The hierarchy of cognitive functions

Why are some people better than others at performing particular cognitive tasks, such as arithmetical operations? Several possibilities present themselves to us. It might be that some people are just better than others generally in doing cognitive tasks. It might be that some people are better than others at mental work that involves numbers and their manipulation and also other related tasks. It might be that some people are just better than others at doing a particular arithmetical task. Of course, there could be other factors that operate on a particular occasion to produce a good, medium or bad score. The answer is that all of these suggestions are correct: there are demonstrable differences between people in overall cognitive ability, in separable domains of cognitive ability, and in specific cognitive tasks. There are also occasion-specific factors affecting mental performance and error of measurement.

When we try to model variance in cognitive tasks – including genetic contributions to the variance – we need to consider the fact that human cognitive variance occurs at these different levels in a hierarchy. At the highest level of the hierarchy, which covers the most general cognitive variation, is what is known as general cognitive ability. It is also sometimes known as general intelligence or just ‘g’. This was discovered by Spearman (Reference Spearman1904) using data from English schoolchildren. Spearman found that there was a tendency for all cognitive tasks to show positive correlations. He hypothesized that people were better at some tasks than others because they were more or less endowed with general cognitive ability and also specific abilities for each particular task. It was realized that this explanation, although accurate, was insufficient. That is because it was empirically demonstrated that some types of tasks correlate more strongly within a particular type of cognitive domain (e.g. spatial or verbal) than with tasks outside of that domain. Therefore, an accommodation was made such that variation in cognitive performance was describable at three levels: at general cognitive ability, at the level of broad cognitive domains, and at the level of specific skills. Probably the first person to describe this hierarchy clearly was Vernon (Reference Vernon1940).

The best evidence for the hierarchy of cognitive function variation came from Carroll (Reference Carroll1993). In this book, he carried out a massive empirical task by re-analysing over 400 datasets from many laboratories that contained a range of mental tasks applied to large samples. These had been gathered over much of the 20th century. Carroll found that all of these datasets conformed to a correlational structure that was best described in a three-level hierarchy. That is, he confirmed Vernon's suggestion that, to understand the correlations among mental tests, people's performance (variance) was best described as a three-level hierarchy with general cognitive ability at the peak, major cognitive domains at the next level, and specific cognitive abilities at the bottom. The g factor at the peak of the hierarchy tended to account for between 40% and 50% of the total variance in mental test performance.

The three-level hierarchy has further implications. It is important to appreciate that the cognitive domains at the second level are very highly loaded on g (Deary et al. Reference Deary, Whalley, Lemmon, Crawford and Starr2000, Reference Deary, Penke and Johnson2010; Deary, Reference Deary2001a–c). That is, they are not independent and they derive much of the variation from g. Thus, the reason that people are good at domains like verbal, spatial, reasoning, speed and other cognitive domains is that they are high on g and also that they have some more specific capability associated with that domain. One should not make the mistake of thinking that ability on a given domain is independent of g. This also applies to the specific skills too: being good at a very specific mental ability is partly to do with that specific skill, partly to do with the domain with which it is associated, and also partly to do with general cognitive ability.

The general cognitive ability factor has several advantages for genetic studies. It has high stability of individual differences across most of the human life course (Deary et al. Reference Deary, Whalley, Lemmon, Crawford and Starr2000). The general factor from different cognitive test batteries ranks people almost identically (Johnson et al. Reference Johnson, Bouchard, Krueger, McGue and Gottesman2004). The general factor from a number of cognitive tests can be extracted and used as a score to indicate people's level of general cognitive ability. This can be done using various multivariate statistical methods, such as principal components analysis (PCA), and exploratory and confirmatory factor analysis. Of course, there arises the question of whether people would be ranked the same or differently when the g factor was based on different cognitive test batteries. This has been studied, and the result is that, when large samples of people have been tested on different cognitive test batteries, the general cognitive factor (g) derived from them correlates very highly, often near to 1; that is, g factors derived from different groups of tests rank people almost identically (Johnson et al. Reference Johnson, Bouchard, Krueger, McGue and Gottesman2004, Reference Johnson, te Nijenhuis and Bouchard2008). This is useful for genome-wide association studies of general cognitive ability. It means that, although different studies have used different cognitive tests, if each has used a sufficient number of sufficiently diverse cognitive tests, then the g derived from each of them may be comparable, and used to indicate a similar trait across studies in meta-analyses. The structural and functional brain correlates of general ability differences are also increasingly well understood (Jung & Haier, Reference Jung and Haier2007; Deary et al. Reference Deary, Penke and Johnson2010). While there are genetic effects – as shown from behavioural genetic studies – at the three levels (Rijsdijk et al. Reference Rijsdijk, Vernon and Boomsma2002; Deary et al. Reference Deary, Johnson and Houlihan2009, Reference Deary, Penke and Johnson2010), much of the genetic influence is on the general mental ability factor. The heritability of general cognitive ability is well established, with an additive genetic contribution that rises from quite low levels in early childhood to over 60% throughout adulthood.

