Do autistic children differ in language-mediated prediction?

Prediction appears to be an important characteristic of the human mind. It has also been suggested that prediction is a core difference of autistic1 children. Past research exploring language-mediated anticipatory eye movements in autistic children, however, has been somewhat contradictory, with some studies finding normal anticipatory processing in autistic children with low levels of autistic traits but others observing weaker prediction effects in autistic children with less receptive language skills. Here we investigated language-mediated anticipatory eye movements in young children who differed in the severity of their level of autistic traits and were in professional institutional care in Hangzhou, China. We chose the same spoken sentences (translated into Mandarin Chinese) and visual stimuli as a previous study which observed robust prediction effects in young children (Mani & Huettig, 2012) and included a control group of typically-developing children. Typically developing but not autistic children showed robust prediction effects. Most interestingly, autistic children with lower communication, motor, and (adaptive) behavior scores exhibited both less predictive and non-predictive visual attention behavior. Our results raise the possibility that differences in language-mediated anticipatory eye movements in autistic children with higher levels of autistic traits may be differences in visual attention in disguise, a hypothesis that needs further investigation.


Introduction
Prediction has become the dominant theoretical framework for explaining the functioning of the human mind (Bar, 2007;Clark, 2013). It is therefore perhaps not surprising that autism, a complex developmental neurodivergent phenomenon characterized by difficulties with social interaction and communication as well as restricted and repetitive behaviors (Frith & Happé, 2005;Lord et al., 2018) has also been related to predictive processing. Sinha and colleagues (Cannon et al., 2021;Sinha et al., 2014), for example, proposed that some typical phenotypes of autism, such as insistence on sameness, sensory hypersensitivities, difficulties interacting with dynamic objects, difficulties with theory of mind, and islands of proficiency (i.e., preserved or enhanced abilities in certain domains, e.g., mathematics, musical performance), can be explained by decreased tendencies to predict. The basic idea is that prediction is a crucial characteristic of human-environment interactions in many seemingly different domains and that reduced predictive processing makes an orderly environment become an overwhelming one. If these ideas are (at least partially) on the right track, and given that psycholinguistic research has established an important role for prediction in language in adults and children, 2 one would expect to observe some evidence for differences in prediction in spoken language processing in autistic children.
Certain features of the visual world paradigm method (i.e., rapid integration of linguistic and visual processing, semiautomatic languagemediated looking behavior, providing unequivocal measures of anticipation, see Huettig et al., 2011 for extensive discussion) make it almost ideally suited to study prediction in language in children. Using this method, Mani and Huettig (2012) observed that even 2-year-olds, just like adults, predict upcoming linguistic input that is a thematic fit to familiar verbs. Upon hearing a familiar verb, for example, 'eat' in sentences such as "the boy will eat the …", typically-developing toddlers showed anticipatory eye movements to (semantically appropriate) edible objects, and looked more towards these objects than towards unrelated distractor objects in a visual display. Given these advantages, several recent studies investigated language-mediated anticipatory eye movements in autistic children. Contrary to the prediction deficit account, Brock et al. (2008) however found no reduced effect of sentence context in adolescents with a diagnosis of autism. Similarly, Bavin et al. (2016) and Zhou et al. (2019) observed largely normal languagemediated anticipatory eye movements in autistic children with low levels of autistic traits.
It is noteworthy that these studies included adolescents or children with low levels of autistic traits only. The possibility remains that young children with higher levels of autistic traits show evidence for the (theoretically) predicted differences in language-mediated anticipatory eye movements (cf. Cannon et al., 2021;Sinha et al., 2014). Two recent studies point in this direction. Venker et al. (2019) found evidence for language-mediated anticipatory processing in autistic children, but noted that this effect was weaker in children with lower compared to larger receptive language skills. Similarly, Prescott et al. (2022) observed evidence that young autistic children can engage in predictive processing, but found that the effect was larger in a neurotypical control group and modulated by receptive language skills.
In the present study, we had a fresh look at prediction abilities in autistic children. Specifically, we investigated language-mediated anticipatory eye movements in children with high levels of autistic traits, who were in professional institutional care in Hangzhou, China. Our main question was whether we can detect any difference in language-mediated prediction between autistic children and typicallydeveloping children, and whether the levels of autistic traits modulate any difference in language-mediated eye movements.

