The Role of Homework in Student Learning Outcomes: Evidence from a Field Experiment

roughly 400-419 students across four sections of introductory economics

University undergraduates at East Carolina University - NOT school-age children

Journal of Economic Education 44(2), 93-109; full text read from a working-paper version marked 'preliminary draft'. Sections were randomised by coin flip to a homework-required arm (n = 201) or a homework-not-required arm. Randomisation checks pass on SAT scores and high school GPA (means 0.04 apart). Compliance separated the arms sharply: over 90% of the treatment group submitted the first assignment against about 14% of the control, and by the final exam under 75% of treatment students versus under 3% of controls had submitted every assignment. THE POPULATION IS THE LIMITATION - these are university students, so the effect size does not transfer to school-age children. The reason it belongs in this topic is the decomposition it permits, which no K-12 study offers.

THE SELECTION COMPONENT IS LARGER THAN THE CAUSAL ONE, and this study is the only place in the homework literature where both are measured on the same students at the same time. 'Requiring homework appears to slightly increase test 1 scores (by about two points), [while] homework completion rates have a much larger effect (between four and six points) on the initial test performance.' The randomised intent-to-treat effect of REQUIRING homework is about 2 points; the observational association with DOING homework is 4 to 6 points. The gap between those two numbers is the size of the selection bias that contaminates every correlational estimate in this topic. The authors' own novel finding is on persistence rather than achievement: retention in the course was over 95% in the treatment arm against under 90% in the control, with a Mann-Whitney test rejecting equality.

Low for the randomised contrast - assignment is by coin flip - and the paper's contribution is to measure the confounded quantity alongside it. The observational 4-to-6-point association is exactly the estimate the archive's premise predicts will be inflated by conscientiousness and ability, and it is inflated by a factor of two to three.

Unreplicated, and the population is wrong for this topic's age range. Its structural finding - that randomised requirement buys much less than self-selected completion - is corroborated by Baş and colleagues' meta-analytic split (student-randomised d = -0.125 against class-randomised d = +0.450) and by Cooper's own corpus (random-assignment subset d = 0.53 against non-random d = 0.83).