Introduction

Autism spectrum disorder (ASD) and psychotic disorders share overlapping symptoms and were historically considered closely related1,2. Early theories even proposed ASD as an early manifestation of psychosis3,4. However, since the publication of the Diagnostic and Statistical Manual of Mental Disorders, third edition (DSM-III)5, ASD and psychotic disorders have been classified as distinct diagnostic categories. ASD is defined as a neurodevelopmental disorder characterized by persistent deficits in social communication and interactions, along with restricted and repetitive patterns of behavior, interests, or activities6. It is typically diagnosed in childhood and tends to remain stable over time or exhibit clinical improvement7. In contrast, psychotic disorders such as schizophrenia are marked by positive symptoms including delusions and hallucinations, as well as negative symptoms such as reduced emotional expression and social withdrawal. These disorders are usually diagnosed in late adolescence or early adulthood and are often associated with a progressive decline in cognitive and functional abilities8,9.

Despite these diagnostic distinctions, ASD and psychotic disorders share significant genetic, pathophysiological, and neurobiological features10,11. Psychotic disorders such as schizophrenia are widely recognized as having neurodevelopmental origins10,12,13. In 30 to 50 percent of cases, childhood-onset schizophrenia is preceded by or comorbid with pervasive developmental disorders. Furthermore, individuals with ASD frequently report psychotic-like experiences (PLEs), which are milder forms of psychotic symptoms, during early adulthood10,14,15. Similarly, ASD symptoms have been observed among adults diagnosed with psychotic disorders16.

PLEs, including attenuated delusions, hallucinations, and negative symptoms, are common in non-clinical populations and are believed to exist along a continuum within the psychosis spectrum. These experiences may predict the later onset of psychotic disorders17,18. Evidence from non-clinical samples also supports an association between autistic traits and PLEs. For instance, higher levels of autistic traits, particularly in relation to speech delay and ritualistic behaviors at age 3, have been linked to increased risk of PLEs by age 1219. In addition, a diagnosis of ASD has been associated with a threefold increase in the likelihood of experiencing PLEs, with each standard deviation increase in autistic traits corresponding to a 17 percent increase in odds20. Positive psychotic symptoms have also been observed in approximately 5.4% of children and adolescents with ASD21. However, the findings are not consistent. A twin study reported weak and non-significant associations between autistic traits and PLEs in youth, suggesting heterogeneity in these relationships15.

Theoretical models have proposed that ASD and psychotic disorders reflect opposite ends of a continuum based in social cognition22. Both conditions are thought to originate from atypical development of the social brain23,24,25,26. According to this model, ASD is characterized by mechanistic cognition, which focuses on physical systems and rules, while psychotic disorders are associated with hyper-mentalistic cognition, involving excessive attribution of mental states to others22,27. Both conditions deviate from typical social cognitive functioning, although in different directions23,28.

These distinctions are reflected in social cognition profiles. In ASD, deficits in theory of mind and social communication contribute to impaired social interaction29,30,31. In contrast, individuals with psychotic disorders often demonstrate altered social cognition, including paranoid misinterpretations of others’ intentions32. Despite these contrasting patterns, both groups show impairments in core social cognitive domains, such as theory of mind, facial emotion recognition, and social and adaptive functioning33,34,35.

Empathy, defined as the capacity to understand and share the emotions of others, is another domain which has been linked to both ASD and psychotic disorders36,37,38. However, findings in psychosis are mixed. Some studies suggest that emotional empathy may be relatively preserved. For example, Berger et al.39 reported no significant differences in emotional empathy between individuals with psychosis and healthy controls, based on both self-report and behavioral measures. Similarly, Aghvinian & Sergi40 found no significant differences in empathy between individuals with high and low levels of schizotypy. Comparative studies examining both conditions also revealed associations with empathy. Pepper et al.41 found that individuals with ASD and those with early psychosis scored lower on empathy measures compared to controls. Likewise, Corbera et al.42 reported general empathy deficits in both groups, although quality of life was strongly predicted by negative symptoms than by empathy impairments.

