Amygdala–midbrain connectivity indicates a role for the mammalian parental care system in human altruism

Abstract

Costly altruism benefitting a stranger is a rare but evolutionarily conserved phenomenon. This behaviour may be supported by limbic and midbrain circuitry that supports mammalian caregiving. In rodents, reciprocal connections between the amygdala and the midbrain periaqueductal grey (PAG) are critical for generating protective responses toward vulnerable and distressed offspring. We used functional and structural magnetic resonance imaging to explore whether these regions play a role in supporting costly altruism in humans. We recruited a rare population of altruists, all of whom had donated a kidney to a stranger, and measured activity and functional connectivity of the amygdala and PAG as altruists and matched controls responded to care-eliciting scenarios. When these scenarios were coupled with pre-attentive distress cues, altruists' sympathy corresponded to greater activity in the left amygdala and PAG, and functional connectivity analyses revealed increased coupling between these regions in altruists during this epoch. We also found that altruists exhibited greater fractional anisotropy within the left amygdala–PAG white matter tract. These results, coupled with previous evidence of altruists' increased amygdala-linked sensitivity to distress, are consistent with costly altruism resulting from enhanced care-oriented responses to vulnerability and distress that are supported by recruitment of circuitry that supports mammalian parental care.

1. Introduction

Why would an individual ever intentionally risk his or her life to benefit a stranger? Risky, life-saving acts on behalf of strangers represent extraordinary manifestations of altruism, broadly defined as a voluntary, costly behaviour aimed at improving the welfare of another individual [1,2]. Altruism can assume several distinct forms, each reflecting distinguishable underlying neurocognitive processes [2–5]. Costly altruism toward strangers, while rare, has been documented across multiple social species, including humans, chimpanzees and dolphins [2,6,7] and is thought to reflect the exploitation of deep and evolutionarily conserved proximal neural mechanisms that evolved to support parental care for distressed or vulnerable young [4,7–9]. At minimum, these mechanisms must support two capacities: to encode cues that signal distress and vulnerability, and to bias the organism toward protective rather than preservative behaviours in response [7].

Recent research has implicated the structure and function of the amygdala in costly altruism [10]. The amygdala is a subcortical neural structure that can rapidly encode non-verbal distress cues [11,12] and is thought to be the point of entry to the parental-care system, through which salient affective input from the sensory cortex and the thalamus is projected to neural networks that support preservative and protective behaviours [7]. Coordinated transmission of information from the amygdala to subcortical regions including the striatum, bed nucleus of the stria terminals, various nuclei of the hypothalamus and periaqueductal grey (PAG) support the generation of caregiving motivation and behaviour following the detection of distress in juveniles or otherwise vulnerable individuals [7,13,14]. Consistent with this, amygdala-mediated sensitivity to distress cues (e.g. fearful expressions) corresponds both to individual variation in altruistic and empathic responding in humans [10,15,16] and to individual variation in parental and alloparental care in other mammalian species [17–19]; among primate species, provision of alloparental care is the single strongest predictor of the frequency of altruistic helping among adults [5].

When distress is detected, selection of a response may be particularly dependent on reciprocal connections between the amygdala and PAG [20–21], an integrative midbrain region that supports various basic survival-related functions that include defensive behaviours, pain modulation, anxiety and reproductive behaviours such as parental care [22,23]. Forebrain projections to the PAG arise primarily from the central nucleus of the amygdala—which also receives reciprocal projections from this region—as well as from the hypothalamus, insular cortex and prefrontal cortex [23]. The PAG plays an essential role in regulating behavioural inhibition in response to amygdala-signalled threat cues [7,24]. This is critical to altruistic responding, which requires overriding defensive responses to amygdala-mediated detection of distress. Rodent studies find that ablations of various sites within the PAG impair normal caregiving behaviour [25–27]. Observationally, activity in the PAG (and amygdala) increases along with protective behaviour in female mice whose pups are under threat [13], and in human mothers exposed to infant cries or images [28]. In humans, activity in this region also increases following induction of compassion or caregiving motivation toward vulnerable or distressed strangers [29–31]. Conversely, unusually low levels of compassion are linked to reduced PAG responsiveness to distress [32,33].

