Neural response to aggressive and positive interactions in violent offenders and nonviolent individuals

Background Due to its severe negative consequences, human violence has been targeted by a vast number of studies. Yet, neurobiological mechanisms underlying violence are still widely unclear and it seems necessary to aim for high ecological validity to learn about mechanisms contributing to violence in real life. Methods The present functional magnetic resonance imaging (fMRI) study investigated the neurofunction of individuals with a history of violent offenses compared with that of controls using a laboratory paradigm requesting individuals to empathically engage in videos depicting provocative aggressive and positive social interactions from a first‐person perspective. Results The contrast of aggressive vs. positive scenarios revealed midbrain activation patterns associated with caudal periaqueductal gray (PAG) in violent offenders; In controls, the rostral PAG was involved. Additionally, only in controls, this contrast revealed an involvement of the amygdaloidal complex. Moreover, in violent offenders the contrast of positive vs. aggressive situations revealed an involvement of areas in the insula, post‐central gyrus and anterior cingulate cortex. Conclusions Our results support findings on the differential role of PAG subdivisions in response to threat and point to altered processing of positive social interactions in violent offenders. They further support the notion that changes in PAG recruitment might contribute to violent individuals “taking action” instead of freezing in case of threatening situations.

Concerning the cortex, studies investigating individuals displaying aggressive behaviour as a trait (for review, see McKinley et al., 2018) or after a traumatic brain injury (Darby, 2017) suggest that deviations in frontal and temporal cortices underlie violence. According to the studies, aggressive behaviour can result from dysfunctional emotion regulation, impaired inhibition of action, difficulties with reaching appropriate moral judgments and with using them as a basis of (re)action (cf. Darby, 2017;McKinley et al., 2018;Raine, 2019).
Within the subcortical parts of the limbic system, the amygdala has been linked to the occurrence of violence due to its role in processing emotionally relevant input, enabling emotional learning and "emotionally triggered" reactions via its interconnectedness with prefrontal and temporal areas as well as the midbrain (cf. Rosell & Siever, 2015). Of midbrain centers, the periaqueductal gray (PAG) has been included into the neurobiological circuitry underlying violence because of its role in freeze, flight and, importantly, fight reactions upon threat (Roelofs, 2017). In line with findings previously revealed by studies investigating the role of the PAG in animals (e.g., Depaulis et al., 1992), human fight and its reactions have been linked with the activation of the caudal, or dorsolateral, subdivision of the PAG (dlPAG; Roelofs, 2017), while activation of the rostral, or ventrolateral, subdivision of the PAG (vlPAG) is proposed to be related to inhibiting fight and flight reactions. Depending on input by the amygdala, activation in the vlPAG blocks fight/flight reactions and initiates freezing instead (Roelofs, 2017).

Different functional mechanisms within the "violence network"?
Research leading to the proposition of a neurobiological network underlying violence has largely been derived from the description of two clinically relevant phenotypes of individuals exhibiting violence: Individuals from the first phenotype are described to act violently without an observable trigger to achieve certain goals, that is, proactively or instrumentally. Individuals of the second phenotype show violence mostly following a trigger (e.g., provocation or frustration), that is, they exhibit violence reactively or impulsively. This phenotype has been associated with a hostile attribution bias, that is, the tendency of individuals to over-attribute hostile intentions in others even in the face of nonhostile social cues (cf., e.g., Card & Little, 2006). Different functional mechanisms of amygdala and frontotemporal activation have been associated with the two phenotypes: A hypoactivation of the amygdala (Lozier et al., 2014in Rosell & Siever, 2015 in combination with a "neuromoral" dysfunction, that is, impairment in the functioning of prefrontal areas implicated in moral decision making (e.g., OFC), is assumed to underlie proactive aggression (Raine, 2019). In contrast, a hyperactivation of the amygdala in response to a trigger in combination with a failure to prefrontally downregulate this heightened activation constitutes an "emotional hyperreactivity" and underlies reactive violence (Rosell & Siever, 2015).
Scientifically relevant and clinically useful as this distinction has been, authors argue for "differences in degree rather than in kind" (Raine, 2019). Indeed, we are confronted with the reality that most vio-lent individuals cannot easily be assigned to one group or the other (Raine, 2019;Rosell & Siever, 2015). Moreover, based on the assumption of a neurobiological network contributing to violence, we must conclude that the occurrence of violence might be a result of impairment in any of the implicated network regions or dysfunctions in their connectivity, leaving the neurobiological characteristics underlying violence still widely unclear.

