The multi-component model proposed by Baddeley & Hitch

Among the most important of theories regarding WM, and one that stands out for its particular influence, is the multi-component theory proposed by ^{Baddeley and Hitch (1974}; ^{Baddeley, 1986}, 2000). This will serve as a main reference for the current work. This theory presents a WM system that includes four main components: two slave subsystems for temporal storage (the phonological loop and the visual-spatial sketchpad), one system of attentional control (the central executive), and one temporal storage unit in charge of coordinating present information active in WM with that of LTM (the Episodic Buffer) (see Figure 1.).

The phonological loop and the visual-spatial sketchpad are subsystems specialized in the transitory processing and maintaining, respectively, of verbal and visual-spatial information. Both of these subsystems are composed of passive storage and a processing unit in order to refresh the information.

Figure 1.  The multi-component model of working memory (^{Baddeley, 2000}). 

On the other hand, the central executive (CE) coordinates and manages other WM components responsible for controlling attentional resources and monitoring the processing of information. It is thus ultimately responsible for the management and regulation of all such activity within the cognitive system. Initially, the theory conceived of CE as that responsible for processing and storage itself. However, later, these functions became more associated with belonging to different components. In this way, the CE became considered as that in charge of resource processing and control, while storage became considered as an additional process dependent on another component of the system, one known as the Episodic Buffer (EB).

The EB was added by ^{Baddeley in 2000} and is characterized as a temporary storage component that uses a multi-dimensional code, is accessible to consciousness, and has limited capacity (Baddeley, Hitch & ^{Allen, 2009}; ^{Allen, Baddeley & Hitch, 2014}). It serves as the “mental space” in which information from the slave subsystems (the phonological loop and visual-temporal sketchpad) is integrated with linguistic and semantic information stored within LTM relevant to whatever task is at hand. It is this function that permits the combination of all information and its integration into multi-dimensional clusters (i.e., “chunks”), which Baddeley has called “episodes.” To do so requires an episodic storage space that, according to recent studies (Baddeley et al., 2009; Baddeley, Allen, & Hitch, 2011), seems to function more passively than that initially proposed (Baddeley, 2000).

Specifically, what Baddeley has recently proposed is that under certain conditions (low level of complexity, common and/or related materials) the processing required to integrate information is basically automated and does not require additional support from the CE. In this way, the maintenance of information deposited as “chunks” in the EB is in no way affected by processing demands (Baddeley et al., 2009). Nor should any interference occur between the cognitive load required to processes information and that it required for its storage in the EB. In this work, Baddeley et al. present four experiments in which, from different conditions and concurrent tasks, they compare the execution of the subjects in the memorization of three types of word sets. From the same set of words, they are constructed: 1) “constrained lists”, are lists of words without syntactic structure, 2) “constrained sentences”, phrases with a certain syntactic meaning (see Table 1). In addition, in some experiment 3) phrases from news are also used, which therefore have a correct syntactic structure and do not use the same words as the other two sets “open sentences”. The objective of using the same word pool is that pro-active interference will minimize the contribution of episodic and semantic long-term memory. Thus, participants are forced to rely on the existing temporary grouping in working memory to support his memory. That is, they must be based on the syntactic structure of the phrases that are constructed with those words. In one case the structure exists and in other cases it is almost non-existent. In this way, the semantic components are limited or equalized to avoid their facilitating effect on the integration of information (binding). As concurrent task, different tasks, such as counting backward tasks, N-back with different conditions, test articulatory suppression, and visuospatial tasks are used in the different experiments.

In general terms, the work shows how despite the use of different concurrent tasks to load the EC, the results are quite stable showing that the best execution occurs when the words are presented in a more recognizable syntactic structure, without finding differences between an oral or visual presentation of the items. Likewise, the highest proportion of semantic errors occur to a greater extent when the syntactic structure is adequate indicating that the reading is carried out in a more superficial way guided by the syntax. The other relevant issue is that the execution in the concurrent task is not affected, so the lack of involvement of the CE in the primary task is confirmed.

Table 1:  Examples of sequences used in ^{Baddeley y cols. (2009}). 

