Professional Documents
Culture Documents
Predicting Educational Attainment: Does Grit Compensate For Low Levels of Cognitive Ability?
Predicting Educational Attainment: Does Grit Compensate For Low Levels of Cognitive Ability?
Predicting Educational Attainment: Does Grit Compensate For Low Levels of Cognitive Ability?
A R T I C LE I N FO A B S T R A C T
Keywords: This study examined the role of cognitive ability in moderating grit's association with educational outcomes.
Grit Using a large, representative sample of young adults, we estimated probit models for the probability of grad-
Cognitive ability uating from high school, enrolling in college, earning any college degree, and earning a bachelor's degree. For
Academic achievement each outcome, the effect of grit (and, alternatively, each lower-order facet) was allowed to differ flexibly with
Educational outcomes
cognitive ability. We found that grit's estimated marginal effect is largely concentrated among students at the
high and low ends of the ability distribution. The low-ability effect is more pronounced when expressed relative
to the ability-specific, baseline probability of success, and the high-ability effect increases with each successive
outcome. The findings are consistent with the notion that high-ability students adopt self-regulated learning
processes that exploit their grit, especially as educational tasks become more challenging. For low-ability stu-
dents, it appears that grit plays a compensatory role.
1. Introduction domains might hold only for individuals with high levels of cognitive
ability or metacognition. This argument was augmented in Credé
As part of a broad effort to identify factors that explain individual (2018) by the conjecture (page 4) that “(i)t is also possible that non-
difference in success, researchers have begun to compile evidence that cognitive attributes such as grit are particularly important for in-
grit is associated with a range of educational outcomes (Bowman, Hill, dividuals at the lower end of the ability spectrum.” We operationalized
Denson, & Bronkema, 2015; Datu & Fong, 2018; Datu, Valdez, & King, these ideas by exploring two non-competing hypotheses: (1) grit plays a
2016; Datu, Yuen, & Chen, 2018; Duckworth, Peterson, Matthews, & synergistic role for high-ability students—especially when educational
Kelley, 2007; Duckworth & Quinn, 2009; Eskreis-Winkler, Shulman, tasks are challenging—because they are able to adopt learning pro-
Beal, & Duckworth, 2014; Rimfeld, Kovas, Dale, & Plomin, 2016; cesses that leverage their grit; and (2) grit plays a compensatory role in
Schmidt, Fleckenstein, Retelsdorf, Eskreis-Winkler, & Möller, 2017; the learning process of low-ability (and, potentially, medium-ability)
Strayhorn, 2014; Sturman & Zappala-Piemme, 2017; West et al., 2016). students; that is, lower-ability students who successfully attain educa-
Recent meta-analytic evidence suggests grit's predictive power might be tional milestones might do so by substituting grit (and other traits) for
small and subsumed by other factors (Credé, 2018; Credé, Tynan, & ability.
Harms, 2017), yet the topic demands further attention because grit—- One reason the potentially moderating influence of cognitive ability
defined as “passion and perseverance for long-term goals” (Duckworth had not been explored prior to the current study is because the litera-
et al., 2007)—has the potential to be instilled in individuals at rela- ture often focused on genius, or eminence. Building on the work of
tively young ages (Duckworth & Gross, 2014). As a result, grit might Galton (1892), samples of high achievers were used to obtain seminal
prove to be a useful mechanism for improving rates of high school and evidence that grit is predictive of success (Duckworth et al., 2007;
college completion as well as grades, test scores, and other measures of Duckworth & Quinn, 2009) and that increased deliberate practice is a
educational achievement. mechanism by which grit affects achievement (Duckworth, Kirby,
In this study, we contributed new evidence on the grit-education Tsukayama, Berstein, & Ericsson, 2011; Ericsson, Krampe, & Tesch-
relationship by determining whether grit's effect on educational out- Römer, 1993). Those samples—which included West Point cadets, Na-
comes differs across the cognitive ability distribution. Credé et al. tional Spelling Bee participants, and Ivy League undergraduates—were
(2017) conjectured that the adaptive effect of grit found for many not suited to learning whether grit is equally predictive for low-ability
⁎
Corresponding author: 410 Arps Hall, 1945 N. High Street, Columbus, OH 43210-1172, United States of America.
E-mail addresses: light.20@osu.edu (A. Light), nencka.2@osu.edu (P. Nencka).
https://doi.org/10.1016/j.lindif.2019.02.002
Received 8 July 2018; Received in revised form 27 January 2019; Accepted 2 February 2019
1041-6080/ © 2019 Elsevier Inc. All rights reserved.
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
and high-ability individuals. In contrast, a small number of subsequent relative paucity of cognitive ability must choose between intensifying
studies of educational outcomes used large samples spanning the po- their use of other inputs, including grit, or failing to achieve. This ar-
pulation-wide cognitive ability distribution (Eskreis-Winkler et al., gument is closely related to the idea of “resource substitution” invoked
2014; Marks, 2016; Rimfeld et al., 2016; West et al., 2016). While those by Damian, Su, Shanahan, Trautwein, and Roberts (2015) and
data could be used to explore the grit-education relationship across a Shanahan, Bauldry, Roberts, Macmillan, and Russo (2014), who found
non-truncated cognitive ability distribution, we were the first to ask: empirical evidence that children from low-SES households were able to
Does grit's relationship to educational outcomes differ for low-ability, improve educational outcomes by substituting Big 5 personality traits
medium-ability, and high-ability students? for the resources they lacked. Whether substituting conscientiousness
and agreeableness for family resources or substituting grit for cognitive
1.1. Cognitive ability as a moderator of grit's relationship to educational ability, the hypothesis is that students who lack a given input exploit
outcomes their comparative advantage as a means of achieving success (Almlund
et al., 2011).
We drew from the educational psychology, personality psychology, It is important to understand that complements and substitutes can
and economics literatures to hypothesize that grit can potentially en- exist side by side. High-ability students can synergistically combine
hance educational outcomes at both the high end and the low end of the their ability and grit in the service of highly effective learning pro-
cognitive ability distribution. Turning first to educational psychology, cesses, while low ability students can compensate for their lack of
the interdependence of grit and cognitive ability can be considered in cognitive ability by relying on grit. There is little question that students
the context of self-regulated learning models that describe metacogni- who possess high levels of cognitive ability and grit will be the most
tive, feedback, motivational, and other processes by which students productive, and will typically surpass minimal educational milestones
affect their own learning (Cohen, 2012; Winne, 2013; Zimmerman, such as graduating from high school and attending college. Low-ability
1986, 1990). In describing self-regulated learners as “self-starters who students will invariably be less productive than their high-ability peers
display extraordinary effort and persistence during learning” (p. 5), no matter how much grit they possess, but those who achieve a given
Zimmerman (1990) effectively characterized them as gritty. Therefore, educational milestone are likely to do so because they made good use of
it was unsurprising that Wolters and Hussain (2015) found empirical their grit. An analogy can be made to farming, where farmers who
evidence that grit—especially perseverance of effort which, along with employ both rich land and good laborers will produce large crops, while
consistency of interest, is one of its two lower-order facets—was farmers with poor soil will often experience crop failure. However,
strongly, positively related with a host of observed indicators of self- those farmers who produce moderate crop levels despite having poor
regulated learning. Their evidence suggests that adoption of self-regu- soil will succeed because they found a way to compensate for their land
lated learning is a mechanism by which grit (or perseverance) promotes deficiency by effectively substituting labor.
