Elsevier

Cognitive Psychology

Volume 51, Issue 1, August 2005, Pages 42-100
Cognitive Psychology

On the capacity of attention: Its estimation and its role in working memory and cognitive aptitudes

https://doi.org/10.1016/j.cogpsych.2004.12.001Get rights and content

Abstract

Working memory (WM) is the set of mental processes holding limited information in a temporarily accessible state in service of cognition. We provide a theoretical framework to understand the relation between WM and aptitude measures. The WM measures that have yielded high correlations with aptitudes include separate storage-and-processing task components, on the assumption that WM involves both storage and processing. We argue that the critical aspect of successful WM measures is that rehearsal and grouping processes are prevented, allowing a clearer estimate of how many separate chunks of information the focus of attention circumscribes at once. Storage-and-processing tasks correlate with aptitudes, according to this view, largely because the processing task prevents rehearsal and grouping of items to be recalled. In a developmental study, we document that several scope-of-attention measures that do not include a separate processing component, but nevertheless prevent efficient rehearsal or grouping, also correlate well with aptitudes and with storage-and-processing measures. So does digit span in children too young to rehearse.

Introduction

Baddeley and Hitch (1974) highlighted a key theoretical construct, working memory (WM), which can be described generally as the set of mechanisms capable of retaining a small amount of information in an active state for use in ongoing cognitive tasks (though it now means different things to different investigators; see Miyake & Shah, 1999). If sufficient information cannot be retained in WM and integrated, it is assumed that various problems cannot be solved, and that reading or language comprehension cannot be completed. An important approach to WM has blossomed, in which experimental and psychometric methods are synthesized (e.g., Conway et al., 2002, Cowan et al., 1998, Engle et al., 1999, Miyake et al., 2001).

Research on WM suggests that the measures used most often to examine individual differences have both strengths and weaknesses. A main type of strength is their strong correlation with intellectual aptitude tests, and a main type of weakness is the difficulty encountered in analyzing and interpreting WM test results. This difficulty stems largely from the reliance on dual tasks in the measurement of WM capacity (which include separate storage-and-processing task components). We will argue that the research literature provides hints that the strengths can be retained without using storage-and-processing measures. We will offer a theoretical framework for doing so, and for measuring WM in a more meaningful way than is found with current measurement practices. The theoretical framework is based on the notion of an adjustable attentional focus and on measures of the storage capacity of attention or its scope. The predictions tested in the present article pertain to the scope of attention, whereas the adjustable nature of the focus allows consistency with other highly relevant research (e.g., Kane, Bleckley, Conway, & Engle, 2001).

We do not judge the success of this endeavor by whether storage-and-processing measures or the proposed alternative, scope-of-attention measures, pick up more variance in aptitude tasks. Rather, success will be judged by whether the variance that is picked up contributes to our understanding of the processes underlying the generally observed relation between WM and intelligence. By detailing some conditions in which tasks that do not include separate storage-and-processing components are successful in correlating with intelligence, we provide information about what mechanisms could or could not be indispensable parts of that relation. The storage-and-processing tasks could be the best WM predictors of aptitudes but still could be relatively undesirable tasks if much of their predictive value comes from mechanisms that are not central to the concept of WM (e.g., domain-specific skill in reading or arithmetic). We examine the relations between traditional WM measurement methods and the ones we favor.

A few papers following Baddeley and Hitch (1974) were crucial in drawing the field’s attention to the strong relation between individual differences in WM performance, on one hand, and individual differences in performance on psychometric indices of scholastic and intellectual aptitudes, on the other (e.g., Case et al., 1982, Daneman and Carpenter, 1980, Kyllonen and Christal, 1990). Search of a popular electronic data base (PsycInfo) carried out on January 30, 2004 showed that the number of articles or dissertations that included both the phrase “working memory” and the phrase “individual differences” in the title or abstract have increased steadily since 1980. In successive 5-year periods between 1980 and 2000, the numbers of entrees were 16, 21, 60, and 113. Based on subsequent entries (97), our projection for 2001–2005 is 160, a tenfold increase from 20 years previous.

The observation of a strong relation between WM and aptitude tasks has been gained at a theoretical cost, though. It is not at all clear why the WM tasks work. Daneman and Carpenter (1980) suggested that, given an assumed structure of WM in which storage and processing shared resources, an adequate test of WM must tax both storage and processing. They therefore designed reading- and listening-span methods in which a processing task (comprehending sentences) was to be carried out interleaved between items in a storage task (retaining the last word of each sentence for later recall). Case et al. (1982) similarly combined counting of displays of objects and recall of the sums (counting span), and Turner and Engle (1989) combined arithmetic with word recall (operation span). Performance levels on these “storage-and-processing” types of tasks are highly interrelated. They correlate well with various mental aptitudes in adults, at a considerably higher level than do simple memory-span tasks, in which a list of items is simply presented for recall on each trial (for a meta-analysis see Daneman & Merikle, 1996). However, as we will discuss, it is far from clear how the storage-and-processing tasks are carried out, what aspects of the tasks account for the high correlations with aptitudes, and whether all such WM tasks operate similarly.

There have been a number of studies applying the logic of storage-and-processing types of tasks and other complex WM tasks in childhood development (e.g., Ashcraft and Kirk, 2001, Bayliss et al., 2003, Case et al., 1982, Daneman and Hannon, 2001, Gathercole and Pickering, 2000, Hitch et al., 2001, Kail and Hall, 2001, Swanson, 1996, Towse et al., 1998). The importance of such study is partly that it can help in predicting and clarifying childhood aptitudes and disabilities (e.g., Swanson, 2005), and partly that it can help in clarifying the processes of WM per se. It can do the latter in several ways. First, cross-age variance provides a wider range of variability in task performance than one finds among adults (a research priority advocated, for example, by Pascual-Leone, 2005). Second, cross-age differences in processing can shed light on the mechanisms of WM. For example, one possible basis of the higher predictive value of storage-and-processing tasks as opposed to single tasks such as digit span is not the inclusion of processing per se, but rather the fact that the processing component interrupts covert verbal rehearsal (cf. Baddeley, 1986) and thereby allows other processes to play a dominant role. Given that young children do not engage in much covert verbal rehearsal, or do so only inefficiently (e.g., Cowan et al., 1994, Cowan and Kail, 1996, Flavell et al., 1966, Gathercole et al., 1994, Guttentag, 1984, Henry, 1991, Hulme and Muir, 1985, Ornstein and Naus, 1978), the superiority of storage-and-processing tasks for young children could be called into question.

