Health Consequences of Retirement Due to Non-Health Reasons or Poor Health

Abstract

Objectives:

This study aimed to contribute to our understanding of the complex linkage between retirement and health by estimating health consequences of retirement transitions that were not driven by health reasons separately from those caused by poor health, while taking into consideration the health differences that exist between individuals who engage in different labor force behaviors.

Methods:

Ten waves of rich data from the U.S. Health and Retirement Study (N = 9,347; 52,658 person-wave observations) were used to estimate within-person associations between retirement transitions and subsequent health, assessed with self-rated health and depressive symptoms. To account for the bidirectional relationship between retirement and health, retiree’s self-reports of the reasons for labor force withdrawal were used to identify and parse out retirement transitions driven by poor health from the retirement transitions that were unrelated to health reasons.

Results:

Retirement transitions were unrelated to subsequent health if the withdrawal from the labor force was driven by non-health reasons, whereas retirement transitions driven by poor health were associated with worse subsequent health. Retirement transitions that were phased through partial retirement were associated with worse health outcomes compared to transitioning from full-time work to complete retirement.

Conclusions:

Study findings suggest that retirement policies designed to prolong working lives may be implemented without adversely influencing health of older individuals, and potentially delay negative health outcomes associated with retirement for some segments of the older population for whom labor force participation is considered more valuable.

Introduction

Population aging in developed countries around the globe, including the United States, has placed issues about retirement at the center of political and social discourse. This is because many of the challenges posed by population aging, such as shrinking of the workforce and fiscal concerns regarding public pensions and healthcare systems are related to, and also partly caused by, how long people choose to work before they withdraw from the labor force. In response to these challenges, many countries instituted a series of policies (e.g., raising statutory retirement ages) to encourage individuals to prolong their work lives (Henkens et al., 2018).

The health of the aging population is of great importance to the success of such policies, in part because an underlying assumption of these policies is that older workers today have the health capital to work longer (Coile et al., 2016). Yet, this assumption is not always valid across the population, as many individuals retire early due to poor health (Fisher et al., 2016). What also requires careful consideration in the context of these policy measures is the health consequences of retirement, because whether retirement helps or hinders the health of individuals has significant implications for public pension as well as healthcare systems. Reflective of these societal and policy concerns, there has been an increasing interest in examining the health consequences of retirement. The growing body of scientific literature on the topic notwithstanding, there is little agreement on whether retirement is associated with an improvement or deterioration in retiree’s health, or if there is no relationship (for a review, van der Heide et al., 2013, Henning et al., 2016).

The inconsistencies may be attributable to substantive and methodological challenges that limit consensus. Most notably, previous studies on the topic are largely limited by research design issues associated with endogeneity bias, including the potential bidirectional relationship between retirement and health. Although a number of recent studies on the topic are based on study designs that aim to at least partially account for endogeneity bias, findings from these studies also remain largely equivocal (Fisher et al., 2016, Nishimura et al., 2018). In addition, the relationship between retirement transitions driven by poor health and subsequent health has garnered little, if any, substantive attention in earlier research, as the issue is mostly seen as a statistical nuisance that only needs to be controlled for in empirical analyses. This is problematic not only because health of retirees who withdrew from the labor force due to health limitations should be of scholarly interest in its own right, but also because those individuals likely constitute the subpopulation of retirees that has the most influence on public programs and healthcare expenditures.

The central aim of this study was to contribute to our understanding of the complex relationship between retirement transitions and subsequent health. Unlike earlier studies on the topic, this study sought to estimate within-person health consequences of retirement transitions that are unrelated to health reasons separately from those driven by poor health, while taking into consideration the health differences that likely exist between individuals who engage in different labor force behaviors. As discussed below, this study focused on health consequences of complete withdrawal from the labor force, paying attention to how and whether the diverse labor force pathways leading to full retirement influence subsequent health. Using ten waves of data from the Health and Retirement Study, this study addressed the following key research questions:

1. What are the health consequences of retirement driven by non-health reasons?

2. What are the health consequences of retirement driven by poor health?

3. Do health consequences of retirement vary depending on retirement patterns?

Concept of Retirement

Retirement is a term commonly used to refer to withdrawal from the labor force in some capacity, but there is little agreement in the scientific literature on how to define retirement (Denton and Spencer, 2009). This is in part due to its unstable nature; namely, retirement is a social institution whose meaning evolves over time and across cultures (Ekerdt, 2010), and further, it is neither an absolute nor an absorbing state for many individuals, as those who retire often un-retire or re-retire (Calvo et al., 2018). The ambiguity of the term can further be attributed to the different but overlapping criteria used to define retirement. These include, but are not limited to, objective indicators such as non-participation in the labor force and reduction in hours worked and/or earnings, as well as a subjective indicator referring to individuals’ description of themselves as retired or not (Denton and Spencer, 2009).

These indicators capture different dimensions of retirement, which are in turn related to various aspects of the retirement-health nexus. For example, non-participation in the labor force, reductions in earnings, receipt of retirement income, and hours worked are linked to social integration, financial security, and time availability for leisure, respectively, all of which are closely related to health (Henning et al., 2016). As focusing on one specific indicator may result in capturing only a limited fraction of the relationship between retirement and health, researchers often use a more complex measurement approach, where retirement is defined using information from multiple indicators (Denton and Spencer, 2009). In this light, retirement in this study was conceptualized using key information from self-reported retirement and labor force status, where retirement was defined conservatively as complete withdrawal from the labor force. That is, individuals were deemed fully retired if they self-reported as completely retired and did not report working for pay in order to fully capture the psychological, social, and economic effects of complete withdrawal from the labor force in relation to subsequent health.