academic achievement. The moderating effect of ability on the grit-education relationship
The question remains: Does the link between grit and learning might depend on the educational outcome being considered.
processes depend on students' cognitive ability levels and, if so, how? Complementarities between ability and grit will enable high-ability
Wolters and Hussain (2015) were unable to address this question em- students to be highly productive; e.g., to master the most difficult
pirically with their sample of 213 college students. Others have sug- concepts and graduate from the most competitive schools. High-ability
gested that grit is more likely to promote self-regulated learning (and, students are likely to attain more modest milestones (e.g., graduating
in turn, academic achievement) among high-ability students because, from high school) with or without the implementation of productive
relative to their low-ability peers, they have better metacognition, are learning processes, so grit might be unrelated to these outcomes; a si-
better able to recognize value and respond to feedback, and are better milar argument is made by MacNamara, Hambrick, and Oswald (2014)
able to employ learning strategies productively (Credé & Kuncel, 2008; regarding the effect of deliberative practice on outcomes that differ in
Credé & Phillips, 2011; Zimmerman, 1990). If this hypothesis holds, task predictability. In contrast, low-ability students might find it ne-
then cognitive ability not only has a direct effect on academic perfor- cessary to substitute grit for cognitive ability to attain even modest
mance, but it also enables students to engage in self-regulated learning educational outcomes. As the outcomes become more challenging,
and deliberate practice (Duckworth et al., 2011; Ericsson et al., 1993) fewer low-ability students will meet with success and the role of grit
that fully utilize their grit. will increase in magnitude relative to the unconditional success rate.
Recent studies in the economics of education made similar argu- Although the discussion thus far has focused on the contrast be-
ments in the context of production functions (Almlund, Duckworth, tween low and high ability students, estimated effects of grit are not
Heckman, & Kautz, 2011; Borghans, Duckworth, Heckman, & ter Weel, necessarily equal for low and high ability students, or absent in the
2008; Cunha & Heckman, 2008). For our purpose, we can consider a middle of the ability distribution. Increasing cognitive ability is likely to
production process in which students combine inputs, including cog- correspond to decreased reliance on resource substitution and increased
nitive ability and grit, to learn a concept. A key aspect of the technology use of synergistic, self-regulated learning processes. How (and whether)
(or learning process) by which inputs convert to output is that inputs grit's effect on educational outcomes differs for low-, medium-, and
can be complements or substitutes. Complementarities between inputs high-ability students is an empirical question—which we addressed, as
imply, for example, that increased levels of cognitive ability increase described in Section 2.3, with a model specification that allows the
the efficiency with which grit leads to output. Proponents of this pro- estimated marginal effect of grit to change with ability in a flexible
duction function framework did not elaborate on the precise reasons for fashion.
such complementarity, but it could arise for the reason we already
provided: high-ability students know how to adopt self-regulated 1.2. Existing evidence on relationships between traits, ability, and education
learning processes that make full use of their grit. Regardless of the
reason for the complementarity, the production function approach Relationships between grit, cognitive ability, and educational out-
supports the hypothesis that grit enhances learning and educational comes are a relatively new area of investigation. In contrast, an ex-
achievement relatively more for high-ability students than for low- tensive literature exists on the role of traits that are closely related to
ability students. grit, such as motivation and conscientiousness. Rather than attempt an
The production function framework also yields the opposite pre- exhaustive review of the literature, we focused on studies that are
diction: grit and ability can serve as substitutes in the learning process. particularly relevant to our analysis.
Students are assumed to behave optimally (given their current in- Conscientiousness—a Big 5 trait with facets that include orderliness
formation) in deciding how to combine inputs, and students with a and self-control (Roberts, Lejuez, Krueger, Richards, & Hill, 2014)—was
143
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
consistently found to have a strong, positive association with academic relationship with educational outcomes exists for both low-ability and
performance measures such as college grade point average. Although high-ability students; and (2) the effect of grit among high-ability stu-
counter-evidence was found on occasion (e.g., Farsides & Woodfield, dents is more pronounced for relatively challenging educational out-
2003), the preponderance of evidence based on multivariate regression comes.
analysis (Conard, 2006; Ivcevic & Brackett, 2014; Noftle & Robins, To achieve these goals, we used data from the 1997 National
2007; O'Connor & Paunonen, 2007; Poropat, 2009; Wolfe & Johnson, Longitudinal Survey of Youth (NLSY97) to construct a large sample of
1995) established that conscientiousness is typically more predictive of individuals for whom we could assess four educational outcomes: high
academic performance than are other Big 5 traits, and that its effect on school graduation by age 20, and college enrollment, completion of any
performance holds when measures of other personality traits and/or college degree, and completion of a bachelor's degree by age 26. In
cognitive ability are included among the regressors. The majority of contrast to studies that examined the grit-education relationship for
studies in this literature shared a characteristic with parallel analyses of small, select samples (Duckworth et al., 2007; Duckworth & Quinn,
the grit-performance relationship: the effect of conscientiousness was 2009), our analysis used a representative sample of students with
often estimated conditional on cognitive ability, but was not allowed to ability levels spanning the population-wide distribution. In addition,
vary with ability. our analysis used a uniform sample of individuals as they advanced
An important group of studies deviated from the conventional through successive educational levels, thus eliminating problems of
strategy by allowing estimated effects of traits to differ with ability. noncomparability across studies that focus on a single educational in-
Ganzach (2000) and Ganzach, Saporta, and Weber (2000) modeled stitution (Bowman et al., 2015; Datu et al., 2016; Duckworth et al.,
educational attainment as a function of cognitive ability, motivation 2007; Duckworth & Quinn, 2009; MacCann & Roberts, 2010: Strayhorn,
(measured as educational expectations), and the interaction between 2014) or a single public school district (Eskreis-Winkler et al., 2014;
ability and motivation. Both studies invoked what Ganzach et al. (2000) West et al., 2016).3
referred to as the “axiomatic” view that performance in academics and For each educational outcome, we used multivariate probit models
other domains is enhanced by the combination of high ability and in which the regressors included a grit score, a cognitive ability test
motivation (Vroom, 1964); both estimated positive coefficients for the score and it square, and interactions between grit and both ability
ability-motivation interaction terms, which supported the prediction of terms. This specification allowed grit's relationship with each outcome
a synergistic relationship between these factors. Logan, Lundberg, Roth, to differ in a flexible manner across the cognitive ability distribution.
