.
Usual retirement misreporting problems are obviated in Supreme Court data because justices are constitutionally protected from firing and sheltered by judicial regulations from workplace pressure to resign. That is, justices’ retirement benefits and pay are fixed by law, their working conditions are free of employer manipulation, and regulations prevent them from receiving gifts, payments, or other material inducements to resign or remain in office (U.S. Const. art. III section 1).3 Justices are famously vocal about their intentions to remain on the Court as long as they wish (e.g., Williams 1990),4 and their behavior is generally consistent with these stated intentions, even in the presence of physical decay and “mental decrepitude” (Garrow 2000). After 1800, 23.3% of all years served on the Court have been served by justices already eligible to retire with pension benefits equal to their full pay as working justices. Historically, 49.5% of all justices died in office, without retiring. In short, the law leaves retirement decisions to justices, evidence suggests that they make those decisions individually, and so Supreme Court data appear to avoid confounding retirement with involuntary unemployment.
The Endogeneity Problem
In addition to measurement problems, analyses of retirement effects are well-known for susceptibility to unrecognized endogeneity and consequent identification and estimation problems (Handwerker 2007; Snyder and Evans 2006). Endogeneity arises because voluntary retirement is, in the language of causal effects, a self-selected “treatment.” Health and vitality, as indicated by mortality hazard and remaining length of life, may affect the decision to select this treatment, even as the treatment may affect various measures of health, vitality, mortality hazard, and remaining life. For example, increases in mortality hazard and decreases in years of remaining life are substantially correlated with self-assessed subjective health (Idler and Benyamini 1997); reduced subjective health is often mentioned as a reason for retirement, even by retirees who are able to work (Bound 1991; Parsons 1980a, b, 1982; Reno 1971; Schwab 1974; Sherman 1985; Sickles and Taubman 1986). If these reciprocal effects exist empirically but are unrecognized analytically, they produce endogeneity bias, even in efforts to estimate only retirement effects on longevity, rather than longevity and retirement effects on each other.
A recent analysis sought to overcome endogeneity problems by using Social Security Administration policy change as an instrumental variable to identify retirement effects (Snyder and Evans 2006; however, see Handwerker 2007). That analysis used an actuarial definition of retirement (i.e., receipt of pension benefits) well-suited to pension fund financial analysis, but not as appropriate for the present purpose of understanding effects of retirement decisions on the individuals who make those decisions. Later in the article, I note that unusual features of Supreme Court justice pension policies, and other peculiarities of those pensions, permit the use of pension qualification (rather than pension receipt) as an instrumental variable to identify retirement effects on health and longevity.
Finally, if mortality risk is determined according to one causal regime before retirement but according to another regime after retirement, then that situation would be described as endogenous switching, and it, too, would represent a form of endogeneity (Mare and Winship 1988; Quandt 1972). Below, I explain why Supreme Court data appear to permit instrumental estimation of retirement effects on subsequent longevity, as well as distinguishing between retirement and involuntary unemployment.
Retirement Effects in Special Populations
Substantively, this article’s focus on Supreme Court justices builds upon findings of occupational differences in mortality and retirement patterns (Crimmins et al. 1994; Fletcher 1983, 1988; Guralnik 1962; Hayward et al. 1989; Hayward and Hardy 1985; Johnson et al. 1999; Kitagawa and Hauser 1973). Further, the concentration on Supreme Court justices extends a body of mortality research and labor force exit studies of very small social groups that are characterized by their members’ high achievement, influence, and power (e.g., Abel and Kruger 2005; Gavrilov and Gavrilova 2001; Korpelainen 2000; McCann 1972; Quint and Cody 1970; Redelmeier and Singh 2001a, b; Sorokin 1925; Treas 1977; Waterbor et al. 1988). These analyses are relevant to ongoing policy debates about both retirement in general (Ashenfelter and Card 2002; Gokhale 2004) and term limits for Supreme Court justices (Calabresi and Lindgren 2006). Even in the unlikely case that Supreme Court retirement and mortality patterns are dissimilar to those patterns in any other social group, and unrelated to retirement, health, and mortality in the general population, Supreme Court demography itself is a topic of perennial popular interest, periodic political significance, long-standing legal importance, and general governmental consequence (Garrow 2000; Toobin 2007; Woodward and Armstrong 1979; USA Today2007; Zernike 2007). Preston (1977) suggested the demographic importance of Supreme Court justice mortality patterns, but the topic has escaped previous demographic study.5 Separately, and apparently unaware of relevant demographic research, a long, contentious, self-critical literature in law and political science examined pre-retirement deaths of Supreme Court justices (see the review and critique by Stolzenberg and Lindgren 2010), but that research did not consider mortality following retirement. Finally, analyses presented here address a key question about nonpecuniary effects of work and employment: Is it economically irrational for justices to work after they become eligible to receive retirement pensions equal to their pay?6 If continued life has sufficient value, and if work prolongs life, then the value received for unpaid work would be apparent.
The next section reviews relevant previous findings and theory and presents hypotheses. Then, methodological issues and data are described, results are presented, and implications of findings are discussed.
The Simple Model
Can the observed statistical relationship of retirement to subsequent health (as indicated by subsequent longevity and mortality hazard) be explained by a “simple model” that lacks retirement effects on subsequent health, mortality, and longevity?7 In this simple model, true health is exogenously determined and tends to decline monotonically over time. Like mortality hazard, true health is unobservable, even to the justice. On average, mortality hazard and true health at any particular time are indicated by time left to live. True health stochastically determines justices’ subjective assessments of their own health. Justices tend to retire when they believe that their true health is poor. In brief, the simple model asserts that true health is a cause of both retirement and mortality. If correct, the simple model would explain the positive correlation between mortality and retirement, without any effect of retirement on health or its indicator, mortality.
