Neural representation of the parent–child attachment from infancy to adulthood

Abstract

Attachment theory is built on the assumption of consistency; the mother–infant bond is thought to underpin the life-long representations individuals construct of attachment relationships. Still, consistency in the individual’s neural response to attachment-related stimuli representing his or her entire relational history has not been investigated. Mothers and children were followed across two decades and videotaped in infancy (3–6 months), childhood (9–12 years) and young adulthood (18–24 years). In adulthood, participants underwent functional magnetic resonance imaging while exposed to videos of own mother–child interactions (Self) vs unfamiliar interactions (Other). Self-stimuli elicited greater activations across preregistered nodes of the human attachment network, including thalamus-to-brainstem, amygdala, hippocampus, anterior cingulate cortex (ACC), insula and temporal cortex. Critically, self-stimuli were age-invariant in most regions of interest despite large variability in social behavior, and Bayesian analysis showed strong evidence for lack of age-related differences. Psycho–physiological interaction analysis indicated that self-stimuli elicited tighter connectivity between ACC and anterior insula, consolidating an interface associating information from exteroceptive and interceptive sources to sustain attachment representations. Child social engagement behavior was individually stable from infancy to adulthood and linked with greater ACC and insula response to self-stimuli. Findings demonstrate overlap in circuits sustaining parental and child attachment and accord with perspectives on the continuity of attachment across human development.

attachment, longitudinal studies, mother–child relationship, ACC

Introduction

Since the discovery of social bonding by Lorenz (1935), the ensuing research of the ethologists (Tinbergen, 1963) and Bowlby’s (1969) adaptation of these findings into a comprehensive theory of human development, the mother–child attachment has been a key lens for understanding human nature, generating extant research across ages, cultures and pathological conditions (Bretherton, 2010; Dykas and Cassidy, 2011; Zimmermann and Iwanski, 2015; Keller, 2016). Among the central propositions of attachment theory is the consistency of attachment across development; the nature of the bond formed between mother and infant defines the trajectory of their relationship and predicts a host of child social, emotional, cognitive and physiological outcomes (Sroufe et al., 1999; Waters et al., 2000; Valadez et al., 2020). In parallel, disruptions to the mother–infant attachment carry long-term negative consequences that are moderated, at least in part, by the consistency of attachment over time (Sroufe, 2005). The consistency hypothesis has been described not only throughout development but also across attachment relationships. Conceptual models (Mikulincer and Shaver, 2005, 2012) and empirical studies demonstrate continuity from the parent–infant to romantic attachment (Selcuk et al., 2010; Göbel et al., 2019) and friendship quality (Freeman and Brown, 2001; Feldman et al., 2013), indicating that the mother–infant bond shapes the individual’s later attachments throughout life (Verhage et al., 2016). Consistency has also been found across measures. Behavioral markers of attachment in infancy (i.e. the ‘strange situation’) have been associated with narrative quality of the parent–adolescent attachment (Hamilton, 2000; Chae et al., 2018) and the adult child’s romantic attachment (Weinfield et al., 2004; Pascuzzo et al., 2013), and the parent’s attachment-related narrative during pregnancy was found to predict the infant’s behavior-based attachment classification (Fonagy et al., 1991) despite a significant difference in measure. The long-term impact of the mother–infant attachment on the adult human appears to draw from its substantial consistency—across the individual’s developmental history, across attachment relationships and across modes of dialogue between partners and measurements of attachment.

More recent studies tapped the brain basis of attachment to describe the neural networks that underpin the parent–infant bond. This line of research typically examined parents’ neural response to own-infant stimuli as compared with unfamiliar infant. To target the attachment context, numerous studies across several labs exposed parents to naturalistic video vignettes of their own parent–child interaction as compared with unfamiliar interaction (Noriuchi et al., 2008; Atzil et al., 2011; Abraham et al., 2014, 2016a; Elmadih et al., 2016). This experimental paradigm is based on two conceptual tenets in ethology and attachment theory. First, that humans form exclusive attachment with their infants, hence their brain reacts distinctly to one’s own child, and second, that in order to study bonding one must go to the natural habitat and observe bonding-related features in the natural ecology (Feldman, 2016). Cumulative evidence from this research has delineated the neural structures that underpin human attachment (Swain et al., 2014; Feldman, 2015b, 2017). The amygdala, ventral tegmental area (VTA) along other subcortical regions of the dopamine network (Strathearn et al., 2009) and hippocampus, areas rich in oxytocin receptors (Boccia et al., 2013; Raam et al., 2017), play a key role in mammalian mothering and are causally involved in bond formation (Insel and Harbaugh, 1989; Oxley and Fleming, 2000; Numan, 2020). In humans, these subcortical regions are similarly implicated in maternal attachment (Strathearn et al., 2009) and are connected via multiple ascending and descending projections with several cortical regions into an integrated caregiving network that sustains human attachment. Among the key cortical structures that underpin attachment are the anterior insula (AI), which supports interoception (Craig, 2003; Zaki et al., 2012; Salomon et al., 2016) and empathy (Singer et al., 2009; Weisz and Zaki, 2018; Lamm et al., 2019), the anterior cingulate cortex (ACC), a higher-order limbic interface of social, affective and representational functions (Bush et al., 2000; Fan et al., 2011), and temporal regions implicated in social understanding, embodiment and mentalization (Frith and Frith, 2003; Mar, 2011; Qin et al., 2020). In anterior temporal regions, dissociable patterns of activation were found in response to one’s parent stimuli, compared to other attachment figures (Laurita et al., 2017). Importantly, while these circuits are not specific to attachment and support multiple affective processes of both social and non-social nature (Eslinger et al., 2021), research has repeatedly implicated these regions in the formation and maintenance of the parent-to-child attachment (Swain et al., 2004, 2007), and, consistent with prior work, we cautiously term it in the following as an ‘attachment network’.

