The domain specificity of working memory is a matter of ability

Highlights

• Ability differentiation is a link of higher ability and stronger impact of domain specific skills.

• Here we provide evidence for ability differentiation in working memory capacity.

• The results inform a fundamental debate about domain generality vs. specificity in cognition.

• The results are in accordance with predictions of process overlap theory.

Abstract

The relative importance of domain-general and domain-specific sources of variance in working memory capacity (WMC) is a matter of debate. In intelligence research, the question of domain-generality is informed by differentiation: the phenomenon that the size of across-domain correlations is inversely related to ability: the lower the ability, the more domain-general the variance. Since WMC and intelligence are related constructs, differentiation might exist in WMC, too. Differentiation in WMC is also predicted by process overlap theory, a recent model of intelligence. We used moderated factor analysis to test for differentiation. The results demonstrate the existence of differentiation in WMC: as capacity increases, variance in WMC becomes more domain-specific. Fluid reasoning (Gf) also contributes to differentiation in WMC: when Gf is lower, WMC variance is more domain-general. There was no significant moderation by crystallized (Gc) and spatial (Gv) ability and Gf only moderated differentiation in WMC but not in short-term memory.

Introduction

Working memory is a psychological construct used to characterize and help further investigate how humans maintain access to goal-relevant information in the face of concurrent processing and/or distraction (Baddeley, 1992). According to its first conception working memory is characterized as a multi-component system which includes domain-specific verbal and spatial “slave” storage systems, as well as a domain-general central executive responsible for attention control (Baddeley & Hitch, 1974).

Even though the model of working memory was initially developed to account for intra-individual phenomena, interest soon arose in measuring individual differences in the capacity of this system. One of the first measures of the capacity of working memory was the reading span task (Daneman & Carpenter, 1980), which requires subjects to read sentences aloud and remember the last word of each sentence for later recall. Another early example is the counting span task (Case, Kurland, & Goldberg, 1982) in which subjects are instructed to count a particular class of items and, after counting aloud, remember and later recall the totals. There are also spatial working memory tasks, such as letter rotation task (Shah & Miyake, 1996) and symmetry span (Kane et al., 2004).

Several of these “complex span tasks” have now been developed to measure working memory capacity (for a review, see Conway et al., 2005). These tasks are thought to be valid measures of working memory capacity because they require access to information in the face of concurrent processing. In contrast, simple memory span tasks (e.g., digit span, word span, letter span), which do not include an interleaved processing task between the presentations of to-be remembered items, are thought to be less ecologically valid measures of working memory capacity (Baddeley and Hitch, 1974, Daneman and Carpenter, 1980, Dempster, 1981).

Besides such progress in measurement, substantial theoretical developments have been made and alternative models have been created since the publication of the original Baddeley and Hitch model (e.g. Cowan, 1999, Oberauer et al., 2003). Virtually all current models of working memory include domain-specific and domain-general processes and in the working memory literature there is considerable debate about their relative importance. In particular, the domain-generality of variation in working memory capacity remains a controversial issue.

One of the most important findings from studies investigating complex and simple span tasks is that variation in complex span is more domain-general than in simple span; across domain correlations are larger in complex than in simple span tasks (Turner & Engle, 1989). This implies that working memory capacity is determined to a larger extent by domain-general processes, relative to domain-specific processes, than short-term memory capacity. Yet the domain-generality of WMC is controversial: although there are larger cross-domain correlations in complex span, other evidence appears supportive of a domain-specific view of individual differences. For instance, Shah and Miyake (1996) found that verbal and spatial working memory predicts verbal and spatial ability better, respectively, arguing for a domain-specific view of individual differences.

Since working memory tasks require parallel storage and processing, observed correlations with other variables may reflect variation in either the storage or the processing components of working memory tasks, or both. Latent variable studies of individual differences in working memory capacity are useful because they are able to decompose storage components (variance common to short-term memory tasks and working memory tasks) from processing components (variance unique to working memory tasks). Kane and colleagues (Kane et al., 2004) applied exactly this method in a latent variable analysis; they decomposed the storage components of complex span tasks and found that while storage processes indeed appear to be more domain-specific, the processes that complex span tasks tap beyond the pure storage and retrieval of information appear to be largely domain-general.

Latent variable studies of working memory have provided additional important results. First, they identified a general factor of working memory, which is generally referred to as “working memory capacity” or WMC (Conway et al., 2002, Conway et al., 2003, Engle et al., 1999). This is the result of all-positive correlations between different working memory tasks. This finding is similar to one of the main findings in the study of intelligence, called the positive manifold: cognitive ability tests with diverse content, ranging from reading comprehension to number series to mental rotation, all correlate positively. This finding is the basis of the general factor of intelligence, g, which explains 40–50% of the variance in cognitive ability tests. The general factor of WMC is similar to the general factor of intelligence since it accounts for the positive correlations between working memory tasks with different content.

There is evidence that the general factor of WMC reflects individual differences in the executive component of working memory, particularly executive attention and cognitive control (Engle and Kane, 2004, Engle et al., 1999, Kane and Engle, 2002, Kane et al., 2001). Also, latent variable studies employing both intelligence tests and working memory tasks revealed that WMC is strongly related to intelligence. Two studies, conducted by different groups of researchers, estimate the median correlation between WMC and nonverbal fluid reasoning (Gf) to be somewhere between r = .72 (Kane, Hambrick, & Conway, 2005) and r = .85 (Oberauer, Schulze, Wilhelm, & Süss, 2005). Thus, according to these analyses, WMC accounts for between half and two-thirds of the variance in Gf. This is substantially higher than the proportion of variance in g, the general factor of intelligence, that is explained by WMC (Ackerman, Beier, & Boyle, 2005).

