How Do Pre-Retirement Job Characteristics Shape One’s Post-Retirement Cognitive Performance?



Download 464.26 Kb.
Page1/6
Date29.01.2017
Size464.26 Kb.
  1   2   3   4   5   6



How Do Pre-Retirement Job Characteristics Shape One’s Post-Retirement Cognitive Performance?
Dawn C. Carr, PhD1

Stanford University


Melissa Castora-Binkley, PhD

University of South Florida


Ben Lennox Kail, PhD

Georgia State University


Robert Willis, PhD

University of Michigan


Laura Carstensen, PhD

Stanford University


November 9, 2015
ABSTRACT

Objectives: This study seeks to examine whether pre-retirement occupational characteristics impact cognitive changes associated with retirement.

Method: Using data from the Health and Retirement Study, we examined a sample of adults age 50 years or older with normal cognitive function over four waves who, at baseline, were working full-time and subsequently either retire (n=721) or remain full time (n=1,296). We adjusted for potential selection bias using propensity scores. Exploratory factor analysis was used to identify two key job factors – intellectual and mechanical – which were coded as low, moderate, or high.

Results: Among retirees, the lowest cognitively complex jobs were related to a significantly greater level of cognitive decline relative to both those who retired from moderate or the highest cognitively complex jobs. Among retirees, low compared to high mechanically complex jobs were associated with significantly less decline. Remaining in full-time work was related to consistent levels of cognitive decline regardless of cognitive or mechanical complexity of one’s job. Among those in the highest cognitively complex and those in moderately mechanical jobs, there were no differences in cognitive decline between continuous full-time workers and retirees.

Discussion: These findings contribute to the growing base of research helping explain how occupational factors influence cognitive changes that occur with aging and retirement. We suggest that scaffolding theory, a recent theory from cognitive psychology and neuroscience, in combination with human capital theory may explain the mechanism underlying our findings.
Key Terms: cognitive performance, retirement, propensity score weighting, Health and Retirement Study

Running Head: Pre-Retirement Job Characteristics and Cognitive Decline

INTRODUCTION

At a population level, there is growing evidence that retirement has a significant, negative impact on one’s cognitive performance in later life. This finding is not merely because those with declining cognition retire while those with more robust cognitive performance continue to work. Rather, several papers find that the negative impact of early retirement on cognition, measured by a test of episodic memory, is causal (Rohwedder & Willis, 2010; Bonsang, Adam, & Perelman, 2012; Celidoni, Bianco, & Weber, 2013). These findings have been interpreted in the context of the long standing “use it or lose it” hypothesis that holds that cognitive declines associated with aging can be reduced by engaging in mental exercise. The negative impact of retirement on cognition then follows from the further hypotheses that the work environment provides more mental stimulation than the home environment and, possibly, that the expectation of early retirement reduces the incentive older workers to exert the mental effort needed to maintain their skills.

While evidence is accumulating that leaving work has a negative impact on the cognitive performance of older people, the mechanisms that underlie this effect have not been fully clarified. In this paper, we build on findings of several recent studies of older adults which suggest that pre-retirement job characteristics shape the degree to which retirement influences changes in cognitive performance. This paper adds to this line of research by estimating how retirement impacts change in cognitive performance over a six-year time span among workers whose jobs vary in complexity in both cognitive and mechancial dimensions.

To help develop hypotheses about the impact of occupational complexity, we draw on recent advances in cognitive psychology and neuroscience that are embodied in the “scaffolding theory of aging and cognition” (STAC) proposed by Park and Reuter-Lorenz 2009). The STAC is motivated by noting that while many components of cognition such as working memory, ability to learn and recall new information and fluid intelligence (i.e., reasoning ability) decline with age, most older adults continue to be able to function quite well despite these declines. Park and Reuter-Lorenz argue that the aging brain develops compensatory scaffolding (i.e. recruitment of additional neural circuitry) to shore up the deteriorating components whose function has become noisy, inefficient or both. Experimental evidence shows that sustained cognitive effort in learning a new complex task has a postive effect on episodic memory (Park et al. 2014)

We argue that the scaffolding theory is consistent with human capital theory. In that theory, an individual’s productivity in a given task depends on reasoning ability (Gf: fluid inteligence) and on the extent of knowledge relevant for that task (Gc: crystallized inteligence) where Gf and Gc tend to be complementary. Early in life, Gf increases the productivity of people in acquiring useful knowledge through schooling, job experience and and other activities (e.g., managing finances, rearing children). Later in life, accumulated knowledge increases the productivity of persons whose reasoning ability has declined. When asked to solve a novel problem, brain imaging studies show that the left pre-frontal cortex of young people lights up, suggesting that Gf is primarly involved in finding a solution. For older people confronted with the same problem, both the left and right lobes light up, suggesting that retreival of knowledge through memory processes as well as reasoning are involved. In addition, the studies find that higher performing older adults show a greater degree of bi-lateral activity than lower performing adults. Cognitively complex jobs plausibly require more mental exercise in order to maintain skills and perform more challenging tasks. This, in turn, stimulates compensatory scaffolding which serves to reduce the decline of episodic memory (Li, Baldassi, Johnson, & Weber, 2012) . Moreover, greater scaffolding may enhance performance in non-work environments, thus lessening the effect of retirement on cognition.