Studies on the neural basis of human attachment lend further support to the consistency hypothesis by showing that other attachments throughout life, including romantic love and friendship, are underpinned by the same neural circuits (Feldman, 2017). The same neural structures that activate in parents in response to attachment reminders also activate when individuals observe their romantic partner, co-parent or close friend (Bartels and Zeki, 2004; Acevedo et al., 2012; Abraham et al., 2017; Laurita et al., 2017), suggesting continuity from parental to romantic to filial attachment. Similar to the parent–infant attachment, the neural basis of other attachments has been linked with oxytocin levels and sensitive inter-partner behavior (Abraham et al., 2017; Scatliffe et al., 2019). Still, very little research assessed the neural underpinnings of the child’s attachment to the parent. For instance, a magnetoencephalography study showed that adolescents exhibited increased activations in a wide cortical cluster of the attachment network, including the insula, cingulate cortex and superior temporal sulcus (STS)/superior temporal gyrus (STG), in response to own mother–child interaction compared to unfamiliar interactions (Pratt et al., 2018). These activations were predicted by mother–child behavioral synchrony in early childhood and decreased in cases of maternal postpartum depression (Pratt et al., 2019), supporting the consistency hypothesis. Similar findings emerged from a study of 10-year-olds’ neural response to own mother’s picture compared with a stranger’s picture in a sample receiving attachment-based intervention in infancy. Findings showed that mother’s picture activated nodes of the attachment network and following attachment intervention children displayed increased activation in the cingulate cortex and hippocampus (Valadez et al., 2020). Similarly, stimuli involving narrative measures of attachment elicited activations in nodes of the attachment network (Buchheim et al., 2006). Yet, despite the reported consistency in the neural network sustaining human attachments, consistency in the adult child’s neural response to cues representing his or her relational history from infancy to adulthood has not been examined.

In this study, we utilized a longitudinal cohort followed for over two decades from infancy to young adulthood. Employing the same paradigm used to test the neural basis of parental attachment (Swain et al., 2007), naturalistic mother–child interactions were videotaped in the home ecology at three age groups: infancy (3–6 months), childhood (10–12 years) and young adulthood (18–24 years). In young adulthood, participants were scanned in functional magnetic resonance imaging (fMRI) while observing these ecological videos spanning three stages in their development in addition to three vignettes of unfamiliar gender- and age-matched dyads. Such longitudinal stimuli spanning their entire attachment history were thought to tap the consistent representations adults construct of their attachment relationship (Goossens et al., 1986; Lewis et al., 2000; Fraley, 2002). Notably, while mothers were not scanned in the current study, we expected that the same regions previously found to underpin the parent’s attachment to the child would also underlie the adult child’s neural response to attachment-related stimuli, including the amygdala, hippocampus, VTA, temporal cortex, insula and ACC (Feldman, 2017), and that these regions would show higher blood oxygen level dependent (BOLD) activity when viewing Self-stimuli across the ages as compared with similar videos of unfamiliar dyads (hypothesis 1). We also examined whether these regions would show consistent, age-invariant responses to attachment-related stimuli from the three age groups presented in adulthood, as compared with unfamiliar interactions (Other) from infancy, childhood and adulthood (hypothesis 2).

Our third hypothesis considered functional connectivity. Studies showed increased connectivity in the parent’s brain in response to attachment stimuli (Leibenluft et al., 2004), and we explored whether Self-stimuli would elicit increased connectivity among nodes that sustain the parent–child attachment. We were particularly interested in the coupling of the ACC with other regions of the attachment network. The ACC is an integrative interface of sensation, cognition, emotion, arousal and neuromodulation (Peterson et al., 1999; Bush et al., 2000; Rolls, 2019; Vassena et al., 2020) and is among the most interconnected hubs in the brain (Margulies et al., 2007), with both upstream and downstream connections (Rushworth et al., 2007; Pavlvlović et al., 2009). The ACC contains evolutionary-conserved reciprocal projections with all subcortical regions of the attachment network, including amygdala, hippocampus, ventral striatum and VTA, which enable the regulation of emotion and motivation within attachment contexts (Beauregard et al., 2001; Burgos-Robles et al., 2019). We hypothesized that the coupling between the ACC and other regions that underpin the parent–child attachment may contribute to consolidation of the long-term representation of attachment relationships (hypothesis 3).

Finally, we tested whether the strength of neural activations in response to Self-stimuli would correlate with the degree of observed maternal sensitivity and child social engagement in the presented interactions. Maternal sensitivity and child social engagement are two key behavioral markers of the mother–child attachment that have shown individual stability from infancy to adolescence (Feldman, 2010) and support child resilience, well-being and adaptation (Ulmer-Yaniv et al., 2017; Feldman, 2020). Since the inception of attachment theory (Ainsworth et al., 1978), maternal sensitivity has been consistently shown to predict attachment security (Meins et al., 2001; Belsky and Fearon, 2002), and activation of nodes in the parental brain has been linked with both maternal sensitivity (Atzil et al., 2011; Kim et al., 2011) and child social engagement (Shimon-Raz et al., 2021). We thus expected that these behavioral markers would correlate with the participants’ neural response to attachment cues and focused on correlations with the ACC and insula (hypothesis 4). Activation of the ACC and insula in the maternal brain was found to correlate with sensitive caregiving (Atzil et al., 2014), and insular activations in the child’s brain were predicted by engaged and synchronous mother–child interactions across development (Pratt et al., 2019; Ulmer Yaniv et al., 2021), suggesting that greater activation of the ACC and insula may correlate with more sensitive and engaged interactions.

Materials and methods

Participants

Participants included 65 young adults (mean age = 20.03 years, s.d. = 2.0, 33 males) who were followed from infancy and participated in two visits at the adult stage, a home visit and a brain imaging session. All participants were healthy, without chronic medical or psychiatric conditions, and completed at least 12 years of education. All children were born to middle-class, two-parent families, and their parents completed high-school education, were above 21 years at the child’s birth, were above poverty line and were screened for psychiatric or psychosocial conditions. All participants were of Israeli–Jewish ethnicity.

Longitudinal study design

Families were recruited in infancy to participate in a longitudinal study on mother–infant attachment and its developmental consequences. Home visits were conducted at three timepoints: infancy (3–6 months), childhood (9–12 years) and young adulthood (18–25 years). Visits were scheduled for the afternoon or evening hours. We were specifically interested in observing naturalistic interactions between mother and child that reflect, as much as possible, their habitual daily interactions. For all interactions, mother and child were videotaped in a face-to-face position. The cameras were placed at 1.2 m from the interacting dyad, and filming tried to capture participants’ faces as much as possible.

At the adult stage, two visits were conducted within a ∼6 week period, including a home visit conducted in the evening hours that lasted ∼3.5 h and a magnetic resonance imaging (MRI) scan session. The study was approved by the Bar-Ilan University’s Institutional Review Board (IRB). Both mother and young adult signed an informed consent and received a gift card of 250 NIS (∼60 USD) for their participation.

Our pre-registered study included two cohorts. Cohort 1 was used for region-of-interest (ROI) definition and included 15 participants who were randomly assigned to this cohort 1. Cohort 2, our main study, included 50 participants with a full dataset. Results reported here are based on data from these 50 participants.

Overall, 81 participants had data from all three stages of the study, including the MRI scan. Ten subjects were discarded due to misplacement in scanner, one for no visual activity and one for bad scan. Four additional subjects were discarded due to problems with the video stream during the scan (video stuck or mistakenly doubled), leaving the final cohort of 65 participants. No differences on demographic information emerged between participants with and without valid data.

