Watching to learn: video engagement and item-level exam performance in a flipped classroom

To assess the association between each student’s engagement with the video-based lectures (VBLs) of a flipped classroom environment and their accuracy on the corresponding final-exam items. Retrospective analysis of two consecutive cohorts (2021/22, 2022/23) of an undergraduate Neurosurgery course taught with a flipped classroom using 51 VBLs (11 topics) hosted on Edpuzzle. For each student and video, percentage viewed and time spent were related to accuracy on the 26 final-exam items, dichotomised (correct vs. error or unanswered). Associations were examined at three levels (item, student, topic) using point-biserial and Pearson correlations, Fisher’s r-to-z test for between-cohort comparison, and logistic regression with standardised predictors, with standard errors also adjusted for the clustering of items within students (cluster-robust estimator and generalised estimating equations). Analytic sample: 51 and 44 students (95 pooled); a pre-specified sensitivity analysis reclassified the excluded students as non-engagers (0% viewed, 0 min; 111 and 114 students). Both metrics were positively associated with accuracy in both cohorts. In the pooled sample, the association was small at the item level (% viewed r = 0.08; time r = 0.07; both p < 0.001) and moderate at the student level (% viewed r = 0.32, p = 0.002; time r = 0.32, p = 0.002). Differences between cohorts were not significant. In the pooled multivariable model, percentage viewed showed the stronger association (OR = 1.18 per standard deviation; p = 0.031), whereas time was no longer significant, reflecting collinearity (item-level r = 0.77; variance inflation factor 2.4). Adjusting standard errors for clustering maintained the univariable associations (% viewed OR = 1.21, 95% CI 1.06–1.38) but rendered the multivariable estimate non-significant (OR = 1.18, 95% CI 0.95–1.45). Reclassifying all excluded students as non-engagers left every association unchanged in direction and slightly stronger (student level r = 0.36; multivariable OR = 1.23, p = 0.004). Individual engagement with VBLs was positively and consistently associated with exam performance, with a moderate effect at the student level and percentage viewed as the more informative indicator. Item-level effects were small ( r = 0.08–0.10, less than 1% of the variance) and should not be interpreted as educationally large. Given the correlational design, greater engagement with the videos is associated with, but cannot be assumed to cause, better performance.