Machine Translated by Google

Cognitive Psychology 41, 49–100 (2000) doi:10.1006/

cogp.1999.0734, available online at http://www.idealibrary.com on

The Unity and Diversity of Executive Functions and Their

Contributions to Complex ''Frontal Lobe'' Tasks: A Latent
Variable Analysis

Akira Miyake, Naomi P. Friedman, Michael J. Emerson,

Alexander H. Witzki, and Amy Howerter

University of Colorado at Boulder

and

Tor D. Wager

University of Michigan

This individual differences study examined the separability of three often postulated executive
functions—mental set shifting (''Shifting''), information updating and monitoring (''Updating''), and
inhibition of prepotent responses (''Inhibi-tion'')—and their roles in complex ''frontal lobe'' or ''executive''
tasks. One hundred thirty-seven college students performed a set of relatively simple experimental
tasks that are considered to predominantly tap each target executive function as well as a set of
frequently used executive tasks: the Wisconsin Card Sorting Test (WCST), Tower of Hanoi (TOH),
random number generation (RNG), operation span, and dual tasking. Confirmatory factor analysis
indicated that the three target executive functions are moderately correlated with one another, but are
clearly separable. Furthermore, structural equation modeling suggested that the three functions

We thank Anna Ficken, Timi Iddings, Silvie Kilworth, Vandana Passi, Juan Quezada, Bob Slevc, and Neal
Wolcott for their help in running the experiments and scoring data. We also thank Greg Carey for his statistical
advice; John R. Crawford, John Duncan, Priti Shah, and an anonymous reviewer for their comments on a draft of
this article; Dan Kimberg and Jim Parker for making the versions of the Wisconsin Card Sorting Test and the Stroop
task pro-grams, respectively, available to us; and Ernie Mross for programming the Tower of Hanoi task. This
research was supported in part by a National Science Foundation (NSF) KDI/LIS Grant (IBN–9873492), a University
of Colorado Council of Research and Creative Work Grants-in-Aid award, and an NSF Graduate Fellowship. A
preliminary version of this research was presented at the 11th Annual Convention of the American Psychological
Society, Denver, Colorado in June 1999.

Correspondence and reprint requests concerning this article should be addressed to Akira Miyake, Department
of Psychology, University of Colorado at Boulder, Campus Box 345, Boulder, CO 80309–0345. Electronic mail may
also be sent to MIYAKE@PSYCH.
COLORADO.EDU.
49
0010-0285/00 $35.00
Copyright ÿ 2000 by Academic Press
All rights of reproduction in any form reserved.
Machine Translated by Google
50 MIYAKE ET AL.

contribute differentially to performance on complex executive tasks. Specifically, WCST performance

was related most strongly to Shifting, TOH to Inhibition, RNG to Inhibition and Updating, and operation
span to Updating. Dual task performance was not related to any of the three target functions. These
results suggest that it is important to recognize both the unity and diversity of executive functions and
that latent variable analysis is a useful approach to studying the organization and roles of executive
functions. ÿ 2000 Academic Press

Cognitive psychology has made considerable progress over the last few decades and has
developed sophisticated theories and models about specific cognitive domains or processes
(such as object perception, word recognition, syntactic parsing, etc.). Despite this headway,
there still remain a number of theoretical issues or phenomena about which little can be said.
According to Monsell (1996), one such ''embarrassing zone of almost total ignorance'' (p. 93)
concerns how specific cognitive processes are controlled and coordinated during the
performance of complex cognitive tasks. In other words, the field still lacks a compelling
theory of executive functions—general-purpose control mechanisms that modulate the
operation of various cognitive subprocesses and thereby regulate the dynamics of human
cognition.

The main goal of the present article is to provide a necessary empirical basis for developing
a theory that specifies how executive functions are organized and what roles they play in
complex cognition. Toward this goal, we report an individual differences study of executive
functions. Specifically, we focus on three of the most frequently postulated executive functions
in the literature—shifting of mental sets, monitoring and updating of working memory
representations, and inhibition of prepotent responses—and specify how separable these
functions are and how they contribute to so-called fron-tal lobe or executive tasks.

Research on executive functions has historical roots in neuropsychological studies of

patients with frontal lobe damage. It has been known for a long time that patients with damage
to the frontal lobes, including the well-known patient Phineas Gage, demonstrate severe
problems in the control and regulation of their behavior and cannot function well in their
everyday lives. Although some of these patients demonstrate remarkably intact performance
on various well-defined cognitive tasks from neuropsychological test batter-ies and IQ tests
(eg, Damasio, 1994; Shallice & Burgess, 1991), they tend to show, as a group, some
impairments on a host of complex frontal lobe or executive tasks. These tasks include, among
others, the Wisconsin Card Sorting Test (WCST) and the Tower of Hanoi (TOH) task and its
variant, the Tower of London task. Although these tasks are complex and poor performance
on them could arise for many different reasons, they have nevertheless become the primary
research tools for studying the organization and roles of executive functions in
neuropsychological studies of brain-damaged patients and, more recently, in individual
differences studies of normal populations from different age groups. In particular, these frontal
lobe or executive tasks
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 51

have provided a basis for many proposals regarding the nature of the
cognitive deficits that frontal lobe patients exhibit as well as the nature of the
control functions that the normal, intact frontal lobes seem to perform.1
One of the most prominent cognitive frameworks that has been associated
with the study of executive functions is Baddeley's (1986) influential multi-
component model of working memory. This model includes three
components, two of which are specialized for the maintenance of speech-based,
phonological information (the phonological loop) and visual and spatial in-
formation (the visuospatial sketchpad), respectively. In addition to these two
''slave'' systems, the model also includes a central control structure called
the central executive, which is considered responsible for the control and
regulation of cognitive processes (ie, executive functions) and is often
linked to the functioning of the frontal lobes. Baddeley (1986) also proposed
that Norman and Shallice's (1986; Shallice, 1988) Supervisory Attentional
System (SAS), originally constructed as a model of attentional control of
behavior in normals as well as neuropsychological patients, may be a candidate-date model
of the central executive.
One important research question that has been a source of controversy in
both neuropsychological and cognitive studies of executive functions are an
issue raised by Teuber (1972) in his review entitled Unity and diversity of
frontal lobe functions and recently revisited by Duncan and his colleagues
(Duncan, Johnson, Swales, & Freer, 1997). Specifically, to what extent can
different functions often attributed to the frontal lobes or to the central execu-tive (or SAS)
be considered unitary in the sense that they are reflections of
the same underlying mechanism or ability?
At least in the early stages of theoretical development, both the central
executive and the SAS had a unitary flavor, without including any distinct
subfunctions or subcomponents. In addition, some recent conceptions of
executive functions suggest that there is some common basis or a unifying
mechanism that can characterize the nature of deficits in frontal lobe patients
or the functions of the frontal lobes (eg, Duncan, Emslie, Williams, John-
son, & Freer, 1996; Duncan et al., 1997; Engle, Kane, & Tuholski, 1999a;
Kimberg & Farah, 1993).
In contrast, there is also some evidence for the nonunitary nature of frontal

1
Despite the fact that the phrases ''frontal lobe'' and ''executive'' are often used
interchangeably, they are not conceptually identical (Baddeley, Della Sala, Gray, Papagno,
& Spinner, 1997). Although there is strong evidence that the frontal lobes may play an important
role in executive control of behavior, some frontal lobe patients do not show any problems
with frontal lobe tasks (Shallice & Burgess, 1991), whereas some patients who have lesions
outside the frontal lobes can demonstrate severe impairments on them (Anderson, Damasio,
Jones, & Tranel, 1991; Reitan & Wolfson, 1994). Such findings suggest that the anatomical
term ''frontal lobe'' and the functional term ''executive'' are not necessarily synonymous.
For this reason, we will use the term ''executive tasks,'' rather than ''frontal lobe tasks,'' for
the rest of the article.
Machine Translated by Google
52 MIYAKE ET AL.

lobe or executive functions (Baddeley, 1996). One line of evidence comes

from clinical observations, which indicate some dissociations in perform-ance
among the executive tasks. For example, some patients may fail on
the WCST, but not on the TOH, whereas others may show the opposite
pattern, suggesting that executive functions may not be completely unitary
(eg, Godefroy, Cabaret, Petit-Chenal, Pruvo, & Rousseaux, 1999; Shallice,
1988).
Another line of evidence for the nonunitary nature of executive functions
comes from a number of individual differences studies, the main focus of the
present article. These studies examined a wide range of target populations,
including normal young adults (Lehto, 1996), normal elderly adults (Lowe &
Rabbitt, 1997; Robbins et al., 1998), brain-damaged adults (Burgess, 1997;
Burgess, Alderman, Evans, Emslie, & Wilson, 1998; Duncan et al., 1997),
and children with neurocognitive pathologies (Levin et al., 1996; Schachar,
Tannock, & Logan, 1993; Welsh, Pennington, & Groisser, 1991). Despite
differences in the target populations, these studies are similar in the sense
that they all employed a battery of widely used executive tasks like the
WCST and TOH and examined how well these tasks correlated with one
another by performing correlation–regression analyzes and, in many cases,
exploratory factor analyzes (EFA). Although details of the results differ from
study to study, a highly consistent pattern that holds across these individuals
differences studies is that the intercorrelations among different executives
tasks are low (usually r .40 or less) and are often not statistically significant.
EFA also tends to yield multiple separable factors (rather than a single
unitary factor) for a battery of executive tasks. The results from these individual
differences studies are often used to argue that the functions of the frontal
lobe or the central executive (or SAS) are not unitary and hence need to be
fractionated.
Although it has provided useful insights, the typical correlational or factor-
analytic approach has several important weaknesses or limitations (Baddeley,
Della Sala, Gray, Papagno, & Spinnler, 1997; Rabbitt, 1997a). One major
weakness is that, although the finding of low correlations among executive
tasks seems to be robust across studies, it is not completely clear whether
such reported lack of correlations is indeed a reflection of the independence
of underlying executive functions (Miyake & Shah, 1999). It is quite possible
that striking differences in nonexecutive processing requirements (eg, lan-
guage and visual processing) have simply masked the existence of
some underlying commonalities among the chosen executive tasks. More
Generally, this issue highlights the so-called task impurity problem, a
particularly vexing issue in studies of executive functions (Burgess, 1997; Phillips,
1997). Because executive functions necessarily manifest themselves by op-
erating on other cognitive processes, any executive task strongly implies
other cognitive processes that are not directly relevant to the target executive
function. For these reasons, a low score on a single executive test does not
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 53

necessarily mean inefficient or impaired executive functioning. Similarly,

low zero-order correlations or multiple separable factors may also not be
due to dissociable executive functions (Miyake & Shah, 1999).
This task impurity problem is further compounded by the observation that
complex executive tasks tend to suffer from relatively low internal and/or
test–retest reliability (Denckla, 1996; Rabbitt, 1997b). Although the reasons
for the low reliabilities are not completely clear, one possibility is that people
adopt different strategies on different occasions (or even within a session)
when performing these tasks. Also, the involvement of executive control
functions is generally considered strongest when the task is novel (Rabbitt,
1997b). Thus, repeated encounters with the task may reduce its effectiveness
in actually capturing the target executive process, thereby yielding low reli-ability.
Regardless of what factors contribute to the reliability problem, an
important point for our current discussion is that measures with low reliabilities
necessarily lead to low correlations with other measures. Thus, low zero-order
correlations among executive tasks could be a reflection of low reliabil-ities of the
measures themselves, rather than a reflection of independence of
underlying executive functions tapped by individual tasks.
Another important problem associated with the reliance on prevalent com-plex
executive tasks like the WCST and TOH is that, despite their wide
acceptance as measures of executive functioning, their construct validities
are not well established (Phillips, 1997; Rabbitt, 1997b; Reitan & Wolfson,
1994). Many popular executive tasks seem to have been validated only to
the rather loose criterion of being somewhat sensitive to frontal lobe damage
(ie, at least some frontal lobe patients show difficulty performing the tasks),
and the precise nature of executive processes involved in the performance
of these tasks is underspecified, to say the least. In other words, there is a
paucity of rigorous theoretical analysis and independent empirical evidence
regarding what these executive tasks really measure.
This clarity of the underlying abilities tapped by these complex executive tasks
is reflected in a proliferation of terms and concepts used to charac-terize the task
requirements of different executive tests. The WCST, for example, has been
suggested by different researchers as a measure of ''mental
set shifting,'' ''inhibition,'' ''flexibility,'' ''problem solving,'' and ''catego-rization,'' just
to name a few. Although these suggestions may sound reason-able at an intuitive
level, no independent testing of them has been reported.
Another related consequence of the uncertainty as to what these executive tests
really measure is the difficulty of interpreting what construct(s) different
factors obtained in many EFA studies of executive functions really represent.
Interpretations given to obtained factors often seem quite arbitrary and post-hoc.
For example, a factor that loaded highly on WCST, verbal fluency, and
design fluency tests was interpreted as a ''Conceptual/Productivity'' factor
by Levin et al. (1996). Although such interpretation difficulties reflect, in
large part, the characteristics of the EFA technique, not knowing what execu-
Machine Translated by Google
54 MIYAKE ET AL.

tive functions these complex tasks really tap is likely another important
reason.2
Taken together, all of these problems seriously undermine the usefulness
of typical correlational, factor-analytic studies for theorizing about the organization
of executive functions and their roles in complex cognition. Although we do not
deny the utility of these methods as exploratory tools,
new approaches that overcome these problems are clearly needed for further
theoretical development. We argue that one such promising approach is la-tent
variable analysis.

THE PRESENT STUDY

In this article, we report an individual differences study of executive functions

that we believe alleviates at least some of the problems that have
plagued the typical individual differences approach. Specifically, we focus
on three executive functions that are frequently postulated in the literature,
carefully select multiple tasks that tap each target executive function, and
examine the extent of unity or diversity of these three executive functions
at the level of latent variables (ie, what is shared among the multiple example tasks
for each executive function), rather than at the level of manifest
variables (ie, individual tasks). In other words, we statistically ''extract''
what is common among the tasks selected to tap a putative executive function
and use that ''purer'' latent variable factor to examine how different executive functions
relate to one another.
As will become clear, this latent variable approach has a number of important
advantages over a more typical individual differences approach. For example, the
emphasis on latent variables (as opposed to manifest variables)
should minimize the task impurity problem. In addition, this study examines,
also at the level of latent variables, how each target executive function con-tributes
to performance on a number of complex executive tasks such as the
WCST and TOH. Such an attempt should provide useful insights as to what
each complex executive task really measures and hence will likely contribute
to the deviation of the construct validity problem as well.