However, the simple model is inconsistent with statistical data and historical narrative. In particular, although the simple model predicts that justices retire when health fades and death approaches, Garrow (2000) described a regular historical pattern in which justices delay or refuse retirement, despite obvious, even lurid, “mental decrepitude” and physical decline. To wit, 49.5% (as of 2006) of former justices never retired, but died in office. Some may have perished while believing themselves to be in good health, but it seems doubtful, if not absurd, to assert that half of all previous justices died in office without prior awareness of impending death. Further, although the simple model asserts that failing health is the primary signal for justices to resign, Stolzenberg and Lindgren (2010: Table 3) found that political circumstances, pension eligibility, and other factors account for more than nine times as much of the variation in the retirement hazard as years-left-alive. In short, the simple model would be a Procrustean bed for Supreme Court retirement and mortality data.
Next, I consider three competing hypotheses about the effects of retirement on post-retirement longevity.
The Null Effects Hypothesis
The null effects hypothesis states that there is no effect of retirement on subsequent mortality hazard and subsequent longevity, on average and other things equal. This hypothesis asserts that any apparent association between retirement timing and subsequent mortality risk is spurious. Although empirical research methods are poorly suited to testing hypotheses of “no effect,” this hypothesis has considerable precedent. For example, Mein et al. (2003) reported that early retirement at age 60 has no effect on physical health. Tsai et al. (2005:997) concluded that, after adding control variables to their analysis of retirement and mortality of Shell Oil employees, “early retirement at 55 or 60 is not associated with increased survival.” After detailed examination of confounding variables and measurement issues, Litwin (2007:739) concluded that “respondents who had prematurely exited the [Israeli] labour force did not benefit from disproportionately longer lives when compared with the respondents who retired ‘on time.’”
The Increased Mortality Hypothesis
According to the increased mortality hypothesis, retirement increases subsequent mortality hazard (and reduces subsequent longevity), on average and other things equal. In an early empirical result, McMahan et al. (1956) found that military personnel live about two years in retirement for every three years served on active duty, suggesting that delayed retirement prolongs life after retirement. Waldron (2001, 2002) reported finding in several large U.S. national data sets that mortality hazard declines as retirement age rises, controlling for current age. Theoretically, the increased mortality hypothesis arises from the observation that, compared with nonparticipation in the labor force, employment tends to intensify social, physical, and mental activity. Increased physical activity reduces the incidence of “depression, fractures, coronary heart disease and mortality” (Wagner et al. 1992:452; Bortz (1984) called these effects “disuse syndrome”; see also Buchner et al. 1992). Wagner et al. (1992:453) speculated that “although most of the evidence available pertains to physical activity, inactivity in other aspects of life—intellectual, social, interpersonal”—reduces physical health, mental health, and longevity. Snyder and Evans (2006) found attenuated mortality among retired workers who returned to work after their Social Security benefits were reduced by a government policy change; they speculated that work at older ages prolongs life by reducing social isolation, and they cited evidence that social contact reduces mortality risk (Berkman 1995; Berkman and Glass 2000; Berkman and Syme 1979; Blazer 1982; Cohen et al. 1997; Colantonio et al. 1993; House et al. 1988; Putnam 2000; Seeman et al. 1987; Zuckerman et al. 1984). Others have reported that any activity, including work, is an antidote to the “powerful adverse effects on physical health and functional status” of depression (Wagner et al. 1992:458; see also Camacho et al. 1991 and Farmer et al. 1988).
The Reduced Mortality Hypothesis
The reduced mortality hypothesis states that retirement reduces subsequent mortality hazard (and increases subsequent longevity), on average and other things equal. Tsai et al. (2005) wrote, “There is a widespread perception that early retirement is associated with longer life expectancy and later retirement is associated with early death.” A competing risks analysis of Danish data reported that “early retirement prolongs survival for men” (Munch and Svarer 2005:17). Mein et al. (2003) reported that early retirement at age 60 was associated with an improvement in mental health, particularly among high socioeconomic status groups. Voluminous evidence suggests that employment exposes many workers to life-shortening health risks (Cameron 2003; Fletcher 1983, 1988; Guralnik 1962; Kitagawa and Hauser 1973; Johnson et al. 1999), or that retirement increases health (Anderson 1985). For office workers like lawyers and judges, the most apparent work-related health and mortality risks are stress and perceived effort-reward imbalance (Cohen and Syme 1985; House et al. 1988, 1990; Marmot and Theorell 1988; Marmot and Wilkinson 1999; Peter et al. 1998, 2002; Siegrist et al. 1990, 1997). The reduced mortality hypothesis reasons that retirement reduces or eliminates work-related exposure to these and other health impediments and their mortality indicators.
Analysis Strategy, Estimation, and Data
This analysis strategy exploits key features of Supreme Court justices’ employment, including the temporal organization of Court work and the federal judicial pension system.