Estimating the effect on cognition of stopping work versus the alternative of continuing to work inherently involves dealing with missing data on the unobserved alternative. Since any given individual can follow only one alternative, it is impossible to know what that person’s cognitive score would have been had he or she followed the other alternative. If it were possible to randomly assign people to a “treatment” consisting of a given pattern of work and retirement, we could estimate the average treatment effect (ATE) by calculating the difference in the mean cognitive change experienced by people who follow each alternative. However, individuals (or their employers) choose which path will be followed; hence, the assignment to a given treatment is non-random.

In this paper, we employ a counterfactual framework involving a comparison of potential outcomes—i.e., mean change in cognition over a six year period—for persons who are are fully employed during their first two waves in the Health and Retirement Study and are fully retired during the next two waves compared to a second group who who work full time during all four waves. Under the assumption that selection into these two groups is random, conditional on observable characteristics measured at baseline, the difference in potential outcomes provides an unbiased measure of the causal average treatment effect of retirement on cognition. We discuss the plausibility of this assumption in the context of describing our econometric model.
Previous Literature

Three recent papers have begun to examine how the complexity of the work environment is related to cognitive change. The first, using data from the Swedish Adoption/Twin Study of Aging (Finkel, Andel, Gatz, & Pedersen, 2009) examined complexity of occupation on cognitive trajectory at retirement. This study found that individuals with occupations involving “high engagement with people” experienced greater improvement in verbal skills up until retirement, but experienced a faster rate of decline following retirement. The authors proposed that taking away work from one’s lifestyle as a key source of mental exercise, i.e., engaging with people, had a detrimental effect on cognitive aging.

A second study of US adults showed that those who engaged in more mentally demanding jobs had higher cognitive function prior to retirement, and experienced less decline in cognitive performance following retirement (Fisher et al., 2014). The authors proposed that these results might stem from individuals with more cognitively complex jobs accumulating greater cognitive resilience through their pre-retirement job from which to off-set the effects of cognitive aging following retirement.

Finally, a study using National Survey of Japanese Elderly longitudinal data (Kajitani, Sakata, & McKenzie, 2013) similarly found that men who have careers that require high mathematical development, reasoning development, and language development experience less decline in memory following retirement. They also observed that jobs high in physical engagement related to greater deterioration in memory loss after retirement.

In combination, these three studies offer compelling evidence that the characteristics of one’s work environment and associated lifestyle play critical roles in one’s cognitive functioning prior to and following the retirement transition. However, these studies did not take into consideration the alternative to retirement – the expected cognitive trajectory had those individuals continued working. Notably, some individuals may experience a hastening of age-related cognitive decline despite continued employment, which may be unrelated to a retirement transition per se and perhaps related instead to pre-retirement occupational factors. Other individuals alternatively may experience little or no decline with or without a retirement transition. In fact, our recent research shows that the effects of work-retirement patterns on cognitive performance are not universal. For some social groups, the retirement transition offers no better or worse effect on cognitive performance than does continuous full-time work (Carr et al., under review). As a result, the effect attributed to retirement in previous research may be related to other factors. The current study seeks to address this by examining the effect of retirement relative to not retiring on cognitive change for those with similar occupational characteristics.
Theoretical Framework

One reason working may offer a better cognitive trajectory relative to retirement is that working provides a more cognitively beneficial lifestyle (Rohwedder & Willis, 2010). In other words, when people retire and stop working, they stop “using it,” and subsequently “lose it,” or rather, they experience more rapid cognitive decline (Foster & Taylor, 1920). It is not necessarily just the capacity to learn that impacts cognitive performance, but the motivation to seek out cognitively engaging opportunities. Being removed from a complex environment, as occurs with retirement, may modify one’s cognitive trajectory because an individual is no longer required to engage in cognitively complex tasks. Some people do not seek out opportunities to maintain their cognitive function after they retire (Schooler, 1984, 1990; Schooler, Mulatu, & Oates, 2004).

One potential mechanism related to the beneficial effects of cognitively complex environments like work could be the building of cognitive capacity throughout one’s life span (even into later life). That is, spending many years in intellectually stimulating or mechanically complex environments – likely related to both educational attainment and occupation factors (Potter, Plassman, Helms, Foster, & Edwards, 2006) – leads to greater neuronal development, and that accumulation of excess neuronal resources, or cognitive reserve, may help people stave off the cognitive losses that come with aging (Fratiglioni & Wang, 2007).

Regardless of the specific mechanisms at play, an individual’s cognitive aging process appears to be influenced by a combination of mental stimulation across the life span (i.e., the tendency of those with greater cognitive function to pursue more complex jobs and activities leading to more significant accumulation of cognitive resources), and the individual and environmental factors that impact one’s cognitive engagement in later life (i.e., the extent to which an individual is capable and motivated to maintain cognitive function in spite of changes to the environment) (Salthouse, 2012; Salthouse, 2006; Salthouse, Atkinson, & Berish, 2003; Salthouse, Berish, & Miles, 2002). Thus, the potential relation between retirement and cognitive decline might be thought of as a response to the way pre-retirement cognitive engagement “habits” adapt to a non-work lifestyle and environment.