Therefore, we hypothesized that extraordinary acts of altruism may result from enhanced care-oriented responses to distress that are supported by recruitment of amygdala–midbrain circuitry. To test this hypothesis, we evaluated activity and connectivity in this circuit in a sample of extraordinary altruists and matched controls as they responded to depictions of others' distress [34]. Altruists had all donated a kidney to a stranger with whom they had no prior relationship. Costs of these donations include extensive pre-operative screening, post-operative pain, small risks to donors' survival and long-term health and financial sacrifices [35]. Such donations are extremely rare, with fewer than 2000 recorded in the USA through 2015 (data from Organ Procurement and Transplantation Network). Altruistic kidney donors consistently describe altruism as the primary motivation for their donations [36] and behave altruistically across a variety of settings [10,37,38]. Whereas previous findings have identified enhanced amygdala reactivity in this population during passive viewing of fearful facial expressions [10], the present research aimed to directly link caring decision-making in this population to caregiving-associated neural circuits through the use of functional magnetic resonance imaging (fMRI) and structural MRI, including diffusion tensor imaging (DTI). Because altruistic responses to distress are thought to be rapid and intuitive [39], we predicted that enhanced altruistic responding would correspond to elevated responsiveness and functional coupling in the amygdala and PAG following even pre-attentively presented non-verbal distress cues, and this pattern would be supported by enhanced structural connectivity between these regions.

During fMRI scanning, altruists and matched controls (see electronic supplementary material, table S1) completed a decision-making paradigm [34] in which they responded to care-eliciting test scenarios that described vulnerable protagonists (electronic supplementary material, table S2). For example, one scenario describes a young girl who is teased about her appearance in the school cafeteria. Half of the scenarios were preceded by distress cues, specifically preattentively (27 ms) presented backward-masked fearful expressions, which signal vulnerability and distress [4]; the other half were preceded by neutral expressions (electronic supplementary material, figure S1). After reading each scenario, participants reported how much sympathy they felt for the protagonist via button press. (This term was selected as more specific than ‘caring’ and because self-reported sympathy during this task correlates at r = 0.90 with desire to help the protagonist [34].) Distractor scenarios that did not elicit care, because the protagonist did not experience distress, were randomly interspersed with test scenarios and were also preceded by pre-attentive facial expressions. The full set of scenarios was validated in a separate sample of 55 participants, who reliably reported more sympathy in response to the care-eliciting scenarios than the distractor scenarios, F_1,54 = 256.39, p < 0.001, and, consistent with prior results [34], reported the most sympathy when the care-eliciting scenarios were preceded by pre-attentively presented fearful expressions relative to neutral expressions, t₅₄ = 11.46, p < 0.001, corroborating the efficacy of combined depictions of vulnerability and distress in eliciting increased care (see electronic supplemental material). In addition to fMRI scanning, participants underwent anatomical MRI scanning and DTI, as well as neurocognitive testing outside the scanner. This testing included self-report measures including the Interpersonal Reactivity Index (IRI) [40], which assesses facets of empathy, and the Psychopathic Personality Inventory (PPI) [41], which assesses traits inversely related to altruism, including Machiavellianism and callousness (see Material and methods).

2. Results

(a). Functional activation and connectivity

We aimed to identify patterns of neural activation that are associated with increases in altruists' caregiving motivation in response to distress and vulnerability. We conducted contrast tests within left and right amygdala regions of interest (ROIs) to compare blood oxygen level-dependent (BOLD) responses during the 3.5 s epoch in which participants rated their sympathy in response to care-eliciting scenarios that were preceded by masked distress cues (fearful > neutral expressions). We included participants' reported sympathy in response to the scenarios (fearful > neutral expressions) as a covariate to identify regions in which increased activity during this epoch corresponded specifically to increases in other-oriented caring responses. Increasing sympathy for the protagonist following pre-attentive distress cues corresponded to increased activation in the left amygdala in altruists relative to controls (x = −27, y = −3, z = −22, k = 23, t₃₂ = 3.73, p < 0.05 clusterwise corrected threshold; figure 1a,b). A similar cluster emerged in the right amygdala (x = 18, y = 1, z = −23, k = 14, t₃₂ = 2.67, p < 0.05 clusterwise corrected threshold). A corresponding whole-brain analysis was also conducted, which revealed that increasing sympathy was associated with increased activation for altruists in a cluster that extended from the midcingulate and posterior cingulate into the PAG, and clusters in bilateral dorsolateral prefrontal cortex, precuneus, striatum, thalamus, bilateral amygdala and bilateral insula, all p < 0.05 clusterwise corrected threshold (electronic supplementary material, table S3, figure S2). No significant clusters were associated with increased sympathy in controls.

Figure 1.