Discrete-Time Event History Models
In the language of causal inference, I seek to measure the effect of a time-related treatment (retirement) on time-related outcomes (mortality hazard and years-left-alive) for those who select the treatment. Accordingly, I apply event history data and methods with accommodations for self-selection (see below). I use discrete-time methods with a one-year time period because Court terms and data are organized annually: justices customarily resign at the end of the Court’s annual term; the Court structures its activities into annual sessions; and Court pension eligibility rules are based on completed years of service and whole years of age. Consequently, dates and times for Supreme Court careers tend to be rounded to whole years; multiple resignations in the same year tend to occur simultaneously; and relevant time-varying political circumstances tend to exist for whole years. Date rounding, co-occurrence of events, and time-varying independent variables are more easily accommodated by discrete-time event history methods than by continuous-time methods (Yamaguchi 1991). Discrete-time methods also accommodate right censoring (Allison 1995), which occurs for justices who die in office. So, I test hypotheses with discrete-time event history models in which the time unit is one year, the unit of analysis is the justice-year, and variables indicate retirement status, death, remaining years of life, and other events and characteristics of a particular justice in a specific year.
My analyses measure retirement effects on two outcome measures: annual mortality hazard and years-left-alive. Annual mortality hazard is the probability that a specific justice who is alive at the start of a particular year dies during that year. Years-left-alive for each justice-year is the number of additional years after the current year until the relevant justice dies. These forward-looking measures differ from lifetime average annual mortality probability or age at death.
To assure that estimates of mortality hazard are in the interval [0,1] for which probabilities are defined, I use maximum-likelihood probit analysis to measure the effect of retirement on mortality hazard.8 To assure that estimates of years-left-alive are nonnegative, I sometimes use a probit transformation (described below) of years-left-alive or a Tobit analysis of years-left-alive (Amemiya 1985). These methods are well-known but not commonly used together.
Endogeneity by Mediation
As explained earlier, endogenous retirement can be represented as endogenous mediation (shown in Fig. 1) or endogenous switching (discussed in the next section and shown in Fig. 2). In endogenous mediation, an endogenous variable, here retired (an indicator for retirement from active service on the Court), mediates some of the effect of exogenous variables on the endogenous outcome variable, here mortality hazard or longevity. Endogenous mediation is the problem for which instrumental variables (IV) estimation is the standard solution (Amemiya 1985).
In IV endogenous mediation models, identification of the effect of an endogenous mediating variable on an endogenous outcome variable requires at least one instrument. The instrument must have direct effects on the endogenous mediator but is restricted to have only indirect effects on the endogenous outcome. Because those requirements must be satisfied logically and substantively, I now discuss reasons to believe that pension eligible—a variable indicating whether a particular justice in a specific year would be eligible for retirement with a federal judicial pension—has a direct effect on retirement but only indirect effects on subsequent mortality measures.
Pension eligible appears to have the necessary direct effect on retired because pension eligibility removes from retirement the tremendous financial disincentive of salary loss. The federal judicial pensions system was enacted in 1869 in specific response to the stated unwillingness of senile, sick, and feeble justices to lose income by retiring from the Court (Yoon 2006). Indeed, justices appear especially likely to lack savings and other income sources because they are legally forbidden to practice law or to receive income related to legal practice, fiduciary duties, writing articles, endorsements, and other business and professional activities while in active service on the Court (5 U.S.C. App. §§ 501–505; http://www.uscourts.gov/RulesAndPolicies/CodesOfConduct/OutsideEmployment.aspx). Stolzenberg and Lindgren (2010) found that pension eligibility increases the annual odds of retirement from the Supreme Court by a multiple greater than 8, on average, net of the retirement effects of the justice’s age, justice’s time served on the Court, the calendar year, and indicators of political climate.
Three conditions justify the identifying restriction that pension eligibility has no direct effect on mortality or years-left-alive. First, mere eligibility for a pension does nothing to health or its mortality indicators; one must actually receive the pension to spend it in ways that affect mortality and longevity. Second, pension eligible is behaviorally distinct from pension receipt: from 1801 to 2006, 23.3% of justice-years served on the Supreme Court were served by justices who were already pension eligible. Finally, pension eligibility is to some extent determined by government policies completely unrelated to the health, mortality hazard, and longevity of particular justices. In particular, there were no pensions for justices until 1869, when retirement pay equal to full-time pay was instituted for former justices over the age of 70 who had 10 or more years of federal judicial service (i.e., service on any federal court). Thus, justices who serve on lower courts before elevation to the Supreme Court might become pension eligible at a younger age than those without previous federal judicial service. Starting in 1954, pensions were awarded to former justices older than 65 years if they had at least 15 years of federal judicial service, and to those older than 70 with at least 10 years of service. In 1984, pension eligibility was extended to former justices older than 65 with at least 10 years of federal judicial service, for whom the sum of years of age and years of service exceeded 80 (“Rule of 80”; see Yoon 2006). For those justices who ever qualify for the pension, the first quartile of the distribution of age at time of first eligibility is 66 years; the median and third quartile are 70 years, and the maximum is 77 years.
For additional consideration of the suitability of qualified for pension as an instrument for retired in Fig. 1, I also estimate IV analyses on the subset of 57 justices who become qualified for pension at some point in their Court tenure. In this subset, pension eligible varies over the tenure of each justice, so qualified for pension can affect the probability of retirement in any justice-year. But, in this subset only, every retiree receives a pension, so qualified for pension cannot possibly serve as a proxy for pension receipt in retirement, which would indicate financial resources available to promote health after retirement.
