ArticlePDF Available

Abstract and Figures

A powerful way of improving one's memory for material is to be tested on that material. Tests enhance later retention more than additional study of the material, even when tests are given without feedback. This surpris- ing phenomenon is called the testing effect, and although it has been studied by cognitive psychologists sporadically over the years, today there is a renewed effort to learn why testing is effective and to apply testing in educational settings. In this article, we selectively review laboratory studies that reveal the power of testing in improving re- tention and then turn to studies that demonstrate the basic effects in educational settings. We also consider the related concepts of dynamic testing and formative assess- ment as other means of using tests to improve learning. Finally, we consider some negative consequences of testing that may occur in certain circumstances, though these negative effects are often small and do not cancel out the large positive effects of testing. Frequent testing in the classroom may boost educational achievement at all levels of education. In contemporary educational circles, the concept of testing has a dubious reputation, and many educators believe that testing is overemphasized in today's schools. By ''testing,'' most com- mentators mean using standardized tests to assess students. During the 20th century, the educational testing movement produced numerous assessment devices used throughout edu- cation systems in most countries, from prekindergarten through graduate school. However, in this review, we discuss primarily the kind of testing that occurs in classrooms or that students engage in while studying (self-testing). Some educators argue
Content may be subject to copyright.
The Power of Testing Memory
Basic Research and Implications for Educational Practice
Henry L. Roediger, III, and Jeffrey D. Karpicke
Washington University in St. Louis
ABSTRACT—A powerful way of improving one’s memory for
material is to be tested on that material. Tests enhance
later retention more than additional study of the material,
even when tests are given without feedback. This surpris-
ing phenomenon is called the testing effect, and although it
has been studied by cognitive psychologists sporadically
over the years, today there is a renewed effort to learn
why testing is effective and to apply testing in educational
settings. In this article, we selectively review laboratory
studies that reveal the power of testing in improving re-
tention and then turn to studies that demonstrate the
basic effects in educational settings. We also consider the
related concepts of dynamic testing and formative assess-
ment as other means of using tests to improve learning.
Finally, we consider some negative consequences of testing
that may occur in certain circumstances, though these
negative effects are often small and do not cancel out the
large positive effects of testing. Frequent testing in the
classroom may boost educational achievement at all levels
of education.
In contemporary educational circles, the concept of testing has a
dubious reputation, and many educators believe that testing is
overemphasized in today’s schools. By ‘‘testing,’’ most com-
mentators mean using standardized tests to assess students.
During the 20th century, the educational testing movement
produced numerous assessment devices used throughout edu-
cation systems in most countries, from prekindergarten through
graduate school. However, in this review, we discuss primarily
the kind of testing that occurs in classrooms or that students
engage in while studying (self-testing). Some educators argue
that testing in the classroom should be minimized, so that valu-
able time will not be taken away from classroom instruction.
The nadir of testing occurs in college classrooms. In many
universities, even the most basic courses have very few tests,
and classes with only a midterm exam and a final exam are
common. Students do not like to take tests, and teachers and
professors do not like to grade them, so the current situation
seems propitious to both parties.
The traditional perspective of educators is to view tests and
examinations as assessment devices to measure what a student
knows. Although this is certainly one function of testing, we
argue in this article that testing not only measures knowledge,
but also changes it, often greatly improving retention of the
tested knowledge. Taking a test on material can have a greater
positive effect on future retention of that material than spending
an equivalent amount of time restudying the material, even when
performance on the test is far from perfect and no feedback is
given on missed information. This phenomenon of improved
performance from taking a test is known as the testing effect, and
though it has been the subject of many studies by experimental
psychologists, it is not widely known or appreciated in educa-
tion. We believe that the neglect of testing in educational circles
is unfortunate, because testing memory is a powerful technique
for enhancing learning in many circumstances.
The idea that testing (or recitation, as it is sometimes called in
the older literature) improves retention is not new. In 1620,
Bacon wrote: ‘‘If you read a piece of text through twenty times,
you will not learn it by heart so easily as if you read it ten times
while attempting to recite from time to time and consulting the
text when your memory fails’’ (F. Bacon, 1620/2000, p. 143). In
the Principles of Psychology, James (1890) also argued for the
power of testing or active recitation:
A curious peculiarity of our memory is that things are impressed
better by active than by passive repetition. I mean that in learning
(by heart, for example), when we almost know the piece, it pays
better to wait and recollect by an effort from within, than to look at
the book again. If we recover the words in the former way, we shall
probably know them the next time; if in the latter way, we shall very
likely need the book once more. (p. 646)
Bacon and James were describing situations in which students
test themselves while studying. We show later that their hy-
potheses are correct and that testing greatly improves retention
of material. However, we need to make a distinction between two
Address correspondence to Henry L. Roediger, III, or to Jeffrey D.
Karpicke, Department of Psychology, Box 1125, Washington Uni-
versity in St. Louis, One Brookings Dr., St. Louis, MO 63130-4899,
e-mail: or
Volume 1—Number 3 181Copyright r2006 Association for Psychological Science
types of effects that testing might have on learning: mediated (or
indirect) effects and direct (unmediated) effects. Let us consider
mediated effects first, because testing can enhance learning in
a variety of ways. To give just a few examples, frequent testing
in classrooms encourages students to study continuously
throughout a course, rather than bunching massive study efforts
before a few isolated tests (Fitch, Drucker, & Norton, 1951).
Tests also give students the opportunity to learn from the feed-
back they receive about their test performance, especially when
that feedback is elaborate and meaningful, as is the case in the
technique of formative assessment, discussed in a later section.
In addition, if students test themselves periodically while they
are studying (as Bacon and James advocated long ago), they may
use the outcome of these tests to guide their future study toward
the material they have not yet mastered. The facts that testing
encourages students to space their studying and gives them
feedback about what they know and do not know are good rea-
sons to recommend frequent testing in courses, but they are not
the primary reasons we focus on in this article. In these cases of
mediated effects of testing, it is not the act of taking the test itself
that influences learning, but rather the fact that testing promotes
learning via some other process or processes. For example, when
a test provides feedback about whether or not students know
particular items and the students guide their future study efforts
accordingly, testing promotes learning by making later studying
or encoding more effective; thus, testing enhances learning by
means of this mediating process.
These examples of mediated effects of testing serve as addi-
tional evidence in favor of the use of frequent testing in edu-
cation. However, our review is focused on direct effects of testing
on learning—the finding that the act of taking a test itself often
enhances learning and long-term retention. In many of the ex-
periments we describe, one group of students studied some set of
materials and then was given an initial test (or sometimes re-
peated tests). Retention of the material was assessed on a final
criterial test, and the tested group’s performance was compared
with that of one or two control groups. In one type of control,
students studied the material and took the final test just as the
tested group did, but were not given an initial test. In a second
type of control (a restudy control), students studied the material
just as the tested group did, but then studied the material a
second time when the tested group received the initial test; in
this case, total exposure time to the material was equated for the
tested and control groups. The typical finding throughout the
literature is that the tested group outperforms both kinds of
control groups (the no-test control and the restudy control) on the
final test, even when no feedback is given after the initial test. In
variations on this prototypical experiment, the effects of several
variables have been investigated (e.g., the materials to be
learned, the format of the initial and final tests, whether or not
subjects receive feedback on the first test, the time interval
between studying and initial testing, and the retention interval
before the final test, to name but a few). As we show, across a
wide variety of contexts, the testing effect remains a robust
The direct effects of testing are especially surprising when
exposure time is equated in the tested and study conditions,
because although the repeated-study group experiences the
entire set of materials multiple times, the students in the tested
group can experience on the test only what they are able to
produce, at least when the test involves recall. Yet despite the
differences in initial exposure favoring the study group, the
tested group performs better in the long term. That the testing
effect is so counterintuitive helps explain why it remains un-
known in education. The direct effects of testing on learning are
not purely a result of additional exposure to the material, which
indicates that processes other than additional studying are re-
sponsible for them. The testing effect represents a conundrum, a
small version of the Heisenberg uncertainty principle in psy-
chology: Just as measuring the position of an electron changes
that position, so the act of retrieving information from memory
changes the mnemonic representation underlying retrieval—
and enhances later retention of the tested information.
In this article, we review research from both experimental and
educational psychology that provides strong evidence for the
direct effect of testing in promoting learning. After presenting
two classic studies, we consider evidence from laboratories of
experimental psychologists who have investigated the testing
effect. As is the experimentalists’ predilection, they have typi-
cally used word lists as materials, college students as subjects,
and standard laboratory tasks such as free recall and paired-
associate learning (see Cooper & Monk, 1976; Richardson,
1985; and Dempster, 1996, 1997, for earlier and somewhat more
focused reviews). Effects on later retention are usually quite
large and reliable. We next consider studies conducted in more
educationally relevant situations. Such studies often use prose
passages about science, history, or other topics as the subject
matter and investigate the effects of tests more like those found
in educational settings (e.g., essay, short-answer, and multiple-
choice tests). Once again, we show that testing promotes strong
positive effects on long-term retention. We also review studies
carried out in actual classrooms using even more complex ma-
terials, and they again show positive effects of testing on learning.
After concluding our review of basic research findings, we
provide an overview of theoretical approaches that have been
directed toward explaining the testing effect, although many
puzzles about testing have not been satisfactorily explained. We
then consider the related approaches of dynamic testing (e.g.,
Sternberg & Grigorenko, 2002) and formative assessment (e.g.,
Black & Wiliam, 1998a), which are both aimed at using tests to
promote learning by altering instructional techniques on the
basis of the results of tests (i.e., mediated effects of testing).
Because testing does not always have positive consequences, we
next review two possible negative effects (retrieval interference
and negative suggestibility) that need to be considered when
using tests as possible learning devices. Finally, we discuss
182 Volume 1—Number 3
The Power of Testing Memory
common objections to increased use of testing in the classroom,
and we tell why we believe that none of these objections out-
weighs our recommendations for frequent testing.
Gates (1917) and Spitzer (1939) published two classic studies
showing strong positive effects of testing on retention. Both were
rather heroic efforts, and so it is unfortunate that neither is ac-
corded much attention in the contemporary literature. Although
other research showing the benefits of testing appeared before
Gates’s work (e.g., Abbott, 1909; Thorndike, 1914), he carried
out the first large-scale study. Gates tested groups of children
across a range of grades (Grades 1, 3, 4, 5, 6, and 8), and, ad-
mirably, he used two different types of materials (nonsense
syllables, the classic stimulus of Ebbinghaus, 1885/1964, and
brief biographies taken from Who’s Who in America). The chil-
dren studied these materials during a two-phase learning pro-
cedure. In the first phase, they simply read the materials to
themselves, whereas in the second phase, the experimenter in-
structed them to look away from the materials and try to recall
the information to themselves (covert recitation). During the
recitation phase, the students were permitted to glance back at
the materials when they needed to refresh their memories. Al-
though this feature of the design relaxed experimental control, it
probably faithfully captured what students do when using a
recitation or testing strategy to study.
