Background
Behavioural difficulties, poor communication skills, and everyday cognitive problems are common in children, and they often co-occur [
1‐
5]. Each set of symptoms is commonly associated with a specific developmental disorder, and as a consequence is typically studied in groups of children in whom such problems are characteristic. For example, behavioural problems such as hyperactivity are typically studied in children with attention deficit hyperactivity disorder (ADHD), while communication difficulties such as poor speech are often studied in children with developmental language disorder (DLD). There are practical advantages to this categorical approach. It defines clear symptom-based criteria to inform practitioner decision-making about diagnoses and interventions. However, it fails to accommodate high rates of comorbidity across developmental disorders [
6‐
10] and substantial heterogeneity within disorders [
7,
11‐
13]. The aim of the current study was to move away from studying links between problems of behaviour, communication, and cognition in discrete groups. Instead, symptom associations are explored in a large heterogeneous sample of children with clinical and subclinical levels of difficulties.
Over the past decade, there has been a broad shift away from diagnosis-specific deficits toward identifying dimensions that cut across disorders conventionally considered to be distinct [
14,
15]. This approach has been applied most widely to adult psychiatric conditions [
16‐
18], but there is widespread recognition of its value for characterising developmental disorders in terms of underlying dimensions of symptoms [
19‐
23]. One of the most common methods for understanding how symptoms are related uses latent variable models, a statistical method that groups variables based on shared variance to derive underlying dimensions of difficulties [
24]. This technique has been used to identify dimensions of phonological and non-phonological skills in children with diagnosed DLD and dyslexia [
25]; and separate latent constructs for inattention and hyperactivity in children with ADHD [
26].
Network analysis offers an alternative approach to understanding symptom interrelations. Instead of identifying underlying dimensions, network models focus on symptom-level associations, allowing for the possibility that symptoms could interact to causally affect and activate one another [
27,
28]. This framework is suited to conceptualising and evaluating the potential origins of comorbidities. For example, pragmatic difficulties characteristic of ASD, and hyperactivity problems common in ADHD often co-occur [
7,
11,
29,
30]. From a network model perspective, this co-occurrence can be conceptualised as arising in a dynamic system in which symptoms traditionally linked with one developmental disorder might trigger and/or maintain symptoms commonly associated with a different disorder.
In the current study network science is used to understand comorbidities between problems with communication, executive functions, and behaviour. Symptoms of communication and behavioural problems co-occur in the general population and in children with developmental disorders such as ADHD [
7,
11,
30,
31], ASD, DLD, and reading difficulties [
20,
31‐
33]. Individuals with communication and/or behavioural problems, including those with diagnosed developmental disorders, often have additional deficits in executive functions (EFs) - the cognitive abilities regulating thoughts and behaviour [
34]. EFs can be broadly construed as two distinct but related neurodevelopmental systems. Cool cognitive-based EFs encompass working memory, planning, and cognitive inhibition, and are associated with academic learning [
32] and attention [
33]; hot executive processes are associated with stronger affective valence and involve the regulation of emotional responses and social awareness [
34].
It is not clear why communication, cognitive, and behavioural problems co-occur in childhood. These areas of difficulties likely share common environmental and genetic influences. Nonetheless, there are several possibilities about how they may interact dynamically. One is that deficits in cool EFs might underlie both behavioural and communication difficulties. Consistent with this, poor working memory has been shown to underpin problems in attention, behaviour, and structural language [
33,
35], and also accounts for the relationship between hyperactivity and pragmatic communication problems [
36]. Alternatively, associations between cool EFs and pragmatics could be mediated by difficulties with hot EFs. Difficulties in inhibitory control could lead to hyperactive-impulsive behaviour and consequently poor social communication skills. Children who regularly display difficult behaviour may have limited opportunities to socialise, and thereby fail to develop good communication skills. In line with this, hyperactive-impulsive behaviours have been shown to account for the relationship between inhibition and the ability to apply pragmatic rules in everyday situations [
37]. Another possibility is that language deficits may directly and/or indirectly impact the ability to regulate cognition, behaviour, and communication [
38,
39]. For example, difficulties with language may lead to peer rejection and academic difficulties, which in turn may trigger behavioural problems.