In Fig. 1, equations (IV1), (IV2a), and (IV2b) summarize the endogenous mediation model of retirement and mortality. Subscript j refers to the jth justice of the Court. Subscript t refers to the tth time period (year). Retiredjt equals 1 if the jth judge is retired at the start of the tth time period; otherwise, Retiredjt equals 0. Functions that can involve nonlinear and nonadditive transformations of variables are represented by f, g, and h. Random disturbances are represented by ε, ϖ, and ζ. Variables agejt, calendar yearjt, pension eligibilityjt, deathjt, and tenurejt are measures of eponymous characteristics or events, measured in whole years, pertaining to the jth individual during the tth time period.
Endogeneity by Switching
The second representation of endogeneity discussed above is endogenous switching (ES). In ES models here, retired and incumbent justices are permitted to experience different causal regimes (parameter values) for mortality hazard and longevity. If a justice is retired, then mortality (or longevity) is unobserved in the equation for incumbents; if a justice is incumbent, then mortality (or longevity) is unobserved in the equation for retirees. Identification is achieved via instrumental variables or nonlinearities.
Estimation and Tests
To constrain estimated hazards to the [0,1] interval for which they are defined, I use probit, IV probit, and selection-corrected probit methods to estimate mortality hazard models. To constrain estimated years-left-alive to the nonnegative values for which it is defined, I use Tobit, IV Tobit, regression with probit transformation of years-left-alive, and IV regression with a probit transformation of years-left-alive (Stolzenberg 2006:56). The probit transformation is as follows: Y is years-left-alive, Ψ is the transformed value of Y, Φ is the normal cumulative distribution function, and Φ−1 is the inverse normal cumulative distribution function, . Transformation back to years is computed from the inverse . Table 1 summarizes this combination of estimation methods and mortality measures.
Although they are not the subject of this article and serve here only as control variables, age, tenure, and calendar year are well-known to have nonlinear effects on mortality and labor force behavior. These nonlinearities are variously described as compression of morbidity (Fries 2005), historical change, decreasing (or increasing) marginal effects, and, in failure-time analysis, the whimsically named, ∪−shaped “bathtub distribution” (Hjorth 1980). Because many mathematical functions virtually duplicate the same values over a fixed domain, it is neither necessary nor possible to distinguish various functions that might produce the same nonlinear effects in a specific data set. Rather, it is sufficient to use log-fractional polynomial transformations of these variables to parsimoniously permit but not require time variables to have nonlinear effects. Log-fractional polynomial transformations are a simple, mathematically well-behaved, and rich generalization of polynomial regression (Gilmour and Trinca 2005; Royston and Altman 1994).
I analyze data on the universe of Supreme Court justices of the United States from 1801 through 2006. Rather than join a debate over the appropriateness of sampling-based significance tests for population data (Leahey 2005), I report standard errors and significance tests for all equation parameters estimated here but take no position on their appropriateness. Because data contain multiple observations per justice, each justice constitutes an observational cluster, and I calculate robust standard errors with first-order Taylor series linearization correction for clustering (Huber-White “sandwich” estimators; Binder 1983). For some analyses in which statistics of interest are population means of analysis forecasts or predictions, and the predictions themselves are based on nonlinear functions of model estimates, ordinary standard errors are not readily available, so I use bootstrapping to calculate them. Although McCullagh (2000) has criticized bootstrapping with clustered data, Feng et al. (1998) and Field and Welsh (2007) found that bootstraps perform well, particularly when the number of clusters is 50 or more. All analyses reported here have more than 50 clusters.
I examine data on all justices of the United States Supreme Court from 1801 through 2006.9 Data are an annual event history data set consisting of one observation for each year in which each justice of the Court was alive, starting in the year in which the justice took office on the Court and ending in the year in which the justice died. These are the data used in Stolzenberg and Lindgren (2010), with one additional justice-year observation for each year that each justice lived after leaving the Court. Table 2 provides descriptive statistics for justice-years used in analyses here, and Table 3 presents historical statistics on all justices of the Court. From 1801 through the end of 2006, 95 justices served on the Court and subsequently died; one (O’Connor) served, retired in 2006, and lives as this article is written; and 9 have neither resigned nor died. Collectively, justices served 1,825 justice-years on the Court and lived 427 justice-years in retirement.10 Two women have served as justices (O’Connor and Ginsburg); at the end of 2006, both live and only O’Connor has retired. Obviously, gender controls are not possible in these analyses, nor are inferences about gender differences. A reader speculated that results would differ if women were excluded from statistical analyses. However, when analyses were redone with Justices O’Connor and Ginsburg omitted from the data, findings were unchanged or virtually identical.
The statistical analyses of longevity are estimated over all 1,971 justice-years after the year 1800 for 91 justices who died before the year 2007 and who either died in office or who resigned from the Court at the age of at least 55 years.11 Analyses of mortality hazard also include justice-years for justices who have not died as of 2006, for a total of 2,132 justice-years. Variables are as follows:
Retired. A dummy (0,1) variable equal to 0 for a justice-year unless the corresponding justice retired or resigned during that year or before starting service the next year.
Death. A dummy (0,1) variable equal to 0 for a justice-year unless the justice died that year.
Year, year1788, and ln(year1788). Year is the calendar year. Year1788 is year – 1788.12 ln(year1788) is the natural logarithm of year1788; the logarithmic transformation improves the fit of some models. I include calendar year to hold constant mortality and retirement trends.
Age, age squared, and age cubed. Age is the age of the justice in years at the start of the justice-year. Probabilities of death and retirement increase with age. In some analyses, I add age squared and age cubed to the analysis to fit nonlinear age effects.