To understand this process, we rely on the Scaffolding Theory of Cognitive Engagement (STAC). According to STAC, brains respond to changes associated with aging through utilization of “scaffolding,” or the development of effective adaptive responses (Park & Reuter-Lorenz, 2009). They write:

Scaffolding is a normal process present across the lifespan that involves use and development of complementary, alternative neural circuits to achieve a particular cognitive goal. Scaffolding is protective of cognitive function in the aging brain, and available evidence suggests that the ability to use this mechanism is strengthened by cognitive engagement, exercise, and low levels of default network engagement.


It is plausible that certain job characteristics, particularly intellectual and mechanical tasks, shape one’s ability to cognitively adapt to age- and environment-related changes. The skills, abilities, and behaviors utilized while engaging in work-related tasks, or during certain job-related training, skills, and education, can be thought of as a form of “scaffolding” that can be honed during one’s career and tapped into during the post-work period. So-called cognitive maintenance following retirement (despite disengagement from work) could also be thought of as cognitive “resilience” because the effects of retirement on cognition are less than expected (Mukherjee et al., 2014). Alternatively, in some cases, a job may be cognitively stimulating enough to maintain cognitive function while working, but not offer sufficient cognitive scaffolding to adapt to the deficit of work-related stimulation in retirement, yielding significant cognitive loss. To that end, it is important to account for level of complexity and the effect on retirement by occupational characteristics between those who retire and those who continue to work when studying cognition as it relates to the retirement process.
Research Question and Hypotheses

Specifically, this research is designed to address the following research question: How do pre-retirement occupational characteristics (i.e., intellectual and mechanical) impact cognitive changes associated with retiring relative to staying engaged in full-time work? Based on empirical evidence and the STAC, we propose three hypotheses. First, we hypothesize that the relationship between retirement and cognitive decline is dependent on the cognitive stimulation of the pre-retirement job. Specifically, those with jobs that require more intellectual engagement will be more resilient and thus, experience less cognitive decline relative to those with jobs that require less intellectual engagement. However, those with jobs that require more mechanical engagement will be less resilient and thus, experience more cognitive decline relative to those with jobs that require less mechanical engagement. Second, we hypothesize that the effect of intellectual and mechanical complexity of work will be less significant for those who continue to engage in full-time work than for those who go on to retire. In other words, we expect that the work “lifestyle” will facilitate maintenance of cognitive performance when people retire, but the absence of a work lifestyle will increase the importance of non-work lifestyles in determining the impact of retirement on cognitive performance.


DESIGN AND METHOD

Our study uses 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 1992 to 2010. These data are ideal for this study because they offer the most comprehensive nationally representative panel data on US older adults available, including information about cognitive performance and work behaviors (Lachman & Weaver, 1997; RAND Center for the Study of Aging, 2008; Smith et al., 2012). For the current study, we include only individuals older than 50.


Selection of Full-Time Workers and Retirees

To test our hypotheses, we selected two samples: full-time workers and retirees. First, to evaluate the effect of pre-retirement job complexity on change in cognitive performance, we began by identifying full-time workers – i.e., those who worked 35 hours or more and self-identified as not retired. From this group, we identified two samples – those who transition from full-time work in wave t to retirement in wave t+1and those who stay working full-time in both waves. We exclude retirees who engage in paid work for two reasons. First, given the focus of this study on the lasting cognitive impact on departing from paid work, those engaging in paid work in retirement are still participating in a “work lifestyle.” While it may be helpful to assess the effect of variations in pathways to retirement on the cognitive decline trajectory, individuals whose labor force status was recorded as “retired” (even if they did engage in part-time paid work) were not consistently asked about their occupation, preventing us from taking into consideration how work tasks changed post-retirement. Additionally, recent research suggests that regardless of how retirement is defined, the relative effect of characteristics of pre-retirement work on post-retirement cognitive performance does not change (Kajitani et al., 2013). Thus, for this study, we choose a conservative definition for retirement, limiting our retiree sample to those individuals who transition from full-time work to complete retirement (i.e., for the first time while participating in the HRS, self-identifying as being retired and working 0 hours per week).

Second, in order to accurately measure cognitive changes in association with a potential retirement transition, we selected specific pre- and post-retirement cognitive performance measures. First, there is evidence that people may begin cognitively disengaging from work in preparation for retirement, and this may result in cognitive decline while working in the period leading up to retirement (Bonsang et al., 2012; Willis, 2013). Thus, to avoid this complication, our baseline cognitive performance occurs two waves prior to retirement, limiting our sample only to those working full-time for two consecutive waves prior to retirement. Second, the long-term adjustment to retirement, with regard to a shift in the cognitive performance trajectory, does not occur until at least one full year following retirement (Bonsang et al., 2012). Thus, to ensure that our post-retirement cognitive performance is observed with an appropriate lag, our post-retirement measure derives from cognitive status at the wave following reported retirement, limiting our sample to only those retirees who continuously remain fully retired in the wave following reported retirement.