Functional activation and connectivity. (a) Sympathy difference scores are associated with greater activation in altruists than controls in the left amygdala in response to care-eliciting scenarios preceded by fearful expressions versus neutral expressions. y = −3. k = 23. Colour bar = t value. Results in the right amygdala did not survive in a subsample of participants with lower movement, and, therefore, are not presented here. (b) Plotting the relationship between left amygdala activation during this epoch and sympathy difference scores confirmed a positive linear slope for altruists. Shading represents 95% confidence intervals. Two control subjects fell outside ±2 s.d. from the mean for the left amygdala BOLD analyses, with parameter estimates of −0.58 and 0.46; excluding these two controls, sympathy difference scores continue to be associated with greater left amygdala activation in altruists during this epoch, (x = −27, y = −3, z = −22, k = 13, t₃₀ = 3.17, p < 0.05 clusterwise corrected threshold). (c) Altruists exhibit greater functional connectivity between the left amygdala and left PAG when rating sympathy for care-eliciting scenarios preceded by fearful expressions. z = −10. k = 35 (2 mm³ voxels). Colour bar = t value.

We next aimed to evaluate functional coupling between the amygdala and PAG in altruists and controls following depictions of vulnerability and distress. Functional coupling was assessed through generalized psychophysiological interaction (gPPI) analyses for which the anatomically defined left amygdala was the seed ROI. The physiological variable was created through extraction of the de-convolved time series from this seed. gPPI analysis controls for functional connectivity with the seed region during other task conditions (including baseline), such that the resulting functional connectivity map is specific to the task condition of interest. A contrast test of functional connectivity with a PAG ROI defined using the results of a recent meta-analysis [23] compared functional coupling with the amygdala in altruists and controls as they reported their sympathy toward vulnerable protagonists following pre-attentively presented distress cues. Results revealed that, during the epoch in which participants rated their sympathy, altruists exhibited increased functional connectivity between the left amygdala and left (but not right) PAG (x = −8, y = −28, z = −12, k = 35, t₃₃ = 2.66, p < 0.05 clusterwise corrected threshold; figure 1c). No cluster emerged in the PAG for which controls showed greater functional connectivity with the left amygdala than altruists.

Post hoc analyses examining movement across the task identified one altruist and three controls for whom at least 15% of the total number of task TRs were censored due to greater than 0.5 mm movement; all functional activation and connectivity analyses were repeated following exclusion of these participants in order to verify the reliability of our results; reported functional connectivity results, and functional activation results in the left amygdala, but not the right amygdala, persisted following their removal (electronic supplementary material, table S4).

(b). Structural connectivity

We investigated whether these functional patterns corresponded to differences in structural connectivity between the amygdala and PAG using probabilistic tractography of DTI data. Voxelwise statistical analysis of fractional anisotropy (FA) was carried out via tract-based spatial statistics (TBSS) [42]. Altruistic kidney donors showed increased FA within a cluster of voxels in the left amygdala–PAG white matter tract, relative to controls (x = −26, y = −24, z = −4, k = 5, t₃₇ = 2.91, p ≤ 0.05 threshold-free cluster enhancement; figure 2). No group differences were observed in the right amygdala–PAG white matter tract, and no area showed significantly higher FA in controls than in altruists.

Figure 2.

Structural connectivity in the left amygdala–PAG tract. (a–c) Greater FA in altruists than controls (red-yellow) is displayed on the mean FA skeleton (green) and the left amygdala–PAG structural connectivity region of interest (blue); peak shown at x = −26, y = −24, z = −4, p = 0.05 threshold-free cluster enhancement, k = 5 (1 mm³ voxels), result is thickened using tbss_fill to aid visualization. (d) FA values extracted from the voxelwise cluster result. Contours represent frequency distributions. Boxplots are displayed with dots representing means.

(c). Behavioural findings

Results from the functional imaging task confirmed that care-eliciting scenarios elicited higher sympathy ratings (M = 2.93, s.d. = 0.41) than affectively neutral distractor scenarios (M = 1.57, s.d. = 0.31), F_1,33 = 421.56, p < .001, η² = 0.927. No main effects of group, F_1,33 = 0.60, p = 0.445, η² = 0.018, or expression, F_1,33 = 0.001, p = 0.974, η² = 0.000, or significant interactions, all p > 0.05, were observed. Further, no group difference in sympathy rating difference scores for care-eliciting scenarios preceded by fearful versus neutral expressions emerged, t₃₃ = 1.33, p = 0.193, (altruist M = 0.01, s.d. = 0.21; control M = −0.13, s.d. = 0.39).