Social context of violence and the present study
In addition to individual characteristics, the situational context is an important aspect of violent impulses actually being put into action.
Specifically, interpersonal provocation and/or proximal threat is considered to be one main reason for human aggression (cf. Anderson & Bushman, 2002;Fehr & Achtziger, 2021;Fehr et al., 2014). Widely established laboratory measures of aggression using noise blasts or money subtraction have been discussed to provoke a revenge-like, minor reactive aggression in the sense of tit for tat (Fehr et al., 2014;Ferguson & Dyck, 2012). However, they might not provoke a defencelike reactive aggression as a response to interpersonal provocation or proximal threat. As such, their expressive power regarding real-life, potentially maladaptive forms aggression might be limited (cf. Chester & Lasko, 2018;Fehr & Achtziger, 2021;Fehr et al., 2014;Ferguson & Dyck, 2012). Thus, in order to improve ecological validity, it seems relevant to (1) investigate individuals who actually have committed real violent offenses and (2) aim for experimental approaches simulating a context in which violence frequently occurs in a quasi-realistic way.
For this purpose, we investigated brain physiological correlates during the presentation of provocative aggressive in contrast to social positive video scenarios in a group of individuals who committed violent offenses and in a group of nonviolent individuals.
Based on the data presented by Fehr et al. (2014), we expected for all participants confronted with provocative aggressive versus social positive scenarios that at iso-cortical level distributed activation patterns occur in several heteromodal association cortices located in the superior parietal, lateral parieto-temporal, lateral occipito-temporal and prefrontal (premotor) brain areas. These were discussed to be associated with a prototypic and lifelong learned perception-actioncycle brain network related to reactive aggressive behaviours as response to (here, quasi-realistic) proximal threat displayed in the video scenarios used as stimulation (see also the concept of the emotional body language network proposed by De Gelder et al. 2006).
Additionally, we expected right inferior frontal brain activation patterns, which have been discussed as an inhibiting instance, as the participants were indeed asked to show no real motor behaviour in the scanner tube, or because they would suppress reactive aggressive motor behaviour driven by concepts of adequate conflict reducing control (cf. Fehr et al., 2014).
We expected specific patterns of activation for violent offenders and nonviolent controls when confronted with the provocative aggressive versus social positive scenarios in regions previously implicated in violence, including the amygdaloid complex and midbrain structures.
As individuals executing violence often have a history of experiencing violence themselves throughout their lives (e.g., Afifi et al., 2019), we assumed that comparing violent offenders' neurofunction during aggressive versus positive social scenarios might reveal different activation patterns in experience-related perception-action cycle systems than the respective contrast in CON individuals.
The amygdaloid complex (AMY) and associated neural networks have been viewed to optionally provide an important instance in both conscious and automatic, preattentive evaluation of emotional context aspects and have especially been implicated in the processing of fear (e.g., LeDoux & Phelps, 2008;Pehlps & LeDoux, 2005). Furthermore, a downregulation or lack of involvement of the amygdaloid complex in emotional context was discussed for psychopathic individuals and/or chronic violent offenders (e.g., Raine, 2019). Thus, we also expected contrasts to reveal specific patterns of activation in the amygdaloid complex for violent offenders and control participants.
The midbrain structure PAG was discussed as one of the most crucial brain areas to be involved in proximal threat-induced flightand fight-behaviours in both animals (e.g., Depaulis et al., 1992) and humans (e.g., Roelofs, 2017). Based on findings by Fehr et al. (2014), we likewise expected midbrain structures, such as the PAG, to be particularly involved in the processing of provocative aggressive scenarios with quasi-realistic proximal threat character when contrasted to social positive scenarios. Respective regional activation patterns were explored for violent offenders and the control group separately.
As working hypotheses based on the data and models provided by Depaulis et al. (1992) and Roelofs (2017), we expected rather caudal, or to say dorsal, midbrain (PAG) activation patterns in violent offenders (facilitating attack behaviours), and rostral, or to say ventral, activation patterns in this region in controls (i.e., inhibiting attack behaviours).