Home observation of mother–child interactions

In infancy (mean age = 4.8 months, s.d. = 1.1), mother and infant were videotaped at home, and instructions were ‘play with your infant as you normally do’ for 5 min. In childhood (mean age = 10.9 years, s.d. = 1.2) and young adulthood (mean age = 20.03 years, s.d. = 2), mother and child engaged in a validated conversation-based positive interaction for 7 min (Feldman et al., 2014; Ulmer-Yaniv et al., 2017, 2018). Interactions were used for offline coding and as fMRI stimuli. For the fMRI task, 2 min from the middle of the interaction was selected.

Coding of social behavior

Videos of mother–child interactions were coded with the Coding Interactive Behavior (CIB) manual (Feldman, 1998). The CIB is a global rating system for social interactions with multiple codes for parent, child and dyadic behavior and several manuals for coding social behavior from newborns to adults that are based, whenever possible, on the same codes and similar conceptual principles. The CIB has been validated in hundreds of studies with infants, children, adolescents and adults across multiple cultures and high-risk conditions. The system has good psychometric properties, including construct validity, test–retest reliability and predictive validity (Feldman, 2012, 2021).

In the current study we used the following two constructs:

Maternal sensitivity—was the averaged codes of mother’s acknowledgment of child communication, constant gaze, warm positive affect, warm vocalization, appropriate range of affective expression and consistent style from the three age groups (alpha = 0.92).

Child positive engagement–was the averaged of codes related to social engagement, affection and trust toward parent, positive affect and involvement from the three stages (alpha = 0.86). The construct includes only child, not adult behaviors. In infancy, codes included child positive affect and child social alertness. In late childhood and young adulthood, in addition to these codes, the construct also includes the following codes: child affection to parent, child trust and openness to parent, child involvement in the conversation and child warmth.

Coding was conducted by two coders blind to other information and trained to 90% reliability. Inter-rater reliability, computed for 20% of the videos averaged 94% (k = 0.87).

MRI data acquisition

MRI data were collected using a 3T scanner (SIEMENS MAGNETOM Prisma syngo MR D13D, Erlangen, Germany) located at the Tel Aviv Sourasky Medical Center. Scanning was conducted with a 20-channel head coil for parallel imaging. Head motion was minimized by padding the head with cushions, and participants were asked to lie still during the scan. High-resolution anatomical T1 images were acquired using magnetization prepared rapid gradient echo sequence: repetition time (TR) = 1860 ms, echo time (TE) = 2.74 ms, field of view (FoV) = 256 mm, voxel size = 1 × 1 × 1 mm, flip angle = 8°. Following, functional images were acquired using echo-planar imaging (EPI) gradient echo sequence. TR = 3000 ms, TE = 35 ms, 44 slices, slice thickness = 3 mm, FoV = 220 mm, voxel size = 2.3 × 2.3 × 3 mm, flip angle = 90°. In total, 300 volumes were acquired during the functional scan, and the first three functional volumes, before signal stabilization, were automatically discarded by the scanner to allow for T1 equilibrium (resulting in 277 volumes). Videos were displayed to subjects inside the scanner, using a projector (Epson PowerLite 74C, resolution = 1024 × 768), and were back-projected onto a screen mounted above subjects’ heads and seen by the subjects via an angled mirror. The stimuli were delivered using windows media player software (Microsoft Corporation). The study was approved by the Bar-Ilan University’s IRB and by the Helsinki committee of the Sourasky medical center, Tel Aviv (ethical approval no. 0161-14-TLV). Before participating, participants signed an informed consent according to protocols approved by the ethics committee of the Tel-Aviv Sourasky Medical Center. Participants received a gift certificate of 300 NIS (∼85 USD) for their participation in the scan session.

Attachment paradigm

Consistent with research in parents, the attachment fMRI paradigm presented subjects with a series of video vignettes while lying in the scanner that included two types of matched stimuli ‘Self’ and ‘Other’. The ‘Self’ Stimuli included a 2 min movie of the subject–mother interactions from each age: infancy, childhood and young adulthood. In the ‘Other’ condition, participants viewed an unfamiliar mother–child dyad matched for gender and age in a similar interaction. Before the first video, participants were presented with a fixation cross for 30 s. Between videos, fixation cross was presented for periods of alternately 15 or 18 s. Total time of the paradigm was 13:51 min. Stimuli were tailor-made for each subject in the home setting. The order of presentation was counterbalanced so that half the subjects viewed themselves first, and half watched the unfamiliar dyad first (see Figure 1 for details); however, the order of age conditions presentation was not randomized in order to present a coherent narrative on the maturation of the mother–child attachment from infancy to adulthood. The paradigm has been validated in several previous studies of the parental brain (Atzil et al., 2011; Abraham et al., 2014).

Fig. 1.

Longitudinal research plan and fMRI paradigm. (A) Experimental procedure. Participants were initially recruited as infants, around 3 months of age, participated again as children and later in young adulthood (current study phase). All sessions included a home visit in which a videotaped interaction of participants and their mothers took place. Video vignettes of interactions were used as fMRI stimuli. (B) Experimental paradigm. Participants were presented with six video vignettes of self and other dyadic interactions. Each age (infancy, childhood and young adulthood) was presented as an interaction of the participant and his mother (or a gender-matched stranger). Clips lasted 2 min each and were previewed by a fixation cross for 30 s. Between videos, fixation cross was presented for periods of alternately 15–18 s. Order of self–other was counterbalanced between participants. (C). Pre-registered regions of interest: ACC (red), insula (green), PHG (purple), amygdala (yellow), temporal cortex (light blue) and DMN (fuchsia).

Data analysis

Data preprocessing

Data preprocessing and data analysis were conducted using BrainVoyager QX software package 20.6 (Brain Innovation, Maastricht, The Netherlands). Preprocessing of functional scans included three-dimensional (3D) motion correction, slice scan time correction, drift removal using a temporal high pass filter and spatial smoothing by a full width at half maximum 6 mm Gaussian kernel. The functional images were then superimposed on two-dimensional anatomical images and incorporated into the 3D dataset trilinear interpolation. The complete dataset was normalized into Montreal Neurological Institute (MNI) space, using ICBM-452 template.

Whole-brain analysis

Multi-subject general linear model (GLM) was computed with random effects, with separate subject predictors, in which the different conditions (videos or fixation) were defined as predictors and convoluted with a standard hemodynamic response predictor. Following, a whole brain, two factors (Attachment (Self/Other) × Age (Infancy/Childhood/Young Adulthood)) repeated measures analysis of variance (ANOVA) was performed. Whole-brain maps were created and voxelwise corrected for false discovery rate (FDR) of q < 0.050 (Benjamini and Hochberg, 1995). For visualization of results, the group contrasts were overlaid on a MNI-transformed anatomical brain scan of a single participant.