In addition to the three-level hierarchical model of cognitive ability differences – with a general cognitive factor at the apex – having much replication in psychometric studies, it should also be addressed whether there is validating biological evidence for such a structure. This has been addressed by a number of authors, and some key examples are given here. These are examples where biological effects have been examined with respect to the hierarchical model of cognitive differences. In each case formal hypothesis testing using structural equation modelling was used to examine influences on g and the other cognitive domains. Thus, for example, the major influence of age is on the g level of the hierarchy with some additional effects on memory and processing speed domains (Salthouse, Reference Salthouse2004). The well-established, modest correlation between brain size – estimated in healthy individuals using magnetic resonance brain scanning – is largely captured by the association between total brain volume and g (MacLullich et al. Reference MacLullich, Ferguson, Deary, Seckl, Starr and Wardlaw2002). Other g–brain associations are reviewed by Deary et al. (Reference Deary, Penke and Johnson2010). In addition, behavioural genetic studies show that the principal genetic influence on cognitive functions is an additive genetic influence on g, and that genetic effects on more specific domains of cognitive function largely derive from the genetic influences on g (Deary et al. Reference Deary, Johnson and Houlihan2009).

In summary, the phenotype of cognitive abilities is described by a three-level hierarchy that captures variation in people's mental functioning at different levels of generality. This framework is a good starting point for genetic studies, which may be aimed at different levels. The evidence to date shows that much of the genetic influence is on general ability, and that like human height will have a complex genetic architecture (Lanktree et al. Reference Lanktree, Guo, Murtaza, Glessner, Bailey, Onland-Moret, Lettre, Ongen, Rajagopalan, Johnson, Shen, Nelson, Klopp, Baumert, Padmanabhan, Pankratz, Pankow, Shah, Taylor, Barnard, Peters, Maloney, Lobmeyer, Stanton, Zafarmand, Romaine, Mehta, van Iperen, Gong, Price, Smith, Kim, Li, Asselbergs, Atwood, Bailey, Bhatt, Bauer, Behr, Bhangale, Boer, Boehm, Bradfield, Brown, Braund, Burton, Carty, Chandrupatla, Chen, Connell, Dalgeorgou, Boer, Drenos, Elbers, Fang, Fox, Frackelton, Fuchs, Furlong, Gibson, Gieger, Goel, Grobbee, Hastie, Howard, Huang, Johnson, Li, Kleber, Klein, Klein, Kooperberg, Ky, Lacroix, Lanken, Lathrop, Li, Marshall, Melander, Mentch, Meyer, Monda, Montpetit, Murugesan, Nakayama, Nondahl, Onipinla, Rafelt, Newhouse, Otieno, Patel, Putt, Rodriguez, Safa, Sawyer, Schreiner, Simpson, Sivapalaratnam, Srinivasan, Suver, Swergold, Sweitzer, Thomas, Thorand, Timpson, Tischfield, Tobin, Tomaszewski, Verschuren, Wallace, Winkelmann, Zhang, Zheng, Zhang, Zmuda, Clarke, Balmforth, Danesh, Day, Schork, de Bakker, Delles, Duggan, Hingorani, Hirschhorn, Hofker, Humphries, Kivimaki, Lawlor, Kottke-Marchant, Mega, Mitchell, Morrow, Palmen, Redline, Shields, Shuldiner, Sleiman, Smith, Farrall, Jamshidi, Christiani, Casas, Hall, Doevendans, Christie, Berenson, Murray, Illig, Dorn, Cappola, Boerwinkle, Sever, Rader, Reilly, Caulfield, Talmud, Topol, Engert, Wang, Dominiczak, Hamsten, Curtis, Silverstein, Lange, Sabatine, Trip, Saleheen, Peden, Cruickshanks, Marz, O'Connell, Klungel, Wijmenga, Maitland-van der Zee, Schadt, Johnson, Jarvik, Papanicolaou, Grant, Munroe, North, Samani, Koenig, Gaunt, Anand, van der Schouw, Soranzo, Fitzgerald, Reiner, Hegele, Hakonarson and Keating2011). It makes sense therefore to target this level of cognition with regard to molecular genetic studies, including in current genome-wide association study (GWAS) approaches described later in this review.

