Early cognitive theories posited that cognitive functions are modular in nature and located within separable but interconnected parts of the brain (Luria, 1970; Shallice, 1988). Within this framework, executive functions have been described as a set of superordinate processes that guide thought and behavior and allow purposive action toward a goal (Miller, 2000). These functions are critical for normal day-to-day cognitive functioning and appear to be particularly susceptible to altered development, injury, and disease. From a traditional cognitive or neuropsychological perspective, executive functions have been thought to comprise a set of distinct cognitive domains that include vigilance, or sustained attention (Pennington & Ozonoff, 1996; Smith & Jonides, 1999); initiation of complex goal-directed behaviors (Lezak, 1995); inhibition of prepotent but incorrect responses (Luna, Padmanabhan, & O’Hearn, 2010; Smith & Jonides, 1999); flexibility to shift easily between goal states (Ravizza & Carter, 2008); planning the necessary steps to achieve a goal (Smith & Jonides, 1999); and working memory, the ability to hold information in mind and manipulate it to guide response selection (Goldman-Rakic, 1996).

These theoretically distinct domains are supported by discrete neural systems (Luria, 1970; Shallice, 1988), which typically include elements of the prefrontal cortex (PFC). Early animal lesion studies provided evidence for PFC involvement in the coordination of complex behaviors, by serving as a temporary store for incoming information, making this information immediately available to guide response selection (Fuster, 1990; Goldman-Rakic, 1987; Jacobsen, 1936). Prefrontal damage in humans also impairs various executive functions, including planning (Owen, Downes, Sahakian, Polkey, & Robbins, 1990; Shallice, 1982, 1988), flexibility (Milner, 1982), response inhibition (Leimkuhler & Mesulam, 1985), and working memory (Milner, 1982).

Early neuroimaging and human lesion studies revealed that the frontal cortex is just one element in a network of spatially distinct regions associated with executive functions (Baddeley & Wilson, 1988). For example, neuroimaging studies of a prototypical working memory task, the n-back paradigm, have consistently shown activated regions in the frontal and posterior parietal cortex and cerebellum (Owen, McMillan, Laird, & Bullmore, 2005). Within this task, Broca’s area and premotor cortex have been associated with subvocal rehearsal processes, while posterior parietal areas were associated with the storage of verbal information (Awh et al., 1996). On tasks that require flexibility, the ability to flexibly switch attention and behavioral responses between different rules is associated with activation of dorsolateral PFC (DLPFC), while switching attention responses between different perceptual features of a stimulus is associated with parietal activation (Ravizza & Carter, 2008).

While traditional theories of executive functions have posited a set of distinct domains supported by at least partially unique brain regions, increasing numbers of functional neuroimaging studies examining diverse executive functions have suggested that these tasks may engage very similar brain networks (e.g., Duncan & Owen, 2000). Recent views of the PFC highlight its role in higher cognitive functions by supporting coordinated activation of multiple brain areas within the “cognitive control network,” including the DLPFC, medial frontal cortex (including the anterior cingulate cortex [ACC]), parietal cortex, motor areas, and cerebellum (Bellebaum & Daum, 2007; Braver, Cohen, & Barch, 2002; D’Esposito, 2007; Fuster, 2002). Furthermore, analyses of functional connectivity in healthy adults revealed that coordinated temporal activation across the network of prefrontal and posterior brain regions is associated with better performance on cognitive control tasks (Fornito, Yoon, Zalesky, Bullmore, & Carter, 2011; Yoon et al., 2008). Miller and Cohen proposed that the PFC supports “cognitive control” by actively maintaining “rules” online in order to evaluate incoming information, as well as internal states to guide response selection toward a current goal (Miller, 2000; Miller & Cohen, 2001). According to this view, cognitive control mechanisms support the range of executive functions, including working memory, selective attention, stimulus–response mapping, and performance monitoring (Carter et al., 1998; Cohen, Dunbar, & McClelland, 1990; Miyake & Shah, 1999; Shallice, 1988), and are not restricted to a particular cognitive domain (Banich, 1997; Smith & Jonides, 1999).