The Three Executive Functions Examined in This Study

We focus on the following three executive functions: (a) shifting between
tasks or mental sets, (b) updating and monitoring of working memory representations,
and (c) inhibition of dominant or prepotent responses. all three
are frequently postulated in the literature as important executive functions

2
Such interpretation difficulties may be exacerbated if an orthogonal rotation technique (a
default option of most statistical programs), which does not allow extracted factors to correlate
with each other, is used when there is good reason to suspect some interfactor correlations
(Fabrigar, Wegener, MacCullum, & Strahan, 1999).
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 55

(eg, Baddeley, 1996; Logan, 1985; Lyon & Krasnegor, 1996; Rabbitt,
1997a; Smith & Jonides, 1999). We chose to focus on these three functions
for several reasons. First, they seem to be relatively circumscribed, lower
level functions (in comparison to some other often postulated executive functions
like ''planning'') and hence can be operationally defined in a fairly
precise manner. Second, for these three executive functions, a number of
well studied, relatively simple cognitive tasks that we believed would primarily
tap each target function were available. Third, and perhaps most impor-tantly,
the three target functions are likely to be involved in the performance of more
complex, conventional executive tests. For example, the
WCST has often been suggested as a test that measures set shifting (for
shifting between sorting principles) as well as inhibition (for suppressing
inappropriate responses). Thus, a good understanding of these three executive
functions may provide a basis for specifying what traditional executive
tests really measure.
Below, we define and review these three executive functions and briefly
discuss the tasks we chose as measures of each executive function. Many
of these tasks seem to involve the frontal lobes, although performance on
they would certainly rely on other brain regions as well. Details of each
tasks are provided under Method.
Shifting between tasks or mental sets (''Shifting''). The first executive
function concerns shifting back and forth between multiple tasks, operations,
or mental sets (Monsell, 1996). Also referred to as ''attention switching''
or ''task switching,'' this ability (henceforth, called ''Shifting'' for short)
has been proposed as a candidate executive function and appears to be
important in understanding both failures of cognitive control in brain-damaged
patients and laboratory tasks that require participants to shift between tasks
(Monsell, 1996). In addition, models of attentional control like SAS (Norman &
Shallice, 1986) often assume that the ability to shift between tasks
or mental sets is an important aspect of executive control.
The tasks we chose to tap the Shifting function are the plus–minus task
(Jersild, 1927), the number–letter task (Rogers & Monsell, 1995), and the
local–global task. All of these tasks require shifting between mental sets,
although the specific operations that need to be switched back and forth are
rather different across tasks; Thus, what the three chosen tasks have in common
is likely to be the Shifting requirement, rather than other idiosyncratic
task requirements not related to the target executive function. Previous stud-ies
have shown conclusively that shifting mental sets incurs a measurable
temporal cost (eg, Jersild, 1927; Rogers & Monsell, 1995), particularly
when the shifting must be driven internally, rather than by external cues
(Spector & Biederman, 1976).
Perhaps the most common explanation of this function is that the Shifting
process involves the disengagement of an irrelevant task set and the subsequent
active engagement of a relevant task set. Although still prevalent, this
Machine Translated by Google
56 MIYAKE ET AL.

conceptualization of Shifting may be too simplistic. Recent work suggests

that, when a new operation (say, subtracting 3) must be performed on a set
of stimuli (eg, a list of two-digit numbers), it may be necessary to overcome
interference or negative priming due to having previously per-formed a different
proactive operation (eg, adding 3) on the same type of stimuli
(Allport & Wylie, in press). Thus, individual differences in the Shifting abil-ity
may not be a simple reflection of the ability to engage and disengage
appropriate task sets per se, but may also (or even instead) involve the ability
to perform a new operation in the face of proactive interference or negative
priming.
Despite the apparent similarity, the notion of Shifting that we focus on in
this article is not synonymous with the abilities involved in spatially shifting
or switching visual attention by making appropriate voluntary eye movements
or covertly moving visual attention. Posner and Raichle (1994) argued
that different neural circuits may mediate the shifting of visual attention and
more executive-oriented shifts that involve, for example, the conscious ful-
filling of instructions, although these networks seem to interact with each
other. More specifically, visual attention shifting may be regulated primarily
by the parietal lobes and the mid-brain (or the ''posterior attention net-work''),
whereas more executive-oriented shifts may be regulated primarily
by the frontal lobes, including the anterior cingulate (or the ''anterior attention
network'').
In fact, there is a growing body of neuropsychological and neurophysio-
logical evidence indicating that shifting between tasks or mental sets in-volves
the frontal lobes, although not necessarily to the exclusion of others
brain regions. For example, an event-related potential (ERP) study has
indicated that shifting between two tasks activated frontal as well as bioccipital
and parietal regions (Moulden et al., 1998). In addition, one key symptom
of frontal lobe impairments, perseveration or repeating the same response
over and over even when it is clearly no longer appropriate, it is often
interpreted in terms of difficulty in shifting mental set (Luria, 1966; Stuss &
Benson, 1986). With regard to the tasks used in the current study, we know
of no neuroimaging studies demonstrating that the frontal lobes are implicated
in the performance on these specific tasks. However, there is some
neuropsychological evidence indicating that patients with damage to the left
frontal lobes demonstrate a significant shifting impairment compared to age-
and IQ-matched controls on a simplified version of the number–letter task,
at least in a task condition that is most similar to the task used in this study
(Rogers et al., 1998).
Updating and monitoring of working memory representations (''Updat-ing'').
The second target executive function, updating and monitoring of
working memory representations (''Updating'' for short), is closely linked
to the notion of working memory (Jonides & Smith, 1997; Lehto, 1996),
which in turn is often associated with the prefrontal cortex, particularly its
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 57

dorsolateral portion (Goldman-Rakic, 1996; Smith & Jonides, 1999). This Updating function requires
monitoring and coding incoming information for relevance to the task at hand and then appropriately revising
the items held in working memory by replacing old, no longer relevant information with newer, more relevant
information (Morris & Jones, 1990). Jonides and Smith (1997) have suggested that this Updating process
may involve ''temporal tagging'' to keep track of which information is old and no longer relevant.

Importantly, this Updating function goes beyond the simple maintenance of

task-relevant information in its requirement to dynamically manipulate the
contents of working memory (Lehto, 1996; Morris & Jones, 1990). That is, the
essence of Updating lies in the requirement to actively manipulate relevant
information in working memory, rather than passively store information.
Consistent with this distinction, recent neuroimaging studies have shown
dissociations in the areas required for relatively passive storage and active
updating: Whereas the simple storage and maintenance of information has
been associated with premotor areas of the frontal cortex and the parietal lobes,
the Updating function, as measured by a complex task like the N-back task, has
been linked to the dorsolateral prefrontal cortex (Jonides & Smith, 1997). In
addition, a proposed component of Updating, namely, temporal sequencing
and monitoring, has also been associated with the frontal lobes (see Stuss,
Eskes, & Foster, 1994, for a review).
The tasks we chose to tap the Updating function are the keep track task
(Yntema, 1963), the letter memory task (Morris & Jones, 1990), and the tone
monitoring task. All three involve constantly monitoring and updating information
in working memory, although the nature of the information that needs to be
updated as well as the goals of the tasks is rather different. To our knowledge,
no studies have linked the keep track and tone monitoring tasks to the frontal
lobes, but a recent PET study has indicated that the updating component of
the letter memory task (with the influence of the storage component subtracted)
is associated most strongly with the left frontopolar cortex (Van der Linden et
al., 1999).
Inhibition of prepotent responses (''Inhibition''). The third executive function
examined in this study concerns one's ability to deliberately inhibit dominant,
automatic, or prepotent responses when necessary (''Inhibition'' for short). A
prototypical Inhibition task is the Stroop task, in which one needs to inhibit or
override the tendency to produce a more dominant or automatic response (ie,
name the color word). This type of Inhibition is commonly labeled an executive
function—for example, Logan (1994) has called it an ''internally generated act
of control'' (p. 190)—and linked to the frontal lobes (eg, Jahanshahi et al., 1998;
Kiefer, Marzinzik, Weisbrod, Scherg, & Spitzer, 1998).

Given that the term inhibition is commonly used to describe a wide variety of
functions at a number of levels of complexity (Kok, 1999), it is important
Machine Translated by Google
58 MIYAKE ET AL.

to note that the conception of Inhibition used here is constrained to the delib-
erate, controlled suppression of prepotent responses. Thus, by Inhibition, we
do not mean inhibition that takes place in typical spreading activation models
or connectionist networks. That type of inhibition usually refers to a decrease
in activation levels due to negative activation (eg, a result of negative connection
weights) and is not necessarily a deliberate, controlled process. Nor
do we mean ''reactive inhibition,'' such as that seen with phenomena like
negative priming or inhibition of return. Reactive inhibition seems to be a
residual aftereffect of processing that is not usually intended (Logan, 1994),
whereas the Inhibition we focus on is a process that is actually intended.
Although these two types of inhibition may share some underlying common-ality
and may be correlated with one another, they are conceptually separable, and
we restrict the notion of Inhibition in this study to deliberate,
intended inhibition of prepotent responses.3
The tasks used to tap the Inhibition ability are the Stroop task (Stroop,
1935), the antisaccade task (Hallett, 1978), and the stop-signal task (Logan,
1994). All require deliberately stopping a response that is relatively automatic,
although the specific response that needs to be inhibited differs across
tasks. Previous research has shown that both the Stroop task (eg, Perret,
1974) and the antisaccade task (eg, Everling & Fischer, 1998) are sensitive
to lesions to the frontal lobes and other types of frontal lobe dysfunction.
Although the stop-signal task has not been examined in a neuropsychological
context, a simpler yet similar ''go–no-go'' task has been shown to strongly
implicate the prefrontal cortex among both children (Casey et al., 1997) and
adults (Kiefer et al., 1998).

Two Central Goals of the Present Study

Previous individual differences studies of executive functions tend to suggest
some level of fractionation of executive functions, but, as we reviewed
earlier, there are several serious problems with interpreting the results of
typical correlational and EFA studies. The present study was designed to
go beyond previous individual differences studies and provide a stronger
assessment of the relationships among the three frequently postulated executive
functions of Shifting, Updating, and Inhibition. More specifically, the
study had two main goals. The first major goal was to specify the extent to
which the three target executive functions are unitary or separable. To the
extent that the three functions represent distinguishable executive functions,

3
One alternative conceptualization of inhibition is in terms of actively ''boosting'' activation
(or maintaining a high level of activation) for the weaker, to-be-selected process, rather
than directly ''suppressing'' the dominant, prepotent process (eg, Kimberg & Farah, 1993).
Both conceptualizations seem plausible at this point as a mechanism involved in the Inhibition
process and are compatible with the results of the present study. Thus, we do not strongly
endorse one conceptualization over the other, although we discuss the construct of Inhibition
in terms of active suppression in this article for the sake of simplicity.
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 59

the second major goal was to specify their relative contributions to more
complex tests commonly used to assess executive functioning.
With respect to the first goal, we used confirmatory factor analysis (CFA)
to specify the degree to which the three postulated executive functions are
separable or share the same underlying ability or mechanism. CFA is similar
to the EFA technique more commonly used in the field (the term ''factor
analysis'' with no modifier typically refers to EFA, rather than CFA). One
major difference, however, is that, whereas EFA finds the one underlying
factor model that best fits the data, CFA allows researchers to impose a par-ticular factor
model and then see how well that statistical model fits the data
(Kline, 1998). In other words, with EFA, one lets the observed data determine
the underlying factor model a posteriori (this characteristic of EFA is
part of the reason the factors extracted with this method do not necessarily
have clear interpretations), whereas with CFA, one derives a factor model
or models a priori on the basis of theoretical considerations and then evaluate
their fit to the data. Thus, CFA is a highly theory-driven multivariate analysis
technique and serves as a valuable tool for specifying the organization
of executive functions.
We used CFA to compare models with one, two, or three factors. figure
1A illustrates the theoretical model that provided the basis for our analysis
(called the ''full, three-factor'' model). Ellipses in the figure represent the
three target latent variables (ie, Shifting, Updating, and Inhibition), whereas
rectangles represent the manifest variables (ie, individual tasks) that were
used to tap the specific functions, as indicated by the straight, single-headed
arrows. The curved, double-headed arrows represent correlations among the
latent variables.
If it is necessary to postulate three separable factors (one for each target
executive function), then this full, three-factor model should provide an ex-
cellent fit to the data, and the correlations among the three latent variables
will provide an estimate of the degree to which the three target functions
are related to one another. In contrast, if the three executive functions
essentially tap the same underlying construct and hence should be considered
uni-tary, then a model with one factor (created by fixing all of the correlations
among the three latent variables to 1.0 so that this alternative model is
''nested'' within the full, three-factor model) should provide an excellent fit
to the data, a fit statistically no worse than the full, three-factor model (be-
cause the full, three-factor model has more free parameters than the one-
factor model, the absolute fit of the one-factor model cannot exceed that of
the three-factor model). Similarly, if two of the target executive functions
(but not the third one) tap a common underlying ability, then a model with
two factors (ie, a model that fixes the correlation between the two unitary
executive functions to 1.0 but lets the other two correlations vary freely)
should provide a fit to the data that is statistically as good as the full, three-
factor model. Finally, if the executive functions are completely independent,
Machine Translated by Google
60 MIYAKE ET AL.

FIG. 1. (A) The theoretical ''full, three-factor'' model used for the confirmatory factor
analysis (CFA). The ellipses represent the three executive functions (latent variables), and the
rectangles represent the individual tasks (manifest variables) that were chosen to tap the
specific executive functions, as indicated by the straight, single-headed arrows. The curved
double-headed arrows represent correlations among the latent variables. Both models depict three
latent constructs, namely, Shifting, Updating, and Inhibition, which are hypothesized to be
correlated but separable. (B) A generic model for the structural equation modeling (SEM)
analysis. This model is identical to the CFA model with the addition of a manifest variable
on the right side that represents a complex executive function measure. In this particular model
(the ''full'' model), the manifest variable on the right has paths from all three latent variables
to estimate the contribution of each to performance on the executive task.
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 61

then a model with no relationships among the three factors (ie, a model
that fixes the correlations among the factors all to zero) should provide a
good fit to the data. Thus, such systematic model comparisons will tell us
the degree to which the three executive functions are separable.
For the second goal, we performed a series of structural equation modeling
(SEM) analyzes to examine how each of the three target executive functions
contributes to performance on a number of executive tasks used in cognitive
and neuropsychological studies: WCST, TOH, random number generation
(RNG), the operation span task, and a dual task. These executive tasks were
chosen primarily because they are frequently used as measures of the
integrity of executive functioning among frontal lobe patients (ie, WCST and
TOH) or measures of central executive functioning among healthy individuals
(ie, RNG, operation span, and dual tasking).
Figure 1B provides an illustration of the logic behind our SEM analyses.
The model is basically the same as the CFA model (Fig. 1A), with the
addition of a manifest variable (ie, an individual executive task) on the right
side of the model and potential paths from each latent variable to this new
manifest variable. By performing SEM analyzes and comparing different
alternative models (eg, models with all three paths, two paths, one path, and
zero paths), we sought to determine which path(s) is (are) really necessary to
fit the data well or which path(s) can be dropped without significantly
worsening the overall data fit. Thus, in this particular application, SEM could be
considered a more elaborate version of multiple regression analysis in which
latent variables serve as predictor variables. In this context, choosing the
best model is analogous to selecting a best-fitting regression model that can
Parsimoniously account for a significant portion of the variance in the
dependent variable with the fewest predictor variables in the equation.
In performing the SEM analyses, we guided our model comparison
process by previous proposals in the literature concerning what each of these
executive tasks really measures. Specifically, we developed a particular
hypothesis (or, in some cases, a particular set of hypotheses) a priori and then
tested the hypothesis (or hypotheses) with SEM. Thus, the SEM analyzes
reported in this article provide an independent empirical test of previous
proposals regarding the nature of executive function(s) tapped by these com-
plex executive tasks.

METHOD

Participants
The participants were 137 undergraduates from the University of Colorado at Boulder who
received partial course credit for taking part in the study. Five additional participants took
part in the study, but their data were not complete for the nine target tasks used to tap the
three executive functions for the following reasons: Two participants were not native speakers
of English and demonstrated marked impairment on certain tasks involving a greater level of
Machine Translated by Google
62 MIYAKE ET AL.

language proficiency; one participant was color-blind and had great difficulty performing the
Stroop task; and two participants were excluded on the basis of outlier analyzes on the nine
tasks to be reported later. Thus, the data from these additional participants were not included
in the CFA and SEM analyses.
For three of the complex executive tasks (ie, WCST, TOH, and dual tasking), some
observations were lost due to equipment malfunction. Hence, the SEM analyzes with these
particular tasks relied on slightly fewer observations (N 134 for WCST and dual tasking
and N 136 for TOH).