This newly formulated model (^{Baddeley et al., 2011}) now places the Episodic Buffer in the core of WM (see Figure 2), involving the CE in the functions of integration and maintenance of information within the EB, but only when the processes that are required for the integration of information (i.e., the processes of “binding”) have not been automated. When these processes are automated, they no longer require the CE's participation (^{Allen & Baddeley, 2009}; Baddeley et al., 2009; ^{Allen, Baddeley & Hitch, 2006}).

Figure 2.  A revised model of working memory in fluid systems (^{Baddeley, Allen & Hitch, 2011}). The new proposal only includes those systems of a fluid character, but the authors have not modified the way in which these systems are related to the crystalized systems in LTM. 

It is in this context that our main interest lies: to examine whether, as ^{Baddeley et al. (2009}, 2011) have laid out, the Episodic Buffer is partially independent of the central executive and, if so, it can operate in certain conditions without requiring its assistance. This brings us to the debate on whether or not the two-basic function of WM (processing and storage) are independent or not. Baddeley's new proposal contrasts somewhat with other theoretical models that argue that both functions, processing and storage, are interrelated. These models assume that in performing any single task, the CE manages a single pool of resources, and that the management of available resources implies a trade-off of between the two functions (^{Barrouillet, Bernardin, Portrat, Vergauwe, & Camos, 2007}; see also, e.g., ^{Daneman & Carpenter, 1980}; ^{Case, 1985}; ^{Turner & Engle, 1989}; ^{Just & Carpenter, 1992}; ^{Anderson, Reder, & Lebière, 1996}).

In particular, our work aims to continue in the line of the studies of ^{Baddeley et al. (2009}) (see Table 1) discussed above, where the storage capacity of the Episodic Buffer was measured on memory tasks for distinct types of word lists that would limit its semantic properties and thus prevent the facilitation effect of the semantic integration of information (i.e., prevent “semantic binding”). These studies have shown how increasing processing load in constrained sentences can presumably occur without any interference in storage, since word maintenance is significantly higher compared to its incidence on the restricted list.

Specifically, our work aims to contrast in a different way the results presented by ^{Baddeley et al. (2009}) where they take as measure of the EB the storage capacity of the subjects in recall tasks on different types of word lists while performing a concurrent task that requires CE resources. As we described above, the idea is to try to show how the processing load in the secondary (concurrent) task can be increased without significantly interfering with the storage and maintenance of the information in the primary task.

In this paper, we explore a new recall task in which participants must comprehensively read a few simple narratives having phrases designed such that understanding requires few attentional resources. In this way, the CE is minimally involved, just as there will exist minimal interference in the creation of “chunks” in the Episodic Buffer (EB). In this way, the CE is minimally involved, because the demand for attention resources by the EB will be minimal, since the texts require the construction of very simple mental models. Simple texts facilitate the organization of information (binding) and probably also the necessary semantic grouping for storage (chunking). In this way, our proposal has similar goals as that of the studies of ^{Baddeley et al. (2009}), but in the opposite direction and with certain differences. Instead of minimizing the effect of the participation of the CE by avoiding the semantic features of sentences and relying and prioritizing the syntactic ones, we propose a task that utilizes a very simple semantic context in order to minimize the participation of the CE. Thus, we intend to measure the passive capacity of the Episodic Buffer through the use of easy-to-understand sentences coordinated in a context of global understanding. Moreover, unlike demands of the classic task, in this way CE resources can be to a greater extent dedicated to the “re-activation” of words previously used in the Analogy span task. This can also be understood in terms of the priming effect, as it “pre-activates” the words in the analogies task that later are “re-activated” in simplified form within the context of the reading. We think, that in any case, the re-activation could be carried out with the support of the situational representation generated during comprehension, as often occurs in everyday reading tasks. As detailed below, this provides a more “ecological” approach of WM functioning in general, and of the EB itself in particular, given all resources available to the participant will be measured in a more natural way.

This contextualization allows that the “binding” and “chunking” processes to be carried out on informative elements with semantic coherence in the context of a text. We think it is important that the measurements be carried out based on cognitive tasks similar to those that participants perform in daily life, in particular the measurement of EB due to its characteristics of integration of the MLP information. Therefore, unlike ^{Baddeley et al. (2009}), we will not use word lists to be “organized and/or grouped”, but semantic contexts (texts) that allow the creation of more significant mental models for participants and facilitate the integration of information (binding processes) and the creation of groups (chunking processes).

