For most of the past century, psychological research has viewed inter-subject variability in behavior as mere statistical noise around a true value, focusing primarily on central tendencies like averages or medians of distributions (Molenaar, Reference Molenaar2004; Nesselroade, Reference Nesselroade2004; Rozin, Reference Rozin2001) – highlighting its problematic status, sometimes this variation is termed a nuisance. Traditional statistical models aimed to predict differences in the average population based on specific experimental manipulations (Molenaar, Reference Molenaar2004; Nesselroade, Reference Nesselroade2004; Rozin, Reference Rozin2001). While these experimental designs and methods have helped to develop important theories across psychological domains, they have significant drawbacks. As noted by Wundt much earlier, experimental manipulations typically explain only a small amount of variance in the data and, relatedly, often lack reproducibility, contributing to a replication crisis across the field (Hedge et al., Reference Hedge, Powell and Sumner2018; Kravitz & Mitroff, Reference Kravitz and Mitroff2023; Siritzky et al., Reference Siritzky, Cox, Nadler, Grady, Kravitz and Mitroff2023). This apparent crisis has led to calls for better characterization of behavior through larger sample sizes, robust measures guided by theoretical assumptions formulated as a priori hypotheses, and increased transparency in the design and conduct of studies. Open science initiatives have contributed to making these changes by supporting data sharing and study preregistration (Collaboration, Reference Collaboration2012, Reference Collaboration2015; Kravitz & Mitroff, Reference Kravitz and Mitroff2023; Siritzky et al., Reference Siritzky, Cox, Nadler, Grady, Kravitz and Mitroff2023).
Individual differences research, on the other hand, relies on large sample sizes and data-driven analysis of variability. Its broad, unconstrained study design enhances our ability to capture behavioral diversity and can produce more replicable findings compared to experimental studies with constrained homogeneous samples. Along with advancements in data science, particularly the big data approach (Adjerid & Kelley, Reference Adjerid and Kelley2018; Gomez-Marin et al., Reference Gomez-Marin, Paton, Kampff, Costa and Mainen2014; Harlow & Oswald, Reference Harlow and Oswald2016), there have been calls for better characterization of interindividual differences across cognitive domains (Adjerid & Kelley, Reference Adjerid and Kelley2018; Eisenberg et al., Reference Eisenberg, Bissett, Zeynep Enkavi, Li, MacKinnon, Marsch and Poldrack2019; Gomez-Marin et al., Reference Gomez-Marin, Paton, Kampff, Costa and Mainen2014; Harlow & Oswald, Reference Harlow and Oswald2016). Researchers have long suspected that deviations from the mean are not random noise but systematic variation. However, small sample sizes in most in-person psychological research have hindered the systematic investigation of these interindividual differences. Over the past couple of decades, individual differences have become a growing interest in psychological research. Personality psychology has spearheaded this movement, developing models of personality types (Ilmini & Fernando, Reference Ilmini and Fernando2017; Patzelt et al., Reference Patzelt, Hartley and Gershman2018; Phan & Rauthmann, Reference Phan and Rauthmann2021), creativity (Lloyd-Cox et al., Reference Lloyd-Cox, Pickering, Beaty and Bhattacharya2023; Mejia et al., Reference Mejia, D’Ippolito and Kajikawa2021; Minai et al., Reference Minai, Doboli and Iyer2021; Saunders & Bown, Reference Saunders and Bown2015), and cognitive styles (Riding & Rayner, Reference Riding and Rayner2013; Zhang, Reference Zhang2002) that could explain differences in perception (Haehner et al., Reference Haehner, Rakhshani, Fassbender, Lucas, Donnellan and Luhmann2023; Zhu et al., Reference Zhu, Li, Zhao and Jiang2018), attention (Subramanian et al., Reference Subramanian, Yan, Staiano, Lanz and Sebe2013), and decision-making (Frolichs et al., Reference Frolichs, Rosenblau and Korn2022; King-Casas et al., Reference King-Casas, Sharp, Lomax-Bream, Lohrenz, Fonagy and Montague2008; Subramanian et al., Reference Subramanian, Yan, Staiano, Lanz and Sebe2013). While this field has provided evidence for interindividual variation in nonclinical groups, much less is known about the role of interindividual variability in neuropsychiatric disorders.
