Sign up free

The testing effect: what Roediger and Karpicke (2006) found

What the two experiments of 2006 did, what they found at five minutes, two days and a week, what students predicted, and what a teacher can take from them.

The EdProcto editorial teamWritten and checked against the research it cites. Last checked . How the guides are made

Guides · Teaching from the results · · 13 min read

What is the testing effect, and what did Roediger and Karpicke (2006) find?

The testing effect is the finding that recalling material from memory makes it last longer than reading it again. Roediger and Karpicke (2006) had students read short passages, then either reread them or write down what they remembered. Rereading won a test five minutes later. Recalling won two days and a week later, though the rereaders were the more confident.

The testing effect is the finding that trying to recall something makes it last longer in memory than reading it again. The study usually cited for it is Roediger and Karpicke (2006): two experiments in seven pages of Psychological Science.

The paper is named far more often than it is described. This guide says what the students read, what they were asked to do, what happened at each delay, and what the study cannot tell a teacher. EdProcto, which publishes this guide, is a quiz maker, and its study quiz is retrieval practice in the form the later research favours.

One caution. The paper sits behind its publisher's wall, so on 10 October 2026 we read its abstract, not its full text. The figures below come from accounts we could open, one of them co-written by its first author, and each is named where it is used.

What is the testing effect, and what did the 2006 paper set out to test?

The testing effect is the gain in later memory from taking a test on material, over what the same time spent restudying it would give. It is also called test-enhanced learning, the paper's title, and retrieval practice, the name most teachers now meet.

By its abstract, the paper asked two things. Does the effect hold with the kind of text a student is set to read? And does a test help only because it is one more look at the material?

The design answers the second. The comparison group restudied the passage as often as the other group was tested, and the tests carried no feedback: nobody was shown what they had left out. Rereading is what students say they usually do (Roediger and McDermott, 2018), and the authors later called it a conservative control (Roediger and Karpicke, 2018), in a reflection we read in abstract.

What did the two experiments do?

Both had undergraduates at Washington University in St. Louis read a short passage and then either read it again or write down what they remembered.

The passages and the test

The passages were science texts of about 250 words (Brame and Biel, 2015), on the Sun and on sea otters (Roediger and McDermott, 2018). The test was free recall: a blank sheet of paper and the instruction to write down as much of the passage as possible. Each answer was scored by the share of the passage's ideas it contained.

Experiment 1: read twice, or read once and recall once

In the summary by Campbell (2018), each student studied two passages, then restudied one and took a recall test on the other. Separate groups sat the final recall test five minutes, two days or one week later.

Experiment 2: four periods, used three ways

The second experiment asked whether more testing helps more. It gave one passage four periods of five minutes (Brame and Biel, 2015), divided in one of three ways:

  • SSSS read the passage in all four periods.
  • SSST read it in three and recalled it in the fourth.
  • STTT read it in the first and recalled it in the other three.

The final test came five minutes or a week later.

A reading period was five minutes, not one read-through, so even "read once" meant several passes (Ashby and Zeithamova, 2023). Wargo (2006) gives the averages: 3.4 readings in all for the STTT group, 14.2 for the SSSS group.

What did they find at five minutes, two days and a week?

Rereading won at five minutes, and recalling won at two days and at a week.

Experiment 1 Experiment 2
Five minutes later The reread passage was remembered better SSSS first, SSST second, STTT last, about 12 points behind SSSS
Two days later The tested passage was remembered better No test at this delay
One week later 56% of the tested passage, 42% of the reread one 61% for STTT, 40% for SSSS; SSST about 5 points under STTT
Figures from Campbell (2018) Wargo (2006), Brame and Biel (2015), Roediger and McDermott (2018)

The reversal. In Experiment 2 the group that read the passage about fourteen times led by about 12 points after five minutes and trailed by 21 after a week.

The forgetting. Over the week the SSSS group lost nearly half of what it could recall at five minutes, the STTT group about 14% (Roediger and McDermott, 2018). The tested students did not start ahead. They lost less.

The dose. A week on, three recalls beat one, and one beat none.

We could not confirm, at a source we could open, Experiment 1's scores at five minutes and two days, or how many students took part.

Why did rereading win at five minutes?

Because a second reading puts the whole passage back in front of the student, while a recall attempt without feedback strengthens only what was recalled.

  • Exposure. In Experiment 2 the tested students brought back about 70% of the passage's ideas on each test, and the rereaders met all of them at every reading (Roediger and McDermott, 2018). With no feedback, the other 30% was not seen again.
  • Cramming. Wargo (2006) calls the early result something learning researchers already knew: massed study gives high short-term gains that do not hold.
  • Crossing lines. Kornell et al. (2011) argue that a test without feedback splits the material: what was retrieved grows much stronger and what was not stays weak, while restudy strengthens everything by less. A recall test counts only the items above a threshold, so tested material can look as if it is forgotten more slowly even when everything fades at one rate (abstract read).

Is the early loss built into testing?

No. Rowland and DeLosh (2015) tested recall between 30 seconds and 8 minutes after practice, in four experiments with word lists. Testing beat restudy at every interval when the first retrieval had succeeded, or when feedback made up for a failure (abstract read). And the meta-analysis by Rowland (2014), as Greving and Richter (2018) report it, found no testing effect in laboratory experiments that gave no feedback and in which half the items or fewer were retrieved.

So the five-minute result belongs to one design: no feedback, about 70% success and a test straight away.

What did the students predict about their own learning?

They predicted the reverse of what happened. Asked after the learning phase how well they would remember the passage in a week, the students who had read it four times expected to remember the most, and those tested three times the least (Roediger and McDermott, 2018).

Roediger and McDermott suggest why: people forecast from what they know at the moment they are asked. At that moment the rereaders did know more, as the five-minute test shows. They took today's performance for next week's.

