ALE meta-analysis of action observation and imitation in the human brain

Data used for the meta-analysis

Functional imaging studies included in the meta-analysis were obtained from the BrainMap database (www.brainmap.org; Fox and Lancaster, 2002, Laird et al., 2005b) and a PubMed literature search (www.pubmed.org, search strings: “mirror neurons”, “imitation”, and “action observation”) on functional magnetic resonance imaging (fMRI) and positron emission tomography (PET) experiments. The literature cited in the obtained papers was also assessed to identify additional neuroimaging studies dealing with action observation or imitation processing. Only studies that reported results of whole-brain group analyses as coordinates in a standard reference space (Talairach/Tournoux, MNI) were analysed, while single-subject reports were excluded. Based on these criteria, 87 articles (reporting 83 fMRI and 4 PET studies) were designated as suitable for meta-analysis. Together, these studies included data from 1289 subjects and reported 139 experiments with 1932 activation foci (Table 1).

The reported tasks were subsumed into two main categories: “action observation” and “action imitation”: 104 experiments reported action observation tasks (1061 subjects, 1390 activation foci), and 35 reported imitation tasks (459 subjects, 542 activation foci). Action observation comprised those experiments in which subjects were instructed to observe the action performed by others without performing their own motor act. In this first analysis, the general action observation brain network was assessed. There are, however, several possible confounds that may influence the analysis across the whole sample of observation and imitation experiments, like effectors, instructions or the involvement of an object. To explore the effects of these potential confounds, we subdivided the studies into several subgroups. These were then analysed separately to reveal the neural correlates of different forms of action observation and compared among each other by contrast and conjunction analyses: observation of hand actions (‘right hand’ (37 experiments), ‘left hand’ (2 experiments), ‘both hands’, or ‘hand not specified’ (23 experiments)), observation of right hand actions, observation of face actions, observation of non-hand actions (either ‘face’, ‘body’, or ‘leg/foot’), observation of object-related hand actions, and observation of non-object-related hand actions. A further analysis was performed within those areas which were found to be consistently active for observation of hand actions: observation of hand actions with instruction ‘passively observe’, and observation of hand actions with instruction ‘observe to imitate’ (Table 2).

Action imitation comprised all those tasks in which subjects were asked to imitate actions performed by a visual model as exactly as possible. As for the action observation category, general effects associated with action imitation were analyzed first. Then, subgroups of the imitation studies were analysed separately for imitation of hand actions (either ‘right hand’ (15 experiments), ‘left hand’ (2 experiments), ‘both hands’, or ‘hand not specified’ (11 experiments)), imitation of right hand actions, and imitation of non-object-related hand actions (Table 2). A subgroup of studies on imitation of object-related hand actions could not been analysed due to an insufficient sample size.

Differences in coordinate spaces (MNI vs. Talairach space) were accounted for by transforming coordinates reported in Talairach space into MNI coordinates using a linear transformation (Lancaster et al., 2007).

Meta-analysis algorithm

Meta-analysis was carried out using the revised version (Eickhoff et al., 2009) of the activation likelihood estimation (ALE) approach for coordinate-based meta-analysis of neuroimaging results (Turkeltaub et al., 2002; Laird et al., 2005a,b). The algorithm aims at identifying areas showing a convergence of activations across different experiments, and determining if the clustering is higher than expected under the null distribution of a random spatial association between the results obtained in the experiments. The key idea behind ALE is to treat the reported foci not as single points, but rather as centers for 3D Gaussian probability distributions capturing the spatial uncertainty associated with each focus. The width of these uncertainty functions was determined based on empirical data on the between-subject and between-template variance, which represent the main components of this uncertainty. Importantly, the applied algorithm weights the between-subject variance by the number of examined subjects per study, accommodating the notion that larger sample sizes should provide more reliable approximations of the ‘true’ activation effect and should therefore be modelled by ‘smaller’ Gaussian distributions (Eickhoff et al., 2009).

The probabilities of all activation foci in a given experiment were combined for each voxel, resulting in a modelled activation map (MA map). Taking the union across these MA maps yields voxel-wise ALE scores describing the convergence of results at each particular location. Since neurophysiologically, activation should predominantly be localized within the grey matter, all analyses were restricted to those voxels where a probability of at least 10% for grey matter could be assumed based on the ICBM tissue probability maps (Evans et al., 1994).

