Early View
ARTICLE
Open Access

Explore your brain: A randomized controlled trial into the effectiveness of a growth mindset intervention with psychosocial and psychophysiological components

Tieme W. P. Janssen

Corresponding Author

Tieme W. P. Janssen

Department of Clinical, Neuro- & Developmental Psychology, Vrije Universiteit Amsterdam, Amsterdam, The Netherlands

Correspondence

Tieme W. P. Janssen, Department of Clinical, Neuro- & Developmental Psychology, Faculty of Behavioural and Movement Sciences, Vrije Universiteit Amsterdam, Van der Boechorststraat 7, 1081 BT Amsterdam, The Netherlands.

Email: [email protected]

Contribution: Conceptualization, Data curation, Formal analysis, ​Investigation, Methodology, Project administration, Supervision, Validation, Visualization, Writing - original draft, Writing - review & editing

Search for more papers by this author
Nienke van Atteveldt

Nienke van Atteveldt

Department of Clinical, Neuro- & Developmental Psychology, Vrije Universiteit Amsterdam, Amsterdam, The Netherlands

Contribution: Conceptualization, Funding acquisition, Methodology, Project administration, Supervision, Writing - review & editing

Search for more papers by this author
First published: 11 December 2022
Citations: 2

Abstract

Background

Although past research demonstrated growth mindset interventions to improve school outcomes, effects were small. This may be due to the theoretical nature of psychosocial techniques (e.g., reading about brain plasticity), which may not be optimally convincing for students.

Aims

To address this issue and improve effectiveness, we developed a growth mindset intervention, which combined psychosocial and psychophysiological components. The latter adds a convincing experience of influencing one's own brain activity, using mobile electroencephalography (EEG) neurofeedback, emphasizing the controllable and malleable nature of one's brain.

Sample

In this randomized controlled trial (RCT), twenty high-school classes (N = 439) were randomized to either the active control condition (no mindset messaging) or our newly developed growth mindset intervention condition (4 × 50 min).

Methods

School outcomes (pre, post, 1-year follow-up) were analysed with Linear Mixed Models (LMM: variable-oriented) and Latent Transition Analysis (LTA: person-oriented).

Results

LMM: students in the growth mindset intervention reported increased growth mindset directly after the intervention (post, d = .38) and at 1-year follow-up (d = .25) and demonstrated a protective effect against deterioration of math grades at 1-year follow-up (d = .36), compared to controls. LTA: we identified three mindset profiles (Fixed, Growth competitive, Growth non-competitive), with more frequent transitions from fixed to one of the growth mindset profiles at 1-year follow-up for students in the growth mindset intervention compared to controls (OR 2.58–2.68).

Conclusions

Compared to previous studies, we found relatively large effects of our intervention on growth mindset and math grades, which may be attributable to synergetic effects of psychosocial and psychophysiological (neurofeedback) components. The person-oriented approach demonstrated more holistic effects, involving multiple motivational constructs.

BACKGROUND

Since the first conception of mindset theory (Dweck, 2000), decades of research and increasing public interest testify to the huge impact of this research field on education. One reason for this success, besides the theoretical underpinnings, is the potential to positively influence mindset and real-world outcomes with interventions. Recently, a lively discussion has started about the efficacy of such growth mindset interventions, for whom it may be more effective, and how to increase effectiveness (Miller, 2019). Modest effect sizes may partly be due to the theoretical nature of psychosocial techniques used in most growth mindset interventions (e.g., reading about brain plasticity), which may not be optimally convincing for students. Here we report the results of a randomized controlled trial (RCT) into a newly developed growth mindset intervention ‘Explore your Brain’, where we aimed to address this issue and increase effectiveness by combining psychosocial and psychophysiological techniques. The latter adds a convincing experience of influencing one's own brain activity, using mobile electroencephalography (EEG) neurofeedback, emphasizing the controllable and malleable nature of one's brain activity.

According to mindset theory, students vary in their implicit beliefs about the malleability of personal attributes, like intelligence or math ability. On a continuum, some believe that intelligence and abilities cannot change much (entity beliefs or fixed mindset), while others believe that they can improve (incremental beliefs or growth mindset). Central to mindset theory is that these beliefs have an organizing function, bringing together goals, beliefs and behaviours in a meaning system (Dweck & Yeager, 2019), which affect self-regulation during learning and ultimately academic achievement and motivation. Overall, people endorsing a growth mindset are characterized by adaptive self-regulation skills, such as setting learning goals instead of performance goals, adopting mastery-oriented strategies instead of helpless-oriented strategies in the face of setbacks, focusing on future expectations of success rather than negative emotions (Burnette et al., 2013) and having positive effort beliefs (Miele et al., 2013; Miele & Molden, 2010). These characteristics are conducive for experiencing a larger sense of control over one's learning process.

The first RCT of a growth mindset intervention was conducted almost two decades ago in the context of stereotype threat in African American college students (Aronson et al., 2002); since then, numerous RCT's followed, widely varying in intervention and study design. For instance, while some interventions lasted only 25 min (Yeager et al., 2014), others lasted up to 600 min (Chao et al., 2017). Likewise, RCT's differed in context (individual versus group), delivery (teacher, researcher or online) and control group (waiting list or active control).

Despite these differences, most interventions include one or two of the following psychosocial components: (1) teaching about brain plasticity and (2) explicitly or implicitly teaching about mindset. The rationale behind brain plasticity classes, is that by explaining how the brain structurally changes with learning, an argument is presented for the malleability of its associated functions (e.g., intelligence, math ability), which supports the notion of a growth mindset. For example, students would learn about how brain areas (e.g., hippocampus) important for specific functions (e.g., visuospatial memory) increase in grey matter volume in taxi drivers, who frequently use visuospatial memory to navigate the city (Maguire et al., 2006). By emphasizing that the amount of effort invested in effective strategies is key to learning and its associated structural brain changes, students realize that brain plasticity and learning are at least to a certain extent controllable. With respect to classes about mindset, several strategies have been used, but most involve techniques borrowed from cognitive-behavioural therapy (CBT) aiming to replace less adaptive cognitions (e.g., fixed mindset) with more adaptive cognitions (e.g., growth mindset).

While all growth mindset interventions share the same proximal goal of stimulating a growth mindset, the distal goals may vary. Most studies have focused on academic achievement, while motivation, academic career choices and well-being have received relatively less attention. Two recent meta-analyses quantified the effects of growth mindset interventions on academic achievement and concluded that overall effect sizes were small, with larger effects for academically at-risk students and economically disadvantaged students (Sarrasin et al., 2018; Sisk et al., 2018). The type of academic outcome measure (e.g., course grade, global point average [GPA], standardized or lab tests) was not a significant moderator for overall intervention effects in Sisk et al. (2018). However, in Sarrasin et al. (2018) there was some tentative evidence for larger intervention effects on math grades in at-risk students relative to non-at-risk students, compared to other achievement domains. This is in line with popular beliefs that STEM (science, technology, engineering and mathematics) subjects like math, are linked to ‘raw’ or ‘innate’ and thus fixed abilities (Leslie et al., 2015), and may therefore be especially sensitive for the effects of growth mindset interventions.

Other relevant outcomes involve motivation and enjoyment in learning, academic career choices and well-being in school. In the same meta-analysis, Sarrasin et al. (2018) looked into motivational outcomes, including constructs that are central to mindset theory (achievement goals, effort beliefs, response to failure) and enjoyment. In line with the academic achievement results, they found a small positive effect size across these motivational outcomes for the growth mindset interventions, benefiting at-risk more than not-at-risk students. With regard to academic career choices, a brief online growth mindset intervention increased advanced mathematics course enrolment, especially in the highest-achieving 25% of schools (Yeager et al., 2019). Lastly, in the past few years there is increasing interest in how mindset is related to psychological well-being. Burnette et al. (2020) conducted a meta-analysis in the context of mental health problems and concluded that a growth mindset relates negatively to distress, and positively to treatment value and active coping. A subsample of 8 studies manipulated mindsets and examined distress longitudinally, demonstrating more promising effect sizes than for academic achievement (as reported in Sisk et al., 2018). More relevant in the school context, a growth mindset seems protective against school burnout symptoms in a recent study (Kim, 2020).

