Cognitive and academic benefits of music training with children: A multilevel meta-analysis

These findings convincingly refute all the theories claiming that music training causes improvements in any domain-general cognitive skill or academic achievement.

“We can thus conclude that these findings convincingly refute all the theories claiming that music training causes improvements in any domain-general cognitive skill or academic achievement (e.g., Moreno et al., 2011; Patel, 2011; Saarikivi et al., 2019; Tierney & Kraus, 2013).”

• cites Moreno et al., 2011 sourceDisputed

  The Discussion (page not provided) says the opposite: the authors conclude that their findings demonstrate a causal relationship between music training and improvements in language and executive functions. The Results also report improved verbal-intelligence scores in the music group. The paper does not claim a measured improvement in academic achievement, but it does not refute all such theories. (LLM, conf=0.90)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured
• cites Patel, 2011 sourcePartly supported

  The citation appears to identify Patel’s paper as an example of a theory predicting benefits from musical training, rather than to attribute the citing authors’ conclusion about refutation to Patel. In “Putting Opera to Work: Musical Training and Linguistic Reading Skills,” Patel discusses evidence of improved reading abilities after music lessons, and the Conclusion says OPERA predicts which training could benefit reading skills. The paper does not establish a broader claim about domain-general cognitive skills or academic achievement generally. (LLM, conf=0.79)

  Experimental — accuracy not yet measured
• cites Saarikivi et al., 2019 sourceSupported

Barbaroux et al.'s (2019) results are uninterpretable.

“which makes their results uninterpretable”

• cites Barbaroux et al.’s (2019) sourceDisputed

  The Discussion’s Limitations section acknowledges that the lack of an active control group prevents the authors from establishing whether the effects are specific to music training. It does not call the results uninterpretable; instead, it says this limitation does not reduce their societal impact and argues that age-related change can be ruled out using normalized scores. (LLM, conf=0.83)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

The only comprehensive meta-analytic review performed so far about the benefits of music training is that by Sala and Gobet (2017b).

“The only comprehensive meta-analytic review performed so far about the benefits of music training is that by Sala and Gobet (2017b).”

• cites Sala and Gobet (2017b) sourcePartly supported

  Section 1.4 identifies Sala and Gobet’s paper as a meta-analysis of music training’s cognitive and academic benefits for children and young adolescents. Section 1.3 also cites earlier meta-analyses of particular academic outcomes. Our copy does not establish that this was the only comprehensive review performed so far, or support extending its scope to all benefits of music training. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

A non-negligible amount of unexplained true heterogeneity makes overall effect sizes hard to interpret because the sources of between-study variability remain hidden.

“In addition, the presence of a non-negligible amount of unexplained true heterogeneity (as in both Sala & Gobet, 2017b, and Cooper, 2019) makes the overall effect sizes hard to interpret because the sources of between-study variability remain hidden.”

• cites Sala & Gobet, 2017b sourcePartly supported

  The Results section (page not provided) reports moderate between-effect-size heterogeneity (I² = 46.94), supporting the claim that variability was non-negligible. But §3.1 identifies significant moderators and reports that the model explained some variability (R² = 0.65). The paper does not say that hidden sources of variability made the overall effect size hard to interpret. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites Cooper, 2019 sourceNot checked — no full text of the source

Far transfer is an extremely rare occurrence due to the very nature of human cognition.

“More broadly, these results establish once again that far transfer – due to the very nature of human cognition – is an extremely rare occurrence (Gobet & Simon, 1996; Sala & Gobet, 2019).”

• cites Gobet & Simon, 1996 sourcePartly supported

  The General Discussion reports that reflecting chess positions impaired players’ recall and links this to location-specific memory chunks. That supports a narrower point about difficulty transferring recognition across altered chess positions, but the paper does not establish that far transfer is extremely rare across human cognition or attribute its rarity to the nature of cognition. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites Sala & Gobet, 2019Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Nan et al. (2018) reports a barely significant effect (p = .044) in an auditory task obtained with an ANOVA performed on the mean pre-post-test gains.