Choosing between specific cognitive functions and tasks in cognitive phenomic analysis

Several criteria have been put forward for choosing appropriate cognitive measures as phenotypes relevant to psychiatric studies (Cannon, Reference Cannon2005; Gur et al. Reference Gur, Calkins, Gur, Horan, Nuechterlein, Seidman and Stone2007; Donohoe et al. Reference Donohoe, Goldberg, Corvin, Goldberg and Weinberger2009). The most obvious criterion, perhaps, is whether and to what extent performance on a given task is heritable. Different classes of relatives share more or less genetic material (e.g. monozygotic twins share 100% of genes, dizygotic twins/siblings 50%, and half-siblings 25%), making it possible to estimate the proportion of individual differences in performance in a population at a given time that are due to genetic differences [termed heritability (h²)]. In considering heritability, the degree to which performance on a particular function or task shares genetic variance with the underlying risk of the disease is also a significant consideration (Glahn et al. Reference Glahn, Curran, Winkler, Carless, Kent, Charlesworth, Johnson, Goring, Cole, Dyer, Moses, Olvera, Kochunov, Duggirala, Fox, Almasy and Blangero2012). While availability of twin and population-based disease registry data has confirmed the importance of heritability for many psychiatric disorders, including autism and the psychoses (Cardno et al. Reference Cardno, Marchall, Coid, MacDonald, Ribchester, Davies, Venturi, Jones, Lewis, Sham, Gottesman, Farmer, McGuffin, Reveley and Murray1999; McGuffin et al. Reference McGuffin, Rijsdijk, Andrew, Sham, Katz and Cardno2003), unavailability of twin data for many specific cognitive tests has meant that heritability has been inferred from the familiality of specific cognitive deficits in healthy relatives of patients. Where twin data have been available (Goldberg et al. Reference Goldberg, Ragland, Torrey, Gold, Bigelow and Weinberger1990, Reference Goldberg, Torrey, Gold, Bigelow, Ragland, Taylor and Weinberger1995 and Toulopoulou et al. Reference Toulopoulou, Picchioni, Rijsdijk, Hua-Hall, Ettinger, Sham and Murray2007 for schizophrenia; Bidwell et al. Reference Bidwell, McClernon and Kollins2007 for ADHD), evidence for heritability of cognitive deficits has generally been noted.

