Cognitive Ability, Personality, and Pathways to Retirement: An Exploratory Study

Transcription

1 Work, Aging and Retirement, 2018, Vol. 4, No. 1, pp doi: /workar/wax030 Empirical Article Cognitive Ability, Personality, and Pathways to Retirement: An Exploratory Study Péter Hudomiet 1, Andrew M. Parker 1, and Susann Rohwedder 1,2 1. RAND Corporation 2. NETSPAR ABSTRACT This article describes an exploratory study that investigates the extent to which two sets of psychological factors, fluid cognitive ability and personality traits, predict late-in-life work and retirement outcomes. Using longitudinal data from the Health and Retirement Study, we first provide a detailed characterization of within-subject work-toretirement pathways, spanning 14 years of data for each individual, and identify the most frequent pathway classes. We found that only 37% of workers followed the standard pattern of retiring completely from a full-time job. We then examined how cognitive ability and personality traits predict these work-to-retirement pathways. We found that individuals with better cognitive ability work longer, both in full- and in part-time jobs, and extraversion is a strong predictor of working longer, mainly in part-time jobs. These results are robust to the inclusion of many covariates, including demographics, health, socioeconomic status, and labor market variables. Although the observed patterns match individuals retirement expectations to some extent, there also seems to be evidence of some surprise. Practical implications and suggestions for future research are discussed. Psychological factors such as cognitive ability and personality traits play important roles in individual decision making, such as choosing healthy lifestyles or avoiding risks (Hampson, Goldberg, Vogt & Dubanoski, 2007; Smith, 2006). Yet, these factors have received relatively little attention in studies of the complex intertemporal decision problems involved in the decisions of how and when to move toward retirement (Barnes-Farrell, 2003; Beehr & Bennett, 2015; Feldman & Beehr, 2011; Fisher, Chaffee, & Sonnega, 2016). Instead, much of the existing literature has focused on the importance of economic incentives, health status, and socioeconomic factors (e.g., Bound, Cullen, Nichols, & Schmidt, 2004; Bound, Stinebrickner, & Waidmann, 2010; French & Jones, 2011; Gruber & Wise, 2004; Gustman & Steinmeier, 2005; Maestas, Mullen, & Strand, 2013). In this article, we investigate empirically how cognitive ability and the Big 5 personality traits are related to individuals retirement pathways the longitudinal pattern of older adults employment, often spanning from full-time work to fully or partially exiting the workforce after standard socioeconomic and health controls are applied. This is an important issue for several reasons. For example, there is mounting financial pressure for people to work longer with the responsibility to provide for spending at older ages having shifted increasingly to the individual. At the same time, the financial pressures on entitlement programs like Social Security and Medicare would be at least somewhat relieved if people worked longer. Therefore, this paper addresses issues relevant to policymakers, who have a vested interest in identifying how to entice people to work longer. Further, studies of economic decision making have recognized that the focus on economic incentives, health, and demographics leaves considerable variation across individuals unexplained. Here, we suggest that psychological factors (including cognitive ability and personality traits) may account for at least some of that unexplained variation in work status in later years. Understanding whether and how psychological variables such as cognitive ability and personality are related to retirement pathways could help policymakers more effectively encourage individuals to work longer by promoting policies that take into account the considerable differences among workers. Such policies could enhance the fit between older workers preferences and the available job opportunities. Similarly, employers may also benefit from better understanding how the personalities and abilities of their employees are related to when and how people retire. Thus, the first step, which is the focus of this paper, is to investigate the extent to which cognitive ability and personality traits are related to various work-to-retirement pathways. THE ROLE OF COGNITIVE ABILITY AND PERSONALITY IN SHAPING RETIREMENT PATHWAYS Research on economic decision making has increasingly acknowledged that factors other than economic incentives and health influence The Author(s) Published by Oxford University Press. For permissions please journals.permissions@oup.com Correspondence concerning this article should be addressed to Susann Rohwedder, RAND Corporation, 1776 Main Street, Santa Monica, CA susannr@rand.org Decision Editor: Gwenith G. Fisher, PhD 52

2 Cognitive Ability, Personality, and Paths to Retirement 53 economic decision making, either in their own right or interacting with economic incentives and health effects (Borghans, Duckworth, Heckman & ter Weel, 2008; Heckman, Stixrud & Urzua, 2006; Knoll, 2011). Although past research has identified many psychological variables that may relate to retirement behavior and outcomes (for a recent review, see Fisher et al., 2016), we focus on personality traits and fluid cognitive abilities. Personality can be defined as a pattern of thoughts, feelings, and behaviors that are largely stable over time and situations (Borghans et al., 2008; Hurd, Duckworth, Rohwedder, & Weir, 2012). Cognitive abilities, which collectively include how people learn, think, reason, and make decisions can be divided into two general types of abilities: fluid and crystallized. Fluid cognitive abilities include reasoning or thinking, and the ability of a person to solve problems, whereas crystallized abilities consist of knowledge and information acquired throughout one s life (Horn & Cattell, 1967). Prior research in cognitive and lifespan developmental psychology has found that fluid cognitive abilities tend to decline with age (Salthouse, 2012; Schaie, 1994); Fluid cognitive abilities, for example, are needed to process complex choices (Henninger, Madden, & Huettel, 2010; McArdle, Ferrer-Caja, Hamagami, & Woodcock, 2002; Salthouse, 1990) and enhance decision-making quality (Bruine de Bruin, Parker, & Fischhoff, 2007; Del Missier, Mäntylä, & Bruine de Bruin, 2012; Del Missier et al., 2013; Peters et al., 2006; Toplak, West, & Stanovich, 2014). Such abilities may be especially important in diverse, intertemporal domains such as retirement (Park, 1999). Theoretical Background Multiple theories in the retirement literature suggest that personality and cognitive ability may be important individual difference characteristics that impact when and how people retire (Fisher et al., 2016), although these