Costs of forced retirement: Measuring the effect of lost work opportunity on health

Abstract

Objective:

Unemployment is a known health stressor that also increases early retirements. This study addresses mixed literature on retiree health and under-reporting of forced retirement to better identify potential health impacts of lost work opportunity.

Methods:

A Lost-work Opportunity Score (LOS) was created using variables from the Health and Retirement Study assessing unemployment, forced retirement, and earlier-than-planned retirement for 2,576 respondents. Reliability and unidimensionality of the score with multivariate regression analyses examined health impacts controlling for demographics and prior health status.

Results:

The LOS possessed unidimensionality with a Cronbach’s Alpha of a=0.76 while predicting self-reported health declines (LOS=2; β =.381, OR=1.464, p<.05) and depression increase (LOS=2; β =.417, OR=1.517, p<.05).

Conclusions:

LOS predicts 46% increased odds of negative self-reported health change after retirement associated with 2 LOS events, with implications to support aging workers.

Introduction

Clear pathways to retirement are shifting, with decreasing protections for workers, fewer retirement pensions, and job churning that result in precarious employment (1, 2). It is estimated that a high percentage of retired workers will also engage in some form of paid employment after official retirement (3), further complicating matters. With the increasing complexity of labor force transitions of older workers (4, 5), it has been difficult to unravel the impact of partial retirements and post-retirement work on health, finances, identity, or adjustment (6). But a look at a related domain of unemployment shows clear evidence of higher rates of overall mortality as well as morbidity across physical and mental health domains that cannot be attributed simply to the loss of income that occurs when individuals stop working involuntarily (7–13). Whether individuals have access to work opportunity at retirement age may be similarly relevant to retiree health.

Given the complexity of retirement as a life stage, researchers are beginning to recognize the importance and challenge of accurately measuring the construct (6, 14). This challenge is even more pronounced when trying to assess involuntary retirements. Over half of early female retirees (aged 55–64) in one analysis had chosen retirement after a documented job loss, deteriorated working conditions, reduced earnings, or an episode of poor health (15). Survey items that ask if retirement was wanted or not are unable to detect these nuanced retirement choices. Mixed-methods comparisons of survey to interviews data found that 17% of individuals who marked retirement as voluntary revealed multiple elements of forced retirement timing in interviews (i.e., retirement prompted by unemployment, temporary layoffs, company buy-outs, workforce reductions, or forced relocations)(16, 17). A deep dive into the Health and Retirement Study data on earlier than planned retirements found 27% of early retirements were still unexplained after adjusting for multiple factors like health change, family situations, and caregiving (18). Measuring this phenomenon is difficult, particularly in cases where retirement is the best choice between bad options.

Yet even as measured with existing imperfect survey tools, involuntary retirements in the US appear to be increasing (19, 20). When older workers are forced into retirement they lose tangible resources like the income they receive from work (21, 22), and also the intangible foundational identity, sense of control, and independence in choosing their retirement path. It is clear that work opportunity overlaps with retirement decisions. Higher unemployment rates are associated with earlier retirements in the US (23, 24). When jobs are difficult to come by, choosing retirement may offer more short-term stability and be less stigmatizing than unemployment (25). Retirement removes workforce re-entry obstacles that are greater for older workers, such as a lower replacement wage, lack of updated skills, or age discrimination (26). Whatever the reason, the numbers suggest that many older adults opt for retirement when a job loss occurs (17, 23), indicating that economic job availability factors influence choices about retirement timing.

The principle of autonomy has been a key domain in quality of life research (27). Though a full discussion on the substrates of autonomy as a health-influencing variable are beyond the scope of this article, choice remains an important consideration in the retirement matrix. Better measurement of this choice metric might help address the debate over the health impact of retirement. Decades of population-level data show that retirement is associated with physical health deterioration in the general population (28–31). Yet planned, voluntary, and pensioned early retirement can result in health gains compared to continued employment (32–34). Addressing the apparent contradictory outcomes, voluntariness of retirement may be key in understanding it’s health impact (35–37). As external factors such as an estimated post-COVID-19 surge in 2.4 million additional retirements (38) continue to influence retirement timing, it is more important than ever to understand the role of worker choice. National survey tools assessing health and retirement ask about forced retirement decisions, but do not adequately capture when older workers are nudged out of the labor force (16). Unraveling the potential health effects for this population begins with better measurement.

