Movement and Emotional Facial Expressions during the Adult Attachment Interview: Interaction Effects of Attachment and Anxiety Disorder

Introduction: Adult attachment is commonly associated with emotion regulation. Less is known about the nonverbal embodiment of adult attachment. Objective: We hypothesized that dismissing attachment is related to less movement and fewer facial expressions of emotions, whereas preoccupied attachment is associated with more negative emotional facial expressions. Moreover, the interaction of attachment and the presence of an anxiety disorder (AD) was explored. Methods: The sample included 95 individuals, 21 with AD without comorbidity, 21 with AD and comorbid major depression (AD-CD), and 53 healthy controls. We analyzed nonverbal behavior during a part of the Adult Attachment Interview (AAI) asking about the family and parental figures. The movements of the interviewees were captured via Motion Energy Analysis. Facial expressions were coded according to the Facial Action Coding System using the OpenFace software. We compared individuals with secure, dismissing, and preoccupied states of mind (assessed with the AAI) with regard to the frequency and complexity of movements and the frequency of the facial expressions such as happy, sad, and contemptuous. Results: As expected, dismissingly attached individuals moved less often and with lower complexity than securely attached. For emotional facial expressions, a main effect of the disorder group and interaction effects of attachment by disorder were found. In the AD-CD group, dismissingly attached patients showed comparatively fewer happy facial expressions than securely attached individuals. Conclusions: Reduced movement specifically seems to be related to dismissing attachment when interviewees talk about significant parental figures. Facial expressions of emotions related to attachment occurred when maladaptive emotion regulation strategies were intensified by a psychological disorder.

Keywords: Diagnostic criteria; Embodiment; Nonverbal behavior; Adult attachment; Anxiety disorders

Adult attachment refers to internal working models of the self and of attachment figures based upon early experiences. These models are thought to influence emotion regulation, interpersonal patterns, and the quality of close relationships in later life [[1]]. Ravitz et al. [[2]] distinguished between categorical and dimensional models of attachment. The Adult Attachment Interview (AAI) [[3]], for example, is a semistructured interview leading to a classification of individuals as secure/free-autonomous (F), enmeshed/preoccupied (E), dismissing (Ds), and disorganized/unresolved trauma (U). A common assessment of attachment-related dimensions is provided by the Experiences in Close Relationships Questionnaire [[4]] that measures attachment anxiety and attachment avoidance related to close relationships. The four-category model developed by Bartholomew and Horowitz [[5]] tried to bridge the 2 attachment models and measurement approaches. One category in the model by Bartholomew and Horowitz [[5]] corresponds to a specific combination of the 2 dimensions. Individuals with a dismissing state of mind, for example, are characterized by low attachment anxiety and high avoidance, whereas individuals with a preoccupied state of mind are characterized by high attachment anxiety and low avoidance. Besides, the attachment model underlying the measure, self-ratings, expert ratings, or (semi-)protective approaches to measure attachment can be distinguished [[2]].

In clinical research, a new diagnostic approach for mental disorders is proposed, which also seems to be attractive for attachment research: using computer vision methods, the nonverbal behavior of patients is assessed and subsequently related to disorder-specific emotion regulation strategies (e.g., [[6]]). Computer vision methods, for instance, allow an automatic, fast, reliable, and valid frame-by-frame coding of movement characteristics (e.g., [[8]]) and emotional facial expressions (e.g., [[9]]) based on digital videos. Because of the automation of coding and the corresponding reduction of personnel expenses and time expenditure, this approach opens up an opportunity for examining larger samples in cross-sectional as well as longitudinal designs [[7]].

For adult attachment, research has shown that attachment anxiety and avoidance are related to different emotion regulation strategies: individuals with high attachment avoidance are prone to deactivating strategies, in terms of low accessibility to negative memories of emotional significance and reduced feelings of anger intensity [[10]]. In close relationships, they have a propensity to avoiding intimacy and depending on others [[12]]. In contrast, for individuals with high attachment anxiety, hyperactivating strategies are typical. They quickly become angry in attachment-relevant situations, often ruminating on anger, and are able to easily access negative emotions and memories. Anger frequently overwhelms other emotions, cognitions, and coping behavior in attachment-relevant situations [[10]].