3. Discussion

Our results provide the most direct evidence to date that conserved parental-care circuits support costly altruism toward strangers. Extraordinary altruists' experienced sympathy following depictions of vulnerability and distress corresponded to heightened activation in both the left amygdala and PAG; these regions also exhibited enhanced functional coupling in altruists relative to controls during this epoch, as well as increased structural connectivity as measured by FA, indicating that these regions may comprise a network that supports care-based decisions in extraordinary altruists. Together, these findings, which identify consistent patterns across behaviourally linked BOLD activity, functional connectivity and structural connectivity analyses and which are consistent with existing literature regarding the basis of caregiving in non-human animal models, converge to support the importance of a subcortical caregiving system supported by amygdala–midbrain connections in the provision of costly altruism toward strangers.

These findings add to accumulating evidence supporting distinct neural pathways underlying distinct forms of altruism. In contrast to the current findings, for example, reciprocal altruism relies primarily on frontal–striatal pathways that mediate expectations of immediate and future reward [3,43,44] and kin-based altruism relies on limbic and cortical pathways that mediate responsiveness to familiar others [45,46]. Findings also represent an important extension of earlier work showing increased amygdala-mediated sensitivity to fearful expressions in altruists [10]. Sensitivity to others' distress is an important contributor to costly helping, but is distinct and dissociable from the care-based motivation that was the focus of the current study. Altruists' demonstrated increased amygdala and PAG activation in this study was specifically associated with other-oriented concern following cues of distress and vulnerability. Thus, beyond indicating that the amygdala simply supports distress sensitivity, the present results indicate that the connections between this region and PAG support these structures' roles in care-based decisions following the detection of distress. That the strength of amygdala–PAG connections corresponds to care-based decision-making in extraordinary altruists also concretely links human altruism to the parental-care networks long hypothesized to support it [2,7]. Fearful expressions like those employed in this paradigm may recruit parental-care networks particularly effectively, given prior findings that they appear morphologically infantile [47], elicit similar patterns of cognitive and behavioural responses as infants' faces [47–49], and are implicitly associated with infants’ faces [48].

The present findings may assist in understanding a seeming paradox of extraordinary altruism, which is that costly sacrifices for strangers by definition entail substantial risks, yet individuals who engage in such behaviours exhibit heightened sensitivity to fear-relevant stimuli, [10]. How can heightened fear sensitivity be associated with seemingly risk-insensitive behaviour? Reciprocal connections between the amygdala and PAG may provide answers. A critical feature of caregiving is the inhibition of behavioural avoidance via pathways that connect the central amygdala and PAG [50]. Such a switch from avoidance to approach characterizes the onset of parental and alloparental caregiving across species [7]. For example, whereas nulliparous female rats actively avoid infants, towards the end of gestation receptors for care-supporting neurotransmitters like oxytocin increase in the amygdala and PAG, causing cues associated with infantile vulnerability and distress to become highly appetitive [7,51,52] and biasing the organism away from preservative behaviours and toward protective behaviours [53,54]. In sum, the present data suggest that amygdala-mediated detection of distress cues, coupled with PAG-mediated inhibition of preservative behaviours, enable protective caregiving responses to emerge.

Our findings are indirectly consistent with the previously postulated role for the neurotransmitters oxytocin and GABA in caregiving and costly altruism. The amygdala and PAG contain high concentrations of oxytocin [55] and oxytocinergic receptors [56,57]. Endogenous and exogenous oxytocin enhances protective behaviours in humans and other mammals [54,58–60] and modulates functional connectivity specifically between the amygdala and a midbrain region proximal to the region identified here (x = −5, y = −25, z = −18) [61]. GABA may also modulate offspring care behaviours through inhibitory signalling in the ventrolateral PAG, which is the location of highly organized afferent amygdala GABAergic input [24]. For example, in rodents, GABA_A receptor antagonism in ventrolateral PAG promotes licking and grooming of pups while decreasing anxiety and aggression [62].

Observed results emerged despite limitations that included constrained sample sizes that reflect the extreme rarity of altruistic kidney donation. Our sample was broadly representative of the total population of these altruists, but further research is needed to confirm our initial findings. Also, both the amygdala and PAG serve multiple sociobehavioural functions in addition to caregiving, including positive emotional experiences, attention and physiological processes such as pain [23,63]. However, our paradigm was designed to maximize the specificity of our conclusions. Our analytic approach for fMRI data included a covariate indexing reported sympathy such that observed patterns of functional activation are limited to those that corresponded to increases in other-oriented sympathy during the task, rather than alternate responses such as self-oriented distress.

