The Effect of Volatility in Linguistic Input on Prediction Behavior in Autistic Toddlers

Abstract

Domain-general prediction differences have been posited as underlying many aspects of the cognitive-behavioral profile in autism. An interesting potential implication of such differences is hyperplasticity of learning – the idea that autistic individuals may privilege more recent input over the accumulation of prior learning. Because real world language input is highly variable, hyperplasticity could have serious ramifications for language learning. To investigate potential hyperplasticity during a language processing task, we administered an experimental anticipatory eye movement (AEM) task to 2- to 3-year-old autistic children and neurotypical (NT) peers. Autistic children’s change in anticipation from before to after a switch in contingencies did not significantly differ from NT counterparts, failing to support claims of hyperplasticity in the linguistic domain. Analysis of individual differences among autistic children revealed that cognitive ability was associated with prediction of the initial, stable contingencies, but neither age nor receptive language related to task performance. Results are discussed in terms of clinical implications and the broader context of research investigating prediction differences in autism.

Lay Summary

Autistic individuals may differ from neurotypical peers when making predictions, especially when the environment is unstable. In this study, we found that 2- to 3-year-old autistic children’s linguistic predictions did not differ from their neurotypical peers, though autistic children with stronger cognitive ability were better at prediction in stable environments. More research will be needed to investigate the relationship of prediction and language skills in autistic individuals.

Autism is a highly heterogeneous condition characterizing 1 in 36 children by age 8 (Maenner et al., 2023). Despite its prevalence, the cognitive and linguistic mechanisms underlying autism remain unclear. In the last decade, theories emerging from the hierarchical predictive coding framework (Clark, 2013; Friston & Kiebel, 2009) have gained considerable attention as a unifying account of the diverse phenotypes across the autism spectrum. Predictive coding characterizes cognitive and perceptual processing as hierarchical, bidirectional interactions between top-down predictions and bottom-up sensory input. Mismatches between top-down predictions and incoming sensory input result in prediction error. The internal predictive model is updated with new learning according to the reliability or “precision” attributed to that error signal. Neurotypical (NT) brains are thought to flexibly adjust the precision of prediction errors to determine whether the input is irrelevant noise or if new learning is warranted.

Grounded in this framework, several theorists have posited that core and associated characteristics of autism result from domain-general differences in predictive processing, though the precise mechanism varies across accounts (Cannon et al., 2021). For example, Pellicano and Burr (2012) suggested that autistic individuals have reduced top-down predictions and thus perceive bottom-up sensory input more accurately compared to NT individuals. Other theories propose that autistic individuals are not as flexible in weighting the precision of prediction errors, resulting in indiscriminate learning even in volatile, noisy environments (Lawson et al., 2014; van de Cruys et al., 2014).

Of particular relevance to the current study, Sinha and colleagues (2014) describe predictive differences in autism in terms of estimation of inter-state conditional probabilities, called the predictive impairment in autism (PIA) hypothesis by these researchers. Similar to other accounts (Pellicano & Burr, 2012; van de Cruys et al., 2014), a potential implication of the PIA is that incoming sensory input will be perceived as more novel, leading to hyperarousal of brain structures thought to be involved in modulation of learning (Sinha et al., 2014). The result is “hyperplasticity” of learning, where current input is overweighted at the expense of aggregated prior experience. This form of hyperplasticity would not be evident when the structure of the input is static, because the current input would be consistent with aggregated prior experience. However, in situations where the input fluctuates – as is characteristic of much of our experience – hyperplasticity might render altered aggregation of learning experiences.

In response to the proliferation of theories proposing prediction differences in autism, researchers have conducted studies across a wide variety of modalities and methodologies to investigate these claims (for a comprehensive review, see Cannon et al., 2021). If hyperplasticity does indeed characterize autistic processing, we would expect the literature to reflect domain-general similar or enhanced predictive sequence learning in autistic individuals when contingencies are stable, but disrupted learning when the probabilistic structure of the input becomes more volatile. Narrowing our search to behavioral studies of predictive sequence learning, previous findings offer partial support for this pattern. Cannon et al. (2021) summarized the literature as follows: studies reporting similarities between autistic and NT groups were those in which antecedents had consistent predictive value relevant to the task and were temporally adjacent or overlapping with the consequence. For example, a common experimental paradigm indexing implicit sequence learning, the serial response time (SRT) task, involves repeated exposure to a sequence in which the relationship between antecedent and consequence is deterministic. Studies utilizing SRT tasks have consistently demonstrated similar performance between autistic individuals and NT peers matched on age or nonverbal cognition (e.g., Brown et al., 2010; Nemeth et al., 2010; Travers et al., 2010; Zwart et al., 2017; for a meta-analysis see Foti et al., 2015).