and Walsh (2017) modeled college exam scores as a function of cog- Our analysis brought to bear arguments made by both proponents
nitive ability, conscientiousness, and the interaction between the two to and critics of existing evidence regarding grit's association with edu-
test the hypothesis of a positive interaction (Maier, 1955). They, too, cational outcomes. The first issue concerns how best to assess the
found evidence of synergistic effects. In a similar vein, Nonis and magnitude of one's estimates. Credé et al. (2017) criticized Duckworth
Wright (2003) modeled college GPA as a function of achievement et al. (2007) and Duckworth and Quinn (2009) for overstating the
striving, ability, and their interaction and, alternatively, situational magnitude of their findings by presenting (large) odds-ratios in a
optimism, ability, and their interaction; they identified positive inter- manner that did not identify the (small) increase in the predicted
action terms using both motivation and optimism.1 probability of success that can be attributed to grit. At the same time,
There is an important difference between these studies and our Credé et al. (2017) noted that “variables that exhibit small to moderate
approach. By including in their regression models an interaction term effect sizes can still be very useful in high-stakes settings because even
between a personality trait and cognitive ability, existing studies con- marginal improvements in individuals' performance can have very
strained the estimated effect of the personality trait on the outcome to meaningful positive effects” (page 11). In light of these arguments, we
increase or decrease linearly with ability.2 That is, they allowed the presented “average marginal effects” of grit computed at various levels
data to identify either a positive (synergistic) effect, as hypothesized, a of cognitive ability, and also expressed those estimated effects relative to
negative (offsetting) effect, or no effect. In contrast, we hypothesized the unconditional probability of the outcome at the given ability level.
that synergistic and offsetting (or, in our terminology, synergistic and Because unconditional probabilities vary dramatically with ability, this
compensatory) effects can exist side by side. As described in Section innovation in the reporting of findings proved to be a critical aspect of
2.3, we specified models in which grit is interacted with both ability our analysis.
and its square, thus allowing its estimated marginal effect to fall and Another point of contention in the grit literature concerns the extent
then rise as ability increases or, alternatively, rise and then fall, rise to which grit and its subscales are distinct from conscientiousness and
monotonically, fall monotonically, or remain unchanged. Existing stu- other Big 5 personality traits. While Credé et al. (2017) concluded from
dies that included interactions served as important precursors to our their meta-analysis that grit is so highly correlated with con-
analysis, but differed from our approach by predicting a positive in- scientiousness as to have virtually no independent, predictive power,
teraction effect only and, in most cases, constraining the specification to several studies demonstrated that grit's relationship with the outcome
disallow nonmonotonic effects. of interest was maintained even after measures of Big 5 personality
traits were included among the controls (Duckworth & Quinn, 2009;
Eskreis-Winkler et al., 2014; West et al., 2016). The extent to which grit
1.3. The current study and Big 5 traits have independent predictive power was also explored
for noneducational outcomes (Eskreis-Winkler et al., 2014; Reed,
The primary goal of this study was to identify the moderating effect Pritschet, & Cutton, 2012). We pursued a similar analysis, although
of cognitive ability on the grit-education relationship for a range of measurement issues (discussed in Section 2.2) limited our contribution
educational outcomes. Specifically, we sought to determine whether to suggestive evidence that correlations between grit and Big 5 mea-
two hypotheses received empirical support: (1) grit's positive sures should be explored more fully across the cognitive ability dis-
tribution.
1
The argument that personal traits might affect outcomes differently at the
low or high ends of the ability distribution was also made in the context of
3
nonlinear, nonconjunctive judgment strategies (Einhorn, 1970; Ganzach & Mendolia and Walker (2015) is a rare example of a prior study that ex-
Czaczkes, 1995). In these studies, a negative (positive) interaction term is amined the predictive power of grit with large-scale survey data. They used
evidence of a conjunctive (disjunctive) strategy. data from the Longitudinal Study of Young People in England to analyze the
2
The exception is Ganzach et al. (2000) who, as part of a robustness check, relationship between grit and being “NEET” (neither in education, employment
added interactions between motivation and ability squared. nor training).
144
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
A final issue that we incorporated is whether overall grit scores (or GED) are typically ineligible to pursue post-secondary education, we
should be replaced with subscale scores for consistency of interest and also examined college-related outcomes for a subsample of 3632 high
perseverance of effort (hereafter referred to as consistency and perse- school graduates.
verance). Duckworth et al. (2007) and Duckworth and Quinn (2009)
provided theoretical and empirical support for the notion that grit is a
2.2. Measures
high-order construct and, as noted in Credé et al. (2017), the majority
of grit-related studies used overall grit scores as predictors. This prac-
2.2.1. Grit
tice was criticized by Credé et al. (2017), whose meta-analysis de-
NLSY97 respondents were administered the Grit-S scale in 2013, at
monstrated that perseverance is more predictive than grit for a range of
ages 28–33, as part of the self-administered portion of the regular in-
outcomes. Studies showing perseverance to be a stronger predictor than
terview. The timing of this measure is potentially problematic because,
consistency for educational outcomes include Abuhassàn and Bates
as discussed below, our outcome measures are based on educational
(2015), Bowman et al. (2015), Datu et al. (2016), West et al. (2016) and
attainment at age 26; therefore, grit was measured 2–7 years after
Wolters and Hussain (2015); see also Sturman and Zappala-Piemme
educational outcomes were observed. Under the assumption that grit
(2017). We conducted a similar comparison and found that estimated
remains stable over this portion of the lifecycle or, at minimum, is
effects of perseverance are often (but not always) substantially larger
unaffected by educational attainment, the “late” measure of grit pro-
than those of grit.
vided by the NLSY97 is appropriate for our analysis. Unfortunately,
these assumptions are unlikely to be valid, so additional research will
2. Methods be necessary to determine whether our findings are robust to changes in
the measurement of grit.5
2.1. Participants We used all four items for the consistency subscale (4 items;
α = 0.74) but only three items from the perseverance scale (3 items;
Our data source was the 1997 National Longitudinal Survey of α = 0.61) to obtain a seven-item overall grit score (7 items; α = 0.75).
Youth (NLSY97), which is a large-scale, longitudinal survey sponsored Following Bowman et al. (2015), we dropped “setbacks don't dis-
by the Bureau of Labor Statistics of the U.S. Department of Labor. The courage me” from the perseverance scale because its inclusion caused a
NLSY97 began in 1997 with a sample of 8984 respondents born in substantial reduction in reliability. All three alphas are virtually iden-
1980–84. This sample combined two independent probability samples: tical to comparable statistics reported for samples of West Point cadets
one designed to be representative of the 1980–84 birth cohort residing in Duckworth and Quinn (2009), and within the ranges reported in the
in the U.S. in 1997 (n = 6748) and an over-sample of black and meta-analysis of Credé et al. (2017). In our sample of 4448 individuals,
Hispanic individuals in this birth cohort (n = 2236).4 Respondents were the two subscales were moderately correlated with each other
interviewed annually from 1997 through 2011 (rounds 1–15) and bi- (r = 0.41, p < .001) and highly correlated with the overall grit score
ennially from 2013 to the present. Interviews were completed for 7141 (r = 0.92, p < .001 for consistency; r = 0.74, p < .001 for persever-
respondents in 2013, yielding a 79.5% retention rate over the first 16 ance).
