Genes Influence Young Children’s Human Figure Drawings and Their Association With Intelligence a Decade Later

7,752 twin pairs (drawing at age 4 in 7,437 pairs; g at 4 in 7,231 pairs; g at 14 in 2,348 pairs)

Twins Early Development Study, England and Wales births 1994-96

Classical twin design on the Draw-a-Child test (McCarthy 1972, 0-12 points for depicting head, eyes, nose, mouth, ears, hair, body, arms, legs, hands, feet and clothing; interrater reliability .93, alpha .79). The only genetically informative study of drawing ability located. It says nothing about instruction, which is precisely why it belongs here: it establishes that the trait an art curriculum would be trying to move has a heritable component and is genetically entangled with general ability a decade later.

Drawing at age 4 is 29% heritable [.22, .35], almost exactly matching g at age 4 (.29 [.27, .32]); g at 14 is 50% heritable [.38, .61]. Drawing at 4 correlates .33 with g at 4 and .20 with g at 14, and the GENETIC correlation between drawing at 4 and g at 14 is .52 [.31, .75]. The authors' summary: "Individual differences in this widespread behavior have an important genetic component and a significant genetic link with g", and they note being "intrigued to find that scores for a single drawing were as heritable as was g estimated from several different indicators." The shared-environment component for drawing is non-trivial - reported around .26 - which is where any family or teaching input would sit, though the study does not measure instruction.

This source is the confound quantified for visual art. Any cross-sectional comparison of children who do and do not take art classes is contaminated by a trait that is 29% heritable at age 4 and genetically correlated with later general ability at .52.

No replication of the drawing-g genetic correlation located. The heritability estimate is a single-cohort UK finding; the variance components reported in secondary extractions do not sum cleanly to 1.0 (A = .36, C = .26, E = .51 appear to mix trivariate and univariate model outputs), so those specific components should be verified against the paper's own tables before being relied on.