School performance and mortality: The mediating role of educational attainment and work and family trajectories across the life course

Abstract

Evidence of a strong negative correlation between adolescent academic performance and mortality points to the importance of not only cognitive, but also non-cognitive, skills in predicting survival. We integrated two bodies of research to evaluate expectations regarding the role of educational attainment and trajectories of employment and marriage experience in mediating relationships between high school class rank and longevity. In particular, we used data from the Wisconsin Longitudinal Study (n = 9,232) to fit parametric mortality models from age 55 to age 77. Multiple mediator models allowed for quantification of the degree to which the association between high school class rank and mortality is mediated by life trajectories and educational attainment. Our results show that high school class rank is a statistically significant and substantively meaningful predictor of survival beyond age 55 and that this relationship is partially, but not fully, mediated by trajectories of employment and marriage experience across the life course. Higher educational attainment also mediates a substantial part of the relationship, but to varying degrees for men and women.

The negative relationship between intelligence, typically measured by IQ, and mortality is well established (e.g., Batty et al. 2009; Deary 2008; Gottfredson and Deary 2004). Yet a study by Hauser and Palloni (2011) found that the IQ-mortality association was reduced to insignificance in models that included a measure of respondents’ rank in their high school class. In the U.S., high school students can be rank ordered according to the grades they receive. The highest grade possible in a course (an “A”) is assigned a value of 4; the lowest grade possible (an “F”) is assigned a value of 0. If a student has grade point average (or GPA) of 4.0 (i.e., perfect marks across all courses taken), he or she would be at the very top of their class, with a rank of 1. The fact that high school class rank is strongly associated with mortality, but only moderately correlated with IQ, implies that most of the association between high school class rank and mortality reflects something other than cognitive ability.

This finding raises important questions about what high school class rank represents and why it correlates so strongly with mortality. The interpretation offered by Hauser and Palloni (2011: i98), while speculative, emphasized high school class rank as a reflection of not only cognitive ability but also conscientiousness, or “responsible, compliant behavior,” which leads individuals to consistently do “the right thing in the right way at the right time and place.” This emphasis on personality and behavior draws upon established theoretical and empirical linkages between conscientiousness, its associated personality characteristics and behaviors, and both mortality and health across the life course (Deary, Batty, Pattie, and Gale 2008; Mirowsky and Ross 1998). If this emphasis on characteristics such as conscientiousness or responsible behavior is in fact appropriate, it becomes important to understand how and why these traits contribute to lower mortality risk.

In this article, we examine the role of trajectories of work and family experience across the life course, and educational attainment, in mediating the relationship between high school class rank and mortality. To develop theoretical expectations, we draw upon two large bodies of heretofore unrelated research. Life course research provides abundant evidence that stable or “more successful” work and family trajectories are associated with lower mortality. A second line of research points to the importance of early-life personality characteristics in shaping work and family experiences across the life course. Together, these two bodies of work suggest that work and family trajectories may play a meaningful role in mediating the observed relationship between adolescent academic performance and mortality.

As shown in Figure 1, our primary goal in this research is to assess whether the indirect pathway linking high school class rank to mortality via more stable or successful life trajectories and higher educational attainment (path BC) explains, or partially explains, the strong relationship between high school class rank and longevity documented by Hauser and Palloni (2011). We are also interested in ascertaining the degree to which the direct pathway between high school class rank and mortality (path A) remains relevant, net of these posited mediators (path BC). In the next section, we motivate our analyses by referencing existing research related to pathways A, B, and C in Figure 1. We then use detailed life-history data from the Wisconsin Longitudinal Study (WLS) to evaluate these relationships.

Figure 1.

Conceptual model linking high school class rank, educational attainment, work and family trajectories, and mortality. In the model, the direct effects of class rank on mortality are captured by pathway A and the indirect effects are captured by the pathway that runs through arrows B and C. In accordance with findings from Hauser and Palloni (2011), the model does not include a direct path from cognitive skills (IQ) to mortality. Instead, we assume the effects of IQ are entirely mediated by class rank. See text for more details.

Background

Intelligence, personality, and longevity

Research in cognitive epidemiology has repeatedly demonstrated that measures of early-life intelligence are associated with longer lives, and better health more generally (Batty, Deary, and Gottfredson 2007; Deary 2008). Explanations for this link include the role of intelligence in predicting educational attainment and other socioeconomic resources that contribute to better health and the role of intelligence as a correlate of “system integrity,” or resilience to factors detrimental to health (Batty and Deary 2004; Deary 2008; Gottfredson 2004). Others have posited that intelligence contributes directly to health and longevity via better health behaviors and health management across the life course (Batty and Deary 2005; Gottfredson and Deary 2004; Gottfredson 2004). Empirical evidence generally supports the view of intelligence as a predictor of socioeconomic attainment and resources that, in turn, contribute to better health (Link, Phelan, Miech, and Westin 2008).

Hauser and Palloni (2011) were not the first to demonstrate a strong relationship between adolescent personality traits (or possible proxies thereof) and longevity, but their findings suggest that a reevaluation of previous studies and interpretations regarding the role of intelligence is warranted. Because two-thirds of the variance in high school class rank is independent of IQ, their results clearly demonstrate the importance of factors other than intelligence (as measured by IQ) in shaping survival. What might these additional factors be? Hauser and Palloni pointed to conscientiousness (a “Big Five” personality characteristic) as one potential candidate. This focus makes good theoretical sense given research on the links between personality and survival. The importance of non-cognitive skills like conscientiousness, or “responsible, compliant behavior,” is now widely recognized in the intelligence-mortality literature, and, as we discuss in greater detail below, has been evaluated over the past several decades by psychologists and others in research on personality, mortality, and its behavioral correlates (Bogg and Roberts 2004; Friedman et al. 1993; Friedman 2008; Shanahan et al. 2014).