To distinguish ‘true’ convergence between studies from random convergence, i.e., noise, the ALE scores were compared to an empirical null distribution derived from a permutation procedure. This null distribution reflects a random spatial association between experiments, while regarding the within-experiment distribution of foci as fixed. Thus, a random-effects inference is invoked, focussing inference on the above-chance convergence between different experiments, not the clustering of foci within a particular experiment. Computationally, deriving this null hypothesis involved sampling a voxel at random from each of the MA maps and taking the union of these values. The ALE score obtained under this assumption of spatial independence was recorded and the permutation procedure iterated 10¹¹ times to obtain a sufficient sample of the ALE null distribution. The ‘true’ ALE scores were tested against the ALE scores obtained under the null distribution and thresholded at a cluster-level corrected threshold of p<0.05 for each separate meta-analysis performed.

Conjunction analysis was carried out to determine the intersection between the meta-analyses on observation and imitation. Results are reported for a corrected p-value of <0.05. Contrast analyses were calculated by means of ALE subtraction analysis, accounting for potential differences in sample size. To increase the specificity of the results, the analysis of differences was restricted to those voxels that showed an effect in main action observation or imitation meta-analyses. The reported contrasts were also thresholded at a corrected p-value of <0.05.

The resulting areas were anatomically labeled by reference to probabilistic cytoarchitectonic maps of the human brain using the SPM Anatomy Toolbox (Eickhoff et al., 2005, 2007). Using a Maximum Probability Map (MPM), activations were assigned to the most probable histological area at their respective locations. Previous studies have provided details about the cytoarchitecture, intersubject variability, and borders of the areas implicated in the current analysis that can be found in the following publications, such as Broca's region (BA 44, BA 45: Amunts et al., 1999), inferior parietal areas (PFop, PFt, PFcm, PF: Caspers et al., 2006, 2008), primary motor cortex (4a; Geyer et al., 1996), premotor cortex (BA 6; Geyer, 2004), primary somatosensory areas (BA 2: Grefkes et al., 2001; BA 1: Geyer et al., 1999, 2000), secondary somatosensory area OP1 (Eickhoff et al., 2006a,b), visual area V5 (Malikovic et al., 2007), superior parietal area 7A, intraparietal area hIP3 (Scheperjans et al., 2008a,b), and intraparietal area hIP1 (Choi et al., 2006).

Individual meta-analyses of action observation and imitation networks

Conjunction and contrast analyses

Methodological considerations

The results of any given neuroimaging experiment are influenced by various study-specific idiosyncrasies, including the experimental design, implementation of the paradigm, task requirements, included subjects and the analysis of the data. Hence, the results of any particular experiment can rarely yield generalisable inference on the cortical substrates of a particular cognitive process. One way to overcome these drawbacks is to integrate the results from several neuroimaging studies by means of quantitative meta-analyses (Eickhoff et al., 2009; Laird et al., 2009; Turkeltaub et al., 2002). Hereby, inference is directed towards identifying those regions where previous experiments showed converging evidence for activation. Significant results in a meta-analysis are achieved if convergence across studies occurs more likely than expected by chance, even though this does not require all or even the majority of the studies to activate at that particular location. Using the revised version of the ALE meta-analysis algorithm (Eickhoff et al., 2009) provided objective modelling of spatial uncertainty relative to sample sizes within different studies and testing for convergence across different experiments. Therefore, possible drawbacks of former coordinate-based meta-analysis approaches (Laird et al., 2005a,b; Turkeltaub et al., 2002) were avoided. Nevertheless, differences in sample size between different meta-analyses (e.g., on action observation and action imitation) may influence the obtained results, in particular with respect to the size of the significant clusters. Furthermore, it has to be noted that meta-analyses on the basis of ALE algorithms only reveal a consistency of activations across studies. Information about strength of a resulting activation cluster is not considered, as these are reported inconsistently and by incompatible measures in the original publications (e.g., percent signal change vs. contrast estimates vs. Z-scores). Consequently, a task which evokes stronger activation in any particular experiment than another may result in less significant and/or extended activation on meta-analyses, if the convergence between studies is less pronounced.