Roisman et al. [[13]] pointed out that the expression of emotions appropriate to the verbal message is a key feature of a well-adapted affective system. Therefore, from an organizational perspective, it can be assumed that securely attached individuals show emotional facial expressions consistent with their (positive and negative) attachment experiences [[13]]. Accordingly, several studies [[13]-[16]] observed that individuals with a dismissing state of mind tended to suppress emotional facial expressions, whereas individuals with a preoccupied state of mind showed discrepancies between verbal and facial expressions of emotions in situations where the attachment system was activated. Studies using the mirror game paradigm [[17]] found that insecurely attached individuals, who showed infrequent movement synchrony, had less motion complexity and less "free" motion behavior. Clinical studies revealed that frequent behavior matching (movement synchrony) of patient and psychotherapist is associated with higher attachment security [[19]] and lower attachment anxiety [[20]] at the end of psychotherapy.

However, when investigating the relationship between nonverbal behavior and adult attachment, it should be kept in mind that there are similarities between nonverbal correlates among individuals with insecure attachment and patients suffering from mental disorders: Slower and fewer movements, for example, are also correlated with depression [[21]] and anxiety disorders [[22]-[25]]. Moreover, depression seems to be linked to infrequent expressions of happiness [[26]], fewer affiliative expressions, and more nonaffiliative expressions [[21]]. It is obvious that the nonverbal correlates of depression and anxiety overlap with the nonverbal correlates of adult attachment. An explanation for this might be that patients with depression [[28]] as well as with anxiety disorder [[29]] predominantly have an insecure attachment. Therefore, it is not clear whether a nonverbal behavior under study is a correlate of attachment or of the mental disorder. Moreover, an interaction effect between mental disorder and attachment ought to be considered. Buchheim and Benecke [[14]] found that insecurely attached patients with anxiety disorder frequently showed a social smile when speaking about early separation experiences. In contrast, securely attached healthy controls frequently showed a genuine smile. Even negative attachment representations were associated with positive expressions. The presence of a mental disorder possibly intensifies the maladaptive emotion regulation strategies of insecurely attached individuals.

The present study was motivated by the findings that attachment-specific emotion regulation strategies can be observed via changes in nonverbal behavior. However, previous studies have not taken a mental disorder into account as a potential confounder. In addition, some studies did not distinguish between attachment anxiety and attachment avoidance (respectively, between individuals with preoccupied vs. dismissing state of mind), although these seemingly are related to different emotion regulation strategies. Using computer vision methods, our aim was to disentangle nonverbal behaviors related to attachment anxiety and avoidance as well as effects of attachment and mental disorder.

A research project dealing with the psychometric properties of adult attachment measures [[30]] offered the opportunity for this secondary data analysis based upon video recorded adult attachment interviews (AAIs). First, we hypothesized that in situations with an activated attachment system (which is supposed to be the case during an AAI), adults with a dismissing state of mind would react with a downregulation of the affect system in terms of showing less motion behavior with fewer complex movements and fewer emotional expressions (hypothesis 1, H1) and that individuals with a preoccupied state of mind react with a hyperactivation of the affect system and would therefore show more negative emotional expressions (e.g., sadness and contempt) (H2). Furthermore, we hypothesized an effect of anxiety disorder in terms of fewer bodily movements and fewer facial expressions of happiness (H3). Finally, we assumed an interaction effect of attachment and mental disorder on nonverbal behavior (H4) with the emotion regulation of insecurely attached individuals (and the related nonverbal correlates) being intensified by the presence of a psychological disorder.

The presented study is a secondary data analysis. The data stem from a study examining the convergent validity of various instruments available for the assessment of adult attachment [[30]]. The study is in accordance with the guidelines for good clinical practice and was approved by the responsible ethics committee.

The primary study included 175 patients and 143 healthy participants. The patients (with panic disorder [ICD-10: F41.0] and/or agoraphobia [ICD-10: F40.00, F40.01] as primary diagnosis) were recruited mainly from the University Hospital Carl Gustav Carus, Dresden, Germany. Inclusion criteria were age between 18 and 65 as well as sufficient language skills to be able to understand the questionnaires and the interview questions. Exclusion criteria were psychological disorders, such as social anxiety disorder (F40.1), or alcohol abuse (F10–F18.1) as a secondary diagnosis. The healthy participants were recruited in Jena, Germany. Exclusion criteria for the controls were insufficient language skills, the presence of a psychological disorder, and psychotherapeutic or pharmacological treatment during the preceding 12 months. The selected patients and healthy participants were matched by sex and age.

