So, where in the brain is intelligence? The commentators provide us with a wide range of insights about our interpretation of neuroimaging studies that speak to this deceptively simple question. Happily and unhappily, the criticisms are relatively mild. On one hand, intelligence research has matured to a point past earlier vehement arguments about whether or not intelligence is a proper subject for scientific investigation. On the other hand, perhaps our parieto-frontal integration theory (P-FIT) model is too tame to invoke strong criticism, at least from other intelligence researchers. Hence, we attempt to be a bit more provocative here for the sake of discussion.

R1. Not a pretty picture

The 19 commentaries generally fall into three categories: Most of them question aspects of our conceptual framework, others focus on specific discussion of P-FIT details, and a few seek to generalize the P-FIT to other theories and domains. Within these categories, most of the commentaries are quite disparate in their emphasis. Therefore, we will discuss each in turn, but first we wish to offer our own pointed criticism of the P-FIT. At best, the P-FIT is a rather bland listing of the brain areas which may be related to intelligence. The important points of our review are that such areas can be identified empirically (after 20 years of neuroimaging research) and that there are multiple areas distributed throughout the brain rather than a single region within the frontal lobe. Whether the P-FIT list is fully accurate is addressed in one way or another by most of the commentaries, but only additional research will firmly establish the relevant areas linking brain and intellectual functioning. The P-FIT suggests that integration among the areas is important for intelligence, but the P-FIT is largely silent about how the areas work and how the areas communicate with each other. Until we can address these issues, the P-FIT is only a modest beginning, as noted by many of the commentators. Brains and minds are beautiful and elegant. As it stands, the P-FIT model is not. We are hopeful that these qualities, necessary for a meaningful theory of the neurobiology of intelligence, will emerge with time.

We begin with acknowledging the historical perspective noted by Hunt, a pioneer in intelligence research. In some ways, our review marks the end of the first 20-year phase of neuroimaging research related to intelligence. As he notes, the studies we reviewed in the target article move well beyond earlier controversies about whether or not there is a neurobiological basis for intelligence, and well past behaviorist dogma that the brain is an unknowable black box. As Hunt suggests, we can now address specific sources of individual differences in intelligence by studying individual differences in brain structure and function. Surely, if Binet or Terman were alive today, they would be avidly applying neuroimaging to move beyond the inherent limitations of psychometrics. The next phase of neuroimaging studies should focus more on multiple measures of intelligence (from psychometrics) and cognition (from cognitive psychology) in using large samples stratified across the range of intellectual ability, gender, and age. Several research groups are doing just this, and are incorporating experimental manipulations of brain areas with drugs, and electrical/magnetic stimulation, to test the role of specific brain regions and networks on intellectual ability. This experimental phase will be more challenging, it will take some time, and it might depend upon imaging, analysis, or data fusion techniques currently unavailable; however, this phase is now poised to begin in earnest based on the progress of the last 20 years. Hunt also reminds us that understanding the genetic aspects of intelligence remains a long and arduous task. Since we wrote our review and described relatively new work on the specific ASPM and Microcephalin genes, new results have been published which dampen enthusiasm that these genes might underlie the relationship between whole brain size and intelligence (Woods et al. Reference Woods, Freimer, De Young, Fears, Sicotte, Service, Valention, Toga and Mazziota2006), although whether these genes or others (Meyer-Lindenberg et al. Reference Meyer-Lindenberg, Straub, Lipska, Verchinski, Goldberg, Callicott, Egan, Huffaker, Mattay, Kolachana, Kleinman and Weinberger2007) are related to the size or functioning of specific P-FIT areas remains to be determined.

R2. Some conceptual clouds – why study the brain?