A crucial backdrop for evaluating and developing growth mindset interventions, is the recent discussion about what constitutes a meaningful effect size (Miller, 2019). In their meta-analysis, Sisk et al. (2018) interpreted the effect size of mindset interventions on academic achievement as not meaningful (overall: d = .08; high-risk students: d = .19), partly based on a comparison with the meta-analytic average effect size (d = .57) for typical educational interventions (Hattie et al., 1996). Yeager et al. (2019), however, argue for the use of other benchmarks, not from laboratory studies, but from the highest-quality field research with objective educational outcomes (Hill et al., 2008; Kraft, 2018). With this benchmark (d = .06), the meta-analytic averages and recent large-scale studies seem more favourable, especially for high-risk subgroups. An effect size of d = .20 could be considered large in this context (Yeager et al., 2019). Another factor of consideration, is the cost-effectiveness of interventions. An effect size of d = .11 for at-risk students with a low-cost intervention (Yeager et al., 2019) lasting less than an hour, requiring no teacher training, is impressive in that respect.

In addition to lowering the costs, another approach is to increase the effectiveness of mindset interventions. We identified a potential way to do so, based on our observation that previous interventions were largely theoretical, neglecting more experiential aspects related to growth mindset. With experiential we mean learning from direct personal experiences (Kolb, 2014), which can act as convincing empirical proof of growth mindset and lead to further internalization. We, therefore, combined psychosocial (theoretical) and psychophysiological (experiential) intervention components with the aim to increase effectiveness. For the psychosocial component, we replicated previous studies, by discussing theoretical topics such as brain plasticity and mindset. For the psychophysiological component, we utilized mobile electroencephalography (EEG) neurofeedback (NFB). With NFB, participants receive real-time feedback of their current brain activity, in the form of visual or acoustic signals and learn to self-regulate (‘control’) this activity using operant conditioning principles. For example, EEG indices of attentional states (e.g., theta/beta ratio) may be represented by the height of a plane flying across the screen (Janssen et al., 2020). NFB is used as a therapeutic tool (e.g., ADHD treatment; Cortese et al., 2016), but also for cognitive enhancement in neurotypical populations and as experimental method (Enriquez-Geppert et al., 2017). More relevant in the context of growth mindset interventions, NFB has been suggested to increase a sense of control and self-efficacy (Linden, 2014), which was recently confirmed (Harris et al., 2021).

In the current study, we aimed to provide a convincing experience of being in control over one's own brain activity using NFB. By emphasizing and experiencing the controllable and malleable nature of the brain, this experience was expected to help with internalizing the more theoretical growth mindset and brain plasticity messages that were discussed in the psychosocial part of our intervention. The importance of self-reference was also pointed out by De Castella and Byrne (2015), who demonstrated that a general belief that abilities can be improved does not necessarily entail believing in the incremental nature of one's own abilities.

We aimed to assess the effectiveness of the intervention by using both a more commonly used variable-oriented approach, and a more novel person-oriented approach. Variable-oriented approaches focus on examining relations among variables, while person-oriented approaches, such as Latent Profile Analysis (LPA), identify latent subgroups of students based on their similarities on a set of variables (indicators). LPA is a data-driven statistical method and is especially useful when there are signs of heterogeneity and inconsistencies in an area of research, which can arise because theorized associations do not hold for the entire population. This may be the case for mindset-based meaning systems (Dweck & Yeager, 2019) that involve a set of other motivational constructs, including effort beliefs and goal orientation. Theorized associations between growth mindset and mastery goals do not always hold, as some students hold mastery and performance goals simultaneously (Molden & Dweck, 2000). Person-oriented studies are well suited to expose such naturally occurring combinations at the individual level, which was recently demonstrated in an LPA study (Yu & McLellan, 2020). Indeed, a subgroup of growth mindset students endorsed performance goals alongside mastery goals (Growth competitive profile), while a subgroup of fixed mindset students did not endorse performance goals (Disengaged profile), providing important nuances to existing theorizing. These new insights lead to a new question: do interventions work better for some mindset profiles than others? This question can be investigated with the longitudinal variant of LPA, which is Latent Transition Analysis (LTA), allowing for a more holistic analysis of mindset intervention effects.

Here we report the results of an RCT with 439 Dutch high-school students, comprising our newly developed growth mindset intervention (4 × 50 min) compared to an active control condition (4 × 50 min). Compared to controls, we hypothesized that students receiving the growth mindset intervention would report larger increases in growth mindset from pre-intervention to direct-post-intervention, and to 1-year follow-up, and show improved grades (GPA, math) and reduced school burnout symptoms (Salmela-Aro et al., 2009) from pre-intervention to 1-year follow-up. We expected relatively larger effect sizes compared to the meta-analytic average (d = .08; Sisk et al., 2018) and the less intensive large-scale interventions (d = .11; Yeager et al., 2019), due to combining psychosocial and psychophysiological components. Moreover, we hypothesized larger effects for students from lower socio-economic status (Sisk et al., 2018). In addition, changes in mindset profiles (Yu & McLellan, 2020) were explored, based on LTA, taking into account the interrelatedness of constructs that play a role in mindset theory (effort beliefs, goal orientation, academic motivation) and naturally occurring combinations between those constructs at the individual level. Although the LTA analysis was more exploratory, we expected transitions from fixed mindset (e.g., disengaged) to growth mindset (e.g., growth competitive) profiles to occur more often in students after receiving the growth mindset intervention compared to controls.

METHODS

Trial design

The reporting of this study follows the CONSORT-SPI 2018 guidelines for social and psychological interventions with cluster extension (Grant et al., 2018). This study is a parallel cluster randomized controlled trial (CRT), stratified for school and educational track in the Dutch secondary school system (ranging from vocational to pre-university). Two schools participated, which both included a range of educational tracks. With an allocation ratio of 1/1, twenty 7th grade classes were randomized to either the growth mindset intervention condition or control condition, see Figure 1 for the consort flow diagram, and Table 1 for an overview of both conditions. The main reason to randomize at the class level, was to avoid contamination, which can dilute the observed differences between conditions and can affect the reliability and validity of the study (Torgerson, 2001). Outcomes were evaluated at the individual level, as the goal of the intervention was to influence mindset at the student level, not at the class level. Linear mixed models were used to account for clustering at the class level, see Analytical Methods. Outcomes were evaluated before (T0), directly after as manipulation check (T1), 1 year later (T2) and 2 years later (T3), with the aim to assess the superiority of the growth mindset intervention compared to the control condition. We disregarded T3, due to COVID-19 complications, which resulted in higher drop-out.