The present study is a secondary analysis which investigated a subsample of individuals. One selection criterion was that the individuals could be assigned to one of 3 groups: (1) healthy controls, (2) patients with panic disorder (ICD-10: F41.0) and/or agoraphobia disorder (ICD-10: F40.0) as primary diagnosis without comorbid disorder, and (3) panic disorder and/or agoraphobia patients as primary diagnosis with comorbid major depression (ICD-10: F33). The allocation ratio was 2:1:1. Matching the participants by sex and age was carried out using the propensity score method. Individuals without a matching partner were excluded. A further exclusion criterion was an interview video unsuitable for the analysis of movements using the Motion Energy Analysis (MEA) [[31]], for example, due to suboptimal lighting conditions or only partial visibility of the participant. The sample of the present study consisted of 53 healthy individuals, 21 patients with panic disorder and/or agoraphobia, and 21 patients with panic disorder and/or agoraphobia and a comorbid major depression (see Fig. 1).

Graph: Fig. 1. Flowchart reflecting sample selection. SCID, Structured Clinical Interview for the DSM-IV; F40, agoraphobia; F41, panic disorder; F3, major depression.

All participants were screened by trained interviewers and supervised by an experienced clinician using the Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders (SCID-I) [[33]]. Furthermore, the participants completed multiple questionnaires. In this study, we only considered sociodemographic variables (age, sex, educational level, and steady relationship) and the Brief Symptom Inventory (BSI) [[34]]. The BSI is a self-rating scale resulting in the Global Severity Index (GSI) that quantifies patients' severity-of-symptoms as a single composite score (Cronbach's α = 0.965).

The AAI [[3]] is a semistructured interview for the assessment of the current attachment representation in adults. The questions mainly address recollections of interviewees' childhood relationship to their parents as well as their experiences with separation, illness, and loss. The verbatim transcripts of the interviews are analyzed by trained and reliable AAI raters (J.B. and A.B.) to classify the speaker's "state of mind with regard to attachment" [[35]] into 4 attachment representations: secure/free-autonomous attachment (F), enmeshed/preoccupied (E), dismissing (Ds), and disorganized/unresolved trauma (U). The AAI has a high test-retest reliability for the 3 main classifications with 78–90% and κ = 0.63 to κ = 0.79 and a high interjudge reliability [[36]]. In the present study, the interrater reliability was κ = 0.43 (n = 32, p < 0.001) with respect to the 4 attachment representations F, Ds, E, and U. However, two-thirds of the interviews were coded by a single rater.

The unresolved attachment representation was recoded as a dismissing or a preoccupied state of mind depending on the second AAI rating by each rater. Otherwise, there would have been 5 subgroups with an n < 5 in the analysis of an interaction between attachment and disorder.

All (potential) participants received oral and written information about the study. Only those who gave written consent were included in the study. Sociodemographic data and psychological measures were assessed prior to admission to treatment, respectively, after inclusion in the group of healthy controls. After inclusion, the attachment interviews were conducted at 2 outpatient clinics (including the interviews of the healthy controls) in small neutrally designed conference rooms. All interviews were performed by a total of 2 trained interviewers (both female, age around 25, education: master's degree in psychology) who both interviewed patients as well as controls.

In the presented video study, we considered nonverbal behavior during 5 interview questions from the early phase of the AAI. The first question (AAI No. 1) addresses the family situation. The second question (AAI No. 3) asks for 5 characteristics of the relationship with the mother. During the third interview question (AAI No. 3a), the participant is requested to explain the last-named characteristic of the mother. Similar to that, the fourth question (AAI No. 4) asks for characteristics of the relationship with the father, and the fifth question (AAI No. 4a) requests an explanation of the last-named characteristic of the father. We obtained 5 video clips per person (95 individuals × 5 questions = 475 clips). The duration of a clip ranged between 0.15 and 2.78 min (M = 1.57, SD = 0.54) and was systematically different in the 5 interview questions (MQ1 = 1.58, MQ2 = 1.63, MQ3 = 1.53, MQ4 = 1.70, and MQ5 = 1.43; F [4, 470] = 3.6, p = 0.007).