Measuring the Episodic Buffer

Our proposal aims to analyze the performance of the EB and its dependence or not on general attentional processes. To do this, we should try to specifically measure the EB, that is, measure the capacity of the EB to store the chunks that are being developed and maintained.

But in addition, we think that this measure should be done in a more natural cognitive context than that involved in the traditional dual-task paradigm (or “concurrent processing”( directed at compromising CE resources. To do this, we use a habitual task of cognitive processing, such as reading comprehension, which requires coordinating the active information in WM with the related information that is stored in the LTM in order to integrate all information and create “semantic fragments” that must maintained simultaneously; that is, a task that necessarily implies a “natural” use of EB. Likewise, we cannot forget the close and well-known relationship between MO measures and reading comprehension (see, for example, ^{Daneman & Merikle, 1996}; ^{Hannon & Daneman, 2001}, ²⁰⁰⁴; ^{Cain, Oakhill, & Bryant, 2004}; ^{Vukovic & Siegel, 2006}; ^{Carretti, Borella, Cornoldi y De Beni, 2009}; ^{García-Madruga et al., 2013}; ^{García-Madruga, Gómez-Veiga, & Vila, 2016}).

In the work we present, a verbal test is used in which the participants must read several texts, taking into account that according to ^{Baddeley (2000}, 2007, ²⁰¹²), the EB must be the mental space in which they must be carried out the processes involved in reading comprehension.

Reading comprehension requires the participation of the processes that carry out the integration of information (binding processes), and the processes necessary to build semantic clusters (chunks) that allow the understanding of the text. It is likely that these “representational chunks” facilitate the recall of the concrete items (words) that have been activated to build them.

Also, to ensure the participation of EB, the memory that subjects must do in this task happens after at least 20 seconds of the last activated keyword. The texts contain an introductory part of the reading to force the delay time thus facilitating the incorporation of more long-term resources (^{Alloway & Ledwon, 2014}; ^{Conway & Engle, 1994}; ^{Cowan, Wood, Nugent & Treisman, 1997}; ^{Cowan 2001}). After the introduction phrases, the texts present a series of incomplete sentences that must be filled in with the previously activated keywords.

The test we propose has been designed based on the Analogies Reasoning Span Test (^{Gutiérrez-Martínez, García-Madruga, Carriedo, Vila & Luzón, 2005}; García-Madruga, Gutiérrez-Martínez, Carriedo, Luzón & Vila, 2007). This is a classical dual-task measure in which participants read aloud a series of simple verbal analogies, find and read out loud the solution, and remember the word-solution of each analogy at the end. In the new test, one second measurement is added, so that the procedure includes two phases which will permit us to take two successive measures of WM: one index based on the cited analogies test show that basically measures CE participation, and another that presumably measures the Episodic Buffer capacity.

To summarize, in addition to the Analogies span task, the new procedure requires participants to perform a second recall task that consists of comprehensively reading short texts that contain some missing words. To reach the correct understanding of the texts, participants must then complete the gaps in the texts with words inferred and stored previously in the analogies task.

Thus, the new procedure consists of two consecutive stages: stage one made up of the analogies span test, which we will call ANALOGY; and a second stage with the new task of short texts, which we will call CONTEX:

1) ANALOGY. Test participants begin by solving a growing series of analogies (2, 3, 4 and 5) the result of which should be stored in memory and recalled at the end of each series in the correct order of appearance (they must recall all stored words in each series; 2, 3, 4 or 5).

2) CONTEX. Then participants are presented with text containing a brief introduction and a number of phrases corresponding to the number of analogies each series contains. Each sentence ends with a blank space that must be filled in with the keywords previously stored during the analogy task, thus offering a second chance to remember them. In this way, the memory for keywords is recorded at two times: one immediately after solving the analogies, and a second delayed memory that must be integrated into the texts (see Figure 3).

Figure 3.  Basic procedural structure of the new test (a level 2 example). 

