Controlled attention and storage: An investigation of the relationship between working memory, short-term memory, scope of attention, and intelligence in children

Abstract

Working memory (WM) has received considerable attention in psychological research, a core finding being a close relationship between WM and measures of complex cognition. However, only a limited amount of studies investigated this relationship in samples of children. This study explored the contribution of storage-and-processing tasks (WM), a measure of the scope of attention (visual array comparison task), and short-term memory (STM) tasks of supraspan length to 275 (8 to 13 years old) school children's fluid and crystallized intelligence. The results showed that a two-factor structure of memory, consisting of a WM (storage-and-processing as well as scope of attention tasks) and a STM factor, showed the best fit. WM was a strong predictor of fluid (Gf) and crystallized (Gc) intelligence both when modeled separately and when modeled as a residual factor controlling for STM variance. Further, STM interacted with age and was unrelated to Gf in children older than 11 years, whereas the effect of WM was not moderated by age. The results suggest that (1) STM and WM are separable but highly-related constructs, (2) secondary memory processes (e.g., storage and retrieval) along with controlled attention are hallmark predictors of intelligence in children, and (3) STM effects on Gf are moderated by age.

Introduction

Working memory (WM) has been commonly referred to as a processing resource of limited capacity that enables the storage and simultaneous manipulation of information (Baddeley, 1986, Engle et al., 1999). WM capacity has usually been measured with complex span tasks that consist of both a storage and a processing component. In contrast, short-term memory (STM) is viewed as the ability to keep a limited amount of information in a passive storage without additional processing (Unsworth & Engle, 2007). STM is usually measured with simple span tasks that require storage of information.

Several studies have shown that complex span tasks correlate more highly with measures of higher order cognition than simple span tasks. For example, Engle et al. (1999) found that after controlling for STM variance in the WM factor, the WM residual was still significantly correlated with measures of fluid intelligence. In contrast, a STM residual was no longer significantly related to fluid intelligence. According to Engle et al. (1999), the residual WM variance corresponds to controlled attention, which is defined as “an ability to effectively maintain stimulus, goal, or context information in an active, easily accessible state in the face of interference” (Kane, Conway, Bleckley, & Engle, 2001, p. 180). However, other studies have found that simple span tasks correlate substantially and nearly as well as complex span tasks with higher order cognition (Colom et al., 2008, Mogle et al., 2008, Tillman et al., 2008). In addition, Ackerman, Beier, and Boyle (2005) reported nearly identical meta-analytic correlations between fluid intelligence (Gf) and WM (r = .50) as compared to Gf and STM (r = .49), respectively. It is therefore not entirely clear whether WM and STM differ in their relationship with intelligence.

There is some evidence that the relationship between WM, STM, and intelligence may be moderated by age and therefore is subject to change across development. Alloway, Gathercole, and Pickering (2006) reported a very high correlation between WM and visuospatial STM factors at ages 4 to 6 (r = .97) that dropped to r = .71 at ages 9 to 11. A possible explanation for these findings could be that because memory span is shorter in younger children, they are required to use effortful retrieval strategies earlier than older children with longer memory spans. According to this position, in younger children, STM tasks should be good predictors of higher cognitive functioning and WM should not explain much additional variance. In older children, however, STM and WM may be seen as separable constructs, and because older children have higher STM spans and a more developed executive system, the importance of STM relative to WM can be expected to decrease. In line with these assumptions, Bayliss, Jarrold, Gunn, and Baddeley (2003) found no significant relationship between a residual WM factor and Gf in a sample of children 7 to 9 years old. Further, Hornung, Brunner, Reuter, and Martin (2011) reported that in children at the end of kindergarten, the relationship between WM and intelligence is primarily explained by STM capacity. However, several findings contradict this position. Shahabi, Abad, and Colom (2014) found that STM predicted individual differences in Gf substantially in both younger (8 years) and older (12 years) children, whereas WM was a more unstable predictor. Further, Swanson (2008) found that when controlling for STM, a residual WM factor remained an important predictor for Gf in children aged 6 to 9 years (see Engel de Abreu et al., 2010, Giofrè et al., 2013 for similar results). To summarize, present research has hitherto not provided unequivocal results on the relationship between WM, STM, and intelligence across development, with the few published studies reaching different, contradictory conclusions.

