Reexamining Empathy in Autism: The Role of Empathic Disequilibrium in Autism and Autistic Traits

Background: While many autistics report feelings of excessive empathy, their experience is not reected by most of the current literature, which typically, but not always, suggests that autism is characterized by intact emotional empathy and reduced cognitive empathy. To try and bridge this gap in empirical ndings and with respect to individuals' experiences, we examined a novel conceptualization of empathy termed empathic disequilibrium, i.e., the imbalance between emotional and cognitive empathy. Empathic disequilibrium was previously found to predict autistic traits in non-autistic population, suggesting it is an important empathy measure. Here, we aimed to extend the generalizability of empathic disequilibrium to the autistic population and to provide a better analytical approach to examine this construct. Methods: We analyzed self-reports of empathy and autistic traits in a large cohort (N = 4,914) of autistic and non-autistic individuals. We applied a polynomial regression with response surface analysis to examine empathic disequilibrium and total empathy as predictors of an autism diagnosis and autistic traits. Results: Total empathy and empathic disequilibrium each predicted autism. There was a higher probability for diagnosis in individuals with lower total empathy, but also in individuals with higher emotional relative to cognitive empathy. Linear and non-linear patterns linked empathy, empathic disequilibrium, and autistic traits and diagnosis, with empathic disequilibrium being more prominent in females. Conclusions: Empathic disequilibrium might allow for a more nuanced and sensitive understanding of empathy and its link with autism. This study provides empirical evidence that empathic disequilibrium is at least as informative as empathy for assessing autism, and offers a novel analytical approach for examining the role of empathy at the phenomenological level. This study provides empirical evidence that ED is at least as informative as empathy for the diagnosis of autism, and for predicting autistic traits in both autistic and non-autistic populations. By offering a novel way to examine the role of empathy in autism, ED promises to scaffold our understanding of the experience of some autistic individuals. By rening our understanding of the link between empathy and autism, targeting ED in future research may provide valuable clinical insights used for prognosis, diagnosis, and interventions in autism. Moreover, it can help delineate the nature of the mechanisms underlying empathy for all individuals.

Impairments in social communication and behaviour are core features of autism. Autism is characterized by social interaction and communication di culties, accompanied by repetitive and restrictive behaviours with onset during early development [19]. Considering the fundamental role empathy plays in social behaviour and communication [2], atypicalities in empathy have been suggested to be a hallmark of autism [7]. EE and CE in autism. Many empirical ndings show impaired CE and intact EE in autism [20,21]. Yet other studies show mixed results [4,22,23]. For example, one study found that young autistic children displayed EE less frequently than non-autistic children [22]. Another study used a common behavioural task to measure CE and classi ed autistic individuals into ve separate subgroups, two of whom did not differ in CE from non-autistic individuals [9]. These mixed ndings hinder our ability to understand the behavioural and biological mechanisms of empathy and its role in autism.
Empathic disequilibrium (ED). It is important to note that almost all studies of empathy in autism examined the role of CE and EE independently of each other. However, CE and EE have also been shown to in uence and regulate each other [24][25][26]. For example, CE-related brain regions interact with EErelated brain regions, particularly during complex social situations in which additional processing is needed to jointly engage EE and CE [26]. As we constantly encounter complex and relatively ambiguous social situations, these studies suggest that maintaining a balance between EE and CE is key for an adaptive and appropriate social response, leading to effective social communication. The possible role of the balance between EE and CE has been overlooked in previous studies, leaving some open questions about what happens when this joint regulation is altered and whether it might explain the mixed ndings above.
For example, some individuals might show average CE, but in combination with higher EE, this creates an empathic imbalance, which might relate to signs of autism. To examine this possibility, we use the term ED, relating to the level of imbalance between CE and EE [10]. We previously found that the level of imbalance between CE and EE, and not each trait independently, predicted autistic traits in non-autistic individuals, even when controlling for their total empathy. Speci cally, we showed that autistic traits were elevated in a group of individuals with relatively higher EE than CE (EE-dominance group) and found that EE-dominance was related to social aspects relevant to autism, such as alexithymia; but not to the restrictive and repetitive interests, such as systemizing [10]. These ndings provide empirical evidence for the notion that an imbalance between CE and EE might contribute to some autistic symptoms [27].
The current work extends these ndings and explores the relevance of ED to clinically diagnosed autistic individuals. Furthermore, in our previous paper, we used a difference score between standardized CE and standardized EE to measure ED while controlling for total empathy. This method has several limitations [28][29][30], but most importantly, it does not allow for the simultaneous investigation of both equity and inequity between EE and CE and their relation to an outcome.
To address these issues, we used polynomial regression with response surface analyses (PRRSA) [28,31]. PRRSA visualizes the three-dimensional (in)congruency between variables and assesses their association with an outcome variable in a statistically valid and comprehensible way. Congruency is assessed via examination of the line of congruence (LOC), representing the degree to which similarity between variables is associated, both linearly and curvilinearly, with an outcome variable (see the blue line in Figure 1). Incongruency is measured using the line of incongruence (LOIC), which examines whether and how the discrepancy between two variables is related to an outcome (see the black line in Figure 1).
In this research, PRRSA allowed the examination of both ED, represented by the LOIC, and total empathy (which comprises both EE and CE), represented by the LOC, while considering the contribution of EE and CE separately. Using PRRSA, we examine whether equilibrium and disequilibrium between emotional and cognitive empathy (measured using validated self-report questionnaires) predict autism diagnosis and autistic traits (separating social and non-social aspects). Based on our previous ndings [10], we hypothesized that both total empathy and ED, favouring EE, will predict autism and traits related to the social aspects of autism. In contrast, ED, favouring CE, will predict non-social aspects related to autism (e.g., systemizing). Furthermore, as ED also shows average sex differences, we expected females on average to show a higher tendency towards EE-dominance, relative to males.