As we have described above, we assume that the first criterion that must fulfill the new task is that it registers the processes specific to the Episodic Buffer with minimum participation by the CE. Accordingly, we developed this task by following an approach that somehow reverses that of ^{Baddeley et al. (2009}). That is, instead of avoiding semantic processes, we used a task that necessarily requires this kind of process in order to achieve comprehension. The key to this is the simplicity of texts used in this new measure and that results from previous studies have not encountered significant differences in neither reading speed nor level of comprehension in adolescent participants (^{Gutiérrez-Martínez et al., 2005}; ^{García-Madruga et al., 2007}). Therefore, it seems normal readers may perform the semantic integration in semi-automatic ways, constructing the representational model of the text in a simple manner, both within and between sentences. Thus, our task makes the binding processes necessary to complete it even more automatic, by using syntactically simple and similar sentences that have less demanding content. However, it is important to note, as pilot studies have shown, that the semantic content of sentences would not allow participants to guess the keywords without prior pre-activation into ANALOGY task, given that those texts were designed with blanks that might be completed with different words to the keywords (keywords were remembered with a success rate of less than 5%).

In sum, for an undemanding task for expert readers (i.e., university-level), we believe that its correct completion would require minimal activation of the CE's attentional resources. This would permit us to attribute any differences found in the new task to the independent or specific capabilities of the Episodic Buffer. Thus, Baddeley's new proposal which affirms the relative independence of EB and its eminently passive character could be tested.

CONTEX, an ecological proposal to measure Episodic Buffer Span

Additionally, our idea also revolves around the possibility of making a more natural or ecological, or, let us say, contextualized index of the Episodic Buffer. We understand that this new contextualization task occurs via two different pathways. First, the words that must be retrieved have been previously activated and stored in the task of analogies, which can be a facilitating element. Also, in the second part of the task, the representational model that must be constructed during the comprehension of the text, requires the activation of previous knowledge that supposes a favorable semantic context to "re-activate" the previously stored keywords. We believe that the activations and chunking caused by comprehensive reading can favor the recovery of stored words.

We have tried to reproduce the context for the task in which WM resources can be utilized naturally, just as when used in daily tasks in which processing and storage functions operate together. In this way, we seek to distance ourselves from procedures that employ concurrent tasks that are not naturally related, and thus serve different purposes. For example, in the classic Reading Span Test (the RST of Danneman & Carpenter, 1980). The subject must read a series of sentences and store the last word of each sentence and retrieve them at the end in their correct order. Both tasks are independent, and the final goal is to remember a sequence of specific words (without an ulterior motive), an uncommon occurrence in daily activities. Indeed, the researcher might choose a different word to remember instead of the last, and the result would remain the same, given that the tasks are not connected. The task can even be solved correctly without completing understanding the sentences; a superficial reading is enough. Nevertheless, comprehensive reading done in a natural context involves storing those important elements that favor comprehension, not the items located in a particular place.

In our proposal, the new test has been designed in order that the information stored in the first task is itself the ultimate goal of the second task: full understanding of the new text. With this, we procure that the characteristic functions of WM, processing and storage, occur naturally in the context of a routine cognitive task. In this case, memory per se is not the final goal of the task, yet it serves a higher purpose, of natural and habitual character in line with completely understanding a text.

In order to analyze whether the processes involved in CONTEX related to the episodic buffer depend or not on the CE resources, the reading time of the sentence was measured to confirm whether it remains stable despite the increase in processing as the task progresses. If in CONTEX the reading time remains stable at all levels, we would have one more indicator of the low intervention of the EC. Other studies have confirmed that in double-task tests that use reading tasks, when storage demand increases, subjects use more time during processing as they must share their resources.

The design of a task that we think measures EB, requires contrasts with other WM tasks and with measures of high-level cognitive processes such as intelligence, reasoning and reading comprehension.

Approach and objectives of the study

The current study has as main objective to analyze if the EB can make the necessary integrations to store and maintain the resulting “fragments” without significant participation of the EC. To do this, we will measure whether the activations required to create the “representational chunks” facilitate the recovery of words that make up those chunks. A daily task will be used to take advantage of the facilitating effect of the context and in which there is a low participation of the CE in the process of organizing information by using very simple texts that require a semi-automated understanding. For this purpose, it is necessary to check whether the context of “reading comprehension” with simple texts does not really increase the processing load, and hence provokes at least a minimum of participation by the CE. If so, and consistent with Baddeley, we can say that the basic functions of WM (processing and storage) are practically independent in these types of tasks in which the processing component is highly automated. It is in the context that we present a new task (CONTEX) that measures WM capacity from those processes specifically related to the Episodic Buffer, and that also meets another one of the objectives outlined above: the assessment of WM in a more natural and ecological way.