Tenure, tenure cubed, and tenure cubed × ln(tenure). Tenure is years of service on the Court. The annual probability of job quitting in the working population is known to first decline as tenure increases and then increase with additional tenure (Stolzenberg 1989). Tenure cubed and tenure cubed × ln(tenure) prove useful transformations for fitting nonlinear tenure effects.
Qualified for pension. A dummy variable equal to 0 unless the justice is eligible for a federal judicial pension.
Years-left-alive. In each justice-year, years-left-alive indicates future longevity, or remaining years of life. Years-left-alive for a justice-year is the difference between the calendar year of the justice-year and the calendar year in which the justice ultimately dies.
Because all of the ES and IV analyses of mortality and longevity require a regression or probit analysis of retired, I report those analyses first, in Table 4. Independent variables in these retirement models are qualified for pension and, to fit expected nonlinear temporal effects, polynomials of age, year1788, and tenure. Because qualified for pension serves as an identifying instrument for retirement, a key result in Table 4 is the expected positive, statistically significant (α ≤ .05, one-tailed, robust, cluster-corrected test) coefficient of qualified for pension. Although probit analysis does (and regression does not) constrain probability estimates to [0,1], probability estimates from these two models are similar, with a Pearson correlation of .8411.
Table 1 defines nine models for estimating the effect of retired on years-left-alive, and the upper panel of Table 5 presents empirical estimates of this effect. See Online Resource 1 for details of analyses.
Analyses That Ignore Endogeneity
Models 1a–1d ignore endogeneity but are presented for comparison to IV and ES analyses. Consistent with the increased mortality hypothesis, Models 1a–1d all indicate negative effects of retirement on future longevity, and all are statistically significant (α ≤ .01, one-tailed, robust, cluster-corrected test). All analyses hold constant functions of age, tenure, and calendar year.
In Model 1a, ordinary regression estimates an average of 3.6 years less remaining life (the coefficient of retired) for those who are retired than for those who are not retired (α ≤ .01, one-tailed, robust, cluster-corrected test).
Model 1b applies a probit transformation to years-left-alive, yielding a coefficient of −0.2847 (α ≤ .01, one-tailed, robust, cluster-corrected test). To express that coefficient in intuitively meaningful terms, I evaluate its effect in years at 11 years-left-alive (the median of years-left-alive), where the retirement effect is 3.74 years less remaining life, on average.
In Model 1c, I estimate separate models of probit-transformed years-left-alive for retired and incumbent justices. For each regression, the regression prediction of probit-transformed years-left-alive is computed for each justice-year, the predictions from each equation are retransformed into years, and the predicted years-left-alive if incumbent is subtracted from the predicted years-left-alive if retired. The mean difference between years-left-alive if retired and years-left-alive if incumbent is 6.60 years less life for the retired than for incumbents (significant, α ≤ .01, one-tailed test, clustered bootstrap standard error, with 1,391 replications).
Model 1d is the Tobit regression of years-left-alive on age, age squared, age cubed, year1788, year1788 squared, tenure, tenure squared, and retired. Model 1d resembles Model 1b but uses Tobit analysis rather than probit transformation to assure that predicted longevity is never negative. The significant (α ≤ .01, one-tailed, robust, cluster-corrected test) coefficient of −3.3338 for retired in Model 1d indicates an average of three-and-a-third fewer years-left-alive for the retired than for incumbents.
IV Regression and IV Tobit Analyses
In Models 3a–3c, instrumental variables estimation is used to accommodate the endogeneity of retirement. Again, all of these analyses hold constant the effects of age, tenure, and calendar year, and all indicate a negative impact of retirement on future longevity, consistent with the increased mortality hypothesis.
Model 3a estimates a coefficient of −13.3562 for retired (significant, α ≤ .05, one-tailed, robust, cluster-corrected test), indicating an average of 13.4 years less remaining life for those who are retired than for those who are not retired.13
Model 3b applies a probit transformation to years-left-alive, as well as IV estimation, yielding a coefficient of −1.0366 (significant, α ≤ .05, one-tailed, robust, cluster-corrected test). The curved, unbroken line in Fig. 3 shows Model 3b estimates of years-left-alive if incumbent or if retired. Other things equal, Model 3b estimates that an incumbent justice with 11 years-left-alive (the median of years-left-alive) would survive 9.24 fewer years if retired.
Model 3c combines IV estimation (to accommodate the endogeneity of retirement) with Tobit analysis to accommodate the restriction of years-left-alive (to nonnegative values). I use Model 3c to predict the years-left-alive if incumbent and the years-left-alive if retired for each justice-year. The mean difference between these predictions is 13.58 fewer years left alive for the retired. Figure 3 plots these predictions in the patterned line. IV probit and IV Tobit predictions shown in Fig. 3 are similar.
Endogenous Switching Analyses
Models 5a and 5b use endogenous switching regression to model the endogeneity of retired effects on years-left-alive.
In Model 5a, years-left-alive is measured in its natural metric, and separate equations, corrected for endogenous selection bias, are estimated for the effects of age, year1788, and tenure on years-left-alive. Each equation is used to predict the years-left-alive for each justice in each justice-year if retired and, separately, if incumbent. The mean difference between these estimates is 5.7903 fewer years of remaining life for the retired than for incumbents. Significance testing is accomplished by clustered bootstrapping, with 1,391 replications (significant α ≤ .01, one-tailed test).
