ArticlePDF Available

Establishing a Role of the Semantic Control Network in Social Cognitive Processing: A Meta-analysis of Functional Neuroimaging Studies


Abstract and Figures

The contribution and neural basis of cognitive control is under-specified in many prominent models of socio-cognitive processing. Important outstanding questions include whether there are multiple, distinguishable systems underpinning control and whether control is ubiquitously or selectively engaged across different social behaviours and task demands. Recently, it has been proposed that the regulation of social behaviours could rely on brain regions specialised in the controlled retrieval of semantic information, namely the anterior inferior frontal gyrus (IFG) and posterior middle temporal gyrus. Accordingly, we investigated for the first time whether the neural activation commonly found in social functional neuroimaging studies extends to these ‘semantic control’ regions. We conducted five coordinate-based meta-analyses to combine results of 499 fMRI/PET experiments and identified the brain regions consistently involved in semantic control, as well as four social abilities: theory of mind, trait inference, empathy and moral reasoning. This allowed an unprecedented parallel review of the neural networks associated with each of these cognitive domains. The results confirmed that the anterior left IFG region involved in semantic control is reliably engaged in all four social domains. This supports the hypothesis that social cognition is partly regulated by the neurocognitive system underpinning semantic control.
Content may be subject to copyright.
NeuroImage 245 (2021) 118702
Contents lists available at ScienceDirect
journal homepage:
Establishing a role of the semantic control network in social cognitive
processing: A meta-analysis of functional neuroimaging studies
Veronica Diveica, Kami Koldewyn, Richard J. Binney
School of Human and Behavioural Sciences, Bangor University, Bangor, Gwynedd, Wales, LL57 2AS, United Kingdom
Cognitive control
Theory of mind
Moral reasoning
Trait inference
The contribution and neural basis of cognitive control is under-specied in many prominent models of socio-
cognitive processing. Important outstanding questions include whether there are multiple, distinguishable sys-
tems underpinning control and whether control is ubiquitously or selectively engaged across dierent social
behaviours and task demands. Recently, it has been proposed that the regulation of social behaviours could rely
on brain regions specialised in the controlled retrieval of semantic information, namely the anterior inferior
frontal gyrus (IFG) and posterior middle temporal gyrus. Accordingly, we investigated for the rst time whether
the neural activation commonly found in social functional neuroimaging studies extends to these ‘semantic con-
trol’ regions. We conducted ve coordinate-based meta-analyses to combine results of 499 fMRI/PET experiments
and identied the brain regions consistently involved in semantic control, as well as four social abilities: theory
of mind, trait inference, empathy and moral reasoning. This allowed an unprecedented parallel review of the
neural networks associated with each of these cognitive domains. The results conrmed that the anterior left IFG
region involved in semantic control is reliably engaged in all four social domains. This supports the hypothesis
that social cognition is partly regulated by the neurocognitive system underpinning semantic control.
1. Introduction
The ability to comprehend and respond appropriately to the be-
haviour of others is essential for humans to survive and thrive. A major
challenge for the cognitive sciences, therefore, is to characterise how
we understand others and coordinate our behaviour to achieve mutu-
ally benecial outcomes, and what can cause this ability to break down
( Frith, 2007 ). There is an indubitable requirement for systems that con-
trol, or regulate, the cognitive processes underpinning social interac-
tions. This is because social interactions are intricate and fraught with
the potential for misunderstandings and faux pas; rst, the everyday
social signals to which we are exposed are typically complex, often am-
biguous and sometimes conicting. This is compounded by the fact that
the meaning of a given gesture, expression or utterance can vary across
contexts ( Barrett et al., 2011 ; Rodd, 2020 ). Moreover, once we have set-
tled upon an interpretation of these signals, we are then faced with the
additional challenge of selecting an appropriate response, and inhibit-
ing others which might, for example, be utilitarian but socially insensi-
tive or even damaging. In order to undergo social interactions that are
coherent, eective and context-appropriate, we must carefully regulate
both our comprehension of, and response to, the intentions and actions
of others ( Binney and Ramsey, 2020 ; Fujita et al., 2014 ; Gilbert and
Burgess, 2008 ; Ramsey and Ward, 2020 ).
Corresponding author.
E-mail address: (R.J. Binney).
Despite there being a wealth of literature describing executive func-
tions involved in general cognition ( Assem et al., 2020 ; Diamond, 2013 ;
Duncan, 2013 , 2010 ; Fedorenko et al., 2013 ; Petersen and Pos-
ner, 2012 ), prominent models of socio-cognitive processing are under-
specied in terms of the contribution and neural basis of cognitive
control mechanisms (e.g., Adolphs, 2009 , 2010 ; Frith and Frith, 2012 ;
Lieberman, 2007 ). For example, Adolphs (2009 , 2010 ) only very briey
refers to cognitive processes involved in ‘social regulation’ and largely
within the limited context of emotional regulation. Likewise, Frith and
Frith (2012) refer to a “supervisory system ”which has the charac-
teristic features of executive control, but its functional and anatomi-
cal descriptions lack detail important for generating testable hypothe-
ses. However, research into specic social phenomena, such as preju-
dice ( Amodio, 2014 ; Amodio and Cikara, 2021 ) and automatic imita-
tion ( Cross et al., 2013 ; Darda and Ramsey, 2019 ) has recently begun
to give the matter of cognitive control greater attention. Of particu-
lar interest has been the contribution of the domain-general multiple-
demand network (MDN), a set of brain areas engaged by cognitively-
challenging tasks irrespective of the cognitive domain ( Assem et al.,
2020 ; Duncan, 2010 ; Fedorenko et al., 2013 ; Hugdahl et al., 2015 ).
