Working memory span task: A methodological review and user’s guide.

Abstract

Working memory (WM) span tasks-and in particular, counting span, operation span, and reading span tasks-are widely used measures of WM capacity. Despite their popularity, however, there has never been a comprehensive analysis of the merits of WM span tasks as measurement tools. Here, we review the genesis of these tasks and discuss how and why they came to be so influential. In so doing, we address the reliability and validity of the tasks, and we consider more technical aspects of the tasks, such as optimal administration and scoring procedures. Finally, we discuss statistical and methodological techniques that have commonly been used in conjunction with WM span tasks, such as latent variable analysis and extreme-groups designs.

Full text

Psychonomic Bulletin & Review
2005, 12 (5), 769-786
Other than standardized instruments, such as intelligence
test batteries, working memory (WM) span tasks, such as
the counting span, operation span, and reading span tasks,
are among the most widely used measurement tools in cog-
nitive psychology. These tasks have come to prominence
not only for their methodological merit, but also because
theoretical advances in the study of human behavior since
the cognitive revolution have placed WM as a central con-
struct in psychology. Methodologically, WM span tasks
have proven to be both reliable and valid measures of WM
capacity (WMC), which we will document below. How-
ever, the larger factor in accounting for their increased use
is simply that WM has become a widely useful, scientifi-
cally fruitful construct. It plays an important role in con-
temporary global models of cognition (e.g., J. R. Anderson
& Lebiere, 1998; Cowan, 1995), and it is purportedly in-
volved in a wide range of complex cognitive behaviors, such
as comprehension, reasoning, and problem solving (Engle,
2002). Also, WMC is an important individual-differences
variable and accounts for a significant portion of variance
in general intellectual ability (Conway, Cowan, Bunting,
Therriault, & Minkoff, 2002; Conway, Kane, & Engle,
2003; Engle, Tuholski, Laughlin, & Conway, 1999; Kane
et al., 2004; Kyllonen, 1996; Kyllonen & Christal, 1990;
Süß, Oberauer, Wittmann, Wilhelm, & Schulze, 2002).
Furthermore, neuroimaging and neuropsychological stud-
ies have revealed that WM function is particularly depen-
dent on cells in the prefrontal cortex of the brain, which
has traditionally held a prominent status in the biological
approach to studying complex goal-directed human be-
havior (Kane & Engle, 2002).
A diverse set of researchers is now using WM as a con-
struct in research programs, as well as measures of WMC
769 Copyright 2005 Psychonomic Society, Inc.
We thank Chris Fraley for assistance in the effect size simulations and
for helpful discussions about prior versions of the manuscript. A.R.A.C. is
now at Princeton University. M.F.B. is now at the University of Missouri.
Correspondence concerning this article should be addressed to A. R. A.
Conway, Department of Psychology, Princeton University, Green Hall,
Princeton, NJ 08544-1010 (e-mail: aconway@princeton.edu).
THEORETICAL AND REVIEW ARTICLES
Working memory span tasks: A methodological
review and user’s guide
ANDREW R. A. CONWAY
University of Illinois, Chicago, Illinois
MICHAEL J. KANE
University of North Carolina, Greensboro, North Carolina
MICHAEL F. BUNTING
University of Illinois, Chicago, Illinois
D. ZACH HAMBRICK
Michigan State University, East Lansing, Michigan
OLIVER WILHELM
Humboldt University, Berlin, Germany
and
RANDALL W. ENGLE
Georgia Institute of Technology, Atlanta, Georgia
Working memory (WM) span tasks—and in particular, counting span, operation span, and reading
span tasks—are widely used measures of WM capacity. Despite their popularity, however, there has
never been a comprehensive analysis of the merits of WM span tasks as measurement tools. Here,
we review the genesis of these tasks and discuss how and why they came to be so influential. In so
doing, we address the reliability and validity of the tasks, and we consider more technical aspects of
the tasks, such as optimal administration and scoring procedures. Finally, we discuss statistical and
methodological techniques that have commonly been used in conjunction with WM span tasks, such
as latent variable analysis and extreme-groups designs.

770 CONWAY ET AL.
in the arsenal of research tools. Within psychology, dis-
cussions of WM are now common in almost all branches
of the discipline, including cognitive, clinical, social, de-
velopmental, and educational psychology. For example,
clinical research has demonstrated that WM is related to
depression (Arnett et al., 1999) and to the ability to deal
with life event stress (Klein & Boals, 2001) and is affected
by alcohol consumption (Finn, 2002). Social psycholo-
gists have revealed that students under stereotype threat
suffer reduced WMC and that WMC mediates the effect
of stereotype threat on standardized tests (Schmader &
Johns, 2003). Also, WMC is taxed and, subsequently,
depleted as a result of interracial interaction for highly
prejudiced individuals (Richeson et al., 2003; Richeson &
Shelton, 2003). In neuropsychology, deficits in WMC may
be a marker of early onset of Alzheimer’s disease (Rosen,
Bergeson, Putnam, Harwell, & Sunderland, 2002). De-
velopmental research suggests that the development of
WMC in children is central to the development of cogni-
tive abilities in general (Munakata, Morton, & O’Reilly,
in press) and that declines in WMC as a result of aging
are central to general cognitive-aging effects (Hasher &
Zacks, 1988). In short, recent research across the disci-
pline implicates WM as a central psychological construct
(for reviews, see Feldman-Barrett, Tugade, & Engle, 2004;
Unsworth, Heitz, & Engle, 2005).
Although the WM construct has been successfully and
appropriately exported from cognitive psychology to other
disciplines, WM tasks have, in our opinion, suffered a bit
in translation. Presumably because of their documented
reliability and validity, WM span tasks, among all the
available measures of WMC, have been embraced most
strongly by researchers outside of cognitive psychology.
Inevitably, as WM is exported to other scientific disci-
plines and as a more varied pool of investigators use WM
span tasks, misconceptions and misuses are bound to in-
crease. In particular, the literature presents inconsistent
information regarding the reliability of WM span tasks,
as well as inconsistent and, in our opinion, problematic
procedures for administration and scoring.
We therefore believe that the time is right for a review of
how and why WM span tasks came to dominate the mea-
surement landscape of WM research and to provide guide-
lines for researchers who would like to use these tasks. That
is, we perceive a need for a thorough review of all aspects
of WM span tasks, from optimal administration and scor-
ing procedures to assessment of reliability and validity.
Another motivating factor is more self-serving. Each au-
thor of this article receives numerous inquiries about these
tasks, including questions about administration, scoring,
assessment of reliability, use of extreme-groups designs,
and so forth, and so we devote considerable time to in-
structing other researchers about the proper use of these
tasks. As such, an important purpose of the present article
is to provide a “user’s guide” for WM span tasks, such that
any interested researcher can read this article, download the
programs from our Web site (http://psychology.gatech.edu/
renglelab/tasks.htm), and use them most appropriately.
We will begin with a historical overview of the devel-
opment of WM span tasks, followed by guidelines for
administration and scoring. The reliability and validity
of the tasks will then be discussed. The tasks will then be
contrasted with other empirical measures of WM func-
tion. Finally, we will discuss common research strategies
that have been used in conjunction with these tasks, such
as latent variable analysis and extreme-groups designs.
The review is limited to considering just three WM span
tasks; counting span (Case, Kurland, & Goldberg, 1982),
operation span (Turner & Engle, 1989), and reading span
(Daneman & Carpenter, 1980). These three span tasks
were chosen because there is much more data from these
tasks than from any of the others (e.g., spatial WM span
tasks; see Kane et al., 2004; Shah & Miyake, 1996), and
the principles outlined here should generalize not only to
other span tasks, but also to other measures of cognitive
ability. Our hope is that this review will serve as an ex-
ample of how to assess measurement instruments within
any particular research domain. We also hope that it lays
bare the importance of taking issues of measurement seri-
ously in one’s research program. We should also note at
the outset that the particular tasks under review here have
been used primarily to investigate individual differences
in healthy young adults, and therefore, our recommen-
dations apply best for similar purposes in similar popu-
lations. That said, we will indicate, where possible, how
these tasks might be modified for other applications.
Finally, we note that the review is as theory neutral as
possible with respect to a particular model of WM and/or
the nature of individual differences in WMC, because our
goal is to review the merits of WM span tasks as research
tools for any researcher, with any theoretical stance. How-
ever, it is simply impossible to review the history of these
tasks, discuss their validity, or even suggest how to score
them without revealing some theoretical bias. Therefore,
at the outset, we will briefly outline our theoretical ap-
proach to WM, for the simple purpose of warning the
reader of any bias that may reveal itself later.
BRIEF THEORETICAL OVERVIEW
OF WM AND WMC
We view WM as a multicomponent system responsible
for active maintenance of information in the face of ongo-
ing processing and/or distraction. Active maintenance of
information is the result of converging processes—most
notably, domain-specific storage and rehearsal processes
and domain-general executive attention. Furthermore, the
extent to which maintenance depends on domain-specific
skills versus domain-general executive attention varies as
a function of individual ability, task context, and ability
× context interactions. For instance, a novice chess player
will rely more on domain-general executive attention to
maintain game information (e.g., recent moves or future
positions) than on domain-specific skills (e.g., learned
strategies and position patterns). In contrast, an expert
chess player typically will rely more on domain-specific