We applied the MEA (8, 31, 32) to each video clip to assess the movement of the interviewee automatically frame by frame. A digital video is a temporally ordered series of digital images (so-called frames). Movements are visible by a change in the color intensity of the pixels. The MEA algorithm converted each video frame to a gray scale and computed the difference between video frame t and frame t + 1. The motion energy of a difference frame is the number of pixels with substantial change of gray intensity comparing a frame with its previous one. As threshold for a substantial change, we used 9 of 256 possible gray intensities. Moreover, only pixels with substantial change in a predefined region of interest (ROI) were considered.

We standardized the resulting motion energy time series (METS) using the size of the ROI so that zero means no motion and 100 means that all the pixels of the considered ROI are activated. This increases the comparability of the time series of different individuals with differently sized ROI. Then, the METS were slightly smoothed using a moving median with a bandwidth of 5 neighboring time points. Based on the smoothed METS, we computed movement frequency and complexity for each clip according to Grammer et al. [[32]]. We implemented the MEA, the standardization and smoothing of METS, and the computation of movement parameters in MATLAB (free download at https://github.com/10101-00001/MEA/). For illustration, in Figure 2, a part of a standardized motion energy time series is shown. Grammer et al. [[32]] called short sequences in which the motion energy is larger than a threshold "bursts." The threshold was 0.1 (i.e., >0.1% of the ROIs must be active to rate an image change as caused by body movement) and separates body movements from very small image changes caused by noise. The length of a burst is the difference between the ending and starting points in time. The size of the burst is the area below the curve. Outcome variables of the present study are movement frequency (points in time with a motion energy larger than the threshold divided by the length of the concerning video clip) and complexity (average number of peaks within a burst) while answering an interview question.

Graph: Fig. 2. Part of a motion energy time series including a burst and showing the characteristics of a burst.

Facial expressions were coded according to the Facial Action Coding System (FACS; 38). First, for the recognition of (a set of) action units (AU), we used the OpenFace software [[9]] which is based on computer vision and machine learning algorithms. The validity of the program was confirmed in terms of a correct identification of an AU even when the person slightly looks down or to the side [[9]]. In the second step, the AU coded by OpenFace were recoded frame by frame into the facial expressions happy, sad, and contemptuous according to FACS [[38]]. Other expressions could not be coded due to the lack of the necessary AU. A fourth measure refers to the general facial activity (the total score is the sum of frequency of the 3 facial expressions). Finally, we computed the frequency of each expression relative to the length of the video clip. Zero means that the facial expression was not observed. 100 means that the facial expression was shown during the entire time of the video clip.

All questionnaires and movement data for the participants of the secondary analysis were completed. For 54 of the n = 95 × 5 = 475 video clips (11.4%), facial expressions could not be coded. This was mainly caused by suboptimal head positions such as looking down so that the facial expressions were not visible. When the facial expressions could not be coded for >70% of the frames of a video clip, the frequency measures for happy, sad, and contemptuous emotions of the corresponding video were set to missing. Missing data were not imputed, since the applied hierarchical linear models allow unbiased parameter estimates under missing random assumptions [[39]], and Little's missing complete random test (considering age, sex, disorder group, attachment group, interview question number, movement frequency of video clip, movement complexity, and frequency of the facial expressions happy, sad, and contemptuous) was not significant (χ2 = 7.136, df = 5, p = 0.211).

First, we described the analyzed sample depending on the disorder group and on attachment. Associations between disorder groups (respectively, the attachment representation) and sample characteristics were investigated with the analysis of variance (ANOVA) or χ2 tests.

Next, we applied a generalized linear mixed model (GLMM) for repeated measures on the nested data (5 video clips per individual). We used a log-linear regression because the dependent variables (movement frequency, movement complexity, and frequency of the expressions happy, sad, contemptuous) are count data and Poisson-distributed accordingly. We included main effects for age, sex, attachment group (categories: secure, dismissing, and preoccupied), and disorder group (healthy controls, patients with anxiety disorder without comorbid disorder, and patients with anxiety disorder and comorbid depression) as well as the interaction between the attachment and disorder groups. For the level-1-residual variables, we assumed a variance-covariance structure of the diagonal type (meaning that the interview questions can have different error variances). Furthermore, on level 2, we included a random intercept (meaning that the nonverbal responses within a person can be different from one interview question to another).

Based on the regression models, we estimated regression-adjusted group means for the attachment groups and the disorder groups. If a main effect for the attachment representation (respectively, the disorder group) was found, we compared the group means pairwise using contrast tests. For these tests, we applied a sequential Bonferroni correction of p values.