MDN activity increases with many kinds of general task demand, includ-
ing working memory load and task switching, and it has been suggested
that this reects the implementation of top-down attentional control and
the optimal allocation of cognitive resources to meet immediate goals
( Duncan, 2013 , 2010 ). The MDN is comprised of parts of the precentral
gyrus, the middle frontal gyrus (MFG), the intraparietal sulcus (IPS), in- .
Received 9 July 2021; Received in revised form 1 October 2021; Accepted 30 October 2021
Available online 4 November 2021.
1053-8119/© 2021 The Author(s). Published by Elsevier Inc. This is an open access article under the CC BY license ( )
V. Diveica, K. Koldewyn and R.J. Binney NeuroImage 245 (2021) 118702
sular cortex, the pre-supplementary motor area (pre-SMA) and the adja-
cent cingulate cortex ( Assem et al., 2020 ; Fedorenko et al., 2013 ), some
of which have been implicated in controlled social processing such as,
for example, working memory for social content ( Meyer et al., 2012 ),
social conict resolution ( Zaki et al., 2010 ), inhibition of automatic im-
itation ( Darda and Ramsey, 2019 ) and mental state inference or the-
ory of mind (ToM) (e.g. Rothmayr et al., 2011 ; Samson et al., 2005 ;
Van der Meer et al., 2011 ). However, there are at least three key unre-
solved questions regarding the role of cognitive control in social cogni-
tion. First, it remains to be seen whether there could be multiple, distin-
guishable types of, and neural systems for, control. Second, it is unclear
whether distinguishable control systems are necessary for all or only
certain social abilities and, third, whether this engagement depends on
specic task demands. Shedding light on these issues has the potential
to generate important new hypotheses regarding social behaviour both
in the context of health and injury/disease.
It has recently been proposed that a relatively specialised form of
cognitive control, termed semantic control , could be particularly impor-
tant for social cognitive processing ( Binney and Ramsey, 2020 ). This
proposal argued that a semantic control system is required during social
cognitive tasks to modulate the retrieval and selection of conceptual-
level knowledge so that it is relevant to the situational context or the task
at hand ( Chiou et al., 2018 ; Jeeries, 2013 ; Lambon Ralph et al., 2017 ).
The reasons why semantic control should be critical for social cogni-
tion are uncomplicated; we retain a vast amount of socially-relevant
knowledge including knowledge about familiar people ( Greven et al.,
2016 ; Hassabis et al., 2014 ), about the structure of and relationship be-
tween social categories and their associated stereotypes ( Freeman and
Johnson, 2016 ; Quinn and Rosenthal, 2012 ), and an understanding
of abstract social concepts, norms and scripts ( Frith and Frith, 2003 ;
Van Overwalle, 2009 ). But only a limited portion of this information
is relevant in a given social instance and it would be computationally
inecient to automatically retrieve it all. For example, there is no need
to retrieve information about someone’s personality traits, or personal
interests and hobbies, if the only task is to pick them out from within
a crowd. Moreover, the types and the scope of information we need to
retrieve to understand and respond appropriately to certain social sig-
nals change according to the context, and irrelevant information could
potentially interfere. Therefore, semantic control should be essential for
limiting potential social errors.
There is a growing body of convergent patient, neuroimaging and
neuromodulation evidence that semantic memory retrieval engages
the semantic control network (SCN) which comprises the anterior IFG
and the posterior middle temporal gyrus (pMTG) ( Badre et al., 2005 ;
Jackson, 2021 ; Noonan et al., 2010 ; Whitney et al., 2012 ). While the
domain-general MDN is also engaged by semantic tasks, and particu-
larly those with high control demands ( Jackson, 2021 ; Thompson et al.,
2018 ), there is evidence to suggest that both the anatomy of the SCN and
MDN and their functional contributions to controlled semantic process-
ing are distinct ( Gao et al., 2020 ; Humphreys and Lambon Ralph, 2017 ;
Wang et al., 2018 ; Whitney et al., 2012 ). In particular, fMRI studies
reveal that the mid- to posterior IFG (pars triangularis and pars oper-
cularis), nodes of the MDN, have been shown to increase activity in
response to increased ‘semantic selection’ demands, a process that is en-
gaged when automatic retrieval of semantic knowledge results in com-
petition between multiple representations which must be resolved (for
example, hearing the word bank might elicit retrieval of the concept of
a riverside and a nancial institution)( Badre et al., 2005 ; Nagel et al.,
2008 ; Thompson-Schill et al., 1997 ). However, this mid- to posterior
IFG region is also engaged by other non-semantic forms of response
competition ( Badre and Wagner, 2007 ; Dobbins and Wagner, 2005 ) and
tests of inhibitory function such as the Stroop task ( Huang et al., 2020 ;
January et al., 2009 ; Nee et al., 2007 ). In contrast, activation of the ante-
rior IFG (pars orbitalis) appears to be more selective to semantic control
demands and driven specically by an increased need for ‘controlled se-
mantic retrieval’, a mechanism that is engaged when automatic semantic
retrieval fails to activate semantic information necessary for the task at
hand, and a further goal-directed semantic search needs to be initiated
( Gold et al., 2006 ; Krieger-Redwood et al., 2015 ; Wagner et al., 2001 ).