Gates (1917) manipulated the amount of time the children
spent reciting by instructing them to stop reading and start re-
citing after different amounts of study time had elapsed. Dif-
ferent groups of children at each age level spent 0, 20, 40, 60,
80, or 90% of the learning period involved in recitation, or self-
testing. Finally, at the end of the period, Gates gave the children
a test, asking them to write down as many items as they could in
order of appearance. He then retested the children 3 to 4 hr later.
Gates’s (1917) basic results are shown in Figure 1, which
shows that in almost all conditions, he obtained positive effects
of recitation. With nonsense syllables, all groups except first
graders showed a strong effect of recitation. For the biographical
materials, all groups showed a recitation effect, but one that was
less dramatic on the initial tests than on the delayed tests. (Note
that first graders were not tested with prose passages because
their reading abilities were so poor.) With prose passages, the
optimal amount of recitation seemed to be about 60% of the total
learning period. Gates concluded that recall attempts during
learning (recitation with restudy of forgotten material) are a good
way to promote learning. He argued that these results had im-
portant implications for educational practice and described
ways to incorporate recitation into classroom exercises (Gates,
1917, pp. 99–104). However, Gates’s work pointed to limitations
of recitation/self-testing, too. First graders did not show the ef-
fect, which suggests that it may occur only after a certain point in
development. Also, with prose passages, the effect of recitation
leveled off and even appeared to drop when the amount of time
spent on recitation exceeded 60%, and consequently study time
was less than 40%. Thus, the data suggest that a certain amount
of study may be necessary before recitation or testing can begin
to benefit learning.
A second landmark study showing positive effects of testing
was carried out by Spitzer (1939) in his dissertation work. His
experiment involved testing the entire population of sixth-grade
students in 91 elementary schools in nine Iowa cities—a total of
3,605 students. The students studied 600-word articles (on
peanuts or bamboo) that were similar to material they might
study in school, and then they took tests according to various
schedules across the next 63 days. Each test consisted of 25
multiple-choice items with five alternatives (e.g., ‘‘To which
family do bamboo plants belong? A) trees, B) ferns, C) grasses,
D) mosses, E) fungi’’). Some students took a single test 63 days
later, whereas others also took earlier tests so that Spitzer could
see what effect these would have on later tests. Several inter-
esting patterns could be discerned in the results, which are
shown in Figure 2. First, the dashed line shows a beautiful
forgetting curve in that the longer the first test was delayed, the
worse was performance on that test. Second, giving a test nearly
stopped forgetting; when students were given a first test and then
retested at a later time, their performance did not drop much at
all (and sometimes increased). Third, the sooner the initial test
was given after study, the better students did on later tests. For
example, Group 2 was tested immediately after study and then a
week later. When tested again 56 days later (day 63), they
showed much better performance than Group 6 (which was not
tested initially until Day 21). In fact, because forgetting had
reached asymptote by Day 21, the first test taken by Group 6 did
not enhance later recall at all. The lesson from Spitzer’s study is
that a first test (without feedback) must be given relatively soon
after study (when the student still can recall or recognize the
material) in order to have a positive effect at a later time.
The studies by Gates (1917) and Spitzer (1939) were among
the most extensive in their times (although see Jones, 1923–
1924, for another impressive study), and in some features the
experimental techniques would not hold up to today’s standards.
However, the essential points Gates and Spitzer made are secure
because later researchers replicated their results. For example,
Forlano (1936) replicated Gates’s work by demonstrating that
testing improved children’s learning and spelling of vocabulary
words, and Sones and Stroud (1940) replicated Spitzer’s (1939)
research, albeit on a smaller scale. However, around 1940, in-
terest in the effects of testing on learning seemed to disappear.
We can only speculate as to why. One reason may be that with the
rise of interference theory (McGeoch, 1942; Melton & Irwin,
1940; see Crowder, 1976, chap. 8), interest swung to the study of
forgetting. For the purpose of measuring forgetting, repeated
testing was deemed a confound to be avoided because, as Figure
2 shows, an initial test interrupts the course of forgetting.
McGeoch (1942, pp. 359–360), Hilgard (1951, p. 557), and
Volume 1—Number 3 183
Henry L. Roediger, III, and Jeffrey D. Karpicke
Deese (1958) all argued against the use of repeated-testing
designs. For example, Deese wrote that ‘‘an experimental study
of this sort yields very impure measures of retention after the first
test, since all subsequent measures are contaminated by the
practice the first test allows’’ (pp. 237–238). This statement is
true for the study of forgetting, but of course, for studying the
effects of tests per se, repeated testing is necessary, and the
‘‘contamination’’ that Deese referred to is the phenomenon of
interest. Nevertheless, leading experimental psychologists’ at-
titude against repeated-testing designs probably halted the
study of testing effects (and the study of phenomena such as
reminiscence and hypermnesia, which also require repeated
testing; W. Brown, 1923; Erdelyi & Becker, 1974; Roediger &
Challis, 1989).
One venerable topic in experimental-cognitive psychology is
how and why learning occurs. The traditional way of studying
learning is through alternating study and test trials. For exam-
ple, in multitrial free-recall learning, students typically study a
list of words (a study trial), recall as many as possible in any
order (a test trial), study the list again, recall it again, and so on
through numerous study-test cycles (e.g., Tulving, 1962). When
data are averaged across subjects, a regular, negatively accel-
erated learning curve is produced (e.g., see Fig. 3, which pre-
sents results of a study we discuss in the next section).
A controversy about the nature of learning erupted in the late
1950s and early 1960s. Some theorists believed that learning of
Fig. 1. Proportion of nonsense syllables and biographical facts recalled by children on immediate and delayed tests as a function of the amount of time
spent reciting the material. Adapted from data reported by Gates (1917).
184 Volume 1—Number 3
The Power of Testing Memory
individual items occurs through an incremental process (the
standard view), and others argued that learning is all-or-none
(Rock, 1957). The incremental-learning position held that each
item in the list is represented by a trace that is strengthened a bit
by each successive repetition; once enough strength is accrued
via repetitions so that some threshold is crossed, an item will be
recalled. The all-or-none position held that on each study trial, a
subset of items jumps from zero strength to 100% strength in a
step function—hence ‘‘all or none.’’ In this view, the fact that
learning curves appear to be smooth is an artifact of averaging,
and performance would actually be all-or-none if the fate of each
item could be examined separately. This controversy about the
nature of the learning process raged on in some circles
throughout the 1950s and into the 1960s and was never com-
pletely decided, although the incrementalist assumption is still
largely built into today’s theories. Tulving (1964) noted that in
one sense the controversy was beside the point, because each
item in such an experiment is perfectly learned when it is first
presented, in the sense that it can be recalled perfectly imme-
diately after its presentation. Thus, learning is always ‘‘all,’’ and
the critical issue is why students forget items on the subsequent
test (i.e., why there is intratrial forgetting).
The reason for bringing up this controversy in the current
context is to examine a hidden assumption. Both the incre-
mentalist and the all-or-none positions make the assumption
that learning occurs during study trials, when students are ex-
posed to the material, and that the test trials simply permit
students to exhibit what they have learned on previous study
trials. This is essentially the same attitude that teachers take
toward testing in the classroom: Tests simply are assessment
devices. An experiment by Tulving (1967) called this assump-
tion into question and helped usher in a new wave of research on
Tulving (1967) had subjects learn lists of 36 words, which
were presented in a different random order on every study trial,
and then take free-recall tests (subjects recalled out loud as
many items as possible in any order, and the experimenter re-
corded responses). In the standard learning condition, students
saw the list, recalled it, saw it, recalled it, and so on for 24 trials.
If S stands for a study trial and T stands for a test trial, then the
standard condition can be represented as STST STST . . . (for a
total of 12 study trials and 12 test trials). Tulving considered
every 4 trials a cycle, for reasons that will be clear when the
other conditions are described. In the repeated-study condition,
each cycle consisted of 3 study trials and 1 test trial (SSST
SSST . . .). If subjects learned only during the study trials, then
by the end of learning, performance should have been much
better in this condition than in the standard condition, because
there were 6 more study trials (18 study trials and 6 test trials
over the six cycles). In the repeated-test condition, each cycle
Fig. 2. Proportion correct on multiple-choice tests taken at various delays after studying. After
studying the passage, each of the eight groups of subjects was given one, two, or three tests on various
schedules across the next 63 days. The solid lines show results for repeated tests for particular
groups, and the dashed line represents normal forgetting as the delay between studying and testing
increases. Adapted from data reported by Spitzer (1939).
Volume 1—Number 3 185
Henry L. Roediger, III, and Jeffrey D. Karpicke
contained 1 study trial followed by 3 consecutive test trials
(STTT STTT . . .), leading to a total of only 6 study trials and 18
test trials during the entire learning phase. By the common as-
sumption that learning occurs only during study trials, subjects
in the repeated-test condition should have been at a great dis-
advantage relative to those in the other two conditions.
The surprise in Tulving’s (1967) research was that the learning
curves of all three conditions looked about the same. For ex-
ample, by the end of the experiment, subjects recalled about 20
words in the standard and the repeated-study conditions, even
though subjects in the repeated-study condition had studied the
words six more times. The subjects in the repeated-test condi-
tion recalled somewhat fewer words, finishing at about 18.5
words. This slight difference is probably partly explained by the
fact that these subjects were deprived of using primary or short-
term memory (Glanzer & Cunitz, 1966). That is, subjects in the
standard and repeated-study conditions had just heard the list
before the very last test trial, so they could use primary memory
to recall the last few items. Subjects in the repeated-test con-
dition could not do this, because they had just had two other tests
before their last test, and so the short-term component of recall
would no longer have been accessible. Given this procedural
difference among conditions, it is remarkable that the learning
curves of the three conditions were so similar. Apparently,
within rather wide limits (6, 12, or 18 study trials), a study trial
can be replaced by a test trial. In other words, just as much
learning occurs on a test trial as on a study trial. Of course, as a
limiting case, there must be some study opportunities before
testing can have an effect (as noted by Gates, 1917), but the
surprise is how wide the variability is. There were only 6 study
trials in the repeated-test condition, and yet final recall was
nearly as good as with 18 study trials (in the repeated-study
condition). In our own research, which we review later (Karpicke
& Roediger, 2006b), we have shown that if long-term retention is
measured after a delay, the repeated-test condition actually
shows better recall than the repeated-study condition, a finding
that is even more counterintuitive given the customary as-
sumptions about the role of study and test trials in learning.
Tulving’s (1967) results seemed hard to believe when they first
appeared, which is probably why so many researchers imme-
diately tried to replicate them with minor variations, creating a
boomlet in testing research that lasted briefly in the early 1970s,
followed by sporadic work thereafter. In the title of their article,
Lachman and Laughery (1968) asked, ‘‘Is a test trial a training
trial in free recall learning?’’ and they answered ‘‘yes’’ from their
data. Other researchers also replicated Tulving’s work, using his
conditions or slight variations thereof (Birnbaum & Eichner,
1971; Donaldson, 1971; Rosner, 1970). One methodological
detail of Tulving’s work and of these replications was unusual.
