Home · C30
C30 killed
Music lessons have not been shown to make children reliably better at thinking skills or school subjects outside music: compared with an active control group the gain is at most small — about zero in the largest meta-analysis's main analysis (g = 0.056; Sala & Gobet 2020), about 0.2 standard deviations when the comparisons in which the control group trained the very skill being tested are dropped (pooled with passive-control studies; Bigand & Tillmann 2022), and 0.28 against active controls for instrumental programmes, with a confidence interval that includes zero (Román-Caballero et al. 2022).
- Status
- killed
- Last change
- 2026-11-05
- Serves
- Q1 (How do humans actually learn, and what helps them learn best?)
- Ages
- foralder-tidig-barndom, grundskola-lag
- Area
- Music and arts training
- Relates to
- C29 relates-to C30 (training one skill does not lift unrelated school results); C30 relates-to C35 (C35 is the narrower restatement after C30 was killed)
- Bounds
- Music training for children, non-music outcomes (verbal, non-verbal, memory, literacy, maths). Evidence on file: Sala & Gobet 2020 (Memory & Cognition 48, DOI 10.3758/s13421-020-01060-2; full text): 54 studies, 254 effects; active controls g = 0.056 (95% CI −0.069 to 0.182), −0.021 in sensitivity analysis; randomised studies with non-active controls g = 0.064; no difference between cognitive and academic outcomes. Challenge already on file: Bigand & Tillmann 2022 (Memory & Cognition, DOI 10.3758/s13421-021-01226-6; accepted manuscript): reanalysing the same data and dropping 21 comparisons where the control trained the very skill tested gave g = 0.208 — the first attack to weigh. Weakened 2026-10-31 (nightly attacks): Bigand & Tillmann 2022 re-read in the published version (Memory & Cognition 50(2), 339–347): the 21 removed comparisons were 18 phonological-training controls tested on phonology, 2 reading-intervention controls tested on reading and 1 visual-art control tested on visual form; the remaining 51 studies (233 effects) gave g = 0.208 (RVE, 95% CI 0.121 to 0.295; RE 0.195), pooled over active and passive controls, with type of control no longer a significant moderator (p = .193); without post-test-only studies g = 0.243. Román-Caballero, Vadillo, Trainor & Lupiáñez 2022 (Educational Research Review 35, 100436, DOI 10.1016/j.edurev.2022.100436; full text): instrumental programmes, 34 samples, 5,998 children, gΔ = 0.26 (0.12 to 0.39); randomised 0.26; active controls 0.28 (−0.06 to 0.61), randomised with active controls 0.23 (−0.13 to 0.58); one bias correction (RVE PET) cut the overall estimate to 0.01. Weakened 2026-11-05 (nightly attacks): Jamey, Foster, Hyde & Dalla Bella 2024 (Cognition 252, 105913, DOI 10.1016/j.cognition.2024.105913; full text, authors' copy): 22 longitudinal studies with controls, 1,734 children aged 3–11 (mean about 5 years 8 months); the 8 trials randomised at participant level with active non-music control training gave g = 0.60 (SE 0.11, 95% CI 0.39 to 0.82; I² = 0%; robust to rho, 0.55–0.60) on inhibition (flanker, go/no-go, Stroop or preschool versions); all 22 studies g = 0.31 (0.19 to 0.44); Egger p > .33. Limits: one inhibition measure per study, picked by the authors as the most suitable; post-tests right after short programmes; no blinding, every study rated 'moderate' quality; the authors note that playing music itself uses inhibition, so near and far transfer blur. Killed 2026-11-05 (Jonatan's decision): inhibitory control is a non-music cognitive result, and the 8 active-controlled randomised trials in Jamey et al. 2024 put it at g = 0.60 with an interval excluding zero, so the kill line below was met as written. The night run's rewrite of the claim and kill line (naming inhibitory control as an exception) was withdrawn; this is the wording that was attacked. The narrower idea that survives is now its own claim, C35.
- What would kill this
- A large randomised trial with an active control group, or a meta-analysis of such trials, finds that music training raises children's non-music cognitive or academic results by about 0.2 standard deviations or more, with a confidence interval that excludes zero in the active-control comparisons themselves (sharpened 2026-10-31 with the weaken: Bigand & Tillmann's 0.208 pools active and passive controls, and Román-Caballero et al.'s active-control 0.28 has an interval from −0.06 to 0.61). Promoted from K49 on 2026-10-29 in the bold expansion Jonatan approved after four empty attack nights (26–29 Oct): new sharp claims in thin areas, each naming a comparison, a rough size or direction and a population. Met 2026-11-05: Jamey et al. 2024 (Cognition, DOI 10.1016/j.cognition.2024.105913), 8 participant-randomised trials with active non-music controls, inhibitory control g = 0.60 (95% CI 0.39 to 0.82).
- Tutor consequence
- none yet
Attack record
2 attacks: 1 narrowed, 1 killed
2026-10-31 narrowed
- Attack
- The "about zero against an active control" result depends on counting comparisons in which the control group trained the very skill being tested (for example phonological training tested on phonological awareness); re-running Sala & Gobet's own data and code without those 21 comparisons gives g = 0.208 (95% CI 0.121 to 0.295) with type of control no longer a significant moderator, and a separate meta-analysis of instrumental programmes puts the active-control estimate at 0.28 (95% CI −0.06 to 0.61).
