Subdimensions of social‐communication behavior in autism—A replication study

Abstract Introduction In order to identify more refined dimensions of social‐communication impairments in autism spectrum disorder (ASD) a previous study applied exploratory and confirmatory factor analyses to diagnostic algorithm scores of the autism diagnostic observation schedule (ADOS), Module 3. A three‐factor model consisting of repetitive behaviors, impairments in ‘Basic Social‐Communication’ and in ‘Interaction quality’ (IQ) was established and confirmed. The current study aimed to replicate this model in an independent sample. To advance our understanding of the latent structure of social communication deficits, previous work was complemented by a probabilistic approach. Methods Participants (N = 1363) included verbally fluent children and young adults, diagnosed as ASD or non‐ASD based on “gold standard” best‐estimate clinical diagnosis. Confirmatory factor analysis examined the factor structure of algorithm items from the ADOS Module 3 and correlations with individual characteristics (cognitive abilities, age) were analyzed. Linear Regressions were used to test the contribution of each latent factor to the prediction of an ASD diagnosis. To tackle large inter‐correlations of the latent factors, a Bayesian exploratory factor analysis (BEFA) was applied. Results Results confirmed the previously reported observation of three latent dimensions in the ADOS algorithm reflecting ‘Restricted, Repetitive Behaviors’, ‘Basic Social‐Communication’ behaviors and ‘Interaction Quality’. All three dimensions contributed independently and additively to the prediction of an ASD diagnosis. Conclusion By replicating previous findings in a large clinical sample our results contribute to further conceptualize the social‐communication impairments in ASD as two dimensional.


INTRODUCTION
The concept of autism spectrum disorder (ASD) has shifted from a "childhood condition" with associated challenges in language and intellectual functioning, to a wider concept of ASD including individuals with only mild symptoms/autistic traits or those who do not show symptoms until later in life (Vivanti & Messinger, 2021). According to the diagnostic conceptualizations in the diagnostic and statistical manual of mental disorders fourth ed (American Psychiatric Association, 1994) as well as in the international statistical classification of diseases, 10th revision (World Health Organization [WHO], 1993), ASD used to be conceptualized as a multi-categorical disorder with different subtypes (childhood autism, asperger syndrome, pervasive developmental disorder not otherwise specified or atypical autism and others). However, these ASD subtypes could not be differentiated as distinct, empirically defined subgroups (Lord et al., 2012a;Walker et al., 2004) but rather overlapping conditions (Macintosh & Dissanayake, 2004;Snow & Lecavalier, 2011). Consequently, ASD is now conceptualized as a hybrid model in the diagnostic and statistical manual of mental disorders, fifth ed.  (American Psychiatric Association, 2013): Dimensional individual differences in symptom severity and general impairment are considered within a categorical umbrella term of ASD (Frazier et al., 2012;Grzadzinski, Huerta, & Lord, 2013). The development towards a dimensional understanding and the broadening of the diagnostic criteria has led to substantial and increasing heterogeneity in the clinical phenotype of ASD (Fombonne, 2020;Hansen, Schendel, & Parner, 2015). This heterogeneity is a challenging problem in research as well as in clinical practice (Lord et al., 2020;Mottron & Bzdok, 2020). The discovery of valid biomarkers is hampered by phenotypic heterogeneity, symptom overlap with other conditions as well as co-occurring conditions in ASD. This is visible in a drop of effect sizes by up to 80% from cognitive, electroencephalogram and neuroanatomical group comparison studies in the past 2 decades (Rødgaard, Jensen, Vergnes, Soulières, & Mottron, 2019).
In order to define the boundaries of ASD many efforts have been made to identify symptom dimensions which has led to the postulation of two symptom domains (social communication and restricted interests and/or repetitive behaviors) in the DSM-5 criteria for ASD.
However, research could not yet identify replicable subdomains within the social-communication domain that may define subgroups of individuals with ASD. Bishop and colleagues argue that more refined dimensions of the social-communication domain of ASD are needed "to elucidate the clinical, nosological and biological boundaries of the multiple disorders associated with social-communication impairment" (Bishop et al. 2016;p. 909). They examined the organizational structure of clinician-observed social-communication deficits with exploratory factor analysis in a sample of 238 children with and without ASD between the ages of 2 and 12 years and found a three-factor model consisting of restricted, repetitive behaviors (RRB) and two separate social-communication behavior dimensions, "Basic Social-Communication" (Basic SOC) and "Interaction Quality"(IQ). This factor structure could be replicated by confirmatory factor analysis (CFA) in an independent sample. The impairments in "Basic SOC" behaviors included items measuring eye contact, facial expressions, gestures and shared enjoyment and were separated from impairments in "Interaction Quality", including items measuring conversation, amount of reciprocal social communication, overall quality of rapport, and quality of social response. While scores in "Interaction Quality" were significantly associated with nonverbal Intelligence quotient (IQ) and male sex in the ASD group and with age in the non-ASD group, scores in "Basic SOC" were not significantly associated with these phenotypic variables but "remarkably intact in children who do not have ASD, even in the presence of significant other impairments or risk factors" (Bishop et al., 2016, p. 913). The authors conclude that basic impairments in nonverbal communication and shared affect seem to be specific to ASD and thus could provide a more specific index of ASD severity, whereas impairments in interaction quality appear to be less specific to ASD and more impacted by other child characteristics and thus may be more relevant for differential diagnoses.
The current study aimed to replicate this suggested two-fold nature of social-communication impairments in a large and independent sample of individuals with ASD and a large sample of individuals with relevant differential diagnoses and other developmental issues. We address two hypotheses originating from the results of Bishop et al. (2016): 1) child characteristics should not be associated with the basic dimension and 2) the factor "Basic SOC" would be more predictive of ASD than "Interaction Quality". We thus tested correlations of age and verbal IQ with the latent subdimensions of social communication in the current sample. Lastly, in order to identify the model that best fit our current data, the methodological approach was complemented by a Bayesian exploratory factor analysis (BEFA).

