Automated Quantification of Head Motion Synchrony in Peer Interactions of Autistic and Non‐Autistic Youth

2.1. Participants

Youth were recruited from the Washington, D.C. metro area to participate in a lab‐based peer interaction as part of two separate studies. For the first study, autistic (n = 33) and neurotypical (n = 103) youth (aged 8–16 years) participated in a lab‐based peer interaction between Fall 2018 and Spring 2020. For the second study, autistic (n = 31) and non‐autistic (n = 69) youth (aged 11–15 years) participated in a lab‐based peer interaction between Summer 2022 and Summer 2025. Youth were recruited from previous participation in research, advertisements on Facebook, outreach at local events, or through the Simons Foundation Powering Autism Research (SPARK). We appreciate obtaining access to recruit participants through the SPARK research match on SFARI Base. All procedures were approved by the University of Maryland Institutional Review Board, and participants and their parents provided informed consent and assent.

All participants had a full‐scale IQ of 80 or higher based on administration of the Kaufman Brief Intelligence Test, 2^nd edition (KBIT‐2) (Kaufman and Kaufman 2004) and were native English speakers. For the first study, neurotypical participants were screened and excluded for any neurological or psychiatric disorders, or first‐degree relatives with schizophrenia or autism. For the second study, non‐autistic participants were screened and excluded for first‐degree relatives with schizophrenia or autism, but participants were not excluded for common co‐occurring conditions with autism (anxiety, depression, obsessive‐compulsive disorder, or ADHD). Because of the differences in exclusion procedures between the two studies, participants from the first study are referred to as “neurotypical”, while participants from the second study are referred to as “non‐autistic”. In the combined sample, participants are referred to as “non‐autistic” because both sets of exclusion criteria are represented.

For the first study, autistic participants had their autism diagnoses confirmed by administration of the Autism Diagnostic Observation Schedule, Second Edition (ADOS‐2; Lord et al. 2012) by a clinician who met research reliability criteria of administration and coding the ADOS‐2. For the second study, autistic participants had their autism diagnoses confirmed by the administration of the Autism Diagnostic Interview‐Revised (ADI‐R; Rutter et al., 2003) by a clinician who met research reliability criteria of administration and coding of the ADI‐R. Research reliability refers to meeting inter‐rater reliability of at least 80% on the ADOS‐2 and at least 90% on the ADI‐R with a trained clinician and coder.

The participants were paired into dyads corresponding to one of three dyad types: autistic participants paired with autistic participants (matched autistic dyads), autistic participants paired with non‐autistic participants (mismatched neurotype dyads), and non‐autistic participants paired with non‐autistic participants (matched non‐autistic dyads). All dyads consisted of participants of the same gender (parent‐reported for study 1, child‐reported for study 2) and similar chronological age (within 1 year or grade level). Dyads were excluded if they had previously met each other outside of the study context. One matched non‐autistic dyad was excluded for meeting this criterion, resulting in a sample of 117 dyads for analysis (66 matched non‐autistic dyads, 38 mismatched neurotype dyads, 13 matched autistic dyads). 224 participants were part of only one dyad. 10 participants were part of two dyads; these participants were part of a dyad in study 1 and then returned in study 2 and participated in a second dyad. Although dyad types were intended to be equal across dyad types, the COVID‐19 pandemic interrupted data collection. Therefore, comparisons with the smaller matched autistic dyad type should be considered preliminary due to the small sample size. Demographic information for the two study samples, including participants’ age, gender, race, ethnicity, and socioeconomic status, is provided in Table 1 (Study 1) and Table 2 (Study 2). The study 1 sample overlaps with a sample used in previously published analyses on smiling synchrony (n = 136) and theory of mind (n = 100) in peer interactions between autistic and neurotypical youth (McNaughton et al. 2024; Alkire et al. 2022). The study was not preregistered. No a priori power analysis was performed.