Materials, Design, and Procedure

All participants completed the nine tasks hypothesized to tap one of the three target
executive functions of Shifting, Updating, or Inhibition, as well as the five complex tasks
commonly used as measures of executive functioning. Task administration was either
computerized (Power Macintosh 7200 computers) or paper-and-pencil. A button box with
millisecond accuracy was employed for the computerized tasks using reaction time (RT)
measures, and a voice key was attached to the button box to record RTs for verbal responses.
The following three tasks were used as the Shifting tasks:
Plus–minus task. The plus–minus task, adapted from Jersild (1927) and Spector and
Bieder-man (1976), consisted of three lists of 30 two-digit numbers (the numbers 10–99
prerandom-ized without replacement) on a single sheet of paper. On the first list, the
participants were instructed to add 3 to each number and write down their answers. On the
second list, they were instructed to subtract 3 from each number. Finally, on the third list, the
participants were required to alternate between adding 3 to and subtracting 3 from the
numbers (ie, add 3 to the first number, subtract 3 from the second number, and so on). The
participants were instructed to complete each list quickly and accurately, and list completion
times were measured by a stopwatch. The cost of shifting between the operations of addition
and subtraction was then calculated as the difference between the time to complete the
alternating list and the average of the times to complete the addition and subtraction lists,
and this shift cost served as the dependent measure.
Number–letter task. In the number–letter task, adapted from Rogers and Monsell (1995),
a number–letter pair (eg, 7G) was presented in one of four quadrants on the computer screen.
The participants were instructed to indicate whether the number was odd or even (2, 4, 6,
and 8 for even; 3, 5, 7, and 9 for odd) when the number–letter pair was presented in either of
the top two quadrants and to indicate whether the letter was a consonant or a vowel (G, K,
M, and R for consonant; A, E, I, and U for vowel) when the number–letter pair was presented
in either of the bottom two quadrants. The number–letter pair was presented only in the top
two quadrants for the first block of 32 target trials, only in the bottom two quadrants for the
second block of 32 target trials, and in a clockwise rotation around all four quadrants for the
third block of 128 target trials. Thus, the trials within the first two blocks required no task
switching, whereas half of the trials in the third block required participants to shift between
these two types of categorization operations. In all trials (plus 10–12 practice trials in each
block), the participants responded by button press, and the next stimulus was presented 150
ms after the response. Similar to the plus–minus task, the shift cost for this task was the
difference between the average RTs of the trials in the third block that required a mental shift
(trials from the upper left and lower right quadrants) and the average RTs of the trials from
the first two blocks in which no shift was necessary.
Local–global task. In the local–global task, a geometric figure often called a Navon figure
(Navon, 1977), in which the lines of the ''global'' figure (eg, a triangle) were composed of
much smaller, ''local'' figures (eg, squares), was presented on the computer screen.
Depending on the color of the figure (either blue or black), participants were instructed to say
out loud the number of lines (ie, 1 for a circle, 2 for an X, 3 for a triangle, and 4 for a square)
in the global, overall figure (blue) or the local, smaller figures (black). Thus, when the colors
of the stimuli changed across successive trials, the participants had to shift from
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 63

examining the local features to the global features or vice versa. A voice key was used to
measure RTs. After 36 practice trials (of which 24 served as voice-key calibration trials),
participants performed one block of 96 target trials, each separated by a 500-ms response-to-
stimulus interval. The target trials were prerandomized, with the constraint that half of the trials
require a switch from local to global features or from global to local features, and the shift cost
was then calculated as the difference between the average RTs for the trials requiring a shift
in mental set (ie, color of stimulus changed) and the trials in which no shift was required (ie,
the color of stimulus remained the same).
The following three tasks were used as the Updating tasks:
Keep track task. In each trial of the keep track task (adapted from Yntema, 1963), partici-
pants were first shown several target categories at the bottom of the computer screen. Fifteen
words, including 2 or 3 exemplars from each of six possible categories (animals, colors, coun-
tries, distances, metals, and relatives), were then presented serially and in random order for
1500 ms apiece, with the target categories remaining at the bottom of the screen. The task
was to remember the last word presented in each of the target categories and then write down
these words at the end of the trial. For example, if the target categories were metals, relatives,
and countries, then, at the end of the trial, participants recalled the last metal, the last relative,
and the last country presented in the list. Thus, participants had to closely monitor the words
presented and update their working memory representations for the appropriate categories
when the presented word was a member of one of the target categories. Before this task
began, participants saw all six categories and the exemplars in each to ensure that they knew
to which category each word belonged and then practiced on a single trial with three target categorie
Participants then performed three trials with four target categories and three with five target
categories, recalling a total of 27 words. The proportion of words recalled correctly was the
dependent measure.
Tone monitoring task. In the tone monitoring task (substantially modified from the Mental
Counters task developed by Larson, Merritt, & Williams, 1988), participants were presented
with four trial blocks, each consisting of a series of 25 tones presented for 500 ms a piece,
with an interstimulus interval of 2500 ms. Each block included a mixed order of 8 high-pitched
tones (880 Hz), 8 medium-pitched tones (440 Hz), 8 low-pitched tones (220 Hz), and 1 tone
randomly selected from the three pitches (for a total of 25 tones). The task was to respond
when the 4th tone of each particular pitch was presented (eg, after hearing the 4th low tone,
the 4th medium tone, or the 4th high tone), which required participants to monitor and keep
track of the number of times each pitch had been presented. For example, if the sequence
was ''low, high, medium, high, high, low, medium, high, low, high,'' then the participant should
have responded to the 4th high tone (italicized) and, if asked at the end of the sequence, they
should also have been aware that his or her mental counters contained 3 low tones, 2 medium
tones, and 1 high tone. In order for momentary mental lapses to have less impact on task
performance, the tone count for each pitch automatically reset to 0 if participants made an
incorrect button press for that pitch (eg, responding after the 3rd high tone), and participants
were informed of this feature before starting the task. Prior to completing the four trial blocks,
participants received a guided training session with a shortened block of 14 tones as well as a
practice block of 25 tones. With four trial blocks and six potential correct responses per block,
the participants could respond correctly a maximum of 24 times. The proportion of correct
responses of this total served as the primary measure.
Letter memory task. In the letter memory task (adapted from Morris & Jones, 1990), several letters from a list were
presented serially for 2000 ms per letter. The task was simply to recall the last 4 letters presented in the list. To ensure
that the task required continues updating, the instructions required the participants to rehear out loud the last 4 letters
by mentally adding the most recent letter and dropping the 5th letter back and then saying the new string of 4 letters
out loud. For example, if the letters presented were ''T, H, G, B, S, K, R,'' the participants should have said, ''T . . .
THG.
. . . TH . . THGB. . . HGBS. . . GBSK. . .
BSKR'' and then recalled ''BSKR'' at the end of the trial. The number of letters presented (5, 7,
9, or 11) was varied randomly across trials to ensure that participants would follow the
Machine Translated by Google
64 MIYAKE ET AL.

instructed strategy and continuously update their working memory representations until the end of each trial. After
practicing on 2 trials with 5 and 7 letters, respectively, the participants performed 12 trials for a total of 48 letters
recalled. The dependent measure was the proportion of letters recalled correctly.

The following three tasks were used as the Inhibition tasks: Antisaccade
task. During each trial of the antisaccade task (adapted from Roberts, Hager, & Heron, 1994), a fixation point
was first presented in the middle of the computer screen for a variable amount of time (one of nine times between
1500 and 3500 ms in 250-ms intervals).
A visual cue (0.4°) was then presented on one side of the screen (eg, left) for 225 ms, followed by the presentation
of a target stimulus (2.0°) on the opposite side (eg, right) for 150 ms before being masked by gray cross-hatching.
The visual cue was a black square, and the target stimulus consisted of an arrow inside an open square. The
participants' task was to indicate the direction of the arrow (left, up, or right) with a button press response. Given
that the arrow appeared for only 150 ms before being masked, participants were required to inhibit the reflexive
response of looking at the initial cue (a small black square) because doing so would make it difficult to correctly
identify the direction of the arrow. The cues and targets were both presented 3.4 in. away from the fixation point
(on opposite sides) and the participants were seated 18 in. from the computer monitor (thus, the total subtended
visual angle from cue to target was approximately 21.4°). The participants practiced on 22 trials and then received
90 target trials. The proportion of target trials answered correctly served as the dependent measure.

Stop-signal task. The stop-signal task (based on Logan, 1994) consisted of two blocks of trials. On each trial in
the first block of 48 trials, used to build up a prepotent categorization response, participants were presented with 1
of 24 words (eg, duck, gun), balanced for both length and frequency, and were instructed to categorize it as either
an animal or nonanimal as quickly as possible without making mistakes. Then, in the second block of 192 trials,
participants were instructed not to respond (ie, to inhibit the categorization response) when they heard a computer-
emitted tone on 48 randomly selected trials, but otherwise to keep per-forming the same categorization task as
before. As recommended by Logan (1994), the instructions emphasized that the participants should not slow down
to wait for possible signals, and whenever slowing was detected, the experimenter reminded them to continue
responding as quickly as possible. The time at which the signal occurred during the stop trial was adjusted for each
participant by taking the mean response time from the first block of trials and sub-tracting 225 ms. In all trials
(including 34 practice trials), the participants viewed a fixation point for 500 ms and were then allowed up to 1500
ms to categorize the target word by button press. The dependent variable for this task was the proportion of
categorization responses for the stop trials.

Stroop task. In the Stroop task (Stroop, 1935), adapted for computer administration, participants were instructed
to verbally name the color of a stimulus as quickly as possible in each trial, with RTs measured by voice key. The
task included 72 trials with a string of asterisks printed in one of six colors (red, green, blue, orange, yellow, or
purple), 60 trials with a color word printed in a different color (eg, BLUE printed in red color), and 12 trials with a
color word printed in the same color (eg, BLUE in blue color), with the different trial types mixed (ie, nonblocked).
The participants also received three short blocks of approximately 10 trials a piece for voice-key calibration and
practice. The dependent measure was the RT difference between the trials in which the word and the color were
incongruent and the trials that consisted of asterisks.

We also administered the following five complex executive tasks: Wisconsin Card
Sorting Test. We used a computerized, speeded version of the WCST developed by Kimberg, D'Esposito, and
Farah (1997). The task required participants to match a series of target cards presented individually in the middle
of the screen with any one of four reference cards shown near the top of the screen. Participants were instructed
to sort the target cards into piles under the reference cards according to one of three categories or stimulus attributes
—color (red, green, blue, or yellow), number (1, 2, 3, or 4), or shape (circle, cross, star, or square)—and were also
told that only one attribute was correct for each target card.
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 65

Each target card appeared until a response was given or for a maximum of 3 s, at which point the
next trial began immediately and the participants received visual feedback (ie, RIGHT or WRONG
appeared below the sorted target card). If the participant did not categorize the target card within
this time constraint, the phrase TIME OUT appeared to the right of the target card in the ensuing
trial. The category (eg, ''color'') stayed the same until the participant correctly performed eight
consecutive sorts, at which point the sorting criterion changed (eg, to ''number''). The participants
were aware that the sorting criterion would change, but they were not explicitly told the exact
number of correctly sorted cards to be achieved before the criterion shifted. After practicing on 30
cards, the main task began and continued until either the participant had successfully achieved 15
sorting categories or the total number of target cards exceeded 288. The main dependent measure
was the number of classical perseverative errors, which was the number of times participants failed
to change sorting principles when the category changed and kept sorting the cards according to
the previous, no longer correct sorting principle.

Tower of Hanoi. In this computerized version of the TOH task, participants were first shown an
ending configuration on a piece of paper, consisting of four disks of varying size positioned on
three pegs, and were given as much time as necessary to study the configuration. When ready,
the participants were shown a different starting configuration on the computer screen and were
instructed to make the starting configuration look like the ending configuration by moving the on-
screen disks with the computer mouse. The instructions emphasized that the participants were to
minimize both the number of moves and the time necessary to accomplish this reconfiguration.
When moving the disks, the participants were required to follow a set of rules commonly imposed
on the TOH task (ie, only one disk can be moved at a time, each disk must be placed on one of
the pegs, and a larger disk can never be placed be on top of a smaller disk). Prior to completing
the two target problems, the participants practiced on an easy two-disk problem and then on two
four-disk problems that each took a minimum of 11 moves to complete (1 tower-ending and 1 flat-
ending). The participants then performed two target four-disk problems that each required a
minimum of 15 moves to complete (1 tower-ending and 1 flat-ending). All problems were taken
from Humes, Welsh, Retzlaff, and Cookson (1997). The dependent measure for this task was the
total number of moves taken to complete the two target problems.

Random Number Generation. In the RNG task, participants heard a computer-generated beep
every 800 ms. Their task was to say aloud a number from 1 to 9 for each beep such that the string
of numbers would be in as random an order as possible. As an illustration of the concept of
randomness (with replacement), the participants were given the analogy of picking a number out
of a hat, reading it out loud, putting it back, and then picking another.
The importance of maintaining a consistent response rhythm was also emphasized during the
instructions, and participants received a brief practice period consisting of 10 beeps.
The valid responses generated during 162 beeps were analyzed using Towse and Neil's (1998)
RgCalc program, which produces many different indices that have been commonly used in the
analysis of ''randomness.'' The measures we initially derived from the data were the turning point
index (TPI), total adjacency (A), runs, Evan's random number generation score (RNG), Guttman's
null-score quotient (NSQ), redundancy (R), coupon score, mean repetition gap (mean RG),
median repetition gap (med RG), mode repetition gap (mode RG; when there were multiple modes,
the smallest was used), phi indices 2 through 7 (phi2–phi7), and analysis of interleaved digrams
(RNG2) (see Towse and Neil for full descriptions of these measures). Because Towse and Neil
argue that these measures tap different aspects of ran-domness, we used a principal components
analysis to reduce the data (with a Promax rotation to allow for correlated factors). More information
about the dependent measures that went into the SEM analyzes for this task is provided under
Results.
Operation span task. In each trial of the operation span task (adapted from Turner & Engle,
1989), participants received a set of equation–word pairs on the computer screen. For each pair,
participants read aloud and verified a simple math equation (eg, for (3 ÿ 4) – 6 5, participants said
''three times four minus six equals five . . . false'') and then read aloud a
Machine Translated by Google
66 MIYAKE ET AL.

single presented word (eg, ''king''). At the end of the trial, the participants recalled all of the words from
the entire set of equation–word pairs, with the instructions stipulating that the word from the last pair
presented should not be recalled first. For example, if there were four sets of equation–word pairs in the
trial, the participants would alternately verify the equation and say the word for each pair and then recall
four words at the end of the trial.
Each equation remained onscreen until either a verification response was given, at which point the
experimenter immediately pressed a response button, or for a maximum of 8 s. Once the equation
disappeared, the word was presented for 750 ms before the next equation was dis-played. The
participants were instructed to begin reading aloud each equation as soon as it appeared and were not
allowed any additional time beyond that needed to solve the equation so that the time for idiosyncratic
strategies such as rehearsal was minimized. After practicing on three trials at set size 2 (ie, two
equation–word pairs), participants performed four target trials at each set size from 2 to 5. The total
number of words recalled correctly (maximum of 56) served as the dependent measure.

Dual task. The dual task required the simultaneous performance of a spatial scanning task (the Maze
Tracing Speed Test, developed by Ekstrom, French, Harman, & Dermen, 1976) and a verbal task (word
generation). Participants first completed as many mazes as possible in 3 min, with instructions to avoid
retracing any lines or removing the pencil from the paper.
Next, participants completed the word generation task for 3 min. In this task, participants were auditorily
presented with a letter every 20 s and instructed to generate as many words as possible that began
with that letter, avoiding proper nouns and function words. In the final dual task condition, participants
performed the maze tracing and word generation tasks simultaneously for 3 min. The letters used for
the word generation task in the individual and dual task conditions were approximately balanced for the
total number of dictionary pages for the letters. Following Baddeley et al. (1997), we used the average
proportion of decrement ob-served in performance from the individual tasks to the dual task, calculated
by the following equation:

Mazesingle Mazedual Word Generationsingle Word Generationdual

Mazesingle Word Generationsingle
.
2

General Procedure
Testing took place in two sessions, administered individually during a 2-week period. Each session
lasted approximately 1.5 h, for a total of 3 h. The stimuli in each of the tasks were balanced for relevant
parameters (eg, an equal number of true/false answers) when appropriate, and the order of the trials
within each task was prerandomized and then fixed for all participants. Also, the order of task
administration was fixed for all participants (with the constraint that no two tasks that were supposed to
tap the same executive function occurred consecutively) to minimize any error due to participant by
order interaction. The tasks administered in Session 1 were (in the order of administration) antisaccade,
number–letter, keep track, stop-signal, local–global, Stroop, and letter memory. Those administered in
Session 2 were (again in the order of administration) plus–minus, tone monitoring, operation span,
RNG, TOH, dual task, and WCST.

Transformations and Outlier Analysis

The distributions of the RT and proportion correct measures for the nine tasks designed to tap the
three target executive functions were skewed and/or kurtotic, requiring transformations to achieve
normality. For RT measures with multiple trials (only correct trials longer than 200 ms were analyzed),
we performed a two-stage trimming procedure. First, upper and lower criteria were determined on the
basis of overall, between-subjects RT distributions, and any
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 67

extreme outliers were replaced with those criterion values. The lower and upper criteria values used in this first
stage of trimming were 300 and 3500 ms for the number–letter task, 500 and 3500 ms for the local–global task,
and 400 and 2000 ms for the Stroop task, respectively.
Next, the within-subject RT distributions were examined for any RTs that were more than 3 standard deviations
(SDs) away from each individual's mean RT for that task, and these observations were replaced with RTs that
were 3 SDs away. Because this trimming procedure could not be applied to the plus–minus task (only one trial per
condition), the RTs in each condition of this task were trimmed by examining the entire between-subjects
distribution and then replacing observations farther than 3 SDs from the mean with a value that was 3 SDs from
the mean. No more than 2.2% of the observations were affected by these trimming procedures in any of the RT-
based tasks. For proportion correct measures, we applied an arcsine transformation, which is useful for creating
more dispersion in ceiling and floor effects, while having little influence for accuracy scores in the range of .20–.80
(Judd & McClelland, 1989). All of the RT and proportion correct measures achieved a satisfactory level of normality
after this trimming/transformation process (see Table 1 for the skewness and kurtosis statistics).

Because the CFA and SEM techniques are sensitive to extreme outliers and careful screening is recommended
(Kline, 1998), we performed bivariate outlier analyzes on the correlations among the nine tasks designed to tap
the three target executive functions.4 Specifically, outliers were identified by computing leverage, studentized t,
and Cook's D values, which assess how much influence a single observation has on the overall results (Judd &
McClelland, 1989).
The effects of removal for any participants with very large values for these statistics (ie, levers greater than .05, t
values greater than |3.00|, or Cook's D values that were much larger than those for the rest of the observations)
were determined for each within-construct correlation, and the data for that participant were excluded from
analysis only if removal substantially changed the magnitude of the correlations. Only two participants were
removed due to these analyses, both of whom greatly affected the correlations within the Inhibition construct.

