Would It Kill You to Retire? Testing Short/Long Term/Recurrent Effects of Retirement on All-Cause Mortality Risk

Abstract

This study traced all-cause mortality risk over the course of retirement and tested whether re-retirement impacts mortality risk differently than the first time. The study differentiated retirement on whether prompted by health (health retirement) or not (non-health retirement). Based on data from 1992 to 2016 Health and Retirement Study (HRS), the sample consists of 7747 women and 7958 men who were working at the baseline. Adjusting for physical health before/after retirement, the discrete-time logit model found increased mortality risk within the first year of non-health retirement only for men, regardless of physical health changes. Re-retirement did not raise mortality risk further. Furthermore, health retirement increased mortality for men and women but substantially less after their surviving the first year. The findings urge future study to explore non-physical pathways of an immediate mortality increase for men in retirement, as well as the monitoring of population trends in health retirement and its antecedents.

Background

Over decades, research on retirement and health has been dedicated ultimately to address one policy question: would retirement policies that prolong working years hurt the health of older adults? A key building block of evidence to address this question, is to document whether retirement is positive, negative, or inconsequential to health (Andel et al., 2015; Bonsang et al., 2012; Ekerdt, 2010; Insler, 2014). Health shocks—unanticipated health changes that prompt retirement (e.g., cognitive/physical difficulties in keeping up with work)—are widely acknowledged, but it has been long contested whether retirement itself impacts health (Bonsang et al., 2012; Coe & Lindeboom, 2008; Insler, 2014).

Studying health changes resulting from retirement is methodologically challenging, chiefly due to the need to discern causal retirement effects from health shocks (Celidoni et al., 2017; Insler, 2014; Oi, 2019). The prior studies have made efforts to address those challenges by directly or indirectly taking into account health changes prior to retirement; yet, the results are largely conflicted even among studies using the same data source (Andel et al., 2015; Coe & Lindeboom, 2008; Finkel et al., 2009; Han, 2021; Insler, 2014; Oi, 2019). A growing body of studies has examined all-cause mortality as a potential health outcome of retirement (Wu et al., 2016). One most recent meta-analysis of dozen studies concludes that, in agreement of other studies, retirement regardless of its timing relative to a statutory retirement age does not affect mortality once health prior to retirement is taken into account (Bozio et al., 2021; Carlsson et al., 2012; Hult et al., 2010; Sewdas et al., 2020).

Another unresolved challenge is theoretical—a lack of consensus on exactly what constitutes “retirement effects” on health. It is most commonly agreed that for one to be considered retired, they need be out of work. To distinguish retirement from an unemployment spell, survey-based studies typically rely on the self-reported status of retirement (Bonsang et al., 2012; Han, 2021; Insler, 2014; Oi, 2019; Wu et al., 2016). In reality, this distinction alone is not adequate to quantify retirement effects on health. The timing of retirement and pathways leading to it are increasingly diverse, which means that years they spend in retirement largely vary among today’s retirees and that a growing number of individuals make multiple transitions into retirement (Ekerdt, 2010; Warner et al., 2010). To extend the literature, this study examines if mortality changes that retirees experience shift over time, and also if mortality changes differ whether individuals enter retirement for the first time or more.

Two approaches constitute this study’s strategy to isolate the extent of health changes following, but not preceding, retirement. First, this study differentiates whether individuals retire due to health or not. During retirement prompted by a health decline, subsequent mortality risk would be likely higher than when individuals retire under other circumstances simply due to a downward health trajectory that began prior to retirement (Wu et al., 2016). Second, this study takes into account an array of physical health measures—difficulties with mobility/daily activities, and morbidities as reflective of health before and after retirement. Logically, it is extremely unlikely that retirement immediately causes changes in these health measures, but rather more plausible that changes in the same measures prompt retirement. By exploiting this asymmetry, this study aims to newly ascertain that retirement effects on health can be immediate, and that resulting changes in mortality are independent of physical health.

Positive/Negative retirement effects on health

Aging and death are inevitable for all, but the literature points out considerable heterogeneity and complexity of individual paths from birth to death (e.g., Andel et al., 2015; Calasanti, 2010; Morack et al., 2013; Mungas et al., 2009). The aging of a population is now a wide-spread phenomena across the globe, and policy initiatives surrounding the fiscal dependency of older adults have become a topic of growing contention (Warner et al., 2010). Any changes to the existing entitlement programs and retirement policies would affect population-level trends in the number of years tax payers spend in retirement as opposed to working, and thus public health consequences of early/delayed retirement have garnered academic and public attention (Han, 2021; Insler, 2014).

While some health shocks are completely unanticipated, health problems could become gradually apparent as a result of underlying poor health. In fact, variability in the timing of retirement was first understood as a function of different paces at which individuals experience health decline to the point that one can no longer work—frail individuals leave the labor force early and subsequently die earlier than those who are able enough to keep working (e.g., Kingson, 1982; Logan et al., 1992; Waldron, 2001). Over the past few decades, a growing number of studies have shown retirement directly affects retirees’ health, not the other way around (Andel et al., 2015; Bloemen et al., 2017; Insler, 2014; Wu et al., 2016). At the same time, there are also studies in a similar quantity that support no causal pathways from retirement to health (e.g., Coe & Lindeboom, 2008; Han, 2021; Sewdas et al., 2020). Among the former, whether retirement benefits or harms health has been also contested.

Retirement effects are conceived as positive from one perspective—retirement relieves or alleviates work-related stress/health hazard and other strains in the body (Andel et al., 2015; Finkel et al., 2009). From another, retirement effects can be negative in a wider range of psychological and ecological pathways. Some of the negative psychological changes associated with retirement include, but not are limited to, identity losses/readjustments and the lack of fulfilling activities/routines (Bonsang et al., 2012; Celidoni et al., 2017; Han, 2021; Insler, 2014; Oi, 2019). Retirement prompts a role switch from being a full-time worker to being a non-worker, whose role is highly dependent on post-retirement circumstances involving linked-lives (e.g., spouses) (Calasanti, 2010). Certain circumstances call for caregiving and homemaking and in others, social isolation and a lack of socially productive activities are not uncommon, all of which potentially create physical and psychological strains (Calasanti, 1996; Denton & Spencer, 2009; Gall et al., 1997; Oi, 2019). These circumstantial changes after retirement are known to differ for male and female retirees. For instance, female retirees may be able to adjust better mentally and socially where their male counterparts struggle; at the same time, they are more likely to be burdened with care giving and emotional care for their family members (Calasanti, 1996; Denton & Spencer, 2009; Gall et al., 1997).

Despite the wide range of available speculations and explanations regarding retirement effects (particularly negative ones), the literature is ultimately unclear whether retirement effects are limited to physical changes in the body. Theoretically speaking, it is very unlikely that retirement itself causes morbidities and functional limitations in a short-term, but is nonetheless plausible over a long period due to said ecological and psychological pathways that involve hazard, stress, and strains. Furthermore, some of those pathways imply that retirement effects are recurrent each time individuals make a transition, and the net of positive and negative health changes likely differs between men and women as some of the ecological and psychological pathways are gendered. These plausible, yet untested, features of retirement effects motivate this study to examine the temporality and recurrence of retirement effects separately for men and women, in combination of accounting for changes in physical health before and after retirement, as expanded below.

Testing the temporality and recurrence of retirement effects

This study defines retirement effects as significant mortality (all-cause) changes over a period of time following retirement independently of age-specific mortality. Furthermore, mortality changes induced by retirement are to be understood as the net of positive and negative impacts on health via various mechanisms discussed above. This study therefore tests two following propositions: (1) retirement impacts mortality differently over time; and (2) retirement significantly impacts mortality immediately after retirement, independently of changes in physical health.

Mechanism behind negative and positive health changes caused by retirement are subject to temporality differently—some may be instantaneous and others take time. In other words, net changes in mortality in retirement differ over time. The second proposition is derived from the speculation that changes in physical health due to retirement, positive or negative, are plausible in a long term but are not immediate. Thus if mortality changes occur shortly after retirement, it likely has little or nothing to do with whatever changes in physical health after retirement but rather to do with changes before. Adjusting for physical health changes around retirement not only accounts for health shocks, but also enables this study to examine if retirement effects occur immediately and independently of physical health changes.

Finally, this study tests the third proposition that retirement impacts mortality risk each time upon entry. Some of the causal mechanisms relating retirement to stress imply that one’s health may be affected each time individuals leave work (Andel et al., 2016; Gall et al., 1997; Mazzonna & Peracchi, 2017). Testing the possible recurrence of short-term mortality shocks/benefits due to retirement will inform on potential changes in the force of mortality in response to growing divergence in retirement pathways at the population level (Fasang, 2012). If immediate mortality effects of retirement are recurrent, a population-level change in the force of mortality among older adults may emerge from the continued divergence of retirement pathways. This study tests these three propositions separately for men and women.