One reason for these inconsistent findings may reside in the cognitive resources required for successfully solving simple span tasks. Importantly, Unsworth and Engle (2006) found that whereas complex span tasks of all list lengths correlated substantially with fluid intelligence, this was true for simple span tasks only when the number of items to be retained exceeded four. However, it is well known from the literature that humans can keep about four distinct entities in mind at the same time (Cowan, 2001). This leads to the differentiation of primary memory (PM) from secondary memory (SM). The purpose of PM is to maintain a distinct number of separate representations active for ongoing processing by continuously allocating attention (Unsworth & Engle, 2007). In contrast, items that have been displaced from PM must be retrieved from SM, which requires a cue-dependent effortful search process that is vulnerable to interference. SM can be measured, for example, by having subjects learn lists of supraspan length, whereas measures of PM assess the scope of immediate memory without activating storage or retrieval processes. In short, PM refers to a maintenance component of memory, whereas SM refers to effortful search and retrieval processes (Unsworth & Engle, 2007).

Both simple and complex span tasks measure PM and SM, but complex span tasks usually rely more strongly on SM than classical simple span tasks with short list lengths (Unsworth & Engle, 2007). It can be assumed that in order to tap SM in complex and simple span tasks to a comparable degree, simple span tasks should be of supraspan length, i.e., they should require a substantial amount of information to be stored, such that effortful retrieval becomes necessary. Based on these ideas, Mogle et al. (2008) investigated to which degree PM, SM, WM, and processing speed predicted fluid intelligence. Mogle et al. (2008) found that WM played no significant role in predicting intelligence when tasks measuring SM were in the model. These authors provide evidence that SM and to a lesser degree, PM are sufficient predictors for fluid intelligence. These results lead to a theoretically parsimonious and powerful explanation for the varying correlations of STM with fluid intelligence across the literature: Simple span tasks are especially good predictors of higher level cognition when performance is measured using supraspan lists and the role of rehearsal is reduced, that is, simple tasks are good predictors to the degree that they capture SM. In line with these results, Maybery and Do (2003) found that verbal and spatial supraspan STM tasks substantially correlated with mathematical ability in a sample of children (r = .50 and .53, respectively). Hence, it can be predicted that simple span tasks of supraspan list length (capturing predominantly SM) should predict higher level cognition to a similar degree as complex span tasks. These findings, however, have been recently challenged in two studies with adults (Shelton, Elliot, Matthews, Hill, & Gouvier, 2010) and children (De Alwis, Hale, & Myerson, 2014), which both reported that it is WM, rather than SM, that uniquely predicts Gf.

Other WM theorists emphasize that cognitive tasks that neither comprise a processing component nor allow rehearsal capture important aspects of WM. In his influential model of WM, Cowan (2005) underscores the key role of the focus of attention for cognition. The focus of attention corresponds to the currently activated content held in working memory, and has a capacity limit of approximately four items. According to Cowan (2005), it consists of long-term memory elements that are in a highly activated state. The scope of attention refers to the amount of information that can be held in the focus of attention for immediate retrieval. Individuals differ in their scope of attention, which is subject to developmental constraints and therefore smaller in younger children than in older children or adults (Cowan et al., 2006, Cowan et al., 2005). Because in contrast to long-term memory, the focus of attention is capacity limited, it is crucial to obtain a measure of this capacity, which forms a gateway of cognitive processing. According to Cowan et al. (2005), WM tasks that neither require a processing subtask nor allow for rehearsal can be good measures of the scope of attention, and hence, of an important component of WM. These authors used a visual array comparison task based on Luck and Vogel (1997). In this task, subjects have to compare two successively-shown arrays of squares of different colors, and to indicate whether a target square in the second array has changed its color or not. Several papers have demonstrated the utility of this paradigm for the determination of individual capacity limits in the scope of attention in both children and adults (Cowan et al., 2006, Wheeler and Treisman, 2002). Further, the work by Cowan et al. (2005) lends support to the view that this task, along with other tasks capturing the focus of attention, is substantially correlated with intelligence.