It can be argued that the storage-and-processing measures are theoretically ambiguous. They have been based on the premise that processing and storage both tap a common resource (see Daneman & Carpenter, 1980). However, within that framework, it has long been thought that some individuals require more of that common resource than others to accomplish a given task because expertise leads to efficient use of the resource (e.g., Case et al., 1982). For example, the interpretation offered by Daneman and Carpenter was that the reading- or listening-span memory score measures how much capacity is available when combined with linguistic comprehension, and therefore provides an index of the efficiency of that comprehension. However, it is possible that there are individual differences in storage itself, even with linguistic processing equated. This might take the form of individual differences in the passive storage of information, as in the phonological loop (Gathercole & Baddeley, 1993), or individual differences in how many unassociated units or chunks can be held in the focus of attention (Cowan, 2001). One person might excel in storage and another in processing, yet both might obtain the same score on a storage-and-processing type of WM task. A WM span score would not indicate which type of person had been tested.

Moreover, there is a fundamental difficulty in interpreting the results of dual tasks (of which, storage-and-processing tasks are one type). Performance on a task can be impaired by a concurrent task if there is a need for the tasks to share a general, cross-domain resource such as attention, specific resources such as verbal or spatial processing, or both of these (for discussion see Cowan, 2001). One way to tell if a general resource exists is to determine whether dual-task interference can be obtained using two tasks that share nothing in the way of more specific processes but, of course, that criterion is difficult to meet with assurance. Some studies have shown little or no dual-task interference (e.g., Cocchini et al., 2002, Duff and Logie, 2001, Farmer et al., 1986) but that proves only that not all tasks require a substantial amount of a general resource such as attention. Other studies have shown dual-task interference even between two tasks differing in modality (Jolicoeur, 1999) and differing in the use of verbal versus non-verbal materials (e.g., Jefferies et al., 2004, Morey and Cowan, 2004, Sirevaag et al., 1989, Stevanovski and Jolicoeur, 2003). Like Cowan (2001) and Kane et al. (2004), we find it most parsimonious to assume that there is a general, amodal attentional resource, which has been assumed by other investigators for very different reasons as well (e.g., Tombu & Jolicoeur, 2003). However, in the literature on dual-task measures of WM, including the storage-and-processing tasks, there has been very little effort to use task combinations in which the storage-and-processing tasks share neither their sensory modality nor the types of coding that they most naturally elicit, so the basis of interference between storage and processing is usually unclear and probably complex.

Making matters worse, the types of domain-specific coding that traditionally have been associated with WM (Baddeley, 1986) may be differentially related to a general resource. Miyake et al. (2001) showed that spatial storage, with or without concurrent processing, tends to be closely related to executive function, in contrast to the separation between phonological storage and executive function. The present paper offers a rationale for measures that are simpler because they concentrate on storage, with intrinsic means to limit rehearsal and grouping.

A wide variety of possibilities for the theoretical interpretation of storage-and-processing WM tasks have been discussed in the literature. There is some evidence that participants do not always engage in attention-sharing between processing and memory maintenance, as one might assume; in some tasks, they appear to switch attention between storage and processing, provided that the processing task included rich semantic cues for retrieval of the memoranda (Copeland and Radvansky, 2001, Cowan et al., 2003, Hitch et al., 2001). There has been some concern that the feature of the processing task that impairs memory performance is the imposition of a time delay during which rehearsal is impossible (Towse et al., 1998) or the imposition of high-frequency retrievals or processing during which rehearsal is impossible or storage is somehow interfered with (Barrouillet et al., 2004, Saito and Miyake, 2004). Other studies have suggested that what is critical is the amount of proactive interference that is already present by the time that long list lengths are used, and individual or age differences in the ability to overcome that interference by inhibiting it (Conway and Engle, 1994, Lustig et al., 2001, May et al., 1999). Bayliss et al. (2003) suggest that domain-specific storage and cross-domain processing work separately, not from a common resource. Given these possibilities, it is important to investigate what we might learn from potentially simpler tasks.

There is already some evidence that some WM tasks that do not include separate storage-and-processing components are nevertheless capable of yielding relatively high correlations with aptitude tests. However, that evidence is still sparse and not well-integrated theoretically. Mukunda and Hall (1992) carried out a meta-analysis of the within-age correlations in adults and children between various WM tasks and various intellectual aptitude tasks, and found that one measure, running memory span (11 tests, R = .40) performed about as well as reading span (11 tests, R = .43) and better than operation span (6 tests, R = .23), counting span (3 tests, R = .28), or digit span (53 tests, R = .22). Running memory span (Pollack, Johnson, & Knaff, 1959) is a procedure in which a list of an unpredictable, long length is presented, the task being to recall as many items from the end of the list as possible after the list terminates. Unfortunately, as in most meta-analyses, different measures had to be drawn from different studies, making the comparability of the measures questionable.

Haarman, Davelaar, and Usher (2003) developed a “conceptual span” task in which list recall was to be organized according to semantic elements (e.g., in an example they offer: “lamp, pear, tiger, apple, grape, elephant, horse, fax, phone,” FRUIT? Correct answer: “apple, pear, grape”). Conceptual span yielded higher correlations than reading span in the prediction of aspects of reading comprehension that required unconnected semantic elements to be retained (e.g., verbal problem-solving; anomaly detection). The premise behind this task appears to be that it is a way to measure attentional capacity; attending to information may be tantamount to selecting objects by categorizing them (see Logan, 2002). Work is ongoing in our laboratory and in other laboratories to determine whether this sort of task correlates well with a wider array of aptitude tasks. Other laboratories, as well, have begun to experiment with a variety of new WM procedures, some of which do not involve a dual task (e.g., Oberauer, Süß, Wilhelm, & Wittmann, 2002).

In the literature, there appears to be no theoretical framework allowing performance on all of these WM tasks to be understood. Toward that end, an adjustable-attention framework will be described. It leads to the advocacy of certain relatively simple WM tasks that might be more easily interpretable than storage-and-processing tasks.