To date, there have been but a few neuroimaging investigations that
have directly questioned the involvement of the SCN in social cogni-
tive processing. Two recent fMRI studies compared activation during
semantic judgements made on social and non-social stimuli and found
that the IFG and pMTG were engaged by both stimulus types ( Binney et
al., 2016b ; Rice et al., 2018 ). Further, Satpute et al. (2014) found that
controlled retrieval, but not selection of social conceptual information
engages the anterior IFG. However, we are not aware of any prior studies
that attempt to examine the engagement of the SCN specically during
tasks that are commonly viewed as social in nature (e.g., ToM tasks).
As a starting point, rather than conducting a novel individual exper-
iment, the present study adopted a meta-analytic approach to extract
reliable trends from large numbers of studies. Meta-analyses of func-
tional neuroimaging data overcome the limitations of individual studies
( Cumming, 2014 ; Eickho et al., 2012 ), which are frequently statisti-
cally underpowered ( Button et al., 2013 ) and vulnerable to eects of
idiosyncratic design and analytic choices ( Botvinik-Nezer et al., 2020 ;
Carp, 2012 ) so that it becomes dicult to distinguish between replicable
and spurious ndings and to generalize the results. Our principal aim
was to determine whether the distributed neural activation commonly
associated with functional neuroimaging studies of social cognition ex-
tends to the neural networks underpinning semantic control (i.e., SCN
and MDN). In order to localise the brain network sensitive to seman-
tic control demands (i.e., semantic retrieval and/or selection), and then
compare and contrast it to networks implicated in social cognition, we
performed an update of Noonan et al. (2013) meta-analysis of seman-
tic control (also see Jackson, 2021 for another updated meta-analysis
of semantic control which additionally investigated the eect of input
We took the approach of investigating multiple sub-domains of so-
cial cognition in parallel because this should allow an assessment of
the extent to which inferences are generalisable, rather than specic
to certain types of social tasks and/or abilities. We chose to focus on
four particular areas of research that target abilities frequently identi-
ed as key facets of the human social repertoire - ToM, empathy, trait
inference, and moral reasoning ( Lieberman, 2007 ; Van Overwalle, 2009 )
– and, for each, we conducted separate meta-analyses of the available
functional imaging data to determine the brain regions consistently im-
plicated. In the case of trait inference, this was the rst neuroimaging
meta-analysis to include studies that used stimuli other than faces (see
Section 2 , and also Bzdok et al., 2011 , and Mende-Siedlecki et al., 2013 ,
for contrasting approaches). In the other three cases, we performed up-
dates of prior meta-analyses ( Eres et al., 2018 ; Molenberghs et al., 2016 ;
Timmers et al., 2018 ).
Further, we conducted an exploratory conjunction analysis aimed
at identifying brain areas reliably implicated in all four social sub-
domains and, thus, a core network for social cognitive processing (also
see Bzdok et al., 2012 ; Schurz et al., 2020 ; Van Overwalle, 2009 ). We
hypothesised that this core network would include parts of the MDN
and the SCN. It is of note that, across all four social sub-domains, we
took a dierent approach to study inclusion and exclusion criteria than
that taken by some prior meta-analyses of general social cognition (e.g.,
Van Overwalle, 2009 ). In particular, we excluded studies investigat-
ing processes associated primarily with the self because social cogni-
tion is, although perhaps only in the strictest sense, about understand-
ing other people. We also excluded studies in which tasks could be
completed based on relatively simple perceptual processing and with-
out a need for deeper cognitive and inferential processes (e.g., emo-
tion discrimination tasks, automatic imitation). This was done in an
attempt to constrain our inferences to be about the neurobiology un-
derpinning cognitive rather than primarily perceptual social processes
(for further detail on this distinction, see Adolphs, 2010 , and Binney and
Ramsey, 2020 ).
V. Diveica, K. Koldewyn and R.J. Binney NeuroImage 245 (2021) 118702
Finally, as a secondary aim, the present study used the meta-
analytic approach to assess whether there are dierences in the neu-
ral networks engaged by implicit and explicit social processing (also
see Dricu and Frühholz, 2016 ; Eres et al., 2018 ; Fan et al., 2011 ;
Molenberghs et al., 2016 ; Timmers et al., 2018 ). This was aimed at
addressing a pervasive distinction in the social neuroscientic litera-
ture between automatic and controlled processes ( Adolphs, 2010 ; Happé
et al., 2017 ; Lieberman, 2007 ), and followed an assumption that implicit
paradigms engage only automatic processes, whereas controlled pro-
cesses are recruited during explicit paradigms ( Sherman et al., 2014 ).
