Bilingualism Predicts Affective Theory of Mind in Autistic Adults

Abstract

Purpose:

This study examined the impact of bilingualism on affective theory of mind (ToM) and social prioritization (SP) among autistic adults compared to neurotypical comparison participants.

Method:

Fifty-two (25 autistic, 27 neurotypical) adult participants (ages 21–35 years) with varying second language (L2) experience, ranging from monolingual to bilingual, completed an affective ToM task. A subset of this sample also completed a dynamic eye-tracking task designed to capture differences in time spent looking at social aspects of a scene (SP). Four language groups were compared on task performance (monolingual autism and neurotypical, bilingual autism and neurotypical), followed by analyses examining the contribution of L2 experience, autism characteristics, and social face prioritization on affective ToM, controlling for verbal IQ. Finally, we conducted an analysis to identify the contribution of SP on affective ToM when moderated by autism status and L2 experience, controlling for verbal IQ.

Results:

The monolingual autism group performed significantly worse than the other three groups (bilingual autism, monolingual neurotypical, and bilingual neurotypical) on the affective ToM task; however, there were no significant differences between the bilingual autism group compared to the monolingual and bilingual neurotypical groups. For autistic individuals, affective ToM capabilities were positively associated with both verbal IQ and L2 experience but did not relate to autism characteristics or SP during eye tracking. Neurotypical participants showed greater SP during the eye-tracking task, and SP did not relate to L2 or autism characteristics for autistic individuals. SP and verbal IQ predicted affective ToM performance across autism and neurotypical groups, but this relationship was moderated by L2 experience; SP more strongly predicted affective ToM performance among participants with lower L2 experience (e.g., monolingual) and had less of an impact for those with higher L2 experience.

Conclusion:

This study provides support for a bilingual advantage in affective ToM for autistic individuals.

Supplemental Material:

https://doi.org/10.23641/asha.25696083

The belief of a bilingual advantage has been used to describe the benefits of speaking a second language (L2) among individuals with autism and neurotypical (NT) development across the life span and is a critical consideration for historically marginalized individuals, for whom bilingualism plays an integral role in their family heritage and lived experiences (Davis et al., 2021). Emerging research on bilingualism in autism has found that some children not only are capable of learning two languages, but also perform at similar and sometimes higher levels in comparison to monolingual peers across a variety of domains (Hastedt et al., 2023). A recent systematic review found that multilingual autistic children may share some advantages of multilingualism with their multilingual non-autistic peers (Gilhuber et al., 2023). In addition to invaluable connections to family and social supports (Davis et al., 2021), research has demonstrated enhanced selective and sustained attention (Gonzalez-Barrero & Nadig, 2019; Sharaan et al., 2022), working memory (Sharaan et al., 2022), semantic verbal fluency (Gonzalez-Barrero & Nadig, 2017), inhibitory controls (Montgomery et al., 2022), cognitive flexibility (Peristeri, Vogelzang, & Tsimpli, 2021), socialization (Hastedt et al., 2023), and nonverbal skills (Valicenti-McDermott et al., 2013) in bilingual/bilingually exposed autistic children when compared to monolingual peers (for a review, see Davis et al., 2022), though there is initial evidence that the benefits of bilingualism for children with autism are more pronounced for low–socioeconomic status families (Peristeri et al., 2022).

These findings are bolstered by research in NT populations that has noted several benefits to being bilingual. In NT populations, bilingualism has been associated with improved metalinguistic awareness and language processing in both children and adults (Bialystok, 2001; Economacou et al., 2023; Teubner-Rhodes et al., 2016), as well as flexible attention and executive functioning (EF) skills in children and adults (Bialystok, 2017; Crivello, 2016; Kovacs & Mehler, 2009), although one study did not find this bilingual advantage in a group of 9- to 10-year-olds without autism (Dick et al., 2019). Of particular interest to the field of autism, bilingualism has also been shown to positively impact theory of mind (ToM) in NT children and adults (Rubio-Fernández, 2016; Schroeder, 2018; Tare, 2010) and, recently, children with autism (Peristeri, Baldimtsi, et al., 2021).

ToM refers to the ability to recognize others' mental states and understand how those mental states are different from one's own viewpoint, as well as from reality (Cook & Bassetti, 2010). For decades, ToM has been identified as a characteristic deficit among autistic individuals (Baron-Cohen et al., 1985; Berenguer et al., 2018; Pagni et al., 2020), with affective ToM skills (including recognizing emotions) receiving considerable attention (Yeung, 2022). In contrast to more cognitive measures of ToM, including first- and second-order false belief tasks, affective ToM is uniquely associated with autism-related social difficulties in autistic youth (Altschuler et al., 2018). Among individuals with autism, challenges in recognizing emotions have been observed for all basic emotions (e.g., anger, happiness, fear), as well as complex emotions (e.g., friendly, interested; Palmer et al., 2023; Yeung, 2022). However, it is also important to consider how the task design may impact results (Yeung, 2022). For example, one group found that when listening to vocal prosody, autistic adults recognized emotions at similar levels as comparison participants, with the exception of surprise (Martzoukou et al., 2017).

When considering challenges in affective ToM faced by autistic individuals, along with findings of a bilingual advantage for ToM among NT individuals, a compelling question is raised: Does speaking an L2 confer a bilingual advantage for affective ToM among autistic individuals? This question is especially important when considering families for whom speaking two languages is not a choice but represents a key component of their daily lives, heritage, and connection to families and other social supports (Davis et al., 2021; Luk, 2023). Despite the myriad of socioemotional benefits that come with having knowledge of an L2, including strengthened positive self-concept and familial bonds, some families are erroneously encouraged to speak one language at home to avoid “confusing” their child (Singh & Bunyak, 2018). Information about the role of L2 on ToM in autistic individuals could further empower families to support their child's bilingual development.

One possible mechanism for the benefits seen in bilingual individuals lies in its relationship to cognitive flexibility. The adaptive control hypothesis suggests that in NT individuals, the brain adapts to dual-language contexts through the strengthening of networks involved in control processes (Green & Abutalebi, 2013), and these changes are modulated by the individual's language experience (DeLuca et al., 2019). Speaking two languages may increase cognitive flexibility that, in turn, supports the ability to switch between multiple languages or perspectives as required in ToM.

Another possible factor contributing to ToM skills is the degree to which individuals prioritize looking at social aspects of their environment, which may occur differently for bilingual compared to monolingual speakers (Fort et al., 2018; Kandel et al., 2016), as well as autistic compared to NT individuals (Chevallier et al., 2015). Bilingual NT adults process faces differently from monolingual NT peers, taking longer to respond in face recognition paradigms (Kandel et al., 2016) and selectively attending to the mouth (Fort et al., 2018; Pons et al., 2015), perhaps as a way to decode the language being used. In autistic monolinguals, selective attention to the mouth has also been associated with enhanced social interaction and communication and fewer autism characteristics (Black et al., 2017; Del Valle Rubido et al., 2018; Klin et al., 2002). Differences in social prioritization (SP) between autistic and NT individuals are well documented, with autistic individuals showing lower SP than NT peers (Chevallier et al., 2015). However, it is not known whether bilingualism or SP contributes to ToM skills in autistic individuals.