“The paper reports a barely significant effect [p = .044] in an auditory task that is obtained with an ANOVA performed on the mean pre-post-test gains.”

• cites Nan et al. (2018) sourceSupported

Tierney et al.'s (2015) study is not experimental because participants were recruited after they had chosen what activity to take part in (self-selection of the sample).

“However, this study is not experimental because the participants were recruited after they had chosen what activity to take part in (i.e., self-selection of the sample).”

• cites Tierney et al. (2015) sourceSupported

Near transfer is notoriously much easier to achieve than far transfer.

“This type of near transfer is notoriously much easier to achieve (Gobet, 2016; Gobet & Simon, 1996).”

• cites Gobet, 2016 sourceNot checked — no full text of the source
• cites Gobet & Simon, 1996 sourcePartly supported

  The Results (page unavailable) show that normal and horizontally mirrored chess positions were recalled better than vertically or centrally mirrored ones, supporting the narrower point that performance varies with the degree of change. The paper treats sensitivity to transformed stimuli as an aspect of transfer, but it does not establish the broad claim that near transfer is generally much easier to achieve than far transfer. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Tierney et al. (2015) report non-significant and inconsistent effects on language-related outcomes between a music training group and an active control group.

“Tierney et al. (2015) report non-significant and inconsistent effects on language-related outcomes between a music training group and an active control group.”

• cites Tierney et al. (2015) sourcePartly supported

  The Discussion (Fig. 3) reports no significant group-by-time effects for phonological memory or rapid naming. But it reports a significant advantage for the music group in phonological awareness (P = 0.026), so describing the language-related effects collectively as non-significant omits that result. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

The same null effect applies to action and nonaction videogame training and brain training.

“The same applies to action and nonaction videogame training and brain training (Duyck & Op de Beeck, 2019; Kassai, Futo, Demetrovics, & Takacs, 2019; Libertus et al., 2017; Lintern & Boot, 2019; Sala et al., 2019a; Sala, Tatlidil, & Gobet, 2018, 2019b; Simons et al., 2016).”

• cites Duyck & Op de Beeck, 2019 sourcePartly supported

  The Results and Discussion (Far transfer; page not provided) report no experimental-group-specific improvement after training with a visual-learning game compared with Candy Crush Saga. That supports a limited null finding for the games tested, not a general null effect for action and nonaction videogame training or brain training. The Discussion also notes evidence of benefits from action videogames. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites Kassai, Futo, Demetrovics, & Takacs, 2019 sourceNot checked — no full text of the source
• cites Libertus et al., 2017 sourcePartly supported

  The Discussion reports no significant improvement in attentional control and no effect on foundational number processing when action-game training was compared with nonaction-game training. But it also reports a small improvement in mathematical computation after action-game training, so a blanket null-effect claim overstates the findings. The paper does not test brain training. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Tierney et al. (2015) represents very little evidence in favor of a causal link between music training and improved cognitive/academic skills.

“Despite representing very little evidence in favor of a causal link between music training and improved cognitive/academic skills”

• cites Tierney et al. (2015) sourcePartly supported

  The Behavioral results report greater gains in phonological awareness for the music group, but no group differences in phonological memory or rapid naming. The Discussion calls the behavioral gains small and notes that participants could not be randomly assigned, limiting causal inference. The study therefore supports the characterization of limited evidence for a literacy-related skill, but it did not directly establish improvements in academic skills more broadly. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

Nan et al.'s (2018) study corroborates the idea that the impact of music training on cognitive/academic skills is slim to null.

“Therefore, this study corroborates the idea that the impact of music training on cognitive/academic skills is slim to null.”

• cites Nan et al. (2018) sourcePartly supported

  The Discussion (page not provided) reports no effect of piano training on general cognitive measures: IQ, attention, and working memory improved similarly in all three groups. But the paper also reports improved word discrimination, described as relevant to early literacy, so it does not support a slim-to-null impact across cognitive and academic skills generally. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Barbaroux et al.'s (2019) study does not implement any type of controls.