Model 5b follows the same procedure as Model 5a, except that the probit transformation is applied to years-left-alive before the analysis, and switching regression estimates are transformed back to years before calculating the difference in remaining life for each justice-year. The mean of that difference is 6.8810 fewer years-left-alive (significant, α ≤ .01, one-tailed test, clustered bootstrap standard error, 1,391 replications) if retired than if incumbent, after holding constant age, tenure, and year1788. In each and every justice-year, predicted years-left-alive if incumbent exceeds its counterfactual, predicted years-left-alive if retired.
Figure 4 is the scatterplot of the ratio of predicted years-left-alive if incumbent to predicted years-left-alive if retired, by age. Curves are fitted by fractional polynomial regression to data for indicated half-century periods. In all periods shown in Fig. 4, the ratio is largest at the youngest ages (and apparently would be even larger below age 55), declines as age increases, and then rises again. As indicated by the graph, for the period 1951–2006, justices who are incumbent at age 65 have twice as many years left alive as those who are retired, other things equal.
Mortality Hazard Analyses
Table 1 defines three probit models for estimating the effect of retirement on annual mortality hazard. Empirical estimates of those effects are shown in the lower panel of Table 5. See Online Resource 1 for analysis details. Results of all analyses are consistent with the increased mortality hypothesis.
Analysis That Ignores Endogeneity
Model 2 is the ordinary probit regression of mortality on year1788, age, age squared, tenure, and retired. The coefficient of retired is 0.4963 (significant, α ≤ .01, one-tailed, robust, cluster-corrected test). The solid line in Fig. 5 graphs the effect of this coefficient on mortality hazard. The distance from the solid line to the lower “equal values” line is the estimated effect of retired on mortality hazard: In the metric of probabilities, according to Model 2, on average, an incumbent justice with an annual mortality hazard of 5% would face a hazard 2.5 times higher, or 12.5%, if retired.
Model 4 is an IV probit analysis of retirement effects on mortality hazard. The coefficient of retired in Model 4 is 0.7538—significantly different from zero (α ≤ .01, one-tailed, robust, cluster-corrected test), and larger than the Model 2 estimate. Based on the Model 4 coefficient, the upper broken line in Fig. 5 shows the IV probit (Model 4) estimate of the retirement effect on mortality hazard. Other things equal, an incumbent with an annual mortality hazard of 5% would face a hazard of 18.6% if retired, according to Model 4.
Analyses of Pension Qualifiers Only
Here, for additional robustness of identifying assumptions, I limit analyses to justices who become pension eligible before departure from the Court. In Table 6, I reestimate the coefficients of retired in that limited sample for analyses 3b, 3c, and 4. These coefficients are approximately equal to and statistically significant at lower α levels than the corresponding coefficients in Table 5. Further, 95% confidence intervals around each coefficient in Table 6 overlap the point estimates for the same quantities in Table 5. In short, findings in Tables 5 and 6 lead to the same conclusions and do not differ meaningfully.
Analyses That Censor Early Retirement Years
In the simple model described earlier, justices simply delay retirement until they perceive that their health is so poor that death is imminent, at which point they resign. If the simple model were correct, then estimates of retirement effects on mortality would weaken or disappear if analyses ignored deaths in the first year or two after retirement. So, for additional robustness, I estimate two additional ordinary probit and IV probit analyses of mortality hazard. The first additional analysis censors data from the first year after retirement. The second analysis censors data from the first two years after retirement. Results are shown in Table 7. In brief, censoring the first year or the first two years of retirement would lead to conclusions that are identical to those drawn from analyses that are not censored.
Endogenous Switching Analysis
In Model 6, retirement effects on mortality hazard are measured with separate selection-corrected probit analyses for retired and incumbent justices. For each justice-year in Model 6, I use observed values of independent variables and estimated parameters from the “retired” equation to calculate the expected mortality hazard if the relevant justice were retired in the corresponding justice-year. Separately, I use the same observed values of independent variables with estimated parameters from the “incumbent” equation to calculate the expected mortality hazard if the relevant justice was incumbent on the Court in the corresponding justice-year. If incumbency occurred in all justice-years, then the mean expected annual hazard would be 0.0433. If retirement prevailed in all justice-years, then mean expected annual hazard would be larger by about one-third, 0.0567, all else equal.14 For parsimony, I calculate the ratio of expected mortality hazard if retired to expected mortality hazard if incumbent. In Fig. 6, that ratio is scatterplotted versus age. Lines in Fig. 6 are obtained by fractional polynomial regression of this ratio on age, fitted separately for four half-century historical periods. In all periods, the ratio declines with increasing age until about age 70, and then increases. Fitted lines indicate average ratios of less than 1 for justices in their 60s and 70s until about 1950. Of the plotted points in Fig. 6, 40.4% indicate a ratio below 1. However, after 1900, the fitted line is always above 1.0. And, in a result not visible from Fig. 6, after 1955, there are no individual justice-years whatsoever for which estimated mortality hazard is lower if retired than if incumbent.
