Adolescents’ homework performance in mathematics and science: Personal factors and teaching practices.

7,725 students (7,451 analysed) in 353 classes and 148 schools

Spanish adolescents in the second year of compulsory secondary education, mean age 13.78, effectively the whole grade cohort of the Principality of Asturias in 2010-11

Journal of Educational Psychology 107(4), 1075-1085. Two-level hierarchical linear model with students nested in classes, so student-level homework time and class-level homework quantity and frequency are estimated separately. Outcomes are independent standardized tests built for the regional evaluation programme (96 maths and 96 science items, IRT-scaled to mean 500 and SD 100, mean alpha .85). This is a between-student cross-sectional design and therefore not causal, but it is the best-powered dose-response description available and it sequences its models so the reader can watch the coefficients move.

THE COEFFICIENT SEQUENCE IS THE FINDING. In mathematics, student-level daily homework minutes enter at B = 0.23 (SE 0.04) with a negative quadratic; adding effort and autonomy takes it to 0.02; adding socio-economic status and sex to 0.01; and ADDING PRIOR GRADES FLIPS IT NEGATIVE at -0.07 (SE 0.04). Science does the same, ending at -0.09 (SE 0.04). What survives the full model is not time at all: doing homework WITHOUT HELP (+0.09, SE 0.01 in maths; +0.08 in science) and the class-level FREQUENCY with which the teacher assigns it (+0.07, SE 0.03 in maths; +0.06 in science). Class-level QUANTITY does not survive (+0.06, SE 0.11). Effort collapses too - from 0.19 (SE 0.01) to 0.04 in maths and from 0.14 to -0.01 in science - and the authors say so directly: 'once prior knowledge is discounted, the homework effort variable presents a small effect in mathematics and no effect in science. This data contradicts the central position of effort in the model of Trautwein et al. (2006), and the conclusions drawn by Dettmers et al. (2010, 2011).' ON DOSE, the paper is the source of the one-hour-a-day rule and its own reasoning is the useful part: 'the top of the curve is between 90 and 100 min per day of homework for all subjects combined. However, the most effective is not always the most efficient' - going from 50 to 70 minutes buys about an eighth of a standard deviation, but from 70 to 90 minutes buys 3% of an SD for two extra hours a week, so 'assigning more than 70 min homework per day does not seem very efficient'. The abstract's conclusion is 'the optimum duration of homework was found to be 1 hr a day'.

High for the raw associations and the paper inadvertently demonstrates it. Prior grades are the single most powerful predictor in the model (B = 0.43, SE 0.01 for maths) and adding them reverses the sign on homework time - which is what happens when a variable was proxying for the child's standing ability all along. Socio-economic status enters at 0.23 before prior grades and falls to 0.11 after, the same story. Nothing in this design separates the child's disposition from the treatment.

The collapse-on-controls pattern replicates in Eren and Henderson's NELS analysis (a homework coefficient falling 73% when one lagged test score is added) and Rønning's Dutch data (-0.12 SD to -0.005 SD when individual covariates are added). Three datasets on three continents, the same behaviour. The 60-to-70-minute optimum is roughly consistent with the OECD's PISA-based statement that beyond about four hours a WEEK the additional time has negligible impact, and it is much lower than Cooper's high-school finding that the relation 'continued to climb unabated' past two hours a night - which is why recorded as mixed.