specific psychological variables have not received a great deal of attention in empirical investigations. Person-environment (P-E) fit theory (Edwards, 1996), in particular, provides important insight into decisions of when to move on from a career job (Feldman & Beehr, 2011; Wang & Shultz, 2010). Person-environment fit (P-E) theory suggests that when there is a positive match between the worker (such as knowledge, skills, abilities, or other characteristics) and their environment (the job or organization in which they work), the worker is more likely to continue working in that job and they will be less likely to retire (Feldman & Beehr, 2011). Personality and cognitive ability are two such characteristics that may be important for determining P-E fit. Changing (declining) cognitive abilities could decrease fit between an individual s abilities and the demands of a job, encouraging that individual to transition out of that job or leave the workforce entirely. Similarly, if and when workers productivity declines, employers may also consider letting these workers go or reassigning them to other positions (Kanfer & Ackerman, 2004). From an economic perspective, fluid cognitive abilities can affect retirement pathways through both labor demand and labor supply channels. Workers with greater cognitive abilities may be able to retain and use their productive skills for a longer time (especially if the job requires high cognitive ability), potentially increasing both their supply of labor as the mental cost of working may be lower and the demand for their labor as employers may be more willing to continue employing them. Similarly, workers with greater cognitive abilities may find it easier to learn new skills and find alternative work arrangements that suit their changing preferences, such as part-time and less demanding jobs. Personality traits have also been shown to predict a wide range of outcomes, including risky activities and health (Hampson et al., 2007; Smith, 2006), lifetime wealth and earnings (Duckworth, Weir, Tsukayama & Kwok, 2012), and financial preparation for retirement (Hurd et al., 2012). A small, but growing literature on the relationship between personality traits and some retirement outcomes (e.g., Angrisani, Hurd, Meijer, Parker, & Rohwedder, 2013; Blekesaune & Skirbekk, 2012; Löckenhoff, Terracciano, & Costa, 2009; McGonagle, Fisher, Barnes-Farrell & Grosch, 2015) found weak and inconsistent relationships between personality and retirement. However, none of these studies considered the entire late-in-life work-to-retirement pathways of workers, which is the focus of our study. Again considering person-environmental fit (Edwards, 1996), life stage and personality might combine to influence the fit between what an individual desires and what a job environment supplies, and consequently influence transition decisions (Wang & Shultz, 2010). We might expect, for example, that less extroverted individuals have less motivation to stay in a job for social reasons, compared to more extroverted individuals. Those with greater openness to experience may seek new opportunities trying their hand at a new career or spending time traveling and those higher in agreeableness may wish to spend more time with family. If this is the case, it should show in observable relationships between detailed retirement pathways, cognitive ability, and personality. Theories regarding retirement adjustment may provide another perspective on how cognitive ability and personality influence retirement pathways. Wang, Henkens, and van Solinge (2011) reviewed the literature on retirement adjustment and suggested a resource-based perspective, wherein diverse resources (e.g., cognitive, motivational, financial, social) facilitate or inhibit psychological comfort (i.e., adjustment) with the current stage of retirement. Considered broadly, reduced cognitive resources, social fit to the work environment, or increased motivation to spend more time with family could reduce comfort with the current condition, promoting change. This change could take the form, for example, of moving from full-time work toward retirement, gradually transitioning out of work, or returning to work. Cognitive ability plays a critical resource role in this perspective, as reduced cognitive capacity could increase discomfort within cognitively demanding work environments, and leaving such an environment could ease adjustment to retirement. Alternatively, greater cognitive abilities could make individuals less likely to work longer, if they are better prepared financially for retirement than workers with lower abilities. Personality, on the other hand, has been understudied with respect to retirement transitions, even though personality variables predict adjustment quality for other types of transitions (Wang & Shultz, 2010). It is conceivable that people with different personalities differ in how they adjust to retirement, a period when social contacts and a sense of purpose are no longer the byproduct of going to work every day. For example, an agreeable person may enjoy and adjust quickly to spending more time with family, whereas an extrovert may get bored and seek to return to work. Other psychological perspectives that could also be relevant include the role of motivation, goals, needs, and values (e.g., Kanfer, Beier, & Ackerman, 2013; Steger & Dik, 2009), social engagement

3 54 P. Hudomiet et al. (e.g., Zhan, Wang, & Shi, 2015), and retirement planning (e.g., Petkoska & Earl, 2009). These perspectives, of course, are not mutually exclusive of those discussed above, but are mentioned as a means of motivating the current interest in how cognitive ability and personality may differentiate retirement pathways. These theoretical perspectives suggest a complex set of roles that psychological factors, such as cognitive ability and personality, can play in the pathways to retirement. Although these are some possible examples, we undertake this investigation in an exploratory manner and do not formulate specific hypotheses about the relationship between personality characteristics and paths from work to retirement. Our exploratory approach, as described in our research questions, is an initial step to understanding their influence, while taking advantage of a unique data source. TRANSITION PATHWAYS FROM WORK TO RETIREMENT In this study, we pay particular attention to the fact that retirement can take many different forms (e.g., full-time to no work, gradual reduction of hours spanning many years, retirement followed by returning to work or multiple transitions) and that individuals retire at different ages. Recent literature finds that nontraditional retirement processes are very common in the United States (Beehr & Bennett, 2015; Kantarci & van Soest, 2008; Maestas, 2010), and it is plausible that personality traits or cognitive abilities affect these pathways in different ways. Because of