Because persistent full time workers are not, by definition, observed making a retirement transition, we use the most recent four-wave period of consistent full-time work for the full-time working sample. For this group, baseline cognition is measured at Time 1, compared with cognitive performance in Time 4 of continuous full-time work.

Third, in order to minimize the potential endogenous effect of declining cognitive status accelerating the decision to retire and therefore, increasing the effect of retirement on change in cognitive performance, we only considered individuals with normal pre-retirement cognitive performance. Specifically, we excluded all individuals who had a cognitive score indicating cognitive impairment during either of the two waves of full-time work preceding potential retirement. Our final pooled sample of retirees included 721 individuals observed over four consecutive waves with complete data, two prior and two following potential retirement. Our total sample of full-time workers included 1,296 individuals. Figure 1 provides the breakdown of our identification of the final samples based on work-retirement patterns.
Measure of Cognitive Status

Cognitive performance is based on a 27-point test. This test derives from the Telephone Interview for Cognitive Status (TICS) (Brandt, Spencer, & Folstein, 1988), which has been validated for use as a screening instrument for cognitive performance (Plassman, Newman, Welsh, & Breitner, 1994; Welsh, Breitner, & Mgruder-Habib, 1993). The TICS is composed of measures of episodic memory (a 10-word immediate and delayed recall test (0 to 20 points)), working memory (a timed serial 7s test (0 to 5 points)), and processing speed (backwards-counting test (0 to 2 points)). The total score ranges from 0 to 27, with higher scores indicating better performance. These tests were administered every two years. 



Cognitive scores were standardized using the average score for HRS respondents ages 51-55: a mean of 17.05, standard deviation corrected for measurement error, equal to 2.46.2 Our outcome variable is the difference in the standardized cognitive score from Time 1 to Time 4. A one-unit change in cognitive score is the equivalent of 4.57 points. A positive score indicates improvement in cognitive performance from Time 1 to Time 4. Negative scores indicate decline. The TICS has validated cut-points differentiating normal cognitive functioning (≥12) from impairment (i.e., those with lower than 12 points) (Crimmins, Kim, Langa, & Weir, 2011; Fisher, Rodgers, & Weir, 2009; Langa et al., 2005).
Job Complexity

HRS respondents’ occupations at each wave of HRS (based on U.S. Census codes) were linked to the O*NET database (via a crosswalk that links U.S. Census codes with the Standard Occupation Classification (SOC) codes in the O*NET) to obtain external occupational-level ratings of job characteristics pertaining to occupation at each wave.3 The O*NET program is the primary source of occupational data in the United States. The O*NET database contains information on standardized occupation-specific characteristics and is publicly available. The O*NET-SOC taxonomy is a set of characteristics for a set of standardized occupations that correspond to the U.S. Census. Each occupation characteristic score is calculated based on a rating scale related to abilities (i.e., the expected abilities required in order to engage in a given job), activities (i.e., the expectation of participation in activities associated with a given job), and contexts (i.e., the situational aspects of day-to-day working associated with a given job). For example, the degree to which a job involves getting information is assessed, with a total score calculated on a range from 0 to 1 based on how often that particular job typically requires getting information.

A total of 36 job-related abilities, activities, and contexts were available for the standard occupations coded in the Health and Retirement Study. To identify meaningful job factors, we used exploratory factor analysis (see Appendix A for the full list). Excluding all items with an alpha score below 0.60, two factors emerged from the remaining 18 items. Using an iterative selection process, we excluded all variables that loaded on both factors, and then systematically removed items until we identified the fewest number of items with the highest alpha score. As shown in Table 1A, the first factor, which we describe as the “intellectual” factor, includes five items: (1) making decisions and solving problems; (2) thinking creatively; (3) coaching and developing others; (4) frequency of decision-making; and (5) freedom to make decisions. The second factor, which we describe as the “mechanical” factor, includes four items: (1) inspecting equipment, structures or material; (2) handling and moving objects; (3) controlling machines and processes; and (4) operating vehicles, mechanized devices or equipment (see Table 1B). The Chronbach’s alpha scores for these factors are 0.952 and 0.958 respectively. The scores for the intellectual measure ranged from 2.417 to 3.724, and the mechanical measure ranged from 0.916 to 2.773.

To get a general sense of how the intellectual and mechanical tertials relate to broader occupational categories, we identified all major occupation categories (an HRS variable that reflects the broad Census categorization for major occupation types) that fell into each tertial. Table 2 shows the breakdown of occupational types, demonstrating that the highest level of the intellectual variable includes primarily individuals in managerial positions (e.g., legislators, CEOs, marketing managers, administrators and officials in the public administration sector, and accountants and auditors). The middle group is composed primarily of individuals with professional specialty positions (e.g., social workers, statisticians, dentists, dieticians and teachers), and secondarily in personal services jobs (e.g., supervisors of welfare service aides, hairdressers, or child care workers), mechanics/repair work, construction, and precision production jobs (e.g., machinists). The lowest cognitive grouping is composed primarily of individuals in sales (e.g., insurance sales occupations and apparel sales clerks) and clerical jobs (e.g., secretaries and typists), and secondarily personal services and operator jobs (e.g., printing machine operators, textile sewing machine operators).