Automatic processes are described as unintentional, eortless, and fast,
whereas controlled processes are deliberate, eortful, and thus slower
( Lieberman, 2007 ; Shirin and Schneider, 1977 ). Some authors have
argued that automatic and controlled social processes are mutually ex-
clusive of one another and draw upon distinct cortical networks, with
the former engaging lateral temporal cortex, the amygdala, ventrome-
dial frontal cortex and the anterior cingulate, and the latter engag-
ing lateral and medial prefrontal and parietal cortex ( Forbes and Graf-
man, 2013 ; Lieberman, 2007 ). However, these dual-process models have
been criticised for over-simplifying both the distinction and the rela-
tionship between automatic and controlled processes ( Amodio, 2019 ;
Cunningham and Zelazo, 2007 ; Ferguson et al., 2014 ; Fidler and Hüt-
ter, 2014 ; Fujita et al., 2014 ; Melniko and Bargh, 2018 ). An alternative
proposal, that we describe above, makes a dierent distinction - one be-
tween representation and control. This neurocognitive model proposes
that social processing relies on a single-route architecture wherein the
degree to which cognitive processing has certain attributes (e.g., speed
or eort) does not reect one system versus another. Instead, it is pro-
posed that it reects the degree to which the control system needs to ex-
ert inuence, upon otherwise automatic activation within the represen-
tational system, in order to meet the demands of a task in an appropriate
and ecient manner ( Binney and Ramsey, 2020 ; Jeeries, 2013 ). If the
dual route model is correct, explicit but not implicit social paradigms
should dierentially engage brain regions associated with cognitive con-
trol demands, including the SCN and MDN. If the single-route model is
correct, then there should be no qualitative dierence in terms of the
network of regions activated by implicit paradigms (ergo automatic pro-
cessing) and explicit paradigms (ergo controlled processing), although
there may be dierences in the magnitude of regional activation.
To summarise, the aims of the present study were as follows: 1) ex-
plore the involvement of domain-general control systems in social cog-
nition; more specically, determine whether social cognitive process-
ing reliably engages brain areas implicated in the controlled retrieval
and selection of conceptual knowledge; and 2) examine the evidence
for dual-route and single-route models of controlled social cognition.
2. Methods
Preregistration and Open Science statement. Following open science ini-
tiatives ( Munafò et al., 2017 ), the current study was pre-registered via
the Open Science Framework (OSF). We adhered to our pre-registered
protocols (available at: ) with a few minor exceptions (see
Section S1 of Supplementary Information (SI) 1 for details). All the
raw datasets are openly-available on the OSF project page (available
at: ) and are accompanied by a range of study character-
istics including details that are not the focus of the present study but
may be of interest in future research (please see Section S1 of SI1 for
a detailed description). Moreover, the input data and output les of all
analyses can be accessed via the OSF page.
In accordance with our pre-registered aims, we performed a com-
prehensive review of published functional neuroimaging studies inves-
tigating four social abilities –Theory of mind (ToM), trait inference,
empathy and moral reasoning - and independent coordinate-based meta-
analyses aimed at characterising the brain-wide neural networks un-
derpinning each. In the case of three of these domains (ToM, empathy
and moral reasoning), we updated earlier meta-analyses ( Eres et al.,
2018 ; Molenberghs et al., 2016 ; Timmers et al., 2018 ), capitalizing on
additional data, and also implementing recommendations for best prac-
tice that became available in a year subsequent to these prior studies
( Müller et al., 2018 ). In the case of trait inference, as far as we are
aware, this was the rst neuroimaging meta-analysis to include studies
that explored potential sources of information beyond face stimuli (for
contrasting approaches see Bzdok et al., 2011 ; Mende-Siedlecki et al.,
2013 ). To localise the brain areas underpinning semantic retrieval and
selection, we also updated a meta-analysis of functional imaging studies
of semantic control by Noonan et al. (2013) . This involved the inclu-
sion of additional data, and improvements in meta-analytic tools which
corrected previous implementation errors that led to the use of liberal
statistical thresholds ( Eickho et al., 2017 ).
To directly address our rst aim, we assessed the degree of overlap
between the neural networks supporting semantic control and those in-
volved in social information processing via a set of formal conjunctions
and contrasts analyses. To address our second aim, where possible, we
contrasted brain-wide activation associated with explicit versus implicit
social cognitive paradigms. Tasks that drew the participant’s attention
to the behaviour/cognitive process of interest were categorised as ex-
plicit, while tasks that used non-specic instructions (e.g., they involved
passive observation of stimuli) or employed orthogonal tasks (e.g., age
judgement) were categorised as implicit. Finally, where sucient rele-
vant information was available, we explored the inuence of task di-
culty on patterns of brain activation.