Also, coordinate-based meta-analyses only use reported peak activations for the analysis, thus discarding a large amount of spatial information from the original statistical parametric images. To address this problem, image-based meta-analyses have been proposed, which use the full statistic images of the experiments (e.g., Schilbach et al., 2008b). While such approaches use more information from the original data, their applicability is quite limited since they require comparable contrast images and error estimates for every included study. That is, image-based analyses may use more data from each individual experiment but the number of experiments that can be included is generally greatly reduced. However, a recent comparison of image- and coordinate-based meta-analyses (Salimi-Khorshidi et al., 2009) revealed good agreement between meta-analyses based on full statistical contrast images and reduced 3D coordinates. Given this evaluation and the difficulties of obtaining full image data from a sufficient amount of published experiments, it seems that coordinate-based approaches such as ALE represent the most practical tool for meta-analyses on neuroimaging data.

An important caveat for the interpretation of meta-analyses is the potential presence of confounding factors in the assessed experiments. Meta-analyses pool across many studies to identify convergent findings while disregarding experiment-specific variability in design and analysis. However, the averaging effects of meta-analyses that allow for the influence of confounding factors to be ignored only pertain to unsystematic study variations. If, however, an additional cognitive process is present in a significant number of the included experiments, the ensuing activations may confound the meta-analysis. In this case, activation in a certain area would not be attributable to the process of interest but to processes that were concurrently present in the included experiments. For example, it has been argued that activation of Broca's area during imitation tasks could result from covert speech (e.g., Brass and Heyes, 2005). Assuming that vocalisation is present in the majority of the imitation experiments, vocalisation-related activity will be indistinguishable from an imitation-related one. This scenario, however, also raises the fundamental question, whether two processes that co-occur consistently in neuroimaging experiments should actually be distinguished from each other. That is, covered vocalisation and the corresponding activation of BA 44 may be an intrinsic part of action imitation rather than a confound that must be excluded. Evidence for such a genuine role of BA 44 in imitation processes, for example, is provided by recent transcranial magnetic stimulation (TMS) and lesion studies, which showed that lesions (artificial or pathological) in BA 44 led to imitation failure (Fazio et al., 2009; Heiser et al., 2003).

Areas involved in both networks

Overall, the current meta-analysis revealed a network for the observation and imitation of actions that expands both hemispheres and reaches far beyond the ‘classical’ mirror neuron areas within ventral premotor and inferior parietal cortex. This view of an “expanded MNS” involving similar areas as revealed by the current meta-analysis, has recently been assumed based on human imaging studies (Fabbri-Destro and Rizzolatti, 2008; Iacoboni, 2009) and with respect to possible homologies to the macaque mirror neuron system (Keysers and Gazzola, 2009). The present meta-analysis could provide further evidence to this discussion regarding the involvement of these networks in processing of observed and imitated actions. The results showed that areas other than the ‘classical’ mirror neuron areas vPMC and rostral IPL were consistently activated across studies, i.e., dPMC, SMA, pMTG, and V5.

Among the commonly activated areas are BA 44 and the rostral IPL/anterior IPS (areas PFt/hIP3). These two regions are thought to be the human homologues of macaque ventral premotor area F5 and rostral inferior parietal areas PFG and PF, i.e., those areas where mirror neurons were discovered using invasive recordings (e.g., Fogassi et al., 2005; Gallese et al., 1996; Rozzi et al., 2008). Activation of these regions by action observation tasks is not surprising, since “activation during action observation” is one of the key properties defining a mirror neuron (e.g., Rizzolatti, 2005). Thus, if BA 44 and the rostral IPL are indeed the homologues of the mirror neuron areas in other primates, they should be activated by action observation tasks. Activation during action imitation, however, is not a typical mirror neuron characteristic. Rather, it has been stressed that monkeys are not able to imitate in a comparable way as humans (e.g., Iacoboni, 2009; Rizzolatti, 2005). In human neuroimaging studies on imitation paradigms, however, robust activation of BA 44 and the rostral IPL have been reported (e.g., Hamilton and Grafton, 2008; Iacoboni, 2009; Iacoboni and Dapretto, 2006; Rizzolatti et al., 2001) and are confirmed in the current meta-analysis. A straightforward explanation for these findings could be provided by the experimental setup of most action imitation experiments. Typically these involve concurrent execution of an observed action, i.e., both properties that define mirror neurons. It has, however, also been argued that potential mirror neurons in the human brain may have an independent relevance for imitation tasks, even though they don’t hold the same function in non-human primates (e.g., Brass and Heyes, 2005; Culham and Valyear, 2006; Heyes, 2001). This view is largely based on conceptualising mirror properties as a matching between sensory input and motor acts (e.g., Kilner et al., 2007; Jakobs et al., 2009) and stressing the importance of such a mechanism for action observation, execution, and crucially also imitation (e.g., Fabbri-Destro and Rizzolatti, 2008). Rizzolatti (2005) moreover stressed the possibility that in particular the caudal aspect of BA 44 may represent the putative homologue of macaque area F5. Our meta-analysis confirms and extends this view in a quantitative analysis over a large number of individual experiments. Since imitation and observation recruited the very caudal aspect of BA 44 at the border to BA 6, the same region was activated during imitation as thought to be a human mirror region (Rizzolatti, 2005). With respect to the parietal cortex, the current meta-analysis could provide new evidence for the discussion of potential homologies between humans and monkeys by showing that human area PFt seems to be most consistently involved in processes that have been ascribed to area PF of the macaque.