However, it is important to note that the ways in which individuals reach full retirement are complex and diverse (Calvo et al., 2018), and it is plausible that such diverse patterns of withdrawal from the labor force have differential effects on health. This issue was addressed in this study by testing whether the health effects of the transition to full retirement varied depending on labor force status (i.e., working full time vs. working part-time vs. being partly retired) immediately preceding complete retirement. A small body of earlier work suggested that engaging in some form of intermediate labor force activity is beneficial for health. For example, Wang (2007) showed that holding a bridge job is predictive of sustaining health in retirement and Dingemans and Henkens (2014) demonstrated that a bridge job buffers the adverse effects of involuntary retirement for life satisfaction.

Prior Research on Retirement and Health

Life-course scholarship has long considered retirement as a major life transition that can have a significant and varied impact on late-life health and well-being, leading to both salubrious and detrimental outcomes (Moen et al., 2000). On the one hand, retirement is often perceived as a key life-course milestone, marking a beginning of a new phase of adulthood that can lead to more time spent engaging in meaningful and healthful activities (e.g., time with family and friends, exercise; Moen et al., 2000). Further, the transition can be construed as an escape from demanding work situations for many older workers, thereby having positive health consequences (van den Bogaard et al., 2016). On the other hand, retirement can also lead to detrimental health and well-being outcomes due to a series of losses in terms of structure for daily routines, meaningful social role and occupational attachment, social relationships, and a source of income (Moen et al., 2000).

Reflective of such competing forces exerting varied influence on health surrounding one’s transition to retirement, as well as the wide range of health outcomes examined across studies (e.g., self-rated health, well-being, physical and mental health, disease and illness), findings from earlier research on the topic remain largely inconsistent (for reviews, see van der Heide et al., 2013, Henning et al., 2016). When health is conceptualized as objective biological and physiological markers (e.g., physical functioning, disability, or chronic illness), health does not show a systematic pattern in response to retirement (van der Heide et al., 2013). The picture tends to become less ambiguous when the evidence is confined to findings from longitudinal studies that examined subjective ratings of one’s health, as studies often found that retirement transitions lead to improvements in self-rated health (Eibich, 2015, van den Bogaard et al., 2016, Hessel, 2016, Coe and Zamarro, 2011, Neuman, 2008). The relatively consistent linkages between retirement transitions and subjective health measures are often explained by justification bias, as retirees often justify their retirement decisions by reporting improved post-retirement health (Kuhn, 2018, p. 5). There is some counter-evidence, however, as other studies reported a detrimental effect of retirement on self-rated health (Calvo et al., 2013, Dave et al., 2006, Behncke, 2012) or no relationship (Eyjólfsdóttir et al., 2019). It is worth noting that studies by Calvo et al. (2013), Dave et al. (2006), and Neuman (2008) were all based on the HRS, yet yielded different findings.

Studies that focus on mental health outcomes, such as depression, also paint an unclear picture. Two studies based on European data reported that retirement is unrelated to depression (Behncke, 2012, Coe and Zamarro, 2011); however, other European studies showed that retirement leads to higher depression risk (Heller-Sahlgren, 2017, Kolodziej and García-Gómez, 2019). Such mixed findings are also found among HRS-based studies: some studies indicated that retirement is detrimental for depression (Calvo et al., 2013, Dave et al., 2006) while others showed that retirement leads to reductions in depressive symptoms (Gorry et al., 2018).

There are several explanations for the inconsistent findings in the literature, one of which is that the lack of consensus may be a reflection of heterogeneous health effects of retirement. Namely, researchers have argued that the relationship between retirement and subsequent health is contingent on contextual factors surrounding the transition, such as one’s job satisfaction, occupational characteristics, and voluntariness of the transition (van den Bogaard and Henkens, 2018, Dave et al., 2006). Although this explanation is gaining more empirical support, what also merits further consideration are the methodological challenges associated with the issue of endogeneity. In particular, a large body of recent studies addressing the issue of endogeneity were based on instrumental variable (IV) approaches (Angrist and Pischke, 2008), a quasi-experimental method employed to provide an unconfounded estimation of the causal relationship between retirement and health (Eibich, 2015, Hessel, 2016, Behncke, 2012, Heller-Sahlgren, 2017, Coe and Zamarro, 2011, Gorry et al., 2018). These studies employed pension eligibility age as the instrument, with the key assumption being that these ages are based on policy and are thus exogenous to health except through their effect on retirement transitions (i.e., exclusion restriction). However, study findings based on IV approaches have also been inconsistent (Nishimura et al., 2018). Further, several researchers have articulated how the exogeneity assumption may be violated under several scenarios (Behncke, 2012, Clouston and Denier, 2017); doing so can lead to biased estimates of the retirement-health association.

In addition, it is worth considering the issue of endogeneity surrounding reverse causality as more than a statistical problem to be controlled in research linking retirement transitions and subsequent health, as its substantive implications have often been overlooked in earlier research. That is, although poor health is widely acknowledged as one of the most important antecedents of retirement (Fisher et al., 2016), potential health outcomes of poor-health-driven retirement have received little scholarly interest to date. The few studies available in the literature showed equivocal findings; one earlier study found that those who were forced to retire due to health problems showed poor retirement adjustment and satisfaction outcomes (van Solinge and Henkens, 2008), while another study reported that retirees who were in poor health while working showed steeper health improvements in retirement (Westerlund et al., 2009).

