Homework — effects by age and by dosage

Verdict

Homework is the largest gap in this archive between how much is claimed and how much has been tested. Something like 900 empirical documents on homework appeared between 1987 and 2003 alone. The number of studies that randomly assigned individual school-age children to more or less homework and measured achievement on an independent test is, as far as this survey can establish, zero.

What exists is a very large correlational literature with a striking age pattern, and a very small causal literature that mostly does not reproduce it. The age pattern is real and replicates three times independently: near-zero or negative in primary school, around r = .25 in secondary. But the one design that removes the child from the equation at primary age — Jerrim and colleagues' student fixed effects across three subjects and 24 countries — returns 0.01 standard deviations per extra ten minutes, precisely estimated, significant in two countries out of twenty-four. And the one design that does it at secondary age — Eren and Henderson's within-student, within-teacher differencing — finds an effect in mathematics and in no other subject, from assigned rather than reported homework, and it moves by a factor of 4.8 depending on which fixed effects are included.

The most useful single number in the topic is not about homework at all; it is about selection. Grodner and Rupp randomised university sections to homework-required or not, and measured both the randomised effect and the observational association in the same students at the same time: requiring homework moved test scores about two points, while doing homework was associated with four to six. The selection component is two to three times the causal one. That is the size of the error in every correlational estimate in this literature.

What the evidence shows

The age gradient — real, replicated, and correlational

What happens when a design controls the student

The entire randomised base

What beats quantity

Hereditarian-lens assessment

Risk: medium, and the reason is worth stating precisely rather than averaging away. The age gradient — the headline finding, and the one everybody quotes — is high-risk: it is built entirely out of between-student correlations between self-reported homework time and achievement. The two estimates that actually test it are low-risk: Jerrim's student fixed effects and Eren and Henderson's within-student, within-teacher differencing both remove the child from the equation, and both come back near zero outside grade-8 mathematics. The verdict rests on that mix, which is what medium means. Read the correlational half as high-risk wherever it is cited alone.

Why the correlational base is compromised, specifically. Time spent on homework is not an input a school administers; it is a behaviour a child produces, and it indexes conscientiousness, prior attainment, parental supervision and family order — every one of them substantially heritable and every one independently predictive of achievement. The literature contains two natural experiments on its own measurement that make this vivid:

• Cooper's reporter analysis. Student-reported homework time correlates r = .25 with achievement. Parent-reported homework time, on the same construct, correlates r = −.02 (95% CI −.10 to .07). Change who fills in the form and the entire association disappears.
• Trautwein's effort-versus-time dissociation. Homework time and homework effort correlate negatively with each other (r = −.12), and predict achievement in opposite directions (time −.14, effort +.16). "Homework time is not a suitable indicator of the effort that students put into their homework." A slow, struggling child and a diligent one are indistinguishable in the data.

And the collapse-on-controls pattern is the archive's premise in its purest form. Three datasets on three continents, none of them talking to each other:

• Eren and Henderson (NELS): homework coefficient 3.56 → 0.97, a 73% fall, from adding one lagged test score.
• Rønning (Dutch PRIMA): −0.12 SD → −0.005 SD from adding individual covariates.
• Fernández-Alonso (Spain): +0.23 → −0.07, a sign flip, from adding prior grades.

One documented absence deserves recording as a finding. A dedicated search across five strategies found no study estimating the heritability of homework time or homework effort, and no study estimating a homework-achievement association within monozygotic twin pairs or within families. The literature's central object is a between-student correlation between self-reported effort-like behaviour and achievement — precisely the estimate a within-family design exists to decompose — and nobody has run one. Adjacent evidence says what it would find: educational achievement is roughly 66% heritable, self-discipline around .49 with shared environment judged unimportant, and roughly two-thirds of the GCSE prediction from conscientiousness-type traits is genetically mediated.