2.2. Peer Interaction Task and Post‐Interaction Questionnaires

Each dyad participated in a 5‐min getting‐to‐know‐you conversation as the first activity as part of a larger peer interaction (Study 1: 25‐min total interaction length, Study 2: 20‐min total interaction length). Data from the 5‐min getting‐to‐know‐you conversation was used for these analyses to focus on unstructured initial conversations and to harmonize data across the two studies.

For the 5‐min conversation, youth sat in chairs approximately 0.7 m across from each other in a room with a small, low table and three video cameras on one side. The experimenter told the youth that they would complete a few short tasks together, and that they should get to know each other before the experimenter came back in five minutes to explain the first task. The interactions were recorded using Canon Vixia HR F80 video cameras at a resolution of 1280 × 720 and 30 frames per second. For the first study, youth had no additional wearable technology. For the second study only, youth wore mobile eye‐tracking glasses attached to a phone that was stored in a fanny pack.

Following the face‐to‐face interaction, youth in both studies moved to separate rooms to complete activities independently, including a six‐item questionnaire describing their perceptions of the interaction (“Interaction Quality”; sum of six items with higher values indicating better quality) (Berry and Hansen 1996). Youth in study 1 also completed a sliding scale question measuring their desire to interact again (‘Desire to Interact Again’; −100 to 100 scale with very positive indicating strong desire to indicate with that partner again). Additional details on wording of the post‐interaction questionnaires is provided in Supplemental Information, and both measures have been previously used in studies of interaction quality in autistic and neurotypical youth (McNaughton et al. 2024). Participants in study 1 also completed additional questions on their discomfort following the interaction, which were summed into a four‐item composite with higher values indicating higher perceived discomfort during the interaction (see Supplemental Information for additional details on wording of items). Some participants did not complete one or more of the questionnaires, because the questions were added to the study protocol after the participants had completed their visit or because the participant chose to end their visit before completing the questions. Four participants in study 1 (n = 1 autistic, n = 3 non‐autistic) did not complete the interaction quality questionnaire (n = 4 because the questionnaire was added to the study protocol after participants completed their visit), 24 participants in study 1 (n = 11 autistic, n = 13 non‐autistic) did not complete the desire to interact again questionnaire (n = 20 because the questionnaire was added to the study protocol after participants completed their visit, n = 4 because participants chose to end their visit before completing the questionnaire), and four participants in study 1 (n = 1 autistic, n = 3 non‐autistic) did not complete the discomfort questions (n = 4 because the questionnaire was added to the study protocol after the participants completed their visit). Youth‐reported interaction quality and discomfort did not significantly differ by dyad type. Youth‐reported desire to interact again was significantly higher in matched non‐autistic dyads compared to mismatched neurotype dyads (B = 36.79, t(54.05) = 2.49, p = 0.02) and significantly higher in matched autistic dyads compared to mismatched neurotype dyads (B = 54.21, t(56.07) = 2.41, p = 0.02), while matched autistic and matched non‐autistic dyads did not significantly differ (B = 17.42, t(56.25) = 0.90, p = 0.37). These findings are consistent with previously reported analyses in an overlapping sample McNaughton et al. 2024).

Dyads were not informed by the research team about whether their peer interaction partner had an autism diagnosis or not. Some participants chose to self‐disclose their diagnosis during the interaction. A total of five autistic participants chose to self‐disclose: two autistic participants in mismatched neurotype dyads, and three autistic participants in matched autistic dyads, two of whom were in the same dyad.

2.3. OpenPose Analysis of Head Motion Coherence

Video recordings of the five‐min getting‐to‐know‐you interaction were analyzed using OpenPose (Cao et al. 2021) software to extract x and y coordinates for 25 body parts, as well as the confidence for the algorithm's location estimate. This labeling process occurred for each member of the dyad for each frame of the video file. The first 9000 frames, corresponding to exactly five minutes at 30 frames per second, were analyzed for each interaction. Based on previous research identifying the importance of head motion synchrony in dyadic conversations (Hale et al. 2020), only the nose point for each participant was used for the present analysis to quantify head motion. Because video footage was obtained in a profile view, analyzing the nose point meant that nodding head motion in the pitch axis of motion was largely represented, consistent with previous work identifying significant evidence of pitch head motion synchrony (Hale et al. 2020). Figure 1 and Video S1 present the analysis workflow for a brief clip from an example dyad.