Although previous research has explored the relationships among autistic traits, PLEs, and empathy, few studies have considered their co-occurrence in the general population. This limitation is important, as the interaction of these traits may significantly affect social and emotional functioning43,44. Additionally, small sample sizes and limited consideration of sex differences have restricted the generalizability of earlier findings15,20. Given that both autistic traits and PLEs are dimensional constructs observed across clinical and non-clinical populations, research in large, community-based samples is essential to advance theoretical understanding and inform clinical practice.

To address these gaps, the present study investigates the associations among autistic traits, PLEs, and empathy in a large, community-based sample of preadolescents, generally defined as individuals between 9 and 12 years of age45. Data were drawn from the Adolescent Brain Cognitive Development (ABCD) study, which includes a demographically diverse cohort of early adolescents. This developmental period is characterized by significant cognitive, emotional, and social changes46.

Using network analysis, the present study aims to examine the associations among autistic traits, PLEs, and empathy47. Network analysis conceptualizes psychological constructs as systems of interacting components, where nodes represent variables and edges reflect statistical associations between them47. A Gaussian graphical model (GGM) was first estimated, in which regularized partial correlations offer insights into the structure of interrelations while controlling for all other variables48. However, GGMs do not provide information about the directionality of these associations. To address this limitation, a Bayesian network analysis with a directed acyclic graph (DAG) was employed. This approach estimates probabilistic directional relationships between variables and may provide insights into the potential causal pathways49.

Network analysis was chosen as the most suitable technique to model complex associations among the key variables in this study. Unlike regression or structural equation modeling (SEM), which typically require the specification of dependent and independent variables based on a predefined theoretical model, network analysis treats variables as mutually interacting components. This data-driven approach allows for a more nuanced understanding of the relationships among constructs without assuming a specific direction.

Both autistic traits and psychotic-like experiences (PLEs) manifest differently by sex. Autistic traits are more commonly observed in males, while females often report greater distress associated with PLEs44,50. Empathy also shows sex-related variation, with females typically exhibiting higher levels of empathic concern than males51. Despite these differences, research examining how sex may influence the interplay among autistic traits, PLEs, and empathy remains limited. By analyzing these associations separately in males and females, and comparing their respective network structures, this study aims to identify potential sex-specific patterns that may enhance our understanding of neurodevelopmental traits and their relation to psychosis-proneness.

The specific aims of the study are to: (1) examine associations among autistic traits, PLEs, and empathy in preadolescents; (2) investigate these associations separately by sex; (3) compare network structures between males and females to identify potential sex differences; and (4) estimate a directed acyclic graph (DAG) to evaluate probabilistic dependencies among these constructs47.

Methods

Participants

The data were drawn from the Adolescent Brain Cognitive Development (ABCD) Study “Curated Annual Release 4.0” (https://nda.nih.gov/abcd)52,53. The ABCD study is a longitudinal multisite investigation that collects clinical, behavioral, neuroimaging, and genetic data from children in the United States. It includes data from over 11,000 children representing a diverse range of demographics. Participants were primarily recruited through a school selection process designed to maximize sample representativeness and minimize selection biases53. In the current study, participants with any neurological diagnoses (such as cerebral palsy, epilepsy, and multiple sclerosis), traumatic brain injury, or missing data were excluded. To control sibling status, only one participant per family was randomly selected (n = 1,915 excluded), resulting in a final sample of 9,214 participants with an age range of 8.92 to 11.08 years. The demographic details of the participants are presented in Table 1. The flowchart of the sample and the measures used in the present study are presented in Fig. 1 and Table 2, respectively. Baseline assessments were used in the present study unless stated otherwise. Parents or guardians provided written informed consent after the study procedures were fully explained, and participants gave their assent before participating. All procedures for the ABCD study were approved by the central institutional review board at the University of California, San Diego (IRB# 160091), and by the institutional review boards of each of the ABCD study sites52,53. The de-identified data used in the present study are available through the National Institute of Mental Health repository (https://nda.nih.gov/) after obtaining approval to access the ABCD Study data.

Table 1 Demographic characteristics of the participants.
Full size table
Table 2 Measures used in the study.
Full size table
Fig. 1
figure 1

Flowchart depicting the determination of the final sample size. Primary measures were drawn from the ABCD study. Participants with missing data were first excluded, followed by the removal of participants not present across all three measures. Study-specific exclusion criteria were then applied. Finally, to control for sibling status, one participant per family was randomly selected.