This Study

Taking advantage of rich longitudinal data from the HRS, this study employed a unique analytical approach to address the issue of endogeneity in estimating the relationship between retirement transitions and health. To account for the bidirectional relationship between retirement and health, retiree’s self-reports of the reasons for retirement were used to identify and parse out retirement transitions driven by poor health from the retirement transitions that were unrelated to health reasons. Further, the potential bias caused by omitted variables were addressed by using within-between random effects (WBRE) models with a robust set of time-varying covariates. Apart from addressing the sources of endogeneity bias, the analytical approaches taken in this study provided an opportunity to shed light on aspects of the retirement-health nexus often overlooked in earlier research. First, disaggregating retirement transitions based on retirement reasons allowed for examining the potential health consequences of withdrawing from the labor force due to poor health. Although the literature suggested that being in the labor force is often more valuable to individuals with health problems and disability compared to their healthy counterparts (Milner et al., 2014, Saunders and Nedelec, 2014), studies that specifically examined the health outcomes associated with retiring due to poor health are rare. In addition, the use of WBRE models helped to elucidate health differences observed among individuals who show different labor force behaviors over time.

Given the mixed evidence and the lack of clear directionality regarding the effects of retirement on subsequent health, no formal hypotheses were provided for the research questions regarding the linkages between retirement transitions and subsequent health. However, based on earlier research showing health benefits associated with engaging in an intermediate labor force activity prior to complete retirement, (Dingemans and Henkens, 2014, Wang, 2007), it was hypothesized that health consequences of retirement would be more detrimental when individuals transitioned to complete retirement from full-time work compared to when they made the transition from part-time work or partial retirement status.

Methods

Data and Study Sample

This study was based on the Health and Retirement Study (HRS), a nationally representative panel study of Americans 51 and older (Sonnega et al., 2014). Data for this study were primarily taken from public-use RAND HRS Longitudinal File 2016 (RAND Center for the Study of Aging, 2020), which is a cleaned and processed version of the raw HRS data that accounted for missingness and inconsistent information across waves. Variables not part of the RAND file (e.g., reason for withdrawal from labor force) were taken from the public-use raw HRS data.

This study employed ten waves of biennial data from 1998 to 2016, as information regarding several key variables was not available prior to the 1998 wave. Due to the use of transition indicators based on labor force status assessed over two consecutive waves (i.e., previous and current; see below), the initial wave at which respondents’ labor force status was observed was excluded from the person-wave observations analyzed in the multilevel models. As all cohorts of participants who were added to the HRS during the observation period were included in the study sample, “baseline” in this study refers to the wave immediately following the initial wave when the first labor force status was observed.

The study sample was restricted to cohort-eligible HRS participants 51 or older at baseline who were 1) in the labor force working in any capacity at the initial wave and 2) remained in the study for at least one subsequent follow-up interview as a non-proxy respondent either as a worker (in any capacity) or as a retiree; 3) participants were excluded from the observation following the wave at which they were identified as being unemployed, disabled, or not in the labor force (e.g., homemaker). These criteria yielded 9,351 participants. Participants were further excluded from the sample if they had missing information for the background characteristics (e.g., race/ethnicity) and/or during waves when they had missing information on any of the wave-specific study variables (equivalent to less than 0.1% of the person-wave observations; an exception was reason for withdrawal from the labor force, see below). The final study sample included 9,347 respondents who provided 52,658 person-wave observations.

Measures

Health.

Self-rated health and depressive symptoms, assessed at each wave, were employed in this study to capture the potential improvements and deteriorations in health related to retirement transitions. Self-rated health was based on the question, “Would you say your health is excellent, very good, good, fair, or poor?” and the responses were recorded on a five-point scale, which was recoded so that higher scores indicate worse health (1 = excellent, …, 5 = poor). Depressive symptoms were assessed using the eight-item version of the Center for Epidemiologic Studies Depression (CES-D) scale. This version included items asking whether the following statements had been true for respondents much of the time during the past week: 1) was depressed, 2) everything was an effort, 3) sleep was restless, 4) felt lonely, 5) felt sad, 6) could not get going, 7) was happy, and 8) enjoyed life (1 = yes). After reverse-coding two positive items (7, 8), affirmative responses for the eight items were summed, with higher scores indicating more symptoms of depression (range: 0–8; Cronbach’s α = .76–.79 across waves). The two health outcomes were treated as continuous measures in the analyses; findings were consistent when the measures were treated as ordinal (for self-rated health) and count (for depressive symptoms; Supplementary Table S1).

Labor force status and transition.

Respondent’s labor force and retirement status at each wave were assessed using information from subjective retirement status and labor force status. Respondents were categorized as retired if they were not working and self-identified as completely retired. Respondents engaging in any form of paid employment were categorized as working. Then, to assess the health effects of labor force-retirement transitions, a set of transition indicators were generated to capture how individuals switched from one labor force-retirement category to another over two consecutive waves. To this end, a one-wave lagged version of labor force status was used, where the lagged (t-1) and current (t) status variables were jointly considered to create the following set of binary indicators a) continued working; b) transitioned to retirement; c) continued retirement, and d) unretirement (see Supplementary Figure S1).