Also as reported previously [10], altruists and controls were distinguished by the factor of self-reported psychopathy most consistently linked to behaviour [64], with altruists scoring lower in Self-Centred Impulsivity than controls, t₃₇ = 2.43, p = 0.020 (altruist M = 118.47, s.d. = 15.51; control M = 133.40, s.d. = 22.10). This difference was driven by group differences in the Blame Externalization, t₃₇ = 2.53, p = 0.016 (altruist M = 19.63, s.d. = 4.00; control M = 25.60, s.d. = 9.50), and Machiavellian Egocentricity, t₃₇ = 2.08, p = 0.045 (altruist M = 34.53, s.d. = 6.00; control M = 39.65, s.d. = 9.00), subscales. As also reported previously, groups did not differ in self-reported empathy, all p > 0.10 for total and subscale scores. The absence of behavioural group differences on trait and state measures of empathy reflects consistently observed limitations of self-report measures of empathy [65], reinforcing the importance of linking self-report variables to variables that are less susceptible to demand and social desirability biases, including real-world behaviour and physiology. For these reasons, the present research focused particularly on neural mechanisms that support caring responses in real-world altruists. The incorporation of a subject population of extraordinary altruists tethers our findings to objectively measured acts of costly altruism toward strangers outside the laboratory. A comparison of such altruistic donors to directed donors might further clarify the specificity of our findings, although directed donations tend to be multiply motivated and, therefore, of less clear relevance to altruistic motivation [66]. Finally, the correspondence among our findings using multiple approaches, including functional and structural imaging, supports the significance of the identified regions in supporting altruistic caregiving. But more precise understanding of the relevant neural processes will require more direct interrogation of the care system, perhaps via genotyping or the use of neurochemical techniques such as ligand binding, which could strengthen conclusions about the molecular-level processes that support such caregiving responses. Results might be further refined by inclusion of PAG localizers in future studies. More direct connections between altruism toward strangers and offspring care could also be made by testing the sensitivity of altruists to images or cries of infants, which have recruited activation in this network in other studies [28]. Considering parental or grandparental status as a moderator variable might be of interest as well.

In sum, the current findings extend existing knowledge about the neural bases of costly altruism toward strangers. Altruistic kidney donation is a costly [35], non-normative [67] behaviour performed to benefit an anonymous, non-kin other, thus meeting the most stringent definitions of altruism [1,2,68]. In interviews, many altruistic donors report that the urge to donate a kidney followed hearing a story featuring a stranger suffering from kidney failure. Our findings are consistent with the possibility that these sympathetic urges in response to depictions of distressed and vulnerable others correspond to the structure and function of a network of regions, including the amygdala and PAG, that serve essential roles in motivating offspring care across mammalian species [4,7,13,14]. Together, these results support the possibility that costly altruism toward strangers, rather than being a wholly inexplicable outcome, may represent an exaptation of the mammalian parental-care system.

4. Material and methods

(a). Participants

Thirty-nine healthy adults between 23 and 56 years old (electronic supplementary material, table S1) took part in this study for monetary payment. Nineteen altruistic kidney donors (seven women) were recruited using mailings and electronic advertisements through local and national transplant organizations. The sample of altruists was limited by the extreme rarity of this behaviour. Only 1265 adults had ever donated a kidney to a stranger in the United States through the end of 2012, when recruitment was completed (0.0005% of the adult US population in 2012; data from the Organ Procurement and Transplantation Network). Because altruists were recruited from across North America, most altruists resided more than a 2-h drive from the university and were provided with airfare and up to two nights' lodging. All altruists had donated a kidney to a stranger unknown to them personally at the time of donation. Sixteen altruists were non-directed donors for whom the recipient was anonymous at the time of donation. The remaining three directed their donations to a specific individual who was known to them at the time of donation but whose need for a kidney they had learned about through, for example, a flier or Internet posting. All donations were verified through independent sources, including transplant centre records or media reports. Using data obtained from the Organ Procurement and Transplantation Network, we confirmed that altruists recruited for this study were representative of the national population of altruistic donors at that time in terms of sex and race (exact ages are not available for the national sample). In addition, 20 healthy volunteers (11 women) were recruited from the local community using fliers, online advertisements, and electronic participant databases including ResearchMatch. Altruists and controls were matched on major demographic variables, including age, IQ, gender, race, handedness, household income, employment status, years of education and parental education (electronic supplementary material, table S1).