To address this gap, we explored the relationship between bilingualism and affective ToM in autistic adults using the “Reading the Mind in the Eyes” Test, Revised Version (RMET; Baron-Cohen et al., 2001). The RMET is a measure of affective ToM in which participants need to infer complex emotions from images of eyes. We prioritized examining affective as opposed to cognitive ToM due to evidence in autistic children that affective ToM was more highly predictive of individual differences in social symptom severity (Altschuler et al., 2018). Additionally, there is strong evidence that monolingual autistic adults experience difficulties in emotion recognition, a domain of affective ToM (Kennedy & Adolphs, 2012; Martzoukou et al., 2017).

The following research questions were examined: (1) To what extent do participants differ on RMET scores based on diagnostic and language group status (i.e., monolingual autism, bilingual autism, monolingual NT, bilingual NT)? (2) To what extent do the following variables contribute to RMET performance: L2 experience, English verbal IQ (EVIQ), SP of faces, and autism characteristics as measured by clinicians and self-report? (3) To what extent does SP of faces predict RMET scores when moderated by diagnostic group (autism or NT) and L2 experience? Finally, for all three research questions, we explored whether the emotional valence of the facial expression (i.e., positive, neutral, or negative) impacted the findings.

For Research Question 1, we hypothesized that (a) autistic participants would perform lower on the RMET than participants in the NT group, given the characteristic challenges with ToM observed in autism (Baron-Cohen et al., 1985; Berenguer et al., 2018; Pagni et al., 2020), but that (b) for those with autism, bilingualism would be associated with better performance on the RMET compared to monolingualism, given the bilingual advantage in ToM among NT bilingual populations (Alqarni & Dewaele, 2020; Lorette & Dewaele, 2019; Schroeder, 2018). For Research Question 2, we hypothesized that among the autism group, more L2 experience and fewer self-reported/clinically observed autism characteristics and greater SP would predict enhanced ToM performance on the RMET (Del Valle Rubido et al., 2018). For Research Question 3, we hypothesized that amount of L2 experience may alter the ways in which individuals prioritize faces during social interactions (Kandel et al., 2016), which may have a cascading effect on how individuals interpret mental states from facial expressions. Finally, for the exploratory analysis of emotional valence, we hypothesized that the results would hold true across all emotional valences (positive, neutral, and negative; Palmer et al., 2023).

Method

Participants

This protocol was approved by the institutional review board at The University of North Carolina at Chapel Hill (UNC-Chapel Hill). Informed consent from participants and legal guardians (if applicable) was obtained. Autistic participants were recruited through the Autism Research Registry at UNC-Chapel Hill. NT participants were recruited via mass e-mails at UNC-Chapel Hill. All participants were recruited as part of a larger study and thus were not specifically recruited based on L2 experience. For the current study, 25 participants were included in the autism group (11 bilingual, 14 monolingual), and 27 participants were included in the NT comparison group (14 bilingual, 13 monolingual).

For all groups, inclusion criteria included the following: (a) spoke English fluently, (b) 21–35 years of age, (c) no sensory deficits (i.e., not blind or deaf), and (d) no significant physical impairments. Additionally, autistic participants had a diagnosis of autism confirmed by (a) a history of clinical autism diagnoses and (b) a total score of 8 or higher on the Autism Diagnostic Observation Schedule–Second Edition (ADOS-2) Module 4 Revised Algorithm (Hus & Lord, 2014), administered by a research-reliable assessor and using standard cutoffs (J.L.K. and R.K.G.). Participants in the NT group also met the following criteria: (a) no self-reported genetic, medical, psychiatric, neurologic, or learning condition upon entry to the study; (b) scored below the cutoff for autism on the Social Communication Questionnaire (SCQ; Rutter et al., 2003); and (c) no clinical indicators of autism characteristics during the research visit. All study tasks occurred in English.

Both autism groups (monolingual and bilingual) were significantly different from both NT groups on the following measures: (a) Social Responsiveness Scale (SRS), where autism groups scored higher, indicating more social interaction impairments, and (b) SP eye-tracking scores, where the autism groups spent significantly less time looking at faces and more time looking at background images. No significant differences were found between the four groups in terms of performance IQ, age, sex, race, or ethnicity (see Table 1). Next, groups were compared examining only participants who completed the eye-tracking task. The results were essentially the same as the full sample. The one exception was that the bilingual NT group achieved significantly higher scores on the full-scale IQ and EVIQ when compared to both autism groups, as opposed to only the monolingual autism group in the full sample (see Supplemental Material S1, Table S3, in the online resource). All analyses thus controlled for EVIQ.

Table 1.

Participant characteristics for the full sample (N = 52).

Variable	Group descriptions				Group comparisons
	Autism language groups		Comparison language groups		Monolingual vs. bilingual		Autism vs. comparison within language groups		Autism vs. comparison across language groups
	MA (n = 14)+	BA (n = 11)+	MC (n = 13)	BC (n = 14)	MA vs. BA	MC vs. BC	MA vs. MC	BA vs. BC	MA vs. BC	BA vs. MC
	+Sample sizes are as indicated, unless otherwise noted below.
	M (SD)				p values for group differences
SRS total T score	72.79 (10.60)	68.64 (12.92)	45.54 (4.79)	51.14 (7.63)	> .99	.33	< .001***	.01**	< .001***	< .001***
ADOS-2 CSS	8.43 (2.38)	8.36 (1.57)	–	–	.30	–	–	–	–	–
WASI-II
Full-scale IQ	106.5 (18.90)	113.36 (15.92)	118.15 (8.72)	124.43 (6.22)	> .99	.18	> .99	.47	.02*	> .99
Verbal IQ	108.14 (18.25)	111.45 (18.07)	117.00 (9.92)	123.21 (7.90)	> .99	.75	> .99	.60	.01**	> .99
Performance IQ	104.00 (21.60)	112.36 (11.56)	114.92 (8.23)	120.00 (7.03)	> .99	.68	> .99	.50	.26	> .99
Social prioritization score	0.07 (0.05) n = 12	0.07 (0.06) n = 8	0.17 (0.05)	0.17 (0.09)	> .99	> .99	.006**	.007**	.003**	.01**
L2 total score	0 (0)	9.16 (9.86) n = 10	0 (0)	10.83 (11.05)	< .001***	< .001***	> .99	> .99	< .001***	< .001***
Age	25.57 (4.01)	25.64 (3.85)	26.69 (4.42)	25.14 (4.15)	> .99	> .99	> .99	> .99	> .99	> .99
SES	39.57 (12.43)	35.68 (12.86)	39.04 (9.16)	38.39 (12.22)	> .99	> .99	> .99	> .99	> .99	> .99
	n (% of group)				p values for group differences
Sex					.50	> .99	> .99	.79	> .99	.98
Female	4 (29)	0 (0)	2 (15)	3 (21)
Male	10 (71)	11 (100)	11 (85)	11 (79)
Race					.97	.86	.11	.86	.22	.76
White	14 (100)	10 (91)	8 (62)	9 (64)
Asian	0 (0)	0 (0)	1 (8)	3 (21)
Asian and White	0 (0)	1 (9)	1 (8)	1 (7)
Black or African American	0 (0)	0 (0)	3 (23)	0 (0)
Unknown or not reported	0 (0)	0 (0)	0 (0)	1 (7)
Ethnicity					.98	.46	> .99	.92	.46	.98
Not Hispanic or Latino	13 (93)	10 (91)	13 (100)	10 (71)
Hispanic or Latino	0 (0)	1 (9)	0 (0)	4 (29)
Unknown or not reported	1 (7)	0 (0)	0 (0)	0 (0)