That is, WMC is more strongly related to the fluid factor of intelligence than to other factors. This is, once again, demonstrably caused by the processing, not the storage component of working memory tasks; when latent variable studies decompose what complex span tasks require beyond storage and retrieval they find that such processing components correlate to a much smaller extent with tests of crystallized intelligence (Gc) or processing speed (Gs) than with fluid reasoning (Gf) (Conway & Kovacs, 2013).

Finally, when one compares complex and simple span in terms of how well they predict fluid intellience (Gf), complex span tasks turn out to be stronger predictors (Conway et al., 2002, Engle et al., 1999, Kane et al., 2004, but see Colom et al., 2006, Unsworth and Engle, 2007). Taken together, these studies demonstrate that: (1) it is the processing component of working memory tasks, mostly reflecting executive processes, that drives the WMC-intelligence relationship, and (2) it is the fluid component of intelligence that correlates most strongly with WMC.

The factorial analysis of intelligence test results is also able to identify a general factor (g), as well as specific factors, and in the intelligence literature there has also been a long-standing debate about domain-generality vs. specificity, and in particular whether g can be identified as a general mental ability permeating all human cognition (Conway & Kovacs, 2013). This debate has been influenced by research on ability differentiation: the phenomenon that across-domain correlations are higher in low ability groups (Blum & Holling, in press; Juan-Espinosa et al., 2006, Kane et al., 2006). Importantly, differentiation is not simply the result of the restriction of range: in high ability groups the correlation between different tests is lower than in low ability groups with equally restricted range (Blum & Holling, 2017). Differentiation, then, means that uni-dimensionality of variance is more applicable in low ability groups than in high ability groups. Thus the question of domain-specificity in intelligence is not independent of the level of intelligence of the sample in question.

A recent theoretical account of human intelligence, process overlap theory (Kovacs and Conway, 2016a, Kovacs and Conway, 2016b), provides an explanation of the positive manifold in intelligence. The theory postulates an overlap of cognitive processes activated by various mental ability tests and working memory tasks. In particular, it is hypothesized that any item or task requires a number of domain-specific as well as domain-general cognitive processes. Domain-general processes responsible for executive attention and cognitive control are central to performance on mental tests as well as working memory tasks since they are activated by a large number of items, alongside with domain-specific processes tapped by specific types of items/tests only.

Process overlap theory draws heavily on the concept of working memory capacity in explaining the positive manifold in intelligence. In fact, it provides an explanation of both positive manifolds, the one in intelligence and the one in working memory. The positive correlations between diverse working memory tasks on the one hand and diverse ability tests on the other are both caused by domain-specific processes overlapping with a set of domain-general executive processes that are tapped by a large number of ability tests and working memory tasks. Since the general factors are statistical accounts of the positive manifolds, process overlap theory provides an explanation of the general factor of WMC as well as g. Moreover, since it proposes that the same pool of domain-general executive processes is tapped by different working memory tasks as different psychometric tests of cognitive ability (especially the ones that measure fluid reasoning), the theory also explains why the general factors of working memory and (fluid) intelligence correlate so strongly.

The theory actually focuses on limitations in its account of the positive manifold. That is, the central processes that are tapped by a large numbers of tasks limit performance in a general way and make errors more likely regardless of the domain-specific processes that are also tapped by the same tasks. This way executive processes function as a bottleneck and can potentially mask individual differences in more specific abilities. This is, according to the theory, the explanation of ability differentiation: it occurs because the lower the ability on central executive processes the lower the probability of correctly solving cognitive tasks, regardless of the level of ability on domain-specific processes.

Differentiation means that the lower the ability of a population, the higher the average correlations between tests; therefore differentiation can also be described as the general factor, g, accounting for more variance at lower levels of ability, whereas in high ability samples more variance is accounted for by domain-specific ability factors.

According to process overlap theory, the same “executive bottleneck effect” that is described above operates in working memory, too. Therefore, it a clear prediction of the theory that differentiation has to manifest itself in WMC as well. This is because the worse the performance of executive processes the more it is likely that executive processes will be the source of error, hence the larger section of the total variance they will account for, relative to specific processes. This prediction is practically agnostic with regard to most actual models of working memory as long as they propose both domain-specific and domain-general sources of variance.

The current study focuses on three specific predictions regarding differentiation in WMC that follow from process overlap theory:

• Ability differentiation occurs in tasks measuring WMC.

• Since executive functions are strongly related to fluid reasoning (Gf), to a much larger extent than to verbal and spatial ability, Gc and Gv, respectively (Conway and Kovacs, 2013, Conway et al., 2011, Unsworth and Engle, 2006), differentiation in WMC is moderated by Gf, but not, or to a much smaller extent by Gc or Gv.

• Since executive processes are tapped by complex span tasks to a much larger extent than by simple span tasks (Engle and Kane, 2004, Unsworth and Engle, 2007), differentiation occurs in working memory, but not to or to a much smaller extent in short term memory.

In the current study we investigated these three predictions. Specifically, in Study 1 we tested prediction 1 using the non-linear differentiation methodology by Tucker-Drob, 2009, Molenaar, Dolan, Verhelst, 2010. Next, in Study 2, we tested predictions 2 and 3 using the moderation methodology of Bauer and Hussong (2009).