When the music's over. Does music skill transfer to children's and young adolescents' cognitive and academic skills? A meta-analysis

38 studies (1986-2016), 40 independent samples, 118 effect sizes, N = 3,085

Children and adolescents aged 3-16 with no prior formal music training and no learning disability or clinical condition. Integrated curricula (music fused with reading instruction) were excluded.

The earlier and smaller of the two Sala & Gobet syntheses; superseded in coverage by their 2020 multilevel meta-analysis but retained because it is the only one of the pair that publishes per-domain pooled effect sizes. Four design moderators were tested jointly and explained 65% of between-study variance (R2 = 0.65), which is the paper's central claim: what predicts a music-transfer effect is the weakness of the design, not the type of outcome or the age of the child.

Overall d = 0.16 [0.09, 0.22], k = 118. The design gradient is monotonic: randomised d = 0.09 [-0.01, 0.18] versus non-randomised 0.23 [0.14, 0.31]; active control 0.03 [-0.07, 0.12] versus do-nothing control 0.25 [0.17, 0.34]; and the cell that is BOTH randomised and active-controlled gives d = -0.12 [-0.27, 0.03], k = 22, while the cell that is neither gives 0.33 [0.23, 0.44]. Age was not a moderator (p = .944). By domain, the two significant cells are memory d = 0.34 [0.20, 0.48] and intelligence d = 0.35 [0.21, 0.49]; literacy is -0.07 [-0.23, 0.09] and mathematics 0.17 [-0.02, 0.36], both null. Phonological processing 0.17 [0.04, 0.29] is significant but small. No publication bias was detected (Begg & Mazumdar p = .433; trim-and-fill removed nothing in any of seven models), so the small pooled positive is not obviously a selective-reporting artefact. The authors are careful about what they are entitled to claim: "Due to the lack of well-designed studies, the question of whether music training enhances children's and young adolescents' intelligence- and memory-related skills is still unanswered."

Addressed by design stratification rather than genetic modelling. The randomised and active-controlled subgroups strip selection out, and the effect goes with it - the same signature the archive finds in the motor-competence and exercise-cognition literatures.

The design-quality gradient reported here did NOT reproduce in the authors' own 2020 multilevel meta-analysis, where randomisation was non-significant in the main model (p = .518), nor in Román-Caballero et al. (2022), where randomised studies gave numerically LARGER effects than self-selection studies. The active-control moderator is the one finding that reproduces across all three syntheses. One further caution the authors flag themselves: the intelligence and memory effects (d ~ 0.35) are dismissed on the ground that only one study (Schellenberg 2004) used randomisation plus an active control for IQ and none did for memory - an argument from absence of evidence.