How Much of a Lecture Video Students Actually Watch Predicts Their Exam Scores

For years, medical educators have embraced the flipped classroom, a teaching model in which students absorb lecture content on their own time through recorded videos and then spend precious face-to-face hours solving problems and discussing cases. The promise is seductive: free the classroom from passive transmission and devote it to active learning. Yet a nagging question has always shadowed this enthusiasm. When an instructor assigns a set of video lectures, how can anyone know whether students actually engage with them, and whether that engagement translates into better performance on the exams that ultimately define their academic fate?

A new study by Raquel Gutiérrez-González and Alvaro Zamarron, neurosurgeons and educators affiliated with La Paz University Hospital and the Autonomous University of Madrid, tackles this question with unusual granularity. Their work, published in BMC Medical Education, links the digital footprint left by each medical student on each video lecture directly to that student’s accuracy on each corresponding final-exam question. Rather than treating engagement as a vague aggregate, the researchers connected the dots at the level of individual videos and individual test items, offering one of the most finely detailed portraits to date of how watching behavior relates to learning outcomes in a flipped medical course.

The setting was an undergraduate Neurosurgery course delivered in a flipped format across two consecutive academic years, 2021/22 and 2022/23. The course relied on 51 video-based lectures covering 11 topics, hosted on the Edpuzzle platform, a tool that allows instructors to track exactly how much of each video each student watched and how much time each student spent with it. In total, the researchers analyzed records from 51 students in the first cohort and 44 in the second, pooling 95 learners. The final examination contained 26 items, each of which could be mapped to the video material that prepared students for it. Every exam answer was dichotomized simply as correct or incorrect, with unanswered questions counted as errors.

The analytical machinery behind the study was correspondingly careful. The authors computed two engagement metrics for every student-video pair: the percentage of the video viewed and the minutes of time spent. They then examined associations with exam accuracy at three distinct levels, the individual item, the individual student, and the topic, using point-biserial and Pearson correlations. To compare the two cohorts, they applied Fisher’s r-to-z test. To model the probability of answering an item correctly, they turned to logistic regression with standardized predictors, and they went a step further by adjusting standard errors for the fact that multiple exam items cluster within each student, using a cluster-robust estimator and generalized estimating equations. A pre-specified sensitivity analysis addressed a subtle problem: students who were excluded from the primary sample might simply have disengaged entirely, so the researchers reclassified them as complete non-engagers with zero percent viewed and zero minutes, expanding the analytic sample to 111 and 114 students across the two cohorts.

The headline finding is consistent in direction across every analysis. Both engagement metrics were positively associated with exam accuracy in both cohorts, and the differences between the two years were not statistically significant. When the two cohorts were pooled, the association proved small at the item level: the correlation between percentage viewed and item accuracy was just 0.08, and between time spent and accuracy a nearly identical 0.07, although both were highly significant at p less than 0.001 given the enormous number of student-item pairs involved. In practical terms, engagement with a specific video explains less than one percent of the variance in whether a specific related exam item is answered correctly.

Zoom out from individual items to whole students, however, and the picture sharpens considerably. At the student level, the correlation between overall percentage viewed and overall exam performance reached 0.32, and the correlation for total time spent was also 0.32, both significant at p equals 0.002. A correlation of that magnitude corresponds to roughly ten percent of shared variance, a moderate effect by the standards of educational research, where countless confounding factors, from prior knowledge to test anxiety to general motivation, muddy every signal. The convergence of the two cohorts on essentially identical effect sizes adds a degree of confidence that this is not a one-off statistical fluke but a reproducible pattern in how students learn from video in this setting.

When both engagement metrics were entered into a single multivariable logistic model on the pooled sample, an instructive subtlety emerged. The two metrics are deeply intertwined, correlating with each other at the item level at 0.77, and this collinearity, reflected in a variance inflation factor of 2.4, meant the model had to arbitrate between two largely redundant signals. Percentage viewed emerged as the stronger and more informative indicator, associated with an 18 percent increase in the odds of answering an item correctly for every standard deviation of additional viewing coverage, a result significant at p equals 0.031. Time spent, by contrast, lost its statistical significance once percentage viewed was accounted for. The lesson for learning-analytics enthusiasts is clear: how completely a student watches matters more than how long, a distinction that raw platform dashboards reporting minutes rarely make.

The authors’ statistical rigor also tempers any temptation toward overclaiming. When standard errors were adjusted for the clustering of exam items within students, the univariable association for percentage viewed held steady, with an odds ratio of 1.21 and a 95 percent confidence interval of 1.06 to 1.38, but the multivariable estimate became non-significant, with a confidence interval spanning 0.95 to 1.45. Meanwhile, the sensitivity analysis that reclassified all excluded students as complete non-engagers left every association unchanged in direction and slightly stronger: the student-level correlation rose to 0.36, and the multivariable odds ratio rose to 1.23 with p equals 0.004. In other words, the most conservative reading of the data still supports a genuine positive association, and the least favorable assumption about missing students does not undermine it.

Equally important is what the study explicitly does not claim. Because the design is correlational and retrospective, greater engagement with the videos is associated with, but cannot be assumed to cause, better performance. It remains entirely possible that the strongest students, already destined for high exam scores, are also the most diligent video watchers, rather than the videos themselves lifting performance. The authors are candid that the item-level effects, at r of 0.08 to 0.10 and less than one percent of variance, should not be interpreted as educationally large. Still, the study’s ethics approval from the Puerta de Hierro University Hospital Ethics Committee, its adherence to the Helsinki Declaration, and its transparent reporting of every statistical caveat place it firmly on the constructive side of the reproducibility ledger.

For educators designing flipped classrooms, the practical implications are tangible. Embedding lectures on platforms like Edpuzzle that log viewing coverage gives instructors a measurable, meaningful signal of student engagement, one that correlates moderately with exam outcomes at the student level and could flag disengaged learners long before test day. The finding that percentage viewed outperforms raw time spent suggests that dashboards should prioritize completion coverage over stopwatch metrics. And the modest item-level effects serve as a healthy corrective to techno-optimism: watching a video closely improves the odds on a related exam question, but no single video transforms an exam. In the end, this Madrid neurosurgery course offers a quiet but valuable data point in the debate over digital learning: students who watch to the end tend to know more, and now educators can see it in the data, one video and one exam question at a time.

Subject of Research: The association between video lecture engagement and exam performance in a flipped undergraduate neurosurgery course

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

Article References: Gutiérrez-González, R., & Zamarron, A. (2026). Watching to learn: video engagement and item-level exam performance in a flipped classroom. BMC Medical Education. https://doi.org/10.1186/s12909-026-10553-8

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10553-8

Keywords: flipped classroom, video-based lectures, medical education, learning analytics, exam performance, student engagement, Edpuzzle, neurosurgery, logistic regression, correlational study, active learning, Watching