Alternative Retirement Paths and Cognitive Performance: Exploring the Role of Preretirement Job Complexity

Abstract

Background and Objectives

Recent research suggests that working longer may be protective of cognitive functioning in later life, especially for workers in low complexity jobs. As postretirement work becomes increasingly popular, it is important to understand how various retirement pathways influence cognitive function. The present study examines cognitive changes as a function of job complexity in the context of different types of retirement transitions.

Research Design and Methods

We use data from the Health and Retirement Study (HRS) to examine change in cognitive function for workers who have held low, moderate, and high complexity jobs and move through distinct retirement pathways—retiring and returning to work, partial retirement—compared with those who fully retire or remain full-time workers. Inverse probability weighted regression adjustment (a propensity score method) is used to adjust for selection effects.

Results

There are systematic variations in the relationships between work and cognitive performance as a function of job complexity and retirement pathways. All retirement pathways were associated with accelerated cognitive decline for workers in low complexity jobs. In contrast, for high complexity workers retirement was not associated with accelerated cognitive decline and retiring and returning to work was associated with modest improvement in cognitive functioning.

Discussion and Implications

Both policy makers and individuals are beginning to embrace longer working lives which offer variety of potential benefits. Our findings suggest that continued full-time work also may be protective for cognitive health in workers who hold low complexity jobs.

Cognitive decline is among the most pressing concerns for both older individuals and aging societies (Corner & Bond, 2004; Institute of Medicine, 2015; Metlife Foundation, 2011). Recent research points to the intriguing possibility that retirement may accelerate cognitive decline (Bingley & Martinello, 2013; Bonsang, Adam, & Perelman, 2012; Fisher et al., 2014; Rohwedder & Willis, 2010). If so, the observation holds enormous importance for individuals as well as societies top-heavy with older citizens: in addition to financial benefits, working longer could mitigate cognitive decline.

A number of important questions remain unanswered. Workplaces vary substantially in cognitive demands (Goh, Pfeffer, & Zenios, 2015). One of the most interesting observations in emerging literature is working longer may provide more benefit to workers in low complexity jobs. In recent years, multiple investigations and research teams from different countries find that preretirement job complexity is associated with postretirement cognitive performance such that retirement does not lead to decline in workers who have jobs that typically require more complex cognitive processing (e.g., Finkel, Andel, Gatz, & Peterson, 2009; Fisher et al., 2014; Grywacz, Segel-Karpas, & Lachman, 2016; Grzywacz, Segel-Karpas, & Lachman, 2016; Lane, Windsor, Andel, & Luszcz, 2017; Oltmanns et al., 2017). Such findings hold after controlling for socioeconomic status, education, health, and a range of other factors that reduce concern that selection effects account entirely for differences. Whereas retiring from jobs that are low complexity is associated with cognitive decline, individuals who retire from high complexity jobs appear to maintain cognitive function, contributing to diverging aging trajectories for the most and least advantaged workers.

The observation is somewhat surprising. High complexity jobs, by definition, provide more cognitive stimulation than low complexity jobs and stimulating environments are good for cognitive functioning (e.g., scaffolding theory of aging and cognition, Park & Reuter-Lorenz, 2009; Reuter-Lorenz & Park, 2014). Thus, leaving a high complexity job could have more dire consequences than leaving a lower complexity job. Just as staying physically active is associated with maintaining physical function as people age, staying cognitively active could maintain cognitive function. This hypothesis is captured in the nearly century-old psychological hypothesis embodied in the adage “use it or lose it.” In short, brains benefit from stimulating environments.

An alternative possibility suggests that individuals in high complexity jobs may maintain function after retirement due to enhanced “cognitive reserves” (Stern, 2002). High complexity jobs regularly expose workers to cognitive challenges that require problem solving. Cognitive reserves not only include bodies of crystallized intelligence, they result in greater flexibility in cognitive processing. Cognitive reserves allow people to better compensate for slower processing and other age-related changes. Cognitive flexibility is especially important in novel situations, such as life transitions.

The concept of “cognitive reserve” is highly compatible with economic models of human capital generation. Investments in education, on-the-job training, and learning-by-doing contribute to individuals’ stock of durable knowledge which then influences occupational choice and subsequent earnings, leveraging increased growth in human capital over the life course (Ben-Porath, 1967). From this perspective, long-term exposure to high complexity jobs protects against cognitive decline. In contrast, workers in lower complexity jobs presumably are less likely to build resources that protect against aging-related cognitive vulnerabilities.

