739704

research-article2018
PSSXXX10.1177/0956797617739704Sisk et al.Mind-Set and Academic Achievement

Research Article

Psychological Science

To What Extent and Under Which 2018, Vol. 29(4) 549­–571

© The Author(s) 2018
Reprints and permissions:
Circumstances Are Growth Mind-Sets sagepub.com/journalsPermissions.nav
DOI: 10.1177/0956797617739704
https://doi.org/10.1177/0956797617739704

Important to Academic Achievement? www.psychologicalscience.org/PS

Two Meta-Analyses

Victoria F. Sisk1, Alexander P. Burgoyne2, Jingze Sun1,

Jennifer L. Butler1, and Brooke N. Macnamara1
1
Department of Psychological Sciences, Case Western Reserve University, and
2
Department of Psychology, Michigan State University

Abstract
Mind-sets (aka implicit theories) are beliefs about the nature of human attributes (e.g., intelligence). The theory
holds that individuals with growth mind-sets (beliefs that attributes are malleable with effort) enjoy many positive
outcomes—including higher academic achievement—while their peers who have fixed mind-sets experience negative
outcomes. Given this relationship, interventions designed to increase students’ growth mind-sets—thereby increasing
their academic achievement—have been implemented in schools around the world. In our first meta-analysis (k =
273, N = 365,915), we examined the strength of the relationship between mind-set and academic achievement and
potential moderating factors. In our second meta-analysis (k = 43, N = 57,155), we examined the effectiveness of mind-
set interventions on academic achievement and potential moderating factors. Overall effects were weak for both meta-
analyses. However, some results supported specific tenets of the theory, namely, that students with low socioeconomic
status or who are academically at risk might benefit from mind-set interventions.

Keywords
mind-set, implicit theories, education, academic achievement, open data

Received 2/21/17; Revision accepted 10/9/17

According to mind-set theory (aka implicit theories; with higher growth mind-sets have more adaptive psy-
Dweck, 2006; Dweck, Chiu, & Hong, 1995), individuals chological traits and behaviors (e.g., positive response
vary in their beliefs about whether human attributes to failure), which lead to greater academic achievement
(e.g., intelligence) are stable or malleable. Individuals (e.g., Dweck, 2000). The theory also suggests that inter-
who believe attributes are stable have fixed mind-sets ventions designed to increase students’ growth mind-
(aka entity theories), whereas those who believe attri- sets will lead to greater academic achievement because
butes are malleable have growth mind-sets (aka incre- there is a “powerful impact of growth mindset mes-
mental theories). According to mind-set theory, holding sages upon students’ attainment” (Boaler, 2013, p. 143).
a fixed mind-set is detrimental for a variety of real-world These ideas have led to the establishment of nonprofit
outcomes, whereas holding a growth mind-set leads to organizations (e.g., Project for Education Research that
a variety of positive outcomes, including weight loss Scales [PERTS]), for-profit entities (e.g., Mindset Works,
(Burnette & Finkel, 2012), reaching international acclaim Inc.), schools purchasing mind-set intervention programs
(Dweck, 2006), and achieving peace in the Middle East
(Dweck, 2012, 2016).
Corresponding Author:
Most frequently, mind-sets are researched in educa- Brooke N. Macnamara, Department of Psychological Sciences, Case
tional contexts (Burnette, O’Boyle, VanEpps, Pollack, Western Reserve University, 11220 Bellflower Rd., Cleveland, OH 44106
& Finkel, 2013). Mind-set theory suggests that students E-mail: brooke.macnamara@case.edu
550 Sisk et al.

(e.g., Brainology), and millions of dollars in funding to Claro, Paunesku, & Dweck, 2016) has suggested that
individual researchers, nonprofit organizations, and for- holding a growth mind-set is especially beneficial for
profit companies (e.g., Bill and Melinda Gates Founda- low-SES students’ academic success.
tion,1 Department of Education,2 Institute of Educational We examined the type of academic achievement mea-
Sciences3). sure because the effect might differ, for example,
Given mind-set theory’s impact on education, we between course grades and standardized tests. Addi-
sought to ask the following questions: tionally, we investigated the possibility that if students
with growth mind-sets are taking more challenging
1. What is the magnitude of the relationship courses (see Romero, Master, Paunesku, Dweck, &
between mind-sets and academic achievement, Gross, 2014), then the relationship could be suppressed
and under which circumstances does the rela- when the measure of achievement also reflects students’
tionship strengthen or weaken? course selection.
2. Do mind-set interventions positively impact aca- Finally, we tested whether publication bias is prob-
demic achievement, and under which circum- lematic within the mind-set-in-education literature. Pub-
stances does the impact increase or decrease? lication bias occurs when some results are systematically
less likely to be published than others (e.g., studies that
To answer these questions, we conducted two meta- find small or null effects; Rosenthal, 1979).
analyses to (a) estimate the sizes of these effects and
whether they are consistent across studies, (b) examine
potential moderating factors, and (c) empirically evalu-
Method
ate the theory. We designed the meta-analysis and report the results
in accordance with the Preferred Reporting Items for
Meta-Analysis 1: The Relationship Systematic Reviews and Meta-Analyses (PRISMA) state-
Between Mind-Sets and Academic ment (Moher, Liberati, Tetzlaff, Altman, & The PRISMA
Group, 2009).
Achievement
Mind-set theory suggests that mind-sets play critical Inclusion criteria, literature search, and coding. We
roles in academic achievement (Rattan, Savani, Chugh, searched for studies for both meta-analyses in a single
& Dweck, 2015). For example, Dweck (2008) stated, search. The criteria for including a study in Meta-Analysis 1
“what students believe about their brains — whether were as follows:
they see their intelligence as something that’s fixed or
something that can grow and change — has profound •• A measure of a belief about one or more human
effects on their motivation, learning, and school attributes (e.g., intelligence) as fixed or malleable—
achievement” (para. 2). In the first meta-analysis, we henceforth mind-set—was collected.
examined the magnitude of the relationship between •• A mind-set measure was collected prior to or with-
mind-sets and academic achievement. out a mind-set intervention.
Next, we investigated potential moderators. We •• A measure of academic achievement—course exam
examined academic risk status because the theory (e.g., midterm exam), course grade, average of
holds that having a growth mind-set is especially impor- course grades (e.g., grade point average, or GPA),
tant for at-risk students (e.g., Paunesku et al., 2015) and or standardized test performance—was collected
students facing situational challenges such as school prior to or without a mind-set intervention.
transitions (e.g., Yeager & Dweck, 2012). According to •• A bivariate correlation coefficient reflecting
the theory, students with growth mind-sets will inter- the relationship between mind-set and aca-
pret struggles as learning opportunities, while students demic achievement was reported, or enough
with fixed mind-sets will be “devastated by setbacks” information was provided to compute this
(Dweck, 2008, para. 2; see also Burnette et al., 2013). effect size.
Similarly, although the theory is not linked to a particu- •• The methods and results were in English.
lar age, some researchers suggest that mind-sets are
particularly influential during the tumultuous period of Mind-set is typically measured using participants’
adolescence when students face new challenges responses to statements such as, “No matter who you
(Blackwell, Trzesniewski, & Dweck, 2007). To assess the are, you can significantly change your intelligence
importance of this moderator, we examined student level” and “You have a certain amount of intelligence,
developmental stage. Additionally, we examined socio- and you can’t really do much to change it” (reverse
economic status (SES) because some research (e.g., scored) using a Likert scale (e.g., Dweck, 2006). The
Mind-Set and Academic Achievement 551

more students agree with statements about the malle- Effect sizes. To measure the magnitude of the relation-
ability of an attribute, the more of a growth mind-set ship, we used the correlation as the measure of effect
they hold. Measures of beliefs about the importance of size. For most studies, the authors reported a Pearson’s
effort without corresponding beliefs about the malle- correlation coefficient; for studies in which the authors
ability of one or more human attributes were not reported group-level comparisons (e.g., students holding
included. Likewise, mind-set of willpower was not a growth mind-set vs. a fixed mind-set), we converted
included because (a) willpower refers to exerted con- standardized mean differences (Cohen’s ds) to biserial
trol rather than an attribute, and (b) mind-set of will- correlations (rbs; Becker, 1986; Hunter & Schmidt, 1990).
power focuses on beliefs about whether willpower is There was not a significant difference in effect sizes
limited or not limited rather than whether an attribute between studies that reported group-level comparisons
is stable or changes with effort. and those that used continuous variables, p = .463. Most
To identify studies meeting these criteria and the studies’ authors coded higher scores on the mind-set
criteria set forth for Meta-Analysis 2, we systematically measure as reflecting more of a growth mind-set. When
searched for relevant published and unpublished arti- authors used a mind-set measure where higher scores
cles in psychology, education, and other disciplines reflected more of a fixed mind-set, we reversed the sign
through October 28, 2016 (for a flowchart designed of the correlation before analyzing the data. We also
according to the PRISMA specifications, see Fig. 1). We reversed the sign of the correlation in the rare cases
also e-mailed authors of articles on mind-set (N = 137) where lower scores on a measure of academic achieve-
and asked that they forward the e-mail to colleagues ment reflected better performance. For instance, in Ger-
who might have conducted relevant studies. Further, many, lower grades reflect better performance. Thus, all
we contacted organizations dedicated to intervention- effect sizes were coded such that a positive correlation
in-education research (e.g., PERTS) to request informa- reflected a positive relationship between growth mind-
tion relevant to our meta-analysis that was not accessible set and academic achievement.
(e.g., unpublished data), and we posted requests for
unpublished data on a Society for Personality and Social Moderator variables
Psychology forum. We accepted new data from these Developmental stage. There were three levels of
calls through January 11, 2017. Following our search developmental stage: children (primary school students),
stop date, we evaluated studies for eligibility and coded adolescents (middle school, junior high school, and high
each study and the measures collected in it for refer- school students), and adults (e.g., postsecondary stu-
ence information, student characteristics, methodologi- dents). Studies that included students in multiple cate-
cal characteristics, and results (the data file is available gories (e.g., students in both primary school and junior
at osf.io/453ds). We included updates to our existing high school in a single sample) were not included in this
records until analyses began on February 1, 2017. moderator analysis.
Our search included 15,867 novel records. After
examining these records and discarding obviously irrel- Academic risk status. There were three levels of aca-
evant ones (e.g., literature reviews, commentaries), we demic risk status: high (at risk of failing; e.g., students
identified 129 studies that met all the inclusion criteria who previously failed courses), moderate (facing a situ-
for Meta-Analysis 1. These studies included 162 inde- ational challenge; e.g., transitioning to a new school, a
pendent samples, with 273 effect sizes and a total sam- member of a stereotyped group under a stereotype threat
ple size of 365,915 students. In cases where authors manipulation), and low (no indicators that students were
reported effects associated with multiple measures of at risk). Each sample was categorized on the basis of
mind-set (e.g., a fixed mind-set scale and a growth the majority (> 50%) of the students in the sample. If we
mind-set scale) or multiple measures of academic could obtain separate effect sizes for each subsample in a
achievement (e.g., GPA and performance on a standard- study based on risk level (e.g., an effect was available for
ized test), we adjusted for dependent samples by using the high-risk students as well as the remaining low-risk
a method based on that of Cheung and Chan (2004, students), we did so and entered those effects as inde-
2008). This method statistically adjusts (lowers) the pendent samples. If effects were available only for the
associated sample size because of dependent effects entire sample and a high-risk subgroup, we replaced the
being partially redundant, which reduces the weight of entire sample with the high-risk subgroup when examin-
these effect sizes in the meta-analysis so as not to overly ing this moderator.
contribute to the model. For a list of studies included We did not code minority students or female students
in the meta-analysis, see the Supplemental Material as academically at-risk samples unless they were under
available online or the file at osf.io/453ds. For addi- a relevant stereotype threat manipulation. While stu-
tional characteristics of Meta-Analysis 1, see Table 1. dents can experience stereotype threat in natural
552 Sisk et al.

