Conscientiousness, Extraversion, College Education, and Longevity of High-Ability Individuals

A. Compulsory Schooling Laws

A popular method of establishing the effect of education on health and longevity is using changes in compulsory schooling laws as an instrument for years of schooling. Results based on this approach differ. Lleras-Muney (2005) finds a large effect of schooling on mortality for the United States, but Mazumder (2008) shows that these effects are not robust to including state-specific time trends. Similarly, analyses by Meghir, Palme, and Simeonova (2018) and Albouy and Lequien (2009) do not find a beneficial effect of schooling on longevity for Sweden and France, respectively. In contrast, van Kippersluis, O’Donnell, and van Doorslaer (2011) use Dutch compulsory schooling laws and find a beneficial effect on mortality for men.⁸

These disagreements may be related to possibly different effects of education for different countries and cohorts. Issues with statistical power and weakness of instruments are other possible explanations. In fact, at least for France and the United States, the instrument appears weak (Galama, Lleras-Muney, and van Kippersluis 2018). Other limitations of the method include possible confounding effects from other progressive reforms, such as public health improvements and school lunch programs, and relying on identification of the effect for a specific group of marginal dropouts from a compulsory school. The population of marginal school dropouts may have a wide range of IQs but can be expected to have lower IQ levels on average. My work complements these results by focusing on a different education level (college education) identified for a different population (high-ability people).

B. Draft Avoidance

Unlike the effect of compulsory education, the effect of college education on longevity is largely unexplored, perhaps since suitable natural experiments are less readily available. An exception is a paper by Buckles et al. (2016), who use the state-by-cohort-level mortality rates of men and identify a strong and statistically significant effect of college graduation on longevity from Vietnam War draft avoidance. Limitations of the Buckles et al. (2016) study include that it identifies the effect only for those who choose more education as a result of the risk of draft, so it is difficult to separate the effect of college from the effect of being drafted to serve in Vietnam.

My work further explores the effect of college education on longevity and complements the paper by Buckles et al. (2016) in a number of ways. First, I use a different methodology, discussed below. Second, I study a different population, and my sample includes women. Third, rather than looking at binary mortality over a 25-year period, I study mortality in continuous time over 56 years of life, which allows me to estimate the survival function and, as a result, the effect on life expectancy and the value of remaining life.

C. Twin Studies

This literature uses twin fixed effects as a source of identification. Studies typically use the total number of years of schooling to maximize statistical power, so they measure the average effect of schooling across various education levels. Very few twin studies are long enough and have a large enough sample size to estimate the effect of education on longevity reliably. Results based on these few studies are mixed.

In particular, based on Swedish twin data, Lundborg, Lyttkens, and Nystedt (2016) find strong effects of education on longevity for both men and women. Results based on Danish twin data vary. Madsen et al. (2010) and Behrman et al. (2011) find no effects, while, based on a different methodology, Van Den Berg, Janys, and Christensen (2015) find an effect on mortality for men but not for women. Savelyev et al. (2022) use U.S. Minnesota Twin Registry (MTR) data and demonstrate a statistically significant effect of schooling on mortality for men.⁹ Again, these different results may be partly related to differing effects of education for different countries.

Limitations of twin studies include a possibility of unobserved confounders behind estimated effects; large attenuation bias due to measurement error in education, which is amplified by taking differences across twins; and difficulties in generalizing results outside of twin populations.

This work complements results of this literature by adding additional evidence for longevity in the United States, which we know very little about from twin studies (Savelyev et al. 2022). Moreover, I study a specific policy-relevant education margin (college education) rather than one year of schooling averaged across different margins.

C. Structural Equation Modeling

An alternative approach to causal inference is to model explicitly the interdependence between observed and unobserved confounders, education, and health-related outcomes (for example, Conti and Heckman 2010). I complement these results by studying longevity, an objective outcome that incorporates all observed and unobserved inputs to the health production function and a unique margin for the quantity rather than quality of life (Becker 2007; Galama and van Kippersluis 2015; Murphy and Topel 2006). As I find no selection into education level on noncognitive skills in this paper, whereas Conti and Heckman (2010) find strong selection, we can conclude that selection on non-cognitive skills strongly depends on the range of IQ. I also extend the methodology by accounting for multidimensional latent noncognitive skills combined with modeling unobserved heterogeneity that is not accounted for by latent noncognitive skills and find that ignoring such unobserved heterogeneity leads to a sizable bias.¹⁰

Following the structural equation approach, Bijwaard, van Kippersluis, and Veenman (2015) and Bijwaard et al. (2015) study the relationship between IQ, education, and mortality. The authors use data from the Netherlands and find an effect of education on mortality and a strong selection into education based on IQ. Bijwaard, van Kippersluis, and Veenman (2015) extend the framework by Conti and Heckman (2010) by accounting for left truncation. The authors note that they could not incorporate a model for noncognitive skills into their analysis due to data limitations, which could possibly bias their results. In contrast, my model incorporates multidimensional non-cognitive skills. After showing that left-truncation issues are negligible for the Terman data, I extend the framework outlined by Conti and Heckman (2010) by accounting for unobserved heterogeneity that is not captured by latent noncognitive skills and show that the lack of this control leads to a large bias. This paper complements that of Bijwaard, van Kippersluis, and Veenman (2015) not only methodologically, but also substantively. First, I study a different population from a different country.¹¹ Second, while the authors look at the effect of completing any schooling above primary school, I examine the effect of college education. Third, while Bijwaard, van Kippersluis, and Veenman (2015) emphasize the effect of IQ, I concentrate on the effects of noncognitive skills conditional on high IQ.

