Subtypes of Aggressive Behavior in Children with Autism in the Context of Emotion Recognition, Hostile Attribution Bias, and Dysfunctional Emotion Regulation

Based on the assumption that a hostile attribution bias mediates the relation between misinterpretation of emotional expressions and aggressive behavior (classical model; Crick & Dodge, 1994, 1996), the present study aimed at identifying the additional impact of dysfunctional emotion regulation expressed as lability and negativity in children with autism spectrum conditions. In reference to the modified SCIP models by Lemerise and Arsensio (2000), we hypothesized (H1) that lability-negativity would predict hostile attribution bias and also would have a direct, and positive impact on the presence of aggressive behaviors (emotion model; compare Fig. 1). In alignment with Farmer and colleagues (2009, 2016), we view aggression as a multifaceted phenomenon expressed by different subtypes of aggression, which potentially have different responses to treatment and prognoses (Connor et al., 1998). We, therefore, aimed to explore interrelated predictors in physical acts of aggression and more complex forms (verbal aggression and covert aggression, see Table 1 for examples) separately in preschool and primary school children with autism and normal intellectual functioning (IQ ≥ 70). Finally, due to the importance of facial expressions as a modality of social judgment (Frith, 2009) and social-cognitive information processing, we used diminished facial emotion recognition as a potential predictor of hostile attribution bias (compare Lemerise & Arsenio, 2000; Russo-Ponsaran et al., 2015).

The present study was part of a registered six-week multicenter, randomized, pragmatic clinical trial testing a tablet-based intervention in children with autism (Kirst et al., 2020; DRKS-ID: DRKS00009337; Universal Trial Number (UTN): U1111-1175–5451). Since no TD children participated in the trial, no comparison group was available for the current study.

Out of 184 screened children with ASC, 82 children were eligible for the RCT trial, from which 60 children (50 males) between 5.0 and 10.11 years (M = 8.0 years, SD = 1.6) fulfilled the sample inclusion criteria. A power analysis revealed that this sample size is sufficient to detect an expected effect size of hostile attribution bias on children’s aggressive behavior of Cohen’s d = 0.33 (see meta-analysis by Verhoef et al., 2019) with 80% power (1- β) at a two-sided 5% α level and emotion regulation as an additional predictor. The inclusion criteria were (1) complete testing data in predictor variables (emotion recognition, emotion dysregulation, hostile attribution bias), (2) intellectual functioning within the normal range (IQ ≥ 70) as assessed by a composite score of the Raven's Colored Progressive Matrices intelligence test (2002) and by the Peabody Picture Vocabulary Test, 4th revision (Dunn & Dunn, 2015), and (3) a clinical consensus ICD-10 (WHO, 1994) diagnosis of childhood autism, Asperger syndrome, atypical autism or pervasive developmental disorder not otherwise specified (PDD-NOS). Diagnosis was established by specialized and experienced multi-professional teams using a variety of measures and clinical judgment. Results of the Autism Diagnostic Observation Schedule (ADOS-G/ADOS-2; Lord et al., 2000, 2015; Merkle et al., 2016) were provided by caregivers or clinicians for 53 participants, who were eligible for the present study. To confirm the ASC diagnosis, the Autism Diagnostic Interview-Revised (ADI-R) short version (Hoffmann et al., 2015) was administered to all participants, and autism symptomatology was further assessed by using the Social Communication Questionnaire (Rutter et al., 2006). Interfering neurological/medical conditions (except for well-treated epilepsy) were ruled out by parental report. The subscale “Aggressive Behavior” of the Child Behavior Checklist (CBCL 4/18; German version; Döpfner & Arbeitsgruppe Deutsche Child Behavior Checklist, 2003) was used to assess clinical severity by age-group comparisons according to gender. Additionally, the frequency of aggressive and auto-aggressive behaviors ranging from 1: “never” to 4: “several times a week” was rated by parents in an unstandardized report.

The children were assessed at three study centers. These were based at Humboldt–Universität zu Berlin, Germany (HU) and at two University Departments of Child and Adolescent Psychiatry and Psychotherapy with specialized outpatient clinics for children/adolescents with ASC in Augsburg (KJPP AUG), Germany, and Vienna (MedUni Wien), Austria. Additional participants were recruited through autism care units and parent organizations in Germany and Austria, as well as through a study website (www.​zirkus-empathico.​de). The assessment of the data reported here took place before the main intervention of the RCT. The questionnaires for parents were provided online on the SoSci-Survey platform (Leiner, 2014) and Lime Survey (Limesurvey GmbH, 2016). The RCT trial received ethical approval from the Ethics Committee at HU (2015/10/07) and the clinical authorities in the two outpatient clinics. Written informed consent was obtained from the children’s legal guardians after receiving a detailed study description. Families received €7/hour as compensation.