Note. SES measured by the Hollingshead Four Factor Index; t tests were conducted for normally distributed variables; Mann–Whitney Wilcoxon tests were conducted for variables that were not normally distributed (ADOS-2 CSS, SRS); Fisher's exact tests were completed for categorical variables (sex, race, ethnicity); all p values were subject to Bonferroni corrections. MA = monolingual autism group; BA = bilingual autism group; MC = monolingual comparison group; BC = bilingual comparison group; SRS = Social Responsiveness Scale; ADOS-2 CSS = Autism Diagnostic Observation Schedule–Second Edition calibrated severity score; WASI-II = Wechsler Abbreviated Scale of Intelligence–Second Edition; L2 = second language; SES = socioeconomic status.

^* p < .05.

^** p < .01.

^*** p < .001.

Materials and Measures

The SCQ (Rutter et al., 2003) is a 40-item questionnaire designed to screen for autism. This questionnaire was administered to exclude participants from the NT group who scored above the cutoff for autism.

The Language Experience and Proficiency Questionnaire (LEAP-Q; Kaushanskaya et al., 2019; Marian et al., 2007) is a validated self-report questionnaire that measures multiple-language experience and proficiency of respondents ranging from 14 to 80 years of age. Per recommendations by the LEAP-Q authors, an exploratory factor analysis (EFA) was completed to explore the underlying factors of the LEAP-Q (Kaushanskaya et al., 2019). The EFA resulted in two factors, which consist of (a) later and (b) early language experiences. The items in each factor were averaged for each participant to create mean language experience scores. All participants indicated that their current dominant language was English, and so, even though this was not the first acquired language for all participants, experiences with English were categorized as first language (L1) and all other languages were categorized as L2 or third language (L3). The two mean factor scores (later and early language experiences) were then summed to create a total language experience score for each language endorsed. Only five participants reported speaking three languages, so L3 scores were not analyzed in subsequent analyses. If participants did not endorse speaking an L2 (i.e., monolingual), their total L2 score was 0. See the Data Analyses section in Supplemental Material S1 for more details about the EFA. The LEAP-Q was used to (a) describe the language experiences of the sample, (b) divide the sample into monolingual and bilingual groups, and (c) include L2 scores as an independent continuous variable to relate to ToM capabilities and SP of faces. Higher L2 scores indicate more experience and proficiency with L2.

The “Reading the Mind in the Eyes” Test, Revised Version (RMET; Baron-Cohen et al., 2001) is a 36-item questionnaire that measures the ability to interpret complex emotions (such as “arrogant,” “convinced,” “aghast,” “fantasizing,” and “despondent”) based on small changes in the eyes and forehead. Participants were provided with written definitions of the words to refer to as needed while completing the task. Higher scores indicate better ability to infer emotions from facial expressions. This task has been shown to differentiate autistic from NT adults, regardless of verbal ability, and is highly related to other validated measures of ToM (Lim et al., 2023). For the exploratory analysis on emotional valence, the RMET items were categorized into three groups: photos displaying positive, neutral, or negative valence (14, 11, and 11 items, respectively). These categories were derived from previous research in which 113 medical students ranked the RMET photos (without viewing the “correct” emotion label) on a Likert scale ranging from very negative to very positive (Palmer et al., 2023). Because these valence categories need to be replicated in future research, analyses using RMET valence variables are considered exploratory, whereas the total RMET score, which has been validated (Lim et al., 2023), is considered the primary variable measuring affective ToM.

The SRS (Constantino & Todd, 2005) is a 65-item self-report questionnaire designed to assess impairments in social interaction. The SRS was used to describe differences between groups in self-reported social interaction impairments and as an independent continuous variable to relate to ToM performance. Higher scores indicate more impairments.

The Wechsler Abbreviated Scale of Intelligence–Second Edition (WASI-II; Wechsler, 1999) is a clinician-administered cognitive assessment, resulting in a full-scale IQ, EVIQ (derived from Vocabulary and Similarities subtests), and performance IQ (derived from Block Design and Matrix Reasoning subtests). These scores were used to describe the sample. Additionally, EVIQ was used as a covariate in all analyses. Because the WASI-II is a cognitive assessment, rather than a full language evaluation, more nuanced domains of language (e.g., receptive and expressive language across semantics and syntax) were not included in the analysis.

The ADOS-2 (Lord et al., 2012) is a semistructured diagnostic interview that allows examiners to observe symptoms of autism and calculate calibrated severity scores (CSSs). Higher CSSs indicate greater autism characteristics. A research-reliable examiner administered ADOS-2 Module 4 to the participants in the autism group to confirm group assignment. The ADOS CSS was also examined as an independent variable to assess contributions of autism symptoms on ToM performance.

The Interactive Visual Exploration task (Chevallier et al., 2015) is a 7-min dynamic eye-tracking activity that has been associated with the magnitude of autism characteristics in previous research and has also been used to differentiate autism and NT groups. For this task, participants viewed 22 silent video clips of 11 sibling pairs engaging in both social and nonsocial play activities. For each video clip, predetermined areas of interest (AOIs) were traced by hand to capture faces and background images. These AOIs changed dynamically over the course of the video to match the progression of each stimulus. The following SP score was calculated based on the participants' total fixation duration to the AOIs: proportion of time spent looking at faces during the social conditions divided by the proportion of time spent looking at background images during the social conditions. Five participants in the autism group were excluded from this analysis due to technical difficulties that prevented them from completing the eye-tracking task (two monolinguals and three bilinguals). SP was used as an independent variable to relate to ToM performance.

The Structured Clinical Interview for DSM-5 Disorders Research Version (SCID; First et al., 2015) is a semistructured interview to identify diagnoses within the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5; American Psychiatric Association, 2013). This tool was used to evaluate both NT and autistic participants for characteristics associated with co-occurring DSM-5 conditions. Eleven autistic participants reported symptoms consistent with at least one of the following co-occurring conditions: attention-deficit/hyperactivity disorder, major depressive episode, bipolarism, schizoid personality disorder, obsessive-compulsive disorder, generalized anxiety disorder, panic disorder, and agoraphobia. Six participants in the NT group reported symptoms consistent with at least one of the following conditions: generalized anxiety disorder, major depressive episode, alcohol use disorder, and eating disorder. These participants were included in the final sample to increase the representativeness of the sample, given the high prevalence of co-occurring conditions among autistic individuals (Mosner et al., 2019), as well as the high prevalence of common psychiatric disorders among the adult population in the United States (Grant et al., 2015, 2017; Gustavson et al., 2018; Hasin et al., 2018). See the Materials and Measures section in Supplemental Material S1 for more details about each of these measures and how the scores were calculated.

Data Analysis

Analyses were completed in SPSS (Version 27) and R (Versions 3.6.3 and 4.3.2).

