Measuring the Impact of Bilingualism on Executive Functioning Via Inhibitory Control Abilities in Autistic Children

89 children (n = 38 autistic, n = 51 non autistic) participated in this study. The inclusion criteria required that all participants were exposed to more than one language, with exposure referring to individuals who spoke and/or received secondary language input at home and/or school (see Table 1 below for the participant dominant language percentage, and the Appendix for further details). However, this exposure varied widely and verbal fluency in one or more languages was not mandatory as a prerequisite for participation. Limits were not placed on IQ scores in the inclusion criteria in order to ensure a more representative autistic sample. However, IQ was included as a covariate in subsequent analyses.

Participating families were recruited throughout Scotland and England through a variety of methods including school networks, charity participant databases, and magazine and poster advertisements. Autistic participants were required to have a clinical autism diagnosis, confirmed by parents. As a further confirmatory measure, autistic children were also assessed on the ADOS-2 (Lord et al., 2012) and all parents completed a Social Communication Questionnaire (Rutter et al., 2003). A total of 33 children completed an ADOS-2, with five children unable to participate as they had very recently completed an ADOS-2, or due to practical constraints at home visitation. Out of this group, three children did not receive an ADOS-2 algorithm score above the likelihood threshold for a diagnosis of autism, but all of the autistic group scored above the SCQ screening threshold. All non-autistic children scored below a 7 on the SCQ, indicating that the two participant groups could be distinguished reliably by diagnostic status.

Detailed demographic information divided by Participant Group is provided in Table 1.

Attentional interference control was measured using reaction time in the Eriksen Flanker Task (Eriksen & Eriksen, 1974), a computer-based spatial attention paradigm. Trials were differentiated between congruent and incongruent, with the latter corresponding to the central target facing in a conflicting direction to peripheral distractors. Reaction times of correct responses were recorded, as output of accuracy has been shown to lead to ceiling effects, exhibiting a reduced variance in collected data (e.g., Duthoo et al., 2014). Cartoon fish stimuli (height: 2 cm) acted as both the target and distractor stimuli for this task, identical apart from pointed gaze direction and screen position. Longer reaction times indicate poorer attentional interference control.

Prepotent response inhibition was measured by accuracy (specifically, false start rate) on the Psychomotor Vigilance Task (Dinges & Powell, 1985) through inaccurate button presses indicating the presence of a target. A red stopwatch (height: 1.8 cm) positioned on the centre of a screen was presented as the target stimulus. The timer counts with numerals upwards in milliseconds before disappearing and being replaced by a reaction time speed, upon the participant responding to the stimulus accordingly. Lower false-positive rates indicate better prepotent response inhibition. A 15-inch Laptop was used to administer both the psychomotor vigilance and flanker task. Inquisit 5 experiment generator software (2016) presented the stimuli and recorded participants’ responses.

EF in the home environment was measured using a parental questionnaire; the Behavior Rating Inventory of Executive Function Second Edition (Gioia et al., 2000). The questionnaire comprises 63 items scored as: often, sometimes and never. The output can be divided into Behavioural Regulation Scales (Inhibit, Shift and Emotional Control) and Metacognition scales (Initiate, Working Memory, Plan/Organise, Organisation of Materials, and Monitor). Higher scores indicate increased levels of executive dysfunction in their respective domain. Only Inhibit and Working Memory subscale scores were analysed in this study.

The bilingual experience of participating children was captured by the short version of the Bilingual Language Exposure Calculator (BILEC) parent questionnaire (Unsworth, 2013). Language exposure was measured by the number of hours each language was used at home, (including after school, at weekends, and during the holidays) during the school day, and with friends. By placing participants on a continuum of language exposure, it was possible to directly compare experience without the loss of variance caused by splitting participants into distinct exposure level groups. Bilingual exposure was calculated from the Bilingual Language Experience Calculator as a percentage ranging from 0 to 100% and measured only through language input.

In order to gather a measure of IQ, participants completed the Wechsler Abbreviated Scale of Intelligence II (WASI-II; Wechsler, 2011). Only the vocabulary (31 items) and matrix reasoning (30 items) subtests were used, sufficient to calculate a partial IQ score and as an estimate of general cognitive ability. IQ limits were not stated within the inclusion criteria in order to permit a representative autistic sample. However, IQ was included as a covariate in subsequent analyses along with age.

Receptive vocabulary was assessed using the British Picture Vocabulary Scale Third Edition (Dunn & Dunn, 2009). The assessment measures age-appropriate receptive vocabulary abilities. Participants are instructed to match a word spoken by the examiner to one of four pictures using non-verbal responses.