Exposition of the theoretical framework depends on an understanding of (1) the work that has been done to link attention to WM, (2) the reasons why that work has not yet produced a meaningful scale of WM capacity for use in examining individual differences in aptitude, and (3) theoretical underpinnings for constructing such a meaningful scale. We discuss these.

Attention is a controversial concept but large-scale treatments of it can be found in the literature (e.g., Cowan, 1995, Luck and Vecera, 2002, Näätänen, 1992, Pashler et al., 2000, Shiffrin, 1988). It is beyond the scope of this article to re-review it. We use the term attention to refer to selective attention, in which some information is selected for processing at the expense of less-than-optimal processing of other information. What is important for the present approach is that at least two dimensions of attention can be discerned: the control of attention and its scope. The recent research literature on the link between attention and WM has focused primarily on the control of attention whereas, we suggest, a meaningful scale of WM capacity depends on an emphasis on the scope of attention.

The control of attention was an important element of early theories of information processing (e.g., Atkinson & Shiffrin, 1968) and is embodied in the central executive component of theoretical conceptions of WM (e.g., Baddeley, 1986, Cowan, 1988, Cowan, 1995). A great deal of recent research has converged on the importance of the control of attention in carrying out the standard type of WM task involving separate storage-and-processing components. Six strands of research on this topic can be enumerated. (1) These WM tasks correlate highly with aptitudes even when the domain of the processing task (e.g., arithmetic or spatial manipulation) does not match the domain of the aptitude test (e.g., reading) (Daneman and Merikle, 1996, Kane et al., 2004). That is to be expected if the correlations are due to the involvement of processes of attention that cut across content domains. (2) An alternative account of the correlations based entirely on knowledge can be ruled out. Although acquired knowledge is extremely important for both WM tasks and aptitude tasks (e.g., Ericsson & Kintsch, 1995), correlations between WM tasks and aptitude tasks remain even when the role of knowledge is measured and controlled for (Hambrick & Engle, 2001). (3) On tasks involving memory retrieval, dividing attention impairs performance in individuals with high WM spans but has little effect on individuals with low WM spans (Rosen & Engle, 1997). This suggests that low-span individuals do not make use of attention in the same way that high-span individuals do within these tasks. (4) Latent variable analyses (Conway et al., 2002, Engle et al., 1999) show that the portion of the variance in storage-and-processing WM tasks that is most responsible for its correlation with aptitude tests is not the portion that is in common with simple list-recall span tasks, but the portion unique to the storage-and-processing tasks. This has been taken to indicate that what is most critical is the ability to retain information in memory even while carrying out processing, an ability that would seem to require the control of attention. (5) Differences between high- and low-span individuals (measured by storage-and-processing WM tasks) have been obtained even in tasks in which the only apparent storage requirement is to hold onto the goal of the task. These differences have been obtained in situations in which there was some type of interference with goal maintenance to be overcome with the help of attentional control (Conway et al., 2001, Kane et al., 2001, Kane and Engle, 2003). (6) Recent research has suggested that the functioning of frontal-lobe areas related to executive control of attention differ between high- and low-span individuals (as measured by storage-and-processing WM tasks). Kane and Engle (2002) reviewed evidence converging on this point. A more recent neuroimaging study by Gray, Chabris, and Braver (2003) explicitly found WM span differences in neural functioning localized in particular frontal areas, which emerged only in situations in which there was a high level of proactive interference to be overcome.

Although the emphasis on the control of attention is supported by the evidence, it does not lead naturally to a meaningful measure of the capacity of WM, if that capacity is defined as a number of items in WM. Indeed, the dependent measure of the ability to maintain a goal differs widely from one test situation to the next (cf. Conway et al., 2001, Kane et al., 2001, Kane and Engle, 2003). We argue that, to measure WM capacity on a common scale, it is necessary to consider the scope of attention instead of its control.

Theoretical work on information processing has long been divided on the role of attention in short-term storage. James (1890) discussed primary memory as the trailing edge of consciousness, thoroughly related to the concept of attention. This was not necessarily true of Hebb’s (1949) concept of short-term storage as a reverberating neural circuit. Miller’s (1956) discussion of the finding that people could retain only about 7 items also seems neutral to the involvement of attention in short-term retention. However, Broadbent’s (1958) conception of a limited-capacity storage faculty seemed to view it as being the direct consequence of attention, given that an attention filter stood between a large-capacity store of information coming directly from the senses and a small-capacity store of information coming after that information was filtered. In contrast, in the conception of WM as a multi-component system (e.g., Baddeley, 1986, Baddeley and Logie, 1999), attention tended to be associated with executive control but not with storage per se; that storage was assumed to be automatic once the information was entered into it, and it was assumed to be time-limited instead of capacity-limited. Cowan, 1988, Cowan, 1995, Cowan, 1999 advocated both attention-free and attention-dependent forms of storage, with only the attention-dependent forms limited in capacity per se.

Broadbent (1975) suggested that there is a form of storage that is limited to 3 items. Cowan (2001) reviewed a great deal of literature in support of the notion that there is a form of storage that typically can include 3–5 separate units, or chunks, of information in normal adults, and proposed that the special form of storage limit may be the capacity of the focus of attention, i.e., the scope of attention. Such a capacity would not replace the attention-free stores of Baddeley (1986) as the two presumably could co-exist (i.e., phonological and visuospatial buffers might exist apart from attention).

At least four related assumptions about this concept of the scope of attention are subsumed under the attention-adjustment hypothesis under investigation here: (1) that there is a limit in the capacity of the focus of attention, (2) that this limit varies between individuals, (3) that measures of this capacity are theoretically and empirically related to storage-and-processing measures of WM, and (4) that the common variance between these measures is related to intellectual aptitude measures. The first of these was the topic of previous work (especially Cowan, 2001, Cowan et al., 2004) whereas the other three are topics of the present investigation. There may be other portions of variance related to specific skills required by a particular measure, especially in the storage-and-processing measures, given that they each include a separate processing component; but our research goal is to determine if there is nevertheless considerable common variance among tasks that can be attributed to the scope of attention.