This paper considers the hypothesis that labor force retirement diminishes mortality-based measures of the health of U.S. Supreme Court justices. Because Supreme Court justices have constitutionally guaranteed freedom to keep their positions as long as they choose, Supreme Court data are unusually well-protected against commonplace confounding of voluntary retirement with unemployment. In addition, since 1869, Supreme Court pensions have equaled Supreme Court salaries, obviating purely financial explanations of retirement effects. Analyses here use IV and ES estimation to accommodate the endogeneity of retirement and various probit and Tobit methods to deal with logical constraints on estimates of times and probabilities. Permutation of these models, methods, and dependent variables provides 12 different tests of the hypothesis that labor force retirement accelerates death. To investigate various possible alternative explanations of findings, additional analyses examine retirement effects after deletion of (a) justices who never qualify for pensions, (b) female justices, or (c) justice-years pertaining to the first one or two years of retirement. Although standard errors are large, as is usual in IV and ES estimation, all tests are statistically significant (α ≤ .05, one-tailed, robust standard error corrected for clustering) and inconsistent with the hypothesis that retirement prolongs life. In particular:
The smallest point estimate of the average effect of retirement on longevity is an average loss of 3.3 years of life; the largest point estimate is an average loss of 13.6 years. For comparison, the current remaining life expectancy of 65-year-old Americans is 18.7 years (U.S. Census Bureau 2007: Table 101).
ES analyses estimate an average annual mortality hazard of 4.3% for justices if incumbent and about one-third higher (5.7%) if retired. The ES hazard analysis implies that if mortality hazards were constant, then justices would live, on average, 5.5 years longer if incumbent than if retired.15 This difference is roughly the same as the difference found in the ES analyses of years-left-alive.
Figure 7 indicates that after 1955, ES analyses estimate that retirement would have increased mortality hazard for every justice in every year. From 1901 until 1955, retirement would have increased mortality hazard, on average, but would have reduced mortality hazard for some justices in some years. From 1851 to 1900, retirement would have reduced average annual mortality hazard for justices between the ages of 67 and 76. And from 1801 to 1850, retirement would have reduced average annual mortality hazard for justices between the ages of 63 and 82.
In endogenous mediation models without correction for endogeneity bias, I estimate that if incumbent justices had a mortality hazard of 5% in a particular year, they would have an average hazard of more than 12% if they were retired in that year, other things equal. In models with IV correction for endogeneity bias, if justices had a 5% mortality hazard if incumbent, then retirement would raise that average hazard to more than 18%.16
A Cautious Conclusion
Have these analyses neglected some exogenous variable that both causes retirement and accelerates mortality? I hope not, but neglected variables are always possible. For example, readers have suggested presidential political party as a possible confounding omitted variable. Stolzenberg and Lindgren (2010) used similar data and methods to consider the effect of presidential political party on the timing of retirement from the Supreme Court, but, regardless of the effect of presidential party on resignations, I am aware of no suggestion anywhere that the political party of the U.S. President could directly affect the mortality hazard of individual Supreme Court justices.17Family caregiving responsibility is also cited as a possible omitted variable because other research has reported that workers sometimes retire to care for sick or disabled relatives (Raymo and Sweeney 2006), and caregiving is found to reduce the health of caregivers (Reinhard and Horwitz 1995). But the caregivers described in that research are mostly middle-aged women with modest financial resources, demographically and economically dissimilar to the mostly male, highly educated, and well-paid individuals who are the past and present justices of the Supreme Court.18 So it seems improbable at best that, for Supreme Court justices, the need to perform unpaid family health care work explains the observed association between justices’ mortality hazard and retirement. In brief, political circumstances and family health care responsibilities do not seem to be omitted variables that rob these analyses of internal validity.
And there are the usual questions of external validity. Supreme Court data appear to avoid some important measurement problems that afflict most other retirement and mortality data, but small numbers always require caution. Further, Supreme Court justices are not average labor force participants, nor is work at the Court comparable to work at construction sites, high schools, coal mines, or grocery stores, to name just a few places where people work. These and other limitations should be considered seriously. At best, my findings suggest general patterns in certain other population segments. For example, Supreme Court justices may resemble others characterized by very high achievement and who hold jobs with high employment security, high job autonomy, pleasant working conditions, low work-related physical demands, and high levels of work satisfaction. Although they form a small part of the U.S. population, such persons are a socially and economically important segment of the labor force. Studies suggest that such workers tend to retire later from work than those who are not so characterized (see Hayward et al. 1989; Raymo and Sweeney 2006; Raymo et al. 2008). It seems reasonable to hypothesize that these talented workers who like to work at their very good jobs may well react to retirement in much the same way as Supreme Court justices. Obviously, more data would be needed before generalizing to larger population segments.
As we await that data, results reported here are added to analyses of other, sometimes larger, portions of the population that find negative effects of retirement on subsequent health and longevity. Much of what is known about work and employment effects on health and longevity has been discovered or tested on seemingly unusual population subgroups, including civil servants in England (Stansfeld et al. 1995), residents of Alameda County, California (Camacho et al. 1991), and the Wisconsin high school graduating class of 1957 (Marks and Shinberg 1997). Finally, Preston (1977:171) aptly noted that low mortality among “elderly leadership groups such as union leaders, Supreme Court justices, and Communist Party officials” contributes to their grip on power, thereby making their longevity more consequential than their small numbers might suggest. How interesting, then, that analyses reported here suggest that, at least for U.S Supreme Court justices, a tenacious grip on power seems to contribute to longevity, even as longevity prolongs their hold on high office.
Thanks for advice and criticism go to James Lindgren, without whom this article would not have been possible, and to Robert Willis, Kenneth Land, anonymous reviewers, and members of the University of Wisconsin–Madison Center for Research on Demography and Ecology. The author is responsible for all remaining errors and omissions.