Statistical Analysis
All of the CFA and SEM analyzes reported below were performed with the CALIS procedure (SAS Institute,
1996), a program that uses the maximum likelihood estimation technique to estimate the specified latent variable
loadings, based on the covariance matrix. For ease of interpretation, the directionality of the dependent measures
was adjusted so that larger numbers always indicated better performance.

In both CFA and SEM, we evaluated the fit of each model to the data by examining multiple fit indices: the ÿ2
statistic, Akaike's Information Criterion (AIC), the standardized root mean-squared residual (SRMR), Bentler's
Comparative Fit Index (CFI), and Bollen's Incremental Fit Index (IFI, also referred to as BL89). We selected these
fit indices because they represent different types: absolute fit indices (AIC and SRMR) as well as Type 2 (IFI) and
Type 3 (CFI) incremental fit indices. Within these classes of fit indices, the ones we selected were those considered
sensitive to model misspecification (ie, models that lack necessary parameters or cluster the variables
inappropriately) while at the same time being relatively insensitive to small sample sizes (ie, N 150; Hu & Bentler,
1995, 1998).

The most common index of fit is the ÿ2 statistic, which measures the ''badness of fit'' of the model compared to
a saturated model. Because the ÿ2 statistic measures the degree to which the covariances predicted by the
specified model differ from the observed covariances,

4
Because such bivariate outlier analyzes were not possible for the five complex executive tasks used in the
SEM analyzes (one cannot calculate such statistics with latent predictor variables), we examined these additional
variables for univariate outliers and replaced observational observations farther than 3 SDs from the mean with a
value that was 3 SDs from the mean. This procedure affected no more than 2.3% of the observations for each
task.
Machine Translated by Google
68 MIYAKE ET AL.

a small value for the ÿ2 statistic indicates no statistically meaningful difference between the covariance matrix
generated by the model and the observed matrix, suggesting a satisfactory fit. The SRMR also assesses ''badness
of fit,'' as it is the square root of the averaged squared residuals (ie, differences between the observed and predicted
covariances). Lower values of the SRMR indicate a closer fit, with values less than .08 indicating a relatively close
fit to the data (Hu & Bentler, 1998). The other fit indices (AIC, CFI, and IFI) are typically used to measure ''goodness
of fit.'' AIC is a modified version of the ÿ2 statistic that takes into consideration the ''complexity'' of the evaluated
model (in terms of degrees of freedom) and penalizes more complex models (ie, models with fewer degrees of
freedom). Lower values of AIC (including negative values) indicate better fit. In contrast, for CFI and IFI, higher
values indicate better fit, as these indices quantify the extent to which the tested model is better than a baseline
model (eg, one with all covariances set to zero). Typically, IFI and CFI values that exceed .90 or .95 are considered
good fits (the values of IFI can exceed 1.0).

In addition to these commonly used indices, we also examined specific indications of fit, such as the magnitudes of
asymptotically standardized residuals, in comparing different alternative models. None of the models we endorse
in our discussion of the CFA and SEM results had large residuals, according to the criteria recommended by
Jo¨reskog and So¨rbom (1989).
To examine if one model was significantly better than another, we conducted ÿ2 difference tests on ''nested''
models. This test entails subtracting the ÿ2 for the fuller model from the ÿ2 for the nested model with a fewer
number of free parameters or larger degrees of freedom (degrees of freedom are calculated with an analogous
subtraction). If the resulting ÿ2 is statistically significant, then the fuller model provides a significantly better fit. For
these and all other statistical tests reported here, we used an alpha level of .05.

RESULTS AND DISCUSSION

Preliminary Data Analysis A

summary of descriptive statistics for the nine measures used to tap the three
target executive functions (ie, Shifting, Updating, and Inhibition) is presented in
Table 1. All of the measures had relatively low skewness and kurtosis
coefficients, and the normalized multivariate kurtosis (Mardia, 1970) for all nine
measures was also quite low: .19 (the normalized multivariate kurtosis did not
exceed .95 for any of the SEM analyses). Internal reli-ability estimates for the
tasks used in the CFA were calculated using either Cronbach's alpha or the
split-half (odd–even) correlation adjusted by the Spearman-Brown prophecy
formula. As seen in Table 1, the reliability estimates for the tasks were relatively
low (except for number–letter and stop-signal), a characteristic often reported
for executive tasks (Denckla, 1996; Rabbitt, 1997b).

The zero-order correlations among the nine measures, provided in Appen-

dix A, were generally low (.34 or lower), consistent with the results from previous
individual differences studies of executive functions. It is important to point out,
however, that the correlations among the nine measures were not uniformly
low; rather, the tasks considered to tap the same executive function tended to
show significant correlations with one another, while correlating not as strongly
with the tasks considered to tap the other executive functions, thus showing
some signs of convergent and discriminant validity.
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 69

TABLE 1
Descriptive Statistics for the Dependent Measures Used in the Confirmatory Factor
Analysis and Structural Equation Models (N 137)

Task Mean (SD) Range Skewness Kurtosis Reliability

Plus–minus 15.5 s (10.8) 7.2 to 51.7 .60 .57 N/Aa

Number–letter 546 ms (250) 45 to 1303 .29 .29 .91b
Local–global 210 ms (160) .63 289 to 709 .56 1.06 .59b
Keep track (.14) [.58 .22 to .95 .06 0.21 .31c
(.11)] .70 [.22 to .81]
Tone monitoring (.26) [.62 .17 to 1.57 .38 .36 .63c
(.18)] .99 [.17 to 1.00]
Letter memory (.13) [.83 .65 to 1.37 .35 0.07 .42c
(.07)] 1.16 [.60 to .98]
Antisaccade (.16) [.91 .69 to 1.57 0.24 .27 .77b
(.07)] .78 [.63 to 1.00]
Stop-signal (.29) [.67 .02 to 1.57 0.08 0.27 .92b
(.20)] 166 [.02 to 1.00]
Stroop ms (60) 50 to 315 .27 0.65 .72b

Note. The data analyzes used arcsine-transformed proportion measures and trimmed RTs.
For the proportion data, the raw proportion statistics are in brackets. a Reliability
could not be calculated for this task because there was only one RT per condition.

b
Reliability was calculated by adjusting split-half (odd–even) correlations with the Spear-
man-Brown prophecy formula. c
Reliability was calculated using Cronbach's alpha.

This pattern suggests that the measures used to tap each target executive
function may have indeed tapped a common underlying ability or func-tion.

To What Extent Are the Three Target Executive Functions Separable?

The first main question we asked in the study was: Are the three targets
executive functions (ie, Shifting, Updating, and Inhibition) distinguishable,
or do they essentially tap the same underlying construct? We addressed this
question with CFA.
The logic of the analysis was as follows: If the three target executive
functions are distinguishable constructs, then the full three-factor model de-
pictured in Fig. 1A should provide a significantly better fit to the data than
either the model that assumes the unity of all three executive functions
(called the ''one-factor'' model) or the models that assume the unity of two
of the executive functions (called the ''two-factor'' models). If the three
executive functions actually are completely unitary and essentially the same
construct, then the one-factor model should provide a fit to the data that is
statistically no worse than the more complex three-factor or two-factor mod-els.
Finally, if the three functions are entirely separate, then the ''three independent
Machine Translated by Google
70 MIYAKE ET AL.

FIG. 2. The estimated three-factor model. Single-headed arrows have standardized factor
loading next to them. The loadings, all significant at the .05 level, are equivalent to standard-ized regression
coefficients (beta weights) estimated with maximum likelihood estimation.
The numbers at the ends of the smaller arrows are error terms. Squaring these terms gives
an estimate of the variance for each task that is not accounted for by the latent construct. The
curved, double-headed arrows have correlation coefficients next to them and indicate significant correlations
between the latent variables.

pendent factors'' model, in which all the interfactor correlations are set to
zero, should provide a fit to the data similar to that of the model in which
the correlations are allowed to vary freely.
The full three-factor model, complete with the estimated factor loadings,
is illustrated in Fig. 2. The numbers next to the straight, single-headed arrows
are the standardized factor loadings, and those next to the curved, double-
headed arrows are the correlations between the factors. In addition, the num-
bers at the ends of the smaller, single-headed arrows represent the error
terms. Squaring these error terms gives an estimate of the unexplained vari-
ance for each task, which could be attributed to idiosyncratic task demands
and measurement error. Note that all the factor loadings listed in Fig. 2 are
equivalent to standardized regression coefficients and can be interpreted
accordingly.
The fit indices for this full three-factor model, summarized in Table 2
(Model 1), were all excellent. Specifically, this model produced a nonsig-
nificant ÿ2 (24, N 137) 20.29, p .65, indicating that the model's predictions
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 71

TABLE 2
Fit Indices for the Full Confirmatory Factor Analysis Model and Reduced Models
(N 137)

Model df ÿ2a AICb SRMRb CFIc IFic

1. Full three-factor 2. 24 20.29 27.71 .047 1.00 1.04

One-factor Two- 27 36.17 17.83 .065 .89 .90
factor models
3. Shifting Updating 4. 25 29.35 20.65 .057 .95 .95
Shifting Inhibition 5. Updating 25 29.17 20.83 .060 .95 .96
Inhibition 6. Independent three 25 24.29 25.71 .052 1.00 1.01
factors 27 47.03* 6.97 .115 .76 .79

Note. The endorsed model is indicated in bold. AIC, Akaike's Information Criterion; SRMR,
standardized root mean-squared residual; CFI, Bentler's Comparative Fit Index; IFI, Bollen's
Incremental Fit Index.
a ÿ2 that were not significant at the .05 level indicate that those models provided reasonable
fits; However, all ÿ2 difference tests indicated that the reduced models (2–6) provided significantly worse fits
than the full model (1).
b
Lower values of AIC and SRMR indicate better fit, with SRMR .08 indicating a close
fit to the data.
c Values above .95 for CFI and IFI indicate good fit.
*
p .05.

tions did not significantly deviate from the current data pattern. In addition, the
values of the AIC and SRMR were quite low (–27.71 and .047, respectively),
whereas the IFI and CFI were well above .95 (1.04 and 1.00, respectively).
Thus, this full three-factor model seems to fit the overall data quite well.
One important issue is whether the three latent variable factors could actually
be considered to be measuring the same underlying ability. As shown
in Fig. 2, the estimates of the correlations among the three latent variables
were moderate, ranging from .42 to .63. The 95% confidence intervals for
the correlations were [.29, .84] for the Updating and Shifting factors, [.30,
.96] for the Updating and Inhibition factors, and [.09, .76] for the Shifting
and Inhibition factors, respectively. Because none of these intervals contain
1.0, we can reject the hypothesis that any pair of the three latent variables
factors is in fact the same construct.
This conclusion was further supported by the direct statistical comparison
of alternative models. We first tested the hypothesis that the three executives
functions are not completely unitary by creating a one-factor model that as-
sumes complete unity of the three target executive functions and comparing
it against the full three-factor model depicted in Fig. 2. Table 2 summarizes
the fit indices for this one-factor model (Model 2), which we created by
fixing the correlations among the three latent variable factors at 1.0 (ie,
perfect correlation). The values of the indices were all poorer than the full
three-factor model. The AIC and SRMR were relatively high (–18 and .065,
respectively), and the IFI and CFI were lower than .95 (.89, and .90, respectively).
Machine Translated by Google
72 MIYAKE ET AL.

tively). In addition, the ÿ2 difference test produced a significant result, ÿ2(3)

15.88, p .01, suggesting that the one-factor model fit the data significantly
worse than the three-factor model did and hence must be rejected.
We also estimated three nested two-factor models in which two of the
three executive functions were assumed to be the same. In these models, two
of the correlations among the three latent variable factors were allowed to
vary and the remaining correlation was set to 1.0. Even though all the fit
indices were respectable for these two-factor models (Models 3–5 in Table
2), the ÿ2 difference tests showed that the full three-factor model provided
a significantly better fit than any of the three two-factor models, all ÿ2 (1)
4.00, p .05. In other words, none of the correlations among the latent
variables could be set to 1.0 without significantly worsening the fit of the
model. These findings further support the notion that the three hypothesized
constructs are indeed separable.
In addition to the above comparisons, we also compared the full three-
factor model to a reduced ''three independent factors'' model in which all
of the correlations among the latent variables were set to zero (ie, the model
in which the three target executive functions are assumed to be completely
independent of one another). The resulting fit indices for this model, shown
in Table 2 (Model 6), were poor, including a significant ÿ2 (p .05, indicating
an unsatisfactory overall fit). The ÿ2 difference test also indicated that
the three independent factors model provided a significantly worse fit than
the full three-factor model, ÿ2(3 26.74, p .001, suggesting that the three
executive functions share at least some commonality and cannot be
considered completely independent.
Taken together, these CFA results suggest that, even though they are
clearly distinguishable, the three latent variables share some underlying
commonality. Thus, the three target executive functions show signs of both unity
and diversity, a point that we consider in more detail under General Discussion.

Which Executive Function(s) Do Complex Executive Tasks Really Tap?

After establishing some separability of the three target executive functions
(ie, Shifting, Updating, and Inhibition) with CFA, we examined the extent
to which these functions contribute to performance on more complex
executive tasks—WCST, TOH, RNG, operation span, and dual tasking—by
per-forming a series of SEM analyses. The descriptive statistics for these
com-plex tasks are listed in Table 3. Although different proposals have been made
regarding what each of these complex tasks really tap, such proposals have
not been independently tested in previous neuropsychological or individual
Differences studies of executive functions and remain highly speculative. in
the SEM analyses, we explicitly tested the previously suggested accounts of
what these executive tasks really measure.
The logic of the analysis is similar to that used in the CFA and centers
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 73

TABLE 3
Descriptive Statistics for the Dependent Measures Used in the Structural Equation Models
(N 137 Unless Noted)

Task Mean (SD) Range Skewness Kurtosis

WCST 32 (12) 15 to 67 .91 .22

Perseverance
Tower of Hanoib 46 moves (12) 30 to 86 1.40 1.94
RNG
Component 1 0 (1) 3.22 to 2.06 0.88 1.07
Component 2 0 (1) 3.02 to 1.66 0.73 .34
Component 3 0 (1) 2.10 to 3.22 .74 .76
Operation span 43 words (6) 30 to 55 .01 0.70
Dual taska .89 (.13) .52 to 1.27 .22 .25

Note. Extreme observations for each task were trimmed to be 3 SDs from the mean. a N 134. b

N 136.

around the comparisons of alternative models. On the basis of previous pro-

posals, we first selected, for each executive task, models that included specific
paths from only one or, in some cases, two latent variables. These a
priori models were then compared against a ''full'' model that included paths
from all three latent variables (illustrated in Fig. 1B). As was the case with
the CFA, a hypothesized ''reduced'' model is considered good if the fit indices
meet the standard criteria and if a ÿ2 difference test indicates that the
model's fit is not statistically worse than the fit of the full model. If multiple
hypothesized models are considered good for any given task, the more parsi-
monious model should be preferred. In addition to the full and hypothesized
reduced models, we also estimated a ''no-path'' model, which included no
paths from the latent variables to the executive task of interest. The preferred
hypothesized model should provide a significantly better fit than this no-path
model if the executive task is related to any of the three executive functions.
For every model tested, we allowed all of the factor loadings into the latent
variables and interfactor correlations to vary (Anderson & Gerbing, 1988).
Hence, the estimated parameters could differ from the values found for the
original CFA (ie, those presented in Fig. 2). We allowed these parameters
to vary (rather than fixing them at the values obtained in the CFA) because,
in addition to examining the path coefficients from the three latent variables
to each target executive task, we also wanted to test the stability of the three-
factor structure supported by the CFA. If the three-factor CFA model we
endorsed earlier is somehow misspecified and the underlying factor structure
is rather unstable, then adding an extra executive task in the model may
cause major distortions to the original factor structure (ie, major changes in
the factor loadings of the individual executive tasks and/or in the correlations
among the three latent variables). In turn, if the paths from the latent variables
to the added complex executive tasks are grossly misspecified, then system-
Machine Translated by Google
74 MIYAKE ET AL.

atic changes from the original factor structure may reveal the nature of the misspecification,
as we will discuss in the case of the operation span task results. We should emphasize,
however, that, for the structural model we endorse for each executive task, the changes in
the parameters were generally quite small. In fact, across the endorsed SEM models, the
individual factor loadings and the latent variable correlations showed an average change of
only .02 and .03, respectively. Such stability in the parameters across differ-ent SEM models
suggests that the factor structure depicted in Fig. 2 was highly reliable, giving further
credence to the results of the CFA.

Below, we describe the results of our SEM analyzes for each complex executive task. For
each task, we first provide a brief description of the pro-posals made previously about what
executive function(s) that task may tap and then discuss the results of the SEM analyzes that
specifically tested these proposals. In the tables that summarize the results for individual
complex executive tasks (Tables 4–8), we report the standardized path coefficients as well
as the ÿ2 statistic, the SRMR, and the IFI for each model we tested.