Importantly, the location of the convergent activation within Broca's area (BA 44, BA 45) differed between action observation and action imitation tasks. Only the caudo-dorsal part of BA 44 was involved in both networks, whereas a higher consistency of activation for imitation was found in a more caudo-ventral aspect of BA 44. In turn, more consistent activation by observation tasks was found in the rostro-dorsal aspect of Broca's region (BA 45). This dissociation has already been noted in previous experiments and was interpreted as deriving from the requirements of forward modelling processes during imitation (e.g., Molnar-Szakacs et al., 2005; Morin and Grèzes, 2008, Brass and Heyes, 2005; Vogt et al., 2007). Furthermore, an explicit model for this differentiation within Broca's region was introduced by Koechlin et al. (2003) and Koechlin and Jubault (2006): Within this model, Broca's region is most likely involved in context specific recognition of stimuli. Further differentiation regarding the amount of cognitive control results in a bipartition: Activation within BA 44 was seen as being responsible for the initiation and termination of simple actions whereas activation in BA 45 was more likely ascribed to the supraordinate aspect of the action (Koechlin and Jubault, 2006). Following this model and the works by Molnar-Szakacs et al. (2005) and Vogt et al. (2007), the differentiation within Broca's region found in the present meta-analysis could be interpreted as follows. Actions shown during the observation experiments were generally more complex, whereas actions in the imitation experiments were kept simpler. This difference is owed to feasibility constraints in the scanner for imitation but not for observation studies. Thus, the dominance of the rostro-caudal part of Broca's region (BA 45) during action observation might result from the processing of more complex movements. As such, there is a high need for integrating and assessing the context of the whole action. This is less the case for the more simple actions used during the imitation experiments. These, however, pose higher needs for control and forward modelling provided by caudal BA 44.

The supplementary motor area (SMA) was also consistently found to be active during action observation as well as action imitation tasks. Whereas action-related activations in BA 44 were linked to, e.g., motor sequence learning, motor imagery, and action preparation (e.g., Binkofski et al., 1999; Johnson-Frey et al., 2003; Krams et al., 1998; Mecklinger et al., 2002) or recognition of abstract motor behaviour and associative motor learning (e.g., Binkofski et al., 2000; Hazeltine et al., 1997; Seitz and Roland, 1992), one function of SMA was seen in temporally sequencing different parts of a complex movement (e.g., Tanji, 1994; Mita et al., 2009). Tankus et al. (2009) ascribed SMA activation to the encoding of speed and direction of a movement. Furthermore, it has been shown that lesions in SMA lead to deficits in sequencing actions (Gentilucci et al., 2000). Following these previous studies, the association of SMA activation with observation and imitation tasks can be interpreted as reflecting the temporal sequencing of the action. After disassembling a complex action into different executable parts, the individual parts have to be put into a temporal sequence to imitate the observed action correctly. For observation alone, this step might as well be necessary to capture all parts of the observed action for subsequent understanding of the action as a whole. This is supported by the notion that the activation within SMA during observation was mainly driven by the observation of non-hand actions which included whole body movements which are much more complex than simple finger and hand movements. To further enlighten the role of the SMA in temporal sequencing, the observation experiments have been subdivided into those with video (i.e., moving) and those with static stimuli, assuming that static stimuli would not require the involvement of the SMA. Both subanalyses revealed a comparable network as the overall observation analysis, with a higher consistency of activations for the video subsample. But since the sample sizes were largely unequal (79 video, 15 static), a potential bias toward the video sample could not fully be excluded. Furthermore, the static sample also involved complex actions which required a disassembly of the actions into different executable parts. Therefore, the need for temporal sequencing, and thus, the involvement of the SMA in this subsample could not completely be ruled out by conceptual reasons, either.