The control-of-attention and scope-of-attention hypotheses are not necessarily in conflict. Individuals who excel at controlling attention could be the same ones who have the largest scope of attention. This could be the case, for example, if attention can be adjusted. When necessary, it might zoom in to hold on to a goal in the face of interference, and perhaps a minimum of related data that is required. However, when there is no interference with the goal and the task has been well-practiced, the focus of attention could afford to zoom out to apprehend multiple items at once, up to its limit. (For a mathematical model of WM based on this concept, see Usher, Haarmann, Cohen, & Horn, 2001; for relevant experimental work see Chen, 2003.) This concept is related to the zoom lens model of attention of Eriksen and St. James (1986). We both propose that a zoomed-out setting has more breadth or covers more objects, but has less intensity or precision of processing of each object, than a zoomed-in setting. The main difference is that we propose that the focus of attention is not specific to visual processing or to its spatial aspects, but covers all modalities and codes. Our concept might also be related to the model of LaBerge and Brown (1989), who propose that there is a gradient of processing that becomes less intense as one gets further from the focus of attention. The gradient can be set narrowly or widely, and it does not matter to us whether a zoom description or a gradient-setting description is used.

As will be discussed, attention cannot be spread infinitely thinly but is limited to about 3–5 chunks of information (Cowan, 2001). It may be that the people who are good at locking attention onto a goal during adversity are the same ones who are good at zooming attention out to apprehend the maximal number of items, or who have the widest attentional focus. If either of these possibilities is true, there should be a strong relation between attentional control and the measured scope of attention. If one assumes that storage-and-processing measures are useful because they tap into attentional control, these measures therefore still should correlate with measures of the scope of attention.

Cowan (1995) reviewed literature suggesting that the scope of attention does not rely primarily upon frontal lobe mechanisms as the control of attention does but, rather, upon parietal lobe mechanisms. Although this distinction is not clean and absolute, frontal lobe damage often results in dysfunctions of the central executive control of attention, whereas parietal lobe damage more often results in dysfunctions of consciousness, such as unilateral neglect and anosognosia (in which an individual shows no sign of awareness of an ostensibly obvious disability, such as paralysis of a limb). Ruchkin, Grafman, Cameron, and Berndt (2003) summarized physiological evidence leading to the idea that the frontal region does not contain the information in WM directly but contains pointers to that information in more posterior regions of the cortex, potentially reinforcing the notion of a difference between a frontal control mechanism and a posterior seat of attention.

It is an open question whether the integrity of the frontal and parietal portions of the attentional system are distinct or whether they function as an integrated system, such that their levels of functioning are strongly correlated among normal individuals. It is also an open question whether individual differences in the measured scope of attention are due primarily to differences in the parietal mechanisms and to an intrinsic limit in the scope of attention, or due primarily to differences in frontal mechanisms and executive control needed to adjust (zoom in or out, and aim) the scope of attention. An unnecessarily zoomed-in focus when there is relatively little interference to be handled would decrease the apparent scope of an individual’s attention in the task. The present correlational investigation does not attempt to resolve these ultimate questions, but it provides an empirical background to investigate them by documenting the relation between WM tasks with very different formats, including scope-of-attention measures that do not include a dual task but, nevertheless, correlate well with aptitude measures.

A critical question is how to measure the scope of attention. Cowan (2001) conceived of this measure as the result of a limited-capacity attentional focus extracting chunks of information from a field of activated features in memory in order to allow an explicit memory response. The form of the activated features could include sensory, phonological, orthographic, visuospatial, semantic, or lexical features held outside of the focus of attention. (Note that common terminology often refers to “automatically activated” memory as synonymous with “passively held” memory.) The act of attending to representations within this activated memory presumably results in the construction of object representations in the focus of attention (cf. Kahneman, Treisman, & Gibbs, 1992). It is the number of objects or chunks that can be extracted and held at one time in the focus of attention that we hope to measure.

There are assumptions that must be met before it can be assumed that this use of attention during retrieval will provide a meaningful measure of the scope of attention. It must be assumed (1) that the objects or chunks of information recalled are identifiable, (2) that the activation of features persists long enough for the maximal extraction of information into the focus of attention, and (3) that the focus of attention does not engage in multiple cycles of retrieval-and-recall from the field of activated features on one trial. We will discuss each of these assumptions in turn, and then show how they apply to tasks that will be used in the present study.

Attention is potentially involved in the reception, maintenance, and retrieval of information. During reception and maintenance, it might be used to recode the items so as to convert them to a smaller number of separate chunks of information. If one receives the digit list 1 3 8 2 4 6, for example, it might be rapidly recoded into the three two-digit numbers 13, 82, and 46 (cf. Miller, 1956). For that reason, Miller’s magical number seven cannot be taken as evidence that seven separate chunks of information are saved. Indeed, Miller does not appear to have intended to make that claim (Miller, 1989), despite the way his 1956 article has often been portrayed. Attention also is needed to initiate a rehearsal loop that preserves the information in a passive store, even if the rehearsal loop then proceeds automatically (cf. Baddeley, 1986).

If we wish to examine how many chunks of information can be formed from the features in activated memory after they are retrieved into the focus of attention at the time of recall, it is necessary to know the nature of information in the activated memory record. Although it is possible to carry out a learning study to estimate how the items are grouped into chunks in serial recall (Cowan et al., 2004), a simpler method is to limit grouping and rehearsal processes during presentation and maintenance of the stimuli. Critically, if familiar items are presented and conditions at encoding and maintenance prevent grouping and rehearsal, then each item constitutes a separate chunk that can be retrieved into the focus of attention. (If the items were not sufficiently familiar, some items might be represented as multiple chunks; and if grouping and rehearsal were not prevented, multiple items might be grouped together to form a smaller number of chunks.) Grouping and at least the initiation of rehearsal presumably require attention (e.g., Guttentag, 1984, Naveh-Benjamin and Jonides, 1984) so, to limit grouping and rehearsal, attention must be diverted or thwarted.

Various procedures appearing to meet that requirement led to convergent estimates of about 4 items (i.e., chunks) recalled in adults, and fewer in children (Cowan, 2001, Cowan et al., 2002). It is beyond the scope of this article to review the many convergent procedures examined in these reviews but, later, the principles will be illustrated with three such tasks used in the present study. There are several ways in which the effects of attention during stimulus encoding and maintenance can be minimized (cf. Cowan, 2001), including diverting attention, making sequences rapid and unpredictable, and presenting a brief array of items.