WORKING MEMORY SPAN TASKS 771
processes and skills to maintain information. However,
even the expert might need to call upon executive atten-
tion under some circumstances, such as playing the game
in particularly demanding situations or under some sort of
cognitive or emotional load.
Similarly, performance on WM span tasks depends
on multiple factors, with domain-specific skills, such as
chunking and rehearsal, facilitating storage and a domain-
general capability allowing for cognitive control and ex-
ecutive attention. A critical aspect of our view, however,
is that WM span tasks predict complex cognitive behavior
across domains, such as reading comprehension, problem
solving, and reasoning, primarily because of the general,
executive attention demands of the tasks, rather than the
domain-specific demands of the tasks (Conway & Engle,
1996; Conway et al., 2003; Engle, Cantor, & Carullo, 1992;
Engle, Tuholski, et al., 1999; Turner & Engle, 1989). As
such, we make a clear distinction between the traditional
concept of short-term memory capacity (STMC) and
WMC. STMC is thought to reflect primarily domain-spe-
cific storage, whereas WMC is thought to reflect primar-
ily domain-general executive attention (Engle, Tuholski,
et al., 1999). More specifically,
By “executive attention” we mean an attention capabil-
ity whereby memory representations are maintained in a
highly active state in the face of interference, and these
representations may reflect action plans, goal states, or
task-relevant stimuli in the environment. Critical to our
view is that, while the active maintenance of information
can be useful in many situations, it is most necessary under
conditions of interference. This is because in the absence
of interference, task-relevant information or goals may be
easily retrieved from long-term memory as needed. Under
interference-rich conditions, however, incorrect informa-
tion and response tendencies, are likely to be retrieved, and
so such contexts set the occasion for the reliance on ac-
tive maintenance of information. (Kane & Engle, 2002,
p. 638)
Also, in claiming that executive attention is domain
general, we make no distinction between verbal WMC
and spatial WMC (or any other domain WMC for that
matter). Kane et al. (2004) recently provided empirical
support for this position, showing that verbal WM span
tasks, such as counting, operation, and reading span, load
on the same factor in a factor analysis as WM span tasks
in which the tasks demand spatial processing and stor-
age. More detailed theoretical reviews of our approach, as
well as further empirical support for the claims presented
here, can be found elsewhere (e.g., Conway et al., 2002;
Conway & Kane, 2001; Conway et al., 2003; Engle, 2001,
2002; Engle, Kane, & Tuholski, 1999; Engle, Tuholski,
et al., 1999; Kane, Bleckley, Conway, & Engle, 2001;
Kane & Engle, 2002; Kane et al., 2004). Finally, because
our view is concerned primarily with the domain-general
attentional aspect of WM, our claims above are consistent
with either structural (e.g., Baddeley, 1986) or functional
(e.g., Nairne, 2002) models of storage.
HISTORY OF WM SPAN TASKS AND
OVERVIEW OF ADMINISTRATION
PROCEDURES
WM span tasks, such as counting span, operation span,
and reading span, were designed from the perspective
of Baddeley and Hitch’s (1974) theory of WM, which
stressed the functional importance of an immediate-
memory system that could briefly store a limited amount
of information in the service of ongoing mental activ-
ity. That is, a WM system would be unlikely to evolve
for the sole purpose of allowing an organism to store or
rehearse information (such as a phone number) while it
was doing nothing else. A more adaptive system would
allow the organism to keep task-relevant information ac-
tive and accessible in memory during the execution of
complex cognitive tasks. WMC measures were, therefore,
created to require not only information storage and re-
hearsal (as do “simple” measures of STMC, such as digit
span or word span), but also the simultaneous processing
of additional information (Case et al., 1982; Daneman &
Carpenter, 1980; Turner & Engle, 1989). Such WM span
tasks interleave the presentation of to-be-remembered tar-
get stimuli, such as digits or words, with the presentation
of a demanding, secondary processing task, such as com-
prehending sentences, verifying equations, or enumerat-
ing an array of shapes.
The reading span task was the first task developed with
the purpose of jointly tapping the storage and process-
ing functions of WM (Daneman & Carpenter, 1980). The
task is essentially a simple word span task, with the added
component of the comprehending of sentences. Subjects
read sentences and, in some cases, verify the logical ac-
curacy of the sentences, while trying to remember words,
one for each sentence presented. The sentences are pre-
sented in groups that typically range in size from two to
six (we refer to a group of sentences as one item).
1
Word
recall is prompted at the completion of an item.
In the original version of reading span (Daneman &
Carpenter, 1980, Experiment 1), subjects were required
to read aloud, at their own pace, sentences presented on
index cards, while remembering the last word of each
sentence for later recall. After a series of sentences, the
subject recalled the to-be-remembered words in the order
in which they had been presented. There were 15 items, 3
each consisting of two, three, four, five, and six sentences
that were 13–16 words in length, and they were presented
in ascending order (i.e., from smallest to largest). Increas-
ingly larger items were presented until the subject failed
to recall all 3 items of a given size. At this point, the ex-
periment was terminated. A subject’s reading span was
the level at which he or she could correctly recall 2 of the
3 items. For example, if a subject were to successfully
recall at least 2 out of 3 two-word items, the experiment
would continue for the subject to attempt 3-word items. If
the subject were then to successfully recall only 1 out of 3
of the 3-word items, the experiment would terminate, and
the subject’s reading span would be 2.

772 CONWAY ET AL.
Daneman and Carpenter (1980) added a true–false
component to the task in their Experiment 2. The subjects
indicated the veracity of each sentence they read by re-
sponding true or false within 1.5 sec of each sentence’s
presentation. Here, the sentences were drawn from gen-
eral knowledge quiz books and covered multiple domains,
including the biological and physical sciences, literature,
geography, history, and current affairs, and were selected
to be of moderate difficulty (e.g., “You can trace the lan-
guages English and German back to the same roots”). Al-
though Daneman and Carpenter (1980) did not monitor
the subjects’ accuracy on the true–false component, the
subjects believed it was an important part of the task. This
prevented the subjects from adopting a strategy of focus-
ing on the final words without devoting much attention to
reading the sentences.
The reading span task used by Engle and colleagues
is somewhat different from Daneman and Carpenter’s
(1980) original version, but the fundamental premise of
the task is unchanged. For instance, Turner and Engle
(1989) created a version of reading span that consisted of
fewer overall items (12 items, 3 each consisting of two,
three, four, and five sentences). Also, whereas Daneman
and Carpenter (1980) tested subjects on the veracity of the
sentences, Turner and Engle tested subjects on whether
the sentences were semantically and syntactically correct
(e.g., “The grades for our finals will classroom the out-
side posted be door”). Turner and Engle administered their
reading span task to small groups of subjects, rather than
individually. The sentences were projected on a screen via
an overhead transparency, and the experimenter used a
sheet of paper to keep all but the current sentence hidden.
The stimuli were simultaneously presented auditorially
by means of a prerecorded cassette, which served to pace
the subjects through the task. Thus, the subjects heard the
sentences while they read them aloud. This procedure var-
ied considerably from the self-paced reading in Daneman
and Carpenter’s (1980) version of the task and afforded
some opportunity for error. Fast readers, especially, might
devise strategies to devote more time to memorizing (i.e.,
coding and rehearsing) the to-be-remembered stimuli.
Recognizing this potential problem, Turner and Engle
tested subjects individually in their Experiment 2. They
also monitored accuracy on the sentence verification
component of the task, and subjects who scored below
80% were excluded from all analyses. This criterion fur-
ther helped ensure that attention was paid to the process-
ing component of the task.
The version of reading span that we have most com-
monly used is one in which the to-be-remembered word is
different from the last word, or any word, in the sentences
(see, e.g., Engle, Tuholski, et al., 1999), and so each sen-
tence is followed by an unrelated word. Subjects still read
the sentences aloud and verify whether the sentence is
semantically or syntactically correct. Here, however, they
are charged with remembering the unrelated word and
not the last word of the sentence. We made this change
because individual differences in reading ability could
lead to differences in the ability to generate the words at
test on the basis of the gist of the sentence (rather than on
the basis of episodic recall). In our latest version of the
task, subjects no longer remember words; instead, they
remember isolated letters that follow each sentence (Kane
et al., 2004).
Does the Secondary Task Need to Involve
Reading?
Daneman and Carpenter (1980) argued that reading is
an integral component of their span task, and in order to
predict reading ability, a WM span task must make use of
reading strategies. Daneman and Carpenter (1980) took
the position that WMC is strategy specific, subsequently
known as the task-specific view (cf. Engle et al., 1992).
On this view, subjects who have developed effective strat-
egies for the processing component of the task will have
greater capacity to devote to storage. Turner and Engle
(1989), however, hypothesized that WMC is independent
of the specific nature of the processing component of the
span task. A highly demanding processing component
is necessary to engage the processing functions of WM
and draw out individual differences in task performance.
Turner and Engle showed that they could predict read-
ing ability with a WM span task that does not involve
the reading of sentences. Their task, the operation span
task, required that subjects solve mathematical operations
while trying to remember words.
There were 84 mathematical operation strings in Turner
and Engle’s (1989) first operation span task. Each string
consisted of a mathematical equation with two arithmetic
operations on one side of the equation and a stated solu-
tion on the other side of the equation. The first operation
was a simple multiplication or division problem and was
followed by a simple addition or subtraction operation.
The stated solution was correct on half of the trials. The
following are examples of a correct and incorrect equa-
tion, respectively: (9/3)  2  1 and (9/3)  2  6.
Turner and Engle’s (1989) operations replaced the sen-
tences in Daneman and Carpenter’s (1980) task, but other-
wise the task demands were largely unchanged. In the first
published version of the task, 12 items were presented,
with 3 items each consisting of two, three, four, and five
operation word pairs, presented in ascending order. The
stimuli consisted of mathematical operations followed by
a to-be-remembered word, drawn from the same normed
set of common four- to six-letter words as that for Turner
and Engle’s version of the reading span task. The task was
administered to small groups of subjects (Experiment 1)
and individually (Experiment 2) by the same means as
the reading span task. When the operation word string
was presented, the subjects read the operation aloud and
verified whether the stated solution was correct or incor-
rect. They then read aloud and remembered the word for
later recall. All intermediate calculations were done si-
lently and without the aid of pencil and paper. Engle et al.
(1992) developed the version of the operation span task
currently used in our laboratories. The primary difference

WORKING MEMORY SPAN TASKS 773
from earlier versions is the manipulation of presentation
order. Rather than presenting reading span and operation
span items in ascending order (items with fewer elements
first), which permitted the subjects to anticipate the num-
ber of words that they would be asked to remember on
any given trial, Engle et al. (1992) randomized the pre-
sentation order, effectively eliminating reliance on any
strategies that come from knowing the size of the memory
set.
2
This modification has the added benefit of decon-
founding item size and buildup of proactive interference,
since recent studies have shown that proactive interfer-
ence builds from trial to trial in WM span tasks (Bunting,
in press; Lustig, May, & Hasher, 2001; May, Hasher, &
Kane, 1999). It also results in a wider range of scores than
does the traditional ascending approach. A potential risk
of this approach, however, is that the early presence of
difficult items may discourage some subjects, particularly
those who are less able, such as children, the elderly, or
patients. (Our advice to researchers working with such
populations is to stress to the subject that perfect recall is
not expected in these tasks.)
Case et al.’s (1982) counting span task has also fre-
quently been used to measure WMC, particularly in
school-aged children (the simplicity of the processing
component—i.e., counting—makes this task ideal for
a variety of populations, including patients, the elderly,
and nonnative English speakers). On the surface, the
commonalities between the counting span task and the
reading span task or the operation span task are not eas-
ily apparent, but the underlying structure is the same for
all three tasks. Whereas most versions of reading span
and operation span tasks require subjects to remember
words, the counting span task involves counting shapes
and remembering the count totals for later recall. In Case
et al.’s version of the counting span task, subjects orally
counted (and pointed their finger at) the green dots pre-
sented against a white background. Yellow dots, inter-
leaved with the green dots, disrupted the visual patterns
of the green dots (Case et al. did not report a range of
values for the number of green or yellow dots presented).
The task presented three items of each size from one to
five, in ascending order.
In a version of the task designed for adults by Engle,
Tuholski, et al. (1999), the stimuli consisted of three items
of each size from one to eight. The visual displays were
made more complex by placing the target shapes among a
field of distractors that shared either the same shape or the
same color (and so counting required conjunctive search,
à la Treisman & Gelade, 1980). Each display consisted of
a random arrangement of three to nine dark blue circles,
one to nine dark blue squares, and one to five light blue
circles. The subject was to count all of the dark blue cir-
cles without pointing. When the final shape was counted,
the subject repeated the total, to signal being finished.
The experimenter immediately presented the next display,
and the subject commenced counting immediately. The
subject recalled the total number of dark blue circles from
each display in serial order.
Summary of Critical Task Components
The reading span, operation span, and counting span
tasks share an underlying structure and are implemented
in much the same way. The tasks are designed to force
WM storage in the face of processing (or distraction), in
order to engage executive attention processes. As such,
the following procedural recommendations should be
kept in mind when these tasks are administered (again,
note that these recommendations are most appropriate for
studies involving healthy young adults).
Immediate and vigilant stimulus presentation
. A
critical feature of the processing component of WM span
tasks is that it interferes with rehearsal. Substantial delays
between stimulus presentations may, therefore, permit re-
hearsal of the to-be-remembered stimuli, thereby making
the task more a measure of STM storage than of WM/ex-
ecutive functioning. Indeed, Friedman and Miyake (2004)
found that an experimenter-paced version of reading span
correlated more strongly with verbal SAT (r  .49) and
reading comprehension (r  .55) than did a subject-paced
version of the task (rs  .18 and .28, respectively). More-
over, several studies have documented, via partial corre-
lations, that the time subjects spend on the processing or
storage components of self-paced span tasks (or very gen-
erously paced tasks) can suppress the correlation between
span and ability (Engle et al., 1992; Friedman & Miyake,
2004; Turley-Ames & Whitfield, 2003). When WM span
tasks are administered, then, each stimulus subsequent to
the first stimulus in an item should be presented imme-
diately upon completion of the preceding stimulus, and
subjects should be instructed to begin acting upon stimuli
immediately.
Individual administration
. Versions of the reading
span task and the counting span task have been designed
for either single-subject or group sessions. Administering
the task to more than 1 subject at a time, however, intro-
duces greater potential for error. Subjects must adequately
attend to the processing component of the task in order for
the processing component to disrupt rehearsal. When a
WM span task is administered to groups, it is more diffi-
cult for the experimenter to observe whether the subject is
attending to the processing task. Moreover, subjects who
are more skilled in the processing task may complete it
more quickly than do others in the same session, leaving
them more time to rehearse the target stimuli.
Sufficient item size
. We have described multiple ver-
sions of the reading span task, for which the range of item
sizes varied from two to five or six. Other published ver-
sions have used an even smaller range (from two to four;
e.g., May et al., 1999). Whereas larger item sizes consid-
erably increase the running time of the experiment, insuf-
ficient item sizes create the potential for ceiling effects
among those subjects in the upper end of the performance
distribution. We consider the range from two to five ele-
ments per item to be adequate for most college student
populations (on the basis of distributions from Conway
et al., 2002, Engle, Tuholski, et al., 1999, and Kane et al.,
2004).