We reduced the number of fit indices in the tables for brevity, but the other indices reported
for the CFAs all showed similar results. The model we en-dorse for each executive task is
highlighted in boldface.
Wisconsin Card Sorting Test. Despite the finding that it is sensitive to some impairments
that do not necessarily involve the frontal lobes (Ander-son, Damasio, Jones, & Tranel, 1991;
Dunbar & Sussman, 1995; Reitan & Wolfson, 1994), the WCST is perhaps the most frequently
used test of executive functions in the neuropsychological populations (particularly patients
with frontal lobe damage). It has also been successfully used (often with some slight
modifications) among normal populations (eg, Kimberg et al., 1997; Lehto, 1996; Levin et al.,
1991). In the literature, the WCST is often conceptualized as a set shifting task (eg, Berg,
1948) because of its requirement to shift sorting categories after a certain number of
successful trials, although this idea has not been independently tested before the best of our
knowledge. In addition, several researchers have considered the hypothesis that the task
requires inhibitory control to suppress the current sorting category and switch to a new one
(eg, Ozonoff & Strayer, 1997).

For these reasons, we evaluated the hypothesis that Shifting or Inhibition or both would
predict WCST performance by testing the two-path model with paths from Shifting and
Inhibition as well as two one-path models with the path from either Shifting or Inhibition. Given
the emphasis in the literature on the set shifting requirement of this task as well as the fact
that the task does not build up a strong prepotent response before a shift is required, we
expected that either a two-path model with both paths from Shifting and Inhibition or a model
with only one path from Shifting would provide the best fit.

We performed our SEM analyzes on the number of perseverative errors,

Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 75

TABLE 4
Fit Indices and Standardized Regression Coefficients for Structural Equation Models with
Wisconsin Card Sorting Test Classic Perseverations (N 134)

Coefficients for specified paths

Model df ÿ2 SRMR IFI Shifting Updating Inhibition

1. Full three paths 30 25.02 2. Two paths from .048 1.05 .33 .01 .09
Shifting 31 25.02 and Inhibition .048 1.06 .33* — .10

3. One path from Shifting 32 25.45 4. One path .048 1.06 .38* — —
from Inhibition 32 30.59 5. No paths 33 37.27 .056 1.01 — — .33*
.076 .96 — — —

Note. The endorsed model is indicated in bold.

p . 10.
*
p .05.

the measure often considered most sensitive to frontal lobe dysfunction.5 The
results are summarized in Table 4. The ÿ2 difference tests indicated that
the two-path model with both paths from Shifting and Inhibition (Model 2)
provided as good a fit as the full three-path model (Model 1), ÿ2 (1) 0.00,
p .10. It also produced a significantly better fit than both the no-path model
(Model 5), ÿ2 (2) 12.25, p .01, and the one-path model with a path from
Inhibition (Model 4), ÿ2 (1) 5.57, p .05, suggesting that this two-path
model provided a good overall fit to the data. However, this model with
paths from both Shifting and Inhibition (Model 2) were not statistically better
than the one-path model with only a single path from Shifting (Model 3),
ÿ2(1) .43, p .10, indicating that the path from Inhibition was really not
making much contribution to the prediction of WCST perseverations once
Shifting ability had been taken into account. This conclusion is also corroborated
by the fact that the full three-path model had a marginally significant
coefficient for Shifting (.33), but a much lower one for Inhibition (.09). Thus,
taken together, the results from the perseveration measure suggest that the
one-path Shifting model is the most parsimonious one and supports the con-
clusion that the Shifting ability is a crucial component of perseverative errors
in the WCST, at least in this sample.
As the existing literature suggests (eg, Anderson et al., 1991; Reitan &
Wolfson, 1994), the WCST is clearly a complicated task that taps various

5
We also examined another dependent measure, the total number of trials necessary to
achieve 15 categories, which is analogous to a standard clinical measure of the number of
categories achieved within a fixed number of cards. Because the correlation between this
measure and the perseveration measure was high (r .79) and showed essentially the same
pattern of results, we report only the SEM analyzes from the perseveration measure in this
article.
Machine Translated by Google
76 MIYAKE ET AL.

cognitive processes and hence cannot be considered selectively sensitive to

frontal lobe impairments per se. For example, Dunbar and Sussman (1995)
have shown that an impairment in the phonological loop, which could arise
from posterior lesions in the left hemisphere, may also lead to perseverations
on the WCST by making it difficult to keep the current category highly
accessible in memory. Despite the complexity of the WCST as a task, the
current results from the SEM analysis demonstrate that the WCST indeed
taps, at least in part, one aspect of executive functioning, Shifting, suggesting
that it may still serve as a useful executive task if proper caution is taken.
Tower of Hanoi. The TOH puzzle, along with the similar Tower of London
puzzle, is frequently described as tapping a ''planning'' ability (eg, Arnett et
al., 1997), an ability that involves mapping out a sequence of moves in
preparation for the task (Morris, Miotto, Feigenbaum, Bullock, & Polkey,
1997). Despite this prevalent conception, the extent to which participants
actually do careful planning in this task is unclear, at least when it is
administered without specification of what strategies to use (as is usually the
case with neuropsychological testing of patients). Indeed, a detailed analysis
of strategies has shown that multiple strategies can be used in solving the
TOH puzzle (Simon, 1975).
According to this analysis, the strategy that is perhaps most closely related
to the notion of ''planning'' and the one actually guaranteed to solve the puzzle
in the minimum number of moves is the goal-recursion strategy, used in some
previous studies of the TOH puzzle (Carpenter, Just, & Shell, 1990).
This strategy involves extensive goal management and requires setting up a
series of subgoals (which, in essence, constitute multiple smaller TOH
puzzles with fewer disks) to achieve the superordinate goal. Despite the
elegance of this strategy, it is highly demanding, as it requires maintaining a
stack of subgoals in working memory. An alternative strategy that is used
more prevalently is the so-called perceptual strategy, which involves simply
making a next move that will bring the current state perceptually closer to the
goal state. This perceptual strategy is much less demanding, and studies of
the TOH have demonstrated that most people tend to favor and spontaneously
adopt the perceptual strategy in the usual implementation and administration
of this task (Goel & Grafman, 1995).
On the basis of this evidence, our main prediction for the TOH puzzle
performance was that the Inhibition factor may play an important role be-
cause, when one is using the perceptual strategy, the major difficulty seems
to come from moves that involve ''goal–subgoal conflicts.'' These conflicts
occur when the optimal action requires moves that take the disk configuration
temporarily further away from the goal state—namely, moves that require one
to transfer a disk in the opposite direction as the end goal (Morris et al., 1997)
and/or to block the goal peg with a disk that must later be cleared (Goel &
Grafman, 1995; Simon, 1975). Making these counterintuitive ''con-flict moves''
likely involves overcoming the natural tendency to make more
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 77

TABLE 5
Fit Indices and Standardized Regression Coefficients for Structural Equation Models with
Tower of Hanoi: Total Number of Moves for the 15-Move Problems (N 136)

Coefficients for specified paths

Model df ÿ2 SRMR IFI Shifting Updating Inhibition

1. Full three paths 30 24.10 .048 1.06 0.14 2. One path from Inhibition 32 24.78 .14 .33
.049 1.07 — — .37* 3. No paths 33 34.87 .071 .98 — — —

Note. The endorsed model is indicated in bold. *

p .05.

obvious, perceptually congruent moves, hence requiring the Inhibition ability

(Goel & Grafman, 1995). Given that we intentionally avoided constraining the
participants' strategies in the present study's implementation of the TOH puzzle
(to simulate typical neuropsychological administrations of this task), we
expected that most participants would use the perceptual strategy and hence
that the Inhibition factor would play some role in predicting the num-ber of
moves they took to solve the target problems.
We tested this hypothesis by estimating a one-path model with a path from
the Inhibition factor, along with the three-path and no-path models.
The results, summarized in Table 5, indicated that the Inhibition path model
(Model 2) provided as good a fit to the data as the three-path model (Model 1),
ÿ2 (2) .68, p .10, and a significantly better fit than the no-path model (Model
3), ÿ2 (1) 10.09, p .01. In addition, the other one-path models we tested (ie,
the ones that include the path from Shifting or Updating, respectively) were
not as good as the Inhibition model, a point corroborated by the path
coefficients in the full, three-path model (Model 1; .14 for Shifting, .14 for
Updating, and .33 for Inhibition). Thus, we found evidence to support the
hypothesis that Inhibition contributes to performance on the TOH puzzle.

To examine the extent to which ''conflict moves'' were indeed responsible

for TOH performance, we analyzed the optimal solution paths for the two
target problems used in this study. The main finding was that, across both
problems, 26% of the moves involved moving a disk in the opposite spatial
direction from its ultimate goal and 33% required blocking a goal peg (these
two types of conflict moves were not completely independent). Thus, conflict
moves constituted a substantial proportion of the moves required for the
optimal solution. In addition, we also analyzed each participant's solution paths
for the two target problems to assess where the first deviation from the optimal
solution occurred, if any deviations did occur. Across both prob-lems, 55% of
the first errors occurred on moves requiring moving a disk spatially away from
its ultimate goal, and 68% occurred on moves that re-
Machine Translated by Google
78 MIYAKE ET AL.

wanted blocking a goal peg. These results indicate that the first moves to
''trip up'' our participants (and, hence, to cause longer solution paths) were
most often conflict moves. Although this analysis is rough and only assessed
the first error each participant made on each problem (after the first deviation
from the optimal sequence, the solution paths differed, making it difficult
to compare across participants), the results still support the view that
the ability to inhibit the tendency to make perceptually congruent yet incorrect
moves is a crucial component of TOH performance.
The important role of the Inhibition ability in solving the TOH puzzle
suggests that, at least in its typical method of administration that encourages
the perceptual strategy, the TOH should not be conceptualized as a ''plan-
ning'' task (Goel & Grafman, 1995). According to recent research (Murji &
DeLuca, 1998), this conclusion may also generalize to an analogous Tower
of London task, which has also been widely used as a ''planning'' task. We
should emphasize, however, that if participants were to use a more de-
manding strategy that requires more extensive goal management (as in the
case of the goal-recursion strategy; Carpenter et al., 1990), then the TOH
task might be related less strongly to the Inhibition factor and more strongly
to the Updating factor, to the extent that Updating also applies to the
management of goal information in working memory.
Random number generation. Within the framework of Baddeley's (1986)
multicomponent model of working memory, the RNG task has been one of
the most frequently used tasks to examine the functioning of the central
executive component. Although a systematic investigation of the underlying
processes for this task has begun only recently (eg, Baddeley, Emslie, Ko-
lodny, & Duncan, 1998; Towse, 1998), several proposals have been made
Regarding what abilities or functions it really taps.
One common proposal emphasizes the importance of suppressing stereo-
typed sequences like counting (eg, 1–2–3–4) to make the produced sequence
as random as possible (eg, Baddeley, 1996; Baddeley et al., 1998),
suggesting that the Inhibition factor may play an important role. Another
proposal suggests that keeping track of recent responses and comparing them
to a conception of randomness is a central aspect of RNG (eg, Jahanshahi
et al., 1998), thus pointing to a role for the Updating factor.
These explanations are not mutually exclusive, and it is entirely possible
that both processes contribute to performance on the RNG task. In fact, a
recent analysis suggests that this might indeed be the case: Towse and Neil
(1998) performed a principal components analysis (PCA) on a set of ran-
domness indices and found that the indices loaded on multiple components.
One of the components had high loadings for the randomness indices that
seem to be to the degree to which stereotype sensitive sequences are pro-
duced, whereas another component had high loadings for the indices that
seem to assess the degree to which each number is produced equally
frequently. Towse and Neil interpreted these components as the ''prepotent
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 79

TABLE 6
Fit Indices and Standardized Regression Coefficients for Structural Equation Models with
Random Number Generation (N 137)

Coefficients for specified paths

Model df ÿ2 SRMR IFI Shifting Updating Inhibition

A. Component 1 (''prepotent associates'')

1. Full three paths 30 25.15 2. One path from Inhibition .047 1.05 .12 0.05 1.06 — — .35
32 25.60 3. No paths 33 36.97 .048 .39*
.073 .96 — — —

B. Component 2 (''equality of response usage'')

1. Full three paths 30 34.71 2. One path from Updating .058 .96 0.08 .52 0.17
32 35.52 3. No paths 33 44.21 .059 .97 — .33* —
.075 .90 — — —

Note. The endorsed models are indicated in bold.

p . 10.
*
p .05.

associates'' component and the ''equality of response usage'' component,

respectively.
We tested Towse and Neil's (1998) interpretations of these randomness
components. Specifically, we first performed a PCA (with an oblique Promax
rotation to allow for the possibility that these components might be corre-
lated) on 15 indices of randomness we derived from the data.6 A three-com-
ponent solution, shown in Appendix B, was obtained that generally replicated
Towse and Neil's results (the three components accounted for 63% of
the total variance). Component 1 was similar to Towse and Neil's ''prepotent
associates'' component, and Component 2 was similar to their ''equality of
response usage'' component. On the basis of their interpretations of these
components, we hypothesized that the present study's Component 1 should
be related to the Inhibition factor, whereas Component 2 should be related
to the Updating factor.7
We evaluated these predictions by estimating two sets of structural mod-
els, one for each of the two RNG components derived from PCA. For our
Component 1 (prepotent associates), the main model we tested had the hy-
pothesized path from the Inhibition factor. As Table 6A indicates, this one-

6
We reversed the direction of all the measures except TPI, mean RG, mode RG, and the phi
indices so that higher numbers indicate better performance (ie, more suppression of usual
counting, more equality of response usage, and less repetition avoidance). We excluded med
RG from analysis because it had only three values, thereby violating the normality assumption.
We also excluded NSQ because of its high correlation with Evan's RNG index, r(135) .90.
7
The prediction for Component 3, comprised primarily of the phi indices, is less clear. The
phi indices are complex measures that seem to quantify how reticent participants are to repeat
a number at given distances (eg, the phi4 measure indicates how many participants avoided
repeating the same number after three intervening numbers, compared to what would be
observed in a random sequence). We did not have a strong a priori hypothesis as to the underlying
Machine Translated by Google
80 MIYAKE ET AL.

path model (Model 2) produced as good a fit to the data as the full three-path
model (Model 1), ÿ2(2) .45, p .10, and a significantly better fit
than the no-path model (Model 3), ÿ2 (1) 11.37, p .001. As can be
inferred from the three coefficients in the full three-path model (Model 1)
in the table, the other single-path models (ie, the ones with the path from
the Shifting or Updating factor, respectively) did not produce as good a fit.
These results confirmed Towse and Neil's (1998) interpretation that a set of
randomness indices that load highly on this component are indeed sensitive
to one's ability to inhibit prepotent responses.
For Component 2 (equality of response usage), the main tested model had
the hypothesized path from the Updating factor. As indicated in Table 6B,
this one-path model (Model 2) produced as good a fit as the three-path model
(Model 1), ÿ2 (2) .81, p .10, and a significantly better fit than the no-path model
(Model 3), ÿ2 (1) 8.69, p .01. Take together with the finding
that the other single-path models (ie, the one with the path from the Shifting
factor or the Inhibition factor) did not provide satisfactory fits, these results
suggest that the randomness indices that load highly on the equality of re-
sponse usage component are indeed sensitive to one's ability to update and
monitor information in working memory.
These results provide supporting evidence for the previously postulated
accounts of the processes underlying RNG. Specifically, RNG draws on mul-
tiple executive functions and requires the Inhibition ability to suppress habit-ual
and stereotyped responses as well as the Updating ability to monitor
response distribution.8 This multidimensionality of the RNG task highlights
the necessity of using multiple randomness indices to evaluate performance
on this task, particularly depending on what aspects of executive functioning
one wishes to examine.
These conclusions are corroborated by a recent neuropsychological study
on RNG (Jahanshahi et al., 1998). This study found that transcranial magnetic
stimulation over the left dorsolateral prefrontal cortex increased the
tendency to produce sequences of numbers adjacent on the number line (an
indication of habitual counting similar to the Adjacency measure used in the
present study, which loaded on the prepotent associates component), without
having any effect on a measure of repetition distance (an indication of cy-cling
similar to the repetition gap measures used in the current study, which
loaded on the equality of response usage component). This dissociation sug-

target executive function(s) for this measure because others have noted that repetition
avoidance is relatively automatic and does not rely on a limited capacity resource (Baddeley et al.,
1998).
8
It has also been suggested that RNG may involve shifting retrieval strategies (Baddeley,
1996). To the extent that the Shifting factor taps this notion, however, the relatively low
Shifting coefficients in the full, three-factor models for both RNG factors indicate that this
function may not play a major role in performance on the RNG task.
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 81

gestures that there might be some neural basis for the separability between the Inhibition and
Updating factors, as far as performance on the RNG task is concerned.