This growing research domain has important implications for theory and practice. First, benefits of working longer must be qualified. Evidence of “mental retirement” in workers who retire from high complexity jobs has not been supported, suggesting that not all individuals may benefit from longer working lives, and likewise, not all individuals may experience cognitive consequences from retiring. Second, findings speak to the potential malleability of cognitive performance; workplaces may prepare individuals for retirement.

Current Study

It is possible that different retirement pathways will be more and less protective for different types of workers. To date, however, research has focused only on the effects of full retirement. The present study was designed to investigate alternative retirement pathways to full retirement and working longer. Exploring alternative retirement pathways is both theoretically intriguing and important since workers and employers are expressing increasing interest in phased retirement (AARP, 2017). Partial retirement (i.e., downshifting to part-time work) or retiring and returning to work may facilitate useful alternatives to full retirement, and potentially influence overall cognitive function.

Hypotheses

Based on prior findings we expect that cognitive effects of retirement will interact with preretirement job complexity. From our conceptual stance, we expect workers who retire from high complexity jobs have greater cognitive reserve and greater access to resources making transitions less challenging regardless of pathway. For low complexity workers, engaging in part-time work after retirement offers continued stimulation, mitigating accelerated cognitive decline imposed by retirement.

Specifically:

• H1: Per earlier research, we hypothesize that cognitive performance will be best preserved in workers retiring from high complexity jobs when compared with low complexity jobs. We hypothesize that retirement pathways will have little effect on cognitive performance for high complexity workers.

Among workers retiring from low complexity jobs, we test two hypotheses:

• H2a: Cognitive decline will be accelerated most by full retirement (i.e., the pathway that produces the least amount of cognitive stimulation);

• H2b: Alternative retirement pathways will offer greater cognitive stimulation, and thus, will help maintain cognitive function.

Research Design and Methods

The study is based on data from the Health and Retirement Study (HRS), a nationally representative longitudinal survey of individuals over age 50 (and their spouses, regardless of the spousal age). We use data from biennial waves of the HRS from 1996 to 2010. These data offer the most comprehensive nationally representative panel data on U.S. older adults available, including cognitive function and work behaviors (Lachman & Weaver, 1997; RAND Center for the Study of Aging, 2014). Although complete HRS data are available through 2014 in most cases, our study uses refined measures of occupational characteristics for which we were granted access through 2010.

Sample

To test our hypotheses, we used a propensity score method (inverse probability weighted regression adjustment) to adjust for selection effects. This approach, described in detail below, allowed us to compare multiple treatment groups to one another. Our sample includes individuals who worked full-time (i.e., 35 hr or more and self-identified as not retired) and subsequently: (a) stopped working completely and self-identified as retired (i.e., full retirement); (b) partially retired by engaging in paid work after retirement, that is, they worked 20 hr or less and reported being “retired” (i.e., partial retirement); (c) retired fully and later returned to work (i.e., retire and return to work); or (d) remained working full-time. To accurately measure cognitive changes in association with retirement, we included pre- and postretirement measures of cognitive function. To circumvent the potential confounding of cognitive disengagement as people approach retirement (Bonsang et al., 2012; Rohwedder & Willis, 2010), we established “baseline” cognitive function scores for participants based on two waves of data collection before retirement. In doing so, we limited the sample to those with two consecutive waves of full-time work prior to retirement. Second, research suggests that cognitive changes correlated with retiring may not be observable for 12 months or longer after retirement (Bonsang et al., 2012). For this reason, alongside our interest in examining the effects of returning to work following retirement, we lagged cognitive function measures by one biennial wave following reported retirement, ensuring that at least 12 months had passed. Finally, we include a continuously employed full-time “control” group, which utilizes the most recent four-wave period of consistent full-time workers. For all groups, baseline cognition is measured at Time 1 (two study waves, or up to 4 years before reported retirement, which is reported at Time 3), and compared with cognitive function at Time 4 (one study wave, or at least 2 years following reported retirement.) To minimize the potential endogenous effect of declining cognitive status and retirement, we excluded all individuals whose cognitive function in the first two study waves indicated cognitive impairment (threshold defined by Crimmins et al., 2011).

Based on these selection criteria, our sample includes: 740 individuals who fully retired, 182 who partially retired, 185 who returned to work after retirement, and 1,187 continuously employed full-time workers.