Language Group Designations

Language status was determined by participants' LEAP-Q total L2 score. Participants who endorsed at least some L2 experience were included in the bilingual group, based on evidence that even limited exposure to an L2 is related to enhanced ToM skills (Fan et al., 2015); all other participants were categorized into the monolingual groups. One autistic participant who endorsed speaking at least some L2 did not fully complete the LEAP-Q and so did not have an L2 score but was included in the analysis addressing group differences.

The following four groups were used: (a) bilingual autism (n = 11), (b) bilingual NT (n = 14), (c) monolingual autism (n = 14), and (d) monolingual NT (n = 13). Each of these four groups was compared to examine differences on the following measures: SRS total T scores, ADOS-2 CSS, WASI-II, SP eye-tracking score, L2 score, number of co-occurring DSM-5 conditions identified on the SCID, age, socioeconomic status, sex, race, and ethnicity. The LEAP-Q responses of the bilingual autism and bilingual NT groups were also compared to provide a description of their language experiences.

Analysis for Research Question 1: Group Differences in RMET

Helmert contrasts were conducted to examine the following comparisons between the language groups on average RMET scores: (a) monolingual autism group compared to the mean of the other three language groups (bilingual autism, monolingual NT, and bilingual NT), (b) bilingual autism group compared to the mean of the two NT groups, and (c) the monolingual NT group compared to the bilingual NT group. EVIQ scores were included as a covariate due to significant differences in EVIQ identified between groups (see Table 1). The data were screened for assumptions. One outlier was identified using Mahalanobis distance screening. When the analysis was run with and without this participant, the results were essentially the same, so the results including the outlier are reported here. Assumptions of linearity, additivity, normality, homogeneity, and homoscedasticity were all met. Cohen's d effect sizes were also calculated to examine the magnitude of RMET score differences between groups. To further explore group differences in RMET scores, Helmert contrasts were completed using RMET valence scores (positive, neutral, negative).

Analysis for Research Question 2: Factors Contributing to RMET

A stepwise selection linear regression was used to determine the best model for predicting RMET scores in the autism group using EVIQ, L2, ADOS, and SRS as predictors. This method begins with the full model with all possible predictors and iteratively uses forward and backward selection strategies to find the combination of predictors that best explain an outcome variable (Draper & Smith, 1998). Three participants met the cutoff for Cook's and Leverage values; however, only one of these participants had multiple outlier indicators (i.e., two of three values used for screening: Mahalanobis distance, Leverage, and Cook's values). This outlier was found to be a highly bilingual participant. Analyses were run with and without this participant. Across both analyses, assumptions of linearity, normality, additivity, homogeneity, and homoscedasticity were all met. Among the autistic participants who completed eye tracking, two-tailed partial correlations were also performed to identify relationships between SP, L2, and autism symptomatology.

To further explore predictors of RMET performance, RMET valence scores (positive, neutral, negative) were included as dependent variables in the linear regression models. When assumptions were checked, two participants had multiple outlier indicators in both the positive and neutral valence conditions; however, no outliers were found for the negative valence condition. Among the outliers in the positive and neutral valence conditions, one outlier was monolingual with relatively low EVIQ, and the other outlier was highly bilingual with relatively high EVIQ. Analyses examining RMET positive and neutral valence were run with and without these outliers. Across all analyses, assumptions of linearity, normality, additivity, homogeneity, and homoscedasticity were met.

Analysis for Research Question 3: The Impact of SP on RMET When Moderated by L2 and Diagnostic Group

To inform analyses involving SP, a 2 × 2 factorial analysis of variance was conducted to examine the relative contributions of diagnosis (autism, NT) and language group (monolingual, bilingual) on SP, controlling for EVIQ. Levene's test for equality of variances indicated that there was a nonsignificant difference in variance of SP between language groups (F = 0.02, p = .89). Sex assigned at birth had a trending main effect on SP task performance (p = .09) and was thus used as a covariate in all analyses involving SP, along with EVIQ.

Next, when controlling for EVIQ and sex, SP was used to predict RMET scores when moderating for diagnostic group (autism or NT) and L2 scores. Data were checked for outliers. Two outliers were found in the NT group, which were the two individuals with the highest L2 scores. Because the focus of the analysis is on the impact of L2, analysis including these two outliers is reported here. See the Results section in Supplemental Material S1 for results excluding the two outliers. Data were checked for assumptions of regression, and no violations were found. The QuantPsyc package was used to center variables and analyze the interaction between variables for predicting RMET scores.

To further explore moderators of RMET performance, RMET valence scores (positive, neutral, negative) were included as dependent variables in the linear regression model. For the models including RMET positive and negative valence scores as the dependent variable, data were checked for outliers and assumptions of regression, and no violations were found. For the model including RMET neutral valence as the dependent variable, two participants had multiple outlier indicators. One participant was monolingual in the autism group with a relatively low EVIQ, while the other participant was highly bilingual in the NT group with a relatively high EVIQ. Analyses examining RMET neutral valence were run with and without these participants. Across all analyses, data were checked for assumptions of regression, and no violations were found.

Results

Language Groups

When LEAP-Q responses were compared, there were no significant differences between the bilingual autism or bilingual NT group for the following characteristics: dominant language (all English), second dominant language (L2; primarily Spanish, but also several other languages), or third dominant language (L3; primarily no L3 spoken, but some endorsed L3 in each group); first acquired language (primarily English, but also Spanish or Swiss German); age that L2 was acquired (ranged from birth to adulthood); proficiency in speaking, understanding, and reading L2 (ranged from very low to perfect); foreign accent (ranged from none to pervasive); preference for reading and speaking English versus L2 (primarily preferred English, but some endorsed approximately a 50:50 split); and current exposure to English versus L2 (primarily exposed to English, but some endorsed equal or more exposure to L2). Both bilingual groups indicated that they were currently exposed to L2 across a variety of activities, such as interacting with friends and family, reading, and listening to music. Both bilingual groups also indicated that a variety of factors helped them learn L2, such as interacting with family and friends, reading, and self-instruction. The same results were found among the eye-tracking subsample. See Supplemental Material S1, Table S1, in the online resource for more details regarding the LEAP-Q responses.

Research Question 1 Results: Group Differences in RMET

After adjusting for EVIQ, the monolingual autism group demonstrated a significantly lower average RMET score than the combined average RMET score of the other three language groups (bilingual autism, monolingual NT, and bilingual NT), t(47) = −3.10, p = .003. No significant differences were found between the bilingual autism group and the combined average RMET scores for the monolingual and bilingual NT groups, t(47) = −0.65, p = .52. Additionally, no significant differences were found between the monolingual and bilingual NT groups' average RMET scores, t(47) = 0.31, p = .76 (see Figure 1). Cohen's d effect sizes are presented in Table 2.

Figure 1.

Average RMET score by language and diagnostic group. Given p values correspond to the significance of the difference in RMET performance between either autism language group and the total comparison group. Figure created in R (Version 3.6.3). BA = bilingual autism; BC = bilingual comparison; MA = monolingual autism; MC = monolingual comparison; RMET = Reading the Mind in the Eyes Test.

Table 2.

Language group comparisons: adjusted RMET scores and effect sizes.