Details are in the caption following the image
Consort flow diagram. Note: The reporting of this study follows the CONSORT-SPI 2018 guidelines for social and psychological interventions with cluster extension (Grant et al., 2018).
TABLE 1. Overview of lesson content in control and growth mindset intervention conditions
Lesson Control condition Growth mindset intervention condition
1 Brain anatomy
  • Introduction Explore your Brain
  • Mirror drawing (1×, without learning curve)
  • Brain anatomy & function (link to brain areas involved in mirror drawing)
  • Make a brain hat
  • Quiz neurological cases: which brain areas were involved?
  • Teaser next lesson (illusions)
Brain plasticity
  • Introduction Explore your Brain
  • Mirror drawing (3× with learning curve)
  • Brain plasticity (link to learning during mirror drawing)
  • Juggling and brain plasticity
  • Connecting mirror drawing, brain plasticity, juggling and school
  • Teaser next lesson (growth mindset)
2 Brain illusions
  • The monkey business illusion (video): selective attention
  • How do these illusions work in the brain?
  • Illusions in daily life
  • McGurk effect (video)
  • Puppeteer (video)
  • What can we learn about the brain?
  • Teaser next lesson (neuroimaging)
Growth mindset
  • Calvin & Hobbes about intelligence and effort (cartoon)
  • Explanation mindset (video)
  • Famous failures (video)
  • Connecting brain plasticity and mindset
  • ABC exercises: growth and fixed mindset reactions in school
  • Teaser next lesson (neurofeedback)
3

Brain imaging

  • Introduction EEG and lesson
  • Mobile EEG application (focus on methods):

    • Real-time view raw EEG
    • Baseline recording
    • Real-time view frequency bands
    • Pro's and con's EEG
    • Real-time view attention index
    • Pro's and con's MRI
    • Choice: info about fNIRS, MEG or ECoG
    • Real-time view meditation index
    • Selfie (photo + EEG + Explore your brain!)

Neurofeedback

  • Introduction EEG + NFB and lesson
  • Mobile EEG NFB application (focus on own influence):

    • Real-time view raw EEG
    • Baseline recording
    • Real-time view attention index
    • NFB: attention or relaxing (choice)
    • NFB: attention (line exercise)
    • NFB: attention (neurons exercise)
    • NFB: attention (reading exercise)
    • NFB choice: line, neurons or reading
    • Selfie (photo + EEG + I am in control!)

4 Brain myths and opportunities
  • Spinning dancer meme: left- versus right-brained people neuromyth
  • Neuromyths in the media
  • Neuromyths quiz
  • Limitless movie trailer: 10% neuromyth
  • Tips for busting neuromyths
  • Neuro opportunities: possible future uses
  • Message in a bottle
SMART goals
  • Connecting lessons 1, 2 and 3
  • Discussion NFB lesson in relation to mindset
  • How to benefit from all 3 lessons at school?
  • Setting a personal goal (SMART)
  • Identifying potential obstacles, thinking of solutions and role of mindset
  • Take-home messages
  • Message in a bottle
  • Note: See archived data for all intervention materials translated in English (DataverseNL).

This trial was preregistered at the Dutch Trial Register (Trial NL7562), approved by the local ethics committee (VCWE-S-18-00149) and conducted in accordance with the Declaration of Helsinki. Harms or unintended effects were monitored, but none were reported by teachers. For our Transparency and Openness statement, see Appendix S1.

Participants

For school selection and eligibility criteria, see Appendix S1. All students in 20 classes were asked to participate in the study (n = 553), with no exclusion criteria to allow stronger generalizability of findings. Parents of eligible students received a flyer and additional information about the study, before deciding to participate. To reduce expectation bias, both the growth mindset intervention and control conditions were explained as lesson series about the brain in the recruitment flyer (‘Explore your Brain’), but with different topics. The control condition was called ‘Opportunities and myths’ and the growth mindset intervention was called ‘Plasticity’, both aimed at ‘…exploring your brain, on your own and together’ to ‘…foster interest in science and learning’. However, the ulterior aim of the study was not communicated at this stage, only in a debriefing after finalizing the study. A sample of 439 students (79%) decided to participate, with both parents and students giving active informed consent. On average, 22 students participated in each class, with a minimum of 12 and a maximum of 29 students. Randomization at class level took place after study enrolment.

Interventions

Development of growth mindset intervention

The growth mindset intervention was developed iteratively over a 1-year period before the RCT, between September 2017 and 2018, incorporating several user tests, and more extensive pilots, approved by the local ethics committee (VCWE-S-18-00014). Development was based on users-as-designers, cocreation and responsible research and innovation (RRI; van Atteveldt et al., 2019) principles. For a more detailed account of the development, see Appendix S1.

Conditions

The control condition and growth mindset intervention condition were matched on duration (4 lessons × 50 min), frequency (1 per week), active participation (~60% of total time: individual, pair, quartet, group activities), graphical style, structural elements (starting with learning goals and activating exercise, take-home messages and teaser at the end for the next lesson, workbook exercises) and when possible on content (e.g., mirror drawing, brain selfie, sending a postcard to future self). While both intervention conditions were about exploring your own brain, the key difference was that only the growth mindset intervention focused on brain plasticity and controllability, mindset and their interrelatedness. In addition, only the growth mindset intervention included elements from cognitive-behavioural therapy (CBT; Fordham et al., 2021), such as the ABC model (Activating event, Beliefs, Consequences) to challenge fixed mindset beliefs and how they affect feelings, behaviour and their consequences. In line with previous studies, principles like ‘saying is believing’ (Yeager et al., 2016), and examples of ‘famous failures’ were incorporated as well in the growth mindset intervention condition.

All lessons were provided by researchers: eight undergraduates with a (research) master in developmental (neuro)psychology or neuroscience, between 28-01-2019 and 22-02-2019 (school 1) and between 11-03-2019 and 05-04-2019 (school 2), during regular mentor hours. Pairs of undergraduates gave both the control condition and growth mindset intervention (to separate classes), to eliminate potential teacher biases. Undergraduates were trained and supervised by the lead investigator of this study, T.W.P. Janssen, and followed a detailed protocol for teaching each lesson. There were no statistical differences in registered attendance of participating students between the growth mindset intervention and control groups (means: 3.61 and 3.66 out of four lessons), F(1, 437) = .398, p = .528. Across all participants, 76.8% attended all 4 lessons, and 94.6% attended 3 or 4 lessons.

The following paragraphs contain brief descriptions of both conditions, but see Table 1 for a more detailed overview, Appendix S1 for a more detailed description, and the archived data for all intervention materials translated in English (stored at DataverseNL). Lessons 1, 2 and 4 comprised the psychosocial component, and lesson 3 the psychophysiological component of the growth mindset intervention.

Control condition

Students learned about brain anatomy (lesson 1), brain illusions (lesson 2), brain imaging techniques (lesson 3) and brain myths and opportunities (lesson 4). Most exercises involved active participation, such as mirror drawing and making a brain hat (lesson 1), discussing various visual illusions, including ones in real-life (lesson 2), seeing, but not influencing, their own brain waves measured with mobile EEG (lesson 3), and participating in a neuromyth quiz and mailing a postcard to their future selves (lesson 4).

Growth mindset intervention condition

Students learned about brain plasticity (lesson 1), growth mindset (lesson 2), experienced influence over their own brain activity using neurofeedback (lesson 3) and learned how everything relates to their own school career (lesson 4). Most exercises involved active participation, such as performing a mirror drawing task (lesson 1), reflecting on former fixed and growth mindset reactions to challenging events at school (lesson 2), influencing a brain correlate of focused attention (theta/beta index, see Appendix S1, technical setup) with mobile EEG neurofeedback (NFB; lesson 3) and formulating SMART goals (Specific, Measurable, Achievable, Relevant and Time-Bound) to implement a growth mindset in school and mailing a postcard to their future selves (lesson 4). Although most lesson content was more general, we included several math examples, because STEM subjects, like math, are linked to ‘raw’ or ‘innate’ and thus fixed abilities: for example, during the main activity in lesson 2 (ABC exercise; see Table 1) and as example of brain plasticity during lesson 3 (neurofeedback; see Figure 2).

Details are in the caption following the image
Neurofeedback application lesson 3 (growth mindset intervention condition). Note: This is a selection of exercises and screens that students received in the neurofeedback application.