ROI definition and preregistration

Based on cohort 1 and a priori theory-based selection, eight ROIs were selected and pre-registered at Open Science Framework: hippocampus and parahippocampal gyrus (PHG), amygdala, VTA, ACC, insula, a temporal cluster encompassing the superior temporal sulcus and gyrus, from the occipito-parietal border to the temporal pole, and the default mode network (DMN) as a network. Pre-registration was made following the completion of data collection and is available at https://osf.io/2ndxr/?view_only=ba738b07cad249e0b1f08c2f458ddb35.

Cohort 1 included a group of 15 subjects (mean age 18.93 years old (s.d. = 0.88), 46.7% males, 86.7% right-handed). Fixed effects multi-subject GLM activation maps were used for ROI definition of the amygdala and thalamus. The DMN was defined based on individual functional connectivity maps with seed in the precuneus, which were superimposed to create a 70% mutual probability map. The temporal cortex region was defined based on the pilot map combined with STS region from the Glasser atlas (Glasser et al., 2016). The insula, ACC and hippocampus–PHG were also taken from the Glasser atlas. VTA was defined by three 5 mm spheres based on coordinates from the literature (Murty et al., 2014). Figure 1 shows the ROIs and Supplementary Figure S1 shows cohort 1 Self > Other map.

Psycho–physiological interaction analysis

Classic pyscho–physiological interaction (PPI) analysis (Friston et al., 1997) was done using PPI plugin for BrainVoyager (V1.30) to asses PPI predictors and confounds as follows: pre-registered ACC was defined as the seed region, and psychological conditions were defined as Self > Other, for all timepoints. For each condition, weight was assigned in such a way that the resulting time course will be zero centered (self adult + 1, self child +1, self infant +1, other adult −1, other child −1, other infant −1). Fixation weight was set to zero.

For each subject, the time course of the ACC ROI was extracted, Z-transformed and then convolved with the hemodynamic response function. Then, it was multiplied TR by TR with the task time course (task time course was based on the protocol associated with the data) to create the ACC PPI predictor. Additionally, for each subject, a psychological regressor, based on the associated protocol, an ACC predictor, based on the ACC time course correlation, and a complementary regressor were created. Additional motion correction predictors were added and Z-transformed. The resulting set of four PPI predictors for each subject were used in a multi-subject GLM analysis. The PPI ACC predictor allows to create a group map of voxels that increase their interaction with the ACC for the self conditions compared to other conditions, over and above what is explained by the task itself (Self > Other contrast; psychological component) and by the global functional connectivity of the ACC (physiological component). Multi-subject GLM analysis was restricted to our pre-registered ROIs, using a mask. Then, the ACC-PPI maps were corrected using Monte Carlo cluster level statistical threshold estimator, with 1000 simulations to estimate cluster-level probabilities (Forman et al., 1995).

Statistical analysis

Statistical analysis was conducted using JASP (Version 0.12.1 for windows, JASP Team, 2020), SPSS (SPSS statistics V25, IBM Corp.) and R version 4.0.0 (R Core Team, 2020) with Tidyverse package (Wickham et al., 2019). Null effects were assessed using Bayesian statistics (Keysers et al., 2020). Greenhouse–Geiser correction was used for sphericity violations. Repeated measures Bayesian ANOVA was used to evaluate the evidence for the null effect found using the standard repeated measures (RM) ANOVA analysis. Of note, throughout the analysis, we used the exclusion Bayes factor (BF). As such, a low value for BF signifies support for the inclusion of the effect (i.e. evidence for the effect): BF <0.33 denotes moderate evidence, BF <0.1 denotes strong evidence and BF <0.03 denotes very strong evidence for the inclusion of the model (Kelter, 2020).

Results

As a first step, we examined the overall brain response to attachment stimuli of naturalistic mother–child interactions vs the baseline fixation condition. A whole-brain map comparing the epochs of audio-visual stimulation to fixation of the analysis dataset (50 subjects; Figure 2), is parallel to the map of cohort 1 dataset (15 subjects; Supplementary Figure S2A). As expected, both maps show wide activations in the visual cortex, spreading to the temporal cortex. Additional activations were observed in the DMN, and in limbic regions such as the PHG, and amygdala. Note that the two maps were highly similar despite differences in the number of subjects.

Fig. 2.

Whole-brain multi-subject GLM map of the analysis dataset (N = 50). random effects GLM, whole-brain visual stimuli map, corrected for multiple comparisons at q(FDR) < 0.05.

Next, we examined the experimental factors using two-factorial ANOVA analysis of Attachment (Self/Other) × Age (Infancy/Childhood/Young Adulthood) on whole-brain activity (N = 50). The ANOVA map of the Attachment main effect (Figure 3A) revealed that the main regions showing differential BOLD responses between the Self and Other conditions include the ACC, thalamus and midbrain. Investigation of Self > Other contrast map indicated stronger activity for the Self condition in the middle ACC (Brodmann area (BA) 32), posterior ACC (BA 24), and a large activation cluster extending from the thalamus ventrally to the brain stem. Additional bilateral activations were found in the visual association regions—peristriate cortex (BA 19), supplementary motor area (SMA) (BA 6) and cerebellum. In addition, a bilateral deactivation was found in the hypothalamus (Supplementary Figure S2, Table 1).

Fig. 3.

Whole-brain two factorial repeated measures ANOVA (Attachment (Self/Other) × Age (Infancy/Childhood/Young Adulthood), analysis dataset (N = 50). (A) Main effect for Attachment, clusters in the ACC, thalamus and midbrain. (B) Main effect for time. Clusters in occipital regions, temporal lobe, parietal cortex and limbic regions. There was no Attachment × Age interaction effect. All maps were FDR corrected (q(FDR) <  0.05) on VMR and projected into an inflated brain for presentation purposes.

Table 1.

Clusters activated in Self > Other contrast, analysis dataset

			Peak cluster voxel
Region	BA	Hemisphere	X	Y	Z	Cluster size	t	P
Cerebellum		R	36	−61	−26	641	4.60	0.000031
Extrastriate cortex	19	R	39	−76	7	496	4.99	0.000008
White matter (WM)		R	33	−64	−5	516	4.52	0.00004
Thalamus to midbrain tegmentum		R, L	6	−22	−5	10 490	5.80	0
SMA	6	R	6	17	61	579	4.93	0.00001
Hypothalamus		R, L	0	2	−15	267	−4.18	0.000118
Ventral/medial ACC	24, 32	L	−3	−1	34	1963	5.21	0.000004
Cortex–cerebellum border		R	0	−82	−23	374	4.31	0.000079
Medial ACC	32	R	3	29	25	321	4.02	0.000198
Cerebellum			−3	−58	−20	295	4.52	0.00004
Cerebellum			−18	−61	−38	523	4.41	0.000056
Fusiform gyrus to Extrastriate cortex	19, 37	L	−39	−58	−5	3083	4.68	0.000023
Cerebellum		L	−30	−85	−26	336	4.77	0.000017

L=left; R=right.

