ABSTRACT
Clustering analyses, a form of unsupervised machine learning, are widely used in research studies aiming to explore the heterogeneity of the multidimensional autism constellation and describe meaningful subgroups of relatively similar autistic individuals. However, clustering is far from objective; selection of dependent measures, the choice of the number of clusters to be described, and interpretation of clusters all require subjective decisions. Furthermore, clustering within the boundaries of autism alone risks ignoring relevant insights from research outside these boundaries. Even more broadly, the purposes of subgrouping must be considered critically; although some subgrouping solutions might help individuals access supports, develop positive identities, and find supportive peer communities, other solutions could conversely restrict access to supports, impose unwanted or pathologising identities, and fracture communities. This chapter suggests that clustering, and diagnostic classification in general, may be most useful when labels are not arranged into a hierarchy of overarching entities and their “subtypes”, but instead when diagnostic labels are allowed to co-occur with one another.
Keywords: autism, heterogeneity, clustering, subgroups, autism constellation