To address the study objective of differentiating health consequences of retirements by whether the transition was driven by poor health, the transitioned to retirement category was decomposed using self-reported information on the reported reason for labor force withdrawal (for a list of HRS measures used, see Supplementary material), where the response options included poor health. Based on this information, retirement transitions were categorized as driven by non-health reasons if the retiree indicated that poor health was not a reason or not an important reason at all for the labor force withdrawal; if poor health served as a reason for the withdrawal to any degree, the retiree was deemed to have retired due to poor health. About 2.5% of the retirement transitions that occurred during the observation period had missing information on the reason for labor force withdrawal. To ensure that the transitions driven by non-health reasons did not include any cases that could potentially be driven by poor health, the transitions with missing information were considered to have been driven by poor health; sensitivity analyses excluding these cases yielded consistent results (Supplementary Table S2). Taken together, participants’ labor force transitions during the observation were captured with the five transition indicators (see Supplementary Figure S1).

Apart from these transition indicators, labor force status assessed at the previous wave (t-1) was used to test the moderating role of retirement patterns. The previous wave labor force status included 1) working full-time (working 35+ hours/week, 36+ weeks/year), 2) working part-time (working less than 35 hours/week or less than 36 weeks/year) without mention of retirement, and 3) partly retired (i.e., working part-time and self-identified as retired).

Controls.

Control variables included both time-invariant covariates (TIC) and time-varying covariates (TVC) that could confound the relationship between retirement transitions and health. TICs included age at baseline, gender (1 = female; 0 = male), race/ethnic status (non-Hispanic white (reference), non-Hispanic black, non-Hispanic other race, and Hispanic), education in years, and occupation type (professional/managerial (reference), clerical/sales/service, blue-collar). TVCs included marital status (1 = married/coupled; 0 = not), household income transformed by the natural log, household wealth transformed by the inverse hyperbolic sine function (Friedline et al., 2015), health insurance coverage (1 = insured; 0 = not), and household size (i.e., number of people in the household).

Analytic Plan

A series of within-between random effects (REWB) models were estimated to address the research questions. REWB models produce estimates that are at least as unbiased as those estimates from the fixed-effects models, which are independent of selection effects attributed to all stable inter-individual differences, both observed and unobserved (Bell and Jones, 2015, Mundlak, 1978). This is achieved by decomposing effect of a time-varying predictor into two separate components, where between-person (BP; level 2) component is estimated with the person-mean for each time-varying predictor and within-person (WP; level 1) component is the deviation from this person-mean at a given occasion (Bell and Jones, 2015). In the context of this study, the BP component compared the health outcomes of one person who continued working over the course of the study to another person who transitioned to retirement due to a non-health reason or due to poor health during the study period. In contrast, the WP component compared the health outcomes of a person who continued working at one time to the same person at a different time when the person transitioned to retirement.

For each health outcome, the following four models were estimated. First, an unadjusted model (Model 1) was estimated, where WP and BP effects of retirement/labor force transitions and time parameters for modeling health trajectories were specified in the model. Specifically, a parameterization approach that captured longitudinal changes in health (i.e., individual change, assessed with time from baseline) while accounting for cross-sectional health differences between those at different ages (assessed with age at baseline) was used; an interaction term between age and time was further added to capture potential age-differences in longitudinal changes in health (Morrell et al., 2009). Random effects were specified for time to partition variance of the longitudinal health outcome into WP and BP components and unstructured covariance structure was used to account for within-person dependency in observations. In Model 2 (i.e., adjusted model), individual- and couple-level covariates measured at baseline and over time were added (Model 2; see Supplementary materials for the multilevel equations). The first two research questions on health consequences of retirement transitions driven by 1) non-health reasons and 2) poor health were addressed by examining WP estimates of the respective transition indicators.

In Model 3, previous labor force status was added to Model 2. The research question regarding the moderating role of retirement patterns was addressed in Model 4 by introducing interaction terms for the retirement transition indicators and labor force status assessed at previous wave. The interaction terms were also decomposed into WP and BP components so that interaction effects were not subject to bias caused by stable omitted characteristics (Schunck, 2013). All analyses were performed using the MIXED function in Stata (Version 16).

Results

Study sample characteristics are presented in Table 1. At baseline, study participants, on average, reported their health to be somewhere between very good and good, and reported a little more than one depressive symptom. At the initial observation, most participants were in the labor force as full-time workers (71.6%), followed by those in partial retirement (15.2%) and those engaged in part-time work (13.2%).

Table 1.

Study Sample Characteristics at Baseline

	Mean	SD
Self-rated ill healtha	2.54	1.01
Depressive symptomsb	1.25	1.75
Labor force statusc, %
Full-time work	71.55
Part-time work	13.23
Partial retirement	15.21
Labor force/retirement transitiond, %
Continued working	56.64
Retirement due to non-health reasons	7.94
Retirement due to poor health	4.41
Continued retirement	28.76
Unretirement	2.25
Age (51–92)	60.40	6.33
Female, %	49.74
Race/ethnicity, %
Non-Hispanic white	72.56
Non-Hispanic black	15.61
Non-Hispanic other	2.58
Hispanic	9.25
Education (in years, 0–17+)	13.10	2.98
Occupation, %
Professional/managerial	33.96
Clerical/sales/service	40.09
Blue-collar	25.95
Married/coupled, %	72.74
Household income (log-transformed)	3.88	1.31
Median income in $1,000	54.14
Household wealth (IHS-transformed)	5.08	2.43
Median wealth in $1,000	134.00
Insurede, %	78.93
Household size	2.38	1.24
Number of observations in the study	5.63	2.81

Notes. Sample N = 9,347. IHS = inverse hyperbole sine.

