Practice Strategies Still Boost Learning in Real Classrooms, Meta-Analysis Finds

Retrieval practice, distributed practice, and interleaved practice genuinely work in real primary and secondary school classrooms, but their benefits are smaller and far more conditional than decades of laboratory research have suggested. That is the central conclusion of a new three-level meta-analysis published in Educational Psychology Review, which pooled 233 effect sizes from 53 experimental studies published between 2006 and 2024, covering 7,303 school students aged roughly 6 to 18. The analysis found an average testing effect of g = 0.37 for retrieval practice, an average spacing effect of g = 0.20 for distributed practice, and an average interleaving effect of g = 0.40 for interleaved practice. All three are smaller than the medium-to-large effects reported in earlier syntheses of laboratory or mixed lab-and-classroom research, yet the authors argue they remain educationally meaningful.

Previous meta-analyses of retrieval practice, eleven of them published since 2010, reported average effects ranging from 0.40 to 0.88, while syntheses of distributed practice reported effects between 0.58 and 0.85 and interleaving syntheses between 0.39 and 0.66. But most of those estimates came from laboratory experiments or classroom research in higher education. Primary and secondary school students are under-investigated populations in this literature, and classroom research differs from laboratory work in ways that could shrink effects: experimental control is weaker, learning materials are more complex, the amount of information to master is greater, prior knowledge is more varied, and procedures span multiple lessons.

To address this, the team applied unusually strict inclusion criteria modeled on Agarwal and colleagues’ 2021 recommendations. Included experiments had to be conducted with typically developing students in regular primary or secondary education, use an active comparison condition, assess knowledge objectively with curricular materials, take place in a school context under supervision, involve individual rather than collaborative practice, and meet basic quality requirements such as random assignment and matched instruction time. The researchers searched ERIC, PsycINFO, and Web of Science, supplemented the search with doctoral dissertations and expert consultation, and screened 2,696 unique records using the active-learning tool ASReview LAB. Of 77 studies assessed in detail, 53 survived, yielding 233 effect sizes nested in 71 experiments.

Publication bias checks were reassuring. Funnel plots for all three strategies were roughly symmetrical, and Egger’s regression tests found no significant indication of publication bias for retrieval practice, distributed practice, or interleaved practice. The dataset also contained many negative and null effects, which the authors note would be unlikely if smaller studies with unfavorable results had been systematically suppressed. Still, they acknowledge a limitation: only one doctoral dissertation out of 300 initially identified met the inclusion criteria, and the team deliberately refrained from contacting researchers to solicit unpublished data, a choice that may constrain how fully publication bias can be ruled out.

The headline numbers come with important statistical context. For retrieval practice, estimated from 120 effect sizes in 31 studies with 3,391 participants, the average effect of g = 0.37 carries a 95 percent confidence interval of 0.18 to 0.56. For distributed practice, estimated from 56 effect sizes in 15 studies with 2,010 participants, the average of g = 0.20 has a confidence interval of 0.03 to 0.36. For interleaved practice, based on 57 effect sizes in 10 studies with 1,902 participants, the average of g = 0.40 spans 0.10 to 0.71. Heterogeneity was substantial to considerable in all three models, meaning individual studies varied widely around these averages. Under Cohen’s conventional benchmarks the effects are small to medium, but the authors point out that Kraft’s 2020 empirical benchmarks for education interventions would classify all three as large.

The moderator analyses reveal where each strategy reliably helps and where it falters. For retrieval practice, four moderators were significant: type of knowledge, posttest format, format matching, and type of learning outcome. The testing effect was robust for factual knowledge (g = 0.53) but not for conceptual, procedural, or mixed knowledge. It was positive for recognition, cued recall, and free recall posttests, but strikingly negative for mixed posttest formats (g = −0.91). It was larger when practice and test formats matched (g = 0.49) than when they did not (g = 0.29), and it appeared for retention and recognition outcomes but not for comprehension or transfer. Crucially, a joint model showed these four moderators were so intertwined that none explained unique variance, a pattern the authors attribute to multicollinearity rather than to any single decisive factor.

The authors interpret this pattern through transfer-appropriate processing: performance is best when the cognitive processing during practice matches what the test later demands. Retrieval practice consolidates memories rather than building coherent understanding, so it pays off when tests ask students to recognize or retain factual material, but not when tests demand comprehension or transfer of interrelated concepts. For those outcomes, students also need generative tasks that construct mental models, not just retrieval tasks that strengthen them. Similarly, the spacing effect was significant when distributed practice was compared with massed practice (g = 0.39) but not with less-distributed practice, and it emerged only when the interstudy interval fell between 10 and 100 percent of the retention interval (g = 0.31), consistent with the idea that spacing works by forcing effortful but successful retrieval.

Interleaving showed its own instructive boundary condition. The interleaving effect was significant when practice included no feedback (g = 0.25) or delayed feedback (g = 0.82, though based on only five effect sizes), but absent when feedback was immediate. The authors explain this through the two-stage framework of sequencing effects: interleaving first demands attention-based discovery of categories and then memory-based consolidation, and immediate feedback after each item may remove students’ need to engage in either process. The finding challenges the widespread assumption that faster feedback is always better, though the authors label their explanation conjectural and call for research testing it directly with methods such as eye tracking.

The authors are careful about the limits of their conclusions. Because relatively few eligible classroom studies exist, the analysis may have lacked power to detect smaller moderation effects, and non-significant moderators should not be read as evidence that they do not matter. Ten moderator-level estimates rested on ten or fewer effect sizes, and the team explicitly flags those as tentative. Initial practice performance was reported for only 105 of 233 effect sizes and effort for only 5 of 71 experiments, preventing analysis of the defining properties of desirable difficulties. The authors recommend that teachers routinely use all three strategies, tentatively favoring retrieval practice for factual retention with matched test formats, spacing schedules within the 10-to-100-percent ratio range, and interleaving with delayed or no feedback, while stressing that a one-size-fits-all prescription must give way to context-dependent instruction tailored to students, content, and the cognitive processes each practice actually engages.

Subject of Research: Effectiveness and boundary conditions of retrieval, distributed, and interleaved practice in authentic primary and secondary school classrooms

Article Title: A Meta-Analysis of the Effectiveness and Boundary Conditions of Retrieval Practice, Distributed Practice, and Interleaved Practice in Authentic School Classrooms

Article References: Ruitenburg, S. K., Goos, M., Lambert, G., Kirschner, P. A., Jarodzka, H., & Camp, G. (2026). A Meta-Analysis of the Effectiveness and Boundary Conditions of Retrieval Practice, Distributed Practice, and Interleaved Practice in Authentic School Classrooms. Educational Psychology Review, 38(1), Article 137. https://doi.org/10.1007/s10648-026-10225-5

Image Credits: AI Generated

DOI: 10.1007/s10648-026-10225-5

Keywords: meta-analysis, retrieval practice, distributed practice, interleaved practice, testing effect, spacing effect, classroom research, primary education, secondary education, feedback, transfer-appropriate processing, desirable difficulties