Homework: what the effectiveness research shows

Homework is the only educational intervention delivered nightly, in millions of homes, by unpaid and untrained staff. Parents supervise it without seeing the lesson it came from. Teachers assign it without seeing the kitchen it lands in. Whole systems disagree about it flamboyantly: some jurisdictions cap it or ban it outright for young children, others pile it on as a badge of rigour. Everyone in the argument claims the research is on their side.

The research itself is calmer than the argument. It has been synthesized repeatedly for over three decades, and the leading scholars in the field once titled a review of it “still much of a mystery” (Trautwein & Köller, 2003). That title is honest about the correlational mess underneath. But two findings rise out of the mess with unusual consistency. Homework’s measurable effect on achievement climbs steeply with age — near zero for young children, moderate for teenagers. And wherever researchers can separate them, the quality of homework beats the quantity of it.

This article walks the evidence in order. First the classic syntheses and their age gradient, then the reasons primary homework underperforms. Then the diminishing returns to minutes, the multilevel studies that dethroned time as a metric, and the design guidance that survives all of it. The stakes are real. Homework consumes hundreds of hours per child per year — hours that belong to childhood if schooling wastes them.

The most argued-about minutes in schooling

Start with what homework should do in theory. Practice strengthens memory, and practice spread over days beats practice crammed into one sitting — the spacing literature is one of the oldest and firmest results in learning science (our corporate library covers it in the forgetting-curve evidence). Homework is, in principle, a nightly spacing machine. It revisits the day’s material after a delay, which is exactly when revisiting pays.

But theory also predicts the failure modes. Practice pays when the task is at the right difficulty, when errors meet feedback, and when the learner actually engages. Homework routinely violates all three. The task is one-size-fits-all. Feedback arrives days later, if ever. And engagement at 8 p.m., unsupervised, varies from deep work to copied answers. So the theoretical question is not “does practice work?” — it does — but whether homework as actually assigned delivers practice worth the name.

That is an empirical question, and it has an unusually long empirical record. The task of settling it fell largely to one researcher, Harris Cooper, whose syntheses bracket the modern literature (Cooper, 1989). What he found made both camps in the homework war half-right.

What the syntheses found

Cooper’s 1989 synthesis pooled roughly 120 studies and asked the plain question: do students assigned homework outperform students assigned none (Cooper, 1989)? On average, yes. But the average concealed the real finding — a gradient by school level. For high-school students the advantage was substantial, on the order of d ≈ 0.6. (Cohen’s d expresses a difference in standard-deviation units; 0.2 is conventionally “small”, 0.5 “medium”.) For middle-school students it was roughly half that. For primary-school children it was close to nothing.

Seventeen years later, Cooper, Robinson and Patall re-synthesized everything published from 1987 to 2003, with tighter methods and harder questions about design quality (Cooper, Robinson & Patall, 2006). The gradient survived intact. The overall correlation between homework and achievement came out at roughly r ≈ 0.24 — and it was carried almost entirely by secondary students. In primary samples the association was near zero. The handful of homework-versus-none comparisons again showed effects on the order of half a standard deviation, and again almost entirely above primary age (Cooper, Robinson & Patall, 2006).

Hold on to what the gradient does and does not say. It does not say primary homework harms children, and it does not say every teenager needs more worksheets. It says the achievement case for homework is an age-dependent case. A policy that assigns nightly homework to seven-year-olds on achievement grounds is spending family evenings on an effect the syntheses cannot find.

Why primary homework moves so little

The gradient demands an explanation, and the synthesists offered several that have aged well (Cooper, 1989). Young children have short attention spans, and an evening assignment competes with fatigue. They lack study skills — the ability to plan, self-check and persist — so unsupervised practice easily becomes practising errors. And teachers assign primary homework for many purposes besides achievement: building routine, informing parents, filling policy quotas. An intervention aimed at four targets at once tends to hit none measurably.

The home multiplies the variance. One seven-year-old sits at a quiet table with a patient parent; another shares a loud room and gets help in a language the worksheet is not written in. Whatever homework adds for the first child, it can subtract for the second — and in a synthesis those families average toward zero (Cooper, Robinson & Patall, 2006). Averaging toward zero is not the same as doing nothing; it can mean doing opposite things to different children.

None of this makes the primary years practice-free territory. It makes them the wrong territory for unsupported written homework. Short reading together, a few retrieval questions with instant feedback, practice that fits inside ten minutes — these respect the attention span and the evidence at once. What the gradient rules against is the long, unsupported worksheet, deployed young, on achievement grounds.

The time trap

Within the secondary years, where homework does work, a second illusion waits: that if some helps, more helps more. The cleanest demonstration against it comes from Spain. Fernández-Alonso and colleagues analysed maths and science performance in a sample of roughly 7,700 grade-8 students, adjusting for background and prior factors (Fernández-Alonso, Suárez-Álvarez & Muñiz, 2015). Adjusted scores rose with daily homework time — up to a peak at roughly 60 to 70 minutes. Beyond it, they bent downward. The students logging two hours a night scored no better, and often worse, than the one-hour group.

The syntheses had seen the same ceiling from another angle: even for high-school students, achievement benefits flatten somewhere beyond roughly 90 to 150 minutes a night (Cooper, Robinson & Patall, 2006). Diminishing returns are the norm in practice of every kind. Tired practice is poor practice, and the fourth worksheet of the evening is rarely met by a fresh mind.

There is a subtler problem with time than mere diminishing returns. Time spent is two variables wearing one name. It measures what the teacher assigned — and how much the pupil struggled. A 40-minute night can mean an ambitious task or a lost child. Any metric that cannot tell those apart will mislead in both directions, which is exactly what the next generation of studies showed.