Evidence connecting high school grades and conscientiousness is also extensive. The literature on academic achievement shows that students’ grades (and, by extension, class rank) are a function of cognitive abilities (e.g., long-term memory, the ability to think abstractly and understand complex ideas, and depth of thought) and non-cognitive skills, including personality characteristics like motivation, resiliency, and self-control.¹ Raw intelligence helps students learn and solve problems with less reliance on formal instruction, whereas self-control and motivation help students study, complete and turn in assignments on time, and behave in a positive manner in the classroom and towards peers (Duckworth et al. 2012). While these individual attributes are all relevant for learning, the weight that teachers accord them when assigning grades may differ. Studies that use measures of IQ and self-control (an important aspect of conscientiousness) within the same model have shown that self-control is the better predictor of GPA, though both variables are important (Almlund et al. 2011). This suggests that grades represent more than just the sum total of a student’s intellectual abilities; they also capture non-cognitive dimensions of academic performance that could be consequential later in life (Poropat 2009).

Non-cognitive skills and longevity (Paths A and BC)

Non-cognitive skills have also been consistently linked to survival and adult health (see, e.g., Bogg and Roberts 2004; Carneiro et al. 2007; Conti and Heckman 2004; Duke and Macmillan 2016; Moffitt et al. 2011; Shanahan et al. 2014). Many studies in cognitive epidemiology show that conscientiousness is a strong predictor of lower mortality risk (e.g., Friedman et al. 1993; Friedman, Kern, and Reynolds 2010), explaining as much or more variation in survival than other well-established correlates such as socioeconomic status or intelligence (Kern and Friedman 2008; Roberts et al. 2007). In studies that consider specific aspects of adolescent conscientiousness, those found to be particularly important for longevity—e.g., dependability, organization, persistence, industriousness (Kern and Friedman 2008)—are all likely contributors to variation in academic performance (Diseth 2003; Heaven and Ciarrochi 2008; Kelly 2008; McAbee and Oswald 2013) and related measures of academic achievement (see, e.g., Furnham et al. 2003; Lounsbury et al. 2003; Noftle and Robins 2007; O’Connor and Paunonen 2007).

Some studies conclude that the relationships between personality and health and mortality are partially mediated by health behaviors (Friedman et al., 1995; Hampson, Goldberg, Vogt, and Dubanosky, 2006; Lodi-Smith et al. 2010; Mroczek, Spiro, and Turiano, 2009), or by educational attainment, marital status, and measures of occupational status in adulthood (Deary, Batty, Pattie, and Gale 2008; Martin, Friedman, and Schwartz 2007). Hauser and Palloni (2011) did not explicitly evaluate the pathways through which adolescent traits may produce the observed association between high school class rank and longevity, but they did show that high school class rank is associated with behaviors such as smoking and binge drinking in mid-life. These findings are consistent with previous research demonstrating positive associations between conscientiousness (and related characteristics such as sense of control) and a range of health behaviors associated with morbidity and mortality (Friedman 2000; Friedman et al. 1995; Kern and Friedman 2011; Lodi-Smith et al. 2010; Roberts and Bogg 2004; Roberts et al. 2007; Ross and Mirowsky 1999).

Non-cognitive skills and life outcomes (Path B)

Efforts to evaluate the ways in which non-cognitive skills assessed early in life are associated with subsequent life success (path B in Figure 1) can be found in several social science disciplines. In sociology, the life-course framework highlights the role of planfulness, “planful competence,” or personal agency more generally, in shaping individual biographies (Clausen 1991; Shanahan, Hofer, and Miech 2003). In economics, emphases on time discounting and early-life non-cognitive skills are very similar. Individuals who discount the value of future well-being less tend to invest more heavily in accumulating human capital and maintaining health, and those who possess more “soft skills” like perseverance and conscientiousness tend to have more successful education and labor market outcomes, net of cognitive skills (Heckman and Kautz 2012; Heckman, Stixrud, and Urzua 2006). Much of this work draws on a long history of related research in psychology that emphasizes concepts such as delayed gratification, goal-directed behavior, planfulness, conscientiousness, and self-control (Hampson 2008; Kern and Friedman 2010; Mischel, Shoda, and Rodriguez 1989).

Empirical evidence of these posited linkages is abundant. Research on career outcomes, including occupational status, income, job performance, and job satisfaction highlights the importance of both cognitive skills and personality characteristics such as conscientiousness, self-discipline, perseverance, and following rules (Bowles, Gintis, and Osborne 2001; Hogan and Ones 1997; Ozer and Benet-Martinez 2006). Similarly, research on planfulness shows this personality trait to be associated with higher levels of socioeconomic attainment, more job satisfaction, and greater marital stability (especially for men) (Clausen 1991; Ozer and Benet-Martinez 2006; Roberts et al. 2007; Roberts and Bogg 2004; Shanahan, Hofer, and Miech 2003).

Educational attainment is another aspect of success that has been linked to adolescent conscientiousness and discipline (Hampson, Goldberg, Vogt, and Dubanoski 2007; Poropat 2009; Shanahan et al. 2014). The same skills that allow students to achieve highly in high school—completing work on time, studying effectively, working well with others, participating actively in class, and behaving appropriately toward instructors and classmates—should also be valuable at the post-secondary level, paving the way for success in college and beyond. This expectation is supported by a large body of research demonstrating a strong (and independent) association between high school grades, college enrollment, college grades, and college completion (see, e.g., Attewell, Heil, and Resel 2010; Bowen, Chingos, and McPherson; Klasik 2012; Roderick et al. 2006; Zwick and Sklar 2005).

Educational attainment, life trajectories, and mortality (Path C)

Path C—linking education, employment careers, and marriage to mortality—has been studied extensively. It is well known that “health disparities in old age cannot be understood without linking them to people’s experiences in early and mid-life” (Herd 2009) and a large body of research on linkages between employment circumstances and mortality has shown that intermittent employment across the life course, self-employment, experience of involuntary job loss, downward occupational mobility, and other aspects of unstable careers are associated with higher mortality relative to careers characterized by stable employment and upward occupational mobility (Hayward et al. 1989; Kitagawa and Hauser 1973; Moore and Hayward 1990; Pavalko et al. 1993; Ross and Mirowsky 1995; Schnittker 2007). These characteristics of less stable or less successful careers are thought to contribute to higher rates of mortality via lower levels of income and wealth, and lack of health insurance coverage (Burgard, Brand, and House 2007; Price and Burgard 2010; Quinlan, Mayhew, and Bohle 2001), and these relationships appear to be stronger for men than for women (Krueger and Burgard 2011; Macintyre and Hunt 1997).