Cognitive function

The outcome measure for this study is cognitive function, which is based on a 20-point test of immediate and delayed word recall, conducted every 2 years. Like Fisher and colleagues (2014), we use this measure because it serves as a standard test of episodic memory. We use the total score, which is the number of correct answers out of 20. We deal with additive common causes that affect both cognition and occupational complexity by using the longitudinal change in the cognitive score from Time 1 to Time 4 as our dependent variable to eliminate unobserved fixed effects that may be correlated with level of cognition. The efficacy of this step is illustrated in Supplementary Appendix A.

Job complexity

Like Fisher and colleagues (2014), we use the O*NET classification of jobs (via a crosswalk that links U.S. Census codes with the Standard Occupation Classification (SOC) codes in the O*NET and the HRS) to obtain external occupational-level ratings of job characteristics pertaining to preretirement occupation. The researchers involved in this study obtained access to these sensitive data for HRS data through the 2010 wave, providing the ability to examine retirement transitions occurring before the 2008 recession. These occupational characteristics provide a series of variables describing occupational characteristic scores using a rating scale for each of three aspects of one’s occupation: (a) abilities (i.e., the expected abilities required to engage in a given job), (b) activities (i.e., the expectation of participation in activities associated with a given job), and (c) contexts (i.e., the situational aspects of day-to-day working associated with a given job). For example, the degree to which an occupational category involves the activity of “getting information” (i.e., observing, receiving, and otherwise obtaining information from all relevant sources) has a total score calculated on a range from 0 to 1 based on how often that job typically requires getting information. A total of 36 job-related abilities, activities, and contexts were available for the standard occupations coded in the HRS (Supplementary Appendix B).

Unlike previous studies, we used exploratory factor analysis to identify factors that explain work complexity. To identify this measure, we included all 36 occupational characteristics measures available. All items with factor loadings below 0.60 were removed, and among the remaining 18 items, we identified unique measures with the highest factor loadings. We used an iterative process to select the fewest items to produce the highest alpha score. The final selection included a total of five remaining items. Based on the items that remained, we describe this factor as “cognitive complexity”: (a) making decisions and solving problems; (b) thinking creatively; (c) coaching and developing others; (d) frequency of decision-making; and (e) freedom to make decisions (Supplementary Appendix C). Cronbach’s alpha for this factor is 0.952, with a range for the total score of 2.417 to 3.724. Due to the methodological approach employed in this analysis, we recoded complexity as terciles—low, moderate, and high.

Covariates

As described below, our methodological approach includes two models: one that accounts for selection into retirement (i.e., “the propensity model”), and a regression model that adjusts for the propensity into a retirement transition and controls for factors that influence changes in cognitive function (i.e., “the outcome model”). Our measures are selected based on robust literature on retirement and cognitive function and extensive sensitivity analyses. We iteratively assessed variables associated with retirement, lifestyle, cognitive and physical health, and employment decisions, identifying a parsimonious and robust set of measures, which we present in our final models (details regarding specific measures tested are available upon request).

Our final models included demographic, health, and economic resource factors. First, demographic covariates included: race (non-Hispanic blacks are “1” and all others are “0”), sex (females are “1” and males are “0”), and educational attainment (four dichotomous variables including: less than a high school education, a completed high school diploma, some college (no degree), and a college degree or more (reference group). Age is a continuous measure, but we also included a dichotomous measure of whether an individual is eligible for early social security (i.e., at least age 62) at Time 3, which has been shown to be an important selection factor into retirement (Bonsang et al., 2012).

Second, for health we included cognitive and physical health factors. Because changes in physical health could initiate a retirement transition or explain a change in cognitive status, we included measures that adjust for potential preretirement health decline. First, we included the cognitive score at Time 2 to account for potential decline in the period leading up to potential retirement. We used a 27-point measure for cognitive function that accounts for speed of processing and memory preceding a retirement transition. We also included a continuous measure of frailty at Time 2, and, to adjust for the potential causal effect of declining health as an impetus for the retirement transition. Frailty (Time 2) is measured following Yang and Lee (2010), as an index based on 30-items: 8 chronic illnesses, 5 activities of daily living limitations, 7 instrumental activities of daily living limitations, 8 depressive symptoms, obesity (i.e., body mass index of 30 or greater), and self-rated health (a 5-point Likert item with higher values indicating better health). We also included a measure for change in self-rated health observed at Time 3 (relative to Time 2). Change in self-rated health was measured such that decline in self-rated health is negative and improvement is positive.