All of the meta-analyses reported below were conducted following
best-practice guidelines recommended by Müller et al. (2018) . This, as
well as several renements to inclusion/exclusion criteria, contributed
to methodological dierences between the present meta-analyses and
those prior meta-analyses upon which the ‘updates’ were based. A sum-
mary of similarities and dierences is provided in Table S1 (SI1) and
further details are given in the sections below.
2.1. Literature selection and inclusion criteria
2.1.1. General approach and criteria
Where possible, relevant functional neuroimaging studies were ini-
tially identied based on their inclusion in a recent prior neuroimaging
meta-analysis. These lists were supplemented via a search on the Web of
Science (WoS) online database ( ) for origi-
nal reports published in the years subsequent, and by searching through
reference lists of said articles. Each WoS search used the terms [‘fMRI’
or ‘PET’], as well as terms uniquely chosen for a given cognitive domain
(see Table 1 ).
A general set of inclusion criteria applied to all our analyses were as
(1) Only studies that employed task-based fMRI or PET to obtain orig-
inal data were included. Studies employing other techniques (e.g.,
EEG/MEG), meta-analyses and review articles were excluded.
(2) Studies were only included if they reported whole-brain activa-
tion coordinates that were localised in one of two standardised
spaces – Talairach (TAL) or Montreal Neurological Institute (MNI)
–or these coordinates were made available on request (see Sec-
tion S1 of SI1). Coordinates reported in TAL space were converted
into MNI space using the Lancaster transform (tal2icbm transform
( Lancaster et al., 2007 ) embedded within the GingerALE software
version 3.0.2; ). Studies exclusively report-
ing results from region-of-interest or small volume correction anal-
yses were excluded because these types of analysis violate a key as-
sumption of coordinate-based meta-analyses ( Eickho et al., 2012 ;
Müller et al., 2018 ).
(3) Studies were only included if they reported activation coordinates
that resulted from univariate contrasts clearly aimed at identifying
the process of interest (e.g., ToM). We included contrasts between an
experimental task and either a comparable active control task or a
V. Diveica, K. Koldewyn and R.J. Binney NeuroImage 245 (2021) 118702
Table 1
Terms used to search the Web of Science database for relevant articles.
Cognitive domain Search terms
Semantic control ‘semantic’, ‘comprehension’, ‘conceptual knowledge’, ‘selection’, ‘retrieval’, ‘inhibition’, ‘control’, ‘elaboration’, ‘uency’, ‘ambiguity’, ‘metaphor’, ‘idiom’
Theory of Mind ‘theory of mind’, ‘ToM’, ‘mentalising’, ‘mentalizing’
Trait inference ‘social judgement’, ‘social evaluation’, ‘social attribution’, ‘trait inference’, ‘impression formation’
Empathy ‘empathy’, plus ‘empath
- corresponding variations (e.g. ‘empathic’)
Moral cognition ‘morality’, ‘moral’, ‘moral decision making’, ‘moral emotion’, ‘harm’, ‘guilt’
N.b., For all ve cognitive domains, the search followed the following format: [fMRI OR PET] AND [term1 OR term2 OR …OR termX].
low-level baseline such as rest or passive xation. Contrasts against
low-level baselines were included in the primary analyses because
they can reveal activity associated with domain-general cognitive
processes that is subtracted out by contrasts between active con-
ditions. This could include semantic processes that are common to
both social and non-social tasks. However, contrasts against low-
level baselines also yield activity associated with dierences in per-
ceptual stimulation and attentional demand. To address this caveat,
we repeated the analyses whilst excluding this subset of contrasts.
The results can be found on the project’s OSF page (available at: ). We excluded contrasts that make comparisons be-
tween components of the process of interest (e.g., aective vs. cog-
nitive ToM; utilitarian vs. deontological moral judgements) because
we were interested in the common, core processes that would be
subtracted out by these contrasts (but see the following paragraph).
(4) Multiple contrasts from a single group of participants (e.g., separate
contrasts against one of two dierent baseline conditions) were in-
cluded in a single meta-analysis as long as they independently met
all other inclusion criteria for the primary analyses. This allowed
maximum use of all available data and enabled us to evaluate the
eect of using dierent types of baseline, for example (see above).
However, it is important to adjust for this ( Müller et al., 2018 ), and
accordingly, we adopted an approach to controlling for within-group
eects ( Turkeltaub et al., 2012 ); specically, sets of activation co-
ordinates from dierent contrasts, but the same participant group,
were pooled. This means that, when we refer to the numbers of ex-
periments, we have counted multiple contrasts from a single partic-
ipant sample as one single experiment. In cases where two or more
published articles contained data from the same participant sam-
ple, we pooled distinct contrasts as above, and excluded duplicates.
This partially explains why the number of experiments in our anal-
yses is lower than in those of some prior meta-analyses. However,
in formal contrast analyses that compare dierent conditions (e.g.,
instructional cue, task diculty), contrasts like these would be sep-
arated, and care was also taken to minimize the dierence in the
number of experiments on either side of the contrast. For example,
if a study reported two contrasts –one implicit and one explicit -
based on the same participant group, only the peaks from the im-
plicit task would be included in the contrast/conjunction analyses if
there were a greater number of explicit than implicit tasks overall
(see Fig. S9 of SI1).