A similarly large literature exists documenting the relationship between education and mortality (see, e.g., Baker 2015; Elo and Preston 1996; Hayward, Hummer, and Sasson 2015; Hummer and Hernandez 2013; Hummer and Lariscy 2011; Kitagawa and Hauser 1973; Lleras-Muney 2005; Mirowsky and Ross 2007; Preston and Taubman 1994; Warren and Hernandez 2007). Researchers have shown that obtaining higher levels of schooling helps people to “acquire better and more stable employment, increase earning power, develop effective agency, attain a greater sense of personal control over their lives, and develop beneficial social connections” (Hummer and Lariscy 2011: 243). These “flexible resources” can then be deployed in a variety of ways, and at various points in the life course, to improve health outcomes and increase longevity (Mirowsky and Ross 2003).

In contrast, much less has been written on marital careers and mortality, and much of what we do know about marriage pertains specifically to marital dissolution. It is clear that the experience of divorce or widowhood is associated with a higher risk of death, or conversely, that stable marriages are positively associated with longevity (Dupre, Beck, and Meadows 2009; Elwert and Christakis 2008; Lillard and Waite 1995; Lund, Holstein, and Osler 2008; Tucker, Friedman, Wingard, and Schwartz 1996). Shorter marriages and marital instability are associated with lower levels of income, wealth, and occupational status, which in turn are associated with higher mortality risk (Dupre, Beck, and Meadows 2009; Hemstrom 1996; Lillard and Waite 1995). Young age at marriage is also associated with higher mortality, reflecting the link between early marriage and subsequent divorce (Dupre, Beck, and Meadows 2009). Some have also found that the association between marital dissolution and mortality is stronger for men than for women (Hemstrom 1996; Zick and Smith 1991).

An important contribution of this earlier research on mortality as a function of marriage and occupational careers is to demonstrate that trajectories of experience across the life course are related to mortality above and beyond marital status and occupational status at older ages. Stated differently, variation in mortality risk reflects not only occupational and marital characteristics at older ages, but also the trajectories of experience across the life course by which individuals came to possess those characteristics. Consistent with a basic tenet of the life course framework, the number, nature, and timing of transitions across the life course and the duration of time spent in different marital and employment states shape differences in well-being at older ages. This insight is a starting point for our study.

To our knowledge, this study is the first to focus explicitly on the role of employment and marriage trajectories in mediating the link between non-cognitive skills in adolescence and mortality. Several studies have examined the extent to which measures of socioeconomic status mediate relationships between mortality and both cognitive and non-cognitive skills (e.g., Batty et al. 2009; Hampson et al. 2007; Kern et al. 2009; Link, Phelan, Miech, and Westin 2008), but this approach is limiting to the extent that trajectories of experience across the life course contribute to variation in later-life well-being above and beyond single, point-in-time measures of those characteristics (Halpern-Manners, Warren, and Raymo 2015). While ours is the first empirical examination, we are not the first to recognize the potential importance of the pathway between adolescent disposition and skills, trajectories of experience and attainments across the life course, and later life well-being. Shanahan, Hofer and Miech (2003: 202) argue that “planfulness in adolescence or in early adulthood may predict indicators of well-being and adjustment in old age by way of life course achievements.” Similarly, Herd (2010: 479) notes that “strong academic performance may affect health indirectly by facilitating improved occupational attainment and higher earnings earlier in the life course, which ultimately influences health across the life course.”

Data

The Wisconsin Longitudinal Study (WLS) is a long-term study of a random sample of 10,317 men and women who graduated from Wisconsin high schools in 1957. WLS “graduates” were interviewed in 1957, 1975, 1993, 2004, and most recently in 2011. The WLS graduate sample is broadly representative of white, non-Hispanic Americans who have completed at least a high school education—a group that includes about two-thirds of all Americans of this generation (Hauser and Roan 2006). In 1993, when most of the surviving graduates were age 53-54, 87% responded to the telephone survey and 71% responded to the mail survey. The corresponding response rates were 81% and 76% in 2004, and 72% and 65% in 2011.

Deaths to graduates have been ascertained in two, complementary ways. Family members or friends who report the death of a graduate in the process of pre-survey tracing or at the time of initial survey contact were asked to provide information about the date and cause of death. This information was then confirmed by matching available information with the Social Security Death Index (SSDI) and National Death Index (NDI). Deaths are also identified via periodic searches of the SSDI and NDI, in which the social security numbers of all graduates are checked against the death index. Data from a recent search show that 2,908 (28%) of the graduates had died before December 2016. Our analyses focus on survival beyond age 55, the age by which nearly all respondents completed the 1993 interview in which much of the life history data was collected. After excluding 467 graduates known to have died before age 55, 613 with missing information on date of birth or death, and 5 who were several years older than the rest of the 1957 graduates, we were left with 9,232 members of the original sample known to have survived to age 55. Of these, 2,902 (31%) had missing data on key variables. This includes 1,142 (12%) who did not respond to the 1993 survey and are thus missing information on life trajectories, 741 (8%) who did respond to the 1993 survey but did not complete the employment history module, and 1,711 (19%) who had missing information on high school class rank or other information collected in the 1957 survey (see Appendix Tables A1 and A2 for further information on the extent and patterning of item-level missingness).² We performed multiple imputation via chained equations (Raghunathan et al. 2001) to impute missing values for all variables for the full sample of graduates known to have survived to age 55 and not missing information on birth date or death date.³

In this sample, 2,180 (24%) died between the ages of 55 and 77, with the mean age of death being 68.5. The fact that we are focusing on relatively early deaths can be seen as either a limitation or a strength of the study. It is a limitation in that a large proportion of our sample (76%) has yet to experience the event of interest (i.e., is right censored). It is a strength in that early death is valuable indicator of disadvantage and thus of particular interest in evaluating the ways in which trajectories of work and family experience across the life course may mediate relationships between adolescent academic performance and health inequalities at older ages.