Data were processed if participants were within the video frame for at least 80% of the video and the interaction was at least five minutes/9000 frames. Data from 10 dyads (n = 1 matched autistic dyad, n = 5 mismatched dyads, n = 4 matched non‐autistic dyads) did not meet this threshold and were not processed further. Once processed, data were considered useable if at least 75% of frames were identified as having nose point confidence ratings for both participants above 0.80. The nose point confidence ratings were assigned on a 0 (lowest) to 1 (highest) scale by the computer algorithm, and the 0.80 cutoff was confirmed as representing a cutoff after which data quality decreased based on comparisons to a subset of files annotated by a human coder. Data from five dyads (n = 2 matched autistic dyads, n = 1 mismatched dyad, n = 2 matched non‐autistic dyads) did not meet this threshold. Data were subject to further manual inspection, and two matched non‐autistic dyads were excluded based on discovery of artifacts. Following these exclusions, 100 dyads were included in analyses across the two studies (study 1: n = 5 matched autistic dyads, n = 16 mismatched dyads, n = 38 matched non‐autistic dyads, study 2: n = 5 matched autistic dyads, n = 16 mismatched dyads, n = 20 matched non‐autistic dyads).

Frame‐by‐frame data were concatenated to form time series of nose position for each participant using in‐house scripts written in R (available on OSF: https://osf.io/tq7su/). Missing data, which occurred due to occlusion of the nose or the nose being outside of the video frame, were imputed using stine imputation to reflect the non‐linear nature of the position data. An average of 40 frames were imputed for each participant (min = 0, max = 1198), with an average maximum consecutive frames imputed per participant of 13 (min = 0, max = 436) across the 9000 frame time series. The time series representing the x and y nose coordinate positional data were then transformed into a frame‐by‐frame change in distance metric using Euclidean distance, yielding one time series of head motion for each participant. The absolute difference between each sequential point on this time series was summed to create a metric of overall motion that was included in analyses as a covariate.

Synchrony was quantified using cross‐wavelet coherence, which allows the time‐frequency nature of head motion to be respected, such that synchrony between slower low frequency head movements (e.g. gradual shifts) is disentangled from synchrony between faster high frequency head movements (e.g. fast nods (Hale et al. 2020)). The head motion time series for each participant was subjected to a wavelet transformation with parameters previously used for analyzing social coordination data (Morlet wavelet, w = 6; (Hale et al. 2020)). Within the dyads, the cross‐wavelet coherence was calculated at 73 frequencies in the 0.0625–4 Hz frequency range. This 0.0625 Hz‐4 Hz frequency range is in line with the frequency (30 Hz) and length (150 seconds per interaction half) of data collection in the present study and consistent with previously identified frequencies of interest for head motion synchrony (Hale et al. 2020). This output was then averaged into six frequency bands to aid interpretation and direct comparison with previous work (0.0625–0.125 Hz, 0.125–0.25 Hz, 0.25–0.5 Hz, 0.5–1 Hz, 1–2 Hz, 2–4 Hz (Fujiwara and Yokomitsu 2021)), resulting in one coherence value per frequency band per dyad.