Sternberg's commentary focuses on the importance of interactions between brain biology and environmental factors. He chastises us for asking the wrong question because he doubts that intelligence resides in the brain, and he provides examples of nonhuman intelligence in the absence of any brain at all. Sternberg credits us with answering instead the question, “Where in the brain are there functions that, when measured in particular ways, show correlations with scores on conventional IQ-related tests?” Two issues are entangled here: the relationship between intelligence test performance and real world adaptation/intelligence, and how to formulate questions about the localization of complex abilities in the brain. Regarding the first point, there is considerable evidence that intelligence test performance and many life adaptations are highly related (Gottfredson Reference Gottfredson and Nyborg2003; Jensen Reference Jensen1998). Regarding the second point, consider the question, “Where in the brain is vision?” Surely the eyes are important and there can be no vision without something to see. Nonetheless, we have been able to investigate how vision works by studying the brain. We take Sternberg's point that knowledge about brain structure and function alone may not be sufficient to answer all questions of interest to intelligence researchers, but we believe that our question is not ill put, and is one of many legitimate questions. The P-FIT is based solely on our review of existing neuroimaging studies, all of which are correlational and use “conventional” tests including but not limited to IQ tests. New neurobiology studies will surely expand the type of measures and research designs used to study intelligence beyond current neuroimaging approaches, but we do not share Sternberg's assertion that there is an “optimal problem” about intelligence and that we should focus only on solving it, whatever it may be. We know of no other scientific field where “optimal problems” are agreed upon and all other problems are deemed “wrong.”

Norgate & Richardson similarly question both the empirical and conceptual basis for a brain-based theory of intelligence, asserting that inconsistencies among studies reviewed show that the correlational basis of the P-FIT must be unreliable. We based the P-FIT on consistencies among the studies as we saw them and ignored many obvious inconsistencies – a necessary step in reviewing a nascent literature. Some questions about the nature of g or IQ are still open in the psychometric domain and the P-FIT bridges psychometric research on these issues with brain research. The alternative explanation of the positive manifold underlying g (van der Maas et al. Reference van der Maas, Dolan, Grasman, Wicherts, Huizenga and Raijmakers2006), for example, is helping to define more possibilities for brain relationships to intelligence and cognitive measures, not fewer. The P-FIT will evolve as we follow these new observations.

Norgate & Richardson also raise an important point about socioeconomic status (SES). Only one of the neuroimaging studies we reviewed included SES measures (Shaw et al. Reference Shaw, Greenstein, Lerch, Clasen, Lenroot, Gogtay, Evans, Rapoport and Giedd2006). This study did not relate SES to brain measures, but rather found that SES was significantly correlated with IQ in their sample of young people (r=–.35, p<.01). Some simple empirical questions that need attention are: In predicting scores on intelligence measures from neuroimaging data, how much variance would be accounted for by adding a measure of SES? Would this be the same amount of variance at all ages or more in younger individuals? Is there a correlation between SES and regional brain volume, especially in P-FIT areas? Along these lines, Noble and colleagues recently reported that fMRI results on a reading-related task differed in children as a function of SES (Noble et al. Reference Noble, Wolmetz, Ochs, Farah and McCandliss2006). The inclusion of SES measures can enrich neuroimaging studies to the extent that they can establish unique sources of variance.

As noted in Hunt's commentary, the P-FIT focuses on brain relationships to intelligence and not on the other aspects noted by Norgate & Richardson. A focus on brain structure and function does not diminish the importance of other factors, nor the importance of interactions between the domains. Researchers working on the cognitive and social aspects of intelligence do not need to diminish the contributions of brain research (or psychometric) perspectives to advance the importance of their theoretical focus. Most research on intelligence is correlational. As we have noted earlier, brain imaging opens the potential for new experimental approaches to a variety of questions about intelligence.

