Learning Styles, Preferences, or Strategies? An Explanation for the Resurgence of Styles Across Many Meta-analyses

17 meta-analyses in the Visible Learning MetaX database, covering ~702 studies, 2,383 effects and ~105,000 students (about 10-15% study overlap between the meta-analyses). Split into 4 matching meta-analyses (143 studies, 401 effects, ~22,172 people) and 13 correlational meta-analyses (559 studies, 1,982 effects, ~82,853 people).

Mixed ages across the learning-styles meta-analytic literature

A second-order review that separates the two things the literature keeps conflating: studies testing the matching hypothesis, and correlational studies relating style scores to outcomes. Included here despite this archive's standing rule against citing Visible Learning averages as evidence - the value is the decomposition, not the aggregate, and the authors are hostile witnesses for a null verdict. Two limits on how far this can be pushed. First, Hattie owns and curates the MetaX database being analysed, so the source pool is not independently assembled; the paper reports no systematic search, no inclusion protocol and no quality coding of the 17 meta-analyses. Second, the headline matching estimate is IMPRECISE: d = 0.04 is reported with se = 0.21, i.e. an interval comfortably spanning zero in both directions, and it is an unweighted average across only four meta-analyses whose own estimates range from about -0.03 to +0.22. The authors are blunt that the underlying literature is bad in both streams - "we have coded and pored over the details in 2,500 meta-analyses and there is not another influence that is marked by such poor quality" - which cuts against treating either number as precise.

Across 17 meta-analyses, studies that actually test the matching hypothesis yield d = .04 (se = .21; 4 meta-analyses, 143 studies, 401 effects), while correlational studies average r = .24 (d = .50; 13 meta-analyses). The naive average across ALL 17 is d = 0.40 - exactly the MetaX grand mean across 400 influences - and this is the number the paper opens with as the thing to be explained, not endorsed. The four matching meta-analyses are not uniformly zero and the spread should be recorded: Garlinger & Frank (1986) d = -0.03 (field-dependent) / +0.04 (field-independent) / +0.03 combined; Kavale & Forness (1987) d = 0.13 overall (auditory 0.14, visual 0.09, kinesthetic 0.18) with 35% of effects negative and only a 6-percentile-rank gain; Phillips et al. (2016) r = 0.11 (d = 0.22) when reflective thinking matched the task versus r = -0.09 (d = -0.18) for intuitive; Peterson & Meissel (2015) "no clear academic achievement advantage". The authors attribute the resurgence of positive-looking meta-analytic results to correlational work that conflates learning STYLES with learning STRATEGIES, document specific errors in the most-cited correlational meta-analysis (Dunn et al. 1995 includes a reported d = 22.29 implying r = 0.996, and misreads a multiple R of 0.87 from LaShell 1986 as a bivariate correlation), and recommend shifting effort to teaching adaptable strategies matched to task complexity. They also record the belief side: Newton & Salvi (2020), 37 studies, 15,405 educators, 18 countries, 89% agreement and no decline over a decade, with pre-service teachers (95%) higher than qualified teachers (88%).

High for the r = .24 correlational stream, which is exactly the point: self-reported style correlates with outcomes because it proxies ability and study behaviour. Low for the d = .04 matching stream.

A commentary-style second-order review ("Reflection on the Field") that re-presents existing meta-analytic averages rather than producing an original pooled estimate, so replication of its own finding is not a meaningful question. The decomposition it reports is consistent with Aslaksen & Loras (2018) and with the crossover null in Clinton-Lisell & Litzinger (2024).