Data and Methods

Data are drawn from the Health and Retirement Study (HRS), one of the longest running longitudinal studies based in the U.S. The study sample is limited to those who entered the HRS while working. Some 15705 HRS participants met the study criteria with information available for the job they had at the baseline, as well as the one they had for the longest.

This study constructed mortality data in a discrete-time format, with each interval representing an observation taken from each HRS Wave that they participated in. For example, the first observation is taken from their first participation in the HRS study, which ranges from the HRS year 1992 to 2010. Subsequent observations for each individual are then taken with a near 2-year interval because the HRS is administered to its panel biennially. The second and third observations mean an observation taken from the HRS follow-up nearly 2 years and 4 years after the baseline HRS year, respectively. In each observation, age (right-censored time) is recorded based on the date they completed their in-person interview. Death status was set in the last observation that they were seen alive and the death date was used to record age in those observations in place of age based on the interview completion date. Individual dates of death (if occurred) were taken from HRS exit surveys, which are separate interviews with close associates of the participants who died (widows, widowers, sons, daughters, and younger siblings). They were asked to provide detailed information regarding their death including the exact date of death to the best of their knowledge. All the other observations were set as censored.

The study sample excludes the participants with incomplete baseline information, and list-wise deletion was applied to subsequent observations with missing information on time-varying variables such as physical health. The study data consist of 93062 observations from the 15705 study members (7747 women and 7958 men). The data that include those dropped cases due to the exclusion criteria and list-wise deletion (full data) are comprised of 99273 observations from 17005 individuals. Based on the study data and full data, the 95% confidence intervals of the estimates for retirement status/time since retirement/age do overlap, which assuages concerns of missing not at random to some extent (see Appendix A). Another auxiliary analysis was conducted to see if the estimates differ when deaths and non-responses (e.g., those who were eligible but did not participate or requested to be dropped from studies for reasons other than death) were simultaneously predicted as competing events via the multinomial logit link. Again, the estimates from this model did not significantly differ from those without adjustments for non-responses (see Appendix B).

Retirement statuses, Age, and Time since retirement

In each observation, the participants’ retirement statuses were ascertained in the following categories: working, retired for non-health reasons (non-health retirement), and retired due to health (health retirement). The participants were first determined whether they were working at the time of observation. Among those who were not working, they were considered retired when they identified themselves as “fully retired” based on self-reports. In 1204 subsequent observations, transitions into unemployment were recorded—that is, they did not work nor did they report that they were retired. Those observations were excluded from the study. Among those who identified themselves as retired, reasons for health were determined with three separate sources of information (Han, 2021). The first source is drawn from a series of questions as to reasons why they were not working at the time of observation. If they refer to “poor health” as a reason, their transition was determined as retirement due to health. The second source is from the question asking to what extent poor health was a determining factor for their retirement. The participants were asked to respond with “very important,” “moderately important,” “somewhat important,” and “not important at all.” The responses excluding “not important at all” were counted for health as a reason for retirement. Finally, the third source of information is taken from disability status—observations where individuals reported their disability status as a reason for being out of the work force were considered as transitions into retirement due to health.

Age and time since retirement are discretized (not continuous) into 3-year intervals. This discretization of both time variables is motivated theoretically and methodologically. Health effects of reaching a statutory retirement age on mortality involve certain pathways that are not often relevant once they age past it. In the United States, reaching early retirement age (62)/full-retirement age (65) is indicative of gaining new access to retirement resources and health care (i.e., Medicare). It is therefore recommended that mortality risk during these age periods is estimated separately (Bonsang et al., 2012; Insler, 2014; Warner et al., 2010). As for time since retirement, health changes within the first year of retirement involve immediate health changes (e.g., psychological adjustments) that may become irrelevant after a while in retirement (Celidoni et al., 2017; Denton & Spencer, 2009). Because of those a-priori suppositions widely supported in the literature, this study treats both time units as discrete, rather than modeling continuous functions of time.

Due to the biennial HRS design, the discretization of time into 1-year intervals is not feasible. The mean lapse time between observations is 2.7, so that many respondents “age up” from one interval in a given observation to another in the next 2–3 years. Three-year intervals of age still capture the meaningful separations of age periods prior to early retirement (50–52, 53–55, 56–58, 59–61), between early and full retirement (62–64), and after full retirement (65–67, 68–70, 71–73, 74–76, 77–79, 80–82, 83–85, 85+). This study also discretizes time since retirement into intervals of the same length to synch with age, and 3 years is the minimal length to assure that both health retirement and non-health retirement are observed in each interval: <1 year, 1–3 years, 4–6 years, and 7+ years. Time after retirement is 4.5 years on average across all observations (not shown), so that those retirement time intervals can be interpreted as “immediate,” “short,” “medium,” and “long,” respectively. For both age and time since retirement, the first decimal was rounded up if 0.5 or above, and down if below 0.5.

Physical Health

This study takes into account three separate pieces of information for physical health: activities of daily living (ADL), mobility limitations, and the morbidity index. The ADL is a number of reported difficulty with (1) eating, (2) bathing, (3) getting dressed, (4) getting in and out of bed (scale of 0–4). Mobility limitations were measured with a series of physical tasks including walking several blocks, walking one block, walking across the room, climbing several flights of stairs, and climbing one flight of stairs (scale of 0–5). An index was created as the number of tasks that the participants had difficulty completing (mobility index). The morbidity index is based on the number of reported health conditions from high blood pressure, diabetes, cancer, lung disease, heart disease, stroke, and arthritis (scale of 0–7). The indexes were taken from each observation to capture changes in one’s physical health prior and after retirement.

Other Covariates

Because of sex differences and gendered processes underlying said psychological/ecological pathways from retirement to mortality, analyses were stratified by men and women (Calasanti, 2010; Case & Paxson, 2005). Within each sex sample, the following variables were included to account for their roles in differentiating those who work from those who leave the work force as well as their mortality: race/ethnicity, education, income, marital status, and physical demands from the longest job they held and from the job they held at the baseline.

Race/ethnicity was categorized in four groups: Non-Hispanic Whites, Non-Hispanic Blacks, Hispanics, and Others. Education was measured with the number of years of formal education. This study took into account income and marital status as time-varying because they were measured in each observation. Total household income includes those from Social Security/pensions/annuities include sources from the partners in residence (for those who were partnered), and log-transformed after adjusted for inflation (0.001 was added to zero income cases). Marital status was measured with four categories: (1) married (including civil union), (2) widowed, (3) divorced/separated, and (4) never married.

Physical strains at work pose not only health risks in the long run, but also make it increasingly difficult for individuals to keep working (e.g., Kang et al., 2019). In other words, the level of physical demands matters not only to retirement decisions but also to mortality. The HRS collected information regarding the characteristics of the job the respondents had at the time of interview. This study pooled three characteristics that are most relevant to physical demands, measured by three following questions: “my job requires lots of physical effort,” “my job requires the lifting of heavy loads,” and “my job requires stopping, kneeling, or crouching.” The respondents were asked to respond to these questions with one of the following responses: all or almost all of the time (4), most of the time (3), some of the time (2), and none/almost none of the time (1). The index was then created by taking the average of these measures.

The HRS recorded one’s longest held job in the census codes, but did not directly measure physical demands of the job. To circumvent the lack of information, this study matched the census-coded occupations in the HRS data with publically available database called the Occupational Information Network (O*Net). O*Net assesses a number of abilities and tasks for each coded occupation by fielding evaluators in person to randomly sampled job sites that are statistically representative of that occupation group. Among available ratings, those relevant to physical demands are—sitting (reverse-coded), standing, climbing ladders, walking/running, kneeing/crouching, using hands, and bending/twisting. The average value across the ratings for each baseline job was taken to create an index. Because measures for physical demands in the longest job and baseline job are on different scales, both indices were transformed into z-scores.

Analytic Strategy

The study conducts the discrete-time survival analysis that estimates mortality risks for a giving individual i at a given observation j (λ_ij) as follows:

λ_{i j} = \exp (α_{i j} + β_{age} \times {age}_{i j} + β_{1 b} \times {timeout}_{i j} + β_{3 - 5} \times {Physical Health}_{i j} + β_{Control} \times {Control}_{i j})

(1)

In Eq (1), a series of βs indicate logged hazard ratios corresponding to their respective covariates via logit link (i.e., the discrete-time logit model). α_ij is equivalent to the baseline hazard when all covariates set to 0. Age and “timeout” indicate discretized 3-year intervals for age and time since retirement with age at 50–52 (age) and time since retirement (timeout) for less than a year as the respective references. β_age estimates represent the hazard ratios associated with specific age intervals for an individual i in the j observation. Similarly, β_1b estimates are the hazard ratios for the three intervals of time after retirement (1–3, 4–6, and 7+ years) in the j observation. The hazard ratios for physical health measures (“Physical Health” in Eq (1)) for an individual i in the j observation (ADL, mobility, and morbidity indexes) are noted as β_3-5.