The diverse array of executive functions has limited our ability to directly test the unitary or modular nature of the underlying brain systems within a single set of experiments. Capitalizing on the unique power of activation likelihood estimation (ALE) meta-analytic tools, this study is the first to synthesize almost 200 published reports, testing the hypothesis that traditional executive functions are supported by a common PFC-related cognitive control network. The ALE meta-analytic approach models three-dimensional coordinates (from reported activations in standard space) as the center of a three-dimensional Gaussian distribution (Laird, Fox, et al., 2005). By combining published data from a wide variety of studies, the ALE method provides the unique opportunity to examine this question in the largest sample of control subjects published to date. Activation likelihood estimation has been has been used to address similar research questions in both healthy and patient samples (Binder, Desai, Graves, & Conant, 2009; Caspers, Zilles, Laird, & Eickhoff, 2010; Chouinard & Goodale, 2010; Dickstein, Bannon, Castellanos, & Milham, 2006; Fusar-Poli et al., 2009; Glahn et al., 2005; Goghari, 2010; Mana, Paillere Martinot, & Martinot, 2010; Minzenberg, Laird, Thelen, Carter, & Glahn, 2009; Molenberghs, Cunnington, & Mattingley, 2009; Owen et al., 2005; Ragland et al., 2009; Richlan, Kronbichler, & Wimmer, 2009; Samson, Mottron, Soulieres, & Zeffiro, 2011; Schwindt & Black, 2009; Spaniol et al., 2009; Turkeltaub & Coslett, 2010; Yu et al., 2010). We hypothesized that healthy adults would show a common pattern of activation across prefrontal (DLPFC, ACC) and parietal regions when performing executive function tasks across multiple domains (see Table 1). Furthermore, we hypothesized that additional areas of domain-specific activation may be observed, but these would occur in addition to the common pattern of activation within the cognitive control network.

Table 1 Definitions of the cognitive domains examined within this study, tasks included within each of the domains, the total numbers of available studies examined, and the total numbers of studies and subjects included in the present analysis, by domain and task

Method

Study selection

A search of the BrainMap database (Fox & Lancaster, 2002; Laird, Fox, et al., 2005) was performed to identify all English-language, peer-reviewed studies that investigated executive function tasks in multiple healthy individuals, ages 18–60 years, using functional magnetic resonance imaging (fMRI) or positron emission tomography (PET). Executive functions were defined as processes that are required in order to regulate or guide other cognitive processes in order to support goal-directed behavior (Minzenberg et al., 2009). For the purpose of this investigation, we examined studies that used task paradigms that are typically considered measures of executive function or cognitive control. As outlined in Table 1, these included measures of vigilance, inhibition, flexibility, planning, working memory, and initiation. Within each study, we included data from healthy individuals on specific contrasts that examined within-group whole-brain activation in response to a task of interest that was compared to an active control task, rather than to rest or fixation. Studies were excluded if the subject pool overlapped with other published studies on smaller subsets of the same sample or included subjects outside of the age range (18–60 years), if the task of interest did not require an appropriate behavioral response (e.g., a buttonpress), or if contrasts with the available coordinate data did not examine a specific executive function or rather examined differences between patients and controls. Table 1 provides the numbers of studies that were available and that met the criteria for inclusion within each domain. The BrainMap database archives the peak coordinates of activations as well as their corresponding metadata, such as the number and diagnosis of the subjects, the analysis technique, the paradigm, and the cognitive domain. Coordinates originally published in MNI space were converted to Talairach space using the Lancaster (icbm2tal) transformation (Laird et al., 2010; Lancaster et al., 2007). Further filtering and meta-analysis of the experiments was carried out using BrainMap’s software applications (Laird et al., 2009), as described below.