^aRanges from 1 ( = excellent) to 5 ( = poor).

^bEight-item Center for Epidemiologic Studies Depression scale.

^cAssessed at initial wave when first labor force status was observed.

^dBased on 52,658 person-wave observations recorded during study period.

^eInsurance coverage through Medicare/Medicaid, either spouse’s employer (current or past), or any other supplemental insurance (1 = insured; 0 = not insured).

During the 18-year observation period, approximately 62% of the participants (n = 5,766) experienced a transition to full retirement (not shown in table). Further, 1,071 participants came out of retirement (i.e., labor force re-entry), 737 of whom made a second transition to full retirement (i.e., re-retirement). While the main analyses included all retirement transitions observed during the study period, supplementary analyses including only the first retirement transitions yielded consistent findings (Supplementary Table S3). Taken together, a total of 6,503 retirement transitions were observed during the study period, where about two-thirds of the transitions were not driven by poor health (n = 4,181) and a third driven were by poor health (n = 2,322).

Multilevel Model Results

Results from the multilevel REWB models are presented in Table 2 (Model A: self-rated health; Model B: depressive symptoms). The pattern of findings was similar in both the unadjusted and adjusted models for the two health outcomes; as such, the adjusted estimates are discussed below. For both health outcomes, it is worth highlighting first the significant and sizable BP differences that were observed across participants who showed different labor force behaviors during the study period. In Model A2, participants who retired due to non-health reasons during the observation period had better overall self-rated health compared to those who remained in the labor force (b = −0.447, p < .001), whereas those who retired due to poor health had worse self-rated health (b = 1.46, p < .001). Similarly, in Model B2, participants who retired due to non-health reasons reported fewer depressive symptoms (b = −0.382, p < .001), whereas those who retired due to poor health reported more depressive symptoms (b = 1.790, p < .001), compared to their counterparts who remained in the labor force during the observation period.

Table 2.

Multilevel Models for Retirement Transitions and Health

	Panel A: Self-rated ill healtha				Panel B: Depressive symptomsb
	Model A1: Unadjusted model		Model A2: Adjusted model		Model B1: Unadjusted model		Model B2: Adjusted model
	B	(SE)	B	(SE)	B	(SE)	B	(SE)
Fixed effects
Within-person (level-1)
Labor force transitionc
Retired: non-health reasons	−0. 011	(0.012)	−0.013	(0.012)	0.013	(0.024)	0.015	(0.024)
Retired: due to poor health	0.222***	(0.016)	0.219***	(0.016)	0.277***	(0.031)	0.278***	(0.031)
Continued retirement	0.032**	(0.012)	0.029*	(0.012)	0.099***	(0.024)	0.097***	(0.024)
Unretirement	−0.048*	(0.020)	−0.049*	(0.020)	−0.072	(0.040)	−0.073	(0.040)
Time	0.071***	(0.002)	0.073***	(0.002)	−0.004	(0.004)	−0.011**	(0.004)
Coupled			0.054***	(0.016)			−0.496***	(0.032)
Household income			−0.004	(0.003)			−0.005	(0.007)
Household wealth			−0.006**	(0.002)			−0.024***	(0.004)
Insured			−0.011	(0.012)			−0.009	(0.025)
Household size			0.011*	(0.005)			0.026**	(0.009)
Between-person (level-2)
Labor force transitionsc
Retired: non-health reasons	−0. 508***	(0.072)	−0.447***	(0.068)	−0.504***	(0.123)	−0.382**	(0.119)
Retired: due to poor health	1.787***	(0.069)	1.460***	(0.067)	2.202***	(0.119)	1.790***	(0.117)
Continued retirement	0.563***	(0.033)	0.451***	(0.032)	0.520***	(0.055)	0.385***	(0.053)
Unretirement	0.215	(0.150)	0.106	(0.140)	−0.361	(0.248)	−0.412	(0.236)
Baseline age	0.003	(0.003)	0.005	(0.003)	−0.021***	(0.005)	−0.013**	(0.005)
Baseline age × time	0.002***	(0.001)	0.002***	(0.001)	0.005***	(0.001)	0.005***	(0.001)
Married/coupled			−0.002	(0.023)			−0.354***	(0.038)
Household income			−0.074***	(0.010)			−0.068***	(0.017)
Household wealth			−0.059***	(0.004)			−0.082***	(0.007)
Insured			0.011	(0.031)			−0.091	(0.052)
Household size			0.026**	(0.009)			0.033*	(0.015)
Female			−0.062***	(0.018)			0.172***	(0.030)
Race/ethnicityd
Non-Hispanic black			0.086***	(0.023)			−0.102**	(0.038)
Non-Hispanic other			0.247***	(0.049)			0.270**	(0.083)
Hispanic			0.083**	(0.030)			0.053	(0.051)
Education			−0.045***	(0.003)			−0.046***	(0.006)
Occupation typee
Clerical/sales/service			0.039*	(0.020)			0.055	(0.033)
Blue-collar			0.016	(0.023)			0.029	(0.039)
Intercept	2.674***	(0.020)	2.584***	(0.043)	1.363***	(0.034)	1.485***	(0.073)
Random effects
Time variance	0.008***	(0.001)	0.008***	(0.001)	0.019***	(0.001)	0.020***	(0.001)
Intercept variance	0.584***	(0.011)	0.499***	(0.010)	1.491***	(0.033)	1.320***	(0.030)
Corr (Time, Intercept)	−0.344***	(0.021)	−0.346***	(0.021)	−0.351***	(0.025)	−0.357***	(0.026)
Residual variance	0.342***	(0.003)	0.342***	(0.003)	1.388***	(0.010)	1.378***	(0.010)
Model Fit
AIC	116,887.689		115,635.32		185,801.348		184,606.543
−2 log-likelihood	116,851.688		115,565.32		185,765.348		184,536.542