Operation span. Along with an analogous reading span task (Daneman & Carpenter,
1980), the operation span task (Turner & Engle, 1989) has been used as a measure of
working memory capacity that strongly implies the operations of the central executive (Engle,
Tuholski, Laughlin, & Conway, 1999b) and is predictive of performance on complex cognitive
tasks, such as reading comprehension tests (Daneman & Merikle, 1996) and complex fluid
intelligence tests. (Engle et al., 1999b). Although there is no clear con-sensus as to what this
task (or other analogous working memory span tasks) really measures (Miyake & Shah,
1999), there are a number of different proposals that relate to the target executive functions
we examined.

The first, perhaps most simple possibility is that the operation span test assesses
participants' abilities to temporarily store and update in-coming information and, hence,
should be related to the Updating factor. A second possibility is that the operation span scores
are related to the Shifting factor (either instead of or in addition to the Updating factor).
Several re-searchers have pointed out that complex working memory span tasks like the
operation span test may require participants to constantly shift back and forth between the
processing and storage requirements of the task (ie, verifying equations and remembering
target words). They further suggest that the ability to efficiently shift between these
requirements may be crucial for, or at least play an important role in, performance on these
tasks (eg, Con-way & Engle, 1996; Towse, Hitch, & Hutton, 1998). According to this view, the
model that includes a path from the Shifting factor to the operation span scores should
provide a good fit to the data.

We tested these hypotheses with a two-path model (paths from Shifting and Updating) as
well as two models with paths from only one of these functions. The results are presented in
Table 7, along with the results of the three-path and no-path models for comparison. The ÿ2
difference tests indicated that the two-path model (Model 2) provided as good a fit to the data
as the full three-path model (Model 1), ÿ2 (1) .08, p .10, and a better fit than the no-path
model (Model 5), ÿ2 (2) 38.49, p .001. Note, however, that the path coefficient from Shifting
in this two-path model is negative (ie, poorer Shifting ability is associated with better operation
span scores). A close examination of the models revealed that this weak negative relationship
was likely to reflect a statistical accommodation of the fact that the relations among the
Shifting tasks and operation span scores were weaker than would be expected on purely
statistical grounds.9

9
Our statistical explanation of this negative coefficient is as follows: Because the Updating
and Shifting latent variables are moderately correlated with each other, even the model with
only a path from Updating would expect at least some slight (but not necessarily significant)
Machine Translated by Google
82 MIYAKE ET AL.

TABLE 7
Fit Indices and Standardized Regression Coefficients for Structural Equation Models with
Operation Span (N 137)

Coefficients for specified paths

Model df ÿ2 SRMR IFI Shifting Updating Inhibition

1. Full three paths 30 25.89 2. Two paths from Updating .048 1.03 0.43 .97* 0.08
31 25.97 and Shifting .049 1.04 0.42 .91* —

3. One path from Shiftinga 32 49.61* 4. One path from .073 .87 .51* — —
Updating 32 32.03 5. No paths 33 64.46* .056 1.00 — .61* —
.101 .76 — — —

Note. The endorsed model is indicated in bold.

p . 10.
*
p .05. a
This model caused two of the paths from the Shifting tasks (number–letter and local–
global) to become nonsignificant and resulted in a Heywood case (ie, a correlation 1),
indicating model misspecification.

This ''statistical accommodation'' interpretation is supported by the

observation that the model with only a path from Shifting (Model 3) provided a
much worse fit than both the full three-path model (Model 1), ÿ2(2) 23.72,
p .001, and the two-path model (Model 2), ÿ2 (1) 23.64, p .001, and
caused major distortions to the Shifting latent variable (ie, the loadings for
the number–letter and local–global tasks dropped below significance, and
the correlation between Updating and Shifting went beyond the upper limit
of 1.0, indicating model misspecification). In contrast, the model with only
a path from Updating (Model 4), while statistically worse than the three-path
model (Model 1), ÿ2 (2) 6.14, p .05, and the two-path model (Model
2), ÿ2(1) 6.06, p .05, did not cause such major distortions to the factor
structure10 and was also clearly better than the one-path Shifting model
(Model 3) in terms of the fit indices (see Table 7). Thus, based on these
statistical reasons as well as the fact that no existing theoretical proposals

correlations between the Shifting tasks and operation span scores. As the correlation matrix
presented in Appendix A indicates, however, the actual correlations were essentially 0 (ie,
only in the .04 to .09 range), thus causing the path coefficient from the Shifting variable to
be negative to accommodate this lack of expected correlations. We are not sure why the
correlations between operation span scores and the Shifting tasks were lower than statistically
ex-pected.
10
It should be noted that this one-path Updating model (Model 4) did cause some modesty
distortion to the factor structure to accommodate the lower than expected correlations between
operation span scores and the Shifting tasks. Specifically, the interfactor correlation between
Shifting and Updating dropped from a CFA estimate of .56 (see Fig. 2) to .40 in the endorsed
one-path Updating model. This distortion, however, was much more modest in magnitude
than that observed for the one-path Shifting model (Model 3) and was not accompanied by
major changes in the pattern of loadings for the nine tasks used to tap the target executive
functions.
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 83

have postulated a negative relationship between operation span performance

and set shifting abilities, we endorse the one-path Updating model (Model
4) as the best one for operation span scores.
These results support the hypothesis that the operation span task primarily
involves the ability to continuously update and monitor incoming informa-tion. This
conclusion is also consistent with the findings of other studies that
have found significant correlations between working memory span tasks and
the letter memory task (Lehto, 1996) as well as the keep track task (Engle
et al., 1999b), two of the Updating measures we used in this study. In con-trast, as
the poor fit of the one-path model from Shifting suggests, we found
no evidence for the proposal that the ability to efficiently switch back and
forth between the processing component (equation verification) and the storage-
age component (word span) is a crucial aspect of the operation span task,
at least to the extent to which the Shifting factor captures the ability to make
such switches.
Dual task. The last complex executive task we examined was dual tasking,
which has been considered a prime example of the type of situation that
implicates the central executive component of working memory (Baddeley,
1996) and has been widely used to study the functioning of the central executive
(eg, Baddeley et al., 1997; Baddeley & Logie, 1999; Bourke, Dun-can, & Nimmo-
Smith, 1996; Hegarty, Shah, & Miyake, in press). Supporting
this claim, a neuroimaging study has shown that simultaneously performing
a verbal task and a visuospatial task activates the prefrontal cortex in addition
to the areas involved in processing verbal and visuospatial information
(D'Esposito et al., 1995). In addition, when compared to performance decrements
on individual tasks, neuropsychological studies have found disproportionately larger
dual task decrements in various patients with suspected executive function deficits,
including traumatic brain injury, frontal lobe lesions,
and Alzheimer's and Parkinson's diseases (see Baddeley et al., 1997, for a
review).
Despite this general agreement that dual tasking involves executive control
processes, there is still no clear consensus on what abilities or specific executive
functions are involved in dual task performance (Miyake & Shah,
1999). One common proposal is that dual tasking involves constantly and
rapidly shifting mental set between tasks (eg, Duncan, 1995). This conception would predict
that the Shifting factor would contribute to dual task per-formance.

We tested this hypothesis by comparing a model with only a path from

Shifting to dual task performance against the three-path and no-path models.
As shown in Table 8, there was no evidence that Shifting contributed to dual
task performance: The model with only a path from Shifting (Model 2) was
no better than the no-path model (Model 3), ÿ2 (1) 0.0, p .10. Further,
the three-path model (Model 1) was also no better than the no-path model,
ÿ2(3) 1.60, p .10, indicating that none of our factors significantly pre-
Machine Translated by Google
84 MIYAKE ET AL.

TABLE 8
Fit Indices and Standardized Regression Coefficients for Structural Equation Models with
Dual Task Performance (N 134)

Coefficients for specified paths

Model df ÿ2 SRMR IFI Shifting Updating Inhibition

1. Full three paths 30 27.41 2. One path from .052 1.03 0.02 .24 0.27
Shifting 32 29.01 3. No paths 33 29.01 .054 1.03 0.01 — —
.054 1.04 —— —

Note. The endorsed model is indicated in bold.

dictated dual task performance.11 These results suggest that dual tasking may
tap an executive function that is somewhat independent of the three targets
functions examined in this study, although null results such as these need
to be interpreted cautiously.
Summary. The results for the SEM analyzes indicate that the three targets
executive functions contribute differentially to performance on the more
complex executive tests. Specifically, Shifting seems to contribute to WCST
performance, Inhibition to TOH performance (at least in the way it is typi-cally
administered), Inhibition and Updating to RNG performance, and Up-dating to
operation span performance. Dual task performance did not seem
to be related to any of the three executive functions examined in this study,
although this result is difficult to interpret.

Alternative Explanations
Although the statistical models evaluated in this article are relatively simple,
the interpretations of the CFA and SEM results critically hinge on our
assumption that the three latent variables in these models indeed succeeded
in tapping the three target executive functions (ie, Shifting, Updating, and
Inhibition, respectively). Any violations of this assumption would seriously
challenge the conclusions we drew from the CFA and SEM results. There-fore,
it is important to rule out alternative explanations that question the
validity of that assumption.
One such alternative explanation is that different models that were not
explicitly tested and reported in this article (eg, a different classification of

11
We also performed the same SEM analyzes on another dependent measure, namely, the
average standardized residuals obtained by separately regressing the maze completion and
word generation scores in the dual task condition on their respective scores from the single
task condition, standardizing both sets of residuals from these analyses, and averaging the
two sets. Although less widely used, this measure has a number of statistical advantages over
the proportion decrement score reported in the main text, such as less error and higher
reliability (Cohen & Cohen, 1983). The SEM results, however, remained virtually identical,
suggesting that the null results might not be due solely to the statistical characteristics of the
proportion decrement scores.
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 85

the nine manifest variables into three latent variables) might produce equiva-lent or even
better fits to the data than the CFA model presented in Fig. 2.
Although it is difficult (as well as unwise) to test all the CFA models possible and show that
the model of our choice is indeed superior to those models, it seems clear that a substantially
different CFA model would not produce as good a fit for several reasons. First, an inappropriate
clustering of the task would be apparent from a number of sources, such as the magnitude
and significance of the specified paths and the magnitudes of the residuals in the fitted
covariance matrix. The CFA model endorsed here (Fig. 2) as well as the SEM model selected
for each complex executive task was free from problems indicated by such sources. Second,
the results of an EFA performed on the same data set, although not as clear-cut as the CFA
results, largely con-formed to the factor structure that we postulated (see Appendix C for the
results).12 Because EFA maximizes fit to the data without any constraints on which tasks
should load on which factors, this finding suggests that a substantially different structure
would not fit the data better.

Closely related is the alternative explanation based on the types of dependent measures
used for different latent variables: Because the three tasks loading on the Shifting factor were
all RT-based measures and the three tasks loading on the Updating factor were all accuracy
measures, the separability of the factors might be due to this methodological artifact. Two
lines of evidence argue against this account. First, if the Shifting and Updating factors were
separable primarily because they were composed of RT-based and accuracy measures,
respectively, then we would expect to see some signs of unsatisfactory model fit, such as
large residuals in the fitted covariance matrix, particularly for the Inhibition measures that
included both RT-based and accuracy measures. There were no such signs in the data,
however. Sec-ond, the SEM results did not conform to the pattern that the Shifting factor
predicted performance on RT-based executive tasks and the Updating factor predicted
performance on accuracy-based executive tasks. Instead, the three latent variables showed
differential contributions to performance on the complex executive tasks in a manner
consistent with our a priori predictions.

12
An examination of the EFA results, presented in Appendix C, indicates that the tone
monitoring task may be related not just to the Updating factor but also to the Inhibition
factor, suggesting that it may not be considered a relatively pure Updating task. In retrospect,
the tone monitoring task does seem to involve an Inhibition ability in that it requires
participants not only to monitor counters for the three tones but also to reset those counters
every time the fourth tone for each pitch occurs. Subjectively, resetting the counter to 0 is
quite difficult and may require an Inhibition ability to overcome the tendency to keep
counting. To make sure that this impurity of the tone monitoring task did not distort the
conclusions we reached, we also estimated the same CFA and SEM models without the
tone monitoring task (ie, only two, rather than three, tasks used for the Updating factor) to
examine the impact of this task being included. The results of these analyzes indicated that
this three-factor model without the tone monitoring task also provided an excellent fit to the
data, and the qualitative conclusions for the model comparisons (both CFA and SEM) remained i
Machine Translated by Google
86 MIYAKE ET AL.

These results suggest that the separability of the three executive functions is
not due to an artifact resulting from the RT–accuracy distinction in the dependent
measures.
Another alternative explanation is that some nonexecutive task requirements
that were common within the three tasks chosen to tap each target executive
function might have driven the underlying factor structure, rather than the
presence of separable executive functions, as we contain. Although this
alternative explanation cannot be completely ruled out on the basis of available
data, we attempted to minimize the influence of idiosyncratic task requirements
by deliberately choosing, for each latent variable, tasks that required the same
executive function but involved quite different specific requirements (eg,
stopping a prepotent eye movement for antisaccade and stopping a prepotent
categorization response for stop-signal). Furthermore, it is not the case that we
simply picked tasks for each function that were essentially the same tasks with
minor parametric variations. Thus, it seems almost impossible to explain the
obtained pattern of CFA and SEM results purely in terms of the commonality
and separability of task requirements other than the three postulated executive
functions.
In summary, the arguments against these alternative explanations are strong;
there is little evidence in the data that suggests a violation of the assumption
that the three latent variables in the CFA and SEM analyzes tapped the intended
target executive functions. Although it is a mistake to take for granted (or
consider it proven) that the latent variables fully captured the intended underlying
functions or abilities (Kline, 1998), these considerations provide strong support
for the view that the latent variables indeed captured the respective target
executive functions.

GENERAL DISCUSSION

In this article, we reported an individual differences study that examined the

organization and roles of three often-postulated executive functions—shifting
between mental sets or tasks (Shifting), updating and monitoring of working
memory contents (Updating), and inhibition of prepotent responses (Inhibition)
—at the level of latent variables, rather than at the level of mani-fest variables
(ie, individual tasks). One primary goal of the study was to specify the degree
of relationship among the three target functions and thereby contribute to the
understanding of the unitary versus nonunitary nature of executive functions.
The second main goal was to examine how the three target executive functions
contribute to performance on more complex executive tasks. The study yielded
clear results with respect to both of these goals.

Regarding the first main goal, the results from the CFA indicated that the
three target functions (ie, Shifting, Updating, and Inhibition) are clearly
distinguishable. The full three-factor model in which the correlations among
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 87

the three latent variables were allowed to vary freely produced a significantly
better fit to the data than any other models that assumed complete unity among
two or all three of the latent variables. The three target executive functions are
not completely independent, however, and do seem to share some underlying
commonality. In the full three-factor model (Fig. 2), the estimates of the
correlations among the three latent variables were moderately high (ranging
from .42 to .63). In addition, this full model provided a far better fit to the data
than the three-factor model that assumed complete independence among the
three latent variables. These results suggest that the three often postulated
executive functions of Shifting, Updating, and In-hibition are separable but
moderately correlated constructs, thus indicating both unity and diversity of
executive functions.
As for the second goal, the results of the SEM analyzes showed that the
executive tasks often used in cognitive and neuropsychological studies are not
completely homogeneous in the sense that different executive functions
contribute differentially to performance on these tasks. Specifically, we found
that the Shifting ability contributes to performance on the WCST. The Inhibition
ability seems to play an important role in solving the TOH puzzle, at least when
no specific instructions for strategies are given and many people are likely to
use the perceptual strategy to perform the task. Producing ran-dom sequences
of numbers in the RNG task seems to depend on multiple abilities, particularly
the Inhibition ability and the Updating ability, which appear to be tapped by
different sets of randomness indices. Finally, the operation span task, a prevalent
measure of verbal working memory capacity, seems to primarily involve the
Updating ability. These results indicate that the Shifting, Updating, and Inhibition
abilities contribute differentially to performance on commonly used executive
tasks, even though they are mod-erately correlated with one another.
Furthermore, the results offer a clear, independent confirmation of some
previously proposed accounts of what these tasks really measure, at least in a
sample of young, healthy college students.
The only complex executive task that did not relate clearly to the three target
executive functions was the dual task. Although such null results need to be
interpreted cautiously, one possibility is that the simultaneous coordination of
multiple tasks is an ability that is somewhat distinct from the three executive
functions examined in this study.
It is important to point out that the current data are based on a restricted
sample of young college students. Therefore, the results may not be completely
generalizable to more cognitively diverse samples, such as those that include
noncollege students, young children, elderly adults, or neurologically impaired
participants. For example, the degree of separability of different executive
functions may be less pronounced among such less restricted samples (eg,
Legree, Pifer, & Grafton, 1996). It could also be the case that different factors
than the ones we reported here contribute more strongly to performance on the
complex executive tasks, possibly reflecting different
Machine Translated by Google
88 MIYAKE ET AL.

strategies adopted by participants or some specific patterns of age-related

changes or neurological impairments in executive functions. Although such
limitations in generalizability across samples are possible, there is also a good
chance of similar patterns of results emerging across different samples, given
that the overall pattern of zero-order correlations we found in this study is fairly
analogous to the patterns obtained from previous individual differences studies
that tested a wide range of target populations (eg, college students, young
children, elderly adults, and brain-damaged patients).