Early-life characteristics

Because the WLS began as a survey of high school seniors, it contains information on intelligence, school performance, and family background that is typically not collected (or is collected retrospectively) in surveys of older populations. Cognitive ability (IQ) is measured using scores on the Henmon-Nelson Test of Mental Ability, a test of general intelligence that was given to all Wisconsin high school students at that time (the measure we use converts students’ raw scores into percentile ranks, with higher ranks corresponding to higher scores)⁴ Class rank is based on grades in classes taken throughout respondents’ high school careers and was obtained from school district administrative records at the time of the 1957 survey. To accommodate graduating classes of different sizes, the version of class rank we use in our analysis converts ranks into percentiles by dividing a student’s rank by the number of students in their high school class, and then subtracting that quantity from 100. So if a student ranked first in a class of 100, her percentile rank would be 100 – [(1/100)x100] = 99. Thus, higher values indicate better (average) grades. In addition to cognitive ability and class rank, our models also include two other measures of family background considered by Hauser and Palloni (2011): parents’ income and farm background (i.e., whether or not the family head’s occupation was classified as farm or nonfarm). These measures were obtained from tax forms filled between 1957 and 1960 and are used to proxy background variables that could influence the processes under study.

Employment histories

The 1993 telephone survey obtained essentially complete employment histories for graduates covering the period 1975 through 1993 (ages 36 through 54 for most graduates). These employment histories are comprised of multiple employment spells—uninterrupted periods of time working for the same employer, including self-employment, with detailed information on employment status and the characteristics of each job held. Based on these data, we produced measures of employment status, full-time employment, occupational earnings, and access to private pension and health insurance coverage at six-month intervals. Because the WLS does not contain individual wage histories, we defined low-wage jobs as those that fall below the median value of occupational earnings—the percentage of people in a given occupation who reported hourly wages of at least $14.30 in the 1990 census (Warren and Hauser 1997).

Marriage histories

The 1975 and 1993 surveys collected information on marital status and marital transitions from age 18. Respondents were asked to provide the month and year in which their most recent marriage began and ended (if applicable), as well as the start and end date for any previous marriages. As with the employment history data, we used this information to construct measures of marital status (married or unmarried) at six-month intervals from age 18 through respondent’s age at the time of the 1993 survey.

Summarizing life history data

Because different techniques for summarizing trajectories of experience across the life course typically result in different numbers of distinct trajectories, often misclassify individual respondents, and are generally sensitive to the nature of the data being modeled, we followed Warren et al.’s (2015) suggestion to compare results across multiple methods. In particular, we employed three different methods for summarizing life-history history data: simple summary measures, group-based trajectory models (Nagin 2005), and grade of membership models (Manton et al. 1994). Although these strategies are quite different in theory and implementation, we use them for the same end: to characterize long-term patterns in respondents’ work and family lives. Below, we present results based on models in which life trajectories were generated using group-based trajectory models, or what is sometimes referred to as latent class growth analysis. Importantly, substantive conclusions are similar if we use a simple measurement strategy (where individuals’ biographies are approximated using a series of simple summary measures) or grade of membership models (which use a general multivariate procedure to produce estimates of how closely affiliated individual trajectories are with a specific trajectory type or “profile”, allowing for partial affiliation with multiple profile types) to summarize employment and marriage history data. Results from these additional analyses are presented in the Online Appendix (Appendix Tables A3-A8), along with a lengthier discussion of each measurement approach.

Educational attainment

Number of years of completed education reported at the time of the 1975 survey was collapsed into three categories: high school (12), some college or an associate’s degree (13-15), and bachelor’s degree or higher (16+).

Gender

Given large gender differences in both employment trajectories and mortality risk in late mid-life, we fit models separately for men and women. There were 4,375 men and 4,857 women in our analytic sample.

Method

To evaluate the conceptual model proposed in Figure 1, we fit a series of parametric survival models with the shape of the mortality hazard assumed to follow a Gompertz distribution between the month that respondents turned 55 and December 2016, when most of the graduates were 76-77 years old.⁵ We fit three sets of models, separately by gender. We begin by replicating the reduced form model in Hauser and Palloni (2011) to assess the extent to which relationships between IQ, high school class rank, and mortality depend upon the age range considered. In particular, we ask whether our findings, based on mortality between the ages of 55 and 77, resemble their findings based on mortality between ages 18-68.

Our second set of models extends Model 1 to include the employment and marriage trajectory measures, as discussed earlier.⁶ To what extent is the expected negative association between high school class rank and mortality (net of IQ and other background characteristics) mediated by employment and marriage trajectories across the life course? To quantify the degree of mediation, we fit multiple mediator models using the inverse odds ratio weighting (IORW) approach proposed by Tchetgen Tchetgen (2013) and later adapted by Nguyen et al. (2016) for use with continuous exposures and time-to-event outcomes with possible censoring.⁷ IORW is a flexible semi-parametric approach that produces estimates of natural direct, natural indirect, and total effects, even in the presence of treatment-by-mediator interactions and multiple mediators.⁸ To implement this approach, we first derive inverse odds ratio weights that summarize the strength and direction of the relationship between class rank and the mediators of interest. We do so by transforming the results obtained from a simple OLS model where class rank, C, is expressed as a function of our mediating variables, M, and the vector of covariates, X, defined earlier:

$$C={\beta }_0+\mathbf{M}{\beta }_1+\mathbf{X}{\beta }_2+\epsilon ,\text{with}\epsilon \sim N(0,{\sigma }^2).$$ (1)

The transformation we use to retrieve the necessary IOR weights is based on an odds ratio function relating class rank to our set of mediators, conditional on covariates:

$$\frac{1}{\text{OR}(C,\mathbf{M}\mid \mathbf{X})}=\frac{1}{e^{[\widehat{{\beta }_1}\times \mathit{C}\times \mathbf{M}]∕{\sigma }^2}}.$$ (2)

Once calculated, these weights can then be used to deactivate the indirect pathways that run through the mediators (by fitting a re-weighted version of the survival model described above that adjusts for the class rank-mediator relationship), isolating the natural direct effect of class rank on mortality (see Tchetgen Tchetgen 2013 for more details). They can also be used to recover estimates of the natural indirect effects (by subtracting the direct effects from estimates of the total effects obtained from an unweighted survival model), with standard errors in all cases obtained via bootstrapping (Pearl 2012). Together, these quantities allow us to make inferences about the nature of the class rank-mortality association and the relative importance of education and work and family roles in mediating the observed relationship.