“Barbaroux et al.’s (2019) study does not implement any type of controls”

• cites Barbaroux et al.’s (2019) sourcePartly supported

  The Methods (Data processing and statistical analyses) explicitly say that no active control group was tested, supporting that narrower criticism. But the same section says the study used age-normalized scores and pre–post comparisons, which the authors describe as similar to a passive control group. The claim overstates the absence of controls by saying there were none of any type. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

The random-effect model yielded very similar estimates: g = 0.176, SE = 0.037, p < .001, τ2 = 0.033.

“The random-effect (RE) model (with Cheung and Chan’s correction) yielded very similar estimates: g¯ = 0.176, SE = 0.037, p < .001, τ2 = 0.033.”

There was no difference between cognitive skills and academic achievement (literacy and mathematics).

“there was no difference between cognitive skills and academic achievement (literacy and mathematics)”

• cites Cooper’s (2019) sourceNot checked — no full text of the source

For non-randomized studies, the Bayes factor BFτ was 0.021.

“BFτ = 0.021 (τ nearly 50 times more likely to be null than positive)”

Randomization explained nearly all the true heterogeneity (τ2 = 0.005, I2 = 7.61%).

“and explained nearly all the true heterogeneity (τ2 = 0.005, I2 = 7.61%)”

For randomized samples, the Bayes factor BFτ was 0.025.

“The Bayes factors were BFg = 0.257 and BFτ = 0.025.”

Fifty-four studies, conducted from 1986 to 2019, met the inclusion criteria.

“We found 54 studies, conducted from 1986 to 2019, that met the inclusion criteria (reported in Appendix A in the Supplemental Online Materials).”

Nineteen of the included studies had never been included in any previous meta-analysis.

“Nineteen of these studies had never been included in any previous meta-analysis.”

For non-randomized studies, the overall effect size was g = 0.246, SE = 0.049, 95% CI [0.140; 0.352], p < .001.

“For non-randomized studies, the overall effect sizes were g¯ = 0.246, SE = 0.049, 95% CI [0.140; 0.352], p < .001”

For non-randomized studies, Vevea and Woods' selection model calculated a more conservative estimate (g = 0.126).

“Vevea and Woods’ (2005) selection model calculated a more conservative estimate (g¯ = 0.126).”

• cites Vevea and Woods’ (2005) sourceNot checked — no full text of the source

For randomized samples, Vevea and Woods' selection model yielded an estimate of g = -0.002.

“Vevea and Woods’ (2005) selection model yielded a similar estimate (g¯ = -0.002).”

• cites Vevea and Woods’ (2005) sourceNot checked — no full text of the source

Music training is ineffective regardless of the type of outcome measure (e.g., verbal, non-verbal, speed-related), participants' age, and duration of training.

“Interestingly, music training is ineffective regardless of the type of outcome measure (e.g., verbal, non-verbal, speed-related, etc.), participants’ age, and duration of training.”

In sensitivity analyses with non-active controls, the random-effect model provided similar results: g = 0.200, SE = 0.041, p < .001, τ2 = 0.024.

“The RE model provided similar results, g¯ = 0.200, SE = 0.041, p < .001, τ2 = 0.024.”

In studies implementing non-active controls, the overall effect size was g = 0.228, SE = 0.045, 95% CI [0.137; 0.320], m = 41, k = 144, df = 30.1, p < .001, τ2 = 0.042, I2 = 43.11%.

“In those studies that implemented non-active controls, the results showed a small and moderately heterogeneous overall effect of music training on cognitive and academic outcomes. The overall effect size was g¯ = 0.228, SE = 0.045, 95% CI [0.137; 0.320], m = 41, k = 144, df = 30.1, p < .001, τ2 = 0.042, I2 = 43.11%.”