the recent rise in the share of individuals working in alternative work arrangements (Katz & Krueger, 2016) such as being on-demand drivers (e.g., Lyft, Uber, etc.), it is conceivable that the role of nontraditional retirement pathways will increase in the future. Moreover, because many older workers have a strong preference for flexible working conditions (Maestas, Mullen, Powell, von Wachter, & Wenger, 2017), it is interesting and important to study workers who experienced flexibility at the end of their careers (at least in number of work hours) and how flexibility relates to cognitive abilities, personalities, and other variables. In the U.S. workforce, the retirement process usually starts at about age 55, when almost nobody is retired yet, and spans into the late 60s or early 70s, by when most workers are retired. Therefore, we use longitudinal data that allow us to follow each individual from age 56 to age 70 to examine work and retirement status over a 14-year period. From detailed study of these individual-level work patterns, we identify the eight most common distinct retirement pathways. Other longitudinal studies of labor market transitions around retirement tend to cover anywhere between 1 to 4 years. Our approach complements these transition-based approaches and has the advantage of capturing much more than just a few years of each individual s retirement process. We use data from the Health and Retirement Study (HRS), which is an ongoing panel study of the 51+ year old U.S. population initiated in The relatively large sample size of the HRS and the fact that it is a long panel (we consider up to 14 years of data for each individual) provides an almost unprecedented opportunity to study detailed retirement pathways in the United States. Once we derived the work-to-retirement pathways for individuals and group them, we relate the retirement pathways to fluid cognitive ability and the Big 5 personality traits. Because cognitive ability and personality traits are largely stable characteristics of individuals (at least in rank order, see our discussion in the data section), in this study we only consider the between-person variation in these variables and we do not analyze how within-person changes (e.g., cognitive decline or changes in personality) affect the work-to-retirement pathways. As we discuss later, future research should investigate how within-person changes in cognitive ability and personality or other characteristics relate to work and retirement behavior over time. Beyond the detailed work-to-retirement pathways, we also consider simpler measures of retirement outcomes, which are more commonly used in the economics literature, such as working past the early or normal retirement ages (e.g., Gustman & Steinmeier, 2005; Nyce, Schieber, Shoven, Slavov, & Wise, 2013). These measures capture elements of the timing of retirement. They also provide the opportunity to check the robustness of our findings to alternative specifications and to have specifications of retirement outcomes that have a direct correspondence to the retirement expectations elicited in the same survey. We use the retirement expectations to examine whether the observed retirement outcomes came as a surprise to individuals, which would indicate that unforeseen changes played a role. In summary this study addresses the following three exploratory research questions: 1. What distinct classes of work and retirement pathways can be identified and how frequently do we observe each class in a cohort of workers? These analyses leverage 20 years of data from the nationally representative Health and Retirement Study (HRS) obtained between 1992 and 2012 and, within that, 14-years of longitudinal data for each respondent to assign them to one of eight distinct retirement pathway categories. How do these pathways relate to simpler measures of retirement outcomes? 2. To what extent and how do fluid cognitive ability and the Big 5 personality traits predict membership in each retirement pathway class, while also accounting for more traditional predictors, including health, socio-economic status, job characteristics, and demographics? 3. To what extent were the observed retirement outcomes expected as opposed to unexpected by individuals with certain fluid cognitive ability or personality traits? METHOD Participants and Data This section summarizes the creation of our analytic data set. The Supplementary Appendix provides further details. The primary U.S. data for studying late-in-life work and retirement comes from the HRS, a longitudinal study of the U.S. population over age 50. The HRS is produced and distributed by the University of Michigan with funding from the National Institute on Aging (grant number NIA U01AG009740). Since 1992, respondents have been interviewed every 2 years, with refresher samples (or cohorts) of year-olds added every 6 years. As a result, work and retirement transitions even spanning several years are well-recorded. Many studies based on the HRS have examined the importance of economic incentives related to pensions and health insurance for retirement, along with the role of shocks in health or socioeconomic status, such as Bound et al. (2010), French and Jones (2011), Gustman and Steinmeier (2005), Kapur and Rogowski (2011), Karoly and Rogowski (2000), Mermin, Johnson, and Murphy (2007), and Sevak, Weir, and Willis (2003). We used the RAND HRS Data, Version O, which was produced by the

4 Cognitive Ability, Personality, and Paths to Retirement 55 RAND Center for the Study of Aging, with funding from the National Institute on Aging and the Social Security Administration. Realized retirement pathways We categorize the longitudinal labor market histories of individuals into eight distinct pathways longitudinal patterns of older adults employment, often spanning from full-time work to fully or partially exiting the workforce. The categories are motivated by earlier literature on bridge jobs (Beehr & Bennett, 2015; Cahill, Giandrea, & Quinn, 2006; Wang & Shultz, 2010), gradual retirement (Kantarci & van Soest, 2008), unretirement (Maestas, 2010), and retirement sequences (Calvo, Madero-Cabib, & Staudinger, 2017), but there are differences. The bridge and gradual employment literatures typically restrict the sample to workers in career jobs by, for example, dropping workers with less than 10 years of tenure, and then investigate workers subsequent retirement patterns. Given that our objective here was to categorize retirement patterns of all workers of a cohort, our preferred specification does not restrict the sample by tenure. The HRS data also allowed us to follow individuals for a longer time than is typical in these literatures, and to look at more detailed retirement pathways (e.g., distinguishing pathways that involve unemployment or disability immediately preceding retirement; and those who never retire). The retirement sequences of Calvo