Pearson’s correlation analyses were calculated to examine the links between aggression subtypes and demographic/clinical characteristics (age, nonverbal/verbal IQ, autism social symptoms (SCQ)). All statistical tests were two-tailed and were conducted pairwise. The Bonferroni–Holm procedure was applied to correct significance thresholds to account for the accumulation of type I error due to multiple comparisons. Reports include corrected significance values (p), and r statistics for Pearson’s r.

As proposed in our hypotheses, and in reference to the modified SCIP models (Lemerise & Arsenio, 2000; Fig. 2), we specified a path model including dysfunctional emotion processes (emotion model), which was compared with the classical version of the SCIP models (classical model; Crick & Dodge, 1994, 1996). The model comparison was used to evaluate if the more complex emotion model explains aggressive behaviors better than the classical model, which only relies on cognitive processes such as emotion recognition and hostile attribution bias. In our specific case, we specified the classical model by facial emotion recognition accuracy predicting hostile attribution bias (as assessed through the FAVK score), which in turn predicts different aggression subtypes (physical aggression, bullying, verbal aggression, and covert aggression as measured by the C-SHARP). The emotion model includes dysfunctional emotion regulation (as assessed by the ERC subscale lability-negativity) as an additional predictor of hostile attribution bias and aggression subtype with hostile attribution bias mediating the relationship between lability-negativity and aggression. Both models were specified through maximum likelihood estimation with robust standard errors and a Satorra-Bentler scaled test statistic. The MLM estimator was used because multivariate normality could not be assumed for every model. Model fit was validated by using model fit indices (comparative fit index, CFI, root-mean-square-error of approximation, RMSEA, and standardized root-mean-square residual, SRMR). The emotion model was compared to the classical model based on CFI comparisons (CFI classical model minus CFI emotion model) with negative delta CFI pointing to a better fit of the emotion model; the cut-off for a meaningful difference was set to -0.002 (Meade et al., 2008). Additionally, a sample-size adjusted Bayesian information criterion (Adj. BIC) and Akaike information criterion (AIC) were used for model comparison, with smaller values indicating better model fit (Merkle et al., 2016). All comparisons were run separately for each form of aggressive behavior to explore the predictive value of the two proposed models for the different subtypes of aggression. Significance thresholds were corrected by applying the Bonferroni–Holm procedure to account for the accumulation of type I error due to multiple comparisons. Reports include corrected significance values (p). All analyses were performed in R (Version 1.3.1073, R Core Team, 2018).

The demographic and clinical characteristics of n = 60 children are displayed in Table 2.

FAVK data was available for 60 participants, and 55 parents rated their child on the C-SHARP aggression assessment. The cut-off (T > 70) for clinically significant aggression on the CBCL subscale was met by 55% of the total sample (n = 33) with the majority (62%) showing aggressive behaviors several times a month (25%), or several times a week (37%). The most prevalent subtypes were covert aggression (M = 9.0, SD = 5.4), bullying (M = 8.9, SD = 6.8), and verbal aggression (M = 8.7, SD = 7.8), while physical aggression showed the lowest prevalence (M = 2.2, SD = 2.5). There was no significant difference between boys and girls for all forms of aggressive behavior (verbal aggression: t(54) = 0.99, p = 0.341; bullying: t(54) = 0.35, p = 0.730; covert aggression: t(54) = 0.62, p = 0.547, physical aggression: t(53) = 1.49, p = 0.154). Analysis revealed that children were more likely to engage in reactive than proactive aggression as reflected by the C- SHARP provocation dimension (Table 2).

After correcting for multiple comparisons, neither age nor nonverbal or verbal IQ or autism social symptomatology (SCQ) correlated significantly with the aggression subtypes or the predictor variables (emotion recognition, hostile attribution bias, lability-negativity) (Table 3).

Path analyses revealed an acceptable to a good fit for the classical model in all aggression subtypes (Table 4). In favor of H1, model comparisons revealed a better model fit for the emotion model when compared to the classical model as indicated through CFI, AIC, and BIC values. This means that the predictive power of the models was enhanced when lability-negativity was included as a second predictor of hostile attribution bias and the respective aggression subtypes.