Across neuropsychiatric disorders, there is larger variability or greater interindividual differences in self-reported or observed behaviors such as personality traits and social skills (Jauk & Kanske, Reference Jauk and Kanske2019) compared to control groups (Lochner & Stein, Reference Lochner and Stein2003; Steinhausen, Reference Steinhausen2009; Wolfers et al., Reference Wolfers, Floris, Dinga, van Rooij, Isakoglou, Kia, Zabihi, Llera, Chowdanayaka, Kumar, Peng, Laidi, Batalle, Dimitrova, Charman, Loth, Lai, Jones, Baumeister, Moessnang and Beckmann2019). These differences, combined (almost inevitably) with smaller sample sizes in clinical studies, present significant hurdles for improving diagnosis and treatment decisions (Beauchaine & Cicchetti, Reference Beauchaine and Cicchetti2016; Jacob et al., Reference Jacob, Wolff, Steinbach, Doyle, Kumar and Elison2019; Wolff et al., Reference Wolff, Jacob and Elison2018). This issue is particularly pronounced in autism spectrum disorder (ASD), where formalizing heterogeneity is seen as critical for linking genetic profiles to neural and behavioral phenotypes (Pelphrey et al., Reference Pelphrey, Shultz, Hudac and Vander Wyk2011; Wolfers et al., Reference Wolfers, Floris, Dinga, van Rooij, Isakoglou, Kia, Zabihi, Llera, Chowdanayaka, Kumar, Peng, Laidi, Batalle, Dimitrova, Charman, Loth, Lai, Jones, Baumeister, Moessnang and Beckmann2019). Autism is marked by significant challenges in social interaction and communication, linked to differences in social perceptual and cognitive abilities across the diagnostic spectrum. The complex and diverse neurodevelopmental characteristics of autism are further compounded by interindividual variability observed in many other neuropsychiatric disorders (Jacob et al., Reference Jacob, Wolff, Steinbach, Doyle, Kumar and Elison2019). This has led to a shift from viewing autism as unidimensional to one seeing it as an overall term including multiple syndromes, each resulting from different etiological pathways (Amaral et al., Reference Amaral, Schumann and Nordahl2008; Geschwind & Levitt, Reference Geschwind and Levitt2007; Insel et al., Reference Insel, Cuthbert, Garvey, Heinssen, Pine, Quinn, Sanislow and Wang2010).
Addressing the heterogeneity in autism is imperative for two reasons. First, it can enhance diagnostic accuracy, which currently relies heavily on clinical observations only (Geschwind & Levitt, Reference Geschwind and Levitt2007; Insel et al., Reference Insel, Cuthbert, Garvey, Heinssen, Pine, Quinn, Sanislow and Wang2010; Wolff et al., Reference Wolff, Jacob and Elison2018). Second, it can facilitate the development of more targeted cognitive-behavioral interventions for core autism symptoms, which remain elusive and rarely based on individual symptoms. However, in order to characterize and formalize the heterogeneity within ASD, several significant obstacles must be overcome. These include (1) increasing sample sizes of autistic individuals from a wider range of cultural and socioeconomic backgrounds, (2) building bridges between autism symptomatology and individual differences in autistic traits in the non-autistic population, and (3) introducing finer-grained dynamic methods to capture behavioral dynamics at the individual level. Here, we briefly describe emerging literature on the first two obstacles, before providing our perspective on how to overcome the third and more challenging one of developing finer-grained assessments and cognitive models of autism symptoms.
With respect to increasing sample sizes of autism research, there have been several initiatives spearheaded by the National Institute of Mental Health to support interdisciplinary research aimed at advancing the understanding of ASD and developing new interventions. The Autism Cluster of Excellence (ACE) initiative brings multiple research sites together jointly to investigate the underlying causes of ASD, including genetic, environmental, and developmental factors. ACE centers foster collaboration among researchers from diverse disciplines, including neuroscience, genetics, psychology, and education. The commitment to data sharing in the ACE program has led to large databases of behavioral and neural data such as the National Database for Autism Research and the Autism Brain Imaging Data Exchange. These data-sharing efforts encompass various types of data, including genetics, genomics, brain imaging, and behavioral assessments. This rich and diverse dataset is invaluable for testing the replicability of findings from smaller, single-site studies, enhancing the robustness and generalizability of research findings in the field of autism (Heinsfeld et al., Reference Heinsfeld, Franco, Craddock, Buchweitz and Meneguzzi2018; Nielsen et al., Reference Nielsen, Zielinski, Fletcher, Alexander, Lange, Bigler, Lainhart and Anderson2013; Payakachat et al., Reference Payakachat, Tilford and Ungar2016).