et al. (2017), which were developed independently from us, also have some similarities with our categories, but the study used fewer categories and a shorter time horizon. We use all available biennial waves of the HRS from 1992 to Our sample consists of individuals working full-time and turning years old between 1992 and We follow eligible respondents for 14 years until they are years old, assigning one of eight retirement pathways to each individual. We exclude respondents who died or left the study before the end of the observation period, as well as those who had too many missing values over this period (see details in the following paragraphs). These categories summarize labor market transitions in the entire 14-year window, involving multiple transitions for many. Figure 1 summarizes the logic we used to define work-to-retirement pathways. First, we assigned labor market statuses to all personwave observations in all HRS waves. If an individual missed a wave, we verified whether the person was deceased using a combination of National Death Index data (linked administrative death records) and HRS survey reports. In each wave, we assigned Missing status to those confirmed as alive who did not complete an interview. Anyone completing a survey and working for pay for more than 35 hours per week and more than 36 weeks per year was considered working full-time. Those working for pay, but for less than full-time hours or weeks, were considered as working part-time. We considered those who reported themselves fully retired as retired, those who reported a disability as disabled, and those who reported actively looking for work in the past 4 weeks as unemployed. Those not fitting any of these categories, such as homemakers, were initially assigned a labor force status of other. Among these other cases, those older than the Social Security full retirement age were recoded to retired, and those below the Social Security normal retirement age and selfidentified as unemployed (but did not necessarily actively look for a job) were recoded as unemployed. The rest remained in the other category. Assign Labor Force Statusfor Each Person-Wave Observation Full-time work, part-time work, unemployed,disabled, retired, deceased,missing,other Fill in MissingLabor Market Status Based on retrospective HRSquestions,including previous jobend-dates,self-reported retirementdates, and wave timing Apply Eligibility Criteria for Study Sample HRS respondents age 55-58sometimebetween 1992 and 1998;working full-timeatbaseline;labor market status stream is observedfor subsequent14 years; drop if deceased,early attritionormulti-wavegap Assign Work-to-Retirement Trajectories 1. full retirement, 2. gradualretirement, 3. unretirement, 4. always full-time work, 5. moves to part-time work, 6. unemployment retirement, 7. disability retirement, 8. complextrajectories Figure 1. Procedure to define work-to-retirement pathways. Next, we filled in missing labor market statuses, whenever possible, using retrospective questions answered in later waves of the HRS, such as the date when a respondent retired, began or ended a job, or became unemployed or disabled. Then we applied the eligibility criteria of the study sample based on age and length of follow-up. For each respondent, we calculated age at the time of each HRS interview and defined the baseline wave as the wave when the respondent s age was closest to 56 years and 6 months. To be eligible, a respondent must: (a) have a baseline age between 55 and 58 years, allowing for a maximum difference of 1.5 years from 56.5; (b) be working full-time at baseline; (c) survive to the end of the 14-year observation window; and (d) have provided sufficient labor force status information to cover the observation period. Out of 4,080 observations that met the age criterion and full-time work requirement at baseline, 554 observations (13.6%) passed away before the end of the observation period, 518 dropped out (12.7%) and 88 (2.2%) participants missed two or more consecutive interviews, without sufficient retrospective information to infer missing labor force status. Michaud and colleagues (2011) show that panel attrition in the HRS only leads to minor changes in the representativeness of the sample by most common population characteristics. Finally, we assigned a work-to-retirement pathway for each eligible respondent. Those re-entering the work force for full- or part-time work after a period in which the participant identified as retired and did not work for pay were assigned a pathway of Unretired. Those who retired completely (i.e., did not work for pay) and did not later re-enter the workforce were assigned one of four pathways. We categorized those

5 56 P. Hudomiet et al. transitioning directly from full-time work to retirement as Full-time to retirement and those who first reduced work hours as Gradual retirement. Separate categories indicated those with a period of disability ( Disability to retirement ) or unemployment ( Unemployment to retirement ) in the wave immediately prior to retirement. Participants not reporting any periods of retirement, and at most one period of unemployment, disability, or absence from the labor force, were separated into two categories: those continuing to work full-time ( Always full-time ) and those reducing their hours to part-time at some point during the study period ( Moves to part-time work ). Because some respondents qualified for multiple pathways, we prioritized categories as follows: Unretirement, Full-time to retirement, Disability to retirement, Unemployment to retirement, Gradual retirement. Very few respondents could not be categorized into any of the previously mentioned retirement pathways and were designated as Complex cases. The most frequent reason was a labor market status of other just prior to retirement. Our procedure focused on labor market statuses at the time of HRS interviews, and as such, we likely miss some changes between waves, such as a retired respondent taking on a part-time job for less than a year. We explored using additional, retrospective HRS data to tease out these intrawave transitions, but ultimately decided that the data quality was insufficient, mainly due to missing or incomplete job start and end dates. Thus, we decided against using these retrospective data to overwrite the higher quality data on contemporaneous labor force statuses. Simpler work-retirement measures For comparison, we also considered four simpler work-retirement measures that are commonly used in the economics literature. We shall explore the relationship between these measures and the one summarizing the 14-year pathway. These measures are: 1. Working full-time after age 62 (the traditional early retirement age). 2. Doing any work for pay after age 62 (including part-time work). 3. Working full-time after age 65 (the traditional normal retirement age). 