- Target
- claim-text
- Source
- Bigand & Tillmann 2022 | Memory & Cognition 50(2), 339–347, DOI 10.3758/s13421-021-01226-6 (published version read 2026-10-31 from the authors' lab site; Crossref-checked); supporting: Román-Caballero, Vadillo, Trainor & Lupiáñez 2022 | Educational Research Review 35, 100436, DOI 10.1016/j.edurev.2022.100436 (full text, Trainor lab copy; Crossref-checked)
- Steelmans to
- an active control is a fair yardstick for far transfer only if the control is not itself being trained for near transfer on the outcome test; once those unbalanced comparisons are removed, the pooled effect sits at the claim's own 0.2 kill threshold and the active/passive difference that carried the "about zero" conclusion disappears, while the other main meta-analysis also finds about 0.25–0.28; so "about zero (below 0.1)" overstates what is known.
- Why
- Nightly-attacks 2026-10-31 Europe/Stockholm. Bigand & Tillmann (published version): removed 18 phonological-training controls tested on phonology, 2 reading-intervention controls tested on reading and 1 visual-art control tested on visual form (p. 343); remaining m = 51, k = 233: RVE g = 0.208 (SE .043, 95% CI 0.121 to 0.295), RE 0.195; moderators randomisation (p = .610) and type of control (p = .193) not significant; without post-test-only studies 0.243; with influential cases also removed 0.234 (Table 3); the 21 removed comparisons themselves gave −0.126 (ns): music did about as well as near-transfer training. Román-Caballero et al.: 34 samples, 5,998 children, gΔ = 0.26 (0.12 to 0.39); active controls 0.28 (−0.06 to 0.61), passive 0.25, no difference (p = .983); randomised with active controls 0.23 (−0.13 to 0.58); RVE PET corrected the overall estimate to 0.01, other bias checks did not (Table 4). Kill withheld: Bigand & Tillmann's 0.208 pools active and passive controls (the active-only estimate is in their supplement, not read), it reuses the same studies, and Román-Caballero et al.'s active-control interval includes zero; no estimate restricted to active controls is reliably at or above 0.2. Not a duplicate (empty ledger before tonight); the Bounds named Bigand & Tillmann as "the first attack to weigh" but the claim text had not been tested against it, so this is not a restatement of a carve-out. Not a preprint. Status alive → weakened. Claim field, Bounds and kill line rewritten in the same turn; brief and diagram core rewritten (C30 has no front-page rule).
- Rewrite
- claim-field+brief
2026-11-05 killed
- Attack
- A meta-analysis restricted to participant-randomised trials with active non-music control training found that music training improved children's inhibitory control by g = 0.60 (95% CI 0.39 to 0.82; 8 trials, I² = 0%), so for at least one non-music thinking skill the active-controlled gain is above 0.2 with an interval that excludes zero — which C30's kill line, as worded, asked for.
- Target
- kill-condition
- Source
- Jamey, Foster, Hyde & Dalla Bella 2024 | Cognition 252, 105913, DOI 10.1016/j.cognition.2024.105913 (full text read 2026-11-05 in the authors' copy of the published article on the Dalla Bella lab site; matches PubMed abstract and Crossref; bioRxiv v3 10.1101/2023.02.08.527718 gives the same numbers)
- Steelmans to
- the earlier meta-analyses pooled many different outcomes, which can hide a real effect on the specific skill music practice exercises; focusing on one theory-driven executive function, in randomised trials with active controls, gives a homogeneous (I² = 0%) medium effect with no funnel asymmetry (Egger p > .33), and in the full set active-control studies even showed larger effects than passive ones — the opposite of an expectancy artefact.
- Why
- Nightly-attacks 2026-11-05 Europe/Stockholm. Full text §3.1.1–3.1.5 and Table 2: 22 longitudinal studies, 1,734 children aged 3–11; 8 RCTs (participant-level randomisation, active control) g = 0.60 (SE 0.11), robust 0.55–0.60 across rho; all 22 studies 0.31 (0.19 to 0.44). Kill withheld: one executive function, with one measure per study picked by the authors as "most suitable"; immediate post-tests after short programmes; no blinding, all studies "moderate" quality; the authors themselves say inhibition is used when playing music, blurring near and far transfer; no academic or broad cognitive outcome in this paper. C30's general statement (school subjects, thinking skills in general) is untouched by it. Not a duplicate (ledger had Bigand & Tillmann / Román-Caballero only); not a restatement of a Bound (Bounds named no executive-function result). Not a preprint (published, peer reviewed). Status stays weakened. Because the sentence would pass about 600 characters, it was rewritten short in the same turn under the short-claim rule: the old size statements moved to Bounds unchanged, and the kill line now names academic results or a broad cognitive measure (not a single executive function chosen per study); flagged for Jonatan. Brief and diagram core rewritten (C30 has no front-page rule).
- Rewrite
- claim-field+brief
Linked sources
4 linked sources (1 supports, 2 limits, 1 challenges)
Supports (1)
Limits (2)
Challenges (1)
Review state: first-pass draft, pending human review.