Related work by Hong, Savelyev, and Tan (2020), based on the Wisconsin Longitudinal Survey data, also suggests an effect of college education on longevity for men, but that paper is based on a different population (high school graduates from Wisconsin) and focuses on estimating contributions of multiple behavioral mechanisms to the total effect on mortality, something that I could not do in this paper due to sample size and data limitations. Unlike this paper, Hong, Savelyev, and Tan (2020) study neither effects of the Big Five nor their confounding role.

Limitations of structural equation methods include reliance on additional structural assumptions and generalizations of the conditional independence assumption. However, using semiparametric models and accounting for unobserved heterogeneity, as I do in this paper, makes these assumptions more plausible. Moreover, my assumptions are consistent with placebo tests.

D. Propensity Score Method

Bijwaard and Jones (2019) use the propensity score method and find that education hardly affects mortality for men born in 1944–1947 in the Netherlands. In a separate paper, Bijwaard et al. (2017) estimate a case-specific education gradient in mortality by combining the propensity score method with family fixed effects. The authors use data on men with brothers from the Swedish Military Conscription agency (1951–83) and find that improving education by one level leads to three to ten additional months of life between ages 18 and 63 and that most of the mortality gains are attributable to mortality from causes that are unrelated to disease (traffic accidents, suicides, etc.). The main limitation of these studies is their reliance on the conditional independence assumption.¹² Substantively, my work complements these papers by estimating the effects of both education and noncognitive skills and studying a very different population.

E. Epidemiological Studies of Schooling, IQ, and Longevity

Batty and coauthors use Swedish conscripts data and show that IQ is associated with all-cause mortality and with mortality from heart disease and suicide and that controlling for education markedly attenuates these results (Batty et al. 2009b). Also, Batty et al. (2007b) show that adolescent IQ is robustly related only to skin cancer. Finally, Batty et al. (2009a) show that IQ has a preventative effect on injury mortality. These results are in line with the rest of the literature, which shows that IQ is negatively associated with mortality and that controlling for education attenuates the association (see Batty et al. 2007a and Calvin et al. 2011 for reviews).

Another source of attenuation is family and genetic influence. Næss et al. (2012) study the confounding effect of a family factor shared by siblings on the association between education and case-specific mortality in adulthood and confirm that shared family factors confound the effect, with varying magnitudes. The authors used data on all Norwegians born 1940–59 who had one or more siblings within the cohort and were alive in 1990.

Given that, based on epidemiological studies, IQ is expected to be a major confounder, it is useful that I am able to control for it twice: (i) by using a high-IQ sample and (ii) by using early IQ as a control. My work complements these studies by estimating the effect of education and socioemotional skills conditional on high IQ.

A. A Model of Schooling and Longevity

I use the full maximum likelihood estimator and the expectation-maximization (EM) algorithm to estimate a system of equations that model relationships between latent noncognitive skills Θ, IQ, schooling D, and the hazard of death λ(t).²² Estimation is conditional on background variables X, which I omit from the equations for the sake of brevity.²³

Let D be a categorical choice of the highest education level obtained in life that takes values from 1 to 3: (1) high school graduate, (2) some college education, and (3) bachelor’s degree or above.²⁴ Denote three binary variables for education levels as D_k = 1[D = k], k = 1, 2, 3. In the ordered logit model, the latent propensity for education, D*, determines the choice of schooling by the following rule: D = 1 if D* ≤ δ₁, D = 2 if δ₁ < D* ≤ δ₂, and D = 3 if D* > δ₂, where δ₁ and δ₂ are thresholds to be estimated and ε_D is logistic distributed. D* is affected by personality Θ and IQ: (1) where ε_D is an idiosyncratic error term, and μ_D is unobserved heterogeneity in D*.

A mixed proportional hazard (MPH) model with nonparametric baseline hazard function λ₀(t) (like in the Cox model) relates the hazard of death after age 30, λ(t), to non-cognitive skills, IQ, and binary variables for education levels, D₁ and D₂: (2) where term μ_λ accounts for unobserved heterogeneity in λ(t).²⁵

To model unobserved heterogeneity μ_D and μ_λ I use the semiparametric heterogeneity model approach, which has been shown to be effective for accurate estimation of both duration models (Heckman and Singer 1984), and recursive nonlinear systems with endogenous variables (Mroz 1999). This method is free from any specific functional form assumptions about the distribution of unobserved heterogeneity. Terms μ_D and μ_λ take on values μ_D1, …, μ_DK, and μ_λ1, …, μ_λK with shared probabilities p₁, …, p_K, so that . The number of latent classes K is determined by the econometrician to maximize statistical fit while avoiding empirical nonidentification (for example, Cameron and Trivedi 2005, p. 624–5).

I relate latent noncognitive skills to their noisy measures through measurement system described by Equation 3, which is necessary for identification of the factor model. The measurement system, which allows for correlations among latent skills, can be described by: (3) where M is a vector of measures generated by a vector of latent factors Θ, α is a matrix of factor loadings that is empirically justified in Online Appendix A, and η is a vector of error terms.