Activation likelihood estimation

We performed a series of coordinate-based meta-analyses of executive functioning using the ALE method (Laird, McMillan, et al., 2005; Turkeltaub, Eden, Jones, & Zeffiro, 2002), in which the voxel-wise correspondence of neuroimaging results is assessed across a large number of studies. The ALE algorithm aims to identify areas showing a higher convergence of findings across experiments than would be expected under a spatially random spatial association. The identified literature coordinates were modeled with a three-dimensional Gaussian probability distribution reflecting the spatial uncertainty of each focus on the basis of an estimation of the intersubject and interlaboratory variability typically observed in neuroimaging experiments. This algorithm limits the meta-analysis to an anatomically constrained space specified by a gray-matter mask and includes a method that calculates the above-chance clustering between experiments (i.e., random-effects analysis), rather than between foci (i.e., fixed-effects analysis), and it also accounts for differences in sample sizes across the included studies (Eickhoff et al., 2009). The probabilities of all foci reported in a given experiment were combined, resulting in a modeled activation map for each experiment, and the union of these probabilities was computed in order to derive voxel-wise ALE values that described the convergence of results across the whole brain. To determine which ALE values were statistically significant, ALE scores were compared with an empirical null distribution reflecting a random spatial association between experiments, thereby estimating convergence between studies rather than the clustering of foci within a particular study.

ALE was performed in Talairach space using GingerALE 2.0 (http://brainmap.org/ale/index.html) to analyze the global set of activation foci for concordance, as well as subsets of foci that corresponded to the cognitive components of interest within executive function. From the set of included studies (Table 2), the results for a global set of within-group activations across all six domains were meta-analyzed to address the primary hypothesis. To examine the foci of greatest concordance across studies, we also performed a conjunction analysis across the three domains in which the data from more than nine studies were available (flexibility, inhibition, and working memory). To examine potential domain-specific patterns of activation, we completed within-group meta-analyses for the domains in which data from more than nine studies were available. The resultant ALE maps were thresholded at a false-discovery rate (FDR)-corrected threshold of p < .05. Images were viewed in Mango (“multi-image analysis GUI”), developed at the Research Imaging Institute in San Antonio (http://ric.uthscsa.edu/mango/).

Table 2 Published studies included in the ALE meta-analysis of executive functions, by domain

Results

Global analysis across all domains

Across all domains (shown in red in Fig. 1; see also Table 3a), large clusters of significant activation were observed within lateral and medial PFC bilaterally, encompassing superior, middle, and inferior frontal gyri including the DLPFC (Brodmann areas [BAs] 9, 46), as well as the ACC (BA 32) on the medial wall. In addition to prefrontal activation, the overall contrast revealed large parietal clusters, including the inferior (BA 40) and superior (BA 7) parietal lobe. This combined frontal–parietal activation is consistent with previous findings related to the cognitive control circuit (Botvinick, Braver, Barch, Carter, & Cohen, 2001; Carter, Botvinick, & Cohen, 1999; Cohen, Botvinick, & Carter, 2000; Yarkoni et al., 2005). Additional activation in frontal regions included the premotor cortex (BA 6), frontopolar cortex (BA 10), and orbitofrontal cortex (BA 11). Activation was also observed in occipital (BA 19) and temporal (BAs 13, 22, 37) regions, which are consistent with processing of the verbal and auditory stimuli, respectively, that are presented as part of the included tasks. Finally, significant activation was found in subcortical structures, including the thalamus, caudate, and putamen, as well as areas of the cerebellum, including the posterior declive and anterior culmen. These findings are consistent with the hypothesis that executive functions are supported by a common set of cortical and subcortical regions within the cognitive control network.

Fig. 1
figure 1

Global analysis of executive function in 193 studies of healthy adults, showing brain regions with significant activation across all executive function domains (red) and the areas of conjunction (green) across the three domains for which data from more than nine studies were available (flexibility, inhibition, and working memory).