Variable	Group scores				Group comparisons
	Autism language groups		Comparison language groups		Monolingual vs. bilingual		Autism vs. comparison within language groups		Autism vs. comparison across language groups
	MA (n = 14)	BA (n = 11)	MC (n = 13)	BC (n = 14)	MA vs. BA	MC vs. BC	MA vs. MC	BA vs. BC	MA vs. BC	BA vs. MC
	M (SD)				Cohen's d effect sizes for group differences
RMET scores adjusted by EVIQ	22.48 (2.89)	24.94 (3.75)	25.80 (2.77)	25.46 (3.04)	0.75	0.12	1.17	0.15	1.00	0.26

Note. RMET = Reading the Mind in the Eyes Test; MA = monolingual autism; BA = bilingual autism; MC = monolingual comparison; BC = bilingual comparison; EVIQ = English verbal IQ.

In addition to overall RMET performance, group differences on RMET valence scores (positive, neutral, negative) were explored, adjusting for EVIQ. The monolingual autism group demonstrated significantly lower average RMET positive valence scores, t(47) = −2.62, p = .01, and neutral valence scores, t(47) = −2.40, p = .02, than the combined average RMET positive valence score of the other three language groups (bilingual autism, monolingual NT, and bilingual NT). When comparing the bilingual autism group to the combined average scores of the monolingual and bilingual NT groups, no significant differences were found on positive valence scores, t(47) = −0.32, p = .75, or neutral valence scores, t(47) = −0.24, p = .81. Additionally, no significant differences were found between the monolingual and bilingual NT groups' average RMET positive valence scores, t(47) = 1.00, p = .32, or neutral valence scores, t(47) = 0.03, p = .97. For the negative valence condition, no significant differences were found between any of the four groups (monolingual autism, bilingual autism, monolingual NT, and bilingual NT; ps > .48).

Research Question 2 Results: Factors Contributing to RMET

In the stepwise selection linear regression analysis including the outlier (highly bilingual individual), the best-fit regression model was significant and indicated that a combination of EVIQ and L2 (but not ADOS or SRS) significantly predicted RMET scores in the autism group, F(2, 21) = 5.47, p = .01, R² = .34, with EVIQ contributing uniquely to the model (see Table 3 and Figure 2). Participants with higher EVIQs were more likely to perform better on the RMET, b = 0.08, t(21) = 2.33, p = .03. Additionally, when accounting for EVIQ, participants with more L2 experience were more likely to perform better on the RMET, b = 0.13, t(21) = 1.67, p = .11 (Figure 2). In the analysis excluding the outlier, similar results were found: The best-fit regression model indicated that a combination of EVIQ and L2 (but not ADOS or SRS) significantly predicted RMET scores in the autism group, F(2, 20) = 6.28, p = .008, R² = .39, with both EVIQ, b = 0.08, t(20) = 2.44, p = .02, and L2, b = 0.23, t(20) = 2.17, p = .04, uniquely contributing to the model (see Supplemental Material S1, Figure S1).

Table 3.

Regression model: autism participants.

Variable	Including outlier: yes Dependent variable: RMET score
	Estimate (SE)	t value	p value	B	pr 2
Intercept	13.57 (3.72)	3.65	.002**	–	–
EVIQ	0.08 (0.03)	2.33	.03*	.43	.17
L2	0.13 (0.08)	1.67	.11	.31	.11
Model	F(2, 21) = 5.47, p = .01**, R2 = .34
	Including outlier: no Dependent variable: RMET score
Independent variables	Estimate (SE)	t value	p value	B	pr 2
Intercept	13.17 (3.66)	3.60	.001**	–	–
EVIQ	0.08 (0.03)	2.44	.02*	.43	.19
L2	0.23 (0.11)	2.17	.04*	.38	.21
Model	F(2, 20) = 6.28, p = .01**, R2 = .39

Note. RMET = Reading the Mind in the Eyes Test; SE = standard error; B = standardized regression coefficient; pr² = proportion of residual variance that is explained by adding that variable to the given model; EVIQ = English verbal IQ; L2 = second language score.

^* p < .05.

^** p < .01.

Figure 2.

Relationships between second language (L2) experience, English verbal IQ (EVIQ), and affective theory of mind measured using the Reading the Mind in the Eyes Test (RMET) within the autism group. The panel on the left depicts the combined impact of EVIQ and L2 on RMET; the panel on the right shows the unique contributions of L2 on RMET scores (controlling for EVIQ). Note that this figure presents findings when the outlier (highly bilingual individual) is included in the analysis. Figure created in R (Version 3.6.3).

Among the autistic participants who completed eye tracking, two-tailed partial correlations found that RMET performance was positively associated with L2 (r = .47, p = .049), above and beyond the effects of EVIQ, in contrast to SP (r = .05, p = .86) and autism characteristics measured via the SRS (r = .05, p = .85) or ADOS (r = .23, p = .36).

The following results were found when RMET valence scores (positive, neutral, negative) were explored as dependent variables in the stepwise selection linear regression model. In the model examining RMET positive valence scores, when outliers were included, the EVIQ variable, t(22) = 2.14, p = .04, best predicted RMET positive valence scores in the autism group, F(1, 22) = 4.56, p = .04, R² = .17. When the two outliers were excluded, a combination of EVIQ, t(19) = 2.38, p = .03, and L2, t(19) = 1.42, p = .17, best predicted RMET positive valence scores in the autism group, F(2, 19) = 4.27, p = .03, R² = .31.

In the model examining RMET neutral valence scores, when outliers were included, the L2 variable, t(22) = 1.55, p = .14, best predicted RMET neutral valence scores in the autism group; however, this model was not significant, F(1, 22) = 2.40, p = .14, R² = .10. This result was essentially the same when outliers were excluded from the model examining RMET neutral valence scores. When examining RMET negative valence scores, the EVIQ variable, t(22) = 4.05, p < .001, best predicted RMET negative valence scores in the autism group, F(1, 22) = 16.43, p < .001, R² = .43.

Research Question 3 Results: The Impact of SP on RMET When Moderated by L2 and Diagnostic Group

When examining the relative contributions of group status on SP, there was no significant interaction between diagnostic status and language group, F(1, 41) = 0.07, p = .80, η² = .002, or a significant simple main effect of language group (p = .91) on SP. The main effect for diagnostic status on SP was significant (p < .001), with NT participants showing greater SP than autistic participants (M = 0.17, SD = 0.7; M = 0.07, SD = 0.06, respectively). When accounting for differences in EVIQ and sex, there was not a significant correlation between SP and L2 (r = .06, p = .82), SRS scores (r = −.10, p = .70), or ADOS scores (r = −.39, p = .12).

Next, when examining moderators in the regression analysis, the overall model of EVIQ, sex, SP, diagnostic group, L2, and interactions between SP with group and L2 significantly predicted RMET, F(7, 38) = 5.80, p < .001, R² = .52. A significant main effect was found for EVIQ, with higher verbal IQ associated with higher RMET scores, b = 0.09, t(38) = 2.92, p < .01. RMET scores were also significantly predicted by an interaction between SP and L2, b = −1.80, t(38) = −2.20, p = .03 (see Table 4). When participants across diagnostic groups reported the lowest levels of L2, SP had the strongest impact on RMET scores, with higher SP predicting higher RMET performance. However, for participants reporting the highest levels of L2, SP did not positively predict RMET scores (see Figure 3). No three-way interactions between SP, L2, and group were found, t(−1.30), p = .20, indicating that there were no significant differences between the autism and NT groups in how L2 and SP were interacting to predict RMET scores.