I allow unobserved heterogeneity μ to be correlated with IQ and Θ. Latent noncognitive skills and IQ are linked to probabilities p₁, …, p_K, through a multinomial logit model: (4)

Using multinomial logit is a standard way to allow for correlation between unobserved heterogeneity and covariates (for example, Huang and Bandeen-Roche 2004). For background controls other than IQ, I use the standard assumption of orthogonality between background controls and unobserved heterogeneity (Van den Berg 2001). The likelihood function for the model is presented in Online Appendix C.

Identification of the model comes from well-established identification of its standard components. I make several standard normalizations and assumptions.²⁶ These include setting the variance of each component of vector Θ to one; setting the sign of the effect of each factor on its first measure so that the interpretation of the factor is positive;²⁷ assuming mutually independent error terms, η and ε_D; making a number of exclusion restrictions in matrix α, which are justified by exploratory and confirmatory factor analysis documented in Online Appendix A.

Second, the recursive nature of Equation 2 relative to the ordered logit Model 1 also contributes to identification by providing a useful exclusion restriction. Schooling decision appears in the right-hand side of the MPH Model 2, but realized life duration, a future outcome, does not appear in the schooling Model 1.²⁸ Our control for unobserved heterogeneity μ allows us to control unobserved characteristics that predict both schooling and longevity. Subjects do not know exactly how long they are going to live when they make schooling decisions, but they may have expectations based on personal characteristics, such as general health. The schooling model is conditional on predictors of longevity, including those unobserved to the econometrician (unobserved heterogeneity). Identification of the semiparametric heterogeneity model that allows for covariate effect on unobserved heterogeneity is shown by Huang and Bandeen-Roche (2004).

Under the assumptions of the model we can interpret results as causal. It is always safe to interpret results as associations conditional on a large number of observables and unobservables. Confounding is possible, but given the quality and wealth of controls, we can expect it to be minor. The first powerful control is selection of the sample on high IQ, which makes all subjects similar with regard to a known major confounder. The second is IQ itself, which is used as a regressor in case it still influences outcomes in such a selective sample. The third is a large set of observables, including multiple early health measures, early educational investments by parents and private tutors, parental background, and cohort. The fourth is a set of well-established latent noncognitive skills represented by the Big Five taxonomy. The fifth is the control for permanent unobserved heterogeneity that is not accounted for by latent skills.²⁹ After taking full advantage of uniquely detailed data, it is hard to think of important confounders that remain uncontrolled.

The problem of reverse causality, which is estimating the effect of longevity on education instead of the effect of education on longevity, is unlikely. Individuals do not know their exact longevity when they make schooling decisions, but they may act on their longevity expectations. I do the following to account for such expectations: (i) controlling for several observed measures of early health of the individual and detailed family background; (ii) using the sample conditional on survival to age 30, thus making the reverse causal case unlikely, since those who died as a student or soon after graduation are excluded;³⁰ and (iii) designing unobserved heterogeneity μ to account for otherwise uncontrolled health-related shocks, attitudes, endowments, or family background.

B. Modeling Health-Related Outcomes and Life Satisfaction

To understand the mechanisms behind the effects of skills and education on longevity, I estimate parameters of the following equations for a number of health-related and life satisfaction outcomes H_p, p = 1,…, P, simultaneously with Equations 1, 3, and 4: (5) where H_p* is defined differently for binary and continuous outcomes. For binary outcomes, H_p* is a latent propensity to make a choice in a logit model. For continuous outcomes, H_p* = H_p in a linear model. Letter ξ_p represents the error term. Other notation is the same as in Equation 2.

Health and health behaviors are likely mechanisms behind the effect of education on longevity and can, as a part of decomposition analysis, be added to the longevity Equation 2. This, however, would change the interpretation of education coefficients. Instead of estimating the total effect of education, as I do now, I would estimate a direct effect conditional on its mediators. This study lacks both statistical power and data about several important mechanisms for a satisfactory decomposition analysis. For example, Terman data have no information about both early- and midlife smoking, an important mechanism (de Walque 2010). Therefore, I limit my analysis to suggesting a number of observed mechanisms. I estimate how education affects variables that can be expected to affect longevity, such as heavy drinking. I leave full decomposition analysis to Hong, Savelyev, and Tan (2020), who use different data with a larger sample size and more detailed observed health behaviors.

C. Placebo Tests

Additional support of the model comes from placebo tests. Instead of outcomes, such as longevity or health at midlife, for which I find strong effects of education, for placebo tests I use eight early health measures, such as low birth weight. Obviously, college education cannot possibly affect early health, but if my results are driven by unobserved third factors, such as unobserved health endowments, it is possible to get statistically significant “effects” of education on early health, which would be an indication of a model that misses important confounders and only identifies associations. Estimating Equation 5 simultaneously with Equations 1, 3, and 4 and using early health outcomes as H_p, I find that none of the college education thresholds affects any of the eight measures of early health.³¹

A. Nonparametric Results

Figure 1 shows the Kaplan–Meier survival curves by education and sex conditional on survival to age 30. For males, higher levels of education correspond to higher probability of survival. For instance, while for high school graduates the survival rate to age 80 is about 22 percent, for men with at least a four-year college degree the survival rate is 56 percent. The log-rank test for the equality of survivor curves is rejected (p = 0.0000).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Kaplan–Meier Survival Function by Education

Notes: Probability of survival is conditional on survival to age 30. Education groups are mutually exclusive and refer to the highest level of education obtained in life; p-values are shown for the log-rank test for the equality of survivor functions. Calculations are based on the Terman data.