The main effect of Age was associated with activation across occipito-temporal regions, mainly in the visual association regions (BA 18) as well as limbic regions (PHG and amygdala) and a parietal cluster (Figure 3B). Random effects GLM maps of the Age contrasts (infancy > childhood, infancy > young adulthood, Supplementary Figure S3) show that visual association regions are activated across all contrasts. There was considerable resemblance between the Adulthood > Infancy and Childhood > Infancy contrast maps, while in the Childhood > Adulthood contrast map, the activations were weaker and sparse. Critically, there was no significant interaction between Attachment and Age at the whole-brain level (Figure 2), lending support to the consistency hypothesis of attachment.

Next, we examined the pre-registered ROIs to test our hypothesis of a network responding to self-related attachment stimuli. Beta weights were extracted from ROIs and analyzed with a 2 × 3 Attachment (Self/Other) × Age (Infancy/Childhood/Young Adulthood) repeated measures ANOVA. We examined the hypothesis that the ROIs would show higher BOLD activity when viewing own attachment videos than when viewing similar videos of others. Indeed, results revealed that across all ROIs, self-related attachment videos elicited stronger BOLD activity, and this difference was significant in all regions except for the DMN. Significant self-related attachment effects were found in the insula (F(1,49) = 4.56, P = 0.03, ηp2 = 0.08, BF = 0.86), PHG (F(1,49) = 5.34, P = 0.02, ηp2 = 0.09, BF = 0.971), temporal cortex (F(1,49) = 5.18, P = 0.027, ηp2 = 0.096, BF = 0.489), ACC (F(1,49) = 9.46, P = 0.003, ηp2 = 0.16, BF = 0.273) and amygdala (F(1,49) = 9.87, P = 0.003, ηp2 = 0.16, BF = 0.093). The extremely low BF in the VTA (F(1,49) = 18.34, P < 0.001, ηp2 = 0.27, BF = 0.003) and thalamus (F(1,49) = 19.81, P < 0.001, ηp2 = 0.28, BF = 3.42 × 10⁻⁵) suggests very strong evidence for the inclusion of the attachment model, whereas in the temporal cortex and insula, the evidence is moderate, and in the PHG, BF is relatively high, suggesting very weak evidence for own attachment model inclusion. As predicted, our pre-registered ROIs showed increased activation for the Self condition compared to the Other condition (Figure 4). Contrary to our hypothesis, the DMN did not show significant differential activation for the Self compared to the Other stimuli, and BF suggested inconclusive evidence for the Attachment model (F(1,49) = 2.69, P = 0.107, BF = 2.054).

Fig. 4.

Attachment main effect in ROIs, analysis dataset, N = 50. All ROIs except the DMN showed Attachment main effect, with higher beta values for the self condition. Note the strong effect in the thalamus. Colored shapes mark the average, whiskers mark SE.

To ensure that the Self–Other effect is not due to a spurious effect related to the random selection of participants to either the ROI definition cohort (cohort 1) or the analysis cohort (cohort 2), we employed a permutation test. Results indicated that the effect is robust and stable across 10 000 different permutations of the subject pool (see Supplementary Materials for the full analysis).

Then, we investigated the consistency of attachment across ages by testing whether these regions would show differential responses to stimuli from the three age groups (Infancy/Childhood/Young Adulthood). The temporal cortex was the only region to show a significant main effect for Age (F(1.5,76) = 6.85, P = 0.004, ηp2 = 0.12, BF = 0.003). Other regions showed strong evidence for exclusion of the age model: insula (BF = 21.71), ACC (BF = 25.57), PHG (BF = 15.48), thalamus (BF = 23.47), VTA (BF = 18.38) and DMN (BF = 12.09); in the amygdala, no such strong evidence was found (BF = 2.22). It thus appears that most nodes of the attachment network were time-invariant, and only one ROI showed sensitivity to age.

Next, we examined interaction effect of Attachment × Age. In the RM ANOVA, there was no significant interaction effect in the whole-brain analysis, as well as in any of the eight pre-registered ROIs (Supplementary Figure S4). This finding suggests that our ROIs did not show differential Self–Other responses for different ages. However, since a lack of significant interaction effect does not provide sufficient evidence for its absence, we employed a repeated measures Bayesian ANOVA to quantify evidence for lack of such interaction (Bayes exclusion factor >10 indicates strong evidence against an interaction between Attachment and Age conditions). BFs in the insula (BF = 11.50), ACC (BF = 81.69), PHG (BF = 12.09), amygdala (BF = 11.39), thalamus (BF = 14.12), temporal cortex (BF = 14.28) and DMN (BF = 9.81) all showed strong evidence for exclusion of the interaction model. In the VTA (BF = 5.96), the evidence was moderate. These findings indicate that in all of our pre-registered regions (apart from the DMN), a significant main effect for own attachment was found, manifested as stronger neural responses to self-within-attachment relationship. However, age-related effects (i.e. age differences in neural response to stimuli) emerged only in the temporal cortex ROI. Critically, there was strong evidence against interaction model across all but one of our ROIs, indicating age invariant brain activity for attachment stimuli.

To capture the ongoing exchange of information between our ROIs during the task, we utilized PPI analysis. Briefly, PPI examines task-specific changes in the correlated activity across different brain regions by identifying voxels that show increased correlation with a seed region within a given psychological context (=condition). The variance explained by the interaction term resulting from the PPI represents explained variance above and beyond the variance accounted for by the main effects of task (Self vs Other) and physiological correlation (e.g. regions that are associated anatomically, activated by a third region, etc.). Hence, significant PPI correlations represent the increasing or decreasing connectivity between regions in response to the task.

In order to investigate such fluctuations, we created a multi-subject random-effect GLM of PPI from an ACC seed to our pre-registered ROIs (for the full list of clusters, see Table 2). The resulting map (cluster threshold corrected, Figure 5) indicated that the left insular cortex had increased correlation with ACC in the Self condition compared to the Other condition. This indicates that during the viewing Self stimuli across the individual’s attachment history, the transfer of information between ACC and insula is increased.

Table 2.

