This kind of connectivity analysis was applied to fMRI data in 59 individuals who also had completed the WAIS IQ test (Song et al., 2008). Usually, fMRI is conducted while the participants perform a cognitive task. As different studies use different cognitive tasks, comparing results is often problematic because each task has its own cognitive requirements that involve different brain areas. In this case, the functional connectivity was determined using fMRI data acquired during a rest condition. In other words, no cognitive task was performed while fMRI data were obtained. The idea was to test whether brain activity at rest might reveal functional connections related to IQ. A consistent pattern of resting-state brain activity has been characterized as a “default network.” That is, the pattern of brain activity when a person is not engaged in a cognitive task tends to be a stable pattern of maintenance involving specific brain areas, rather than a completely random pattern of uncorrelated, chaotic activity.
In this study, the seed was placed in a part of the frontal lobes corresponding to where Brodmann areas (BAs) 46 and 9 come together (see PFIT Figure 3.7); one seed was in each hemisphere. In the first step of the analyses, the resting-state functional connectivity between the seeds and the rest of the brain was determined statistically by correlating the blood flow value in the seeds to blood flow values in all other voxels. Several statistically significant connections were identified. As expected, some connections between the frontal seeds and other brain areas were stronger than others (i.e., stronger correlations). In the second step, the strength of the connections was correlated to IQ scores. The strongest IQ correlations were for connections between areas noted in the PFIT model. Moreover, this study indicated that individual differences in resting-state default network activity were related to IQ.
Soon thereafter, several studies reported the use of a better statistical method for inferring brain networks and how they relate to intelligence. The method is called graph analysis (Reijneveld et al., 2007; Stam & Reijneveld, 2007), a more mathematically sophisticated approach that determines how every voxel (also called a node in graph analysis) is correlated to all other voxels and how strong the connections are (connections are called edges). Graph analyses can be computed on structural or functional imaging data. Some nodes are hubs with many connections to other nodes. Networks in the brain tend to be “small-world” connections in that most clustering of connectivity is around adjacent brain areas or “neighborhoods.” There are also connections among more distant regions in the brain through hubs that are connected to each other; these are the so-called rich clubs (van den Heuvel et al., 2012). Small-world networks tend to allow more efficient transmission of information across shorter distances with less wiring (white matter fibers) and rich clubs foster faster communication across more distant brain areas. These networks develop at different times and rates from infancy through early adulthood. The factors that influence how networks develop are not yet understood, but they likely are related to individual differences in cognitive abilities. Graph analysis is illustrated in Textbox 4.1.
Textbox 4.1: Graph analysis
Graph analysis is a mathematical tool that is used to model brain connectivity and infer networks. The idea is to establish how each voxel in a brain image is correlated to all other voxels throughout the brain. These connections, called edges, can be computed for structural or functional images. A voxel, or a cluster of voxels, that show correlations to many other voxels is called a hub. Hubs that show correlations to many other hubs are called rich clubs. The strength of any connection is determined by the magnitude of the correlation between voxels or hubs. The efficiency of any connection can be estimated by determining its length. Most of the brain has local connectivity in that many nearby voxels are connected to each other via a neighborhood hub. This makes for efficient information transfer. Rich clubs connect more distant brain areas and this makes for faster communication. This is illustrated in Figure 4.1 from van den Heuvel and Sporns (2011). Psychometric test scores can be correlated to the strength of hubs and connections to indicate which brain networks are related to intelligence, as described in Section 4.1.
Figure 4.1 Brain connections determined by graph analysis. The red nodes show brain areas with many connections (larger nodes indicate more connections). Blue lines called edges indicate the strength of connections among areas (thicker lines indicate stronger connections; dark blue lines show rich club connections to other brain areas).
Adapted with permission from van den Heuvel and Sporns (2011).
Van den Heuvel and his colleagues applied graph analysis to fMRI data collected during a resting state in a small sample of 19 adults (van den Heuvel et al., 2009). They calculated a measure of global efficient communication among multiple brain areas based on the overall length of pathway connections. This measure was inversely correlated with IQ scores. In other words, higher IQ scores were related to shorter pathways indicative of greater efficiency of information transmission within the entire brain. Path length of frontal–parietal connections had the strongest inverse correlations to IQ. Similarly, another group (Song et al., 2009) reported a graph analysis targeted specifically at the default network and how it differed between high- and average-IQ subgroups (N = 59). They too found that differences in the overall global efficiency of connections in the default network were related to IQ. The high-IQ group showed greater efficiency. A different research group reported a graph analysis of global efficiency using fMRI obtained from 120 participants (Cole et al., 2012). After a whole-brain analysis, they reported that efficient connections involving only the left dorsal lateral prefrontal cortex with other frontal–parietal connections were correlated to intelligence test scores. Other researchers used graph analysis on resting-state EEG data in 74 participants and reported that efficient connections centered in the parietal lobe were most correlated to intelligence test scores (Langer et al., 2012).
