Does Growth Mindset Work? The Largest Trial Found 0.10 of a Grade Point, and the Bias-Corrected Meta-Analysis Found Nothing

Two things are true about growth mindset and they are usually reported separately, which is why the question feels unresolved.

The best-designed large trial found a small, real effect on a specific group of students under specific conditions. The most rigorous meta-analysis found that once you account for study quality and bias, the average effect across the literature is not distinguishable from zero.

Both of those results are worth having, and neither supports what schools were told.

The observational relationship is weak

Start with the correlation, because the intervention rests on it: do students who believe intelligence is malleable actually do better?

Sisk, Burgoyne, Sun, Butler and Macnamara (2018), in Psychological Science 29(4), 549-571, ran two meta-analyses. The first pooled 273 effect sizes across 129 studies, N = 365,915, on the association between mindset and academic achievement.

The pooled correlation: r̄ = .10, 95% CI [.08, .13].

That is a real association and a small one. A correlation of .10 means mindset accounts for about one per cent of the variance in achievement.

The interventions are weaker still, on average

Their second meta-analysis covered the interventions: 43 effect sizes from 29 studies, N = 57,155.

Pooled effect: d̄ = 0.08, 95% CI [0.02, 0.14].

But the subgroup results are where the theory earns something back, and the authors say so:

Group	Intervention effect
All students	d̄ = 0.08 [0.02, 0.14]
Low socioeconomic status	d̄ = 0.34 [0.07, 0.62]
High academic risk	d̄ = 0.19 [0.02, 0.36]

Their conclusion is exactly this shape: “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.”

A universal school programme delivering d = 0.08 is a poor use of a school’s time. A targeted intervention delivering d = 0.34 to disadvantaged students is a different proposition.

The largest trial, and it was delivered online

Yeager and 24 co-authors (2019), in Nature 573, 364-369, ran the National Study of Learning Mindsets: 12,490 ninth-grade students across 65 schools in a nationally representative US sample, with 6,320 in the lower-achieving target group.

The intervention was two self-administered online sessions of about 25 minutes each, delivered roughly 20 days apart during school hours. Under an hour in total, at essentially zero marginal cost.

The result for lower-achieving students: +0.10 grade points in core classes, 95% CI [0.04, 0.16], a standardised effect of 0.11, p = 0.001. Enrolment in advanced mathematics also rose overall.

Two features make this the strongest study in the field. It was pre-registered, with independent data collection and processing and a blinded Bayesian corroboration. And it looked for the conditions under which the effect appears rather than reporting an average: the intervention worked “when the behavioural norm that surrounded students was supportive of the growth mindset belief system”, chiefly in medium- and low-achieving schools.

That is a considerably more interesting finding than “growth mindset works”. It says a one-hour online message shifts grades slightly, for students who are struggling, in schools where their peers will not punish them for trying.

The bias correction

Macnamara and Burgoyne (2022), in Psychological Bulletin, conducted the most methodologically aggressive review to date: 63 studies, 79 independent samples, 97,672 participants.

Subset	Effect
All studies	d̄ = 0.05 [0.02, 0.09]
After correcting for publication bias	not significant
Studies that actually changed mindset (13 studies, N = 18,355)	d̄ = 0.04 [-0.01, 0.10], n.s.
Highest-quality evidence (6 studies, N = 13,571)	d̄ = 0.02 [-0.06, 0.10], n.s.

The third row is the sharpest. If you restrict to studies that verified their intervention actually shifted students’ mindsets - the theory’s own precondition - the effect on achievement disappears.

Their conclusion is unusually direct for a meta-analysis:

“Apparent effects of growth mindset interventions on academic achievement are likely attributable to inadequate study design, reporting flaws, and bias.”

And they report a finding that belongs in any assessment of this literature: researchers with financial incentives published significantly larger effects than those without. Growth mindset is a substantial commercial category - training, curricula, consultancy - and the studies produced by people with a stake in it found bigger effects.

Holding all of it at once

The temptation is to pick a side. The honest reading takes all four results.

The correlation is real and small. Untargeted interventions do very little. A large, pre-registered, well-controlled trial found a small effect for lower-achieving students in supportive school contexts. And a bias-corrected meta-analysis of the whole literature finds the average effect indistinguishable from zero, with higher-quality studies producing smaller effects - the standard signature of a literature whose apparent effect is an artefact of method.

Those can all be true. A small real effect in a specific population can coexist with an average of zero across a body of work containing many underpowered studies with flexible analyses, some produced by people selling the programme.

What is not supportable is the version schools were sold: that teaching children intelligence is malleable produces broad achievement gains. Sisk and colleagues could not find that. Macnamara and Burgoyne found the appearance of it dissolving under quality controls. Yeager and colleagues, in the field’s best trial, found a tenth of a grade point in one subgroup under one condition.

There is a reason this belongs on a site about digital pressure, and it is the delivery. The largest and best test of this idea was an hour of online material, which is why it could reach 12,490 students in 65 schools. That is the model for the entire category of scalable, low-cost, software-delivered educational interventions - and the honest result, from the most careful version of it, is a tenth of a grade point for the students who were struggling. Which is not nothing, and is a long way from a transformation. The same shape appears when hardware was distributed at scale and when a single tutoring figure was used to justify a product category.

Limitations and disputes

• This is a live and heated dispute. Yeager and Dweck have published responses to the meta-analytic critiques, and Macnamara is an author on both the 2018 and 2022 critical papers. Readers should know the critical literature is substantially one research group’s programme.
• Small effects at population scale are not worthless. A d of 0.11 delivered for under an hour at near-zero cost has a favourable cost-benefit profile that a raw effect size does not capture, and the 2019 authors argue this explicitly.
• The subgroup findings in Sisk et al. have wide confidence intervals - the low-SES estimate runs from 0.07 to 0.62.
• “Financial incentives” is the meta-analysts’ own coding, not an allegation about any individual study.
• We have not assessed growth mindset outcomes other than academic achievement, such as persistence, wellbeing or course selection, beyond the advanced-mathematics enrolment finding.
• Nothing here is a claim about whether intelligence is malleable. That is a separate question. This is about whether teaching students that it is raises their grades.
• This is not advice to schools about whether to run such programmes.