The current approach uses network modelling to estimate associations and conditional independence across symptoms. Clustering methods are then used to identify closely inter-connected symptoms that may, or may not, correspond to areas of difficulty commonly associated with categorical diagnoses such as ADHD or DLD. This novel approach allows us to identify where symptoms sit in the network (i.e., which symptoms sit together), and to quantify the importance of different symptoms within the network. This includes identifying bridging symptoms that have multiple strong links across problem domains/clusters. Symptoms linking problem domains may reflect causal processes and/or shared aetiological influences. The conceptual interpretation of bridging symptoms identified in cross-sectional networks is not straightforward but one possibility is that such symptoms may spread activation across the system and may be the origins of the comorbidities observed on a dimensional level [
40]. All analyses are cross-sectional, exploratory, and data-driven. However, on the basis of co-morbid symptom presentations reported in the literature, it is predicted that multiple direct associations would emerge across pragmatic and behavioural difficulties (hot EFs), with additional connections between everyday cognitive abilities (cool EFs) and structural language skills.
Discussion
Network analysis was used, for the first time, to explore the co-occurrence of symptoms commonly observed across developmental disorders such as ADHD, DLD, ASD, and reading difficulties. Inter-symptom associations across communication, behavioural, and cognitive difficulties were modelled in a large heterogeneous sample of children. Four empirically-derived clusters of symptoms emerged corresponding to: 1) structural language and learning; 2) pragmatic abilities and peer relations; 3) behavioural and emotional difficulties (hot EFs); and 4) cognitive skills (cool EFs). Hot and cool EFs were directly related, as were structural and pragmatic language skills. Cool EFs were directly linked with learning but not with the formal use of language in communication.
Problems with pragmatic communication and peer relationships were directly connected to emotional and behavioural difficulties (hot EFs). In turn, behavioural difficulties were directly related to cool EFs. This was manifested in the high bridge centrality rankings of shifting and hyperactivity, both of which were in the hot EFs cluster. Their centrality in the network reflected associations with symptoms in the pragmatic communication cluster (e.g. hyperactivity – inappropriate initiation (0.13, 95% CI [0.07–0.18], non-zero = 97%); shifting – interests (0.27, 95% CI [0.22–0.34], non-zero = 100%) and the cognitive skills cluster (e.g. hyperactivity – inattention (0.36, 95% CI [0.31–0.44], non-zero = 100%); shifting – working memory (0.09, 95% CI [0.07–0.16], non-zero = 69%). These findings are consistent with hypotheses suggesting that social communication difficulties can arise as a downstream consequence of hyperactive-impulsive behaviours that are themselves underpinned by poor cool EFs [
35]. The network structure cannot provide evidence for the direction of these associations, but the observed paths may imply that cool EFs (e.g. inattention, working memory) may not directly influence peer relations and pragmatic skills, but can lead to difficulties in these domains via the activation of behavioural problems (e.g. hyperactivity, shifting) [
37,
62].
Contrary to expectations, structural language skills and cool EFs were not directly linked. It has been previously reported that cool EFs such as verbal short-term and working memory underpin language development [
63]. However, no such links were found in the network. One possibility is that the development of structural communication skills is constrained by phonological processing, which was not included in the current study, and not by executive function [
25,
64]. Links between both clusters and learning may reflect how cognitive deficits (e.g. working memory, inattention) and structural language skills (e.g. syntax, semantics) limit learning in different ways. Language impairments arising from phonological difficulties might be more closely related to literacy problems [
25], with cool EF problems impacting on broader aspects of learning [
32]. An alternative possibility is that cool EF deficits constrain learning, and this impairs the acquisition of structural language abilities [
31]. It is further possible that the direct associations were not observed due to potential poor specificity of the regularisation methods used.