Shared genetic variation between cognitive and illness phenotypes is a key concept of the ‘endophenotype’ approach. As originally hypothesized, endophenotypes – measurable components located along the pathway between genotype and disease, such as those derived from cognitive and neuropsychological measures – were suggested to be phenotypically and genetically simpler than the more complex disease syndrome, hence leading to more powerful – and successful – genetic analysis (Gottesman & Gould, Reference Gottesman and Gould2003). The use of cognitive phenotypes as neuropsychiatric endophenotypes has been reviewed extensively (Glahn et al. Reference Glahn, Bearden, Niendam and Escamilla2004; Goldberg & Weinberger, Reference Goldberg and Weinberger2004; Cannon, Reference Cannon2005; Meyer-Lindenber & Weinberger, Reference Meyer-Lindenberg and Weinberger2006; Gur et al. Reference Gur, Calkins, Gur, Horan, Nuechterlein, Seidman and Stone2007; Walters & Owen, Reference Walters and Owen2007; Donohoe et al. Reference Donohoe, Goldberg, Corvin, Goldberg and Weinberger2009; Corvin et al. Reference Corvin, Donohoe, Hargreaves, Gallagher and Gill2012). Several questions about the nature and use of cognitive endophenotypes remain. For example, Gottesman's original hypothesis that cognitive phenotypes would represent genetically ‘simpler’ constructs than genetically complex psychiatric conditions has not generally been supported – cognitive constructs like intelligence appear to be themselves, like other human traits including height and weight, genetically complex (Bilder et al. Reference Bilder, Howe, Novak, Sabb and Parker2011). Furthermore, whether cognitive functions are mediators or moderators of genetic effects on illness remains unclear; psychiatric GWAS studies to date suggest that while the increased risk associated with some common variants is also associated with variation in cognitive function [e.g. calcium channel, voltage-dependent, l type, alpha 1c (CACNA1C); Zhang et al. Reference Zhang, Shen, Xu, Chen, Cheng, Zhai, Gu, Bao, Chen, Wang, Deng, Ji, Liu, Li, Dong and Chen2012] as either a moderator/mediator of genetic risk or as a pleiotropic effect, other common variants appear to affect risk largely independently of cognition [e.g. neurogranin (NRGN) in schizophrenia; Donohoe et al. Reference Donohoe, Walters, Morris, Da Costa, Rose, Hargreaves, Maher, Hayes, Giegling, Hartmann, Möller, Muglia, Moskvina, Owen, O'Donovan, Gill, Corvin and Rujescu2011]. Finally, although high genetic overlap between cognitive deficits and illness has been reported (e.g. overlapping genetic effects of about 92% between intelligence quotient and schizophrenia; Toulopoulou et al. Reference Toulopoulou, Picchioni, Rijsdijk, Hua-Hall, Ettinger, Sham and Murray2007), population-based studies suggest that the actual overlap may be much smaller (Fowler et al. Reference Fowler, Zammit, Owen and Rasmussen2012). Whether because of these factors, or for pragmatic reasons such as availability of much larger diagnostically phenotyped samples than cognitively phenotyped samples, cognitive phenotype studies in psychiatric illness have more typically been used for following up psychiatric GWAS signals than for discovery. Sometimes termed ‘a reverse endophenotyping’ strategy, the approach here has been to use neuropsychological measures to characterize the effects of already GWAS-identified risk variants on cognition (for a recent review, see Corvin et al. Reference Corvin, Donohoe, Hargreaves, Gallagher and Gill2012).

Irrespective of whether cognitive phenotypes are used for the purposes of genetic discovery or characterization of already identified risk variants, a further criterion for genetic studies of cognitive functions and associated measures is that they can be quantitatively measured in a reliable manner. That cognitive phenotypes should have enhanced reliability as compared with that afforded by diagnostic categories is an important assumption of the intermediate phenotype approach (Gottesman & Gould, Reference Gottesman and Gould2003; Bearden & Freimer, Reference Bearden and Freimer2006). Thus, the inter-rater reliability (the consistency of scores across raters) and test–retest reliability (the consistency of scores over time) of cognitive phenotypes have been widely scrutinized. In patient studies, the importance of state independence (that scores are relatively independent of fluctuations in clinical symptoms) and independence from medication effects has been a particular focus in this regard, particularly for schizophrenia. These studies suggest that while cognitive deficits are somewhat correlated with clinical symptoms (for example, negative symptoms in schizophrenia), the correlation is low (r ² < 0.3) and the amount of variance shared by these variables appears to be small. In factor analysis, cognitive function (as measured in terms of memory and attention) often emerges as a separate factor from clinical symptoms (Donohoe & Robertson, Reference Donohoe and Robertson2003; Good et al. Reference Good, Rabinowitz, Whitehorn, Harvey, DeSmedt and Kopala2004; Donohoe et al. Reference Donohoe, Corvin and Robertson2006; Lipkovich et al. Reference Lipkovich, Deberdt, Csernansky, Sabbe, Keefe and Kollack-Walker2009). Furthermore, changes in cognition following trials of either medication or cognitive remediation are only weakly associated with changes in clinical presentation (Davidson et al. Reference Davidson, Galderisi, Weiser, Werbeloff, Fleischhacker, Keefe, Boter, Keet, Prelipceanu, Rybakowski, Libiger, Hummer, Dollfus, López-Ibor, Hranov, Gaebel, Peuskens, Lindefors, Riecher-Rössler and Kahn2009; Wykes et al. Reference Walters and Owen2011). Finally, data derived from large-scale studies of individuals at risk for developing psychosis, either by virtue of prodromal symptomatology (clinical high risk) or genetic predisposition (genetic high risk), suggest that neurocognitive impairment is present regardless of current symptom status and likelihood of later conversion to psychosis (Seidman et al. Reference Seidman, Giuliano, Meyer, Addington, Cadenhead, Cannon, McGlashan, Perkins, Tsuang, Walker, Woods, Bearden, Christensen, Hawkins, Heaton, Keefe, Heinssen and Cornblatt2010). Taken together, these data indicate that cognitive phenotypes are, in addition to being heritable, likely to be stable and generally independent of fluctuations in clinical symptomatology.