Rae points out that the P-FIT focus on the neuroimaging aspects of intelligence is, in fact, not focused enough because data from molecular and systems biology are not yet available. She notes some of the conceptual difficulties inherent in neuroimaging studies of intelligence, acknowledges that we must nonetheless start somewhere, and remarks that we have been suitably cautious in “sieving the data down to groups of mutually inclusive spots.” She notes that the P-FIT is silent about how effective integration among brains areas happens, a key point. Indeed, our very first PET (positron-emission tomography) imaging study, for example, found inverse correlations between cerebral glucose metabolic rate and scores on the Raven's Advanced Progressive Matrices (RAPM; a high g test of nonverbal abstract reasoning) taken during the PET procedure (Haier et al. Reference Haier, Siegel, Nuechterlein, Hazlett, Wu, Paek, Browning and Buchsbaum1988). We interpreted this as showing that the less hard a brain was working, the better it solved the test problems which became known as the brain efficiency hypothesis. A number of studies have examined “brain efficiency” in various ways and generally support a role for efficient allocation of cognitive resources in intelligent individuals (see Neubauer & Fink Reference Neubauer and Fink2003; Neubauer et al. Reference Neubauer, Fink and Schrausser2002; Reference Neubauer, Grabner, Freudenthaler, Beckmann and Guthke2004; Rypma et al. Reference Rypma, Berger, Prabhakaran, Bly, Kimberg, Biswal and D'Esposito2006). We also note our recent study in which performance on intelligence tests was positively correlated with parietal and inversely correlated with frontal brain N-acetylaspartate (NAA), particularly in women, suggesting some level of sex-mediated biochemical optimization across the brain as underlying intellectual capacity (Jung et al. Reference Jung, Haier, Yeo, Rowland, Petropoulos, Levine, Sibbitt and Brooks2005). As Rae notes, new imaging technology using diffusion tensor imaging (DTI) and connectivity analyses will help determine how effective communication occurs among brain areas; molecular and systems biology surely play important roles, as demonstrated by the work she cites on bioenergetic capacity. We have proposed that neuronal mitochondria function (or even the number of mitochondria and variation in their structure) may be important in understanding how brain energy is created and consumed during cognition (Haier Reference Haier and Nyborg2003).

Rae also properly cautions that the P-FIT may characterize some but not all groups; we completely agree. We see the P-FIT areas somewhat like a tool kit. Some people have more, different, or better tools to apply to cognition and problem solving, forming the basis for individual differences in intelligence. Which subsets of the P-FIT areas work best for which individuals in which cognitive domains are open questions. Even studies of this kind, however, fall short of Rae's warning: we critically need more research on molecular and systems levels, in addition to investigating the location of “spots” in the brain. Here we have the full range of conceptual comment on the P-FIT's focus, from too much brain biology (Sternberg and Norgate & Richardson) to not enough.

Like Rae, Schmithorst notes the central importance of the concept of connectivity among brain areas implied in the P-FIT and the paucity of data available to test hypotheses about relationships among areas, despite the availability of sophisticated statistical methods to do so. He details an important conceptual distinction between effective connectivity and functional connectivity. The former refers to how much one brain area influences another area; the latter refers only to there being a connection or relationship between areas (i.e., “correlation”). Both can be studied effectively, especially in large samples. Schmithorst also points out that each of the neuroimaging studies we reviewed used a single imaging method. He notes the importance of using a combination of imaging techniques in the same sample in future studies. We know of several projects where this is under way and we completely agree that this will provide powerful data for understanding intelligence. We also agree that, in its current form, the P-FIT lacks sufficient specificity for how each area might contribute to intelligence, and the likely interactions of age and sex to such contributions. Again, we are partial to thinking about the P-FIT from the analogy of a tool kit. The current state of imaging research reviewed in the target article provides a tentative listing of brain “tools,” but additional research and analysis techniques of increasing sophistication, as nicely summarized by Schmithorst, will provide the empirical basis for detailing how, when, and for whom specific “tools” are applied.

R3. Darker conceptual clouds – What is “intelligence”?

Blair raises the most challenging conceptual critique of the P-FIT, simply and starkly: “general intelligence is a mathematical abstraction, not a thing in itself.” In early schizophrenia research, a claim was made that schizophrenia was a myth based on an abstract social construct of normality and not a medical illness or disease. A psychiatrist involved in adoption studies once remarked that if schizophrenia was a myth, it was a myth with a genetic component. However intelligence is defined, there is a genetic and, therefore, a biological component which is evident in neuroimaging studies. Blair's challenge (Blair Reference Blair2006) is based upon a careful distinction between general intelligence and fluid intelligence. Lee, Choi, & Gray (Lee et al.) also point out that the P-FIT may be more related to fluid intelligence than to general or crystallized intelligence. Both commentaries note that fluid intelligence refers to what some researchers think of as fundamentally “working memory/executive cognitive abilities.”