4. Doing any work for pay after age 65. All of these variables were defined on the same sample as the detailed pathways. For comparability with the detailed retirement pathways, we only used labor force status information of individuals between the target age (62 and zero months or 65 and zero months) and age 71 and zero months (which is approximately the age of respondents in the last wave that was used to create their detailed retirement pathways). Retirement expectations The HRS asks respondents about their subjective probability of working full-time past age 62 and a separate question about working fulltime past age 65. Thinking about work in general and not just your present job, what do you think the chances are that you will be working full-time after you reach age 62/65? Previous research by Hurd (2009) and our own investigation reported in the Supplementary Appendix show evidence that these subjective probability reports are strong predictors of subsequent realizations. Fluid cognitive ability We assess fluid cognitive ability using a 27-point score of working and episodic memory, developed using measures available on the HRS (Crimmins, Kim, Langa, & Weir, 2011; McArdle, Fisher & Kadlec, 2007), which are closely linked to fluid intelligence and decision-making abilities (Del Missier et al., 2013), and include immediate word recall (of a list of 10 words), delayed word recall (same list), backward counting (from 20 to 10), and serial 7s (which asks individuals to start with 100 and sequentially subtract 7 five times) (Breitner et al., 1995; Ofstedal, Fisher, & Herzog, 2005). We focus on these cognitive tasks because, unlike other HRS cognitive ability measures, they were implemented in earlier waves and therefore provide a consistent measure of cognitive ability across all waves. Serial 7s and backward counting were introduced in 1996, and immediate and delayed word recall measures were consistently used in the 1996 and later waves. The sum of the correct items on the four cognitive tasks creates a score ranging from 0 to 27 points at each wave (Crimmins et al., 2011; Ofstedal et al., 2005) that we consider a measure of fluid cognitive ability. We experimented with using alternative measures such as the WAIS vocabulary score that is one measure of crystallized intelligence. Vocabulary is only assessed once from people age 50 65, so we decided against combining it with the other measures that are available in all waves. Moreover, the vocabulary score did not predict retirement outcomes, above and beyond the 27-point scale. Timing of measurement Theoretical considerations imply that fluid cognitive ability should ideally be measured (a) before retirement, because some studies show that retirement may affect cognition (Bonsang, Adam, & Perelman, 2012; Mazzonna & Peracchi, 2012; Rohwedder & Willis, 2010), and (b) at about the same age for all individuals, because fluid cognitive ability declines with age. So for our purposes, we would like to observe fluid cognitive ability at baseline, at about age 56, for all individuals in our analytical sample. Due to variation in interview dates and the fact that longitudinally comparable measures of cognition only became available with the 1996 wave the ideal measure is not immediately available, albeit close. We therefore construct it from the rich data in the HRS. Using the observation that fluid cognitive ability tends to be rank-order stable (i.e., even though it tends to decline with age, a person s rank in his or her cohort tends to remain the same) we estimate age profiles of fluid cognitive ability on all comparable person-wave observations available in the HRS. From the estimated relationship, we obtained each individual s individual score predicted for age 56. Formally, we first took the 27-point score for all individuals from all waves when the person was between ages 50 and 61 (that is, before or shortly after retirement for most respondents). Second, we regressed the score on a quadratic polynomial of age; then we took the personspecific mean of the residuals of this regression. Finally, we normalized the score to age 56 (baseline age) by adding back the predicted values of the regression for age 56. We call this score the age-adjusted personspecific mean of fluid cognitive ability. Missing cases were imputed with multiple imputation (see the following sections). Personality traits The Big 5 personality taxonomy, derived from the Midlife Development Inventory (MIDI; Lachman & Weaver, 1997; see also Costa & McCrae, 1992a, 1992b), distinguishes five high-level

6 Cognitive Ability, Personality, and Paths to Retirement 57 personality factors. Borghans et al. (2008, Table 1, p. 983) provide a useful summary of these dimensions: Conscientiousness involves the degree to which a person is willing to comply with conventional rules, norms, and standards. Neuroticism reflects the degree to which a person experiences the world as threatening and beyond his/her control. Openness to experience captures the degree to which a person needs intellectual stimulation, change, and variety. Extraversion involves the degree to which a person needs attention and social interaction. And Agreeableness reflects the degree to which a person needs pleasant and harmonious relations with others. An expanded set of psycho-social variables was first asked of one half of the HRS sample in 2004 in the Leave-Behind Questionnaire (Smith et al., 2013), and has since been measured in alternating sample halves every 2 years. The HRS Big 5 personality measures were introduced to the HRS in Respondents rate themselves using a scale from not at all to a lot on 26 adjectives representing the Big 5 personality traits. Extraversion, for example, includes outgoing, friendly, lively, active, and talkative. In 2010 and 2012, HRS introduced additional adjectives, but we use the 26-item list that is consistent over time. Similarly to the cognitive ability measures, we use age-adjusted person-specific mean scores, and missing cases were imputed by multiple imputation. We examined the correlations among the Big 5 personality scores, because such correlations might make it challenging to disentangle the individual effects of the different scores in regressions with all scores appearing at the same time as predictors. See Supplementary Appendix for details. Covariates Demographic covariates include gender, race (white, black, or other race), ethnicity (Hispanic or non-hispanic), marital status (single or nonsingle), and education (high school dropout, high school graduate, college dropout, college graduate), and the labor market status of spouses (works or not). Age is not used directly in our regressions, because the work-to-retirement pathway is defined based on the entire labor histories of individuals in a predefined age range. The timevarying marital status and spousal work status enter the models twice: at the baseline wave (around age 56) and around age 66 (or the