The Evidence on Teaching

A Meta-Analysis of Pair-Programming in Computer Programming Courses

Umapathy, K., & Ritzhaupt, A. D. · 2017

grade Cmeta-analysisindependentunclear
Sample
18 manuscripts, 28 independent effect sizes, N = 3,308 students
Population
Undergraduate computing courses, mostly US.
Design
Aggregates a mix of randomised and non-randomised course-level comparisons. The outcomes are course-graded assignments and exams — and on assignments the work is produced BY THE PAIR, which mechanically favours the pair condition and makes that estimate the least trustworthy of the three. Heterogeneity is high; the authors report effects varying significantly across studies. Read at abstract depth, so the per-domain estimates and confidence intervals were not obtained.
Key findings
Pair programming beat solo programming on programming assignments, exams and passing rates, with effect sizes in the 0.41-0.64 range, and showed no advantage on affective measures. The exam and pass-rate effects are the informative ones because they are individually produced; the assignment effect is partly an artefact of joint submission.
Genetic confound
Low-to-medium. Some pooled studies randomise pairing; others compare self-selected sections.
Replication notes
The pooled primaries are largely small single-institution course studies, and no registered replication of the pooled estimate was located.
DOI / URL
10.1145/2996201

Effects

OutcomeMetricValueMeasureTimingVsHorizonClass
Programming assignments, exams and passing rateseffect size (mixed d/g)0.41-0.64 across the three domainsresearcher-designedend of courseactive-alternativeend-of-treatmentdomain-skill
Affective measures (satisfaction, confidence)effect sizeno advantage for pair programmingresearcher-designedend of courseactive-alternativeend-of-treatmentnon-cognitive

Cited by