The results in this task will be contrasted with another two measures of WM used in various studies, and which mainly report on the participation of processes more closely related with the CE (the Reading Span Test-RST, and the Reasoning Span Test based on Analogies-ANALOGY). Thus, we think that the new task measures the most specific processes of the EB, while the other measures of MO measure processes of the CE of a more general nature.

In addition, this paper also includes a reading comprehension test in order to analyze the specific task of reading texts containing the new measure. Finally, to provide even stronger validity of the WM measures, we have included two higher-level cognitive variables, intelligence and reasoning, given that numerous studies have illustrated the close relationship WM has with these variables (see, for example, ^{Cornoldi, 2006}; ^{Oberauer, Schulz, Wilhelm & SüB, 2005} on fluid intelligence; and Kane & ^{Engle, 2002}, ^{García-Madruga et al., 2007} on reasoning). Although it is not the main interest of the work, we think it is important to confirm that the tasks of WM maintain the usual relationships with the measurements of the high-level cognitive processes.

Thus, our hypotheses can be divided into two groups. The first two related to the primary objective of the work, analyze the relationship between the Episode Buffer and the Central Executive. And the last two hypotheses will focus on the validity of the tests used, analyzing whether the usual reactions found in the literature between intelligence, reasoning and reading comprehension are maintained, taking special interest in the behavior of the new task. Thus, 4 main hypotheses are specified:

Participants

The sample consisted of 60 college freshman psychology students with an average age of 24.1 years (SD = 8.5 months). The participants were rewarded for their participation with a bonus towards their qualifying practicum in developmental psychology. A total of 6 participants were removed from the initial sample as for various reasons they did not complete all of the tasks.

Measures

Working memory

- Reading Span Test (RST)

The RST test, designed by ^{Daneman and Carpenter (1980}), applied in its Spanish version (^{Elosúa et al., 1996}), consists in reading a series of 12-word long sentences. The subject must remember the last word of each sentence composing the series and in the right order of appearance. The task includes various levels that increase progressively. The initial series requires remembering 2 words from two sentences in correct order, up to a maximum of 5 words from 5 sentences (you can see an example of level 2 in Table 2). Each level is composed of 3 series. Points are awarded for correct answers when recalling all of the words in correct order, and for incomplete answers when remembering the words but in a different order. Omitting or substituting any one key word is considered an error. Minimum score is 2 and the maximum is 5.8. (Further details about this test, as well as the criteria for awarding points, can be found in ^{Gutierrez-Martínez et al., 2005}.) All of the WM tests were scored by following this same procedure.

Table 2:  Example of a level 2 RST. 

- Test of Contextualized Span (PA-CONTEX)

The newly designed procedure we have named PA-CONTEX involves solving and remembering the word-solutions of a few simple analogies (see Figure 3). After the first recalling of these key words as described above, the participant must then read a few texts containing blanks the participant is to fill out with the words previously stored. To facilitate identifying the results at each stage, we have named the points obtained in the first memory task ANALOGY, and those obtained in the second task CONTEX.

To prepare for this set of tasks, a few pilot studies were run in order to confirm that the texts themselves did not lend “guessing” or intuiting the key words. Furthermore, in order to create very simple texts, low difficulty words across the frequency spectrum were always used. Similarly, the average time participants took to read each text was homogenized (around 15 seconds), as well as the approximate time each participant took to fill out each word blank (5 seconds). These 20-second delays with respect to the first memory (in the first phase) may involve more long-term resources than otherwise the phonological loop or the visual-spatial sketchpad, without refreshing, can handle (see ^{Alloway & Ledwon, 2014}; ^{Conway & Engle, 1994}; ^{Cowan et al., 1997}; ^{Cowan 2001}). Thus, as Baddeley has proposed, the maintenance of this information must occur in the Episodic Buffer with the help of LTM. The criterion followed to parameterize the extent of the sentences was to equalize the length of sentences found in texts with those used in the Spanish version of the RST (an average of 12 words). The full content of the task is detailed in the Appendix.