Regarding the mechanical variable, the highest mechanical group is composed primarily of individuals who work in mechanical, construction, precision production, and operator jobs. The moderate mechanical group is composed of individuals primarily in sales positions and health and personal services jobs. The lowest mechanical group is composed of individuals who are in managerial, professional specialty, and clerical positions.
Covariates

Demographic covariates included gender, race (an indicator of whether an individual is non-Hispanic white (reference group), non-Hispanic black, Hispanic, or another race), and age (a continuous measure at Time 3 because our selection required individuals to be at least 50 at potential retirement), education (in years). Because changes in health status could initiate a retirement transition or a change in cognitive status, we include measures that adjust for potential pre-retirement health decline: (a) raw cognitive score at Time 2, a continuous measure of frailty at Time 2; and (b) to adjust for the potential causal effect of declining health as an impetus for the retirement transition, we also include a measure for change in self-rated health observed at Time 3 (relative to Time 2). Frailty (at 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 (Radloff, 1977), obesity (i.e., body mass index of 30 or greater), and self-rated health (a five point likert item with higher values indicating better health.

Given the significant relationship between physical engagement behaviors and cognitive performance (Ahlskog, Geda, Graff-Radford, & Petersen, 2011; Hindin & Zelinski, 2012; Langlois et al., 2013), we include a dichotomous measure of frequency of moderate physical engagement (1 = every day, 2 = > once per week, 3 = once per week, 4 = one to three times per week, 5 = never). Unfortunately, this measure was only available consistently at Time 4 for the entire pooled sample (i.e., HRS included a consistent measure of moderate engagement beginning in 2002).

Econometric Model

Our goal is to estimate the effect on the trajectory of cognitive performance of a worker’s decision to choose full retirement versus continuing full-time work among workers whose longest jobs varied in intellectual complexity or, alternatively, in mechanical complexity. Following Rohwedder and Willis (2010), we call this the “mental retirement effect” or the MRE. Ideally, our estimate of the MRE could be interpreted as causal, representing the potential loss (or gain) in cognitive performance that a given worker could expect if he or she were to fully retire rather than continue working full time. Taken literally, it is impossible to achieve this ideal even in a hypothetical randomized controlled trial in which young adults are randomly assigned to occupations and subsequently assigned to full retirement when they reach a randomly chosen age sometime after, say, their mid-50s. The reason, of course, is that people only live their life once and, consequently, one cannot estimate the counterfactual MRE at the individual level.

Using the language of the “Rubin causal model” (Holland 1986),the best we could achieve in this hypothetical experiment is to estimate the mean “potential outcomes” or POMs of people assigned to retirement or continued work for a given level of complexity, j, of the occupation to which they been assigned. The mental retirement effect is then just the difference between these POMs:


  1. ,

where corresponds to the intellectual complexity or, alternatively, the mechanical complexity of the occupation based on O*Net Data.

Obviously, this hypothetical experiment is impossible to conduct for practical and ethical reasons. In addition, it would be economically inefficient since many people would be assigned to occupations for which they lack the requisite abilities, education or interests and be assigned to retirement when they prefer to continue working or conversely. (Think of a physics department in which the janitor is assigned to give course lectures and Einstein is assigned to empty the waste baskets.) Self-selection in competitive labor markets tends to create the most productive matches between the skills, capabilities and preferences of individuals and the employers’ demands for workers to conduct particular tasks (Roy, 1951; Willis and Rosen, 1979). From this point of view, the most relevant MREs to estimate are for people in the occupations they have actually chosen. Indeed, the results from the hypothetical experiment would likely be highly misleading because many people would be assigned to tasks they dislike or cannot perform—a mismatch that the designers of the randomization may observe imperfectly, if at all—possibly causing changes in cognitive performance that do not occur among well-matched workers.

While self-selection helps us choose the most relevant comparisons to make in judging how the effect of retirement on the trajectory of cognition is influenced by occupational complexity, it creates important challenges to our ability to obtain unbiased measures of . Self-selection throughout life in the level and type of education, occupational choice, labor supply and in other areas of life such as marriage, fertility, residential location, etc. create heterogeneity among individuals that is likely to be correlated with both the level of cognitive performance when they enter the HRS and with their subsequent retirement decision. Some of this heterogeneity can be controlled using observed covariates in the data, but we also need to worry about unobserved heterogeneity.

Our econometric model attempts to address these issues. As shown in Figure 1, the sample we study consists of N persons who are working full-time at Time 1 and Time 2. Of these, R individuals are completely retired at Time 3 and Time 4 and W people continue working full time in Time 3 and Time 4. Individuals with other patterns of work and retirement are excluded.

Denoting the parameters that pertain to members of the two groups by the superscripts R and W, a person’s cognitive score at Time 1, is assumed to be a linear function of a set of observed variables, ; a person-specific fixed effect, , that captures unobserved heterogeneity and an iid measurement error term, , with mean zero and variance , that captures the difference between cognitive scores on the HRS and the latent variable, “true ability.”

Thus, the observed test scores at Time 1 of people in groups R and W, conditional on the complexity of their longest occupation, vary according to both observable variables,



; unobserved factors embodied in and measurement error:

(2) , and

(3) .

