UCLA

Minimally verbal children are considered the enigmatic and, unfortunately, the neglected end of the autism spectrum. This subpopulation has likely garnered this title due to their exclusion from research studies, which has inevitably affected their evidence base. The paucity of proper measurement tools that sensitively and accurately assess behaviors has been one limiting factor towards the improved knowledge of these children. Short of creating and validating a new measurement tool for this subpopulation, this study took an alternative and more immediate approach: examine an existing social communication measure (ESCS) with repurposed quantitative methodologies, item response theory (IRT) and machine learning algorithms (CART and random forests). The final sample consisted of 453 minimally verbal children culled from four different intervention studies. The IRT models analyzed the frequency of social communication gestures from the ESCS and returned an objective difficulty hierarchy regarding initiations of joint attention and behavior regulation gestures. The best-fitting and final model was a zero-inflated negative binomial model (ZINBM) which determined that joint attention gestures were, on average, more difficult than behavior regulation gestures. Joint attentional shows and gives were especially tough, and behavior regulation reaches were the easiest gestures for this sample. The ZINBM separately modeled children with some gestures and children who did not present with any gestures and determined that behavior regulation reaches and gives were likely the first gestures a child will eventually exhibit among children with no gestures. Classification and regression trees (CART) were used to understand the clinical meaning behind frequencies of social communication gestures. Influential cut points were identified by the recursive partitioning algorithm of CART and determined which frequencies were able to classify children into more or less robust language outcomes at baseline. On average, a single behavior regulation point was sufficient to classify children into more robust language outcomes. For many of the trees, responding to bids joint attention around one-third of the time or more was also predictive of more robust language outcomes. Variable importance was examined with a random forest algorithm, which matched the results from the classification trees. Overall, this study demonstrated that the use of IRT and CART yielded additional information, beyond traditional scoring and analytic techniques, regarding the presentation of social communication gestures among minimally verbal children. This study also discussed the methodological contributions and potential future applications of IRT and CART within this field.