Working memory training in typically developing children: A multilevel meta-analysis

m = 41 studies, k = 393 effect sizes, N = 2,375 children

Typically developing children (no learning disability or clinical condition), studies from 2007 to 2019. Sample mean age 8.63 years (median 8.69, quartiles 6.00 and 9.85, range of study mean ages 4.27-15.40); the authors caution that adolescents are thinly represented, so results generalise less well to ages 12-16.

The most directly on-scope meta-analysis in the brain-training literature: typically developing children only, robust variance estimation with hierarchical weights, estimating both within-study (omega-squared) and between-study (tau-squared) true heterogeneity, with active vs nonactive control coded as a moderator. Transfer distance is coded on Noack et al.'s taxonomy - "nearest" (same or near-identical memory task) vs "less near" (different task, same Gsm construct) vs far (everything else). Publication bias handled three ways: influential-case analysis, trim-and-fill (L0 and R0), and Vevea & Woods selection models. Grade caveat for anyone recomputing: inclusion required a control group not doing adaptive WM training and a pre-post design with non-self-selected participants, but did NOT require explicit randomisation - so the pool is stronger than typical but not uniformly RCT. The load-bearing contrast is nevertheless within the experimental literature (active vs nonactive control).

Near transfer is real and scales with overlap; far transfer is zero once placebo is controlled. Near transfer overall g = 0.389, decomposing into nearest transfer g = 0.468 and less-near transfer g = 0.261 - a dose-of-overlap gradient, not a capacity gain, and the gradient survives publication-bias correction (trim-and-fill drops nearest transfer to 0.356; the selection model drops less-near transfer to 0.196). Far transfer to fluid intelligence, processing speed, mathematics and language is g = 0.092 overall, and that small significant pooled value is entirely a control-type artefact: g = 0.139 against nonactive controls (0.097-0.116 after bias correction) versus g = 0.032 against active controls (RVE 95% CI -0.073 to 0.138, p = .517; the random-effects estimate quoted in the abstract is 0.001, SE 0.055, p = .982, tau-squared = 0.000, and the selection model returns -0.002). No significant differences across far-transfer outcome types (all pairwise p = 1.000) or training-task types, and age was not a moderator. All models showed null or low true heterogeneity, fully explained by control type and statistical artefacts - which forecloses the "it works for some programme or subgroup" escape and contradicts the claim that the field has produced mixed results. At follow-up both effects are non-significant: far transfer g = 0.006 and near transfer g = 0.239 (RVE p = .059). The authors conclude there is no reason to keep investing in WM training research with typically developing children.

Low. Randomised or pre-post controlled contrasts; the informative comparison (active vs nonactive control) is within the experimental literature.