Table 4.

Double moderation analysis across diagnostic groups including outliers.

Independent variables	Dependent variable: RMET Score
	Estimate (SE)	t value	p value	B	pr 2
Intercept	−3.54 (2.09)	−1.692	< 0.001***	—	—
EVIQ	0.09 (0.03)	2.92	< 0.01**	0.38	0.18
Sex	0.81 (1.07)	0.76	0.45	—	0.01
SP	6.59 (6.83)	0.97	0.34	−0.44	0.12
Diagnostic group	1.36 (1.13)	1.20	0.24	0.20	0.13
L2	−0.07 (0.07)	−1.04	0.32	−0.17	0.01
SP × diagnostic group	16.39 (13.82)	1.19	0.24	0.39	0.01
SP × L2	−1.80 (0.82)	−2.20	0.03*	−0.26	0.06
Model	F(7, 38) = 5.80, p < 0.001***, R2 = 0.52

Note. RMET = Reading the Mind in the Eyes Test; EVIQ = English verbal IQ; sex = male or female (male as reference group); SP = social prioritization of faces; diagnostic group = autism or neurotypical comparison; L2 = second language score; B = standardized regression coefficient; pr² = proportion of residual variance that is explained by adding that variable to the given model.

^* p < 0.05,

^** p < 0.01,

^*** p < 0.001.

Figure 3.

Impact of social prioritization of faces on Reading the Mind in the Eyes Test (RMET) performance and its moderation by second language (L2) experience. Figure created in R (Version 4.3.2).

To further explore RMET performance, RMET valence scores (positive, neutral, negative) were examined as dependent variables in the moderation analyses. When examining moderators of RMET positive valence, the overall model was significant, F(7, 38) = 4.97, p < .001, R² = .48. A significant main effect was found for sex, t(38) = −2.10, p = .04, in that men were more likely to have higher RMET positive valence scores. Significant main effects were also found for SP, t(38) = 2.09, p = .04, and L2, t(38) = 02.21, p = .03. An interaction between L2 and SP significantly predicted RMET positive valence, t(38) = −2.33, p = .03, in that SP more strongly predicted RMET positive valence scores among individuals with lower L2 levels (e.g., monolingual) but was less predictive of RMET positive valence scores among individuals with higher L2 levels (e.g., bilingual).

When examining moderators of RMET negative valence, the overall model was significant, F(7, 38) = 5.10, p < .001, R² = .48. A significant main effect was found for EVIQ, t(38) = 3.62, p < .001, with higher EVIQ predicting higher RMET negative valence. Similar to the RMET positive valence condition, a significant interaction was found between SP and L2, t(38) = −2.52, p = .02, in that SP more strongly predicted RMET negative valence scores when participants had low L2 levels (e.g., monolingual) but was less predictive when individuals had higher L2 levels (e.g., bilingual). Across both models examining RMET positive and negative valence, no three-way interactions between SP, L2, and group were found, indicating that there were no significant differences between the autism and NT groups in how L2 and SP were interacting to predict RMET positive and negative valence scores (ps > .44).

In the RMET neutral condition, when the two outliers were included in the analysis, the overall model did not significantly predict RMET neutral scores, F(7, 38) = 1.30, p = .28, R² = .19, nor were there any significant main effects (ps > .1). Essentially the same results were found when the two outliers were excluded (ps > .1).

Discussion

In the first study to examine the relationship between bilingualism and ToM in autistic adults, we found evidence for a bilingual advantage in affective ToM in autistic adults reporting a variety of L2 experiences and proficiency. This expands upon similar work in autistic children, which observed enhanced cognitive ToM in bilingual compared to monolingual autistic youth (Peristeri, Vogelzang, & Tsimpli, 2021). The task used to measure affective ToM, the RMET, requires participants to look at a photo of a person's upper face, infer what complex emotion the person is feeling, and select the correct emotion from a choice of four written words (Baron-Cohen et al., 2001).

As hypothesized for Research Question 1, autistic participants performed significantly worse on this task than participants in the NT group; however, when the autism group was divided into monolingual and bilingual speakers, the bilingual autism group identified complex facial expressions at similar levels of accuracy as the monolingual and bilingual NT groups when controlling for EVIQ, whereas the monolingual autism group performed significantly lower than the bilingual autism group and both NT groups. Further supporting this finding, effect sizes revealed small differences between the bilingual autism group and NT groups on ToM performance but large effect size differences between the monolingual autism group and all other groups.

In an exploratory analysis, these findings held true when facial expressions were examined in different valence categories (i.e., positive, neutral, negative). Specifically, the bilingual autism group identified all categories of facial expressions at similar levels as the monolingual and bilingual NT groups. Although the monolingual autism group identified negative facial expressions at similar levels as the other three groups, they performed significantly lower at identifying positive and neutral facial expressions as compared to the bilingual autism group and both NT groups.

In line with the hypothesis for Research Question 2, when controlling for EVIQ, autistic adults who reported more L2 experience tended to have higher ToM performance than those with low L2 experience. Together, a combination of EVIQ and L2 explained 34%–39% of the variance in ToM performance among autistic adults, with L2 explaining 11%–21% of the residual variance (depending on whether an outlier was excluded). However, in contrast with the hypothesis for Research Question 2, self-reported and clinically observed autism characteristics in autistic adults did not contribute to the model above and beyond EVIQ and L2 experience. EVIQ and L2 experience were consistently the most predictive for ToM performance in this group, more so than how frequently adults prioritized faces when viewing a social scene or how many characteristics of autism an individual reported or demonstrated. These findings indicate that stronger EVIQ and L2 experience in autistic individuals contributed to an increased ability to infer complex emotions from the upper face. In an exploratory analysis of emotional categories (positive, neutral, and negative valence), similar patterns emerged: Stronger EVIQ and L2 experience best predicted the ability of autistic adults to interpret positive facial expressions; however, only EVIQ best predicted the ability to interpret negative facial expressions.

Finally, supporting the hypothesis for Research Question 3, L2 experience moderated the relation between SP and affective ToM performance, when controlling for EVIQ. The complete model significantly explained 52% of the variance in ToM scores. Across all groups, among participants with the lowest levels of L2 (e.g., monolingual), more frequent eye gaze toward faces while watching videos of social interactions predicted how well participants were able to complete the task. This finding suggests that, as would be expected, more time spent gazing at faces corresponds to better ability to infer complex emotions during the affective ToM task, regardless of whether a person has autism or not. However, when participants across both groups reported higher levels of L2 (e.g., more bilingual experience), prioritizing faces while viewing social scenes had less of an impact on how well participants performed on the affective ToM task. This pattern held true across facial expressions with positive and negative valence.

Taken together with findings from Research Questions 1 and 2, these results suggest that L2 experience enhances affective ToM skills among autistic adults to the level of NT adults, while also reducing the relative importance of gazing at faces as a factor in identifying complex emotions. Positive facial expressions, in particular, may play an important role in bilingualism, given that bilingual autistic adults were better able to identify these emotions when compared to monolinguals and that bilingual experiences significantly contributed to identification of positive facial expressions.