In this article, the utility of a broadened construct of unemployment occurring at retirement age is explored which recognizes the complexities of work-access factors prompting aging workers to leave the workforce. “Work opportunity,” or rather the loss of it, is highlighted over forced retirement as the employment construct more accurately reflecting the complex transitions in work roles that occur with aging. The present analysis examines a Lost Work Opportunity Score (LOS) incorporating unemployment, forced retirement, and unplanned early retirement for predictive validity to examine whether lost work opportunity has measurable impacts on retiree health. A higher score in LOS is hypothesized to result in health declines, consistent with research linking lack of autonomy and involuntary retirement to negative health change.

Methods

Data

The current study used the Health and Retirement Study (HRS) data collected by the National Institutes of Health Institute on Aging. The HRS is a longitudinal survey administered biennially to adults aged 50 years and older residing in the US and includes many factors relevant to employment and retirement decisions (39)(40). It has multiple data files available for public use from three decades of survey administration. This study utilized HRS data files from the waves 2004 to 2016 and from the RAND Corporation multi-wave longitudinal files (41) with 2,459 HRS participants providing necessary information for this study. Using the 2004 survey wave, participants were included if they were aged 55 or greater and participated in the labor force either as employed or looking for work in that year regardless of age, excluding only those already outside the labor force on who were on sick leave, disabled, or retired. Designed as a secondary analysis of a national dataset, the project received a non-human subjects research determination during ethics review. The HRS survey administration for LOS relevant items used skip-patterns that preclude answering some items (e.g., a reported labor force status of “retirement” skips the question for planned age of retirement). The variable with the most limited data availability was a question categorizing retirement as voluntary or partially forced, with a range of 128 to 807 valid responses in any given survey wave.

LOS Measure

LOS was created by incorporating three variables - unemployment, involuntary retirement, and earlier than planned retirement (16). The LOS score for each survey wave was created by summing up the scores of the three variables described below. The LOS score ranged from 0 to 3 for any given wave from 2004 to 2016.

Unemployment status was drawn from the survey question, “Are you working now, temporarily laid off, unemployed and looking for work, disabled and unable to work, retired, a homemaker, or what?” Any wave in which the answer was “laid off” or “unemployed” for the labor force status was coded as a +1 for LOS using the prior survey wave data, as an indicator of recent unemployment and allowing for additive lost work opportunity if there was a report of involuntary retirement in the current survey wave; otherwise coded as 0.

Involuntary retirement was assessed by the question, “Thinking back to the time you (partly/completely) retired, was that something you wanted to do or something you felt you were forced into?” Either an answer of “forced” or “part wanted/part forced” was coded as a +1 for LOS for that survey wave; otherwise coded as 0.

Planned age of retirement was based on the question “At what age do you plan to stop working?” in the prior survey wave. The actual retirement age was calculated as the year when individuals reported retirement as the labor force status in any survey wave minus the individual’s birth year. When the actual retirement age minus the planned retirement age yielded a negative number (which was constituted as earlier than planned retirement), a +1 for LOS was recorded for that survey wave; otherwise coded as 0.

The highest LOS score among all LOS scores from waves 2004 to 2016 was defined as the total LOS. The total LOS score could be 0, 1, 2 or 3.