For women, we can see no difference between the Kaplan–Meier survival curves by education level, and the log-rank test cannot be rejected (p = 0.9874). This result is preserved when I estimate the main model, which controls for both observed and unobserved heterogeneity. Cognitive and noncognitive skills show no effects for women either, either individually or jointly. These sex differences are consistent with results found by Savelyev and Tan (2019), who show that effects of education and noncognitive skills on health behaviors are substantially more numerous and stronger for men than for women of the high-IQ population. Conti and Heckman (2010) study the heterogeneity of the effect of education on health-related outcomes, including smoking, poor health, and obesity, and find that the effect of education for men of the general population tends to become stronger with IQ, while for women it tends to become weaker. These tendencies are consistent with a strong effect for men and no effect for women for the Terman sample of high-ability individuals.

B. Model-Based Results by Sex

As I find no relationships of interest for women, the rest of the paper focuses on results for men. Corresponding statistically insignificant main model results for women are presented in Online Appendix B.

C. Main Model Estimates and the Role of Controls

Table 4 contains estimates of the main model for men (see Model 1), as well as estimates of simplified models (see Models 2–5).³² Column 1 shows coefficients of the main MPH model of mortality. Both incomplete college and high school education lead to higher mortality relative to completing a college degree. The estimates suggest that high school graduates are 2.6 (= exp(0.974)) times more likely to die at any moment of time than those with a bachelor’s degree, and those with some college education are 1.7 (= exp (0.525)) times more likely to die compared to college graduates.

We can also see that conscientiousness and extraversion show negative direct effects of similar magnitude on the hazard of death. As effects of conscientiousness and extraversion on education are at best weak, direct effects of these skills are close to their total effects.³³ I do not find a statistically significant effect of openness, but positive sign of the estimate is consistent with adverse effects of openness on health behaviors and life satisfaction, which is documented below.

Column 2 shows coefficients for the main ordered logit model of education. The only statistically significant result is a positive coefficient for IQ. It is remarkable that IQ shows a strong effect on education even though subjects were selected on IQ above 140. One reason for finding no statistically significant effect of noncognitive skills on education is that I control for unobserved heterogeneity. A comparison of Columns 2 and 4 shows that without controlling for unobserved heterogeneity, the model shows a statistically significant relationship between conscientiousness and education. This finding is likely driven by an unobserved confounding factor.

Another reason why controlling for unobserved heterogeneity is important is that estimated coefficients differ substantially with and without such a control. These results are in line with a well-known property of the MPH model that ignoring unobserved heterogeneity leads to a downward (by absolute value) bias of estimates (for example, Van den Berg 2001). Indeed, the average downward bias due to omitting μ in the MPH model is about 20 percent (compare statistically significant coefficients in Columns 1 and 3).

Models 3, 4, and 5 show results of incremental omission of variables from the model: noncognitive skills, cognitive skills, and background variables. Interestingly, as a result of successive changes from Model 1 to Model 5, coefficients decrease and then increase again and get closer to coefficients in the main model, so that the coefficient for high school degree in Model 5 is biased 16 percent relative to that in Model 1, while the coefficient for some college is biased only by 3 percent. This peculiar result could not be foreseen prior to estimation and might be specific to this particular population.

Another unforeseen result of Table 4 is that the selection effect is at best small relative to selection found in the literature for the general population. The omission of all non-cognitive skills changes estimated education coefficients by 5 percent on average.³⁴ The likely reason for this result is either weak or nonexistent links between noncognitive skills and education for this population, as we see in Column 2. The omission of IQ, a single skill variable that does matter for education, changes education coefficients by 6 percent on average.

Table 4 also offers evidence that the proportional hazard assumption behind the main MPH model is valid. Column 9 shows p-values for the proportional hazard (PH) test. The PH hypothesis cannot be rejected for either individual variables or jointly.³⁵

Results of the MPH model of all-cause mortality are robust to the exclusion of individuals with specific major causes of death, as documented in Online Appendix B.³⁶ In another robustness check documented in Online Appendix B, I demonstrate that latent factors of agreeableness and neuroticism that are proxied by 1940 measures are not predictive of longevity, both individually and jointly, but greatly increase the degrees of freedom. These factors can be dropped for greater parsimony and to avoid mixing childhood and adulthood skills. The lack of association between these two factors and longevity is in line with previous research based on the same data (Martin, Friedman, and Schwartz 2007). Also, the literature provides no strong prior for either positive or negative effects of these traits (see Chapman, Roberts, and Duberstein 2011).³⁷

D. Model-Estimated Survival Curves and a Comparison with the General Population

Figure 2 shows model-predicted survival curves by education and compares them with a survival curve for white men of the general population, the 1910 birth cohort. First, predicted curves are very similar to nonparametric Kaplan–Meier curves in Panel A of Figure 1, which is another representation of the result that for this sample the combined selection effect on observables and unobservables is negligible.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Model Prediction of the Survival Function by Education and Comparison with General Population, Males

Notes: Calculations are based on the Terman data and Census for Disease Control (CDC) (Arias 2012) for the survival of the 1910 cohort of white men.