Participants:
Participants were 1,905 individuals diagnosed with autism and 3,009 non-autistic controls (see Table 1 for descriptive statistics) recruited via the Cambridge Autism Research Database (CARD). Participants self-reported their diagnosis (including speci c details, such as date of diagnosis, which is used as a validity check for diagnostic status), age, and birth sex. Autistic and non-autistic participants then completed a battery of questionnaires. The non-autistic group showed elevated (yet in the typical range) autistic traits, and 193 individuals exceeded the autism cut-off of the Autism-Spectrum Quotient [32], suggesting that this group, although undiagnosed, shows slightly elevated features of autism.

Measures:
Empathy. Empathy was measured using the Empathy Quotient (EQ) [7]. The questionnaire consists of 60 items (40 empathy items and 20 ller items) on a 4-point scale. On each empathy item, a person can score 2, 1, or 0. Two three-factors structures are commonly used in the EQ to tap cognitive, emotional, and social skills aspects of empathy [33,34]. To decide which of the two classi cations provide the best t for the speci c data used, we conducted con rmatory factor analysis using lavaan package in R [35]. This analysis revealed a reasonable t for both Lawrence's 28-items three factors and Muncer and Ling's 15items three factors, but as the latter showed better t indices (see Table 2), we chose to calculate EE and CE scores using Muncer and Ling's classi cation [33]. Con rmatory factor analysis and model t parameters of Lawrence and Shaw, and Muncer and Ling Empathy Quotient classi cation.
We did not include the subscale tapping social skills (which is part of the original three-factor classi cation) as it does not directly relate to EE and CE. Using these classi cation, EE and CE were found to be positively correlated (r = 0.59, p < 1x10 −100 ). Following Fleenor et al. [36] recommendation and in line with our previous study [10], EE and CE were standardized (separately). Moreover, to create an easily interpretable measure, before dividing both measures by the standard deviation of the total sample, both EE and CE were centered based on the mean of the control group. Thus, the scores re ect the standardized score of CE and EE, relative to the mean of the non-autistic population.
Autistic traits. Autistic traits were measured using the Autism Spectrum Quotient (AQ) [32]. This questionnaire consists of 50-items measuring autistic traits in the general population. Responses are scored using a binary system, where an endorsement of the autistic trait (either mildly or strongly) is scored as 1, while the opposite response is scored as 0. Scores are then summed up leading to a maximum score on the AQ of 50. The AQ can also be divided according to ve domains: 'social 'skills', 'attention 'switching', 'attention to 'detail', ''communication', and ''imagination'. We also measured systemizing, which is the drive to analyze or construct systems, and is an autism-related feature of the non-social domain of autism [37,38]. Systemizing was measured using the Systemizing Quotient [39], a 60 items (40 systemizing items and 20 ller items) questionnaire with a 0-2 rating scale, with higher scores representing higher systemizing.
Statistical analyses: Sex differences analysis. A 2x2 ANOVA was conducted examining sex, diagnosis, and their interaction as predictors of ED. ED was calculated by subtracting standardized CE from standardized EE [10].
Response surface analysis of empathy. To examine ED and its applicability in autism, we applied PRRSA [28,31]. PRRSA tested both the linear and curvilinear pattern of total empathy, de ned as the LOC, and of ED, de ned as the LOIC, using a polynomial regression between EE and CE as described using the following equation (1) : To interpret the surface of the polynomial regression, regression coe cients are used to extract four surface parameters, as follows: 1. The linear association between total empathy and an outcome variable (a1 = b1 + b2).
Therefore, a1 and a2 re ect the association between total empathy and an outcome, while a3 and a4 re ect the association between ED and an outcome.
Autism prediction using PRRSA. We rst wanted to examine if the polynomial regression surface created using EE and CE and its derived total empathy (i.e., a1 and a2) and ED (i.e., a3 and a4) parameters predict autism diagnosis. To do so, we conducted a polynomial logistic regression with autism diagnosis as a binary outcome. We also examined whether the surface parameters differed between the sexes. Age was used as a covariate.
Autistic traits prediction using PRRSA. Using PRRSA, we also examined whether total empathy and ED predicted autistic traits and whether surface parameters differed between autistic and non-autistic individuals. To do so, we conducted a polynomial regression analysis using EE and CE for AQ and SQ as outcome variables (separately). Differences in surface parameters were investigated between autistic and non-autistic individuals. Age and sex were used as covariates. We also conducted polynomial regression analyses for each of the ve AQ subscales separately (see supplementary information).
All analyses were carried out using R v4.0.3 'stats' package [40]. RSA plots were produced using the RSA package in R [41].