Because Tulving wanted to equate the time of study and test
trials, and because he made the presentation rate for words
rather fast in the study trials, the duration of the test trials was
short. He presented the 36 words at a 1-s rate during study trials,
and so he also gave subjects only 36 s to recall the words during
test trials. Even with spoken recall, this is a short time to recall
36 words even if they are well learned. In light of later work
examining how free recall unfolds over time, tests lasting this
long might greatly underestimate the amount of knowledge
subjects have acquired (e.g., Roediger & Thorpe, 1978). The
short recall time may also explain why subjects were able to
recall only about 20 of 36 words after 24 study or test trials; in all
probability, they simply did not have time to recall all they knew.
We (Karpicke & Roediger, 2006b) recently conducted an
experiment with Tulving’s three conditions (standard, repeated-
study, and repeated-test), but using 40 words and a 3-s rate of
presentation, so that the accompanying tests lasted 2 min and
time on study trials and recall tests remained equated. We ex-
amined learning curves and compared the conditions on the five
common test positions out of the total of 20 study and test trials.
That is, every 4th trial was a test trial for all three conditions
(standard: STST . . . ; repeated-study: SSST . . . ; and repeated-
test: STTT . . .), so we could directly compare recall on the 4th,
8th, 12th, 16th, and 20th trials across the three conditions. We
also eliminated short-term memory effects that would normally
disadvantage the repeated-test condition by using Tulving and
Colotla’s (1970) method of separating short-term from long-term
memory effects. (Watkins, 1974, concluded that this technique
was the best method for this purpose.) Finally, we provided a
Fig. 3. Proportion of words recalled across trials in standard, repeated-
study, and repeated-testing conditions. The shorthand condition labels
indicate the order of study (S) and test (T) periods. Data are from Kar-
picke and Roediger (2006b).
186 Volume 1—Number 3
The Power of Testing Memory
delayed test 1 week later to examine lasting effects of the three
study schedules on long-term retention.
Our basic results during the learning phase are shown in
Figure 3, which indicates recall from secondary memory across
tests in the three conditions (Karpicke & Roediger, 2006b). It is
clear that subjects in the repeated-test condition were at a dis-
advantage early in learning (on Trials 4 and 8), but quickly
caught up to the repeated-study condition, so that there was little
difference between these two conditions later in learning (Trials
12, 16, and 20). However, the standard group performed better
than the other two groups over the last four tests (and this dif-
ference was statistically significant). Thus, we replicated Tul-
ving’s (1967) basic result that learning curves for these three
conditions are remarkably similar, although we did find a dif-
ference favoring the standard condition. The advantage for the
standard condition probably arose because a study trial just after
a test trial serves as feedback for what students do not know (they
can recognize words they failed to recall and focus their study
efforts on these items), and the standard condition had more test
trials followed immediately by study trials than the other con-
ditions did. As Izawa (1970) observed, test trials potentiate new
learning on the next study trial. We discuss the role of feedback
later in this article.
As noted, we (Karpicke & Roediger, 2006b) also measured
performance after a 1-week delay. Subjects were given 10 min to
recall and at the end of every minute drew a line under the last
word recalled, which permitted us to measure how recall cu-
mulates across time (see Wixted & Rohrer, 1994). Figure 4
shows the result, and it is apparent that from the very first minute
of the final test period, subjects in the repeated-study condition
performed worse than those in the other two conditions. At the
end of the recall period, subjects in the standard and repeated-
test conditions recalled 68% and 64% of the 40 words, re-
spectively, whereas those in the repeated-study condition re-
called only 57% of the words (this was a significant difference
from the other two conditions, which did not themselves differ).
Thus, despite the fact that the subjects in the repeated-study
condition had studied the list 15 times 1 week earlier and those
in the repeated-test condition had studied it only 5 times, de-
layed recall was greater for the latter group. This outcome again
shows the power of testing in improving long-term retention.
Although the results just reported are striking, other, earlier
experiments also showed testing effects in free recall. For ex-
ample, Hogan and Kintsch (1971) reported two experiments
showing the advantage of test trials over study trials in promoting
long-term retention. In one experiment, they had some students
study a list of 40 words four times, with only short breaks be-
tween presentations of the lists. A second group studied the list
once and then took three consecutive free-recall tests (similar to
a single cycle in the repeated-test condition of Tulving’s, 1967,
experiment). Both groups returned 2 days later for a final test.
The pure-study group recalled 15% of the words, whereas the
group that received only one study trial but three tests recalled
20%. A single study trial and three tests produced significantly
better recall than did studying the material four times.
Repeated Testing and Selective Re-Presentation of
Forgotten Material
Thompson, Wenger, and Bartling (1978) replicated Hogan and
Kintsch’s (1971) results, again using 40-word lists, but with two
new twists that deserve special mention. In addition to condi-
tions with four study trials (repeated-study condition) and one
study trial and three tests (repeated-test condition), they in-
cluded a condition in which subjects studied the list once, re-
called it, studied only those words they failed to recall, recalled
the entire list again, and so on for three more study-test episodes
with the study lists becoming shorter and shorter. This test/re-
presentation condition mimicked a variation of what students
are often told to do in study guides: study the material, test
themselves, restudy items they missed, and so on until they
achieve perfect mastery (this guidance is similar to what Gates’s,
1917, subjects were instructed to do). However, note that the
subjects of Thompson et al. were instructed to recall the entire
list on each test trial, not just the items they restudied in the
previous study phase. Besides adding this condition to Hogan
and Kintsch’s (1971) design, Thompson et al. also included final
tests 5 min after the learning phase and 2 days later. (Retention
interval was manipulated between subjects, so the 5-min test
would not influence the 2-day test.)
Table 1 summarizes the results Thompson et al. (1978) ob-
tained. It is clear that on the 5-min test, the group that had only
one study trial but repeated tests had the poorest recall. The
group that only studied the lists did next best, but the group that
Fig. 4. Cumulative recall on a final retention test given 1 week after initial
learning. Results are shown separately for standard, repeated-study, and
repeated-testing conditions. The shorthand condition labels indicate the
order of study (S) and test (T) periods. Data are from Karpicke and
Roediger (2006b).
Volume 1—Number 3 187
Henry L. Roediger, III, and Jeffrey D. Karpicke
was tested with re-presentation of the missed items performed
best of all. However, 2 days later, the situation changed. Al-
though the test/re-presentation group still did best, the repeat-
ed-test group slightly outperformed the repeated-study group.
Looking at these results another way, subjects in the repeated-
study condition showed dramatic forgetting over 2 days (meas-
ured either as the difference between 5-min and 2-day recall or
as a percentage of 5-min recall; see Loftus, 1985). Although
subjects in the repeated-study condition forgot 56% of what they
originally could recall, those in the test/re-presentation condi-
tion forgot 26%, and subjects in the repeated-test condition
showed the least forgetting, just 13%. This outcome shows that
the advice in study guides appears to be accurate: Students
should study, test themselves, and then restudy what they did not
know on the test. However, in a later experiment, we (Karpicke
& Roediger, 2006b, Experiment 2) showed that the fact that
Thompson et al. required recall of the entire list during each test
was critical to this outcome. If students in the test/re-presen-
tation condition are required to recall only the items that were
presented in the preceding re-presentation study phase, they
display rather poor recall on a delayed test. Repeated testing of
the whole set of material is critical to improve long-term re-
In sum, the results of Thompson et al. also show the power of
testing for enhancing long-term retention: Both tested groups
recalled more on the delayed final test than the group that only
studied the word lists, without initial testing. On the delayed test
in this experiment, the advantage of repeated testing over re-
peated studying was rather small (Thompson et al., 1978),
probably because of the relatively brief amount of time given to
subjects to recall on the initial tests. Nevertheless, the effect has
been replicated by Wheeler, Ewers, and Buonanno (2003). In
their second experiment, subjects studied a 40-word list either
five times (repeated-study condition) or one time with four
consecutive recall tests (repeated-test condition). Final free-
recall tests were given to different groups of subjects either 5
min or 1 week later. The results are shown in Figure 5, which
reveals a huge advantage for massed study on the immediate
test, but a significant reversal on the test given a week later. This
result and others like it are even more surprising when one
considers that in the repeated-study condition, subjects are
presented with all 40 words in the list on each trial, whereas in
the repeated-test condition, they are reexposed only to those
words that they can recall (only about 11 of the 40 words in this
experiment). Thus, the overwhelmingly greater number of ex-
posures in the repeated-study condition improved performance
only on a relatively immediate test. After a 1-week delay, sub-
jects in the repeated-test condition outperformed those in the
repeated-study condition despite having studied the material
only once. Once again, the power of testing is clear. In a later
section, we review evidence that the same pattern holds for re-
call of text materials like those used in educational settings
(Roediger & Karpicke, 2006).
The experiments we have just discussed compared conditions
with several recall tests and conditions in which students re-
peatedly studied the material. Wheeler and Roediger (1992)
investigated whether multiple tests are more beneficial than a
single test, and also gave subjects fairly lengthy initial recall
tests (unlike most of the experiments reviewed thus far). In some
Proportion Correct in Immediate and Delayed Recall in Thompson, Wenger, and Bartling’s
(1978) Experiment 2
Test Difference
(5 min 48 hr)
forgetting5 min 48 hr
Repeated study (SSSS) .50 .22 .28 56
Repeated test (STTT) .28 .25 .03 13
Repeated test and re-presentation (ST
) .60 .44 .16 26
Note. Percentage forgetting was calculated as follows: [(recall at 5 min recall at 48 hr)/recall at 5 min] 100. S 5
study period; T 5test; T
5test with re-presentation of forgotten items.
Fig. 5. Proportion of words recalled on immediate (5-min) and delayed
(7-day) retention tests after repeated studying or repeated testing. Data
are estimated from Wheeler, Ewers, and Buonanno (2003).
188 Volume 1—Number 3
The Power of Testing Memory
conditions, subjects heard a story that named 60 particular
concrete objects. A picture of each object was shown on a screen
the first time the object was named in the story, and subjects
were told that they would be tested on the names of the pictures.
After presentation, control subjects were dismissed from the lab
and asked to return a week later. Another group of subjects took
one 7-min recall test and left, and a third group received three
recall tests before being permitted to leave. All subjects re-
turned a week later for a final recall test. The results are shown in
Table 2. On the initial test, subjects in the single-test condition
recalled 53% of the items; control (no-test) subjects would
presumably have recalled about the same number of items had
they been tested, so this estimate was used to measure forgetting
in that condition. Subjects in the three-test condition recalled
61% of the items on their third test; their recall was higher than
that of subjects in the one-test condition because recall often
increases upon such repeated testing, a phenomenon called
hypermnesia (Erdelyi & Becker, 1974; Roediger & Thorpe,
1978). Final recall after a week was 29% in the no-test condi-
tion, 39% in the one-test condition, and 53% in the three-test
condition. Clearly, forgetting (as either a difference or a pro-
portion) was inversely related to the number of immediate tests,
with subjects exhibiting 13% forgetting after three tests, 27%
forgetting after one test, and 46% forgetting after no tests. In a
sense, subjects who received three tests were completely im-
munized against forgetting, because they recalled the same
number of pictures after a week that subjects in the single-test
condition recalled a week earlier (53%). The two extra tests in
the repeated-testing condition maintained performance at a high
level 1 week later.