Attention can be diverted from the stimuli to be remembered during their presentation. The classic example of this is in tests of memory for the unattended channel in selective listening with dichotic presentation (e.g., Glucksberg and Cowen, 1970, Norman, 1969). The assumption is that a stream of sensory memory for the most recent events forms even for unattended stimuli, and that attention can be switched to the past few seconds of that sensory memory stream to allow the conversion of that information to a categorized form for recall (Broadbent, 1958).

A sequence of items can be presented too quickly for the items to be rehearsed or grouped. This is especially effective if the number of items to be presented is unpredictable, so that each item cannot be classified as occupying a certain slot within a known list structure. Although attention has not been diverted from the stimuli, it is ineffective at producing grouping or rehearsal. The classic example of this sort of technique is running memory span (Pollack et al., 1959), which will be analyzed at greater length shortly.

Last, a spatial array of information containing too many items to be apprehended at once can be presented. Perhaps the first such procedure was carried out by Jevons (1871), who picked up a small handful of beans and threw them on the table to be enumerated as quickly as possible. Using himself as the subject, he found that only a small number could be counted. A better-controlled example is the classic study by Sperling (1960), who presented an array of characters to be recalled. The main point of the study was to use partial-report cues to determine how much information persisted in sensory memory but, without a partial-report cue to enable attention at encoding, only about 4 items could be recalled.

Although the scope of attention presumably limits how many chunks of information can be extracted from the activated memory field, or apprehended, at once, there is another important limit that must be considered. The activated memory record could fade before more information is drawn into the focus of attention. However, at least two types of findings argue against a limitation in the persistence of activated features as the source of the memory limit in procedures reviewed by Cowan (2001): constant capacity across different decay conditions, and insensitivity to stimulus-array duration. These will be described in turn.

In conditions in which the activated features are short-lived, it appears that the similarity in capacity limits has held up despite very different decay parameters. For example, whereas Sperling (1960) found that the benefit of a partial-report cue (which allows sensory memory to be used efficiently) was effective only up to a fraction of a second, Darwin, Turvey, and Crowder (1972) carried out a similar procedure with a spatio-temporal array of spoken digits and found that a partial report cue was effective up to about 4 s. Thus, decay of the relevant sensory memory trace was much slower in audition. Nevertheless, in both modalities, the whole-report limit was about 4 items. It might be possible to account for these results with a theory in which the auditory modality has both slower sensory decay than vision and, for some reason, commensurately slower retrieval of information from sensory memory. However, a simpler hypothesis is that retrieval into the focus of attention is fixed across modalities, at about 4 separate items in the typical, normal adult.

Second, in several array procedures (e.g., Luck and Vogel, 1997, Sperling, 1960) the duration of the array has been varied from less than 100 ms to about a half second, with little or no change in the resulting memory limit. This seems inconsistent with the view that the cause of a limited capacity is insufficiency in the duration of the temporary memory record upon which attention is focused.

The third assumption has to do with retrieval from activated memory into the focus of attention. There is the theoretical possibility that the contents of attention are recalled, after which the participant’s focus of attention returns to the activated memory field again for a second cycle of retrieval, or for multiple cycles. If this happened, then the memory response would have to be taken as an overestimate of the scope of attention. There are several arguments against this, and in favor of the suggestion that the focus of attention can consult activated memory only once. Theoretically, the act of recall may interfere with the activated-memory record (Cowan, 2002), or there may be a phenomenon analogous to inhibition of return (Posner, Rafal, Choate, & Vaughan, 1985), which can occur not only for spatial locations but also for previously attended objects (Tipper, Driver, & Weaver, 1991).

For unattended auditory sequences, the acoustic memory record persists for a number of seconds (Cowan et al., 1990, Cowan et al., 2000, Darwin et al., 1972, Glucksberg and Cowen, 1970, Norman, 1969). If the focus of attention could repeatedly access that record, it would be difficult to explain the limit to about 4 items recalled from unattended lists (Cowan, 2001, Cowan et al., 1999). Also, some procedures include a cue to examine only one object within the memory record of an array (e.g., Luck & Vogel, 1997) or list (Cowan, Johnson, & Saults, in press), yet a similar capacity limit of about 4 items is obtained.

Finally, in procedures in which recall is from the long-term memory record, so that the memory representation does not decay, we see that the pattern of recall looks different than when recall is from a short-term record. Recall from a long-term memory record includes much more than the 4-chunk limit, though they appear in bursts of about 4. This is the case in recall from a semantic category (Broadbent, 1975, Graesser and Mandler, 1978), from a sequence of digits memorized by a mnemonic expert (Wilding, 2001), and from chess boards that remain present while they are copied (Gobet & Simon, 2000). The present theoretical suggestion is that a single retrieval process occurs when retrieval is from an activated memory field, whereas multiple retrievals can occur when retrieval is from long-term memory, though both share the same capacity limit for each retrieval.

We argue that three tasks that will be used in Experiment 1 appear to conform to the requirements of good measures of the scope of attention. The first operates by diverting attention at the time of the presentation of items to be recalled; the second, by presenting these items at a rapid rate with an unpredictable endpoint; and the third, by presenting these items in a simultaneous array.

In a series of experiments (Cowan et al., 1990, Cowan et al., 1999, Cowan et al., 2000) we have examined memory for spoken lists that were ignored while a silent, visual task requiring phonological processing was carried out. This sort of task serves the same purpose as memory for the ignored channel in selective listening (Glucksberg and Cowen, 1970, Norman, 1969) except that it avoids the problem of acoustic masking between channels. The finding in all of these studies was that there is a memory record of the ignored speech channel but that, in comparison to memory for attended speech, successful memory for ignored speech requires a shorter retention interval between presentation and test (⩽2 s for maximum performance). Cowan et al. (1999) judged the capacity of memory with short test delays and found that it was about 3.5 items in adults, regardless of the list length, and smaller in children.

In the main procedure of Cowan et al. (1999), participants carried out a task in which the learned name of a picture in the center of the computer screen rhymed with the learned name of one of four peripheral pictures (to be selected with a mouse click, followed by a new central picture to be judged). Meanwhile, lists of spoken digits were to be ignored. No response was required for most such lists but, once a minute or so, the rhyming game was replaced by a digit-recall response screen shortly after the onset of the most recent list. By attracting attention to the sensory memory of the list only after it ended, rehearsal and grouping of the list during its presentation could be minimized and memory performance (recall of digits by keypress) presumably was based on the post hoc conversion of items from an auditory sensory memory record into a categorical form in the focus of attention. Each digits was scored as correct only if it was recalled in the correct serial position, so memory of the binding between digits and serial positions in the list was required for correct responding.