Outcomes

Outcomes were evaluated before (T0/pre), directly after as manipulation check (T1/post), and 1 year later (T2/1-year follow-up), with the aim to assess the superiority of the growth mindset intervention compared to the control condition. Primary outcomes involved mindset, school burnout symptoms, mindset-related constructs (effort beliefs, goal orientation, academic motivation) and academic achievement, with socio-economic status (SES) as moderator. The manipulation check at T1 focused only on the mindset measure, as this was the main mechanism that we targeted with the growth mindset intervention. Manipulation checks are critical for establishing causal inferences (Alferes, 2012). Table 2 shows a brief overview of the outcome variables, but see Appendix S1 for more details.

TABLE 2. Overview of outcome variables
Variable Source Nr. items Range α
Growth mindset (T0,1,2) De Castella and Byrne (2015) 8 8–48 .81 to .87
School burnout (T0,2) Salmela-Aro et al. (2009) 9 9–54 .83 to .84
Positive effort beliefs (T0,2) Tempelaar et al. (2015) 9 9–54 .60 to .64
Goal orientation a (T0,2) Elliot and Murayama (2008) 12 3–18 .50 to .88
Academic motivation b (T0,2) Ryan and Connell (1989) 17 5–20, 3–12 .52 to .75
Academic achievement c (T0,2) GPA, math grade n/a 0–10 n/a
Socio-economic status d (T0) RIVM (2017) n/a n/a n/a
  • Note: See Appendix S1 for detailed descriptions of the outcome variables.
  • a Subscales: performance avoidance (PAV), performance approach (PAP), mastery avoidance (MAV) and mastery approach (MAP).
  • b Only the ‘external regulation’ subscale (5 items) and ‘intrinsic motivation’ subscale (3 items) were used.
  • c Dutch, English, Math, Geography, History and Biology (Global Point Average; GPA).
  • d Based on postal code.

Sample size

Based on an a-priori power analysis (G*Power; Faul et al., 2007) with a RM-ANOVA design (2 groups, 3 measurements), using an expected effect size of f = .1 (small, based on previous RCTs), α error probability .05, power .80 and correlation between repeated measures of .50, we needed a total sample size of N = 164. However, due to the COVID-19 pandemic, which started between T2 and T3, we only report the T0–T2 results. A sensitivity power analysis with only two measurements, showed that a total of N = 200 participants were needed. To account for potential drop-outs and intracluster correlation (ICC) that can reduce the effective N, we planned to include at least 300 participants (Trial NL7562), but as many as possible to increase power. The final sample of N = 439 was well above this goal. See Analytical Methods for effective N, based on ICC and design effects.

Randomization

Randomization of classes (cluster allocation) was realized using the RANDARRAY function in excel, which generated random numbers between 0 and 1, with 15 decimal places. For each of the two schools, a random number array was generated for the participating classes within each school (n = 8, n = 12). Because we used stratification for both school and educational track, with the additional restriction of even frequencies within each educational track, we could make duo's of two classes within the same educational track of the same school. The class with the highest random number was allocated to the growth mindset intervention, while the class with lowest random number was allocated to the control condition. All participants within each class were invited to participate, see Participants.

The following order of events was implemented in this RCT: enrolment of schools, enrolment of classes, consent of individual students, random allocation, start interventions. Allocation was concealed throughout the RCT for participating students, parents and class mentors, but not the researchers. Although we did not explicitly mention the allocation, it was possible for participants to identify the allocated intervention, based on the participant information they received during enrolment. However, this information only described the interventions as ‘two different courses about the brain’ focusing either on ‘Opportunities and myths’ (control condition) or ‘Plasticity’ (growth mindset intervention), concealing the actual aims of the growth mindset intervention (stimulating a growth mindset). The lead investigator, T.W.P. Janssen, generated the random allocation sequence, enrolled the classes and assigned clusters to the interventions.

Analytical methods

Variable-oriented approach

A Linear Mixed Model approach was used in SPSS 26.0 (IBM Corp, 2019) to account for nested data, with Maximum Likelihood (ML) as estimation method, for each of the outcomes measures (dependent variables: mindset, school burnout symptoms, GPA, math grade) at T0 and T2. For missing longitudinal data due to drop-out or non-response (11.4%), Missing at Random (MAR) can be assumed if incomplete cases are retained in the multilevel model using ML estimation (Hox et al., 2017a). Using a repeated measures mixed model design is, therefore, the preferred method over ANCOVA (Hox et al., 2017b). We followed the principle of Intention-To-Treat Analysis (ITT), meaning that we analysed all available data, irrespective of compliance (whether students received all four lessons). ITT analysis is the most valid but conservative estimate of the true intervention effect (Gupta, 2011).

First, we specified a random intercept model (no predictors) with three levels (repeated measures, student, class), in order to assess whether clustering at the class level was present, and whether it should be taken into account in the analyses of the outcomes measures (ICC; mindset = .03; school burnout symptoms = .01; GPA = .13; math grade = .13). Based on the ICC and average cluster size of 21.95, the design effect (1 + (n − 1) × ICC) was calculated to be 1.702 for mindset and 1.176 for school burnout symptoms (<2; Muthen & Satorra, 1995), which means there was no meaningful clustering at the class level. However, the design effect for GPA (3.828) and math grade (3.643) showed meaningful clustering, which reduces the effective N, and thus statistical power, for GPA (439/3.828 = 115) and math grade (439/3.643 = 120). For consistency, all multilevel models included class as third level, to account for clustering.

We used a diagonal covariance structure for the repeated measures, which provided a better fit compared to other covariance structures. Subsequently, the following predictors were added to the model as fixed effects: Time, Condition, Gender, SES, Time × Condition, Time × Condition × SES and Time × Condition × Gender. As random effects we included intercepts for student (Level 2) and class (Level 3). We used a p-value of <.05 as indicator of significance. Effect sizes for intervention effects (Time × Condition) were calculated as Cohen's d, by dividing the unstandardized beta (representing the difference between conditions in change over time) with the pooled standard deviation of the raw outcome variable. Effect size interpretation was based on both statistical conventions (small: .20, medium: .50, large: .80; Cohen, 1988) and in comparison to the mindset literature.

Person-oriented approach

Within the social–cognitive model of achievement motivation (Dweck & Leggett, 1988), several constructs, such as goal orientation and effort beliefs, are closely interrelated with each other and with mindset. Recent data-driven work by Yu and McLellan (2020), demonstrated several motivational profiles, with two profiles displaying alternative combinations of mindsets and goal orientations that were not in line with the main tenets of the theory. To acknowledge both the interrelatedness of mindset-related constructs and potential heterogeneity in how they are combined across individuals, we applied Latent Transition Analysis (LTA) in Mplus Version 8.6 (Muthén & Muthén, 1998-2017), to explore whether the growth mindset intervention influenced transitions between motivational profiles from T0 to T2.

We included eight indicator variables: growth mindset, positive effort beliefs, performance approach, performance avoidance, mastery approach, mastery avoidance, external regulation and intrinsic motivation. For latent profile analysis (LPA) and LTA, we transformed the data based on the proportion of maximum scaling (‘POMS’) method, which transforms each scale to a metric from 0 (=minimal possible) to 1 (=maximum possible), by first making the scale range from 0 to the highest value and then dividing the scores by the highest value (Moeller, 2015). Contrary to standardization, this maintains the proportions of the absolute distances between the observed response options.