Clusters of PPI with ACC seed Self > Other map, restricted to pre-registered ROIs

				Peak cluster voxel
Correlation	Preregistered ROI	BA	Hemisphere	X	Y	Z	Cluster size	t	P
Positive	Insula	13	L	−27	8	−14	550	4.51	0.00004
	Temporal cortex	22	L	−48	−7	−17	110	3.43	0.001252
	PHG		L	−37	−25	−17	81	3.04	0.003801
	Amygdala	53	R	33	−1	−20	42	2.85	0.006412
	VTA		L	−2	−13	−11	28	3.39	0.001379
	Insula	47	R	40	23	−11	17	2.61	0.011836
	VTA		R	3	−13	−10	16	2.76	0.008136
	Insula	13	R	36	14	−8	7	2.68	0.010094
	Insula	47	R	40	20	−14	4	2.55	0.014035
	Insula	47	R	36	26	−8	4	2.63	0.011298
Negative	Temporal cortex	22	R	63	−34	19	179	−3.08	0.003413
	Temporal cortex	21	L	−63	−36	1	53	−3.04	0.00381
	Temporal cortex	39	R	42	−49	25	46	−2.76	0.008204
	DMN	10	R	3	62	10	38	−2.91	0.005434
	ACC	32	R	6	38	16	26	−2.90	0.005549
	DMN	10	R	6	65	22	23	−2.81	0.007161
	ACC	9	R	3	44	19	9	−2.72	0.008941
	Temporal cortex	39	R	39	−57	28	2	−2.52	0.015008
	ACC	9	R	9	38	19	1	−2.51	0.015254

L=left; R=right.

Figure 5.

Psychophysiological interaction analysis, with ACC seed for Self > Other attachment contrast of 50 participants (analysis dataset). Map was corrected using a Monte Carlo cluster-level estimation. A cluster in the left insula, in orange (the insula ROI is outlined in green), shows voxels that increase their interaction with the ACC (seed region) for the self conditions.

Finally, we explored the relationship between the magnitude of ACC and insular activations during the Self conditions and the independent coding of maternal and child’s social behavior during the interaction. ACC beta values for the Self condition were positively correlated with child social engagement (r = 0.375, P = 0.007; Bonferroni corrected, Figure 6), but not with maternal sensitivity (r = 0.242, P = 0.091). Similarly, insula beta values for the Self condition were moderately correlated with child social engagement (r = 0.280, P = 0.049; uncorrected) (Figure 6), but not with maternal sensitivity (r = 0.237, P = 0.097). These findings suggest that the more the child showed initiation of social communications, positive affect, warmth, motivation and involvement during interactions with the mother across development, the greater the ACC and insula activated in response to viewing Self stimuli. Notably, no significant correlations were found between child social engagement and neural response to the Other condition in the ACC (r = 0.179, P = 0.168) and insula (r = 0.062, P = 0.667), highlighting the associations between these paralimbic structures and attachment representations.

Fig. 6.

Child social engagement correlation with activation during Attachment–Self conditions in the ACC and insula, analysis dataset, N = 50. Pearson’s correlation; ROIs are shown in blue.

Discussion

How individuals construct a stable representation of the mother–infant attachment that guides their social and emotional life has been a core issue across multiple theories ranging from attachment theory to psychodynamic ‘object relations’ models (Winnicott, 1971) and perspectives on socialization and morality (Baumrind, 1967; Hoffman, 1970; Emde, 1992; Maccoby, 1992). Terms such as ‘internal working models’ (Bowlby, 1969) or ‘representations of interactions that have been generalized’ (Stern, 1995) were coined to indicate that familiar mother–child behavioral patterns are abstracted to form representations of the relationship that are first sensorimotor-based and then, with development, gain symbolic and representational components (Bretherton and Bates, 1984) and shape other attachments throughout life. Such far-reaching consensus across theories underlines the centrality of the mother–infant attachment and describes how it gains consistency over time and becomes the most influencing factor in the child’s social life, emotional well-being and stable sense of self.

Neuroscientific investigations into the brain basis of attachment were based on the assumption that the centrality of attachment, its key evolutionary role and its far-reaching consequences must have a distinct signature in the brain. Studies uncovered a group of regions in the human brain that were repeatedly shown to underpin the mother’s attachment to her infant and charted its evolutionary origins in mammalian maternal care and comparability across different human attachments (Swain, 2008; Kim et al., 2016; Feldman, 2017, 2020, 2021). Within this framework, our study is the first to address the brain basis of the child-to-parent attachment within a longitudinal design, and our findings lend further support to the consistency hypothesis and tap a new angle in its application. First, we show consistency in the neural underpinnings of the child’s attachment to the parent with the network that supports the parent’s attachment to the child. Second, we demonstrate consistency in the neural representation of attachment in the adult brain to stimuli depicting his or her attachment relationship from infancy to adulthood, as seen by the findings that most regions of the attachment network exhibited time-invariant activations. Such findings are in line with the aforementioned theories and with research in animal models (Meaney and Champagne, 2001). Our pre-registered ROIs showed greater activations to stimuli representing own mother–child attachment (‘Self’) as compared to a similar attachment relationship of an unfamiliar mother–child pair (‘Other’) in the same regions that activate in the parents’ brain in response to own parent–child interaction compared to unfamiliar interaction (Ranote et al., 2004; Noriuchi et al., 2008; Musser et al., 2012; Feldman, 2015b; Rigo et al., 2019). These included all regions previously shown to activate when parents or romantic partners observe their attachment targets, including subcortical regions implicated in mammalian caregiving, such as amygdala and hippocampus, as well as cortical areas involved in human affect and reward and nodes of the salience network, including the ACC, insula and temporal regions implicated in simulation and mentalization.

In addition to selective activation to own attachment, we found that neural response to attachment stimuli across the 20-year span were age-invariant. Our Bayesian analysis indicated strong evidence for lack of interaction between the Age and Attachment factors in the ROIs, indicating that the greater activation to the ‘Self’ stimuli observed within an attachment context did not differ as a function of age. This is particularly striking as stimuli differed substantially across ages. While the interaction of a 3-month-old infant and that of a 20-year-old adult with the mother differ on any possible sensory and mental attachment–related parameter—looks (body, face), verbal content, temporal distance, self-similarity, affective expression or mentalization—it appears that the identification of the stimuli with one’s own self interacting with the attachment figure trumps these differences, and activations in the attachment nodes were mainly consistent from infancy to adulthood. These findings are the first to show that representation of the primary attachment by adults is consistent across their entire relational history, lending some support to theoretical models postulating that the consistency of attachment is a possible mechanism that underpins the pervasive impact of early attachment on human adults (Dykas and Cassidy, 2011; Keller, 2016). However, a caveat that must be remembered is that our study showed consistency in young adult’s response to stimuli representing different epochs of their attachment history. A full demonstration of continuity across ages would necessitate repeated neural measurements, which are beyond the scope of the current data. Our study highlights the need to include a longitudinal perspective and multiple brain-imaging sessions in social neuroscience studies, a practice that is rarely presented in the literature. Another important argument is that this consistency may relate to domain-general circuits of affect, motivation or salience that participate in the processing of such stimuli; however, the consistently higher activations to own attachment reminders as compared to similar stimuli of unfamiliar pair suggest that greater resources are allocated to attachment representations, possibly as they attract more attention and trigger greater affect and reward.