An alternative link function, complementary log-log (Gompertz), was considered as it was widely used for mortality analysis. The logit function was selected for the main analysis because of its relative ease in interpreting coefficients and also no significant difference in the fit found between the two link functions based on the likelihood ratio test (p = 0.23, not shown).

λ_{i j} = \exp [α_{i j} + β_{age} \times {age}_{i j} + β_{1 a} \times {Retired}_{i i} + β_{1 b} \times ({timeout}_{i j}, {Retired}_{i i}) + β_{3 - 5} \times {Physical Health}_{i j} + β_{Control} \times {Control}_{i j}]

(2)

Equation (2) introduces one’s retirement status in the j observation, which can be working (reference group), health retirement, or non-health retirement. β_1b*(timeout_ij, Retired _ii) parametrizes mortality effects of “timeout” (β_1b) to be different for health/non-health retirement by each time interval, compared to working (by definition, “timeout” is 0 as long as one is working). β_1a estimates are hazard ratios for health retirement and non-health retirement for less than a year (i.e., referred to as the first year of retirement), with the same reference (working). β_1a and β_1b estimates together quantify mortality changes associated with health retirement and non-health retirement over “timeout,” as opposed to working.

Fitting equation (2) to the mortality data enables the testing of the first two propositions. The first proposition specifies that β_1a estimates for non-health retirement significantly differ from β_1b estimates, meaning that non-health retirement affects mortality within the first year differently than after. This study evaluates whether that is the case by testing against equality between β_1a and β_1b estimates. The second proposition is that non-health retirement within the first year affects mortality in similar magnitude, regardless of physical health measured before and after the transition. The present study evaluates this proposition by testing for the equality of β_1a estimates before and after adjusting for the physical health measures.

Finally, the third proposition is that retirement effects on mortality are greater when retired for the second time or more (i.e., re-retirement). To test this proposition, the model (Eq (2)) was fitted to two sets of data. The first set excludes observations where individuals re-retired (41203 observations for the men, 41416 for the women). The second retains those observations of “re-retirement,” but excludes first-time transitions into retirement. Note that every individual in the second set is also present in the first set but observations taken during their first-time retirement were dropped in the second set (30021 observations for the men, 28786 for the women). β_1a/β_1b estimates for non-health retirement based on the first set are contrasted against those on the second set, to examine if they differ as the third proposition specifies.

Because equation (2) assumes that β_1a, and β_1b are independent of age, retirement effects are estimated to be the same regardless of age at which they leave work. The particular age interval of interest, as mentioned above, is before/after age 62 (early retirement)/65(full retirement). As with age-specific mortality, the effects of retirement on morality are likely to be unique in those age periods (Bonsang et al., 2012). Thus, this study conducted a series of tests to see if β_1a/β_1b estimates differ before/after reaching age 62/65. The first and second propositions are therefore considered for their applicability to age periods dichotomized as before and after age 62/65. Because almost all re-retirement transitions occurred around this age period, the age-dependency of β_1a and β_1b was not tested when evaluating the third hypothesis.

Results

Descriptive statistics shown in Table 1 provide insight into the diversity of work histories that the data contain. For both men and women, the rate of retirement increased in the later observations and health retirement was more common than other reasons (non-health retirement) in any given observation. Note that all rates shown in Table 1 are based on the number of available observations within each jth observation as the denominator (e.g., 5915 was used to calculate the rate of “working” in the sample among the second observations). The rate of widowhood was seen higher in the later observations while that of marriage steadily declines, and widowhood was more common for women than for men in any given observation.

Table 1.

Descriptive statistics of the male and female samples.

Male	1st	2nd	3rd	4th	5th	6th	7th	8th	9th	10th	11th	12th	13th
N	7958	6056	5653	5170	3857	3553	3322	2587	2355	2096	1464	1270	1071
Age	55.26	57.25	59.13	61.10	63.26	65.14	67.06	69.38	71.14	73.32	75.14	76.68	78.58
	(5.35)	(5.16)	(5.09)	(5.01)	(5.15)	(5.11)	(4.91)	(4.94)	(4.68)	(4.47)	(3.94)	(3.76)	(3.66)
Timeout of work	0.00	0.09	0.31	0.62	1.19	1.86	2.67	3.75	4.63	5.89	7.19	8.16	9.47
(years)	(0.00)	(0.34)	(0.82)	(1.39)	(2.06)	(2.77)	(3.48)	(4.26)	(4.91)	(5.66)	(6.36)	(6.86)	(7.53)
Deceased	0.01	0.02	0.03	0.02	0.04	0.04	0.04	0.07	0.08	0.06	0.10	0.10	0.00
Morbidities	0.56	0.66	0.77	0.90	1.01	1.15	1.30	1.46	1.58	1.71	1.85	1.93	2.06
	(0.77)	(0.82)	(0.88)	(0.94)	(0.98)	(1.03)	(1.07)	(1.13)	(1.15)	(1.16)	(1.16)	(1.17)	(1.20)
Mobility limitations	0.29	0.36	0.44	0.49	0.59	0.67	0.73	0.86	0.89	1.01	1.09	1.16	1.27
	(0.77)	(0.85)	(0.95)	(1.02)	(1.12)	(1.19)	(1.23)	(1.31)	(1.33)	(1.42)	(1.47)	(1.48)	(1.54)
ADLs	0.28	0.35	0.43	0.47	0.56	0.65	0.70	0.82	0.85	0.95	1.03	1.10	1.19
	(0.73)	(0.81)	(0.90)	(0.95)	(1.04)	(1.11)	(1.14)	(1.22)	(1.23)	(1.29)	(1.35)	(1.34)	(1.39)
Logged HH income	10.89	10.95	10.84	10.52	10.40	9.97	9.57	9.25	9.07	8.36	7.88	7.86	7.61
	(2.77)	(2.60)	(2.70)	(3.43)	(3.60)	(4.06)	(4.56)	(4.83)	(4.89)	(5.75)	(6.08)	(6.02)	(6.24)
Married	0.80	0.85	0.85	0.84	0.85	0.85	0.84	0.84	0.83	0.83	0.80	0.79	0.77
Widowed	0.01	0.02	0.02	0.03	0.03	0.04	0.05	0.06	0.07	0.08	0.10	0.12	0.14
Divorced/Separate	0.11	0.10	0.10	0.10	0.09	0.08	0.08	0.07	0.07	0.07	0.07	0.06	0.07
Never married	0.05	0.04	0.03	0.03	0.03	0.03	0.03	0.02	0.02	0.02	0.02	0.02	0.02
Working	1.00	0.90	0.82	0.75	0.63	0.55	0.48	0.40	0.36	0.30	0.25	0.21	0.18
NH retirement	0.00	0.06	0.11	0.15	0.23	0.29	0.34	0.40	0.43	0.46	0.49	0.53	0.55
Health retirement	0.00	0.04	0.08	0.11	0.14	0.17	0.20	0.21	0.22	0.25	0.27	0.27	0.28
Never left	1.00	0.93	0.86	0.81	0.76	0.72	0.69	0.66	0.65	0.63	0.62	0.61	0.60
Left once	0.00	0.07	0.14	0.18	0.23	0.26	0.29	0.30	0.31	0.32	0.32	0.33	0.33
Left 2nd+	0.00	0.00	0.00	0.00	0.01	0.02	0.03	0.04	0.04	0.05	0.06	0.06	0.07
Female	1st	2nd	3rd	4th	5th	6th	7th	8th	9th	10th	11th	12th	13th
N	7747	5915	5580	5170	3747	3528	3352	2614	2427	2211	1616	1466	1277
Age	55.01	56.56	58.45	60.42	62.60	64.54	66.54	68.78	70.56	72.73	74.34	76.11	78.13
	(4.69)	(4.94)	(4.90)	(4.78)	(5.12)	(4.96)	(4.83)	(4.79)	(4.55)	(4.22)	(3.39)	(3.31)	(3.31)
Timeout of work	0.00	0.11	0.39	0.76	1.41	2.23	3.07	4.25	5.37	6.70	8.05	9.28	10.76
(years)	(0.00)	(0.36)	(0.91)	(1.50)	(2.22)	(2.98)	(3.67)	(4.36)	(4.96)	(5.65)	(6.20)	(6.66)	(7.24)
Deceased	0.01	0.01	0.02	0.01	0.02	0.02	0.03	0.04	0.05	0.03	0.05	0.07	0.00
Morbidities	0.62	0.74	0.86	0.97	1.06	1.18	1.30	1.41	1.51	1.62	1.66	1.75	1.85
	(0.78)	(0.83)	(0.88)	(0.94)	(0.96)	(1.00)	(1.04)	(1.06)	(1.09)	(1.09)	(1.09)	(1.11)	(1.12)
Mobility limitations	0.62	0.66	0.74	0.82	0.91	0.97	1.04	1.10	1.21	1.30	1.30	1.50	1.55
	(1.05)	(1.08)	(1.16)	(1.22)	(1.32)	(1.31)	(1.36)	(1.37)	(1.41)	(1.48)	(1.44)	(1.53)	(1.55)
ADLs	0.60	0.65	0.72	0.79	0.88	0.93	0.99	1.05	1.15	1.23	1.24	1.41	1.47
	(1.00)	(1.02)	(1.08)	(1.15)	(1.23)	(1.23)	(1.26)	(1.27)	(1.30)	(1.35)	(1.32)	(1.38)	(1.39)
Logged HH income	10.69	10.58	10.31	10.00	9.72	9.18	8.87	8.32	8.03	7.33	6.78	6.89	6.31
	2.55	2.79	3.34	3.85	4.20	4.86	5.07	5.47	5.70	6.37	6.76	6.54	7.03
Married	0.63	0.63	0.61	0.60	0.60	0.59	0.57	0.56	0.54	0.51	0.48	0.42	0.38
Widowed	0.08	0.10	0.11	0.13	0.16	0.18	0.20	0.24	0.26	0.29	0.34	0.39	0.44
Divorced/Separate	0.22	0.21	0.22	0.22	0.19	0.19	0.19	0.17	0.17	0.16	0.15	0.15	0.15
Never married	0.07	0.06	0.06	0.06	0.04	0.04	0.04	0.04	0.04	0.04	0.04	0.04	0.04
Working	1.00	0.87	0.78	0.70	0.59	0.50	0.42	0.34	0.28	0.23	0.19	0.16	0.13
Retired	0.00	0.07	0.12	0.16	0.24	0.29	0.34	0.40	0.44	0.48	0.51	0.53	0.55
Health retirement	0.00	0.06	0.10	0.14	0.18	0.21	0.24	0.26	0.28	0.29	0.30	0.31	0.33
Never left	1.00	0.90	0.83	0.78	0.73	0.69	0.65	0.63	0.61	0.59	0.58	0.58	0.57
Left once	0.00	0.10	0.17	0.22	0.26	0.29	0.31	0.33	0.34	0.34	0.35	0.35	0.35
Left 2nd+	0.00	0.00	0.00	0.01	0.02	0.03	0.04	0.05	0.05	0.07	0.07	0.07	0.08
	Male	SD	Female	SD
Whites	0.62		0.67
Black	0.22		0.15
Hispanic	0.13		0.13
Other	0.04		0.04
Education	13.05	2.96	13.00	3.30
Physical baseline	0.03	0.98	−0.12	1.05
Physical longest	0.18	0.91	−0.25	1.07