Notes. Sample N = 9,347 (52,658 person-wave observations). All level-2 continuous variables were grand-mean centered.

^aRanges from 1 ( = excellent) to 5 ( = poor).

^bEight-item Center for Epidemiologic Studies Depression scale.

^cReference category: continued working.

^dReference category: non-Hispanic white.

^eReference category: Professional/managerial.

^*p < .05,

^**p < .01,

^***p < .001.

The research questions addressing the linkages between retirement transitions (driven by non-health reasons and poor health) and subsequent health were addressed with the WP estimates of the respective transition indicators. Adjusting for the BP differences, retirement transitions were associated with differential changes in subsequent health depending on whether the transition was driven by poor health. In Model A2, participants did not show significant changes in subsequent self-rated health when they transitioned to retirement due to non-health reasons (b = −0.013, p = .297) compared to when they continued to work, whereas worse self-rated health was observed when they withdrew from the labor force due to poor health (b = 0.219, p < .001). Similar results were found with regards to depressive symptoms; within the same person, retirement transitions driven by poor health were associated with an increased level of depressive symptoms compared to when the person continued to work (b = 0.278, p < .001), but such change was not observed when the retirement transition was unrelated to health reasons (b = 0.015, p = .540). As noted earlier, sensitivity analyses that took into consideration the functional form of the outcome measures, missing information for the reason for retirement, and repeated retirement transitions among some participants yielded consistent results (see Supplementary Tables S1–S3).

The Role of Retirement Patterns

In Models A3 and B3 (see Table 3), the indicators for previous labor force status were added to the adjusted models for self-rated health (A2) and depressive symptoms (B2), respectively. The estimates for retirement transition indicators remained largely unchanged in these models. The within-person estimates for previous labor force status were indicative of better self-rated health and fewer depressive symptoms reported when participants were in partial retirement in the previous wave compared to when they had worked full-time. Additionally, participants also reported fewer depressive symptoms (but no differences in self-rated health) when they had worked part-time in the previous wave relative working full-time.

Table 3.

Moderation Effects of Retirement Patterns

	Panel A: Self-rated ill health				Panel B: Depressive symptomsb
	Model A3: Main effects		Model A4: Interaction effects		Model B3: Main effects		Model B4: Interaction effects
	b	(SE)	B	(SE)	b	(SE)	B	(SE)
Fixed effects
Within-person (level-1) effects
Labor force transitionc
Retired: non-health reasons	−0.011	(0.012)	−0.037*	(0.016)	0.018	(0.024)	−0.040	(0.033)
Retired: due to poor health	0.220***	(0.016)	0.213***	(0.021)	0.281***	(0.031)	0.261***	(0.042)
Previous Labor force statusd
Part-time work	−0.006	(0.014)	−0.013	(0.015)	−0.073**	(0.027)	−0.083**	(0.030)
× Retired: non-health reasons			0.045	(0.038)			0.126	(0.075)
× Retired: due to poor health			0.015	(0.047)			−0.073	(0.093)
Partial retirement	−0.032**	(0.012)	−0.046***	(0.014)	−0.084***	(0.024)	−0.120***	(0.027)
× Retired: non-health reasons			0.055*	(0.024)			0.121*	(0.048)
× Retired: due to poor health			0.018	(0.033)			0.084	(0.065)
Random effects
Time variance	0.008***	(0.001)	0.008***	(0.001)	0.020***	(0.001)	0.020***	(0.001)
Intercept variance	0.498***	(0.010)	0.497***	(0.010)	1.319***	(0.030)	1.318***	(0.030)
Corr (Time, Intercept)	−0.346***	(0.021)	−0.345***	(0.021)	−0.357***	(0.026)	−0.356***	(0.026)
Residual variance	0.342***	(0.003)	0.342***	(0.003)	1377***	(0.010)	1377***	(0.010)
Model Fit
AIC	115,629.459		115,630.588		184,592.716		184,590.601
−2 log-likelihood	115,551.458		V 115,536.588		184,514.716		184,496.602

Notes. Sample N = 9,347 (52,658 person-wave observations). Estimates are fully adjusted for all study variables shown in Table 2; time-varying covariates (including interaction terms) were decomposed into within-and between-components in the models.

^aRanges from 1 ( = excellent) to 5 ( = poor).

^bEight-item Center for Epidemiologic Studies Depression scale.

^cReference category: continued working; categories for continued retirement and unretirement not shown.

^dReference category: full-time work.

^*p < .05,

^**p < .01,

^***p < .001.

