Autism spectrum disorder is a highly heterogeneous diagnosis. When a child is referred to autism services or receives a diagnosis of autism spectrum disorder it is not known what their potential adult outcomes could be. We consider the challenge of making predictions of an individual child’s long-term multi-facetted adult outcome, focussing on which aspects are predictable and which are not.
We used data from 123 adults participating in the Autism Early Diagnosis Cohort. Participants were recruited from age 2 and followed up repeatedly through childhood and adolescence to adulthood. We predicted 14 adult outcome measures including cognitive, behavioral and well-being measures. Continuous outcomes were modeled using lasso regression and ordinal outcomes were modeled using proportional odds regression. Optimism corrected predictive performance was calculated using cross-validation or bootstrap. We also illustrated the prediction of an overall composite formed by weighting outcome measures by priorities elicited from parents.
We found good predictive performance from age 9 for verbal and non-verbal IQ, and daily living skills. Predictions for symptom severity, hyperactivity and irritability improved with inclusion of behavioral data collected in adolescence but remained modest. For other outcomes covering well-being, depression, and positive and negative affect we found no ability to predict adult outcomes at any age. Predictions of composites based on parental priorities differed in magnitude and precision depending on which parts of the adult outcome were given more weight.
Verbal and non-verbal IQ, and daily living skills can be predicted well from assessments made in childhood. For other adult outcomes, it is challenging to make meaningful predictions from assessments made in childhood and adolescence using the measures employed in this study. Future work should replicate and validate the present findings in different samples, investigate whether the availability of different measures in childhood and adolescence can improve predictions, and consider systematic differences in priorities.