(5) Only studies that tested healthy participants were included. Con-
trasts including clinical populations or pharmacological interven-
tions were excluded.
(6) Only research articles published in English were included.
2.1.2. Theory of mind
This meta-analysis was built upon that of
Molenberghs et al. (2016) and only included studies that were
specically designed to identify the neural network underpinning ToM
processes (i.e., they employed tasks involving inferences about the
mental states of others, including their beliefs, intentions, and desires).
Therefore, studies that looked at passive observation of actions, social
understanding, mimicry or imitation were not included, unless tasks
included a ToM component. Unlike Molenberghs et al. (2016) , we
excluded studies investigating irony comprehension (e.g., Wang et al.,
2006 ) because ToM might not always be necessary to detect non-
literal meaning in language ( Ackerman, 1983 ; Bosco et al., 2018 ;
Pexman, 2008 ) and studies that employed interactive games (e.g.,
Rilling et al., 2008 ). These latter studies are commonly designed to
investigate the degree to which ToM is engaged under dierent task con-
ditions rather than distinguish activation associated with ToM from that
related to other processes. Moreover, unlike Molenberghs et al. (2016) ,
we excluded studies that employed trait inference tasks as these were
considered separately (see Section 2.1.3 ).
The Molenberghs et al. (2016) search was inclusive of fMRI studies
published prior to July 2014 and yielded 144 independent experiments
(1789 peaks) contributing to their analysis. We performed a WoS search
for further original fMRI and PET studies conducted between August
2014 and March 2020, and a search of PET studies published prior to
July 2014. We then applied our inclusion criteria to both newly iden-
tied studies and those analysed by Molenberghs and colleagues (see
Table S1 of SI1 for further dierences in criteria). In the end, we found
136 experiments with a total number of 2158 peaks and 3452 partic-
ipants that met our criteria for inclusion (see Fig. S1of SI1 for more
details regarding the literature selection process; and Table S1 of SI2
for a full list of the included experiments).
2.1.3. Trait inference
Studies were included in the meta-analysis if they employed tasks
that required the participants to infer the personality traits of others
based on prior person knowledge or another’s appearance and/or be-
haviour. Whereas the types of mental states typically inferred in ToM
tasks are transitory in nature (e.g., relating to immediate goals or the in-
tentions behind a specic instance of behaviour), traits are coherent and
enduring dispositional characteristics of others (i.e., personality traits;
Van Overwalle, 2009 ). Previous meta-analyses ( Molenberghs et al.,
2016 ; Schurz et al., 2014 ) of ToM have included tasks requiring trait
inferences. However, it has been suggested that personality trait infer-
ences dier from mental state inferences in terms of likelihood and speed
of processing, and hold a higher position in the hierarchical organisation
of social inferential processes ( Korman and Malle, 2016 ; Malle and Hol-
brook, 2012 ). In line with this proposal, we maintained a distinction
and performed separate analyses. Moreover, previous imaging meta-
analyses of trait inference were limited to studies that used face stimuli
( Bzdok et al., 2011 ; Mende-Siedlecki et al., 2013 ). However, trait infer-
ences can be made on the basis of many dierent sources of information,
including physical appearance, behaviour and prior knowledge about
others ( Uleman et al., 2007 ). To our knowledge, the present attempt is
the rst to include studies that required participants to make trait in-
ferences based on facial photographs, behavioural descriptions or prior
person knowledge. We excluded any studies that asked participants to
make inferences about transitory mental states, including basic emo-
tions. We also excluded studies that did not use a subtraction approach,
but rather investigated brain activity that varied parametrically with the
levels of a pre-dened trait dimension (e.g. Engell et al., 2007 ). Finally,
we excluded studies that included emotional face stimuli to avoid con-
ating brain activity related to trait inference with that associated with
emotion recognition and processing.
V. Diveica, K. Koldewyn and R.J. Binney NeuroImage 245 (2021) 118702
We performed a WoS search of studies published before March 2020
and reference-tracing to identify relevant studies for inclusion in the
meta-analysis. A total of 40 experiments with 523 peaks and 732 par-
ticipants were found to meet the criteria for inclusion (Fig. S2 of SI1;
Table S2 of SI2).
2.1.4. Empathy
This meta-analysis was built upon that of Timmers et al. (2018) and
only included studies that were specically designed to identify the neu-
ral network underpinning empathy by employing tasks asking partici-
pants to observe, imagine, share and/or evaluate the emotional or sen-
sory state of others. The task denition was kept identical to previous
meta-analyses on empathy ( Fan et al., 2011 ; Timmers et al., 2018 ). We
also made a distinction between tasks eliciting empathic responses to
other people’s pain and those investigating empathic responses to oth-
ers’ aective states.