Furthermore, the posterior middle temporal gyrus/superior temporal sulcus (pMTG/STS), anterior and dorsal to V5, was consistently involved in action observation and imitation. This region is known to be involved in the processing of biological motion (e.g., Buccino et al., 2001; Puce and Perrett, 2003; Morris et al., 2008). Since the majority of all action observation and imitation experiments included in the present meta-analysis featured the display of video clips showing natural human movements, the activation of the pMTG/ STS is well explained by this role.

Extrastriate visual area V5 has been involved within both networks revealed by the present meta-analysis. Activations in V5 have been reported in previous studies due to recognition and early processing of visually presented motion stimuli (e.g., Seymour et al., 2009; Thompson et al., 2009; Vaina et al., 2001). In the context of action observation and imitation, the involvement of area V5 could be explained in line with these previous reports, serving as an encoder of the dynamic aspect of the movement.

A part of the fusiform gyrus was also involved in both networks, most probably the fusiform face area / fusiform body area (FFA/FBA). The name of this region refers to the involvement of FFA and FBA in recognition of faces and bodies. (e.g., Downing et al., 2006; Kitada et al., 2009). In the current meta-analysis, activation in FFA/FBA was primarily found during observation of face actions and, more generally, non-hand actions, which also involved, e.g., the whole body. No activation in this region was found for the observation of hand actions. The same holds true for the imitation sample: Whereas the total analysis which also contained imitation of face actions revealed activation in FFA/FBA, the analysis of imitation of hand actions did not reveal such an activation (there, the activation is located more rostro-dorsally). Thus, in both networks, FFA/FBA most likely serves as an encoder of facial and body stimuli. Furthermore, both networks only involved right FFA/FBA which is also in line with recent studies on the lateralization of visual perception areas, arguing in favour of a specialization of hemispheres with respect to different levels of processing which results in a specialization of the right hemisphere for tasks where spatial metrics and conjoining features play an essential role, like in the recognition of faces and bodies (e.g., Willems et al., 2009; Umiltá et al., 1985; for review: Dien, 2009).

Both action observation and imitation were also robustly associated with activations of the primary somatosensory cortex (SI). While an involvement of sensorimotor cortices during action observation has been demonstrated in a recent study explicitly dealing with this issue (Gazzola and Keysers, 2009), other studies provide evidence that somatosensory cortical regions also respond to the sight of touch (Blakemore et al., 2005; Carlsson et al., 2000; Keysers et al., 2004). But still, the neurobiology of this phenomenon remains elusive. Given that primary or unimodal sensory cortices such as SI are driven by modality-specific thalamic input, these activations should be attributable to top-down modulation from associative areas. One interpretation for SI activation during action observation is that this region may act as a simulator of “what it could feel like to act as seen.” This idea of SI as providing a proprioceptive and tactile matching of seen actions has recently been advanced (Gazzola and Keysers, 2009; Keysers and Gazzola, 2009), saying that an action needs to be mapped onto one's own sensorimotor system to fully understand the motor components of the observed action. It could be speculated that this simulatory processes in SI might be coordinated by the ventral premotor cortex (BA 44 and adjacent BA 6) which has been assumed to be responsible for forward modelling processes, especially during imitation experiments (Molnar-Szakacs et al., 2005).

Neural correlates of action observation

The action observation network, as delineated by the present meta-analysis of 104 functional neuroimaging experiments, spanned both hemispheres in a largely symmetrical manner, consisting of frontal, parietal, and posterior temporal areas as assumed previously (e.g., Culham and Valyear, 2006; Fadiga et al., 2005; Lui et al., 2008; Rizzolatti and Craighero, 2004). The involvement of frontal premotor, parietal, and extrastriate visual areas within this network was also further supported by transcranial magnetic stimulation (TMS) studies. It has been shown that transient inactivation (“virtual lesions”) over these areas may result in an impaired action observation ability (for review, e.g., Fadiga and Craighero, 2004), e.g., for the discrimination of biomechanically possible actions (Candidi et al., 2008) or for the correct rearrangement of a sequence of actions (Gangitano et al., 2008).

A main question of our analysis regarding the organization of this network relates to the effect of possible confounds such as effectors, use of an object, or instructions given to the subject.