Exclusion criteria for all participants included current use of any psychotropic medication, current mood, anxiety, thought, eating or developmental disorder (e.g. depression, bipolar disorder, autism, psychosis, eating disorder), history of head injury or neurological illness, IQ < 80 (as assessed using the Kaufman Brief Intelligence Test, Second Edition) [69], and pregnancy or other contraindications to safe MRI scanning, including metal fragments or implants. One altruist had a current diagnosis of attention deficit disorder, which was not an exclusionary diagnosis. Controls were excluded if they reported having ever volunteered to donate an organ to any individual (not including consenting to become a deceased organ donor) or if they had registered to be a bone marrow donor (in order to exclude potentially similar altruistic behaviour). Twelve altruistic kidney donors were also on the bone marrow registry, thus our groups were distinguished by this measure of costly altruism as well. All study procedures were approved by the Internal Review Board at Georgetown University in Washington, DC, and all participants provided written informed consent before testing.

Two altruists and one control were excluded from functional MRI analyses due to a computer monitor error that affected the timing of the pre-attentive facial expressions. This resulted in an fMRI subsample of 35 participants, 17 of whom were altruists and 18 of whom were matched controls. Within this subsample, altruists and controls did not differ on demographic variables reported for the full sample in electronic supplementary material, table S1, all p > 0.05. These three excluded participants were retained in the DTI analyses, as was one control who did not complete the fMRI task, yielding the full sample of participants. Following the stated exclusions, our fMRI sample was non-identical to the sample of 39 participants reported in our prior study of altruistic kidney donors [10]. As described in Procedures, the fMRI paradigm described here was a distinct task from that reported previously. The present task evaluated caring decision-making following the presentation of masked subliminal (27 ms) fearful facial expressions and short scenarios. The prior study was focused on sensitivity to supraliminal (2 s) fearful and angry expressions.

(b). Procedures

Volunteers completed a 90-min online survey assessing exclusion and inclusion criteria, demographic variables, self-reported psychopathic traits via the PPI Revised (PPI-R) [41], and self-reported empathy via the IRI [40]. Eligible volunteers were screened by telephone to confirm eligibility. Researchers coordinated altruists' travel to and lodging at Georgetown University to enable on-site testing. To ensure groups were matched, eligible controls completed screening that included assessments of IQ, demographic variables, psychological history, medications and MRI compatibility before MRI scanning. After confirmation of eligibility, controls completed cognitive testing and MRI scanning.

(c). Neuroimaging acquisition and task

MR images were acquired with a 3 T Siemens Tim Trio scanner (Siemens Medical Solutions) and a 12-channel phased-array head coil. Functional data were collected using a T2*-weighted echo-planar imaging sequence (46 3.0 mm transversal slices; 64 × 64 matrix; repetition time, 2500 ms; echo time, 30 ms; field of view, 192 mm²; 3.0 × 3.0 × 3.0 mm voxels). The first four volumes of each run were excluded from analysis to account for magnet stabilization. High-resolution T1-weighted anatomical images were also acquired (three-dimensional magnetization prepared rapid acquisition gradient echo; 176 1.0-mm axial slices; field of view, 250 mm²; repetition time, 1900 ms; echo time, 2.52 ms; 246 × 256 matrix). DTI data were collected using two runs of an echo-planar pulse sequence (repetition time, 6300 ms; echo time, 86 ms; 2.5 × 2.5 × 2.5 mm voxels; diffusion directions, 30; b value, 1000 s mm⁻²; number of b value = 0 s mm⁻² images, 5).

Participants completed four runs of the task, each lasting 8 min 22 s. Each run featured 20 brief scenarios that each appeared for 9 s, and were either care-eliciting (i.e. featuring a protagonist who was the target of aggressive or callous behaviour) or not, with 10 of each scenario type included per run (electronic supplementary material, table S2). Scenarios were adapted from a previous version of this paradigm [34]. Each scenario was followed by a 3.5-s prompt in response to which participants reported their sympathy using a 1 (no sympathy at all) to 4 (a lot of sympathy) scale. Participants responded via an MRI-compatible button box with four buttons, each corresponding to a point on the rating scale, such that a response of one required pushing the first button, a response of two required pushing the second button, and so on. Participants' understanding of responding via the button box was tested prior to beginning the task. Each scenario was preceded by four facial expressions presented in the context of a lexical decision task (deciding whether a string of letters formed a real or nonsense word via button press), which participants were told was the focal task, following Marsh & Ambady [34] (electronic supplementary material, figure S1). Word type was presented randomly. Affectively neutral letter strings were selected from the ANEW database [70]. Each string of letters was presented for 1.5 s, and was preceded by either a 27 ms fearful or neutral facial expression [71], which was immediately backward masked by an 80 ms scrambled neutral face. Facial expressions were two males and two females from the Pictures of Facial Affect set [72]. Participants completed four expression–mask–letter string sequences before each scenario. Within each block, all expressions were either fearful or neutral. This task structure resulted in a total of 80 scenarios, 40 featuring care-eliciting protagonists and 40 that were distractor scenarios. Of each type of scenario, 20 were preceded by fearful expressions and 20 by neutral expressions. The task was validated in a separate sample of 55 participants prior to the study (see electronic supplemental material, table S5). Presentation of the four scenario types was randomly ordered within each run. Jittered fixations, randomly varying from 1 to 4 s preceded and followed each scenario–question sequence. Each run concluded with a final 12-s fixation. Runs 1 and 2 were separated from runs 3 and 4 by a 10-min DTI scan.