Multiple direct paths linked the structural and pragmatic communication clusters, but their associations with other symptoms in the network differed. This supports the view that these dimensions of communication are linked but could have distinct origins [
65]. As suggested in the network, impairments in pragmatic language may arise through both structural language problems [
66] and social/behavioural difficulties [
67]. Structural language abilities appear to be more closely tied to learning, and indirectly to cool EFs.
Centrality indices highlighted important roles for working memory, language coherence, and the appropriate use of context in communication. Working memory and language coherence had multiple connections both within and outside their own clusters. Symptoms with multiple connections across problem areas may potentially interact with other areas of difficulty and may be the origin/consequence of co-occurrences [
40]. Working memory bridged the cool EF and structural language clusters via learning (0.12, 95% CI [0.10–0.19], non-zero = 76%). The role of working memory in learning is well-established [
68]. Working memory also linked the cool EF and pragmatic clusters via coherence (0.11, 95% CI [0.07–0.16], non-zero = 93%), and shared direct paths with hot EFs (shifting: 0.09, 95% CI [0.07–0.16], non-zero = 69%). Working memory did not rank highest on bridge strength, but its highest overall strength and expected influence, together with its direct associations with all symptoms within the cool EFs cluster and direct links with symptoms from all other clusters, provide evidence about its potential role as an area of difficulty that may spread activation across the network. Although causal conclusions on the basis of correlational network models are not warranted, working memory is important for holding information in mind, focusing attention, and ignoring distractions. This may explain how impairments activate difficulties in other problem areas such as producing coherent speech/narratives and shifting on to novel activities. Considering working memory as a transdiagnostic risk factor for developmental disorders fits with data from the mental health field [
69].
Coherence in communication (pragmatic abilities and peer relations cluster) was associated with all three symptoms of structural communication (edge weights range: 0.08–0.28, non-zero range: 70–100%) and four symptoms in its own cluster (edge weights range: 0.07–0.19, non-zero range: 72–100%). Poor language coherence may therefore activate, or be the consequence of, multiple other communication problems. In contrast, the high centrality of the use of context when communicating reflected multiple links within the pragmatics and peer relations cluster (edge weights range: 0.09–0.23, non-zero range: 90–100%) and a single association with the structural cluster (semantics: 0.12, 95% CI [0.07–0.17], non-zero = 97%). Symptoms with strong connections in the same cluster may be core symptoms of this problem area [
40]. Consistent with this, the appropriate use and interpretation of language in relation to context is central to the definition of pragmatics, and difficulties in this specific ability are reported to differentiate across children with primarily structural versus pragmatic difficulties [
70].
Limitations & future directions
There are several limitations to this study. First, the symptom relationships identified are based on parent ratings that are designed to capture aspects of functioning distinct from those measured in lab-based assessments of cognition [
71]. Parent ratings provide ecologically valid assessments of children’s day-to-day functioning across different situations [
71], but are known to be subject to reporter bias. To test whether bias affected the results, the current analyses were conducted both including and excluding reports flagged as overly negative or inconsistent. The overall pattern of results did not differ between analyses, indicating that such reporter bias is unlikely to be the cause of the overall network. In addition, the identification of clusters of symptoms that align with theoretical constructs provides further validation that the ratings were meaningful as they effectively distinguished between different aspects of functioning. It will be interesting to explore whether a similar network structure emerges with objective assessments of language, behaviour, and cognition.
The communities of symptoms identified depend on the chosen community detection algorithm and the stability of the input network. The algorithm applied here is reported to identify network communities with an acceptable level of accuracy, and the input network was stable in terms of the strength and number of connections across symptoms. Furthermore, the problem areas present in the network map on to broad latent constructs identified in previous studies of these comorbidities [
3], suggesting some level of agreement across methods.