Participants
Thirty-five children previously diagnosed with autism, aged between 3 and 9 years (28 boys, 7 girls), took part in the experiment. All children were enrolled in a professional day care center, specializing in treating children with autism spectrum disorder in Hangzhou, China. Data from seven children could not be analyzed due to extensive track loss (see the Data analysis section). The remaining sample comprised 28 children previously diagnosed with autism (mean age = 5.3, SD = 1.4, range = 3-8; 22 boys, 6 girls). All children underwent an assessment of their developmental traits and behavior by completing the Chinese version of the Psychoeducational Profile (PEP-3; Lam & Rao, 1993). The PEP-3 provides composite scores for children's communication and motor abilities, and (adaptive) behavior (see Table 1, higher PEP-3 scores reflect less severely affected individuals).
In addition, 34 typically-developing children, aged between 3 and 9 years (23 boys, 11 girls), were tested. All these children were recruited from pre-school centers and primary schools in Hangzhou, China. One child had to be excluded due to track loss (applying the same criterion as for the autistic children). The final sample comprised 33 children (mean age = 6.3, SD = 1.6, range = 3-9; 22 boys, 11 girls). None of the typically-developing children had a history of developmental or neurological disorders (these children did not complete the PEP-3).
The dominant language of all children was Mandarin Chinese. All children were tested at the Hangzhou Autism Center in Hangzhou, China. Ethical approval for the study was granted by the Institutional Ethical Review Board of the Hangzhou Autism Center. Written consent for taking part in the study was provided by the children's caretakers.

Materials
We chose the same spoken sentences and visual stimuli as Mani and Huettig (2012) and included a control group of typically-developing children. Replicating Mani and Huettig (2012) with young Chinese children allowed us to establish the suitability of the materials for Chinese children. A native speaker of Mandarin Chinese translated the German stimulus sentences used by Mani and Huettig (2012) into Mandarin. The materials comprised twelve target nouns embedded in predictive and non-predictive sentence contexts (e.g., the Chinese translation equivalent of 'The boy eats/loves the big cake'; see Appendix). Speech stimuli were produced by a female native speaker of Mandarin. The mean duration of the sentences was 3606 ms (SD = 325 ms). Onsets and offsets of verbs and target nouns in the sentences were marked using Praat (Boersma & Weenink, 2002). On average, verbs and target nouns started at 1057 ms (SD = 182 ms) and 2722 ms (SD = 413 ms), respectively, into the spoken sentences. The time between onset of the verb and onset of the target was on average 1665 ms (SD = 381 ms).
The same pictures as in Mani and Huettig (2012) were used. These were photographs of objects commonly known by children aged between 3 and 9. Each item was associated with a picture of the target noun and a distractor picture. Labels for the target and distractor images were semantically and associatively unrelated.

Procedure
The children were tested individually using an Eyelink 1000 eyetracker sampling at 1000 Hz. Viewing distance was held constant between 55 and 60 cm. The eye-tracker was calibrated, and the children were instructed to listen to the sentences carefully and asked to not move their eyes off the screen. Such look-and-listen tasks have been successfully used with adults, young children, and clinical populations (Huettig et al., 2011). The spoken sentences were presented through loudspeakers. The 24 sentences were distributed across two experimental lists such that one target only occurred once on each list. The children were randomly assigned to one list and were presented with all 12 trials on that list. Each test trial began with a red dot moving around the screen in a circle at 500 ms intervals to capture the children's attention. The dot landed in the center of the screen and, after 1500 ms, was replaced with the two objects (each 250 × 250 pixels), one presented on each side of the screen, 512 pixels apart. Participants had 3 s to inspect the displays prior to the acoustic onset of the verb (cf. Mani & Plunkett, 2010). Areas of interest (300 × 300 pixels) were defined around target and distractor pictures. Eye movements were coded as fixations, saccades and blinks by the algorithm provided in the eyetracker software.

Data pre-processing
Data from seven autistic children and one control child were excluded from the analysis due to extensive track loss. These children had >50% of trials with no fixation to one of the objects during the critical analysis period (verb onset to target word onset; cf. Venker et al., 2013). Thus, data from 61 children (28 autistic children, 33 control) were available for analysis.  Fig. 1 suggests a bias in looks to the target object in the predictable condition emerging shortly after verb onset. In contrast, in the non-predictable condition, more looks to the target than to the distractor were made only after target word onset. Autistic children looked more at the target object than at the distractor after target word onset in both predictable and non-predictable conditions but not during the onset-verb-onsettarget window (neither in the predictable nor in the non-predictable condition).