In sensitivity analyses, the random-effect model provided similar estimates: g = 0.149, SE = 0.035, p < .001, τ2 = 0.024.

“The random-effect (RE) model provided similar estimates: g¯ = 0.149, SE = 0.035, p < .001, τ2 = 0.024.”

In non-active controls, the random-effect model provided similar results: g = 0.201, SE = 0.041, p < .001, τ2 = 0.023.

“The RE model provided similar results, g¯ = 0.201, SE = 0.041, p < .001, τ2 = 0.023.”

For non-randomized studies, the trim-and-fill analysis retrieved two missing studies with the L0 estimator, with adjusted estimate g = 0.211, 95% CI [0.095; 0.328].

“the trim-and-fill analysis retrieved two missing studies with the L0 estimator (adjusted estimates g¯ = 0.211, 95% CI [0.095; 0.328])”

For active controls, the Bayes factors largely favored the null hypothesis (BFg = 0.063 and BFτ = 0.006).

“The Bayes factors largely favored the null hypothesis (BFg = 0.063 and BFτ = 0.006).”

A small overall effect size is observed in studies employing neither active controls nor random assignment (g = 0.246).

“Conversely, a small overall effect size is observed in those studies employing neither active controls nor random assignment (g¯ = 0.246).”

In active controls, the random-effect model provided similar results: g = 0.090, SE = 0.060, p = .136, τ2 = 0.032.

“In fact, the RE model provided similar results, g¯ = 0.090, SE = 0.060, p = .136, τ2 = 0.032.”

The overall effect size of the RVE intercept model was g = 0.184, SE = 0.041, 95% CI [0.101; 0.268], m = 54, k = 254, df = 38.36, p < .001, τ2 = 0.041, I2 = 43.16%.

“The overall effect size of the RVE intercept model was g¯ = 0.184, SE = 0.041, 95% confidence interval (CI) [0.101; 0.268], m = 54, k = 254, df = 38.36, p < .001, τ2 = 0.041, I2 = 43.16%.”

Five effect sizes were found to be significantly inflating the true heterogeneity in the non-active-control sensitivity analysis.

“Five effect sizes were found to be significantly inflating the true heterogeneity.”

For randomized samples, the Bayes factor BFg was 0.257.

“The Bayes factors were BFg = 0.257 and BFτ = 0.025.”

For non-randomized studies, the Bayes factor BFg was 217.840.

“The Bayes factors were BFg = 217.840 (g¯ far more likely to be non-null than null)”

Randomization was the only statistically significant moderator in the moderator analysis (p = .042).

“Randomization was the only statistically significant moderator (p = .042)”

For non-active controls, Vevea and Woods' selection model calculated a similar estimate (g = 0.119).

“Vevea and Woods’ (2005) selection model calculated a similar estimate (g¯ = 0.119).”

• cites Vevea and Woods’ (2005) sourceNot checked — no full text of the source

Duration of training was not a significant moderator for hours, weeks, or sessions.

“Duration of training was not significant either (p = .266, p = .952, and p = .662 for hours, weeks, and sessions, respectively).”

Small statistically significant overall effects of music training on cognitive and academic outcomes are obtained only in studies implementing no random allocation of participants and employing non-active controls (g ≈ 0.200, p < .001).

“Small statistically significant overall effects are obtained only in those studies implementing no random allocation of participants and employing non-active controls (g¯ ≈ 0.200, p < .001).”

In sensitivity analyses excluding three studies with large baseline IQ differences and influential cases, the overall effect size was g = 0.166, SE = 0.041, 95% CI [0.083; 0.249], m = 51, k = 235, df = 34.9, p < .001, τ2 = 0.036, I2 = 40.62%.

“The overall effect size of the RVE intercept model was g¯ = 0.166, SE = 0.041, 95% CI [0.083; 0.249], m = 51, k = 235, df = 34.9, p < .001, τ2 = 0.036, I2 = 40.62%.”