METHOD Participants
The data used for the present study represent a subsample extracted from an existing research database of the ASD-Net, a state-funded research network (Kamp-Becker et al., 2017). Datasets stem from individuals who had been referred to ASD specialized Key points 1. In a large clinical sample of verbally fluent children and adolescents with and without ASD, we could replicate the finding of two separable subdimensions of socialcommunication impairment 2. A Bayesian exploratory factor analysis supports the conceptualization of social-communication impairments in ASD as two-or even multi-dimensional 3. The "Basic SOC" subdomain represents a replicable and stable factor that seems appropriate as an index of ASD symptom severity 4. Identifying subdimensions of social-communication impairment and subtypes along the autism spectrum may support future research efforts trying to link neurobiological mechanisms to specific types of behaviors outpatient clinics for a diagnostic assessment due to a suspicion of ASD. To assemble a representative sample of individuals who seek an investigation of ASD, the presence of a clinical suspicion of ASD was the general inclusion criterion. Subjects were either diagnosed as having ASD or the diagnosis was excluded (non-ASD) based on "gold standard" best-estimate clinical diagnosis ( Table 1. Participants' data were collected retrospectively from the respective medical record and analyzed anonymously. The procedure was approved by the local ethics committee (Az. 92/20) and due to the retrospective nature of data collection and analysis of anonymized data, the need for informed consent was waived by the ethics committee. All methods were performed in accordance with the relevant institutional and international research guidelines and regulations.
The study included 1363 datasets (4-27 years of age, n = 557 with ASD, 9.5% female, n = 806 with other mental disorders, 12.8% female), all evaluated with the ADOS Module three during the diagnostic process.
IQ scores, including full scale IQ, verbal IQ and non-verbal IQ, were available for 902 participants (82.1%). A statement regarding the intellectual level from the clinical record (average, above average, borderline, mildly impaired, etc. following ICD-10 categories) was available for another 6 cases (6.2%). Preliminary analyses on group differences regarding age and IQ and ASD symptoms are reported in Table 2.

Measures
The

Statistical analyses
Following ADOS conventions codes of seven and eight were recoded to 0. Codes of three were re-coded to 2. Missing values were excluded listwise (12 cases had missing values in the ADOS data und were thus excluded from CFA and BEFA). Group characteristics were explored via t-Tests. In order to test the latent dimensions suggested by Bishop et al. (2016), a CFA was performed with a robust weighted least square mean and variance adjusted estimator and oblique rotation (goemin) to allow factor correlations.
Factor number and factor-item-assignment was based on the model of Bishop et al. (2016) including three factors ("Basic SOC", '"Interaction Quality" and '"RRB"). The ADOS item "quality of social overtures" (QSOV) was excluded from the CFA following the exclusion by Bishop et al. (2016). Model fit was assessed by a combination of parameters: The goodness-of-fit index χ 2 is sensitive to sample size and was therefore combined with the root-meansquare error of approximation (RMSEA), the comparative fit index (CFI) and the Tucker-Lewis index (TLI). Evaluation of model fit was based on criteria used by Bishop et al. (2016) (Hu & Bentler, 1999). The model fit of the three factor model was finding the optimal model among candidate models) by modeling parameters as probability distributions (Eddy, 2004) and producing consistent and intuitive estimates of posterior probabilities. Data are only discarded from the model if they have an exceedingly small probability (≤0.02) of loading onto any factor (Conti et al., 2014). The current study used the R package BayesFM (Piatek, 2019). A sparse model with uninformative priors (Huang & Wand, 2013) was specified as to estimate the optimal data-driven factor solution. Following a predictive analyses for the prior specification for the number of latent factors and the covariance matrix of the latent factors, the factor model was estimated using the parameters presented in the supporting information (Table S1). All other parameters were set to default values (see BayesFM documentation for details).  Table S2.