First, the factor structure of the indicator variables were verified using exploratory structural equation model (ESEM; Asparouhov & Muthén, 2009), separately for T0 and T2. A good model fit was determined by a Comparative Fit Index (CFI) value of .95 or higher, a Root Mean Square Error of Approximation (RMSEA) value of .06 or lower, and a Standardized Root Mean Square Residual (SRMR) of .08 or lower (Hu & Bentler, 1999). Second, to determine the best fitting LPA models for T0 and T2 separately, solutions with one to eight profiles were explored on several fit metrics: log likelihood, Akaike's Information Criterion (AIC), Bayesian Information Criterion (BIC), Sample-Adjusted BIC (SABIC), entropy, smallest class %, Lo, Mendell and Rubin (LMR) test and bootstrap likelihood ratio test (BLRT). In line with Ferguson et al. (2020), model retention decisions were based on a wide range of fit metrics and interpretability of profiles. Third, the retained models for T0 and T2 were compared on the number and type of profiles. As we were interested in transitions between the same profiles from T0 to T2, measurement invariance was assumed by fixing the indicator thresholds over time. This assumption considers the latent profiles to have similar meaning over time. This assumption was checked by qualitatively comparing the latent profiles at T0 and T2, and by exploring changes in model fit between the unconstrained (non-invariant) and constrained (invariant) models. Fourth, LTA was performed in four steps: (a) without the covariate Condition, (b) with Condition only predicting latent profiles at T0, (c) with Condition also predicting transitions between profiles from T0 to T2, and (d) the most complex model, by also including interactions. The best fitting model was chosen based on a chi-squared difference test. Fifth, odds ratios and 95% CI were calculated for transitions between profiles from T0 to T2 for the growth mindset intervention compared to the control condition, and significance was assumed if the 95% CI did not include 1.

RESULTS

Descriptive statistics

At pre-intervention (T0), there were no significant differences between the control condition and growth mindset intervention groups on demographics, questionnaire data or academic achievement, demonstrating successful randomization, see Table 3.

TABLE 3. Group characteristics at pre-intervention (T0)
Condition, N = 439
Control Intervention Group difference
n = 217 n = 222 F (df) p
Demographic data
Age (years) 12.73 (.48) 12.73 (.46) .04 (1424) .848
SES .26 (.88) .32 (1.00) .37 (1434) .543
Gender (M/F) 110/107 120/102 χ2 = .50 (1) .504
Questionnaires
Growth mindset 34.84 (6.19) 34.87 (6.28) .00 (1424) .955
School burnout 27.33 (7.72) 26.77 (7.93) .54 (1423) .464
Positive effort beliefs 41.46 (4.58) 41.79 (5.12) .51 (1424) .475
MAP 14.87 (2.04) 14.99 (1.98) .37 (1424) .545
MAV 13.23 (2.76) 13.11 (2.92) .19 (1424) .666
PAP 11.38 (3.26) 11.40 (3.16) .00 (1424) .949
PAV 11.00 (3.91) 11.26 (3.73) .48 (1424) .488
External regulation 12.84 (2.29) 12.59 (2.31) 1.22 (1422) .269
Intrinsic motivation 6.99 (1.91) 7.14 (1.91) .62 (1422) .433
Academic achievement
GPA 6.66 (.63) 6.65 (.66) .04 (1437) .844
Math grade 6.58 (1.15) 6.51 (1.32) .31 (1437) .578
  • Note: There were no significant differences between the control and growth mindset intervention groups pre-intervention (T0), demonstrating successful randomization.
  • Abbreviations: GPA, global point average; MAP, mastery approach; MAV, mastery avoidance; PAP, performance approach; PAV, performance avoidance; SES, socio-economic status.

Evaluation psychophysiological component

We assessed whether the implementation of the neurofeedback application (lesson 3) in the growth mindset intervention group was successful based on whether (1) students had objective control over their own brain activity, as indicated by a decrease in theta/beta index compared to their resting baseline, (2) objective and subjective control were related, and (3) students were more convinced that they could influence their own brain directly after finishing lesson 3.

First, objective control was assessed by comparing theta/beta index during resting baseline with each neurofeedback exercise. The results showed that students were able to lower theta/beta index by deliberately increasing their focused attention during the line exercise, F(1, 203) = 28.18, p < .001, η p 2 $$ {\eta}_p^2 $$  = .12, during the neurons exercise, F(1, 202) = 71.31, p < .001, η p 2 $$ {\eta}_p^2 $$  = .26, during the free choice exercise at the end, F(1, 202) = 60.98, p < .001, η p 2 $$ {\eta}_p^2 $$  = .23, but not during the reading exercise, F(1, 203) = 1.59, p = .209, η p 2 $$ {\eta}_p^2 $$  = .01. Except for the reading exercise, these are all large effect sizes (> η p 2 $$ {\eta}_p^2 $$  = .14). Although the reading exercise was expected to be most ecologically valid, students told us that it was very difficult to simultaneously focus on reading the text and focus on the auditory feedback, which may explain why this exercise was less successful. Second, to examine whether subjective and objective control were related, we calculated the average objective control, by averaging theta/beta index for the exercises that worked as intended at the group level (line, neurons and choice) minus resting baseline. Immediately after lesson 3, we also asked students how much subjective control they experienced during the neurofeedback exercises on a 0–100 scale, see Appendix S2: Figure 1 (it follows the same pattern as objective control). Average objective and subjective control were strongly correlated, r(201) = −.420, p < .001, meaning that having more objective control was also experienced that way. Finally, 91.7% indicated they were more convinced that they could influence their own brain after lesson 3.

Manipulation check

Within one week after finishing the interventions, we checked whether the growth mindset intervention affected the proposed mechanism (mindset). In line with a successful manipulation, we found a significant Time × Condition interaction, F(1, 450.89) = 11.98, p < .001. The positive slope, β = 2.33, 95% CI [1.01, 3.65], indicates that the growth mindset intervention resulted in 2.33 points more growth mindset endorsement from pre (T0) to direct-post (T1) than the control condition. This is a small-medium effect size (d = .38).

Variable-oriented approach

Four linear mixed models were used to test the hypotheses concerning the expected superiority of the growth mindset intervention compared to the control condition, one for each outcome measure (mindset, school burnout symptoms, GPA and math grade). Table 4 and Figure 3 contain the main results. For three out of four outcomes (mindset, GPA, math grade), there were significant Time × Condition interactions. Concerning the three-way interactions, SES and gender did not moderate the Condition effects for any of the outcomes. There were, however, main effects of gender on all outcome measures: girls reported less growth mindset endorsement, more school burnout symptoms and had higher academic achievement (GPA and math grade). Lastly, students from higher SES backgrounds reported higher growth mindset endorsement, irrespective of Time or Condition. The following paragraphs focus on the specific effects of the intervention.

TABLE 4. Linear mixed model outcomes
Fixed effects Mindset School burnout GPA Math grade
F (df) p F (df) p F (df) p F (df) p
Time .01 (1, 398.17) .917 63.07 (1, 397.88) <.001 71.31 (1, 415.96) <.001 28.29 (1, 419.33) <.001
Condition .03 (1, 25.25) .872 .27 (1, 23.36) .608 .03 (1, 21.66) .865 .09 (1, 23.52) .771
Gender 3.92 (1, 468.56) .048 5.85 (1, 471.19) .016 39.51 (1, 484.26) <.001 32.09 (1, 491.78) <.001
SES 4.78 (1, 311.07) .030 .05 (1, 234.33) .827 1.77 (1, 485.89) .184 2.16 (1, 500.46) .142
Time × Condition 3.89 (1, 414.04) .049 1.02 (1, 413.52) .313 4.85 (1, 439.60) .028 9.26 (1, 450.28) .002
Time × Condition × SES .01 (1, 438.17) .918 .68 (1, 438.98) .409 .44 (1, 462.46) .506 .01 (1, 477.57) .907
Time × Condition × Gender .07 (1, 427.56) .800 .25 (1, 426.76) .621 3.08 (1, 459.20) .080 .84 (1, 471.79) .361
  • Note: df, degrees of freedom; GPA, global point average; SES, socio-economic status.
Details are in the caption following the image
Raw means for the control condition and growth mindset intervention condition for pre-intervention (T0) and 1-year follow-up (T2). Note: Grades were on a 1–10 scale. Note that the figure represents raw means with standard error (SE) bars and is only used to show the direction of the raw effects (not representing the statistical results directly).