Second, we can see a sizable difference in counterfactual survival by education. Conditional on survival to age 30, the chances to survive to age 86 is 40 percent for college graduates, 22 percent for those with some college education, and 10 percent for high school graduates.

Finally, we can see that the general population of 1910 birth cohort of men has about a 17 percent chance to survive to age 86, which is higher than for Terman high school graduates but lower than for Terman subjects with some college education. The result that Terman subjects with a high school degree live shorter than men from the general population, despite having superior intelligence and education, is a puzzle. I address this puzzle below in a section on the mechanisms and provide evidence that this result is consistent with life dissatisfaction and adverse health behaviors of Terman high school graduates likely arising from a major mismatch between their extraordinary cognitive ability and types of careers available to high school graduates.

E. Effects of Skills on the Hazard of Death

I find that both conscientiousness and extraversion have statistically significant effects on longevity and that these effects are not driven by the choice of education, as we can foresee from Model 1 of Table 4.

A nonnegative association between conscientiousness and longevity is consistent with a substantial literature (see Kern and Friedman 2008 for a survey). In particular, Friedman et al. (1993) found a beneficial association for high-ability men based on the same data. The authors used the Cox PH model with one to seven degrees of freedom. I show that this qualitative result survives controlling for detailed observed and unobserved heterogeneity. My total effect estimate implies that an increase in conscientiousness by one standard deviation (SD) leads to a 17 percent decline {[1 – exp (–0.182)] · 100%}in the hazard of death, a result that is statistically significant at the 5 percent level.³⁸

While extraversion generally shows mixed associations with longevity, studies that use sociability and related facets of extraversion, as in this paper, find positive associations with longevity (see Chapman, Roberts, and Duberstein 2011 for a survey). Previous studies based on the Terman data did not find any associations between extraversion/sociability and longevity (for example, Friedman et al. 1993), but they did not account for IQ, background controls, unobserved heterogeneity, or measurement error in measures of personality. Thus, the strong effect of extraversion on longevity that I find is a novel result for a high-ability population and a contribution to the ongoing debate about the role of extraversion in longevity production. One standard deviation increase in extraversion leads to a 13 percent decline {[1 – exp(–0.135)] · 100%} in the hazard of death, a result that is statistically significant at the 10 percent level.³⁹

I find no statistically significant effects of openness and IQ on longevity. As mentioned, I also find no evidence of longevity effects of agreeableness and neuroticism.

F. Effects of Skills and Education on Life Expectancy

Panel A of Table 5 shows counterfactual life expectancies at age 30 by education level (see Column 1). It also shows that the model-predicted sample average is close to the sample average implied by nonparametric Kaplan–Meier estimates. Column 2 shows differences between counterfactual life expectancies by education level relative to high school education. Conditional on survival to age 30, some college education adds 4.7 additional years of life, while a four-year college degree or above adds 9.8 additional years of life, a remarkably high benefit.⁴⁰ The same column shows that changing conscientiousness and extraversion by one standard deviation leads to 2.0 and 1.6 additional expected years of life, respectively.

The baseline survival function is nonparametrically estimated as part of the MPH model until age 86, which allows estimation of the main model survival function until that age. However, for a life expectancy calculation I need full support of the longevity distribution. To solve this problem, I extrapolate the baseline survival function to age 100 using census data for men of the cohort born in 1910. As a robustness check, I compare this extrapolation with a linear extrapolation. As the upper tail of the survival curve is expected to be convex, the linear extrapolation is, arguably, an upper bound. I also use a Gompertz–Makeham extrapolation (for example, Brown, Liebman, and Pollet 2002), an alternative approach to using census survival rates. Results based on these three methods are similar. The standard deviation of estimates based on these three methods of extrapolation is shown in Column 7 and can serve as an approximate estimate of the error introduced by the extrapolation. For the effect of bachelor’s degree (9.8 years), two standard deviations give us a confidence width of about one year.⁴¹

G. Effects on Skills and Education on the Value of Remaining Life

In Panel B of Table 5, longevity benefits are evaluated in 2016 USD for a statistical person, using the Murphy and Topel (2006) methodology. The present value of additional life expectancy at age 30 compared to high school graduates is $500,000 for some college education and $930,000 for four years of college. These numbers represent the value of increase in life expectancy induced by education for a contemporary 30-year-old man if effects estimated in this paper were applicable to him. This high longevity gain is estimated as a lower bound, as it does not include the value of increased quality of life in the form of superior general and mental health over the lifespan.⁴² The valuation of one standard deviation of conscientiousness and extraversion are $160,000 and $110,000, respectively. Thus, the effects of education and skills are economically significant. Unlike Panel A, Panel B shows almost identical results for all three methods of extrapolation. The robustness of the value of remaining life comes from discounting the future: the extrapolation only affects ages above 86, which are heavily discounted at age 30.⁴³

H. Effects of Education and Skills on Health and Health-Related Outcomes

Table 6 shows effects of education and skills on health and health-related outcomes. Models presented in the table are similar to the longevity model discussed above. Just as for longevity, Model 5 for behavioral and health outcomes is estimated jointly with the Models 1, 3, and 4 conditional on the same observable and latent controls. Estimation is performed outcome by outcome.

The observed effect of education on longevity is supported by several plausible mechanisms documented in Table 6. Indeed, we observe statistically significant positive effects of education on marriage, memberships in organizations, occupation as a professional, and general health and negative effects on heavy drinking of alcohol and mental difficulty.

