Dozens of earlier posts on this site cite Roediger and Karpicke's 2006 experiments on the testing effect, for the half that suits a rehearsal company: a week after learning, students who had practised recalling a passage remembered around 50 per cent more than students who had re-read it. None of them mentions the other half. Five minutes after learning, re-reading was ahead, and that changes when a repetition should happen.
What the study actually did
Roediger and Karpicke gave Washington University undergraduates two short prose passages each. One passage they read a second time; on the other they took a free-recall test, with no feedback. Everyone then took a final recall test on both after five minutes, two days or one week.
In Experiment 1 the five-minute test favoured re-reading: 81 per cent of the passage recalled against 75 for the tested passage. After two days the order had flipped, 68 against 54. After a week it was 56 against 42, a difference the authors report as d = 0.83.
Experiment 2 moved the comparison between groups. One group studied a passage four times, one studied three times and took one test, one studied once and took three tests. At five minutes the pure-study group led, 83 per cent to 78 to 71. A week later the order reversed exactly: 61 for the repeated-test group, 56 for the single test, 40 for the re-readers, who had read the passage 14.2 times to the tested group's 3.4.
The authors' summary: "Repeated studying improved performance relative to repeated testing on final tests given after a 5-min retention interval, but the effect reversed on delayed tests."
The half we kept quoting
Dozens of earlier posts on this site mention this study. The usual sentence says retrieval produced around 50 per cent better retention, and it is accurate: Experiment 2, one week on, 61 against 40. None of them mentions that the same experiment, measured five minutes on, says the opposite. The result does not flatter rehearsal, and those posts left it out.
That stops here. This post sets the study out in full, and it is the page the shorter mentions should point to rather than repeating the convenient half. It is not the first sweep of this kind: in September we found an unsourced training-spend figure on our own pages and replaced it with the official survey. Half a finding is not a finding.
Confidence did not track the result in either study
In Experiment 2, Roediger and Karpicke also asked students how well they expected to remember the passage a week later. The group that had only re-read was the most confident. In the authors' words: "Although students in the repeated-study condition predicted they would perform very well a week later (relative to those in the other conditions), they actually performed the worst."
Our own research has a parallel. In the randomised 2020 UCL study behind our method, co-founder Ben Laumann had 57 participants learn the same feedback model by rehearsal with a role-play coach, by slideshow or by video, with delivery rated blind by two independent assessors. Rehearsal came out ahead: Cohen's d of 0.75 against slides and 1.08 against video. Participants also rated their own confidence on a 1 to 7 scale before the input, after it, and after the simulated conversation. Slides: 4.46, 5.00, 5.39. Video: 4.69, 5.13, 5.38. Practice: 4.27, 5.33, 5.53. Confidence rose with time in every group; the format made no significant difference (p = .958), nor did the interaction of time and format (p = .541). Source: Laumann (2020), UCL, unpublished research paper, Figure 4.
The two studies measure different things, memory for prose and the delivery of a feedback conversation, and a non-significant difference is not proof of sameness. The narrower parallel still holds: in both, what people reported about their own learning did not track the measured result. Design and limits are on our research page.
What our study cannot say about a week later
One limit matters most here. Our study scored delivery once, immediately after the intervention, and the practice condition was a single live coaching session of at most 20 minutes over a video call, not our current programme. In Roediger and Karpicke's terms, we measured at five minutes. We did not measure at a week, so the study cannot say whether rehearsal's lead grows, holds or shrinks. It found rehearsal ahead at the horizon that favoured the passive method in their data, and that is as far as it goes: one study of one skill, not replicated.
Two kinds of repetition, and only one is for retention
Here is the design consequence. In our new manager programme, day two closes with five one-to-one conversations, each with a professional actor as the direct report. Each takes around fifteen minutes: five playing it, five on the debrief, five rewinding to the moment that mattered and playing it again. That replay is a same-day repetition. What it is for is set out on the programme page: the first attempt shows the default, the debrief makes it visible, and the second attempt shows whether the manager can do something else when the pressure comes back. Making the skill last is a different job, and the five-minute result says a same-day repeat will look good whichever way you run it. What the replay shows is read in the room, not on a form afterwards.