774 CONWAY ET AL.
SCORING
The scoring of WM span tasks is a neglected topic in
the research literature, and this is unfortunate because dif-
ferent scoring procedures not only may affect the rank
order of subjects, but also may have implications for data
analyses. Generally, scoring measures of cognitive be-
havior are considered to be straightforward and simple
processes. However, in dual-task situations, such as WM
span tasks, there are two sources of data: one from the
processing component of the task and one from the stor-
age component. In operation span, for example, multiple
data points might be collected, such as accuracy on the
math problems, time spent processing the math problems,
and recall of the words.
3
Correlational evidence from studies on adults supports
the common procedure of not considering processing
performance in the WM span score. First, processing ac-
curacy is typically close to ceiling, because task instruc-
tions emphasize processing-task accuracy to ensure that
subjects are attending to the secondary task. Second,
despite this near-ceiling accuracy, performance on the
processing component usually correlates positively with
performance on the storage component: Subjects who re-
call the most target items also perform most accurately on
the processing task (Kane et al., 2004; Waters & Caplan,
1996). Thus, there is typically no evidence for process-
ing/storage trade-offs.
In the traditional scoring of WM span tasks, the subject
is assigned a quasi-absolute span score (e.g., Daneman
& Carpenter, 1980; Waters & Caplan, 1996). The task
begins with an item consisting of two elements and con-
tinues until the subject’s accuracy falls below a threshold.
Once this threshold is reached, testing is discontinued,
and the last item size recalled with a specific probability
(say, four out of five items) is the span score. The under-
lying assumption here is that items with a given load or
demand “meet” a person with a given ability. The person
is either able to solve an item or not, and so item difficulty
and person ability are on the same scale. If item difficulty
exceeds the subject’s ability, the probability of correct re-
sponse is low, and if the subject’s ability exceeds item dif-
ficulty, the probability of correct response is high.
A problem with these absolute scoring methods is
that the difficulty of a span item may vary on many di-
mensions, thus threatening span reliability across dif-
ferent tasks (or different versions of the same task). For
example, other things being equal, longer sentences in a
reading span task should decrease the quantity of recalled
words (see Towse & Hitch, 1995; Towse, Hitch, & Hutton,
1998, 2002). Similarly, the display duration for individual
sentences, or the semantic similarity of the stimuli, could
have an influence on recall performance (see Copeland &
Radvansky, 2001). Thus, there are several ways in which
various instantiations of the same WM span measure
might yield different “span” values for the same person.
A second problem in studies in which absolute scores are
used is that, by simply estimating the item size at which a
subject falls below a given threshold (and then ending the
task), information on all other trials is discarded. Here,
the scores can take only one of very few values, usually
somewhere between 2 and 6, greatly limiting the sensitiv-
ity of the measure (see Oberauer & Süß, 2000). We there-
fore suggest that absolute span scores are inappropriate
for individual-differences research.
Consider, instead, a much simpler scheme. Correct
responses to individual elements within an item are as-
signed one number, and all other responses are assigned
a different number (e.g., correct  1 and incorrect  0),
with no distinction among different types of errors. That
is, errors are not classified as omissions or commissions,
as more or less erroneous, or as indicating any particular
cognitive process. For items of various sizes, there are
varying numbers of observations. For an item with six
elements, there are three times as many responses as for
an item with two elements. Other things being equal, then,
we may assume that items with more elements are more
reliable indicators than are those with fewer elements, be-
cause longer items rely on more instances of the behavior
of interest.
In a next step, the data are aggregated, and here there
are several possible procedures from which to choose. To
illustrate, consider the performance of the fictional but re-
alistic subject depicted in Table 1 (correct serial recall of
individual elements is presented for various items across
various tasks). This person is performing rather well, but
even on items with only two elements, not all the elements
are correctly recalled in serial order. Moreover, on items
imposing a higher memory load, the person sometimes
recalls fewer elements than on items with a smaller load.
Also note that the person is performing somewhat differ-
ently on the different tasks.
In order to assign a score to this person, some deci-
sions need to be made. Should credit be given if the re-
call of elements was correct but there were errors on the
associated processing component of the task? Should
full or partial credit be given if some, but not all, of the
elements were recalled in the correct serial position?
Table 1
Results From Three Working Memory Span Tasks for Person A
No. of
Elements Item No.
Counting
Span
Operation
Span
Reading
Span
21222
2212
3222
31333
2333
3332
41432
2423
3234
51544
2542
3 3 4 3
Note—Each cell represents the number of elements recalled correctly
for that item.

WORKING MEMORY SPAN TASKS 775
Should a higher weight be assigned to items with a higher
memory load (i.e., more elements)? In the WM literature,
these questions generally have been raised implicitly, if at
all, but from a psychometric perspective it is crucial that
these scoring decisions are theoretically and empirically
informed.
Our own answer to the first question is straightforward.
We assign credit to elements recalled despite errors made
on the processing component of the task (for the reasons
we discussed previously). That said, we do strive to ensure
that accuracy on the processing component of the task is
near perfect. If accuracy on the processing component of
the task falls below a certain level (typically, 85%), the
entire data set for that subject is discarded.
The latter decisions require more discussion. The first
question is whether partial-credit scoring, in which credit
is given to partly correct items, is superior to all-or-nothing
scoring, in which credit is given only to completely correct
items (i.e., where all elements are recalled in the correct se-
rial position). Considering our fictional subject in Table 1,
a partial-credit procedure would give some points for items
with a memory load of five, but the all-or-nothing scor-
ing procedure would not. The second question is whether
all items should count the same or whether those with a
higher memory load should contribute more to the overall
score. Counting all items equally is done by scoring each
item as a proportion of correctly recalled elements per
item, regardless of item size (e.g., recalling one element
from a two-element item would count as much as recall-
ing two elements from a four-element item—i.e., .50).
These proportions are then averaged. In contrast, giving
a higher weight to items with a higher load is done by
computing the mean of all correctly recalled elements (ir-
respective of item size). Note that these two decisions,
one regarding partial versus all-or-nothing credit, and one
regarding equal weighting (or unit weighting) versus load
weighting, are orthogonal to each other. Consequently,
we will consider four scoring procedures that cross these
solutions: partial-credit unit scoring (PCU), all-or-nothing
unit scoring (ANU), partial-credit load scoring (PCL),
and all-or-nothing load scoring (ANL).
The results for our fictional subject, for the four scoring
procedures, are summarized in Table 2. For unit scoring,
PCU expresses the mean proportion of elements within
an item that were recalled correctly, and ANU expresses
the proportion of items for which all the elements were
recalled correctly. For load-weighted scoring procedures,
PCL represents the sum of correctly recalled elements
from all items, regardless of whether the items are per-
fectly recalled or not (also without respect to serial order
within items), and ANL represents the scoring method we
have most often used, reflecting the sum of the correctly
recalled elements from only the items in which all the ele-
ments are recalled in correct serial order.
Although load scoring is rather uncommon in psycho-
metrics, it is perhaps the most frequently applied method
for span measures. This tradition can be traced as far back
as Ebbinghaus (1897), who used a digit span task as an
individual-differences measure and applied a partial-
credit load-weighted scoring procedure. However, load-
weighted scoring is rarely used in psychometrics, because
there simply is no good reason to assign a greater weight
to harder items. That is, all items within a task, such as
WM span, are supposed to measure the same underlying
ability, such as storage in the face of concurrent process-
ing; they just discriminate at different points along the
ability distribution. Moreover, a typical consequence of
load-weighted scoring is positive skew: Individual dif-
ferences in the upper half of the ability distribution are
inflated, relative to the lower half of the ability distribu-
tion. Obviously, normal distributions are to be preferred
in correlational studies.
Table 2
Results From Four Scoring Procedures for Three
Working Memory Span Tasks of One Fictitious Subject
Scoring Procedure Counting Span Operation Span Reading Span
Partial-credit unit scoring (1  1  1
 1  1  1
 1  1  .5
 1  1  .6)
 11.1/12  .93
(1  .5  1
 1  1  1
 .75  .5  .75
 .8  .8  .8)
 9.9/12  .83
(1  1  1
 1  1  .67
 .5  .75  1
 .8  .4  .6)
 9.72/12  .81
All-or-nothing unit scoring (1  1  1
 1  1  1
 1  1  0
 1  1  0)
 10/12  .83
(1  0  1
 1  1  1
 0  0  0
 0  0  0)
 5/12  .42
(1  1  1
 1  1  0
 0  0  1
 0  0  0)
 6/12  .50
Partial-credit load scoring (2  2  2
 3  3  3
 4  4  2
 5  5  3)
 38/42  .90
(2  1  2
 3  3  3
 3  2  3
 4  4  4)
 34/42  .81
(2  2  2
 3  3  2
 2  3  4
 4  2  3)
 32/42  .76
All-or-nothing load scoring (2  2  2
 3  3  3
 4  4  0
 5  5  0)
 33/42  .79
(2  0  2
 3  3  3
 0  0  0
 0  0  0)
 13/42  .31
(2  2  2
 3  3  0
 0  0  4
 0  0  0)
 16/42  .38

