Gestures beat drawing for statistics novices, study finds

College students learning introductory statistics from short instructional videos performed better when they reenacted the instructor’s hand gestures than when they reenacted the instructor’s drawings, but only if they started out with relatively little knowledge of the material. That is the central result of a randomized study published in Cognitive Research: Principles and Implications, which compared two of the most common embodied teaching practices side by side and found that their effectiveness depended on how much prior knowledge learners brought to the lesson. For students who already knew the concepts reasonably well, the two approaches were statistically indistinguishable.

The study, led by Icy Yunyi Zhang and colleagues, recruited 188 undergraduates who were enrolled in introductory statistics and data science courses at two large public institutions in the same metropolitan area in the United States, one highly selective and research-intensive, the other access-oriented. Participants were randomly assigned to one of two conditions. In the gesture condition, 92 students watched and then physically reproduced gestures demonstrated in instructional videos. In the drawing condition, 96 students watched and then reenacted drawings. The lesson content was identical across conditions: nine short video clips, ranging from 22 seconds to 1 minute 27 seconds, teaching the general linear model, the empty model, the two-group model, and sum of squares as a measure of model fit.

The instructional design was tightly controlled. The videos in each condition were identical in narration and differed only in the actions visually depicted, filmed with an overhead camera showing the instructor’s hand working over a worksheet. In one representative segment, the narration explained that a coefficient of roughly 25 represented the difference between the model predictions of two groups; drawing participants watched the instructor draw a line between those predictions, while gesture participants watched the instructor move a hand up and down across that distance. After each clip, participants read the narrated sentence aloud from a screen and then reenacted the demonstrated action on their own worksheet, a watch-read-enact sequence the researchers chose to keep speech and action conceptually aligned while managing cognitive load.

The theoretical backdrop is embodied cognition, a family of theories holding that cognition is grounded in the interplay between body and mind, and that grounding abstract ideas in action and perception can support learning. Abstract STEM concepts are often described as idealized: stripped of real-world complexity and expressed in generalized forms, which makes them broadly applicable but harder to grasp because the connection to concrete experience is less apparent. Embodied pedagogies, defined as instructional methods involving bodily actions and representations, are thought to help by connecting such concepts to concrete bodily experience. Prior studies have shown benefits for gesturing, directed actions, hand-drawing, and object manipulation, as well as for merely observing actions, but also that mismatched movements can hinder understanding, so it is not movement itself but how action maps onto conceptual structure that matters.

Gesture and drawing, though both bodily, differ in a key representational way. Gestures are transient movements in which the action itself constitutes the representation, often highlighting motion or relational change over time. Drawing uses movement to produce a persistent external artifact that can be inspected and revised, with motor activity serving mainly to create a stable visuospatial representation. The authors note that their study does not test the isolated effect of any single mechanism, such as embodied representation or representational permanence, but rather compares two instructional formats that differ in how action externalizes relational structure. Narration, content, and timing were held constant across conditions.

A distinctive methodological contribution was how the researchers classified prior knowledge. Rather than using a median split of their own sample, they recruited a separate sample of 62 college students with no prior training in modeling-based statistics, gave them brief definition-based descriptions of the target concepts, and administered the same ten-item true-false pretest. That novice sample averaged 6.65 correct out of 10, so the team used a cutoff of 7: main-study participants scoring below 7 were classified as lower in prior knowledge, those at 7 or above as higher. The authors emphasize this cutoff is not a chance-performance benchmark; it was intended to mark a boundary between performance consistent with guessing plus surface familiarity and performance suggesting stronger initial understanding. By this criterion, 43 participants fell into the low-prior-knowledge group and 145 into the high group, a imbalance the researchers attribute to their participants all being actively enrolled in relevant coursework. A logistic regression confirmed that prior knowledge group did not differ significantly by condition, supporting the random assignment.