Study participants
Participants were recruited via social media and social services and  First, 2015). Two participants of the VIOL group were accommodated in a correctional facility.
CON participants did not report any history of psychiatric or neurolog-

Procedure
Participants were asked to attentively watch and empathically engage

Behavioural data analyses
Video scenarios were rated after the scanner session according to two scales introduced by Bradley and Lang (1994) Collins, 1994) and smoothed using an isotropic Gaussian kernel (full width half maximum = 8 mm) prior to further analysis. Global effects (grand mean scaling over all volumes) were removed from the functional MRI data, and a high-pass filter (128 s) was applied to remove low-frequency signal drifts.
Design matrices for GLM processing included four regressors in all: one regressor for each of the three 5 s taking emotional scenario categories provocative aggressive, social positive and neutral scenarios, and one more for the between-trial fixation cross displayed between 4 and 8 s. All trial events were modeled by the standard hemodynamic response function.

2.4.2
Contrasting and statistical processing of functional imaging data The preprocessed data sets were analyzed by calculating a t-statistic for different contrasts between stimulus categories. In the present Bradley and Lang (1994); furthermore, group-differences for Reactive-Proactive Aggression Questionnaire (RPQ) (c, left panel) and CFT (c, right panel) values were illustrated study, contrasts between reactive aggressive (A) and social positive (P) scenario-types were reported. Second-level random effects analyses (Holmes & Friston, 1998) were performed on individual contrast images to identify the main task effects by means of a one-sample T-test. A statistical threshold of p < .001 (corrected by an ad hoc determined lowered significance threshold, k ≥ 10 voxel cluster size) was applied to identify significant activation clusters. Statistical MNIcoordinates of peak activations were converted from the SPM8-output into Talairach space with a transformation algorithm, and a reference template based on the Talairach-atlas (Talairach & Tournoux, 1988) was used to determine the respective anatomical regions. Common activation patterns for A versus P scenario types were investigated by conjunction analyses (p < .05; FDR-corrected, k ≥ 5) (see Friston et al., 2005). All above-listed statistical analyses were separately performed for the VIOL and CON groups. For a detailed comparison between second-level random effects analyses with and without CFTvalues as covariate, see Tables S3.1-S3.4 in the Supporting Information Appendix).

F I G U R E 2 Post hoc evaluation of the video-stimuli concerning (a) VALENCE and (b) INTENSITY according to
To inspire future research, exploratory correlations between contrast images reactive aggressive versus social positive (a > p) video scenarios and both RPQ total values (see Supporting Information Appendix 1 ) and valence ratings (see Supporting Information Appendix 2) were calculated for selected regions of interest (ROIs), which is related to emotional stimulus processing (i.e., areas in the limbic system and brainstem).

Self-reported aggression
T-tests for independent samples revealed significantly higher selfreported overall, reactive and proactive aggression measured with the RPQ in VIOL compared to CON participants (all p < .002, see Table 1 and Figure 2c, left panel).

Behavioural data
There were no group differences in INTENSITY ratings. However There were group-specific differences in EC-related VALENCE ratings (for detailed illustration and respective t-values, see Figure 2a).

FMRI data
3.3.1 Reactive-aggressive (A) versus social-positive (P) interaction scenarios At an iso-cortical level, in both groups the contrast A versus P revealed distributed activation patterns in superior parieto-temporal, occipitotemporal, precentral and premotor areas (for illustration, see  Table 2, upper part, and for a detailed list of activation foci, coordinates and t-values, see Table 3). At the sub-cortical level, both groups showed partially overlapping (conjunct) and partially distinct distributed activation patterns in the cingulate gyrus, insula and midbrain (VIOL individuals caudal and dorsal and CON rather rostral and ventral) in the anterior PAG (see Figure 3a, lower panel, left two section views for detailed illustration) regions. Furthermore, both groups showed the involvement of cerebellar brain areas. VIOL individuals specifically recruited parts of the globus pallidus (see Figure 3a, lower panel, section view for detailed illustration), and CON specifically recruited areas in and adjacent to the amygdaloidal complex (i.e., parahippocampal gyrus and uncus, see Figure 3a, lower panel, middle right section view for detailed illustration).