Santarnecchi and colleagues also used graph analysis based on resting-state fMRI data obtained from 207 individuals across a wide age range and found IQ scores were related to connections distributed around the brain, including PFIT areas (Santarnecchi et al., 2014). Both strong local connectivity and weak distant connectivity were found, but their study added a new and surprising observation: IQ was most related to the strength of the weaker long-distance connections than to the stronger shorter connections. These researchers also reported quite a clever experiment using graph analysis and “damage” created mathematically. First, they constructed a measure of brain resilience based on functional connectivity related to IQ scores. Then they tested the impact of “damage” to specific areas or to random ones (Santarnecchi et al., 2015a). They concluded that higher intelligence was related to brain resilience to targeted damage and that the key areas were consistent with the PFIT. Supporting this general conclusion, recall from Chapter 2 that the Val/Met gene related to BDNF might play a role in preservation of IQ after traumatic brain injury.
Using the same sample of 207 and fMRI data, this same group has reported a different type of connectivity analysis based on the functional correlations between the same brain areas in the right and left hemispheres (Santarnecchi et al., 2015b). This is called homotopic connectivity and the results were counter-intuitive. Higher IQ was correlated to brain areas that show weaker inter-hemispheric homotopic connectivity, suggesting that decreased inter-hemisphere communication is related to higher intelligence. Several of the homotopic areas are included in the PFIT, but this study adds a new dimension of inter-hemisphere communication. Age and sex differences were also reported. For example, higher-IQ females showed less homotopic connectivity in the prefrontal cortex and posterior midline regions. Younger participants (below age 25) with higher IQs also showed increased homotopic connectivity patterns. These age and sex analyses were done on smaller subsamples so must be viewed with caution, but they illustrate the potential importance of using these variables as a matter of routine. Because the results are based on resting-state data, one wonders wh
ether functional homotopic relationships with IQ might be even stronger if based on fMRI during a cognitive task. Wonder no longer. In a 2014 study of 79 participants, networks were identified from both resting-state fMRI and fMRI during problem-solving of items from the Raven’s test (Vakhtin et al., 2014). This is the only imaging study of intelligence to date that investigated both resting-state and task activation conditions in the same subjects. Connectivity was determined with a statistical technique called independent component analysis prior to the homotopic analysis reported subsequently (Santarnecchi et al., 2015b). Functional connectivity during the problem-solving overlapped with functional connectivity during the resting state. The overlapping networks were consistent with the PFIT.
The PFIT framework also has received strong support from other methods of voxel-wise analyses of brain connectivity (Shehzad et al., 2014) and from both an evolutionary perspective (Vendetti & Bunge, 2014) and a developmental perspective (Ferrer et al., 2009; Wendelken et al., 2016); both perspectives emphasize the importance of parietal/frontal connectivity for reasoning ability. The PFIT hypothesis about subnetworks underlying different cognitive functions has received strong support from experiments using non-verbal reasoning tasks during standard fMRI (Hampshire et al., 2011). Overall, it is clear that results from numerous network analyses that use different methods converge substantially and support the existence of intelligence-related networks that are distributed across the brain. The findings are consistent generally with the PFIT framework, although that framework is subject to modification and elaboration, or even disproof, as new data emerge.
Many analyses identify networks mathematically irrespective of actual brain anatomy. White matter fibers are the tangible structural units of the brain that transmit information from one area to another. Figure 4.2 shows white matter fiber throughout the brain as determined by DTI. See Animation 4.1 showing DTI of the white matter connections between left and right hemispheres (the corpus callosum) on this book’s website (www.cambridge.org/us/academic/subjects/psychology/cognition/ neuroscience-intelligence). The thickness of the corpus callosum has been related to intelligence (Luders et al., 2007).
Figure 4.2 White matter fibers throughout the brain assessed by DTI. Seed refers to a spot selected for determining connections from that spot to other areas (see Animation 4.1 on this book’s website, www.cambridge.org/us/academic/subjects/psychology/cognition/neuroscience-intelligence) (courtesy Rex Jung).
One group of researchers (Li et al., 2009) reported a graph analysis specifically of white matter connections to assess brain efficiency. In the previous chapter, we introduced DTI as a special variety of MRI that assessed the integrity of white matter. The Li group used DTI in 79 young adults. Among other findings, global white matter efficiency was greater in the high-IQ subgroup. They concluded that “… higher intelligence scores corresponded to a shorter characteristic path length and a higher global efficiency of the networks, indicating a more efficient parallel information transfer in the brain … Our findings suggest that the efficiency of brain structural organization may be an important biological basis for intelligence.”
Another research group assessed white matter with DTI in 420 older adults (Penke et al., 2012). They did not find that any one white matter tract was highly correlated to intelligence scores. However, they reported that 10% of the variance in intelligence test scores could be explained by a general factor of global white matter integrity computed from all tracts combined. This effect was due entirely to a factor of information processing speed. Another group (Haasz et al., 2013) reported similar findings in middle-aged and older adults. Other researchers calculated white matter/intelligence correlations separately for males and females in a small sample of 40 young adults (Tang et al., 2010). The pattern of correlations with IQ differed between the sexes. Although their sample sizes were too small for generalization, from the perspective of individual differences and known sex differences in the brain (Luders et al., 2004, 2006), there is a strong argument for always computing separate analyses for males and females, especially when both groups are matched for intelligence.