(d). Analysis of behavioural data

Mean ratings of sympathy were calculated for the four scenario types. The difference between sympathy ratings for care-eliciting scenarios preceded by fearful versus neutral expressions was calculated to index the degree to which the fearful expressions elicited increased sympathy in individual participants, and this was a covariate of interest in fMRI functional activation analyses. Self-reported empathy measured via the IRI was calculated for each of the four seven-item subscales, and self-reported psychopathy measured via the PPI-R was calculated for the major factors of Self-Centred Impulsivity and Fearless Dominance. Group means were compared via independent sample t tests. Sympathy ratings were analysed via repeated-measures ANOVA with group as a between-subjects factor.

(e). Analysis of neuroimaging data

(i). Average functional activation analyses

Functional data were preprocessed and analysed according to the general linear model, using analysis of functional neuroimages (AFNI) [73]. The four runs of the task were concatenated, despiked, motion-corrected and spatially smoothed using a 6 mm FWHM Gaussian filter. TRs with greater than 0.5 mm frame displacement were censored during preprocessing. Functional data were aligned to the anatomical grid, transformed to a Talairach space template (the detailed TT_N27 template created from 27 scans of one individual), and masked with an extents mask to exclude voxels without valid data at every TR for every run, helping to control for false activations. Ten regressors were created to model task events: fearful expression blocks, neutral expression blocks, care-eliciting scenarios preceded by fearful expressions, care-eliciting scenarios preceded by neutral expressions, distractor scenarios preceded by fearful expressions, distractor scenarios preceded by neutral expressions and four regressors for question blocks (corresponding to each of the four preceding scenario types). Fixation trials were modelled implicitly, baseline was modelled by a first-order function, and motion artefacts were modelled using the six estimated rigid-body motion parameters. Boxcar regressors representing the occurrence of each block type were convolved with a canonical hemodynamic response function, scaled to an amplitude of 1.

Group-level analyses were limited to a brain mask defined by voxels with functional activation shared by at least 50% of participants. Cluster size thresholds were calculated for a corrected clusterwise p threshold of 0.05 for this group mask (whole-brain clusterwise corrected p = 0.05: 350 contiguous voxels at uncorrected p = 0.05), and also anatomical left and right amygdala masks (left amygdala clusterwise corrected p = 0.05: 12 contiguous voxels at uncorrected p = 0.05; right amygdala clusterwise corrected p = 0.05: 14 contiguous voxels at uncorrected p = 0.05), using 10 000 Monte Carlo simulations conducted via 3dClustSim in AFNI (October 2016 AFNI 16.3.03 version of 3dClustSim, utilizing a Gaussian plus mono-exponential spatial autocorrelation function). This version of 3dClustSim includes updates to decrease the risk of false positives, in response to Eklund et al. [74]. Amygdala masks were defined using the DD_Desai_MPM atlas, a maximum probability map of a combination of Freesurfer subcortical parcellations and cortical parcellations from the Destrieux [75] atlas created by Daniel Glen and Rutvik Desai in Talairach space for AFNI. MNI coordinates of peak t statistics within significant clusters are reported, transformed from Talairach space. Average group neural activation in whole-brain and ROI analyses were compared via independent sample t tests, with reported sympathy included as a covariate of interest, as described in Results.