Conscientiousness is known to have a strong association with beneficial health-related outcomes (Roberts et al. 2007). Conscientious people tend to delay gratification, plan for the future, and act towards their goals (John and Srivastava 1999). These characteristics boost health-beneficial choices. Not surprisingly, Table 6 shows that conscientious subjects are less likely to drink heavily, are more likely to get married, and have better general and mental health. These results suggest mechanisms behind the relationship between conscientiousness and longevity and support a causal interpretation.

The effect of extraversion on longevity is also supported both theoretically and empirically. Extraversion, which is a propensity to be active and social, may help create social skills and networks of friends, which, in turn, boost both mental health and earnings.⁴⁴ ,Gensowski (2018) shows a strong effect of extraversion on earnings with the same data. However, it is unclear from the literature whether earnings (or, in general, wealth) affect health and longevity and, if yes, whether the effect is positive, negative, or nonexistent.⁴⁵

As for other mechanisms, Table 6 shows that extraversion is associated with one adverse health behavior, heavy drinking, which might be a side effect of greater socialization (Column 1). However, positive effects on general and mental health suggest that negative contribution of heavy drinking is compensated by other channels. I document a positive effect of one such possible channel, stable marriage (Column 2).

As is also discussed in a companion paper on health behaviors by Savelyev and Tan (2019), openness for high-ability men shows no single beneficial health-related outcome and a number of adverse ones. In particular, we see from Table 6 that heavy drinking increases, marriage probability decreases, and mental and general health decrease. However, the effect on general health fades after age 40, which might be the reason why we do not observe any statistically significant adverse effect of openness on longevity.

I. Compression of Morbidity

The beneficial effects of education, conscientiousness, and extraversion not only on longevity but also on health over the lifecycle I find are consistent with the controversial compression of morbidity hypothesis—whether the fraction of the lifespan spent in good health increases with longevity (for example, Cutler, Ghosh, and Landrum 2014). Compression of morbidity has important practical implications. First, it implies that the highly valuable superior quantity of life is accompanied by greater quality of life, thus making increases in longevity even more beneficial for individuals (Murphy and Topel 2006). Therefore, an increase in longevity though investments in skills and education does not create an army of disabled people, but people who both are healthier and live longer. Finally, the compression of morbidity lowers the cost of healthcare relative to the case of no compression.

These results are consistent with the compression of morbidity. As we see from Table 6, the effect of schooling and extraversion on general health are economically and statistically significant and persistent for ages 40–76.⁴⁶ Conscientiousness increases general health at ages 40 and 62, an effect that declines with age and becomes statistically insignificant by age 76. Schooling increases mental health at age 40. Conscientiousness and extraversion increase mental health at both ages 40 and 50. Thus, greater longevity that is induced by growing education and noncognitive skills is associated with superior general and mental health in both middle- and advanced ages.

Savelyev and Tan (2019) analyze similar health-related outcomes in a companion paper, which is based on the same Terman data but does not study longevity. Apart from other contributions relative to earlier works by psychologists (for example, Friedman et al. 1995; Martin, Friedman, and Schwartz 2007), Savelyev and Tan (2019) demonstrate that the key conditional associations between skills, education, and health-related outcomes survive controlling for familywise error, an error that originates from testing multiple hypotheses. In this paper, the aim of modeling health and health behaviors is different: I support the compression of morbidity hypothesis and provide evidence on the mechanisms using models that are fully consistent with the main longevity model.⁴⁷ In particular, Table 6 demonstrates results conditional on unobserved heterogeneity μ, which was not controlled for in previous literature (for example, Friedman et al. 1995; Savelyev and Tan 2019).

J. Effects of Education and Skills on Life Satisfaction

Table 7 is based on the same model as Table 6 but with different outcomes. It shows effects of education and personality on life and job satisfaction. Columns 1–8 suggest that a lower level of education, especially a high school degree, leads to lower probability of job and life satisfaction, lower control over choice of occupation, sense of wasted intellectual abilities, and hindered life due to inadequate education.

Another result of this table is that conscientiousness is associated with greater work satisfaction, availability of outlets for mental capabilities, and job satisfaction, whereas openness has negative effects on the very same outcomes. Combined with results of Table 6, which showed increased heavy drinking combined with negative effects on marriage, mental, and general health, we can conclude that openness is counterproductive for this cohort with respect to multiple observable health-related outcomes. In line with these results, the estimate of the total effect of openness on the hazard of death is positive, but not precisely determined.⁴⁸

K. Understating Sizable Effects of Education on the Hazard of Death and Life Expectancy

The first question is whether estimated sizable effects represent causal relationships. As I argue in Section IV, while confounding and reverse causality cannot be completely ruled out, for this particular data they are unlikely to drive the results.

The second question is how to explain sizable effects of education for men: 4.7 additional years for some college education and 9.8 years for a bachelor’s or above conditional on living to age 30. The estimated effect is comparable with the effect found by Lleras-Muney (2005), who uses U.S. compulsory schooling laws as a source of identifying variation and finds that an additional year of compulsory schooling causes a 1.7-year increase in life expectancy conditional on living to age 35. Conditional on a strong assumption of a constant effect of schooling over different educational thresholds, this would imply about seven years for an additional four years of education, which is the difference between four-year college and high school. Since we find no effect for women, our effect of college for a person of random sex is about five years (0.5 × 9.8 + 0.5 × 0). This is less than the seven years implied Lleras-Muney’s result, which is based on a sample of pooled sexes. As mentioned, it is a strong assumption that the effect of compulsory schooling for a marginal school dropout is the same as the average effect of college education for a person with an IQ above 140, but it is of interest that our complementing results are of comparable magnitude.