Table 3 Brain regions (Brodmann areas in parentheses) with significant activation within healthy adults from (a) a combined meta-analysis across all six executive function domains and (b) a conjunction meta-analysis for domains with more than nine included studies (flexibility, inhibition, and working memory)

Results of the conjunction analysis (shown in green in Fig. 1; see also Table 3b) across the three domains for which the data from more than nine studies were available (flexibility, inhibition, and working memory) revealed similar patterns of common activation in cognitive-control-related frontal and parietal regions, including the DLPFC (BAs 9, 46), anterior cingulate (BA 32), inferior (BAs 39, 40) and superior (BA 7) parietal lobe, and precuneus (BA 19). The results of these analyses can be examined through an interactive viewer at http://carterlab.ucdavis.edu/research/ale_analysis.php.

Domain-specific within-group analysis

Flexibility

For tasks that examined flexibility, similar patterns of activation were observed in frontal and parietal regions supporting the cognitive control network (see Fig. 2 and Table 4), including the DLPFC (BAs 9, 46), cingulate (BAs 32, 24), as well as superior (BA 7) and inferior (BA 40) parietal lobe. Activation was also observed in additional prefrontal (BAs 6, 10, 11), occipital (BA 19), and temporal (BAs 13, 37) regions.

Fig. 2
figure 2

Domain-specific analysis showing patterns of common and distinct activation across the working memory (red; 78 studies), inhibition (green; 79 studies), flexibility (blue; 21 studies), and initiation (yellow; 9 studies) domains

Table 4 Brain regions (Brodmann areas in parentheses) with significant activation within healthy adults for tasks within the flexibility domain

Inhibition

As is shown in Fig. 2 (see Table 5), tasks that require inhibition were associated with activation in frontal and parietal cognitive-control-related regions, including DLPFC (BAs 9, 46), ACC (BA 32), and superior (BA 7) and inferior (BA 40) parietal lobe. Such tasks also elicited activation in other prefrontal (BAs 6, 10), occipital (BA 19), and temporal (BA 13) regions. Activation of subcortical regions included the caudate, thalamus, putamen, and cerebellar declive.

Table 5 Brain regions (Brodmann areas in parentheses) with significant activation within healthy adults for tasks within the inhibition domain

Working memory

Working memory tasks elicited the common pattern of frontal–parietal activation associated with the cognitive control network (see Fig. 2 and Table 6), including the DLPFC (BAs 9, 46), cingulate (BAs 32, 24), and parietal lobe (BAs 7, 40). A consistent pattern of activation was also observed in prefrontal (BAs 6, 10), occipital (BA 19), temporal (BAs 13, 37), and subcortical (thalamus, caudate, putamen, cerebellar declive) regions.

Table 6 Brain regions (Brodmann areas in parentheses) with significant activation within healthy adults within the working memory

Other domains

Domain-specific analyses for the planning and vigilance domains were not possible, due to the small number of studies available for inclusion within the ALE analysis (four and two studies, respectively). Although the number of studies for the initiation domain was also small (n = 9), the results are presented here as a preliminary analysis of site-specific activation within this domain. In contrast to the pattern of frontal–parietal activation observed in the other three domains, initiation tasks were associated with a pattern of activation primarily in frontal regions, including the DLPFC (BA 46), middle (BA 10) and inferior (BA 47) frontal, anterior cingulate (BA 32), and motor (BA 6) regions, with no observed activation in parietal regions (see Fig. 2, Table 7). Activation was also observed in the superior (BA 21) and middle (BA 22) temporal, occipital (BA 17), and subcortical (putamen, caudate, cerebellar declive and culmen) regions, in a manner similar to other executive domains.