Outcome Measures and Covariates

The outcome measures in this study included self-reported health status using a single question, “Would you say your health is excellent, very good, good, fair, or poor?” Self-rated health was dichotomized as good to excellent or poor to fair health (42). A higher score indicated poorer health. A second measure used the HRS modified version of the Center for Epidemiological Studies-Depression (CES-D) scale score. Respondents were asked, “Now think about the past week and the feelings you have experienced. Please tell me if each of the following was true for you much of the time during the past week.” Prompts included nine items: 1) felt depressed, 2) felt that everything you did was an effort, 3) sleep was restless, 4) were happy, 5) felt lonely, 6) enjoyed life, 7) felt sad, 8) could not get going, and 9) had a lot of energy, with items 4, 6 and 9 reverse-coded. Given the skew of this scale toward low values, scores were dichotomized into low depressive symptoms (0–2) and high depressive symptoms (3–9) based on a score of 3+ as a clinical indicator of depression (43). A higher score indicated greater depression as a negative indicator.

Covariates included race (White, Black/African American, or Other), ethnicity (Hispanic or Not Hispanic), gender, age, education coded from 1–5 (less than high school = 1, GED = 2, high school graduate = 3, some college = 4, and college graduate and above = 5), and income. Income was defined as the total household income in the final survey wave. These covariates were drawn from the literature as relevant to health status outcomes (44).

LOS Validity Analysis

Prior validity analysis related to the LOS construct provide support for convergent and discriminant validity (16). In the current study, additional validity analysis was performed on the three variables (i.e., unemployment, involuntary retirement, and early retirement) that constructed the LOS score, the covariates, and the two outcome measures, examining means and standard deviations for continuous variables and percentages for categorical variables, and checking if the data were sufficiently normally distributed. There were no violations of the normality of the error terms found when analyzed with both the quantile plots and the Shapiro-Wilk’s test for normality (results available upon request). Principal component factor analysis was conducted to identify factor loadings of the three variables that constituted the LOS score.

LOS Outcome Analysis

Goodness of fit and ordinal multiple regression were conducted to examine the effect of LOS with health variables. Self-rated health as reported in 2004 was the baseline health measure and was included in the final regression model along with the sociodemographic factors described above with outcome measures from 2016 data for self-reported health and CES-D scores. Variable correlations are available upon request; covariates entered in the final model were not highly correlated (<.30). In the regression analysis, age and income were treated as continuous variables, and gender, race, ethnicity, education, and health outcomes were treated as categorical variables.

Generally, the following model specifications were used:

$${\mathit{\text{Health}}}_{\mathit{\text{post}}}={\beta }_{\mathit{0}}+{\beta }_{\mathit{1}}\mathit{\text{LOS}}+{\beta }_2{\mathit{\text{Health}}}_{\mathit{\text{pre}}}+\sum {\beta }_{j}Xj+\epsilon$$

Health_post is the respondent’s self-reported health level from the 2016 year-end survey period. The factor X includes other potential determinants of Health_post that may correlate with the LOS variable (e.g., sociodemographic and economic factors), and ε is the error term. The consistency and efficiency of β₁ can be improved by including ex-ante health status at a baseline timepoint when using longitudinal datasets (45). Hence, we further estimated by including β₃ as the variable Health_pre where Health_pre is the measure of the subject’s self-reported health status in the first survey period.

Three logistic regression models were run for each of the two outcome variables (health and depression). For dichotomized health, model 1 included the 2016 wave’s self-reported health as the outcome, and the total LOS as the predictor; model 2 included the same variables as in model 1 with the addition of age, gender, race, ethnicity, income, and education level as covariates; model 3 included all of the variables in model 2 with the addition of the 2004 wave’s self-reported health as covariate. For CES-D as the dichotomized outcome, its three logistic regression models were identical to those regression models for the self-reported health except for the outcome variable being the 2016 wave’s CES-D measured mental health.

Results

The unweighted sample of HRS participants included in the final analysis was 54.7% female (N = 1,344) (see Table 1). The median year of birth of participants was 1943, with ages ranging from 67–96 in 2016. The majority of the sample identified as White (80.8%, N = 1987), 13.8% as Black/African American (N = 339), 5.4% (N = 133) identified race as “Other”; and 9.4% (N = 232) identified their ethnicity as Hispanic. The education mean score was 3.53, which falls into the “Some College” category. Income was calculated as total household reported income and ranged from $0 to $2,146,928, with a median income of $48,000.