Darden, Gilleskie, and Strumpf (2018) show that the longevity cost of lifetime smoking is about 4.3 years, about a half of our estimated effect of education. However, we can expect education of people with high ability to affect longevity not only through smoking but also through many other mechanisms of longevity (for example, Savelyev and Tan 2019). A number of suggested mechanisms are documented here: reduced heavy drinking, stable marriage, increased social interactions, higher likelihood of having a professional occupation, superior mental and general health, and multiple measures of superior job and life satisfaction. As I argue below, job and life satisfaction may play a special role for this unusual cohort.

Two traditional explanations of the effect of education on health are (1) productive and (2) allocative efficiencies. Productive efficiency refers to larger health output from given amounts of endogenous (choice) inputs (Grossman 2000). Allocative efficiency refers to better allocation of investments due to better information about the true effects of health investments (Kenkel 2000). Cutler and Lleras-Muney (2008) offer another classification of mechanisms: (i) higher income, (ii) safer jobs, (iii) higher value of life, (iv) better health knowledge and superior cognitive skills, (v) lower discount rate and increased risk aversion, (vi) higher rank in the society, and (vii) larger social networks, which provide financial, physical and emotional support. Sociologists (Ross and Wu 1995) suggest three categories of explanations: (A) work and economic conditions, (B) social-psychological resources, and (C) health lifestyle.

The mechanisms mentioned above (1, 2, i–vii, A–C) cover many possible channels and partially overlap. The Terman sample provides evidence consistent with the following channels among those named above: higher income/economic conditions (Gensowski 2018), higher rank in society (measured by professional occupation), higher value of life (measured by greater life satisfaction), larger social networks/social-psychological resources (measured by number of memberships in organizations), and healthier lifestyles (measured by no heavy drinking and stable marriage). However, the channel though safer jobs is unlikely for this sample.⁴⁹

One channel related to effects of education on life and job satisfaction documented in Table 7 deserves special attention for Terman subjects. This mechanism is separate from productive and allocative efficiencies and can be viewed as a special case of the effect through the value of life. I call this suggested mechanism “human potential–education matching.” One role that college education plays is that it helps select individuals with high potential, such as those with high IQ, and provides them with complementary skills that open doors to markets for high-skilled jobs. These high-skilled jobs provide a level of challenge that is a good match for people with high potential. There is anecdotal evidence that people who fail to realize their high potential may end up abusing substances or otherwise neglecting their health and living shorter lives. I provide statistical evidence consistent with this intuition, suggesting a channel through which IQ can increase the effect of education on health. So far, the literature suggested the opposite effect due to substitution of IQ for education in health production (Auld and Sidhu 2005).

The Terman data are consistent with the hypothesis that a mismatch between education and IQ decreases health and longevity. First, the survival curve of Terman high school graduates is below the survival curve of the general population of U.S. white men of the same cohort (see Figure 2), even though an average white man from this cohort had both lower intelligence and lower education than Terman’s high school graduates.⁵⁰ I suggest an explanation for this paradox: Terman’s high school graduates had a larger mismatch between human potential and their education than an average man from the general population, which likely resulted in less self-care for their health.

Second, Table 7 shows effects of high school and some college education relative to completing a four-year college degree on job and life dissatisfaction. We can see that lower education levels lead to a dislike of current occupation (Columns 1, 3, 7, and 8), lack of control in the choice of work (Column 2), lack of intellectual challenge (Columns 5 and 6), and hindered life due to inadequate education (Column 4). All these effects take place for those with a high school degree, and the majority of them also hold for those with an incomplete college education. These satisfaction results are consistent with effects of education on life outcomes from Table 6—a smaller likelihood of belonging to professionals (Column 4), a higher likelihood of heavy drinking (Column 1), and greater difficulties with mental and general health (Columns 5–9).

Sociologists have investigated the hypothesis that education improves health through work satisfaction, suggesting that high-skilled jobs tend to be more rewarding and provide superior control over an employee’s life. However, the results of this literature are inconclusive, and intrinsic work rewards among the employed are found to be small, inconsistent, and not always positive (Ross and Wu 1995). Perhaps one reason behind this controversy could be that it is good matching between education and personal potential that matters, rather than the level of education alone. Education that overmatches personal potential could be counterproductive for life satisfaction; the person could become overeducated for a number of jobs that could otherwise be a good fit, but still not a good match for challenging high-level jobs. In addition, jobs that require more education do not necessarily provide more control over a person’s life. For instance, a self-employed contractor with a high school degree can be expected to experience superior control over their life than a low-level office worker with a college degree. However, in the case of Terman high school graduates, the mismatch between intellectual ability and education level is major and one-directional, and the evidence for the job dissatisfaction effect is major.