To address the second point, building bridges between autism symptomatology in individuals diagnosed with autism and autistic traits in the non-autistic population, there is a need to investigate autistic traits in larger and more diverse samples of individuals with an autism diagnosis alongside individuals without a diagnosis. Measuring self-reported autistic traits alongside objective measures of autistic symptom domains across diagnosed and undiagnosed individuals would shed light on the convergence of autism symptomatology and autistic traits in the general population. Initiatives like the Simons Foundation Autism Research Initiative provide an important resource for recruiting larger and heterogeneous groups of individuals with an autism diagnosis for behavioral studies. This initiative provides a valuable resource for researchers aiming to investigate ASD across different stages of life, from infancy through adulthood. The Simons Foundation Powering Autism Research for Knowledge (SPARK) initiative has aggregated a cohort of over 50,000 individuals with ASD and their families, who are interested in contributing to research. Researchers can access this large and diverse pool of participants nationwide for online behavioral studies. They can additionally apply to utilize existing phenotypic and genetic data on their recruited participants. SPARK has become increasingly popular for recruiting participants across developmental stages and levels of functioning, allowing studies to investigate variability in autism symptoms, including the co-occurrence of certain autism risk genes in phenotypes (Gaugler et al., Reference Gaugler, Klei, Sanders, Bodea, Goldberg, Lee, Mahajan, Manaa, Pawitan, Reichert, Ripke, Sandin, Sklar, Svantesson, Reichenberg, Hultman, Devlin, Roeder and Buxbaum2014; Grove et al., Reference Grove, Ripke, Als, Mattheisen, Walters, Won, Pallesen, Agerbo, Andreassen, Anney, Awashti, Belliveau, Bettella, Buxbaum, Bybjerg-Grauholm, Bækvad-Hansen, Cerrato, Chambert, Christensen, Churchhouse and Børglum2019; Matoba et al., Reference Matoba, Liang, Sun, Aygün, McAfee, Davis, Raffield, Qian, Piven, Li, Kosuri, Won and Stein2020; Myers et al., Reference Myers, Challman, Bernier, Bourgeron, Chung, Constantino, Eichler, Jacquemont, Miller, Mitchell, Zoghbi, Martin and Ledbetter2020; Wilfert et al., Reference Wilfert, Turner, Murali, Hsieh, Sulovari, Wang, Coe, Guo, Hoekzema, Bakken, Winterkorn, Evani, Byrska-Bishop, Earl, Bernier, SPARK, Zody and Eichler2021), as well as a more thorough investigation of the influence of sex and gender differences in autism (Dillon et al., Reference Dillon, Kanne, Landa, Annett, Bernier, Bradley, Carpenter, Kim, Parish-Morris, Schultz and Wodka2023; Fombonne et al., Reference Fombonne, Green Snyder, Daniels, Feliciano and Chung2020; Saré & Smith, Reference Saré and Smith2020). Large online studies leveraging such databases are beginning to shed light on the similarities and differences in autistic traits between people with autism or their family members and those without a close relative diagnosed with autism (Bora et al., Reference Bora, Aydın, Saraç, Kadak and Köse2017; Ruzich et al., Reference Ruzich, Allison, Smith, Watson, Auyeung, Ring and Baron-Cohen2016).