776 CONWAY ET AL.
To empirically compare these various scoring proce-
dures in action, we reanalyzed the data from Kane et al.
(2004). In a study with 236 subjects from both university
and community samples, Kane et al. (2004) administrated
three verbal WM tasks: operation span (12 items of two
to five elements), counting span (15 items of two to six
elements), and reading span (12 items of two to five ele-
ments). Table 3 presents internal consistencies, as indi-
cators of task reliability, for the four scoring methods in
these three tasks. Partial-credit scoring procedures show
a clear advantage, and within these, unit-weighted scoring
has a slight advantage over load-weighted scoring.
The correlations among the scores for all four scoring
procedures within a task rely on the same initial informa-
tion (whether or not an individual element was recalled
correctly). Consequently, the correlations among scoring
procedures necessarily are very high. Within the all-or-
nothing and partial-credit scoring procedures, all cor-
relations are .98 or higher for the three tasks. However,
correlations between all-or-nothing scores and partial-
credit scores within a task are substantially lower (al-
though still high, ranging from .87 to .93). On the basis of
this information there is no relevant difference between
load weighting and unit weighting, once one commits to
partial-credit versus all-or-nothing scoring. However, cor-
relating partial-credit scores with all-or-nothing scores
shows substantial deviations from perfect correlations
within all of the tasks, and so researchers’ decisions re-
garding these options should be considered carefully and
justified theoretically.
In summary, established procedures of assigning absolute
spans have various shortcomings, and so scoring procedures
that exhaust the information collected with a task should be
used instead. Because empirical results favor partial-credit
scoring, we prefer it over all-or-nothing scoring. Our prefer-
ence between unit-weighted and load-weighted procedures
is less strong. The empirical results—including approxi-
mation of normal distributions—do not strongly favor one
of these procedures over the other. However, one might
favor unit-weighted scoring because it follows established
and sound procedures from psychometrics.
RELIABILITY
Reading span, operation span, and counting span have
been administered to literally thousands of subjects in
over a hundred independent studies. One conclusion that
can be drawn from this body of research is that measures
obtained from these tasks (span scores) have adequate re-
liability. That is, irrespective of what WM span tasks are
supposed to measure, evidence suggests that they mea-
sure, with reasonable accuracy, whatever it is that they
actually measure. For example, estimates of reliability
based on internal consistency, such as coefficient alphas
and split-half correlations, which reflect the consistency of
participants’ responses across a test’s items at one point in
time, are typically in the range of .70–.90 for span scores,
where values can range from 0 (no reliability) to 1 ( per-
fect reliability). As a specific example, with a sample size
of 236, Kane et al. (2004) observed coefficient alphas of
.78 for reading span, .80 for operation span, and .77 for
counting span. This indicates that subjects who responded
with the correct answer for one set of span stimuli in these
tasks (e.g., equation word pairs in operation span) tended
to respond with the correct answer on the others (and vice
versa). Therefore, span scores were reliable in the sense
that there was consistency in responding across items
within the task at one point in time. Internal consistency
estimates of similar magnitudes have been reported in a
number of other large-scale studies, including Conway
et al. (2002), Engle, Tuholski, et al. (1999), Hambrick and
Engle (2002), Miyake, Friedman, Rettinger, Shah, and
Hegarty (2001), and Oberauer, Süß, Schulze, Wilhelm,
and Wittmann (2000).
Evidence also suggests that WM span tasks are reliable
in the sense that the rank order of span scores are stable
across time. In adults, test–retest correlations of approxi-
mately .70–.80 have been observed for operation span and
reading span, over minutes (Turley-Ames & Whitfield,
2003), over weeks (Friedman & Miyake, 2004; Klein &
Fiss, 1999), and even over 3 months (Klein & Fiss, 1999).
4
In children, Hitch, Towse, and Hutton (2001) found op-
eration span to be slightly less reliable over a year (.56),
but reading span was more acceptable (.71). Importantly,
Hitch et al. also found that the original administrations of
operation and reading span predicted number skills and
verbal skills measured 1 year later and that the second
administrations of the span tasks accounted for very little
additional variance beyond the first.
Although two studies have shown less adequate test–
retest reliability for the reading span task, ranging from
.50 over weeks to .40–.65 over months (MacDonald,
Almor, Henderson, Kempler, & Andersen, 2001; Waters
& Caplan, 1996), the observation that span scores cor-
relate strongly with various other measures provides ad-
ditional evidence for their reliability. This is because the
correlation between two measures is limited by reliabil-
ity. More specifically, the correlation between any two
measures (x and y) cannot exceed the square root of the
product of their reliabilities. That is,
rrr
xy xx yy
≤⋅
(
)
sqrt .
Therefore, when span scores from one WM span task cor-
relate with span scores from another span task, the impli-
cation is that the span scores must have some degree of
reliability—given that if a measure has zero reliability,
Table 3
Internal Consistency for Three Working Memory
Span Tasks With Four Scoring Procedures
Task PCU ANU PCL ANL
Counting span .768 .668 .763 .673
Operation span .814 .698 .808 .701
Reading span .788 .697 .776 .699
Note—PCU, partial-credit unit scoring; ANU, all-or-nothing unit scor-
ing; PCL, partial-credit load scoring; ANL, all-or-nothing load scoring.

WORKING MEMORY SPAN TASKS 777
correlation with all other measures will necessarily be
zero. To illustrate, in Kane et al.’s (2004) study, the cor-
relation between operation span and reading span was .69.
This indicated that for either task, reliability must have
been at least .69. As other examples, Conway and Engle
(1996) and Lehto (1996) found correlations among op-
eration span tasks that varied in processing difficulty in a
range from .70 to .80, suggesting, again, that the reliabil-
ity of operation span is at least .70.
One way that this evidence for reliability can be under-
stood is in terms of classical test theory, first proposed by
Spearman (1904). Briefly, the basic assumption of classi-
cal test theory is that a single score on a test—an observed
score (x)—consists of two components. The true score (t)
is assumed to reflect stable aspects of the trait (or traits)
that the test measures, whereas error (e) is conceptualized
as a random fluctuation in scores. That is,
x = t + e.
Correspondingly, the total variance of scores on a test (σ
x
2
)
is decomposed into true-score variance (σ
t
2
) and error
variance (σ
e
2
). That is,
σσσ
x
222
=+
te
.
Finally, reliability (r
xx
) is interpreted as the proportion of the
total variance that is attributable to true-score variance:
r
xx x
=−
(
)
1
22
σσ
e
.
Or conversely, the proportion of the total variance that
is attributable to error variance is equal to one minus the
reliability:
σσ
e
22
1
xxx
r=− .
Within this framework, it can be demonstrated that
span scores are influenced more by stable true scores than
by error. Consider again the coefficient alphas observed
by Kane et al. (2004) for reading span, operation span,
and counting span: .78, .80, and .77, respectively. A coef-
ficient alpha—the average reliability resulting from all
possible split-half correlations for a test—indexes error
due to factors operating at a given point in time, including
momentary fluctuations of attention or mood, fatigue, and
so forth. Thus, it follows that the proportion of the total
variance in scores due to such random factors was 22%
for reading span (1  .78), 20% for operation span (1 
.80), and 23% for counting span (1  .77).
In sum, as with any psychological instrument, no WM
span task is free of measurement error. In other words, no
WM span task is perfectly reliable. However, it is clear
that WM span tasks do a reasonable job of measuring ac-
curately whatever it is that they measure. As evidenced by
acceptable reliability estimates, as well as by moderate
to strong correlations with other measures, it is evident
that span scores are influenced by something stable, with
a minor contribution of error due to random fluctuations
in scores. Next, we will consider the question of what this
stable something is.
VALIDITY
Much of cognition is ballistic, in that one thought
leads to the next through automatic activation. Atten-
tion is often captured by events in the environment and
by thoughts that intrude into consciousness. Those per-
ceptions and thoughts, in turn, lead inexorably to other
thoughts. However, the solution to life’s problems often
requires that such automatically elicited thoughts, asso-
ciations, and captured attention be resisted and thought
be directed or controlled. We have argued that this ability
to control attention and thought represents the common
construct measured by tests of WMC. The evidence is
quite clear that there are abiding individual differences in
the ability to control attention and thought and that those
differences are reflected by WM span tasks. It is also be-
coming clear that, in addition to the abiding individual
differences in ability to control cognition, a host of other
variables, from drunkenness to depression, also influence
this ability, and those variables also manifest their effect
on WM span tasks.
Performance on WM span tasks correlates with a wide
range of higher order cognitive tasks, such as reading and
listening comprehension (Daneman & Carpenter, 1983;
Daneman & Merikle, 1996), language comprehension
(King & Just, 1991), following oral and spatial directions
(Engle, Carullo, & Collins, 1991), vocabulary learning
from context (Daneman & Green, 1986), note taking in
class (Kiewra & Benton, 1988), writing (Benton, Kraft,
Glover, & Plake, 1984), reasoning (Barrouillet, 1996; Kyl-
lonen & Christal, 1990), hypothesis generation (Dough-
erty & Hunter, 2003), bridge playing (Clarkson-Smith &
Hartley, 1990), and complex-task learning (Kyllonen &
Stephens, 1990).
WM span measures predict performance on lower level
attention and perception tasks as well. For instance, in
comparing individuals who score in the upper and lower
quartiles on such tasks, lower quartile individuals (1) have
difficulty resisting the attention capture of an exogenous
cue in the antisaccade task (Kane et al., 2001; Unsworth,
Schrock, & Engle, 2004), (2) have difficulty constraining
their attention to discontiguous regions of space (Bleck-
ley, Durso, Crutchfield, Engle, & Khanna, 2004), (3) are
slower to constrain their focus of attention in a flanker
task with incompatible distractors (Heitz & Engle, 2004),
(4) make many more errors in a Stroop task (Kane &
Engle, 2003), and (5) are more vulnerable to proactive
interference (Kane & Engle, 2000). All of these find-
ings point to the idea that the central construct measured
by WM span tasks is the ability to control attention and
thought.
Whereas the literature described above demonstrates
the importance of the construct measured by WM span
tasks for what might be thought of as cold cognition, more
recent studies point to the importance of this construct
for hot cognition as well. Barrett et al. (2004) reviewed a
large literature connecting individual differences on WM
span measures with dual-process theories of mind and