Research shows that retirement is strongly correlated with retirement and health status of a spouse (e.g., Moen, Kim, & Hofmeister, 2001). Thus, we included a dichotomous measure for marriage with “1” coded for those who are married and “0” for those not married. We included three interaction measures related to marriage such that individuals who are unmarried are coded “0” and those married had a value. We also included married interacted with spousal cognitive function (measured based on total recall), married interacted with total years of education (based on a continuous measure of years of education from 0 to 17), and married interacted with work status of the spouse (i.e., individuals are coded “1” if they have a working spouse and “0” otherwise).

We included several measures of financial resources: a measure of total household wealth (measured in quintiles at baseline), weekly personal income (logged), and a dichotomous measure of whether an individual has a defined benefit pension, and whether an individual has a defined contribution pension.

Finally, we included controls related to timing of data collection and birth cohort. We controlled for number of months between Time 1 and Time 4 to account for variations in the timing of data collection. Our sample includes three birth cohorts: the HRS cohort (reference), the War Babies cohort, and the Early Baby Boomers. Each are dichotomous measures.

Although data about psychosocial factors are available in the HRS, they are collected on half of the sample on a rotating basis every 4 years beginning in 2006. Our sample selection process did not allow us to use any of these data. Table 1 provides covariates that appear in the outcome and selection models (see description below).

Table 1.

Descriptive Characteristics and Identification of Variables in Models

	Total sample (N = 2,295)		Cont full-time worker (N = 1,187)		Partially retire (N = 182)		Retire/return to work (N = 185)		Fully retiree (N = 740)
	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Min	Max	Sig
Change in total recall	−0.63	3.26	−0.33	3.31	−1.099	2.990	−0.659	3.272	−0.977	3.194	−15	13
Non-Hispanic blacka,b	0.11		0.11		0.082		0.162		0.095		0	1	***
Educational credentiala,b
Less than high school	0.09		0.09		0.071		0.097		0.095		0	1
High school	0.33		0.28		0.308		0.324		0.409		0	1	*
Some college	0.24		0.25		0.181		0.227		0.224		0	1
College	0.35		0.38		0.440		0.351		0.272		0	1	+
Baseline agea,b	58.08	4.35	56.95	4.38	59.467	3.741	58.584	4.065	59.434	3.995	47	80	**
Age 62+ at retirementb	0.35		0.23		0.511		0.443		0.485		0	1	***
Femalea,b	0.49		0.48		0.555		0.535		0.480		0	1
Preretirement cognitive scorea,b	17.80	3.35	17.74	3.47	18.088	3.049	17.951	3.234	17.768	3.256	5	27	+
Frailty Indexa,b	−0.12	0.10	−0.11	0.09	−0.112	0.087	−0.119	0.111	−0.127	0.101	−0.778	0	***
Change in self- rated healtha,b	−0.09	0.81	−0.11	0.77	0.044	0.712	−0.049	0.905	−0.091	0.885	−4	3	***
Marrieda,b	0.69		0.68		0.687		0.686		0.707		0	1
Married × Spouse total recall scorea,b	7.98	5.94	7.95	5.97	8.121	6.021	7.881	5.971	8.031	5.868	0	20
Married × Spouse years of educationa,b	9.23	6.55	9.28	6.70	9.456	6.725	8.751	6.369	9.207	6.310	0	17
Married × Spouse worksa,b	0.49		0.501		0.473		0.422		0.468		0	1
Total household wealth (quintiles) b	5.54	2.83	5.51	2.99	5.742	2.864	5.497	2.619	5.549	2.605	1	10	***
Individual weekly income (logged) b	6.63	0.68	6.65	0.72	6.672	0.724	6.622	0.582	6.597	0.634	0.731	10.882	***
Has a defined benefit pensionb	0.43	0.50	0.34	0.47	0.478	0.501	0.508	0.501	0.553	0.498	0	1
Has a defined contribution pensionb	0.50	0.50	0.53	0.50	0.451	0.499	0.481	0.501	0.459	0.499	0	1
Months baseline to Wave 4a,b	72.51	3.90	73.04	4.35	72.247	3.334	71.827	3.241	71.912	3.239	59	86	***
Birth cohort groupb
HRS	0.54		0.33		0.79		0.77		0.11		0	1	***
War babies	0.25		0.30		0.17		0.19		0.20		0	1	***
Early baby boomers	0.20		0.36		0.04		0.04		0.03		0	1	***

^aItems used in the outcome model.

^bItems used in the propensity model.

***p < .001; **p < .01; *p < .05; +p < .10 (significance indicates significant differences across groups, based on a one-way ANOVA test).

