Does Homework Improve Academic Achievement? A Synthesis of Research, 1987–2003

Six studies that manipulated homework versus no homework (2-6 classes and 36-131 students each, effective sample sizes 5.2-15.8 after adjusting for clustering); 69 homework-achievement correlations from 32 documents; plus multiple-regression and structural-equation studies. Drawn from ~900 empirical documents indexed 1987-2003

US students, kindergarten through grade 12; the naturalistic correlational estimates are dominated by large national surveys of secondary students

Review of Educational Research 76(1), 1-62 — the anchor synthesis of the homework literature and the source of nearly every number quoted in the public debate. Four design families are synthesised separately, which is the paper's virtue: (a) homework imposed versus withheld, (b) non-random assignment with statistical control, (c) naturalistic time-on-homework correlations with controls, (d) structural equation models. The critical facts about family (a), the only causal family, are buried in Table 3 and are decisive for how the headline should be read. All six studies were UNPUBLISHED. Only one assigned students within classes (McGrath 1992, and even there the report contradicts itself about whether assignment was random or alphabetical); the others randomised or assigned whole classrooms — as few as one classroom per arm — and then analysed at the STUDENT level, ignoring clustering. Every outcome was a UNIT TEST written by the teacher or the textbook publisher, aligned to the homework content; durations were a single unit or about three weeks. The one study that also administered a standardized instrument (Meloy 1987, a shortened Iowa Test of Basic Skills language subtest) reported POSITIVE effects on its unit tests and NEGATIVE effects on the standardized test, in both the third- and fourth-grade samples. Cooper et al. themselves conclude "we await studies that individually permit strong conclusions establishing the productive impact of homework on achievement."

THE AGE GRADIENT IS THE FINDING, and it is sharp. Under fixed-error assumptions the correlation between time on homework and achievement is r = .25 (95% CI .25/.25) for grades 7-12 but r = -.04 (95% CI -.06/-.02) — significantly NEGATIVE — for grades K-6, Q(1) = 710.68, p < .0001. Under random-error assumptions it is r = .20 (95% CI .17/.22) secondary versus r = .05 (95% CI -.03/.13) at elementary, i.e. not different from zero. The pooled correlation across all grades is r = .24 fixed / r = .16 random (unweighted mean .14, median .17), and 19 of 69 correlations were negative. WHO REPORTS THE HOMEWORK TIME MATTERS AS MUCH AS THE CHILD'S AGE: student-reported time correlates r = .25 with achievement, parent-reported time r = -.02 (95% CI -.10/.07) — the whole effect lives in the self-report. THE CAUSAL EVIDENCE IS SIX UNPUBLISHED STUDIES ON TEACHER-MADE UNIT TESTS: d ranging .39 to .97, weighted mean d = .60 (d = .63 using effective sample sizes at an assumed ICC of .35); the three studies with some form of random assignment give d = .53 (95% CI .29/.79) fixed and .54 (.26/.82) random, the two non-random ones d = .83 (95% CI .37/1.30). THE MEASURE-TYPE INFLATION IS ADMITTED IN THE PAPER: Cooper (1989) reported d = .21 for the same comparison; the 2006 figure is d = .60, and the authors' own explanation is that "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." THE RIGHT COUNTERFACTUAL IS NEVER TESTED AFTER 1987: Cooper (1989) put homework against in-school supervised study at d = .09, and this synthesis "found no study conducted since 1987 that carried out a similar type of comparison." On dosage, Cooper (1989) found the high-school relation appears only above about 1 hour per week and keeps rising to the top measured band (2+ hours a night), while for junior high it appears at small doses and disappears entirely beyond 1-2 hours a night.

HIGH for everything except the six unit-test experiments, and those are too small and too aligned to carry the verdict. Time spent on homework is a behaviour, chosen by the child and enforced by the parent, and it indexes conscientiousness, prior attainment, parental education and parental supervision — all substantially heritable and all independently predictive of achievement. The parent-report result is the tell: the same construct measured by a different reporter yields r = -.02, which is what one expects when the correlation is carried by a self-perception that co-varies with the outcome rather than by hours of work. The reverse-causation channel runs at the same time and in the opposite direction — struggling students take longer over the same assignment — which is the standard explanation for why the elementary coefficient is negative.

The correlational core replicates its own predecessor closely: Cooper (1989) reported r = .19 fixed across all grades and grade-band correlations of about zero in elementary, r = .07 in middle grades and r = .25 in high school; this synthesis, on a disjoint 1987-2003 literature, returns r = .24 overall with the same monotone gradient. Two independent syntheses of different studies finding the same age pattern is the strongest claim in the topic. What did NOT replicate is the experimental magnitude — d = .21 became d = .60 — and the authors attribute the change to the outcome measure rather than to the intervention, which makes the larger number the less trustworthy one.