The relationship between n-back performance and matrix reasoning — implications for training and transfer

Abstract

We have previously demonstrated that training on a dual n-back task results in improvements in fluid intelligence (Gf) as measured by matrix reasoning tasks. Here, we explored the underlying mechanisms of this transfer effect in two studies, and we evaluated the transfer potential of a single n-back task. In the first study, we demonstrated that dual and single n-back task performances are approximately equally correlated with performance on two different tasks measuring Gf, whereas the correlation with a task assessing working memory capacity was smaller. Based on these results, the second study was aimed on testing the hypothesis that training on a single n-back task yields the same improvement in Gf as training on a dual n-back task, but that there should be less transfer to working memory capacity. We trained two groups of students for four weeks with either a single or a dual n-back intervention. We investigated transfer effects on working memory capacity and Gf comparing the two training groups' performance to controls who received no training of any kind. Our results showed that both training groups improved more on Gf than controls, thereby replicating and extending our prior results.

Introduction

Fluid intelligence (Gf) is defined as a complex human ability that allows us to adapt our thinking to new cognitive problems or situations for which we cannot rely on previously acquired knowledge (e.g. Carpenter, Just, & Shell, 1990). Gf is considered critical for a wide variety of cognitive tasks (Engle et al., 1999, Gray and Thompson, 2004), and it seems to be one of the most important factors in learning (Deary et al., 2007, Neisser et al., 1996, Rohde and Thompson, 2007, te Nijenhuis et al., 2007). There is considerable agreement that a substantial proportion of the variance in Gf is hereditary (Baltes et al., 1999, Cattell, 1963, Gray and Thompson, 2004), but it has also been shown that social class and age moderate the heritability of Gf (Haworth et al., 2009, Turkheimer et al., 2003). Although high heritability in principle does not preclude alteration of Gf through environmental factors or interventions (Jensen, 1981), evidence of such alteration has been sparse. However, there is now accumulating evidence showing that certain interventions seem to increase performance in Gf tasks (Buschkuehl & Jaeggi, 2010), although the mechanisms that underlie such change are not well understood (Basak et al., 2008, Jaeggi et al., 2008, Klingberg et al., 2005, Klingberg et al., 2002, Rueda et al., 2005, Tranter and Koutstaal, 2007). For example, we have shown that the n-back task can be used as a training vehicle to improve performance on matrix reasoning tasks which are commonly used as a typical measure of Gf (e.g. Gray and Thompson, 2004, Kane and Engle, 2002, Snow et al., 1984). In our study, subjects were pretested on measures of Gf, after which they were given up to four weeks of daily training on a dual n-back task (Jaeggi, Buschkuehl, Jonides, & Perrig, 2008). The dual n-back task consisted of a position that was pseudo-randomly marked on a computer screen in each stimulus frame which subjects had to match for spatial position to the stimulus presented n frames back in the sequence. Simultaneously with the spatial task, subjects had to process an auditory stream of stimuli in which a single letter was presented in each auditory frame that had to be matched to the letter that appeared n items ago. The value of n was matched for the spatial and verbal tasks, both of which required responses. The level of n changed during the experiment according to the participants' performance to keep overall task difficulty approximately constant. Following training, subjects were given non-overlapping items from an instrument measuring Gf. The results showed that training on a dual n-back task yielded improvements in Gf relative to a control group that did not train.

Why was this training regimen successful? That is, what mechanisms drive such transfer effects? We believe it is critical that the training and the transfer tasks share overlapping cognitive processes for transfer to succeed. Thus, we think that the gain in Gf emerges because the processes that are engaged by the training task also mediate performance in Gf tasks. We proposed that the framework by Halford, Cowan, and Andrews (2007) might serve as a useful model to understand why Gf can be improved by means of a working memory task. Their claim is that working memory and intelligence share a common capacity constraint, which is driven by attentional control processes. Other authors have come to a related conclusion (Gray et al., 2003, Kane et al., 2004), and in particular, Carpenter, Just, and Shell (1990) have proposed that the ability to derive abstract relations and to maintain a large set of possible goals in working memory accounts for individual differences in typical tasks that measure Gf. The underlying neural circuitries provide additional evidence for the shared variance between working memory and Gf in that both seem to rely on similar neural networks, most consistently located in lateral prefrontal and parietal cortices (Gray et al., 2003, Kane and Engle, 2002). Thus, it seems plausible that the training of a certain neural circuit might lead to transfer to other tasks that engage similar or at least overlapping neural circuits. Indeed, recent evidence shows that transfer occurs if the training and the transfer task engage overlapping brain regions, but not if they engage different regions (Dahlin, Neely, Larsson, Backman, & Nyberg, 2008; see also Persson & Reuter-Lorenz, 2008).