Search Features
(through October 28, 2016)
• Searching electronic databases (ERIC, PsycINFO, PubMed, ProQuest Dissertations & Theses) and Google Scholar, using
combinations of the following search terms: implicit theory, mindset, self theory, lay theory, intelligence, academic,
education, Dweck
• Scanning curriculum vitaes of implicit-theory researchers
• Scanning reference lists in publications on implicit theories
Search

• Scanning blogs and websites discussing implicit theories

• Scanning tables of contents in relevant journals
• Searching past conference programs for relevant studies
• Sending e-mail requests to authors (N = 137) of articles on implicit theories requesting unpublished data
• Posting requests on a Society for Personality and Social Psychology forum

Records Searched Records After Duplicates Removed

(N = 24,766) (N = 15,858)

Criteria For Study Inclusion: Meta-Analysis 1 Criteria For Study Inclusion: Meta-Analysis 2
• Must include a measure of students’ mind-sets • Must include a mind-set intervention for students
assessing belief about the malleability of a human • Must include a measure of academic achievement
attribute (e.g., intelligence) • Grade average (e.g., GPA) • Course grade
Inclusion Criteria

• Must include a measure of academic achievement • Course exam score • Standardized test score
• Grade average (e.g., GPA) • Course grade • Must include a comparable control group
• Course exam score • Standardized test score • Must report an effect size reflecting the difference in
• Must report the bivariate correlation coefficient academic achievement between the treatment and the
between mind-set and academic achievement or control groups following the mind-set intervention or
enough information needed to compute this effect size information needed to compute this effect size
• Must report methods and results in English • Must report methods and results in English

Abstracts Screened Abstracts Excluded Full-Text Articles Evaluated for Eligibility

(N = 15,858) (N = 15,495) (N = 354)

Full-Text Articles Evaluated but Excluded From One or Both Meta-Analyses (N = 207)
• No measure of students’ mind-sets (meeting our criteria) was collected (Meta-Analysis 1)
• No baseline (preintervention) measure of students’ mind-sets (meeting our criteria) was collected (Meta-Analysis 1)
• No measure of academic achievement (meeting our criteria) was collected (Meta-Analysis 1, Meta-Analysis 2)
• No baseline (preintervention) measure of academic achievement (meeting our criteria) was collected (Meta-Analysis 1)
Eligibility

• No postintervention measure of academic achievement (meeting our criteria) was collected (Meta-Analysis 2)
• No mind-set intervention (meeting our criteria) was administered directly to students (Meta-Analysis 2)
• No (comparable) control group was reported (Meta-Analysis 2)
• Not enough information about the measures or methods was reported to evaluate (Meta-Analysis 1, Meta-Analysis 2)
• No (relevant) quantitative results are reported (Meta-Analysis 1, Meta-Analysis 2)
• A bivariate correlation coefficient (r) between students’ mind-sets and academic achievement was not reported, and not
enough information was provided to calculate r (Meta-Analysis 1)
• A standardized mean difference (d ) between the treatment and control group on academic achievement after intervention
was not reported, and not enough information was provided to calculate between-groups d ( Meta-Analysis 2)
• The data appear to overlap with data from another study, or no original results reported (Meta-Analysis 1, Meta-Analysis 2)
• Methods and results were not in English (Meta-Analysis 1, Meta-Analysis 2)

Meta-Analysis 1 Meta-Analysis 2
• 123 records included • 27 records included
Included

• 129 studies included • 29 studies included

• 162 independent samples • 38 independent samples
• 273 effect sizes • 43 effect sizes
• N = 365,915 • N = 57,155

Fig. 1. Flow chart of the literature search and study coding. Three articles were included in both Meta-Analysis 1 and Meta-
Analysis 2 (i.e., 147 unique records are included in the two meta-analyses). GPA = grade point average.
Mind-Set and Academic Achievement 553

Table 1. Descriptive Characteristics for Meta-Analysis 1 varied across studies. Third, levels of achievement vary
across studies such that academic risk for minority stu-
Number
of effect Number of dents and women is not constant. That is, low-achieving
sizes participants minority students do not have the same level of aca-
Study characteristic (k = 273) (N = 365,915) demic risk as high-achieving minority students. Fourth,
minority status can be confounded with SES, which we
Developmental stagea
coded separately. The effect of student ethnicity and
Children 50 8,118
gender on the relationship between growth mind-set
Adolescents 126 332,240
and academic achievement is an important research
Adults 89 21,673
question. However, we did not have the level of detail
Academic risk statusb
Low 208 346,043
that would allow us to conduct a meaningful moderator
Moderate 55 19,215 analysis on ethnicity or gender as risk factors.
High 6 218
Socioeconomic statusb Socioeconomic status. There were two levels of SES:
Low 33 173,614 low SES (e.g., students qualified for reduced-price lunch)
Not low 62 27,160 and not low SES (i.e., middle class or higher). Each study
Mind-set type was categorized on the basis of the majority (> 50%) of
Intelligence 167 335,560 the students in the sample. Studies not reporting student-
Other attribute (e.g., math ability) 106 30,355 level SES were not included in the SES moderator analysis.
Academic achievement measure
Course exam 15 9,318 Type of academic achievement measure. There were
Course grade 51 11,384 four levels of academic achievement measure: standard-
Average grades (i.e., grade point 82 46,986 ized test (e.g., Iowa Test of Basic Skills, SAT), and three
average) pertaining to course performance—course exam (e.g., final
Standardized test 125 298,227 exam score), course grade (e.g., math course grade), and
Laboratory measures 24 2,121 cumulative or current GPA. When studies included mul-
Publication status tiple standardized test scores (e.g., verbal SAT, quantitative
Published 116 323,040
SAT, total SAT), we used the combined score when avail-
Unpublished 157 42,875
able. When studies included multiple course performance
a
For this characteristic, some effect sizes are excluded because they measures, we used the measure that provided the most
included a wide age range. comprehensive measure of academic achievement. That
b
For this characteristic, one or more effect sizes were excluded is, we used GPA when available because this provides the
because this information was not provided in the study.
most information about a students’ course performance.
Likewise, we used course grades over course exams. We
environments, we cannot know whether this occurred did not include enrollment status (full time vs. part time)
in each study without stereotype threat being measured or number of absences as academic achievement mea-
or manipulated. Some mind-set researchers have cat- sures because they are not readily comparable with course
egorized minority students or female students as aca- performance or standardized test performance.
demically at risk and analyzed their results separately Some studies included measures of academic
without measuring or manipulating stereotype threat. achievement administered by a researcher (e.g., prac-
However, this categorization was not suitable in the tice questions on the GRE, a researcher-designed
present meta-analysis for multiple reasons. First, the course-relevant test) as a proxy for academic achieve-
cultural context varies across studies (Meta-Analysis 1, ment. We did not include researcher-designed tests as
for instance, includes studies from over 20 different course exams if they were irrelevant to students’ course-
countries; see the Supplemental Material for more work (e.g., trivia quizzes, reading comprehension of
detail). Cultural conceptualizations of ethnic minorities the mind-set stimulus, worksheets on topics described
vary across countries, such that identifying and catego- as outside students’ curricula). We present the results
rizing at-risk groups in each sample to compare across with and without laboratory measures because perfor-
studies would not be feasible or precise. Second, the mance on these measures does not contribute to stu-
educational contexts vary across studies in such a way dents’ academic records.
that the association between academic risk and minority
status is likely not consistent. In other words, the type Developmental stage as a moderator of mind-set on
and degree of challenge that minority students face GPA. If students with a growth mind-set select more chal-
depends on their educational environment, which lenging courses or schools, it is possible that their GPAs
554 Sisk et al.

would not be significantly higher than fixed mind-set not correct individual effect sizes for the attenuating
students taking easier classes, leading to the relationship effect of measurement error (i.e., measurement unreli-
between mind-set and GPA being suppressed, especially ability), because very few studies in the meta-analysis
for older students who have more opportunities for course reported a reliability estimate for mind-set. However,
selection. This suppression would affect only GPA, which measures of mind-set have typically been found to have
reflects students’ course selections. It would not suppress acceptable reliability greater than .80 (see, e.g., Dweck
the effect on course exams or course grades because all et al., 1995). If we assume reliability of .80, the meta-
students in the sample are in the same course. It would analytic average correlation between mind-set and aca-
not suppress the effect on standardized tests because, if demic achievement is r = .12, 95% CI = [.09, .14].
anything, students taking more challenging courses will Figure 2 shows that 157 of the 273 effect sizes (58%)
be better prepared for standardized tests than students are not significantly different from zero. Another 16
not exposed to higher-level material. We therefore also effect sizes (6%) are significantly different from zero
examine the interaction between mind-set and develop- but negative, indicating that growth mind-sets were
mental stage on GPA. associated with worse academic achievement. The
remaining 100 effect sizes (37%) are significantly dif-
Meta-analytic procedure. The meta-analysis involved ferent from zero and positive, indicating that growth
four steps. The first step was to obtain correlations mind-sets were positively associated with academic
between mind-set of a human attribute and academic achievement. As can be seen in Figure 2, the effect sizes
achievement, along with their sampling error variances. are not consistent across studies. The I2 statistic speci-
The second step was to search for extreme values. We fies the percentage of the between-studies variability
defined outliers as effect sizes whose residuals had z in effect sizes that is due to heterogeneity rather than
scores of 3 or greater. There were no outliers for this random error. The I2 statistic, I2 = 96.29 (τ2 = .025),
meta-analysis. The third step was to use random-effects demonstrated a very large proportion of heterogeneity
meta-analysis modeling, which assumes meaningful dif- in the effect sizes, indicating that the true effect of a
ferences across studies; estimate the meta-analytic mean given study could be substantially higher or lower than
distribution of effects and heterogeneity in the effect the meta-analytic average. We investigated the source
sizes; and then test whether some of the heterogeneity of this heterogeneity through the moderator analyses
was predictable from moderator variables, using mixed- reported next.
effects meta-analysis modeling. The final step was to per-
form publication bias analyses. Moderator analyses
We used the Comprehensive Meta Analysis Version Student factors. The developmental stage of the stu-
2 (Borenstein, Hedges, Higgins, & Rothstein, 2005) soft- dents was a statistically significant moderator, Q(2) =
ware package to conduct the meta-analyses. We used 72.84, p < .001. Eight effect sizes associated with samples
Comprehensive Meta Analysis, the R Project for Statisti- with a wide age range encompassing more than one of
cal Computing (www.r-project.org), the p-curve web the developmental levels (e.g., children and adolescents)
application (http://www.p-curve.com/app), and the were excluded from this analysis. The average correla-
p-uniform web application (https://rvanaert.shinyapps tion between mind-set and academic achievement was
.io/p-uniform) to conduct the publication bias analyses. r = .19, 95% CI = [.16, .23], p < .001, for children; r = .15,
(See also the Supplemental Material.) 95% CI = [.12, .18], p < .001, for adolescents; and r = .02,
95% CI = [−.005, .05], p = .110, for adults. Post hoc follow-
up analyses were conducted to determine the source of
Results the difference using a corrected alpha of .05/3 = .017 for
The model consists of 273 effect sizes. Effect sizes are multiple comparisons. Adults differed significantly from
weighted by the inverse of the variance, which includes both adolescents, Q(1) = 39.89, p < .001, and children,
both between-studies variance and within-study vari- Q(1) = 58.37, p < .001. Adolescents and children did not
ances. Within-study variance accounts for the sample differ significantly from each other, Q(1) = 3.47, p = .063.
sizes, such that smaller studies are given less weight Academic risk status was not a significant moderator,
while larger studies are given more weight in the model. Q(2) = 0.22, p = .895. Four effect sizes did not have a
The majority of effect-size–associated adjusted Ns are sample description other than age and location and
≥ 90. The mean adjusted N associated with this model’s thus were removed from this analysis. The average cor-
effect sizes is 1,429. relation between mind-set and academic achievement
The meta-analytic average correlation (i.e., the aver- was r = .11, 95% CI = [.08, .13], p < .001, for low-risk
age of various population effects) between growth students; r = .11, 95% CI = [.07, .16], p < .001, for mod-
mind-set and academic achievement is r = .10, 95% erately at-risk students; and r = .08, 95% CI = [−.04,
confidence interval (CI) = [.08, .13], p < .001. We did .21], p = .196, for highly at-risk students.
Mind-Set and Academic Achievement 555