When the learning environment is more volatile, evidence for diagnostic group differences is more mixed. For example, Manning et al. (2017) manipulated the probability of a visual reward location such that the learning environment was either stable or volatile, and found that school-age autistic children did not differ in their learning rate compared to age- and cognition-matched NT peers in either condition. Likewise, adding sensory noise to an implicit learning paradigm did not lead to group differences in prediction in a study of 3-year-olds with high or low likelihood of autism diagnosis (Ward et al., 2022). Another study manipulating input probability found that autistic adolescents’ anticipatory looking during action sequences was modulated by the predictability of the sequence steps, differing from NT comparisons only in overall looking time to action goals (Ward et al., 2021). By contrast, other studies did report diagnostic group differences in prediction of less stable contingencies. For example, Amoruso et al. (2019) found that autistic and NT children were able to make goal-based predictions when provided with kinematic cues, but when kinematics were ambiguous, autistic children’s predictions relied less on co-occurrence probabilities than NT children. Another study suggesting differences in predictive association learning employed a violation of expectation paradigm (Greene et al., 2019). Eighty percent of the trials contained accurate cue-target sequences, while 20% of the trials violated the expected cue-target sequence. On those violation trials, autistic adolescents looked less to the cue-predicted target compared to the NT group. Taken together, behavioral evidence in nonverbal sequence-learning tasks partially fits the pattern predicted by hyperplasticity as described by Sinha et al. (2014). That is, autistic individuals do demonstrate similar predictive learning when input is stable, but whether prediction differences are present in volatile learning environments is less clear.

While the research on prediction in autism has largely focused on nonverbal abilities, prediction is also a key component of linguistic processing (Zarcone et al., 2016). Given the domain-general claims of the PIA, hyperplasticity should also be evident in linguistic contexts. However, predictive language processing in autism has to date only been investigated with paradigms involving familiar, meaningful, and stable predictive contingencies. For example, autistic adolescents (Brock et al., 2008), school-age children (Bavin et al., 2016; Hahn et al., 2015), and preschool children (Prescott et al., 2022; Venker et al., 2019; Zhou et al., 2019) looked more quickly to target nouns when they are preceded by a semantically-informative verb (e.g., “eat the cake”) than a neutral verb (e.g., “find the cake”). Most studies reported largely similar looking behavior between autistic and NT groups matched on various characteristics (Brock et al., 2008; Bavin et al., 2016; Hahn et al., 2015). Two studies of preschool-age children found a smaller effect of semantic information overall in the autistic group when compared to same-age children (Zhou et al., 2019) and younger, receptive language-matched children (Prescott et al., 2022), though both demonstrated similar processing efficiency between autistic and NT groups. Verbal statistical learning tasks can also offer insight into deterministic contingency learning in the linguistic domain. Similar performance between autistic children and NT peers of either similar age and nonverbal cognition or language ability has been demonstrated across a variety of tasks, such as word segmentation (Haebig et al., 2017; Mayo & Eigsti, 2012), artificial grammar learning (Brown et al., 2010), and cross-situational word learning (Hartley et al., 2020; Venker, 2019). In sum, these findings indicate that autistic children can learn stable linguistic contingencies in experimental settings. However, the hyperplasticity claim of the PIA cannot be fully addressed without investigation of predictive sequence learning when input is inconsistent, which to date has not yet been examined in the linguistic domain.

The need for further investigation of autistic children’s predictive behavior in volatile linguistic learning environments is further highlighted by a recent study by Hu and colleagues (2023) which found a disassociation between linguistic and non-linguistic statistical learning in autistic children. Despite comparable performance on non-linguistic statistical learning tasks, the autistic children demonstrated weaker learning compared to the NT group across two linguistic statistical learning tasks as measured by composite scores based on both online learning and learning outcomes (Hu et al., 2023). In one of the two linguistic tasks (an auditory syllable triplet-learning task), autistic children did not differ from NT peers in online learning (as measured by mean change in reaction time) but did demonstrate weaker offline retrieval of learning (as measured by mean accuracy), possibly due to the inclusion of foil stimuli. These study findings raise the possibility that linguistic prediction in unstable learning environments may not follow the same pattern as previous studies of non-linguistic prediction in autistic children. While hyperplastic learning could potentially help explain some areas of strength for autistic individuals (e.g., lower susceptibility to false memories; Griego et al., 2019), a domain-general processing style characterized by hyperplasticity could manifest in deleterious effects on language. Acquiring a language involves not only detecting simple associations among units in the input, but also aggregating learning across instances over time, abstracting across the variability of those instances, and generalizing that learning to new situations (Sandhofer & Schonberg, 2020). If input is always perceived as novel, the shared features of that observation with prior learning may be overlooked and thus affect abstraction and generalization processes. As such, hyperplasticity may also help explain results of previous studies reporting autistic differences in tasks such as categorical induction and shape bias (Kelley et al., 2006; Naigles et al., 2013; Potrzeba et al., 2015; Tek & Naigles, 2017).