Mindset

In line with the manipulation check directly after the intervention (T1), one year later the growth mindset intervention group continued reporting increased growth mindset compared to the control condition group. The positive slope, β = 1.62, 95% CI [.01, 3.25], indicates that the growth mindset intervention resulted in 1.62 points more growth mindset endorsement from pre (T0) to 1-year follow-up (T2) than the control condition. This is a small effect size (d = .25), which is approximately 2/3 of the effect size directly after the interventions (manipulation check).

School burnout symptoms

In contrast to our hypothesis, the growth mindset intervention did not affect school burnout symptoms from T0 to T2. Rather, both the control and growth mindset intervention groups reported an increase in school burnout symptoms, as indicated by the positive slope of Time, β = 4.82, 95% CI [3.63, 6.02], which is a 4.82 points increase from pre (T0) to 1-year follow-up (T2).

GPA and math grade

In line with our hypotheses, the growth mindset intervention group demonstrated better academic achievement (GPA and math) at follow-up (T2) compared to the control condition group. Figure 3 shows that this is because of a less steep decline in grades, rather than a larger increase in grades for the growth mindset intervention, similar to other studies (Blackwell et al., 2007). For GPA, the positive slope, β = .14, 95% CI [.01, .26], indicates that the growth mindset intervention resulted in .14 points higher GPA from pre (T0) to 1-year follow-up (T2) than the control condition. This is a small effect size (d = .22). For math grades, the positive slope, β = .42, 95% CI [.15, .69], indicates that the growth mindset intervention resulted in .42 points higher math grade from pre (T0) to 1-year follow-up (T2) than the control condition. This is a small-medium effect size (d = .36). To test whether the GPA effects were actually driven by the more pronounced effects of math grades, we re-analysed GPA based on five instead of six subjects (without math). This seemed to be the case, as the Time × Intervention interaction for the adapted GPA without math was no longer significant, F(1, 453.05) = .38, p = .537. Overall, these results show a selective, relatively large (compared to the literature) protective effect of the growth mindset intervention on math grades (less steep decline).

As an additional exploratory analysis, we were interested to test whether changes in growth mindset were predictive of changes in math grades. We, therefore, added a time-varying predictor, which was mindset at T0 and T2, to the linear mixed model for math. The three-way interaction, Time × Condition × Mindset, was significant, F(1, 344.87) = 7.08, p = .008. Pearson correlations showed that only for the growth mindset intervention group, changes in mindset and math grades were significantly related, r(194) = .171, p = .016 (see Figure 4), but not for the control group, r(176) = −.06, p = .423. This finding provides indirect evidence that the proposed mechanism, changing mindset, was responsible for improved academic achievement in students receiving the growth mindset intervention.

Details are in the caption following the image
Scatterplot showing correlation between changes in mindset (T2-T0) and changes in math grade (T2-T0) for the growth mindset intervention group. Note. T0 = pre-intervention, T2 = 1-year follow-up; N = 196 (complete data).

Person-oriented approach

Exploratory structural equation model (ESEM)

ESEM supported the factor structure of the eight indicator variables based on most fit indices: T0 (CFI = .942, RMSEA = .039 and SRMR = .027), T2 (CFI = .961, RMSEA = .035 and SRMR = .026).

LPA model fitting

To determine the best fitting LPA models for T0 and T2, solutions with one to eight profiles were explored, see Table 5 for a summary of all fit metrics. In line with Ferguson et al. (2020), a holistic approach was used to support model retention decisions, utilizing both a wide range of fit metrics and interpretability of profiles.

TABLE 5. LPA model fit indices for T0 and T2
Model Log likelihood AIC BIC SABIC Entropy Smallest class % LMR p-value BLRT p-value
T0
1 1170.38 −2308.77 −2243.89 −2294.67
2 1348.64 −2647.27 −2545.91 −2625.25 .84 32.60 <.001 2 > 1 <.001 2 > 1
3 1440.17 −2812.33 −2674.48 −2782.37 .82 14.60 .01 3 > 2 <.001 3 > 2
4 1496.42 −2906.84 −2732.50 −2868.96 .78 15.70 .58 3 > 4 <.001 4 > 3
5 1526.28 −2948.57 −2737.74 −2902.75 .80 3.80 .119 4 > 5 <.001 5 > 4
6 1548.64 −2975.27 −2727.95 −2921.53 .78 8.00 .279 5 > 6 <.001 6 > 5
7 1580.61 −3021.23 −2737.42 −2959.55 .80 3.80 .279 6 > 7 <.001 7 > 6
8 1606.63 −3055.27 −2734.97 −2985.66 .82 4.00 .488 7 > 8 <.001 8 > 7
T2
1 908.30 −1784.60 −1721.18 −1771.95
2 1080.72 −2111.44 −2012.35 −2091.68 .82 40.10 <.001 2 > 1 <.001 2 > 1
3 1169.81 −2271.62 −2136.85 −2244.73 .80 22.10 .08 3 > 2 <.001 3 > 2
4 1214.17 −2342.33 −2171.90 −2308.33 .82 2.60 .14 3 > 4 <.001 4 > 3
5 1258.25 −2412.49 −2206.39 −2371.38 .81 2.30 .080 5 > 4 <.001 5 > 4
6 1297.39 −2472.77 −2231.00 −2424.54 .81 2.30 .107 5 > 6 <.001 6 > 5
7 1303.23 −2466.46 −2189.01 −2411.11 .82 .50 .752 6 > 7 <.001 7 > 6
8 1297.39 −2436.77 −2123.65 −2374.31 .84 .00 .240 7 > 8 <.001 8 > 7
  • Note: The LMR test and the BLRT compare the current model to a model with k-1 profiles. The retained models are depicted in bold.
  • Abbreviations: AIC, Akaike's Information Criterion; BIC, Bayesian Information Criterion; BLRT, bootstrap likelihood ratio test; LMR, Lo–Mendell Ruben; LPA, latent profile analysis; SABIC, Sample-Adjusted BIC.

For T0, based on the elbow plots of the Log likelihood, AIC, BIC and SABIC, model three was in between large increases in model fit (models 1–2) and flattening of model fit (models 4–8). The LMR test was significant for model three (p < .001), which means the three-profile model is a better representation than the two-profile model. Entropy was high (>.80), BLRT was significant (<.001), and the smallest profile included more than 5% of the sample. As the model fit did not improve considerably for the four-profile solution, we retained model 3.

For T2, the elbow plots for Log likelihood, AIC and SABIC also showed the largest improvement up to model three, but also continued improvements up till model six. However, the smallest class sizes were below 5% for models 4–8. For model three, LMR was near-significant, which means the three-profile model is a better representation than the two-profile model. Entropy was high (>.80), BLRT was significant (<.001), and the smallest profile included more than 5% of the sample. As the model fit did not improve considerably for the four-profile solution, we retained model 3.

As we established an equal number of three profiles for T0 and T2, the next step was to qualitatively compare the three profiles on similarities and differences, and to test the assumption of measurement invariance. Appendix S2 contains line graphs, means of the indicator variables for T0 and T2, and the number of students in each profile. While the profiles show clear resemblance for both time points on six out of eight indicator variables, this is less the case for external regulation and intrinsic motivation. At T2, these two indicators become more similar for the three profiles, with all three profiles scoring relatively higher on external regulation compared to intrinsic motivation. This instability in profiles is reflected in a better fitting unconstrained model compared to the constrained model, χ 2(120) = 589.82, p < .001. As the profiles are similar on most of the other indicator variables, we decided to continue with the constrained model by fixing the thresholds over time for LTA, but to be more cautious in interpreting profile differences on external regulation and intrinsic motivation.