interview rounds. Each data collection effort received approval from Given that NLSY97 grit data appeared not to have been used prior to
Institutional Review Boards and complied with federal laws governing this study, we conducted confirmatory factor analysis to establish the
human subjects research. scale's psychometric validity. Perseverance and consistency were esti-
The NLSY97 provides detailed, longitudinal records of individuals' mated as latent, potentially correlated constructs measured by three
educational activities, employment experiences, family formation and and four questions, respectively. Our estimated, standardized factor
fertility, household income and assets, geographic mobility, and much loadings were 0.46, 0.79, and 0.50 for the three perseverance questions
more. It has been widely used in the social sciences to study a myriad of and 0.56, 0.63, 0.64, and 0.75 for the four consistency questions; si-
topics, including educational outcomes. However, we are unaware of milar factor loadings were found by Duckworth and Quinn (2009). The
prior studies that used the self-reported grit measures for any purpose, two-factor model fit the data significantly better than the one-factor
let alone as predictors of educational outcomes. model: χ2(13, N = 4448) = 306.34, p < .001 versus χ2(14,
Most of our analysis was based on a sample of 4448 respondents N = 4448) = 940.40, p < .001; Δχ2(1) = 634.06, p < .001. Addi-
who satisfied four selection criteria. First, they must have been ad- tional fit indices revealed a satisfactory fit for the two factor model:
ministered the Armed Services Vocational Aptitude Battery (ASVAB), RMSEA = 0.071 (90% CI = 0.064–0.078), CFI = 0.96. These statistics
which provided our cognitive test scores; 7093 of the original sample of suggest a fit that is comparable to or better than those found by
8984 (79%) met this criterion. In addition, they must have been ad- Duckworth and Quinn (2009) for five alternative samples.
ministered the Grit-S (n = 5323 of 7093 respondents) and the Ten-Item
Personality Inventory (TIPI) of Big 5 personality traits (n = 4828 of 2.2.2. Cognitive ability
5323 respondents). Finally, we required that respondents reported their We used an age-adjusted, percentile achievement test score com-
enrollment and degree attainment in sufficient detail that their puted by NLSY97 staff as our measure of cognitive ability. This test
achievements could be identified at ages 20 and 26 (n = 4448 of 4828 score was based on the arithmetic reasoning, mathematical knowledge,
respondents). Although respondents were as old as 36 when inter- paragraph comprehension, and word knowledge components of the
viewed in 2015–16 (the last interview round for which we had data), an Armed Services Vocational Aptitude Battery (ASVAB), which was ad-
age 26 cutoff gave us the largest possible sample of individuals who had ministered to NLSY97 respondents in 1997, when respondents were
largely (albeit not entirely) completed their education. ages 12–17. The ASVAB was not a high-stakes test for NLSY97 re-
We used this uniform sample of 4448 individuals to identify the spondents, but each individual received a $75 participation fee for
effect of grit on the probability of completing several educational out-
comes. Because individuals who did not receive a high school diploma
5
Although the potential limitation of our grit measure is undeniable, one
advantage of the survey design is that respondents were unlikely to be thinking
4
The NLSY97 is one of several cohorts collectively referred to as the National about their educational experiences when self-reporting their grit. Most re-
Longitudinal Surveys. See https://www.nlsinfo.org for information about the spondents were far removed from school by 2013, and those with schooling
NLS and the NLSY97. The NLSY97 Technical Sampling Report can be viewed at activities to report in 2013 subsequently discussed many other topics (in a
https://www.nlsinfo.org/sites/nlsinfo.org/files/attachments/121221/ survey often lasting well over an hour) before being administered the Grit-S
TechnicalSamplingReport.pdf scale.
145
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
completing the test. The four-item score that we used was designed to 2.2.4. Additional controls
correspond to an Armed Forces Qualification Test (AFQT) score, and is The remaining controls in our regression models included binary
hereafter referred to as an AFQT score. indicators of whether the individual was male (mean = 0.48), black
AFQT scores reflect achievement as much as innate ability (mean = 0.26), and Hispanic (mean = 0.19), plus measures of both the
(Borghans, Golsteyn, Heckman, & Humphries, 2016; Cascio & Lewis, father's and mother's highest grade completed. We included these
2006), but have been found to be highly predictive of numerous aca- measures in light of extensive evidence that educational outcomes
demic and labor market outcomes (Almlund et al., 2011; Altonji, differ by sex, race, ethnicity, and socioeconomic status, the latter of
Bharadwaj, & Lange, 2012; Ganzach, 2014; Kearney & Levine, 2016). which represents opportunity to perform (Musu-Gillette et al., 2016;
We further validated this measure's predictive capability by computing Shanahan et al., 2014; Sirin, 2005). Moreover, we would not want es-
its correlation with a range of achievement test scores drawn from timates of the moderating effects of ability to reflect effects of omitted
NLSY97 respondents' college transcripts. For each correlation, we used demographic factors that are correlated with ability. It was unnecessary
a sample consisting of all respondents for whom an AFQT score and the to control for age because all outcomes were defined with respect to a
given achievement test score were available. We found correlations of fixed point in the lifecycle (age 20 for high school completion, and age
r = 0.74 (p < .001, n = 1193) for AFQT and SAT math scores, 26 for the college-related outcomes).
r = 0.76 (p < .001, n = 1192) for AFQT and SAT verbal scores,
r = 0.82 (p < .001, n = 1132) for AFQT and ACT scores, r = 0.75
(p < .001, n = 869) for AFQT and PSAT verbal scores, and r = 0.74 2.2.5. Educational outcomes
(p < .001, n = 868) for AFQT and PSAT math scores. We considered four educational outcomes representing increasingly
A small, negative correlation was found between grit and AFQT challenging achievements: earn a high school diploma by age 20, enroll
scores (r = −0.06, p < .01, n = 4448). A similar negative correlation in a two-year or four-year college by age 26, earn any undergraduate
was found for consistency and AFQT scores (r = −0.08, p < .001, college degree (associate's or bachelor's) by age 26, and earn a bache-
n = 4448) while no significant correlation was found for perseverance lor's degree by age 26. We required that enrollment and educational
and AFQT scores (r = 0.01, p ≥ .05, n = 4448). These findings indicate attainment be reported through age 26 for inclusion in our sample, so
that AFQT scores are largely unrelated to grit and its lower-order facets all four outcomes were observed for all 4448 sample members. Sample
and can be viewed as independent predictors of educational outcomes. members who did not receive a high school diploma would typically be
ineligible to enter college and complete a college degree, so we per-
formed regression analysis for the three college-related outcomes for a
2.2.3. Big 5 personality traits sub-sample of 3632 high school graduates as well as for the full sample.
For select specifications, we added measures of the Big 5 personality For each outcome, we created a binary indicator equal to one if the
traits (conscientiousness, emotional stability, agreeableness, extraver- educational milestone was attained and zero otherwise. Summary sta-
sion, and openness). Scores were based on the ten-item personality tistics for all outcomes and select regressors, for both samples, were
inventory (TIPI) developed by Gosling, Rentfrow, and Swann Jr. summarized in Table 1. Although this table shows means and standard
(2003), which was administered to NLSY97 respondents in 2008, when deviations for “raw” grit, AFQT, and TIPI scores, we used standardized
they were ages 23–28. scores (mean = 0, SD = 1) for the regression analysis.
Unfortunately, these measures of Big 5 traits suffer from two po-
tential problems. First, they were based on a ten-item scale (two items
per trait), which yields less reliable measures of personality traits than a Table 1
richer scale (Gosling et al., 2003). Second, they were measured five Sample means and standard deviations for select variables.
years earlier than grit and, therefore, do not necessarily reflect per-
Full Sample HS Graduates
sonality traits that prevailed when educational outcomes were de-
termined or when grit was measured.
