Near and far transfer: Is music special?

Reanalysis of Sala & Gobet's (2020) own dataset - 54 studies, 254 effect sizes, N = 6,984 - using their published OSF R script

Same as the reanalysed meta-analysis; children aged 3-16

The strongest adversarial check available on the archive's preferred deflationary source, and it deserves weight because it is the good kind of critique: the authors downloaded Sala & Gobet's data and script from OSF and re-ran them, so the disagreement cannot be about the numbers. Graded C - it is a reanalysis-commentary, not new data, and the authors are music-cognition researchers arguing for their own field, which is the direction of bias to expect.

Three specific charges. (1) The randomisation moderator is fragile: it was non-significant in Sala & Gobet's main model (p = .518) and became significant (p = .042) only at a sensitivity step where - per lines 737-739 of their own script - the analysis switched from one multi-moderator model to three separate single-moderator models. Re-specified consistently it gives p = .08. The influential-case procedure also flags nine effect sizes rather than the five removed; dropping all nine leaves g = 0.203, p < .0001, with tau-squared = 0 and randomisation non-significant (p = .194). (2) The active-control comparison is not equidistant: music-related outcomes were stripped from the music groups but the equivalent was not done for control groups, so a phonologically trained control tested on phonological awareness is being scored on NEAR transfer while the music group is scored on FAR transfer using the same test - which, as the authors put it, "leads to underestimation of the effect of musical training." (3) Post-test-only studies were treated as having zero baseline difference. Re-running with those 21 mismatched effect sizes removed and post-test-only studies excluded gives g = 0.234 [0.141, 0.327] with tau-squared = 0.017, and neither randomisation nor control type is a significant moderator. Their most pointed sub-analysis: on the 21 excluded effect sizes, music's FAR transfer versus language training's NEAR transfer on the same tests is g = -0.126 [-0.350, 0.099], p = .17 - statistically indistinguishable.

Not addressed. The reanalysis inherits whatever selection remains in the underlying studies; the disagreement is about effect-size coding, not about confounding.

Sala and Gobet have published no reply. A search of all 75 papers citing this one found none authored by either, and their 2023 restatement (Gobet & Sala, Perspectives on Psychological Science) does not mention Bigand & Tillmann or Román-Caballero. That restatement gives second-order figures for music in typically developing children of g = 0.19 naive and g = -0.02 under active controls with bias correction, i.e. the position is unchanged. The near/far coding charge therefore stands unanswered rather than refuted, and readers should treat it as an open methodological dispute.