Table 7 Preliminary data for brain regions (Brodmann areas in parentheses) with significant activation within healthy adults for tasks within the initiation domain

Discussion

Using a meta-analytic approach, we examined 193 neuroimaging studies of tasks divided according to classic executive function domains, creating the largest sample of healthy adults to date. We sought to provide evidence that discrete executive functions (initiation, inhibition, working memory, flexibility, planning, and vigilance) are supported by a shared, superordinate network that has been previously associated with cognitive control. Results of the combined analysis across domains showed that executive functions are indeed associated with increased activity in this common cognitive control network (Bellebaum & Daum, 2007; Botvinick et al., 2001; Carter et al., 1999; Cohen et al., 2000; D’Esposito & Postle, 2002; Yarkoni et al., 2005), which includes the DLPFC (BAs 9, 46), frontopolar cortex (BA 10), orbitofrontal cortex (BA 11), and anterior cingulate (BA 32). Additional concurrent regions of activation included the superior and inferior parietal (BAs 7, 40), occipital (BA 19), and temporal (BAs 13, 22, 37) cortex, as well as subcortical areas including the caudate, putamen, thalamus, and cerebellum. These conclusions were further supported by a conjunction analysis across the three domains in which data from more than nine studies were available (flexibility, inhibition, and working memory), which revealed a similar pattern of common activation in cognitive-control-related frontal and parietal regions. Although the present analysis did not directly examine the functional connectivity of these brain regions during each task, previous studies of cognitive control (Fornito et al., 2011; Yoon et al., 2008) have consistently shown task-related increases in functional connectivity between the DLPFC and the network of brain regions shown here.

These results provide additional evidence that a superordinate cognitive control network supports executive functions across a range of “domains” previously considered to be distinct, including flexibility, working memory, initiation, and inhibition. As proposed by Miller and Cohen (2001), it has been common to stress the distributed nature of the network that supports cognitive control functions, as well as the unique functional contributions by specific regions within the network. Within this framework, elements of the network may be differentially engaged, depending on the task demands. For example, previous studies (Spreng, Stevens, Chamberlain, Gilmore, & Schacter, 2010; Vincent, Kahn, Snyder, Raichle, & Buckner, 2008) have shown that the frontoparietal control network is engaged across multiple goal-directed activities, flexibly engaging the default-mode network to support autobiographical planning, or engaging the dorsal attention network to support visual spatial planning. Similarly, demands for specific goal- or task-context-related activity may be associated with stronger engagement of the PFC, and demands for maintaining information over longer periods of time may lead to more sustained network activity (Dosenbach et al., 2006; Yarkoni, Barch, Gray, Conturo, & Braver, 2009). Its connectivity with sensory and motor regions, including the cerebellum, allows the DLPFC to play a central role in the maintenance of the rules for action, as well as response selection and inhibition (Asaad, Rainer, & Miller, 2000; Bellebaum & Daum, 2007; Watanabe, 1990, 1992). The ACC and related medial frontal regions are considered to support cognitive control by detecting conditions, such as processing conflicts, that indicate the demand for control, which then leads to the engagement of the DLPFC (Egner & Hirsch, 2005; Kerns et al., 2005; MacDonald, Cohen, Stenger, & Carter, 2000). Furthermore, parietal activation is considered to provide the DLPFC with information on stimulus salience and learned stimulus–response pairings, while the DLPFC is thought to support its ability to shift attentional focus according to the demands of the task at hand (Bunge, Hazeltine, Scanlon, Rosen, & Gabrieli, 2002; Bunge, Kahn, Wallis, Miller, & Wagner, 2003; Miller & Cohen, 2001; Posner & Petersen, 1990).