When considering the role of EVIQ, which indexes knowledge of both vocabulary and metalinguistic skills, it is important to consider how both components may impact complex emotion attribution in autistic adults. In terms of vocabulary, participants needed knowledge of emotion words to complete the ToM task. Written definitions of the answer choices (i.e., four complex emotions) were provided to participants in case they were unfamiliar with the words; nonetheless, it seems clear that stronger English vocabulary would help participants understand their answer choices. Indeed, the task used in this study has been shown to rely in part on receptive and semantic language skills (Pavlova & Sokolov, 2022). Beyond supporting understanding of emotion vocabulary, it may be that English verbal skills support affective ToM through additional mechanisms. Greater EVIQ may reflect stronger metalinguistic skills, which are enhanced in bilingual individuals (Abu Rabia, 2019). Future studies should consider the extent to which specific domains of language (e.g., receptive and expressive semantics, syntax, metalinguistics) contribute to affective ToM performance.

Importantly, a combination of EVIQ and L2 significantly predicted affective ToM scores better than EVIQ alone, when both including and excluding an outlier from the analysis. Additionally, L2 and SP significantly interacted to predict affective ToM performance while controlling for EVIQ, providing further support that affective ToM skills measured on the RMET relied on more than verbal IQ skills alone. However, future research would benefit by examining the relationship between L2 experience and measures of ToM that rely less on language ability to provide greater insight into the mechanisms that support perspective taking and emotion recognition in autism (Abu Rabia, 2019).

In terms of the role of L2, the bilingual advantage in affective ToM for adults with autism dovetails with the small body of research among bilingual and bilingually exposed autistic children, which has found enhanced abilities in the following areas: selective attention when shifting sets during the dimensional change card sort (DCCS) task, but not during activities of daily life (Gonzalez-Barrero & Nadig, 2019); sustained attention, interference control, flexible switching, and working memory when rated by parents, but not by teachers (Sharaan et al., 2022); semantic fluency when generating words in a category (i.e., producing more correct words; Gonzalez-Barrero & Nadig, 2017); adaptive socialization skills (Hastedt et al., 2023); cooing; gesturing to make requests (i.e., pointing, leading caregivers to a desired object); and simple pretend play (i.e., pouring; Valicenti-McDermott et al., 2013). Similar to the current results are the studies conducted by Gonzalez-Barrero and colleagues examining selective attention (2019) and semantic fluency (2017) among elementary school–aged monolingual and bilingual autistic and NT children. Findings revealed an advantage for bilingual autistic children relative to monolingual peers during both the structured DCCS task and word generation task, but no bilingual advantage among the NT children (Gonzalez-Barrero & Nadig, 2017, 2019). Similarly, Peristeri, Vogelzang, and Tsimpli (2021) observed heightened ToM and EF skills among bilingual autistic children compared to monolingual autistic children. The present findings also indicated a bilingual advantage among autistic adults, and importantly, this protective effect was found even when the bilingual autism group included participants with very low levels of L2 experience and proficiency. This finding mirrors a study of NT youth, in which even children who had limited L2 exposure and proficiency demonstrated enhanced performance on a ToM task relative to monolingual peers (Fan et al., 2015).

Possible Mechanisms for Enhanced Affective ToM Skills Among Autistic Bilinguals

The literature in non-autistic bilingual populations has proposed three possible mechanisms for the bilingual advantage seen in ToM skills: (a) metalinguistic awareness, (b) EF, and (c) sociopragmatic accounts (Schroeder, 2018). Examining research within each of these factors will help elucidate bilingual characteristics and experiences of autistic individuals that may, in turn, contribute to the development of ToM and, specifically, the ability to interpret complex facial expressions.

According to metalinguistic accounts of ToM, comprehending the symbolic nature of language strengthens the understanding that an event itself is distinct and separable from an individual's perception of the event and that there could be many different perceptions about the same event (Schroeder, 2018). When monolingual autistic children demonstrate stronger functional and symbolic play skills, whether naturally or through play-focused interventions, they tend to make gains in L1 communication skills (Charman et al., 2003; Dykstra et al., 2012), which then contribute to later L2 skills (Gutiérrez-Clellen et al., 2012; Schmid & Yılmaz, 2018). Bilingually exposed autistic toddlers also demonstrate enhanced functional play (i.e., simple pretend play that is a precursor to symbolic play) when compared with monolingual peers (Valicenti-McDermott et al., 2013), raising the possibility that that metalinguistic awareness (e.g., understanding the symbolic nature of language) could be positively impacted by L2 exposure. Future research could examine the symbolic play and metalinguistic skills of autistic youth to examine the role of metalinguistic awareness in the development of bilingualism and ToM skills.

There is also evidence to suggest that autistic bilinguals may have greater selective attention than their monolingual peers (Gonzalez-Barrero & Nadig, 2019). If enhanced selective attention to social stimuli and related predictive coding systems were identified as mediators between bilingualism and ToM in autistic individuals, it would lend support to an EF account of a bilingual autism advantage. To explore this speculation, it would be valuable for future research to examine both behavioral and neurological mechanisms of selective attention and predictive coding systems of autistic bilinguals, particularly during social experiences, and how these systems relate to L2 experiences and complex emotion recognition, including different categories of emotions. In the current study, participants identified negative facial expressions at similar levels, regardless of group or bilingual abilities; however, bilingual experiences appeared to play a stronger role in identifying positive facial expressions. It may be that negative facial expressions are more salient than positive facial expressions (Palmer et al., 2023). Indeed, in previous studies, a stronger negativity bias has been found in some autistic adults compared to neurotypical adults, where participants were more likely to attribute negative valence to neutral expressions and neutral valence to positive expressions (Eack et al., 2015), although these findings are inconsistent across studies (Bergman et al., 2020). Having greater accuracy in identifying negative emotions could be protective, as attending to these expressions may be especially important for an individual's safety during social interactions, whereas accurate identification of positive emotions has implications for developing social relationships (Eack et al., 2015). Nuanced investigations of factors that impact the identification of emotions in ASD continues to be an important area for future research, including the interplay between bilingualism and selective attention to relevant social cues.

Another benefit of selective attention to social cues is that these cues could help individuals make social predictions, increasing motivation to seek out more positive social interactions (Schultz, 2000, 2015; Schultz et al., 1997). Autism is characterized by decreased attention to social stimuli, such as eye gaze, which may reflect a lack of motivation for social stimuli (Chevallier et al., 2012) or altered salience processing of social stimuli (especially in comparison to environmental sensory stimuli; Bast et al., 2023). Both of these theories are supported by neuroimaging studies indicating deficits in social and nonsocial reward processing (Bottini, 2018; Clements et al., 2018; Dichter, 2018), as well as differences in sensory processing that impact emotion recognition (Patterson et al., 2021). The socio-pragmatic theory for the bilingual advantage in ToM posits that ToM development is based on bilingual social experiences (Schroeder, 2018). Motivation could, therefore, play an integral role in both EF and socio-pragmatic accounts of a bilingual ToM advantage. More research is needed to explore the extent to which motivation plays a role in the EF account of a bilingual autism advantage in ToM (i.e., through selective attention and reward processing systems) and the socio-pragmatic account (i.e., through engaging in motivating social activities).