Testing the factor model
Confirmatory factor analyses were used to replicate the factor structure suggested by Bishop et al. (2016). We found a threedimensional model with acceptable model fit and a two-dimensional model with inferior but still acceptable model fit. Figure S1 in the supporting information materials depicts the structure of the models with two und three latent factors.
The Basic SOC-factor correlated high with the Interaction Qualityfactor (r = 0.92, p < 0.001) and even in a high range with the RRBfactor (r = 0.72, p < 0.001). The Interaction Quality-factor also correlated in a high range with the RRB-factor (r = 0.71, p < 0.001).
A two-factor model also showed acceptable fit indices:  Table 3.

Comparison of the models via a Chi squared difference test
showed superiority of the three-factor model over the two-factor model (χ 2 diff = 32.12; df = 2, p < 0.001), although differences in the fit indices seemed negligible.

Bayesian exploratory factor analysis
The specified BEFA model returned a posterior probability of 85.1% for a 5-factor model. Two almost identical models with 5 factors showed posterior probabilities of pmp m1 = 49.3% and pmp m2 = 35.8%. Figure 1 shows the factor structure of both five factor models. Factor solutions with less factors were unlikely given the observed data (K = 3, pmp = 9.4%, K = 4, pmp = 5.5%).
We had re-entered the ADOS algorithm item "Quality of Social Overtures" (QSOV) back into the analysis after it had been discarded by Bishop et al. due to unclear factor loadings on the Basic SOC and interaction quality subdimensions. It was argued that this items seems to capture aspects of behaviors that rely on skills in both dimensions (see Bishop et al., 2016, p. 912). In our BEFA analysis this items again switched between both subdimensions loading on a factor related to basic SOC in m1 and on a factor related to interaction quality in T A B L E 3 Results from the confirmatory factor analysis: factor loadings for the three factor solution and two factor solution

Amount of Reciprocal Social Communication [ARSC]), and Interaction
Quality appropriateness " (QSR and OQR). "Basic SOC eye " correlated with "Basic SOC ges " (r = 0.94) and "Interaction Quality conv " correlated with "Interaction Quality appropriateness " (r = 0.92). All other factors were also intercorrelated as shown in a correlation matrix in Table 5.

Associations between ASD symptom dimensions and individual characteristics
In the full sample (ASD as well as non-ASD), the RRB factor was negatively correlated with nonverbal IQ, meaning that individuals with higher IQ showed less repetitive behaviors. In the non-ASD group this association was also found for verbal IQ. In the ASD group "RRB" was also negatively correlated to age -the older the child, the less repetitive behaviors were observed. Furthermore, in the ASD group but not in the non-ASD group, "Interaction Quality" was negatively correlated with nonverbal and verbal IQ meaning that higher IQ was associated to less impaired interaction quality.
The "Basic SOC" dimension showed only low and nonsignificant correlations with IQ. In the non-ASD group higher age was associated with better "Basic SOC" skills. Sex was not correlated to any dimension in neither ASD nor non-ASD individuals. Correlation coefficients are presented in Table 6. All coefficients represent small effects according to Cohen's interpretation of effect sizes.
Correlations of the five BEFA dimensions fully resembled the pattern of correlation with the three symptom dimensions and are thus not further described.

Associations with ASD diagnosis
The Interaction Quality appropriateness was the only factor that was not predictive of diagnosis.

DISCUSSION
The results of this replication study amended by a BEFA confirms that items from the ADOS-2 diagnostic algorithm need to be par- Associations of "Interaction Quality" with child characteristics from the ASD group were restricted to IQ measures, no associations were found in the non-ASD group. Unlike previous findings by Bishop et al. (2016)

Strengths and limitations
A key strength of this study is the considerable sample size and the inclusion of a clinical comparison group that comprises individuals with mental disorders that are relevant differential diagnoses to ASD.
In consideration of the "replication crisis" (Lewandowsky & Oberauer, 2020) replicability is fundamental especially for research on the heterogeneous autism spectrum. Our presented results add substantial evidence to the proposed subdimensions of socialcommunication impairment in ASD.
A limitation of the current study -as has been a limitation of

CONCLUSION
Social communication is a multidimensional construct that can be influenced by individual, contextual, and other factors and thus requires a more precise approach (Bishop et al., 2016). By replicating previous findings in a large clinical sample our results support the efforts to further conceptualize the social-communication impairments in ASD as two-or even multi-dimensional. A better understanding of different types of social communication impairments will promote the identification of behaviorally relevant subgroups within ASD.

ACKNOWLEDGMENTS
The authors would like to thank Sarah Wittkopf, Gerti Gerber, Imke Garten and Marie Kollarczyk for their assistance in the conduct of this research, all clinicians who collected the data and all patients who participated in the study. This work was funded by the German Open access funding enabled and organized by Projekt DEAL.

CONFLICT OF INTEREST
The authors have declared that they have no competing or potential conflicts of interest.

ETHICAL CONSIDERATIONS
Participants' data were collected retrospectively from the respective medical record and analyzed anonymously. The procedure was approved by the local ethics committee (Az. 92/20) and due to the retrospective nature of data collection and analysis of anonymized data, the need for informed consent was waived by the ethics committee. All methods were performed in accordance with the relevant institutional and international research guidelines and regulations.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available on request from the corresponding author -pending approval of the coauthors. The data are not publicly available due to privacy restrictions.