Estimating the accuracy of early life predictors of 11-year language outcome: A replication study

• In Event: Language and Communication Across Development

Abstract

Background: Early identification of language disorders enables provision of support to those most in need of intervention. A previous study identified 6 preschool indicators that predict 11-year language outcome with 75% sensitivity and 81% specificity using data from the Early Language in Victoria Study (ELVS). These predictors related to sentence complexity, vocabulary and gross motor skills (see Table 1). This study aims to replicate the original predictor study using data from the Longitudinal Study of Australian Children (LSAC) birth cohort.
Method: LSAC recruited a representative birth cohort of 5107 infants across Australia born between March 2003–February 2004 and followed them up in biennial waves thereafter. Between 0-3 years of age, LSAC collected parent surveys, direct assessments and investigator observations relating to the child. A selection of families who were retained at wave 6 (73.7% of baseline participants) were selected to participate in the Child Health CheckPoint, a one-off detailed cross-sectional assessment nested between LSAC waves 6 and 7 when the children were aged 11-12 years. At the Child Health CheckPoint visit, 1441 children completed the sentence recall task of the Clinical Evaluation of Language Fundamentals, 4th Edition as a potential indicator of Language Disorder. We used SuperLearner an ensemble machine learning algorithm to estimate the accuracy of the set of 6 questionnaire items identified by the previous ELVS study in predicting 11-year language outcome (dichotomized in the LSAC cohort as sentence recall score >1.5SD below the mean).
Preliminary results: The current results include the 1174 participants for whom the 6 predictors from wave 2 (see Table 1) and the outcome, the sentence recall score, are available (1128 with typical and 46 with low language). At 2-3 years of age, this set of 6 predictors yielded 72% sensitivity (95% confidence interval (CI): [57,84]) and 71% specificity (95% CI: [68,73]) for predicting low language at 11 years. We conducted sensitivity analyses, using alternative outcome measures (Academic Rating Scale – Language and Literacy score at age 12 years and the Rice Test of Grammaticality Judgement prime score at age 14 years) and cut-offs for dichotomizing the outcome (1.25 and 2SD below the mean). These yielded lower accuracy (AUC<70%), which could be due to differences in the outcome measures and uncovers difficulties in accurately measuring the outcome of interest, namely Language Disorder.
Next steps: We are also using random forests to identify whether any additional variables relating to families’ socioeconomic status, parents’ or children’s general health and development arise as predictive of language outcome.
Implications: We have identified a set of 6 parent-reported questions relating to children’s morphosyntax, vocabulary and motor skills that predict with >71% accuracy which children will have language difficulties in late childhood, replicated across two population-based cohorts. This could be used by researchers to identify young children at risk of persisting Language Disorder, for example to recruit into early intervention trials.

Authors

• Loretta Gasparini, University of Melbourne
  Presenting Author

• Daisy A. Shepherd, The University of Melbourne

• Jing Wang, The University of Melbourne

• Melissa Wake, Murdoch Children’s Research Institute

• Angela T. Morgan, Murdoch Children's Research Institute