The utility of several cognitive functions as endophenotypes have been extensively investigated, and to a reasonable extent supported across illness, including memory function (both episodic and working memory), and various aspects of attentional control and executive function. Evidence of the utility of these individual cognitive phenotypes and measures in relation to specific neuropsychiatric disorders has been reviewed extensively, including Gur et al. (Reference Gur, Calkins, Gur, Horan, Nuechterlein, Seidman and Stone2007) and Donohoe et al. (Reference Donohoe, Goldberg, Corvin, Goldberg and Weinberger2009) for schizophrenia, Glahn et al. (Reference Glahn, Bearden, Niendam and Escamilla2004) for bipolar disorder, MacQueen & Frodl (Reference MacQueen and Frodl2011) for major depressive disorder, Bellgrove & Mattingly (Reference Bellgrove and Mattingley2008) for ADHD, and Abrahams & Geschwind (Reference Abrahams and Geschwind2010) for autism. A particular challenge for utilizing these cognitive constructs in GWAS studies, however, is regarding how best to combine test scores between datasets. The requirement for adequately large samples powered to undertake genome-wide analysis means that data need to be combined across datasets, usually from a large number of research groups. However, because of differences between groups in the cognitive measures employed, combining datasets is extremely challenging. Even where similar cognitive and test constructs have been employed between sites – e.g. using the continuous performance test (CPT) to measure attentional control – different versions of the test (e.g. CPT-AX: Identical Pairs, CPT-IP; Degraded-Stimulus, CPT-DS) usually result in non-identical phenotypes between sites.

Univariate versus multivariate analysis

A third consideration for cognitive genetics approaches is whether to use a multivariate, rather than univariate, approach. The power of a genetic analysis (e.g. linkage or GWA) to detect an effect can be greatly enhanced by collapsing large cognitive batteries into cognitive domains that can be achieved via factor analysis. The advantage of factor analysis is threefold: (1) the maximum amount of cognitive information is captured by relatively few constructs; (2) reducing a large number of variables to smaller number of latent constructs limits the need for conservative multiple comparison correction; and (3) if employing confirmatory factor analysis measurement error can be accounted for. A potential disadvantage of this approach is that if the grouped measures do not share common genetic aetiology it actually reduces the power to find an effect. Thus the development of factor models should be guided by the covariance within the cognitive data and the degree to which that covariance can be explained by shared genetic effects (current efforts to address these issues are described by Bilder et al. Reference Bilder, Howe, Novak, Sabb and Parker2011).