Thus, to provide a stronger test of the PIA and claims regarding hyperplasticity, the current study examined autistic children’s prediction behavior in response to variable linguistic input. Two- to three-year-old autistic children and younger, cognitive ability-matched NT children participated in an eye-gaze task eliciting anticipatory eye movements (AEMs) as an index of linguistic prediction. On each trial, children would see a pair of objects in greyscale and hear one of two adjective-noun pairs (happy kitty, silly birdie). When the noun was spoken, the corresponding object on the screen would light up and move, providing a visual reward for correct anticipatory eye movements. The adjective-noun contingencies were stable for first eight trials, and then switched for two trials before returning to the original pairings for the next eight trials, followed by another switch. The dependent variable was children’s logit probability of looking (as measured by eye gaze) to the target (noun referent) during an anticipation window after the onset of the predictive adjective and before the onset of the target noun. The design of this task allowed for comparison of predictions (adjective noun) based on cumulative probability versus recency. If autistic children’s predictions are more influenced by recent input than cumulative experience (suggesting hyperplasticity), then we would expect group differences in AEMs when the input is more volatile (i.e., after the switch in predictive contingencies). Children also completed standardized autism diagnostic, cognitive ability, and language assessments to explore the relationship of individual child characteristics to prediction within the autistic group. Given that prediction differences are posited to be domain-general, we expected both cognitive and language ability to relate to autistic children’s performance on the linguistic prediction task. We also expected autistic children’s tasks performance to improve with age. We asked the following research questions:

1. Do 2- to 3-year-old autistic children demonstrate prediction differences indicative of hyperplasticity as compared to NT peers? In particular, given the structure of this task, do they show lingering effects of a brief switch in contingency relative to NT peers?

2. What are the effects of individual differences in age, receptive language, and cognitive ability on autistic children’s prediction in this linguistic task?

Method

Participants

Participants included in this study were 42 autistic children (9 female; M=30.98 months, SD=3.20) and 47 NT children (14 female; M=21.09 months, SD=3.37) though subsample sizes varied across analyses (see Group Comparison). Children were recruited as part of a larger longitudinal study with a first visit when autistic children were 2- to 3-years-old and a second visit one year later (see Ellis Weismer & Saffran (2022) for conceptual framework; Prescott et al. (2022) for prior experimental work following similar protocol with adaptations in the current study for remote assessment). The data reported in this study were collected during the first visit. The research protocol was approved by the university Institutional Review Board, and written informed consent was provided by a legal guardian for all participants. Participants were recruited from the local community and surrounding region. Participant demographics and characteristics can be found in Table 1. All children were from monolingual English-speaking households and had no known uncorrected hearing or vision impairments. Children were assigned to the autistic group if they received DSM-5 autism diagnoses (see assessment measures below) from an experienced clinical psychologist on our research team, working in collaboration with a certified speech-language pathologist with autism experience. Autistic children with and without concomitant diagnoses of cognitive and/or language delay were included in the autistic group. Exclusionary criteria for the autistic group included known metabolic disorders, cerebral palsy, congenital rubella syndrome, fetal alcohol syndrome, neurofibromatosis, hypoxic encephalopathy associated with prematurity, progressive neurological disorders, and seizure disorders. Children included in the NT group demonstrated typical development as measured by standardized assessment and parent report, including overall language and cognitive abilities (no features of autism, developmental delays, motor deficits, or seizures).

Table 1.

Participant characteristics and demographic information for children with data included in the first switch (trials 7, 8, 11, 12) and the second switch (trials 17, 18, 21, 22).

First Switch (Trials 7. 8, 11, 12)
	Autistic (n=25)	NT (n=32)	Group Comparison
	Mean (SD) Median range	Mean (SD) Median range
Participant Characteristics
Age (Months)	31.32 (2.94) 32 24–35	21.53 (3.73) 21 16–35	W=765, p<.001
Auditory Comprehension
Raw Score	18.42 (6.75) 18 8–41	29.47 (7.18) 27 19–52	W=72.5, p<.001
Standard Score	59.88 (15.04) 55.5 50–117	110.59 (15.28) 110 81–150	W=23.5, p<.001
Expressive Communication
Raw Score	21.96 (6.01) 21 12–38	31.47 (12.78) 27.5 23–96	W=103.5, p<.001
Standard Score	71.75 (14.82) 69 50–111	110.47 (14.00) 105 91–142	W=35, p<.001
Cognitive Ability
Raw Score	33.00 (7.53) 31 20–49	36.03 (4.90) 36 29–51	W=234, p=.008
Standard Score	87.24 (13.42) 84 64–116	108.50 (8.88) 108 94–131	t(39.64)=−6.84, p<.001
Adaptive Behavior	69.56 (9.90) 67 49–97	97.84 (9.48) 97.5 83–130	W=16, p<.001
Level of Autism Traits
Total Score	38.22 (4.22) 39 28.5–45	—	—
Demographic Information
Maternal Education (Years)	14.04 (2.41) 13 11–19	17.34 (1.54) 18 14–20	W=120, p<.001
Race	0 American Indian/Alaska Native 0 Asian 0 Native Hawaiian/Pacific Islander 1 Black or African American 5 More than one race 19 White	0 American Indian/Alaska Native 0 Asian 0 Native Hawaiian/Pacific Islander 0 Black or African American 1 More than one race 31 White
Ethnicity	2 Hispanic or Latino 23 Not Hispanic or Latino	0 Hispanic or Latino 32 Not Hispanic or Latino