Different locations of the activations when observing actions performed by different effectors raised the question of a possible somatotopic organization within the involved areas. Buccino et al. (2004b) reported a somatotopy within the fronto-parietal part of the observation network, with observation of mouth movements activating most ventral parts (BA 44 and rostral IPL, respectively), observation of foot actions more dorsal parts, and observation of hand actions in between. With focus on the lateral premotor cortex, similar findings were reported by Sakreida et al. (2005) as well as Wheaton et al. (2004). Besides these findings on visual action processing, Gazzola et al. (2006) found a comparable somatotopical arrangement of activations in the premotor cortex for the processing of action sounds, indicating a topic arrangement of concepts rather than sensory representations. In contrast, a recent meta-analysis on action observation by Morin and Grèzes (2008) did not find a clear somatotopical arrangement of activations within the lateral premotor cortex. By comparing MNI coordinates and counting the number of hits in the macroanatomically defined lateral premotor cortex and BA 44 for different effectors, they found association of activations within BA 44 slightly more often for observation of whole body and leg movements than for observation of mouth or finger movements. In contrast, the meta-analysis by Van Overwalle and Baetens (2009) did report a somatotopic arrangement comparable to that found by Buccino et al. (2004b) and confirmed in the present, more extended analysis.

Our meta-analysis on action observation revealed a bilateral network with pronounced involvement of the lateral premotor and parietal cortex, which was confirmed to be largely independent of the effector by subanalyses on observation of hand actions, right hand actions, and non-hand actions. Contrasting observation of hand and non-hand actions, however, revealed a notable difference with regard to possible somatotopy: whereas observation of non-hand (i.e., whole body, face, and leg) actions were more associated with activation in BA 44, observation of hand actions was more consistently associated with activations in a more dorsal part of premotor cortex (BA 6). For the parietal lobe, our meta-analysis did not provide such a possible somatotopical arrangement: whereas observation of hand actions was consistently associated with activations within parietal cortex, the observation of non-hand actions was not. The difference to the results of Morin and Grèzes (2008) might be caused by the difference in sample size, which was considerably larger in our study, or the applied method. Nevertheless, it must be pointed out that meta-analyses may not be ideally suited to investigate somatotopy since pooling of data from very different studies could diminish or even delete such effects (Morin and Grèzes, 2008), especially when somatotopic organization is not very pronounced, as in the parietal cortex (e.g., Buccino et al., 2004b; Rizzolatti and Arbib, 1998).

Another potential influencing factor for the organization of the action observation brain network is the involvement of an object within the observed action. Separating the experiments on observation of hand actions into object-related and non-object-related ones revealed a major difference: whereas activation within the temporo-occipital cortex (pMTG, V5) was consistently found within both subanalyses, activation within the fronto-parietal part of the observation network was mainly driven by observation of object-related actions.

It has been proposed that activation in these regions reflects visually guided feedback control of an action (e.g., Shmuelof and Zohary, 2007). This hypothesis, however, was mainly inferred from imitation or grasping studies. The involvement of the SPL and adjacent IPS in somatosensory and visuomotor integration, reaching movements in particular, as well as object recognition has frequently been demonstrated in neuroimaging studies (e.g., Battaglia-Mayer and Caminiti, 2002; Grèzes and Decety, 2001; Hahn et al., 2006; Pellijeff et al., 2006; Rizzolatti and Matelli, 2003, Buccino et al., 2001). Moreover, it is also supported by lesion studies of patients suffering from optic ataxia, a syndrome with deficits in the online control of visually guided actions (e.g., Glover 2003). It was assumed that these superior and adjacent intraparietal areas form a human parietal reach region (e.g., Connolly et al., 2003), referring to the comparable region in macaques (for review, e.g., Grefkes and Fink, 2005). Other authors, however, reported the parietal cortex active also for the observation of non-object-related actions (e.g., Montgomery et al., 2007). These discrepant findings indicate that there apparently is no strict and exclusive neurophysiological distinction between object and non-object-related actions. Rather, the type of the observed movement and/or its spatio-temporal properties may drive neurons in some grasp-related areas.

For the lateral premotor cortex additional strong association with actions aiming at a certain target have been found. This was also interpreted as providing additional information to the visuomotor integration process required for object-related actions (for review, e.g., Hoshi and Tanji, 2007). This correlation was supported by a recent meta-analysis on the involvement of the premotor cortex in different types of action observation (Morin and Grèzes, 2008), which revealed a less consistent activation within premotor areas during non-object-related actions. Buccino et al. (2004b), however, reported that observation of object-as well as non-object-related actions activates lateral premotor areas to a comparable degree.