2.4. Statistical Analysis

To establish evidence of group‐level coherence, data from true pairs were statistically compared to time‐scrambled data from those same pairs. That is, wavelet transformations from the first half of individual A's true time series and the first half of individual B's true time series, as well as wavelet transformations from the second half of individual A's true time series and the second half of individual B's true time series were each subjected to a cross‐wavelet transformation and averaged to extract a single “true” coherence value. Wavelet transformations from the first half of individual A's time series and the second half of individual B's time series, as well as from the second half of individual A's time series and the first half of individual B's time series were subjected to a cross‐wavelet transformation and then averaged to extract a single “pseudo” coherence value. Six paired t‐tests were performed to compare data from these true pairs and pseudo pairs, one test for each of the frequency bands, with false discovery rate correction for multiple comparisons across 6 frequency bands (0.0625–0.125 Hz, 0.125–0.25 Hz, 0.25–0.5 Hz, 0.5–1 Hz, 1–2 Hz, 2–4 Hz). Analytic results for the comparison between true and pseudo coherence in the original 73 frequency values obtained from the wavelet analysis before averaging into the six frequency bands are presented in Supplemental Information (Figure S2).

For the remaining analyses on dyad type differences in synchrony and relations between synchrony and interaction enjoyment, we report coherence in the frequency band(s) that showed significant differences between true and pseudo pairs. These analyses are confirmatory analyses testing our primary hypotheses. We provide details on frequency bands that did not display evidence of significant synchrony in the Supplemental Information for completeness of reporting.

To assess dyad type differences in synchrony, linear models were performed with dyad type (matched autistic, mismatched neurotype, matched non‐autistic) as the predictor of interest and coherence in a given frequency band as the outcome of interest (model details in ‘Dyad Type Differences in Head Motion Synchrony Supplemental Information ’). Age, gender, data quality (proportion of frames missing from analysis/below a confidence level of 0.8), overall motion, and study (study 1 or 2) were included as covariates based on potential for confounds. Results were robust to inclusion/exclusion of covariates. Follow‐up Bayesian ANOVAs were performed in JASP v0.95.1 (JASP Team 2025) to quantify the evidence for and against dyad type differences (Van Den Bergh et al. 2020; Van Doorn et al. 2021).

To assess relations between synchrony and interaction outcomes (interaction quality, desire to interact again, discomfort), multilevel models were constructed treating youth‐reported interaction enjoyment as nested within dyads, and intercepts were modeled as randomly varying across dyads. The discomfort outcome was square‐root transformed to better meet normality assumptions. Coherence in a given frequency band was the predictor of interest, and time‐scrambled pseudocoherence, age, gender, dyad type (matched autistic/mismatched neurotype/matched non‐autistic), data quality (proportion of frames missing from analysis/below a confidence level of 0.8), overall motion, and study (study 1 or 2, for the interaction quality outcome that was collected across both studies) were included as covariates of no interest. Results were robust to inclusion/exclusion of covariates. Each outcome (interaction quality, desire to interact again, discomfort) was treated as its own family of tests as prior work has shown these outcomes may reflect separable constructs (Berridge et al. 2009). Analyses were performed in R (R Core Team 2020) using the lmerTest package (Kuznetsova et al. 2017).

3.1. Evidence of Significant Head Motion Synchrony in Naturalistic Peer Interactions

To test for evidence of significant head motion synchrony in naturalistic peer interactions, we compared coherence for true dyads and time‐swapped pseudo dyads using paired t‐tests for each of the six coherence bands across all 100 dyads, with false discovery rate correction applied for six coherence bands. There was significant evidence of coherence in the 0.0625–0.125 Hz (t(99) = 2.77, p _corrected = 0.02), 0.25 Hz–0.5 Hz (t(99) = 3.66, p _corrected = 0.002), and 1–2 Hz (t(99) = 2.51, p _corrected = 0.03) frequency bands (Figure 2). There was no significant evidence of coherence in the 0.125–0.25 Hz frequency band (t(99) = 2.02, p _corrected = 0.07), 0.5–1 Hz (t(99) = 1.14, p _corrected = 0.31), or 2–4 Hz (t(99) = 0.18, p _corrected = 0.86) frequency bands.