3.3.2
Social positive (P) versus reactive aggressive (A) interaction scenarios At iso-cortical level, in both groups the contrast P versus A revealed distributed activation patterns in premotor, medial parietal (particularly in precuneus) and occipital (particularly in cuneus), occipito-temporal (left fusiform gyrus) and temporal brain regions (for illustration, see Figure 3b, left and right panels; for an overview of recruited brain regions, see Table 2, lower part; and for a detailed list of activation foci, coordinates and t-values, see Table 4). At the sub-cortical level, both groups showed activation patterns in the cingulate cortex (particularly right lateral in the middle part and bilateral in the posterior part) and in the parahippocampal gyrus. VIOL individuals also recruited areas in the postcentral gyrus, the insula and the anterior cingulate.

Supplementary analyses
Tables S3.1-S3.4 in Appendix 3 in the Supporting Information show detailed analyses with and without CFT (i.e., intelligence measures) as a covariate. The results were comparable and did not affect the key conclusions of the present study.
Despite the fact that correlation analyses of fMRI data with external variables can at best be seen as an exploratory approach and need to be interpreted with caution due to the correlative nature of the fMRI data, we also provide data on correlation analyses between contrast images (a > p) in emotionally relevant ROIs and both RPQ (total) and valence rating scores (see Appendices 1 and 2 in the Supporting Information). A further discussion of these data would go beyond the scope of the present study; therefore, these data can be used as basis for meta-analytic research and as an inspiration for future studies on the topic.

DISCUSSION
In the present study, two groups of participants, that is, individuals, who had committed violent offenses, and control individuals, who had not, were examined using fMRI while watching stimuli selected from a stim-  Note. List of brain regions showing at least one focus of activation for the contrast A versus P (upper panel) and P versus A (lower panel), separately for violent individuals (VIOL), controls (CON) and the conjunction (CONJ) between both groups. L means left hemispheric and R means right hemispheric foci in the respective anatomical regions. *) = Within-group-and conjunction analyses produce an inconsistent picture due to the different underlying statistical complexities of the applied algorithms (t-test vs. conjunction analyses). Detailed information about coordinates and t-values are listed in an online supporting document in Tables S1 and S2. Note: Anatomical regions, peak activation t-values and Talairach-coordinates for contrast reactive aggressive (a) versus social positive (p) scenarios, separately for violent individuals (VIOL) and control participants (CON) and conjunction{null} analyses including contrast a > p of both groups; H = hemisphere: L = left, R = right, all statistics p < .001, uncorrected, minimum voxel cluster size k = 10 voxels.

TA B L E 3 Detailed fMRI data for contrast A versus P
but not in the respective conjunction analyses including both groups.
The present fMRI-analyses were focused on contrasts between reactive aggressive and social positive scenarios.

Behavioural data and the validity of the experimental approach
While there were no group differences in intensity ratings of the scenarios, valence ratings differed between groups: VIOL rated reactive aggressive scenarios less negative than CON. This finding might indicate a reduced fear response in violent individuals on a behavioural level, which might result from an amygdala hypoactivation in this group (cf. Lozier et al., 2014;Rosell & Siever, 2015). This interpretation is supported by our neurobiological findings revealed by the contrast reactive aggressive versus positive interaction, indicating amygdala involvement in CON but not in VIOL participants (see Section 4.2). Moreover, there were pronounced, yet nonsignificant differences between groups: VIOL rated neutral scenarios more negative and social positive scenarios less positive in comparison to CON. Thus, VIOL may have experienced nonthreatening social interaction scenarios more negative than nonviolent individuals, which might express a general mistrust toward social interactions possibly due to life-long negative experiences (cf. Afifi et al., 2019). However, VIOL compared to CON produced larger distributions in valence ratings for all scenario-types also indicating more heterogeneous, and thus idiosyncratic, emotional processing.
Summing up, behavioural data confirmed the validity of the applied stimuli. In accordance with the taxonomy of Bradley and Lang (1994), the evaluation of the stimuli revealed higher intensity ratings for social positive and reactive aggressive as compared to neutral scenarios.
Reactive aggressive scenarios were rated most intense. Valence ratings revealed positive ratings for positive, negative ratings for reactive aggressive and ratings in-between for neutral scenarios in both groups.