The experiments we have reviewed in this section all involved
free-recall tests or slight variations of free-recall tests. Tulving
(1967), among other researchers, showed that within very broad
limits, a free-recall test permits as much learning as restudying
material. However, later research showed a more complicated
picture: Repeatedly studying material is beneficial for tests
given soon after learning, but on delayed criterial tests with
retention intervals measured in days or weeks, prior testing can
produce greater performance than prior studying. In the case of
delayed recall, test trials produce a much greater gain than study
trials. Of course, there must be at least one study opportunity for
testing to enhance later recall, but many of the experiments we
have discussed used only one study trial followed by several
tests and yet demonstrated an advantage in delayed recall for
this condition over one in which there were multiple study trials
(e.g., five study trials and no tests in Wheeler et al., 2003).
Testing reduces forgetting of recently studied material, and
multiple tests have a greater effect in slowing forgetting than
does a single test (Wheeler & Roediger, 1992). We consider
theoretical accounts of these data in a later section, but first we
review selected experiments from a different tradition of testing
When a person learns names to go with faces, or that caballo
means ‘‘horse’’ in Spanish, or that 8 9572, or that a friend’s
telephone number is 792-3948, the task is essentially one of
paired-associate learning. Of course, in the laboratory, paired-
associate learning is often studied using word pairs that may
vary in association value (chair-table or chair-donkey) or non-
word-word pairings (ZEP-house), among many other variations.
This task, first used in experiments by Calkins (1894), has been
a favorite for studying testing effects. In addition to mimicking
many learning situations with which people are faced in daily
life, the task is especially tractable in the laboratory. When used
to investigate the testing effect, the task makes it possible to
manipulate the interval between study and test of a specific pair,
and presentation or withholding of feedback can also be easily
accomplished. In this section, we briefly review literature show-
ing testing effects in paired-associate learning and then turn to
the issue of spaced testing in continuous paired-associate tasks.
Testing Effects in Cued Recall and Paired-Associate Tests
Estes (1960) began research on testing effects in paired-asso-
ciate learning, and this work has been carried forward by other
researchers. For example, Allen, Mahler, and Estes (1969) had
subjects study a list of paired associates either 5 or 10 times and
then take no, one, or five tests on the items. One day later, the
subjects were given a final retention test in which they were cued
with the stimulus (the left-hand member) of the pair and asked to
recall the response. Allen et al. found a modest benefit of
studying the list 10 times relative to studying it 5 times, but the
effects of initial testing were much larger, with final test per-
formance in both study conditions increasing directly as a
function of the number of initial tests (see Table 3). Final test
performance of subjects who studied the list 5 times and were
tested once was equivalent to that of subjects who studied the list
Proportion of Pictures Recalled Immediately After Study and 1
Week Later in Wheeler and Roediger (1992)
(immediate delayed)
1 week
No test (.53)
.29 .24 46
One test .53 .39 .14 27
Three tests .61
.53 .08 13
Note. Percentage forgetting was calculated as follows: [(immediate recall
recall at 1 week)/immediate recall] 100.
Because subjects in this condition did not take an immediate test, the per-
formance of subjects in the one-test condition was used to estimate their likely
performance so that their forgetting could be measured.
This proportion is
taken from the third test.
Volume 1—Number 3 189
Henry L. Roediger, III, and Jeffrey D. Karpicke
10 times and received no initial test. This outcome led Allen et
al. to conclude that taking a single test was as effective for long-
term retention as 5 additional study trials. Izawa, in particular,
has continued this line of research and produced a large body of
work (e.g., Izawa, 1966, 1967, 1970; see Izawa, Maxwell, Hay-
den, Matrana, & Izawa-Hayden, 2005, for a recent summary of
this program of research). Izawa has referred to test trials as
potentiating future learning and presented a mathematical
model of how this process might operate, although this model is
specific to repeated study-test trials (Izawa, 1971).
In a rather different tradition, Jacoby (1978) had subjects
study word pairs (e.g., foot-shoe) and then either restudy the pair
(foot-shoe) or take a simple test in which they had to generate
the right-hand member of the pair when given the left-hand
member and a fragmented form of the right-hand member ( foot-
s__e). Further, the second occurrence of the pair (either re-
studied or tested) was either immediately after the pair had in-
itially been studied or after a delay filled with 20 intervening
pairs. Many different pairs were presented in these four condi-
tions (restudy or test after either a short or a long delay). At the
end of the experiment, subjects received a final test in which
they were given only the left-hand cue word and were asked to
recall the right-hand target ( foot-????). The results on this final
test showed that prior testing with the fragment ( foot-s__e) led
to better retention than restudying the intact word pair ( foot-
shoe), once again demonstrating that testing can be better than
restudying materia l even when the ‘‘test’’ seems quite simple. In
addition, recall on the final test was much better when the initial
test had been delayed by 20 intervening items than when it
occurred immediately after study of the pair. Jacoby argued that
when the test occurred immediately after the study phase, the
effortful processing that usually occurs during memory retrieval
was short-circuited, and the test lost its potency. We return to
this issue later.
Jacoby’s (1978) experiment is often cited as a pioneering
study of the generation effect (the fact that generating material
often leads to better recall or recognition than reading the same
material; see also Slamecka & Graf, 1978), a phenomenon re-
lated to the testing effect. The fragment cues led to high levels of
recall (above 90%) on the initial tests in Jacoby’s experiment,
but other researchers using standard cued-recall tests that do
not produce such high initial recall levels have also demon-
strated positive effects of testing on later retention of paired-
associate material (Carrier & Pashler, 1992; Kuo & Hirshman,
1996; McDaniel & Masson, 1985). Tests during paired-associate
learning greatly reduce forgetting (Runquist, 1986), and the
effects are increased when feedback is given for items that are
missed on the tests (see Cull, 2000; Pashler, Cepeda, Wixted, &
Rohrer, 2005). Thus, the testing effects observed in free recall
also hold in paired-associate learning.
Spaced Retrieval Practice With Paired Associates
We now focus on a practical question raised by Landauer and
Bjork (1978). Given that testing generally improves retention
relative to restudying, they asked if the schedule of testing
matters. If a subject learns an A-B pair (where A might be horse
and B caballo), what is the best sequence of testing to promote
long-term retention? Perhaps testing should occur soon after
learning and be repeated in a massed fashion, because multiple
tests promote better retention than a single test. Massed testing
immediately after study would also permit errorless retrieval on
the repeated tests. But perhaps spacing tests over intervals of
time is a better schedule, because spaced practice is known to
benefit retention in the long term (e.g., Glenberg, 1976; Melton,
1970; for a review, see Cepeda, Pashler, Vul, Wixted, & Rohrer,
2006). However, if tests are spaced at equal intervals, then
delaying an initial test after studying a pair (in a spaced
schedule) may lead to forgetting. Thus, Landauer and Bjork
made the case for an expanding schedule of testing. In this
scheme, a first test occurs immediately after an A-B pair is
presented, to ensure that subjects can recall B when given A.
Then, a longer span of time (with more studied and tested items
presented) occurs before A is presented again for a test, and a yet
longer time occurs before a third test, and so on. The idea behind
expanding retrieval schedules is to gradually shape production
of the desired response so that it can be retrieved out of context,
at a long delay (the analogy is to shaping of responses in operant
Of course, if an expanding schedule of repeated retrieval
shows an advantage over massed testing, this advantage might
accrue simply because the expanding schedule, unlike the
massed schedule, involves spaced presentations (Rea & Mo-
digliani, 1985). For this reason, Landauer and Bjork (1978)
tested expanding and equal-interval schedules matched on the
average spacing between tests. For example, if the expanding
schedule was 1-5-9 (the numbers refer to the number of trials
intervening between successive tests of A-B after its study), then
the appropriate equal-interval schedule was 5-5-5, which on
average produced the same amount of spacing, but distributed
equally. Expanding retrieval practice is thought to be an optimal
schedule for long-term retention because success is high on an
immediate test and then the spacing implemented on the ex-
panding tests gradually increases the difficulty of retrieval at-
tempts, encouraging better later retention.
Landauer and Bjork (1978) reported two experiments that
compared these schedules in paired-associate learning (first
Proportion of Final Cued Recall on a 24-Hr Retention Test as a
Function of Different Levels of Initial Study and Number of Tests
on Day 1 (from Allen, Mahler, & Estes, 1969)
Number of initial tests
None One Five
5 study trials .58 .66 .82
10 study trials .65 .81 .88
190 Volume 1—Number 3
The Power of Testing Memory
name–surname pairs in one experiment and name-face pairs in
the other). No feedback or correction was given to subjects if
they made errors or omitted answers. Landauer and Bjork found
that the expanding-interval schedule produced better recall
than equal-interval testing on a final test at the end of the ses-
sion, and equal-interval testing, in turn, produced better recall
than did initial massed testing. Thus, despite the fact that
massed testing produced nearly errorless performance during
the acquisition phase, the other two schedules produced better
retention on the final test given at the end of the session. How-
ever, the difference favoring the expanding retrieval schedule
over the equal-interval schedule was fairly small at around 10%.
In research following up Landauer and Bjork’s (1978) original
experiments, practically all studies have found that spaced
schedules of retrieval (whether equal-interval or expanding
schedules) produce better retention on a final test given later
than do massed retrieval tests given immediately after presen-
tation (e.g., Cull, 2000; Cull, Shaughnessy, & Zechmeister,
1996), although exceptions do exist. For example, in Experi-
ments 3 and 4 of Cull et al. (1996), massed testing produced
performance as good as equal-interval testing on a 5-5-5
schedule, but most other experiments have found that any
spaced schedule of testing (either equal-interval or expanding)
is better than a massed schedule for performance on a delayed
test. However, whether expanding schedules are better than
equal-interval schedules for long-term retention—the other part
of Landauer and Bjork’s interesting findings—remains an open
question. Balota, Duchek, and Logan (in press) have provided a
thorough consideration of the relevant evidence and have shown
that it is mixed at best, and that most researchers have found no
difference between the two schedules of testing. That is, per-
formance on a final test at the end of a session often shows no
difference in performance between equal-interval and expand-
ing retrieval schedules.
For example, Balota, Duchek, Sergent-Marshall, and Roe-
diger (2006) compared expanding-interval retrieval tests with
equally spaced tests and massed tests in three groups of sub-
jects: young adults, healthy older adults, and older adults with
Alzheimer’s disease. They presented items twice (to ensure that
patients encoded them) and then employed massed testing for
some items (0-0-0), equal-interval testing for others (3-3-3), and
expanding-interval testing for still others (1-3-5). A final test
occurred at the end of the session. During acquisition, all three
groups showed the highest level of performance on the massed
tests, the next best performance on the expanding-interval tests,
and the worst performance on the equal-interval tests. This last
outcome was due to the relatively long lag before the first test for
the equal-interval condition. However, despite these differences
during acquisition, on the final test at the end of the session,
there was no difference between the equal-interval and ex-
panding-interval conditions for any of the three groups (although
recall in both these conditions was superior to that in the mas-
sed-test condition). Carpenter and DeLosh (2005) showed sim-
ilar effects in learning of name-face pairs, except that on their
final test they found a slight benefit for an equal-interval con-
dition over an expanding-interval condition.