To facilitate interpretation, the reference trajectory in all models is the one that is associated with stability or “success” (e.g., stable employment, stable marriage, higher-paying occupations, access to health insurance and private pension coverage). We fit models separately for each of the trajectory generating approaches (and by gender) to assess the extent to which results may depend upon the method used to summarize the life history data and to allow for variation across a theoretically important demographic variable. Our third set of models includes educational attainment in order to evaluate its role as a mediator in the class rank-survival association.

Results

Descriptive statistics, summarized across the imputed data sets, are presented separately by gender in Table 1. These figures are for the analyses using trajectories generated from group-based trajectory models (descriptive statistics for analyses based on simple summary measures of trajectory data and grade of membership models are presented in Appendix Tables A1 and A2, respectively). Table 1 shows that women tended to do better than men in high school and that they experienced more complex life trajectories, especially with respect to employment. A majority of men were consistently employed full-time across mid-life in jobs that provided access to both health insurance and private pension coverage. Most women were also consistently employed across mid-life (56%), but in jobs that were more heterogeneous with respect to full-time status and access to health insurance and private pension coverage. Although both groups tended to marry early and remain stably married, a non-trivial number either never married (6% of men and women), married later in life (26% of men, 23% of women), or moved in and out of marriages across the life course (14% of men and 12% of women).

Table 1.

Descriptive statistics, by sex, group-based trajectory models.

Variable	Men	Variable	Women
Died (0 = No, 1 = Yes)	0.28	Died (0 = No, 1 = Yes)	0.20
Farm background (0 = No, 1 = Yes)	0.17	Farm background (0 = No, 1 = Yes)	0.17
IQ (Henmon-Nelson test)	56.72	IQ (Henmon-Nelson test)	56.54
High school rank	43.90	High school rank	57.51
Parent’s income	64.16	Parent’s income	63.13
Educational attainment		Educational attainment
High school	0.56	High school	0.69
Some college	0.14	Some college	0.13
College or more	0.29	College or more	0.18
Employment status trajectory		Employment status trajectory
Consistently employed	0.91	Consistently employed	0.56
Not consistently employed	0.09	Exited employment	0.12
		Entered employment later	0.15
		Never employed	0.17
Low-paying occupation trajectory		Low-paying occupation trajectory
Never in a low-paying occupation	0.41	Never in a low-paying occupation	0.36
Always in a low-paying occupation	0.42	Always in a low-paying occupation	0.43
Occasionally in a low-paying occupation	0.17	In a low-paying occupation later	0.10
		In a low-paying occupation earlier	0.12
Access to health insurance trajectory		Access to health insurance trajectory
Always covered by health insurance	0.72	Always covered by health insurance	0.38
Never covered by health insurance	0.14	Never covered by health insurance	0.38
Usually covered by health insurance	0.14	Occasionally covered by health insurance	0.23
Access to pension coverage trajectory		Access to pension coverage trajectory
Always covered by private pension	0.69	Always covered by private pension	0.34
Occasionally covered private pension	0.31	Never covered by private pension	0.45
		Covered by private pension later	0.21
Full-time employment trajectory		Full-time employment trajectory
Always employed full-time	0.89	Always employed full-time	0.40
Not always employed full-time	0.11	Never employed full-time	0.36
		Entered full-time employment	0.25
Marital status trajectory		Marital status trajectory
Married later – stably married	0.26	Married early – stably married	0.59
Married early – stably married	0.53	Later marriage – some dissolution	0.23
Never married	0.06	Never married	0.06
Unstably married	0.14	Unstably married	0.12
N	4,375	N	4,857

Note: Figures averaged across five multiply imputed data sets. See text for more details.

Table 2a presents the results of three multivariate models for men. Model 1 shows that high school class rank had a strong negative association with mortality beyond age 55, but that none of the other early-life measures, including IQ, are significantly associated with men’s mortality; this result closely replicates Hauser and Palloni’s (2011) earlier findings.⁹ Model 2 shows that, in some cases but not others, more stable or successful life courses were associated with lower mortality. Holding all else constant, men who were stably married and who always had access to pension coverage had lower mortality than their counterparts with less stable marriages and only occasional coverage. Other trajectories of job characteristics (i.e., earnings, health insurance, full-time and consistent employment) were not associated with the risk of death. Importantly, the inclusion of these trajectory measures attenuated the statistically significant relationship between high school class rank and mortality only slightly (from −0.179 to −0.164). A more dramatic change is evident in Model 3, which includes a measure of educational attainment (as observed in 1975). After controlling for education, the coefficient for high school class rank was reduced by nearly 50 percent relative to our baseline specification, though it did retain significance at the .05 level. Although we present more formal analyses below, we take this to mean that education acts as a mediating variable.

Table 2a.

Estimated coefficients from Gompertz hazard models, men, group-based trajectory models.