3.2. Dyad Type Differences in Head Motion Synchrony

To test for dyad type differences in head motion synchrony, we conducted three linear models for each of the frequency bands with significant evidence of coherence as the outcome and dyad type (matched autistic, mismatched neurotype, matched non‐autistic) as a predictor of interest. There were no significant differences in head motion synchrony between dyad types in the three frequency bands that demonstrated significant evidence of above‐chance coherence: 0.0625–0.125 Hz, 0.25–0.5 Hz, or 1–2 Hz (Figure 3; Tables S1–3). Follow‐up Bayesian analyses to quantify support for and against an effect of dyad type revealed moderate evidence against an effect of dyad type in the frequency bands that demonstrated evidence of above‐chance coherence (Table S7).

3.3. Relations Between Head Motion Synchrony and Interaction Enjoyment

To test for relations between head motion synchrony and interaction enjoyment, we conducted multilevel models with head motion synchrony in a given frequency band as the predictor of interest. For the frequency bands with above‐chance evidence of coherence, separate models were constructed for each of the three outcomes of interest, which were administered at different studies: interaction quality (including participants from study 1 and 2, study included as covariate), desire to interact again (participants from study 1), and discomfort (participants from study 1). For the six‐item interaction quality outcome, head motion synchrony in the 0.0625–0.125, 0.25–0.5 Hz, and 1–2 Hz bands did not significantly predict interaction quality (Table S8,S11).

For the desire to interact again sliding scale, head motion synchrony in the 0.25–0.5 Hz frequency band significantly negatively predicted desire to interact again (B = −583.07, t(40.57) = −2.91, p = 0.01; Table S9), such that youth reported less desire to interact again with their peer partner after interactions with greater head motion synchrony (Figure 4A). Head motion synchrony in the 0.0625–0.125 and 1–2 Hz bands did not significantly predict desire to interact again (Table S12). Findings remained significant with false discovery rate correction for the three above‐chance coherence bands.

For the discomfort measure, head motion synchrony in the 0.25–0.5 Hz (B = 5.58, t(49.07) = 2.72, p = 0.01; Table S10) frequency band significantly positively predicted discomfort, such that youth reported experiencing greater discomfort with their peer partner during interactions with greater head motion synchrony (Figure 4B). Head motion synchrony in the 0.0625–0.125 and 1–2 Hz bands did not significantly predict discomfort (Table S13). Findings remained significant with false discovery rate correction for the three above‐chance coherence bands.

Exploratory models were also constructed including statistical interaction terms between dyad type and head motion synchrony to predict interaction outcomes. There were no significant statistical interactions between dyad type and head motion in predicting interaction quality, desire to interact again, or discomfort that remained after correction for multiple comparisons across frequency bands (ps>0.05). Models were also constructed for relations between synchrony and interaction outcomes in each of the frequency bands that did not display significant above‐chance evidence of synchrony (0.125–0.25 Hz, 0.5–1 Hz and 2–4 Hz) for completeness of reporting (Tables S11–S13).

4.1. Head Motion Synchrony Naturally Occurs at Above‐Chance Levels in Youth Conversations

We found evidence of above‐chance synchrony in frequency bands ranging from 0.0625–2 Hz. This finding of significant synchrony in more gradual, lower frequency head movements (in contrast to high frequency movements) is consistent with previous findings from adults engaged in conversational tasks (Falk et al. 2023; Hale et al. 2020), supporting the existence of spontaneous social coordination across development (Rauchbauer and Grosbras 2020).

Because of the camera location and selection of the nose point for analysis, the analyzed head motion time series largely represent nodding and activity in the “pitch” axis of head motion. Previous work has suggested that these low frequency nods may be indicative of gaze following behaviors (Falk et al. 2023), such that low frequency synchrony could indicate gaze synchrony between partners. Overall, our results reinforce the idea that synchrony occurs in specific types of head movements, not across all head movements.

The use of computer vision methods to capture significant evidence of synchrony in similar frequency bands to previous work with wearable devices (Falk et al. 2023; Hale et al. 2020) supports the use of these methods for quantifying naturally occurring synchrony. Non‐invasive computer vision methods may be particularly important for populations with sensory sensitivities, such as autistic individuals, and research questions that would be affected by participants being aware of the wearable devices or feeling them impact their motion patterns. Furthermore, because this analytic pipeline uses video recordings that are commonly obtained in research and clinical settings, our approach is economical and practical for use on pre‐existing datasets, presenting practical advantages for many developmental science questions of interest.