Black (2008) – S4 M1
Schullo (1996) – S6 M2
Black (2008) – S3 M3
Black (2008) – S8 M1
Black (2008) – S8 M2
Black (2008) – S3 M1
Dvorak (2014) – S1 M2
Black (2008) – S6 M2
Dvorak (2014) – S2 M3
Froehlich, Martiny, Deaux, Goetz, & Mok (2016) – S1
Schullo (1996) – S2 M2
Black (2008) – S7 M1
Black (2008) – S7 M3
Gaultney (1989) – S1 M2
Schullo (1996) – S8 M2
Dvorak (2014) – S2 M2
Gaultney (1989) – S1 M3
Schullo (1996) – S7
Black (2008) – S5 M4
Cain, Leonard, Gabrieli, & Finn (2016) – M1
Cain et al. (2016) – M2
Schullo (1996) – S8 M1
Black (2008) – S6 M1
Black (2008) – S4 M2
Schullo (1996) – S6 M1
Schullo (1996) – S2 M1
Gaultney (1989) – S1 M1
Benningfield (2013) – S3
Delavar, Ahadi, & Barzegar (2011) – M1
West (2016)
West et al. (2016) – M2
Dvorak (2014) – S2 M1
Claro, Paunesku, & Dweck (2016)
Gaultney (1989) – S2 M4
Shih (2007)
Da Fonseca (2009) – M2
Gaultney (1989) – S1 M4
Romero, Master, Paunesku, Dweck, & Gross (2014)
Black (2008) – S7 M4
Froehlich et al. (2016) – S8
Linehan (1998) – M3
West et al. (2016) – M1
Black (2008) – S3 M2
Chen & Pajaras (2010) – M2
Gaultney (1989) – S2 M2
Hazard (1997) – M3
Law (2009)
Schullo (1996) – S4 M2
Black (2008) – S5 M3
Chen & Pajaras (2010) – M1
Cury, Elliot, Da Fonseca, & Moller (2006) – M2
Gaultney (1989) – S4 M4
Linehan (1998) – M1
Kraft & Grace (2016) – M2
Black (2008) – S3 M4
Parker (2016)
Riley (2003)
Timpone & Hostutler (2012)
Bettinger, Ludvigsen, Rege, Solli, & Yeager (2016)
King (2012) – M1
Blackwell, Trzesniewski, & Dweck (2007) – S2 M2
Black (2008) – S7 M2
Volpe (2016) – M3
Faria & Fontaine (1997) – M2
Froehlich et al. (2016) – S4
Kraft & Grace (2016) – M1
Schullo (1996) – S1 M1
Volpe (2016) – M2
Ehrlinger, Hartwig, et al. (2016a)
Chen (2012) – M1
Diseth, Meland, & Breidablik (2014) – M2
Ehrlinger, Hartwig et al. (2016b)
Eskreis-Winkler et al. (2016) – S2 M2
Luo, Lee, Ong, Ming Wong, & Fah Foo (2014) – M3
Stipek & Gralinski (1996) – M2
Stipek & Gralinski (1996) – M4
Zhao & Wang (2014)
Volpe (2016) – M1
Gonida, Kiosseoglou, & Leondari (2006)
Wu & Kraemer (2016)

–1.00 0.00 1.00

Fig. 2 (continued on next page)
556 Sisk et al.

Cury et al. (2006) – M1

Delavar et al. (2011) – M2
Eskreis-Winkler et al. (2016) – S2 M1
Haimovitz, Wormington, & Corpus (2011)
Mouratidis, Michou, & Vassiou (2016)
Black (2008) – S1 M4
MacGyvers (1993) – S2
Benningfield (2013) – S4
Eskreis-Winkler et al. (2016) – S3 M2
Gaultney (1989) – S2 M3
Hendricks (2012)
Shih (2007) – M2
Stipek & Gralinksi (1996) – M1
Williams, Ari, & Dortch (2011)
Bagley (2016) – S2
Dai & Cromley (2014) – M1
Gaultney (1989) – S2 M1
Yeager et al. (2016)
Eskreis-Winkler et al. (2016) – S1 M2
MacGyvers (1993) – S1
Chen (2012) – M2
Eskreis-Winkler et al. (2016) – S1 M1
Jones, Wilkins, Long, & Wang (2012)
Luo et al. (2014) – M1
Black (2008) – S1 M3
Volpe (2016) – M4
Tarbetsky, Collie, & Martin (2016) – S1
Gubi (2012) – M2
Tarbetsky et al. (2016) – S2
Froehlich et al. (2016) – S2
Schwinger, Steinmayr, & Spinath (2016)
Chen & Wong (2015)
Eskreis-Winkler et al. (2016) – S3 M1
Gaultney (1989) – S4 M1
Priess-Groben & Hyde (2016)
Renaud-Dube (2015)
Rheinschmidt & Mendoza-Denton (2014) – S3 M1
Riyaz (2013) – M1
Shively & Ryan (2013) – S2
Schullo (1996) – S5 M1
Boazman (2010) – S2 M2
Dupeyrat & Mariné (2005) – M1
Linehan (1998) – M2
Riyaz (2013) – M6
Stipek & Gralinski (1996) – M3
Tucker-Drob, Briley, Engelhardt, Mann, & Harden (2016) – M2
Kappes, Stephens, & Oettingen (2011)
Bagley (2016) – S1
Broome (2001) – S1
Benningfield (2013) – S2
Diseth et al. (2014) – M1
Guich (2007) – M1
Kench, Hazelhurst, & Otulaja (2016)
Dvorak (2014) – S3 M2
Blackwell et al. (2007) – S1 M2
King (2012) – M2
Da Fonseca (2009) – M1
Dupeyrat & Mariné (2005) – M2
Martin, Nejad, Colmar, & Liem (2013) – M1
Tucker-Drob et al. (2016) – M1
Benningfield (2013) – S1
Fleming (2007)
Rheinschmidt & Mendoza-Denton (2014) – S2 M1
Uchida (2004)
Bagley (2016) – S3
Faria & Fontaine (1997) – M1
Hazard (1997) – M1
Hazard (1997) – M2
Rudig (2014)
Harpalani (2005)
Blackwell et al. (2007) – S1 M1
De Castella & Byrne (2015)
Leondari & Gialamas (2002)
Stipek & Gralinski (1996) – M5
Ziegler, Fidelman, Reutlinger, Vialle, & Stoeger (2010)
Black (2008) – S5 M2
Kappes, Oettingen, & Mayer (2011)
Shell, Soh, Flanigan, & Peteranetz (2016) – M4
Black (2008) – S2 M1
Dai & Cromley (2014) – M2

–1.00 0.00 1.00

Fig. 2 (continued on next page)
Mind-Set and Academic Achievement 557

Riyaz (2013) – M2
Stipek & Gralinski (1996) – M7
Kornilova, Chumakova, & Izmailova (2015) – S1
Gaultney (1989) – S3 M1
Hotulainen & Telivuo (2014)
Shively & Ryan (2013) – S1
Shell et al. (2016) – M1
Luo et al. (2014) – M4
Riyaz (2013) – M5
Riyaz (2013) – M7
Schnedecker (1997) – M1
Stipek & Gralinski (1996) – M6
Blake (2015)
Edwards (2014)
Howell (2009)
Shell et al. (2016) – M2
Pepi, Alesi, & Geraci (2004)
Ravenscroft, Waymire, & West (2012)
Howell (2009) – M1
Luo et al. (2014) – M2
Riyaz (2013) – M3
Stump, Husman, & Corby (2014) – M1
Schroder, Dawood, Yalch, Donnellan, & Moser (2016) – M1
Froehlich et al. (2016) – S3
Zientek, Yetkiner Ozel, Fong, & Griffin (2013)
Gaultney (1989) – S4 M3
Stipek & Gralinsky (1996) – M8
Ziegler & Stoeger (2010)
Dickhäuser, Dinger, Janke, Spinath, & Steinmayr (2016)
Furnham, Chamorro-Premuzic, & McDougall (2002)
Fillmore (2015)
Hwang, Reyes, & Eccles (2016) – M1
Kennett & Keefer (2006)
Magno (2012) – M2
Miller (2010) – M1
Feldman, Chandrashekar, & Wong (2016)
Gaultney (1989) – S3 M4
Gaultney (1989) – S4 M2
Holden, Moreau, Greene, & Conway (2016)
Lindsay (2006)
Rheinschmidt & Mendoza-Denton (2014) – S1
Riyaz (2013) – M4
Schnedecker (1997) – M2
Ryan et al. (2007) – M2
Shell et al. (2016) – M3
Boazman (2010) – S2 M1
Greene, Oswald, & Pomerantz (2015)
Hwang et al. (2016) – M2
Ryan et al. (2007) – M1
Robin & Pals (2002) – M2
Boazman (2010) – S1 M2
Magno (2012) – M1
Martin et al. (2013) – M2
Northrop (2014) – M1
Northrop (2014) – M2
Black (2008) – S2 M2
Ehrlinger, Mitchum, & Dweck (2016)
Miller (2010) – M2
Ehrlinger & Conlon (2016)
Callahan, Schroder, & Moser (2015)
Dvorak (2014) – S3 M3
Blackwell et al. (2007) – S2 M1
Clevenger (2013)
Dinger, Dickhäuser, Spinath, & Steinmayr (2013)
Stump et al. (2014) – M2
Tempelaar, Rienties, Giesbers, & Gijselaers (2015) – M1
Tempelaar et al. (2015) – M3
Flanigan, Peteranetz, Shell, & Soh (2015) – M2
Tallman (2000)
Schullo (1996) – S1 M2
Boazman (2010) – S1 M1
Guich (2007) – M2
Kornilova, Kornilov, & Chumakova (2009)
Rheinschmidt & Mendoza-Denton (2014) – S2 M2
Black (2008) – S5 M1
Gaultney (1989) – S3 M2
Tempelaar et al. (2015) – M2
Schroder et al. (2016) – M2
Cordell-McNulty (2009)
Flanigan et al. (2015) – M1

–1.00 0.00 1.00

Fig. 2 (continued on next page)
558 Sisk et al.

Ehrlinger & Brewer (2016)

Adatitomo (2015)
Matheson (2015)
Solotruk (2013)
Ehrlinger & Dweck (2016)
Lewis (1998)
Kornilova et al. (2015) – S2
Dvorak (2014) – S1 M1
Rheinschmidt & Mendoza-Denton (2014) – S3 M2
Black (2008) – S1 M1
Macdonald (2016)
Black (2008) – S1 M2
Cadwallader (2009)
Schullo (1996) – S3 M1
Froehlich et al. (2016) – S5
Robins & Pals (2002) – M1
Froehlich et al. (2016) – S6
Gubi (2012) – M1
Schullo (1996) – S5 M2
P’Pool (2012) – M1
Moser, Moran, Schroder, Donnellan, & Yeung (2013)
Gaultney (1989) – S3 M3
Broome (2001) – S2
Ehrlinger (2016)
Gallagher (2006)
Northrop (2014) – M3
Dvorak (2014) – S3 M1
Froehlich et al. (2016) – S7
Schullo (1996) – S3 M2
Bergen (1991)
P’Pool (2012) – M2
Shedlosky-Shoemaker & Fautch (2015)
Schullo (1996) – S4 M1

–1.00 0.00 1.00

Fig. 2. Correlations between growth mind-set and academic achievement. Correlations (squares) and 95% confidence intervals (error bars)
are displayed for all effects entered into Meta-Analysis 1. The diamond on the bottom row represents the meta-analytically weighted mean
correlation coefficient. For studies with multiple independent samples, the result for each sample (S1, S2, etc.) is reported separately. Simi-
larly, for studies with multiple measures, the result for each measure (M1, M2, etc.) is reported separately. Multiple measures were adjusted
for dependency. See the Supplemental Material available online for details on all references.