Full size image

Measures

Psychotic-like experiences: The psychotic-like experiences (PLEs) of participants were assessed using the Prodromal Questionnaire-Brief Child Version (PQ-BC)54. This scale has demonstrated adequate construct validity and internal reliability in the ABCD sample55. Participants responded to 21 items with “yes” or “no.” For each endorsed item, children were asked, “Did it bother you?” For each “yes,” participants indicated how much it bothered them on a scale from 1 to 5, with 1 indicating least bother and 5 indicating maximum bother. Following previous research, a total PLEs score for each item was calculated by summing the endorsed symptoms, weighted by the level of distress (1 indicating no distress and 6 indicating maximum distress)55,56. The PLEs items were categorized into three factors based on previous research: thought delusions (items 1, 5, 8, 12, 14, and 18), unusual or grandiose delusions (items 4, 7, 15, and 16), and hallucinations (items 2, 3, 9, 10, 11, 17, 19, and 20)56. The remaining three items (6, 13, and 21) were not included in the analysis, as it is unclear to which factor they should be assigned56,57. The internal consistency of PQ-BC in the present sample was 0.82.

Autistic traits: Autistic traits were assessed using the abbreviated version (11 items) of the Social Responsiveness Scale, Second Edition58. The Short Social Responsiveness Scale (S-SRC) was administered in year 1, with parents responding to items on a four-point scale ranging from “Not true” (1) to “Almost always true” (4). Following previous research59, the two-factor structure of the S-SRC corresponding to the DSM-5 was utilized in the present research: social and communication interactions (items 1, 2, 4, 6, 7, and 8) and restricted and repetitive behaviors (items 3, 5, 9, 10, and 11). The total score for both subscales of the S-SRC was calculated by summing all items. For the present sample, the internal consistency of the S-SRC total scale was 0.85.

Empathy: The ABCD study did not include direct measurements of empathetic behaviors and emotions; therefore, questions indicative of empathy were extracted based on self-report of the participants. Previous research in the ABCD study has used a specific set of questions from parent reports as indicators of empathy in participants60. In the present study, this approach was adapted by selecting the items that had previously been validated as reflective of empathetic behaviors and emotions based on participant responses60. Three items were taken from the Youth Prosocial Behavior Survey (YPBS)61, which is part of the Strengths and Difficulties Questionnaire: (1) “I try to be nice to other people. I care about their feelings”, (2) “I am helpful if someone is hurt, upset, or feeling sick”, and (3) “I often offer to help others (parents, teachers, children)”. These items were rated on a scale from 0, indicating “Not true,” to 1, indicating “Somewhat true,” and 2, indicating “Certainly true.” The total score for the three items was calculated by summing these scores. For the present sample, the internal consistency of the YPBS total scale was 0.58. While this value falls below the commonly accepted threshold, it is important to note that Cronbach’s alpha is sensitive to the number of items in a scale. Given that the YPBS is a brief measure consisting of only three items, lower alpha values are not uncommon62.

Data analysis

The data analysis was conducted using the open statistical software R63. First, the mean, standard deviation, and normality (skewness and kurtosis) of social and communication interactions, restricted and repetitive behaviors, thought delusions, grandiose delusions, hallucinations, and empathy were assessed. All variables, except for empathy, did not meet the normality assumption according to the benchmarks of skewness > 2 and/or kurtosis > 764. These results are presented in Supplementary Information (Table 1). Consequently, the guidelines for conducting the psychological network analysis were followed65, and a nonparanormal transformation was applied to all variables using the R package huge66.