Thus far, we have reviewed studies comparing expanding- and
equal-interval retrieval over a relatively narrow range of possi-
ble spacing schedules. Logan and Balota (in press) used a va-
riety of expanding schedules and compared them with
appropriate equal-interval schedules in younger and older
adults. In younger adults, they found that recall at the end of the
session was no better for expanding- than for equal-interval
testing, but they did find an advantage for expanding-interval
retrieval among older adults. However, Logan and Balota also
gave subjects a 24-hr delayed test and discovered that initial
equal-interval testing produced better recall on this test than did
the expanding-interval testing schedule. This outcome occurred
despite the fact that expanding-interval retrieval produced
better recall during initial acquisition and (for older subjects) on
the test at the end of the first day.
We recently obtained a similar result (Karpicke & Roediger,
2006a), using pairs consisting of vocabulary words and their
meanings (e.g., sobriquet-nickname). We tested subjects in
massed (0-0-0), equal-interval (5-5-5), and expanding-interval
(1-5-9) conditions during acquisition, and then subjects were
given a final test either 10 min or 2 days after the learning
session. At both retention intervals, the spaced-practice con-
ditions produced better recall than massed practice. On the 10-
min test, we replicated Landauer and Bjork’s (1978) results by
showing that expanding-interval retrieval produced a modest
benefit relative to equal-interval retrieval. However, after 48 hr,
we found the opposite pattern of results: Items in the equal-in-
terval condition were recalled better than items studied under
an expanding-interval schedule. We replicated this pattern of
results in a second experiment in which subjects were given
feedback after each test trial during the learning phase.
Our results (Karpicke & Roediger, 2006a) and those of Logan
and Balota (in press) indicate that in some circumstances,
equal-interval retrieval practice may promote greater long-term
retention than expanding-interval retrieval practice. We have
argued that the factor responsible for the advantage of equal-
interval practice is the placement of the first retrieval attempt:
The longer interval before the first test demands more retrieval
effort and leads to better retention (this argument is similar to
what Jacoby, 1978, concluded). Other research with paired as-
sociates has shown that increasing the delay before an initial test
promotes later retention, even though success on the initial test
often decreases with increasing delays (e.g., Jacoby, 1978;
Modigliani, 1976; Pashler, Zarow, & Triplett, 2003; Whitten &
Bjork, 1977). In the equal-spacing conditions used by Logan
and Balota (in press) and by us (Karpicke & Roediger, 2006a), as
well as by other researchers, the first retrieval attempt occurred
after a brief delay. However, the hallmark of expanding-interval
retrieval practice is an initial retrieval attempt immediately
after studying, to ensure high levels of recall success. Indeed,
Volume 1—Number 3 191
Henry L. Roediger, III, and Jeffrey D. Karpicke
performance on this massed initial test is often nearly perfect,
most likely because the test involves retrieval from primary or
short-term memory. However, retrieval from primary memory
usually does not produce benefits for later retention (see also
Craik, 1970; Madigan & McCabe, 1971). Thus, equally spaced
practice may lead to benefits for long-term retention because of
the delayed initial test, and current research is aimed at clari-
fying why certain spacing conditions are more or less effective
for learning (see Balota et al., in press).
Many of the testing effects found with free-recall tests hold true
in paired-associate learning. Tests promote better retention than
do additional study trials with paired associates, and repeated
tests provide even greater benefits. In addition, paired associ-
ates have been used to investigate whether a particular type of
testing schedule is optimal for long-term retention. Most of the
research has indicated that spaced retrieval practice leads to
better retention than massed practice, but the evidence is mixed
regarding whether expanding-interval retrieval is a superior
form of spaced retrieval. The most recent evidence points to the
conclusion that expanding-interval retrieval may not benefit
long-term retention, as was originally thought, because the in-
itial test in an expanding schedule appears too soon after study,
rendering it ineffective for enhancing learning. Although the
efficacy of expanding and equally spaced schedules remains an
open issue, the research we have reviewed shows that delaying
an initial retrieval attempt and spacing repeated tests often will
boost later retention with paired-associate materials.
Many of the testing effects we have discussed so far have been
observed in psychology laboratories, and the effects have been
obtained with materials commonly used in the lab, such as lists
of words or unrelated word pairs. Some exceptions do exist.
Positive effects of testing have been found in experiments using
foreign-language vocabulary words (e.g., Carrier & Pashler,
1992), materials taken from test-preparation books for the
Graduate Record Examination (Karpicke & Roediger, 2006a;
Pashler et al., 2003), and general knowledge questions
(McDaniel & Fisher, 1991). The two classic studies by Gates
(1917) and Spitzer (1939) also used educational materials, but
these examples aside, the majority of the research on testing
effects has used materials that are not found in educational
settings. Moreover, the limited range of materials most likely is
part of the reason why the testing effect is not widely known in
education and has not been incorporated into educational
practice. One can therefore wonder, does the testing effect
generalize to educationally relevant materials and test formats?
The answer to this question is ‘‘yes,’’ and in this section, we
review research using prose materials and then focus on the
effects of different types of tests often used in schools (e.g.,
short-answer questions and multiple-choice tests).
Testing Effects With Prose Materials
One area of research related to the testing effect has shown that
answering questions while reading textbook material often fa-
cilitates comprehension and retention of the material. The be-
ginning of research on such adjunct questions is attributed to
pioneering studies by Rothkopf (1966), who referred to brief
questions placed at different points throughout an instructional
text as ‘‘test-like events.’’ The effects of adjunct questions on
learning were investigated intensively until the 1980s (see
Hamaker, 1986), but have received little attention since. Re-
search on adjunct questions showed that they often facilitate
retention and comprehension of text material and also pointed to
two other important conclusions. First, questions that follow a
text promote better retention than questions that appear in ad-
vance of the text or interspersed throughout the text. Second,
answering questions that accompany a text will often en-
hance later performance on related questions (see also Chan,
McDermott, & Roediger, in press, which is discussed later). We
mention the research on adjunct questions only briefly because
that literature has been extensively reviewed elsewhere (see
R.C. Anderson & Biddle, 1975; Crooks, 1988; Hamaker, 1986;
Rickards, 1979). Although the results indicate that these test-
like events do facilitate learning of prose materials, it is not clear
how often students actually answer questions that accompany
texts or how closely adjunct questions approximate the condi-
tions of actual classroom tests.
Recently, we (Roediger & Karpicke, 2006) have investigated
the testing effect taking an approach aimed at integrating the
research tradition from cognitive psychology, which we have just
reviewed (e.g., Hogan & Kintsch, 1971; Thompson et al., 1978;
Wheeler & Roediger, 1992), with educational research that has
focused on learning of more complex prose materials. In our
experiments, we had college students study prose passages
covering general scientific topics. Depending on the condition
to which a passage was assigned, the students then either re-
studied the entire passage or took a free recall test in which they
were asked to write down as much as they could remember from
the passage (this test was similar to an essay test in school
contexts). The students were not given any feedback about their
test performance (i.e., they did not restudy the material after the
test), but were given ample time (7 min) to study the passage in
the restudy condition and to take the recall test in the test
condition (as mentioned earlier, the brief amount of time given to
subjects in previous experiments probably attenuated the pos-
itive effects of testing). Finally, 5 min, 2 days, or 1 week after the
learning session, different groups of students took a final free-
recall test that was just like the recall test given initially. The
results of the experiment are shown in Figure 6. After 5 min,
restudying produced a modest benefit over testing (81% vs. 75%
of the passage recalled), but the opposite pattern of results was
192 Volume 1—Number 3
The Power of Testing Memory
observed on the delayed retention tests. After 2 days, initial
testing produced better retention than restudying (68% vs.
54%), and an advantage of testing over restudying was also
observed after 1 week (56% vs. 42%). The results conceptually
replicate earlier experiments using free recall and paired-as-
sociate learning of lists and generalize them to educational
We conducted a second experiment to investigate the effects
of repeated studying and repeated testing on later retention
(Roediger & Karpicke, 2006). Subjects studied passages during
four separate periods (SSSS), studied during three periods and
took one recall test (SSST), or studied during one period and took
three tests (STTT). They took a final recall test either 5 min or 1
week after this learning session. The results, which are shown in
Figure 7, reveal that after 5 min, recall was correlated with re-
peated studying: The SSSS group recalled more than the SSST
group, who in turn recalled more than the STTT group. However,
on the 1-week retention test, recall was correlated with the
number of initial tests: The STTT group recalled more than the
SSST group, who in turn recalled more than the SSSS group. In
terms of proportional measures of forgetting (which take into
account differences in the level of original learning), the SSSS
group showed the most forgetting (52%), followed by the SSST
group (28%), and the repeated-testing group (STTT) showed the
least amount of forgetting (10%) over 1 week.
Our results (Roediger & Karpicke, 2006) demonstrate the
powerful effect testing has in enhancing later retention, and
confirm and extend with prose materials the earlier findings with
word-list materials. In addition, we investigated the subjects’
experience after repeated studying or repeated testing by asking
them to predict how well they thought they would remember the
passage in the future. These predictions were inflated after re-
peated study, relative to the testing conditions, even though
repeated studying produced the worst long-term retention (see
Dunlosky & Nelson, 1992, for a similar result). This finding
suggests that students may prefer repeated studying because it
produces rapid short-term gains, even though it is an ineffective
strategy for long-term retention.
Testing effects have also been found using educationally
relevant test formats, such as short-answer and multiple-choice
tests. In another experiment (Agarwal, Karpicke, Kang, Roe-
diger, & McDermott, 2006), we had students study textbook
passages and then complete short-answer tests on some of the
passages. An initial short-answer test enhanced retention on a
final short-answer test given 1 week later, relative to studying the
passage without taking the test. We also investigated the effects
of giving students feedback about their test performance. Pro-
viding feedback (by having students restudy the passage) en-
hanced retention to a greater extent than testing alone, but the
effectiveness of feedback depended on when it occurred. In one
condition, students were shown the passage while they took the
test. This condition was similar to open-book testing commonly
used in education and also similar to taking notes while reading.
Subjects in this condition had access to feedback continuously
during the test. In another condition, students took the test and
then were given the passage and instructed to look over their
responses (a delayed-feedback condition). Although the im-
mediate-feedback condition produced the best performance on
the initial test (not surprisingly), the delayed-feedback condi-
tion promoted better long-term retention. The results of this
study are analogous to those obtained with motor learning tasks
(see Schmidt & Bjork, 1992) and suggest that students should
Fig. 6. Mean proportion of idea units recalled from a prose passage after
a 5-min, 2-day, or 1-week retention interval as a function of whether
subjects studied the passages twice or studied them once before taking an
initial test. Error bars represent standard errors of the means. From
Roediger and Karpicke (2006).