The coefficients and in (2) and (3) capture the potentially different effects of observable variables that characterize a person’s life history depending on whether the person belongs to the group of people who are fully retired by Time 3 or to the group who continue working through Time 3 and Time 4. Also note that the average cognitive score at Time 1 of persons in groups R and W may differ because “selection on observables” causes the distribution of to differ between those who retire and those who do not. Likewise, it is possible that “selection on unobservables” causes the mean value of to differ between persons in the two groups. In analogy with equations (2) and (3), the levels of cognition at Time 4 in the R and W groups are given by:

(4)

(5) .

We are interested in estimating the effect of retirement on cognition for each complexity category by comparing the change in cognition between Time 1 and Time 4 for those who retire with the change in cognition for those who continue working by occupational complexity. Within the potential outcomes framework, the potential outcome for individuals in Group R is obtained by subtracting equation (2) from equation (4) and taking the expected value to obtain:

(6) ..

Similarly, the POM for Group W is obtained by subtracting equation (3) from equation (5) and taking expected values to obtain:

(7)

where and . Note that unobserved heterogeneity, , is eliminated by these subtractions so that person-specific POMs are only subject to selection on observables, , and that measurement error has no effect; i.e., .

If the selection of individuals into the R and W groups and into occupations with low, medium and high levels of complexity were random, we could simply apply the formula in equation (1) to obtain an unbiased estimate of the mental retirement effect for persons in occupations of each degree of complexity by calculating. There are two major threats that we need to address before we can claim to have estimated the causal effect of retirement: reverse causation and self-selection on observables.

Reverse causation would occur if cognitive change is a significant cause of retirement. Worry about reverse causation has motivated the use of IV methods in papers such as Rohwedder and Willis (2010), Bonsang, et al. (2012) and others. As described earlier, we have tried to guard against reverse cognition by eliminating people with low cognitive scores from our analytic sample and by including the cognitive score just before retirement at Time 2 as a control variable. In addition, we note that there is little evidence in the literature to suggest that cognitive decline is an important cause of retirement. For example, Rohwedder and Willis (2010, Figure 5) display the OLS regression line between cognition and retirement in their cross-national sample. It has nearly the same slope as the IV relationship that they estimate. We believe that it is reasonable to assume that reverse causation is not present in our analysis.

The existence of self-selection on observables is much more plausible for occupational choice and retirement decisions. For example, occupational choice is strongly related to education and retirement both directly and indirectly through economic and health status. Suppose we estimate the mean cognitive change, , from data on people in group R whose longest job was in a highly complex occupation and imagine that education is the only covariate in. In order to calculate the causal effect of retirement on cognitive change, we would like to compare the mean cognition of people in group R with people in group W with the same level of education in order to isolate the causal effect of retirement on cognitive change equal to where ed is the common mean education and the sign and magnitude of determines the sign and magnitude of . Since it is likely that continuing workers, on average, have more education than those who retire, failure to correct for selection on education will lead to an overstatement of the mental retirement effect.

Fortunately, there are a number of approaches in the statistical and econometric literature on treatment effects and the related literature on missing data that allow us to correct for selection on observables. If, for the moment, we ignore self-selection into occupations, we need only deal with the binary decision to fully retire or continue working full time. The method we use combines two popular approaches: regression adjustment and inverse probability weighting (Curtis, Hammill, Eisenstein, Kramer & Anstrom, 2007). Regression adjustment of sample means uses covariates that predict selection into the R and W groups to make the distribution of covariates in the two groups comparable, in analogy to our univariate example of education. Inverse probability weighting is motivated by the recognition that data on potential outcomes of Group R, had they continued working full time, is missing and, conversely, potential outcomes if they had retired are missing for members of Group W. Both approaches make use of a propensity model that estimates the probability that an individual is a member of Group R or Group W (Abadie & Imbens, 2012). It been shown that this approach has a “doubly robust” property such that unbiased estimates of treatment effects can be obtained even if either the POM model or the propensity model (but not both) is misspecified (Wooldridge, 2007).

Recently, (Cattaneo 2010) has extended this approach to allow estimation of multi-valued treatment effects using semi-parametric methods and this approach is implemented in parametric form in the ipwra option of the teffects command in Stata 14 (StataCorp, 2013).4 Using this command, we estimate six POM equations described in equations (6) and (7) in order to estimate the three Mental Retirement Effects for persons in low, medium and high complexity jobs described in Equation (1). The probabilities that individuals self-select into one of these six states are estimated using a multivariate logit propensity model. This model is estimated for two alternative definitions of occupational complexity: intellectual complexity and mechanical complexity.

RESULTS

We now turn to our estimates of the potential outcome means, and



, and the mental retirement effects,, measured by the average treatment effect on the change in cognition effect over a six year period beginning two waves before full retirement and ending one wave after retirement. These effects are estimated for three levels of intellectual complexity of the individual’s job at baseline, reported in Table 4, and by three levels of mechanical complexity of the baseline job reported in Table 5. We also report the results of the auxiliary potential outcome and propensity equations in Appendix Tables A1 and A2.