Statistical Analysis

A large-scale experimental research design is the only way to definitively determine which work/retirement pathway causes the best cognitive outcomes. However, it is not possible to randomly assign individuals to a work or retirement pathway. Scholars have used various methods to attempt to best address these issues for studies on the cognitive effects of retirement using observational data. For instance, Fisher and colleagues (2014) used latent growth curve modeling to assess the effects of job complexity on the impact of full retirement among a sample of individuals initially working. This approach allows for longitudinal analysis of both intra- and inter-individual changes over time. Others have addressed this problem using instrumental variable estimation (e.g., Bonsang et al., 2012, Rohwedder & Willis, 2010). Given the goals of our study, we used a different set of strategies to decrease the potential bi-directional relationship between retirement pathways and changes in cognitive function.

First, as described earlier in the description of our sample section, we included only individuals who were: (a) initially full-time workers; (b) did not have cognitive deficits before retirement; and (c) had data for four consecutive waves of data. Second, we used propensity score methods to simulate a quasi-experimental study for use with observational data. We know that occupations and retirements are not randomly assigned, so like Carr, Kail, and Rowe (2018) and Grool and colleagues (2016), we use inverse probability weighted regression adjustment (Cattaneo, 2010; Cattaneo, Drukker, & Holland, 2013) as implemented in the ipwra option of the teffects command in Stata 14 (StataCorp, 2013) to estimate cognitive function if an individual continued working full time rather than the following the retirement path actually taken. The estimator adjusts for key observed factors that influence selection into particular occupations and retirement pathways on changes in cognitive function (Holland, 1986; Imbens & Rubin, 2010), using a multinomial logistic regression model to predict selection into each specified group. This method has a doubly robust property, meaning that unbiased estimates of treatment effects can be calculated even if the outcome model or the propensity model (but not both) is misspecified (Wooldridge, 2007). However, the estimated treatment effects may be biased if there are unobserved confounders that affect occupational choice or retirement decisions and also affect cognitive change. We examined sensitivity of our results to variations in the variables included in the econometric models. Given our goal of investigating how treatment effects vary by job complexity and retirement paths, instrumental variable methods are not feasible.

We addressed our research questions by using the teffects model to estimate potential outcomes (POMs) defined as the mean 6-year change in cognition score from baseline wave t to four study waves later at t+3. Separate POMs are estimated based on each of the four retirement pathways first (Model 1) and then within 3 × 4 = 12 groups classified by occupational complexity (high, medium, low) and work/retirement path (continuous full-time work, partial retirement, retiring and returning to work, fully retiring; Model 2). Average treatment effects (ATEs) are calculated as the difference between the POM of the treated group and the POM of the control group. For example, the predicted effect for full retirement relative to continued full time work after two survey waves for those who held low complexity jobs at baseline is equal to POM(full retirement, low) − POM(full-time work, low). Under the assumption of no confounding due to unobservables, inverse propensity and regression adjustments make the treatment and control groups comparable (Abadie & Imbens, 2012).

Results

Table 2 presents results for Model 1, with calculated POMs for each of the four pathways, and differences between each pathway relative to all others (based on ATEs). Results indicate that remaining a continuous full-time worker is associated with less cognitive decline relative to fully retiring (ATE = 0.613 fewer points; p < .001). Partial retirement and retiring and returning to work were not associated with statistically different cognitive losses than remaining full-time. Partial retirement was not associated with a better outcome than full retirement. However, individuals who retired and returned to work experienced less decline than those who fully retired (0.706 fewer points, or 70% less decline; p < .05).

Table 2.

Estimated Changes in Cognitive Function for Different Retirement Pathways Relative to Continuous Full-Time Work

Work/ retirement group	Change in cognitive function score (POM)	Difference relative to full-time work (ATE)		% Difference relative to full-time work	Difference relative to partial retirement (ATE)		% Difference relative to partial retirement	Difference relative to return to work (ATE)		% Difference relative to return	Difference relative to full retirement (ATE)		% Difference relative to full retirement	95% CI	95% CI
Continuous full-time worker	−0.39				0.37	NS		−0.09	NS		0.61	***	61% less	−1.24	−0.78
	0.11				0.26			0.31			0.16
Partially retire	−0.76	−0.37	NS					−0.46	NS		0.25			−1.22	−0.30
	0.23	0.26						0.37			0.26
Retire/return to work	−0.30	0.09	NS		0.46	NS					0.71	*	70% less	−0.87	0.26
	0.29	0.31			0.37						0.31
Fully retire	−1.01	−0.61	***	156% more	−0.25	NS		−0.71	*	235% more				−0.61	−0.17
	0.12	0.16			0.26			0.31

Note: “Change in Cognitive function” is the potential outcome mean (POM) calculations for each group and the “Difference Relative to ...” are the estimated average treatment effects (ATE) for retirement treatment groups relative to the associated control group. Statistical significance indicates significant differences relative to the control group (i.e., continuous full-time workers). All calculations use inverse probability weighted regression adjustment. The variables used in the propensity models and the outcome models are given in Table 1.

