Factor Analysis of DSM-IV, DSM-5, and Other Models of Symptom Structure in Toddlers with Autism Spectrum Disorder

Background: Little agreement exists about the exact structure of autism spectrum disorder (ASD) symptoms, and the DSM-IV conceptualization has received little empirical support. As such, several other models have been proposed, including alternative three-factor models (Boomsma et al., 2008; Georgiades et al., 2007; vanLang et al., 2006), more parsimonious two-factor models (Frazier et al., 2008; Gotham et al., 2007; Snow et al., 2009), and a one-factor model (Constantino et al., 2004). However, none has emerged as the best fitting, perhaps due to variations in sample size and characteristics, and age range. In particular, examining the phenotypic structure with a broad age range likely fails to account for developmental changes in symptom presentation. Symptom structure has not yet been examined specifically in toddlers, although it is crucial to understand the phenotype early in development.

Objectives: To compare the relative fit of DSM-IV, DSM-5 and other proposed models of autism symptom structure using confirmatory factor analysis (CFA) in a large, representative sample of toddlers. 

Methods: Children were recruited from the FIRST WORDS^® Projects at Florida State University and University of Michigan Autism and Communication Disorders Center. Children included had one or more administrations of the Autism Diagnostic Observation Schedule-Toddler Module (ADOS-T: Lord, Luyster, Gotham, & Guthrie, in press) between 12-30 months, and received a best-estimate diagnosis of ASD (total observations=387; unique cases=199). ADOS-T algorithm and play items provided indicators of autism symptoms for CFAs. 

Results: Five models were specified: (1)one-factor, (2)two-factor DSM-5, (3)two-factor DSM-5 including play items in the social communication domain, (4)three-factor DSM-IV, and (5)three-factor model proposed by VanLang et al. (2006). A series of CFAs examining the relative fit of the specified models was run using Mplus software (Muthen & Muthen, 1998).  Weighted Least Squares Mean and Variance Adjusted was used for estimation, as it is preferred when modeling ordinal data. DSM-IV, DSM-5, and VanLang models demonstrated good fit according to RMSEA (i.e., values ≤.08), CFI, and TLI values (i.e. values≥.95). Further comparison of fit using AIC and BIC, generated by rerunning models using Maximum Likelihood, indicated that the DSM-5 model demonstrated the best fit. Item loadings on factors and estimates of the relationships between factors will be discussed.

Conclusions: Results indicated the two-factor structure proposed for DSM-5 provided the best fitting model for a large, representative sample of toddlers. These findings lend support to the use of two distinct domains to characterize features measured by the ADOS-T, suggesting that although symptom presentation changes throughout development, factor structure in toddlers is similar to what has been documented in older children (Frazier et al., 2008; Gotham et al., 2007; Snow et al., 2009). The performance of the two-factor DSM-5 model is also consistent with studies that find that social and communication deficits represent just one domain, while repetitive/unusual language may be most likely to load together with repetitive and stereotyped behaviors. Finally, the goodness of fit provided by the models, which utilized observational ADOS-T items as indicators, supports the utility of this newly developed tool to capture the core components of the ASD phenotype in toddlers.