The heterogeneous nature of the sample was suited for investigating the possibility of transdiagnostic symptom-level associations across comorbid difficulties. Nonetheless, it is also possible that disorder-specific associations may have been masked by combing multiple diagnostic groups. The current study was too small in size to formally test this possibility, but coarse checks based on comparing a network for the largest diagnostic group in the sample (those with ADHD) to a network based on the full sample showed acceptable correspondence. An important direction for future research will be to compare networks both within and across diagnoses.
The network model revealed bridging symptoms, which may provide important insights about the origins of comorbidities observed on a dimensional level, and which could be interpreted as candidate targets for interventions. However, in order for such interventions to be successful the temporal order of activation in relation to connected symptoms is of key importance. The targeted difficulty should be the cause rather than the effect of other symptoms. The current analytical framework does not afford such conclusions – central symptoms may cause other symptoms or may be the consequence of those other symptoms. Furthermore, the estimated relationships may not signal interacting areas of difficulties and may instead reflect shared item content or aetiological influences that were not included in the model. To disentangle these possibilities, it is important to evaluate whether bridging symptoms play a causal role in the co-occurrence of problem domains and to test whether interventions targeting these symptoms reduce the activation of difficulties in other areas of functioning. The incorporation of neurological, genetic, and environmental factors into network models is another important step that could provide important insights about the origins of comorbidities and their potential dynamic interactions.
Conclusion
The co-occurrence of pragmatic communication and behavioural problems was observed in a large heterogeneous sample of children with a broad range of difficulties. This suggests these comorbidities extend beyond specific diagnostic groups (e.g. ADHD, ASD). On a practical level, these findings highlight the importance of considering potential language and communication problems among children presenting with difficult behaviour, and vice versa.
Pragmatic communication skills might be indirectly influenced by cognitive skills through mediating role of behavioural regulation. In line with this, working memory and was identified as a bridging symptom, suggesting it may spread difficulties across different domains of functioning. Working memory, inattention, and structural language difficulties, were associated with learning, in line with many previous reports.
The data presented here provide one of the first large-scale applications of network modelling to symptoms associated with a range of developmental disorders. Akin to developments in psychiatry, this investigation suggests that there might be utility in shifting away from conceptualising developmental disorders as nosological entities. Transdiagnostic approaches can enable the discovery of shared liabilities and are suited for investigating the possibility that different developmental difficulties may cause one another: each problem may be the starting point for the activation of other symptoms. Using network modelling to conceptualise developmental comorbidities as arising in dynamic causal systems can provide insights into the nature of these comorbidities and may help researchers and clinicians to formulate specific hypotheses about potential causal mechanisms and intervention strategies.
Acknowledgements
The Centre for Attention Learning and Memory (CALM) research clinic is based at and supported by funding from the MRC Cognition and Brain Sciences Unit (CBSU), University of Cambridge. The lead investigators are Joni Holmes, Susan Gathercole, Duncan Astle, Tom Manly, Kate Baker, and Rogier Kievit. Data collection is assisted by a team of researchers and PhD students at the CBSU that includes Joe Bathelt, Giacomo Bignardi, Sarah Bishop, Erica Bottacin, Lara Bridge, Annie Bryant, Sally Butterfield, Elizabeth Byrne, Gemma Crickmore, Fánchea Daly, Tina Emery, Laura Forde, Delia Fuhrmann, Andrew Gadie, Sara Gharooni, Jacalyn Guy, Erin Hawkins, Agniezska Jaroslawska, Amy Johnson, Elise Ng-Cordell, Sinéad O’Brien, Cliodhna O’Leary, Joseph Rennie, Ivan Simpson-Kent, Roma Siugzdaite, Tess Smith, Stepheni Uh, Francesca Woolgar, and Mengya Zhang. The authors wish to thank the many professionals working in children’s services in the South-East and East of England for their support, and to the children and their families for giving up their time to visit the clinic.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.