Notes: NH = Non-health.

In Table 1, age and time since retirement are listed as continuous, rather than in the discrete format used for the main analysis. As defined above, age indicates when last seen alive in each observation. In case of death, age at death based on the reported date was stored in the last observation where they were last seen alive, and death status was set to 1. For example, the rate of death is 1% in the first observation in both samples, which means 1% of those who participated in the baseline did not make it to the second observation due to death. Attrition rates increased for the later observations at a faster pace for the men than for the women. For instance, in the 11^th observations, the attrition rate is 10% and 5% for the male and female samples, respectively. The mortality rate in the 13^th observations is 0, so that age recorded in those observations serves as the end of observation period for a given individual (right-censoring). The average age of death (not shown) is 72 in both samples, shorter than the current life expectancy since birth in the United States (78) primarily because deaths were not observed for many who survived the observation period, and also because the eligibility of HRS is being aged 50 or over.

For descriptive purposes, the study tracks who in the sample never left the labor force, left once, and left twice or more by each observation. By the end of 13^th observation, nearly 60%/57% of the male/female sample never left the work force during the observation period, and 33/35% of them left the work force at least once. The remaining respondents (7%/9%) had a sequence of returning and then leaving the work force, up to 5 times (exiting twice consisted of the majority). Although not shown, a small minority of those who left the work force once (10%/11% for male and female, respectively, not shown) returned to work later and their second exit was never observed. In all three physical indices, a steady increase is seen for the later observations.

Tables 2 and 3 (2 for the men and 3 for the women) summarize hazard ratios predicting mortality based on variant models of equation (2) with different variable specifications. As stated, all models were fitted to the female and male samples separately. For the sake of brevity, β_age estimates were omitted from these tables and listed in (Appendix D Male Female) instead. The age-specific probabilities of death were plotted and shown in Figure 1 based on those estimates from the full model (the column titled “timeout × Retired + Physical”). In both samples, mortality risk clearly follows a well-known pattern of mortality.

Table 2.

Estimated Hazard Ratios (HR) predicting mortality with four different model specifications fitted to the male sample.

	Timeout			Retired Status			Timeout × Retired			Timeout × Retired + Physical Health
Male	HR	SE	P	HR	SE	P	HR	SE	P	HR	SE	P	95% CI
β_a NH retirement				1.511	0.111	0.000	1.404	0.230	0.038	1.347	0.206	0.046	1.030	1.871
β_b\ 1–3 years	2.034	0.155	0.000				1.618	0.168	0.000	1.540	0.162	0.000	1.253	1.893
4–6 years	1.706	0.151	0.000				1.526	0.179	0.000	1.382	0.165	0.007	1.094	1.745
7+ years	1.454	0.115	0.000				1.359	0.130	0.001	1.150	0.113	0.154	0.949	1.394
β_a health retirement				3.219	0.243	0.000	4.713	0.621	0.000	2.750	0.368	0.000	2.116	3.574
β_b\ 1–3 years							3.585	0.346	0.000	1.971	0.199	0.000	1.617	2.403
4–6 years							2.833	0.331	0.000	1.451	0.178	0.002	1.140	1.846
7+ years							2.534	0.259	0.000	1.338	0.141	0.006	1.088	1.645
v. Married Widowed	1.010	0.104	0.923	0.982	0.104	0.860	0.986	0.103	0.890	0.952	0.104	0.648	0.769	1.178
Divorced/Separated	1.264	0.122	0.015	1.214	0.119	0.049	1.210	0.118	0.050	1.220	0.118	0.040	1.010	1.474
Never married	1.524	0.228	0.005	1.508	0.228	0.007	1.522	0.228	0.005	1.596	0.267	0.005	1.150	2.217
HH income	0.983	0.005	0.001	0.988	0.005	0.014	0.986	0.005	0.007	0.989	0.005	0.035	0.979	0.999
Years of education	0.942	0.009	0.000	0.956	0.009	0.000	0.956	0.009	0.000	0.970	0.011	0.005	0.950	0.991
v. White Black	0.650	0.078	0.000	0.659	0.079	0.000	0.662	0.078	0.000	0.752	0.087	0.014	0.600	0.944
Hispanic	1.225	0.103	0.016	1.188	0.102	0.046	1.192	0.102	0.040	1.278	0.111	0.005	1.078	1.515
Others	0.788	0.174	0.282	0.792	0.176	0.294	0.788	0.174	0.282	0.742	0.174	0.202	0.469	1.173
Morbidities										1.258	0.033	0.000	1.195	1.325
Mobility limitations										1.432	0.141	0.000	1.181	1.735
ADLs										0.974	0.107	0.810	0.786	1.207
Physical DemandsBaseline job										0.985	0.031	0.629	0.926	1.048
Longest job										1.004	0.032	0.897	0.943	1.069
Constant	0.009	0.004	0.000	0.007	0.003	0.000	0.007	0.003	0.000	0.004	0.002	0.000	0.002	0.010

Notes: Parameters with a p-value below 0.05 are in bold. NH = Non-health.

Table 3.

Estimated Hazard Ratios (HR) predicting mortality with four different model specifications fitted to the female sample.