Both shortcomings are expected to reduce the correlations between
Mean SD Mean SD
grit and each Big 5 measure. We found correlations of r = 0.37,
r = 0.22, r = 0.09, r = 0.12 and r = 0.09 (p < .001, n = 4448 in all Outcomes
cases) between grit and conscientiousness, emotional stability, agree- Earn high school diploma 0.82 – 1.00 –
ableness, extraversion, and openness, respectively. These can be com- Enroll in college 0.65 – 0.76 –
pared to correlations reported by Duckworth and Quinn (2009, p. 169) Earn any college degree 0.36 – 0.43 –
Earn Bachelor's degree 0.28 – 0.35 –
for a sample of 1554 adults who were contemporaneously administered Covariates
the Grit-S scale and the 44-item Big Five Inventory (John & Srivastava, 1 if male 0.48 – 0.47 –
1999) as part of an online survey: r = 0.77 (p < .001), r = −0.40 1 if black 0.26 – 0.24 –
(p < .001), r = 0.24 (p < .001), r = 0.20 (p < .001), and r = 0.06 1 if Hispanic 0.19 – 0.17 –
Father's highest grade completed 12.72 2.84 12.98 2.89
(p ≥ .001) (n = 1554 in all cases). Clearly, the NLSY97-based correla-
Mother's highest grade completed 12.76 2.75 13.04 2.76
tions are smaller than previously-reported correlations based on a AFQT score (A)a 49.07 29.06 54.10 28.00
preferred Big 5 measure for conscientiousness, agreeableness, and ex- Grit score (G)a 26.97 4.28 27.03 4.21
traversion. Consistency scorea 14.35 3.16 14.35 3.10
The question, of course, is how the inclusion of nonoptimally- Perseverance scorea 12.62 1.87 12.68 1.84
Conscientiousnessa 5.68 1.11 5.70 1.08
measured Big 5 traits is expected to affect regression estimates. When Emotional stabilitya 5.01 1.31 5.10 1.26
Big 5 controls are omitted, the estimated relationship between grit and Agreeablenessa 4.99 1.11 5.04 1.10
the educational outcome represents both the “true” effect of grit and the Extraversiona 4.70 1.36 4.76 1.35
effect of omitted Big 5 traits with which grit is correlated. When Big 5 Opennessa 5.50 1.07 5.51 1.04
Sample size 4448 3632
controls are included, any downward bias in their estimated effects due
to mismeasurement is likely to produce upward bias in the estimated Note. Additional regressors include A2, G∙A and G∙A2. Predictions in Table 2
parameter for grit (or for any included variable that is correlated with used estimates from regressions based on the full sample; predictions in Table 3
the Big 5 traits). In short, scenarios exist in which the estimated grit were based on the high school graduation sample.
effect will be overstated with or without Big 5 controls. a
Means and standard deviations for raw scores were reported in this table;
standard scores (mean = 0, SD = 1) were used in all regressions.
146
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
2.3. Analytic procedure difference between these values is the estimated marginal effect of grit
on the individual's probability of college graduation at the median ability
To model the probability of each (binary) educational outcome, we level. The mean of these individual estimates is the estimated average
used probit models, which are identical to logistic regression models marginal effect.
except that the distribution of errors is assumed to be normal rather In Section 3, we tabulated estimated, average marginal effects of
than logistic. Probit and logit models are known to deliver virtually grit not only at the median (50th percentile) level of ability, but also at
identical estimates except in extreme tails of the distributions the 10th and 90th percentiles; estimates at other ability levels were
(Maddala, 1986), which we confirmed for our findings. discussed but not tabulated. For estimates based on specification (2),
Our primary probit model was specified as follows: which does not allow the marginal effect of grit to differ with ability,
we computed only one average marginal effect, using each sample
Yi ∗ = β0 + β1Gi + β2 Ai + β3 Ai 2 + β4 Gi ∙Ai + β5 Gi ∙Ai 2 + β6 Zi + εi (1) member's actual ability level.
where Yi ∗
is the latent dependent variable for individual i, Gi and Ai are To interpret the magnitude of each average marginal effect, we also
standardized scores for grit and cognitive ability, respectively, and Zi is reported them relative to a baseline probability. If we were to find, for
a vector of additional controls. Our observed outcomes, Yi, take the example, that a one-unit (1 SD) increase in grit is predicted to raise the
value one if Yi ∗ > 0 and zero otherwise; we estimated the binomial chance of earning a college degree by 3.5 percentage points for students
probability Pr(Yi ∗ > 0) using maximum likelihood estimation. in the 10th percentile of the ability distribution and by 6.4 percentage
After experimenting with several specifications, we settled on this points for students in the 90th percentile, a naïve interpretation would
“quadratic interaction” model because the estimated marginal effect of suggest that grit is more beneficial among high-ability students than
Gi, ∂Yi ∗/∂Gi = β1 + β4Ai + β5Ai2, is a quadratic function of Ai.6 This among low-ability students. However, high-ability students are 10
functional form allows the estimated marginal effect of grit to be higher times more likely than their low-ability counterparts to graduate from
for low and high levels of ability than for intermediate levels, but does college. We accounted for such differences by dividing each average
not constrain such a u-shaped pattern to emerge. If the relationship marginal effect by the unconditional probability of the given educa-
between grit and each outcome proved to be monotonically increasing tional outcome at the given point in the ability distribution. To compute
or decreasing in ability or even independent of ability, our model would the denominators (baseline probabilities), we used subsamples of in-
identify those patterns. dividuals whose actually ability levels were “close to” the target levels.7
In addition to (1), we also used a less flexible specification: In our example, this adjustment revealed that a one-unit (1 SD) increase
in grit is predicted to raise the graduation rate by 50% relative to the
Yi ∗ = γ0 + γ1Gi + γ2 Ai + γ3 Ai 2 + γ6 Zi + ui . (2) baseline (0.035/0.07) at the 10th percentile, but by only 9% relative to
the baseline (0.064/0.70) at the 90th percentile.
Specification (2) constrains the marginal effect of Gi, ∂Yi ∗/∂Gi = γ1,
to be invariant to ability, but is identical to (1) in all other respects. This
constrained model is similar to what was used in prior studies of the 3. Results
grit-education relationship.
We estimated alternative versions of (1) and (2). First, we included Table 2 contains estimated “average marginal effects” of grit and
the short set of “additional controls” described in Section 2.2 in the perseverance at three ability levels for our sample of 4448 students.
vector Z. Second, we added the Big 5 traits to Z. Third, we reestimated (See Appendix Table A1 for estimated effects based on consistency, and
each specification (with and without controls for Big 5 traits) after re- Tables A2–A5 for probit model estimates used to compute all marginal
placing grit scores with scores for perseverance and, in turn, con- effects.)