Given evidence that shared genetic factors influence g, cognitive domains and specific cognitive traits, searching for genes influencing g is intrinsically a multivariate analysis. Such an analysis is designed to localize loci associated with each level of the three-level hierarchy. In this context, focusing on g should be more effective than searching for the genetic influences of each of the specific cognitive tests alone, as the index of g is typically more reliable than individual test scores and a single phenotype is assessed (compared with many), reducing the need for correction for multiple comparisons. However, there is an alternative view. It is possible that the g construct is exceptionally polygenetic, with large numbers of genes of very small effect (Lanktree et al. Reference Lanktree, Guo, Murtaza, Glessner, Bailey, Onland-Moret, Lettre, Ongen, Rajagopalan, Johnson, Shen, Nelson, Klopp, Baumert, Padmanabhan, Pankratz, Pankow, Shah, Taylor, Barnard, Peters, Maloney, Lobmeyer, Stanton, Zafarmand, Romaine, Mehta, van Iperen, Gong, Price, Smith, Kim, Li, Asselbergs, Atwood, Bailey, Bhatt, Bauer, Behr, Bhangale, Boer, Boehm, Bradfield, Brown, Braund, Burton, Carty, Chandrupatla, Chen, Connell, Dalgeorgou, Boer, Drenos, Elbers, Fang, Fox, Frackelton, Fuchs, Furlong, Gibson, Gieger, Goel, Grobbee, Hastie, Howard, Huang, Johnson, Li, Kleber, Klein, Klein, Kooperberg, Ky, Lacroix, Lanken, Lathrop, Li, Marshall, Melander, Mentch, Meyer, Monda, Montpetit, Murugesan, Nakayama, Nondahl, Onipinla, Rafelt, Newhouse, Otieno, Patel, Putt, Rodriguez, Safa, Sawyer, Schreiner, Simpson, Sivapalaratnam, Srinivasan, Suver, Swergold, Sweitzer, Thomas, Thorand, Timpson, Tischfield, Tobin, Tomaszewski, Verschuren, Wallace, Winkelmann, Zhang, Zheng, Zhang, Zmuda, Clarke, Balmforth, Danesh, Day, Schork, de Bakker, Delles, Duggan, Hingorani, Hirschhorn, Hofker, Humphries, Kivimaki, Lawlor, Kottke-Marchant, Mega, Mitchell, Morrow, Palmen, Redline, Shields, Shuldiner, Sleiman, Smith, Farrall, Jamshidi, Christiani, Casas, Hall, Doevendans, Christie, Berenson, Murray, Illig, Dorn, Cappola, Boerwinkle, Sever, Rader, Reilly, Caulfield, Talmud, Topol, Engert, Wang, Dominiczak, Hamsten, Curtis, Silverstein, Lange, Sabatine, Trip, Saleheen, Peden, Cruickshanks, Marz, O'Connell, Klungel, Wijmenga, Maitland-van der Zee, Schadt, Johnson, Jarvik, Papanicolaou, Grant, Munroe, North, Samani, Koenig, Gaunt, Anand, van der Schouw, Soranzo, Fitzgerald, Reiner, Hegele, Hakonarson and Keating2011), making it very difficult to identify a single genetic target (gene). In contrast, individual tests may have relatively less complex genetic architectures, affording them the potential for gene discovery. Eventually empirical evidence will determine which of these alternatives is more likely. Unfortunately, support for either hypothesis is largely missing at this time.

GWAS era and challenges of power

An extensive literature now exists on cognitive studies of single genes associated with increased illness risk for neuropsychiatric diseases with childhood onset (e.g. autism, ADHD), and adult onset (e.g. schizophrenia, bipolar disorder, Alzheimer's disease). With few exceptions (e.g. apolipoprotein E), however, identifying a specific role in cognition for individual gene variants across these disorders has been challenging (Payton, Reference Payton2009; Corvin et al. Reference Corvin, Donohoe, Hargreaves, Gallagher and Gill2012). In the majority of cases, individual variants have not been robustly associated with the illness phenotype, and no clear functional variants have been identified, making direct comparison of studies difficult. These difficulties have been further hampered by the differences in the cognitive measures investigated. What is perhaps most striking about this literature is how closely the cognitive phenotype findings recapitulate the illness phenotype findings in term of the small proportion of variation explained (about 1–2%) and the frequent lack of replication. This, together with evidence of a significant negative correlation between observed effect and sample size (Rose & Donohoe, Reference Rose and Donohoe2012), highlights the necessity of incorporating large samples in order to optimally detect these expected small effects.

Recent (and relatively affordable) technological advances in genotyping platforms have resulted in a move from single-gene studies towards genome-wide association studies using platforms that can assay more than 1 million genomic markers. This GWAS approach has allowed for a relatively hypothesis-free scan of the entire human genome, thereby markedly increasing the genetic information. GWAS platforms offer optimal coverage of common variation for single nucleotide polymorphism (SNP)-based association analyses and simultaneously capture structural variation, such as copy-number variation through the use of intensity data analyses without requiring the implementation of a secondary technology. Imputation, or prediction, strategies capitalize on what is known about the correlation among SNPs and provide meaningful estimates of genomic variation that is not directly genotyped on GWAS platforms and can further improve resolution. These advantages notwithstanding, GWAS methodology has certain challenges, especially for cognitive phenomics.

The use of a genome-wide, hypothesis-free strategy implies the need to handle large quantities of data while attempting to maintain adequate statistical power. While the standard in the field of research for statistical significance has long been set at a p value <0.05, the use of hundreds of thousands, or millions, of statistical tests in GWAS analyses has necessitated multiple-testing corrections that require conservative α values: for example, a p value threshold in and around 5 × 10⁻⁸ derived from a Bonferroni correction for about 550 000 observations is often used, although the p value will be smaller for >1 million SNPs. This, together with the likelihood of individual common variants explaining only a modest proportion of variation in cognition, highlights the need for extremely large sample sizes (Manolio, Reference Manolio2010).