Tables 1 and 2 show the descriptive statistics for the primary study, the sample of the secondary analysis, and list the examined groups. Compared to the sample of the primary study, we found that the examined sample included more female participants (χ2 = 4.005, p = 0.045, V = 0.112). Regarding age, education, distribution of attachment representations, and impairment (GSI), no significant selection effects were found. Comparing the 3 disorder groups examined in this video study, we found that the healthy controls had the lowest impairment (η2 = 0.573, p < 0.001). Furthermore, there was a significant association between attachment representation and education (χ2 = 7.336, p = 0.026, V = 0.278). Securely attached individuals had the highest proportion of individuals with high school education. Regarding the other baseline variables, no significant differences were found between the attachment groups.

Graph: Table 1. Descriptive statistics for the disorder groups

Graph: Table 2. Descriptive statistics for the subgroups with different states of mind related to attachment

In the next step, the associations between nonverbal behavior, attachment groups, and psychological disorder were examined. The statistics of GLMM are displayed in Table 3. Age and sex had no significant effect on movement characteristics and facial expressions. The attachment group was a significant predictor for the frequency of body movements but not for movement complexity and emotional facial expressions. The frequency of the facial expressions happy and sad could be significantly predicted for the disorder group. Interaction effects of attachment and disorder groups were found for the facial expressions happy and contemptuous.

Graph: Table 3. Statistics of the generalized linear mixed model

Table 4 lists the regression-adjusted averages of the attachment and disorder groups. Pairwise comparisons of the means revealed that individuals with a dismissing state of mind show less movements than those with a secure (here and in the following, c means a contrast, p values are corrected according to Bonferroni: c = 13.8, SE = 4.9, p = 0.015) and a preoccupied state of mind (c = 11.2, SE = 4.6, p = 0.028). Regarding the disorder groups, we found that patients with anxiety disorder with comorbid depression more often showed the facial expression "sad" than the healthy controls (c = 3.0, SE = 1.0, p = 0.013) and more often the facial expression "happy" than the patients with an anxiety disorder without comorbidity (c = 0.8, SE = 0.3, p = 0.010). Accordingly, we observed facial activity (the total score) for patients with comorbid depression more often than for healthy controls (c = 5.7, SE = 1.9, p = 0.008) and patients with anxiety disorder without comorbidity (c = 4.3, SE = 2.0, p = 0.056, only on a trend level).

Graph: Table 4. Regression-adjusted means of movement characteristics and facial expressions depending on the disorder group (above) or attachment group (below)

As mentioned above, we found significant interaction effects between attachment and disorder groups for the frequency of the facial expressions happy and contemptuous as well as for the total mimic activity. For these variables, we compared the regression-adjusted means of individuals with autonomous, dismissing, and preoccupied state of mind separately for each disorder group. When applying a Bonferroni correction, only in the group of anxiety patients with comorbid depression, we found significant differences between the attachment groups: securely attached individuals showed more happy facial expressions than individuals with a dismissing state of mind (c = 2.5, SE = 1.0, p = 0.034). Accordingly, securely attached individuals had a higher facial activity (the total score) than individuals with a dismissing state of mind (c = 15.7, SE = 5.9, p = 0.025) in this disorder group.

Besides testing our hypotheses, we examined the length of the answer to rule out that it biased our results related to movements and facial expressions of emotions. The average length of an answer over all 5 interview questions and all participants was M = 94.4 s (SD = 32.4, range: 9–167 s). When applying the HLM used for the analysis of movements and facial expressions, we found no significant effect for sex, age, attachment, disorder group, and the interaction of attachment and disorder (all p > 0.4). However, the variance of random effect was significant (Z = 2.52, p < 0.05) indicating that the answer length is different from question to question. Furthermore, we used the answer length as a predictor in the analyses of movements and facial expressions and found no significant effect (all p > 0.099). Therefore, we did not consider the answer length in the main analyses.

This study examined nonverbal behavior in situations when the attachment system was supposedly activated, that is, during a segment of the AAI. Video frame by video frame, motion behavior, and facial expression of emotions were coded with computer vision methods. We compared adults with free-autonomous, dismissing, and enmeshed/preoccupied state of mind with respect to attachment classified using the AAI. In contrast to recent studies, we controlled for the presence of an anxiety disorder and comorbidity of depression in order to disentangle the effects of attachment and mental disorder.