Gaussian graphical models

First, three Gaussian graphical models (GGM) were estimated for the entire sample, males and females separately, incorporating the two autistic traits subscales: social and communication interactions, and restricted and repetitive behavior, along with the three PLEs: thought delusions, grandiose delusions, and hallucinations, and empathy. In a GGM network, edges between the nodes indicate conditional relationships while controlling for all other nodes in the network67. The regularized partial correlations in the GGM network were estimated using the graphical Least Absolute Shrinkage and Selection Operator (graphical LASSO)68 in combination with the Extended Bayesian Information Criterion (EBIC)69. This procedure ensures that all remaining non-zero edges in the GGM network are robust and meaningful70. To estimate and visualize the GGM networks, qgraph and bootnet packages in R were used with default parameters67,70. In a GGM, each variable is defined as a node and each pairwise association between the variables as an edge. Red edges indicate negative and blue edges indicate positive associations in the networks, respectively. Thicker edges indicate higher associations between the nodes.

The one-step expected influence (EI) was computed to estimate the centrality of nodes in networks. For a specific node, EI is defined as the summed weight of edges shared with the remaining nodes of the network71. EI is preferable to other measures of centrality, as it considers the negative associations in the network. Higher EI indicates greater centrality and thus greater significance in the network72. In addition to EI, the predictability was also estimated for the three networks using the mgm package in R73. Predictability indicates the degree to which a given node can be explained by all other nodes in the network73. The significant differences in male and female networks were assessed using the Network Comparison Test (NCT). The NCT assesses the significant differences in three properties of networks, overall structure, strength of specific edges, and global strength74.

The stability and accuracy of the networks was estimated by bootnet package. Correlation stability coefficients (CS) assessed the stability of the networks and bootstrapping of edge weights and centrality was done to assess the accuracy. For edge weights and centrality, 2,500 bootstrap iterations were done at α level 0.05, and 95% confidence interval. Thereafter, a node-dropping subsetting bootstrap technique was used to determine the stability of centrality and calculate the CS coefficient of the networks. The CS coefficient value of more than 0.25 is considered acceptable and should be preferably higher than 0.570.

Bayesian network (directed acyclic graph)

Second, for the entire sample, a directed acyclic graph (DAG) was computed to indicate a potential causal structure based on the probabilistic dependencies of the nodes in a network75. The DAG was estimated by running the hill-climbing algorithm within the R package bnlearn76. The bootstrap function in bnlearn learns the structural elements of the network by removing, adding, and reversing the edges to optimize the goodness-of-fit target score, i.e., the Bayesian Information Criterion (BIC). This bootstrap function involves an iterative procedure of randomly restarting the process with various possible connections between node pairs, disturbing the network system, and applying 50 different random restarts to circumvent local maxima49. In line with previous research, 100 perturbations were performed for each restart77,78. As the iterative process of restarts and perturbations progresses, the function returns the best fitting network with the optimal BIC value49.

To ensure the stability of the resulting DAG, the guidelines for implementing DAGs in psychological research were followed, and 10,000 bootstrapped samples with replacement were computed to create a network for each sample, which were then averaged to produce a final network structure49. This involves a two-step approach. First, the frequency of each edge’s appearance in the 10,000 bootstrapped networks was ascertained. The optimal cutoff method of Scutari and Nagarajan79 for retaining edges was used, yielding both high sensitivity and specificity. Second, the direction of each surviving edge in the 10,000 bootstrapped networks was determined, an edge from node X to node Y was considered if it appeared in at least 51% of the bootstrapped networks80.

The averaged networks were visualized in two ways. First, the thickness of the edges represents relative BIC values. Thus, higher edge weights indicate greater importance, and removing a thick edge from the network would be more damaging to the model fit77,80. In the second visualization of the network, edge weights represent directional probabilities. Higher edge weights indicate a greater probability of direction. Thus, a thick edge from node X to node Y appeared in a larger proportion of the averaged 10,000 bootstrapped networks than a thin edge pointing from node Y to node Z.

Results

Overall sample: regularized partial correlation network

A zero-order product moment correlation plot of the entire sample is given in Supplementary Information (Fig. 1). The GGM of the entire sample is given in Fig. 2A. The regularized partial correlation matrix of the entire sample is given in Table 3. The autistic trait subscale and PLEs subscales were strongly associated with each other. Several pairwise connections between the nodes in the network are particularly notable. Empathy was negatively associated with the social and communication interactions (edge weight = − 0.037), and with the restricted and repetitive behaviors (edge weight = − 0.059) of autistic traits. Empathy further had a negative connection with the grandiose delusions (edge weight = − 0.012) of the PLEs subscale. There was no connection between the thought delusions and hallucinations of PLEs, and empathy. Positive associations were observed between the social and communication interactions and hallucinations (edge weight = 0.035), and hallucinations and restricted and repetitive behaviors (edge weight = 0.022).