Blair's skepticism, however, is focused more on whether there is a neural basis of “general intelligence” to discover. He argues that theories of general intelligence must include consideration of experience and development. We have no argument with his position. The P-FIT is based on existing neuroimaging studies which have the limitations enumerated by us and other commentators. Chief among these is the reliance on single measures of intelligence rather than on a standard battery, which could provide analyses testing models of the general factor of intelligence (i.e., g) and on models of general intelligence (as we defined in our review; see also Colom's commentary). Early on, we attempted to incorporate test difficulty and/or levels of intelligence among subjects into our neuroimaging paradigms (e.g., Haier et al. Reference Haier, White and Alkire2003b; Larson et al. Reference Larson, Haier, LaCasse and Hazen1995), but much more needs to be done in studying these interactions (see Lee et al. Reference Lee, Choi, Gray, Cho, Chae, Lee and Kim2006). As indicated previously in this response, we will not be surprised if some subset of the P-FIT are more relevant than others, depending on what the tasks are and who the subjects are. Moreover, the subsets may also depend on which definition of intelligence is applied. How P-FIT areas develop and how sensitive they may be to experience or training are open questions central to evolving more detailed neural theories. In our view, it is still too early to rule out a neural basis for a general factor of intelligence independent of a neural basis for specific cognitive abilities.

Noveck & Prado also address fundamental questions of definition and caution us specifically that intelligence and reasoning are not the same thing. They argue that there is no advantage to study a “large” domain like intelligence over a subset domain like “reasoning,” and that a correlational-based individual difference approach, especially one dependent upon Wechsler Adult Intelligence Scale (WAIS) subtests, does not have any advantage over the investigation of component parts of reasoning as studied in the cognitive neuroscience tradition. However, as we noted in response to Sternberg's concept of an “optimal” problem, we are dubious that any one approach, either philosophical or experimental, is so superior over another in all instances that research should be limited to that approach alone. The potential merging of traditional psychometric research on individual differences with traditional cognitive psychology based on neuroimaging is welcome and exciting, and it is unlikely that either approach will subsume or diminish the other. Both will evolve to explain different aspects of higher cognitive function. For example, teachers are interested in why one child reasons better than another. Empirical investigations of this question may well require several well-integrated approaches, and perhaps the answer will have something to do with general intelligence, the g factor, and/or the P-FIT. Would it not be interesting, for instance, to know if cognitive imaging experiments designed to study components of reasoning show the identical results in subjects selected for high or low intelligence? What non-intuitive insights and conjectures might emerge if the results differed for men and women?

Going further than Noveck & Prado, commentators Cohen Kadosh, Walsh, & Henik (Cohen Kadosh et al.) focus on a single cognitive component in their commentary. This component is response selection and they make a strong case for its central role in any definition of intelligence. They note that response selection is the “interface between perception and action that allows one to choose the most adequate response among alternatives.” They are concerned that the parietal-frontal network is not specific to intelligence, but rather to the response selection problem they see as common to all intelligence measures (see our remarks on Naghavi & Nyberg's commentary further on). We are intrigued with Cohen Kadosh et al.'s notion but we find it unlikely that this single component alone is the cognitive key to intelligence. Separating out response selection from other components of cognition (e.g., reasoning) should be a problem amenable to neuroimaging research.

Colom calls attention to the inconsistencies in P-FIT areas derived from different imaging modalities, as well as the problem of defining the construct of human intelligence. He concludes that many of the inconsistencies among the neuroimaging studies we reviewed have their basis in that every measure used in these studies comprises several cognitive abilities along with some amount of the g factor, producing heterogeneous imaging results. One might conclude, therefore, that the component abilities should be studied separately, but Colom instead approaches the problem by focusing on the g factor. The practical implication of this approach is that future neuroimaging studies would do well to include a diverse battery of cognitive tests so that g can be extracted.

The commentary by Wilke bridges the disparate arguments made by Colom favoring focus on a general factor and the arguments by Noveck & Prado and Cohen Kadosh et al. favoring a focus on component abilities. Wilke understands that the P-FIT provides a framework of “several collaborating (and competing?) network nodes, with strengths in one (or more) explaining individual performance advantages.” This is consistent with the tool kit analogy for the P-FIT that we discussed earlier. Moreover, like Colom, Wilke takes the view that the P-FIT has identified a broad pattern based on available data, and that it is time to apply more sophisticated approaches both to global definitions of intelligence and to a possibly more promising effort for “an exact characterization of the cognitive functions” underlying performance on such intelligence tests. (Noveck & Prado also suggest this.) Wilke's caution about generalizing from cases of pediatric brain injury is particularly important (see our remarks on the Varley and Zaidel commentaries below).