As hypothesized (H1), individuals with a dismissing state of mind showed less movements and less movement complexity during the attachment interviews than individuals with a secure attachment representation. These results are in line with Feniger-Schaal and Lotan [[40]], who studied movements during a mirror dance game. In their study, attachment anxiety was associated with infrequent directional movements. Likewise, Talia et al. [[41]] observed less contact-seeking behavior (e.g., avoiding emotional proximity) among dismissing patients, and Petrowski et al. [[42]] reported that insecurely attached individuals need more attention from the attachment figure to regulate their negative emotions during exposure to social stressors. The findings of the present study suggest that deactivating of emotion regulation is mainly triggered by the first interview questions of AAI. These processes are measurable in terms of suppression of body movements and less complex body movements.

In contrast to our expectations, the frequency of specific facial expressions did not differ between the attachment groups. In particular, for individuals with a preoccupied state of mind, we could not observe more negative facial expressions (H2). In contrast, Mikulincer and Orbach [[10]] revealed that in individuals with preoccupied attachment, anger frequently overwhelms other emotions. Accordingly, Roisman et al. [[13]] reported that individuals with a preoccupied state of mind frequently showed facial expressions of negative emotions. A reason might be that Roisman et al. [[13]] considered the entire AAI, whereas our study focused only on the first 5 interview questions regarding the relationship to mother and father. Moreover, Roisman et al. [[13]] did not statistically control for the presence of a mental disorder, as we did in the present study. Despite our findings, we still believe that individuals with preoccupied state of mind have a hyperactivation of the emotion regulation in situations with an activated attachment system. However, it seems that corresponding nonverbal correlates are not visible during the first part of the AAI. Similarly, Dozier and Kobak [[43]] observed no relationship between the attachment dimension deactivation/hyperactivation and a rise in skin conductance during the interview answers of the first 4 AAI questions. However, individuals with preoccupied state of mind showed physiological responses when the interviewer addressed separation, rejection, or changes in relationships. Future studies should examine the nonverbal behavior during the entire AAI and try to identify which interview question leads to strong nonverbal responses according to the hyperactivating emotion regulation of the interviewee. Therefore, the presence of a psychological disorder should also be considered, since in the present study emotional facial expressions were mainly related to the disorder and not to the attachment group.

In line with many other studies [[22]-[25]], we found that the presence of a mental disorder affects the frequency of facial expressions of emotions (H3). Most of the differences were observed in the group of patients with anxiety disorder and comorbid depression. In concordance with our expectations, this group, for example, showed more sad facial expressions than the control group. At the same time, we found no main effect of the attachment group on the frequency of emotional facial expressions. It seems that the comorbid depression has a stronger impact on the affective system (and as a result on facial expressions during the AAI) than the attachment. This finding corresponds with physiological experiments with positive, negative, neutral, and (socially) threatening images as stimulus material, whereas the skin conductance reactivity (SCR) did not differ between healthy controls and patients with anxiety disorder; patients with anxiety disorder and comorbid depression showed a significantly more intense SCR [[44]]. However, we did not find a main effect of mental disorder on movement characteristics. This is contrary to Girard et al. [[21]], who found a correlation between depressive symptoms and reduced head movements. Similarly, Renneberg et al. [[27]] observed less emotional facial activity in depressed patients watching movies. The difference can be partly explained by the fact that none of these studies controlled for attachment. Also, the context was different (e.g., watching movies vs. attachment interview). Therefore, future studies should examine the nonverbal behavior of patients in different contexts. Altmann [[31]] found, for example, that 2 interacting students synchronized their body motions more often in positive situations than in negative ones. This possibly holds also for nonverbal behavior related to attachment.

Furthermore, in concordance with Buchheim and Benecke [[14]], we found an interaction of the attachment and the disorder group (H4). In the subgroup of patients with anxiety disorder and comorbid major depression, securely attached individuals showed happy facial expressions more often than individuals with a dismissing state of mind. We assume that some nonverbal correlates of adult attachment emerge only when the maladaptive emotion regulation is intensified by the presence of an anxiety disorder. Future research should deepen the understanding of the attachment-disorder interaction. The presented study suggests that there is much more interaction between attachment and depression rather than between attachment and anxiety.