The third and most challenging obstacle is developing finer-grained assessments and cognitive models of autism symptoms that can help bridge the gap to animal models and genetics, on the one hand, and to real-world behavioral outcomes, on the other hand. This need is not unique to studying autism but is much needed across neuropsychiatric domains. Indeed, researchers have acknowledged the need for a computational psychiatry approach to precision phenotyping, increasing the precision with which we characterize certain sub-phenotypes of the autism spectrum (Tiego et al., Reference Tiego, Martin, DeYoung, Hagan, Cooper, Pasion, Satchell, Shackman, Bellgrove and Fornito2023). A more nuanced approach to characterizing and quantifying the observed behavioral differences and their biological correlates can deepen our understanding of neuropsychiatric disorders and their symptoms (Friston et al., Reference Friston, Stephan, Montague and Dolan2014; Hitchcock et al., Reference Hitchcock, Fried and Frank2022; Huys et al., Reference Huys, Maia and Frank2016; Montague et al., Reference Montague, Dolan, Friston and Dayan2012; Wang & Krystal, Reference Wang and Krystal2014). In this respect, computational psychiatry describes mathematical approaches to quantitatively analyze the complex interactions across biobehavioral system levels within and between neuropsychiatric disorders (Frässle et al., Reference Frässle, Yao, Schöbi, Aponte, Heinzle and Stephan2018; Karvelis et al., Reference Karvelis, Paulus and Diaconescu2023; Petzschner et al., Reference Petzschner, Weber, Gard and Stephan2017; Stephan & Mathys, Reference Stephan and Mathys2014; Wiecki et al., Reference Wiecki, Poland and Frank2015). The hope of computational psychiatry is to identify nuanced patterns of behavior as well as their underlying cognitive mechanisms and neural implementation. This latter goal can be achieved through cognitive computational modeling.
Computational modeling can reveal clinically meaningful individual differences
Cognitive computational models have been widely used in the field of cognitive neuroscience (Castelfranchi & Falcone, Reference Castelfranchi and Falcone2010; Farrell & Lewandowsky, Reference Farrell and Lewandowsky2018; Kriegeskorte & Douglas, Reference Kriegeskorte and Douglas2018; Lewandowsky & Farrell, Reference Lewandowsky and Farrell2010; Pitt et al., Reference Pitt, Myung and Zhang2002; Sun, Reference Sun and Sun2008). These models can be conceptualized as formal mathematical translations of theoretical assumptions. These mathematical models can reveal internal, unobservable states that govern behavioral output (Baker et al., Reference Baker, Saxe and Tenenbaum2009; Baker & Tenenbaum, Reference Baker and Tenenbaum2014; Gluck et al., Reference Gluck, Stanley, Moore, Reitter and Halbrügge2010; Just et al., Reference Just, Carpenter and Varma1999; Wolpert et al., Reference Wolpert, Doya and Kawato2003). By translating cognitive processes into mathematical terms and testing them on an individual level, cognitive computational models have revealed individual differences in a wide array of decision-making contexts. Previous studies have differentiated decision-making of more risk averse and risk seeking individuals (Daw et al., Reference Daw, Gershman, Seymour, Dayan and Dolan2011; Jacob et al., Reference Jacob, Wolff, Steinbach, Doyle, Kumar and Elison2019; Levy, Reference Levy2017; Pushkarskaya et al., Reference Pushkarskaya, Tolin, Ruderman, Henick, Kelly, Pittenger and Levy2017, Reference Pushkarskaya, Tolin, Henick, Levy and Pittenger2018), impulsive versus more deliberate individuals who plan ahead (Blankenstein et al., Reference Blankenstein, Peper, Crone and van Duijvenvoorde2017; Kable & Glimcher, Reference Kable and Glimcher2007, Reference Kable and Glimcher2010; Kurzban et al., Reference Kurzban, Duckworth, Kable and Myers2013), or those that learn from environmental feedback itself or through imitation of others, versus individuals that are more likely to engage in metacognition and represent the task structure (Charpentier et al., Reference Charpentier, De Neve, Li, Roiser and Sharot2016, Reference Charpentier, Aylward, Roiser and Robinson2017, Reference Charpentier, Iigaya and O’Doherty2020; Feher da Silva et al., Reference Feher da Silva, Lombardi, Edelson and Hare2023; Ramsey et al., Reference Ramsey, Kaplan and Cross2021; Vélez & Gweon, Reference Vélez and Gweon2021). Cognitive computational models can enhance our understanding of behavioral dynamics over time, enabling a more nuanced characterization of individual differences and providing a fundamentally dynamic perspective on cognitive variability within individuals (Schurr et al., Reference Schurr, Reznik, Hillman, Bhui and Gershman2024).