Table 1:

Demographics

LOS Scorea	N	% or Median (Range)
0	603	24.5%
1	1235	50.2%
2	521	21.2%
3	100	4.1%
Gender
Male	1115	45.3%
Female	1344	54.7%
Race
White	1987	80.8%
Black/African American	339	13.8%
Other	133	5.4%
Ethnicity
Not Hispanic	2227	90.6%
Hispanic	232	9.4%
Education
Less than high school	310	12.6%
GED	92	3.7%
High school graduate	705	28.7%
Some college	629	25.65
College graduate and above	723	29.4%
Household Income	2459	$48,000 ($0–$2,146,928)
Self-reported Health 2004
Fair to Poor	290	12.5%
Excellent to Good	2032	87.5%
CES-D Mental Health 2004
High Depression (3–9)	320	13.8%
Low Depression (0–2)	2002	86.2%

^aNote: Highest LOS score from any survey wave (2006–2016)

The average self-reported health for the sample was 2.39 in 2004 and 2.77 in 2016, indicating that for the sample as a whole, self-reported health declined slightly in the ten-year period, but remained better than the 2016 mean for the full HRS national sample of 2.95 (see Table 1 for dichotomized health scores). The mean for the sample 2016 CES-D was 1.11 indicating the sample reported less depression than national averages (HRS 2016 national sample = 1.54). No lost work opportunity was reported by 24.5% (N = 603) of the sample, 50.2% (N = 1235) reported one LOS event, 21.2% (N = 521) experienced two LOS events, and 4.1% (N = 100) experienced all three LOS events (see Table 1).

The total LOS score had a Spearman’s rho correlation of r_s = 0.02 with race, r_s = 0.01 with ethnicity, r_s = −0.09 with education, r_s = 0.16 with gender, r_s = −0.15 with age, and r_s = −.021 with income. The low correlations between LOS and demographic variables indicate demographic factors did not have an overly large influence on LOS as a construct, with income, education, and gender more likely to influence LOS scores. The correlation between age (recorded as the birth year so that a lower number is an older age) and planned retirement age never rose above r = −.26 across survey waves, showing a tendency for older individuals to state plans for an older retirement age.

The Spearman’s rho correlations between individual LOS factors (unemployment, involuntary retirement, earlier than planned retirement) and the LOS score ranged from r_s = 0.48 to r_s = 0.62 across the waves of variable construction, indicating related but distinct factors. Principal component factor analysis for the LOS score Eigenvalues suggested unemployment explains 33.3% of the variance, voluntariness of retirement accounts for 35.2%, and earlier than planned retirement for 31.5% of the variance. All variables loaded onto a single component for the LOS scores for years 2006, 2008, 2012, and 2016 (see Table 2). In 2010 and 2014, unemployment loaded more substantially on a second component, indicating a related but distinct construct (eigenvalues values available upon request). Internal consistency of the scale was computed with a Cronbach’s Alpha for the LOS score a = 0.76 and the Guttman Split-Half coefficient as 0.71, indicating adequate levels of consistency and reliability.

Table 2:

Factor Analysis Component Matrix

Survey Year		LOS Factor
		Unemployment	Forced Retirement	Earlier than Planned Retirement
2006		.660	.443	.791
2008		.551	.587	.716
2010	Component 1	−.010	.748	.749
	Component 2	1.000	.019	−.005
2012		.435	.670	.757
2014	Component 1	−.112	.722	.767
	Component 2	.918	.375	−.218
2016		.418	.594	.726

In Model 1, univariate analysis using simple logistic regression demonstrated the LOS score had a direct relationship with worse self-reported health for groups with 1 (β = .252, OR = 1.287, p < .05), 2 (β = .611, OR = 1.842, p < .001), or 3 (β = .882, OR = 2.415, p < .001) LOS events (see Table 3). Model 2 included demographic variables resulting in minimal changes to the estimates on LOS (see Table 3). Model 3 indicated the LOS significantly predicted self-reported health for groups with 2 (β = .381, OR = 1.464, p < .05) or 3 (β = .654, OR = 1.923, p < .01) LOS events relative to no lost work opportunity (see Table 3). For 2 LOS events, this equated to a 46% increased odds of a negative health report relative to no lost work; for 3 LOS events, this equated to an 92% increased odds of a negative health report relative to no lost work.

