Notable heterogeneity in the configuration, severity, trajectory, and treatment response of both core diagnostic and co-occurring psychiatric and behavioural symptoms in individuals with Autism Spectrum Disorder (ASD) is well established (1, 2). Lack of insight into the sources contributing to this heterogeneity has hampered progress towards understanding etiological mechanisms, developing effective treatments and predicting outcomes (3, 4). We (3) and others (5) have suggested that a more fine-grained understanding of atypical sensory features which encompass hyper-reactivity, hypo-reactivity, and unusual sensory interests may offer a promising approach to parse heterogeneity in ASD. Indeed, a range of studies have linked atypical sensory features with restricted and repetitive behaviours (6, 7), socio-communicative impairments (8, 9), anxiety (10, 11), behavioural and sleep problems (12, 13), and adaptive functioning (14). Further, several studies have identified the existence of potentially informative sensory-based subgroups among individuals with ASD (11, 15–18). However, statistical and methodological limitations have precluded the field from fully capitalising on the potential utility of sensory features for explaining the heterogeneity of ASD (3).
Within the broader field of psychopathology, different methodological approaches have been put forward to investigate phenotypic heterogeneity (19). Taxometric methods aim to address the question whether individual differences should be conceived as continuous traits (i.e. participants differ in degree of an observed behaviour) or in terms of typologies (i.e. participants belong to either of two qualitatively different types or latent subgroups; 20). While the former has been addressed using various methods of factor analysis (FA), the latter has been tested using cluster analysis and Latent Class Analysis (LCA). In FA, latent factors capture the common content among test items and variability between participants is assumed to arise because of inter-individual differences on these factor variables. The latent factor(s) in FA can be thus construed as a dimensional quantity upon which individuals differ in degree. Conversely, cluster analysis or LCA adopts a categorical view to explain heterogeneity: categorical latent classes or subgroups are assumed to capture variability between participants and individuals are classified based on their similarities in response pattern on a set of item variables. The term subgroup will be used throughout to refer to latent classes or clusters. In cluster analysis or LCA, variation between individuals is therefore assumed to relate to a difference in kind, and derived subgroups may differ qualitatively (i.e. subgroups present with a qualitatively different profile) or quantitatively (i.e. a high- or low-scoring subgroup).
To date, these two methodological approaches have been separately applied to investigate variability in sensory features in ASD. A limited but growing number of studies have used FA methods to delineate the underlying structure of sensory features as measured by commonly used parent-report questionnaire measures, including the Short Sensory Profile (SSP; 21, 22), Sensory Behavior Questionnaire (SBQ; 23), and Sensory Experiences Questionnaire Version 3.0 (SEQ-3.0; 24). Of these measures, the SSP (25) is one of the most widely used parent-report/caregiver questionnaire measure of sensory features in ASD (26) and has been used in large multi-centre collaborative projects such as the Autism Speaks Autism Treatment Network (27) and the EU-AIMS Longitudinal European Autism Project (28). The few existing studies that have applied FA to investigate sensory features in individuals with ASD as measured by the SSP have however produced inconclusive results, suggesting either a six- (21) or nine-factor structure (22) that only partially resembled the originally proposed seven factor structure (25). Reasons for these inconsistencies may relate to differences in sample size and age compositions, as well as the use of different FA techniques of varying specifications. In addition, some of the newly hypothesised constructs featured too few items to be psychometrically or clinically useful (22). This suggests that it is currently not clear what the exact structure/taxonomy of sensory features in ASD is, which also limits previous studies that have utilised the SSP in subgrouping approaches.