In sensitivity analyses with non-active controls, the overall effect size was g = 0.226, SE = 0.045, 95% CI [0.133; 0.319], m = 40, k = 139, df = 29.2, p < .001, τ2 = 0.041, I2 = 42.96%.

“When non-active controls were used, the overall effect size was g¯ = 0.226, SE = 0.045, 95% CI [0.133; 0.319], m = 40, k = 139, df = 29.2, p < .001, τ2 = 0.041, I2 = 42.96%.”

For non-randomized studies, the trim-and-fill analysis retrieved three missing studies with the R0 estimator, with adjusted estimate g = 0.189, 95% CI [0.068; 0.310].

“Three missing studies were retrieved with the R0 estimator (adjusted estimates g¯ = 0.189, 95% CI [0.068; 0.310]).”

The RE model provided similar results for active controls, g = -0.010, SE = 0.035, p = .787, τ2 = 0.

“The RE model provided similar results, g¯ = -0.010, SE = 0.035, p = .787, τ2 = 0.”

For randomized studies, the overall effect size was g = 0.064, SE = 0.065, 95% CI [-0.116; 0.244], p = .381.

“for randomized studies, the relevant statistics were g¯ = 0.064, SE = 0.065, 95% CI [-0.116; 0.244], p = .381”

No other study features (e.g., age, duration of training, and outcome measure) seem to have any influence on the effect sizes.

“No other study features (e.g., age, duration of training, and outcome measure) seem to have any influence on the effect sizes – not even the outcome measures.”

The observed true between-study heterogeneity in studies employing non-active controls was accounted for by a few extreme effect sizes and the type of allocation of participants to the groups.

“Therefore, the observed true between-study heterogeneity in the studies employing non-active controls was accounted for by a few extreme effect sizes and the type of allocation of participants to the groups.”

When random allocation was employed, the overall effect was near-zero.

“Thus, when random allocation was employed, the overall effect was near-zero.”

When transfer occurs, it is often limited to the degree to which the source and target domains share contents.

“Moreover, when it does occur, transfer of skills is often limited to the degree to which the two domains (source and target) share contents.”

The low observed true heterogeneity in active controls (τ2 = 0.025, I2 = 23.10%) is very likely to be spurious.

“The latter finding confirms that the low observed true heterogeneity (τ2 = 0.025, I2 = 23.10%) is very likely to be spurious.”

Different values of within-study effect-size correlation (ρ) did not affect the results for active controls (g = -0.021, τ2 = 0).

“Different values of within-study effect-size correlation (ρ) did not affect the results (g¯ = -0.021, τ2 = 0).”

Music-trained twins do not possess a higher IQ than non-music-trained co-twins.

“Notably, music-trained twins do not possess a higher IQ than non-music-trained co-twins.”

• cites Mosing, Madison, Pedersen, and Ullén (2016) sourceNot checked — no full text of the source

Different values of within-study effect-size correlation (ρ) did not significantly affect the results.

“Different values of within-study effect-size correlation (ρ) did not significantly affect the results (g¯ range 0.184– 0.185, τ2 = 0.041).”

No thorough sensitivity analysis was performed in previous meta-analyses.

“Finally, no thorough sensitivity analysis was performed (e.g., outlier analysis and multiple publication bias analysis).”

Studies using active controls or implementing random assignment report homogeneous null or near-zero effects (g = -0.021-0.064, τ2 ≤ 0.005).

“Specifically, those studies using active controls or implementing random assignment report homogeneous null or near-zero effects (g¯ = -0.021–0.064, τ2 ≤ 0.005).”

Music training fails to produce solid improvements in all the examined cognitive and academic skills equally.

“Altogether, these results indicate that music training fails to produce solid improvements in all the examined cognitive and academic skills equally.”

Only a low amount of publication bias is observed in the models (about 0.100 standardized mean difference at most).

“Finally, only a low amount of publication bias is observed in the models (about 0.100 standardized mean difference at most), which is in line with the near-zero effect sizes estimated.”