***p < .001; **p < .01; *p < .05.

Table 3 presents results for Model 2, which examined change in cognitive function for the retirement paths as a function of preretirement job complexity. To address our first hypothesis, we explored whether individuals with high complexity jobs experienced fewer consequences in relation to retirement regardless of retirement transition. This hypothesis was supported, but surprisingly, retiring and returning to work was associated with a small improvement in cognitive function, and the best cognitive function outcome overall.

Table 3.

Estimated Changes in Cognitive Function for Different Retirement Pathways Relative to Continuous Full-Time Work by Complexity Level

Work/retirement group		N	Change in cognitive function score (POM)	Difference relative to full-time work (ATE)		% Difference relative to full- time work	Difference relative to partial retirement (ATE)		% Difference relative to partial retirement	Difference relative to return to work (ATE)		% Difference relative to return	Difference relative to full retirement (ATE)		% Difference relative to full retirement	95% CI	95% CI
Low complexity	Continuous full-time worker	316	−0.31				1.84	***	86% less	2.04	***	87% less	1.31	***	81% less	−0.72	0.1
			0.21				0.50			0.44			0.30
	Partially retire	47	−2.15	−1.85	***	592% more				0.20	NS		−0.52	NS		−3.04	−1.25
			0.46	0.50						0.59			0.50
	Retire/ return to work	53	−2.35	−2.04	***	657% more	−0.20	NS					−0.73	+		−3.10	−1.60
			0.38	0.44			0.59						0.44
	Fully retire	447	−1.6	−1.31	***	423% more	0.52	NS		0.73	+					−2.03	−1.21
			0.21	0.30			0.50			0.44
Moderate complexity	Continuous full-time worker	278	−0.53				0.65	+		0.74	+		0.20	NS		−0.90	−0.16
			0.19				0.35			0.39			0.27
	Partially retire	87	−1.18	−0.65	+					−1.40	**	649% more	−0.45	NS		−1.76	−0.60
			0.30	0.35						0.45			0.35
	Retire/ return to work	80	0.22	0.74	+		1.40	**	118% less				0.95	*	129% less	−0.45	0.88
			0.34	0.39			0.45						0.39
	Fully retire	451	−0.73	−0.20	NS		0.45	NS		−0.95	*	440% more				−1.11	−0.35
			0.19	0.27			0.35			0.39
High complexity	Continuous full-time worker	146	−0.66				−0.14	NS		−1.09	**	253% more	0.04	NS		−1.13	−0.19
			0.24				0.47			0.35			0.42
	Partially retire	48	−0.52	0.14	NS					−0.96	*	221% more	0.18	NS		−1.31	−0.03
			0.40	0.47						0.47			0.53
	Retire/ return to work	52	0.43	1.09	**	152% less	0.96	*	182% less				1.13	**	162% less	−0.05	0.91
			0.25	0.35			0.47						0.42
	Fully retire	290	−0.70	−0.04	NS		−0.18	NS		−1.13	**	262% more				−1.37	−0.03
			0.34	0.42			0.53			0.42

Note: “Change in Cognitive function” is the potential outcome mean (POM) calculations for each group and the “Difference Relative to ...” are the estimated average treatment effects (ATE) for retirement treatment groups relative to the associated control group. Statistical significance indicates significant differences relative to the control group (i.e., continuous full-time workers). All calculations use inverse probability weighted regression adjustment. The variables used in the propensity models and the outcome models are given in Table 1.

***p < .001; **p < .01; *p < .05.

To address our second set of hypotheses, we examined the effects of those retiring from low cognitive complexity jobs. Our results do not support either hypothesis. Significant decline was observed for low complexity workers who retired fully, retired partially, or retired and returned to work. Partial work or returning to work offered no protection from cognitive decline. Indeed, retiring and returning to work was associated with greater decline than fully retiring (ATE = 0.725 points greater), though this effect did not quite reach significance.