Within the cognitive control network, it is likely that network-level subdivisions also exist and may be differentially engaged in the same manner. For example, Dosenbach, Fair, Cohen, Schlaggar, and Petersen (2008) proposed discrete circuits within this broader network that support task-sustained versus transient aspects of control, and that these networks may be differentially engaged across different forms of executive functions. Similarly, Braver, Paxton, Locke, and Barch (2009) emphasized that cognitive control has proactive and reactive elements. Proactive control may also depend more on sustained activity in the cognitive control network and, to the degree that these systems may be segregated, they may be differentially engaged during executive functions. A study of the degree to which systematic differences exist in the engagement of discrete elements (regions or subnetworks) of the cognitive control networks across different executive function domains is beyond the resolution of this meta-analysis, and our understanding of this issue will be informed by future experimental studies, particularly those that include direct measures of intraregion connectivity or network dynamics across task demands.

While the use of quantitative meta-analytic methods allowed us to examine executive functions across a variety of tasks and domains within the largest sample of healthy adults to date, it is important to recognize that these findings are limited by the quality of the data available in the extant literature. Activation likelihood estimation requires the reporting of imaging data in three-dimensional coordinates in a standard brain space. Therefore, this analysis did not include studies in which such data were not reported for relevant contrasts (e.g., a within-subjects contrast related to the primary effect of interest in healthy controls), analyses that focused on particular regions of interest, or studies that reported negative findings, as the ALE method does not allow for the modeling of null results (Li, Chan, McAlonan, & Gong, 2010). Furthermore, the lack of appropriate contrasts, such as contrasting an active task with rest or fixation, reduced the number of studies available for inclusion within each domain. However, the use of an active control condition is essential in order to isolate the cognitive process of interest in subtraction contrasts (Stark & Squire, 2001). Our approach to this analysis integrated findings from both fMRI and O15 PET studies, and the ALE method does not account for the potential influence of the different physiological signals associated with these two methods. Additionally, this method does not account for differences in behavioral performance across tasks or the influence of demographic factors, although the sample was restricted to studies that examined a specific age range (18–60 years). While all available studies within the BrainMap database were considered for this analysis, studies that were not included in the database at the time of the analysis have been omitted. Furthermore, this meta-analytic method does not allow for the weighting of results on the basis of levels of statistical significance or the numbers of activation foci that may have been reported by some studies within this investigation (Li, Chan, McAlonan, & Gong, 2010). Although Gaussian blurring of the coordinates will have tended to remove per-study bias of the peak activation localizations, noise within the data might have influenced the study results (M. G. Berman et al., 2010). Finally, our definition of executive functions was based on a traditional view that is often used in cognitive or neuropsychological research (Lezak, 1995; Luria, 1970; Shallice, 1988), and the use of other definitions might have altered the domains examined.

In conclusion, the present study used the meta-analysis of a very large number of published fMRI data sets to examine whether traditional taxonomies of executive functions purporting discrete modular cognitive domains are supported by a superordinate cognitive control system that is engaged during the performance of a range of executive function tasks. Our results suggest that a frontal–cingulate–parietal–subcortical cognitive control network is consistently recruited across a range of traditional executive function tasks. Further research investigating the contributions of modular (e.g., prefrontal) versus shared elements (e.g., frontal–parietal connectivity) of the cognitive control network will inform our understanding of common and unique patterns of impairment in traditional executive functions that are often associated with various brain disorders. Novel approaches to investigating the function of different component systems using single methodologies (e.g., resting state; Deshpande, Santhanam, & Hu, 2011) or combined methodologies (e.g., EEG and fMRI; Debener et al., 2005) have the potential to elucidated the complex brain dynamics underlying cognitive control. Further studies will be needed to make explicit the precise functional contributions of each individual element of the cognitive control network, as well as to understand the complex interactions between network nodes to support coordinated, goal-directed behavior. Through increased understanding of the function of modular components within this network, along with their anatomical connections and functional interactions, we will be able to more effectively investigate the mechanisms by which aberrant behavior or clinical symptoms may result from dysfunction in individual regions or in their connectivity within the broader network (Menon, 2011). Additional research on the relationship between various imaging modalities (e.g., resting state, task-related fMRI, or diffusion tensor imaging) will also help us to uncover ways in which discrete brain systems interact to support complex cognition and behavior.