Neural correlates of reactive aggressive versus social positive scenarios
At a cortical level, we found a large overlap between VIOL and CON in precentral gyrus, premotor, perceptual parieto-temporal and occipitotemporal brain regions. Taken together with the findings by Fehr et al. (2014), the stimulus material seems to validly provoke activation in previously proposed experience-related, context dependent actioncycle systems. Additionally, both groups showed insula and postcentral gyrus involvement, which have previously been associated with mental states of aversion and pain expectancy (Decety, 2010;Fan et al., 2011;Gu et al., 2010). Contrary to our expectations, we did not find a recruitment of right inferior-frontal regions previously discussed as an inhibiting instance (cf. Fehr et al., 2014). Neither did we find Note: Anatomical regions, peak activation t-values and Talairach-coordinates for contrast social positive (p) versus reactive aggressive (a) scenarios, separately for violent individuals (VIOL) and control participants (CON), and conjunction{null} analyses including contrast p > a of both groups; H = hemisphere: L = left, R = right, all statistics p < .001, uncorrected, minimum voxel cluster size k = 10 voxels. support for differential patterns of activation in experience-related action-perception cycles (Fehr et al., 2014)  The following activation patterns were found in the midbrain and brainstem: CON showed rostral midbrain (PAG) activation and activations in the reticular formation (see Siegel & Victoroff, 2009), whereas VIOL showed distributed caudal midbrain (PAG) involvement. The PAG of the midbrain is assumed to be a key structure in the neural processing and initiation of defence-, flight-and fight-related behaviours (Roelofs, 2017). Caudal PAG involvement in VIOL individuals fits in well with the proposed function of this subdivision initiating fight and flight reactions by studies investigating both, humans (Roelofs, 2017) and animals (Depaulis et al., 1992). In contrast, CON showed rostral midbrain activation associated with PAG, which was implicated in the initiation of freezing (Roelofs, 2017). Thus, these results support the notion that differences in PAG recruitment might contribute to violent individuals rather "taking action" instead of freezing in threatening situations, and thus may play an important role in the execution of violent behaviour.
Furthermore, the present data indicate activation patterns related to the amygdaloid complex only in the CON but not in the VIOL group, that is, suggesting hypoactivation rather than hyperactivation of the amygdala in this sample of violent offenders, as had been linked to the proactive aggressive phenotype. This is in line with the findings in similar samples, for example, in chronic offenders (e.g., Raine, 2019), and underlines the possible relevance of amygdala hyporeactivity in the occurrence of severe forms of violence. Importantly, this might be the case even in situations where reactive aggression is provoked. Reactive aggression has previously rather been associated with amygdala hyperreactivity, for example, in a study applying a money-subtraction paradigm in violent offenders (da Cunha-Bang et al., 2017). The diverging findings in violent offenders might be explained by the use of different experimental approaches, which possibly investigate different kinds of reactive aggression (tit for tat versus defence-like aggression in a situation more closely resembling threat). However, the findings still remain somewhat unclear and support the idea of going beyond an either/or distinction between reactive and proactive aggression in order to investigate the occurrence of violence. Moreover, as the activation of the amygdaloidal complex during threatening situations was found to be a prerequisite for projecting to the rostral PAG to initiate freezing (Roelofs, 2017), reduced amygdala activation in VIOL might explain the lack of rostral PAG activation and consequently, a reduced ability to freeze in threatening situations. However, it is also possible that VIOL individuals' amygdala activation was comparably high during both positive and reactive aggressive interaction scenarios and would therefore not reveal neurofunctional differences in the respective con- trast. Yet, VIOL individuals' valence ratings indicated that they perceived aggressive scenarios as negative and positive videos as positive (although the difference between positive and aggressive scenarios was smaller in VIOL than in CON). Moreover, in combination with the revealed PAG activation during the aggressive versus positive contrast, the interpretation of an amygdala hypoactivation during threatening situations possibly contributing to a tendency toward fight and flight reactions seems plausible.
In addition, VIOL seemed to link midbrain behavioural impulses to basal ganglia structures such as the globus pallidus, and therefore might run the risk to perform sub-cortically driven behaviours that are not justified by emotional and/or contextual adequacy via limbic structures such as the amygdala. In sum, our findings concerning the reactive aggressive versus positive contrast give rise to the assumption of a differential activation pattern in individuals with a history of violent offenses in structures involved in the evaluation of emotional stimuli (amygdala) and in the preparation of defence and action (PAG, basal ganglia), which might enhance fast and potentially impulsive behavioural reactions. This might also indicate a reduced fear response in individuals exhibiting violence.