The digit list length was individually adjusted and four list lengths per individual were used: a length equal to the participant’s predetermined digit span (using attended lists), and lists 1, 2, or 3 digits shorter than that. The main results of that study, reproduced in Fig. 1, shows that the capacity limit for ignored lists (solid lines) was fairly constant across list lengths and increased with age. A similar finding was obtained when individuals were compared at absolute list lengths, as opposed to lengths determined relative to span (not shown).

At least two points must be established before it can be inferred that these solid lines reflect the scope of attention at the time that the recall cue is presented: first, that the limit in the recall of ignored information is not a result of residual attention to the list during its presentation and, second, that the limit is not a result of a sensory memory deficit. Control conditions and experiments establish both of these points. Regarding residual attention, there was a control condition in which there was no rhyming task and the digit lists were attended (Fig. 1, dashed lines). In contrast to the ignored-speech condition, it can be seen that the number correct did not remain constant across list lengths, but grew with list length. The stark difference from the results of the ignored-speech condition indicate a clear signature of attention during stimulus presentation in this task. In control conditions in which the visual task was to be carried out alone, just before and just after the ignored-speech task, performance closely matched the level on the same task in the presence of speech to be ignored. There were no tradeoffs between the visual and auditory tasks; no indication that individuals scoring better on memory for ignored speech did worse on the rhyming task, nor any indication within participants of better speech memory on trials with poorer rhyming-task performance (see Cowan et al., 2002). It would be difficult to attend to the boring and repetitive spoken lists, most of which require no response, during the more interesting rhyming game, because of habituation of orienting to the sounds (e.g., Cowan, 1988, Sokolov, 1963).

Second, the suggestion that sensory memory decay might have been the limiting factor does not agree with the findings from the short test delays that Cowan et al. (1999) used. Cowan et al., 1999, Cowan et al., 2000 found pronounced bow-shaped serial position functions in memory for ignored speech, so sensory memory for even the early serial positions seemed to remain available at the time of list-recall testing. Cowan et al. (2000) examined the serial position effects across test delays and found severe forgetting of information from both the primacy and the recency portions of the list, in contrast to the stability of primacy effects across test delays that is found for attended lists (Jahnke, 1968). This finding reinforces the conclusion that sufficient sensory memory was available at short test delays so that performance limits were not due to sensory memory decay. Also, Cowan et al. (2000) found that, with the list length equal to a predetermined span for each individual, there was no overall age difference in the loss of information across test delays. (There was a difference at the final serial position but it was not enough to result in a significant age difference in list-wide decay.) In sum, there appears to be a constant capacity for digits within ignored lists; the capacity appears to be related to the scope of attention when it is focused on the sensory memory record of the list, as opposed to sensory memory decay; and this capacity changes with development during childhood.

In running memory span (Pollack et al., 1959), participants receive many verbal items in a list that ends at an unpredictable point, whereupon the items at the end of the list are to be recalled. One might think that running memory span possibly could be carried out through a very active process in which the participant retains the most recent k items and continually updates the retained set by dropping the least-recent item to make room for the newest item. Given that serial order recall is required, the relative serial positions of the items in the retained set would have to be continually updated, also. According to an alternative proposal, though, participants wait passively until the list ends and then retrieve items from the automatically activated memory stream (e.g., from sensory memory).

Two studies greatly strengthen the latter interpretation when the items are presented quickly. First, Mayes (1988) provided evidence that it is reasonable to believe that sensory information can persist for a sufficient duration in this task. He presented lists of spoken or printed digits at a rate of 900 ms/digit in a running span procedure. An advantage for spoken digits over printed digits was obtained at each of the last 7 serial positions, which suggests the presence of an auditory-modality-specific memory code at these positions (for discussion of modality effects see Nairne, 1990, Penney, 1989).

Second, Hockey (1973) gave participants instructions to process the running-span stimuli either in a passive manner (a request that was further elaborated), or in an active manner, with instructions to “Concentrate on the items as they arrive, trying to form them into groups of three” using a rhythmic type of rehearsal. The digits were recorded and presented at rates of 1, 2, and 3 digits per second. The outcome was a very clear crossover interaction of presentation rate and strategy. At slower rates, an active strategy was advantageous whereas, at faster rates, a passive strategy became advantageous. At the fastest rate across 10 serial positions, there was about a 0.3- to 0.4-item advantage for a passive strategy over an active strategy. In our experiments we have used digitally compressed speech to achieve a rate of 4 digits per second so that an active strategy cannot be used to advantage. Although attention is directed to the list during its presentation, it is rendered ineffective in producing rehearsal or grouping, so that the results are functionally equivalent to memory for ignored speech.

The visual-array comparison task of Luck and Vogel (1997), like the task of Sperling (1960), presents a simultaneous array of objects to be remembered on every trial, typically too many to be combined into a smaller number of groups in the time available. Unlike the other tasks we have examined, and unlike Sperling’s task, it is designed in such a way that there is only one response to be made on every trial, avoiding the possibility of output interference. On every trial in the condition that we use, an array of colored squares was presented briefly, followed after an inter-stimulus interval by a second array similar to the first. One square in the second array was encircled (from the onset of that array) and, if the arrays differed at all, it was in the color of the encircled square. The task was to indicate whether the arrays were the same or different. A color could be used more than once in an array so that the location of colors, and not merely their inclusion or omission from the array, had to be remembered for successful performance. Performance fell off as a function of the number of squares in the array above 3 or 4.

A simple model of performance (Cowan, 2001) can be used to estimate how many of the squares from the first array were held in WM. The model is explained in detail in Appendix A. The basic idea is that, if the participant recalls the color of the square in the first array that was at the location corresponding to the encircled square in the second array, then he or she answers correctly; otherwise, he or she guesses. This measure provides capacity estimates in the same range as a wide variety of other tasks reviewed by Cowan (2001) and is fairly constant across array set sizes above the capacity limit of 3–4 items.