By applying the emotion model, and after correcting for multiple comparisons (Table 4, Fig. 3), we found hostile attribution bias being positively predicted by emotion recognition accuracy (standardized estimates with confidence intervals, b = 0.283, p = 0.032, [0.074, 0.496]) and lability-negativity (b = 0.594, p < 0.001, [0.390, 0.797]). Hostile attribution bias was a significant positive predictor of verbal aggression (b = 0.545, p < 0.001, [0.375, 0.715]) and covert aggression (b = 0.540, p < 0.001, [0.308, 0.772]), but not of bullying (b = 0.332, p = 0.124, [0.030, 0.634]) and physical aggression (b = 0.126, p = 1.00, [-0.227, 0.478]). Lability-negativity had a direct positive effect on verbal aggression (b = 0.272, p = 0.004, [0.113, 0.430]), bullying (b = 0.403, p = 0.008, [0.143, 0.662]), and covert aggression (b = 0.356, p = 0.016, [0.116, 0.596]), but not on physical aggression (b = 0.321, p = 0.108, [0.036, 0.607]). Hostile attribution bias partly mediated the relationship between lability-negativity and verbal, or respectively, covert aggression.

According to the SCIP, the first step of social-cognitive information-processing is the correct encoding and processing of others’ emotions to make moral judgments (Chester & Langdon, 2016; Lemerise & Arsenio, 2000). Previous studies identified positive relationships between emotion recognition and later SCIP steps (e.g., problem identification and solution, goal clarification, response selection) in TD children with and without mild intellectual impairments (Bauminger et al., 2005; Meyer et al., 2006; Schultz et al., 2004; van Nieuwenhuijzen & Vriens, 2012) and children with autism (Russo-Ponsaran et al., 2015). In line with this, a lack of cognitive empathy (understanding of others' emotions; Dziobek et al., 2008) was related to higher aggressive tendencies in some studies with TD children (Euler et al., 2017; Mayberry & Espelage, 2007). Hence, it was rather unexpected that children with better emotion recognition skills were rated as having a more pronounced tendency to attribute hostile intent to others in our autistic sample. Interestingly, Pouw et al. (2013) reported a positive relationship between self-rated cognitive understanding of others’ emotions and aggression levels in children with autism but not TD children. They argued that emotional content of any kind, such as when correctly understanding the emotions of others, could activate empathic arousal, which is perceived as aversive (personal distress) due to dysfunctional emotion regulation and therefore triggers aggressive behaviors. However, from a longitudinal perspective, it seems plausible, that the causal relationship between the variables is reversed: a child might have developed a tendency to attribute hostile intent in the first place due to frequent negative social experiences (e.g., being teased, or being excluded because of autism-related social impairments; see Ziv et al., 2014 for further causes). As a consequence, the child might have trained emotion recognition skills/cognitive empathy more intensively to detect potentially hostile or aggressive cues early (compare hypervigilance to hostile cues, Helmsen et al., 2012) to prevent negative experiences. Indeed, Embregts and van Nieuwenhuijzen (2009) found boys with autism and mild intellectual impairment to strongly focus on negative and emotional information in video-presented vignettes of social situations. Longitudinal studies are therefore needed to further disentangle the complex relationship between emotion understanding and aggression (compare Quan et al., 2019).

We observed the proposed interplay between dysfunctional emotion regulation and hostile attribution bias for verbal aggression (e.g., saying hurtful things, insulting others) and covert aggression (e.g., physically teasing others against their will, sneering at others). Thus, the revised SCIP models (here: emotion model, Lemerise & Arsenio, 2000) seem to be a valid approach for explaining these more complex aggression subtypes in children with autism. Interestingly, due to the missing impact of hostile attribution bias on bullying and physical aggression in our sample, the revised SCIP models might not be informative for the more physically or overtly expressed aggression subtypes. This is surprising given studies demonstrating significant impairments in SCIP operations in children with autism when compared with TD children including hostile intent attribution in ambiguous situations (e.g., Flood et al., 2011; Russo-Ponsaran et al., 2015; Ziv et al., 2014), and studies demonstrating relations between hostile intent attribution and aggressive behavior in TD children (Martinelli et al., 2018; Verhoef et al., 2019). However, Helmsen et al. (2012) reported no association between hostile intent attribution and aggression in TD children. A study by Coy et al. (2001) further found that preschool boys with oppositional defiant disorder were no more likely to attribute hostile intentions in ambiguous situations than boys of the control group. Finally, bullying, which is defined as malicious actions to strategically harm another person in order to gain or preserve power or reputation (Volk et al., 2017), is thought to arise from deficiencies, or persistent biases in the early stages of the SCIP (Crick & Dodge, 1999). In contrast to this view, current studies (e.g., Guy et al., 2017) do not support that TD bullies make more hostile attributions in response to ambiguous social information, which would indicate biases in early SCIP operations.