Fig. 7. Mean proportion of idea units recalled on a final test 5 min or 1
week after learning as a function of learning condition. The shorthand
condition labels indicate the order of study (S) and test (T) periods. Error
bars represent standard errors of the means. From Roediger and Karpicke
Volume 1—Number 3 193
Henry L. Roediger, III, and Jeffrey D. Karpicke
delay feedback or reviewing their answers until after completing
a test in order to optimize later retention.
Nungester and Duchastel (1982) investigated the effects of
multiple-choice and short-answer tests on later retention of a
prose passage. In their experiment, one group of subjects
studied the passage and then took an initial test in which half of
the questions were short-answer questions and half were five-
alternative multiple-choice questions. Another group of sub-
jects studied the passage and then reviewed portions of it, and a
third group studied the passage only once. All the students re-
turned 2 weeks later for a final retention test, in which each
question was in the alternate format relative to the initial test
(i.e., items that were initially tested in short-answer format were
tested in multiple-choice format on the final test, and likewise
initial multiple-choice questions were tested as short-answer
questions on the final test). Nungester and Duchastel found that
reviewing the passage enhanced retention relative to just study-
ing it once, but taking the initial test led to the best retention.
This testing effect was found for both the multiple-choice and
the short answer-test formats (see also LaPorte & Voss, 1975). In
addition, in a follow-up to this original experiment, Nungester
and Duchastel had the same subjects take another multiple-
choice retention test 5 months after the initial learning session
(see Duchastel & Nungester, 1981). The pattern of results was
identical on this 5-month test, with the initially tested group
performing better than the study-once and study-twice groups.
Nungester and Duchastel’s work provides a compelling dem-
onstration that the testing effect persists over very long retention
intervals (see also Butler & Roediger, in press,and Spitzer, 1939).
Transfer of Testing Effects Across Different Test Formats
The research just described shows that both short-answer and
multiple-choice tests produce positive testing effects on later
retention. Other research on testing effects with prose materials
has investigated whether certain types of tests (e.g., essay, short-
answer, or multiple-choice) are more effective than others for
enhancing retention, or whether a particular test format facili-
tates later performance only for that test format. These issues
have also been addressed in laboratory research on the effects of
recall tests on performance on later recognition tests (e.g.,
Darley & Murdock, 1971; Lockhart, 1975; Wenger, Thompson,
& Bartling, 1980) and the effects of recognition tests on later
recall (e.g., Mandler & Rabinowitz, 1981; Runquist, 1983; see
also Carpenter & DeLosh, 2006; Hogan & Kintsch, 1971). In
this section, we review studies that have used educational ma-
terials to investigate the effects of different test formats.
To address the issue of whether testing effects are greater with
certain types of tests than with others, we again return to the
work of Duchastel and Nungester. Although these researchers
carried out several investigations of the testing effect in the early
1980s, their work is rarely cited in discussions of the testing
effect. In one study, Duchastel (1981) gave some students an
initial short-answer or multiple-choice test on a prose passage
and then a final short-answer test 2 weeks later. Both types of
initial tests produced better long-term retention than studying
alone, but taking the initial short-answer test promoted superior
retention 2 weeks later on the final short-answer test. Thus, this
work provides evidence that perhaps short-answer tests yield
greater testing effects than multiple-choice tests (but see Du-
chastel & Nungester, 1982, for a somewhat different conclusion).
Glover (1989) had students study a prose passage similar to
the one used by Duchastel and Nungester (1982). Two days after
studying the passage, the students took a free-recall test, a cued-
recall (fill-in-the-blank) test, or a recognition test that involved
identifying whether statements had or had not been in the
original passage. Two days later, the students took a final free-
recall, cued-recall, or recognition test. Glover found that taking
the initial free-recall test produced the best final retention, re-
gardless of the format of the final test, and the cued-recall test
produced better retention than the recognition test on both the
final cued-recall test and the final recognition test. Glover’s
study indicates that recall tests promote greater retention than
recognition tests, which is also a conclusion generally reached
by researchers studying testing effects in word-list paradigms.
However, one oddity in Glover’s study was that scores on the
free-recall test were consistently higher than scores on the cued-
recall test, which indicates that subjects could recall more in
free recall than they did on Glover’s cued-recall test, a result
directly in contrast to the results of fundamental research on
human memory (e.g., Tulving & Pearlstone, 1966). This strange
aspect of Glover’s data is most likely an artifact of the type of
questions asked on the cued-recall test, which somehow led to
subjects being able to recall more in free recall than they could
express on the cued-recall test. Thus, Glover’s results should be
interpreted with some caution.
Recently, Kang, McDermott, and Roediger (in press) reex-
amined the testing effect with short-answer and multiple-choice
tests in a study with better control of test content, to try to ensure
that the same information was being tested by the two formats.
They also examined transfer across test format and examined the
role of feedback on a first test in enhancing the testing effect.
The students studied articles from Current Directions in Psy-
chological Science, and after each article, they took a short-an-
swer or a multiple-choice test. We consider Experiment 2, in
which subjects received feedback after the tests, a procedure
that equates exposure to information for multiple-choice and
short-answer tests. In addition, in a control condition, the stu-
dents read statements from the articles after reading them; these
statements were the same as the items that were tested in the
other two conditions, again to equate exposure to the informa-
tion. Three days later, the students took a final test in either
a short-answer or a multiple-choice format. The initial short-
answer test produced the best retention for both final-test for-
mats (results consistent with those of Glover, 1989). Butler and
Roediger (in press) and McDaniel, Anderson, Derbish, and
Morrisette (in press) have reported similar outcomes.
194 Volume 1—Number 3
The Power of Testing Memory
Clearly, the work using educationally relevant materials has not
resolved all the questions concerning the effect of test format.
However, some conclusions are warranted. In virtually all the
experiments, taking an initial test led to better later retention
than not taking a test or than engaging in a period of additional
study. The testing effect is secure. Most evidence points to the
conclusion that tests involving production of information (essay
and short-answer tests) produce greater benefits on later tests
than do multiple-choice tests, which involve recognition of a
correct answer among alternatives. The literature is not totally
consistent on this point, however, so it remains a hypothesis for
further investigation. One problem is that performance is usu-
ally much higher on initial multiple-choice tests than on initial
short-answer tests; unless feedback is given to equate exposure
to answers, multiple-choice tests may have an advantage over
short-answer tests simply for this reason. Kang et al. (in press)
found that a short-answer test (with feedback) produced a
greater testing effect than did a multiple-choice test (also with
feedback), regardless of the format of the final test. A greater
testing effect for production tests than for recognition tests
would be similar to the generation effect during study of mate-
rial. That is, generating or producing material during study
usually creates greater retention than reading the material (Ja-
coby, 1978; Slamecka & Graf, 1978).
The experiments we have described show that the testing effect
generalizes to educationally relevant materials (e.g., prose
passages) and to test formats like those used in education (e.g.,
short-answer and multiple-choice tests). Nonetheless, most of
the studies described so far have been carried out in the labo-
ratory, and one can still ask whether the testing effect general-
izes to actual classroom situations. Several differences between
the laboratory and the classroom may lead to different results in
these two contexts. For example, the amount of information that
students are responsible for learning is much greater in the
classroom than in the laboratory (even when the laboratory
materials include prose passages taken from educational text-
books). Also, the to-be-learned materials in the classroom are
presented in a variety of ways—in textbooks, in lectures, in class
discussions, and so on. Students also differ greatly in the amount
of studying they do before exams, in how soon they begin
studying (relative to when exams occur), in their interest in the
course material, and in their motivation to learn. All these fac-
tors are typically controlled in well-designed experiments, but
they are free to vary in the classroom. In this section, we review
evidence from classroom studies of the testing effect. This evi-
dence shows that despite the differences between psychology
laboratories and school classrooms, the testing effect is a robust
phenomenon in educational settings, and frequent testing in the
classroom improves students’ learning.
Although classroom studies of frequent testing date back to
the 1920s (Deputy, 1929; Maloney & Ruch, 1929), relatively few
systematic studies have been carried out since that time. Ban-
gert-Drowns, Kulik, and Kulik (1991) conducted a meta-anal-
ysis of 35 classroom studies (22 published, 13 unpublished),
carried out from 1929 through 1989, that manipulated the
number of tests given to students during a semester. All of the
studies compared a frequently tested group of students against a
control group of students who received fewer tests. Bangert-
Drowns et al. obtained the studies from the Educational Re-
sources Information Center (ERIC) and Dissertation Abstracts
databases, and only studies in which the frequent-testing and
control groups received identical instructions were included in
the meta-analysis. Twenty-eight of the studies were carried out
in college classrooms, and 7 were carried out in high school
classrooms. Most of the classes covered math and science, but
some covered other topics (e.g., reading, government, law), and
the tests were conventional classroom tests, such as multiple-
choice and short-answer tests (though Bangert-Drowns et al. did
not analyze different test formats separately). The criterial
measure for all studies was performance on a final examination
given at the end of the class.
The majority of the studies Bangert-Drowns et al. (1991) in-
cluded (29 of 35, 83%) found positive effects of frequent testing,
and the mean effect size (standardized mean difference, d) was
.23. Five of the studies found negative effects, and 1 study found
no difference between frequent testing and the control condi-
tion. There was great variation in the number of tests given
during the semester, with the number of tests in the control group
ranging from 0 to 15, and number of tests in the frequent-testing
group ranging from 3 to 75. To investigate the effects of in-
creasing the number of tests during a semester-long class,
Bangert-Drowns et al. fit the data from the frequent-testing and
control conditions to a regression equation predicting the size of
the effect (indicating gains in learning due to testing) from fre-
quency of testing. The function they obtained, showing the re-
lation between the number of tests given during the semester-
long class and the expected effect size, is displayed in Figure 8,
which shows that performance on the final test increased as a
negatively accelerated function of the number of tests given in
class. Most notably, giving just 1 test produced a big gain rel-
ative to giving no tests at all, and subsequent repeated tests
added to these gains in learning. (Of course, unlike the exper-
imental studies described earlier, the repeated-testing studies in
this meta-analysis involved testing different sets of material, not
the same set of material repeatedly.) Bangert-Drowns et al. noted
that in 11 studies in which the control group received no tests,
the effect size comparing the frequent-testing and control condi-
tions was .54. However, when the control group received at least
1 test, the effect size dropped to .15. The implication is that in-
cluding a single test in a class produces a large improvement in
final-exam scores, and Figure 8 shows that gains in learning con-
tinue to increase as the frequency of classroom testing increases.
Volume 1—Number 3 195
Henry L. Roediger, III, and Jeffrey D. Karpicke
One other result from this meta-analysis (Bangert-Drowns
et al., 1991) is worth noting. Four of the studies reported stu-
dents’ attitudes toward the amount of testing in their classes, and
all four studies found that the students who were tested fre-
quently rated their classes more favorably (in course ratings at
the end of the semester) than the students who were tested less
frequently. We return to this point later.