The Autism Diagnostic Observation Schedule, 2nd edition (ADOS-2; Lord et al., 2012) is a standardised assessment tool that is semi-structured. It is used to measure social and communication behaviours which contribute towards a diagnosis of autism. Activities are administered from one of four modules. Modules are selected as a result of language and developmental level.

The SCQ-Lifetime (Rutter et al., 2003) is a parent-administered questionnaire which may function as an initial screening measure for autism. This measure accounts for the developmental history of the individual. Scores greater than 15 suggest higher-than-average levels of autistic traits.

Ethical approval was obtained by the University of Edinburgh Psychology Research Ethics Committee (336-1718/5). Informed consent from parents and caregivers was recorded electronically before the appointment and signed physically on the day of assessment. Children were also asked to provide verbal assent prior to participation. Data was collected in a single session by the research team. Data was collected through two-hour home visits so as to maximise ecological validity and ensure time for a break. As a reward, children received a bag of sensory toys upon completion of the session.

Within the task battery, most participants first completed the Psychomotor Vigilance Task and then immediately after, the Flanker Task. Parents completed the BRIEF within two weeks prior to the home visit. Full detail on the administration protocol for the Flanker and Psychomotor Vigilance tasks are included in the Supplementary Appendix Materials.

Flanker Task data were processed prior to analysis to exclude trials in which participants scored inaccurately (13.84%) or responded within 200 ms or less (0.12%). PVT data were not treated prior to analysis. For all independent variables, data were standardized to have a mean of 0 and a standard deviation of 1, allowing all analyses to be compared on the same scale.

To capture individual variability, we used mixed effects modelling to examine the influence of bilingual exposure on inhibitory control tasks. We used backwards selection to find the best fitting regression models for our data, beginning with models that incorporated all variables as fixed effects. For mixed models, we also incorporated by-participant random intercepts. We removed predictor variables through a backwards stepwise selection procedure based on the model that favoured the Akaike Information Criterion (AIC). Marginal R squared for fixed effects, and where appropriate conditional R squared for both fixed and random effects, were calculated for each final model to act as a measure of model fit and variance explained. Our only constraint on model selection was the conservative decision to always incorporate Age and IQ into the regression, as these variables were not matched between participant groups, and each can potentially modulate task performance. All analyses were conducted in R with mixed models fit using the lme4 package (Version 3.3.1; Bates et al., 2015).

For each experimental task, our initial mixed effects model included fixed effects of Group (autistic/non-autistic), Bilingual Exposure, Age and IQ. For the Psychomotor vigilance task, where the outcome variable (false starts) was binary, we used a mixed logistic regression, and incorporated a two-way interaction term between Participant Group and Bilingual Exposure. For the Flanker Task, where RTs was the outcome variable, we used a mixed linear model, and included an additional predictor of Trial Type (congruent/incongruent) with a three-way interaction term between Participant Group, Bilingual Exposure and Trial Type. Both models additionally included by-participant random intercepts.

Due to the standardised scoring of the BRIEF, simple multiple linear regression was deployed. BRIEF Inhibit and Working Memory subscale scores were fitted with multiple linear regression using Participant Group (autistic/non-autistic), Bilingual Exposure, Age and IQ as independent variables. Also included was a two-way interaction term between Participant Group and Bilingual Exposure.

We used first order Spearman correlations (robust to the presence of extreme values) to test associations between parent reports and EF from the experimental tasks. As multiple analyses were run, a Bonferroni adjustment was applied (alpha value of 0.05/8 = 0.006).

Underlying assumptions were checked and validated for all analyses. The Flanker Task violated the assumptions of normality and homogeneity of variance. To resolve, outliers were removed, using a cut-off threshold of 2.5 standard deviations from the mean. Different transformations of the dependent variable were then tested with a log transformation selected as the most appropriate adjustment. Upon being rechecked, data passed visual checks for normally distributed residuals and demonstrated constant variance. The pattern of results reported here was the same when run with unadjusted Flanker Task data.

For the PVT Task outliers were similarly removed, considered as observations greater than or equal to Cook’s D of 4/n and calculated from simple logistic regression. Upon visual inspection only influential values were removed by this process.

For BRIEF analyses outliers were removed, specified as greater than or equal to Cooks D of 4/n with additional visual leverage checks undertaken. Results further excluded 3 participants whose parents failed to score as Acceptable on the BRIEF Inconsistency Scale.

The best fitting model incorporated predictors for Age, IQ, Bilingual Exposure and Trial Type with a summary presented in Table 3. Increased Age and IQ were associated with shorter RTs, whereas greater bilingual exposure resulted in marginally longer RTs. Incongruent trials elicited longer RTs than congruent trials. Participant Group was not included in the best model, nor did it have a significant effect.