	Timeout			Retired Status			Timeout × Retired			Timeout × Retired + Physical Health
Female	HR	SE	P	HR	SE	P	HR	SE	P	HR	SE	P	95% CI
β_a NH retirement				1.271	0.128	0.017	0.885	0.223	0.629	0.851	0.213	0.518	0.521	1.389
β_b\ 1–3 years	1.724	0.175	0.000				1.305	0.191	0.068	1.190	0.179	0.247	0.887	1.598
4–6 years	1.560	0.177	0.000				1.304	0.206	0.094	1.167	0.189	0.341	0.849	1.604
7+ years	1.364	0.136	0.002				1.176	0.143	0.182	0.920	0.113	0.500	0.723	1.172
β_a health retirement				3.155	0.312	0.000	6.293	0.976	0.000	3.469	0.551	0.000	2.540	4.737
β_b\ 1–3 years							3.183	0.394	0.000	1.675	0.219	0.000	1.297	2.163
4–6 years							2.724	0.379	0.000	1.370	0.199	0.030	1.030	1.821
7+ years							2.522	0.303	0.000	1.148	0.143	0.269	0.899	1.466
v. MarriedWidowed	1.149	0.098	0.103	1.107	0.095	0.236	1.105	0.094	0.239	1.002	0.087	0.986	0.845	1.187
Divorced/Separated	1.371	0.126	0.001	1.253	0.117	0.015	1.238	0.115	0.022	1.070	0.102	0.476	0.888	1.289
Never married	1.408	0.235	0.040	1.356	0.224	0.066	1.352	0.224	0.069	1.225	0.212	0.241	0.873	1.721
HH income	0.981	0.005	0.000	0.986	0.005	0.012	0.984	0.005	0.004	0.989	0.006	0.043	0.978	0.999
Years of education	0.933	0.013	0.000	0.949	0.013	0.000	0.949	0.013	0.000	0.971	0.015	0.050	0.943	0.999
v. WhiteBlack	0.470	0.082	0.000	0.469	0.082	0.000	0.471	0.082	0.000	0.554	0.099	0.001	0.391	0.786
Hispanic	1.021	0.093	0.821	0.954	0.088	0.611	0.958	0.088	0.637	0.942	0.090	0.535	0.782	1.136
Others	0.556	0.184	0.075	0.527	0.178	0.058	0.538	0.179	0.062	0.512	0.171	0.045	0.266	0.985
Morbidities										1.448	0.046	0.000	1.361	1.540
Mobility limitations										2.101	0.209	0.000	1.730	2.553
ADLs										0.611	0.070	0.000	0.488	0.764
Physical DemandsBaseline job										0.968	0.038	0.411	0.895	1.046
Longest job										1.016	0.046	0.723	0.930	1.110
Constant	0.003	0.002	0.000	0.003	0.002	0.000	0.003	0.002	0.000	0.001	0.001	0.000	0.000	0.004

Notes: Parameters with a p-value below 0.05 are in bold. NH = Non-health.

Figure 1.

Age-specific probabilities of death based on the female (left) and male (right) samples.

The first model (titled “timeout”) is a nested model that only includes the three intervals of time since retirement, with a less than 1 year as the reference (β1b). In both samples, mortality risk increased in the 1–3 year interval (by 103% and 72% for men and women, respectively) and remained significantly higher as opposed to working (hazard ratios significantly above 1) in the subsequent intervals. In the second model (titled “retired status”), the hazard ratios for retirement statuses β1a were estimated in place of β1b. In both samples, compared to working, non-health retirement significantly increased mortality risk (by 51.1% and 27.1% for men and women, respectively). As expected, health retirement significantly increased mortality risk (by 222% and 216% for men and women, respectively), The third model (titled “timeout × Retired”) then estimates each interval of time since retirement including the first year, compared to working, but without adjusting for the physical health indices. Beginning with the male sample, non-health retirement is associated with significantly higher hazard ratios across the intervals (e.g., 40.4% increase within the first year and 35.9% after 7 years of more). Retirement effects did not significantly shift after the first year as they failed to reject equality at the p value of 0.648, 0.858, and 0.753 for 1–3, 4–6, and 7+ years, respectively. In the female sample, non-health retirement did not significantly impact mortality across time since retirement—the hazard ratios ranging from 0.885 to 1.305. Equality between β_1a and β_1b estimates held as well, as those estimates did not reject it at the p-value of 0.244, 0.252, and 0.406 for 1–3, 4–6, and 7+ years, respectively.

The next model (titled “timeout × Retired + Physical Health”) represents the full implementation of equation (2), by adding the physical measure indices. Non-health retirement is still associated with significantly higher hazard ratios at large in the male sample—with a 34.7%/38.2% increase within the first year and 4–6 years compared to working. Tests failed to reject equality between the β_1a/β_1b estimates from this model and those not adjusted for physical health, with the lowest p-value being 0.102 for the 7+ year interval. As seen in the previous model, the β_1a and β_1b estimates in the female sample indicate no-significant change in mortality associated with non-health retirement.

Using the estimates from the models “timeout × Retired” and “timeout × Retired + Physical Health,” Figure 2 illustrates mortality patterns by plotting predicted probabilities of death over time since retirement for working, non-health retirement, and health retirement. To clarify, mortality risk associated with working is fixed with time since retirement at 0, but is extended over the period with a horizontal line to facilitate comparisons against mortality plots for retirement and health retirement. The probabilities were estimated counter-intuitively with one regard—no age-based progression of mortality risk and it is instead fixed at the 62–64 age interval to better contrast the mortality patterns in retirement. In the male sample (top), an excess mortality of non-health retirement over working is shown across time since retirement, but not in the female sample (bottom), regardless of physical health measures.

Figure 2.

Probabilities of death specific to each time since retirement (t2), based on the female (left) and male samples (right). Notes: The age-specific hazard fixed at the interval 62–64. Those on the left are unadjusted for the physical health measures (ADL, mobility, morbidities) as shown in the column titled “timeout × Retired” in Tables 2 and 3. Those on the right are adjusted as shown in the column titled “timeout × Retired + Physical health” in Tables 2 and 3.

Although not related to the three propositions in question, the results also show that health retirement significantly increased mortality within the first year and that it did so to a lesser extent in subsequent years in both samples. Equality between β_1a and β_1b estimates for health retirement was in fact rejected at the p-value of 0.01 and less than 0.001 for 3–6 and 7+ years for men, and at the p-value less than 0.001 for all intervals for women.

The models imposed the aforementioned assumption that β_1a, and β_1b are the same regardless of the age interval in which retirement transitions occurred. To test this assumption, a series of likelihood ratio tests were conducted to see if β_1a and β_1b estimates significantly differ before and after reaching 62/65 separately. Age dependency of β_1a and β_1b was tested by comparing the model fit of four variant models against the age-independent model (timeout × Retired + Physical Health, referred henceforth as V1). The first variant estimated only β_1a as different before/after age 62/65 (V2). The second variant estimated only β_1b as age-dependent (V3). In the third variant, both β_1a and β_1b were estimated as age dependent (V4). The final variant modeled the three-way interaction—time since retirement, retirement statuses, and age dependency (V5).

The results are summarized in Table 4, where × indicates which component of age dependency is estimated in each variant: β_1a, β_1b, and the three-way interaction (joint). The results show that, age-dependency was not found significant in all variants before/after age 62. As for age 65, V2 (β_1a) shows a better model fit in both samples. V4 did not improve the model fit over V2, and nor did V5 over V4. In other words, the initial risk associated with non-health/health retirement differed, depending on when it occurred before/after age 65, but there was no significant age dependency in how mortality shifted over time since retirement. Age did not affect mortality patterns over time since retirement regardless of having been retired for health or other reasons (no three-way interaction).

Table 4.

Likelihood ratio tests for the age dependency of β_a, β_b and, β_c for age 62 and 65.

Male age <62 or 62+
	β_a	β_b	Joint		Log Likelihood Difference	P
V1
V2		×		vs V1	1.920	0.781
V3	×			vs V1	3.010	0.492
V4	×	×		vs V2	3.330	0.512
V5	×	×	×	vs V4	3.120	0.321
Male age <65 or 65+
	β_a	β_b	Joint		Log likelihood difference	P
V1
V2		×		vs V1	10.620	0.000
V3	×			vs V1	6.100	0.210
V4	×	×		vs V2	5.380	0.303
V5	×	×	×	vs V4	2.290	0.515
Female age <62 or 62+
	β_a	β_b	Joint		Log likelihood difference	P
V1
V2		×		vs V1	4.380	0.100
V3	×			vs V1	2.930	0.403
V4	×	×		vs V2	2.800	0.424
V5	×	×	×	vs V4	2.560	0.109
Female age <65 or 65+
	β_a	β_b	Joint		Log likelihood difference	P
V1
V2		×		vs V1	7.760	0.012
V3	×			vs V1	2.260	0.520
V4	×	×		vs V2	3.020	0.388
V5	×	×	×	vs V4	1.910	0.591

Notes: × indicates the interaction with the indicator variable of age 62/65 (1 = 62/65 or older, 0 = 61/64 or younger).