These mixed results might in part be due to differences in methodology such as different measurements of emotion processes, hostile attribution bias, and aggressive behavior (see Helmsen et al., 2012). Furthermore, it may be relevant to operationalize hostile attribution bias analog to the aggression subtype in focus. In their meta-analysis, Martinelli et al. (2018) found physically aggressive TD children to attribute hostile intent especially in response to physically provocative situations (e.g., when being hit with a ball). In contrast, children engaging in relational aggression (infliction of harm via actual or threatened damage to, or control of, relationships; Crick & Grotpeter, 1995) primarily displayed relational hostile attribution bias (e.g., in response to vignettes targeting ambiguous social situations like not being invited to a friend’s birthday). The items of the FAVK subscale that were used here (e.g., thinks that many people do not like him/her and have a hostile attitude towards him/her; often feels unfairly treated) seem to address hostile intent attributions, which are more closely associated with complex aggression subtypes such as verbal or covert aggression than with physical aggression. Since physical hostile attribution bias was not specifically targeted here, the assumption that a tendency to attribute hostile intent might have an impact on the relationship between lability-negativity and physical aggression in children with autism should be reevaluated with a broader set of hostile attribution bias items.

Furthermore, our results for bullying, with lability-negativity having an impact on this subtype while hostile attribution bias does not, underline the interpretation of the C-SHARP bullying subscale by its authors Farmer and Aman (2010, 2011). Based on their findings in children with autism, they suggested that the items of the bullying subscale (e.g., throwing objects at others, invading personal space) might not represent malicious actions intended to harm other persons in this population, but rather impulsive, socially inadequate responses to stressful environmental conditions. More plastically, the “children engage in physical ‘communication’ when frustrated” (Farmer & Aman, 2010, p.278) because they are incapable of alternative actions (Mazza et al., 2017) due to autism-related social skills impairments (e.g., difficulties to communicate desires, or personal needs in adequate ways). Therefore, we could potentially conclude that our results for the bullying subscale might generally account for simple physical acts of aggression towards others, with lability-negativity being a prominent predictor. Additional predictors related to social interaction and communication impairments potentially having an impact on later SCIP operations (e.g., response access/construction; response decision, see Fig. 2) should be investigated in future research.

Surprisingly, the C-SHARP subscale, which explicitly targets physical aggression, was not associated with lability-negativity. Besides its questionable reliability in our sample (Cronbach’s Alpha = 0.65; but 0.74 in Farmer & Aman, 2010), it might be that the low physical aggression rates (M = 2.18, SD = 0.33), with 17 children (32%) showing no physical aggression at all, resulted in low variance and therefore insufficient statistical power to detect the proposed relations in the rather small sample (n = 54). These low physical aggression rates might be due to a low representation of children with intellectual impairment, limited language ability, and low adaptive functioning; factors which are associated with an increased risk for aggressive behavior for individuals with autism (Hill et al., 2014; Mazefsky et al., 2013). Farmer et al. (2015) found physical aggression being related to lower IQ levels in autistic children, while more complex aggression subtypes (verbal/covert aggression) were associated with higher IQ, better adaptive behavior, and older age. However, we did not observe correlations between demographic/clinical variables (autism symptom severity, age, verbal/nonverbal IQ) and the aggression subtypes in our sample, which shows a relatively narrow age range and (high) IQ level when compared to Farmer et al. (2015). We might conclude that, especially the more physically expressed subtypes (here: bullying and physical aggression), should be targeted with carefully designed longitudinal studies to disentangle a range of different potential predictors (dysfunctional emotion regulation, hostile attribution bias, lack of social skills, etc.) in larger samples under the theoretical perspective of the SCIP models.

By using a cross-sectional mediation approach, the developmental trajectories and directionality of the relations between risk factors (here: emotion recognition, hostile attribution bias, lability-negativity) leading to aggressive behaviors cannot be disentangled sufficiently to fully understand causal relationships. According to Cole and Maxwell (2003, 2007) mediation consists of causal processes that unfold over time. Thus, using cross-sectional approaches to mediation typically generate substantially biased estimates of longitudinal parameters (Maxwell & Cole, 2007). As pointed out by Helmsen et al. (2012), it is most likely that the relationship between emotion regulation, social information processing, and aggressive behavior is bidirectional. Therefore, longitudinal studies are needed to investigate the causal direction of these relationships. Second, we cannot rule out that observer biases confounded relationships between the different constructs. However, we had to largely rely on parent questionnaires due to the young age of the children. We encourage future studies to use more objective measures to assess emotion regulation abilities and hostile attribution bias (e.g., pictorial interviews using vignettes, compare Helmsen et al., 2012; Mazza et al., 2017; Ziv et al., 2014). Lastly, we have not included a typically developed comparison group, given that we relied on data from an RCT including only children with autism. Thus, we cannot make inferences about the specificity of the reported results. However, much is known about factors predicting aggressive behavior in TD, which we sought to supplement with insights from autism in the current work as a preliminary step. Nevertheless, future research should compare autistic to TD children to investigate between-group differences in the pattern of the interplay of these socio-emotional predictors of aggression.