Presumably, the arrays are presented too briefly for participants to encode items verbally and recall them as a list (e.g., top-left-blue, middle-green, and so on). We believe that participants must extract information from a visuospatial record into the focus of attention before the presentation of the second array (see Cowan, 2001). There are several relevant findings. Luck and Vogel (1997) found no effect of a 2-digit memory load on visual-array comparison performance. Morey and Cowan (2004) replicated that effect using 2-digit loads repeatedly spoken aloud, and also found no effect of overt, repeated recitation of the participant’s own 7-digit telephone number during the visual-arrays task. Presumably, then, memory of the squares is not assisted by a verbal rehearsal process. Morey and Cowan found that recitation of a random 7-digit memory load did impair performance, presumably because that task taxes attention.

Cocchini et al. (2002) carried out another experiment in which, in some conditions, a visual task was combined with an auditory memory load. Unlike Morey and Cowan (2004), Cocchini et al. suggested that their data could be explained without reference to a cross-domain resource. As in Morey and Cowan, there were conditions in which verbal stimuli were presented before non-verbal stimuli and tested after the non-verbal stimuli, so that the verbal stimuli served as a memory load. It was emphasized that only a small effect of load was obtained. However, performance on the verbal memory load was 80–90% correct in these conditions. In contrast, in Morey and Cowan (2004), the verbal memory load was repeated correctly on only 45% of the trials, a much more difficult load. This may account for the difference in effects of the load in the two studies.

It might still be theoretically possible that the verbal and non-verbal tasks of Morey and Cowan (2004) share some mechanism other than attention. It presumably could not be verbal rehearsal because a great deal of research on articulatory suppression (see Baddeley, 1986) indicates that recitation of the verbal digit load should prevent the application of verbal rehearsal to the visual arrays. It theoretically could be a visuospatial form of rehearsal. Logie, Della Sala, and Wynn (2000) found that visual codes play a role in verbal recall. However, they used visual stimuli whereas Morey and Cowan used spoken stimuli for the digit load. Moreover, Stevanovski and Jolicoeur (2003) found considerable interference in a similar procedure with a simple tone-discrimination task between arrays. It is difficult to explain this effect on the basis of a shared visuospatial rehearsal component (presumably absent from tone discrimination), so it does appear likely that the visual-array comparisons depend to some extent on retention of visual items in the focus of attention during the inter-array interval.

Recent neurophysiological studies with similar visual-array comparison tasks strengthen the assumption that they rely on a capacity-limited, categorized memory for objects in the array. Vogel and Machizawa (2004) provided a cue to attend to colored squares on one side of the screen while ignoring those on the other side and used the extra event-related electrical activity on the side of the scalp contralateral to the attended field as an index of WM maintenance. This activity was found during the time between the first and second arrays and its magnitude matched the behavioral capacity limit. Like behavioral capacity, it increased up to a maximum of 3–4 items and then increased no more. The individual differences in electrical activity (measured as the change in activity between trials with 2 vs. 4 items in the array) correlated with the behaviorally measured capacity at r = .78. Todd and Marois (2004) measured fMRI and found that the visual-array comparison task caused capacity-limited activity in two small brain areas, the intra-occipital and intra-parietal sulci, during the inter-array period. The activity increased as a function of array size only up to a limit of 3–4 items, which again matched the behavioral limit. (Other areas responded in a load-dependent manner but the relation to the behavioral limit was not as clear as for these areas.) The results suggest that the scope of attention in this task is limited not because of a failure of attentional control, but because of an inherent limitation in how many objects can be included in WM at once.

It is important to note that a theory of WM capacity limits based primarily on the scope of attention might be able to explain results from more traditional measures of WM, too. For simple spans, there is the expectation of developmental changes in this regard. For adults, individual differences in rehearsal and the concomitant grouping of items play a role that may mask the influence of individual differences in the scope of attention. However, for young children, who cannot engage in sophisticated rehearsal strategies (Flavell et al., 1966), simple span tasks may well serve as adequate measures of the scope of attention.

The success of storage-and-processing spans may not be for exactly the presumed reason (e.g., Daneman & Carpenter, 1980). Instead, the critical point may be that the processing task prevents continual rehearsal and grouping of the information to be stored during the stimulus presentation. In this case, the items to be recalled must be retrieved as separate chunks (one per item) from the activated memory representation, into the focus of attention at the time of recall.

We completed two developmental experiments with a range of WM tasks and scholastic and intellectual aptitude tasks. The first experiment concentrated on WM tasks that have been tested before (although never in the same study) and a set of applied measures of scholastic ability. The second experiment provided further refinement; it concentrated on WM tasks modified to be more comparable to one another, and a set of verbal and non-verbal aptitude measures drawn from intelligence tests. The main issue is whether measures of WM designed to examine the scope of attention perform in a manner comparable to measures that involve storage and processing together, even though the scope-of-attention measures do not impose a simultaneous dual task. The scope-of-attention measures will be viewed as important even if they do not provide higher correlations with aptitudes than storage-and-processing measures, provided that they pick up much of the common variance of aptitude tasks without as much variance from specific skills that play a role in storage-and-processing spans, but are theoretically distinct from WM (e.g., language comprehension, arithmetic, and counting abilities).

We used not only digit span, but also two storage-and-processing measures of WM (listening span and counting span) and four measures of the scope of attention (memory for ignored speech, in Experiment 1; running memory span, in both experiments; visual-array comparisons, in both experiments; and a tone-sequence analogue to visual-array comparisons, in Experiment 2). In Experiment 1, we used separate measures of aptitude in adults (high school grade percentiles and the American College Test, or ACT) and children (the Cognitive Abilities Test, or CAT). In Experiment 2, we used two verbal and two non-verbal intelligence measures: the vocabulary and pattern-analysis subtests of the Stanford-Binet intelligence scale (Thorndike, Hagen, & Sattler, 1986), the Peabody Picture Vocabulary Test (Dunn & Dunn, 1997), and Ravens’ Progressive Matrices (Raven, Raven, & Court, 1998).

Section snippets

Expectations

Our main expectations for the study are of three types.

