Embracing complexity in neurodevelopment - Part 3
Updated: Feb 25, 2022
Welcome back! This is a blog series on transdiagnostic research approaches to better understand childhood learning difficulties. The last article looked at how a researcher narrows down their research question. This month we cover how to choose research participants.
In a transdiagnostic study, you are faced with two choices. Do you want your findings to be interpretable within the supposed boundaries imposed by the current canonical diagnostic system (most commonly DSM-5)? Or do you want to generate new insights about the prevalence and the co-occurrence of different learning difficulties in the wider population? The recruitment of your participants will depend on this choice.
Recruitment of children based on their diagnosis
One strategy you might consider is to choose your sample according to predefined diagnostic groups. Is that not counterintuitive when conducting a transdiagnostic study? Actually, this is how transdiagnostic research into behaviour, cognition, and neurobiology first began. For some researchers, if it isn’t diagnostic, it can’t be transdiagnostic. For example, Kushki et al. (2019) recruited children with diagnoses of ADHD, autism, or Obsessive Compulsive Disorder (OCD) in addition to a group of comparison children (with no learning difficulty). However, unlike the majority of studies looking at neurodevelopmental disorders which compare the ‘control’ group with every ‘case’ group, the authors decided to let the data ‘speak’. By using clustering techniques (described in detail in subsequent articles of this blog series), new groups were defined in terms of inattention levels, social skills and brain structure, and not in terms of diagnostic labels. And surprise: there was little or no overlap between these new groups and the groups defined by diagnostic labels. Recruiting children with formal diagnostic labels can therefore be very useful for challenging the role these labels play in differentiating children.
Another strategy for including transdiagnostic elements into a conventional diagnostic study has been to include children with multiple co-occurring diagnoses. This approach has been useful for testing whether children diagnosed with both ADHD and autism present with unique features or simply the sum of ADHD features and autism features. Findings from these types of studies are clear: there is no single way of defining a co-occurring disorder because children typically display various combinations of features associated with different disorders.
Yet, constraining recruitment to the diagnostic framework will always limit the transdiagnostic aspirations of any study. First, it reinforces the mainstream practice of using these labels as a legitimate way of sampling the population and automatically inflates the differences between groups by using stringent criteria for exclusion. Secondly and most importantly, it excludes all undiagnosed children who experience barriers to learning but have never been formally diagnosed, either because they do not fit the ‘diagnostic standard’, or because of systematic barriers to diagnosis (e.g. race, ethnicity, socio-economic status).
Recruitment of children with identified difficulties (or ‘functional recruitment’)
There are several recruitment alternative to using diagnostic labels, each one with a very specific purpose in mind that should match your research question.
The first way is to recruit a larger and more heterogeneous population of children with identified difficulties, whether they conform to standard diagnostic criteria or not. For example, many researchers in the 4D lab, alongside their own projects, are responsible for running the CALM (Centre for Attention, Learning and Memory) project which intends to recruit children based on functional needs rather than formal diagnosis. The CALM project recruits children referred by specialist educational and/or clinical services. Typically, less than 40% of these children have received a formal diagnosis of autism, dyslexia or ADHD. Studies analysing the CALM sample have consistently shown that the subgroups of children identified with novel transdiagnostic analytical methods (e.g. community detection in Bathelt et al., 2018 or self-organising maps in Siugzdaite et al., 2020: see next blog for further details) did not match their formal diagnosis status very well (e.g. the subgroup of children with elevated inattention and hyperactivity did not match the subgroup of children formally diagnosed with ADHD).
The issue with functional recruitment is the ascertainment bias associated with the likelihood of referral. Indeed, this likelihood increases due to several factors: one being the co-occurrence of difficulties (a child with both language and motor difficulties is more likely to be referred than one with a single difficulty), the other being environmental and cultural (e.g. parents with greater financial resources or from specific cultures are more likely to seek help in dedicated educational or clinical services than parents with financial difficulties).
In this study, Roma Siugzdaite showed that diagnostic labels were unrelated to the children's cognitive abilities. If they were, we would see children diagnosed with a specific disorder (depicted as blue nodes in the figures) grouped together and not spread across the network.
See her blog for a full description of this study.
One way to reduce this bias is to continue relaxing the recruitment process and get rid of both diagnostic and functional criteria. This is the second and most extreme way to go beyond the diagnostic labels. These types of studies tremendously reduce sampling bias as they only keep criteria such as year of birth (birth cohort study: MCS), geographical location (e.g. SCALES, ALSPAC) or being a twin (e.g. twinsUK). One of the purposes of such unselective recruitment-based studies is to furnish population-level information of neurodevelopmental difficulties, such as prevalence or range of severity. Population-representative cohorts also more directly challenge the current diagnostic labels by testing more optimal classification of children based on their daily difficulties identified through routine tests (e.g. questionnaires for parents or teachers) rather than specific diagnostic tests.
A study like SCALES which recruited a population-representative cohort of 7000 children living in the same county and born on the same years was able to expose a fundamental flaw in using the typical IQ test to identify variations in cognitive scores for children with learning difficulties. For instance, Norbury and colleagues (2016) showed that IQ tests failed to identify 75% of children experiencing difficulties with language. On the contrary, SCALES calculated that 10% of their cohort had language difficulties with only a small minority formally diagnosed with language impairments.
Population-level cohorts also have the advantage of being (at least most of the time) longitudinal, i.e. children are asked to be tested at different stages of their development. The Millenium Cohort Study sampled over 18,000 children, all born in 2000, and assessed them 8 consecutive times so far: at 9 months, 3, 5, 7, 11, 14, 17 and 20 years old. This type of study offers invaluable new insights on how complex interactions of hundreds of factors can shape the inner and outer life of a child over time. For instance, the TEDs longitudinal study of UK twins revealed that the number of preschool difficulties (breadth of risk) better predicted language and reading difficulties eight years later, in early adolescence, than the magnitude of preschool difficulties (depth of risk) (Hayiou-Thomas et al., 2021). It is worth mentioning that while such longitudinal studies offer a way to account for the temporal heterogeneity in the nature and the severity of difficulties across major developmental stages, they often don’t have the granularity of measures that we need in order to study how behavioural, cognitive and brain characteristics interact with each other in groups of children with different types of difficulties.
In conclusion, there are several ways of choosing a sample for transdiagnostic studies, each fulfilling a different purpose that must match your research question as best as possible. Incorporating diagnostic labels does not necessarily go against the core principle of the transdiagnostic approach but allows a more direct way of challenging existing diagnostic boundaries. However, we must keep in mind that this type of recruitment (by diagnosis or functional) inevitably introduces biases (selection bias or ascertainment bias, respectively).
Finally, population-representative cohorts offer the possibility of studying large numbers of children in the community with subtle but nonetheless significant needs. New analytical methods must be mapped out to accommodate these large-scale studies and fully exploit this flow of data. This will be the topic of our next blog: leveraging the dimensional approach to uncover new latent factors or dimensions hidden in this mass of data. Stay tuned!