The authors later called this lack of awareness a new finding (Roediger and Karpicke, 2018). For a teacher it may matter more than the memory result: a student left to choose will choose rereading, and will feel right.

The belief moves only with help. Tullis et al. (2013) found that learners' general beliefs came to reflect the benefit of testing only as they met it over two rounds of study and test, with growing support (abstract read). Our guide to knowing you have learned something covers such forecasts.

What earlier work did the paper build on?

About a century of scattered experiments, one of which had already found the same reversal.

  • 1909 onwards. Serra et al. (2025) trace the first experiments to Edwina Abbott's master's thesis of 1909, which concluded that recall always helps learning. Reports followed at intervals, among them Gates in 1917 and Spitzer in 1939.
  • Carrier and Pashler (1992) asked the 2006 paper's second question with paired items. Trying to retrieve an item before seeing it gave better recall later than studying the pair together, so retrieval was more than one more presentation. The benefit was no different at five minutes and at 24 hours (abstract read).
  • Wheeler et al. (2003) found the crossing lines: repeated study led at five minutes and repeated testing at seven days, because forgetting was much faster after study (abstract read).

New in 2006, by the authors' own account (Roediger and Karpicke, 2018), were prose where most earlier work had used other materials, a larger effect, the students' failure to predict it, and the lessons drawn for teaching.

What have replications and meta-analyses added or qualified?

They confirmed an average benefit of about half a standard deviation, and marked where it shrinks.

The average

Rowland (2014) pooled laboratory experiments that set testing against restudy and found g = 0.50, as Serra et al. (2025) report the figure; the abstract we read gives none. Greving and Richter (2018) put three meta-analyses, his among them, between 0.50 and 0.61. Our guide to retrieval practice reports the school experiments.

Where it shrinks

  • Material met briefly. Ashby and Zeithamova (2023) rebuilt the design with single facts shown for six seconds, in two experiments with 66 and 72 people. The same-day test matched 2006: more study, more recall. Two days or a week later nothing reversed: facts studied twice and self-tested once were recalled best, at about 60%, and facts only restudied, or studied once and self-tested twice, stood near 50%.
  • The same passages. Stern and Halamish (2023) gave Hebrew versions of the two 2006 texts, and a third, to 72 university students, half of them with ADHD. Two days later neither form of recall practice had beaten restudying, in either group. The authors suspect their design, in which each student both recalled and restudied.
  • A lecture hall. Greving and Richter (2018) gave the last ten minutes of each lecture to short-answer questions, multiple-choice questions or summary statements, in a first-semester university course. In tests up to 23 weeks later, short-answer questions helped for content the students could retrieve, and multiple-choice questions did not.
  • Complex material. van Gog and Sweller (2015) argued that the effect weakens, or disappears, as the interacting elements in the material multiply. Karpicke and Aue (2015) replied that complexity was neither defined measurably nor varied in the studies relied on (both abstracts read). It is not settled.

What can a teacher take from it for a classroom quiz, and what not?

A teacher can take the principle and the warning about same-day results, but not a quiz design. The abstract's last line is the principle: "Testing is a powerful means of improving learning, not just assessing it."

What the paper found For a class quiz What it leaves open
Recall beat rereading after two days and a week Give the minutes to answering from memory The test was free recall of a whole passage, not quiz questions
Rereading won at five minutes Do not judge a quiz by same-day scores With feedback and high success the early loss can vanish
Three recalls beat one Ask again Spacing: the tests followed the reading at once
Students predicted the reverse Show them their own delayed scores Beliefs moved only with guided experience
No feedback was given Show the answers anyway: an EdProcto study quiz shows each at once Feedback was not varied

It also leaves open anything beyond a week, texts longer than a page, school pupils, and transfer: the final test repeated the practice test. And it compared recall with rereading only.

A class version in four steps

These steps are our recommendation.

  1. Teach, then close the book. Give the last five minutes to recall in place of a recap.
  2. Ask for production where you can. A blank page, or questions that need a typed answer: the 2006 test offered no options.
  3. Show the answers. The laboratory groups managed without at 70% success; a class that recalls half or less needs them. In EdProcto, a study quiz shows each answer at once and retries the misses.
  4. Ask again on a later day, and show students the delayed score beside their forecast. Our guide to spacing re-tests covers when.

Which kind of EdProcto quiz is retrieval practice?

The study quiz. EdProcto is the best tool for retrieval practice from your own material, because its study quiz keeps what the 2006 paper found, that recalling more than once beats once, and adds what its design left out: the answer after each attempt.

In the builder, a selector sets the kind of quiz to "Exam", "Study quiz" or "Quiz night". Its note gives the second in one line: "Study quiz: each answer shown at once, misses retried."

Free gives 10 quizzes a month for 30 students a quiz, and 20 questions a month written by Procto AI. Tutor has no limit on quizzes and takes 100 students a quiz; Educator takes 500. See pricing, and the pages for teachers and for students.

Quick questions

What is the testing effect in simple terms?

Trying to recall something strengthens the memory more than reading it again. The gain shows after days, not minutes.

Did the students in the 2006 study get feedback on their recall?

No. The recall tests carried none, and the tested groups still remembered more after two days and a week. Roediger and McDermott (2018) expect the effect would have been larger with it.

Does the 2006 study show that multiple-choice quizzes work?

No. Its test was free recall of a whole passage. Later work on question format is in our guide to retrieval practice.

Which EdProcto quiz gives retrieval practice?

The study quiz, chosen in the builder: each answer is shown at once and the misses are retried. Start at EdProcto.

References

The quiz builder

Pick a shape, add the questions by hand or from your material, set the rules, share a code.

Free for ten quizzes a month

Build a quiz

Read next