sistency. All specifications were estimated for four alternative out- To clarify what the numbers in Table 2 represent, consider the es-
comes: high school graduation, college enrollment, any college degree timate of 0.013 in the “all” column of the first row. This estimate was
completion, and bachelor's degree completion. We also estimated spe- based on specification 2 described in Section 2.3, with high school
cifications for the three college-related outcomes with a subsample of graduation as the outcome and grit as the key regressor; it indicates that
individuals who completed high school. After evaluating a wide array a one-SD increase in grit raises the predicted graduation probability by
of estimates, in Section 3 we reported a subset that best illustrated the a statistically significant 1.3 percentage points. The corresponding
key findings. “relative effect” indicates that the 1.3 percentage point marginal effect
Because estimated coefficients are difficult to interpret directly in represents a 1.5% increase relative to the baseline graduation prob-
probit models, we computed “average marginal effects” of grit (and ability (across all ability levels) of 0.82. The remaining three columns in
perseverance and consistency) at select levels of ability. For example, the first row were based on specification 1, which augments specifica-
we assigned each sample member an AFQT score corresponding to the tion 2 by adding interactions between grit and cognitive ability and its
sample median and actual values for all other regressors, and computed square. Identical probit estimates were used to compute the marginal
his/her predicted probability of earning a college degree. We then re- effects in the 10th, 50th, and 90th percentile columns; only the AFQT
computed each individual's predicted probability after increasing his/ score used for the computations changed across columns. Estimates in
her grit level by one unit (1 SD, or 4.28 points on a 35-point scale). The the next two pairs of rows in Table 2 were obtained in identical fashion,
except perseverance scores replaced overall grit score in each regres-
6 sion. Big 5 personality traits were added to the regression controls to
Alternative specifications included (a) adding G2, with and without inter-
obtain the second pair of perseverance-based estimates.
actions between G2 and each ability term; (b) adding A3 and G∙A3; (c) extending
to cubics in both A and G; and (d) replacing A, A2, G∙A and G∙A2 in (1) with Focusing first on the “earn high school diploma” outcome in
binary indicators of the ability quartile and interactions between quartile in- Table 2, the “all” column indicates that the estimated marginal effect of
dicators and G. Each alternative specification yielded similar patterns to what grit is small (1.3 percentage points) when it is not allowed to differ with
we reported in section 3. However, because the correlation between A and G cognitive ability. However, the next three columns show a much higher
moves from strongly positive at low ability to weakly negative at high ability estimated effect of 2.9 percentage points (5.3% relative to the baseline)
(e.g., r = 0.17, p < .001, n = 1211 for ability quartile 1; r = −0.07, p < .05, among students with AFQT scores at the 10th percentile, and smaller
n = 1211 for quartile 4), the estimated marginal effect of grit on each outcome
was slightly smaller (larger) at low (high) levels of ability the more flexibly we
controlled for A. Note that (1) includes “main effects” of ability (Ai and Ai2) in 7
We used subsamples of individuals with AFQT scores in percentile intervals
addition to their interactions with Gi to avoid attributing the grit effect to a [0−20], [40–60], and [80–100] to compute baseline probabilities for the 10th,
direct, uncontrolled ability effect. 50th, and 90th percentiles, respectively.
147
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
148
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
149
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
appear to have been mismeasured (as discussed further in Section 4.1), as definitive, and in this case it is particularly important that additional
we provided compelling evidence that the “grit versus conscientious- research be conducted to confirm or refute our findings.
ness” debate would benefit from considering relationships between Additional limitations of our analysis could be addressed in future
these (and other) measures at different points in the ability distribution. NLSY97-based research. The NLSY97 collected high school and college
transcripts for subsets of respondents, so future studies could explore
4.1. Limitations of our analysis grit's relationship to grades and test scores, albeit for smaller samples
than the ones used in the current study. The current study did not ex-
Large-scale surveys offer numerous advantages for the analysis of plore the extent to which grit-educational relationships depend on in-
grit's relationship to educational outcomes, but they also impose lim- dividuals' demographic characteristics, nor did it consider the many
itations. Key limitations of the NLSY97 were three-fold. First, the Grit-S noneducational outcomes reported in the NLSY97 related to family
scale was administered when respondents were ages 28–33. While formation, labor market experiences, asset accumulation, and much
temporal separation between administration of the grit scale and edu- more.
cational activities might be advantageous in some respects, the concern
is that self-reported grit at this stage in the lifecycle might be condi-
tioned on earlier educational experiences. Additional evidence on the 5. Conclusions
extent to which self-reported grit changes over time is needed to de-
termine whether the survey's relatively “late” collection of grit scores In this study, we used data from the 1997 National Longitudinal
posed problems for our analysis. Survey of Youth to determine whether grit's marginal effect on educa-
Second, measures of the Big 5 personality traits were obtained via tional attainment varies with cognitive ability. Our analysis had the
the administration of the Ten Item Personality Inventory when re- following key features: (a) we analyzed a sequence of increasingly
spondents were ages 23–28. These measures are problematic because challenging educational outcomes for a large, uniform, representative
(a) the 10-item scale lacks reliability; (b) responses can be conditioned sample of individuals; (b) grit's relationship with each outcome was
on completed educational experiences (as with the grit measure); and allowed to differ flexibly across the cognitive ability distribution; and
(c) the Big 5 and grit scales were not administered contemporaneously. (c) estimated, marginal effects of grit (as well as perseverance and
Both timing issues would affect our analysis if core personality traits consistency) at each ability level were expressed relative to baseline
evolve throughout early adulthood. While evidence exists that changes probabilities of success at those same ability levels.
in personality traits over adulthood are not important enough to affect The finding that grit is positively associated with college-related
research findings (Cobb-Clark & Schurer, 2012) the preponderance of outcomes among high-ability students is consistent with the notion that
evidence, as summarized by Roberts, Walton, and Viechtbauer (2006), high-ability students adopt self-regulated learning processes that
suggested that conscientiousness and other traits can change by sub- leverage their grit. For these students, grit and cognitive ability com-
stantial amounts. Overall, the concern is that our analysis misstated plement each other in the learning process, especially when tasks are
grit's (perseverance's) relationship with each outcome due to an in- relatively challenging. The finding that grit is also positively associated
ability to correctly “net out” the independent effects of Big 5 traits. with all educational outcomes at the opposite end of the ability dis-
Third, although NLSY97 data enabled us to define a sequence of tribution is perhaps more surprising, and is consistent with the hy-
educational outcomes for a large, representative sample, the survey did pothesis that low-ability students substitute grit for ability in the
not provide details on college engagement and satisfaction, time allo- learning process. This does not imply that grit is as valuable as ability:
cation, institutional support, and other aspects of educational experi- relatively few low-ability students succeed in attending college and
ences that previous studies (e.g., Bowman et al., 2015; Datu et al., 2016; earning college degrees, but among those who do attain success, grit
Eskreis-Winkler et al., 2014) brought to bear. might an important factor. These findings suggest that interventions
We conducted our analysis despite these limitations to provide a designed to augment grit among low-ability students could have sub-
“first look” at whether grit's relationship to educational outcomes varies stantial benefits—and that additional evidence on the moderating ef-
across the cognitive ability distribution. No single study can be viewed fects of cognitive ability is well worth pursuing.
Appendix A
Appendix Table A1
Predicted change in the probability of alternative educational outcomes by cognitive ability percentile (Supplement to Tables 2–3).
150
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
Note: N = 4448 in the full sample; N = 3632 in the sample of individuals who earn a high school diploma. Estimates in bold are statistically significant at
p < .05. Specifications reported here exclude measures of Big 5 traits. See Tables 2–3 for additional notes and baseline probabilities.
†
p < .20 for two-tailed test of the null hypotheses that the difference between this estimate and the corresponding 50th percentile estimate (in the same row)
is zero.