The repetition that decides whether the behaviour is still there weeks later has to sit at a distance, in the job, and it has to be a retrieval rather than a re-read: the manager holding the conversation again, not re-reading the model. Lacerenza and colleagues' 2017 meta-analysis of leadership training, across 335 independent samples, found spaced training sessions and practice among the design features associated with more effective training. A working template for the eight weeks after a delivery day is in our 90-day plan; the wider case is in does a one-day training last.
What to do with it
Three things follow. Do not evaluate the day on the day; in Roediger and Karpicke's data the five-minute test favoured re-reading, the method that trailed a week later, and in our own study confidence rose in every format. Treat a same-day replay as a way of seeing the default, not as what makes the skill last. Put the repetition that counts at a distance from the room, make it a performance rather than a review, and measure what was observed, not what was reported. How to run that without a research department is in how to measure behaviour change.
After Five Minutes, Re-reading Wins: The Takeaways
In Roediger and Karpicke's 2006 experiments, re-reading beat retrieval practice on a test five minutes later; two days and a week later, retrieval was ahead, 61 per cent against 40 after a week in Experiment 2. The students who had only re-read were also the most confident. For training design the lesson is timing and measurement: a same-day repeat looks good whatever you do, the repetition that decides retention sits days or weeks later, and neither can be judged by how ready people feel.
- Roediger and Karpicke (2006): five minutes after learning, re-reading beat retrieval practice (81 against 75 per cent in Experiment 1); after a week, retrieval was ahead (56 against 42 in Experiment 1, 61 against 40 in Experiment 2).
- In Experiment 2, the students who had only re-read predicted the best recall and delivered the worst.
- In our randomised 2020 UCL study, self-rated confidence rose in every format with no significant difference between them, while blind-rated delivery favoured rehearsal (d = 0.75 against slides, 1.08 against video).
- Our study measured immediately after the intervention and says nothing about retention weeks later.
- A same-day replay shows the default and whether a second attempt differs. The repetition that decides retention sits days or weeks later, in the job, as a performance rather than a re-read.
- Dozens of earlier posts on this site cite the study and none mentioned the five-minute result; this post sets it out in full.
Frequently Asked Questions
Does the testing effect mean practice is always better than reviewing?
No. In Roediger and Karpicke's experiments, re-reading produced better recall on a test five minutes after learning, and only on the delayed tests, at two days and a week, did retrieval practice pull ahead. The effect is about long-term retention. Anyone quoting the one-week figure without the five-minute one is quoting half the study.
Why does confidence rise even when the method is weaker?
Because learners base their predictions of future performance on what produces rapid short-term gains, as Roediger and Karpicke put it. In Roediger and Karpicke's data, the students who had re-read the passage most often were the most confident and remembered the least a week later. In our own randomised UCL study, self-rated confidence rose in all three formats with no significant difference between them, while blind-rated delivery differed. Neither study supports using confidence ratings as evidence that training worked.
When should a rehearsal be repeated?
Twice, for two different reasons. A same-day replay, as in our new manager programme, shows the default and whether a second attempt differs. The repetition that decides whether a behaviour holds has to happen days or weeks later, in the job, as an attempt rather than a review. Lacerenza and colleagues' 2017 meta-analysis found spaced sessions and practice among the design features associated with more effective leadership training.
Sources: Roediger and Karpicke (2006), Test-Enhanced Learning, Psychological Science 17(3) · Lacerenza, Reyes, Marlow, Joseph and Salas (2017), Leadership training design, delivery, and implementation: a meta-analysis, Journal of Applied Psychology 102(12) · Sidestream, The Study Behind the Method · Sidestream, New Manager Training · Sidestream, Does One-Day Training Last?