This use of the word objective follows Bound (1991) and the definition in The American Heritage Dictionary of the English Language (1996): a synonym for “observable” or perceived “by someone other than the person affected.” Shang and Goldman (2008:409) compared their predicted life expectancy measure (“predicted life expectancy with some noise”) to several health condition and health risk factor measures, thereby providing evidence of the criterion validity of future longevity as an indicator of current health.
This current social science research usage is consistent with the common-language definition of retirement as “withdrawal from one’s occupation, business, or office” [emphasis added] (American Heritage1996), although it differs from some other definitions. For example, the U.S. Current Population Survey (CPS) accepts, solely as an expedient, jobless respondents’ description of their labor force status as “retired” if they are at least 50 years old; thus, CPS respondents who are coded as “retired” include persons who would be classified as disabled, unemployed, or otherwise if full and accurate information were available (Polivka and Rothgeb 1993:24; see also Rones 1985). This social science definition of retirement differs from actuarial and financial accounting definitions, which usually include only recipients of money payments from pension funds (Society of Actuaries 1992).
Justices can be, but none have been, terminated from office for treason, bribery, or serious crimes.
Justice Thurgood Marshall is reported to have stated for publication, “I have a lifetime appointment and I intend to serve it. I expect to die at 110, shot by a jealous husband” (Williams 1990).
Preston (1977:171) wrote, “Mortality levels obviously have a major influence on the structure of other elderly leadership groups such as union leaders, Supreme Court justices, and Communist Party officials.” Thus, determinants of these anomalous mortality levels are of interest for the identical reason.
This question was asked informally, by Gary S. Becker, of U.S. Federal Judge Richard Posner (personal communication, May 8, 2007).
I thank Robert Willis for suggesting this approach.
I use probit rather than logit or gompit methods (see Manton et al. 1994) for consistency with my use of Tobit, instrumental variables probit, and endogenous switching methods.
I started with a database supplied to Professor James Lindgren by Professor Albert Yoon (see Yoon 2006), based on information obtained from the Administrative Office of the U.S. Courts (Federal Judicial Center n.d.). Lindgren checked some of those data against various sources, including the Congressional Record; corrected errors; and added more data from the Federal Judicial Center (n.d.) and the U.S. Supreme Court (2006) for 1789–1868 and 2003–2006. I added post-retirement data for justices who did not die in office.
From establishment of the Supreme Court in 1789 until the end of 2006, 110 justices served a total of 1,895 justice-years on the court and lived 457 post-resignation justice-years.
Years-left-alive is unobservable for justices still alive as this research is done.
Subtracting 1,788 from the calendar year preserves all information and avoids rounding problems that occurred in initial analyses with STATA version 8, which used calendar year.
A skeptical reader proposed that the estimated effect of retired on mortality must be weaker when estimated with IV analyses than when estimated in analyses that ignore endogeneity. Online Resource 2 uses simulation to examine that concern and finds no support for it.
1/3 ≈ 31% = (0.0567 – 0.0433) / 0.0433. Estimates based on justice-years for which the justice’s age is at least 55. If all ages are included, then the mean hazard if retired is 0.0516 and the mean if incumbent is 0.0380.
Based on the geometric distribution. If mortality is geometrically distributed with annual mortality probability of p, then the expected years until death is 1/p. In this case, 5.458 = 1 / 0.0433 – 1 / 0.0567
For comparison, a recent study found that smoking two or more packs of cigarettes a day (compared with never smoking) would raise a 5% mortality hazard among nonsmokers to a 15.8% hazard. That smoking effect is about midway between my instrumented and uninstrumented probit estimates of the effect of retirement on one-year mortality hazard. So, even the smallest of the IV point estimates of hazard effects of retirement can be characterized as comparable to the effects of heavy smoking on one-year mortality. Of course, length of exposure matters, too: smoking typically starts well before retirement, so lifetime effects of smoking would be much greater than lifetime effects of retirement, even if the annual hazard rate effects of smoking and retirement were identical. Here, I compute probability effects of smoking from Rogers et al. (2005:272), who reported that the largest estimated logistic regression coefficient for a dummy variable for smoking two or more packages of cigarettes a day, compared with never having smoked, is 1.274.
A reader asked for this paper “to convince readers why voluntary retirement is a rational decision at all.” However, this article is an analysis of an effect of retirement on mortality, not an examination of the causes of retirement. See Stolzenberg and Lindgren (2010) for analysis of retirement and death in office by Supreme Court justices.
Available information indicates that only one Supreme Court justice has cited caregiving as a reason for retirement (Sandra Day O’Connor). But Justice O’Connor has placed her husband in a nursing facility, where he is attended by professional caregivers (Zernike 2007). Further, it is impossible to know the correspondence between O’Connor’s actual and stated reasons for retirement, and it is difficult to know what effect, if any, her caregiving responsibilities might have on her mortality, as she remains alive at this writing. In addition, Supreme Court justices are well-paid, so there is reason to believe that they could buy caregiver services in place of their own labor, as Justice O’Connor has done. And, further yet, evidence suggests that Supreme Court justices differ markedly from most contemporary caregivers: according to a 1997 report (National Alliance for Caregiving and AARP 1997:10), caregivers are disproportionately female (74%), less than 50 years of age (64%), have low household income (median $35,000), and lack professional or graduate education (91%). In short, almost everything that is known about caregiving effects on caregivers applies to a population segment that is very dissimilar to Supreme Court justices. Finally, I repeated all analyses in this article after omitting Justice O’Connor and Justice Ginsburg from the data. The omission of these two justices produced no change in findings and virtually no change in estimates.