We hypothesized that those with the most intellectually complex jobs will experience less cognitive decline with retirement relative to those with the least intellectually complex jobs. The results presented in Table 4a strongly support this hypothesis. The first row of Table 4a shows that the average cognitive decline (POM) for a person who retires is much greater for persons in low complexity jobs than in high complexity jobs (-0.742 compared to -0.256) with the decline for moderate complexity in between (-0.449). The second row shows that cognitive decline shows little variation with intellectual complexity for those who continue working (-0.282, -0.216, -0.300 respectively for low, moderate, and high complexity). The difference between the POMs for the retired and full time workers in the third row provides causal estimates of the effect of retirement on cognition. Workers who retire from the least complex jobs suffer a highly significant decrease of -0.460 standard deviations compared to what they would expect had they continued full time work. Retirement from a job with moderate intellectual complexity causes a smaller marginally significant decline of -0.233 standard deviations while there is no effect on cognition of retirement from a highly intellectually complex job. Figure 2 presents these results in graphical form with effect sizes corrected for test-retest measurement error as described in footnote 2.

Corresponding results for jobs classified by mechanical complexity are presented in Table 5 and in graphical form in Figure 3. The causal effects of retirement are negative, but smaller and less significant for job complexity on this dimension than for intellectual complexity of the job. However, as is clear the Figure 3, in contrast to the case of intellectual complexity, the negative effect of retirement on cognition is greatest for retirement from jobs with high mechanical complexity, and least for those with moderate mechanical complexity.

To further clarify the observed cognitive changes with respect to level of intellectual and mechanical complexity, Figure 4 depicts the percentile change at Time 4 relative to Time 1. Relative to the normally distributed cognitive performance of healthy 50-54 year olds in the population (for which our data are normed), cognitive performance declines associated with aging occur over the 6-year time frame, but in differing amounts. Assuming that cognition is normally distributed, a worker who has a low intellectually complex job at Time 1 and 2 was at the 50th percentile of the baseline distribution at Time 1, would be expected to fall to the 23rd percentile six years later if they retired, but only the 39th percentile if they continued working, whereas a worker with a high intellectually complex job would only decline to the 40th percentile if they retired and would decline at a slightly higher rate (down to about the 30th percentile) if they continued working full-time. For those with high mechanically complex jobs, retirement would predict a decline to the 22nd percentile versus a decline only to the 38th percentile with continuous full-time work. For those with low mechanically complex jobs, retirement would predict a decline to the 35th percentile, and continuous full-time work the 43rd percentile.

DISCUSSION AND CONCLUSION

Fluid intelligence—the capacity to think and reason—as well as the cognitive mechanics that underpin this capacity including working and episodic memory, processing speed and other components of intelligence or executive function all follow a declining path from early adulthood until death in a process called “normal cognitive aging.” At advanced ages, dementing diseases such as Alzheimer’s disease begin to disable a significant fraction of the elderly, but most people will die with normal cognition and, conversely, very few older workers under the age of 75 have experienced the onset of dementia. Recently, cognitive psychologists and neuroscientists have begun to ask why it is that many older people can function so effectively at work, in the family and in society despite the dramatic declines in measures of fluid intelligence they have suffered since they were young. Indeed, they note, the puzzle is deepened when one recognizes that most of the leaders in government, business, education and the military tend to be drawn from the older part of the working population in almost every society.

Cognitive scientists believe that a clue to the resolution of this puzzle is to be found in the observation that, unlike fluid intelligence, crystallized intelligence—the accumulation of knowledge and wisdom—tends to continue to increase over the life cycle. To an economist, crystallized intelligence is essentially the same thing that economists call human capital (McArdle and Willis, 2011). Human capital is durable knowledge that is acquired through investment in education, on-the-job training and learning-by-doing by combining effort, knowledge and cognitive ability. Because it is durable, the stock of human capital tends to grow over the life course, although toward the end of working life it may decline as the rate of investment falls below the rate of depreciation (forgetting) and the rate of obsolescence (failure to replace obsolete knowledge with current best practice).

Economists emphasize both pecuniary and non-pecuniary incentives to invest in the accumulation of human capital. They also note that the productivity of human capital tends to become more specialized as the person matures because of the incentive to acquire skills necessary to conduct the tasks required by one’s occupation. It also seems likely, especially in the knowledge economy, that an individual’s current productivity depends on his or her ability to carry out a complex novel task by accessing relevant pieces of knowledge—resident in their brain, books, the web or another person’s brain—and use their fluid intelligence to assemble those pieces in ways that are relevant to the execution of the task. In short, although fluid intelligence and crystallized intelligence are distinct components of intelligence, they complement each other in helping individuals accomplish valued goals during their hours of work and also in their hours outside of work in dealing with their family and friends, enjoying hobbies, leisure activities and so on.

The scaffolding theory of Park & Reuter-Lorenz (2009) provides intriguing links between the theory of fluid and crystallized intelligence in psychology, the analysis of brain structure and function in neuroscience and the theory of human capital in economics. It suggests that when a person solves a novel task, new neural circuits are created in the brain. If a circuit continues to be accessed, it will become a long-lasting part of the person’s mental capacity, but if it is not used it will be destroyed by the ongoing pruning process by which the brain maintains an efficient allocation of its scarce resources. Thus, scaffolding theory provides an underlying mechanism for the “use it or lose it” hypothesis that, in turn, inspired the mental retirement hypothesis of Rohwedder and Willis (2010) that we investigate in this paper.