The definition of intelligence continues to be a major concern, as these commentaries demonstrate. Psychometric approaches to definition issues may be reaching an inherent limit because no intelligence measures are based on a ratio scale with an actual zero point. Jensen has proposed shifting away from psychometrics to a focus on mental speed assessed with reaction time (RT) measures to overcome this problem (Jensen Reference Jensen2006), an approach he calls mental chronometry. Many studies show that higher intelligence scores are associated with faster reaction times. Jensen makes the following observation as a general rule: “The lower the grade of measurement used to represent the variables of interest, the more their quantitative description and analysis must depend upon complex statistical methods” (Jensen Reference Jensen2006, p. ix). He further notes that the lowest grade of measurement is typical in the behavioral and social sciences because it is based on ordinal scales. Chronometry, however, is based on a true ratio scale. Moreover, he notes that a chronometric g can be extracted from RT batteries and this is related both to psychometric g and to biological variables. Therefore, it will be important for future imaging studies of intelligence to include RT measures on a variety of tasks as well as psychometric batteries. The use of RT also bridges intelligence and cognitive research even more closely (see Thoma et al. [Reference Thoma, Yeo, Gangestad, Halgren, Davis, Paulson and Lewine2006] for an excellent example).

R4. Some rays of light on the P-FIT spots

As we began to characterize the results of our review, we realized that the P-FIT areas also appear regularly in cognitive neuroimaging studies. A comprehensive review of this literature by Naghavi & Nyberg was particularly helpful (Naghavi & Nyberg Reference Naghavi and Nyberg2005). As they point out in their commentary, cognitive studies have revealed a robust frontal-parietal network involved in several fundamental functions, including attention, working memory, and episodic memory. Our model, which is based solely on imaging studies of higher-order, global measures of intelligence, shows essentially the same brain network. Naghavi & Nyberg argue that the apparent multi-functionality of this network may be understood conceptually as a framework for integration and control of information. They regard the frontal-parietal network as uniquely suited for these functions, and this is consistent with the P-FIT role for intelligence. However, they caution us that this system may well be a more general backbone for bringing order to diverse brain functions rather than a more specific neural basis for intelligence. We are inclined to propose that intelligence may be defined as the degree to which the frontal-parietal network integrates and controls the flow of information throughout the brain. The sequence of flow through these areas and the efficiency of the flow at various points may be parameters that can be measured with neuroimaging methods, and correlated to measures of intelligence (g or otherwise). This data-driven conceptual merging of two empirical traditions, the psychometric and the cognitive, will advance our understanding of the neural basis of intelligence.

Prabhakaran & Rypma summarize some of their results from fMRI studies, which are more or less consistent with the P-FIT, but with some distinctions. They are inclined to see a more general posterior interaction with the prefrontal cortex than one focused on the parietal lobe. They argue that posterior brain areas are recruited for separate and varying task demands while the prefrontal cortex generally integrates among component task demands. When they divide subjects on the basis of fast or slow reaction time to a visual search task, they find that the slower group shows stronger evidence of frontal-parietal connection. For the faster (more efficient?) group, there is more independence between frontal and posterior areas. They also report evidence that efficient brain function was a key variable in that poor performers during a word generation task showed increased activity in frontal areas; a rest condition showed increased activity in parietal areas. Thus, Prabhakaran & Rypma are already working with the next generation of neuroimaging research designs described earlier, and more investigations of this type will be of obvious consequence to the P-FIT model.

Two commentaries, in addition to Wilke's, advise caution in the interpretation of neuropsychological data from brain-damaged individuals to support the neuroimaging findings in normal individuals which form the basis for the P-FIT. Varley notes that damage can cause reorganization of brain functions so that cognitive performance can be sustained, citing examples about language and aphasia. Intelligence, she notes, may be characterized by neural (and behavioral) flexibility and plasticity. We know very little yet about the dynamic cognitive reorganization manifested within brain tissue as a patient regains or improves a behavioral or cognitive skill. Sequential neuroimaging in brain damaged patients, soon after the injury and periodically during recovery (in tandem with cognitive rehabilitation as specific functions improve), could be quite informative in articulating the constraints that the P-FIT might place on how and where such skills are integrated into a larger cognitive framework. Zaidel makes a similar point and discusses examples based on performance of WAIS subtests, which she notes are difficult to link with specific brain areas. We tried to find support for the P-FIT in the lesion literature and concede that this is more complex than it may appear at face value. We favor Zaidel's hypothesis that, consistent with discussions above, “Abundant connectivity among widely localized neuronal centers would be expected for above-average performance on intelligence tests, whereas for average intelligence only limited connectivity may be sufficient.” It would seem that the P-FIT areas would be places to start connectivity analyses (see commentaries by Rae and Schmithorst). Whether the P-FIT applies to the other “intelligences” as listed by Zaidel depends on issues of definition discussed above and remains an open question.