The moderating role of retirement patterns was tested in the subsequent models by introducing the interaction terms between retirement transition and previous labor force indicators. For both health outcomes, significant WP interaction effects were found when partial retirement preceded full retirement driven by non-health reasons, but in the opposite direction of the hypothesis. That is, individuals reported worse self-rated health (b = 0.055, p = 0.024) and more depressive symptoms (b = 0.121, p = 0.012) when they transitioned from partial retirement to full retirement compared to when they transitioned from full-time-work to full retirement.

Discussion

The topic of retirement presents many challenges that make it difficult to study. One is its “famously ambiguous” definitional concept (Ekerdt, 2010), one that is becoming even more complex due to the de-standardization of how individuals withdraw from the labor force (Calvo et al., 2018). Examining health consequences of retirement adds several layers of complexity due to the bi-directional relationship that underlies labor force transitions and health. This study sought to shed light on the complex picture by examining within-person linkages between retirement transitions and subsequent health, while taking into consideration the between-person health differences that likely exist between individuals who show different labor force behaviors. Study findings based on a large sample of national data from the HRS spanning almost two decades provided unique insight into the association between labor force status and transitions and two indicators of health.

The linkages between retirement transitions and subsequent health outcomes as found in this study were substantially different depending on whether the transition was driven by poor health. Specifically, the findings showed that retirement transitions were unrelated to subsequent health if the withdrawal from the labor force was driven by non-health reasons, whereas retirement transitions driven by poor health were associated with worse subsequent health, as assessed with both self-rated health and depressive symptoms. In addition and contrary to the hypothesis, retirement transitions driven by non-health reasons that were “phased” through partial retirement were associated with worse health outcomes compared to when an “abrupt” transition was made from full-time work to full retirement (Calvo et al., 2018). These findings can be understood in the context of the broader literature on labor force behavior and health in later life.

When Retirement is not Driven by Poor Health

By and large, earlier studies were interested in uncovering health consequences of retirements that were unrelated to health, regardless of whether the issue of endogeneity was addressed. On the one hand, the null relationship found in this study parallels a recent review by Henning et al. (2016) as well as a meta-analysis which suggested that retirement is a ‘neutral’ life event with respect to psychological well-being (Luhmann et al., 2012). On the other hand, the finding is at odds with results from a stream of studies that found that retirement leads to better health of the retiree, especially when assessed with subjective health (Nishimura et al., 2018, van den Bogaard et al., 2016).

Several explanations are possible. First, earlier studies defined retirement differently; researchers often define retirement as working less than a threshold number of annual hours or leaving a career job, regardless of actual work status (e.g., 1,200 hours; Coe and Zamarro, 2011). In such cases, retirement transitions may include being in a process of gradual retirement through reducing one’s work hours, and as a consequence, the health benefits of retirement found in these studies could in part be driven by a higher level of job-satisfaction and psychological well-being associated with working less than full-time in later life (Booth and Van Ours, 2008, Nikolova and Graham, 2014). Consistent with this view, an Australian study employing a within-person analytic approach reported that transitioning from a stable employment to a more “casual” form of employment (with no paid leave entitlements or fixed hours) resulted in improvements in mental health among workers aged 55–64 years (LaMontagne et al., 2014).

Second, earlier studies may not have adequately accounted for person-level characteristics when estimating the health effects of retirement transitions, which may have biased the findings. Unlike earlier studies, the BP health differences underlying labor-force behavior were explicitly accounted for in this study (see also Clouston and Denier, 2017). As seen in Table 2, individuals who showed different labor force-retirement transition behaviors were observed to have substantially different underlying health characteristics, such that those who retired due to non-health reasons during the course of the observation period had better overall health compared to their counterparts who continued working.

Third, many earlier studies (including those based on the HRS) that found positive health consequences of retirement were based on IV approaches in attempts to overcome the issue of endogeneity. Yet, pension eligibility ages used as the instrument may not be completely exogenous to health, for example, in such circumstances where people who retired due to poor health waited until they reached the pension age and then retired. This may be the case for some participants in this study sample; when age at retirement was plotted for the cases with information on the timing of retirement, the age-distribution of poor-health-driven retirements had a spike at 62 (i.e., Social Security eligibility age for early retirement in the U.S.), similar to spikes seen at ages 62 and 65 for non-health-related retirements (Supplementary Figures S2; also, see below for discussion on Figure S3). These ages also coincide with perceived milestone ages in the U.S. (Toothman and Barrett, 2011), in which case reaching such a milestone may also have non-negligible direct psychological and health effects. Finally, it is important to note that the results based on IV approaches estimate local average treatment effects that are only generalizable to individuals whose labor force withdrawal is driven by reaching pension eligibility ages. As such, the positive health consequences of retirement often observed in IV studies may partly be driven by an on-time life-course transition that can benefit well-being (e.g., Umberson et al., 2010). As a partial test of this explanation, indicators denoting the first receipt of Social Security benefit upon reaching early retirement age (i.e., 62; Supplementary Table S4) and full retirement ages (i.e., 65–66 based on birth cohort; Supplementary Table S5) were created and added to the analytic models. Adding these indicators did not alter the substantive findings reported in this study; yet, the findings from the analyses indicated that the receipt of first retirement benefit upon reaching pension eligibility ages (i.e., 62) had a direct salubrious effect on depressive symptoms above and beyond the effects of retirement transitions, whereas it moderated the health consequences of non-health-driven retirement transition on self-rate health in a manner that benefitted health in retirement (Supplementary Tables S4, S5).

