Evaluating and Extending Multivariate-Multistage Screening Models to Early Identify Children at High Risk for Developing Reading Difficulties

Multivariate-Multistage models developed to identify children who are at early risk for developing reading difficulties (RD) have been shown to accurately identify true-positives and true-negatives of RD risk while also producing acceptable rates of false-positives and false-negative. However, these models have been developed and tested on relatively small research populations with a rich diversity of child-level measures. In this presentation we extend the current models to population-level databases (e.g., Progress Monitoring Research Network collected by the State of Florida and the DIBELS Nationwide Database) with far fewer child-level measures to evaluate the potential of these screening models to accurately identify children in need of early reading intervention. The primary question addressed in this presentation is whether the current models are robust enough to accurately identify children who exhibit early risk for developing reading difficulties when only sparse data is available at the child level. We conclude that sparse data networks typically collected by schools often lack the information needed to allow accurate identification of children who are in need of early reading intervention using common model building techniques. However, we highlight and illustrate a new set of analytic procedures (e.g., machine learning, Bayesian models, constellation models, latent class models) that may allow sparse data networks to produce more acceptable identification rates for use with sparse data structures typically collected by schools.