4.2. Dyad Types Do Not Differ in Head Motion Synchrony

We found no evidence for differences in head motion synchrony across dyad types (matched autistic, matched non‐autistic, mismatched neurotype). Previous findings have pointed to motion synchrony differences between mismatched neurotype dyads and matched non‐autistic dyads (Carnevali et al. 2024; Glass and Yuill 2023). Specifically, previous work investigating conversations between adults has demonstrated reduced head and whole‐body motion synchrony in mismatched neurotype dyads compared to matched non‐autistic dyads (Georgescu et al. 2020; Koehler et al. 2022), and previous work investigating social interactions between children and adult experimenters has found reduced head and whole‐body synchrony for autistic children interacting with experimenters compared to non‐autistic children interacting with experimenters during structured tasks (Noel et al. 2018; Zadok et al. 2022).

One important difference between our work and prior work is that our study is the first to investigate head motion synchrony in youth peer interactions. Youth display different speaker and listener behaviors when interacting with peers compared to adults (Turkstra 2001), highlighting the importance of assessing social interactions in the social context of interest. Future research could more directly test differences between adult‐child and peer interactions by having children engage in separate interactions with both a peer and an adult partner, enabling within‐subjects comparisons of synchrony in these two distinct social environments.

Another difference is that, in the present study, youth engaged in an unstructured conversation, while other research has used more structured tasks (Noel et al. 2018; Zadok et al. 2022). Task context may impact synchrony. One recent study examined unstructured play interactions between children and an adult experimenter and found that dyad type differences in head motion synchrony varied based on whether there was joint engagement in the interaction (X. Chen et al. 2023), highlighting that within‐interaction dynamics play an important role in synchrony differences. Previous research in adults has also demonstrated that head motion synchrony differs across tasks (Falk et al. 2023). Interestingly, previous research using an unstructured conversation similar to our approach also did not find dyad type differences in synchrony (Efthimiou et al. 2025), while research using directed discussion prompts or semi‐structured conversations did find dyad type differences in synchrony (Georgescu et al. 2020).

Finally, previous work on synchrony has utilized windowed cross correlations to quantify the presence and degree of synchrony. Windowed cross‐correlations focus on the time‐domain, such that they identify synchrony across the time series as a whole rather than within different frequency bands (Fujiwara and Daibo 2016), and have previously provided valuable insight into the presence of dyad type differences in synchrony (Georgescu et al. 2020). Furthermore, parameter choices in the windowed cross correlation, such as the amount of lag in the time domain at which changes can be quantified as “synchronous” (e.g. 5 seconds (Georgescu et al. 2020; Koehler et al. 2022)), influence how synchrony is identified within a time series. The absence of differences using cross‐wavelet methods of synchrony quantification could reveal that underlying synchrony differences between autistic and neurotypical dyads are not strong in the frequency dimension, highlighting how different approaches to quantifying synchrony provide a more complete picture of the nature of synchrony differences between autistic and neurotypical individuals.

4.3. Head Motion Synchrony Predicts Decreased Interaction Enjoyment and Increased Discomfort

We identified negative relations between head motion synchrony and interaction enjoyment, which contrasts with previous findings of positive social outcomes (increased rapport, closeness, etc.) in response to experimentally‐induced synchrony, for a range of body motions (Hove and Risen 2009; Mogan et al. 2017; Rennung and Göritz 2016; Vicaria and Dickens 2016). Our findings further contrast with a previous study identifying positive relations between rapport and spontaneous motion synchrony in a face‐to‐face conversation, although this study investigated whole‐body synchrony, not head synchrony specifically (Fujiwara et al. 2020). An advantage of our approach using cross‐wavelet coherence is that it points to specific synchrony patterns that link to reduced enjoyment, as we find these relations specifically in low frequency head movements, not in high frequency rapid nods, and future work could supplement these findings with additional manual coding of specific movements.