Poor-health Driven Retirement Transitions

In contrast, this study found that individuals showed deleterious health outcomes after they retired due to health problems, even after adjusting for the sizable between-person health-differences (i.e., the significantly worse self-rated health observed for individuals who retired due to poor health during the study period compared to those who continued working; see Table 2). Although the within-person estimates for poor-health-driven retirements were robust and sizable in the direction indicating worse health, it is important to note that the findings pertaining to retirement transitions driven by poor health are subject to a couple of alternative interpretations. First, the findings may indicate that individuals who completely withdrew from the labor force due to poor health experienced further health deterioration following retirement. Labor force participation serves as an important source of social integration, which provides a number of psychosocial resources (e.g., role identity, sense of community) essential to health in later life (Moen et al., 2000). Research indicated that the loss of these resources was not easily replaced in retirement when one retires due to health reasons, suggesting that the meaning of work may be especially greater for older adults in compromised health conditions (Moen et al., 2000, Milner et al., 2014). A recent HRS-based study examining the link between retirement timing and mortality supported this view, where prolonged working life was associated with survival benefits for unhealthy retirees, defined as individuals who self-reported that health was an important reason for retirement (Wu et al., 2016). Additional work is needed to explicate the potential health benefits of working in compromised health.

However, it is also possible that the worse health outcomes observed in the wave following health-related retirement transitions were in part capturing effects of health conditions that precipitated the labor force withdrawal in the first place, given that unexpected and severe health events (e.g., cardiovascular disease; CVD) may trigger retirements in later life. While this possibility cannot be completely ruled out under the current study design, supplementary analyses were conducted to take into consideration the potential confounding effects of incident CVD events (assessed as heart attack and stroke), which could have influenced both the retirement transition as well as the health outcomes. Supplementary models adjusting for the significant effects of incident CVD found results consistent with the main findings, and the health consequences of retirement transitions did not vary depending on whether respondents reported an incident CVD around the time of retirement transitions (Supplementary Table S6; findings were consistent when other health conditions, such as cancer, diabetes, and arthritis were additionally considered). Key findings were also substantively consistent when respondents’ self-reported health changes since the previous wave was employed as a potential confounder for the relationship between retirement transitions and subsequent health (Supplementary Table S7). It is worth noting that the majority (88%) of the health-related retirement transitions did not co-occur with an incident CVD event within the same two-year interval, and that these retirement transitions occurred most frequently at age 62 (Supplementary Figure S3), suggesting to a degree that individuals with chronic health problems were adjusting and postponing retirement until when they were eligible for early retirement.

Limitations

First, health outcome measures used in this study have shortcomings. Depressive symptoms are different from clinically diagnosed depression. Self-rated health, despite its widely-acknowledged utility for assessing global health and mortality risk, only provided a general measure of health and did not provide any insight into how retirement may be related to a specific health outcome. Employing objective markers of health (e.g., anthropometric measures and other biomarkers) that capture the continuum of health (and not merely absence of disease) would be a fruitful area of future research. Second, although this study employed methodological approaches to minimize endogeneity bias, it is not possible to completely rule out bias caused by response error and time-varying omitted variables unaccounted for in the analyses. Nor is it possible to fully decompose the bi-directional associations between retirement and health, especially as it pertains to poor-health-driven retirement transitions. Employing fine-grained health data with shorter intervals, possibly from administrative data sources, may represent a promising approach to study health-related retirement transitions. Finally, findings from this study have limited generalizability outside of the U.S., as institutional and cultural differences across countries play an important role in the retirement-health associations (Nishimura et al., 2018).

Conclusions

This study contributed to the literature in several ways. Using longitudinal HRS data, the current study examined health consequences of retirement transitions while employing a novel approach to address the issues of endogeneity. Importantly, the use of self-reported information on retirement reasons to estimate health outcomes of non-health-related retirement transitions separately from those driven by poor health is one of the key contributions. Unlike methodological approaches taken in earlier research, use of self-reported reason for retirement allows for addressing the potential reverse-causality in estimating the retirement-health nexus without relying on untestable assumptions. Exploring health outcomes associated with poor-health-driven retirement transitions, an understudied yet important research area, while addressing potential confounding by employing rich information available from the HRS, is another contribution. Given the considerable health disadvantage observed among individuals who made a health-related labor force withdrawal, identifying health-promoting factors for older individuals working in compromised health, as well as factors promoting retirement adjustment and quality of life for those individuals who are no longer able to work due to poor health, represent fruitful areas of future research.

The current study found no evidence for health benefits of retirement. Rather, the overall findings pertaining to labor force transitions and health provide suggestive evidence of the opposite, where retirement transitions and continued retirement were either unrelated to health or associated with worse subsequent health. In contrast, continued working, and to a certain degree, unretirement, appeared to be associated with improved health outcomes. Taken together, these findings suggest that retirement policies designed to prolong working lives may be implemented without undermining retiree’s health, and potentially result in delaying negative health consequences of retirement for some segments of the population. In turn, it may mean that fiscal concerns regarding public pensions, as well as the healthcare systems, could be alleviated with an increase in formal retirement ages. Fostering an age-friendly work environment and providing flexible work options, as well as designing jobs more suitable for older adults, including those in compromised health, may help to ensure overall success of retirement policies.

Research Highlights.

• Assessed linkages between retirement transitions and subsequent health.

• Retirements driven by non-health reasons were unrelated to subsequent health outcomes.

• Retiring due to poor health was associated with worse subsequent health.

• Findings support retirement policies designed to prolong working lives.