Longitudinal Associations Between Homework and Examination Performance in Engineering Mathematics

Generative artificial intelligence has renewed questions about whether take-home homework remains associated with performance on supervised assessments. This retrospective longitudinal study analyzed 759 student records from 15 cohorts of an engineering mathematics course (2009–2025). Annual Pearson correlations between homework and quiz, midterm-examination, and final-examination scores were calculated with 95% confidence intervals and pooled within three periods using inverse-variance-weighted Fisher transformations. Course-grade outcomes were treated separately because they could incorporate homework. Homework was positively associated with all three supervised outcomes in every cohort, although effect sizes varied across years. Pooled homework–final-examination correlations were 0.75 before COVID-19, 0.68 during 2020–2022, and 0.78 during 2023–2025. The difference between the latter periods was nonsignificant, Z = 1.70, p = 0.089, and remained nonsignificant under random-effects pooling (p = 0.303) and after excluding 25 all-zero records (p = 0.075). Annual final-examination correlations rose from 0.45 in 2020 to 0.71 in 2021 and 0.81 in 2022, indicating that the increase preceded 2023. Thus, homework remained informative about performance on supervised assessments, but the temporal pattern did not show a discrete change beginning with post-ChatGPT availability. These findings support homework as complementary assessment evidence while showing that calendar-based comparisons cannot identify the educational effects of generative AI.