Socioeconomic status was not a significant moderator, effect sizes where the measure of academic achieve-
Q(1) = 1.48, p = .223. The 95 effect sizes associated with ment was a laboratory-based measure did not change
reported student-level SES were included in this analysis. the overall results. The overall meta-analytic average
The average correlation between mind-set and academic correlation with these effect sizes removed was r = .11,
achievement was r = .17, 95% CI = [.10, .23], p < .001, 95% CI = [.08, .13], p < .001 (compare with r = .10, 95%
for low-SES students, and r = .12, 95% CI = [.09, .16], CI = [.08, .13], p < .001, when including these effect
p < .001, for middle-class and higher students. sizes). Excluding laboratory-based effect sizes did
change the pattern of results for the academic achieve-
Academic achievement measure. The measure of aca- ment measure moderator. Without laboratory measures,
demic achievement used was not a statistically significant the relationship between mind-set and academic
moderator, Q(3) = 6.18, p = .103. The average correlation achievement varied significantly on the basis of the
between mind-set and academic achievement was r = academic achievement measure used, Q(3) = 13.12,
.08, 95% CI = [.01, .15], p = .027, for studies that used a p = .004. Specifically, removing researcher-designed
course exam; r = .13, 95% CI = [.09, .16], p < .001, for course exams lowered the average correlation between
studies that used a course grade; r = .08, 95% CI = [.05, mind-set and course exam performance, r = .04, 95%
.11], p < .001, for studies that used GPA; and r = .12, 95% CI = [−.02, .11], p = .178 (compare with when these
CI = [.09, .15], p < .001, for studies that used a standard- effect sizes were included, r = .08, 95% CI = [.01, .15],
ized test. p = .027). The correlation between mind-set and aca-
Twenty-four effect sizes reflected the relationship demic achievement was similar for standardized tests
between mind-set and a measure of academic achieve- regardless of whether laboratory-based effect sizes
ment that was laboratory based. These included were excluded, r = .14, 95% CI = [.11, .17], p < .001, or
researcher-designed tests supposed to reflect compre- included, r = .12, 95% CI = [.09, .15], p < .001. Post hoc
hension of course-specific content (coded as a course follow-up tests using an adjusted alpha of .008 for mul-
exam) and standardized tests and portions of standard- tiple comparisons revealed that when laboratory-based
ized tests administered by researchers in a laboratory measures were removed, course exams and GPA dif-
setting (coded as standardized tests). Excluding the 24 fered significantly from standardized tests, Q(1) = 7.14,
Mind-Set and Academic Achievement 559

p = .0075, and Q(1) = 8.36, p = .004, respectively. No versus unpublished studies. The second analysis tested
other pairwise comparisons were significant at the .008 whether selective reporting of results (i.e., p-hacking)
level, all ps > .029. was responsible for significant effects in the published
literature. We tested this via p-curve analysis. The third
Developmental stage as a moderator of mind-set on analysis tested whether our meta-analysis is overestimat-
GPA. If students with growth mind-sets select more chal- ing the meta-analytic mean effect size due to missing
lenging courses, we would expect two patterns of results. unpublished data we were unable to obtain and include.
First, the relationship between mind-set and academic We tested this via Egger’s regression.
achievement would not be suppressed for children who
typically have little control over their course selection, Moderator analysis. We tested whether published stud-
somewhat suppressed for adolescents who have more ies, on average, report larger effect sizes than studies that
course selection opportunities, and most suppressed remain unpublished. Unpublished studies included man-
for adults who have the most opportunities for course uscripts in preparation to submit, manuscripts submitted
selection. The other pattern of results we would expect if but not yet accepted, conference papers and posters, and
students with growth mind-sets are selecting more chal- studies and manuscripts that have remained unpublished.
lenging courses is that the relationship between mind-set Studies may remain unpublished for multiple reasons. For
and academic achievement will be suppressed when the example, a study may remain unpublished because its
measure of academic achievement is GPA, because GPA methodology is weak. That is, the author may choose to
reflects performance in the students’ selected courses. not submit the study for publication, or the study may be
The relationship should not be suppressed for this rea- rejected for publication because of methodological short-
son when the measure is course grade or course exam comings. Alternatively, studies may remain unpublished
because, in these cases, all students are taking the same because they found small effects, null effects, or effects
courses. The relationship should also not be suppressed in the nonpredicted direction. That is, studies with results
for this reason when the measure is standardized test that do not strongly support a particular hypothesis may
performance because, if anything, students exposed to not be submitted for publication or may be rejected from
higher-level material should perform better than students publication. If this occurs, the overall effect size is overes-
taking less challenging courses. timated within the published literature.
Only four effect sizes were associated with children’s A moderator analysis revealed that the 157 correla-
mind-set and GPA and thus were excluded from this tions between mind-set and academic achievement
analysis as were samples associated with a wide range from unpublished studies (median study sample size =
of ages (e.g., children and adolescents in the same 122) were not significantly different from the 116 cor-
sample). The difference between adolescents and adults relations from published studies (median study sample
was significant, Q(1) = 17.01, p < .001, with adolescents size = 245), Q(1) = 0.65, p = .420. The average correla-
exhibiting a stronger relationship between mind-set and tion between mind-set and academic achievement was
GPA, r = .12, 95% CI = [.07, .16], p < .001, than adults, r = .09, 95% CI = [.07, .12], p < .001, for unpublished
r = .002, 95% CI = [−.03, .03], p = .892. This pattern of studies and r = .11, 95% CI = [.08, .14], p < .001, for
results supports the suppression hypothesis. However, published studies.
the relationship between mind-set and GPA did not
differ from the relationship between mind-set and other p-curve analysis. We tested whether the source of
measures of academic achievement where course selec- published significant effects was due to p-hacking; that
tion is unlikely to affect the relationship, Q(3) = 6.18, is, selective reporting of results (e.g., when authors con-
p = .103. Additionally, the relationship between mind- duct multiple analyses on the same data set but report
set and academic achievement was lowest for adults only significant effects) or collecting data until a nonsig-
across all measures of academic achievement (r = .02, nificant effect becomes significant (though see Bishop &
95% CI = [−.005, .05], p = .110), including those where Thompson, 2016).
course selection is unlikely to affect this relationship. The logic underlying the p-curve analysis is as fol-
lows. A p-curve depicts the distribution of statistically
Publication bias analyses. Publication bias threatens significant (p < .05) p values for a set of studies
the validity of published research by masking small and (Simonsohn, Nelson, & Simmons, 2014a). The p-curve
null effects. We conducted three types of publication bias analysis tests the skew of the p-value distribution. Stud-
analyses. The first analysis tested whether studies finding ies demonstrating true effects will yield a right-skewed
weak effect sizes were less likely to be published than p-curve, indicating more lower significant p values (e.g.,
studies finding stronger effects. We tested this via mod- p = .001) than marginally significant p values (e.g., p =
erator analysis comparing effect sizes from published .049). Studies demonstrating null effects will yield a
560 Sisk et al.

uniform (i.e., “flat”) p-curve. Studies with predominately was Z = −16.14, p < .0001, and the result of the full
p-hacked effects will yield a left-skewed p-curve, indi- p-curve test was Z = −16.20, p < .0001. These results
cating more marginally significant p values than lower suggest that the p-curve is significantly right-skewed,
significant p values (Simonsohn, Nelson, & Simmons, indicating evidential value. These results were corrobo-
2014a, 2014b). The results of this test should be inter- rated by the robustness analysis (half: Z = −21.16, p <
preted with caution because p-curve estimates may be .0001; full: Z = −20.25, p < .0001). The primary and
affected by large amounts of heterogeneity (van Aert, robustness p-curve analyses estimated that after cor-
Wicherts, & van Assen, 2016). recting for selective reporting, the included studies had
We conducted the p-curve analyses in two steps. an estimated power of 99%. For additional figures and
First, we classified effect sizes as either published or an index of the effect sizes that were entered into each
unpublished. If authors provided data that were not analysis, see the Supplemental Material.
reported in a published study, the effect sizes calculated
using those data were classified as unpublished for the Egger’s regression. We tested whether our meta-anal-
purposes of this analysis. Simonsohn et al. (2014b) rec- ysis was affected by missing studies. We found a consid-
ommended that when choices among multiple effects erable number of unpublished studies to include in the
must be made to adhere to a prespecified selection rule meta-analysis. However, during our search process, we
for the first (i.e., primary) and second (i.e., robustness) became aware of unpublished studies that appeared to
analyses (Simonsohn et al., 2014a). When there were fit our inclusion criteria that we were unable to obtain.
multiple effect sizes from the same study included in To examine how much these missing studies potentially
the meta-analysis, we randomly selected one effect size affected our meta-analysis, we conducted Egger’s regres-
from each study for the primary analysis and then ran- sion (Egger, Smith, Schneider, & Minder, 1997). If a meta-
domly selected a different effect size for the robustness analysis is unaffected by publication bias, larger studies’
analysis. This process ensured that studies with a large effect sizes will cluster around the mean effect size, while
number of effect sizes did not have undue influence smaller studies’ effect sizes (containing more sampling
on the analyses. Next, we ran primary and robustness error) will be randomly dispersed around the mean (i.e.,
p-curve analyses for the published effect sizes using higher and lower than the mean). If a meta-analysis is
the p-curve web application (http://www.p-curve.com/ affected by publication bias, smaller studies will con-
app). tribute significantly more effect sizes with higher-than-
The results of the primary p-curve analysis are pre- average effects than smaller-than-average effects. Egger’s
sented in Figure 3. The result of the half p-curve test regression was significant, B0 = −2.98, 95% CI = [−3.58,

100 Observed p-curve

Power estimate: 99%, CI = [98%, 99%]
83% Null of no effect
Tests for right-skewness: p full < .0001, p half < .0001

75 Null of 33% power

Percentage of Test Results

Tests for flatness: p full > .9999, pbinomial = .9968

50

25

7% 7%
3%
0%
0

.01 .02 .03 .04 .05

p value

Fig. 3. Primary p-curve analysis for Meta-Analysis 1.