Experiment 2: Intelligence measures and WM capacity

A second experiment was conducted to overcome limitations in the first one. In Experiment 1, the scholastic measures could only be examined in children and adults separately and they were applied measures. In the second experiment, we included two measures of verbal ability, the Peabody Picture Vocabulary Test (PPVT: Dunn & Dunn, 1997) and the vocabulary subtest of the Stanford-Binet Intelligence Scales (Thorndike et al., 1986), and two measures of non-verbal ability, Raven’s Progressive

General discussion

There is now a considerable variety of studies in which relationships between different types of WM procedures and intellectual aptitudes, and their development, have been assessed (e.g., Ashcraft and Kirk, 2001, Booth et al., 2000, Caplan and Waters, 1999, Conway et al., 2002, Cowan et al., 2003, Daneman and Hannon, 2001, Daneman and Merikle, 1996, Engle et al., 1999, Fry and Hale, 1996, Gathercole and Pickering, 2000, Haarman et al., 2003, Hedden and Park, 2003, Hitch et al., 2001; Hutton &

References (148)

  • M. Daneman et al.

    Individual differences in working memory and reading

    Journal of Verbal Learning & Verbal Behavior

    (1980)
  • C.J. Darwin et al.

    An auditory analogue of the Sperling partial report procedure: Evidence for brief auditory storage

    Cognitive Psychology

    (1972)
  • S. Glucksberg et al.

    Memory for nonattended auditory material

    Cognitive Psychology

    (1970)
  • F. Gobet et al.

    Five seconds or sixty? Presentation time in expert memory

    Cognitive Science

    (2000)
  • R.E. Guttentag

    The mental effort requirement of cumulative rehearsal: A developmental study

    Journal of Experimental Child Psychology

    (1984)
  • E. Jefferies et al.

    Automatic and controlled processing in sentence recall: The role of long-term and working memory

    Journal of Memory and Language

    (2004)
  • D. Kahneman et al.

    The reviewing of object files: Object-specific integration of information

    Cognitive Psychology

    (1992)
  • P.C. Kyllonen et al.

    Reasoning ability is (little more than) working-memory capacity!

    Intelligence

    (1990)
  • J.R. Anderson et al.

    Atomic components of thought

    (1998)
  • M.H. Ashcraft et al.

    The relationships among working memory, math anxiety, and performance

    Journal of Experimental Psychology: General

    (2001)
  • Baddeley, A. D. (1986). Working memory. Oxford Psychology Series #11. Oxford: Clarendon...
  • A. Baddeley

    The magic number and the episodic buffer

    Behavioral and Brain Sciences

    (2001)
  • A. Baddeley et al.

    The phonological loop as a language learning device

    Psychological Review

    (1998)
  • A. Baddeley et al.

    Working memory

  • A.D. Baddeley et al.

    Working memory: The multiple-component model

  • P. Barrouillet et al.

    Time constraints and resource sharing in adults’ working memory spans

    Journal of Experimental Psychology: General

    (2004)
  • C. Basak et al.

    Subitizing speed, subitizing range, counting speed, the Stroop effect, and aging: Capacity differences and speed equivalence

    Psychology & Aging

    (2003)
  • D.M. Bayliss et al.

    The complexities of complex span: Explaining individual differences in working memory in children and adults

    Journal of Experimental Psychology: General

    (2003)
  • S. Bentin et al.

    The effects of aging and first-grade schooling on the development of phonological awareness

    Psychological Science

    (1991)
  • M.K. Bleckley et al.

    Individual differences in working memory capacity predict visual attention allocation

    Psychonomic Bulletin & Review

    (2003)
  • J.R. Booth et al.

    Developmental differences in visual and auditory processing of complex sentences

    Child Development

    (2000)
  • D.E. Broadbent

    Perception and communication

    (1958)
  • D.E. Broadbent

    The magic number seven after fifteen years

  • S. Cahan et al.

    Age versus schooling effects on intelligence development

    Child Development

    (1989)
  • D. Caplan et al.

    Verbal working memory and sentence comprehension

    Behavioral & Brain Sciences

    (1999)
  • Z. Chen

    Attentional focus, processing load, and Stroop interference

    Perception & Psychophysics

    (2003)
  • G. Cocchini et al.

    Concurrent performance of two memory tasks: Evidence for domain-specific working memory systems

    Memory & Cognition

    (2002)
  • R.A. Conway et al.

    The cocktail party phenomenon revisited: The importance of working memory capacity

    Psychonomic Bulletin & Review

    (2001)
  • A.R.A. Conway et al.

    Working memory and retrieval: A resource-dependent inhibition model

    Journal of Experimental Psychology: General

    (1994)
  • Conway, A. R. A., Bottoms, B. L., Nysse, K. L., Haegerich, T. M., & Davis, S. L. (unpublished). Working memory capacity...
  • D.E. Copeland et al.

    Phonological similarity in working memory

    Memory & Cognition

    (2001)
  • N. Cowan

    Evolving conceptions of memory storage, selective attention, and their mutual constraints within the human information processing system

    Psychological Bulletin

    (1988)
  • Cowan, N. (1995). Attention and memory: An integrated framework. Oxford Psychology Series, No. 26. New York: Oxford...
  • N. Cowan

    An embedded-processes model of working memory

  • N. Cowan

    The magical number 4 in short-term memory: A reconsideration of mental storage capacity

    Behavioral and Brain Sciences

    (2001)
  • N. Cowan et al.

    Constant capacity in an immediate serial-recall task: A logical sequel to Miller (1956)

    Psychological Science

    (2004)
  • Cowan, N., Elliott, E. M., & Saults, J. S. (2002). The search for what is fundamental in the development of working...
  • Cowan, N., Johnson, T. D., & Saults, J. S. (in press). Capacity limits in list item recognition: Evidence from...
  • N. Cowan et al.

    Covert processes and their development in short-term memory

  • N. Cowan et al.

    Properties of memory for unattended spoken syllables

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (1990)
  • Cited by (887)

    View all citing articles on Scopus

    Some of the data from the conventional WM measures (counting-, listening-, and digit-span tasks) and the scholastic tests in Experiment 1 were reported by Cowan et al. (2003), which focused on response timing. These measures are included here in order to examine their relation to less conventional tasks in the same study that have not been reported previously: the running-span, ignored-speech, and visual-array tasks that we used to assess the capacity of the focus of attention. This study was conducted with funding from NIH Grant HD-21338 awarded to the first author. We thank Jebby Arnold, Ryan Brunner, Troy Johnson, Matt Moreno, and Jennifer Norris for excellent assistance and we thank Moshe Naveh-Benjamin and Jeff Rouder for helpful comments.

    1

    Andrew Conway is now at Princeton University.

    View full text