closest available age) of the respondent. This is how we account for the effect of changes in the controls on the retirement pathways. The HRS has various measures of health. We used self-reported health, number of limitations in activities of daily living, number of limitations in instrumental activities of daily living, mental health, and survival expectations (see RAND-HRS, 2016, for technical details). Selfreported health was reported on a 5-point scale: 1 = Excellent, 2 = Very good, 3 = Good, 4 = Fair, or 5 = Poor, which we collapsed into a 3-point scale by merging categories 1 and 2 and 4 and 5, respectively. Subjective survival expectations were measured using an age-adjusted, person-specific mean of the subjective probability of living to 75 or more. The other health indicators were measured at baseline and at age 66. Job-related control variables include baseline reports of having a defined benefit pension plan, a defined contribution pension plan, and private health insurance, as well as self-employment status, major occupations (eight categories), and job tenure. Assets and income control variables include log hourly wage in the main job and log total household assets at baseline; as well as log annual earnings and log annual other income in the year before baseline. In all cases, zero values were replaced by $1 to be able to take logs, but we included indicators of zero values in each regression model. These control variables also contained missing values in a handful of cases for which we used simplified imputation methods (replacement by values from the closest available wave when possible or by the modal response). Analysis Plan We used ordinary least squares (OLS) linear-probability regressions and multinomial logit models of retirement pathways. Although the variables capture information from an extended, 14-year window of individuals life spans, the modeling techniques were cross-sectional because we collapsed each individual s longitudinal information (e.g., retirement pathways) into a single record. Missing cognitive ability and personality traits were imputed with Stata s multiple imputation program. Multiple imputation assumes Table 1. Distribution of Retirement Pathways Based on 14-Year Follow-Up and Its Relationship to Simpler Work Retirement Measures Distribution Percent working after certain ages by retirement pathway N % Any work after 62 Full-time work after 62 Any work after 65 Full-time work after 65 Full retirement 1, Gradual retirement Unretirement Always full-time work Moves to part-time work Unemployment retirement Disability retirement Complex pathways Total 2, Sample includes HRS participants who were years old between 1992 and 1998 and who were full-time workers at baseline. Full retirement: transitioned directly from full-time work to retirement; Gradual retirement: reduced work hours before full retirement; Unretirement: re-entered the work force after a period full retirement; Always full-time work: continued to work full-time; Moves to part-time work: reduced hours but never retired; Unemployment retirement: Moved to retirement from unemployment; Disability retirement: Moved to retirement from disability; Complex pathways: Cases that do not fit into any other categories. See the data section for details.

7 58 P. Hudomiet et al. that the data are missing at random. That is, the relationships between cognitive ability, personality, and the retirement pathways are the same in the (unobserved) missing and nonmissing samples after the control variables are applied. We used multivariate normal imputation with 40 data sets. The imputation models included all control variables from the regressions, as well as the main dependent variables: indicators of the retirement pathways, and indicators that the individuals worked full-time or did any work for pay after age 62 and 65. Cognitive ability was missing (and imputed) mostly because some individuals first answered these questions after age 61 (recall that we only use values measured before age 61). There were 208 missing cases (7.13% of the regression sample). The personality score was missing (and imputed) mostly because some individuals left the sample after our window of investigation (around age 70), but before they answered the personality questions. There were around 320 missing cases (depending on the trait; around 11% of the regression sample). RESULTS Distribution of Realized Work-to-Retirement Pathways The first two columns of Table 1 show the distribution of work-toretirement pathways observed in the HRS data. Fewer than 40% of the sample went through the conventional retirement process of retiring directly from a full-time job; 13.6% took a part-time job before retiring completely; 16.9% retired completely, but later re-entered the labor force (they unretired ); and 25.7% did not retire before the end of our study period (around age 70). Of those who never retired, 4 of 10 (9.8% of all cases) worked in a full-time job all along, while the rest moved to a parttime job. The other three retirement pathways were much less frequent. Columns 3 through 6 show the relationship between the pathways and the simpler work-retirement measures (working past age 62 or 65). A percentage of 45.9 of those who retired from a full-time job reported working after age 62, and 22% worked after age 65. Compared to this group, a larger fraction of those who gradually retired reported working; though a relatively smaller fraction of them reported fulltime work. Specifically, 80.8% of those gradually retiring reported to work after age 62, but only 26.8% in a full-time job; and while 48.2% of them held a job after age 65, only 7.5% of them worked full-time. Interestingly, the majority of those who unretired, did so into a parttime job. While 86% of them worked after age 65, only 25.4% held a full-time position after this age; and the corresponding numbers for doing any work or full-time work after age 62 were 96.6% and 44.3%, respectively. Naturally, 100% of those who always engaged in fulltime work reported working full-time after age 62 and 65; and 100% of those who moved to part-time work, but never retired reported to do any work for pay after age 62 and 65. Those who retired after unemployment had somewhat smaller than average work probabilities, especially in full-time jobs (34.3% after age 62 and 13.2% after age 65). People who retired after disability had very low attachment to the labor force; only 5.9% were working for pay after age 62. Finally, those with complex retirement pathways had close to average propensities to have a job, and below-average propensities to hold a full-time job (36.5% after age 62 and 20.3% after age 65). The Effect of Psychological Factors on Retirement Pathways To begin with, we show simple descriptive statistics of retirement patterns in the samples with above and below median values of fluid cognitive ability and personality traits based