Computational modeling can inform the links between autistic traits and social functioning
With respect to autism specifically, recent research has highlighted the utility of computational modeling in studying autistic traits within the general population. This is especially important because autistic traits are known to exist on a continuum within the general population (Robinson et al., Reference Robinson, Koenen, McCormick, Munir, Hallett, Happé, Plomin and Ronald2011). Some studies suggest that clinically relevant autistic traits are an extension (or end point) of that continuum (Constantino & Todd, Reference Constantino and Todd2003; Ronald & Hoekstra, Reference Ronald and Hoekstra2011; Skuse et al., Reference Skuse, Mandy and Scourfield2005), while other studies suggest a discontinuity between autistic traits in the general population and those of individuals with an autism diagnosis (Abu-Akel et al., Reference Abu-Akel, Allison, Baron-Cohen and Heinke2019; Frazier et al., Reference Frazier, Youngstrom, Sinclair, Kubu, Law, Rezai, Constantino and Eng2009; Peralta & Cuesta, Reference Peralta and Cuesta2007). Irrespective of these two opposing positions, exploring autistic traits in the general population provides several advantages and can critically inform research on ASD. This claim is warranted by the observation that nonclinical groups with high autistic traits exhibit a higher degree of social functioning (De Groot & Van Strien, Reference De Groot and Van Strien2017) than individuals with ASD. It is, therefore, possible to examine the cognitive profiles or behavioral strategies that contribute to greater social functioning despite high autistic traits. Exploring autistic traits in neurotypical individuals also makes it possible to avoid the confounding effects of comorbid conditions that co-occur with ASD. Furthermore, it enables researchers to study larger cohorts (De Groot & Van Strien, Reference De Groot and Van Strien2017), making it possible to control confounding effects or specifically examine their interaction with autistic traits as a question of interest. The Broad Autism Phenotype questionnaire (Hurley et al., Reference Hurley, Losh, Parlier, Reznick and Piven2007), for instance, describes behavioral and cognitive tendencies that are less severe but rather stable characteristics, similar in nature to those found in individuals with an ASD. Importantly, autistic traits seem to scale with the genetic risk for ASD. Relatives of individuals with ASD without a diagnosis have been shown to exhibit more autistic traits than those without a close relative with an ASD diagnosis (Piven et al., Reference Piven, Palmer, Jacobi, Childress and Arndt1997).
Recent studies that have leveraged computational modeling better to define cognitive mechanisms that co-occur with high autistic traits have produced promising avenues for autism research. For example, one study used a hierarchical Bayesian modeling framework to examine the integration of nonsocial and social cues through a reward-based learning task. They found that more pronounced autistic traits in a group of healthy control subjects were related to less integration of social cues in decision-making. Computational modeling further demonstrated that performance differences between individuals with low versus high autistic traits were not due to an inability to process the social stimuli (gaze direction) and their causes, but rather to the extent to which participants relied on social information to infer the nonsocial cue (Sevgi et al., Reference Sevgi, Diaconescu, Henco, Tittgemeyer and Schilbach2020). Another recent large-scale study found that high autistic traits were associated with reduced goal emulation during observational learning. This means that participants with higher autistic traits were more prone to imitate the observer but showed a reduced tendency to represent the overall goal of the observed person, which corresponded to the reward structure of the task (Wu et al., Reference Wu, Oh, Tadayonnejad, Feusner, Cockburn, O’Doherty and Charpentier2024). This emerging literature on social decision-making and autistic traits suggests that high autistic traits do not amount to a general inability to process social cues (Sevgi et al., Reference Sevgi, Diaconescu, Henco, Tittgemeyer and Schilbach2020). They do, however, have a nuanced influence on the extent to which different types of social information are used, which may result in less adaptive outcomes.
