Reading Difficulties Identification: A Comparison of Neural Networks, Linear, and Mixture Models

Abstract

ABSTRACT Purpose We aim to identify the most accurate model for predicting adolescent (Grade 9) reading difficulties (RD) in reading fluency and reading comprehension using 17 kindergarten-age variables. Three models (neural networks, linear, and mixture) were compared based on their accuracy in predicting RD. We also examined whether the same or a different set of kindergarten-age factors emerge as the strongest predictors of reading fluency and comprehension difficulties across the models. Method RD were identified in a Finnish sample (N ≈ 2,000) based on Grade 9 difficulties in reading fluency and reading comprehension. The predictors assessed in kindergarten included gender, parental factors (e.g., parental RD, education level), cognitive skills (e.g., phonological awareness, RAN), home literacy environment, and task-avoidant behavior. Results The results suggested that the neural networks model is the most accurate method, as compared to the linear and mixture models or their combination, for the early prediction of adolescent reading fluency and reading comprehension difficulties. The three models elicited rather similar results regarding the predictors, highlighting the importance of RAN, letter knowledge, vocabulary, reading words, number counting, gender, and maternal education. Conclusion The results suggest that neural networks have strong promise in the field of reading research for the early identification of RD.