on the raw data in Table 2. This dichotomization is only for presentational purposes for this table; continuous measures are used in all other analyses. This table ignores any missing values in cognitive ability and personality. We will deal with these when moving to the regression framework. Fluid cognitive ability is strongly related to retirement patterns. Those with higher cognitive abilities are more likely to work longer, more likely to never retire, less likely to retire from a full-time job, and less likely to retire after disability. For example, 59.1% of people with high cognitive ability had a job after age 65, while only 51.6% did so in the group with low ability; and the fraction of those who did not retire in the study period is 29.1% and 21.6% in the two groups, respectively. Extraversion seems to be strongly related to having part-time jobs, but less related to having full-time jobs. For example, in the groups with high versus low values of extraversion, 60% versus 51.8% of workers had any job after age 65, and 33.4% versus 30.9% had full-time jobs. Openness to experience is somewhat positively and neuroticism is somewhat negatively related to having part-time jobs. Agreeableness and conscientiousness are less strongly related to retirement patterns, at least based on these simple cross-tabulations. Table 3 shows OLS linear probability models of the simpler workretirement measures. The table includes all demographic, education, health, labor market, asset, and income control variables; and the missing cognitive ability and personality traits are imputed by multiple imputation. For conciseness, we only report regression coefficients for cognitive ability and the five personality traits. Full results, including all covariates, are available in the online Supplementary Appendix. Respondents with greater fluid cognitive ability, those who are more extraverted and those who score lower on agreeableness are significantly more likely to work after age 65. The coefficients on cognitive ability and extraversion are statistically significant even at the 1% level. A one standard deviation increase in fluid cognitive ability increases the probability of working after age 65 by 4.3 percentage points from the mean of 55.9% (an increase of about 8%), and the corresponding number for extraversion and agreeableness are 4.3 and 3.1 percentage points. Fluid cognitive ability is also a strong positive predictor of working full-time after age 65, but none of the personality variables are statistically significant. These patterns are similar in the regressions of working after age 62. The other personality traits less consistently predict working after age 65. Conscientiousness may be positively related to having a full-time job after age 62, but less so after age 65. Neuroticism and openness to experience are not statistically significant in any of the regressions. To investigate the retirement process in a more nuanced way, we estimated multinomial logit models of the detailed retirement pathways. Table 4 shows the average marginal effects of the psychological factors on the probabilities of each retirement pathway. The multinomial logit results are consistent with and expand on the results provided by the preceding OLS analyses. Greater fluid cognitive ability increases the probability of never retiring (both staying in full-time jobs and moving to part-time jobs) and decreases the probability of retiring from a full-time job before age 70. Greater extraversion decreases the probability of retiring directly from a full-time job. It also increases the probabilities of moving to a part-time job and not retiring (until age 70), while having no effect on staying in full-time jobs until age 70. Agreeableness is not statistically significant in any of the pathways (except the relatively infrequent and hard-to-interpret

8 Cognitive Ability, Personality, and Paths to Retirement 59 complex category), though the point estimates suggest a weak positive effect on retiring from a full-time job. Neuroticism, conscientiousness, and openness to experience have weak and insignificant effects on the retirement pathways. Altogether, none of the variables of interest were statistically significant predictors of gradual retirement or unretirement. The point estimates suggest that extraversion may be a weak positive (nonsignificant) predictor of unretirement and a weak negative predictor of gradual retirement, while agreeableness may be a weak negative predictor of unretirement. It appears that cognitive ability and extraversion mainly predict working longer, but conditional on retiring before age 70, they do not predict the type of retirement paths strongly. The Effect of Psychological Factors on Retirement Expectations Although a systematic investigation of potential retirement mechanisms is beyond the scope of this paper, we ask whether the observed retirement outcomes came as a surprise to individuals. Table 5 shows OLS regressions of the subjective probability (i.e., expectation) of working full-time past age 62 and 65 measured at the baseline wave together with regressions of actually working in a full-time job after age 62 or 65. We also test if the coefficients are significantly different by running regressions of created variables that subtract expectations (measured on a 0 1 scale) from actual retirement outcomes (0/1 Table 2. Distribution of Work Retirement Measures and Retirement Pathways by Cognitive Ability and Personality Traits Panel A Cognitive ability Neuroticism Extraversion Low High Low High Low High A.1. % Who did Any work after age Full-time work after age Any work after age Full-time work after age A.2. Distribution of retirement pathways Full retirement Gradual retirement Unretirement Always full-time work Moves to part-time work Unemployment retirement Disability retirement Complex pathways N 1,356 1,356 1,301 1,301 1,302 1,302 Panel B Agreeableness Conscientiousness Openness to experience Low High Low High Low High B.1. % Who did Any work after age Full-time work after age Any work after age Full-time work after age B. 2. Distribution of retirement pathways Full retirement Gradual retirement Unretirement Always full-time work Moves to part-time work Unemployment retirement Disability retirement Complex pathways N 1,302 1,301 1,301 1,300 1,299 1,299 Sample includes HRS participants who were years old between 1992 and 1998 and who were full-time workers at baseline. Panels A.1. and B.1. show the percent of individuals working after age [x] by cognitive ability and personality traits (low vs. high score). Panels A.2. and B.2. show the percent distribution of retirement pathways by the same variables and the column entries sum to 100.0%. Low and high values of cognitive ability and personality traits indicate values below and above the median respectively. Cases with missing cognitive ability scores and personality traits are not included.

