This is the latest article of our blog series authored by researchers of the 4D lab, Duncan Astle and Sue Fletcher-Watson.
The purpose of this blog series is to make you more familiar with the different questions a researcher faces when studying neurodevelopment and applying a transdiagnostic approach. In the two last articles, I described how we choose our research questions and our population sample. This article will cover one of the analytical approaches developed as an alternative to the methods we typically apply in case-control designs which compare a group of children meeting criteria for a standard neurodevelopmental disorder diagnosis with either another diagnostic group or a typical group. This alternative method is most commonly referred to as the dimensional approach.
Relying on continuous dimensions that span the range from typical to atypical functioning: the example of the Research Domain Criteria Initiative (RDoC)
As described in the first article of this blog series, the co-occurrence of behavioural and cognitive difficulties in neurodevelopmental disorders is the rule rather than the exception (Gilger and Kaplan, 2001). Why does this happen? It’s partly because doctors make a clinical diagnosis based on what they can observe from someone’s visible behaviour, and match their observations to a list of criteria laid down in the Diagnostic and Statistical Manual of Mental Disorders (DSM-5; American Psychological Association, 2013). By design, the DSM is primarily concerned with identifying characteristics that tend to cluster together, determining when the features appear, and whether they resolve, recur, or become chronic. Criticism of the DSM has been ongoing for the last 20 years, particularly regarding its failure to demonstrate real boundaries between related DSM syndromes, or between a DSM syndrome and normality, or its disregard for the context in which a syndrome occurs. The RDoC initiative is a response to these long-term criticisms. It started in 2010 and the ‘revolutionary’ principle about it was to rely on multiple continuous dimensions that span the range from typical to atypical functioning rather than strict categories of observable cognitive or behavioural signs. Indeed, across both typically-developing and ‘neurodivergent’ children, the range of cognitive ability and behaviour spans several levels of functioning. This supports the idea of the continuity of clinical features rather than their binarization (typical vs atypical).
The RDoC was an essential initiative and laid down an overarching theory which helped researchers to come up with new types of datasets and new transdiagnostic methodologies. However, it had a fundamental flaw: the measures used to make a better classification of neurodevelopmental disorders were defined beforehand. In the cognitive domain, for instance, it specified six different measures or dimensions: attention, perception, working memory, declarative memory, language, and cognitive control (Cuthbert & Insel, 2013). Yet, it has already been established through experimental psychology and neuropsychology that these dimensions have several subconstructs (i.e. more precise dimensions) that can be efficiently measured and distinguished from one another. For instance, working memory has been decomposed into distinct storage and executive components (Baddeley, Allen & Hitch, 2011), and language into distinct elements including vocabulary, word recognition, comprehension and pragmatic aspects of communication (Language and Reading Research Consortium, 2015).
How do we define new dimensions and which level in the hierarchy should we aim for?
A fundamental goal of the transdiagnostic approach is to identify these dimensions relating to cognition, behaviour or neurobiology based on the data themselves. We have multiple ways of doing this.
New dimensions can be first defined at the highest level of the hierarchy, i.e. the phenotype (observable features) such as children’s language development, academic achievement, social functioning or behavioural control. A range of phenotypic dimensions have been identified in the last ten years using a range of statistical techniques which attempt to identify the smallest number of dimensions that can parsimoniously explain how a set of measured features covary together, i.e. how features change together and at which extent they move in the same or opposite direction. Data reduction methods such as covariance analysis, exploratory factor analysis or latent class models can all identify the largest ‘sources’ of variation in a range of different features. For example, it is possible that variation in six observed variables mainly reflect the variations in two unobserved (or underlying) variables. Looking at how a range of measures such as questionnaire data or performance scores at experimental tasks co-vary together have led to the discovery of several dimensions. In particular, ‘Hyperactivity and impulsivity’, ‘Inattention’, ‘Social communication’, ‘Executive functioning’, ‘Phonological processing’ have repetitively been identified in several studies using one or another data reduction method (see Astle et al., 2021 for a review).