Fig. 2
figure 2

Regularized partial correlation network and expected influence for the entire sample. 2A, regularized partial correlation network of the entire sample. 2B, expected influence centrality values corresponding to the network shown in panel 2A.

Full size image
Table 3 Regularized partial correlation matrix of the entire sample (upper triangle).
Full size table

The EI of the entire sample network is given in Fig. 2B. Hallucinations and thought delusions of PLEs had the highest EI, which was followed by the social and communication interactions of autistic traits in the network. Empathy was the least influential node in this network. The mean node predictability of the entire sample network was 0.26, meaning that, on average, 26% of the variance in each node was explained by the nodes directly connected to it. The highest predictability in this network was observed for hallucinations (0.36), followed by thought delusions (0.35), implying that these nodes are strongly influenced by other nodes in the network. The correlation stability of edge weights, CS [cor = 0.7] = 0.75 and EI CS [cor = 0.7] = 0.75, reflected that the network was highly stable. The bootstrapped edge stability and EI stability of the entire network are given in Supplementary Information (Figs. 2 and 3, respectively).

Male sample: regularized partial correlation network

The male regularized partial correlation network, and its expected influence (EI) are presented in Fig. 3A and B, respectively. In the male network, empathy was negatively associated with both social and communication interactions (edge weight = − 0.031) and restricted and repetitive behaviors (edge weight = − 0.039) of autistic traits. Empathy also had a negative connection with grandiose delusions (edge weight = − 0.010) related to psychotic-like experiences (PLEs). Social and communication interactions had two positive edges with thought delusions (edge weight = 0.026) and hallucinations (edge weight = 0.028) of PLEs. Restricted and repetitive behaviors were positively associated with hallucinations (edge weight = 0.021). Centrality analysis of the male network indicated that hallucinations had the highest expected influence (EI), while empathy was the least influential node in this network.

Fig. 3
figure 3

Regularized partial correlation network and expected influence for the male sample. 3A, regularized partial correlation network of the male sample. 3B, expected influence centrality values corresponding to the network shown in panel 3A.

Full size image

Female sample: regularized partial correlation network

The female regularized partial correlation network and its expected influence (EI) are presented in Fig. 4A and B, respectively. Empathy was negatively associated only with hallucinations (edge weight = − 0.021). Empathy also had negative associations with social and communication interactions (edge weight = − 0.33) and restricted and repetitive behaviors (edge weight = − 0.037). The centrality analysis indicated that thought delusions had the highest influence, while empathy was the least influential node.

Fig. 4
figure 4

Regularized partial correlation network and expected influence for the female sample. 4A, regularized partial correlation network of the female sample. 4B, expected influence centrality values corresponding to the network shown in panel 4A.

Full size image

Table 4 presents the regularized partial correlation matrices, with the upper triangle corresponding to the male sample and the lower triangle corresponding to the female sample. The mean predictability of the male network was 0.28, with the greatest predictability for the hallucinations node (0.38). In the female network, the mean node predictability was 0.23, with the greatest predictability for thought delusions (0.34). Both networks were highly stable, as indicated by the correlation stability of EI (CS [cor = 0.7] = 0.75) and edge weights (CS [cor = 0.7]) for the male and female networks.

Table 4 Regularized partial correlation matrices.
Full size table

Sex differences between male and female networks

To assess sex differences in network structure, the Network Comparison Test (NCT) was conducted after estimating separate networks for male and female subsamples. The NCT is a permutation-based statistical procedure that evaluates whether networks differ significantly in terms of: (1) overall network structure (i.e., configuration of edges), (2) global strength (i.e., the sum of absolute edge weights, reflecting overall connectivity), and (3) individual edge weights74.