778 CONWAY ET AL.
the ways in which those differences should be manifest in
studies of emotional control and social cognition. Brewin
and Beaton (2002), thinking of individual differences in
WMC as a causal factor, showed that high WM span in-
dividuals were better at suppressing intrusive thoughts
than were low WM span individuals. However, reductions
in WMC can also be thought of as a result of events that
control attention and thought. For instance, Klein and
Boals (2001) found that individuals who reported more
life event stress scored lower on operation span than did
low life event stress individuals. The interpretation was
that stressful events captured attentional resources, which
reduced ability to perform the WM span task.
Perhaps the most elaborate example in which a span task
has been used to measure the effect of a hot cognition vari-
able on WM resources is an elegant article by Schmader
and Johns (2003) on stereotype threat. Such a threat oc-
curs when a relevant stereotype or social stigma is primed
in the context of a performance situation and leads to a
reduction in performance. Schmader and Johns reasoned
that stereotype threat might have its effect on subsequent
performance through reduction in available WMC. They
had women complete both a WM span task and a stan-
dardized math test under stereotype threat or nonthreat
conditions. Women in the stereotype threat condition did
worse on the WM span task and on the standardized math
test. More important, a mediation analysis supported the
contention that reduction in WMC was responsible for the
decreased performance on the math test.
As this mediation analysis suggests, WMC is some-
times viewed as a cause and sometimes as an effect. Early
investigations of immediate memory performance almost
always treated memory capacity as an outcome variable—
that is, the dependent measure in an experimental design.
In contrast, more recent work, particularly individual-
differences studies, has tended to treat WMC as a more
stable trait and to use scores on WM span tasks as predic-
tors of some other outcome measure (e.g., using WMC to
predict intelligence). Until the cognitive and biological
mechanisms underlying performance of WM tasks are
better understood, we argue that treating WMC as either a
cause or an effect is warranted as long as one recognizes
the limitations of any one investigation (e.g., recogniz-
ing the inherent limitations of correlational data). That
said, we do view WM span scores as reflecting both stable
interindividual variation and more state-dependent intra-
individual variation. Although some researchers might
take issue with the former argument about stability, there
is direct evidence of stability from the strong test–retest
reliability data discussed earlier, as well as indirect evi-
dence of stability, which comes from the fact that WM
span scores are strongly related to general fluid intelli-
gence, which itself is relatively stable across the lifespan
(Conway et al., 2003).
In summary, measures of WMC, such as counting, op-
eration, and reading span, show considerable construct
validity insofar as they predict performance on a wide
array of tasks for which control of attention and thought
are important. Importantly, construct validity implies not
only convergent, but also discriminant validity. We have
documented above the convergent validity of WM span
tasks, such that they correlate extremely well with each
other and correlate well with performance on tests of
more complex cognition that purportedly depend upon
WM. In addition, WM span tasks reveal discriminant va-
lidity in that they do not predict performance on tasks that
appear to reflect relatively automatic processing, such as
the prosaccade condition in the antisaccade task (Kane
et al., 2001) or recall/recognition in the absence of inter-
ference (Cantor & Engle, 1993; Conway & Engle, 1994;
Kane & Engle, 2000; Rosen & Engle, 1997, 1998). Fi-
nally, WM span tasks also diverge from more traditional
“simple” span tasks in their predictive validity—that is,
in their ability to successfully predict complex cognition.
It is to this distinction between WM span tasks and other
tests of immediate memory that we will now turn.
WM SPAN TASKS VERSUS
OTHER WM TASKS
Ample empirical work has demonstrated the impor-
tance of the processing demand of WM span tasks by
contrasting their predictive utility with STM span tasks,
which present only to-be-recalled items and no additional
processing task. In short, this research shows that WM
span tasks tend to be stronger predictors of general intel-
lectual ability than are STM span tasks, and STM span
tasks account for no unique variance in general cognitive
ability after variance related to WMC is accounted for
(Conway et al., 2002; Engle, Tuholski, et al., 1999). Sub-
jects in these studies completed multiple tests of verbal
STM, all using word stimuli, as well as reading span, op-
eration span, and counting span tasks. Confirmatory fac-
tor analysis and structural equation modeling techniques
demonstrated that the variance shared among the STM
tasks was closely related to the variance shared among the
WM tasks; that is, the constructs of STM and WMC were
correlated. However, the correlations were not strong
enough to suggest that STM and WM tasks measure the
same construct, and moreover, only the WMC construct
shared unique variance with standardized tests of Gf.
These studies will be discussed in more detail below (see
the Latent Variable Analysis section), but for now they
indicate that in order to most effectively measure WMC, a
task must include a demanding secondary task to compete
with information storage.
Questions remain open, however, regarding the re-
quired structure of a WMC task. Must the processing task
present stimuli additional to those in the storage task, or
can additional processing be required on the target items
themselves? Must the presentation of target memory items
alternate regularly with the presentation of the secondary
processing task, or can the entire set of memory items be
presented together, prior to a processing task? As well,
the specific research goals of the investigator might influ-
ence the choice of WMC task. For example, to what extent

WORKING MEMORY SPAN TASKS 779
is it important to the question at hand to tap primarily
domain-specific storage processes versus domain-general
executive attention?
There is mixed evidence regarding the first question, of
whether the secondary task must present actual stimuli to
be processed, rather than simply requiring some mental
transformation of the target memory items. Engle, Tuhol-
ski, et al. (1999) tested subjects on the backward word
span task, in which target words were recalled in the re-
verse order from that in which they had been presented,
in addition to traditional WM and STM span tasks. Note
that the backward span’s “processing” requirement was
only a mental transformation, and not consideration of
new stimuli, to interfere with the primary storage task.
Factor analyses showed that backward word span grouped
itself with the STM tasks, rather than with the WM span
tasks, indicating that a mental transformation alone is not
enough to turn an immediate-memory task into a WMC
task (see also Hutton & Towse, 2001). In contrast, Ober-
auer et al. (2000) found that simple transformation span
tasks seemed to measure the same construct as did WM
span. They tested subjects in a backward digit span task
and an alpha span task (Craik, 1986), in addition to read-
ing span and counting span tasks. Alpha span required
recall of target words in alphabetical order, rather than
in their presentation order, and so, like backward span,
presented a secondary processing task without second-
ary stimuli. Here, the correlation between reading and
counting span (r  .66), was only slightly stronger than
their correlations with the transformation tasks (mean
r  .60), suggesting that all of these span tasks reflect a
single construct. The source of the discrepancy between
the Engle and the Oberauer findings is not obvious, so
further research must determine the importance of inter-
fering stimuli, in addition to interfering processing, to the
measurement of WMC.
With respect to the second open question about the
structure of WM tasks, the limited research regarding the
regular interleaving of memory and processing stimuli in
WMC tasks does not suggest it to be a critical variable.
For example, Kane and Engle (2000) found that subjects
identified as having high or low WMC (via a quartile split
on the operation span task) performed quite differently
on a Brown–Peterson-like task as proactive interference
built throughout the task. The subjects recalled three con-
secutive lists of 10 words each, with all words drawn from
the same category (e.g., animals) and with recall of each
list preceded by a demanding processing task. Here, then,
the secondary task followed, rather than alternated with,
the target memory items. High- and low-span subjects re-
called equivalent numbers of items on the first list, but
low-span subjects recalled fewer items than did high-span
subjects on subsequent lists. Thus, Brown–Peterson tasks
may tap the WMC construct after several stimulus lists
are presented and recalled, allowing proactive interfer-
ence to challenge retrieval. Indeed, Oberauer et al. (2000)
and Oberauer, Süß, Wilhelm, and Wittmann (2003) tested
subjects in Brown–Peterson-like tasks, using 15–30 lists
of target digits or words, and found that they correlated
with WM span measures with mean rs of .59. In fact, their
WM span, transformation span, and Brown–Peterson tasks
(backward digit and alpha span) all loaded onto a single
factor: a unitary WMC construct.
Although WM span, transformation span, and Brown–
Peterson tasks are structurally heterogeneous in subtle
ways, they all present subjects with lists of two to seven
target items (often in a predictable sequence of list
lengths) and require recall of each list, in turn, following
some interfering task.
5
It may not be surprising, then, that
they appear to measure the same underlying construct.
However, other candidate WMC tasks in the literature,
such as running span, keeping-track, and n-back tasks,
present quite different cognitive demands. Specifically,
these more dynamic tasks of immediate memory require
subjects to monitor a continuous stream of stimuli, often
of uncertain length, and to respond according to only a
subset of the stimuli presented. The subjects in these tasks
must, therefore, continuously update their mental repre-
sentation of the target items while also dropping now-
irrelevant items from consideration. So, like WM span
tasks, some demanding processes are required in addition
to storage.
More specifically, the running-memory span task
(N. S. Anderson, 1960; Pollack, Johnson, & Knaff, 1959;
Waugh, 1960) presents stimuli in lists of unknown length,
and subjects must recall only the last n items (the pre-
specified, variable memory load). Thus, the subjects re-
tain only the most recent n items that are presented and
continuously drop items from the maintenance/rehearsal
set once the list length exceeds n. Similarly, the keep-
ing-track task (Yntema & Mueser, 1960, 1962) presents
a list of items, of unknown length and from n categories
(the memory load), and subjects retain only the most re-
cent exemplar of each category. Finally, the n-back task
(Kirchner, 1958; Mackworth, 1959; Moore & Ross, 1963)
presents a list of items in which the subject must continu-
ously report whether each item matches the one that had
appeared n items ago in the stream (n typically ranges
from 1 to 4). In a two-back task, for example, subjects
must continuously maintain the last 2 items in the list, up-
dating this memory set with each new item and dropping
out the least recent one.
Unfortunately, very little research has contrasted these
dynamic WM tasks with other WMC or STMC tasks. We
know of no relevant studies on running span, but the very
similar keeping-track task does appear to be a valid index
of WMC. Engle, Tuholski, et al. (1999) used exploratory
and confirmatory factor analyses to test whether a num-
ber of different immediate-memory tasks tapped either
the STMC or the WMC. The keeping-track task had rea-
sonably high loadings on the WMC factor (consisting of
WM span tasks), low loadings on the STMC factor, and
correlations with fluid intelligence scores of similar mag-
nitude to those in the WM span task. Similarly, Oberauer
and colleagues (Oberauer et al., 2000; Oberauer et al.,
2003) found strong correlations among WM span tasks,