The meta-analysis of Bangert-Drowns et al. (1991) is lacking
in some important respects. For instance, the authors did not
analyze possible differences between test formats, nor did they
include any information about what kind of feedback students
received on their tests. In addition, most (29) of the studies in-
cluded in the analysis did not randomly assign students to the
frequent-testing or control conditions. Nevertheless, the impli-
cations of this meta-analysis are important: The testing effect
works in the classroom, and students react favorably to frequent
testing in their courses.
Leeming (2002) recently reported that giving a brief test each
day in college courses on introductory psychology and on
learning and memory improved students’ final grades (relative to
the grades in other courses he taught without daily testing).
Leeming began each class period with a 10- to 15-min test that
included about seven short-answer questions. After each test, he
spent 2 to 3 min discussing the correct answers with the students
(i.e., giving immediate feedback) before starting the lecture.
Thus, a typical semester-long class that met 2 days a week could
involve 22 to 24 exams. Leeming reported that the final grades
in his courses with this exam-a-day procedure were better than
the final grades in previous versions of the same courses that he
had taught without daily testing (80% vs. 74% for the intro-
ductory psychology course and 89% vs. 80% for the learning and
memory course). In addition, near the end of the course, Leem-
ing had some introductory psychology students take a reten-
tion test covering material that had not been discussed in class
for at least 6 weeks (thus, the retention interval before the test
was approximately 6 weeks). Leeming compared students in the
frequent-testing course with students in other sections of the
introductory psychology course that did not involve daily testing
and found that students in the frequent-testing course performed
better on this test than did students in the other sections.
Leeming’s (2002) report provides yet another example of how
frequent testing in the classroom can enhance students’ learn-
ing. Also, students in the frequent-testing classes completed a
questionnaire about the procedure at the end of the course. The
responses indicated that, overall, students liked the frequent-
testing procedure. Although the majority of students agreed that
they were skeptical about the procedure at the beginning of the
course, they also indicated that they studied more frequently in
this class than in other classes with fewer tests and believed that
they learned more. The majority of students also said they liked
daily testing and would choose frequent testing over fewer exams.
One problem with these classroom studies, noted earlier, is
that they lack some of the controls included in laboratory ex-
periments, such as random assignment of students to tested
versus nontested conditions. Recently, McDaniel et al. (in press)
were able to overcome the problem of random assignment by
instead randomly assigning different items to the tested and
nontested conditions in a within-subjects design. They had
volunteer students enrolled in a brain and behavior course take
weekly 10-min quizzes during the semester. The quizzes were
administered and scored over the Internet and included short-
answer questions or multiple-choice questions. Some individual
statements or facts that were not included on the quizzes were
presented to the students for them to reread, and other items
were not reexposed to the students at all (no-exposure control
condition). After completing each quiz, the students were given
feedback about their performance on each question. The stu-
dents also took two unit tests during the semester and then a
cumulative final exam at the end of the course. Some items that
appeared on the quizzes were repeated on the later criterial
tests, and other items on the criterial tests had not been on a quiz
(the items in the no-exposure control condition). However, the
items that were repeated from the quizzes were worded differ-
ently when they appeared on the criterial tests.
McDaniel et al. (in press) observed similar patterns of results
on the unit tests and final exams. Being reexposed to the facts
(restudy, multiple-choice quiz, or short-answer quiz) produced a
modest benefit over not being reexposed to them (no-exposure
control condition), and both of the quiz conditions produced
better performance on the unit and final tests than the restudy
condition. Although taking multiple-choice quizzes produced
better performance than studying the statements, short-answer
quizzes produced even greater gains on the criterial tests. Thus,
Fig. 8. Expected effect size for classroom testing as a function of the
number of tests given during a semester-long course. From ‘‘Effects of
Frequent Classroom Testing,’’ by R.L. Bangert-Drowns, J.A. Kulik, and
C.L.C. Kulik, 1991, Journal of Educational Research,85, p. 96. Copy-
right 1999 by Heldref Publications. Reprinted with permission of the
Helen Dwight Reid Educational Foundation.
196 Volume 1—Number 3
The Power of Testing Memory
the results of this classroom experiment converge with the re-
sults of other experiments in demonstrating the effectiveness of
frequent testing for enhancing learning. Further, they confirm
that short-answer tests produce greater testing effects than
multiple-choice tests, supporting the results of the laboratory
studies of Butler and Roediger (in press), Glover (1989), and
Kang et al. (in press).
Classroom studies often lack the control over variables found in
laboratory studies. Nonetheless, the meta-analytic study by
Bangert-Drowns et al. (1991) reviewing the literature on fre-
quency of classroom testing, Leeming’s (2002) work in his own
courses, and the within-subjects, within-course experiment of
McDaniel et al. (in press) all point to the same conclusion, that
the testing effect does generalize to the classroom.
Prior reviews of the literature by Dempster (1996, 1997) iden-
tified two theories to account for the positive effects of testing on
learning. He referred to these theories as the amount-of-
processing hypothesis and the retrieval hypothesis (see also
Glover, 1989). In this section, we evaluate and expand upon
these two theories and provide additional explanations to ac-
count for the data we have reviewed. We first consider the idea
that the testing effect is merely a result of additional exposure to
material during the test (i.e., the amount-of-processing hy-
pothesis), or more specifically, that testing simply leads to
overlearning of a portion of the to-be-learned materials. As we
have noted throughout this review, the bulk of evidence about
the testing effect leads us to reject these ideas. Next, we discuss
several ideas emphasizing that tests enhance learning via re-
trieval processes that reactivate and operate on memory traces
either by elaborating mnemonic representations or by creating
multiple retrieval routes to them (Bjork, 1975; McDaniel &
Masson, 1985), and we discuss the related notion of creating
‘‘desirable difficulties’’ for learners, an idea championed by
Bjork (1994, 1999; see also Bjork & Bjork, 1992). Finally, we
consider the concept of transfer-appropriate processing (e.g.,
Blaxton, 1989; Morris, Bransford, & Franks, 1977; Roediger,
1990) and how it can be applied to the testing effect.
Additional Exposure and Overlearning
One idea that we sketched at the outset of our review is that a test
provides additional exposure to the tested material, and that this
extra exposure is responsible for the testing effect (an idea
suggested by Thompson et al., 1978). We believe that the evi-
dence is inconsistent with this simple explanation. The probable
reason this idea arose is that many experiments on the testing
effect have compared a condition in which students study ma-
terial and then take a delayed final test with a condition in which
subjects study, take an initial test, and then take the delayed
final test. The latter condition shows better performance on the
final criterial test—the testing effect—but this design con-
founds the effects of testing with the effects of total exposure
time. Other experiments we have reviewed have equated expo-
sure to the material in the two conditions (by re-presenting
material for study in the control condition) and have still ob-
tained robust testing effects. In fact, the usual restudy control
condition provides a greater (rather than equal) exposure to the
material, because in the testing condition subjects are reex-
posed only to the material that they could produce on the test.
This suggests that some process other than additional exposure
is responsible for the effect.
Nevertheless, some authors have argued that the testing effect
simply reflects overlearning of items practiced on the test (e.g.,
Slamecka & Katsaiti, 1988; Thompson et al., 1978), concluding
that it is not the process of retrieval per se that promotes later
retention, but rather overlearning of a subset of the materials.
This explanation, however, encounters problems explaining why
additional studying produces better retention in the short term
than repeated testing does, even though testing produces better
long-term retention (e.g., Roediger & Karpicke, 2006; Wheeler
et al., 2003). That is, repeated studying apparently leads to
‘‘overlearning’’ on immediate tests, but this initial overlearning
does not translate into greater long-term retention because the
testing conditions show better recall than the repeated-study
conditions on delayed tests. In short, the additional-exposure, or
overlearning, account predicts a main effect at all retention
intervals and cannot explain the interaction that has been ob-
tained in several experiments. Finally, an account of the testing
effect based on additional exposure to, or overlearning of, the
material practiced on the test does not provide an explanation
for how tests can facilitate later retention of related material that
was not tested (Chan et al., in press). We agree with previous
researchers (Dempster, 1996; Glover, 1989) that accounts of the
testing effect based on additional processing or overlearning are
not satisfactory.
One other problem related to the exposure-overlearning ac-
count of the testing effect is also worth addressing. Some in-
vestigators may worry that the testing effect is nothing more than
the result of some sort of item-selection artifact because subjects
themselves select which items are recalled on an initial test. The
logic would be as follows: Some items are inherently easier than
other items (for whatever reason), and those easy items are re-
called on an initial test and then again on the final test, pro-
ducing the illusion that the test has caused learning when all it
did was show that easy items can be recalled twice. That is, the
‘‘easy’’ items receive additional practice through the test and are
better recalled later than items in the nontested control condi-
tion, in which they were not selected and practiced (see Mo-
digliani, 1976, for discussion). However, this account cannot
explain many important phenomena in the literature, such as the
crossover interactions observed as a function of retention in-
terval (e.g., Roediger & Karpicke, 2006; Wheeler et al., 2003).
Moreover, procedures developed to estimate and remove item-
Volume 1—Number 3 197
Henry L. Roediger, III, and Jeffrey D. Karpicke
selection effects (when initial test performance differs across
conditions) demonstrate that testing facilitates learning even
when item-selection effects are present in the data. For example,
Modigliani (1976) showed that increasing the delay before an
initial test led to increasingly greater effects of testing (Jacoby,
1978; Karpicke & Roediger, 2006a), and when the enhancement
effects due to testing were mathematically separated from item-
selection effects, the positive effects of delaying the initial test
were attributed entirely to enhancement effects, whereas item-
selection estimates remained invariant across the delays (and
were quite negligible to begin with). Other procedures for
handling item-selection problems were developed by Lockhart
(1975) and Bjork, Hofacker, and Burns (1981) and show simi-
lar results. To conclude, the testing effect is not simply a
result of additional exposure, or overlearning, or item-selection
Effortful Retrieval and Desirable Difficulties
If additional exposure and overlearning cannot explain the
testing effect, then the alternative is that some aspect of the
retrieval process itself must be at work. This is what Dempster
(1996) called the retrieval hypothesis. A variety of ideas about
how retrieval may affect later retention have been advanced,
although they may be describing the same process in somewhat
different words. Various writers have argued that retrieval effort
causes the testing effect (e.g., Gardiner, Craik, & Bleasdale,
1973; Jacoby, 1978). Alternatively, retrieval may increase the
elaboration of a memory trace and multiply retrieval routes, and
these processes may account for the testing effect (e.g., Bjork,
1975, 1988; McDaniel, Kowitz, & Dunay, 1989; McDaniel &
Masson, 1985). We consider these ideas in turn, but note that
they need not be mutually exclusive.