Taxometry is another popular way to test if a categorical or dimensional account provides a better fit to a specific set of data. It has notably been used with questionnaire data asking parents to assess their children’s behaviour and showed that both hyperactive/impulsive and inattention are distributed continuously across the population (with or without an ADHD diagnosis) as opposed to discrete difficulties that could easily differentiate ADHD children from typically-developing children (Marcus & Barry, 2012).
We can also define dimensions that sit at an intermediate position between the phenotype and its underlying neurobiological mechanisms, called the endophenotype. Unlike observable features in behaviour that we tend to categorise, an endophenotype corresponds to a quantifiable measure on a continuous scale, not directly accessible to clinical practitioners unless specifically tested with standardised tests. Performance results on these tests can indicate an underlying problem at a lower level of the hierarchy, for instance neurotransmitter functions at the level of synapses. For example, a go-no go task which measures the time taken to inhibit a response has been related to the concept of delay aversion which finds its mechanism deep inside the brain, in an area called the striatum which regulates reward processing (Castellanos & Tannock, 2002). Delay aversion has been related to lesions of the striatum but also excessive presence of the striatal transporter responsible for transporting dopamine, linked to a polymorphism in the gene DAT1 (Madras et al., 2002). People with ADHD typically displayed longer reaction times at the go-no go task. Interestingly, it has also been shown that children with ADHD showing slow reaction time were more likely to have first-degree relatives with ADHD than children with ADHD showing normal reaction time (Crosbie & Schachar, 2001), showing a potential genetic causal mechanism between DAT1 polymorphism and ADHD features. These kind of experimental tasks, offering quantitative metrics on a dimensional scale, rather than discrete categories like the presence or the absence of a particular behavioural feature, is crucial if we want to get closer to mechanistic biological processes. Testing both parents and children affected by a neurodevelopmental disorder can give us precious clues as to which metrics and underlying dimensions to use to identify an endophenotype. The search for endophenotypes holds a lot of promise as it would allow us to bridge the gap between high-level feature presentation and low-level genetic variability and thus differentiate potential diagnoses that present with similar phenotypes or features.
Advantages and limitations of the dimensional approach
If we define neurodevelopmental disorders according to these broad transdiagnostic dimensions, what we see is that all dimensions encompass several disorders (Astle et al., 2021).
By establishing these continuous dimensions that characterise neurodevelopmental processes, we favour broad assessment of features that frequently co-occur, and we allow a more flexible recruitment of individuals with a complex combination of ‘dimensional’ difficulties, each presenting with graduated levels of functioning. Dimensional methods thus seem to provide a more efficient guide to assessing the needs of individual children than conventional diagnostic boundaries, with the potential for developing and delivering personalised interventions. By capturing more variability between children, these dimensions might also be closer to underlying mechanisms, or at least more useful for identifying these mechanisms.
However, researchers must make several thoughtful and evidence-based methodological decisions while conducting dimensional studies. For instance, the choice of variables is a fundamental step and will eventually determine the depth of dimensions, e.g. relatively broad levels (e.g., phonological, inattention, pragmatic communication), or narrower constructs such as components of phonological processing. Yet, this choice heavily depends on the available data and the type of sample. The type of measures included in the analysis might induce unforeseen common variance for instance, whether it is a binary response questionnaire or a score represented as a frequency value. Notably, this can lead to poor generalisability in the case of diagnostic-based or functional recruitment due to skewing of behavioural or cognitive scores and shrinking of variance along very specific variables depending on the recruitment criteria. On the contrary, if the population sample is representative of the broader population, dimensions characteristic of ‘atypical’ subsamples might be lost, subdued by dimensions representing the ‘typical’ majority of the population-representative dataset.
The next article will describe one solution that researchers have found to avoid this last scenario by modelling dimensional approaches in subgroups of the available population sample. These are referred to as clustering techniques. Stay tuned!