Second Switch (Trials 17, 18, 21, 22)
	Autistic (n=15)	NT (n=27)	Group Comparison
	Mean (SD) Median Range	Mean (SD) Median Range
Participant Characteristics
Age (Months)	29.80 (3.32) 30 24–35	20.85 (3.58) 20 16–35	W=387, p<.001
Auditory Comprehension
Raw Score	17.20 (7.71) 18 8–41	27.19 (6.49) 27 19–44	W=32, p<.001
Standard Score	59.87 (16.62) 54 50–117	104.30 (16.59) 107 81–150	W=21.5, p<.001
Expressive Communication
Raw Score	21.87 (6.95) 20 13–38	30.74 (13.79) 26 23–96	W=67, p<.001
Standard Score	73.07 (16.40) 71 50–111	104.89 (11.19) 102 88–130	W=25, p<.001
Cognitive Ability
Raw Score	29.93 (6.87) 31 20–49	34.74 (4.89) 33 29–51	W=97, p=.005
Standard Score	82.67 (11.32) 83 64–113	105.33 (5.85) 106 94–120	t(18.24)=−7.23, p<.001
Adaptive Behavior	66.13 (9.88) 67 49–82	96.89 (9.17) 94 87–130	W=0, p<.001
Level of Autism Traits
Total Score	40.27 (4.28) 41 30–45	—	—
Demographic Information
Maternal Education (Years)	13.40 (2.23) 12 12–18	16.96 (1.76) 16 12–20	W=53.5, p<.001
Race	0 American Indian/Alaska Native 0 Asian 0 Native Hawaiian/Pacific Islander 0 Black or African American 3 More than one race 12 White	0 American Indian/Alaska Native 0 Asian 0 Native Hawaiian/Pacific Islander 0 Black or African American 1 More than one race 26 White
Ethnicity	2 Hispanic or Latino 13 Not Hispanic or Latino	1 Hispanic or Latino 26 Not Hispanic or Latino

Note. Descriptive statistics for the autistic and NT children included in analyses of the first switch and the second switch. Receptive language was measured by PLS-5 Auditory Comprehension subtest scores. Expressive language was measured by PLS-5 Expressive Communication subtest scores. Cognitive ability was measured by DAYC-2 cognitive domain scores. Adaptive behavior was measured by Vineland-3 Adaptive Behavior Composite standard scores. Level of autism traits was measured by CARS-2 total raw score (where 15–29.5 indicates minimal-no autism traits, 30–36.5 indicates mild-moderate autism traits, and 37 and higher indicates high autism traits; Schopler et al., 2010). p values were obtained by two-sample, two-tailed Welch’s t-tests unless score distributions for either group failed the Shapiro–Wilk normality test (Shapiro & Wilk, 1965). If distributions were non-normal, a non-parametric Wilcoxon rank sum test with continuity correction was used (Wilcoxon, 1945). PLS-5 data were missing for one autistic participant.

Standardized Measures

All standardized measures were selected for their validity and reliability in remote administration due to the COVID-19 pandemic. Autism diagnoses were based on psychologist administration of the Toddler Autism Diagnostic Interview – Revised (ADI-R; Kim & Lord, 2012), Brief Observation of Symptoms of Autism (BOSA; Dow et al., 2022), and Childhood Autism Rating Scale – Second Edition (CARS-2; Schopler et al., 2010) and best estimate clinical diagnosis. All NT children received passing scores (‘low risk’ scores of 0–2) on the Modified Checklist for Autism in Toddlers - Revised with Follow-Up ( M-CHAT-R/F; Robins et al., 2014) to rule out symptoms of autism. Children in both groups received the cognitive domain of the Developmental Assessment of Young Children – 2^nd Edition (DAYC-2; Voress & Maddox, 2013) from which raw scores were used as a measure of cognitive ability for group matching and to explore individual differences. The Vineland Adaptive Behavior Scales, Third Edition (Vineland-3; Sparrow et al., 2016) Comprehensive Parent/Caregiver Form was used to provide information about adaptive functioning. Caregivers also completed a background form pertaining to developmental history and demographic information. The clinician-administered Preschool Language Scales – 5^th Edition (PLS-5; Zimmerman et al., 2011) was used to evaluate receptive and expressive language.

Eye-Gaze Task

Procedure

The experiment was conducted in-person in a research lab. Children sat on caregivers’ laps, approximately 60 cm from the screen in a sound-attenuated booth. Caregivers wore opaque glasses to avoid biasing their child’s looking behavior and were instructed not to talk to their child or otherwise direct their attention. Children were asked to watch and listen to the “movie.” Visual stimuli were presented on the 55-inch screen and auditory stimuli were presented at 65 dB from a speaker under the screen. Below the screen, a Canon Vixia HFG10 video camera with a 30 Hz frame rate recorded children’s eye movements. The experimenter observed the video recording in real-time from a Mac computer outside the booth. The experiment was programmed in E-Prime 3.0 (Psychology Software Tools, Pittsburgh, PA) on a PC computer and lasted approximately 2 minutes.