Mind-Set and Academic Achievement 561

−2.39], t(271) = 9.92, p < .001, suggesting that the meta- group (active control, passive control, fixed mind-set).
analysis is likely overestimating the relationship between If studies using passive control groups have the largest
mind-set and academic achievement. However, Egger’s effects, this suggests that exposure to treatments might
regression is prone to Type I errors when heterogeneity drive the effect rather than growth-mind-set interven-
is high (Sterne et al., 2011), as is the case in the cur- tions per se. Alternatively, if growth mind-sets are ben-
rent meta-analysis. Thus, this result should be interpreted eficial for academic achievement and fixed mind-sets
with caution. are detrimental, we should see the largest effect when
the comparison group is a fixed-mind-set condition.
We examined the type of intervention to test whether
Discussion
interactive (e.g., saying-is-believing) interventions are
The meta-analytic average correlation between growth more effective than passive interventions. The number
mind-set and academic achievement was very weak— of intervention sessions was examined as a continuous
r = .10. This result is almost identical to the meta- variable to test whether there is a linear additive effect
analytic average correlation found between mind-set of intervention exposure. We included mode of inter-
and achievement across achievement domains: r = .095 vention (computerized, in person, reading materials,
(Burnette et al., 2013). However, the overall effect is combination) to test whether certain modalities are
overshadowed by the high degree of heterogeneity. more effective than others. For interventions at least
Moderators were limited in accounting for this vari- partially administered in person, we further classified
ance. Academic risk status and SES did not affect the whether administers were teachers, researchers, or
relationship. Developmental stage moderated the rela- both. We include intervention context (integrated in
tionship, though the effect remained weak for all sub- the classroom, outside regular classroom activities)
groups and nonsignificant for adults. This pattern held because some researchers have suggested that mind-set
when examining only GPA as the outcome, and GPA interventions might be context dependent (Yeager &
did not differ from other measures of academic achieve- Walton, 2011).
ment. Thus, there is limited evidence for a suppression We examined whether studies included a manipula-
effect due to students with growth mind-sets potentially tion check and whether the manipulation check was
selecting more challenging courses. successful. We included these measures because if
Growth-mind-set interventions in education are mind-set interventions are a scalable treatment, we
predicated on the relationship between mind-sets and should expect most manipulation checks to be success-
academic achievement. However, it is possible that ful, and if mind-set interventions are generally effective,
despite generally weak relationships between students’ we would expect null results only when manipulation
naturally held mind-sets and academic achievement, checks are unsuccessful.
interventions promoting growth mind-sets might still We also investigated factors related to the measure
be effective, especially for certain subgroups. We exam- of academic achievement: intervention-achievement
ined the effectiveness of growth-mind-set interventions measure interval and type of academic achievement
on academic achievement next. measure. If mind-set interventions are susceptible to the
fadeout effect (Protzko, 2015), we should expect stron-
ger effects the shorter the intervention-achievement
Meta-Analysis 2: The Effect of
measure interval. In contrast, if mind-set interventions
Growth-Mind-Set Interventions on interact with recursive processes (Yeager & Walton,
Academic Achievement 2011) the effects should be sustained (or enhanced)
Growth-mind-set interventions have been suggested as with additional time. Finally, we tested whether publica-
a way for students to earn higher grades and score tion bias is problematic within the mind-set intervention
higher on standardized tests (see mindsetscholarsnet literature.
work.org/learning-mindsets/growth-mindset/). To
examine the effectiveness of these interventions, we Method
estimated the standardized mean differences in aca-
demic achievement between students who received a As with Meta-Analysis 1, we designed the meta-analysis
growth-mind-set intervention and students who did not. and report the results in accordance with the PRISMA
To investigate potential moderators, we tested the statement (Moher et al., 2009).
same three student-related factors as in Meta-Analysis
1: developmental stage, academic risk status, and SES Inclusion criteria, literature search, and coding. The
as well as control- and intervention-related method- criteria for including a study in Meta-Analysis 2 were as
ological factors. We examined the type of control follows:
562 Sisk et al.

•• A growth mind-set treatment, henceforth inter- using a method based on that of Cheung and Chan
vention, where the primary goal was to increase (2004, 2008). For 8 studies, assignment to condition
students’ belief that one or more human attributes occurred at the classroom or school level rather than
(e.g., intelligence) can improve with effort was at the student level. In these cases, the assumption of
administered directly to students. independence is violated, and calculations of variance
•• A control group (active, passive, or fixed-mind- are inappropriately small if not adjusted. Artificially
set condition) was included. small variances increase the chance of a Type 1 error,
•• A measure of academic achievement—course and in the case of meta-analyses, extend too much
exam (e.g., midterm exam), course grade, average weight to those effect sizes. In these cases, we adjusted
of course grades (e.g., GPA), or standardized test (increased) the variance associated with their effect
performance—was collected. sizes by taking into account the design effect (i.e.,
•• An effect size reflecting the difference between multiplying Kish’s, 1965, deff formula, using the typical
the mind-set intervention group and the control intraclass correlation for school effects, ρ = .10; Hox,
group on one or more measures of academic 1998, with the student-level variance) to find the oper-
achievement after the intervention was reported, ating variance. For a list of studies included in the
or enough information was provided to compute meta-analysis, see the Supplemental Material available
this effect size. online. For additional characteristics of Meta-Analysis
•• The methods and results were in English. 2, see Table 2.

Some studies administered combined interventions, Effect sizes. To measure the magnitude of the effective-
for example, a growth-mind-set intervention immedi- ness of the intervention, we used Cohen’s d as the mea-
ately followed by another intervention to a single group sure of effect size. Ideally, we would have estimated the
before measuring academic achievement. We do not difference in gain scores between the treatment and con-
include these effects because we cannot know the trol groups (e.g., Melby-Lervåg, Redick, & Hulme, 2016).
extent to which the mind-set content is contributing to However, only a third of the studies provided enough
the effect. Some studies reported results including stu- information to calculate this difference. Therefore, except
dents who were randomly assigned to the mind-set when a study reported a significant pretest difference, we
intervention condition, but did not receive the mind-set use the standardized mean difference posttreatment
intervention. We included effect sizes reflecting only scores, which could cause a bias in the effect sizes (see
the difference between students who received the e.g., Melby-Lervåg et al., 2016). Positive Cohen’s ds indi-
mind-set intervention and controls. We excluded two cated that the group receiving a growth-mind-set inter-
effect sizes, d = −0.87 and d = −0.65, from two studies vention performed higher on a measure of academic
conducted by Mendoza-Denton, Kahn, and Chan (2008). achievement than students in the control group.
Mendoza-Denton and colleagues (2008) designed these
studies to reverse any positive effects from a growth- Potential moderators
mind-set intervention (relative to a fixed-mind-set con- Student factors. As with Meta-Analysis 1, there were
dition) when participants received a stereotype lift. They three levels of developmental stage: children (primary
found that students in a fixed-mind-set condition per- school students), adolescents (middle school, junior
formed better on an academic achievement measure than high school, and high school students), and adults (e.g.,
students in a growth-mind-set condition when experi- postsecondary students). Studies that included students
encing stereotype lift. Presumably, this occurred because in multiple categories (e.g., elementary school students
fixed mind-sets reinforce the fixedness of stereotyped junior high school students in a single sample) were not
group differences, which ameliorated self-doubt, anxiety, included in this moderator analysis.
and other disruptive processes (Mendoza-Denton et al., As with Meta-Analysis 1, there were three levels of
2008; see also Walton & Cohen, 2003). These effect sizes academic risk status: high (at risk of failing; e.g., students
were excluded because their inclusion in the meta- who previously failed courses), moderate (facing a situ-
analysis could suppress an overall positive effect. ational challenge; e.g., transitioning to a new school, a
We identified 29 studies that met all the inclusion member of a stereotyped group under a stereotype threat
criteria. We coded each study and the measures col- manipulation), and low (no indicators that students were
lected in it for reference information, student charac- at risk). Each sample was categorized on the basis of the
teristics, methodological characteristics, and results (the majority (> 50%) of the students in the sample. If we
data file is available at osf.io/453ds). These studies could obtain separate effect sizes for each subsample in
included 38 independent samples, with 43 effect sizes a study based on risk level, we did so and entered those
and a total sample size of 57,155 students. As with effects as independent samples. For two studies contrib-
Meta-Analysis 1, we adjusted for dependent samples uting three effect sizes, we could not obtain separate
Mind-Set and Academic Achievement 563

Table 2. Descriptive Characteristics for Meta-Analysis 2 Table 2. (continued)

Number of Number of Number of Number of

effect sizes participants effect sizes participants
Study characteristic (k = 43) (N = 57,155) Study characteristic (k = 43) (N = 57,155)
Developmental stagea Outside regular classroom 38 55,098
Children 2 181 activities
Adolescents 27 48,991 Intervention-achievement
Adults 13 7,871 measure intervalc
Academic risk statusa Immediate (same session) 5 533
Low 17 3,801 Short interval 32 56,180
Moderate 18 8,664 Long interval 4 292
High 5 1,960 Manipulation checkd
Socioeconomic statusa Not included 15 44,484
Included 28 12,671
Low 7 577
Significant increase after
Not low 8 4,596
intervention
Control group
No 13 7,409
Active 26 11,365 Yes 15 5,262
Passive 11 45,267 Mind-set type
Fixed-mind-set condition 6 523 Intelligence 33 54,002
Intervention type Other attribute (e.g., math 10 3,153
Passive 13 1,355 ability)
Feedback 1 1,589 Academic achievement measure
Interactive 29 54,211 Course exam 3 628
Mode of intervention Course grade 4 2,083
Computerized training 16 11,581 Average of course grades 15 10,564
Reading material 8 1,441 (i.e., grade point average)
In-person trainingb 14 43,681 Standardized test 21 43,880
By teachers 7 43,141 Laboratory measures 12 889
By researchers 5 649 Publication status
By both teachers and 5 162 Published 25 6,180
researchers Unpublished 18 50,975
Other (e.g., business 2 181 a
For this characteristic, one or more effect sizes were excluded
partners) because this information was not provided in the study.
b
Both computerized and 5 452 For this characteristic, the subgroup effect sizes and sample sizes sum
in-person training to the total sample size for in-person training plus both computerized
and in-person training.
Intervention context c
For this characteristic, two effect sizes were excluded because the
Integrated into regular 5 2,057 study collected data on grades after a short and long interval but did
classroom activities not identify which of these was reported in the results.
d
Measures of pre- and postintervention mind-set were included in the
(continued) study.

effect sizes for each subsample but could obtain a sepa- reporting student-level SES were not included in the
rate effect size for an at-risk subsample (i.e., we obtained SES moderator analysis.
the full sample effect size and an at-risk subsample’s
effect size). In these cases, when examining academic Control and intervention method factors. There were
risk status, we report the results in two models: one with three levels of control group type: active control (i.e., pla-
the full-sample effect sizes and one where the full-sam- cebo control), passive control (e.g., no contact control),
ple effect sizes are replaced with the at-risk students’ and fixed-mind-set condition (i.e., students in the com-
effect sizes. As with Meta-Analysis 1, we did not code parison group were given a fixed-mind-set intervention).
ethnic minorities or women as at risk unless they were Students in active (placebo) control groups engaged in
under a stereotype threat manipulation. similar activities and amounts of contact with administra-
As with Meta-Analysis 1, there were two levels of tors but without the content of a hypothesized effective
SES: low SES (e.g., students qualified for reduced-price treatment. Active controls did not consist of other treat-
lunch) and not low SES (i.e., middle-class or higher). ments designed to be effective in improving academic
Each study was categorized on the basis of the majority achievement. When multiple control groups were used
(> 50%) of the students in the sample. Studies not in a study, we used the active control whenever pos-
564 Sisk et al.