Results
The raw data and analysis code can be found on OSF: https://osf. io/qzcd8/. For the analysis, we considered looks to targets and distractors (and track loss as missing data). We divided each trial into three windows of the same size. The 'critical window' started at verb onset plus 233 ms (the approximate time it takes to program and launch a saccadic eye movement in young children, Mani & Plunkett, 2010) and ended at target onset (+ 233 ms). For each trial, we extracted a 'baseline window' of the same size as the critical window, which ended at verb onset (+233 ms), and a 'label window', which started at target onset (+ 233 ms). The goal of this approach was to compare fixation behavior during the baseline window, where no linguistic information about the upcoming target had been provided yet, to fixation behavior during subsequent time windows (the three analysis windows are highlighted through shading in the fixation plots). In comparison to the baseline window, we predicted an increased likelihood of looks to the target compared to the distractor during the label window, as the target object is referred to in the speech. We also predicted more looks to the target than to the distractor during the critical window (compared to the baseline) in the predictable condition, as the information becoming available on hearing the verb could be used to predict the upcoming target.
For each of the three windows, we summed the duration of looks to the target (i.e., target gaze duration). We used beta-regression rather than logistic regression, because multiple target looks within a single trial are not independent of one another, and because this allowed the use of proportions directly without weighting them by window length (i. e., our approach automatically normalized for durational differences among auditory stimuli). Gaze durations were analyzed separately for critical and label windows, using beta regression as implemented in R package glmmTMB (Brooks et al., 2017). As fixed-effect predictors, we included an intercept, 'Group' (treatment-coded such that autistic children = 1 and control = 0), 'Condition' (treatment-coded such that predictable = 1 and unpredictable = 0) and their interaction. In addition, an offset term was included, consisting of the trial-specific (logittransformed 3 ) gaze proportion in the baseline window. We included a maximal random-effects structure (Barr et al., 2013) as long as the model could still converge, assuming a diagonal random-effects covariance structure. Function buildglmmTMB from R package buildmer (Voeten, 2023) was used to find the maximal feasible randomeffects structure. In both the critical window and labeling window, the maximal random-effects structure turned out to be feasible and was hence used; this means a by-participants random intercept and random slope for Condition, and a by-items random intercept and random slopes for Group, Condition, and their interaction ( Table 2). The analyses statistically confirmed that the autistic children gazed less at the target object than at the distractor in the critical window in the predictable condition compared to the control children (PR = 0.27, t = − 2.25). Moreover, we observed robust evidence that both autistic and control children fixated the target more than the distractor after target word onset (PR = 2.87, t = 3.51). There was also a difference between the groups on predictable trials in the labeling window (most likely a continuation of the same effect observed in the critical window, PR = 0.25, t = − 2.17; cp. The same effect in the critical window where PR = 0.27. 4 In post-hoc follow-up analyses, we assessed the extent to which predictive looks by autistic participants correlated with their PEP-3 composite scores. Note that autistic children also looked less (Fig. 1) at the target object than control children in the neutral condition (i.e., the 'label effect', looking at an object once it is mentioned in unfolding speech, which is one of the most robust effects in the psycholinguistic literature). This observation suggests that language-mediated prediction in autistic children may be a general deficit in visual attention. We thus investigated the relationship between looks of autistic children to the target object in the non-predictable condition (labeling window) with the amount of anticipatory looks to the target object during the predictive window (reflecting their prediction skills) and examined whether this relationship was moderated by the children's (1) Communication skills, (2) Motor skills, and (3) (Adaptive) behavior scores on the PEP-3. Model predictions for all autistic children with PEP-3 scores and all items for the non-predictable labelling window and the critical 3 If a trial had a baseline gaze-duration proportion of exactly zero or one, this was clamped to 0.001 or 0.999, respectively. 4 At the request of an anonymous reviewer, we included both participants' age, andnested within an interaction with Groupthe autistic children's PEP-3 scores to the model. Autistic children's age and their PEP-3 scores were significantly negatively correlated (Communication: r = − 0.54, p = .002; Motor Skills: r = − 0.54, p < .001; Adaptive Behavior: r = − 0.64, p < .001). We assessed whether these correlations led to collinearity in the model by means of generalized variance-inflation factors (gVIFs, computed using function check_ collinearity from R package performance). Whereas the gVIFs for the original models were all well below 2 (indicating weak collinearity), those of the model with both age and PEP-3 scores exceeded 20, suggesting a very high degree of collinearity. On the basis of these results and the fact that PEP-3 scores were only available for a subset of 20 autistic children, PEP-3 scores could not be included in the model. The 'critical window' and 'label window' models that contained (mean-centered) age and its interactions with all other terms in the model revealed qualitatively the same results as the main analysis model. That is, in the predictive window, we continued to observe reduced fixations by the autistic children relative to the controls (PR = 0.24, t = − 2.43), and in the labeling window we continued to observe significant target preferences overall (PR = 2.50, t = 3.38) and the same interaction between Group and Condition (PR = 0.14, t = − 3.37). Age did show an effect in a significant three-way interaction with Group and Condition (PR = 0.40, t = − 2.49) in the labelling window, such that older autistic children showed a smaller target bias than younger autistic children, which is consistent with the notion that higher levels of autistic traits were more strongly expressed in the older than in the younger children (viz. the negative correlation between age and PEP-3 scores).
window in the predictable condition were obtained, resulting in 240 data points (20 participants, 12 items) that fully represent the model fit for those children. Then, for each of the three PEP-3 composite scores of Communication, Motor Skills, and (Adaptive) Behavior, we performed a median split, resulting in six subsets stratified by PEP-3 score type and high/low performance on that PEP score. For each of these, we computed the correlation between the predictions from the critical window and those in the labelling window, resulting in six correlation coefficients. The correlations were computed using repeated-measures correlations, 5 based on R package rmcorr (Bakdash & Marusich, 2022). Items 6 were included in the correlations as random factor. Bonferroni corrections were applied to correct for the increased family-wise error rate in the p-values from these repeated tests (Table 3). Fig. 2 provides a visualization of the correlations. In this figure, the random effect for Item has been partialed out. The overall correlation between the fixation behavior in both windows was r = 0.30, F(1,227) = 22.62, p < .001, suggesting that in the autistic children, prediction ability (critical window) correlated positively and moderately with visual attention to the labeled target object (labeling window) in the non-predictable condition. Importantly, this relationship was moderated by the children's performance on the PEP-3: The correlation between prediction and visual attention to the (non-    5 We thank an anonymous reviewer for suggesting to adopt this mixed-effects approach in our correlation tests. 6 It was not necessary, and in fact not possible, to also include participants as a random factor in these correlations, as the PEP-3 scores we tested are all between-subjects, and hence would be partialed out if in addition a byparticipants random effect had been included. predictable) target was statistically significant and relatively strong in the children with higher Communication, Motor Skills, and (Adaptive) Behavior scores, and not statistically significant and relatively weak in children with lower scores.

