A Mixed Models Approach to Analyzing Large Cohorts of Natural Conversational Data from Individuals with ASD

Background: Natural, unstructured conversation represents a significant challenge for individuals with autism spectrum disorder (ASD), but relatively little is known about how speech and language unfold in an uncontrolled context; less yet is known about the myriad interacting factors that likely influence dynamic conversation in ASD, such as age, sex, and cognitive ability. Until recently, most studies lacked sufficient power to assess all of these contributors in a single cohort. Here, we report on a growing corpus of natural conversations between individuals with and without ASD and naïve confederates. This corpus will ultimately provide the field with a multi-faceted view of conversational interactions in typical development (TD) and in individuals with ASD, and lay the foundation for new metrics to assess social interaction that allow us to track change across ages, sexes, and ability levels.

Objectives: Use mixed effects models to assess the contributions of diagnostic status, age, IQ, and sex to speech/language features produced during naturalistic conversations with naïve confederates.

Methods: Sixty-nine individuals aged 5-59 years with ASD (N=48, 15 female, mean age=14.93y, mean IQ = 96) and TD (N=21, 15 female, mean age=21.89y, mean IQ=107) participated in a naturalistic ~5 minute “get to know you” conversation with one of 13 undergraduate confederates (10 female). Language was transcribed and time-aligned using XTrans (LDC). Average speaking rate, words per minute, word length, response latency, and turn length were calculated. LIWC was used to calculate social category words (Tausczik & Pennebaker, 2010). The contributions of diagnostic status (ASD, TD), chronological age, sex (M, F), and full-scale IQ to speech and language features were assessed using linear mixed effects models, with confederate ID included as a random effect to account for individual variability in naïve interlocutors.

Results: Linear mixed models revealed that participants spoke faster as they got older (t=3.52, p<.001). Older participants produced more words (t=5.39, p<.001), as did individuals with higher IQ estimates (t=3.05, p=.003). Participants with ASD used slightly longer words than TD participants (t=2.12, p=.04). Older participants had longer response latencies (t=2.30, p=.02), produced more overlapping speech (t=2.40, p=.03), and took longer turns (t=4.54, p<.001). Participants with higher IQ estimates also took longer turns (t=3.48, p<.001). Social topics were discussed more often by older participants (t=2.42, p=.02) and by female participants (t=2.40, p=.02).

Conclusions: A multi-faceted understanding of natural conversational interactions in ASD will inform the next generation of highly granular intervention response metrics. We anticipate adding an additional 20 participants to this cohort by May, 2018, as well as including additional speech/language metrics and an analysis of confederate speech.