780 CONWAY ET AL.
Brown–Peterson tasks, and versions of the keeping-track
task developed by Salthouse (Salthouse, 1995; Salthouse,
Babcock, & Shaw, 1991), in which different numbers of
screen locations, rather than taxonomic categories, are
monitored for the most recent items presented there.
Thus, the keeping-track task and, perhaps by analogy, the
running-memory span task currently appear to be valid
measures of WMC, along with WM span and Brown–
Peterson tasks.
6
Although the n-back task is arguably the
current gold standard measure of WMC in the cognitive
neuroscience literature (for a review, see Kane & Engle,
2002), almost no behavioral research has been conducted
to validate it. The only study that has compared n-back
with other immediate-memory tasks (Dobbs & Rule,
1989) found the two-back task to correlate more strongly
with simple digit span than with a Brown–Peterson task
(rs  .27 and .14, respectively). Given that the two-back
task correlated only modestly with a one-back task (r 
.38), its correlation with digit span may be considered
relatively high. Thus, the n-back task may be a more ap-
propriate indicator of the construct measured by STMC,
rather than by WMC tasks, but more research is obviously
needed.
Finally, we should note that the particular research
goals of the investigator typically influence which tasks
are used to measure WMC. For instance, some research-
ers are more interested in testing hypotheses about the
mechanisms underlying storage and, as a result, may pay
little attention to the attentional aspects of WM tasks,
whereas others are more interested in testing hypotheses
about the mechanisms underlying executive attention and
cognitive control and, as a result, may pay little attention
to the storage demands of the task. Although this state
of affairs is understandable, it is also quite problematic,
because tasks as diverse as operation span and n-back are
referred to as WM tasks in the literature, as if they come
from the same class, yet very little data exists to assess the
extent to which they tap similar constructs.
LATENT VARIABLE ANALYSIS
All of the above-mentioned measures of WMC, includ-
ing operation span, reading span, and counting span, suf-
fer from the fact that no single task is a perfect measure
of the construct it ostensibly represents. For example, the
operation span task measures WMC but, most likely, also
taps mathematical ability, motivation, and word knowl-
edge, among other factors. Similarly, the reading span task
measures WMC but, certainly, also verbal ability. Also, as
was noted in the section on reliability above, despite the
strong reliability of WM span tasks, they are not perfectly
reliable. Thus, despite being valid and strongly reliable
measures of WMC, WM span tasks are not perfect or pro-
cess pure. Given such imperfection, an optimal research
strategy is to administer multiple WM span tasks and then
use the average (or weighted average) of scores on all the
tasks as the measure of WMC (note that this is the same
logic frequently applied by experimental psychologists;
rather than measuring performance with one trial, they as-
sess performance with multiple trials, and then a measure
of central tendency is taken).
Latent variable analysis is a statistical approach in
which multiple measures of a construct are administered
and then a latent variable is derived from the common
variance among those measures. For example, in our own
work, we have derived a latent variable, which we label
working memory capacity, from the common variance
among counting span, operation span, and reading span
(Conway et al., 2002; Engle, Tuholski, et al., 1999; Kane
et al., 2004). Conceptually, the latent variable represents
only the variance that is common among the three tasks
and removes task-specific factors. Statistically, the latent
variable is a predictor of the manifest (task) variables, and
the strength of each predictor is a function of the correla-
tions among the manifest tasks and the individual reli-
abilities of the tasks. Ideally, a set of manifest variables
predicted by the same latent variable will have good reli-
ability and be moderately to strongly correlated.
Counting span, operation span, and reading span, as a
group, are particularly suited for latent variable analysis,
because they are all moderately correlated with one an-
other, suggesting that they are indeed tapping a similar
construct, yet are not mere replicas of one another. From
a measurement standpoint, this is an ideal situation; when
a construct is measured with imperfect tools, it is best to
use multiple, reliable measures that do not replicate one
another. Correlations among the three span tasks typically
range from .40 to .60, suggesting that they are indeed tap-
ping some common process or ability but also suggesting
that they are not identical (it is also worth noting here that
the correlations are not diminished when spatial WM span
tasks are considered; see Kane et al., 2004).
The main benefit of latent variable analysis is that a
more “pure” measure of WMC can be derived from three
span tasks than from one. As such, the predictive power
of latent variables is better than that of individual mani-
fest variables. To demonstrate this point, we analyzed the
correlations between individual span tasks and Raven’s
advanced progressive matrices (a prototypical measure
of general fluid intelligence) and contrasted these corre-
lations with those between latent variables derived from
the span tasks and Raven’s matrices. The results of this
analysis are reported in Table 4. The correlations between
the latent variables and Raven’s matrices are considerably
higher than the correlations between the manifest vari-
ables and Raven’s matrices. Also note that the correlations
from the latent variable analyses are much more stable
across studies than are the correlations from the manifest
variables.
Another benefit of having multiple measures per con-
struct is that multivariate outliers can be detected and the
impact of such outliers can be controlled. For example,
suppose that a subject is particularly anxious about math
and, therefore, performs very poorly on the operation span
task, despite the fact that he or she may “truly” have an
above-average WMC. In this scenario, the subject might
score well above average on counting span and reading
span but well below average on operation span. Given that

WORKING MEMORY SPAN TASKS 781
all three tasks tend to be correlated, this subject’s multi-
variate profile would be flagged as an outlier. Multivariate
outliers can be detected by calculating the Mahanalobis
distance (or d
2
) for each case in a data set. The d
2
repre-
sents the distance a case is from the centroid of a multi-
variate distribution. For example, Figure 1 represents the
three-way relationship between counting span, operation
span, and reading span from Kane et al. (2004). Each line
in the figure represents the Mahanalobis distance for one
case. Furthermore, assuming multivariate normality, one
can calculate the probability with which a case with such
a distance would be observed. Outliers can be detected
and deleted on the basis of this information.
A final benefit of latent variable analysis (and factor
analysis, for that matter) is that latent variable scores (i.e.,
factor scores) can be calculated for each subject and then
used as a dependent measure in experimental designs. For
example, in Kane et al.’s (2004) study, some subjects were
college students, whereas others were community dwell-
ers. If one were interested in group comparisons, factor
scores could be created for each individual in the data set,
and these scores could be entered into an analysis as the
dependent measure.
In sum, latent variable analysis is a powerful research
strategy that has proven particularly beneficial in exam-
ining individual differences in WMC. Of course, these
methods are expensive and time consuming. Multiple
measures of each construct are needed, and large samples
are required. When tasks are chosen for a latent variable
analysis, it is important to select tasks that are reliable
and also reveal moderate to strong correlations with one
another. The overall correlation matrix for the study also
should indicate both convergent and discriminant validity,
meaning that measures that purportedly measure the same
construct should converge on one another and diverge
from the rest of the pack. In determining sample size,
there are many factors to consider, such as the number of
tasks used, the quality of each univariate distribution, and
the number of parameters being estimated in the latent
variable model. Although there are no hard and fast rules
for determining sample size for a latent variable design
or for structural equation modeling, it has been suggested
Table 4
Correlations Between Multiple Measures of Working Memory Capacity
and Raven’s Advanced Progressive Matrices
Task
Engle, Tuholski, et al. (1999)
(N  133)
Conway et al. (2002)
(N  120)
Kane et al. (2004)
(N  236)
Counting span .32 .38 .25
Operation span .34 .20 .32
Reading span .28 .15 .30
Latent variable .44 .40 .37
Note—Kane et al. (2004) used only the odd-numbered items from the Raven’s and imple-
mented a strict time limit, which is not customary. These modifications to Raven’s might
account for the slightly lower correlations in that study.
Figure 1. The three-way relationship between counting span (CSPAN),
operation span (OSPAN), and reading span (RSPAN) from Kane et al.
(2004), N ⴝ 236. Each line represents Mahanalobis distance for each
case in the data set. The x-, y-, and z-axes represent scores on OSPAN,
RSPAN, and CSPAN, respectively.

782 CONWAY ET AL.
that in any situation more than 100 subjects are necessary
and 10 times the number of manifest variables is also de-
sired (Kline, 1998).
EXTREME-GROUPS DESIGNS
Extreme-groups designs refer to situations in which
a continuous variable is categorized and only categories
representing the upper and lower ends of the distribution
of the continuous variable are represented. In the case of
WM span tasks, the most common extreme-groups de-
sign is one in which the upper and lower quartiles of a
distribution of WM span scores are categorized as high
and low span, respectively. The process of categorizing
a continuous variable is considered problematic among
most statisticians for obvious reasons. First, information
and power are lost, because less variability is captured by
categories than by a continuum. Second, subjects who are
not equal on some ability or trait are treated as if they are
equal. Third, subjects can easily be misclassified, due to
measurement error. We will not review these issues here.
A more in-depth discussion of the problems associated
with categorization can be found in Cohen (1983).
Despite these problems, extreme-groups designs are
common in the WM literature (e.g., the present authors
have collectively published over a dozen experiments with
extreme-groups designs). In this section, we will justify
the use of extreme-groups designs but will caution re-
searchers to use them only in certain situations and to be
aware of the interpretive problems that they can create.
First, it is necessary to explicitly state that, according
to all theoretical accounts of WM that we are familiar
with, WMC is assumed to be normally distributed in a
population of healthy subjects and, therefore, should be
measured with an instrument that can produce a continu-
ous normal distribution. In terms of creating a normal
distribution, WM span tasks are successful, particularly
if they are scored with the partial-credit unit-weighting
method, as was discussed above. For example, using this
scoring procedure, Kane et al. (2004) observed normal
distributions (as indicated by standard deviation, skew,
and kurtosis) for each counting span, operation span, and
reading span. The Kane et al. (2004) distributions will
serve as a good reference point for other researchers who
would like to use these tasks, because they were created
using our most recent versions of the span tasks (which
are available on our Web site) and because the sample
was quite large (N  234) and represents both college
student and nonstudent populations. Thus, if a researcher
is concerned that he or she is suffering from a restriction
of range in WMC (e.g., if the research is being conducted
with college students from an elite institution), he or she
could check the distribution against the ones observed in
Kane et al. (2004).
As was mentioned above, the ideal research approach is
to sample the entire range of WMC. This is especially true
if the goal of the research is to estimate, in the popula-
tion, the magnitude of the relationship between WMC and
performance of some other task (or in some experimental
context). However, if the goal is simply to test whether
a relationship exists or does not exist, a more efficient
approach is to compare extreme groups. As Underwood
(1975) has suggested, most nomothetic psychological
theories make at least tacit predictions about individual
differences; thus, confirmed predictions regarding indi-
vidual differences may give a theory a “go-ahead signal,”
whereas failed predictions may refute the theory, in which
case “there is no alternative but to drop the line of theoret-
ical thinking” (p. 130). Hence, Underwood famously con-
cluded that “individual differences may indeed be used as
a crucible in nomothetic theory construction.” Critical to
our argument is that it is often the mere presence or ab-
sence of a relationship that is in question, rather than the
magnitude of the relationship.
Of course, one hazard of extreme-groups designs is
that the observed span  treatment interaction might be
a Type 1 error due to the sampling of extreme groups or
might be an overestimation of the true relationship between
WM span and the treatment. In order to illustrate this point,
we simulated experiments in which the entire distribution
of span was used and calculated the effect size (R
2
) for a
simulated span effect. We then simulated the same effect,
using an extreme-groups design, in which the upper and
lower quartiles were used as groups and an ANOVA was
conducted. Figure 2 demonstrates the extent to which the
extreme-groups design overestimates the “true” effect size
(i.e., the effect size that would have been observed had
the experiment been run on the continuum). As the figure
illustrates, extreme-groups designs tend to modestly over-
estimate effect size, particularly for moderate effects.
Figure 2. Simulated effect sizes (R
2
) for situations in which the
entire distribution of span is used (continuous; x-axis), as com-
pared with situations in which extreme groups are used ( y-axis).