Table 3:

Multivariate analyses of factors impacting self-reported physical health

	Model 1	Model 2	Model 3
Lost Work Opportunity (Comparison LOS = 0)
LOS=1	0.252(0.127)*	0.141(0.132)	0.119(0.138)
LOS=2	0.611(0.145)***	0.421(0.153)**	0.381(0.160)*
LOS=3	0.882(0.237)***	0.674(0.248)**	0.654(0.259)**
Gender (Comparison = Male)
Female		−0.162(0.105)	−0.206(0.110)
Ethnicity (Comparison = Non-Hispanic)
Hispanic		0.522(0.172)**	0.334(0.182)
Race (Comparison = White)
African American		0.358(0.139)**	0.304(0.146)*
Other Race		0.345(0.214)	0.395(0.223)
Birth Year		−0.015(0.010)	−0.023(0.011)*
Education (Comparison = Less than HS)
GED		−0.313(0.264)	−0.174(0.280)
High School Graduate		−0.554(0.156)***	−0.357(0.167)*
Some College		−0.677(0.163)***	−0.456(0.173)**
College Graduate or above		−1.127(0.179)***	−0.831(0.190)***
Household Income 2016 (per 1000)		−0.001(0.000)***	−0.001(0.000)*
Household Income 2004 (per 1000)			−0.001(0.000)
Self-Reported Health 2004 (Comparison = Good to Excellent)
Fair to Poor Health			1.728(0.135)***

^***< .001,

^**< .01,

^*< .05

Estimates are log odds, Italicized are the Standard Error

In assessing the relationship with mental health, a LOS score of 2 (β = .417, OR = 1.517, p < .05, model 3) or 3 (β = .633, OR = 1.883, p < .05, model 3) lost work opportunity events similarly predicted CES-depression scores, with significance across all three models (see Table 4). For model 3, this suggested that a lost work score of 2 was associated with 52% greater odds of having the high depression score relative to no lost work opportunity. A lost work score of 3 was associated with 88% greater odds of having high depression.

Table 4:

Multivariate analyses of factors impacting CES-D mental health score

	Model 1	Model 2	Model 3
Lost Work Opportunity (Comparison LOS = 0)
LOS=1	0.330(0.180)	0.201(0.183)	0.185(0.187)
LOS=2	0.676(0.200)***	0.453(0.207)*	0.417(0.211)*
LOS=3	0.985(0.310)***	0.691(0.320)*	0.633(0.327)*
Gender (Comparison = Male)
Female		0.294(0.142)*	0.274(0.145)
Ethnicity (Comparison = Non-Hispanic)
Hispanic		0.157(0.231)	−0.046(0.239)
Race (Comparison = White)
African American		0.208(0.185)	0.150(0.190)
Other Race		0.397(0.272)	0.426(0.275)
Birth Year		−0.008(0.014)	−0.017(0.014)
Education (Comparison = Less than HS)
GED		0.025(0.323)	0.186(0.336)
High School Graduate		−0.570(0.206)**	−0.346(0.214)
Some College		−0.731(0.218)***	−0.466(0.227)*
College Graduate or above		−0.745(0.229)***	−0.391(0.243)
Household Income 2016 (per 1000)		−0.002(0.000)*	−0.002(0.000)
Household Income 2004 (per 1000)			−0.001(0.000)
Self-Reported Health 2004 (Comparison = Good to Excellent)
Fair to Poor Health			1.307(0.162)***

^***< .001,

^**< .01,

^*< .05

Estimates are log odds, Italicized are the Standard Error

Discussion

This paper evaluated a novel measurement concept relevant to aging research and work policy by expanding the construct of involuntary retirement into the broader domain of lost work opportunity. Findings offer evidence to support current reformulations of the aging worker dilemma, holding that work and retirement are two sides of a unified construct (46). The LOS incorporates three measurement opportunities contained in the HRS to assess voluntariness and choice in these labor force transitions, with a unidimensionality that supports the relatedness of work opportunity and retirement and measuring these as a unified construct. The unidimensionality of the LOS score augments the demonstrated internal consistency and previous validation for discriminant and convergent validity (16). These findings suggest some utility in expanding the view of forced retirement into a broader matrix of lost work opportunity.