Reasoning

The test of reasoning used was that taken from the studies of ^{Gutiérrez-Martínez et al., (2005}), by selecting conditional deductive reasoning problems of an either factual (if A then B) or counterfactual (if A had been, then would have been B). Participants were presented with the four classical conditional inference types (Modus Ponens, Affirmation of the Consequent, Denial of the Antecedent, and Modus Tollens) The test consists of two examples of each kind of statement, resulting in a total of 16 statements randomly ordered for each participant: 2 statements for each of 4 inference types (8) and two examples of each (8 x 2 = 16). Final results were obtained from directly scoring the results of each correct answer.

Table 3 shows an example of the content of each of the statements.

Table 3:  Example of statements in the Reasoning task. 

Procedure

A within-subject design was used in which each participant completed all of the tasks. In the first session paper-and-pencil based tests of intelligence, reasoning and reading comprehension, in this order were conducted in groups. The second session was completed individually, and participants completed the tests of WM (RST and PA-CONTEX: ANALOGY + CONTEX). The WM texts were administered on a computer using the E-Prime software (^{Schneider, Eschman, & Zuccolotto, 2002}). The order of administering the two WM tests was alternated.

Data analysis

The reliability of the tests has been estimated in terms of “internal consistency” by means of Cronbach's alpha coefficient.

The validity of the construct of the new task will be analyzed based on the differences in time to read within each level and between levels, and the total scores obtained in WM tests. In order to test for significance in performance differences, we applied the parametric Student t-test for related samples.

The expected linear relationship between the tests was estimated by calculating the Pearson correlation coefficient between the total scores. This correlation analysis was also performed with respect to the criterion measures (General Intelligence, Reasoning and Reading Comprehension) in order to verify the predictive capacity of the tests studied.

In order to contrast the relationship between WM and the criteria variables, we carried out a confirmatory factorial analysis. Previously, the degree of association between variables was analyzed by “Bartlett's test of sphericity” and a measure of sampling adequacy by the “Kaiser-Meyer-Olkin” (KMO).

Finally, to assess the predictive power of WM measures in relation to our criteria variables, three stepwise multiple regression analyses were performed.

Hypothesis 1 and 2

Reading times of sentences making up WM tasks were recorded in order to test the possible independence between processing (secondary task) and storage (primary task). This was done to confirm whether reading time is affected by increasing the load on storage. Table 4 illustrates the mean reading times for levels 2 and 4 (only a small number of participants reached level 5, thus it was not considered). By using the difference between these levels as criteria, RST and ANALOGY times show a significant increase (RST: t = -12.26; ANAOLOGY: t = -6.67; df = 53 and p < .00 in both cases). This data confirms that there is a greater load when the task level is increased (from 2 to 4). However, as expected, the new task (CONTEX) shows no significant reading times differences between levels (t = -1.90, df = 53, p > 0.5), and thus we can say that processing load is similar across both task levels.

Table 4:  Mean differences in sentence reading (Student'). 

On the other hand, while the sentences used in RST and CONTEX have the same length, the time participants spent reading on the RST (M = 4774.19ms., SD = 668.21ms) is significantly less than spent in CONTEX (M = 5344.91ms., SD = 940.38 ms), which suggests that reading was done more superficially (t = -5.38, df = 53, p < .01). Other studies have also found that RST allows a superficial reading of sentences freeing up resources in processing to facilitate storage. (^{Gutiérrez-Martínez et al., 2005}; ^{García-Madruga et al., 2007}).

We should note that we identified outlier data more than three standard deviations above the mean, as per convention. Seven outliers were identified and each was replaced by the upper limit (the mean plus three standard deviations).

In order to test whether the WM measures indicated the utilization of different system processes, we analyzed the descriptive statistics, the mean (M) and standard deviation (SD) of which can be seen for all measures in Table 5. The data shows that scores were highest for the RST and lowest for ANALOGY, making the latter the most complex task (hypothesis 2a). These results agree with those found in other studies (^{García-Madruga et al., 2007}; ^{Gutiérrez-Martínez et al., 2005}). The new measure, CONTEX, occurs in intermediate position and has a score close to that of RST. Reliability measures of CONTEX also reflect a high degree of internal validity for this measure (Cronbach's alpha = .78).

Table 5:  Descriptive statistics of measures.  