‡
p < .05 for two-tailed test of the null hypotheses that the difference between this estimate and the corresponding 50th percentile estimate (in the same row)
is zero.
Appendix Table A2
Probit estimates for the “earn high school diploma” outcome used to compute predicted marginal effects in Tables 2 and A1.
Regressor 1 2 3 4 5 6 7 8
AFQT score (A) 0.582 0.590 0.579 0.583 0.574 0.576 0.581 0.588
(0.031) (0.032) (0.031) (0.032) (0.032) (0.032) (0.031) (0.032)
A 2
−0.051 −0.041 −0.051 −0.041 −0.045 −0.034 −0.056 −0.051
(0.029) (0.030) (0.029) (0.029) (0.030) (0.030) (0.029) (0.030)
Key predictor (P)a 0.058 0.022 0.082 0.023 0.044 −0.028 0.028 0.023
(0.025) (0.038) (0.024) (0.036) (0.025) (0.037) (0.025) (0.039)
P∙A 0.019 0.008 −0.004 0.029
(0.030) (0.027) (0.026) (0.031)
P∙A2 0.046 0.062 0.067 0.018
(0.031) (0.028) (0.027) (0.032)
1 if Male −0.158 −0.156 −0.157 −0.154 −0.141 −0.137 −0.157 −0.155
(0.049) (0.049) (0.049) (0.049) (0.051) (0.051) (0.049) (0.049)
1 if Black 0.239 0.242 0.238 0.239 0.233 0.233 0.249 0.251
(0.062) (0.062) (0.062) (0.062) (0.063) (0.063) (0.062) (0.062)
1 if Hispanic 0.189 0.189 0.190 0.187 0.180 0.175 0.194 0.196
(0.071) (0.071) (0.071) (0.071) (0.071) (0.071) (0.071) (0.071)
Father's highest grade completed 0.035 0.035 0.035 0.035 0.033 0.033 0.035 0.035
(0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011)
Mother's highest grade completed 0.039 0.039 0.039 0.039 0.040 0.040 0.040 0.040
(0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011)
Conscientiousness 0.069 0.073
(0.027) (0.027)
Emotional stability 0.102 0.101
(0.027) (0.027)
Agreeableness 0.044 0.046
(0.027) (0.027)
Extraversion 0.064 0.066
(0.026) (0.026)
Openness −0.067 −0.068
(0.027) (0.027)
Constant 0.275 0.272 0.284 0.279 0.285 0.282 0.271 0.267
(0.159) (0.159) (0.159) (0.159) (0.160) (0.160) (0.159) (0.159)
Note: N = 4448. AFQT, grit (including perseverance and consistency), and Big 5 scores are standardized to have mean = 0 and SD = 1. Standard errors are in
parentheses. Columns 1, 3, 5 and 7 correspond to the “all” estimates and columns 2, 4, 6, and 8 correspond to the remaining (ability-specific) estimates in Tables 2
and A1; columns 5–6, which include Big 5 measures, correspond to the perseverance rows in Table 2 denoted by ‘a.’
a
The key predictor (P) is either grit, perseverance, or consistency as indicated in the column heading.
Appendix Table A3
Probit estimates for the “enroll in college” outcome used to compute predicted marginal effects in Tables 2 and A1.
Regressor 1 2 3 4 5 6 7 8
AFQT score (A) 0.716 0.723 0.714 0.714 0.704 0.704 0.713 0.722
(0.028) (0.028) (0.028) (0.028) (0.028) (0.028) (0.028) (0.028)
(continued on next page)
151
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
Regressor 1 2 3 4 5 6 7 8
A 2
−0.065 −0.056 −0.065 −0.063 −0.057 −0.055 −0.071 −0.062
(0.026) (0.027) (0.026) (0.026) (0.026) (0.027) (0.026) (0.026)
a
Key predictor (P) 0.090 0.049 0.139 0.117 0.103 0.079 0.039 0.005
(0.022) (0.032) (0.021) (0.031) (0.023) (0.033) (0.022) (0.032)
P∙A 0.014 −0.008 −0.012 0.031
(0.023) (0.022) (0.022) (0.024)
P∙A2 0.051 0.023 0.025 0.047
(0.026) (0.024) (0.025) (0.026)
1 if Male −0.348 −0.346 −0.348 −0.348 −0.366 −0.365 −0.346 −0.344
(0.044) (0.044) (0.044) (0.044) (0.047) (0.047) (0.044) (0.044)
1 if Black 0.366 0.370 0.366 0.366 0.349 0.348 0.380 0.386
(0.057) (0.057) (0.057) (0.057) (0.058) (0.058) (0.057) (0.057)
1 if Hispanic 0.371 0.371 0.374 0.372 0.359 0.357 0.377 0.379
(0.066) (0.067) (0.066) (0.066) (0.067) (0.067) (0.066) (0.066)
Father's highest grade completed 0.067 0.067 0.067 0.067 0.067 0.067 0.067 0.067
(0.010) (0.010) (0.010) (0.010) (0.010) (0.010) (0.010) (0.010)
Mother's highest grade completed 0.043 0.042 0.042 0.042 0.041 0.041 0.043 0.043
(0.090) (0.090) (0.090) (0.090) (0.090) (0.090) (0.090) (0.089)
Conscientiousness 0.038 0.040
(0.024) (0.025)
Emotional stability 0.096 0.096
(0.025) (0.025)
Agreeableness −0.009 −0.008
(0.024) (0.024)
Extraversion 0.046 0.047
(0.023) (0.023)
Openness 0.018 0.018
(0.023) (0.023)
Constant −0.771 −0.776 −0.761 −0.761 −0.752 −0.752 −0.772 −0.778
(0.152) (0.152) (0.153) (0.153) (0.153) (0.154) (0.152) (0.152)
Note: N = 4448. AFQT, grit (including perseverance and consistency), and Big 5 scores are standardized to have mean = 0 and SD = 1. Standard errors are in
parentheses. Columns 1, 3, 5 and 7 correspond to the “all” estimates and columns 2, 4, 6, and 8 correspond to the remaining (ability-specific) estimates in Tables 2
and A1; columns 5–6, which include Big 5 measures, correspond to the perseverance rows in Table 2 denoted by ‘a.’
a
The key predictor (P) is either grit, perseverance, or consistency as indicated in the column heading.
Appendix Table A4
Probit estimates for the “earn any college degree” outcome used to compute predicted marginal effects in Tables 2 and A1.