In parallel to the above work, several studies have attempted to characterise heterogeneity in sensory features by identifying more homogeneous groups of individuals via different types of cluster analyses and Latent Class Analyses (LCA) approaches (for a review see 29). To date, cluster analytic studies have proposed anywhere from two to five subgroups using a range of different measures including the SSP (11, 17, 30), Sensory Experiences Questionnaire (SEQ; 15), Adolescent/Adult Sensory Profile (AASP; 31), Sensory Profile (SP; 32), Infant Toddler Sensory Profile (ITSP; 33) and Sensory Profile 2 (SP-2; 34, 35). These instruments differ widely in the type of informant (i.e. self- vs. proxy-based), intended target population use (i.e. infants, children or adolescents/adults), sensory domains assessed, and their psychometric properties (26, 36). In addition, most studies have been limited by sample sizes and did not consider multiple developmental and clinical variables, leaving unanswered questions about the clinical correlates of sensory subgroups. Thus, it is not surprising that existing studies lack a clear consensus on the number of purported sensory subgroups in ASD, their frequency and profile, as well as associated clinical and demographic correlates. Despite some of these differences, two sensory subgroups were consistently identified: those with predominantly mild sensory features (i.e. referred to as ‘sensory adaptive’ or ‘perceptive-adaptable’) and those with marked impairments across all or most of sensory domains (i.e. termed ‘Sensory severe’, ‘Generalized sensory difference’, or ‘Sensorimotor’). Relevant to the current study, research using the SSP has identified either three sensory subgroups that differ in their severity of anxiety symptoms, but not on age, expressive language or social-communicative symptoms associated with ASD (11), or four sensory subgroups. The former was found in one study to differentiate in terms of age and level of adaptive behaviour (30), and in another study in age and non-verbal IQ, but not gender or ASD symptoms (17). At least some of these inconsistencies are likely to be related to the varied choice of sensory measures employed across studies, as well as the different age and size of the sample studied (29).
While both FA and LCA approaches have been useful to further characterise sensory features in ASD, these taxometric procedures presuppose that sensory atypicalities either fall exclusively along a continuum from mild to severe or that individuals can be categorised into a finite number of discrete homogeneous entities or subgroups. Thus, the major limitation of FA is that it does not allow to classify individuals into groups, which is critical both in terms of informing clinical decision making, but also for advancing neurobiological and genomic research and precision medicine approaches in ASD (37). The major limitation of LCA and the categorical approach more broadly is that subgroups do not consider the range in severity and impairment within and across classes. Factor Mixture Modelling (FMM; 38) is a flexible hybrid model that combines LCA and FA approaches by simultaneously modelling the underlying structure to be both categorical and dimensional. The structure is considered categorical since the model allows for stratification of individuals into discrete subgroups while allowing for heterogeneity in the severity of the underlying trait within these groups through the use of continuous latent variables. This approach is particularly useful since it does not have the limitations of the two conventional taxometric procedures and it allows to directly compare different models of symptom structures. Indeed, FMM has been successfully applied to assess core diagnostic symptom structures in Attention-Deficit/Hyperactivity Disorder (ADHD; 39, 40) and ASD (41–44). However, previous efforts in ASD focussed either on the two symptom dimensions of social communication and interaction and restricted and repetitive behaviours (RRB; 41, 42, 43, 45) or on empathy and systemising (44).
Despite the utility of this approach, FMM has not been used to characterise sensory features in ASD. Therefore, our study sought, for the first time, to apply FMM to compare dimensional, categorical, and dimensional-categorical hybrid structures of sensory features in a large and well-characterised sample of individuals with ASD. Our aim was to clarify whether the structure of sensory features in ASD can be best conceptualised either by (1) a continuum on which individuals differ in severity, (2) sensory subgroups that display either quantitative or qualitative differences in their sensory profiles, or (3) a multidimensional factor model composed of sensory subgroups that differ in both their severity within and across groups along continuous factor scores. In addition, we aimed to further characterise identified groups in terms of potential differences in age, gender distribution and IQ, as well as how they relate to individual differences in social communication and RRB symptoms, co-occurring symptoms of anxiety and ADHD, and adaptive functioning. With few exceptions, previous subgrouping studies have not simultaneously examined potential associated clinical variables. It remains therefore unclear how sensory subgroups also differ in other aspects of core ASD and co-occuring symptoms, as well as adaptive functioning.