As expected, significant differences between the scores on all measures was found, as reflected in the comparison of means using Student's t-test: between RST and ANALOGY (t = 9.98, df = 53; p < .01), between RST and CONTEX (t = -2.51, df = 53; p = .01), and between two measures of our new procedure, ANALOGY and CONTEX (t = 7.44, df = 53; p < .01). However, we think that the most relevant result to confirm our hypotheses is that the participants recalled in CONTEX words they had not been able to recall in ANALOGY (hypothesis 2b). These results reflect a certain independence between the two measures and confirm that the results of the new procedures designed here in both phase 1 and 2 are indeed different in some way. In fact, CONTEX gets scores closer to RST than to ANALOGY, see Table 5 (hypothesis 2c).

Hypothesis 3 and 4

Inter-correlations between WM measures are highly significant, indicating that despite differences in their scoring, they are founded upon an important common base (see Table 6). We emphasize the level of significant obtained by CONTEX (< .001), especially, with respect to other WM measures.

Correlations between criteria variables reflect the relationships that may exist between them, as we did expect. In particular the test of intelligence correlated very significantly with that of reasoning and textual comprehension, the highest of which was between reasoning and intelligence and, on the other hand, the lowest of which was between reasoning and comprehension.

Also, as expected, the correlations of WM measures with the higher-level cognitive variables were also positive and significant in every case (see Table 6). Among them, data again confirm CONTEX as having the highest correlation across all cases. The intelligence test (RAVEN) is more highly correlated with CONTEX than it is with any other measure, and the only one that is highly significant. Our reasoning measure (REASONING) is more highly correlated with CONTEXT, even if very similar to ANALOGY, and is least correlated with RST (in line with other studies, see ^{García-Madruga et al., 2007} and ^{Gutiérrez-Martínez et al., 2005}). As the for the reading comprehension test (PROLEC), the data also support our predictions: RST obtains the lowest correlations and CONTEX obtains the highly significant correlational indices.

Table 6:  Pearson's Correlations between measures.  

In order to deepen the understanding of the relationship between WM and the criteria variables, we also carried out a confirmatory factorial analysis using principal components to evaluate the fit of a two-factor model. However, the sample size means that these results must be taken as merely exploratory. We think that in this model the WM tasks should cohere together, on the one hand; and other the other, our criteria variables, especially intelligence and reasoning, as these are tasks that assess more general characteristics of a higher order. Studies confirm that these are related but independent variables.

Prior to carrying out the factorial analysis, the degree of association between variables was analyzed by Bartlett's test of sphericity, indicating that the Chi Squared index forms the best fit to the data (X ² = 77.307, df = 15, p = .00). Likewise, a measure of sampling adequacy, the KMO (Kaiser, Meyer and Olkin) reflects a score of .778, over and above the recommended score for factorial analysis, which is .75.

The two-factor generated by the model explain more than 63% of the variance (see Table 7). Using Varimax rotation shows Factor 1 to be basically made up of WM measures and explains 34% of the variance, with CONTEX being the greatest contributor to this factor. Factor 2 explains 29% of the variance, with RAVEN being the measure that makes up the largest contribution to this component. It is worth noting that CONTEX still participates in Factor 2 with a weight of .34, while the other WM measures are below .15, again indicating that CONTEX has more predictive power relative to the others.

Table 7:  Factorial Analysis with extraction through an analysis of Principal Components. Method of rotation: Kaiser's Varimax rotation. 

Finally, to assess the predictive power of WM measures in relation to our criteria variables, three step-wise multiple regression analyses were performed. The main variables were WM measures, and intelligence, reasoning, and textual comprehension were dependent variables. Previously, we calculated an index of multi-collinearity (VIF), which resulted in less than two for all cases (and below the index of 10 cited by Kleinbaum), and illustrated a condition index below 15 (and below the limit of 20 calculated by Belsley to indicate a range of weak co-linearity).

Regression analysis generated a model that explained 16% of the variance (F(1, 52) = 9.579, p < .01; Error = 4.43) for the intelligence measure, with CONTEX the only independent variable introduced (B = 2.36; β = .394, p < .01). For reasoning, the model generated explained 13% of the variance (F(1, 52) = 7.80, p < .01; Error = 206), again with CONTEX the only participatory independent variable (B = .099; β = .361; p < .01). The same goes for textual comprehension, explaining 13% of the variance (F (1, 52) = 7.478; p < .01; Error = 1.469), with CONTEX the only variable (B = .691; β = .355, p < .01).