Results
Sex differences in ED. Before applying PRRSA, we examined whether males and females differ on average in ED, and whether sex interacts with diagnosis.  Figure 2.
Response surface analysis of empathy. We next examined how total empathy and ED predict autism diagnosis and autism-related traits using PRRSA models. Residuals of all the models tested were normally distributed.
Predicting autism diagnosis. The overall polynomial regression model predicted autism diagnosis (R 2 = 0.52, p < 1x10 −100 ) in males and females (see Table 3 and Figure 1). Total empathy -Lower total empathy was associated with an autism diagnosis, showing both a linear (a1) and a curvilinear (a2) association. The linear effect of total empathy was stronger for females than for males (t = -3.95, p = 0.00008).
ED -ED also signi cantly predicted autism, with an effect size that was similar to that of total empathy. The probability for autism diagnosis was higher for individuals whose EE was higher than their CE (negative a3). A signi cant curvilinear association shows that autism probability increases more sharply as ED increases (positive a4).
Total empathy -Lower total empathy was associated with higher AQ scores in autistic and non-autistic individuals, showing linear (a1) and curvilinear (a2) associations. The curvilinear association for total empathy was stronger for non-autistic than for autistic individuals (t = -3.3, p = 0.001).
ED -A linear association between ED and autistic traits was found for both autistic and non-autistic individuals, with higher EE-dominance predicting higher autistic traits (negative a3). A curvilinear association of ED and autistic traits was also found for non-autistic individuals, which differed from the non-signi cant curvilinear effect of ED in autistic individuals (t = -4.43, p = 0.00001).
PRRSA for each of the ve AQ subscales were also examined and are reported in the supplementary information.
Systemizing. To examine the non-social domain of autism, we also measured systemizing -the drive to analyse and construct systems, using the SQ [39]. Autism diagnosis was associated with higher SQ score (β = 0.235, p = 3x10 −14 ), and males showed higher SQ scores than females (β = -0.14, p = 4x10 −25 ). Age was also associated with systemizing (β = 0.056, p = 0.00001). The overall model of empathy was found to be predictive of systemizing traits (R 2 = 0.26, p < 1x10 −100 ) in autistic and non-autistic individuals. See details in Table 5 and Figure 4. Parameters of polynomial regression with response surface analysis (PRRSA) of EE and CE, predicting Systemizing-Quotient score in autistic and non-autistic individuals.
*p < 0.05, **p < 0.005, ***p < 0.0005 Total empathy -In the autistic population, the curvilinear, but not linear, association was signi cant, with higher SQ scores predicted by high or low total empathy. In contrast, although not signi cantly different from the autistic group (see 'group 'comparison' statistics in Table 4), non-autistic individuals also showed signi cant linear and curvilinear association of small sized effects between total empathy and systemizing.
ED -In the autistic group, the curvilinear association was again signi cant, with higher SQ scores predicted by high or low ED. Although only nominally signi cant, linear association between ED and SQ showed a tendency towards higher SQ scores for autistic individuals whose CE is higher than EE. In the non-autistic group, ED was associated linearly and curvilinearly with ED, with a tendency towards higher CE than EE predictive of SQ score.