Stimuli

Auditory stimuli were recorded by a female, native English speaker using child-directed intonation and consisted of two adjectives (“happy,” “silly”) and two nouns (“kitty,” “birdie”) presented in pairs. Each token was recorded separately such that co-articulation could not be used to facilitate prediction of upcoming nouns. Visual stimuli included side-by-side cartoon images of a bird and a cat which appeared within white boxes on a grey background (Figure 1). At the beginning of each trial, the images were static and appeared in greyscale. After 1000 ms of silence with the images in greyscale visible on screen, children heard one of the two predictive adjectives (“happy” or “silly”). At 2500 ms, children heard the consequent noun (“kitty” or “birdie”) as the corresponding image simultaneously appeared in full color and moved. The “kitty” appeared in orange, while the “birdie” appeared in blue. The movement of the target image was a slight side-to-side rocking over an approximately 20-degree range that occurred about three times within 900 ms. The target image appeared in color and moved at 2500 ms regardless of the location of children’s gaze at the time. Trials ended after 3400 ms followed by a 1000 ms inter-trial pause with a blank screen (Figure 2). For the first 8 trials, “happy” was consistently paired with “kitty” and “silly” was consistently paired with “birdie”. Trial order and target image side were pseudorandomized, though all participants received the same trial order due to an error in the pseudorandomization of the second version of the task. In trials 9 and 10, the contingencies were switched, such that “happy” predicted “birdie” and “silly” predicted “kitty”. In trials 11–18, the pairings reverted to the original contingencies. Trials 19 and 20 constituted a second switch, matching the contingencies presented in trials 9 and 10. Trials 21 and 22 reverted to the original pairings (Figure 3).

Figure 1.

Example of study trial.

Figure 2.

Trial timeline.

Figure 3.

Task Structure. Contingency switch trials are in bold.

Data Analysis

Details of data processing and data cleaning procedures can be found in Supplementary Materials.

Analytic Approach

To determine whether autistic children updated their predictions after the contingency switch based on recent input (suggestive of hyperplasticity) or did not (suggestive of use of cumulative probability), we focused our analysis on the two trials immediately preceding the contingency switch (trials 7 and 8 in the first switch and trials 17 and 18 in the second switch) and the two trials immediately following the contingency switch (trials 11 and 12 in the first switch and trials 21 and 22 in the second switch). We intentionally did not analyze the switch trials themselves (trials 9 and 10 in the first switch and trials 19 and 20 in the second switch) because children had not yet seen the contingency switch in anticipatory window of the first switch trial (i.e., trial 9 in the first switch, trial 19 in the second switch) but had seen the contingency switch once by the anticipatory window of the second switch trial. Due to this unavoidable dissimilarity, it was not possible to average these two trials, making them difficult to incorporate into the analyses. We specified an anticipatory window which began 300 ms after the onset of the adjective to accommodate the time required for children to launch a saccade (Alahyane et al., 2016) and ended at the onset of the noun, in total lasting from 1300–2500 ms from the trial onset (Figure 2). Each 33-ms frame was given a code of 1 if the child was looking to the target image (i.e., the image predicted by cumulative probability) and 0 if the child was looking to the non-target image. Gaze shifts and looks away from visual stimuli were also coded but removed prior to analysis as part of data cleaning (see Supplementary Materials).

Group Comparison

Our initial approach for establishing a basis for diagnostic group comparison was to match NT participants to autistic participants based on cognitive ability. We conducted a 1:1 bootstrap matching procedure based on DAYC-2 cognitive domain raw scores with a caliper of 2 points, resulting in groups of n=31. A detailed explanation of this procedure can be found in Pomper et al. (2019). However, we considered successful anticipation of the contingencies prior to the switch essential for comparison to post-switch trials for the first research question. The initial set of analyses using these matched groups revealed that the matched NT group did not learn the pre-switch contingencies in either the first or second switch (ps>.05). Full model results for these analyses can be found in Supplementary Materials (Table S1 and S2) but were not considered interpretable due to the lack of evidence for learning by the NT group during the pre-switch condition.

Therefore, in order to establish a common baseline from which to measure change from pre-switch to post-switch between groups for the first research question, we returned to the full dataset and created new groups of NT and autistic children. The new groups included only those children with an average proportion of looks to target image vs. non-target image during the analytical window greater than .50 (i.e., chance) in the pre-switch trials of either the first switch (trials 7 and 8), the second switch (trials 17 and 18) or both. In doing so, we were able to establish a common baseline from which to compare pre-switch to post-switch change between diagnostic groups. In total, data from 42 autistic children and 47 NT children were included in the present study. Of those, 25 autistic children and 32 NT children met criteria for inclusion in analyses for the first switch, while 15 autistic children and 27 NT children met criteria for inclusion in analyses for the second switch. It is worth noting that these groups of children partially (but not completely) overlapped; 19 children (7 autistic children, 12 NT children) had data included in analyses for both the first and second switch. For the second research question, we sought to understand child characteristics associated with the full range of performance among autistic children in this task. Therefore, the full sample of autistic children who passed exclusionary and data cleaning criteria (n=42) were included in analyses for the second research question.