The estimates for V2 in the male and female samples are listed in Tables 5 and 6 (5 for the men and 6 for the women). In the male sample, β_1a and β_1b hazard ratios at age 64 or younger are shown as well as β_1a estimates for age 65+ indicate hazard ratios relative to those for age 64 or less (under × age 65+). For instance, the hazard ratio of non-health retirement within the first year is 0.914, and 0.997 for 2–4 years when retired before age 65. After age 65, the hazard ratio of non-health retirement was higher by 72.1% across all intervals (0.914*1.721 and 0.997*1.721 provide the net hazard ratios over working for age 65+). A similar pattern is seen in the female sample. After age 65, the hazard ratio for non-health retirement increased by 76.5% across all intervals, however, with the p-value higher than 0.05 (0.056). Generally, non-health retirement heightened mortality risk after age 65, to a greater extent than before 62.

Table 5.

Estimated Hazard Ratios (HR) predicting mortality with the age-dependent model (V2) and with two subsets of data (Retired second⁺ time and Retired first time) in the male sample.

	V4					Retired 2nd + time vs working					Retired 1st time vs working
Male	HR	SE	P	95% CI		HR	SE	P	95% CI		HR	SE	P	95% CI
β_a NH retirement	0.914	0.226	0.717	0.563	1.485	1.478	0.496	0.244	0.766	2.851	1.306	0.249	0.161	0.899	1.898
β_b\ 1–3 years	0.997	0.197	0.988	0.677	1.468	1.480	0.309	0.060	0.983	2.229	1.532	0.177	0.000	1.222	1.921
4–6 years	0.837	0.184	0.420	0.544	1.289	1.596	0.392	0.057	0.986	2.583	1.301	0.170	0.043	1.008	1.680
7+ years	0.679	0.148	0.075	0.443	1.040	1.037	0.272	0.890	0.620	1.735	1.104	0.112	0.330	0.905	1.346
β_a health retirement	2.982	0.479	0.000	2.177	4.086	1.213	0.447	0.600	0.589	2.496	3.220	0.449	0.000	2.449	4.232
β_b\ 1–3 years	2.183	0.305	0.000	1.660	2.870	1.029	0.252	0.908	0.637	1.661	2.253	0.237	0.000	1.833	2.769
4–6 years	1.651	0.269	0.002	1.199	2.271	1.256	0.315	0.364	0.768	2.055	1.507	0.202	0.002	1.159	1.961
7+ years	1.562	0.251	0.005	1.140	2.139	1.027	0.284	0.923	0.597	1.766	1.327	0.144	0.009	1.072	1.642
X age 65+
β_a NH retirement	1.721	0.370	0.012	1.129	2.623
β_a health retirement	0.861	0.132	0.332	0.637	1.164
v. MarriedWidowed	0.954	0.104	0.662	0.771	1.180	0.909	0.163	0.597	0.640	1.293	0.929	0.111	0.538	0.735	1.175
Divorced/Separated	1.217	0.118	0.042	1.007	1.472	1.289	0.184	0.075	0.975	1.705	1.223	0.124	0.047	1.003	1.491
Never married	1.600	0.268	0.005	1.151	2.223	1.457	0.362	0.130	0.896	2.371	1.694	0.264	0.001	1.248	2.299
HH income	0.989	0.005	0.040	0.979	0.999	0.994	0.011	0.561	0.973	1.015	0.987	0.005	0.016	0.976	0.997
Years of education	0.971	0.011	0.006	0.950	0.992	0.953	0.015	0.002	0.924	0.983	0.970	0.011	0.007	0.950	0.992
v. WhiteBlack	0.743	0.086	0.010	0.592	0.932	0.601	0.119	0.010	0.408	0.886	0.759	0.091	0.021	0.601	0.959
Hispanic	1.275	0.111	0.005	1.075	1.512	1.511	0.184	0.001	1.190	1.918	1.203	0.111	0.046	1.003	1.442
Others	0.744	0.174	0.207	0.470	1.178	1.150	0.319	0.615	0.668	1.980	0.742	0.184	0.229	0.457	1.206
Morbidities	1.257	0.033	0.000	1.194	1.324	1.305	0.055	0.000	1.203	1.417	1.258	0.035	0.000	1.192	1.328
Mobility limitations	1.426	0.139	0.000	1.178	1.728	1.435	0.269	0.050	1.001	2.071	1.396	0.148	0.002	1.133	1.719
ADLs	0.976	0.107	0.826	0.788	1.210	0.918	0.189	0.675	0.613	1.373	0.998	0.118	0.988	0.792	1.258
Physical demandsBaseline job	0.982	0.031	0.559	0.923	1.044	1.023	0.048	0.635	0.932	1.122	0.960	0.032	0.221	0.898	1.025
Longest job	1.002	0.032	0.957	0.941	1.067	1.026	0.048	0.584	0.936	1.124	0.982	0.033	0.590	0.919	1.049
Constant	0.004	0.002	0.000	0.002	0.009	0.005	0.002	0.000	0.002	0.012	0.005	0.002	0.000	0.002	0.010

Notes: Parameters with a p-value below 0.05 are in bold. NH = Non-health.

Appendix A

Estimated Hazard Ratios (HR) predicting mortality with the age-dependent model (V2) and with two subsets of data (Retired second⁺ time and Retired first time) in the female sample.

	V2					Retired 2^nd+ time vs working					Retired 1^st time vs working
Female	HR	SE	P	95% CI		HR	SE	P	95% CI		HR	SE	P	95% CI
β_a NH retirement	0.563	0.187	0.084	0.294	1.080	0.413	0.303	0.228	0.998	1.736	0.957	0.252	0.869	0.572	1.604
β_b\ 1–3 years	0.750	0.219	0.325	0.423	1.330	1.144	0.334	0.645	0.646	2.026	1.187	0.196	0.300	0.858	1.641
4–6 years	0.691	0.212	0.230	0.378	1.263	1.186	0.406	0.618	0.606	2.322	1.136	0.201	0.470	0.804	1.606
7+ years	0.629	0.158	0.073	0.394	1.051	1.001	0.281	0.998	0.578	1.734	0.908	0.116	0.447	0.707	1.165
β_a health retirement	3.698	0.704	0.000	2.546	5.371	2.691	1.012	0.008	1.287	5.623	3.695	0.625	0.000	2.653	5.148
β_b\ 1–3 years	1.823	0.320	0.001	1.292	2.571	1.210	0.327	0.481	0.712	2.057	1.834	0.252	0.000	1.401	2.400
4–6 years	1.524	0.295	0.030	1.042	2.227	0.816	0.261	0.525	0.436	1.527	1.534	0.235	0.005	1.136	2.070
7+ years	1.305	0.251	0.166	0.895	1.903	0.784	0.220	0.386	0.453	1.358	1.208	0.156	0.143	0.938	1.554
X age 65+
β_a NH retirement	1.765	0.536	0.053	0.974	3.199
β_a health retirement	0.886	0.173	0.536	0.605	1.299
v. MarriedWidowed	0.998	0.086	0.985	0.843	1.183	0.803	0.121	0.143	0.598	1.077	1.021	0.094	0.820	0.853	1.223
Divorced/Separated	1.064	0.101	0.515	0.883	1.281	0.907	0.138	0.520	0.673	1.222	1.111	0.110	0.289	0.915	1.349
Never married	1.215	0.210	0.259	0.866	1.705	1.347	0.325	0.217	0.839	2.162	1.238	0.223	0.236	0.869	1.764
HH income	0.989	0.005	0.040	0.979	0.999	0.994	0.011	0.561	0.973	1.015	0.987	0.005	0.016	0.976	0.997
Years of education	0.971	0.011	0.006	0.950	0.992	0.953	0.015	0.002	0.924	0.983	0.970	0.011	0.007	0.950	0.992
v. WhiteBlack	0.553	0.098	0.001	0.391	0.784	0.701	0.180	0.165	0.424	1.158	0.535	0.103	0.001	0.366	0.781
Hispanic	0.940	0.089	0.514	0.780	1.132	1.299	0.187	0.069	0.979	1.723	0.912	0.092	0.363	0.748	1.112
Others	0.509	0.170	0.043	0.264	0.979	0.780	0.325	0.551	0.345	1.763	0.512	0.185	0.064	0.252	1.039
Morbidities	1.257	0.033	0.000	1.194	1.324	1.305	0.055	0.000	1.203	1.417	1.258	0.035	0.000	1.192	1.328
Mobility limitations	1.426	0.139	0.000	1.178	1.728	1.435	0.269	0.050	1.001	2.071	1.396	0.148	0.002	1.133	1.719
ADLs	0.976	0.107	0.826	0.788	1.210	0.918	0.189	0.675	0.613	1.373	0.998	0.118	0.988	0.792	1.258
Physical demandsBaseline job	0.982	0.031	0.559	0.923	1.044	1.023	0.048	0.635	0.932	1.122	0.960	0.032	0.221	0.898	1.025
Longest job	1.002	0.032	0.957	0.941	1.067	1.026	0.048	0.584	0.936	1.124	0.982	0.033	0.590	0.919	1.049
Constant	0.004	0.002	0.000	0.002	0.009	0.005	0.002	0.000	0.002	0.012	0.005	0.002	0.000	0.002	0.010

Notes: Parameters with a p-value below 0.05 are in bold. NH = Non-health.