In addition, scaffolding theory provides clues about the neurological and psychological mechanisms that underlie the formation of forms of crystallized intelligence embodied in the specialized skills and knowledge that a worker brings to the labor market and that constitute his or her stock of human capital. The theory implies that people who perform complex and novel tasks lay down more circuitry than people who do simpler, repetitive tasks. Importantly, we conjecture that the combinatorics of bringing together different relevant pieces of knowledge involved the tasks that a worker performs during a career in a complex occupation may spill over into increased capacity to engage in and enjoy a wide variety of activities in non-work activities that can be pursued during retirement. Conversely, engagement in less complex, repetitive activities in low skilled jobs is likely to have been learned early in the career and have little transferability to non-work activities. The structure of working life may help with the maintenance of cognitive capabilities, but workers in low skilled jobs are likely to have developed relatively little scaffolding to buffer the effects of leaving work unless they have developed independent interests in cognitively challenging activities that they pursue during retirement.

Our empirical analysis of the differential effects of full time retirement on the trajectory of cognition is quite consistent with previous empirical studies that suggest that pre-retirement job characteristics shape the cognitive performance trajectory following retirement and with the theoretical predictions of scaffolding theory, especially, as it is augmented by insights from theory and evidence from the human capital literature in economics. We find clear evidence of differential “mental retirement effects” (MREs) when intellectual complexity is used as the criterion for occupational complexity and the change in cognition is measured over a six year time span beginning two HRS waves before retirement and ending one wave after retirement. We find that full time workers in low complexity jobs have a predicted decline of cognition for those who fully retire that is three times larger than it is for those who continue working full-time. We also estimate a smaller, marginally significant MRE effect for workers in moderately complex jobs, with a cognitive decline that is twice as great if they retire than if they continue working. However, among workers in highly skilled jobs, retirement has no effect on cognition. (Indeed, retirees in this group have a very small and insignificant cognitive advantage relative to full-time workers.)

Interestingly, the differentials in MRSs by intellectual complexity occur entirely because of differential negative impacts of retirement on cognition. Among those who continued working full-time, decline is essentially invariant across the different categories of complexity. This pattern suggests that the productivity of human capital accumulated during the working career is more transferable to non-work settings, the greater the intellectual complexity of the job. It also suggests that reduced mental effort in maintaining one’s skills as the date of retirement approaches (what Rohwedder and Willis (2010) called the “on-the-job retirement effect”) is not responsible for the MRE effect. If it were, we would expect that those with the most skills in the most complex occupations would have the greatest scope to decrease their human capital investment in anticipation of retirement.

We used an alternative classification of occupational complexity based on mechanical complexity rather than intellectual complexity. By construction, the two classifications are essentially orthogonal in the sense that they involve non-overlapping O*NET characteristics. On this criterion, we find the opposite pattern of differential by complexity than we found for the intellectual complexity classification. The largest cognitive declines occur among workers in the most complex occupations and the smallest occur for those in the least complex occupations. In addition, we see greater declines with increased complexity for both those who continue working and for those who retire, although the retirement effects are much larger.

We conjecture that these differences in the patterns of cognitive decline for mechanical complexity compared to intellectual complexity might also be explained by scaffolding theory. It seems plausible that workers in jobs involving complex machinery or equipment develop knowledge that is not very transferable to the non-work environment they encounter when they retire. In addition, it seems likely that workers in such jobs face greater dangers that their skills will become obsolete due to changing technology than workers in intellectually complex jobs and, therefore, have an incentive to reduce their investment effort in maintaining their skills as the date of retirement draws near.

The validity of our empirical findings and their interpretation as causal effects, as we explain at length in the paper, depend on the assumption that differences between people who self-select into different occupations and make different decisions about work and retirement can be adequately controlled with observable variables. Our focus on change in cognition rather than level of cognition strengthens the plausibility of the assumption of selection observables—also known as conditional ignorability or unconfoundedness—by eliminating person-specific fixed effects. But ultimately, the plausibility of our findings depends on whether similar effects are found using different data and different methods. In this respect, we see much scope for using data generated by psychologists and neuroscientists as well as additional analysis of survey data such as HRS. There is even potential for combining the different types of data as discussed in “What is a representative brain?” (Falk, et al., 2013). While this paper addresses the complexity and characteristics of the work environment, we also believe that further study of the non-work environment and its impact on cognition should be a high priority for future research.
Funding:

This work was supported by funding from the Alfred P. Sloan Foundation.

Table 1A: Intellectual Complexity Items




Item-Test Correlation

Alpha

Making Decisions and Solving Problems

0.977

0.925

Thinking Creatively

0.957

0.937

Coaching and Developing Others

0.957

0.931

Frequency of Decision-Making

0.880

0.957

Freedom To Make Decisions

0.914

0.947

Test scale




0.952



Download 464.26 Kb.

Share with your friends:
  1   2   3   4   5   6




The database is protected by copyright ©ininet.org 2020
send message

    Main page