Model Fitting

For the first research question, we specified two-level (time interval within subject) logistic regression models in which target look (1 = target image, 0 = non-target image) was modeled against time (units coded by second) at the trial level over the analytical window from 1300 ms to 2500 ms after trial onset. Time was centered at the midpoint of the analytical window (1900 ms). Logit-linear models were specified at level 1 such that the logit probability of target looks was a linear effect of time, trial type and the interaction of time and trial type (contrasts: 0=pre-switch, 1=post-switch). We included diagnostic group at level 2 as a binary subject-level variable. In separate analyses, we coded diagnostic group so simple effects within each group could be interpreted in reference to the intercept (contrasts: 0=NT, 1=autism for the NT model, 1=NT, 0=autism for the autistic model). The full by-subject random effects structure including subject-level random effects for the intercept and all level 1 predictors was specified. Because we had no specific hypotheses about change in looking behavior from early in the experiment to later in the experiment, separate models were fit for the first research question to analyze anticipatory looking behavior before and after the first switch (trials 7, 8, 11, 12) and the second switch (trials 17, 18, 21, 22) respectively.

For the second research question, we specified two-level logistic regression models (time within subject) which included data from all autistic children tested in the experimental task who passed exclusionary criteria. The level 1 model regressed the logit probability of looking to target (1 = target image, 0 = non-target image) on time at the trial level (from 1300 ms to 2500 ms after trial onset, centered with units corresponding to seconds), trial type (0=pre-switch, 1=post-switch) and the interaction of time and trial type. Subject (level 2) variables included age (in months, mean-centered), cognitive ability (as measured by DAYC-2 cognitive domain raw scores, mean-centered), and receptive language (as measured by PLS-5 AC raw scores, mean-centered). A full by-subject random effects structure including subject-level random effects for the intercept and all level 1 predictors was specified. As in the analyses for the first research question, we analyzed data from the first switch (trials 7, 8, 11, 12) and second switch (trials 17, 18, 21, 22) separately.

Results

Research Question 1: Diagnostic Group Differences in Prediction

The focal variable for the first research question was the interaction of diagnostic group and trial type. The diagnostic group (0=NT vs. 1=autism) x trial type (0=pre-switch vs. 1=post-switch) interaction was not significant in either the first switch (γ=−0.97, SE=0.88, t=−1.11, p=.272; Figure 4) or the second switch (γ=0.30, SE=1.13, t=0.27, p=.793; Figure 5) indicating no detectable group differences in looking behavior change from pre-switch to post-switch. However, when the autistic group served as the reference group (contrasts: 1=NT, 0=autism), there was a significant reduction in target looking from pre-switch to post-switch in the first switch (γ=−1.63, SE=0.66, t=−2.48, p=.016) but this effect was not significant in the second switch (γ=−1.69, SE=0.91, t=−1.86, p=.071). When the NT group served as the reference group (contrasts: 0=NT, 1=autism), there was not a significant reduction in target looking from pre-switch to post-switch in the first switch (γ=−0.66, SE=0.58, t=−1.14, p=.261) though this effect was significant in the second switch (γ=−1.99, SE=0.67, t=−2.98, p=.005). The only significant group effect in these models was the three-way interaction of diagnostic group, trial type, and time in the second switch only (γ=7.06, SE=2.95, t=2.40, p=.021). When the NT group served as the reference group, the time x trial type interaction in the second switch was significant and negative (γ=−5.50, SE=1.77, t=−3.11, p=.003) while the time x trial type interaction was not significant when the autistic group served as the reference group (γ=1.56, SE=2.36, t=0.66, p=.512). This finding suggests that, compared to the NT group, the autistic group showed less of a decrease in the logit probability of looking to target in relation to time for post-switch trials compared to pre-switch trials in the second switch. Model intercepts were significant across all models (ps<.001) indicating children in both groups on average anticipated target images in the pre-switch trials in both the first and second switch, though these findings were expected due to our decision to include only children whose average proportion looking to target exceeded .50. No other effects were significant in either model (ps > .05). Full model results can be found in Supplementary Materials (Table S3, S4, S5, and S6).

Figure 4.

Profile plot of raw proportion looks to target images across the analysis window (1300–2500 ms after trial onset) before and after the first switch for autistic and neurotypical (NT) children. Pre-switch trials (7 and 8) are in green, and post-switch trials (11 and 12) are in blue.

Figure 5.

Profile plot of raw proportion looks to target images across the analysis window (1300–2500 ms after trial onset) before and after the second switch for autistic and neurotypical (NT) children. Pre-switch trials (17 and 18) are in green, and post-switch trials (21 and 22) are in blue.