Based on the estimates from V2, Figure 3 shows the predicted probabilities of death plotted over the intervals of time since retirement. The top left pane of Figure 3 plotted the predicted probabilities of death among men, with the base hazard fixed at age 56–58, which means health retirement/non-health retirement occurred before age 65. In the top right pane, predicted probabilities of death for retirement statuses at age 65 or later are plotted for men, with the base hazard fixed at age 65–67. The bottom pane is structured similarly for women. In both samples, non-health retirement before age 65 did not increase their mortality risk. An increase in mortality associated with non-health retirement at age 65 or older is seen only for men.

Figure 3.

Probabilities of death specific to each time interval for timeout (time since retirement), based on the female (left) and male samples (right) adjusted for the physical measures. Notes: Those in the top pane are estimated with the age-specific hazard fixed at 56–58. Those on the right are with the age-specific hazard fixed at 65–67. The hazard ratios associated with retirement statuses are estimated differently whether aged before 65 or after 65 at the time of retirement (see V2 in Tables 5 and 6). Y-axes for Age < 65 Figures are scaled differently from the rest, for better readability (0–0.05).

In the final part of analysis, the third proposition was tested by contrasting estimates from the model (timeout × Retired + Physical Health, V1) with β_1a,b fitted to the two data sets defined above. For both samples, β_1b estimates based on the first data set (“working vs. retirement for the first time” in Tables 5 and 6) show similar patterns of mortality compared to the main analysis data. β_1b estimates for non-health retirement are not significant at all for both samples in the second set (“working vs. retirement for the second + time”). Regardless, for both male and female samples, there is no significant indication based on the 95% intervals of β_1b estimates that non-health retirement, or even health retirement for that matter, heightened the risk of mortality for the second + time, any more than the first time.

Discussion

Despite growing public/private/academic interests in health impacts of retirement, the causal linkage remains elusive due to conflicting findings in prior research. Some studies found no causal effects (Carlsson et al., 2012; Hult et al., 2010; Sewdas et al., 2020). For others that did, they are divided over whether retirement is harmful or beneficial to health (Andel et al., 2015; Bloemen et al., 2017; Insler, 2014; Wu et al., 2016). Instead, this study conceptualizes retirement effects on mortality as the net of positive and negative changes in health that are subject to temporality during retirement—some changes may be immediate and others manifest in a longer term. In the attempt to better understand what retirement does and does not do to health, this study tests whether they are temporal, related to physical health changes, and/or recurrent. The three propositions were formulated and tested—(1) retirement significantly impacts mortality differently over time; (2) retirement significantly impacts mortality immediately after retirement, independently of changes in physical health, and (3) retirement recurrently impacts mortality.

In short, different conclusions are drawn for the men and women. For men, the first proposition is rejected as equality holds between the first-year estimates and those in the subsequent years. Retirement does negatively impact mortality independently of changes in physical health within the first year (the second proposition supported), but not recurrently when retiring for the second+ time (the third proposition rejected). The findings do not support any of the propositions among women. In both male and female samples, mortality risk increased significantly for health retirement among men and women, and the extent of mortality risk for health retirement tapered off significantly after the first year. Mortality patterns associated with non-health retirement were age-dependent for men, and that the patterns above were applicable only when they retire at age 65 or later.

What do these all mean? First, it can be confidently suggested that negative retirement effects on mortality take hold in a few years, in ways that do not involve physical morbidities including functional issues with mobility/activities of daily living. One plausible explanation, is that non-health retirement induces psycho-social changes for men, such as identity losses/readjustments and the lack of fulfilling activities/routines (Bonsang et al., 2012; Celidoni et al., 2017; Han, 2021; Insler, 2014; Oi, 2019). Exposure to negative health changes is likely episodic and unique to the first-time transition of non-health retirement, based on the lack of support for the third propositions. This increase in mortality does not worsen and also shows little sign of waning for at least several years (Figures 2 and 3). For future studies, exploring pathways involving negative psychological changes that disproportionately affect men is a promising approach consistent with the findings.

Age dependency of non-health retirement effects on mortality is also worth exploring. Non-health retirement at age 62 or younger does not make any difference and it is not until age 65 or older that it increases mortality risk. Taken together, the findings suggest that a transition for non-health reasons during the cusp of early and full retirement likely circumvents or attenuates otherwise detrimental changes in health. It may be that transitioning past the normal retirement age involves unique psycho-social processes including planning and logistical/psychological preparations (Bonsang et al., 2012; Ekerdt, 2010).

An increase in mortality associated with non-health retirement means that working at an older age is more protective against mortality. This finding is comparable with other studies based on the same data source. For instance, one study found that prolonging retirement by working reduces the risk of mortality (Wu et al., 2016). Also the study’s estimate for retirement effects on mortality is within the bound of the meta-analysis based on 12 studies (Sewdas et al., 2020). In the study by Sewdas et al. (2020), it was estimated that retirement, compared to working, increases the risk by 56% (HR 95% CI = 1.41–1.73). This estimate is statistically comparable to a 51% increase for men based on this study (In Table 2, under “Retirement Status”). One point of divergence between the two studies, is that this study found a significant mortality increase even after adjusting for health prior to retirement while the other did not. This could be attributed to a number of reasons, but chief among them are that a) the meta-study did not distinguish health retirement from retirement; and/or that b) health adjustments made in this study are limited strictly to physical health.

Methodologically, the findings highlight the confounding of health selection. It is critical to distinguish retirement transitions on whether prompted by health (i.e., health retirement) or by other reasons (Han, 2021). Frail individuals are “pushed” into retirement and are more likely to be subsequently deceased due to their poor health (Kingson, 1982; Mazzonna & Peracchi, 2017). Accounting for health before and after retirement is regarded as an adequate method to adjust for health selection into retirement (Sewdas et al., 2020). However, this study suggests otherwise, as the effects of non-health retirement could be still confounded by mortality attrition over time. An increase in mortality risk associated with health retirement/non-retirement is immediate, but unlike non-health retirement, the extent of increase following health retirement is substantially less past the first year. Conditioned on survival shortly after retirement, mortality risks for health retirement and non-health retirement tend to converge, as shown in Figures 2 and 3. The most plausible explanation is that many of those who are too frail to work were dropped due to mortality within the first year.

On a relevant note, the study emphasizes that more studies are needed to draw a definitive conclusion regarding the recurrence of retirement effects. The presented estimates that contrast the mortality risk between re-retirement and working are not adjusted for the retirees’ survival upon the first-time transition into retirement. It is very plausible that the estimates in this study are biased downward towards the null. Health among those who re-retired was likely more robust, as those who with poor health had been selected out shortly after the first-time retirement. The reader is therefore reminded that the conclusions drawn for the third proposition are preliminary. Nonetheless, the findings assuage those concerns that increasingly divergent retirement pathways are detrimental to public health (O’Rand, 2006), characterized by the rise of unretirement (returning to work after a period of retirement) and re-retirement (returning to work then leaving again).

Gender differences in health consequences of retirement have been reported by prior research, although with no clear consensus. Based on some prior studies, immediate negative effects of retirement that male retirees experience may be attributed to psychological distresses (e.g., identity losses and isolation) that women tend to cope better with in retirement (Calasanti, 1996; Gall et al., 1997; Quick & Moen, 1998). It could be also argued that the results for male and female respondents differ not because of gendered processes undercutting ecological/psychological pathways, but because deaths were more likely to be reported by surviving spouses of men than of women. This speculation, while plausible, might not be the case with the study data, as the rate of non-responses does not differ significantly for the male and female respondents (Appendix C).

There are some key limitations that are not effectively redressed as they are beyond the scope of this study. First, the discretization of time into 3-year intervals is inherent in the biennial HRS design. Despite some substantive motivations for discretization, the time-based estimates in this study imply that change in mortality is constant within each time interval, which is rather an assumption. Secondly, the findings of this study could be further contextualized for policy implications with trend analyses. It remains to be seen whether mortality changes associated with non-health retirement are ubiquitous across different birth cohorts. Last but not least, further exploration of pathways from non-health retirement to mortality is clearly needed to formulate interventions.