In an effort to surmount this problem, several large-scale consortia have been organized across numerous medical and neuropsychiatric disorders, which represent collaborative efforts to merge independent datasets. In addition to increasing sample sizes, optimal power can be achieved through the use of information in the public domain, which can provide for targeted selection of the smallest number of SNPs required to tag common variation across the genome through the use of linkage disequilibrium (LD) data in reference samples.

There have been several genome-wide association studies conducted that directly targeted normal human cognition, each of which highlight the importance of large sample sizes and replication. The first GWAS study of cognition (Papassotiropoulos et al. Reference Papassotiropoulos, Stephan, Huentelman, Hoerndli, Craig, Pearson, Huynh, Brunner, Corneveaux, Osborne, Wollmer, Aerni, Coluccia, Hanggi, Mondadori, Buchman, Reiman, Caselli, Henke and de Quervain2006) focused on episodic memory in 351 young adults from Switzerland and reported a significant effect of the gene, KIBRA (kidney- and brain-expressed protein), on free recall performance 5 min and 24 h after initial word presentation. This result was replicated in a second independent cohort of subjects from the USA. Subsequent cognitive GWAS studies have also suggested that KIBRA could influence aspects of cognitive function; Need et al. (Reference Need, Attix, McEvoy, Cirulli, Linney, Hunt, Ge, Heinzen, Maia, Shianna, Weale, Cherkas, Clement, Spector, Gibson and Goldstein2009) reported that in a study of over 1000 subjects recruited from college campuses, 10 genes achieved nominal significance for association with specific aspects of cognition, including a SNP (but not the same SNP as originally reported by Papassotiropoulos et al. Reference Papassotiropoulos, Stephan, Huentelman, Hoerndli, Craig, Pearson, Huynh, Brunner, Corneveaux, Osborne, Wollmer, Aerni, Coluccia, Hanggi, Mondadori, Buchman, Reiman, Caselli, Henke and de Quervain2006) in KIBRA with verbal learning and memory as assessed with the Cambridge Neuropsychological Test Automated Battery. Other cognitive GWAS studies (e.g. Seshadri et al. Reference Seshadri, DeStefano, Au, Massaro, Beiser, Kelly-Hayes, Kase, D'Agostino, Decarli, Atwood and Wolf2007; Butcher et al. Reference Butcher, Davis, Craig and Plomin2008) have been unable to observe these effects, however, though substantive differences in both the phenotypic and genotypic assessment strategies hamper the interpretation of these data.

Luciano et al. (Reference Luciano, Hansell, Lahti, Davies, Medland, Räikkönen, Tenesa, Widen, McGhee, Palotie, Liewald, Porteous, Starr, Montgomery, Martin, Eriksson, Wright and Deary2011) describe the results of a collaborative meta-analysis of data derived from three cohorts with a total sample size of 2379 individuals. The primary outcome measure of this study was focused on the domain of processing speed, a lower-level cognitive construct that shares a large proportion of genetic variance with higher-order processes (Luciano et al. Reference Luciano, Wright, Geffen, Geffen, Smith and Martin2004). The primary results indicated no individual marker (SNP) that reached genome-wide statistical significance. More recently, we undertook a GWAS of general intelligence in a sample of 3511 healthy individuals (Davies et al. Reference Davies, Tenesa, Payton, Yang, Harris, Liewald, Ke, Le Hellard, Christoforou, Luciano, McGhee, Lopez, Gow, Corley, Redmond, Fox, Haggarty, Whalley, McNeill, Goddard, Espeseth, Lundervold, Reinvang, Pickles, Steen, Ollier, Porteous, Horan, Starr, Pendleton, Visscher and Deary2011). General intelligence in this study was separated into crystallized intelligence and fluid intelligence, based on a PCA of the overlapping (but non-identical) measures available in the five cohorts. Genome-wide analyses of SNP data in this study indicated that genetic variants in LD with common SNPs account for 40–50% of the variation in general intelligence. We furthermore observed, using gene-based analysis, a genome-wide significant association between general fluid-type intelligence and variation in the formin-binding protein 1-like (FNBP1L) gene. Taken together, these cognitive GWAS studies highlight the expectation that small effect sizes for individual loci will be the norm, reinforce the advantages of utilization of comparable genotyping platforms, and emphasize the need for careful consideration of the precise cognitive phenotype for examination.