Timmers et al. (2018) included studies published before December
2017, totalling 128 studies with 179 contrasts (1963 peaks). We iden-
tied additional original studies conducted between January 2018 and
March 2020 via a WoS search and subsequently applied our inclusion
criteria to all, including those analysed by Timmers et al. (2018) (see
Table S1 of SI1 for further dierences in criteria). This resulted in a
yield of 163 experiments with a total number of 2691 peaks and 4406
participants (Fig. S3 of SI1; Table S3 of SI2). Empathy for pain was inde-
pendently investigated in 93 of these experiments, empathy for aective
states was independently explored in 69 experiments, and 9 experiments
concurrently explored both empathy for pain and emotions in the same
2.1.5. Moral reasoning
This analysis updated a previous meta-analysis conducted by
Eres et al. (2018) and included studies that employed tasks designed
to investigate judgements and decision-making based on moral values.
In line with Eres et al. (2018) , studies that did not specically have a
morality component were not included. For example, studies investi-
gating judgements regarding adherence to social expectations but not
moral values (e.g., Bas-Hoogendam et al., 2017 ) were excluded.
Eres et al. (2018) ’s search was restricted to fMRI studies and covered
the period before February 2016 yielding 123 contrasts (989 peaks). We
expanded this list via a WoS search for original fMRI and PET studies
published between March 2016 and March 2020, and a search for PET
studies published before March 2016, and then applied our inclusion
criteria (see Table S1 of SI1 for dierences in criteria). This resulted
in a yield of 68 experiments with a total number of 884 foci and 1587
participants (Fig. S4 of SI1; Table S4 of SI2).
2.1.6. Semantic control
In this meta-analysis, we sought to extend an earlier meta-analysis
conducted by Noonan et al. (2013) (also see Jackson, 2021 ). In line with
theirs, this analysis only included studies that were specically investi-
gating semantic processing, and that reported contrasts that reected
high > low semantic control within a semantic task, or comparisons be-
tween a task requiring semantic control and an equally demanding exec-
utive decision in a non-semantic domain. We excluded studies with a fo-
cus upon priming without an explicit semantic judgement (e.g., primed
lexical decision), bilingualism, episodic memory, or sleep consolidation.
Noonan et al. (2013) ’s search covered the period between January
1994 and August 2009 and yielded 53 studies (395 peaks) that met their
criteria for inclusion in their analysis. We performed a WoS search for
original studies published between September 2009 and March 2020,
and reference-tracing, and then applied our inclusion criteria to both
newly identied studies and those analysed by Noonan et al. (2013) .
This produced a yield of 92 experiments with a total number of 971
peaks and 1966 participants that met the criteria for inclusion in our
analysis (Fig. S5 of SI1; Table S5 of SI2).
2.2. Data analysis
We performed coordinate-based meta-analyses using the revised ac-
tivation likelihood estimation (ALE) algorithm ( Eickho et al., 2012 ,
2009 ; Turkeltaub et al., 2012 ) implemented in the GingerALE 3.0.2
software ( ). We used the GingerALE software to
conduct two types of analysis. The rst were independent dataset anal-
yses, which were used to identify areas of consistent activation across
particular sets of experiments. These analyses were performed only on
the experiment samples with a recommended minimum of 17 exper-
iments in order to have sucient power to detect consistent eects
and circumvent the possibility of results being driven by single experi-
ments ( Eickho et al., 2016 ). The ALE meta-analytic method treats re-
ported activation coordinates as the centre points of three-dimensional
Gaussian probability distributions which take into account the sam-
ple size ( Eickho et al., 2009 ). First, the spatial probability distribu-
tions of all coordinates reported were aggregated, creating a voxel-wise
modelled activation (MA) map for each experiment. Then, the voxel-
wise union across the MA maps of all included experiments was com-
puted, resulting in an ALE map that quanties the convergence of results
across experiments ( Turkeltaub et al., 2012 ).The version of GingerALE
used in the present study tests for above-chance convergence between
experiments ( Eickho et al., 2012 ) thus permitting random-eects
Following the recommendations of Müller et al. (2018) , for the main
statistical inferences, the individual ALE maps were thresholded using
cluster-level family-wise error (FWE) correction of p < .05 with a prior
cluster-forming threshold of p < .001 (uncorrected). Cluster-level FWE
correction has been shown to oer the best compromise between sensi-
tivity to detect true convergence and spatial specicity ( Eickho et al.,
2016 ). However, we subsequently applied an additional and more con-
servative threshold at the voxel level (FWE corrected at p < .05). This
level of thresholding suers from decreased sensitivity to true eects,
but has the advantage of allowing an attribution of signicance to
each voxel and thereby increases the spatial specicity of inferences
( Eickho et al., 2016 ). The FWE-corrected cluster-level and voxel-height
thresholds were estimated using a permutation approach with 5000 rep-
etitions ( Eickho et al., 2012 ). None of the meta-analyses that we up-
dated had used the recommended cluster-level FWE or the FWE height-
based correction methods.