For those in moderately complex preretirement jobs, there were no differences between full retirement and full-time continued work. Retiring and returning to work, however, was associated with less decline than fully retiring (ATE = 0.946 points, or 129% less decline). Partial retirement was not protective.

The estimated changes in cognitive function (i.e., POM calculations and robust standard errors) are shown in Figure 1. The significance shown in this figure represents significant differences relative to full retirement.

Figure 1.

Estimated 6-year change in cognitive function estimates by work/retirement group and preretirement job complexity (low, moderate, and high). Notes: Significance indicates a significant difference in cognitive function relative to full retirement; ***p < .001, **p < .01, ⁺p < .1. Positive numbers indicate an estimated improvement in cognitive function relative to Time 1, and negative numbers indicate a decline in cognitive function relative to Time 1.

Discussion and Implications

With longer working lives on the horizon, there is a growing need to understand how alternative retirement paths affect cognitive function. This article builds on recent literature that suggests that retirement may have a negative effect on cognitive performance or, conversely, that continued work may protect cognitive function. Our contribution is to provide evidence on whether the effect of work on cognition is uniformly negative across alternative pathways in the transition between lifetime careers and full retirement and, furthermore, whether the cognitive complexity of the occupations that individuals pursue before retirement affects the impact of work or retirement on changes in cognitive performance.

The methodology we used attempted to replicate results that would be obtained by two different (hypothetical) longitudinal randomized trials with two statistical models. Model 1 considered an assignment of full-time workers to four conditions: continuous work and three different retirement pathways with weights based on regression and propensity adjustments to balance the samples in each condition as if the assignment were random. The results are consistent with previous research: full retirement was related to significant cognitive losses relative to what would be expected if an individual remained engaged in full-time work. Regarding alternative retirement paths, we also found that partial retirement did not relate to significantly greater losses than continued full-time work, and, surprisingly, a return to work was associated with better outcomes than full retirement, but not as good as continuous full-time work.

Model 2 considers a 12-way assignment in which pathways are estimated for those with jobs involving low, medium and high levels of cognitive complexity. Our first hypothesis was supported and is consistent with models of cognitive reserve. We found that cognitive performance was best preserved in workers retiring from high complexity jobs when compared with low complexity jobs and that retirement pathways had little effect on cognitive performance for high complexity workers.

We also considered two hypotheses related to low complexity work. We found full retirement had a significantly smaller negative effect than either partial retirement or retiring and returning to work. Contrary to our hypothesis, postretirement work was associated with steeper decline than full retirement. This pattern is consistent with the cognitive reserve model of decline. Full retirement may place fewer demands on cognitive functioning than more complex trajectories.

Reverse causality is a key concern for research examining the relation between cognitive function and retirement. We attempted to control for the effects of the level of cognition by using the difference in cognition more than 6 years, including a measure for preretirement cognition in Wave 2, and applying propensity score methods. A careful analysis of the impact of health and cognition on employment of older workers using data from both HRS and its English counterpart, ELSA, concluded that “…cognition is not a key driver of employment at these ages” (Blundell et al., 2017, p. 4). Thus, we think it is unlikely that reverse causation explains our results.

By definition, our study shows average effects across groups of individuals based on potential work and retirement pathways. However, there are significant variations across individuals that are likely to be explained by a variety of factors. For example, cognitive changes in association with retirement may be related to economic factors (wealth, pension benefits, wage rates, working conditions, skill requirements), health (disability, disutility of work, life expectancy), or leisure preferences (time with family and grandchildren, bucket list of foreign travel, volunteer activities, hobbies, and intellectual interests). Under the most favorable assumptions (i.e., conditional independence), we may interpret our estimated retirement “effects” as the weighted average of the difference in the heterogeneous potential outcomes of the people in the treatment and control groups where the composition of these two groups are balanced by regression adjustment and propensity weighting rather than randomization. Interpreted in terms of Rubin causality, our findings for the entire sample of full-time workers at Time 1 allow us to conclude that, on average, continued work tends to protect against cognitive decline but it does not provide any clues about why this effect occurs, for whom or what changes in economic, health or other policies would have in protecting older workers.

The value of dividing the sample of initially full-time workers by type of work/retirement transition in Model 1 and then further partitioning it by cognitive complexity of occupation in Model 2 is that the comparison of average effects across subsamples can offer clues about underlying mechanisms influencing cognitive changes at retirement. However, because POMs and ATEs within subgroups are still just averages of heterogeneous effects generated by a variety of mechanisms, our results still do not allow us to make unqualified statements about “the effect” of retirement on cognition. Nonetheless our results suggest that the most important mechanisms are not those associated with varying paths to retirement. Rather, they point to mechanisms related to the cognitive challenges that individuals faced during their working careers that are captured in the occupational complexity measure.