The network structure invariance test revealed a significant difference between the male and female networks (M = 0.091, p = 0.002), indicating that the overall configuration of edges differed between the groups. Global strength also differed significantly between males (1.70) and females (1.53), with a test statistic of S = 0.169 (p = 0.016), suggesting that the male network was more strongly connected overall. Regarding individual edge differences, only one edge showed a statistically significant difference after Benjamini–Hochberg correction: the edge between social and communication interactions and restricted and repetitive behaviours (E = 0.091, p = 0.015). These results indicate that the overall network structure and global connectivity differ between males and females. Specifically, males show a more strongly connected network. The only individual connection that significantly differs is between social and communication interactions, and restricted and repetitive behaviors, which is stronger in males.

Directed acyclic graph

Figure 5 depicts the directed acyclic graph (DAG) generated by averaging 10,000 bootstrapped networks. In this DAG, the thickness of the edges indicates their significance to the overall network structure; thicker edges represent greater importance for model fit. The most important edges in the network connect thought delusions to hallucinations (BIC of -1741.45) of psychotic-like experiences (PLEs), social and communication interactions to restricted and repetitive behaviors (BIC of − 1668.57) associated with autistic traits, and hallucinations to grandiose delusions (BIC of − 274.30) of PLEs.

Fig. 5
figure 5

Directed Acyclic Graph (DAG) with arow thickness indicating the importance of each arrow to the overall network model fit.

Full size image

Figure 6 depicts the directional probabilities of edge weights in the 10,000 averaged bootstrapped networks. Greater values signify a greater likelihood that the edge points in the depicted direction. In this DAG, the thickest edges point from social and communication interactions to restricted and repetitive behaviors (directional probability of 0.58) of autistic traits, thought delusions to grandiose delusions (directional probability of 0.56), and hallucinations to restricted and repetitive behaviors (directional probability of 0.56) of PLEs. Table 5 presents the change in BIC values and directional probabilities of the two DAGs.

Fig. 6
figure 6

Directed Acyclic Graph with arrow thickness indicating directional probability.

Full size image
Table 5 BIC values and directional probabilities of arrows in the DAGs.
Full size table

Thought delusions emerged at the top of the model with two dependent nodes: hallucinations and grandiose delusions, meaning that thought delusions had no incoming edges (in-degree = 0) but had two outgoing edges (out-degree = 2). The occurrence of hallucinations and grandiose delusions is more likely to depend on the presence of thought delusions. Hallucinations had three direct descendants: grandiose delusions, social and communication interactions, and restricted and repetitive behaviors. Restricted and repetitive behaviours had three parent paths through social and communication interactions, hallucinations, and empathy. Lastly, empathy had one parent paths occurring through social and communication interactions, suggesting that empathy depends on the presence of these factors rather than vice versa.

Autistic traits and PLEs: regularized partial correlation network

Lastly, associations between autistic traits and PLEs in the entire sample were examined. The regularized partial correlation network is presented in Supplementary Information (Fig. 4), with EI of the network in Supplementary Information (Fig. 5), and the corresponding regularized partial correlation matrix provided in Supplementary Information (Table 2). Overall, increased autistic traits correlated with elevated distress across all three subscales of PLEs. Hallucinations were positively associated with both social and communication interactions (edge weight = 0.036), and restricted and repetitive behaviors (edge weight = 0.023). Among all nodes, hallucinations exhibited the highest expected influence, whereas grandiose delusions had the lowest.

Discussion

In this study, network analysis was used to examine the complex interrelations among autistic traits, PLEs, and empathy in a large, community-based sample of preadolescents. Three networks were estimated, one for the entire sample and two stratified by sex (male and female). The network for the full sample revealed several key patterns that shed light on both the distinct and overlapping features of these constructs.

Autistic traits and PLEs formed two separable clusters across all networks, suggesting conceptual and phenotypic distinctiveness despite their known overlaps10,11. Within the entire sample GGM, empathy was negatively associated with both social and communication interaction difficulties, and with restricted and repetitive behaviors, indicating that higher empathy scores were linked to lower overall levels of autistic traits. Additionally, empathy was negatively associated with grandiose delusions, suggesting that higher empathy scores correspond to lower distress from these experiences. This finding is consistent with evidence that deficits in emotional and cognitive empathy are characteristic of psychotic-spectrum conditions37,42. In contrast, no associations were found between empathy, and hallucinations and thought delusions, which may reflect the relatively abstract and self-referential nature of this PLEs subtype, involving cognitive distortions that are less directly tied to social-emotional processing81.

