The relationship between cognitive ability and chess skill: a comprehensive meta-analysis

Alexander P. Burgoyne and Giovanni Sala (co-first authors), Fernand Gobet, Brooke N. Macnamara, Guillermo Campitelli, David Z. Hambrick. Intelligence 59:72–83, doi 10.1016/j.intell.2016.08.002. Read as the author's accepted manuscript from LSE Research Online (eprint 102241, CC BY-NC-ND). The published version may differ. A 2018 corrigendum exists and was not read (ScienceDirect, not open). Read by researcher R6 (capability), 6 October 2026. Provenance: papers/capability/burgoyne2016_cognitive_ability_chess.provenance.json.

What was read

• Read in full: every line of the 58-page manuscript: abstract, introduction, method, results, moderator analyses, publication bias, additional analyses, general discussion, references, Appendix A (forest-plot data for every model), Tables S1a–S1h and the funnel-plot captions.
• Not read: the figures as images (bar charts and forest plots); their values are in the text and the appendix numbers.

What they did

• Question: does cognitive ability correlate with chess skill? Is the link weaker at high skill (Ericsson's "circumvention" claim) or in adults?
• Studies: 2,287 records searched up to March 2016 → 19 studies, 26 independent samples, 82 effect sizes, N = 1,779.
  • Ability was organised by the Cattell–Horn–Carroll model: fluid reasoning (Gf), comprehension–knowledge (Gc), short-term/working memory (Gsm), processing speed (Gs), and full-scale IQ. Content was also coded as visuospatial, numerical or verbal.
  • Chess skill: Elo rating (ranked samples) or a chess test (unranked samples).
  • Random-effects models, moderator tests on Gf, trim-and-fill for publication bias.
• Samples:
  • Elo (7 studies): weighted mean 2,018 (SD 177), range 1,311–2,607.
  • Full-scale IQ (5 studies): weighted mean 120.5, mean SD 14.8.

Main results (verified)

• All four broad abilities correlate with chess skill, modestly:

  Ability	r̄	95% CI	Variance explained
  Gf	.24	[.18, .30]	6%
  Gc	.22	[.11, .32]	5%
  Gsm	.25	[.13, .37]	6%
  Gs	.24	[.08, .39]	6%
  Average (≈ g)	.24	[.19, .28]	6%

  • Full-scale IQ: r̄ = .10 [−.19, .38], n.s., from 6 effect sizes. This was driven by one elite subsample (Bilalić et al. 2007, r = −.51); without it, .24 (p = .015).

• Moderators of Gf:

  • Skill: ranked samples r̄ = .14 (2%) against unranked .32 (10%); Q(1) = 8.37, p = .004.
  • Age: adults .11 (1%) against youth .32 (10%); Q(1) = 9.83, p = .002.
  • Mean rating ≥ 2000 against < 2000: −.10 against .10, n.s.
  • Ranked adults .11 against ranked youth .27, n.s.

• By content: numerical ability r̄ = .35 (12%), verbal .19 (3%), visuospatial .13 (2%).

  • Visuospatial: ranked .05 against unranked .25; adults .03 against youth .24.

• Publication bias: asymmetry on both sides across models, "little evidence" of systematic suppression.

• Cited evidence:

  • The chance of becoming an International Master or Grandmaster was .24 when starting chess at age 12 or younger, against .02 when starting later (Gobet & Campitelli 2007).
  • General mental ability's validity for job performance does not fall with experience (Schmidt & Hunter).
  • Ericsson (2014) claims acquired mechanisms "circumvent the role of any basic general cognitive capacities … at the expert level".
  • The authors suggest circumvention may depend on the task: possible in chess, not in dynamic tasks such as sight-reading.

Limits (the authors')

• Range restriction: the SD of Elo in ranked samples was 148, against 200 for the rating system.
• Confounds: skill level is confounded with age (all adult samples were ranked, most youth samples unranked) and with the skill measure (ranked always rated, unranked always tested). The weaker correlation at high skill cannot be separated from age or measurement.
• No correction for unreliability. The authors estimate it is small: .10 → .11 at reliability .90.
• Small numbers of effect sizes for Gc, Gsm, Gs and IQ. No moderator tests were possible there. One corrigendum is unread.

What it means for Kurisutina (inference)

• General capacity explains little of the differences among skilled adults. Gf explains about 1–2% of skill variance in ranked or adult samples; content (chunks, templates, practice; see gobet2000_five_seconds_or_sixty.md, macnamara2014_practice_meta.md) explains much more. In a replica, an expert's level is carried by the slot's content, not by a capacity setting.
• Capacity still matters for learning and among novices (youth and unranked samples: about 10%).
  • It also matters where the task leaves no room to circumvent it (cited: sight-reading).
  • And it matters through selection: ranked samples had a mean IQ of about 120 with a population-like SD. Chess may draw from the upper range already.
• So a capacity setting is needed mainly for:
  • the person's learning rate (how fast new content forms);
  • performance on novel problems outside their trained domains.
  • It is not the lever for their in-domain level. That fits a design where the base's capacity range is sized to reach the most capable person, while per-person level comes from content.
• "How far general ability still predicts at high skill": r ≈ .1 in this meta-analysis, possibly attenuated by range restriction. It is not zero, but it is small.