LTA results

The three profiles were labelled (1) Fixed, (2) Growth non-competitive and (3) Growth competitive, see Figure 5 for a line graph of the profiles across both time points and Appendix S3 for the means on the indicator variables. The first profile was characterized by less growth mindset endorsement (more towards the fixed side of the mindset scale), less positive effort beliefs and less mastery goals. The second and third profiles both endorsed growth mindset and positive effort beliefs more than the first profile, while they differed in goal orientation: The Growth non-competitive profile was characterized by low endorsement of performance goals and high endorsement of mastery goals. In contrast, students in the Growth competitive profile endorsed both performance and mastery goals.

Details are in the caption following the image
Three-profile model LTA: Line graph comparing profiles on indicator variables. Note: Data on Y-axis were transformed based on the proportion of maximum scaling (‘POMS’) method, which transforms each scale to a metric from 0 (=minimal possible) to 1 (=maximum possible), by first making the scale range from 0 to the highest value and then dividing the scores by the highest value (Moeller, 2015). Contrary to standardization, this maintains the proportions of the absolute distances between the observed response options. POMS = [(observed − minimum)/(maximum − minimum)].

The best fitting LTA model included the effects of the covariate Condition on profile membership at T0, T2 and the transition probabilities between T0 and T2, compared to a model without Condition influencing transition probabilities, χ2(2) = 589.82, p < .001.

Table 6 contains the transition probabilities, and Table 7 contains the covariate effects (Condition) on transition (T0–T2) probability odds ratios. Students who were in the Fixed profile (1) and subsequently received the growth mindset intervention, were 2.677 (95% CI [1.261, 5.680]) times more likely to transition to the Growth Non-Competitive profile (2) and 2.577 (95% CI [1.238, 5.364]) times more likely to transition to the Growth Competitive (3) profile, 1 year later, compared to students who received the control condition. In the case that students were in one of the growth mindset profiles (2 and 3) and subsequently received the growth mindset intervention, they were less likely to transition to the Fixed profile (1), 1 year later, compared to students who received the control condition (respectively: OR = .374, 95% CI [.176, .793] and OR = .388, 95% CI [.186, .808]). All these effects were significant, as the 95% CI did not include 1.0.

TABLE 6. Latent profile and transition probabilities for T0 and T2
Latent classes T0 Transition matrix T2
1 2 3
1 .135 .684 .156 .160 .287
2 .277 .173 .703 .124 .324
3 .588 .251 .185 .564 .388
  • Note: This table contains the probabilities of each latent class (1 = Fixed; 2 = Growth non-competitive; 3 = Growth competitive) at T0 and T2, and the transition probabilities between T0 and T2. For example, P(T2 = 1) = .135 × .684 + .277 × .173 + .588 × .251.
TABLE 7. Covariate effects (condition) on transition (T0–T2) probability odds ratios
Latent classes T2
1 2 3
T0 1 1.000 (1.000, 1.000) 2.677 (1.261, 5.680) 2.577 (1.238, 5.364)
2 .374 (.176, .793) 1.000 (1.000, 1.000) .963 (.485, 1.911)
3 .388 (.186, .808) 1.039 (.523, 2.062) 1.000 (1.000, 1.000)
  • Note: Transition table of covariate effects (0 = control condition; 1 = growth mindset intervention condition) on odds ratio (OR) with 95% confidence intervals (CI) for T0 to T2. Larger OR means more students move to that class relative to staying in the class for the growth mindset intervention compared to the control condition. Smaller OR means less students move to that class relative to staying in the class for the growth mindset intervention compared to the control condition. Effects (in bold) are significant if 95% CI does not include 1.0.

DISCUSSION

In recent years, a discussion has been unfolding about the effectiveness of growth mindset interventions, and more specifically, about what constitutes a meaningful effect (Miller, 2019). Depending on the benchmarks used, such interventions are typified as hardly meaningful (Sisk et al., 2018) to more favourable (Miller, 2019). From a cost-effectiveness perspective, brief and low-cost growth mindset interventions (Yeager et al., 2019) are a promising direction to take. In addition to lowering the costs, another approach is to increase the effectiveness of mindset interventions. We identified a potential way to do so, based on our observation that previous interventions were largely theoretical, neglecting more experiential aspects related to growth mindset. We, therefore, combined psychosocial (theoretical) and psychophysiological (experiential) intervention components with the aim to increase effectiveness. For the latter we used neurofeedback, to provide a convincing experience of being in control over one's own brain activity. By emphasizing and experiencing the controllable and malleable nature of the brain, this experience was expected to help with internalizing the more theoretical growth mindset and brain plasticity messages that were discussed in the psychosocial part of our intervention. We used both variable- and person-oriented approaches to assess the effectiveness of the intervention, with the latter allowing us to better capture the heterogeneity in mindset-based meaning systems. Overall, the results of this Randomized Controlled Trial (RCT) with 439 Dutch adolescents demonstrate relatively large effects of our newly developed intervention on growth mindset (d = .25), math grades (d = .36) and mindset profiles one-year post-intervention, irrespective of SES.

One still open question in the literature is whether growth mindset interventions are susceptible to the fadeout effect, or whether effects are sustained (or enhanced) over time due to positive recursive processes (Sisk et al., 2018). Although we could not investigate this question for academic achievement, the effects on mindset seem to support the fadeout effect: the effect size one year after the intervention (d = .25) was approximately 2/3 of the effect size directly after the intervention (d = .38; when we only measured mindset, as manipulation check). Nevertheless, these results show some robustness to fadeout effects as the mindset changes were still detectable one year later, in combination with higher math grades. Our follow-up time of one year is a unique contribution to the literature, as only few studies have investigated intervals longer than four months (4 of 43 effect sizes; Sisk et al., 2018).

With respect to academic achievement, we demonstrated a selective, relatively large protective effect of the growth mindset intervention on math grades (less steep decline). In first instance, effects seemed unselective, with a small effect size advantage for the growth mindset group on GPA (d = .22); however, after re-analysing GPA based on five subjects (without math), the effect was no longer significant. The selectivity of our findings are not fully unexpected, as STEM subjects like math seem especially mindset-sensitive and for that reason are considered main academic achievement measures in previous studies (Blackwell et al., 2007). Another explanation for the selectivity of our findings, is that for similar reasons, we emphasized math examples more than other subjects during the intervention. This may have skewed the effect of the intervention towards math more than other subjects. Either way, for growth mindset interventions to have a broader impact on academic achievement, we need to better understand why intervention effects do not always transfer to other subjects, besides math.

Another important open question in the literature, is whether the effects of growth mindset interventions on achievement are specific, in other words whether improvements in academic achievement are caused specifically by changes in mindset. Despite our best efforts to match the control condition with the growth mindset intervention condition, it is impossible to precisely manipulate only one factor (mindset), such as in placebo-controlled RCTs. However, we found indications for the specificity of our findings, namely: (1) experimental evidence showed that the manipulation of mindset was successful directly post-intervention and (2), indirect correlational evidence showed that changes in mindset were related to changes in math grades at 1-year follow-up, only for students who received the growth mindset intervention. These are unique findings, considering that 35% of previous studies did not report a manipulation check, and of those reported, almost half failed to demonstrate a successful manipulation – and even more puzzling – those who reported a successful manipulation did not demonstrate significant effects on academic achievement in contrast to studies with failed manipulation checks (Sisk et al., 2018).