No higher-order interactions were identified as significantly predicting reaction time, meaning that the difference between congruent and incongruent trials did not significantly vary based on Age, IQ, or Bilingual Exposure. Bilingual exposure was significant using approximate p-values (0.047), however upon calculating more precise p values based on conditional F-tests with Kenward-Roger approximation for the degrees of freedom, bilingual exposure failed to reach significance (0.051).

The best fitting model incorporated Bilingual Exposure, Age, and IQ as significant predictors on the likelihood of producing a False Start. A summary of this model is shown in Table 4. Increased Bilingual Exposure (see Fig. 1) and Age were associated with a lower likelihood of false start rate. Higher IQ was associated with a higher likelihood of false start button presses. Again, Participant Group was not found to be significant, nor included in the best model.

A multiple regression (n = 74) was run to predict inhibitory control scores using Age, IQ, Participant Group, and Bilingual Exposure. Only IQ and Participant Group were selected as predictors in the best model which significantly predicted inhibitory control scores, F(2,71) = 36.90, p = 0.00, R² = 0.51. An identical control model (N = 74) was run with Working Memory scores as the dependent variable. The best model predicted Working Memory scores, F(2,71) = 41.58, p = 0.00 R² = 0.54, with IQ and Participant Group significantly contributing as independent variables. In each model, lower IQ corresponded to higher BRIEF scores (more executive dysfunction), and non-autistic participants’ scores suggested higher levels of executive function skills than the autistic group. Results from both models are presented in Table 5 and illustrated in Fig. 2a, b.

Spearman correlations assessed the strength of relationship between EF in experimental tasks alongside BRIEF scores. Table 6 presents the results of the correlations. BRIEF questionnaire Inhibit and Working Memory subscale scores are correlated with accurate incongruent Flanker Task reaction times and the total number of false starts, listed by group. No correlations are significant upon applying a Bonferroni correction.

A main aim of this investigation was to explore the relationship between bilingualism and executive inhibitory processes in autistic and non-autistic children. The results suggest that accurately managing behavioural inhibition of motor functions is modulated through bilingual language use: the probability of initiating a PVT false start on an experimental trial was reduced by increasing bilingual exposure and this effect was consistent regardless of autism diagnostic status. We also found a strong correlation between parent reports of inhibitory control and PVT false start rate, but this finding was not significant. However, analysis suggests that bilingual exposure does not meaningfully impact inhibitory thought processes when it comes to visual selection and the suppression of irrelevant information. Although an effect of increasing bilingual exposure was included in the best FT model as slowing reaction time of attentional control, the factor is non-significant. In addition, the absence of group effects on performance on experimental tasks was surprising and may suggest that the lack of group difference is a result of bilingualism mitigating EF impairments for autistic children. Our results were similar to that of Sharaan et al. (2020) who also found no effect of diagnosis group on PVT false start rate. However, more surprising was that their study found no main effect of language group on false start rate, although they did identify an interaction where autistic bilinguals showed significantly lower mean false starts than autistic monolinguals. This is somewhat in contradiction to the results of our study where bilingual exposure increased performance regardless of diagnostic status. The reason for the discrepancy could be due to a number of factors including power and approach to analyses (mixed modelling vs ANOVA), differences in group conceptualisation (bilingual exposure vs monolingual/bilingual groups) or IQ inclusion criteria (no exclusion vs only participants who score as intellectually average or above).

Why might bilingualism interact with one aspect of inhibitory control but not another? Generally, it is agreed that inhibitory control is not a cohesive construct, but instead a collection of cognitive processes grouped by a shared function (Roberts et al., 2011). Differences between these cognitive processes exist in terms of how mental processes are applied to tasks, and in brain circuitry (Mostofsky et al., 2003), thus explaining why bilingualism may modulate operations separately. The two functions can be teased apart in a number of ways. First, inhibitory control involving oculomotor response and deployment of visual attention (as in the Flanker Task) is executed independently of inhibitory control involving more physical motor regulation (as in the PVT) (Aron et al., 2004). Second, selective filtering of attention whilst overlooking irrelevant distractors (as in the Flanker Task) is goal directed and therefore requires top-down processing (Pinto et al., 2013), whereas regulation of impulses (as in the PVT) is typically automatic, incorporating some elements from bottom-up processes. Lastly, conflict tasks like the Flanker Task have a unique relationship with working memory (Carlson et al., 2002) that is not present with delay tasks like the PVT, in the sense that perceptual features are held and internally manipulated via the selection of a desired mental representation. Consequently, it is possible that our two measures are underpinned by distinct cognitive mechanisms that interact with bilingualism differently.