WORKING MEMORY SPAN TASKS 783
Thus, extreme-groups designs can be cost efficient, but
they can pose interpretive problems. We therefore caution
researchers to use extreme-groups designs sparingly and
to be aware of the problems documented here. We also
advocate replicating effects that are observed in extreme-
groups designs, as well as the use of converging methods
to provide support for the theoretical conclusions derived
from extreme-groups design experiments.
An additional positive aspect of extreme-groups de-
signs is worth mentioning here. When the upper and
lower quartiles of a distribution of WM span scores are
classified as high and low span, respectively, subjects
are rarely misclassified as high when they should have
been classified as low, and vice versa. To demonstrate this
point, we reanalyzed the data from Conway et al. (2002).
Table 5 illustrates how many subjects would have been
classified as high span on one span measure and as low
span on another. As the table illustrates, a very small per-
centage of subjects (approximately 8%) would have been
misclassified had a quartile design been used with these
distributions. To further illustrate this point, we examined
the consistency of quartiles across the three span tasks in
Conway et al. (2002). In this analysis, we first created four
quartiles on the basis of a z score composite representing
the average of counting, operation, and reading span. We
then examined how many subjects were classified in the
correct quartile (i.e., the same quartile they were classi-
fied by with the composite score) when only one span
task was considered and then when two span tasks were
considered. The results of this analysis are presented in
Table 6. This analysis suggests that quartile efficiency
(i.e., the extent to which a subject is classified in the cor-
rect quartile) is significantly better when two span tasks
are considered than when just one span task is considered.
On the basis of this analysis, we recommend using at least
two span tasks to assess WMC, whenever possible.
Finally, we would like to note here that median split de-
signs are not acceptable, for two simple reasons: (1) There
is no reason to categorize subjects when the entire con-
tinuum has been sampled, and (2) misclassification of
subjects is more likely in a median split design than in an
extreme-groups design. Recall that in the Conway et al.
(2002) data, as illustrated in Table 5, only 15 out of a
possible 180 classifications were mismatches (8% of the
cases) when classifications were based on quartile splits.
A parallel analysis examining the consistency of median
split classifications with the same data reveals that 25%
of the cases were mismatches.
CONCLUSION
WM is a central construct in cognitive psychology. Fur-
thermore, WMC is an important individual-differences
variable in differential approaches to understanding human
behavior. WM span tasks, such as counting span, opera-
tion span, and reading span, are reliable and valid mea-
sures of WMC. They have proven to be extremely useful
research tools in cognitive psychology and, more recently,
in other branches of psychology. In an attempt to maxi-
mize the future utility of these tasks, we have documented
here all the relevant information we have gathered in our
15 years experience with the tasks.
Our hope is that the review above has demonstrated
the reliability and validity of WM span tasks. As well,
Table 5
Number of Subjects Classified as High Span on One Span Task
and as Low Span on Another Span Task
High Span on:
Low Span on: Counting Span Operation Span Reading Span
Counting span 0 4 3
Operation span 2 0 2
Reading span 1 3 0
Note—Data are from Conway, Cowan, Bunting, Therriault, and Minkoff (2002).
N  120 (therefore, the maximum number of misclassifications per cell is 30).
Table 6
Consistency of Quartiles When One or Two Working Memory (WM) Span Tasks Are Used
to Assess WM Capacity, Relative to a Standard in Which Three WM Span Tasks Are Used
Quartile z Score Composite CSPAN OSPAN RSPAN
CSPAN &
OSPAN
CSPAN &
RSPAN
OSPAN &
RSPAN
1 30 19 22 21 24 24 27
2 30 7 16 13 20 20 20
3 30 15 18 11 21 23 18
4 30 21 19 21 26 27 25
Efficiency for all quartiles .52 .63 .55 .76 .78 .75
Efficiency for upper and lower quartiles .67 .68 .70 .83 .85 .87
Note—Each cell represents the number of subjects classified in the correct quartile (i.e., the same quartile in which they
were classified by the z score composite). Efficiency refers to the proportion of subjects classified correctly. CSPAN, count-
ing span; OSPAN, operation span; RSPAN, reading span.

784 CONWAY ET AL.
we hope that other researchers will now be able to use
these tasks to their full potential, by adopting the optimal
administration and scoring procedures. We also encour-
age other researchers to conduct latent variable analyses
whenever possible. We caution researchers to use extreme-
groups designs sparingly and with great care. Finally, we
do not suggest that these WM span tasks are, or should be,
the gold standard measures of WMC. Instead, we hope
that our explicit documentation of WM span tasks, their
genesis, and their development will inspire new task de-
velopment and analysis, which hopefully will result in
even better measurement of WMC and cognitive abilities
in the future.
REFERENCES
Anderson, J. R., & Lebiere, C. (1998). The atomic components of
thought. Mahwah, NJ: Erlbaum.
Anderson, N. S. (1960). Poststimulus cuing in immediate memory.
Journal of Experimental Psychology, 60, 216-221.
Arnett, P. A., Higginson, C. H., Voss, W. D., Bender, W. I., Wurst,
J. M., & Tippin, J. M. (1999). Depression in multiple sclerosis: Re-
lationship to working memory capacity. Neuropsychology, 13, 546-
556.
Baddeley, A. D. (1986). Working memory. Oxford: Oxford University
Press.
Baddeley, A. D., & Hitch, G. (1974). Working memory. In G. A.
Bower (Ed.), The psychology of learning and motivation (Vol. 8,
pp. 47-89). New York: Academic Press.
Barrett, L. F., Tugade, M. M., & Engle, R. W. (2004). Individual
differences in working memory capacity and dual-process theories of
the mind. Psychological Bulletin, 130, 553-573.
Barrouillet, P. (1996). Transitive inferences from set-inclusion re-
lations and working memory. Journal of Experimental Psychology:
Learning, Memory, & Cognition, 22, 1408-1422.
Benton, S. L., Kraft, R. G., Glover, J. A., & Plake, B. S. (1984).
Cognitive capacity differences among writers. Journal of Educational
Psychology, 76, 820-834.
Bleckley, M. K., Durso, F. T., Crutchfield, J. M., Engle, R. W., &
Khanna, M. M. (2004). Individual differences in working memory
capacity predict visual attention allocation. Psychonomic Bulletin &
Review, 10, 884-889.
Brewin, C. R., & Beaton, A. (2002). Thought suppression, intelli-
gence, and working memory capacity. Behaviour Research & Ther-
apy, 40, 923-930.
Bunting, M. F. (in press). Proactive interference and item similarity
in working memory. Journal of Experimental Psychology: Learning,
Memory, & Cognition.
Cantor, J., & Engle, R. W. (1993). Working-memory capacity as long-
term memory activation: An individual-differences approach. Journal
of Experimental Psychology: Learning, Memory, & Cognition, 19,
1101-1114.
Case, R., Kurland, M. D., & Goldberg, J. (1982). Operational ef-
ficiency and the growth of short-term memory span. Journal of Ex-
perimental Child Psychology, 33, 386-404.
Clarkson-Smith, L., & Hartley, A. A. (1990). The game of bridge as
an exercise in working memory and reasoning. Journal of Gerontol-
ogy, 45, P233-P238.
Cohen, J. (1983). The cost of dichotomization. Applied Psychological
Measurement, 3, 249-253.
Conway, A. R. A., Cowan, N., Bunting, M. F., Therriault, D., &
Minkoff, S. (2002). A latent variable analysis of working memory
capacity, short term memory capacity, processing speed, and general
fluid intelligence. Intelligence, 30, 163-183.
Conway, A. R. A., & Engle, R. W. (1994). Working memory and re-
trieval: A resource-dependent inhibition model. Journal of Experi-
mental Psychology: General, 123, 354-373.
Conway, A. R. A., & Engle, R. W. (1996). Individual differences in
working memory capacity: More evidence for a general capacity
theory. Memory, 4, 577-590.
Conway, A. R. A., & Kane, M. J. (2001). Capacity, control and con-
flict: An individual differences perspective on attentional capture. In
C. Folk & B. Gibson (Eds.), Attraction, distraction and action: Mul-
tiple perspectives on attention capture (pp. 349-372). Amsterdam:
Elsevier.
Conway, A. R. A., Kane, M. J., & Engle, R. W. (2003). Working
memory capacity and its relation to general intelligence. Trends in
Cognitive Sciences, 7, 547-552.
Copeland, D. E., & Radvansky, G. A. (2001). Phonological similarity
in working memory. Memory & Cognition, 29, 774-776.
Cowan, N. (1995). Attention and memory: An integrated framework.
Oxford: Oxford University Press.
Cowan, N., Towse, J. N., Hamilton, Z., Saults, J. S., Elliott, E. M.,
Lacey, J. F., et al. (2003). Children’s working-memory processes:
A response-timing analysis. Journal of Experimental Psychology:
General, 132, 113-132.
Craik, F. I. M. (1986). A functional account of age differences in mem-
ory. In F. Klix & H. Hagendorf (Eds.), Human memory and cognitive
capabilities (pp. 409-422). Amsterdam: North-Holland.
Daneman, M., & Carpenter, P. A. (1980). Individual differences in
working memory and reading. Journal of Verbal Learning & Verbal
Behavior, 19, 450-466.
Daneman, M., & Carpenter, P. A. (1983). Individual differences in
integrating information between and within sentences. Journal of
Experimental Psychology: Learning, Memory, & Cognition, 9, 561-
584.
Daneman, M., & Green, I. (1986). Individual differences in compre-
hending and producing words in context. Journal of Memory & Lan-
guage, 25, 1-18.
Daneman, M., & Merikle, P. M. (1996). Working memory and lan-
guage comprehension: A meta-analysis. Psychonomic Bulletin &
Review, 3, 422-433.
Dobbs, A. R., & Rule, B. G. (1989). Adult age differences in working
memory. Psychology & Aging, 4, 500-503.
Dougherty, M. R. P., & Hunter, J. (2003). Probability judgment and
subadditivity: The role of working memory capacity and constraining
retrieval. Memory & Cognition, 31, 968-982.
Ebbinghaus, H. (1897). Über eine neue Methode zur Prüfung geistiger
Fähigkeiten und ihre Anwendung bei Schulkindern [On a new method
of testing mental abilities and its application with school children].
Zeitschrift für die Psychologie, 13, 401-459.
Engle, R. W. (2001). What is working memory capacity? In H. L. Roed-
iger, III, J. S. Nairne, I. Neath, & A. M. Surprenant (Eds.), The nature
of remembering: Essays in honor of Robert G. Crowder (pp. 297-
314). Washington, DC: American Psychological Association Press.
Engle, R. W. (2002). Working memory capacity as executive attention.
Current Directions in Psychological Science, 11, 19-23.
Engle, R. W., Cantor, J., & Carullo, J. J. (1992). Individual dif-
ferences in working memory and comprehension: A test of four hy-
potheses. Journal of Experimental Psychology: Learning, Memory, &
Cognition, 18, 972-992.
Engle, R. W., Carullo, J. J., & Collins, K. W. (1991). Individual dif-
ferences in working memory for comprehension and following direc-
tions. Journal of Educational Research, 84, 253-262.
Engle, R. W., Kane, M. J., & Tuholski, S. W. (1999). Individual dif-
ferences in working memory capacity and what they tell us about
controlled attention, general fluid intelligence and functions of the
prefrontal cortex. In A. Miyake & P. Shah (Eds.), Models of working
memory: Mechanisms of active maintenance and executive control
(pp. 102-134). New York: Cambridge University Press.
Engle, R. W., Tuholski, S. W., Laughlin, J. E., & Conway, A. R. A.
(1999). Working memory, short-term memory and general fluid intel-
ligence: A latent variable approach. Journal of Experimental Psychol-
ogy: General, 128, 309-331.
Finn, P. R. (2002). Motivation, working memory, and decision making:
A cognitive–motivational theory of personality vulnerability to alco-
holism. Behavioral & Cognitive Neuroscience Reviews, 1, 183-205.