Research Question 2: Child Characteristics Associated with Prediction

For the second research question, we were interested in the relationship of child characteristics to autistic children’s performance in this linguistic prediction task. Specifically, we were interested in the relationship of child characteristics to anticipation of predictive contingencies (i.e., looking to target before the switch) and the relationship of child characteristics to hyperplasticity (as measured by change in target looking from pre-switch to post-switch). Therefore, because pre-switch trials served as the reference condition in this analysis (i.e., 0=pre-switch, 1=post-switch), the focal variables were the main effects of each child characteristic and the interaction of each child characteristic with trial type (pre-switch vs. post-switch). Characteristics examined in this analysis included age, cognitive ability, and receptive language. In the first switch, there was a significant main effect of cognitive ability (γ=0.27, SE=0.11, t=2.44, p=.020) though this effect was not significant in the second switch (γ=−0.15, SE=0.13, t=−1.14, p=.261). No other effects related to child characteristics were significant in either switch, (ps>.05). Full model results can be found in Supplementary Materials (Table S7 and S8).

In addition to the adjective-noun contingency, another variable that could have informed children’s predictions was the side (right or left) on which the target image appeared. Due to an error in the pseudo-randomization, target image side was partially (but not fully) confounded with adjective in the task design. We thus conducted a post-hoc investigation of trials in which target image side makes a strong probabilistic prediction to determine whether or not side is a stronger predictor than the linguistic target noun. Results indicated that image side was not the driving predictive cue for either diagnostic group. Full details of this analysis and results can be found in Supplementary Materials.

Discussion

This study represents a first test of the hyperplasticity claim as hypothesized by the PIA theory (Sinha et al., 2014) in the linguistic domain. By employing an AEM paradigm manipulating the strength of the predictive relationship between units in linguistic input, we were able to investigate whether linguistic sequence learning in young autistic children could be characterized as “hyperplastic” (Sinha et al., 2014). In our experimental design, children were taught two stable adjective-noun contingencies (“happy kitty” and “silly birdie”) over 8 trials. Those contingencies were switched in trials 9 and 10 such that “silly” predicted “kitty” and “happy” predicted “birdie.” The pattern returned to the original contingencies in Trial 11. As such, the cumulative probability diverged from recency in trials 11 and 12, allowing us to determine whether children made predictions based on cumulative learning or the most recent input. This pattern was then repeated such that trials 19 and 20 represented a second switch in the input. We expected the NT children to make predictions based on their cumulative experience with the adjective-noun contingencies and therefore make predictions after the switch that aligned most with cumulative probabilities. By contrast, if autistic children demonstrated hyperplasticity of learning, we could expect their post-switch predictions to reflect the most recent input (i.e., the switched contingencies). Finally, because prediction differences are posited to be domain general and underlie many aspects of the autistic phenotype over development, we expected children’s prediction task performance to relate to language, cognitive ability, and age.

Our findings do not support the presence of prediction differences in young autistic children reflective of hyperplasticity. Autistic children did not significantly differ from NT peers in pre-switch to post-switch looking behavior in either switch. During the first contingency switch, autistic children did show a significant reduction in looking to target in the post-switch trials relative to the pre-switch trials (following the expected pattern for hyperplasticity) while NT children did not, but this pattern flipped by the second switch such that NT children showed a significant reduction in target looking from pre-switch to post-switch while autistic children did not. The only evidence for group differences observed in this task was in the extent to which the probability of looking to target changed relative to time over the course of the trial between pre-switch and post-switch trials in the second switch. However, the direction of this effect suggested that autistic children showed less of a decrease in target looking over time after the switch relative to before the switch compared to NT peers. Together, these findings contradict the hyperplasticity claim of the PIA as it pertains to linguistic prediction. Our results align with several studies of nonverbal prediction in volatile environments which reported null group effects (Manning et al., 2017; Ward et al., 2021 between autistic and NT groups; Ward et al., 2022 between groups with high vs. low likelihood of autism) though other studies have found differences at the diagnostic group level (Amoruso et al., 2019; Greene et al., 2019). These conflicting results may be attributed to the different populations (i.e., children, adolescents, differing NT comparison groups) and/or methodologies characterizing each study. The current study adds to this body of evidence by extending investigations to the linguistic domain. Our research group and others have previously compared autistic children’s online predictive processing of stable, meaningful, and familiar linguistic contingencies to NT peers (i.e., Prescott et al., 2022; Zhou et al., 2019) finding no differences in prediction efficiency between groups. The present study built upon this work by investigating prediction of novel, unstable linguistic contingencies designed specifically to determine whether autistic children demonstrate hyperplasticity. Taken together with the extant evidence, the null results of the present study call into question the claims of the PIA including the domain-generality and universality of prediction differences across the autism spectrum and hyperplasticity of learning.

The second aim of this study was to explore factors underlying individual differences in prediction among autistic children. Model results indicated that autistic children with higher cognitive ability were better able to learn the stable predictive contingencies before the first switch, providing evidence for the importance of cognitive ability for linguistic predictive sequence learning. However, the effect of cognitive ability was not significant in the pre-switch trials of the second switch later in the experiment. It is possible that some children – even those with higher cognitive ability – might have become fatigued or bored over the course of this experiment. Alternatively, the first switch might have resulted in lingering effects on children’s predictions later in the experiment that were unrelated to cognitive ability. Together with the lack of interaction between cognitive ability and trial type (pre-switch vs. post-switch) in either switch, these findings suggest that cognitive ability is related to stable linguistic contingency learning among autistic children but may be less relevant to modulation of predictive behavior in response to volatile input.