Increased attention to nonverbal cues, especially facial expressions, may be particularly important for those who are learning a new language; L2 learners may selectively attend to the mouth to support comprehension of an unfamiliar language, in contrast to relying only on auditory information, while the ability to correctly identify facial expressions in others would likely provide additional context that would aid an L2 learner in their understanding of the new language. Indeed, increased attention to the area directly beneath the eyes has been associated with increased performance on face and emotion recognition tasks (Peterson & Eckstein, 2012). As such, SP to both the eyes and mouth of others may be important factors underlying L2 development. This possibility is relevant in the context of the current study, given that higher L2 experience reduced the extent to which SP of faces predicted affective ToM performance. The eye-tracking task used in this study only allowed for examination of faces as a whole. Future studies could extend these findings by examining nuanced parts of the face, such as eyes and mouth, to discover more precise patterns of how L2 and SP interact with one another and, ultimately, influence an individual's understanding of complex facial expressions. This is an important consideration in autism especially, as differences in salience processing, motivation, and attention may interact in ways that promote alternative strategies for promoting affective ToM.

Another consideration for the socio-pragmatic account of ToM in bilingualism relates to sociolinguistic awareness, or knowing when to code-switch between languages based on the needs of their listener (Yu et al., 2021). Previous research has found that, among typically developing 4-year-olds, sociolinguistic awareness is related to bilingualism and also predicts performance on false belief tasks (Cheung et al., 2010). To foster sociolinguistic awareness, it is possible that adults scaffold their child's understanding of which language to use (e.g., Spanish with abuela, English with “the teacher”)—a skill that directly relates to ToM and could be strengthened by interpreting other's facial expressions when speaking.

Interestingly, although the majority of bilingual autistic participants reported that they (a) were currently exposed to their L2 much less than English (less than the 40%–60% level of exposure recommended for proficiency; Gonzalez-Barrero & Nadig, 2018) and (b) had less than adequate L2 speaking proficiency, the bilingual autism group still demonstrated enhanced abilities to recognize complex emotions as compared to the monolingual autism group. This suggests that even limited experiences with another language may positively impact affective ToM skills.

Future research should explore which qualities of social interactions contribute to sociolinguistic awareness and, subsequently, development of affective ToM. For example, bilingual interactions could vary based on the timing of when L2 is learned (e.g., simultaneous vs. sequential with L1), the teaching strategies that are used to support L2 development, and the environments in which L2 is encountered (e.g., one language per environment; multiple languages per environment, but one primary language per speaker; or multiple languages per environment and speaker, requiring intense code-switching; Luk, 2023). Each scenario has the potential to impact sociolinguistic awareness and affective ToM skills in different ways.

This study had several limitations. First, because the participants were adults who had at least average IQ, it is unknown whether a bilingual ToM advantage would be seen in other age groups or those with lower intellectual functioning levels. Additionally, the current study did not prospectively examine the development of English, L2, and ToM over time, so we cannot make concrete conclusions about the relative contributions of one construct to the other. During the study visit, a subset of participants in each group reported symptoms of co-occurring DSM-5 conditions. Future studies are needed to determine if these results would be replicated among adults without co-occurring conditions; however, including these participants allowed for a more representative sample, especially given the high frequency of co-occurring conditions among autistic individuals (Mosner et al., 2019). Another limitation is the small sample size of each language group; however, the fact that the groups demonstrated significant differences, even at this small sample size, is notable.

Self-reported L2 experiences and proficiency levels provided valuable information about the context in which individuals acquired and were currently using their languages. However, direct measurements of L2 proficiency, in addition to self-report, would also be informative to include in future studies. Additionally, because the current study included a range of language profiles, including many autistic participants who first learned English followed by their L2, future research should examine the extent to which bilingualism enhances ToM among targeted language profiles. To address health care disparities, it is particularly important for research to focus on the bilingual development of children from culturally and linguistically diverse families, who could benefit from speaking both their family's heritage language and the language of mainstream society. While there were no group differences in racial background in this study, it is also important to note that there was limited racial/ethnic diversity in our autism sample. There is some evidence that sex also impacts language development and use (Wallentin, 2020) and should thus be a focus in future research examining sex differences in the relationship between language and ToM in autism. For example, in the current study, sex was identified as a factor in identifying positive facial expressions; however, given that the sample did not include any bilingual autistic women, future studies are needed to elucidate this relationship.

Importantly, the task used to measure ToM displayed static black-and-white images of the upper face, rather than providing a more naturalistic measure of ToM performance. Two studies of bilingualism in autistic individuals found that bilingual advantages did not generalize across settings, such as to parent reports of daily routines (Gonzalez-Barrero & Nadig, 2019; Sharaan et al., 2022). Given broader difficulties with generalization among autistic individuals (de Marchena et al., 2015), it will be important for future studies to consider the extent to which bilingualism impacts the ToM performance of autistic individuals in “real-life” settings.

Finally, the exploratory results pertaining to emotional valence (positive, neutral, negative) should be interpreted cautiously, given that the classification of these categories still requires replication (Palmer et al., 2023). Future research should continue to examine whether bilingualism in autism confers an advantage in affective ToM skills differentially based on the type (Martzoukou et al., 2017) or valence (Palmer et al., 2023) of emotional stimuli, as difficulties in emotion recognition in autism may not encompass all emotions.

In summary, bilingual autistic adults who reported a range of L2 experiences outperformed monolingual peers on a ToM measure of complex emotion recognition, performing at similar levels as monolingual and bilingual NT groups. Furthermore, when controlling for EVIQ, more experiences and proficiency with L2 among autistic individuals (a) predicted better ability to interpret complex facial expressions and (b) reduced the relative importance of eye gaze toward faces for interpreting complex facial expressions. Together with previous research about EF (Gonzalez-Barrero & Nadig, 2019), semantic fluency (Gonzalez-Barrero & Nadig, 2017), and early nonverbal communication and play skills (Valicenti-McDermott et al., 2013) in bilingual and bilingually exposed autistic children, these findings support emerging evidence of a bilingual advantage in autistic individuals and extend findings of a bilingual advantage in ToM skills found in NT populations (Schroeder, 2018) and autistic children (Gonzalez-Barrero & Nadig, 2019; Peristeri, Baldimtsi, et al., 2021). Future research should examine mechanisms facilitating this bilingual advantage in autistic individuals, along with interventions for supporting the bilingual development of autistic children from culturally and linguistically diverse families at risk for health care disparities, so that children can benefit from the rich social and linguistic support of their families and communities.

Compliance With Ethical Standards

Ethical approval. All procedures performed involving human participants were in accordance with the ethical standards of the institutional review committee and with the 1964 Helsinki declaration and its later amendments. The study was approved by the institutional review board at The University of North Carolina at Chapel Hill.

Informed consent. Informed consent was obtained from all individual participants included in the study.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Funding Statement

This research was supported by National Institute of Mental Health Grant R21 MH110933. Assistance with recruitment was provided by the Clinical Translational Core of the UNC Intellectual Developmental Disabilities Research Center Grant P50 HD103573 (awarded to Gabriel Dichter). Funding sources had no direct involvement in the study design, collection, analysis, or interpretation of data; the writing of the article; or the decision to submit the article for publication.