Another kind of study that examines brain networks is based on patients with brain lesions and the pattern of cognitive deficits that result. Prior to the advent of neuroimaging, the study of brain lesion patients was a primary, if inexact, source of data for inferring brain/intelligence relationships. Neuroimaging advanced this approach by providing exact localization of lesions and mapping correlations between cognitive test score deficits and brain parameters. For example, Glascher and colleagues assessed primary factors of intelligence including g in a sample of 241 neurological patients with brain damage (Glascher et al., 2009, 2010). The main finding was that damage in frontal and parietal areas was related to deficits in the g-factor and that other intelligence factors (verbal comprehension, perceptual organization, and working memory) showed deficits when damage occurred in different parts of the frontal–parietal network (see Figures 4.3 and 4.4). Brain/intelligence relationships were also tested by other researchers with structural MRI, fMRI, and DTI in a small sample of people with Shwachman–Diamond syndrome, a rare genetic disorder characterized in part by various cognitive impairments (Perobelli et al., 2015). They found brain abnormalities consistent with the PFIT.
Figure 4.3 3D renderings show cortical and subcortical regions with a statistically significant relationship (red/yellow) between lesion location and the g-factor (top row). Bottom rows: Axial (horizontal) slices are shown for a more detailed inspection.
Reprinted with permission, Glascher et al. (2010, figure 2, p. 4707).
Figure 4.4 Effect of lesion location on four indices of mental ability. Row (A) is perceptual organization, (B) verbal comprehension, (C) working memory, and (D) processing speed. The red/yellow colors show where the location of lesions significantly interferes with scores on the indices. The graphs on the right show the mean difference on each index score between patients with and without lesions at the area of maximum effect (white arrow on the 3D projection).
Reprinted with permission, Glascher et al. (2009, figure 2, p. 684).
As we see, the PFIT framework has generated support from a number of studies and new data are providing some potential refinements to it. For example, one research group expanded neuroimaging beyond the cortex to subcortical areas (Burgaleta et al., 2014). They analyzed the shape of several subcortical structures based on MRI in 104 young adults who had completed a battery of cognitive tests. Fluid intelligence scores, highly correlated to the g-factor, were related to the morphology of the nucleus accumbens, caudate, and putamen, all in the right hemisphere only. These areas and the morphometry of the thalamus were also related to the factor of visual–spatial intelligence. Another study reported that the volumes in the basal ganglia were correlated to different intelligence factors and there were some sex differences (Rhein et al., 2014). Both these studies require replication, but they expand the PFIT framework to subcortical areas. There is also some evidence regarding the PFIT in children. One new study reports that efficient structural brain networks related to the PFIT are related to perceptual reasoning and to one high g-loaded measure in a sample of 99 children aged 6–11 years old (Kim et al., 2016).
In a comprehensive report, a German group of researchers completed a detailed meta-analysis of neuroimaging studies of intelligence through 2014 with the explicit purpose of testing the PFIT (Basten et al., 2015). In their final analysis, they only considered studies where individual differences in intelligence could be assessed directly; studies of average group comparisons were excluded. Jung and Haier had included both kinds of studies and their PFIT analysis was based on a qualitative assessment of areas common across studies. The German group compared structural and functional imaging results in an empirical voxel-by-voxel analysis (VBM as described in Chapter 3) to identify common brain areas related to intelligence across 28 studies totaling over 1,000 participants. They concluded that the results generally supported th
e primary involvement of the parietal–frontal network They also found evidence that suggested revising the PFIT to include areas of the posterior cingulate/precuneus, caudate and midbrain. Their revised framework is shown in Figure 4.5. These suggested changes need replication.
Figure 4.5 Brain areas related to intelligence from a 2015 review are shown on lateral (left) and medial (right) surfaces of the brain. ACC, anterior cingulate cortex; PCC, posterior cingulate cortex; PFC, prefrontal cortex; (pre) SMA, (pre-) supplementary motor area; VBM, voxel-based morphometry. Reprinted with permission, Basten et al. (2015).
Whereas the PFIT was a good start, a more advanced model is necessary so that more specific predictions can be tested. Frameworks like the original and revised PFIT have conceptual problems related to a reliance on correlations that are fundamentally not interpretable regarding cause and effect between brain measures and cognitive measures (Kievit et al., 2011). One promising possibility for addressing this limitation that might advance the study of “neuro-g” may be the use of analyses based on multiple indicators and multiple causes (Kievit et al., 2012). These advanced statistical approaches, too complex to detail here, can generate more specific hypotheses about brain variables and they are particularly important for the large data sets. They offer the potential for clarifying the weight of evidence regarding how brain physiology specifically relates to cognitive measures, especially for identifying different brain variables relevant for individual differences in cognitive test performance (see Section 4.3).
The Neuroscience of Intelligence Page 15