Contrary to our expectations, age and receptive language were not significantly related to any aspect of autistic children’s performance on this linguistic prediction task. There are several possible explanations for these surprising findings. First, our sample may not have been sufficiently large to observe subtler effects of these characteristics. Moreover, we may have lacked sufficient variability in our sample to detect effects, particularly with respect to age. It is also possible that effects of these characteristics might emerge in experimental tasks that are more complex or engaging than the present study, which employed a relatively simple and repetitive design. With regard to receptive language, it is also possible that children’s predictions during this task were largely superficial, based on phonetic form rather than semantic representations. The adjectives and nouns were intentionally semantically unrelated in this task and may therefore index only a portion of the processes involved in real-world linguistic prediction. For example, studies have demonstrated that predictions about upcoming nouns based on semantically informative verbs are related to receptive language ability in young autistic children (Prescott et al., 2022; Venker et al., 2019). An interesting future line of research could investigate the effect of prediction errors in semantically constrained contexts in autistic children and its relationship to language ability. Finally, it may be the case that prediction and receptive language ability have a longitudinal relationship, whereas our current data capture skills at only one time point. Our research group will further investigate this possibility in future studies.

Several limitations must be noted for the present study. First, children in our sample were all monolingual English speakers, primarily white, and non-Hispanic or Latino. Thus, these results may not generalize to children from all backgrounds and should be examined with a more representative sample (Girolamo et al., 2023). Second, the NT group had significantly higher maternal education (p<.001) than the autistic group, likely due to differences in our recruiting process for each group (i.e., NT children were primarily recruited from the local community, while autistic children were recruited from across the broader region in order to increase the sample size). However, the fact that the NT children did not outperform the autistic group suggests this difference was likely unrelated to study results. Next, while we intentionally selected adjectives for their semantic neutrality with regard to the target nouns, it should also be noted that using familiar words in this experiment could have introduced bias if any children associated an adjective with one noun more than the other. Future studies could employ novel words to eliminate any risk of bias. Additionally, our use of a standardized assessment yielded only a broad measure of cognitive ability, limiting our ability to pinpoint specific facets of cognition that related to prediction. Future studies might consider employing more fine-grained measures to investigate the precise cognitive abilities most relevant to prediction in autism.

We also note the limitations of the present study resulting from below-chance pre-switch contingency learning in the initial cognitive ability-matched NT group. In order to draw conclusions about children’s anticipation after the contingency switches, it was imperative that they showed evidence of having learned the initial contingency pairs. Therefore, we returned to the full tested sample and included only children whose proportion looks to target images during the analysis window of the pre-switch trials exceeded .50. While this approach enabled us to more fairly compare change in anticipatory looking from pre-switch to post-switch between groups, it also resulted in groups that differed significantly on all child characteristic variables measured, including receptive language, age, and cognitive ability (Table 1). Therefore, we cannot rule out the possibility that autistic children might show differences reflective of hyperplasticity when compared to NT peers matched on one or more of these variables. In particular, future research might compare autistic children at an older age where NT children matched on cognition would be more likely to successfully learn pre-switch predictive contingencies. Age-matched non-autistic peers with developmental delay may also provide an interesting basis for comparison in future studies. Finally, we acknowledge the choice of a .50 threshold as a proxy for successful pre-switch contingency anticipation at the individual level as somewhat arbitrary. As such, we conducted post-hoc analyses replicating the approach of the first research question with a .55 threshold to examine the effect of stricter criteria for pre-switch anticipation. The results mirrored the pattern of the primary findings and are included in Supplementary Materials (Table S9, S10, S11, and S12).

A final limitation of this study was that in analyzing the two switches separately, our findings did not allow for any interpretation of longitudinal, experiment-level change in proportion target looking from the first switch (trials 7, 8, 11, 12) to the second switch (trials 17, 18, 21, 22). While not a goal of the present study, the evolution of young autistic children’s predictive looking to linguistic stimuli over the course of an experiment given multiple contingency switches poses an interesting open research question that may be addressed in future work.

Conclusions

This study did not detect prediction differences indicative of hyperplasticity among autistic children relative to NT peers when linguistic input is more volatile. Among autistic children, cognitive ability was positively related to stable contingency learning (i.e., before the first switch), while age and receptive language were not related to any aspect of task performance measured. In sum, these findings do not provide support the framework proposed by Sinha and colleagues (2014) in the linguistic domain for autistic children represented by this sample. However, future work might investigate autistic children’s prediction behavior compared to non-autistic peers of similar age or cognitive ability, over longer time periods, and in semantically-constraining contexts to more richly characterize autistic children’s linguistic prediction in volatile environments. Clinically, a future line of research may be warranted to determine how best to support autistic children with lower cognitive abilities who may have more difficulty learning predictive contingencies from linguistic input. More work will be needed to specify the effect of prediction errors in autistic children’s language learning and determine whether altering the predictability of input could improve learning in a therapeutic context.