Discussion
In the present study, typically-developing young children showed a large prediction effect, replicating Mani and Huettig (2012) with young Chinese participants and Mandarin materials. In contrast, autistic children (as a group) did not anticipate the target (but directed preferential looks at it once it was mentioned). This finding does not necessarily contradict previous studies that found evidence for anticipatory eye gaze in autistic children, because these studies were carried out with adolescents or autistic children with low levels of autistic traits. Our autistic participants were young children who differed in the severity of autistic traits and were in professional institutional care. Autistic children in our study also looked less at the target (e.g., the cake) than control children when it was mentioned in the non-predictable condition ("The boy loves the big cake"). Most strikingly, autistic children with higher scores for Communication, Motor Skills, and (Adaptive) Behavior were the ones with a higher tendency of both predictive eye gaze and overt visual attention to non-predictable targets.
Do some autistic children show reduced language-mediated anticipatory eye-movements? Superficially the answer is yes. As a group, autistic children in the present study showed strikingly less anticipatory eye gaze in a task where typically-developing young children strongly anticipated. Looking more closely at our results, however, it becomes clear that the underlying relationship between autism and (languagemediated) prediction is a very complex one. First, and perhaps unsurprisingly, it makes little sense to lump all autistic children together. When it comes to language-mediated anticipatory eye movements not all autistic children are alike. The tendency to engage in anticipatory eye movements appears to be related to the severity of autistic traits in communication, motor skills, and (adaptive) behavior. Second, and even more informative, is the finding of a possible link between generalmediated attention (looking at the cake when the word 'cake' acoustically unfolds) and language-mediated prediction (looking at the cake Fig. 2. Repeated-measures correlations between looks in the predictable condition of the critical window and looks in the labeling window in the non-predictable condition of the labeling window, for the autistic children for whom PEP-3 scores were available. The first panel visualizes the 'raw' correlation (r = 0.3, p < .001). The other panels show low and high median splits of each of the three PEP scores, as indicated by the two labels above each panel. The shaded area around the trend line indicates the 95% CI.
when hearing the word 'eat'). The higher the autistic traits of the autistic children according to their communication, motor, and (adaptive) behavior scores, the less they tended to engage in typical predictive as well as non-predictive visual attention behavior. The finding that not only communication skills but also motor and (adaptive) behavior scores showed such a correlation raises the possibility that it reflects more than a receptive language difference (cf. Prescott et al., 2022;Venker et al., 2019): apparent differences 7 in language-mediated anticipatory eye movements in autistic children with higher levels of autistic traits may at least partly be differences in visual attention in disguise. 8 Another useful direction for further research would be to explore to what extent the anticipation differences in autistic children are a secondary consequence of differences in the efficiency of processing of the speech signal (Fernald et al., 2006(Fernald et al., , 1998. Given the challenges inherent in research with autistic children, we suggest that further investigation with a longitudinal design (cf. Goswami, 2015;Huettig et al., 2018) and collaborative efforts involving many labs (cf. Frank et al., 2017) would prove most fruitful.

Declaration of Competing Interest
None.

Data availability
Our data are freely available at the OSF long provided on the title page.