	Model 1		Model 2		Model 3
Variable	Coeff.	p < ∣ z∣	Coeff.	p < ∣ z∣	Coeff.	p < ∣ z∣
High school rank	−0.179	0.000	−0.164	0.000	−0.103	0.014
IQ (Henmon-Nelson test)	0.027	0.455	0.052	0.166	0.086	0.023
Farm background	−0.065	0.426	0.054	0.521	−0.080	0.337
Parent’s income	−0.049	0.310	−0.043	0.391	−0.013	0.801
Educational attainment
High school (ref)					—	—
Some college					−0.185	0.038
College or more					−0.448	0.000
Employment status trajectory
Consistently employed (ref)			—	—	—	—
Not consistently employed			0.172	0.167	0.211	0.093
Low-paying occupation trajectory
Never in a low-paying occupation (ref)			—	—	—	—
Always in a low-paying occupation			0.123	0.076	0.028	0.693
Occasionally in a low-paying occupation			−0.061	0.548	−0.113	0.276
Access to health insurance trajectory
Always covered by health insurance (ref)			—	—	—	—
Never covered by health insurance			−0.031	0.767	−0.044	0.669
Usually covered by health insurance			0.116	0.279	0.092	0.387
Access to pension coverage trajectory
Always covered by private pension (ref)			—	—	—	—
Occasionally covered private pension			0.157	0.050	0.153	0.056
Full-time employment trajectory
Always employed full-time (ref)			—	—	—	—
Not always employed full-time			−0.123	0.250	−0.102	0.337
Marital status trajectory
Married later – stably married (ref)			—	—	—	—
Married early – stably married			0.062	0.250	0.024	0.767
Never married			0.602	0.000	0.609	0.000
Unstably married			0.489	0.000	0.465	0.000
Constant	−7.504	0.000	−7.812	0.000	−7.707	0.000
Gamma	0.006	0.000	0.006	0.000	0.006	0.000

Notes: The reference category for farm background is "not from a farm background". IQ and high school class rank have both been standardized to have a mean of 0 and a SD of 1, and parental income has been logged. N = 4,375. See text for more details.

In Table 2b, the magnitude of the high school class rank coefficient in Model 1 for women (−0.193) is similar to that for men. The results of Model 2 are also similar to those for men: stable marriage and stable employment were both associated with a lower risk of mortality. Trajectories of occupational earnings and access to pension and health insurance were not related to mortality, but women who were not consistently employed full-time had a lower risk of death than those who were always in full-time work. This pattern was not observed in the analysis of men, presumably due to differences in the meaning of work and non-work for men and women in this cohort. Higher education was also significantly associated with lower mortality for women, but including education in Model 3 did not reduce the class rank coefficient as dramatically as it did in the previous table. This suggests that the various components of path BC, and their relative importance as mediators, may vary by gender.

Table 2b.

Estimated coefficients from Gompertz hazard models, women, group-based trajectory models.

	Model 1		Model 2		Model 3
Variable	Coeff.	p < ∣z∣	Coeff.	p < ∣z∣	Coeff.	p < ∣z∣
High school rank	−0.193	0.000	−0.166	0.000	−0.148	0.001
IQ (Henmon-Nelson test)	0.050	0.239	0.052	0.244	0.062	0.163
Farm background	−0.373	0.002	−0.320	0.002	−0.316	0.002
Parent’s income	−0.060	0.323	−0.063	0.309	−0.046	0.462
Educational attainment
High school (ref)					—	—
Some college					0.134	0.165
College or more					−0.303	0.009
Employment status trajectory
Consistently employed (ref)			—	—	—	—
Exited employment			0.517	0.000	0.502	0.000
Entered employment later			0.330	0.015	0.322	0.017
Never employed			0.248	0.024	0.250	0.024
Low-paying occupation trajectory
Never in a low-paying occupation (ref)			—	—	—	—
Always in a low-paying occupation			0.115	0.229	0.069	0.469
In a low-paying occupation later			−0.132	0.396	0.172	0.268
In a low-paying occupation earlier			0.092	0.460	0.045	0.719
Access to health insurance trajectory
Always covered by health insurance (ref)			—	—	—	—
Never covered by health insurance			0.057	0.695	0.075	0.603
Occasionally covered by health insurance			0.012	0.925	0.027	0.833
Access to pension coverage trajectory
Always covered by private pension (ref)			—	—	—	—
Never covered by private pension			0.154	0.236	0.146	0.255
Covered by private pension later			−0.026	0.848	−0.044	0.745
Full-time employment trajectory
Always employed full-time (ref)			—	—	—	—
Never employed full-time			−0.274	0.020	−0.279	0.018
Entered full-time employment			−0.173	0.150	−0.173	0.148
Marital status trajectory
Married early – stably married (ref)			—	—	—	—
Later marriage – some dissolution			0.142	0.087	0.174	0.039
Never married			0.547	0.000	0.605	0.000
Unstably married			0.427	0.000	0.430	0.000
Constant	−7.889	0.000	−8.185	0.000	−8.203	0.000
Gamma	0.007	0.000	0.007	0.000	0.007	0.000

Notes: The reference category for farm background is non-farm background. IQ and high school class rank have both been standardized to have a mean of 0 and a SD of 1, and parental income is logged. N = 4,857. See text for more details.

To facilitate evaluation of the magnitude of differences in longevity by adolescent academic performance, we present cumulative survival probabilities for men and women at the 25^th and 75^th percentile of the gender-specific high school class rank distributions. Figure 2 shows these survival probabilities from age 55 to age 77, separately by sex for each of the three models in Table 2. Cumulative survival probabilities are evaluated at the mean value of continuous variables (i.e., IQ and parents’ income) and the omitted values of categorical variables (i.e., farm background and all trajectory measures). From these figures, we can see that higher high school class rank is associated with longer lives in all models, that the magnitude of this difference is very similar for men and women (7-8 percentage points at age 77), that the longevity gap shrinks (especially for men) when life trajectories are introduced in Model 2, and that controlling for educational attainment further reduces the longevity gap (again, especially for men).

Fig. 2.

Cumulative probability of survival for those at the 25th (red) and 75th (blue) percentiles of high school class rank, by age, gender, and model specification. All survival probabilities were evaluated at the mean value of continuous variables and the omitted values of categorical variables. Model 1 replicates the reduced-form specification used by Hauser and Palloni (2011), which included controls for parental income and farm background. Model 2 adds measures describing respondents’ trajectories of work and family experiences, as derived from group-based trajectory models. Model 3 includes a measure of educational attainment, plus all of the variables from Models 1 and 2. See text for more details.