Testing predictive validity and consistent with previous research on declining health among involuntary retirees, the LOS score of 2 was associated with 46% increased odds for a negative self-reported health change and 52% increased odds for depression. Results were robust across models accounting for sociodemographic factors and prior self-reported health, with all models trending in the expected direction relative to more or less LOS events. This LOS formulation can offer improved precision over a single forced retirement variable in predicting health outcomes, as it incorporates more information on lost work opportunity as a known factor implicated in health change. Late-career unemployment as a single factor predicted negative mental health outcomes after retirement (47), while the LOS as a more comprehensive assessment of lost work opportunity at retirement age reveals a stronger association with negative self-reported health, reflecting more explanatory power.

In the final analysis, expanding involuntary retirement into a lost work construct created improved measurement but may not be the best approach to measurement going forward. Missing data on key variables resulted in the analysis incorporating just 8% of the total HRS sample. The study goal to use existing national survey data to improve measurement of occupational choice in retirement timing was only minimally accomplished. The HRS variables selected to inform forced retirement were subject to skip patterns and low response rates. For future research, an analytic approach is recommended that will better capture the complexity of the lost work opportunity dynamic that occurs at retirement age, while taking advantage of variables and survey items with fewer missing data.

This analysis did not directly assess health or non-work factors as causes of early retirement. The lag in time between the prior health measurement and the date of retirement may potentially miss issues of reverse-causality in which poor health prompts early retirement. We made the analytic decision to categorically code some variables that could potentially be treated as continuous (number of years retiring earlier than planned, self-reported health, and unemployment prior to retirement). Alternative approaches to consider include differential item functioning or an item-focused tree approach which can help resolve data loss from grouping (48). Attrition and non-response were not accounted for in this longitudinal analysis of the HRS data set, with other research showing that failure to account for attrition can result in under-representing the strength of an association (49). Incorporating these or other analytic strategies could offer additional insights in future research on this construct.

The study goal was to use measurement approaches to explore whether work opportunity related to forced retirements has impacts on retirement health. Better psychometrics will help inform ongoing research and policy as to the relevance of this relationship of pre-retirement work access to post-retirement health. Active aging initiatives (such as incentivized on-the-job training, retraining older workers, and adaptive workplaces) are used to increase work opportunity for older adults (50). Yet these approaches may fall short when paired with demographic considerations on the type of workers who benefit from extended work opportunity and which workers need alternate supports. It is evident from the recent COVID-19 pandemic that some demographic groups are hit harder by lost work opportunity than others, with black workers without a college degree showing the highest rates of unexpected COVID-19 early retirements (51). Working longer may work for some groups (often the college-educated), but public policy must consider the needs of all workers (e.g. aging adults in unstable and physically demanding work)(46). It has been suggested that any movement toward a working longer policy must be balanced with the opportunities for good jobs and protections for disability and job inequities (46). The current analysis finding negative health implications associated with loss of work opportunity as retirement approaches can help individuals and agencies prioritize work policy to the needs of the aging worker.

Learning Outcomes:

• Forced retirement is difficult to measure in cases where people choose to retire as their best choice between bad options, yet the distinction may be important for assessing health effects of forced retirement.

• A measure of lost work opportunity incorporating earlier-than-planned retirements, late career unemployment, and involuntary retirements (LOS) from survey data identifies a 46% increase in odds for negative self-reported health change for retirees reporting at least two of three LOS experiences, compared to those reporting no lost work opportunity.

• Results suggest that autonomy over retirement timing and working options can be important for the health of older adults, indicating a need for better measurement, education, programming, and policy to support aging worker opportunities.