Beyond associations with empathy, the network revealed notable positive links between autistic traits and PLEs. Specifically, social and communication difficulties as well as restricted and repetitive behaviors, were positively connected with hallucinations, while restricted and repetitive behaviors were also linked with grandiose delusions. These associations suggest that higher levels of autistic traits may correspond to increased distress related to PLEs. A previous study using network analysis in individuals with psychotic disorders and familial risk for psychosis (siblings) also observed a positive association between autistic traits i.e., communication skills and paranoia in both networks82. Another study using meta-analysis found a positive association between schizotypal traits and autistic traits83. These patterns support the overlap theory, which posits that autistic traits and psychosis-prone experiences partially converge at phenotypic and mechanistic levels82,84. Importantly, hallucinations emerged as the most central node in the full-sample network. This is consistent with prior work identifying hallucinations as a core feature in the emergence and progression of psychotic symptoms18,85 and suggests their potential role as a bridge linking autistic and psychotic traits.

These findings were further substantiated by DAGs, which provided additional insights into the possible directionality among variables. In the DAGs, hallucinations received input from thought delusions and exerted directed influences on both grandiose delusions and autistic traits. In turn, social and communication interactions directly influenced empathy and restricted and repetitive behaviours. This pattern may suggest a potential indirect pathway through which PLEs, particularly hallucinations may influence empathy via autistic traits of social and communication interactions. Accordingly, empathy appears to function as a downward construct relative to both PLEs and social and communication interactions. The convergence between the GGM and DAG results enhances confidence in the observed patterns and highlights the added value of incorporating both undirected and directed models in network-based research49.

Taken together, these findings offer important insights. The observed associations between autistic traits and empathy may provide tentative support for the diametric model of autistic and psychosis, which proposes that these conditions represent opposing extremes in social cognitive functioning22. The negative edges linking empathy with both social and communication difficulties and restricted and repetitive behaviors are consistent with the hypo-mentalizing profile typically associated with autistic. Although autistic traits were positively associated with certain PLEs, particularly hallucinations and grandiose delusions, empathy remained inversely linked to autistic traits and demonstrated weak or inconsistent associations with PLEs. This dissociation suggests that while autistic and psychotic traits may share surface-level behavioral similarities, they may diverge in their underlying social cognitive architecture. This interpretation is reinforced by the DAG findings, which showed a directional pathway from social and communication interactions to empathy, rather than the reverse. This pattern may be consistent with theories suggesting that reduced theory of mind in autistic contributes to difficulty with empathy86, whereas psychotic symptoms such as hallucinations or delusions may reflect hyper-mentalizing or distorted social cognition that does not necessarily entail reduced empathic capacity28.

The observed co-occurrence of autistic traits and PLEs likely reflects shared neurodevelopmental vulnerabilities10,11. While these traits were positively associated in the network, they appear to represent partially overlapping but distinct constructs shaped by common genetic and environmental risk factors11,85. As such, their co-expression should not be interpreted as evidence for a unified underlying pathology.

The centrality of hallucinations across both GGM and DAG models underscores their potential role as a transdiagnostic marker. Hallucinations may disrupt both perceptual and social cognitive processes, thereby contributing to empathy and reinforcing social withdrawal. This could trigger a maladaptive feedback loop in which reduced empathy exacerbates both autistic traits and psychotic-like experiences, ultimately impairing social and functional outcomes9,87.