Even though our understanding is still limited, the results reported here may have implications for designing and selecting targeted interventions for children with autism and comorbid aggression as well as for future research on the topic. First, our study showed the important role of emotion regulation for verbal and covert aggression as well as for bullying. Thus, emotion regulation competencies (e.g., awareness of own emotions, impulse/anger control, functional emotion regulation strategies) should be given priority in therapy (compare Helmsen et al., 2012; or novel technology-based approaches like “Zirkus Empathico”; Kirst et al., 2020). This might also strengthen autistic children to better deal with negative arousal potentially induced by others’ emotional displays (Kliemann et al., 2013; Pouw et al., 2013), which could, in turn, enable more fruitful training of understanding others’ socio-emotional cues. Given our results, emotion regulation competencies may also diminish hostile attribution biases, and thus exert additional beneficial effects on the reduction of externalizing behavior via this indirect route. Since particularly emotionally engaging social situations were found to elicit the automatic and emotional processes that activate hostile attribution bias, interventions should assess and target biases in similar and naturalistic situations (Verhoef et al., 2019). Additionally, the specific pattern of aggressive behavior in children with autism should be carefully identified for each patient to allow individualized interventions. Beyond assessing the most prevalent aggression subtypes and their function for the individual, the nature of hostile attribution biases should be examined to allow customized and effective interventions. Behaviors summarized by the bullying and potentially by the physical subscale might be effectively reduced by strengthening social skills in addition to emotion regulation strategies, while more complex aggression subtypes such as verbal and covert aggression could be targeted by identifying and modifying aggression-promoting attitudes, thought patterns, and response tendencies towards others through cognitive-behavioral approaches. Indeed, interventions modifying SCIP in TD children (e.g., Hudley & Graham, 1993; Lochman & Wells, 2002) have been proven relatively effective (Kazdin, 2003).

Finally, our behavioral findings of the interplay between hostile attribution bias, emotion regulation, and abnormal behavior should be further investigated from different perspectives (e.g., socio-cognitive, developmental, neurobiological) in autism samples. So far, a prominent role of emotion dysregulation and related personal traits such as impulsivity in moderating the relationship between social cognition and aggression has been demonstrated cross-sectionally (e.g., Musher-Eizenman et al., 2004) and longitudinally in TD individuals. For example, Blandon et al. (2010) and Halligan et al. (2013) reported a causal role for problematic regulation of negative emotions at age one and, respectively, two, and the etiology of externalizing psychopathology at age five, and seven. By using a longitudinal approach in a large adolescent sample (N = 585), Fite et al. (2008) found impulsivity moderating the relationship between cognitions (here: positive endorsement of aggressive responses in hypothetical, ambiguous situations) at age 11–13 and aggressive behavior at age 14–17. Interestingly, only moderately to highly impulsive individuals showed a significant association between aggression-prone cognitions and aggressive behavior. Likewise, Goldweber et al. (2011) suggested that individual differences in executive functions (here: inhibiting behavior, shifting attention, and controlling emotions) may account for stability in aggressive social information processing (SIP). They found children aged 7–13 years with a stable aggressive SIP pattern exhibiting more executive functions problems than children who showed a decline in aggressive SIP over one year.

From a neurobiological perspective, the medial prefrontal cortex (mPFC) as the “cortical control board” (Xu et al., 2019, page 2), has been found to play an essential role for emotion regulation and, among others, for sociability (Xu et al., 2019). In addition to autism, abnormal activity in the mPFC has been shown for other psychiatric disorders (e.g., depression, anxiety, schizophrenia, addiction; see review by Xu et al., 2019). Identifying the specific pattern of cortical activation in response to emotion regulation processes inherent to autism may help to differentiate between autism and potentially co-occurring psychiatric disorders (e.g., depression, anxiety). Furthermore, localizing distinct cortical areas in the mPFC related to autism-specific deficits in emotion regulation processes (compare findings for major depression, Rive et al., 2013) and studying their connections to regions involved in higher-order socio-cognitive processing may result in a more in-depth understanding of the interrelated abnormalities underlying aggressive behavior in some individuals with autism.