R5. New fronts to consider

Three commentaries address how the P-FIT may generalize to other domains. Roring, Nandagopal, & Ericsson (Roring et al.) provide an interesting commentary on skilled and expert performance. They review studies which show that expert and skilled performance generally is unrelated to individual differences on traditional intelligence tests. We have no issue with their conclusion that future neuroimaging studies of intelligence and mental ability should address skill development in beginners and experts, especially over time. In fact, we were among the first to use neuroimaging to study brain changes associated with the development of visuospatial skill (Haier et al. Reference Haier, Siegel, MacLachlan, Soderling, Lottenberg and Buchsbaum1992a) and how those changes related to intelligence (Haier et al. Reference Haier, Siegel, Tang, Abel and Buchsbaum1992b), so we are quite interested in this perspective. Currently, we are undertaking a new multimodal neuroimaging study to track network connectivity changes associated with making the transition from novice to expert.

Kirov's commentary calls attention to the role that sleep and consciousness may play in intelligence. His summary of the sleep data is intriguing, but we need more data to determine if all the subcortical areas implicated in sleep are also implicated in intelligence. The lack of subcortical areas in the P-FIT has troubled us, although we do make reference to the “mission critical” nature of these subcortical structures as compared to the individual differences manifested in the cerebral cortex. Subcortical areas allow interesting inferences about neurotransmitter systems, which surely must be implicated in neural theories of intelligence. We have published several studies which use anesthetic drugs to manipulate states of consciousness during neuroimaging in the hope of identifying the subcortical areas, and the neurotransmitter systems, relevant for consciousness (Alkire & Haier Reference Alkire and Haier2001; Alkire et al. Reference Alkire, Pomfrett, Haier, Gianzero, Chan, Jacobsen and Fallon1999). As far as we are aware, there is not yet any confirmation that the brain systems relevant for consciousness studies are also implicated in intelligence studies.

Demetriou & Mouyi compare the P-FIT to a model of intelligence based on psychometrics and development (i.e., the PSY-DEVO model). After noting some conceptual points of agreement between the two models, they apply the “So what?” test. Overall, their commentary challenges whether the areas identified in the P-FIT could be the only areas necessary for intelligence, since so many cognitive processes and networks must be involved. For us, this is an empirical as opposed to a rhetorical question. It may be that the list of P-FIT areas will expand as future neuroimaging studies of intelligence become available. Existing studies, however, support the idea that some brain areas are more relevant than others for the components underlying higher-order cognition and these areas are consistent with the P-FIT (see Naghavi & Nyberg's commentary). Undoubtedly, many networks are necessary even for much simpler brain functions than those active during reasoning and problem solving. Each of the P-FIT areas easily could be part of multiple networks, but it should be possible to address which networks and areas define a core for intelligence.

R6. Conclusion

Although we have expressed our view in favor of parsimonious, elegant approaches at this early stage of neuroimaging studies of intelligence, Nyborg finally reminds us that the world is exceedingly complex. We admire his attempt at integrating findings across multiple domains at different levels of analysis. His commentary tantalizes us with a possible approach to resolving the persistent dualist dilemma, but we cannot relate his large-scale conceptual model to the relatively small-scale empirically based P-FIT. We are of a mind that a deep understanding of brain/mind relationships will come when we can describe the underlying complexities with elegant clarity. We are inspired by those who study the first few nanoseconds of the Big Bang at the creation of the universe billions of years ago. If they can seek to describe this unseen moment with precision and clarity, surely we have a chance to capture the essence of a neural basis of intelligence.