Neural correlates of social positive scenarios versus reactive aggressive
Contrasting social positive and reactive aggressive scenarios revealed highly reliable activation patterns predominantly distributed in medial parieto-occipital (precuneus and cuneus) and parieto-temporal brain regions, parahippocampal areas and premotor cortex. This pattern of activation might reflect a perception-action network related to episodic memories on stereotypic positive scenarios associated with respective pro-social behaviours as already shown by Fehr et al. (2014).
Interestingly, VIOL likewise showed activation patterns in the insula (at other sites compared to those activated by reactive aggressive scenarios in both groups), anterior cingulate and postcentral areas. anterior cingulate cortex activation have been associated with processing salience and prediction errors (Uddin, 2015;Jahn et al., 2014), that is, processing stimuli which stand out or deviate from one's expectations, respectively. To violent individuals, positive social interactions might stand out in which these types of interactions might contradict life-long experiences and learned expectations (see Afifi et al., 2019).
This could represent a neural correlate of the hostile attribution bias, which has been related to reactive aggression (Card & Little, 2006).

Final conclusion
First, the present data regarding the contrast between reactive aggressive and social positive interaction scenarios indicated an involvement of the amygdala in control participants, but not in violent individuals in the context of interpersonal provocation. However, as our findings also hint at a somewhat ambivalent processing style of positive interaction scenarios in violent offenders, when considering our within-group comparisons, we cannot preclude the possibility that amygdala activation is high in violent individuals during both scenarios. Yet, the specific PAG recruitment in violent offenders revealed by the reactive aggressive versus positive contrast renders this interpretation rather unlikely. Thus, our findings point to a hypoactivation of the amygdala (see Raine, 2019) playing a significant role in violence even when the situation is provocative, that is, usually eliciting reactive aggression, which has previously been linked with amygdala hyperactivation. In line with Raine, our findings point to the need to see beyond the theoretical distinction between proactive and reactive aggression in order to take a step toward understanding the complexity of violence in the real life.
Second, in VIOL participants the caudal part, while in CON participants the rostral part of the midbrain, associated with the PAG, was recruited during the processing of reactive aggressive scenarios. Based on the model described by Roelofs (2017) Third, based on the contrast between social positive and reactive aggressive interaction scenarios, insula recruitment in the VIOL group together with larger distribution and lower positivity ratings for social positive interaction stimuli led us to conclude that positive interactions might be experienced somewhat ambivalently by violent individuals, which might be caused by a general distrust in social interactions (cf. Lawrence & Hodgkins, 2009) and result in the tendency to interpret social interactions more negatively and provokingly. Thus, these findings argue in favor of a hostile attribution bias in violent individuals (cf. Card & Little, 2006), which might not necessarily be characterized by amygdala hyperactivation as has been proposed to underlie aggression in the reactive aggressive phenotype (Rosell & Siever, 2015). However, insula and anterior cortex recruitment indicated by this contrast might also point to violent individuals experiencing positive social interactions as more salient in comparison with negative interaction scenarios (cf., Uddin, 2015). This might be due to expectancies or schemata that social interactions turn out negatively and that positive social interactions are unlikely to happen due to repeated past and current negative interpersonal experiences in these individuals (see Taubner et al., 2017).

Limitations and perspectives
Contrary to the idea of a neuro-moral or top-down regulatory dysfunction in violent individuals (cf. Raine, 2019), we did not find differential activation patterns in VIOL and CON at a cortical level. As this might be related to our experimental design, the validity of the present experimental approach might benefit from adding a decisional component.
The sample size of N = 25 violent and N = 21 nonviolent individuals in the study was low, which limits the explanatory power as well as the generalizability of findings.
Future studies should aim for experimental designs that enable the investigation of different conflict-related interactions such as relational, pro-active, instrumental and security-related aggression (cf. Blair, 2010) in order to gain further insight in the usefulness of a distinction between aggression types on the level of the individual or rather identify other relevant factors (e.g., context-dependency, rather than dependency on individual characteristics; cf. Fehr & Achtziger, 2021).
Therapeutic interventions might benefit from training violent individuals in being able to establish a "freezing phase" during situations, which are experienced as threatening, as is promoted in mentalizationbased therapy (Bateman & Fonagy, 2013).