But overlapping processes and neural circuits might not be the only prerequisites for transfer. We believe that the training task has to be very carefully designed in a certain way to promote transfer. First, a successful training task must minimize the development of strategies that are specific to the task in question because the object of training must be changes in the information processing system, not changes in the way one particular task is performed (cf. Ericsson & Delaney, 1998). Second, we think that it is very important to keep a persistently high level of training demand while also considering inter-individual performance differences. This can be achieved by using an adaptive training method that continuously adjusts the current training difficulty to the actual performance of each subject. Third, we argue that it is necessary to stress the information processing system during training, for example by taxing more than one input modality at a time or by having the subject engage in two tasks simultaneously (Oberauer, Lange, & Engle, 2004). As we have shown in our work, the dual n-back training paradigm is a task that fulfills these requirements and subsequently leads to the predicted transfer effects (Jaeggi et al., 2007, Jaeggi et al., 2008). Nevertheless, although various versions of the n-back task are widely used in research, only few studies have examined the processes involved in n-back performance (e.g. Hockey and Geffen, 2004, Jaeggi et al., 2010, Kane et al., 2007). Therefore, little knowledge is available about the cognitive processes that mediate performance in this task and consequentially, about the processes underlying n-back training that eventually promote transfer to Gf. In addition, although the n-back task is commonly regarded as a measure of working memory, its concurrent validity is still open to question (Jaeggi et al., 2010, Jarrold and Towse, 2006, Kane et al., 2007, Oberauer, 2005). For example, research from Kane's lab as well as our own work suggests that the n-back task and more traditional measures of working memory capacity (e.g. reading span or operation span tasks) do not share a great deal of common variance, although they independently predict performance in Gf tasks (Jaeggi et al., 2010, Kane et al., 2007). This is in line with findings from training on the n-back task which leads to improvements in Gf (Jaeggi, Buschkuehl, Jonides, & Perrig, 2008), but not in measures of working memory capacity (Jaeggi et al., 2008, Li et al., 2008). Therefore, we do not know whether training on an n-back task results in transfer to Gf due to an improvement in basic working memory processes, or whether there are other processes that are better predictive of such transfer.

Study 1

The main goal of Study 1 was to document the results of a correlational study investigating the relationship between the n-back task and selected cognitive tasks chosen so that they might reveal factors that underlie the transfer effect that we have observed by training on the dual n-back task. Based on our own work and Kane's work, we included measures of matrix reasoning and a measure of working memory capacity (Jaeggi et al., 2010, Kane et al., 2007). Further, as we were interested in investigating the transfer potential of a simpler n-back task version as well as the dual n-back task, we included both single and dual n-back task versions. There are four reasons to investigate the transfer potential of a single n-back task: First, the dual n-back task is relatively new and not much is known about its constituent processes (Jaeggi et al., 2007, Jaeggi et al., 2008, Jaeggi et al., 2009, Jaeggi et al., 2003). Second, the dual n-back task is inherently complex, and so it is not easy to disentangle the underlying processes. Third, the dual n-back task includes an obvious task-switching component (i.e., going back and forth between the two stimulus streams that must be tracked). This task-switching component might contribute to increased reasoning performance because in many matrix reasoning problems, it seems important to be able to switch back and forth between different representations. However, it is not at all clear that task-switching processes are an essential component of Gf; thus, if task-switching processes are not critical to matrix reasoning, then a single n-back task should correlate just as well with Gf as a dual n-back task. Finally, the dual n-back task is very challenging for participants, thereby restricting its range of application mainly to healthy young adults. We know from our previous research that a frequently used and well-established single n-back task recruits similar neural networks to a dual n-back task (Jaeggi et al., 2003), and also, that single n-back tasks share common variance with Gf tasks (e.g. Gray et al., 2003, Hockey and Geffen, 2004, Jaeggi et al., 2010, Kane et al., 2007). Thus, we investigated the relationship between single n-back performance and measures of Gf, and whether and how this relationship is different from that of the dual n-back task and Gf. We also investigated the role of working memory capacity, hypothesizing that working memory capacity predicts performance in n-back tasks, however, to a lesser extent than Gf.