With this study we set out to develop a more effective growth mindset intervention by combining psychosocial and psychophysiological intervention components. In reference to the literature, the effect size for math grades (d = .36) is relatively large compared to the meta-analytic average (overall: d = .08; high-risk students: d = .19; Sisk et al., 2018), and a low-intensity intervention for lower-achieving adolescents (d = .11; Yeager et al., 2019). Although the design of this study does not allow to estimate the contributions of both components separately, the relatively larger effect size that we found may be attributable to synergetic effects of the psychosocial and psychophysiological components. The visceral and personal experience of being in control over one's own brain activity during the neurofeedback exercises (psychophysiological component), may have helped with internalizing the theoretical content about brain plasticity and growth mindset (psychosocial component), by acting as convincing ‘empirical’ proof for the controllable and malleable nature of one's very own brain functions. This interpretation is supported by (1) a recent study, which demonstrated that neurofeedback can increase the sense of control and self-efficacy (Harris et al., 2021), (2) the importance of self-reference in relation to mindset (De Castella & Byrne, 2015) and (3) literature about experiential learning (Kolb, 2014).

An important question is whether the larger mean effects of our intervention outweigh the additional costs and intensity. The psychophysiological component is evidently the most innovative, but also expensive addition to the growth mindset intervention. We used commercial mobile EEG systems, costing €240 ($274) per headset. A possible scenario for a school would be to purchase one or more systems, which could be used repeatedly to cover larger numbers of students (which would bring the cost down per student). To make this a better value proposition, the EEG systems could be used as an educational tool as well and provide opportunities for teachers and students to gain hands-on experience in neuroscience research (Azeka et al., 2020; Janssen et al., 2021). Our developed neurofeedback application will be shared free of charge and is available to download via DataverseNL.

In addition to the variable-oriented approach, we assessed the effectiveness of our growth mindset intervention using a person-oriented approach, to capture both the heterogeneous nature and interrelatedness of constructs that play a role in mindset theory (effort beliefs, goal orientation, academic motivation). We identified three profiles: (1) Fixed, (2) Growth non-competitive and (3) Growth competitive. Similar profiles were identified in a previous study, although we did not identify a fourth profile of disengaged students (Yu & McLellan, 2020). In this previous study, the growth profiles showed more adaptive patterns with higher motivation and math grades. Our third profile, the Growth competitive, deviates from most variable-oriented research and theorized associations (Burnette et al., 2013), which relate growth mindset with mastery goals (for an exception, see Molden & Dweck, 2000): students in the Growth competitive profile endorsed performance goals alongside mastery goals, instead of only mastery goals. In line with our hypotheses, students in the growth mindset intervention were more likely to transition from the Fixed mindset profile to the more adaptive growth mindset profiles one year later, compared to students in the control condition, with very similar odds ratios for the Growth non-competitive and Growth competitive profiles (OR = 2.7 and 2.6). In addition, students in the growth mindset intervention were less likely to transition from either growth mindset profile to the Fixed mindset profile (OR = .37 and .39). These results add to the literature, by demonstrating more holistic effects of the intervention on mindset-based meaning systems, involving multiple motivational constructs. Another notable observation is that the same growth mindset intervention can apparently stimulate two different types of growth mindset meaning systems, varying in goal orientation. This leads to new questions: why do some students transition to a more competitive profile, and others not? Could these profiles have different long-term trajectories?

We could not confirm two other hypotheses, namely the expected intervention effects on school burnout symptoms and moderation of intervention effects by SES. First, we expected a reduction in school burnout symptoms (or protection against further deterioration) based on recent correlational and experimental evidence for positive effects of growth mindset on psychological distress (Burnette et al., 2020; Kim, 2020). Rather, we found a general increase in school burnout symptoms in both the control and growth mindset intervention groups from pre to one-year follow-up. This may be partly explained by the fact that our intervention was implemented at the start of a major transition period, in the first half year of high school. Our results are in line with other longitudinal studies during early adolescence, which showed a general increase in school burnout symptoms (Wang et al., 2015), a reduction in school engagement (Wang et al., 2015; Wang & Eccles, 2012) and declines in motivation and academic performance (Eccles et al., 2009). The effects of our intervention may thus not generalize to the mental health domain, or such effects may be more delayed, depending on a positive recursive process.

Lastly, we could not demonstrate more beneficial effects for students from lower SES backgrounds, in contrast to other research (Sarrasin et al., 2018; Sisk et al., 2018). Although more objective than self-report, our SES measure based on postal code may not have been sensitive enough, as it was based on only part of the postal code (4 of 6 digits). Caution should also be paid to directly comparing SES indicators in the Netherlands with for example the US, and it may be the case that SES was relatively high in this sample compared to other countries.

Strengths, limitations, future studies

This study had both strengths and limitations. Strengths were the involvement of stakeholders in the development of the intervention, a successful randomization, high implementation fidelity (94.6% attended 3 or 4 lessons), high questionnaire completion rate at one-year follow-up (85%), a successful manipulation check (directly after the intervention) and successful neurofeedback implementation (evidence for both objective and subjective control, 91.7% convinced that they could influence their own brain), the addition of a well-being measure and the combination of variable- and person-oriented approaches. There were several limitations as well to take into account. First, the reliability of the academic motivation questionnaire scales was low, which may have contributed to the violation of the measurement invariance assumption in the LTA. For this reason, we decided to be cautious in interpreting any profile differences on these scales. Second, the impact of the COVID-19 pandemic refrained us from exploring the two-year follow-up, due to high drop-out rates and it's confounding effects. Another drawback of our design is that we are not able to dissociate the relative contributions of the psychosocial and psychophysiological components to the intervention effects. Lastly, although we found a relatively large effect size compared to the literature, this should be interpreted with caution, as the comparison is indirect.

CONCLUSION

Overall, our newly developed growth mindset intervention, which combined psychosocial and psychophysiological intervention components, demonstrated relatively large effects on growth mindset (d = .25) and mindset profiles (transition to more adaptive growth-oriented profiles), and a protective effect on math grades (d = .36) one-year post-intervention, compared to the literature (d = .08; Sisk et al., 2018). Against our hypotheses, we did not find effects on school burnout symptoms, nor a moderating effect of SES. We did find indirect indications for the specificity of our findings: (1) the manipulation check was successful directly post-intervention, which is critical for establishing causal inferences (Alferes, 2012) and (2), changes in mindset were related to changes in math grades only for students who received the growth mindset intervention.

AUTHOR CONTRIBUTIONS

Nienke van Atteveldt: Conceptualization; funding acquisition; methodology; project administration; supervision; writing – review and editing. Tieme W. P. Janssen: Conceptualization; data curation; formal analysis; investigation; methodology; project administration; supervision; validation; visualization; writing – original draft; writing – review and editing.

ACKNOWLEDGEMENTS

This work was supported by a European Research Council (ERC) Starting Grant 716736 (BRAINBELIEFS). We are grateful for the contributions to this research by the following undergraduate students, Reine Ramaekers, Evy Sens, Fleurie Reeuwijk, Jamie Hoefakker, Lisanne Berends, Lodewike de Groot, Jikke Boelen Keun, Paul Bethlehem, Patricia Dreier Gligoor, Alfa Sanoveriana, Roos de Jong, Iris Koning, Hülya Yilmaz, Chaîmae el Hassouni, biology teacher Anna Tuenter, PhD students Smiddy Nieuwenhuis and Sibel Altikulaç, the main contact persons in the participating schools Annette van Oosten and Astrid Kwakernaak, and all the participating students. Special thanks go to Creative Learning Lab (Waag Society, Amsterdam): Taco van Dijk, Marloeke van der Vlugt, Thomas Asscheman and Karien Vermeulen.

    CONFLICTS OF INTEREST

    The authors declare no conflicts of interest.

    DATA AVAILABILITY STATEMENT

    This study's design and hypotheses were preregistered; see https://www.trialregister.nl/trial/7562. Materials have been made publicly available at DataverseNL and can be accessed at https://doi.org/10.34894/EULELM. Data and analysis code for this study are available by emailing the corresponding author.