One possible explanation for the time-invariant representation of attachment across a 20-year span can be drawn from perspectives on the ‘self’ and its neural underpinnings, as our paradigm presented participants with stimuli of the ‘self–within–attachment relationships’ where the attachment context is inseparable from observing the self, a practice consistent with research on the parent-to-child attachment. Studies on the neural representation of the ‘self’ indicate that it combines the bodily (Blanke, 2012; Salomon, 2017), social (Decety and Sommerville, 2003; Yeshurun et al., 2021) and narrative (Christoff et al., 2011; Peer et al., 2015) aspects of the ‘self’ into an integrated whole (Northoff et al., 2006) that, similar to attachment, is considered a stable organizational construct across time and place (Allen and Tsakiris, 2018; Rubianes et al., 2020). It has been repeatedly suggested that the sense of self and its neural foundations develop within the attachment context (Rochat, 2003; Ciaunica and Crucianelli, 2019; Montirosso and McGlone, 2020) which integrates its bodily, social and salience-attributing features (Allen and Tsakiris, 2018; Ciaunica and Crucianelli, 2019; Montirosso and McGlone, 2020). Our paradigm, included several key aspects previously addressed in neuroscientific studies on the self (Northoff et al., 2006; Peer et al., 2015), including visual and auditory self-recognition (Uddin et al., 2005; Qin and Northoff, 2011; Salomon et al., 2012, 2020), autobiographical memory, affective self-relevance (Kelley et al., 2002; Schäfer et al., 2020) and ‘self-relatedness’ factors (mother’s face, mother’s voice, my home). These were encored within the ‘narrative self’, the ‘story’ the individual builds of his or her personal development, attachment relationships and childrearing environment (Gergen and Gergen, 1988; Miller et al., 1990). Our findings accord with perspectives suggesting that the development of the self and its bodily functions are embedded within the mother–infant bond (as initially suggested Freud, 1923 in his article ‘The ego and the id (1923)’) and, over time, gain consistency of representation through ongoing transactions between the consolidation of the self and internal models of attachment (Fonagy et al., 2007; Feldman, 2015a). However, since we did not compare stimuli of the mother–child attachment with those of the self alone from the three age groups, these points are presented as suggestions and should be treated with caution. Much further research is needed to examine the ongoing relationship between the neural underpinnings of the self and those of attachment relationships.

Stimuli of one’s own attachment elicited not only greater activations but also increased inter-region coherence. The PPI analysis indicated that our seed region in the ventral ACC showed increased connectivity with the insula during the observation of own attachment stimuli. The ACC is a key hub that receives exteroceptive inputs from all primary sensory and associative cortices, which are integrated into a higher-order percept (Beckmann et al., 2009). The ACC contains projections to the amygdala, striatum, and VTA and these connections mark stimuli with salience and reward value (Fujiwara et al., 2009). The ACC also links with the hippocampus to create higher-order representations of memorized events (Rolls, 2019). Finally, upstream projections of the ACC to the OFC enable the processing of reward value and the prioritizing of action on the basis of such valuation (Burgos-Robles et al., 2019).

It has long been known (Papez, 1937) that the ACC plays a key role in affective processing, which underpins the formation of attachment bonds. The ACC is implicated in evaluative emotional processing (Esslen et al., 2004), assessment of the motivational value of stimuli (Fujiwara et al., 2009), generation of behavioral emotional response (Etkin et al., 2011) and top–down monitoring of affective information (Carter et al., 2001). The ACC integrates social functions relevant for the formation of human attachments. It contains neurons that respond specifically to cues related to ‘self’ vs ‘nonself’ (Sturm et al., 2013) and integrates sensory, cognitive and affective information into a coherent percept that prioritizes motivation (Porter et al., 2019; Lee and Reeve, 2020), forms predictions, regulates affect (Ochsner et al., 2009), consolidates memories (Restivo et al., 2009; Vetere et al., 2011) and shapes the individual’s mode of operation in social contexts (Krill and Platek, 2009; Vassena et al., 2017). Such integration of functions related to self, affect and social processing renders the ACC as a key region for the representations of attachment and their consistency over time.

Extensive reciprocal projections connect the ACC with the insula, which monitors interoceptive signals from the body. These projections connect insula-monitored interoceptive signals with ACC-controlled exteroceptive cues into an integrative interface that enables embodiment, emotional mirroring, and higher-order representation (Craig, 2009; Seth, 2013; Park et al., 2018). Both the ACC and insula contain layer V Von Economo neurons that afford rapid communication among the two regions as well as with other upstream or downstream targets (Allman et al., 2011) and both contain areas of overlap between self and others’ pain (Corradi-Dell’Acqua et al., 2016; Smith et al., 2021), which mark them as regions for the interplay of connection and separation between self and partner that define the basis of attachment bonds throughout life. Our results on the increased connectivity between the ACC and insula during the processing of self–within–attachment stimuli suggest that this interface becomes more functionally coupled in response to attachment reminders in order to provide the neural underpinnings for the consistent representation of attachment relationships.

Interestingly, our seed region in the ACC was also found to play an important role in representation of the self. Several meta-analyses (Northoff et al., 2006; Qin et al., 2020) indicated that midline cortical structures, including our seed region, are integrated across the ventral and dorsal midline to sustain the multiple dimensions of the self, including proto-self, self-‘qualia’, bodily self, facial self-recognition and mental self (Northoff and Bermpohl, 2004; Uddin et al., 2007; Northoff and Panksepp, 2008; Moran et al., 2009). The ventral part of the midline cortical structures includes an area of overlap among the post-genual ACC, our seed region, the ventro-medial prefrontal cortex (vmPFC) and the medial orbitofrontal cortex (mOFC), and is particularly linked with the narrative self, which matures in the context of the mother–infant attachment (Northoff et al., 2006; Araujo et al., 2013; Salomon et al., 2014). Notably, a recent meta-analysis (Qin et al., 2020) differentiated on the basis of all available neural studies of the self between three levels: the bodily—the interoceptive monitoring of one’s body, the environmental—the self-relevant exteroceptive signals and the mental—the representational dimension of the self. While the insula is represented at all three levels of the self, the ventral ACC is implicated only in the mental level. This level expands the representations of the self beyond the bodily or immediate sensory into autobiographical memory, personal perspective and self-reference. This autobiographical ‘narrative’ self assembles from the affective components of early attachments and integrates into a stable concept. Again, the parallels and reciprocal relationships between the neural underpinnings of self and the brain basis of attachment require much further research, and we present these thoughts only as directions for future research, as our study did not separate the neural response to the self from the neural response to attachment reminders and the associations between self and attachment could not be empirically tested.