Even with these caveats that limit our understanding of “hows and whys” of retirement effects, the findings do shed light on the broad policy question posed earlier; would retirement policies prolonging work years be detrimental to the health of older adults? The findings are in alignment with other studies that suggest a no (Bozio et al., 2021; Han, 2021; Sewdas et al., 2020). In the context of policy, the findings call for a nuanced approach. While prolonging labor force participation itself may not adversely affect health, health retirement consists of a significant portion of all transitions out of the labor force, which in turn drive the force of mortality among retirees as previously suggested (Coe & Lindeboom, 2008; Han, 2021; Sewdas et al., 2020). In other words, frail/worsening health is a dominant factor in retirement decisions for many old adults today. Echoing concerns raised by prior research (Dudel & Myrskylä, 2017), any policy changes to incentivize labor force participation by limiting social security entitlements are not recommended without attention to vulnerable subgroups of the population that are prone to retirement due to their poor health.

In conclusion, these findings collectively call for a better understanding of near-immediate negative changes in health upon retirement that disproportionately affect men, as well as the monitoring of population trends in health retirement and its antecedents.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Katsuya Oi

Author Biography

Katsuya Oi is Assitant Professor of Sociology at Northern Arizona University. His research revolves around the intersection between social psychology and aging. He is often seen working on his research remotely at his favorite tea shop, Steep in Flagstaff, AZ.

Appendix A

Appendix B

Appendix C

Appendix D

References

Andel

Finkel

Pedersen

N. L.

(2016). Effects of preretirement work complexity and postretirement leisure activity on cognitive aging. Journals of Gerontology Series B: Psychological Sciences and Social Sciences, 71(5), 849–856. https://doi.org/10.1093/geronb/gbv026.

Andel

Infurna

F. J.

Rickenbach

E. A. H.

Crowe

Marchiondo

Fisher

G. G.

(2015). Job strain and trajectories of change in episodic memory before and after retirement: Results from the Health and Retirement Study. J Epidemiol Community Health, 69(5), 442–446. https://doi.org/10.1136/jech-2014-204754

Bloemen

Hochguertel

Zweerink

(2017). The causal effect of retirement on mortality: Evidence from targeted incentives to retire early. Health Economics, 26(12), e204–e218. https://doi.org/10.1002/hec.3493.

Bonsang

Adam

Perelman

(2012). Does retirement affect cognitive functioning? Journal of Health Economics, 31(3), 490–501. https://doi.org/10.1016/j.jhealeco.2012.03.005.

Bozio

Garrouste

Perdrix

(2021). Impact of later retirement on mortality: Evidence from France. Health Economics, 30(5), 1178–1199. https://doi.org/10.1002/hec.4240.

Calasanti

(1996). Gender and life satisfaction in retirement: An assessment of the male model. The Journals of Gerontology Series B: Psychological Sciences and Social Sciences, 51(1), S18–S29. https://doi.org/10.1093/geronb/51b.1.s18.

Calasanti

(2010). Gender relations and applied research on aging. The Gerontologist, 50(6), 720–734. https://doi.org/10.1093/geront/gnq085.

Carlsson

Andersson

Michaëlsson

Vågerö

Ahlbom

(2012). Late retirement is not associated with increased mortality, results based on all Swedish retirements 1991–2007. European Journal of Epidemiology, 27(6), 483–486. https://doi.org/10.1007/s10654-012-9696-4.

Case

Paxson

(2005). Sex differences in morbidity and mortality. Demography, 42(2), 189–214. https://doi.org/10.1353/dem.2005.0011.

10.

Celidoni

Dal Bianco

Weber

(2017). Retirement and cognitive decline. A longitudinal analysis using SHARE data. Journal of Health Economics, 56, 113–125. https://doi.org/10.1016/j.jhealeco.2017.09.003.

11.

Coe

Lindeboom

(2008). Does retirement kill you? Evidence from early retirement windows.

12.

Denton

F. T.

Spencer

B. G.

(2009). What is retirement? A review and assessment of alternative concepts and measures. Canadian Journal on Aging, 28(1), 63–76. https://doi.org/10.1017/S0714980809090047

13.

Dudel

Myrskylä

(2017). Working life expectancy at age 50 in the United States and the impact of the Great Recession. Demography, 54(6), 2101–2123. https://doi.org/10.1007/s13524-017-0619-6.

14.

Ekerdt

D. J.

(2010). Frontiers of research on work and retirement. Journals of Gerontology Series B: Psychological Sciences and Social Sciences, 65(1), 69–80. https://doi.org/10.1093/geronb/gbp109.

15.

Fasang

A. E.

(2012). Retirement patterns and income inequality. Social Forces, 90(3), 685–711. https://doi.org/10.2307/41682674.

16.

Finkel

Andel

Gatz

Pedersen

N. L.

(2009). The role of occupational complexity in trajectories of cognitive aging before and after retirement. Psychology and Aging, 24(3), 563. https://doi.org/10.1037/a0015511.

17.

Gall

T. L.

Evans

D. R.

Howard

(1997). The retirement adjustment process: Changes in the well-being of male retirees across time. The Journals of Gerontology Series B: Psychological Sciences and Social Sciences, 52(3), P110–P117. https://doi.org/10.1093/geronb/52b.3.p110.

18.

Han

S. H.

(2021). Health Consequences of Retirement Due to Non-Health Reasons or Poor Health. Social Science & Medicine, 273(1), 113767. https://doi.org/10.1016/j.socscimed.2021.113767

19.

Hult

Stattin

Janlert

Järvholm

(2010). Timing of retirement and mortality-A cohort study of Swedish construction workers. Social Science & Medicine, 70(10), 1480–1486. https://doi.org/10.1016/j.socscimed.2009.10.043.

20.

Insler

(2014). The health consequences of retirement. Journal of Human Resources, 49(1), 195–233. https://doi.org/10.3368/jhr.49.1.195.

21.

Kang

M.-Y.

Myong

J.-P.

Kim

H.-R.

(2019). Job characteristics as risk factors for early retirement due to ill health: The Korean longitudinal study of aging (2006-2014). Journal of Occupational Health, 61(1), 63–72. https://doi.org/10.1002/1348-9585.12014.

22.

Kingson

E. R.

(1982). The Health of Very Early Retirees. Social Security Bulletin, 45(1), 3.

23.

Logan

J. R.

Ward

Spitze

(1992). As old as you feel: Age identity in middle and later life. Social Forces, 71(2), 451–467. https://doi.org/10.2307/2580019.

24.

Mazzonna

Peracchi

(2017). Unhealthy retirement? Journal of Human Resources, 52(1), 128–151. https://doi.org/10.3368/jhr.52.1.0914-6627r1.

25.

Morack

Ram

Fauth

E. B.

Gerstorf

(2013). Multidomain trajectories of psychological functioning in old age: A longitudinal perspective on (uneven) successful aging. Developmental Psychology, 49(12), 2309. https://doi.org/10.1037/a0032267.

26.

Mungas

Reed

B. R.

Farias

S. T.

DeCarli

(2009). Age and education effects on relationships of cognitive test scores with brain structure in demographically diverse older persons. Psychology and Aging, 24(1), 116. https://doi.org/10.1037/a0013421.

27.

(2019). Does Gender Differentiate the Effects of Retirement on Cognitive Health? Research on Aging, 41(6), 575–601. https://doi.org/10.1177/0164027519828062.

28.

O’Rand

(2006). Stratification and the life course: Life course capital, life course risks, and social inequality Handbook of aging and the social sciences (pp. 145–162). Elsevier.

29.

Quick

H. E.

Moen

(1998). Gender, employment and retirement quality: A life course approach to the differential experiences of men and women. Journal of Occupational Health Psychology, 3(1), 44. https://doi.org/10.1037/1076-8998.3.1.44.

30.

Sewdas

De Wind

Stenholm

Coenen

Louwerse

Boot

Van Der Beek

(2020). Association between retirement and mortality: Working longer, living longer? A systematic review and meta-analysis. J Epidemiol Community Health, 74(5), 473–480. https://doi.org/10.1136/jech-2019-213023.

31.

Waldron

(2001). Links between early retirement and mortalitySocial Security Administration. Office of Policy, Office of Research.

32.

Warner

D. F.

Hayward

M. D.

Hardy

M. A.

(2010). The retirement life course in America at the dawn of the twenty-first century. Population Research and Policy Review, 29(6), 893–919. https://doi.org/10.1007/s11113-009-9173-2.

33.

Odden

M. C.

Fisher

G. G.

Stawski

R. S.

(2016). Association of retirement age with mortality: A population-based longitudinal study among older adults in the USA. Journal of Epidemiology and Community Health, 70(9), 917-923. https://doi.org/10.1136/jech-2015-207097.