The data of our meta-analysis on 104 individual experiments also provide evidence that activation in the fronto-parietal part of the action observation network may not only be related to the observation of an action per se but also particularly involved in the (implicit) processing of object features and their integration within the observed motor act. For the parietal part of the network, this is in line with the concept of a human parietal reach region. For the frontal part, a stronger association with object- and goal-directed actions was assumed when considering one of these areas as a possible human homologue of macaque area F5 since macaque mirror neurons more strongly responded to such actions as compared to non-goal- and non-object-related actions (e.g., Morin and Grèzes, 2008). Our meta-analysis results support this notion, providing a further aspect for future research on such possible homologies between humans and macaques (Morin and Grèzes 2008; Nelissen et al., 2006).

For this frontal part of the network, another notable involvement was found: this part of the cortex, together with the temporo-occipital visual areas, was constantly involved when passively observing a movement, but also when intending to imitate the observed movement. Furthermore, the primary motor cortex was consistently activated during active observation. In an early study, Grèzes et al. (1999) studied a possible differentiation between active and passive observation. They also found increasing activity within premotor cortex and on the precentral gyrus (presumably primary motor cortex), but also in inferior and superior parietal cortices, which was interpreted as reflecting the information processes needed for subsequent action. The results of the present meta-analysis did not show any involvement of the parietal cortex. But the samples of active and passive observation experiments considerably differed in sample size. Thus, even with the meta-analysis algorithm covering for such differences, such larger difference could still have introduced potential bias to the present analysis. This could have led to detection failure of parietal activations during active observation since only 8 experiments could have been involved in this analysis. Thus, the present meta-analysis provides first hints that especially primary and premotor areas might consistently be involved in active observation, whereas involvement of the parietal cortex could not finally be resolved.

Another consistently activated region during action observation was the dorso-lateral premotor cortex (dPMC; BA 6). Activation of this region was also found consistently in imitation experiments, but the exact location differed between the observation and imitation sample, leading to no common activation being detected in the conjunction analysis.

Summarizing previous reports, recent reviews suggested that the dPMC is involved in learning appropriate motor responses based on arbitrary cues (Chouinard and Paus, 2006), and thus, motor planning and preparation (Hoshi and Tanji, 2004). Furthermore, it was proposed that the dPMC integrates different pieces of sensorimotor information to formulate the appropriate motor program (Hoshi and Tanji, 2007). Given this current knowledge on the dPMC, we would assume that within the action observation and action imitation networks, the dPMC might provide the composition of the appropriate motor program during movement preparation. Such a step should be required during action observation particularly to understand the observed action (e.g., Grafton and Hamilton, 2007), and certainly, for the realization of the observed action by imitation.

Neural correlates of action imitation

The action imitation network as revealed by the present meta-analysis recruited frontal, parietal, and temporo-occipital areas as previously assumed in qualitative reviews (e.g., Brass and Heyes, 2005; Heyes, 2001; Iacoboni, 2005, 2009; Turella et al., 2009a,b).

One issue of controversial discussion is a possible lateralization within activations of the action imitation network since previous studies have provided conflicting evidence on this issue. Since imitation is one form of higher-order motor processing, it could be assumed that it recruits a bilateral brain network rather than showing a hemispheric lateralisation (e.g., Iacoboni and Dapretto, 2006). Support for this assumption is provided by functional (e.g., Aziz-Zadeh et al., 2006a; Molnar-Szakacs et al., 2005) and virtual lesion studies (e.g., Aziz-Zadeh et al., 2002; Heiser et al., 2003), arguing in favour of a bilateral organization of an imitation network, in particular for frontal premotor areas.

Predominant right hemispheric activations during imitation have also been reported, e.g., within right occipito-temporal junction (Binkofski et al., 2000; Iacoboni et al., 1999, 2001), even disregarding the used hand (Aziz-Zadeh et al., 2006a), or in temporal and frontal areas for right hand movements (Biermann-Ruben et al., 2008) as well as imitation of emotional faces Carr et al., 2003).

Other studies, however, reported a dominance of left hemispheric areas during imitation tasks, which was interpreted in reference to the lateralisation of language functions (e.g., Aziz-Zadeh et al., 2004; Rizzolatti and Arbib, 1998; Iacoboni et al., 2005). Moreover, Goldenberg and Karnath (2006) argued in favour of a left-lateralisation of imitation-related processes based on lesion studies.