The present meta-analysis has succeeded in explaining all the observed true heterogeneity.

“the present meta-analysis has succeeded in explaining all the observed true heterogeneity.”

There is an inverse relationship between studies' design quality and the magnitude of the effect sizes.

“The inspection of true heterogeneity shows that there is an inverse relationship between the studies’ design quality and magnitude of the effect sizes.”

Suboptimal modeling choices produce biased estimates.

“In brief, such suboptimal modeling choices produce biased estimates.”

The results were not affected by modeling choices (i.e., ρ values and procedure for modeling nested data).

“The results were not affected by modeling choices (i.e., ρ values and procedure for modeling nested data).”

Previous meta-analyses did not employ multilevel modeling.

“First, no multilevel modeling was employed.”

For randomized samples, all publication bias analyses estimated a substantially null overall effect.

“Regarding the randomized samples, all the publication bias analyses estimated a substantially null overall effect.”

A small amount of publication bias was still detected in the non-randomized studies.

“Thus, a small amount of publication bias was still detected.”

There is no need to postulate any explanatory mechanism in the absence of any genuine effect or between-study variability.

“In fact, there is no need to postulate any explanatory mechanism in the absence of any genuine effect or between-study variability.”

Some effect sizes in previous meta-analyses were incorrectly calculated because of a mistake in the reporting of results in Rickard et al. (2012).

“Also, some of the effect sizes were incorrectly calculated because of a mistake in the reporting of the results in one of the primary studies (Rickard et al., 2012; personal communication).”

• cites Rickard et al., 2012 sourceNot checked — no full text of the source

No between-study true heterogeneity was present in the data for non-randomized samples and passive controls.

“these results showed that no between-study true heterogeneity was present in the data.”

Schellenberg (2019) has shown how correlational evidence is often used by scholars to incorrectly infer causal relationships between engagement in music and non-music outcomes.

“For example, Schellenberg (2019) has shown how correlational evidence is often used by scholars to incorrectly infer causal relationships between engagement in music and non-music outcomes.”

• cites Schellenberg (2019)Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Recent correlational studies have confirmed that music engagement is not associated with domain-general cognitive skills or academic performance.

“Moreover, recent correlational studies have confirmed that music engagement is not associated with domain-general cognitive skills or academic performance.”

The correlation between amount of music training and reading ability in adults disappears when domain-general cognitive skills are controlled for.

“Swaminathan, Schellenberg, and Venkatesan (2018) demonstrate that the correlation between amount of music training and reading ability in adults disappears when domain-general cognitive skills are controlled for.”

• cites Swaminathan, Schellenberg, and Venkatesan (2018) sourceNot checked — no full text of the source

There is no reason to support the hypothesis that music training boosts cognition or academic skills.

“Therefore, there is no reason to support the hypothesis that music training boosts cognition or academic skills.”

Once pre-existing differences in overall cognitive function are ruled out, the correlation disappears.

“Once pre-existing differences in overall cognitive function are ruled out, the correlation disappears (Swaminathan & Schellenberg, 2019).”

• cites Swaminathan & Schellenberg, 2019 sourceNot checked — no full text of the source

The impact of music training on cognitive and academic skills is null.

“Rather, all the evidence points toward the opposite conclusion, that is, that the impact of music training on cognitive and academic skills is null.”

Music aptitude, rather than the amount of music training, predicts fluid intelligence in a sample of adults.

“Swaminathan, Schellenberg, and Khalil (2017) show that music aptitude, rather than the amount of music training, predicts fluid intelligence in a sample of adults.”

• cites Swaminathan, Schellenberg, and Khalil (2017) sourceNot checked — no full text of the source

The observed correlation between music training and particular domain-general cognitive/academic skills is a byproduct of previous abilities.

“These findings corroborate the hypothesis according to which the observed correlation between music training and particular domain-general cognitive/academic skills is a byproduct of previous abilities.”