WORKING MEMORY SPAN TASKS 785
Friedman, N. P., & Miyake, A. (2004). The reading span test and its
predictive power for reading comprehension ability. Journal of Mem-
ory & Language, 51, 136-158.
Hambrick, D. Z., & Engle, R. W. (2002). Effects of domain knowl-
edge, working memory capacity, and age on cognitive performance:
An investigation of the knowledge-is-power hypothesis. Cognitive
Psychology, 44, 339-384.
Hasher, L., & Zacks, R. T. (1988). Working memory, comprehen-
sion, and aging: A review and a new view. In G. H. Bower (Ed.), The
psychology of learning and motivation (Vol. 22, pp. 193-225). San
Diego: Academic Press.
Heitz, R. P., & Engle, R. W. (2004). Focusing the spotlight: Individ-
ual differences in visual attention control. Manuscript submitted for
publication.
Hitch, G. J., Towse, J. N., & Hutton, U. (2001). What limits children’s
working memory span? Theoretical accounts and applications for
scholastic development. Journal of Experimental Psychology: Gen-
eral, 130, 184-198.
Hutton, U. M. Z., & Towse, J. N. (2001). Short-term memory and
working memory as indices of children’s cognitive skills. Memory,
9, 383-394.
Kane, M. J., Bleckley, M. K., Conway, A. R. A., & Engle, R. W.
(2001). A controlled-attention view of working-memory capacity.
Journal of Experimental Psychology: General, 130, 169-183.
Kane, M. J., & Engle, R. W. (2000). Working memory capacity, proac-
tive interference, and divided attention: Limits on long-term memory
retrieval. Journal of Experimental Psychology: Learning, Memory, &
Cognition, 26, 333-358.
Kane, M. J., & Engle, R. W. (2002). The role of prefrontal cortex in
working memory capacity, executive attention, and general fluid intel-
ligence: An individual-differences perspective. Psychonomic Bulletin
& Review, 9, 637-671.
Kane, M. J., & Engle, R. W. (2003). Working-memory capacity and
the control of attention: The contributions of goal neglect, response
competition, and task set to Stroop interference. Journal of Experi-
mental Psychology: General, 132, 47-70.
Kane, M. J., Hambrick, D. Z., Tuholski, S. W., Wilhelm, O., Payne,
T. W., & Engle, R. W. (2004). The generality of working memory ca-
pacity: A latent-variable approach to verbal and visuo-spatial memory
span and reasoning. Journal of Experimental Psychology: General,
133, 189-217.
Kiewra, K. A., & Benton, S. L. (1988). The relationship between in-
formation processing ability and notetaking. Contemporary Educa-
tional Psychology, 13, 33-44.
King, J., & Just, M. A. (1991). Individual differences in syntactic pro-
cessing: The role of working memory. Journal of Memory & Lan-
guage, 30, 580-602.
Kirchner, W. K. (1958). Age differences in short-term retention of
rapidly changing information. Journal of Experimental Psychology,
55, 352-358.
Klein, K., & Boals, A. (2001). The relationship of life event stress
and working memory capacity. Applied Cognitive Psychology, 15,
565-579.
Klein, K., & Fiss, W. H. (1999). The reliability and stability of the
Turner and Engle working memory task. Behavior Research Methods,
Instruments, & Computers, 31, 429-432.
Kline, R. B. (1998). Principles and practice of structural equation
modeling. New York: Guilford.
Kyllonen, P. C. (1996). Is working memory capacity Spearman’s g?
In I. Dennis & P. Tapsfield (Eds.), Human abilities: Their nature and
measurement (pp. 49-75). Mahwah, NJ: Erlbaum.
Kyllonen, P. C., & Christal, R. E. (1990). Reasoning ability is (little
more than) working-memory capacity?! Intelligence, 14, 389-433.
Kyllonen, P. C., & Stephens, D. L. (1990). Cognitive abilities as de-
terminants of success in acquiring logic skill. Learning & Individual
Differences, 2, 129-160.
Lehto, J. (1996). Are executive function tests dependent on working
memory capacity? Quarterly Journal of Experimental Psychology,
49A, 29-50.
Lustig, C., May, C. P., & Hasher, L. (2001). Working memory span
and the role of proactive interference. Journal of Experimental Psy-
chology: General, 130, 199-207.
MacDonald, M. C., Almor, A., Henderson, V. W., Kempler, D., &
Andersen, E. S. (2001). Assessing working memory and language
comprehension in Alzheimer’s disease. Brain & Language, 78, 17-
42.
Mackworth, J. F. (1959). Paced memorizing in a continuous task.
Journal of Experimental Psychology, 58, 206-211.
May, C. P., Hasher, L., & Kane, M. J. (1999). The role of interference
in memory span. Memory & Cognition, 27, 759-767.
Miyake, A., Friedman, N. P., Rettinger, D. A., Shah, P., &
Hegarty, M. (2001). How are visuospatial working memory, execu-
tive functioning, and spatial abilities related? A latent-variable analy-
sis. Journal of Experimental Psychology: General, 130, 621-640.
Moore, M. E., & Ross, B. M. (1963). Context effects in running mem-
ory. Psychological Reports, 12, 451-465.
Morris, N., & Jones, D. M. (1990). Memory updating in working mem-
ory: The role of the central executive. British Journal of Psychology,
81, 111-121.
Munakata, Y., Morton, B., & O’Reilly, R. C. O. (in press). De-
velopmental and computational approaches to variation in working
memory. In A. R. A. Conway, C. Jarrold, M. J. Kane, A. Miyake, &
J. N. Towse (Eds.), Variation in working memory. Oxford: Oxford
University Press.
Nairne, J. S. (2002). Remembering over the short-term: The case against
the standard model. Annual Review of Psychology, 53, 53-81.
Oberauer, K., & Süß, H.-M. (2000).Working memory and interference:
A comment on Jenkins, Myerson, Hale, and Fry (1999). Psychonomic
Bulletin & Review, 7, 727-733.
Oberauer, K., Süß, H.-M., Schulze, R., Wilhelm, O., & Wittmann,
W. W. (2000). Working memory capacity: Facets of a cognitive ability
construct. Personality & Individual Differences, 29, 1017-1045.
Oberauer, K., Süß, H.-M., Wilhelm, O., & Wittmann, W. W. (2003).
The multiple faces of working memory: Storage, processing, supervi-
sion, and coordination. Intelligence, 31, 167-193.
Pollack, I., Johnson, L. B., & Knaff, P. R. (1959). Running memory
span. Journal of Experimental Psychology, 57, 137-146.
Richeson, J. A., Baird, A. A., Gordon, H. L., Heatherton, T. F.,
Wyland, C. L., Trawalter, S., et al. (2003). An fMRI investiga-
tion of the impact of interracial contact on executive function. Nature
Neuroscience, 6, 1323-1328.
Richeson, J. A., & Shelton, J. N. (2003). When prejudice does not
pay: Effects of interracial contact on executive function. Psychologi-
cal Science, 14, 287-291.
Rosen, V. M., Bergeson, J. L., Putnam, K., Harwell, A., & Sunder-
land, T. (2002). Working memory and apolipoprotein E: What’s the
connection? Neuropsychologia, 40, 2226-2233.
Rosen, V. M., & Engle, R. W. (1997). The role of working memory
capacity in retrieval. Journal of Experimental Psychology: General,
126, 211-227.
Rosen, V. M., & Engle, R. W. (1998). Working memory capacity and
suppression. Journal of Memory & Language, 39, 418-436.
Salthouse, T. A. (1995). Differential age-related influences on mem-
ory for verbal-symbolic and visual-spatial information? Journal of
Gerontology, 50B, P193-P201.
Salthouse, T. A., Babcock, R. L., & Shaw, R. J. (1991). Effects of
adult age on structural and operational capacities in working memory.
Psychology & Aging, 6, 118-127.
Schmader, T., & Johns, M. (2003). Converging evidence that stereo-
type threat reduces working memory capacity. Journal of Personality
& Social Psychology, 85, 440-452.
Shah, P., & Miyake, A. (1996). The separability of working memory
resources for spatial thinking and language processing: An individual
differences approach. Journal of Experimental Psychology: General,
125, 4-27.
Spearman, C. (1904). “General intelligence,” objectively determined
and measured. American Journal of Psychology, 15, 201-293.
Süß, H.-M., Oberauer, K., Wittmann, W. W., Wilhelm, O., &
Schulze, R. (2002). Working-memory capacity explains reasoning
ability—and a little bit more. Intelligence, 30, 261-288.

786 CONWAY ET AL.
Towse, J. N., & Hitch, G. J. (1995). Is there a relationship between
task demand and storage space in tests of working memory capacity?
Quarterly Journal of Experimental Psychology, 48A, 108-124.
Towse, J. N., Hitch, G. J., & Hutton, U. (1998). A reevaluation of
working memory capacity in children. Journal of Memory & Lan-
guage, 39, 195-217.
Towse, J. N., Hitch, G. J., & Hutton, U. (2002). On the nature of the
relationship between processing activity and item retention in chil-
dren. Journal of Experimental Child Psychology, 82, 156-184.
Treisman, A. M., & Gelade, G. (1980). A feature-integration theory of
attention. Cognitive Psychology, 12, 97-136.
Turley-Ames, K. J., & Whitfield, M. M. (2003). Strategy training and
working memory task performance. Journal of Memory & Language,
49, 446-468.
Turner, M. L., & Engle, R. W. (1989). Is working memory capacity
task dependent? Journal of Memory & Language, 28, 127-154.
Underwood, B. J. (1975). Individual differences as a crucible in theory
construction. American Psychologist, 30, 128-134.
Unsworth, N., Heitz, R. P., & Engle, R. W. (2005). Working mem-
ory capacity in hot and cold cognition. In R. W. Engle, G. Sedek,
U. Hecker, & D. N. McIntosh (Eds.), Cognitive limitations in aging
and psychopathology (pp. 19-43). New York: Oxford University
Press.
Unsworth, N., Schrock, J. C., & Engle, R. W. (2004). Working
memory capacity and the antisaccade task: Individual differences
in voluntary saccade control. Journal of Experimental Psychology:
Learning, Memory, & Cognition, 30, 1302-1321.
Waters, G. S., & Caplan, D. (1996). The measurement of verbal
working memory capacity and its relation to reading comprehension.
Quarterly Journal of Experimental Psychology, 49A, 51-74.
Waugh, N. (1960). Serial position and the memory-span. American
Journal of Psychology, 73, 68-79.
Yntema, D. B., & Mueser, G. E. (1960). Remembering the present
states of a number of variables. Journal of Experimental Psychology,
60, 18-22.
Yntema, D. B., & Mueser, G. E. (1962). Keeping track of variables
that have few or many states. Journal of Experimental Psychology,
63, 391-395.
NOTES
1. For our discussion of the administration and scoring of WM span
tasks, it is important to distinguish between elements, items, and tasks.
Elements are the individual stimuli that have to be recalled. This is the
lowest level of observation. The recall of an element is either correct or
incorrect. Items include various numbers of elements. Usually, there is
an experimental manipulation of how many elements an item includes.
The number of elements in an item has often been labeled set size, and
typically varies between two and seven. A number of homogeneous
items make up a task. In psychometrics, a task is usually called a test.
We prefer the term task, in order to maintain continuity with the experi-
mental, rather than the psychometric, origins of this research tradition.
2. Engle, Tuholski, et al. (1999) also changed the stimuli by adding
the word is before the operation word string. Stating the operation in the
form of a question emphasized that the subjects were to indicate aloud
whether or not the stated solution was correct.
3. Additional data can also be obtained from the recall portion of
WM span tasks. Although it is not a very common procedure, evidence
suggests that overall recall durations, as well as the pause time between
recall of words, can add independent variance to that offered by span
scores in predicting ability (Cowan et al., 2003).
4. Klein and Fiss (1999) found that scores on the operation span task
markedly increased from the first to the second administration of the
test, indicating a practice effect. This result is important for researchers
interested in training or intervention-type manipulations that require a
pretest/posttest design with a control group.
5. Most WM tasks require serial recall. However, the original reading
span task was essentially a constrained free recall task, in that subjects could
recall the sentence-final words in any order, as long as they did not recall
the last word first. More recently, Friedman and Miyake (2004) instructed
subjects to recall in order, but if they could not, then to just avoid recalling
the last word first. Thus, strict serial recall is not always required.
6. There is some evidence that running-memory span may have more
in common with STM than with WMC: It is highly vulnerable to the
effects of articulatory suppression and background speech (Morris &
Jones, 1990), which are hallmarks of the operation of the phonological
loop, or verbal STM system (see Baddeley, 1986).
(Manuscript received June 18, 2004;
revision accepted for publication February 24, 2005.)