(ii). Functional connectivity analyses

gPPI analyses were conducted using the generalized PPI toolbox [76] in SPM8 (Wellcome Trust Department of Cognitive Neurology). Functional images were slice-time corrected, realigned, coregistered to anatomical scans, normalized to 2.0 × 2.0 × 2.0 mm voxel size in MNI space using parameters calculated during segmentation of anatomical scans, and smoothed using a 6 mm Gaussian kernel. Task-specific functional connectivity with the left amygdala was estimated using gPPI analysis [76]. To avoid statistically biasing gPPI analyses, the amygdala seed region was independently anatomically defined using the AAL atlas [77]. A design matrix was created for each participant that included the stimulus time series for each of the 10 conditions and six motion parameters, which were convolved with a haemodynamic response function to create psychological regressors. Non-parametric clusterwise inference was computed via the SnPM13 toolbox (http://warwick.ac.uk/snpm) in SPM8, at an FWE corrected threshold of p = 0.05 at an uncorrected p = 0.05, using 10 000 permutations. Reported results are from the output of these SnPM simulations. PAG ROIs were defined as 5-mm radius spheres centred at the average coordinates for left and right PAG activation (x = ±4, y = −29, z = −12) from a recent meta-analysis of PAG functional neuroimaging studies [23], in order to examine brainstem activation most likely to be associated with PAG function. These thresholds were applied to the ROI independent sample t tests conducted on the gPPI results.

(iii). Structural connectivity analyses

Standard preprocessing was performed using FMRIB's Diffusion Toolbox (FDT) within the FSL software package (http://www.fmrib.ox.ac.uk/fsl/). Diffusion images were skull-stripped, eddy current corrected and visually inspected for artefacts. Diffusion images were registered to T1 structural space using a linear transformation and to MNI152 standard space using a nonlinear transformation.

White matter regions of interest between the amygdala and PAG were created using probabilistic tractography with multiple fibre orientations [78]. Right and left amygdala and PAG regions of interest (same as the functional connectivity analysis) were transformed to subject diffusion space and tractography was run separately for each hemisphere using default parameters (curvature threshold = 0.2, maximum steps = 2,000, step length = 0.5 mm, fibthresh = 0.01, loop check = on, constrained by FA = off). To exclude erroneous voxels, exclusion masks were created to eliminate paths traveling: (1) rostral to the amygdala; (2) caudal to the PAG; (3) dorsal to the PAG; (4) through the hypothalamus (3D box with MNI coordinates: x, y, z = −10:10, −16:6, 0:−20); and (5) through the stria terminalis (box with MNI coordinates: x, y, z = −22:22, −22:14, 4:22). Ten thousand samples per amygdala seed voxel were modelled and results were normalized for the number of samples reaching the target PAG mask (divided by waytotal) and thresholded at 10%. White matter ROI masks were then defined in standard space by including voxels that showed overlap in at least two-thirds of subjects' tractography results.

Voxelwise statistical analysis of the FA data was carried out using TBSS, which helps ensure an objective and interpretable comparison between subjects [42]. Subjects' FA images were obtained with dtifit and subsequently nonlinearly registered to a 1 × 1×1 mm MNI152 coordinate space using the FMRIB58_FA target. A template FA skeleton was then derived from all subjects and thresholded at 0.4. This threshold was found to sufficiently exclude voxels in deep gray matter regions (i.e. thalamus) from being included in the analysis (similar group difference results were observed when thresholding FA at 0.2 and 0.3). Subject FA data were then projected onto the skeleton and tested for group differences using an unpaired t test in the left-hemisphere white matter ROI. Within FSL's randomise tool [79], threshold-free cluster enhancement [80] was applied to correct for multiple comparisons, which used 5000 random permutations to build a null distribution of voxel statistical values and correction threshold was set at ≥95th percentile of the distribution (corrected cluster result is actually the 94.9th percentile, which we considered significant due to our strong a priori hypothesis being based on the functional connectivity results). Cluster results are displayed expanded onto the mean FA image for visualization (figure 2).

Ethics

All study procedures were approved by the Internal Review Board at Georgetown University in Washington, DC, and all participants provided written informed consent before testing.

Data accessibility

Data are available through the Open Science Framework (http://osf.io/z3pvm) [81].

Authors' contributions

A.A.M. designed the research; K.M.B.H., E.M.C., S.A.S., L.M.M., J.W.V. and A.A.M. performed the research; K.M.B.H., K.O. and M.S. analysed the data; K.M.B.H., K.O. and A.A.M. drafted the manuscript. All authors gave final approval for publication.

Competing interests

We have no competing interests.

Funding

This project was supported by John Templeton Foundation grant no. 47861 to A.A.M. and National Institutes of Health/National Center for Advancing Translational Sciences grant no. 1KL2RR031974-01 to J.W.V.