Our findings suggest that any bilingual advantage for inhibitory control is evident under specific conditions only: namely through the emergence of sustained attention where the behavioural regulation of motor impulses is required. In contrast, our results cast doubt on the idea that bilinguals’ exercise of inhibitory control, via activation of a target language and suppression of a non-context appropriate language, benefits selective visual attention. The findings accentuate the need to use multiple measures when evaluating a cognitive ability as a means of gathering converging evidence.

These results do not contradict the possibility that children with stronger executive skills may be more likely to learn multiple languages, and thus offer more positive response overall to bilingual input, explaining potential biases in our sample. However, our inclusion criteria were as inclusive as possible, including children with learning disabilities and children with a wide range of language competency, giving some credence to a causal relationship between bilingual exposure on EF. A more counterintuitive finding is that as the IQ of participants increased, the likelihood of producing false starts also increased. Researcher observations during data collection suggest that children with higher IQs may have been more engaged with the PVT task in general, and as a result prone to accidental button presses through trying to predict when the target would appear as quickly as possible. In contrast, children with lower IQs demonstrated less engagement during data collection and would often look away from the task until the target was clearly fixed on the screen, despite previously being told they should complete the task as quickly and accurately as possible. This result would perhaps not have arisen had we implemented an inclusion criterion based on IQ above a certain threshold. Furthermore, although the WASI-II has been found to characterise meaningful construct validity (Canivez et al., 2009), caution must be taken so as not to over-interpret the results from a brief and albeit rapid estimate of general functioning which cannot replace a full intellectual quotient.

The hypothesis that non-autistic people outperform autistic peers on measures of attentional interference control and prepotent response inhibition is not supported by the results. Here, we find no significant difference in performance between the two groups on either the FT or PVT. This is despite substantial evidence and consensus within the academic community that executive dysfunction is widespread in autism and may underlie the core presentation used in diagnosis (e.g., Russell, 1997; Xie et al., 2020). For example, various studies indicate that autistic children perform poorly on distractor interference (e.g., Geurts et al., 2008) and on prepotent response inhibition (e.g., Bíro & Russell, 2001), with the latter being linked to repetitive and stereotyped behaviour (Schmitt et al., 2017).

The fact that an autistic disadvantage is often found in tasks of attentional interference control and prepotent response inhibition, but not in our sample of participants subject to bilingual exposure, may suggest a bilingualism effect compensating autistic difficulties in EF. This supports the idea that bilinguals’ regular engagement with inhibitory thought processes and language switching can be protective against behavioural and cognitive inflexibility (Faja & Darling, 2018). Yet, this finding is not replicated in terms of parent perception of EF—in this case significant relations were found with autism diagnostic status and IQ only. Still, various studies have suggested that caregivers may not always measure the skills of their children accurately (Mattson et al., 2020) and uniting the differences between experimental outcomes and parental perception continues to pose problems for forging consistent patterns of results.

The results of this study are specific to considering bilingualism as exposure to more than one language in terms of listening and should be generalised as such. The decision was taken to include only bilingual input into analyses, rather than output, because the autistic participants in this study encompass a considerable range in ability level, from highly verbal to minimally verbal with learning difficulties. Certainly, it is possible that listening to a second language demands fewer mental resources than actively speaking and this could be expected to confer different or weaker cognitive benefits. In this context, the observed impact of bilingual exposure on prepotent response inhibition in our data is perhaps surprising, and important.

One potential drawback of our decision to study bilingualism as a continuum, rather than through the traditional design of comparing bilingual and monolingual groups, is that a latent effect of bilingualism could be biasing the results in the sense that multiple language exposure may have enhanced EF throughout the sample. A further limitation is that our sample didn’t include trilinguals. Remarkably, trilinguals have demonstrated several distinct cognitive differences compared to bilinguals. For example, trilingual adults are less likely to develop a mild cognitive impairment compared to bilinguals and therefore may be bettered safeguarded against cognitive aging (Perquin et al, 2013). Could then a trilingual group have produced different results to our recruited bilinguals? Possibly, but it has already been shown that while outperforming monolinguals, trilinguals do not show an advantage on the Flanker Task compared to bilinguals (Schroeder & Marian, 2017). Nevertheless this would have been interesting to capture with the inclusion of the bilingual exposure effect found in the PVT. Lastly, standardised clinical assessments may have provided more valid and replicable results. However, the measures for this project were adapted from a larger study which aimed to engage non-autistic and autistic participants alike through activities in their own home, and therefore the measures used was not purely clinical.

Future research related to inhibitory control should clarify the mechanism underlying the engagement and disengagement of attention. Otherwise, investigation is still required to ascertain the effects of bilingual exposure on various aspects of EF for autistic children as at this time such literature is scarce and our findings cannot be generalised globally.