Quality over quantity

The correlational literature had a structural flaw, and naming it changed the field. Trautwein and Köller’s “mystery” review pointed out that homework operates at two levels at once (Trautwein & Köller, 2003). Classes differ from each other in how much homework is assigned — a teacher decision. Pupils differ within a class in how long they take and how hard they try — a pupil variable. Studies that mash the two levels together produce answers that mean nothing, because the same variable can point opposite ways at each level. The fix is multilevel modelling: statistics that keep the class question and the pupil question separate.

Run properly, the levels tell different stories. In Trautwein’s multilevel analyses, homework frequency at the class level predicted achievement: classes given regular, short assignments outperformed. Pupil effort predicted achievement too. But pupil time — the variable every survey asks about — predicted achievement weakly, and often negatively, within classes (Trautwein, 2007). The longer a pupil sat over the same assignment as classmates, the worse the signal. Time on homework, at the pupil level, is substantially a struggle detector.

Then came the quality evidence. Dettmers, Trautwein and colleagues followed German secondary classes over time and measured how pupils rated their homework: well-selected, thought-provoking tasks versus busywork (Dettmers et al., 2010). Classes with high-quality homework gained achievement over the school year, and their pupils were more motivated to do the work. Quantity carried no comparable weight. The paper’s title states the field’s mature position in six words: homework works if homework quality is high.

The ten-minute rule

Practical guidance has to compress all of this into something a school can print, and the standard compression is Cooper’s ten-minute rule: homework should total roughly ten minutes per grade level per night (Cooper & Valentine, 2001). Ten minutes for a first-year pupil, rising to something under two hours by the end of high school. The numbers are not magic. They are a ceiling that tracks the age gradient — trivial dosage where effects are trivial, real dosage where effects are real — and stays under the ceiling where returns bend negative.

Used properly, the rule is a cap, not a quota. It does not oblige a primary teacher to find ten minutes of worksheet; it forbids the sixty-minute pile-up that happens when four teachers assign twenty minutes each without talking to each other. That coordination failure is common in secondary schools, invisible to every individual teacher, and exactly what a whole-school dosage policy exists to prevent (Cooper & Valentine, 2001).

What homework is for

The last strand of the literature asks the design question: what is any given assignment for? Epstein and Van Voorhis catalogued the purposes teachers actually hold: practising taught material, preparing for new material, extending skills to new contexts, and connecting school with parents. Most homework fails, they argued, because its purpose was never chosen (Epstein & Van Voorhis, 2001). An assignment designed for practice looks different from one designed to involve a parent; an assignment designed for nothing looks like a worksheet.

Their “more than minutes” program built interactive assignments — tasks that structure a short conversation with a family member instead of assuming help will happen. In their studies, those assignments showed better completion and family involvement (Epstein & Van Voorhis, 2001). The general lesson holds beyond any one program. The home is not a classroom, and assignments that pretend otherwise select for the homes that already resemble one. Purpose-built homework treats the home as what it is, and gives parents a defined, doable role — audience, quizzer, signature — rather than the impossible role of substitute teacher.

What the evidence doesn’t show

The homework literature is broad but shallow in places, and its limits belong next to its findings.

• Primary purposes are under-measured. Near-zero achievement effects do not test the habit-building and parent-connection goals primary homework usually claims; those outcomes are asserted far more often than measured (Cooper, 1989).
• The gradient is mostly correlational. True experiments are scarce and small; stronger students attract and complete more homework, and no synthesis can fully unpick that (Cooper, Robinson & Patall, 2006).
• The optimum is one system’s. The 60–70-minute peak comes from one country, one grade and one model; treat it as a shape, not a prescription (Fernández-Alonso, Suárez-Álvarez & Muñiz, 2015).
• Effort cannot be assigned. Effort predicts achievement, but it is a pupil state driven by motivation and task quality — a target, not a lever (Trautwein, 2007).
• Quality lacks a ruler. “Well-selected tasks” are rated by pupils after the fact; no validated instrument yet lets a school audit homework quality prospectively (Dettmers et al., 2010).
• Abolition has never been tested. No credible large trial has removed or doubled homework at system scale; the loudest policy positions rest on the thinnest slice of the evidence (Trautwein & Köller, 2003).

Where the evidence stops

1. 1Primary purposes are under-measured
2. 2The gradient is mostly correlational
3. 3The optimum is one system’s
4. 4Effort cannot be assigned
5. 5Quality lacks a ruler
6. 6Abolition has never been tested

© 2026 FUTURE PROOF™

Homework by the evidence

Compressed into practice, the literature gives schools and families five rules that fit on one page.

Set dosage by age, and cap it. Little or none in the primary years on achievement grounds; real but bounded amounts for teenagers, with the ten-minute rule as a whole-school ceiling that four subject teachers cannot accidentally breach (Cooper & Valentine, 2001).

Assign for a named purpose. Every assignment should answer “what is this for?” before it leaves the classroom — practice, preparation, extension or parent connection — and be designed for that answer (Epstein & Van Voorhis, 2001).

Choose frequency and quality over volume. Regular short assignments of well-selected tasks beat occasional long ones; the class-level evidence is consistent on this (Trautwein, 2007); (Dettmers et al., 2010).

Read time as a signal, not a virtue. The pupil who takes triple the intended time is not being conscientious; they are drowning quietly. Collect intended time versus actual time, and treat the gap as a referral (Trautwein, 2007).

Fill the minutes with retrieval, not rereading. Whatever the dosage, spaced retrieval practice is the content the learning science pays for — short questions answered from memory, with feedback, spread across nights (Cooper, Robinson & Patall, 2006). Homework’s format is negotiable. What fills it should not be.