The Unity and Diversity of Executive Functions Revisited The

main results from the CFA analyzes indicate that executive functions may be
characterized as separable but related functions that share some un-derlying
commonality. Thus, as Teuber (1972) suggested in his review of frontal lobe
functions more than a quarter of a century ago, the results point to both unity and
diversity of executive functions and indicate that both of these aspects need to be
taken into consideration in developing a theory of executive functions (see also
Duncan et al., 1997).
Concerning the unity of executive functions, the results of the present study
are compatible with a fair number of theoretical proposals that note some ''family
resemblance'' or common mechanisms across different executive functions or
functions putatively performed by the frontal lobes (eg, Duncan et al., 1996, 1997;
Engle et al., 1999a; Kimberg & Farah, 1993). The moderately high intercorrelations
among the three target executive functions raises one important theoretical
question, however. What might the source(s) of the commonality be? Although
precisely specifying the nature of the un-derlying commonality is beyond the
scope of this article and awaits future research, at least two explanations seem
possible.
First, although the nine tasks used in the CFA were each chosen to tap one
target executive function, it is likely that they did share some common task
requirements, particularly the maintenance of goal and context informa-tion in
working memory. Working memory plays a prominent role in several existing
theoretical accounts of executive functions, in which the crucial role of the frontal
lobes is hypothesized to be the active maintenance of goals, plans, and other
task-relevant information in working memory (Engle et al., 1999a, 1999b; Kimberg
& Farah, 1993; O'Reilly, Braver, & Cohen, 1999; Pennington, Bennetto, McAleer,
& Roberts, 1996). For example, Engle et al. (1999a, 1999b) recently proposed
that a crucial component of working memory capacity is ''controlled attention,''
which is a domain-free atten-tional capacity to actively maintain or in some cases
suppress working mem-ory representations. In their account, any situations that
involve controlled processes (such as goal maintenance, conflict resolution,
resistance to or sup-pression of distracting information, error monitoring, and
effortful memory search) would require this ''controlled attention'' capacity,
regardless of the specifics of the tasks to be performed. Thus, the ability to keep
goal-related
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 89

and other task-relevant information active in working memory during controlled

processing could be the basis for the observed commonality among
the three executive functions.
Another possible explanation is that the three target executive functions
all involve some sort of inhibitory processes to operate properly. For example,
one could argue that the Updating function may require ignoring irrelevant
incoming information and also suppressing no longer relevant information.
Similarly, the Shifting function may require deactivating or suppressing
an old mental set to switch to the new set. Although conceptually separable,
This type of inhibition (what one might call inhibition or suppression of irrele-
vant or unnecessary information or mental sets) may be related to the deliber-
ate, controlled inhibition of prepotent responses that we focused on in the
reported study. Thus, all three target functions may share some inhibitory
process, which in turn might have led to the moderate correlations among
the three executive functions. Although this account is vague in terms of
what the notion of ''inhibition'' really means, it deserves further investigation,
given that the theoretical proposals that emphasize inhibition as a basic
unit of working memory and executive control processes have become
increasingly popular in the literature (eg, Dempster & Corkill, 1999; Zacks &
Hasher, 1994).
As for the diversity of executive functions, the results of the present study
are also quite compatible with a substantial body of research noting the ap-
parent neuropsychological or correlational dissociations of executive func-tions
reviewed earlier. Furthermore, the results also support recent theoretical
attempts to fractionate the central executive (eg, Baddeley, 1996; Badde-ley &
Logie, 1999) or the SAS (eg, Stuss, Shallice, Alexander, & Picton,
1995), both of which tended to have a unitary flavor in their earlier conceptualizations.

One important question that needs to be considered regarding the diversity

of executive functions is how best to classify separable executive functions.
In this article, we have taken a rather pragmatic approach, focusing on three
of the most frequently postulated functions in the literature. Our choice of
the three functions were not arbitrary, however. We chose these functions
because they seemed relatively basic (or at least more basic than prevalently
mentioned higher level concepts like ''planning'') and have often been used
to explain performance on complex executive tasks like the ones we exam-
ined in this study. The CFA and SEM results demonstrate that our strategy
was successful and that examining the organization of executive functions
at this level of analysis has merit, at least at this early stage of executive
function research.
Despite this success, we are not claiming that the three investigated
executive functions are the only executive functions, nor would we suggest that
they are anything like the fundamental units or primitives of cognition. Our
exploration of the diversity of executive functions is only a first step, and
Machine Translated by Google
90 MIYAKE ET AL.

There are a number of important issues that need to be addressed in future research to
better characterize the nature of separability or diversity of executive functions.

First, although our choice of the three target functions in this study seemed a reasonable
one, it is certainly not exhaustive and there are other important relatively basic functions that
need to be added to the current list. One such function, suggested by the SEM results for
dual task performance, is the coordination of multiple tasks (Baddeley, 1996; Emerson,
Miyake, & Ret-tinger, 1999), which may be somewhat separable from the three functions
examined in this study. Second, the relationship between these relatively basic executive
functions and more complex concepts like ''planning'' needs to be examined. If a combination
of relatively basic functions can account for more complex executive functions (eg, a
combination of Shifting and Updating), then it helps make the classification of executive
functions less chaotic. Finally, although the current level of analysis might be the most useful
at this moment, it is also possible that the target functions we consider here can be
decomposed into more basic component processes. Such a finer level of analysis seems to
have been adopted by Stuss et al. (1995) in their effort to fractionate the SAS. Although this
finer level of analysis faces the difficulty of selecting tasks that primarily tap one target
process and hence may not lend itself readily to individual differences analyses, it appears
a theoretically worthwhile approach to pursue.

In summary, although there are many more issues that need to be explored with respect
to the organization of executive functions, the current results, together with some recent
theoretical proposals (Duncan et al., 1996, 1997), help reconcile the controversy regarding
the ''unitary versus nonunitary'' nature (or ''unity versus diversity'') of executive functions. A
simple dichotomy will not suffice, and both aspects must be taken into account.

Implications of the Latent Variable Approach for Studying

Executive Functions

In the reported study, we examined the organization and roles of three often postulated
executive functions at the level of latent variables, rather than at the level of manifest
variables. This latent variable approach has several important advantages over a more
common approach of relying on zero-order correlations and EFA, particularly in the context
of studying exec-utive functions.

First, the latent variable approach can alleviate the task impurity problem, namely, that
commonly used executive tasks are highly complex and typically place heavy demands on
not just executive processes of interest, but also nonexecutive processes within which the
executive processing requirement is embedded. The latent variable approach circumvents
this problem by statistically ''extracting'' what is common across multiple tasks that all involve
the same target processing requirement and analyzing the relationship-
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 91

ships among different executive functions in terms of these ''purer'' factors.

Second, this approach may also help solve another common problem in
executive function research, namely the construct validity problem. Most
proposals of what each executive task measures have tended to be speculative
and not independently tested, but, as we demonstrated in our SEM analyses,
the latent variable approach can provide a useful way to characterize the
nature of specific executive functions involved in complex executive tasks.
These advantages are particularly important, given that the executive tasks
are generally associated with low reliability (Denckla, 1996; Rabbitt, 1997b)
and are thereby constrained to yield low intercorrelations. There may be too
much error to test specific hypotheses or theoretical proposals if one analyzes
the correlations at the level of individual tasks.
Our success in being able to ''extract'' common factors for each of the
three latent variables is encouraging for executive function research, particularly
in the context of a recent remark by Rabbitt (1997b): ''In our laboratory,
we have been unable to find any commonality of individual differences in
'inhibition' between each of a wide variety of logically identical but superficially
dissimilar Stroop-like tasks. That is, we can find no evidence that the
ability to inhibit responses across a range of different tasks is consistently
greater in some individuals than in others'' (pp. 12–13). At the level of zero-
order correlations, the underlying commonality might not be obvious, but
analyzing the data at the level of latent variables may increase the chances
of revealing the common structure if it exists.
In addition to these advantages for individual differences studies of executive
functions, the latent variable approach provides important implications
for other lines of research on executive functions, including testing of brain-
damaged patients in neuropsychological settings and neuroimaging studies
of executive functions.
First, the results of this study suggest that it is important to systematically
administer multiple executive tasks to understand the nature of sparing and
impairments in a patient's executive functioning. Given that executive functions
are separable and that different executive functions contribute differentially to
various executive tasks, simply relying on prevalently used tasks
like the WCST and TOH as general measures of executive functioning does
not enough. Although the generalizability of the current results to
neuropsychological populations needs to be carefully evaluated first, it is important
to be aware of the underlying separable functions and assess the patient's
profile of executive functioning by taking into consideration which task taps
which executive function(s) (see Miyake, Emerson, & Friedman, in press,
for further discussion of the implications of the latent variable approach for
clinical assessment).
Second, the current results also have interesting implications for
neuroimaging studies of executive functions. So far, such studies have focused on
complex executive tasks like the WCST, TOH, and RNG to examine the
Machine Translated by Google
92 MIYAKE ET AL.

neural basis of executive functions, particularly the involvement of the front-

tal lobes in performance on those tasks. In almost all these cases, each study
reports the pattern of brain activation observed for one executive task,
locating a set of specific brain regions that are considered important for
certain processes (eg, ''planning'' for TOH). However, one major danger of
relying on just one task to infer the neural implementation of a specific target
executive function is that, even though a clever subtraction method is used
to isolate a target process, it is still possible that the isolated process includes
other nonexecutive processes specific to that particular task. An interesting
alternative is to consider multiple tasks (two or more) that are known to
share the same underlying target process (perhaps as a result of an
independent latent variable analysis study) and then examine the common regions
of activation across these tasks. Although it may be more costly and time-
consuming, this latent variable approach to neuroimaging may also help
illuminate the degree of commonality or separability of different executive
functions at the level of brain implementation or functioning.
 A
APPENDIX

Noted)
Unless
137
(N
Measures
15
the
for
Coefficients
Correlation
Pearson

1 10 11 12 13 15
14
Machine Translated by Google

taska
Dual
15.
span
Operation
14.
2
13.
1
Component
RNG
12.
TOHb
11.
perseverationa
WCST
10.
track
Keep
4.
letter
Number–
2.
Stroop
9.
signal
Stop-
8.
Antisaccade
7.
memory
Letter
6.
monitoring
Tone
5.
global
Local–
3.
minus
Plus–
1. —

0.03
.09
0.07
.20*
.08
.26*
.07
.11
.15
.24*
.22*
.23* —
.32*

—
.32*
0.02
.10
.09
.13
.17
.11
.19*
.08 —
.12

0.04
0.09
0.05
.05
.07
.01
.18*
.06
.11
.21*
.00

0.09
.41*
.29*
.03
.13
.09
.11
.10
.12
.34* —
.27*

—
.22*
.26*

—
.15
.34*
.28*
.06
.11
.14
.19*
.18*
.16
.04
.09 —
.19*

0.08
.16
.02
.24*
.21*
.15
.20*

—
.10
0.01

.20*
.13
.01
.18*
.11
.12
.17
.08

—
.18*
0.08 —
0.02
0.05
0.18*
.06
.16
.13
.04
.12 —
.02

—
.10
.17* —
.13

.12
0.03 .06
0.16 —
0.14
0.09
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS

generation.
number
random
RNG,
Hanoi;
of
Tower
TOH,
Test;
Sorting
Card
Wisconsin
WCST,

136.

N
*

to bNote.
p
93
Machine Translated by Google
94 MIYAKE ET AL.

APPENDIX B
Loadings for the Principal Components Analysis of 15 RNG Measures
(N 137)
Component

Measure 1 2 3

TPI .92 0.06 0.15

TO .89 0.14 0.21
Runs .86 0.16 0.01
RNG .85 .16 .13
R .06 .86 .22
Coupon score 0.03 .81 .02
Mean RG 0.05 .65 0.19
Mode RG 0.01 .53 0.32
Phi6 .02 0.52 .37
Phi7 .18 0.48 .32
Phi3 .02 .01 .84
Phi4 0.04 0.22 .78
Phi2 0.24 .00 .71
Phi5 0.04 0.34 .63
RNG2 .49 .36 .50

Correlations
—
Component 1
.02 —
Component 2
.04 0.06 —
Component 3

Note: An oblique Promax rotation was used to obtain estimates of correlations among the
components. These three components were the only ones with eigenvalues larger than 1, and
the Scree plot also suggested the three-component solution. TPI, turning point index; A, total
adjacency; RNG, Evan's random number generation score; R, redundancy; mean RG, mean
repetition gap; mode RG, mode repetition gap; phi2–7, phi indices; RNG2, analysis of inter-leaved digrams.
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 95

APPENDIX C
Factor Loadings for the Exploratory Principal Factor Analysis of the
Shifting, Updating, and Inhibition Tasks ( N 137)
Factor

Measure 1 2 3

Plus–minus .40 .22 .05

Number–letter .58 0.11 .16
Local–global .58 .07 0.17
Keep track 0.01 .58 0.05
Tone monitoring .01 .22 .35
Letter memory .04 .57 .05
Antisaccade .08 .07 .44
Stop signal .09 0.09 .38
Stroop 0.12 .09 .43

Correlations
Factor 1 —

Factor 2 .39 —

Factor 3 .30 .42 —

Note: An oblique Promax rotation was used to obtain estimates of interfactor correlations.
These three factors were the only ones with eigenvalues larger than 1 in the unreduced correlation matrix,
and the Scree plot also suggested the three-factor solution.

REFERENCES

Allport, A., & Wylie, G. Task switching, stimulus-response bindings, and negative priming.
In S. Monsell & J. Driver (Eds.), Control of cognitive processes: Attention and performance XVIII.
Cambridge, MA: MIT Press. [in press]
Anderson, J.C., & Gerbing, D.W. (1988). Structural equation modeling in practice: A review
and recommended two-step approach. Psychological Bulletin, 103, 411–423.
Anderson, S.W., Damasio, H., Jones, R.D., & Tranel, D. (1991). Wisconsin card sorting
performance as a measure of frontal lobe damage. Journal of Clinical and Experimental
Neuropsychology, 13, 909–922.
Arnett, P.A., Rao, S.M., Grafman, J., Bernardin, L., Luchetta, T., Binder, JR, & Lobeck,
L. (1997). Executive functions in multiple sclerosis: An analysis of temporal ordering,
semantic encoding, and planning abilities. Neuropsychology, 11, 535–544.
Baddeley, A.D. (1986). Working memory. New York: Oxford Univ. Press.
Baddeley, A. D. (1996). Exploring the central executive. Quarterly Journal of Experimental
Psychology, 49A, 5–28.
Baddeley, A., Della Sala, S., Gray, C., Papagno, C., & Spinnler, H. (1997). Central testing
executive functioning with a pencil-and-paper test. In P. Rabbitt (Ed.), Methodology of
frontal and executive function (pp. 61–80). Hove, UK: Psychology Press.
Baddeley, A., Emslie, H., Kolodny, J., & Duncan, J. (1998). Random generation and the
executive control of working memory. Quarterly Journal of Experimental Psychology,
51A, 819–852.
Baddeley, A.D., & Logie, R.H. (1999). Working memory: The multicomponent model. in
Machine Translated by Google
96 MIYAKE ET AL.