The results showed a clear interaction. A two-way analysis of variance on overall posttest performance, measured with 22 questions including 15 multiple-choice and seven open-response items graded by blind coders, produced a significant condition-by-prior-knowledge interaction, F(1, 184) = 6.04, p = .015, with the full model explaining roughly 8 percent of posttest variance. Among low-prior-knowledge learners, the gesture group averaged 14.47 points (SD = 3.59) against 11.44 (SD = 4.74) for the drawing group, a significant difference. Among high-prior-knowledge learners, gesture averaged 14.81 (SD = 4.48) and drawing 15.70 (SD = 4.74), a nonsignificant difference that the authors argue is unlikely to reflect a ceiling effect given the score distributions. Separately, prior knowledge itself predicted posttest performance, with higher-knowledge students scoring higher overall, while condition alone did not.

Exploratory analyses added a complicating layer involving engagement. Participants rated how engaged they felt during the reenactment activity on a 1-to-7 scale, and this perceived engagement did not differ between conditions. But when the researchers fit a model including engagement, they found a significant three-way interaction between condition, prior knowledge, and engagement. At low levels of perceived engagement, the original pattern held and even reversed for higher-knowledge learners: lower-prior-knowledge students performed better after gesturing, while higher-prior-knowledge students performed better after drawing. At average engagement, the condition-by-knowledge interaction remained significant but neither within-group comparison did, and at high engagement the interaction disappeared entirely. Because engagement was measured rather than experimentally manipulated, the authors stress that this exploratory finding should be interpreted descriptively rather than causally.

The theoretical interpretation draws on simulation-based accounts of embodied cognition, which propose that effective embodied pedagogies require learners to construct multimodal representations integrating sensory, motor, and conceptual information. For novices, gesture may more directly couple a concept’s structure to bodily movement, whereas drawing introduces an additional representational layer: learners must coordinate bodily action, a developing visual representation, and its intended conceptual interpretation. The authors situate their results along a continuum of representational directness, in which object manipulation may be more direct than gesture, which in turn is more direct than drawing, a framing that helps reconcile their findings with earlier work in which low-knowledge children benefited more from manipulating objects than from gesturing. They also caution that the study cannot establish a causal interpretation of any specific mechanism.

The practical implications are framed carefully. The authors suggest instructors may want to prioritize gesture-based instruction for novices while reserving drawing-based approaches for more advanced learners, and they argue the findings demonstrate the value of mechanism-aware, learner-centered embodied pedagogies in STEM education. The context matters for generalizability: introductory statistics enrolls roughly 1.5 million US students annually by one cited estimate, and the intervention was designed as a one-hour laboratory activity resembling a mini-lecture aligned with students’ actual coursework, which the researchers view as a strength for ecological validity. At the same time, they acknowledge several limitations. The study design, hypotheses, and analytic plans were not preregistered. Recruiting students from real courses constrained the sample size, and the prior-knowledge measure captured only content-specific knowledge, not other learner characteristics such as working memory or spatial reasoning that might also moderate pedagogical effects. The watch-read-reenact structure and the absence of a no-movement control mean the findings compare gesture and drawing only under this structured implementation, and future work could manipulate speech-action synchrony, include movement-free controls, test a factorial design separating sensorimotor engagement from visuospatial permanence, and extend the paradigm to authentic classrooms across age groups and longer time spans.

Subject of Research: A randomized laboratory study comparing gesture-based and drawing-based statistics instruction and the moderating role of learners' prior knowledge.

Article Title: Drawing or gesturing? Prior knowledge moderates the impact of embodied pedagogies

Article References: Zhang, I. Y., Zhao, Z., Yu, R., Tan, J., & Son, J. Y. (2026). Drawing or gesturing? Prior knowledge moderates the impact of embodied pedagogies. Cognitive Research: Principles and Implications, 11(1), Article 52. https://doi.org/10.1186/s41235-026-00740-y

Image Credits: AI Generated

DOI: 10.1186/s41235-026-00740-y

Keywords: embodied cognition, gesture, drawing, statistics education, prior knowledge, STEM learning, instructional video, randomized study, college students, data science education, cognitive science, pedagogy