The likely mechanisms responsible for the substantial negative effect of retirement on cognition for workers in the least complex occupations and the lack of effect for those in the most complex groups can be found in theories and evidence in economics, psychology, and neuroscience, and involve behavior over the entire life course from birth to postretirement. In economics, human capital theory (Becker, 1964) posits that individuals develop both market and nonmarket skills through parental investment in child rearing and formal schooling followed by on-the-job investments in developing skills required to be productive in their chosen occupation. Recent estimates of the parameters of the Ben-Porath model by Polachek, Das, and Thamma-Apiroam (2015) show that cognitively complex careers (i.e., high cognitive ability, elastic response of effort to rewards, heavier weight on future rewards) tend to relate to lower rates of depreciation in skills than those in cognitively less complex careers. This implies that the cognitive abilities embodied in a worker’s human capital are more durable for those in more complex occupations. In addition, the vast human capital literature shows that human capital also has large effects in nonmarket activities such as health and leisure activities.

These theoretical and empirical findings from economics are congruent with the century old “use it or lose it” hypothesis from psychology as well as much more recent theories and evidence from psychology and neuroscience. If retirees with a history of cognitively complex jobs have developed a large stock of human capital that is useful in both market and nonmarket spheres, they may be motivated to pursue more cognitively (and perhaps physically) stimulating and novel activities in retirement for the pleasure or satisfaction (i.e., utility) they provide. Conversely, if people who followed low complexity careers arrive at retirement age with a low stock of capital that is depreciating, they will be impeded from learning to enjoy novel and stimulating activities that would help them to maintain their cognitive abilities. There is growing evidence that engaged lifestyles in later life maintain cognitive function (Hertzog, Kramer, Wilson, & Lindenberger, 2008). In addition, lifelong experience associated with education and work create “cognitive reserve” (Stern, 2002) or “scaffolding” (Park and Reuter-Lorenz, 2009) by building more complex and more durable neural networks that embody knowledge (or crystallized intelligence), which helps older people compensate for decreasing fluid abilities.

This study has limitations that should be considered when interpreting the findings. The kinds of activities individuals engage in while at work, and whether these findings are a consequence of nonwork lifestyle behaviors are not known. For instance, it is plausible that individuals in high complexity jobs seek out more cognitively engaging activities outside of work, particularly when seeking activities to replace work following retirement. In other words, instead of going to work, individuals in high complexity jobs may engage in stimulating activities such as remaining socially engaged, cognitively engaged, and physically engaged. Individuals who retire from work in low complexity jobs may not have the resources or interests to engage in stimulating lifestyles that protect against cognitive decline. For instance, we know that many older adults spend a significant amount of time watching TV, an activity that does not offer social, cognitive, or physical engagement (Depp, Schkade, Thompson, & Jeste, 2010).

In addition to these unknown factors, because of small sample sizes, we often lack statistical power to obtain reliable estimates of differences in average effects of alternative retirement pathways within complexity groups (as given in Table 3 and illustrated in Figure 1). This means, for example, that we do not have sufficient data to explore why a return to work appears to have a small positive effect for workers in moderate and high complexity jobs while it has the largest negative effect of any pathway for low complexity workers. This could be the result of selection bias that our methods are unable to address, like different reasons for undertaking work after retirement (e.g., financial pressure vs seeking rewarding experience). Alternatively, it is possible that characteristics of low complexity jobs that help maintain cognition (e.g., keeping a regular work schedule and old friendships with coworkers) are not present in part time jobs available to older low skilled workers.

In conclusion, we believe that the present findings pointing to beneficial effects of working longer for workers in low complexity occupations and a non-negative retirement effect for high complexity workers are consistent with economic, psychological, and neurological theory and evidence. These novel findings reveal important information about the combined effects of preretirement job complexity and alternative retirement paths to cognitive function. Future research should examine specific mechanisms responsible for the connection between work, retirement, and cognition.

Funding

This work was supported by funding from the Alfred P. Sloan Foundation. R. Willis acknowledges support from National Institute on Aging grant (P01 AG2026571). L. L. Carstensen acknowledges support from National Institute on Aging grant (NIA R37-8816).

Conflict of Interest

None reported.