The second set of analyses, conjunction and contrast analyses, were
also performed in GingerALE and were aimed at identifying similarities
and dierences in neural activation between the dierent sets of stud-
ies. The conjunction images were generated using the voxel wise mini-
mum value ( Nichols et al., 2005 ) of the included ALE maps to highlight
shared activation. Contrast images were created by directly subtract-
ing one ALE map from the other to highlight unique neural activation
associated with each dataset ( Eickho et al., 2011 ). Then, the dier-
ences in ALE scores were compared to a null-distribution estimated via
a permutation approach with 5000 repetitions. The contrast maps were
thresholded using an uncorrected cluster-forming threshold of p < .001
and a minimum cluster size of 200 mm
In addition, we performed post-hoc analyses to investigate if the
clusters of convergence revealed by the ALE analyses were driven by
experiments featuring specic characteristics of interest (i.e., type of
instructional cue, task diculty). To this end, we examined the list of
experiments that contributed at least one peak to each ALE cluster and
compared the number of contributing experiments featuring the char-
acteristic of interest (e.g., explicit vs implicit processing) by conduct-
ing Fisher’s exact tests of independence and post-hoc pairwise compar-
isons (using False Discovery Rate correction for multiple comparisons)
in RStudio Version 1.2.5001 ( RStudio Team, 2020 ).
A full list of the conrmatory and exploratory analyses we conducted
can be found in Section S3 of SI1.
V. Diveica, K. Koldewyn and R.J. Binney NeuroImage 245 (2021) 118702
Fig. 1. Binary whole-brain ALE maps showing statistically signicant convergent activation resulting from independent meta-analyses of ToM studies ( N = 136),
trait inference ( N = 40), empathy for pain ( N = 85) and emotions ( N = 69) and moral reasoning ( N = 68). The ALE maps were thresholded using an FWE corrected
cluster-extent at p < .05 with a cluster-forming threshold of p < .001 (red) and an FWE corrected voxel-height threshold of p < .05 (yellow). The lateral views, which
show projections on the cortical surface, are accompanied by brain slices at the sagittal midline and also coplanar with the peak of the left IFG cluster observed
across all social domains ( X = 39; Table S1.5).
3. Results
3.1. The “Social brain
3.1.1. Theory of mind
Convergent activation across all 136 ToM experiments was found in
13 clusters (see Fig. 1 a and Table S1.1.1 of SI3) located within the bi-
lateral middle temporal gyrus (MTG) (extending anteriorly towards the
temporal poles and also in a posterior and superior direction towards
the superior temporal gyrus (STG) and angular gyrus (AG) in both hemi-
spheres), bilateral IFG, bilateral dorsal precentral gyrus, ventromedial
prefrontal cortex (vmPFC), dorsomedial prefrontal cortex (dmPFC), pre-
SMA, precuneus, left fusiform gyrus and left and right cerebellum. All
these clusters survived both the height-based and extent-based thresh-
olding. A cluster in the posterior cingulate cortex (PCC) survived height-
based thresholding but did not survive extent-based thresholding. These
results are largely consistent with those of Molenberghs et al. (2016) ,
with the dierence being that they did not nd activation in SMA, left
fusiform gyrus or cerebellum. In order to address concerns regarding the
validity of some other popular ToM tasks ( Heyes, 2014 ; Quesque and
V. Diveica, K. Koldewyn and R.J. Binney NeuroImage 245 (2021) 118702
Rossetti, 2020 ), we conducted a separate supplementary meta-analysis
that was limited to the subset of ToM experiments that employed false
belief tasks (see Section S3.1 of SI1; Table S1.1.2). This analysis revealed
convergent activation in similar temporo-parietal and medial frontal re-
gions to the inclusive ToM analysis but did not implicate the lateral
frontal cortex.
3.1.2. Trait inference
The ALE meta-analysis revealed convergent activation across 40 ex-
periments in 8 clusters ( Fig. 1 b, Table S1.2) implicating the bilateral
IFG, dmPFC, vmPFC, PCC, right pMTG (extending to AG), left AG and
left anterior MTG. Voxels from all clusters, except for those in the right
pMTG and vmPFC, survived the more conservative height-based thresh-
3.1.3. Empathy
The ALE meta-analysis of all 163 empathy experiments revealed
16 clusters of convergent activation (Fig. S7a; Table S1.3.1), includ-
ing in the bilateral IFG (extending towards the insula), SMA, dmPFC,
bilateral posterior inferior temporal gyrus (ITG), right pMTG, bilateral
supramarginal gyrus (SMG), left inferior parietal lobule (IPL), bilat-
eral occipital cortex, bilateral amygdala, left thalamus, left caudate and
brainstem. These clusters survived both the height-based and extent-
based thresholding, except for the anterior dmPFC and right pMTG clus-
ters which survived extent-based thresholding only. One cluster in the
right cerebellum survived height-based thresholding but did not survive
cluster extent-based thresholding. These areas were also implicated by
Timmers et al. (2018) . In contrast, however, we did not nd convergent
activation in the left posterior fusiform gyrus, left SMG (although we
found a cluster slightly more posterior and inferior), left anterior ITG,
right TP, precuneus, middle cingulate gyrus, and