The quantitative results of the present meta-analysis argue in favour of a bilateral activation pattern for action imitation. Most of the included imitation experiments involved the imitation of hand movements with either the right or an unspecified hand (29 out of 35). Since action imitation contains a major component of motor execution, it could have been assumed that this imbalance would result in a dominance of left hemispheric activations, for frontal motor areas in particular. Instead, activations within these areas were evenly found in both hemispheres. Our data are thus in line with the idea of imitation being a higher-order motor process supported by a bilateral network as assumed by Iacoboni and Dapretto (2006).

Within this context, one idiosyncrasy of the imitation network could be noted: only the subanalysis ‘imitation of right hand actions’ revealed consistent activation within the right pMTG while imitation in general and imitation of hand actions did not. The importance of the right pMTG/STS for imitation processes was first pointed to by Iacoboni et al. (2001), based on an imitation study on right hand actions. In this study, activation of the pMTG for imitation was even stronger than the pure observation of the respective action.

Furthermore, it has to be noted that imitation studies included in the present meta-analysis involved online (n = 24) as well as delayed imitation experiments (n = 11). Comparison of these two subsets revealed largely comparable brain networks between both subsamples and compared to the imitation analysis including all experiments. Moreover, direct comparison revealed a higher consistency of activations for delayed as compared to simultaneous imitation in all of the activated brain areas. This could most likely be interpreted as resulting from a higher difficulty of the delayed imitation paradigms, thus a higher cognitive demand, as compared to the online imitation paradigms (e.g., Buccino et al., 2004a). A comparable effect is known for the imagery of action where strong activations can be found in premotor and visual areas in particular, in the absence of a visual model. The imagination of an action also requires a higher cognitive demand, thus resulting in stronger activations (e.g., Creem-Regehr and Lee, 2005; David et al., 2006; Grafton et al., 1996; Johnson-Frey et al., 2005). Thus, further investigation seems required to detect possibly subtle differences between on- and off-line performances of imitative behaviour.

Reference to recent meta-analyses

Two recent smaller meta-analyses (Molenberghs et al., 2009; Van Overwalle and Baetens, 2009) also reported a largely bilateral activation pattern for imitation tasks. Both these meta-analyses, however, used a region of interest (ROI)-based approach and assessed only activations which fell within predefined regions of the lateral premotor and parietal (Molenberghs et al., 2009) and posterior temporal cortex (Van Overwalle and Baetens, 2009). The definition of these ROIs was based either on estimates of the location of anatomical areas using the Talairach atlas (Molenberghs et al., 2009) or on manually delineated ones based on previous knowledge from the literature (Van Overwalle and Baetens, 2009) which could have confounded the results. Since the delineation of cortical areas should only reliably be possible by means of cytoarchitectonic investigation (not by means of macroanatomical anatomy; Amunts et al., 2007; Zilles et al., 2002), the areal definition within these previous studies might potentially introduce bias towards a misinterpretation of areal boundaries. Our meta-analysis used a different approach, assessing the action imitation network as a whole, without any a priori assumptions or focus on ROIs.

Since, in general, our findings on the action imitation network are well in line with those of previous meta-analyses as discussed above, the present meta-analysis could confirm and amend the findings of previous smaller analyses using an unbiased quantitative algorithm to synthesise results from a larger sample of primary studies.

One major difference to the analysis of Molenberghs et al. (2009) relates to the involvement of BA 44 within the action imitation network which is controversially discussed in the literature (e.g., Brass and Heyes 2005; Iacoboni 2005). The difference between our and Molenberghs' result might on the one hand be due to a methodological difference. Since we used the revised version of the ALE algorithm (Eickhoff et al., 2009), potential drawbacks of previous approaches which were used by Molenberghs et al., could be avoided. On the other hand, activation within BA 44 might have failed reaching significance in their analysis, especially considering the fact that the respective activation is located at the very caudal part of BA 44 (as stated above). This fact might have resulted in a failure of detection in an ROI-based approach as performed by Molenberghs. Furthermore, we were able to include a larger amount of imitation experiments within our analysis. For the ROI within BA 44 in Molenberghs’ analysis, an even smaller amount of activations was found since several studies did not seem to report respective activation foci within their predefined ROI. This small number of activation foci provides difficulties for the interpretation of their negative result for BA 44. The larger sample of studies within our analysis increased the power of the ALE analysis. By objectively analysing reported activations without any preallocation to a certain ROI, our analysis was able to find activation within BA 44, with a major focus in its most caudal aspect. This provides further evidence for the role of BA 44 in imitation as stated in the section about the areas involved in both networks.