Since adult attachment is a predictor for psychotherapy process [[46]] and outcome [[47]], therapists need to consider the attachment representation of the patients [[41]]. However, the AAI (including conduct and coding) is highly time-consuming and require long training. Therefore, new methodological approaches for psychotherapists would be helpful in order to identify the attachment representation based on nonverbal behavior in psychotherapy sessions. Our study showed that assessing adult attachment based on the nonverbal behavior is a promising approach.

Notwithstanding, one of the limitations is that the primary study did not aim at frame-by-frame coding of the videos. As a result, the recording setting was not standardized (e.g., different camera angle and zoom). This technical aspect was one main reason for the sample reduction, from N = 318 (primary study) to N = 95 subjects (presented secondary analysis). Because of the selection of the participants, the results of the secondary analysis cannot be generalized to the sample of the primary study. Despite the comparatively large sample size of N = 95, small subsamples resulted for the examination of the interaction between the attachment representation and the disorder group. Furthermore, we considered only the first 5 interview questions of the early phase of the AAI. The attachment system might be activated to a larger degree during the later phases of the interview (e.g., when rejection or changes in relationship are addressed; [[43]]). A further limitation is the comparatively low interjudge reliability of the AAI (κ = 0.43 based on n = 32). Moreover, the length of the answer to the interview question might have an impact. Therefore, we considered the nonverbal behavior in relation to the length of the answer and could show that the results related to movement and facial expression variables were independent of the length of the answer and that neither the attachment nor the disorder group affected the length of the answer.

The present study has also several strengths. The participants were selected carefully based upon standardized clinical interviews. The sample size (N = 95, respectively, 475 video clips) was comparatively large. Movements and facial expressions were coded frame-by-frame (25 measures per second) with highly reliable and objective computer vision methods. The disorder groups were matched regarding age and sex, since randomization was practically impossible. The associations between nonverbal behavior and attachment were examined using GLMM, controlling for the disorder group, age, and sex as possible confounders and taking the nested data structure into account (multiple interview questions per person). Hence, we were able to exclude the presence of an anxiety disorder (healthy controls, anxiety patients, and anxiety patients with comorbid depression), age, and sex as reasons for the differences in nonverbal behavior when comparing attachment groups.

Overall, our findings suggest that adult attachment is not only related to internal working models and emotion regulation. In a situation with an activated attachment system, dismissing attachment is "expressed" by reduced body movements, which can be understood as a nonverbal correlate of a deactivating emotion regulation. In contrast, facial expressions seem to be more affected by the presence of a mental disorder than by adult attachment. However, we found multiple interaction effects suggesting that mental disorder and adult attachment cannot be completely disentangled. The nonverbal behavior during the attachment interviews seems to have been the result of emotion regulation, incorporating facets of mental disorder and attachment, which can amplify or weaken each other.

Future studies should deepen the understanding of the interplay between attachment and mental disorder by considering additional disorder groups and studying nonverbal behavior experimentally in various situations (e.g., an attachment interview and conversations with anxiety-related contents). In addition, the interplay of the nonverbal behavior and physiological parameters as well as the influence of the interviewers (e.g., effects of nonverbal synchronization of conversation partners [[48]]), should be taken into account.

The authors declare that the research was conducted ethically in accordance with the World Medical Association Declaration of Helsinki. Ethical approval for the study was obtained from the Ethics Committee of Jena University Hospital, Jena, Germany (ID 3060-02/11). Informed consent was obtained from the participants. The participation in the study was voluntary.

The authors have no conflicts of interest to declare.

This study was supported by the Deutsche Forschungsgemeinschaft (DFG), FKZ PE 1804/2-1 and STR 306/22-1.

The present study is a secondary data analysis. The primary study was designed by Katja Petrowski and Bernhard Strauss. The recruitment of the entire sample, interviewing and recording of the interviews, and the questionnaire survey were conducted by Sashi Singh and Susan Schurig. For the secondary data analysis, Uwe Altmann formulated the research goals, planned the study, developed the study design, and managed the data. Catharina Friemann, Theresa S. Frank, Mareike C. Sittler, and Désirée Schoenherr were responsible for the coding of nonverbal behaviors. Uwe Altmann conducted the statistical data analyses and wrote the first draft of the manuscript. All authors read, commented, and approved the final manuscript.

Bernhard Strauss and Katja Petrowski contributed equally to this manuscript.