Discussion
In this study we investigated the independent role of ED and empathy in predicting autism diagnosis and autistic traits in autistic and non-autistic individuals. In line with our hypotheses, both total empathy and ED predicted autism, with a higher probability for diagnosis in individuals with lower total empathy and in individuals with higher emotional than cognitive empathy. Our data suggest linear and non-linear patterns linking empathy, ED, and autism diagnosis in autistic and non-autistic individuals; such complexity was also apparent in predicting autistic traits. We also found that a tendency towards EE-dominance (higher emotional than cognitive empathy) is more related to the social domain of autism (e.g., as measured by the social-related subscales of the AQ), while a tendency towards CE-dominance (higher cognitive than emotional empathy) is more related to the non-social domain (e.g., as measured by the SQ). In addition, while the relationship between ED and autism holds for both sexes, females across diagnostic groups showed a greater tendency towards higher EE than CE.
Investigating ED and total empathy simultaneously allowed us to show that both aspects of empathy are informative of autism diagnosis and autistic traits. Thus far, studies examining EE and CE separately resulted in inconsistent ndings, suggesting each component of empathy by itself is not sensitive enough to characterize autism [4,9,22,23]. The current approach to investigating empathy takes into account the relationship between EE and CE, and as such may shed light on some of these mixed ndings. Based on our ndings, mean differences between autistic and non-autistic individuals in CE or EE do not re ect the role of empathy in autism to the fullest. Indeed, we nd that beyond overall empathy, the probability for autism diagnosis is associated with higher EE relative to CE (i.e., a tendency towards EE-dominance).
How would such an imbalance manifest? A person with ED towards EE-dominance might understand others' emotional states (CE) at the typical level, but her/his ability to experience and share in these emotions (EE) will be relatively higher. Smith [27] suggested that this state would cause overarousal, as the individual becomes overwhelmed with the other's emotions, resulting in the cognitive and behavioural characteristics of autistic individuals, which constitute an adaptive response to overarousal. This conceptualization coincides with the inner experience of some autistic individuals reporting "excess of empathy" [42]. Future research will need to validate this notion and examine it as a possible mechanism of action underlying the relationship between ED and autism.
The idea that empathy might be linked to overarousal in autism is also re ected by the non-linear associations between empathy (both total empathy and ED) and autism diagnosis and some autistic traits (see supplementary information for details). This nding is in line with the suggestion, although rarely examined empirically, that non-linear models are better suited for describing empathy in a nuanced way [43,44]. In addition, some researchers suggest that extreme (high or low) levels of empathy can lead to overarousal and worsen psychological functioning [43,45]. If this is the case in autism -where problems in emotion-regulation are common [46, 47] -emotional dysregulation may be driven by ED or extreme levels of total empathy.
Regarding ED, our data show that the two types of ED (EE-dominance versus CE dominance) predict different domains of autistic behaviour in both autistic and non-autistic individuals: a tendency towards EE-dominance is associated with the social domain of autism, while a tendency towards CE-dominance is associated with the non-social domain of autism (such as systemizing). We found the same patterns in a non-autistic population [10]. Therefore, these results replicate our previous study and extend the generalizability and utility of the ED concept to autistic individuals.
Furthermore, the differences within the autistic group between individuals with a tendency towards EE compared to those with a tendency towards CE dominance highlights the heterogeneity characterizing the autistic spectrum [48]. It might also provide a new basis for stratifying autistic individuals, which is a means for understanding the heterogeneity of autism [49,50] We also observed average sex differences in relation to ED in autistic and non-autistic populations. These differences were prominent in autistic individuals showing that while both autistic males and females displayed higher EE than CE, this effect was more pronounced in autistic females. This is in line with our hypothesis [10] that ED might be of particular relevance to autistic females, a relatively under-studied population [51,52], and could help differentiate a female presentation of autism.

Limitations
Our study has several limitations. First, the sample used in our study consists of autistic individuals with average or above average intelligence (re ected indirectly in being able to participate in online studies) and so does not represent the entire autistic spectrum. Online studies also lead to an ascertainment bias, re ected by the relatively high proportion of autistic females in the sample, which is not representative of the typical higher male-to-female ratio in autism [51]. This limits our conclusions regarding sex differences in ED between autistic males and females. Furthermore, the non-autistic group also included family members of diagnosed individuals, suggesting the ndings might be more representative of the broad autism phenotype, i.e., people who carry genetic liability for autism and/or display milder phenotypic features [53,54]. Yet even in this population as a comparison group, we see signi cant differences with the autistic group.
Second, all measures used in our study are self-report questionnaires. Although these measures are validated and correlate with other behavioural measures [7,32,34,39], they primarily re ect the participants' perception of their own functioning and ability. However, while observational methods offer rich information, empathy is largely an internally-experiential process that cannot be inferred from behaviour alone [55], suggesting that self-report measures are valuable tools for understanding empathy.

Conclusions
This study provides empirical evidence that ED is at least as informative as empathy for the diagnosis of autism, and for predicting autistic traits in both autistic and non-autistic populations. By offering a novel way to examine the role of empathy in autism, ED promises to scaffold our understanding of the   Sex and diagnosis differences in empathic disequilibrium (ED). The mean of each group appears in red. 95% con dence intervals of each group are depicted. Positive values of ED (on the x-axis) represent higher cognitive than emotional empathy. Negative values represent higher emotional than cognitive empathy. The dashed line represents the point of no difference between cognitive and emotional empathy.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. ReexaminingempathyinautismSIMolecularAutism.docx