sible because this meets a higher scientific standard. If tion was administered). When studies included measures
an active control was not available, we used the pas- of academic achievement at multiple time points follow-
sive control over a fixed-mind-set condition comparison ing the intervention, we used the longest interval avail-
because a fixed-mind-set condition is often theorized to able within the same academic term as the intervention.
lower academic achievement (thus, in these cases, it is For example, if a study examined students’ academic
unclear whether the growth mind-set improves academic achievement on a course exam a week after the inter-
achievement or whether the fixed mind-set lowers it). vention and again on the course final exam 2 months
Intervention type has three levels: passive (students after the intervention, we used the effect size for the final
read a document or watch a video on how human attri- exam. The longest interval within the same academic
butes are malleable), feedback (students are given feed- term was chosen when available to give the interven-
back on their performance in terms of growth mind-set), tion time to positively affect study habits and response
and interactive (e.g., participants read materials and to failure within the same academic context as the inter-
then write an essay about how intelligence can be devel- vention. However, it is also possible that longer intervals,
oped or participate in an in-class discussion). If passive regardless of whether the academic term changes, will
and feedback interventions are as effective as interactive increase the effect. Thus, when conducting the interven-
interventions, this suggests that effective interventions tion-academic achievement interval moderator analysis,
can be implemented with few resources and with a light we also analyzed the effect sizes from the longest interval
touch (see Yeager, Walton, & Cohen, 2013, for a discus- between the intervention and the measure of academic
sion of stealthily implementing interventions). achievement regardless of whether the measure of aca-
Intervention length was a continuous variable based demic achievement occurred in a subsequent term.
on the number of intervention sessions. If intervention As with Meta-Analysis 1, the measure of academic
effectiveness increases with the number of intervention achievement had four levels: course exam (e.g., mid-
sessions, then this suggests a positive dose-response rela- term exam), course grade (e.g., math grade), grade
tionship. In contrast, if intervention effectiveness decreases average (e.g., GPA), and standardized test (e.g., Iowa
with the number of intervention sessions, this could be Test of Basic Skills). We also examined this modera-
due to students perceiving the repetition as a message tor excluding laboratory measures of academic
that they need help, undermining the credibility of the achievement (e.g., administering practice problems
growth-mind-set intervention (see Yeager et al., 2013). from the GRE) that are not part of a student’s aca-
Mode of intervention had four levels: computerized demic record.
training (e.g., Brainology computer program), reading
mind-set materials only (e.g., reading how intelligence Meta-analytic procedure. The meta-analysis involved
can change with effort), in-person training (structured four steps. The first step was to obtain the standardized
discussion or lecture), or a combination of modes (i.e., mean difference (Cohen’s d) in academic achievement
computerized training and in-person training). Addi- between students who received a growth-mind-set inter-
tionally, for in-person training, we examined who vention and students in the control group, along with
administered the intervention. There were three levels their sampling error variances. The second step was to
to this moderator: researcher, teacher, or both. search for extreme values. There were two outliers—
Intervention context had two levels: integrated into effect sizes whose residuals had z scores of 3 or greater
regular classroom activities (e.g., teacher provides (ds = −0.9554 and 1.5053); we Winsorized these values to
mind-set feedback or fosters discussion of mind-set in z scores equaling 2.99 (ds = −0.9050 and 1.0960, respec-
class) or outside regular classroom activities. We also tively). The third step was to use random-effects meta-
examined whether studies included a manipulation analysis modeling, which assumes meaningful differences
check of the mind-set intervention. To be coded as across studies, to estimate the meta-analytic mean distri-
having a manipulation check, the study needed to bution of effects and heterogeneity in the effect sizes,
include a pre- and postintervention measure of mind- and then to test whether some of the heterogeneity was
set. Of the studies that included a manipulation check, predictable from moderator variables using mixed-effects
we examined how many were successful. meta-analysis modeling. The final step was to perform
publication bias analyses.
Academic-achievement-measure-related factors. There As with Meta-Analysis 1, we used the Comprehensive
were three levels of intervention-achievement measure Meta-Analysis (Version 2; Borenstein et al., 2005) soft-
interval: immediate (within the same session as a mind-set ware package to conduct the meta-analyses. We used
intervention), short interval (within 4 months [approximately Comprehensive Meta-Analysis, the R Project for Statisti-
a semester’s time] of the mind-set intervention), and long cal Computing (www.r-project.org), the p-curve web
interval (longer than 4 months after the mind-set interven- application (http://www.p-curve.com/app), and the
Mind-Set and Academic Achievement 565

p-uniform web application (https://rvanaert.shinyapps Student factors. The developmental stage of the stu-
.io/p-uniform) to conduct the publication bias analyses. dents was not a significant moderator, Q(1) < 0.01, p =
(See also the Supplemental Material.) .999. Only two effect sizes associated with children were
available. For one effect size, sample age information was
unavailable. These three effect sizes were not included
Results
in this analysis. Growth-mind-set intervention did not
The model consists of 43 effect sizes. Effect sizes are significantly improve academic achievement relative to
weighted by the inverse of the variance, which includes controls either for adolescents, d = 0.08, 95% CI = [−0.01,
both between-studies variance and within-study vari- 0.17], p = .090, or for adults, d = 0.08, 95% CI = [−0.02,
ances. Within-study variance accounts for the sample 0.17], p = .123.
sizes, such that smaller studies are given less weight, Academic at-risk status was not a significant modera-
while larger studies are given more weight in the model. tor, Q(2) = 0.67, p = .715. Sample descriptions were
The majority of effect-size–associated adjusted Ns are unavailable for three effect sizes, and thus these effect
≥ 90. The mean adjusted N associated with this model’s sizes were not included in this analysis. Growth-mind-set
effect sizes is 1,664. intervention did not significantly improve academic
Figure 4 shows that 37 of the 43 effect sizes (86%) achievement relative to controls for low-risk students,
are not significantly different from zero. One effect size d = 0.06, 95% CI = [−0.01, 0.12], p = .109; for moderately
is significantly different from zero but negative, indicat- at-risk students, d = 0.08, 95% CI = [−0.04, 0.19], p = .177;
ing that students receiving a growth-mind-set interven- or for highly at-risk students, d = 0.17, 95% CI = [−0.11,
tion had significantly worse academic achievement than 0.45], p = .231. In two studies (with 3 effect sizes),
students in the control conditions. The remaining 5 authors provided separate effect sizes for subsamples of
effect sizes (12%) are significantly different from zero students who were high risk. When replacing these three
and positive, indicating that students receiving a effect sizes and variances of the full samples (where the
growth-mind-set intervention had significantly greater majority were not high risk) with these subsamples, the
academic achievement than students in the control moderator remains nonsignificant, Q(2) = 2.18, p = .335.
groups. However, the high-risk group (8 effect sizes) then dem-
The meta-analytic average standardized mean differ- onstrated a borderline significant effect of growth-mind-
ence (i.e., the average of various population effects) in set intervention, d = 0.19, 95% CI = [0.02, 0.36], p = .031.
academic achievement between students receiving a Low-risk students and moderate-risk students did not
growth-mind-set intervention and students in control benefit from a growth-mind-set intervention, d = 0.05,
groups is d = 0.08, 95% CI = [0.02, 0.14], p = .010. When 95% CI = [−0.02, 0.12], p = .162, and d = 0.09, 95% CI =
the original outlying effect sizes were entered in lieu [−0.04, 0.23], p = .162, respectively.
of the Winsorized effect sizes, the overall effect did not SES was a significant moderator, Q(1) = 4.76, p =
differ (d = 0.08, 95% CI = [0.02, 0.15], p = .010), nor did .029. Student-level SES was not reported for 28 effect
the pattern of results for the moderator analyses, with sizes, and thus these effect sizes were not included in
the exception that the effect for unpublished studies this analysis. Growth-mind-set intervention did not
was no longer significant (see the Supplemental Mate- improve middle-class and upper-class students’ aca-
rial for complete results with outliers). demic achievement, d = 0.03, 95% CI = [−0.06, 0.11],
As illustrated by the I2 statistic, which specifies the p = .538. However, for those from low-SES households
percentage of the between-studies variability in effect (7 effect sizes), academic achievement was significantly
sizes that is due to heterogeneity rather than random higher for students who received growth-mind-set inter-
error, there was a medium amount of heterogeneity in ventions relative to controls, d = 0.34, 95% CI = [0.07,
the effect sizes, I2 = 43.15 (τ2 = .010), indicating that 0.62], p = .013.
the true effect of a given study could be somewhat
lower or higher than the meta-analytic average. We Control and intervention-related factors. Control-group
investigated the source of this heterogeneity through type was not a significant moderator, Q(2) = 2.96, p = .228.
the moderator analyses reported next. Academic achievement was similar between students who
received a growth-mind-set intervention and students who
Moderator analyses. Williams (2012) recommended at received a fixed-mind-set condition, d = 0.27, 95% CI =
least five cases per subgroup to perform a moderator [−0.05, 0.59], p = .100. There was also no effect when the
analysis in meta-analyses. We therefore did not include control group was passive, d = 0.02, 95% CI = [−0.05, 0.10],
any groups with fewer than five effect sizes when con- p = .522. A borderline significant difference was observed
ducting moderator analyses. See Table 2 for the number when the control group was an active control (i.e., pla-
of effect sizes per group. cebo control), d = 0.08, 95% CI = [0.01, 0.16], p = .034.
566 Sisk et al.

Wilson (2009) – S2
Good, Aronson, & Inzlicht (2003) – S1 M1
Yeager et al. (2014) – S3
Good et al. (2003) – S2 M1
Blackwell, Trzesniewski, & Dweck (2007)
Mendoza-Denton, Kahn, & Chan (2008) – S2
Good et al. (2003) – M2
Donohoe, Topping, & Hannah (2012)
Aronson, Fried, & Good (2002)
Yeager et al. (2014) – S1
Yeager, Lee, & Jamieson (2016)
Saunders (2013)
Rienzo, Wolfe, H., & Wilkinson (2015) – M2
Ehrlinger, Mitchum, & Dweck (2016)
Bagès, Verniers, & Martinot (2016)
Mendoza-Denton et al. (2008) – S1
Wilkins (2014) – M2
Yeager et al. (2014) – S2
Anderson, Lammers, Nunnley, & Davis (2016)
Bostwick (2015)
Yeager, Romero, et al. (2016) – S2
Fabert (2014)
Burnette et al. (n.d.)
Dommett, Devonshire, Sewter, & Greenfield (2013) – S5
Broda (2015)
Lin-Siegler, Ahn, Chen, Fang, & Luna-Lucero (2016)
Outes, Sanchez, & Vakis (2016) – M1
Holden, Moreau, Greene, & Conway (2016)
Outes et al. (2016) – M2
Dommett et al. (2013) – S2
Paunesku et al. (2015)
Rienzo, Wolfe, & Wilkinson (2015) – M1
Wilkins (2014) – M1
Yeager, Romero, et al. (2016) – S1
Dommett et al. (2013) – S3
Gauthreaux (2015) – M2
Wilson (2009) – S1
Zonnefeld (2015)
Sriram (2014)
Gauthreaux (2015) – M1
Dommett et al. (2013) – S6
Dommett et al. (2013) – S4
Dommett et al. (2013) – S1

–2.00 –1.00 0.00 1.00 2.00

Fig. 4. Standardized mean differences (Cohen’s ds) in academic achievement between students receiving a growth-mind-set intervention
and students in the comparison group. Cohen’s ds (squares) and 95% confidence intervals (error bars) are displayed for all effects entered
into Meta-Analysis 2. The size of the square represents the effect size’s meta-analytic weight. The diamond on the bottom row represents
the meta-analytically weighted mean Cohen’s d. For studies with multiple independent samples, the result for each sample (S1, S2, etc.)
is reported separately. Similarly, for studies with multiple measures, the result for each measure (M1, M2, etc.) is reported separately.
Multiple measures were adjusted for dependency. See the Supplemental Material available online for full details of all references.
Mind-Set and Academic Achievement 567