L. Placebo Tests

As part of the specification testing of my model, I run placebo tests for eight early health outcomes. I use the same model (1, 3, 4, 5) as I use for health-outcomes in midlife. This model has the same structure as my main longevity model except that instead of the hazard of death I use early health variables as outcomes one by one. Despite no control for family wise error rate, none of the p-values are statistically significant, even at the 10 percent level (see Table 8).

M. Robustness to a Two-Step Estimation

One common concern about using a full maximum likelihood model in research involving latent skills is that, in the case of model misspecification, latent skills might be primarily driven by outcomes rather than the measurement system. I show in Online Appendix B that this concern is not pertinent to my results. Following Heckman, Humphries, and Veramendi (2018), I estimate only the measurement system in the first step and then estimate the maximum likelihood model while fixing factor loadings of the measurement system to the values estimated in the first step. The results are almost identical to those based on full maximum likelihood.⁵¹ With this concern shown to be alleviated, full maximum likelihood is a preferred estimator due to its superior efficiency and the lack of a need to correct standard errors for a two-step estimation.

N. Data Limitations

I use a unique and detailed data set that has early measures of skills and a long follow-up, but it is not without disadvantages. One disadvantage is its modest sample size. In addition, a number of important potential mechanisms, such as smoking and exercise, have not been measured in early life or in midlife.⁵² Because of the modest sample size and missing data on potentially important mechanisms, I lack both statistical power and key information for decomposing the estimated total effects on longevity to a sum of underlying mechanisms. As a result, I only provide suggestive evidence based on Table 6 and 7 and leave decomposition analysis for a separate paper based on a different data set (Hong, Savelyev, and Tan 2020).

Another limitation is the lack of measurement of childhood agreeableness and neuroticism. However, I show that agreeableness and neuroticism measured in adulthood do not predict longevity, suggesting that this disadvantage is unlikely to be crucial, which is in line with the literature.

Finally, having a censoring point for mortality in 1991 is another disadvantage. I account for the right censoring in the model, leading to internally consistent estimates before the censoring point. However, calculations of life expectancy and the value of remaining life required extrapolating the baseline survival function to more advanced ages. As shown, additional error introduced by this extrapolation is estimated to be small. A likely reason for an accurate extrapolation is that it starts from an advanced age of 86.

O. External Validity

The trade-off of analyzing quality longitudinal mortality data is the need to consider whether results are generalizable to a more contemporary population. I argue that for high-ability males we can expect similar qualitative results for later cohorts. Indeed, we know from the literature that education and skills still determine lifestyles for the current population, which, in turn, still determine their longevity. Quantitatively, however, effects might be different for many reasons, including significant progress in health knowledge and public health improvements over the course of the 20th century.

Contemporary cohorts have better knowledge of the importance of health behaviors, such as smoking, maintaining a healthy diet, and exercise. For instance, while students today learn about the adverse effects of smoking at school, earlier cohorts lacked widely available consensus knowledge about smoking before the Surgeon General Report (Terry et al. 1964), at which point the Terman cohort was, on average, 54 years old.

Despite the lack of consensus knowledge about the dangers of smoking before the 1950s (de Walque 2010), substantial health information was widely distributed and available in 1922, the year when the survey began and respondents were about 12 years old. For example, “Rules of the Health Game,” a public health pamphlet that was published in 1922 and targeted children, included personal hygiene, long sleep, eating vegetables and fruits, drinking at least four glasses of water a day, and playing part of every day outdoors (Allensworth et al. 1997). Alcohol was not mentioned, but given that Prohibition took place from 1920 to 1933, the adverse effects of alcohol were well publicized at the time. Taking other drugs and obesity-related behaviors were not mentioned, but those were much less prevalent for this cohort.

The direction of quantitative change of the effect for later cohorts is unclear. On the one hand, since educated and conscientious people act on their health knowledge, the effects of education and conscientiousness might be even stronger for later cohorts than what is estimated for the Terman cohort, as health knowledge grows over time. Larger incidence of obesity and the use of narcotics for current cohorts could increase the effect. On the other hand, smaller incidence of smoking for current cohorts may diminish the effect.

For females, the results should be considered historical since women’s emancipation over the 20th century led to a considerably wider variety of lifestyles and health behaviors for women than before. Labor markets and the purpose of education also drastically changed for women. A wider variety of lifestyles together with greater health knowledge and labor force participation could create health effects of education and skills even if they were not strong or existent for earlier cohorts of women.

Presently, intelligent people tend to finish college in larger numbers than before. However, low-income students with high human potential are less likely to enroll in college than their higher-income counterparts. Even after being accepted into college, low-income students are much more likely to choose not to attend (Castleman and Page 2014). Moreover, college dropout rates vary substantially by socioeconomic status (Bowen, Chingos, and McPherson 2009). My paper shows that it is wrong to assume that high cognition substitutes for education in producing health and longevity. Moreover, I argue that IQ may even increase the effect of education on health-related outcomes. This makes mismatches between human potential and education even worse than we may think from lost productivity considerations alone.

The effects reported here might be similar to those for a larger population of able people, not only to extraordinarily able people. Indeed, understanding the best inputs into health production, such as a healthy diet, definitely requires cognitive skills, but not necessarily the abilities of a genius. The direction of effect change once the IQ gets smaller is unclear though. Results by Auld and Sidhu (2005) suggest larger effects for lower IQs. However, results of Conti and Heckman (2010) for men and the human potential–education matching mechanism suggested in this paper allow for a decline in the effect of education with a decline of the IQ.