One explanation for why tests that require production, or re-
call, of material lead to greater testing effects than tests that
involve identification, or recognition, is that recall tests require
greater retrieval effort or depth of processing than recognition
tests (Bjork, 1975; Gardiner et al., 1973). Bjork (1975) argued
that depth of retrieval may operate similarly to depth of
processing at encoding (e.g., Craik & Tulving, 1975), and that
deep, effortful retrieval may enhance the testing effect. As
already discussed, increasing the spacing of an initial test—
which can be assumed to increase retrieval effort—promotes
better retention (Jacoby, 1978; Karpicke & Roediger, 2006a;
Modigliani, 1976), so long as material is still accessible and able
to be recalled on the test (Spitzer, 1939) or feedback is provided
after the test (Pashler et al., 2003). This positive testing effect
probably reflects greater retrieval effort on delayed tests.
Other evidence from different sorts of research also leads to
the general conclusion that retrieval effort enhances later re-
tention. Gardiner et al. (1973) asked students general knowl-
edge questions and measured the amount of time it took them to
answer the questions. At the end of the session, they gave sub-
jects a final free-recall test on the answers. The longer it took
subjects to produce the answer to a question (indicating greater
retrieval effort), the more likely they were to recall the answer on
the final test (see also Benjamin, Bjork, & Schwartz, 1998). In a
similar line of research, Auble and Franks (1978) gave subjects
sentences that were initially incomprehensible (e.g., The home
was small because the sun came out) and varied the amount of
time before they provided a key word that made the sentences
comprehensible (igloo). They found that the longer subjects
puzzled over the incomprehensible sentences (making an ‘‘effort
toward comprehension’’), the greater their retention of the sen-
tence on a final test. These studies demonstrate the positive
effects of retrieval effort on later retention, and the testing effect
reflects another example of retrieval effort promoting retention.
Other experiments have examined the multiplexing of re-
trieval routes by using the technique of varying cues given on a
first test to examine how the type of retrieval on the first test
affects performance on a second test given later (e.g., Bartlett,
1977; Bartlett & Tulving, 1974; McDaniel et al., 1989; Mc
Daniel & Masson, 1985). The general finding is that the nature
of the cues on the first test can affect how much that test
enhances performance on the second test (although in some
case, the exact nature of the experimental design matters; see
McDaniel et al., 1989, p. 434). For example, McDaniel and
Masson (1985) manipulated whether studied words were pro-
cessed with semantic or phonemic encoding tasks, the typical
levels-of-processing manipulation (Craik & Tulving, 1975).
Soon after study, subjects were given cued-recall tests with
phonemic or semantic cues, and the cues either matched or
mismatched the type of initial encoding. Subjects took a final
cued-recall test 24 hr later. (There were also conditions in which
items were tested only on the second test, to assess the testing
effect.) McDaniel and Masson found that the testing effect that
appeared on the second test was greater when the cues for the
first test mismatched the original encoding and yet successful
retrieval occurred than when the cues on the first test and the
type of encoding matched. This result can be understood as due
to an increase in the types of retrieval routes that permit access
to the memory trace (or perhaps a multiplexing of the features of
the memory trace itself).
Recently, Jacoby and his colleagues have obtained direct
experimental evidence for different depths of retrieval in a
memory-for-foils paradigm (Jacoby, Shimizu, Daniels, & Rhodes,
2005; Jacoby, Shimizu, Velanova, & Rhodes, 2005). In this
type of experiment, subjects encode material under shallow
or deep encoding conditions. During a first recognition test,
subjects discriminate between old words that were studied un-
der either the shallow or the deep conditions and new items (foils
or lures). They are later given a second recognition test that
assesses memory for the foils on the first test. For college stu-
dents, having taken the first recognition test with the meaning-
fully studied (or deeply studied) items enhanced recognition of
foils on the later test, compared with having taken the first test
with the shallowly studied items. Interestingly, older adults did
198 Volume 1—Number 3
The Power of Testing Memory
not show this difference (Jacoby, Shimizu, Velanova, & Rhodes,
2005), but for present purposes, the critical aspect of these
studies is that manipulation of the depth of retrieval on the first
test produced a large effect on recognition of the foils on the later
test among younger adults.
Bjork and Bjork (1992) developed a theory to explain the
testing effect and other effects of retrieval effort. They distin-
guished between storage strength, which reflects the relative
permanence of a memory trace or permanence of learning, and
retrieval strength, which reflects the momentary accessibility of
a memory trace and is similar to the concept of retrieval fluency,
or how easily the memory represented by the trace can be
brought to mind. Their model assumes that retrieval strength is
negatively correlated with increments in storage strength; that
is, easy retrieval (high retrieval strength) does not enhance
storage strength, whereas more effortful retrieval practice does
enhance storage strength and promotes more permanent, long-
term learning. However, because students often use the fluency
of their current processing (retrieval strength) as evidence about
the status of their current learning (e.g., see Jacoby, Bjork, &
Kelly, 1994), they may elect poor study strategies. That is,
students may choose strategies to maximize fluency of their
current processing, even though conditions that involve non-
fluent processing may be more beneficial to long-term learning.
For example, students may prefer massed study (or repeated
rereading) because it leads to fluent processing, although other
strategies (such as spaced processing or effortful self-testing)
would lead to greater long-term gains in knowledge.
Bjork (1994, 1999) has referred to techniques that promote
long-term retention even though they slow initial learning as
desirable difficulties and has argued that teachers should focus
on creating desirable difficulties for students in order to enhance
their learning. Techniques such as spaced practice (relative to
massed practice) and delayed feedback (relative to immediate
feedback) constitute desirable difficulties. We have argued that
relative to studying, testing also constitutes a desirable difficulty
(Roediger & Karpicke, 2006). Repeated testing tends to slow
initial learning relative to repeated studying (as evidenced on
final tests at a short retention interval), but testing promotes far
greater long-term retention (e.g., see Fig. 7).
Not surprisingly, people often do not voluntarily engage in
difficult learning activities, even though such activities may
improve learning. To give but one relevant example, Baddeley
and Longman (1978) trained postal workers on typing and
keyboard skills under massed- or spaced-practice conditions.
The subjects reported that they preferred the massed-practice
condition (and some refused to participate in further spaced-
practice training), even though spaced practice promoted far
better retention than massed practice. In many contexts, con-
ditions that lead to rapid gains in initial learning will produce
poor long-term retention, and likewise, conditions that make
learning slower or more effortful often enhance long-term re-
tention, with the testing effect being an example of the latter
scenario. To the extent that students monitor and guide their
learning on the basis of the fluency of their current processing,
they may fall prey to illusions of competence, believing that their
future performance will be greater than it really will be (see
Bjork, 1999; Jacoby et al., 1994; Koriat & Bjork, 2005, in press).
Because repeated testing is more effortful than repeated study-
ing, students may choose not to test themselves while learning,
and likewise, teachers may choose not to give many tests in their
classes. Implementing test-enhanced learning as a desirable
difficulty remains a challenge for education.
Transfer-Appropriate Processing
The concept of transfer-appropriate processing is also useful in
understanding the testing effect, although it should be seen as
perhaps incorporating some of the ideas discussed earlier in this
section at a more general level. Encoding may emphasize many
different strategies and types of processing, such as rote or
meaningful processing, as described in the levels-of-processing
tradition (Craik & Tulving, 1975), or item-specific (focused on
isolated facts) or relational (focused on relating ideas) process-
ing, as described in a different framework (Hunt & McDaniel,
1993). The idea behind transfer-appropriate processing is that
performance on a test of memory benefits to the extent that the
processes required to perform well on the test match encoding
operations engaged during prior learning (Morris et al., 1977;
see also Kolers & Roediger, 1984; McDaniel, Friedman, &
Bourne, 1978). Thus, the same study strategies or processes of
encoding that may greatly aid performance on one type of test
may have no effect or even an opposite effect on a different type
of test that emphasizes different types of information or
processing (e.g., Blaxton, 1989; Fisher & Craik, 1977). The idea
is similar to the encoding-specificity principle (Tulving &
Thomson, 1973) and emphasizes the critical relation between
encoding and retrieval processes. The concept of transfer-ap-
propriate processing has been applied to a wide array of phe-
nomena. For example, Roediger, Weldon, and Challis (1989)
argued that transfer-appropriate processing is critical for un-
derstanding differences between performance on explicit and
implicit memory tests (see also Blaxton, 1989; Roediger, 1990).
McDaniel (in press) pointed out that all situations in which
information is learned and then expressed through tests or ac-
tions involve transfer. He noted that although the idea of
transfer-appropriate processing seems obvious in prospect, in
practice it is often violated. He used the example of a teacher
who encourages excellent classroom study strategies that permit
deep understanding of the core concepts of the subject and how
they relate to one another, but then gives students a multiple-
choice test emphasizing recognition of isolated facts and won-
ders why the students perform so poorly. In this case, relational
processing strategies (although they may be good for long-term
retention) are poor for the specific test that the instructor gives.
Thomas and McDaniel (in press) provided experimental evi-
dence to bolster this point. Educators make the same point about
Volume 1—Number 3 199
Henry L. Roediger, III, and Jeffrey D. Karpicke
standardized tests; such tests may assess what is easy to measure
rather than the complex skills students may develop in class.
In applying transfer-appropriate processing to education, the
key question is what knowledge and skills the instructor wants
the students to know when they leave the course. One goal would
be being able to retrieve the information when it is needed, and
retrieval practice is critical to developing this skill. Taking tests
allows students to engage in retrieval operations during learning
and thus to practice the same skills needed to enhance subse-
quent retrieval. Such retrieval practice in taking tests permits
greater retention than does engaging in additional encoding
operations such as repeated reading (Roediger & Karpicke,
2006). Transfer-appropriate processing provides an explanation
for why taking memory tests often enhances performance on
later memory tests, especially when effortful retrieval is re-
quired. The results we have reviewed show that testing under
conditions of effortful retrieval has a greater transfer effect on
later test performance than testing under conditions of easy
retrieval. Of course, another educational goal is to have students
transfer information learned in courses to new problems they
face later in their jobs, but this kind of distant transfer is more
difficult to study although it remains a target for future research
(see Barnett & Ceci, 2002, for a review).
We believe that the concept of transfer-appropriate processing
offers an intuitive explanation for the somewhat counterintuitive
testing effect, and for this reason, the concept may be useful in
helping educators understand why taking tests should benefit
learning—testing leads students to engage in retrieval pro-
cesses that transfer in the long term to later situations and
contexts. However, we note one drawback to this approach.
One prediction that may be drawn from transfer-appropriate
processing is that performance on a final test should be best
when that test has the same format as a previous test. As we have
shown, the general finding is that recall tests promote learning
more than recognition tests, regardless of the final test’s format
(e.g., Kang et al., in press). This result needs confirmation
through additional experiments, but if it is true, it would seem to
be good news for educators, because it would lead to a
straightforward recommendation for educational practice.
Nonetheless, the same outcome (e.g., better transfer from a
short-answer test than from a multiple-choice test to a later
multiple-choice test) may be construed as inconsistent with
transfer-appropriate processing. However, it may not be incon-
sistent with the broader idea embodied in transfer-appropriate
processing. If, for example, a final multiple-choice test requires
effortful retrieval and a prior short-answer test fostered such
effortful </