Intervention type was not a significant moderator, measures of mind-set to test whether the growth-mind-
Q(1) = 0.58, p = .447. Only one effect size used feed- set intervention effectively increased growth mind-set
back (weekly growth mind-set feedback with students’ (i.e., no manipulation checks). Interestingly, the effect
quiz grades) as the manipulation, and thus this effect of a growth-mind-set intervention was significant when
size was removed from this moderator analysis. The no manipulation check was administered, d = 0.18, 95%
effectiveness of a growth-mind-set intervention on aca- CI = [0.05, 0.31], p = .005, but not significant for studies
demic achievement was not significant when the inter- that employed a manipulation check, d = 0.04, 95%
vention was passive (e.g., reading about growth CI = [−0.03, 0.10], p = .249. The difference between these
mind-set without writing a reflection), d = 0.02, 95% two groups of studies was borderline significant, Q(1) =
CI = [−0.16, 0.19], p = .852, but demonstrated effective- 3.95, p = .047.
ness when the intervention was interactive (e.g., read- For the 28 effect sizes associated with studies that
ing about growth mind-set and then writing a reflection), did employ a manipulation check, almost half (46%)
d = 0.09, 95% CI = [0.02, 0.16], p = .011. failed to observe a significant difference between pre-
Intervention length was not a significant moderator, and postintervention measures of mind-set. The effec-
Q(1) = 0.12, b = −0.005, 95% CI = [−0.03, 0.02], p = .734. tiveness of the growth-mind-set intervention was
The number of sessions ranged from 1 to 10. Increasing borderline significant only when the manipulation check
the number of growth-mind-set-intervention sessions failed, d = 0.05, 95% CI = [0.001, 0.09], p = .044, indicat-
neither increased nor decreased the impact on aca- ing that students’ growth mind-sets had not changed
demic achievement. following the intervention. Growth-mind-set interven-
Mode of intervention was a significant moderator, Q(3) = tions were not effective for the studies where the manip-
9.33, p = .025. Growth-mind-set interventions were not ulation check succeeded, d = 0.02, 95% CI = [−0.11,
effective when administered via computer programs, 0.15], p = .771; Q(1) = 0.18, p = .672.
d = 0.03, 95% CI = [−0.03, 0.08], p = .409; in person, d =
0.06, 95% CI = [−0.12, 0.25], p = .517; or via a combination Factors related to academic achievement measures.
of modes, d = 0.27, 95% CI = [−0.04, 0.59], p = .092. The When using effect sizes associated with the greatest amount
intervention was effective when students read growth- of time between the intervention and measure of academic
mind-set materials, d = 0.20, 95% CI = [0.09, 0.30], p < achievement within the same semester, if available, the
.001. Post hoc follow-up tests were conducted to examine interval between the growth-mind-set intervention and
the source of this heterogeneity with an adjusted α of the measure of academic achievement was not a signifi-
.008 (.05/6). The follow-up tests revealed that mind-set cant moderator, Q(1) = 2.41, p = .121. Two effect-size
interventions administered via reading materials were intervals were ambiguous. Only four effect sizes mea-
significantly more effective than when administered via sured solely long-term (> 4 months) academic achieve-
computer programs, Q(1) = 7.75, p = .005. No other pair- ment following the interventions. These 6 effect sizes
wise comparisons were significant, all ps > .133. were excluded from this analysis. The effectiveness of
When interventions were administered in person growth-mind-set interventions was not significant regard-
(solely or as part of a combination), growth-mind-set less of whether academic achievement was measured
interventions remained ineffective regardless of whether within the same session, d = 0.35, 95% CI = [−0.03, 0.72],
the intervention was administered by a teacher, d = p = .070, or within 4 months of the intervention, d = 0.05,
−0.01, 95% CI = [−0.14, 0.12], p = .882; a researcher, 95% CI = [−0.01, 0.11], p = .126.
d = 0.34, 95% CI = [−0.14, 0.82], p = .167; or both, d = Nine effect sizes were associated with studies that
0.27, 95% CI = [−0.23, 0.77], p = .296; Q(2) = 2.80, p = measured academic achievement at two time points,
.246. Two effect sizes were associated with administra- once within the same semester as the intervention and
tors other than a teacher or researcher and were thus once following that semester. For seven of these effects,
removed from this analysis. the second measure was administered more than 4
The context in which the intervention was imple- months after the intervention. For two of these effects,
mented was not a significant moderator, Q(1) = 2.52, the second measure was administered following the
p = .112. Growth-mind-set interventions were not effec- semester of the intervention, but within 4 months of
tive when the intervention was integrated into regular the intervention. When replacing the seven short-term-
classroom activities, d = −0.14, 95% CI = [−0.43, 0.14], interval effect sizes with their long-term-interval coun-
p = .327. However, when the interventions were admin- terparts, replacing the two short-term effect sizes with
istered outside regular classroom activities, the effect their longest interval (though still short-term) counter-
was significant, d = 0.09, 95% CI = [0.03, 0.16], p = .003. parts, and adding in the four long-term-interval effect
Fifteen of the 43 effect sizes (35%) were associated sizes that were previously excluded because there were
with studies that did not report pre- and postintervention fewer than five, we found that interval was still not a
568 Sisk et al.

significant moderator, Q(2) = 3.92, p = .141. The two this simulation was 18.2%. Therefore, the results of the
effect sizes mentioned previously where the interval p-curve analyses are inconclusive. See the Supplemental
was ambiguous were not included in this analysis. Material for the results, additional figures, and an index
Short-term academic achievement (within 4 months of of the effect sizes that were entered into the analyses.
the intervention) remained nonsignificant, d = 0.05,
95% CI = [−0.002, 0.10], p = .057. Long-term academic Egger’s regression. During our search, we became
achievement (measured more than 4 months after the aware of multiple unpublished studies for which we
intervention) was also nonsignificant, d = 0.19, 95% could not access the methods or results. As one example,
CI = [−0.02, 0.39], p = .072. when we requested results for an unpublished interven-
The type of academic achievement measure was not tion discussed in a presentation, the researcher declined
a significant moderator, Q(1) = 0.03, p = .862. Only to provide this information on the grounds that replica-
three effect sizes were associated with a course exam tion attempts had failed. The researcher also declined to
grade, and only four effect sizes were associated with provide access to the results of the failed replications.
a course grade. We did not include these seven effect Despite being aware of missing studies, we found that
sizes in this moderator analysis. Growth-mind-set inter- the funnel plot was approximately symmetrical, suggest-
ventions were borderline significant when the measure ing that our meta-analysis was unaffected by missing
of academic achievement was GPA, d = 0.07, 95% CI = studies with weaker-than-average effect sizes, B0 = 0.38,
[0.002, 0.14], p = .045. Growth-mind-set interventions 95% CI = [−0.20, 0.97], t(41) = 1.33, p = .192.
were not significant when the measure of academic
achievement was performance on a standardized test,
Discussion
d = 0.09, 95% CI = [−0.07, 0.24], p = .276.
Twelve effect sizes were from standardized tests (or Some researchers have claimed that mind-set interven-
portions of standardized tests) administered by research- tions can “lead to large gains in student achievement”
ers in a laboratory setting (coded as standardized test). and have “striking effects on educational achievement”
Excluding these 12 effect sizes did not change the over- (Yeager & Walton, 2011, pp. 267 and 268, respectively).
all results, d = 0.06, 95% CI = [0.01, 0.11], p = .012, or Overall, our results do not support these claims. Mind-
the pattern of results for the academic achievement set interventions on academic achievement were non-
moderator. Measure of academic achievement remained significant for adolescents, typical students, and
a nonsignificant moderator, Q(1) = 0.07, p = .796, and students facing situational challenges (transitioning to
growth-mind-set interventions remained ineffective a new school, experiencing stereotype threat). How-
when the measure of academic achievement was per- ever, our results support claims that academically high-
formance on an actual standardized test, d = 0.09, 95% risk students and economically disadvantaged students
CI = [−0.05, 0.24], p = .213. may benefit from growth-mind-set interventions (see
Paunesku et al., 2015; Raizada & Kishiyama, 2010),
Publication bias analyses. We conducted the same although these results should be interpreted with cau-
three types of publication bias analyses as in Meta- tion because (a) few effect sizes contributed to these
Analysis 1. results, (b) high-risk students did not differ significantly
from non-high-risk students, and (c) relatively small
Moderator analysis. The median sample size associ- sample sizes contributed to the low-SES group.
ated with unpublished studies was 270 (compared with The results do not support the claim that mind-set
66 for published studies). A moderator analysis revealed interventions benefit both high- and low-achieving stu-
that the 18 effect sizes associated with unpublished stud- dents (e.g., see Mindset Scholars Network 4). Mind-set
ies (d = 0.05, 95% CI = [0.004, 0.10], p = .032) were not interventions are relatively low cost and take little time,
significantly different from the 25 effect sizes associated so there may be a net benefit for students’ academic
with published studies (d = 0.11, 95% CI = [0.002, 0.22], achievement. However, there may be a detriment rela-
p = .045), Q(1) = 1.02, p = .313. tive to fixed-mind-set conditions when students are
confident in their abilities (Mendoza-Denton et al.,
p-curve analysis. Only four statistically significant 2008). Regardless, those seeking more than modest
results were available to be included in the primary effects or effects for all students are unlikely to find
analysis (p-curve excludes unpublished results and them. To this end, policies and resources targeting all
nonsignificant results). We performed a simulation in students might not be prudent.
R (www.r-project.org; Simmons & Simonsohn, 2017) to Regarding methodological moderators, interactive inter-
assess the power of p-curve to detect right-skew for four ventions produced a significant effect in line with mind-set
studies, each with 16.5% power. The estimated power of theory (Yeager et al., 2013). However, other results were
the p-curve analysis to detect right-skew on the basis of confusing. For example, there was no significant
Mind-Set and Academic Achievement 569

difference between students in growth-mind-set versus manipulation checks to ensure that mind-set interven-
fixed-mind-set conditions or when the treatment group tions are influencing students’ mind-sets. If mind-set
was passive—the effect was significant only when com- manipulations are not demonstrating an influence on
pared with active controls. As another example, the students’ mind-sets (as was found in nearly half the
effect was significant for studies that did not report studies including manipulation checks), then the mecha-
manipulation checks while nonsignificant for studies nism affecting any observed change in achievement is
with manipulation checks. Further, of studies that either due to chance or due to mediating variables.
reported manipulation checks, almost half failed, sug- Additionally, while the results that supported mind-set
gesting that the interventions had no impact on students’ theory were not strong, it is possible that unmeasured
mind-sets. Most surprising, the effect was significant factors are suppressing effects or that imperfect control
when the manipulation checks failed but null when the of the intervention in the classroom buffers the effects
manipulation checks succeeded. This suggests that “suc- (Yeager & Walton, 2011). Alternatively, mind-set inter-
cessful” interventions may not be attributable to stu- ventions might need to be combined with other inter-
dents’ mind-sets. Manipulation checks are critical for ventions to increase effectiveness. From a theoretical
establishing causal inferences (Alferes, 2012). perspective, further investigations into potential media-
tors and moderators might yield important discoveries
General Discussion about the nature of human beliefs, the role of educa-
tional interventions, or both.
Mind-sets and their implications for academic achieve- However, from a practical perspective, resources
ment have received substantial attention from the media might be better allocated elsewhere than mind-set inter-
(e.g., PBS, Time, NPR; see Paul, 2013; Eisenberg, 2005; ventions. Across a range of treatment types, Hattie,
Smith, 2014, respectively), funding agencies, educators, Biggs, and Purdie (1996) found that the meta-analytic
and government institutions. For example, in 2013, the average effect size for a typical educational intervention
White House convened a special meeting entitled on academic performance is 0.57. All meta-analytic
“Excellence in Education: The Importance of Academic effects of mind-set interventions on academic perfor-
Mindsets.” Boaler (2013) summarized the impact as the mance were < 0.35, and most were null. The evidence
“mindset revolution that is reshaping education.” suggests that the “mindset revolution” might not be the
Part of the reshaping effort has been to make fund- best avenue to reshape our education system.
ing mind-set research a “national education priority”
(Rattan et al., 2015, p. 723) because mind-sets have
“profound effects” on school achievement (Dweck, Action Editor
2008, para. 2). Our meta-analyses do not support this D. Stephen Lindsay served as action editor for this article.
claim. Effect sizes were inconsistent across studies, but
most analyses yielded small (or null) effects. Overall, Author Contributions
the first meta-analysis demonstrated only a very weak
V. F. Sisk developed the study concept with input from J. L.
relationship between mind-sets and academic achieve-
Butler and B. N. Macnamara. V. F. Sisk and B. N. Macnamara
ment. Similarly, the second meta-analysis demonstrated developed the methodologies with input from J. L. Butler.
only a very small overall effect of mind-set interventions V. F. Sisk, A. P. Burgoyne, and J. Sun performed effect-size
on academic achievement. data collection with input from B. N. Macnamara and J. L.
However, not all mind-set research makes broad Butler. B. N. Macnamara and V. F. Sisk coded moderators and
claims. Some research focuses on specific tenets of the conducted calculations for Meta-Analysis 1. V. F. Sisk and
theory regarding how mind-sets affect individuals fac- A. P. Burgoyne coded moderators and conducted calculations
ing challenges, hypothesizing effects only for specific for Meta-Analysis 2. B. N. Macnamara performed the data
groups of students. Some subgroup results from the analyses. A. P. Burgoyne and B. N. Macnamara performed the
present meta-analyses supported these hypotheses, publication bias analyses. B. N. Macnamara drafted the intro-
such as the significant effects for academically high-risk duction, Method, Results, and Discussion sections of the manu-
script. B. N. Macnamara and A. P. Burgoyne drafted the results
students and low-SES students. Other subgroup results
for the publication bias analysis. V. F. Sisk, A. P. Burgoyne,
did not support these hypotheses, such as null results and J. L. Butler provided critical revisions. All authors approved
for students facing situational challenges and adoles- the final version of the manuscript for submission.
cents. Still other subgroup results suggest that standards
are needed for implementing intervention studies and
interpreting the results. Declaration of Conflicting Interests
Moving forward, researchers interested in mind-sets’ The author(s) declared that there were no conflicts of interest
effects on academic achievement should institute with respect to the authorship or the publication of this article.
570 Sisk et al.