A. Miyake & P. Shah (Eds.), Models of working memory: Mechanisms of active

maintenance and executive control (pp. 28–61). New York: Cambridge Univ. Press.
Berg, E. A. (1948). A simple objective technique for measuring flexibility in thinking. Journal of
General Psychology, 39, 15–22.
Bourke, P.A., Duncan, J., & Nimmo-Smith, I. (1996). A general factor involved in dual-task
performance decrement. Quarterly Journal of Experimental Psychology, 49A, 525–545.
Burgess, P.W. (1997). Theory and methodology in executive function research. In P. Rabbitt
(Ed.), Methodology of frontal and executive function (pp. 81–116). Hove, UK: Psychol-
ogy Press.
Burgess, P.W., Alderman, N., Evans, J., Emslie, H., & Wilson, B.A. (1998). The ecological
validity of tests of executive function. Journal of the International Neuropsychological
Society, 4, 547–558.
Carpenter, P.A., Just, M.A., & Shell, P. (1990). What one intelligence test measures: A
theoretical account of the processing in the Raven Progressive Matrices Test. Psychological
Review, 97, 404–431.
Casey, BJ, Trainor, RJ, Orendi, JL, Schubert, AB, Nystrom, LE, Giedd, JN, Castellanos, FX,
Haxby, JV, Noll, DC, Cohen, JD, Forman, SD, Dahl, RE, & Rapoport, JL (1997). A
developmental functional MRI study of prefrontal activation during performance of a go-no-
go task. Journal of Cognitive Neuroscience, 9, 835–847.
Cohen, J., & Cohen, P. (1983). Applied multiple regression/ correlation analysis for the behavior
ioral sciences (2nd ed.). Hillsdale, NJ: Erlbaum.
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.
Damasio, A. R. (1994). Descartes' error: Emotion, reason, and human brain. New York:
Grosset/Putnam.
Daneman, M., & Carpenter, P. A. (1980). Individual differences in working memory and reading.
Journal of Verbal Learning and Verbal Behavior, 19, 450–466.
Daneman, M., & Merikle, P. M. (1996). Working memory and language comprehension: A
meta-analysis. Psychonomic Bulletin & Review, 3, 422–433.
Dempster, F.N., & Corkill, A.J. (1999). Individual differences in susceptibility to interference and
general cognitive ability. Acta Psychologica, 101, 395–416.
Denckla, M. B. (1996). A theory and model of executive function: A neuropsychological
perspective. In GR Lyon & NA Krasnegor (Eds.), Attention, memory, and executive function
(pp. 263–278). Baltimore, MD: Brookes.
D'Esposito, M., Detre, JA, Alsop, DC, Shin, R.K., Atlas, S., & Grossman, M. (1995).
The neural basis of the central executive system of working memory. Nature, 378, 279–
281.
Dunbar, K., & Sussman, D. (1995). Toward a cognitive account of the frontal lobe function:
Simulating frontal lobe deficits in normal subjects. Annals of the New York Academy of
Sciences, 769, 289–304.
Duncan, J. (1995). Attention, intelligence, and the frontal lobes. In MS Gazzaniga (Ed.), The
cognitive neurosciences (pp. 721–733). Cambridge, MA: MIT Press.
Duncan, J., Emslie, H., Williams, P., Johnson, R., & Freer, C. (1996). Intelligence and the frontal lobe: The organization
of goal-directed behavior. Cognitive Psychology, 30, 257–303.

Duncan, J., Johnson, R., Swales, M., & Freer, C. (1997). Frontal lobe deficits after head injury:
Unity and diversity of function. Cognitive Neuropsychology, 14, 713–741.
Ekstrom, R.B., French, J.W., Harman, H.H., & Dermen, D. (1976). Manual for kit of factor-
referenced cognitive tests. Princeton, NJ: Educational Testing Service.
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 97

Emerson, M.J., Miyake, A., & Rettinger, D.A. (1999). Individual differences in integrating and
coordinating multiple sources of information. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 25, 1300–1321.
Engle, R.W., Kane, M.J., & Tuholski, S.W. (1999a). Individual differences in working memory
capacity and what they tell us about controlled attention, general fluid intelli-gence, 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 Univ. Press.
Engle, RW, Tuholski, SW, Laughlin, JE, & Conway, ARA (1999b). Working memory, short-
term memory, and general fluid intelligence: A latent variable approach. Journal of
Experimental Psychology: General, 125, 309–331.
Everling, S., & Fischer, B. (1998). The antisaccade: A review of basic research and clinical
studies. Neuropsychologia, 36, 885–899.
Fabrigar, LR, Wegener, DT, MacCallum, RC, & Strahan, EJ (1999). Evaluating the use of exploratory factor analysis
in psychological research. Psychological Methods, 4, 272–299.

Godefroy, O., Cabaret, M., Petit-Chenal, V., Pruvo, J.-P., & Rousseaux, M. (1999). Control functions of the frontal
lobes: Modularity of the central-supervisory system? Cortex, 35, 1–20.

Goel, V., & Grafman, J. (1995). Are the frontal lobes involved in ''planning'' functions?
Interpreting data from the Tower of Hanoi. Neuropsychologia, 33, 623–642.
Goldman-Rakic, P.S. (1996). The prefrontal landscape: Implications of functional architecture
for understanding human mentality and the central executive. Philosophical Transactions
of the Royal Society of London, 351, 1445–1453.
Hallett, P. E. (1978). Primary and secondary saccades to goals defined by instructions. Vision
Research, 18, 1279-1296.
Hegarty, M., Shah, P., & Miyake, A. Constraints on using the dual-task methodology to specify
the degree of central executive involvement in cognitive tasks. Memory & Cognition. [in
press]
Hu, L.-T., & Bentler, P. M. (1995). Evaluating model fit. In RH Hoyle (Ed.), Structural equation
modeling: Concepts, issues, and applications (pp. 76–99). Thousand Oaks, CA: Sage.

Hu, L.-T., & Bentler, P.M. (1998). Fit indices in covariance structure modeling: Sensitivity to
underparameterized model misspecification. Psychological Methods, 3, 424–453.
Humes, G.E., Welsh, M.C., Retzlaff, P., & Cookson, N. (1997). Towers of Hanoi and London:
Reliability and validity of two executive function tests. Assessment, 4, 249–257.
Jahanshahi, M., Profice, P., Brown, R.G., Ridding, M.C., Dirnberger, G., & Rothwell, J.C.
(1998). The effects of transcranial magnetic stimulation over the dorsolateral prefrontal cortex on suppression
of habitual counting during random number generation. Brain, 121, 1533–1544.

Jersild, A. T. (1927). Mental set and shift. Archives of Psychology, Whole No. 89.
Jonides, J., & Smith, E.E. (1997). The architecture of working memory. In M. D. Rugg (Ed.),
Cognitive neuroscience (pp. 243–276). Cambridge, MA: MIT Press.
Jo¨reskog, KG, & So¨rbom, D. (1989). LISREL 7: A guide to the program and applications
(2nd ed.). Chicago: SPSS.
Judd, C. M., & McClelland, G. H. (1989). Data analysis: A model-comparison approach. San
Diego, CA: Harcourt Brace Jovanovich.
Kiefer, M., Marzinzik, F., Weisbrod, M., Scherg, M., & Spitzer, M. (1998). The time course
Machine Translated by Google
98 MIYAKE ET AL.

of brain activations during response inhibition: Evidence from event-related potentials in a

go/no go task. NeuroReport, 9, 765–770.
Kimberg, D.Y., D'Esposito, M., & Farah, M.J. (1997). Effects of bromocriptine on human subjects
depend on working memory capacity. NeuroReport, 6, 3581–3585.
Kimberg, D.Y., & Farah, M.J. (1993). A unified account of cognitive impairments following frontal
lobe damage: The role of working memory in complex, organized behavior. Journal of
Experimental Psychology: General, 122, 411–428.
Kline, R. B. (1998). Principles and practice of structural equation modeling. New York:
Guilford.
Kok, A. (1999). Varieties of inhibition: Manifestations in cognition, event-related potentials
and aging. Acta Psychologica, 101, 129–158.
Larson, G.E., Merritt, C.R., & Williams, S.E. (1988). Information processing and intelli-gence:
Some implications of task complexity. Intelligence, 12, 131–147.
Lehto, J. (1996). Are executive function tests dependent on working memory capacity? Quar-
terly Journal of Experimental Psychology, 49A, 29–50.
Legree, P.J., Pifer, M.E., & Grafton, F.C. (1996). Correlations among cognitive abilities
are lower for higher ability groups. Intelligence, 23, 45–57.
Levin, H.S., Culhane, K.A., Hartmann, J., Evankovich, K., Mattson, A.J., Harward, H., Ringholz,
G., Ewing-Cobbs, L., & Fletcher, J.M. (1991). Developmental changes in performance on
tests of purported frontal lobe functioning. Developmental Neuropsychol- ogy, 7, 377–395.

Levin, HS, Fletcher, JM, Kufera, JA, Harward, H., Lilly, MA, Mendelsohn, D., Bruce, D., &
Eisenberg, H.M. (1996). Dimensions of cognition measured by the Tower of London and
other cognitive tasks in head-injured children and adolescents. Developmental
Neuropsychology, 12, 17–34.
Logan, G. D. (1985). Executive control of thought and action. Acta Psychologica, 60, 193–
210.
Logan, G. D. (1994). On the ability to inhibit thought and action: A user's guide to the stop signal
paradigm. In D. Dagenbach & TH Carr (Eds.), Inhibitory processes in attention, memory,
and language (pp. 189–239). San Diego, CA: Academic Press.
Lowe, C., & Rabbitt, P. (1997). Cognitive models of aging and frontal lobe deficits. In P.
Rabbitt (Ed.), Methodology of frontal and executive functions (pp. 39–59). Hove, UK:
Psychology Press.
Luria, A. R. (1966). Higher cortical functions in man. New York: Basic Books.
Lyon, GR, & Krasnegor, N.A. (Eds.). (1996). Attention, memory, and executive function.
Baltimore: Brookes.
Mardia, K. V. (1970). Measurements of multivariate skewness and kurtosis with applications.
Bio-metrika, 57, 519–530.
Miyake, A., Emerson, MJ, & Friedman, NP Assessment of executive functions in clinical settings:
Problems and recommendations. Seminars in Speech and Language. [in press]
Miyake, A., & Shah, P. (1999). Toward unified theories of working memory: Emerging general
consensus, unresolved theoretical issues, and future research directions. In A. Miyake & P.
Shah (Eds.), Models of working memory: Mechanisms of active maintenance and exec-
utive control (pp. 442–481). New York: Cambridge Univ. Press.
Monsell, S. (1996). Control of mental processes. In V. Bruce (Ed.), Unsolved mysteries of the
mind: Tutorial essays in cognition (pp. 93–148). Hove, UK: Erlbaum.
Morris, N., & Jones, D. M. (1990). Memory updating in working memory: The role of the central
executive. British Journal of Psychology, 81, 111–121.
Machine Translated by Google
UNITY AND DIVERSITY OF EXECUTIVE FUNCTIONS 99

Morris, R.G., Miotto, E.C., Feigenbaum, J.D., Bullock, P., & Polkey, C.E. (1997). The effect of
goal-subgoal conflict on planning ability after frontal- and temporal-lobe lesions in humans.
Neuropsychologia, 35, 1147–1157.
Moulden, DJA, Picton, TW, Meiran, N., Stuss, DT, Riera, JJ, & Valdes-Sosa, P.
(1998). Event-related potentials when switching attention between task-sets. Brain and
Cognition, 37, 186-190.
Murji, S., & DeLuca, J.W. (1998). Preliminary validity of the cognitive function checklist:
Prediction of Tower of London performance. Clinical Neuropsychologist, 12, 358–364.
Navon, D. (1977). Forest before trees: The precedence of global features in visual perception.
Cognitive Psychology, 9, 353–383.
Norman, D. A., & Shallice, T. (1986). Attention to action: Willed and automatic control of
behavior. In RJ Davidson, GE Schwartz, & D. Shapiro (Eds.), Consciousness and self-
regulation: Advances in research and theory (Vol. 4, pp. 1–18). New York: Plenum.
O'Reilly, R.C., Braver, T.S., & Cohen, J.D. (1999). A biologically based computational model of
working memory. In A. Miyake & P. Shah (Eds.), Models of working memory: Mechanisms
of active maintenance and executive control (pp. 375–411). New York: Cambridge Univ.
Press.
Ozonoff, S., & Strayer, D.L. (1997). Inhibitory function in nonretarded children with autism.
Journal of Autism and Developmental Disorders, 27, 59–77.
Pennington, B.F., Bennetto, L., McAleer, O., & Roberts, R.J., Jr. (1996). Executive functions and working memory:
Theoretical and measurement issues. In GR Lyon & NA Kras-negor (Eds.), Attention, memory, and executive
function (pp. 327–348). Baltimore: Brookes.

Perret, E. (1974). The left frontal lobe of man and the suppression of habitual responses in
verbal categorical behavior. Neuropsychologia, 12, 323–330.
Phillips, L. H. (1997). Do ''frontal tests'' measure executive function? Issues of assessment and
evidence from fluency tests. In P. Rabbitt (Ed.), Methodology of frontal and executive
function (pp. 191–213). Hove, UK: Psychology Press.
Posner, M.I., & Raichle, M.E. (1994). Images of mind. New York: Sci. Am.
Rabbitt, P. (Ed.). (1997a). Methodology of frontal and executive function. Hove, UK: Psychol-
ogy Press.
Rabbitt, P. (1997b). Introduction: Methodologies and models in the study of executive function.
In P. Rabbitt (Ed.), Methodology of frontal and executive function (pp. 1–38). Hove, UK:
Psychology Press.
Reitan, R.M., & Wolfson, D. (1994). A selective and critical review of neuropsychological deficits
and the frontal lobes. Neuropsychology Review, 4, 161–197.
Robbins, TW, James, M., Owen, AM, Sahakian, BJ, Lawrence, AD, McInnes, L., & Rabbitt, PMA
(1998). A study of performance on tests from the CANTAB battery sensitive to frontal lobe
dysfunction in a large sample of normal volunteers: Implications for theories of executive
functioning and cognitive aging. Journal of the International Neuropsychological Society,
4, 474–490.
Roberts, R.J., Hager, L.D., & Heron, C. (1994). Prefrontal cognitive processes: Working memory
and inhibition in the antisaccade task. Journal of Experimental Psychology: General, 123,
374–393.
Rogers, R.D., & Monsell, S. (1995). Costs of a predictable switch between simple cognitive
tasks. Journal of Experimental Psychology: General, 124, 207–231.
Rogers, R.D., Sahakian, B.J., Hodges, JR, Polkey, C.E., Kennard, C., & Robbins, T.W.
(1998). Dissociating executive mechanisms of task control following frontal lobe damage
and Parkinson's disease. Brain, 121, 815-842.
Machine Translated by Google
100 MIYAKE ET AL.

SAS Institute (1996). SAS/ STAT software: Changes and enhancements through release 6.11.
Cary, NC: SAS Inst.
Schachar, R.J., Tannock, R., & Logan, G. (1993). Inhibitory control, impulsiveness, and attention
deficit hyperactivity disorder. Clinical Psychology Review, 13, 721–739.
Shallice, T. (1988). From neuropsychology to mental structure. New York: Cambridge Univ.
Press.
Shallice, T., & Burgess, P.W. (1991). Deficits in strategy application following frontal lobe damage
in man. Brain, 114, 727–741.
Simon, H. A. (1975). The functional equivalence of problem solving skills. Cognitive Psychol- ogy,
7, 268–288.
Smith, E.E., & Jonides, J. (1999). Storage and executive processes in the frontal lobes. Science,
283, 1657–1661.
Spector, A., & Biederman, I. (1976). Mental set and mental shift revisited. American Journal
of Psychology, 89, 669–679.
Stroop, J.R. (1935). Studies of interference in serial verbal reactions. Journal of Experimental
Psychology, 18, 643–662.
Stuss, D.T., & Benson, D.F. (1986). The frontal lobes. New York: Raven Press.
Stuss, D.T., Eskes, G.A., & Foster, J.K. (1994). Experimental neuropsychological studies of frontal
lobe functions. In F. Boller & J. Grafman (Eds.), Handbook of neuropsychology (Vol. 9, pp.
149–185). Amsterdam: Elsevier Science.
Stuss, D.T., Shallice, T., Alexander, M.P., & Picton, T.W. (1995). A multidisciplinary approach to
anterior attentional functions. Annals of the New York Academy of Sciences, 769, 191–211.

Teuber, H.-L. (1972). Unity and diversity of frontal lobe functions. Acta Neurobiologiae Ex-
perimentalis, 32, 615–656.
Towse, J. N. (1998). On random generation and the central executive of working memory.
British Journal of Psychology, 89, 77–101.
Towse, J.N., Hitch, G.J., & Hutton, U. (1998). A reevaluation of working memory capacity
in children. Journal of Memory and Language, 39, 195–217.
Towse, J.N., & Neil, D. (1998). Analyzing human random generation behavior: A review of methods
used and a computer program for describing performance. Behavior Research Methods,
Instruments, & Computers, 30, 583–591.
Turner, M. L., & Engle, R. W. (1989). Is working memory capacity task dependent? Journal of
Memory and Language, 28, 127–154.
Van der Linden, M., Collette, F., Salmon, E., Delfiore, G., Degueldre, C., Luxen, A., & Franck, G.
(1999). The neural correlates of updating information in verbal working memory.
Memory, 7, 549–560.
Welsh, M.C., Pennington, B.F., & Groisser, D.B. (1991). A normative-developmental study of
executive function: A window on prefrontal function in children. Developmental
Neuropsychology, 7, 131–149.
Yntema, DB (1963). Keeping track of several things at once. Human Factors, 5, 7–17.
Zacks, R. T., & Hasher, L. (1994). Directed ignoring: Inhibitory regulation of working memory. In D.
Dagenbach & TH Carr (Eds.), Inhibitory processes in attention, memory, and language (pp.
241–264). San Diego, CA: Academic Press.
Accepted December 2, 1999