The relationship between homework time and achievement is not universal: evidence from multilevel analyses in 40 countries

231,759 students in 9,791 schools across 40 countries (PISA 2003 mathematics)

15-year-olds in 40 countries

School Effectiveness and School Improvement 20(4), 375-405. READ AT ABSTRACT DEPTH ONLY - the publisher paywalls the article and Semantic Scholar's abstract is elided, so no coefficient tables were obtained and this record carries that debt. It is included because it is the canonical statement of the level-of-analysis problem across the largest available sample: school-average homework time and individual homework time are modelled simultaneously in 40 separate countries. Multilevel correlational, no instrument, hence grade C.

From the ERIC abstract: there is 'a positive association between school-average homework time and mathematics achievement' in nearly all countries, but 'this relationship diminished when researchers accounted for socioeconomic background and school track. At the individual student level, the connection between homework duration and achievement varied considerably across nations, with no universal pattern emerging.' Jerrim and colleagues, whose full text was read, characterise it as finding 'mixed results for 40 countries... positive relationship in some countries, but a negative one in others'. The frequently quoted claim that the individual-level association is negative in 12 of the 40 countries could NOT be verified in any fetched text and is not recorded here as established.

High. Both levels of analysis are between-unit comparisons on observational survey data. The school-level association is confounded by track and intake, which the authors themselves show absorbs most of it; the student-level association is confounded by ability, conscientiousness and family, which nothing in the design removes.

The level-of-analysis dissociation replicates directly in Trautwein's German analyses, where school-level homework time falls 93% once track is controlled while student-level time is negative and grows more negative with controls, and in Fernández-Alonso's Spanish multilevel model, where class-level assignment frequency survives full controls and student-level time flips negative. Three datasets, one structure.