Supplemental Material review of implicit theories and self-regulation. Psy­

chological Bulletin, 139, 655–701.
Additional supporting information can be found at http://
Cheung, S. F., & Chan, D. K.-S. (2004). Dependent effect
journals.sagepub.com/doi/suppl/10.1177/0956797617739704
sizes in meta-analysis: Incorporating the degree of inter-
dependence. Journal of Applied Psychology, 89, 780–791.
Open Practices Cheung, S. F., & Chan, D. K.-S. (2008). Dependent correla-
tions in meta-analysis: The case of heterogeneous depen-
dence. Educational and Psychological Measurement, 68,
760–777. doi:10.1177/0013164408315263
All data have been made publicly available via the Open Science
Claro, S., Paunesku, D., & Dweck, C. S. (2016). Growth mind-
Framework and can be accessed at osf.io/453ds. The complete
set tempers the effects of poverty on academic achieve-
Open Practices Disclosure for this article can be found at http://
ment. Proceedings of the National Academy of Sciences,
journals.sagepub.com/doi/suppl/10.1177/0956797617739704.
USA, 113, 8664–8668. doi:10.1073/pnas.1608207113
This article has received the badge for Open Data. More infor-
Dweck, C. S. (2000). Self-theories: Their role in motivation,
mation about the Open Practices badges can be found at http://
personality, and development. East Sussex, England:
www.psychologicalscience.org/publications/badges.
Psychology Press.
Dweck, C. S. (2006). Mindset: The new psychology of success.
Notes New York, NY: Random House.
1. See, for example, the PERTS press kit (www.perts.net/press_kit). Dweck, C. S. (2008, Winter). Brainology: Transforming stu-
2. See the U.S. Department of Education’s Press Releases page dents’ motivation to learn. Independent School. Retrieved
(www.ed.gov/news/press-releases/us-department-education- from https://www.nais.org/magazine/independent-
announces-first-ever-skills-success-grants-and-initiative-sup school/winter-2008/brainology/
port-learning-mindsets-and-skills). Dweck, C. S. (2012). Mindsets and human nature: Promoting
3. See the website of the Institute of Education Sciences (ies change in the Middle East, the schoolyard, the racial divide,
.ed.gov/funding/grantsearch/details.asp?ID=1728). and willpower. American Psychologist, 67, 614–622.
4. A relevant quote from the “Improving Student Outcomes and doi:10.1037/a0029783
Expanding Educational Opportunity” section on the Mindset Dweck, C. S. (2016). The remarkable reach of growth mind-
Scholars Network’s FAQ page regarding this claim is, “Learning sets. Scientific American Mind, 27, 36–41. doi:10.1038/
mindsets have been shown to be beneficial at every level: from scientificamericanmind0116-36
students struggling academically in middle school to under- Dweck, C. S., Chiu, C. Y., & Hong, Y. Y. (1995). Implicit
graduates at highly selective universities” (http://mindsetschol theories and their role in judgments and reactions: A word
arsnetwork.org/learning-mindsets/faq/). from two perspectives. Psychological Inquiry, 6, 267–285.
Egger, M., Smith, G. D., Schneider, M., & Minder, C. (1997).
Bias in meta-analysis detected by a simple, graphical test.
References British Medical Journal, 315, 629–634.
Alferes, V. R. (2012). Methods of randomization in experimen- Eisenberg, D. (2005, November 6). How to help them succeed.
tal design. Thousand Oaks, CA: Sage. Time. Retrieved from http://content.time.com/time/maga
Becker, G. (1986). Correcting the point-biserial correlation for zine/article/0,9171,1126743,00.html
attenuation owing to unequal sample size. The Journal of Hattie, J., Biggs, J., & Purdie, N. (1996). Effects of learning
Experimental Education, 55, 5–8. skills interventions on student learning: A meta-analysis.
Bishop, D. V. M., & Thompson, P. A. (2016). Problems in Review of Educational Research, 66, 99–136.
using p-curve analysis and text-mining to detect rate of Hox, J. (1998). Multilevel modeling: When and why. In I.
p-hacking and evidential value. PeerJ, 4, Article e1715. Balderjahn, R. Mathar, & M. Schader (Eds.), Classification,
doi:10.7717/peerj.1715 data analysis, and data highways (pp. 147–154). Berlin,
Blackwell, L. S., Trzesniewski, K. H., & Dweck, C. S. (2007). Germany: Springer.
Implicit theories of intelligence predict achievement across Hunter, J. E., & Schmidt, F. L. (1990). Dichotomization of
an adolescent transition: A longitudinal study and an inter- continuous variables: The implications for meta-analysis.
vention. Child Development, 78, 246–263. doi:10.1111/ Journal of Applied Psychology, 75, 334–339. doi:10.1037/
j.1467-8624.2007.00995.x 0021-9010.75.3.334
Boaler, J. (2013). Ability and mathematics: The mindset revo- Kish, L. (1965). Survey sampling. New York, NY: Wiley.
lution that is reshaping education. Forum, 55, 143–152. Melby-Lervåg, M., Redick, T. S., & Hulme, C. (2016). Working
Borenstein, M., Hedges, L., Higgins, J., & Rothstein, H. (2005). memory training does not improve performance on mea-
Comprehensive Meta-Analysis (version 2) [Computer soft- sures of intelligence or other measures of “far transfer”:
ware]. Englewood, NJ: Biostat. Evidence from a meta-analytic review. Perspectives on
Burnette, J. L., & Finkel, E. J. (2012). Buffering against weight Psychological Science, 11, 512–534. doi:10.1177/1745691
gain following dieting setbacks: An implicit theory inter- 616635612
vention. Journal of Experimental Social Psychology, 48, Mendoza-Denton, R., Kahn, K., & Chan, W. (2008). Can
721–725. doi:10.1016/j.jesp.2011.12.020 fixed views of ability boost performance in the context
Burnette, J. L., O’Boyle, E. H., VanEpps, E. M., Pollack, J. M., of favorable stereotypes? Journal of Experimental Social
& Finkel, E. J. (2013). Mind-sets matter: A meta-analytic Psychology, 44, 1187–1193. doi:10.1016/j.jesp.2008.03.005
Mind-Set and Academic Achievement 571

Moher, D., Liberati, A., Tetzlaff, J., Altman, D. G., & The Simonsohn, U., Nelson, L. D., & Simmons, J. P. (2014b).
PRISMA Group. (2009). Preferred reporting items for sys- p-curve and effect size: Correcting for publication bias
tematic reviews and meta-analyses: The PRISMA state- using only significant results. Perspectives on Psychological
ment. PLOS Medicine, 6(7), Article e1000097. doi:10 Science, 9, 666–681.
.1371/journal.pmed1000097 Smith, T. (2014, March 17). Does teaching kids to get ‘gritty’
Paul, A. M. (2013, June 11). The science of smart: Eight ways help them to get ahead? NPR. Retrieved from https://www
of looking at intelligence. PBS. Retrieved from http:// .npr.org/sections/ed/2014/03/17/290089998/does-teach-
www.pbs.org/wgbh/nova/blogs/secretlife/blogposts/the- ing-kids-to-get-gritty-help-them-get-ahead
science-of-smart-eight-ways-of-looking-at-intelligence/ Sterne, J. A. C., Sutton, A. J., Ioannidis, J. P. A., Terrin,
Paunesku, D., Walton, G. M., Romero, C., Smith, E. N., N., Jones, D. R., Lau, J., . . . Higgins, J. P. T. (2011).
Yeager, D. S., & Dweck, C. S. (2015). Mind-set inter- Recommendations for examining and interpreting fun-
ventions are a scalable treatment for academic under- nel plot asymmetry in meta-analyses of randomised con-
achievement. Psychological Science, 26, 784–793. trolled trials. British Medical Journal, 343, Article d4002.
doi:10.1177/0956797615571017 doi:10.1136/bmj.d4002
Protzko, J. (2015). The environment in raising early intelli- van Aert, R. C. M., Wicherts, J. M., & van Assen, M. A. L. M.
gence: A meta-analysis of the fadeout effect. Intelligence, (2016). Conducting meta-analyses based on p values:
53, 202–210. Reservations and recommendations for applying p-uniform
Raizada, R. D., & Kishiyama, M. M. (2010). Effects of socio- and p-curve. Perspectives on Psychological Science, 11,
economic status on brain development, and how cogni- 713–729.
tive neuroscience may contribute to levelling the playing Walton, G. M., & Cohen, G. L. (2003). Stereotype lift. Journal of
field. Frontiers in Human Neuroscience, 4, Article 3. Experimental Social Psychology, 39, 456–467. doi:10.1016/
doi:10.3389/neuro.09.003.2010 S0022-1031(03)00019-2
Rattan, A., Savani, K., Chugh, D., & Dweck, C. S. (2015). Williams, R. (2012, May). Moderator analyses: Categorical
Leveraging mindsets to promote academic achievement: models and meta-regression. Paper presented at the
Policy recommendations. Perspectives on Psychological annual Campbell Collaboration Colloquium, Copenhagen,
Science, 10, 721–726. doi:10.1177/1745691615599383 Denmark.
Romero, C., Master, A., Paunesku, D., Dweck, C. S., & Gross, Yeager, D. S., & Dweck, C. S. (2012). Mindsets that promote
J. J. (2014). Academic and emotional functioning in mid- resilience: When students believe that personal charac-
dle school: The role of implicit theories. Emotion, 14, teristics can be developed. Educational Psychologist, 47,
227–234. doi:10.1037/a0035490 302–314. doi:10.1080/00461520.2012.722805
Rosenthal, R. (1979). The file drawer problem and tolerance Yeager, D. S., & Walton, G. M. (2011). Social-psychological
for null results. Psychological Bulletin, 86, 638–641. interventions in education: They’re not magic. Review
Simmons, J. P., & Simonsohn, U. (2017). Power posing: of Educational Research, 81, 267–301. doi:10.3102/003
P-curving the evidence. Psychological Science, 28, 687–693. 4654311405999
Simonsohn, U., Nelson, L. D., & Simmons, J. P. (2014a). Yeager, D. S., Walton, G., & Cohen, G. L. (2013). Addressing
P-curve: A key to the file-drawer. Journal of Experimental achievement gaps with psychological interventions. Phi
Psychology: General, 143, 534–547. Delta Kappan, 94(5), 62–65.