Regressor 1 2 3 4 5 6 7 8
AFQT score (A) 0.610 0.617 0.606 0.608 0.607 0.609 0.610 0.616
(0.027) (0.027) (0.027) (0.027) (0.028) (0.028) (0.027) (0.027)
A2 −0.005 −0.001 −0.007 −0.006 −0.006 −0.005 −0.008 −0.004
(0.025) (0.026) (0.025) (0.025) (0.026) (0.026) (0.025) (0.026)
Key predictor (P)a 0.120 0.065 0.133 0.097 0.082 0.045 0.086 0.035
(0.021) (0.030) (0.022) (0.032) (0.023) (0.033) (0.022) (0.031)
P∙A −0.010 −0.001 −0.010 −0.001
(0.024) (0.023) (0.023) (0.024)
P∙A2 0.067 0.040 0.044 0.059
(0.025) (0.024) (0.024) (0.025)
1 if Male −0.315 −0.313 −0.313 −0.312 −0.333 −0.332 −0.316 −0.314
(0.043) (0.043) (0.043) (0.044) (0.047) (0.047) (0.043) (0.043)
1 if Black 0.060 0.065 0.068 0.070 0.050 0.051 0.069 0.075
(0.058) (0.058) (0.058) (0.058) (0.059) (0.059) (0.058) (0.058)
1 if Hispanic 0.046 0.045 0.051 0.051 0.050 0.049 0.050 0.050
(0.064) (0.064) (0.064) (0.064) (0.065) (0.065) (0.064) (0.064)
Father's highest grade completed 0.069 0.069 0.068 0.069 0.070 0.070 0.068 0.069
(0.010) (0.010) (0.010) (0.010) (0.010) (0.010) (0.010) (0.010)
Mother's highest grade completed 0.043 0.043 0.042 0.042 0.043 0.043 0.043 0.043
(0.010) (0.010) (0.010) (0.010) (0.011) (0.011) (0.010) (0.010)
Conscientiousness (0.093) (0.093) (0.093) (0.093) (0.093) (0.093) (0.093) (0.092)
0.122 0.122
Emotional stability (0.025) (0.025)
0.081 0.082
Agreeableness (0.025) (0.025)
−0.026 −0.025
Extraversion (0.023) (0.023)
0.017 0.017
(continued on next page)
152
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
Regressor 1 2 3 4 5 6 7 8
Note: N = 4448. AFQT, grit (including perseverance and consistency), and Big 5 scores are standardized to have mean = 0 and SD = 1. Standard errors are in
parentheses. Columns 1, 3, 5 and 7 correspond to the “all” estimates and columns 2, 4, 6, and 8 correspond to the remaining (ability-specific) estimates in Tables 2
and A1; columns 5–6, which include Big 5 measures, correspond to the perseverance rows in Table 2 denoted by ‘a.’
a
The key predictor (P) is either grit, perseverance, or consistency as indicated in the column heading.
Appendix Table A5
Probit estimates for the “earn bachelor's degree” outcome used to compute predicted marginal effects in Tables 2 and A1.
Regressor 1 2 3 4 5 6 7 8
AFQT score (A) 0.684 0.687 0.680 0.684 0.690 0.695 0.682 0.685
(0.031) (0.030) (0.030) (0.031) (0.032) (0.032) (0.031) (0.030)
A2 −0.004 0.003 −0.006 −0.007 −0.003 −0.005 −0.008 0.002
(0.028) (0.028) (0.028) (0.028) (0.028) (0.029) (0.028) (0.028)
Key predictor (P)a 0.131 0.075 0.147 0.114 0.081 0.048 0.094 0.039
(0.023) (0.032) (0.024) (0.034) (0.025) (0.035) (0.023) (0.032)
P∙A 0.020 −0.010 −0.022 0.041
(0.027) (0.028) (0.029) (0.026)
P∙A2 0.053 0.040 0.045 0.046
(0.027) (0.027) (0.027) (0.027)
1 if Male −0.322 −0.318 −0.320 −0.319 −0.317 −0.316 −0.323 −0.319
(0.046) (0.046) (0.046) (0.046) (0.050) (0.050) (0.046) (0.046)
1 if Black 0.084 0.093 0.094 0.096 0.095 0.096 0.092 0.103
(0.063) (0.063) (0.063) (0.063) (0.064) (0.064) (0.063) (0.063)
1 if Hispanic 0.076 0.078 0.083 0.082 0.095 0.093 0.079 0.081
(0.070) (0.070) (0.070) (0.070) (0.072) (0.072) (0.070) (0.070)
Father's highest grade completed 0.080 0.080 0.080 0.080 0.081 0.082 0.080 0.080
(0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011) (0.011)
Mother's highest grade completed 0.059 0.059 0.057 0.057 0.059 0.059 0.059 0.059
(0.012) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012) (0.012)
Conscientiousness 0.166 0.167
(0.027) (0.027)
Emotional stability 0.067 0.067
(0.026) (0.026)
Agreeableness 0.003 0.003
(0.025) (0.025)
Extraversion 0.035 0.034
(0.024) (0.024)
Openness −0.054 −0.053
(0.025) (0.025)
Constant −2.414 −2.427 −2.398 −2.403 −2.464 −2.468 −2.408 −2.419
(0.177) (0.178) (0.177) (0.177) (0.179) (0.180) (0.177) (0.177)
Note: N = 4448. AFQT, grit (including perseverance and consistency), and Big 5 scores are standardized to have mean = 0 and SD = 1. Standard errors are in
parentheses. Columns 1, 3, 5 and 7 correspond to the “all” estimates and columns 2, 4, 6, and 8 correspond to the remaining (ability-specific) estimates in Tables 2
and A1; columns 5–6, which include Big 5 measures, correspond to the perseverance rows in Table 2 denoted by ‘a.’
a
The key predictor (P) is either grit, perseverance, or consistency as indicated in the column heading.
References Bowman, N. A., Hill, P. N., Denson, N., & Bronkema, R. (2015). Keep on truckin’ or stay
the course? Exploring grit dimensions as differential predictors of educational
achievement, satisfaction, and intentions. Social Psychological and Personality Science,
Abuhassàn, A., & Bates, T. (2015). Distinguishing effortful persistence from con- 6, 639–645.
scientiousness. Journal of Individual Differences, 36, 205–214. Cascio, E., & Lewis, E. (2006). Schooling and the armed forces qualifying test: Evidence
Almlund, M., Duckworth, A. L., Heckman, J., & Kautz, T. (2011). Personality psychology from school-entry laws. Journal of Human Resources, 41, 294–318.
and economics. In E. Hanushek, S. Machin, & L. Woessmann (Vol. Eds.), Handbook of Cobb-Clark, D., & Schurer, S. (2012). The stability of big-five personality traits. Economics
the economics of education. 4. Handbook of the economics of education (pp. 1–181). Letters, 115, 11–15.
Amsterdam, The Netherlands: Elsevier Science. Cohen, M. T. (2012). The importance of self-regulation for college student learning.
Altonji, J., Bharadwaj, P., & Lange, F. (2012). Changes in the characteristics of American College Student Journal, 46, 892–902.
youth: Implications for adult outcomes. Journal of Labor Economics, 30, 783–828. Conard, M. A. (2006). Aptitude is not enough: How personality and behavior predict
Borghans, L., Duckworth, A. L., Heckman, J. J., & ter Weel, B. (2008). The economics and academic performance. Journal of Research in Personality, 40, 339–346.
psychology of personality traits. Journal of Human Resources, 43, 972–1059. Credé, M. (2018). What shall we do about grit? A critical review of what we know and
Borghans, L., Golsteyn, B. H. H., Heckman, J. J., & Humphries, J. J. (2016). What grades what we don't know. Educational Researcher, 20, 1–6.
and achievement tests measure. Proceedings of the National Academy of Sciences of the Credé, M., & Kuncel, N. R. (2008). Study habits, skills, and attitudes: The third pillar
United States of America, 113, 13354–13359. supporting collegiate academic performance. Perspectives on Psychological Science, 3,
153
A. Light, P. Nencka Learning and Individual Differences 70 (2019) 142–154
154