One interpretation of these findings could be that our metric of synchrony captures markers of synchronized discomfort, which then link to reduced enjoyment. We found relations between a composite of youth‐reported discomfort in the interaction and increased synchrony in the interaction. In the getting‐to‐know‐you context with an unfamiliar peer, higher synchrony in low frequency bands could capture aspects of synchronized head motion like mutual “cringes” of discomfort, shared awkward laughing, or synchronized gazing at other parts of the room, which then relate to increased youth‐reported discomfort. Future work that measures moment‐to‐moment youth‐reported discomfort in the interaction could better identify these markers of discomfort and link them to synchrony over the time series of the interaction.

Our findings also fit in with a growing body of work emphasizing that more synchrony is not always better (Mayo and Gordon 2020). Instead, rapport may be influenced by other factors like divergence, novelty, and exploration, rather than exclusively synchrony (Ravreby et al. 2022; Speer et al. 2024). Furthermore, the role of synchrony may depend on context, such that reduced synchrony may help with problem solving or indicate conversation ease. In a problem solving task, weaker full‐body synchrony in interacting adult dyads predicted better task performance (Abney et al. 2015). Increased lower‐frequency head motion synchrony was associated with conversations with greater background noise, which could be interpreted as a marker of conversation difficulty (Hadley and Ward 2021). The importance of synchrony in context is also seen in the developmental origins of synchrony; that is, moderate, rather than high or low, synchrony in mother‐child interactions at 4 months predicts secure attachment at 12 months (Beebe et al. 2010). Together, these studies emphasize roles for reduced synchrony in relationship development, conversational ease, problem solving, and novelty, all of which may support links between reduced synchrony and increased interaction enjoyment.

From this variety of perspectives, synchrony emerges not as a universal advantage, but as a marker with differential impacts across contexts. In this context of a brief getting‐to‐know‐you interaction between previously unacquainted youth, we found that head motion synchrony is associated with less interaction enjoyment and more discomfort. Our study highlights the importance of investigating social phenomena in naturalistic contexts and directly linking features of interactions to youth reports, as findings from experimental inductions of synchrony may not generalize to the phenomenon of naturally occurring synchrony. Because automated approaches are increasingly deployed to identify differences between interactions involving autistic and non‐autistic individuals, it important to consider the meaning of the features and how they relate back to participants’ experiences of the interaction. This validation of these conversational features will lead to the best understanding of youth peer interactions for autistic and non‐autistic youth.

4.4. Limitations and Future Directions

Given the small sample of matched autistic dyads, findings including this dyad type should be viewed as preliminary. Previous work has identified reduced motion synchrony between matched autistic and matched non‐autistic dyads (Georgescu et al. 2020; Stoit et al. 2011), therefore our null findings of differences between matched autistic and matched non‐autistic dyads may reflect the fact that the study was underpowered to identify differences. Including larger samples of matched autistic dyads in future work is crucial to understand autistic youth sociality across partner types, as there are differences in interaction enjoyment and interaction features between matched autistic and mismatched neurotype youth peer interactions (Y.‐L. Chen et al. 2021; McNaughton et al. 2024).

Future work could also use further approaches to quantify social interaction dynamics. For example, in line with recent theoretical calls to consider the dynamics of increased and reduced synchrony across the time course of the interaction, measures of metastability could be used to better calculate flexibility or entrance in and out of synchrony (Mayo and Gordon 2020; Tognoli et al. 2018). Additionally, future work could also take into consideration the multimodal nature of social interactions by integrating head motion synchrony with other constructs (e.g. turn taking, gaze etc.) to more fully characterize conversation dynamics (Fauviaux et al. 2024). These approaches will provide a richer characterization of social dynamics in peer interactions between autistic and non‐autistic youth, with further implications for understanding characteristics of interactions related to enjoyment and discomfort.