Analyses of autism groups have inherent flaws
Among the uncountable findings reported in the vast neuroimaging literature of autism, there is no single one that would qualify as ‘unique’ in the sense that it would pinpoint a brain feature found in every single individual with autism but in no individual without the condition. In addition, at the neurobiological level, there is no ‘autism’ in the singular.
One may debate whether the ‘autism spectrum disorder’ of the latest edition of the Diagnostic and Statistical Manual of Mental Disorders, defined by behavioral (not neurobiological) criteria, is appropriately referred to in the singular. However, when it comes to neurodevelopmental causes and pathways, the evidence overwhelmingly indicates plurality. There are many different combinations of causes and many different pathways of neurodevelopmental differences that may result in a child receiving a diagnosis of autism. Once this is understood, the expectation of a unique brain difference is entirely unrealistic.
This conclusion may sound pessimistic. Does it imply that decades of neuroscience and neuroimaging research have been futile? Not at all. However, it does imply that a whole generation of neuroimagers (including myself), while acquiring important data, may have looked at them the wrong way.
The question has almost always been: What is the difference between an autism group and a comparison group (usually typically developing)? But this question makes little sense when we know that any autism sample is probably composed of individuals with divergent neurodevelopmental histories. With small samples, this will lead to skewed findings (reflecting the sample composition) that will not be replicated in a different sample.
Large samples that have become available in recent years are not by themselves a solution. A difference at the group level (say 1,000 people with autism and 1,000 typically developing participants) may be hugely ‘significant’ given ample statistical power, but it may tell us little about critical brain differences in any single individual in the large autism cohort. This is because in group-level tests, differences found across many individuals in a group will reach ‘significance,’ even if they are minimal in magnitude.
On the other hand, a brain feature differing dramatically from neurotypical, but in opposite ways across different individuals with autism, may not be identified in group-level analyses at all, even though it may be critical for explaining behavioral features at the level of the individual or of subgroups. Negative findings in neuroimaging comparisons between autism and other developmental conditions such as ADHD can be equally explained: Critical brain differences diverging across different subsets within each condition may obscure findings at the group level.
As a way forward, we need statistical approaches for the identification of clusters, or subsets of autism participants, which usually rely on data-driven techniques. In this approach, availability of large samples will indeed be beneficial because they may permit the identification of variants of autism even if these are relatively uncommon.