Finally, to quantify the extent to which life trajectories and educational attainment mediate the relationship between high school class rank and mortality, we present the results from our mediation analysis. Figure 3 shows the total effect broken down into the natural direct effect of high school class rank on mortality (path A) and the natural indirect effect via educational attainment and life trajectories (path BC). The results from Model 2 show that differences in occupational and marital trajectories explain one-fourth of the association between high school class rank and mortality risk for men, but only 7% of the association for women. In this model, the direct and indirect paths (paths A and BC) are significantly different from zero for men (p < 0.01 for the direct path; p < 0.10 for the indirect path), but only the direct path is significant for women (p < 0.01).¹⁰ For both men and women, educational attainment appears to play a more important role than life trajectories in mediating the relationship between high school class rank and mortality. In Model 3, we find that 59% of the association between adolescent academic performance and mortality among men operates via life trajectories and educational attainment; among women the same figure is 31%. In this model, both the direct and indirect paths are statistically different from zero for men (p < .01) and women (p < .10).¹¹ In general, these findings conform to our hypotheses regarding mediators of the class rank-mortality relationship (particularly education), and highlight the gender-specific ways in which mediation occurs.

Fig. 3.

Direct and indirect effects of high school class rank on mortality, by model specification and gender. Direct effects are shown in grey; indirect effects are shown in red. Model 2 includes work and family trajectories as mediating variables; model 3 includes work and family trajectories and respondents’ educational attainment, as in Tables 2a and 2b. All models also include controls for parental income and farm origin. Direct and indirect effects were estimated using the inverse odds ratio weighting approach proposed by Tchetgen Tchetgen (2013). See text for more details.

Discussion

This study was motivated by evidence that high school class rank is a powerful predictor of longevity (Hauser and Palloni 2011). Because inclusion of class rank in models of all-cause mortality reduces the well-studied link between intelligence and lower mortality to statistical insignificance, it is important to understand what predictors of longevity (other than intelligence) high school class rank reflects. Starting from Hauser and Palloni’s (2011) suggestion that class rank could serve as an indicator of conscientiousness, we integrated two distinct bodies of literature to posit that better academic performance in high school contributes to lower mortality via its association with higher educational attainment and more stable employment and marriage trajectories across the life course.

Our results provide some support for this hypothesis, though to varying degrees by gender. For men, the amount of mediation (path BC in Figure 1) exceeded 25% when we restricted our list of mediators to work and family trajectories, but approached 60% when we expanded it to include respondents’ educational attainments. For women, approximately a third (30%) of the total association was mediated by life trajectories and educational attainment, but less than a tenth was mediated by just the trajectories. This suggests that, especially for women in the WLS, the pathways linking class rank and mortality depend on other variables not included in our models. Additional mechanisms worth exploring include trajectories of health behaviors across the life course, trajectories of spouses’ employment characteristics and experiences, and other aspects of socioeconomic status not captured in our analyses.

What do our findings mean for life course theory and research? In addition to providing support for Hauser and Palloni’s (2011) earlier claims regarding the relationship between high school class rank and survival, we also identified a set of plausible mechanisms connecting the two variables. Our results suggest that the same characteristics and behaviors that lead individuals to achieve highly in high school also have benefits for survival during middle and later life—and that these benefits accrue, in part (and, again, to varying degrees by gender), through people’s subsequent experiences at school, at work, and in their family lives. Although we have not sought to ascertain what these traits and behaviors are, we agree with Hauser and Palloni (2011) that conscientiousness could play an important role. Students who are more dependable and responsible in high school tend to achieve highly—and tend to be more dependable and responsible after high school—causing the effects of class rank to reverberate (in multiple health-enhancing ways and through multiple mechanisms) across the life course and into older adulthood. We see this hypothesis as broadly consistent with key tenets of the life course perspective, which place an emphasis on life pathways and interconnections across life stages (Elder et al. 2003).¹²

We believe that our findings also have potential policy implications. In the U.S., interventions designed to foster non-cognitive skills have received considerable attention (Kautz et al. 2017). Researchers now believe that non-cognitive skills are key predictors of adult success and that intervening early in the life course to develop these skills is an especially efficient way to reduce inequality (see, e.g., Heckman 2006). Our findings provide one more, longer-term reason to think about the value of early-life interventions. If survival to older ages is indeed a function of non-cognitive skills that develop earlier in life (as our results would seem to suggest), and if these skills also have positive implications for individuals’ education and work/family lives (as our results also suggest), then the benefits of intervening may be greater, more far-reaching, and more multi-faceted than previously anticipated.

To refine our analyses further, researchers should consider incorporating a broader array of trajectory types using more dynamic modeling strategies that allow for the co-evolution of multiple trajectories. Life-course theory highlights dependencies that exist between people (the concept of “linked lives”) and across work and family domains (the concept of “interlocking trajectories”). The survival chances of individuals in our sample may be linked, for example, not only to their own experiences in the labor market, but also to the experiences of their partner and others around them—and these “spillover” effects could have implications for a broad array of outcomes (e.g., for trajectories of marital quality, for trajectories of employment and job quality, for trajectories of mental and physical health, for trajectories of health behaviors, and for trajectories of childbearing and parenting) (see, e.g., Gathmann et al. 2020). Grappling with this level of complexity within the context of a single model is no easy task, but we are optimistic that advances in multi-trajectory modeling and network analysis will soon facilitate this type of research (Valente and Pitts 2017; Nagin et al. 2018).

Gender differences in effect mediation also deserve further attention. We suspect that the diminished role of work and family trajectories among women in our analysis is a function of the WLS cohort and the historical context in which it was embedded. Women’s labor force participation expanded rapidly during the 1960s and 1970s (as women in the WLS moved through their 20s and 30s), but that does not mean that long-term stable employment was an expectation among women in the WLS or that they primarily approached higher education as a way to achieve economic success (Goldin 2006). These unique contextual characteristics may have attenuated the relationship between class rank, education, and work among women, potentially explaining the lack of indirect effects. Follow-up studies of younger cohorts (who are only just now aging into retirement and out of their working years) and of subpopulations not included in the WLS (especially racial/ethnic minorities and those who did not complete high school) will be important for assessing this hypothesis.