9 60 P. Hudomiet et al. Table 3. Linear Probability Models of Cognitive Ability and Personality Predicting Working After Certain Ages Controlling for Demographics, Education, Health, Labor Market Variables, Assets, and Income N = 2,920 Work after age 62 Work after age 65 Any work Full-time work Any work Full-time work Predictor [1] [2] [3] [4] Fluid cognitive ability [0.011]*** [0.012]* [0.012]*** [0.011]*** Neuroticism [0.010] [0.012] [0.011] [0.011] Extraversion [0.013]*** [0.015] [0.014]*** [0.014] Agreeableness [0.012]*** [0.014] [0.014]** [0.014] Conscientiousness [0.011] [0.013]** [0.012] [0.012] Openness to experience [0.013] [0.014] [0.014] [0.013] Mean outcome Note. N = 2,920. *p <.10; **p <.05; ***p <.01. Standard errors in brackets. Weighted by the baseline sample weights. Sample includes HRS participants who were years old between 1992 and 1998 and who were full-time workers at baseline. Missing cognitive ability and personality scores are multiply imputed using 40 data sets including all variables in the regressions and using multivariate normal imputation. Sample restriction: Survived and remained in the HRS seven waves after baseline and no multiwave gaps in labor force status. The outcome variable is defined as any indication of work in HRS waves when the person s age is at least 62/65 years and zero months. The psychological variables are all standardized to have zero mean and standard deviation of one in the sample. Demographic control variables: gender, race (black, other race vs. white), Hispanic indicator, and marital status (single vs. not) at baseline and at age 66, person has a working spouse at baseline and at age 66. Education: high school dropout, college dropout, college graduate vs. high school graduate. Health control variables: Self-reported health is fair/poor vs. excellent/very good/ vs. good at baseline and at age 66, number of functional limitations and number of instrumental functional limitations at baseline and at age 66, subjective probability of living to 75 or more (age-adjusted person-specific average), CESD depression score at baseline and at age 66. Labor market control variables measured at baseline: had a DB plan through his main job, had a DC plan through his main job, had private health insurance through his main job, occupations (eight categories), self-employment indicator, job tenure. Assets and income variables measured at baseline: Log hourly wage in the main job, log annual earnings during the previous calendar year (zero imputed if zero earnings), an indicator of zero earnings, log annual other income during the previous year (zero imputed if zero other income), an indicator of zero other income, log total household wealth (zero imputed if nonpositive), an indicator of nonpositive wealth. The complete output is available in the Supplementary Appendix. binary variable). Fluid cognitive ability predicts expecting to work full-time after age 62 in a similar way to actually working after age 62. The two coefficients are similar (0.021 and 0.019) and their differences are not statistically significant. Thus, cognitively more able individuals were more likely to expect to work after 62 and they were also more likely to do so. The pattern is different for working after age 65. Even though cognitively more able individuals are more likely to work in full-time jobs after age 65, they did not expect to work longer when they were in their 50s. Agreeableness and neuroticism do not significantly predict either retirement expectations or actual retirement or their differences. Interestingly, extraverted individuals are somewhat less likely to expect to work longer, but they are somewhat more likely to do so in the end. Even though these effects are not statistically significant, the difference is significant at the 5% level in the regressions of working full-time after age 65. Openness to experience has a positive effect on expectations, even though it did not predict actual retirement outcomes. People who scored high on this measure described themselves as creative, adventurous, and broad-minded. While it seems reasonable that these people would expect to work full-time longer, it is interesting to find that they then fail to do so. The difference between the coefficients from the expectations and the realizations regression is only statistically significant for working after 62, however. In summary, the observed retirement outcomes match individuals retirement expectations to some extent, but there also seems to be evidence that unforeseen changes play a role. It is worth noting, however, that retirement expectations like many other variables are measured with noise which could hamper the detection of systematic relationships. DISCUSSION Summary of Results The main purpose of this study was to investigate whether fluid cognitive abilities and Big 5 personality traits predicted detailed retirement pathways of individuals observed in a large and long panel data set, the HRS, when standard socioeconomic and health controls were applied. First, we used data from the HRS, spanning 20 years overall and 14 years for any given individual from a cohort, to identify eight categories of work-to-retirement pathways. These categories reflected distinctions both made in the literature (e.g., unretirement, bridge jobs) and observed in the data. We found that considering detailed retirement pathways is important, because less than 40% of workers in our sample followed the standard pattern of retiring directly and completely from a full-time job. A substantial proportion of respondents followed patterns that included first transitioning to part-time work, leaving the workforce and then subsequently re-entering, and remaining in the workforce (either full- or part-time) past age 70. This coincides with prior literature (e.g., Maestas, 2010; Beehr and Bennett, 2015) who also found that most U.S. workers went through less conventional paths of retirement. Smaller numbers retired from unemployment or disability