Boundaries & what critics say

• Cooper's causal corpus does not support the number that gets quoted from it. The d = .60 comes from six studies, none of them published, with 36 to 131 students and effective sample sizes of 5.2 to 15.8 after adjusting for clustering. Several randomised whole classrooms — in one case a single classroom per arm — and then analysed at the student level. Every outcome was a unit test written by the teacher or the textbook publisher, aligned to the homework content, over one unit or about three weeks. The one study that also administered a standardized instrument reported positive effects on its unit tests and negative effects on the standardized test.
• Cooper says the measure explains the number, in his own paper. The 1989 synthesis put the same comparison at d = .21, and the 2006 update at d = .60. The authors' explanation: "four of the studies in the earlier synthesis used standardized tests as the outcome measure. In the current synthesis, all of the studies used unit tests, a measure more closely aligned with the content of the homework assignments." This is the archive's 2× measure-type inflation rule, confirmed by the meta-analyst, and it means the smaller, older number is the trustworthy one.
• The right counterfactual has not been tested since 1986. Cooper (1989) compared homework with in-school supervised study and found d = .09 — and "in-school study proved superior" at elementary level. The 2006 synthesis "found no study conducted since 1987 that carried out a similar type of comparison." Every effect size in the modern literature is against doing nothing, not against the next-best use of the same time.
• The class-level and student-level coefficients have opposite signs, and conflating them is the field's characteristic error. Trautwein's PISA analysis: school-level homework time enters at +24.67 and falls to +1.76 once school track is controlled — a 93% collapse — while student-level time is −3.45 and grows to −7.84 with controls. Dettmers finds the same across 40 countries. Fernández-Alonso finds the same in Spain. "Schools where students do more homework score higher" is mostly a statement about tracking.
• The dosage literature does not agree with itself. Cooper (1989) found the high-school association "continued to climb unabated to the highest measured interval (more than 2 hours per night)." Fernández-Alonso finds a peak at 90-100 minutes and argues 70 is the efficient point. The OECD puts the plateau at about four hours a week. And the only student-randomised dosage trial tested a maximum of 45 minutes a week, so it cannot speak to any of this.
• Ozyildirim's time trend should unsettle everyone. The homework-achievement effect in TIMSS runs −0.270 (1999), 0.036 (2003), 0.251 (2007), 0.439 (2011), 0.525 (2015). A parameter that swings from strongly negative to strongly positive in sixteen years, on the same instrument in the same countries, is not a causal constant. It is a marker of something else that changed.
• The inequality claim is contested by the better design. Rønning's headline is that homework raises within-class dispersion — the 85th-15th percentile gap widens by 0.18 — because "the weakest pupils are unaffected by homework" while the top pulls away. Jerrim, using student fixed effects across 24 countries, explicitly fails to replicate it: "homework may not prejudice poor students nor benefit rich ones (unlike the results of Rønning, 2011)."
• The one randomised student-level dosage trial is a vendor white paper. The Sparx trial is genuinely well designed — matched quadruples randomised within class, double-blind, 96% complete data — and it is not peer reviewed, publishes no effect size, and was produced by the company selling the platform. Under this archive's rules an unreplicated developer-led positive cannot establish anything; it is recorded because a documented near-absence of randomised evidence is itself the finding.
• Parental help is null to negative in three independent syntheses, which is a boundary rather than a harm claim: the amount of help is confounded with the child needing it. What survives is style — autonomy-supportive help — and doing homework unaided, which is one of only two variables standing after full controls in the best-powered dose study.

Practical guidance

• Primary school: do not assign homework to raise achievement. The best-identified estimate is a precise zero across 24 countries, the correlational estimate is negative, Cooper's own comparison found supervised in-class study superior, and homework time is falling across the OECD without visible cost. If you assign it for habit or for parental visibility, say that is the reason.
• Secondary school: cap it early and assign it often. The plateau arrives around an hour a night or four hours a week. Frequency of assignment survives every set of controls in every multilevel study; quantity does not. Short and regular beats long and occasional, which is also what the spacing literature predicts.
• Change what the homework is before changing how much there is. The two clean randomised trials in this area both did that and both beat the entire quantity literature: interleaving the same problems produced d = 0.83 on a delayed unannounced test, and adding immediate feedback and hints produced 0.18 SD on a standardized state test across 43 schools.
• In mathematics, assigned homework is the one place the causal evidence holds up — about 0.18 SD per additional weekly hour assigned, at grade 8. In science, English and history the same design finds nothing, and there is no cross-subject spillover.
• Never treat homework hours as a measure of effort or of virtue. They correlate negatively with effort, negatively with prior achievement, and the whole association vanishes when a parent rather than the child reports the number.
• Do not use homework as an equity intervention in either direction. The claim that it helps disadvantaged children is unsupported; the claim that it harms them rests on one study that a better design failed to replicate.

Open questions

• The obvious trial has never been run: randomise classrooms within schools to a homework and a no-homework condition for a full year, measure on an independent standardized test, and follow for two. Every number in this topic would move.
• Nothing here is genetically informative. A within-monozygotic-twin-pair analysis of homework time and achievement would settle in one study what forty years of between-student correlations cannot. The data exist; the analysis does not.
• Why does the TIMSS homework effect swing from −0.27 to +0.53 in sixteen years? Nobody has looked, and until someone does, the secondary-school gradient should be held loosely.
• What is the right counterfactual? Not one study since 1986 has compared homework with the same time spent in supervised study at school, which is the comparison a school actually faces.
• Is the maths-only result real? It is the single positive finding from the best-identified secondary-school design, and it moves by a factor of 4.8 with the fixed effects. It needs a replication before it carries any weight.