Finally, we would like to see a direct assessment of conscientiousness and the mechanisms that connect it to survival. Our analyses show that the class rank-mortality relationship operates, in part, through lifelong trajectories of work and family experiences and educational attainment, but the causal effect of class rank itself is likely to be minimal. What may matter more, as we and others have theorized, are the behaviors and dispositions that give rise to academic success. Replicating our approach using direct and well-validated measures of “Big Five” personality characteristics (including conscientiousness)—or other theoretically relevant personality and/or behavioral traits ascertained early in life—would further clarify the causal processes at play in our analysis (and in other analyses examining the etiology of survival), while also pointing the way to possible interventions. We are hopeful that this line of work will move us one step closer to understanding the full set of cognitive and non-cognitive factors that produce variation in human mortality, and the intervening (and potentially gender-specific) life-course processes through which their effects operate.

Highlights.

• High school class rank is a significant and substantively important predictor of survival beyond age 55, even net of IQ.

• The relationship between class rank and survival is partially mediated by trajectories of employment and marriage across the life course.

• Educational attainment also mediates the class rank-mortality relationship, but to varying degrees for men and women.

Appendix

TECHNICAL APPENDIX (to be provided online)

In this appendix, we provide a brief discussion of the three approaches we used to summarize respondents’ long-term trajectories of work and family experiences. Our goal is not to provide a full technical description of each method; instead we seek to (1) identify the assumptions that underlie each approach and (2) provide further details about our implementation. For a more comprehensive treatment, the interested reader should consult Bollen and Curran (2006), Collins and Lanza (2010), Manton et al. (1992), Nagin (2005), and Warren et al. (2015).

Simple summary measures

The easiest way to describe individual biographies is to generate a series of summary variables that record (1) whether or not individuals’ ever experienced a particular work or family circumstance and (2) if they did experience it, when it first occurred and/or how long it lasted. Although this strategy may seem overly-simplistic (and ad hoc), it does not appear to perform much worse (or better) than other more advanced approaches, including those described below (Warren et al. 2015). In our analysis, we constructed five summary measures of employment based on biannual data on respondents’ employment circumstances, and four summary measures of respondents’ family lives based on detailed marital history information. These measures are as follows: number of employment spells; number of years employed; whether the respondent ever had access to employer-provided health insurance; whether they ever had access to private pension; whether they ever lost a job involuntarily; number of marriages; age at first marriage (earliest 25%, the middle 50%, the latest 25%, or never married); whether the respondent ever got a divorce; and whether or not they were ever widowed. We followed previous WLS research by defining employment spells as uninterrupted (except for seasonal workers) periods of time working for the same employer, including self-employment. This should not be confused with distinct job spells, which entail specific work responsibilities, and which may or may not change during the course of employment with a single employer (i.e., within spells).

Group-based trajectory models (GBTM)

Our summary measures are useful and informative, but they are also fairly crude. GBTMs—which use a discrete set of trajectories to encode information about the timing, duration, and sequencing of individuals’ experiences—provide a more sophisticated alternative. The basic estimating equation can be expressed most compactly in SEM notation (Bollen and Curran 2006):

$$y={\sum }_{g=1}^G{P}_i^{(g)}\left[{\Lambda }^{(g)}{\eta }^{(g)}+{\epsilon }^{(g)}\right],$$ (1)

where y represents a longitudinal sequence of measurements for some attribute of interest (in our case, different types of work and family experiences time ordered across the life course); $P_i^{(g)}$ is the probability that WLS respondent i belongs to latent group g (such that $P_i^{(g)}\ge 0$ and ${\sum }_{g=1}^G{P}_i^{(g)}=1$; Λ is a matrix of constants and factor loadings; η is a vector of growth parameters that includes intercepts, slopes, and quadratic terms; and ε is a vector of error terms. The superscripts in Eq. (1) indicate that the growth parameters that define the shape of each trajectory are free to vary across the g trajectory groups (note that there is no residual within-class variance, implying that individuals’ work and family experiences are assumed to be homogeneous within groups). This feature is what allows each trajectory to have a distinct functional form.

To identify the most appropriate number of groups, we followed standard practice by fitting models in a stepwise fashion, incrementally increasing numbers of latent trajectories. We then compared the model fit obtained using several statistical criteria (e.g., BIC, sample-size adjusted BIC, AIC, the entropy index, the Lo-Mendell-Rubin LR test, and the Vuong-Lo-Mendell-Rubin LR test). When formal criteria failed to converge on a common solution, we gave preference to specifications that produced adequately populated classes (e.g., no fewer than 5 percent of all cases) with substantively distinct trajectory groups.

After settling on a preferred number of groups, the parameter estimates obtained from Eq. (1) can be used to generate a respondent’s posterior probability of group membership, denoted P(j ∣ y). The posterior probability records the likelihood that a respondent with the observed sequence of measurements y belongs to trajectory group g. In this respect, the posterior probability provides a criterion for assigning respondents to their most likely trajectory group and for assessing the precision with which the model fits the data (Nagin 2005).

Grade of membership models (GoM)

GoMs are similar to the GBTMs described above, but the notation is somewhat different:

$$\mathrm{P}\left(y=1_j\mid g\right)={\sum }_{g=1}^G{m}_g{\lambda }_{gjl}.$$ (2)

Here, P(y = 1_j ∣ g) is the probability that the value of dichotomous variable y is equal to 1 for a given polytomous outcome j (e.g., the age-specific observations of employment or marital status in our study) conditional on membership in class (or type) g. m is the so-called GoM score, which range from 0 to 1 and denote individual i’s degree of (partial) membership in group g. As with GBTMs, m_g ≥ 0 and ${\sum }_{g=1}^G{m}_g=1$. Rather than “latent classes” or “latent trajectories”, g are referred to as “pure types” or “extreme profiles” in the GoM literature, reflecting the emphasis on partial membership in different states or classes. λ_gjl are the estimated probabilities of response 1 for outcome j for hypothetical individuals who belong completely to a single pure state, g. Given the focus on partial membership (or grade of membership) in multiple states, GoM models do not provide estimates of the probability that an individual belongs to a single profile. We therefore use GoM scores (g) to represent employment and marital status trajectories rather than using the estimates of partial membership to assign individuals deterministically to one profile or pure type, as some scholars have done (e.g., Berkman, Singer, and Manton 1989; Cassidy, Pieper, and Carroll 2001).