A meta-analysis of teaching and learning computer programming: Effective instructional approaches and conditions

139 interventions, 375 effect sizes

K-12 and higher education programming learners, international; ages not separable at abstract depth.

The companion to the same team's 2019 transfer meta, and the anchor source for the domain-skill leg of this subject. Three separate syntheses — programming interventions per se, visualisation and physicality, and dominant instructional approaches. The "per se" arm pools mostly small pre-post and weak quasi-experimental contrasts with treatment-aligned outcome measures, which is why g = 0.81 sits well above anything a controlled contrast produces, and the confidence interval is enormous (0.42 to 1.21). Read at ABSTRACT depth: the article is CC-BY licensed but every open route failed, so the moderator tables and the measure-type breakdown are unverified. That is real debt on the topic's anchor.

Programming interventions per se g = 0.81 (95% CI 0.42-1.21). Visualisation g = 0.44 (0.29- 0.58); physicality g = 0.72 (0.23-1.21); dominant instructional approaches g = 0.49-1.02. The moderator result is the real story and it is deflationary: effect sizes differed only MARGINALLY between the instructional approaches and conditions. Standouts were collaboration within metacognition instruction, problem-solving instruction outside regular lessons, short-term physicality, and Scratch-based visualisation.

Medium. The pool is dominated by quasi-experimental and pre-post designs, so selection into treatment is present in much of the base. It is not the main threat — measure alignment is.

The magnitude is echoed by independent teams (Hu 2024 g = 0.62; Simonsmeier 2025 d = 0.73), but the specific approach-contrasts it reports are contradicted wherever anyone has tested them head to head — most importantly Xu et al. (2019), whose block-versus-text pooled estimate is a null with detected publication bias.