Computational modeling can help to specify differences in the cognitive mechanisms underlying behavioral phenotypes
Cognitive computational models have also revealed differences in the cognitive mechanisms of individuals with an ASD diagnosis. In line with previous theories (Baron-Cohen et al., Reference Baron-Cohen, Leslie and Frith1985), some studies have shown that individuals with ASD have difficulties in representing their partners’ intentions. Yoshida and colleagues (Yoshida et al., Reference Yoshida, Dziobek, Kliemann, Heekeren, Friston and Dolan2010), for instance, employed a stag-hunt game to characterize unobservable computational processes implicit in social interactions and to measure whether individuals prefer smaller individual versus larger joint rewards. They found that the decisions of autistic individuals were less guided by inferring their partners’ beliefs and instead were guided by a fixed strategy compared to non-autistic control participants. Autistic individuals who showed less mental state inference had greater symptom load. Similarly, a study on social learning showed that autistic adolescents relied less on social knowledge and feedback to learn about peers’ preferences (Rosenblau et al., Reference Rosenblau, Korn, Dutton, Lee and Pelphrey2021). However, this study specifically tested whether autistic teens could learn about the preferences of non-autistic peers. Given that autistic and non-autistic adolescents may have differing preferences, these disparities could contribute to the “double empathy problem” – a phenomenon where misunderstandings between autistic and non-autistic individuals arise from mutual challenges in understanding each other’s perspectives (Milton, Reference Milton2012).
Other studies using economic exchange tasks have shown more nuanced differences in autistic individuals, which result in more adaptive behavior. For instance, autistic individuals have been shown to assess their partner’s cooperation history more accurately and reciprocate less when partners are untrustworthy (Maurer et al., Reference Maurer, Chambon, Bourgeois-Gironde, Leboyer and Zalla2018). In accordance with this finding, a study on information sampling for cooperation showed that autistic adolescents had lower overall expectations (i.e., priors) about their partner’s reciprocation tendencies (Liu et al., Reference Liu, Shah, Ma and Rosenblau2024). Given the range of trustworthy and untrustworthy agents, the autistic priors more accurately reflected the overall trustworthiness distribution of their potential partners. Moreover, autistic adolescents shared less often with untrustworthy agents than the non-autistic sample, which suggests that they are less prosocial than non-autistic adolescents. Moreover, this type of strategic interaction was less related to social skills in the autism group. Participants’ task behaviors were less strongly associated with social skills as measured by the social responsiveness scale (SRS) compared to the non-autistic group, in which the SRS was the strongest predictor of cooperation tendencies. These findings suggest differences between strategic decision-making in economic exchange and non-economic social interactions in autistic individuals. In non-autistic groups, individuals’ strategic choices may be more reflective of their mentalizing abilities and prosocial tendencies.
In conclusion, computational approaches offer a promising way to identify nuanced behavioral patterns and their underlying cognitive mechanisms. Studies leveraging both advances in building big data platforms and computational modeling hold promise for better characterizing autism phenotypes. This could pave the way for defining autism subtypes more clearly and examining their etiology, developmental trajectories, and comorbidities. Moreover, these advances can predict responses to therapeutic interventions, leading to more personalized and effective treatment plans (Collin et al., Reference Collin, Gebhardt, Golebiewski, Karaderi, Hillemanns, Khan, Salehzadeh-Yazdi, Kirschner, Krobitsch, Consortium and Kuepfer2022; Johnson et al., Reference Johnson, Wei, Weeraratne, Frisse, Misulis, Rhee, Zhao and Snowdon2021). Finally, we can combine insights from larger studies on autistic traits in the general population and those that investigate variability in symptoms in individuals diagnosed with ASD to inform commonalities and differences between populations with high autistic traits and no ASD diagnosis and those with a diagnosis. It could also help to systematically investigate the roles of sex, gender, and gender diversity in populations with high autistic traits and ASD diagnoses, given emerging evidence of the importance of examining sex differences in autism (Bölte et al., Reference Bölte, Neufeld, Marschik, Williams, Gallagher and Lai2023; Loomes et al., Reference Loomes, Hull and Mandy2017). This special issue focuses on how an individual differences approach, rather than the focus on central tendencies, can deepen our understanding of autistic traits and symptoms. It emphasizes the importance of investigating variability in autistic traits and phenotypes to refine autism classification. The issue highlights cutting-edge methods suited to capturing heterogeneity in behavioral and brain function and showcases big data approaches for analyzing large samples as well as computational modeling approaches that can expose how differences in cognitive mechanisms underlie meaningful behavioral variability.
Open access
