The cognitive benefits of learning computer programming: A meta-analysis of transfer effects.

105 studies, 539 effect sizes, N = 9,139 (4,544 treatment, 4,595 control). Mean study n = 87 (SD 72, median 66, range 14-416). Interventions averaged 25 hours (range 2-120, median 20)

Mostly primary and secondary students in Asia and North America; the corpus is dominated by publications from the 1980s and 1990s Logo era, with a second cluster in the 2010s.

The best paper in this literature, and it contains its own refutation. A three-level random-effects meta of experimental and quasi-experimental studies with a control group; pre-post control designs predominate but groups were mostly not matched, so most are quasi-experiments. About 70% of effect sizes came from UNTREATED (passive) control groups. Two structural facts about the corpus deserve to be read out loud: 87.6% of studies measured far transfer ONLY — they never checked whether the children learned to program — and NO study in 105 reported a delayed follow-up, so the maximum measured horizon in the entire corpus is the end of treatment. The authors' own recommended "optimal design" (measure programming skill and cognitive skill, with treated and untreated controls, plus a follow-up) describes a study nobody has run.

Headline overall transfer g = 0.49 (95% CI 0.37-0.61); near transfer g = 0.75 (0.39-1.11, from only 13 studies and 19 effect sizes); far transfer g = 0.47 (0.35-0.59). The decisive moderator collapses the headline: far transfer against TREATED (active) control groups is g = 0.15 against g = 0.64 for untreated controls, Q_M(1) = 40.12, p < .001. Publication bias is present — Egger t(537) = 4.10, p < .001; trim-and-fill 0.43; published g = 0.58 against grey literature 0.43. After removing ten influential effect sizes (g = 2.10 to 8.63), standardized tests give g = 0.33 against unstandardized 0.49 (p = .03), and the school-achievement effect drops from 0.28 to 0.22 and becomes non-significant. Literacy is flat at g = -0.02 (-0.12 to 0.08). Near transfer with random assignment is g = 0.29 against 0.95 without. Every correction moves the same direction.

Medium for the pooled estimate — most inputs are quasi-experiments where selection into treatment is possible. Low for the moderator that matters: active-versus-passive control is an experimental feature, not a family characteristic.

Its two significant moderators — publication status and control-group treatment — reproduce Liao & Bright (1991) exactly, 28 years apart. Its headline far-transfer estimate has not survived the randomised trials that followed: two using validated computational-thinking instruments returned d ≈ 0.00, and the largest randomised test against a content-matched mathematics alternative returned negative effects.