How can tasks measuring the scope of attention be reconciled with the framework of PM and SM advanced by Unsworth and Engle (2007)? Because the scope of attention refers to a limited amount of highly-activated memory traces, one might suppose that the visual array comparison task described above measures primarily PM. However, according to Cowan et al. (2005), tasks measuring the scope of attention load on a single factor together with complex span tasks. Because rehearsal is difficult in visual array comparison tasks, they further conceptually differ from classical simple span tasks. Hence, it is of theoretical interest to locate the scope of attention in latent variable models including both complex span tasks as well as simple span tasks of supraspan length. The scope of attention should be statistically separable from complex and simple supraspan tasks in case it measures primarily PM (e.g., Shipstead, Redick, Hicks, & Engle, 2012). The studies by Mogle et al. (2008) and Cowan et al. (2005) provide some first results concerning this issue. However, Mogle et al. (2008) investigated a selected adult sample, whereas Cowan et al. (2005, Experiment 2) did not include any supraspan STM measure in their analysis.

Another important open research question pertains to the domain-specificity or generality of WM. Whereas the domain-specific WM model assumes different WM factors across content domains (e.g., verbal/numeric vs. spatial), the domain-general model postulates a single WM factor that spans across content domains. Several researchers have found support for a domain-general WM model (e.g., Colom et al., 2003, Kane et al., 2004). However, some results support the notion of separate factors of WM that are highly correlated. For example, Jarvis and Gathercole (2003), testing the WM model suggested by Baddeley (1986) in a sample of 11 and 14-year-old children, found evidence for a verbal WM factor and nonverbal WM factor that were highly correlated. In the study by Süß, Oberauer, Wittmann, Wilhelm, and Schulze (2002), a visuospatial and a verbal–numerical WM factor correlated at r = .80. In this context, Engle et al. (1999) suggested a hierarchical model of WM that assumes both a domain-general factor as well as domain-specific residual factors (cf. Hornung et al., 2011). It should be noted that numerous studies supporting the domain-specificity of WM used relatively homogeneous samples, which can result in overfactorization (Shah & Miyake, 1996). In contrast, samples representing a broader range of the population (e.g., Kane et al., 2004) rather found support for a more domain-general model of WM.

Current hierarchical models of intelligence assume different ability factors. Whereas Gf is commonly defined as the ability to reason under novel conditions, another important facet of intelligence, crystallized intelligence (Gc), is related to academic achievement or cultural knowledge based on already learned knowledge (cf. Haavisto & Lehto, 2004). It can be hypothesized that Gc depends on storage or maintenance of known, activated information to a much larger degree than on executive functioning as compared to Gf (cf. Swanson, 2008). Further, the relationship between Gf and Gc can be assumed to decline with age because children's knowledge base tends to be standardized by school curricula (Schweizer & Koch, 2002).

The present study pursued three main goals. Firstly, I wanted to shed light on the structure of WM, STM, and the scope of attention, respectively, as well as intelligence in a sample of children by means of latent variable models. Although several studies have provided detailed results in adult samples, studies with children are still sparse and contradictory. Secondly, it was my goal to clarify whether WM, STM, and the scope of attention were comparable in predicting facets of intelligence. While numerous studies show that Gf strongly depends on WM, it is conceivable that Gc depends more strongly on STM. And thirdly, I investigated whether the prediction of intelligence by WM, STM, and scope of attention was moderated by age.