Connectivity between the ACC and AI, which subserves key functions such as interoception and affective processing (Craig, 2009; Seth, 2013), appears to play an important role in the parent-to-child attachment. Connectivity of the ACC and AI in the parental brain has been associated with parent–infant synchrony in infancy and predicted the child’s emotion regulation and cortisol reactivity in preschool and lower behavior problems at 7 years (Abraham et al., 2016b; Abraham and Feldman, 2018). The coupling between the ACC and insula in response to attachment stimuli shown in our study anchors the representation of the mother–child attachment in the bodily and non-verbal sensory, as seen by the insular involvement, yet integrates this interoceptive level into adult representation, as seen by our seed region in the ACC (Morita et al., 2014). Insular activations in the maternal brain are thought to provide external-regulatory function for the infant’s emerging ability to recognize his/her own bodily signals and, over time, develop interoceptive representations (Atzil et al., 2018). Furthermore, as part of the sociotemporal brain (Schirmer et al., 2016), the insula monitors the duration and patterns of social events, including the early patterns of mother–infant non-verbal synchronous interactions that later expand into symbolic and verbal exchanges that are individually stable and provide the background for consistent attachment relationships from infancy to adulthood (Ulmer Yaniv et al., 2021).

It is interesting to note that the same brain regions sustaining human attachment overlap with the so-called ‘interoceptive network’, particularly, as recent models on interoception suggest that viewing the self often triggers activation in this ‘interoceptive network’ (Chen et al., 2021). For adults, viewing the self while interacting with the mother as an infant may be an especially strong reminder of caregiving and bodily contact. It has been suggested (Chen et al., 2021) that interoceptive information is first processed in the brainstem nucleus of the solitary tract and then projects to the thalamus, from where it is relayed to higher targets: the amygdala, insula and ACC. As seen in our data (Figure 2), the same nodes of the interoceptive network were found here to differentiate Self from Other’s attachment: thalamus-to-brainstem, amygdala, hippocampus, insula and ACC. This suggests some overlap between the attachment network and the interoception network and raises the possibility that the primary attachment and the experience of caregiving provide the earliest context for the infant’s capacity to first sense, then identify and finally form representations of signals from the body. Further study on the overlap in the brain basis of interoception and attachment may be a fruitful avenue of research and may shed further light on this important topic.

Finally, the magnitude of ACC and insular activations correlated with the degree of child social engagement, which defines the degree of positive involvement, motivation for social connection and initiation of social communication the child exhibits during interactions with the mother across the 20-year span. Child social engagement is an important feature of the mother–child interaction that has received significantly less research as compared with maternal sensitivity, albeit its role in shaping social-emotional competencies and the social brain appears to be just as critical. Child social engagement is an individually stable disposition that develops on the basis of both maternal sensitivity and the child’s temperamental sociality (Feldman et al., 2010; Feldman, 2021). No study, to our knowledge, tracked the expression of children’s social engagement during interaction with the mother across two decades, and thus, our findings that show stability in this orientation from infancy to adulthood are novel and important and likely represent the longest timespan for which such stability in children’s behavioral social orientation is reported. This suggests that features of the infant’s behavior within the first social relationship may persist throughout life, and thus, high-risk conditions associated with dampened child social engagement, such as maternal postpartum depression or high contextual risk, should receive intervention that help mothers increase infant social engagement already in early infancy. Notably, activation of the ACC and insula was not significantly related to maternal sensitivity, the key variable in attachment research that often views attachment from the mother’s perspective (van Ijzendoorn et al., 1995), but with the child’s own behavior as an infant, child and adult. Interestingly, while maternal sensitivity has been repeatedly shown to serve as a buffer against harsh rearing conditions, beginning in late childhood the child’s social engagement charts a unique pathway to resilience, independent of the maternal path (Halevi et al., 2017), suggesting that children’s own social behavior becomes more important beginning in late childhood. Our findings indicate that in adulthood, the neural representation of the mother–child attachment, particularly the paralimbic interface that becomes functionally coupled in response to attachment cues, shows closer associations with markers of the ‘self’ and its contribution to the relationship, including the degree of engagement, valuation and motivation for the interaction, than with the mother’s behavior. These results highlight, again, the ongoing relationship between representations of the ‘self’ and internalization of the primary attachment as they mutually evolve across development and cohere into a time-invariant representation.

Several study limitations should be considered. First, we did not measure mothers’ neural response to the same stimuli, and our ROIs were based on prior research with mothers using the same naturalistic paradigm. Only a design that includes both mother and child can definitively demonstrate that own attachment stimuli trigger similar activations in both mother and child. Second, similar to all neural studies of the self, it is possible that self-related stimuli are allotted more attentional resources than non-self-relevant stimuli. However, this attentional account would also suggest that the novel attachment stimuli from the earlier ages would probably elicit greater attention; still our data show no difference between ages, suggesting that these findings do not stem from differential attentional engagement. Third, stimuli presentation order was counterbalanced for ‘attachment’ (Self vs Other), but not for ‘Age’ presentation order. This stemmed from our desire to present a coherent narrative account of the mother–child relationship from infancy to adulthood and describe the unfolding of the attachment relationship across the individual’s developmental history. Still, the lack of counterbalance in age is a study limitation and should be taken into account when interpreting the Age effect. Additionally, as in all ecological studies, our stimuli varied on numerous visual and auditory properties. Despite these limitations, we found consistent activation across our pre-registered ROI, which speak to the robustness of the effect above and beyond the specific stimuli. Much further research is needed to characterize the development of the neural basis of attachment across ages and relationships, understand its impact on the consolidation of the self and representation of the bodily milieu, and tease apart the impact of culture, context, habit and risk conditions on the maturation of the neural representation of attachment bonds throughout life.

Acknowledgements

We would like to thank the mothers and children for their participation and cooperation.

Funding

The study was supported by the Simms/Mann Chair to Ruth Feldman.

Conflict of interest

The authors declare no competing financial interests.

Supplementary data

Supplementary data is available at SCAN online.

Author notes

Adi Ulmer-Yaniv, Shani Waidergoren, Roy Salomon and Ruth Feldman contributed equally to this study.