To explore potential sex differences, separate networks were estimated for males and females. The network invariance test revealed a significant difference in overall structure between male and female networks, suggesting that the pattern of associations among the constructs differed by sex. The male network also demonstrated significantly greater global strength, indicating that the connections among constructs in this network were more strongly interrelated. Among males, empathy was negatively associated with social and communication difficulties, restricted and repetitive behaviors, and grandiose delusions. Among females, empathy was negatively linked with social and communication difficulties, restricted and repetitive behaviors, and hallucinations. These findings align with prior research suggesting sex-based differences in the expression and neurobiological correlates of both autistic traits and psychotic symptoms88,89. The edge weight between social and communication interactions and restricted and repetitive behaviors was stronger in males compared to females. These sex-based findings reflect both global and local network connectivity. The stronger association between core autistic traits in males may reflect sex-specific mechanisms in the behavioural expression of autistic traits. These differences could be driven by neurodevelopmental, phenotypic, or assessment-related factors. These findings are critical for considering sex as an important factor in understating the relationships of autistic traits, PLEs, and empathy.

Implications and limitations

The present study offers several theoretical and clinical implications. Preadolescence is a critical developmental period characterized by significant psychological and biological, making it an important window for examining the interrelations among autistic traits, psychotic like experiences, and empathy46. To explore these complex relationships, network analysis was employed, providing nuanced insights into how these factors interact within a community-based sample of preadolescents. Autistic traits exist on a continuum and are observable not only in clinical populations but also within the general community. Similarly, subclinical PLEs are relatively common in adolescents from non-clinical samples17. Although PLEs are typically transient and benign, persistent or increasingly pronounced experiences can signal heightened vulnerability to subsequent mental health difficulties90. Thus, while not always clinically significant, these phenomena warrant attention due to their potential to foreshadow later psychological challenges. Consistent with prior literature42, empathy was negatively associated with both autistic traits and psychotic like experiences in this study. Additionally, a positive association between autistic traits and psychotic like experiences was observed, suggesting partial overlap in the underlying mechanisms contributing to these phenomena.

The study also yields several clinical insights. Network theory posits that nodes with broad influence on connected nodes, so called hub nodes, can amplify network activity, and that targeted intervention on these hubs may disrupt maladaptive cascades47. The findings indicate hallucinations and social communication difficulties as central nodes within the network, suggesting that interventions targeting these features may help reduce exacerbation of symptoms across related domains. Moreover, thought delusions, which function as influential drivers in DAGs, may also contribute significantly to network dynamics and thus represent additional targets for intervention. Furthermore, this study emphasizes the relevance of empathy in relation to both autistic traits and psychotic like experiences during preadolescence.

Several limitations warrant consideration. First, empathy was assessed using a self-report measure, which may be subject to bias and may not capture empathic behaviors expressed in diverse social contexts, particularly among peers. Second, empathy was treated as a unidimensional construct, precluding investigation of potentially important distinctions between emotional, cognitive, and other empathy subtypes. Third, the study focused exclusively on positive psychotic experiences, limiting the breadth of the psychosis spectrum evaluated. Fourth, the cross-sectional design restricts causal inference regarding the directionality of observed associations. Fifth, the age range of the participants was narrow. Sixth, the S-SRS was based on parent reports, and the YPBS had low internal consistency. These findings should be interpreted with caution as the reliability of YPBS may affect the robustness of the results. The current study does not interpret or endorse the view that autistic participants are inherently deficient in empathy. It is important to emphasize that the concept of empathy lacks clear theoretical definitions and precise measurement91. Milton92 further introduced the concept of the “double empathy problem”, which suggests that autistic and non-autistic individuals mutually misunderstand each other. Therefore, difficulties in empathy may exist on both sides. Lastly, neurodevelopmental and psychiatric disorders were not formally diagnosed, and the ABCD sample in the present study may not be representative in terms of socioeconomic status93. Therefore, these limitations may limit the generalizability of the findings.

Future research should incorporate more objective and multi contextual assessments of empathy and differentiate between empathy subtypes to provide a richer understanding of its role. Including both positive and negative psychotic experiences would offer a more comprehensive view of psychosis-related traits. Longitudinal designs are essential to elucidate causal pathways and developmental trajectories, as empathy and psychotic like experiences follow distinct patterns of maturation94. Examining these relationships across a broader developmental span, including early adulthood, would further clarify potential shifts in the interplay among empathy, autistic traits, and PLEs over time. Finally, given that the current sample comprised non-clinical participants, future studies should extend these investigations to clinical populations to validate and potentially generalize these findings in individuals with diagnosed autistic spectrum or psychotic disorders.