45 – Innovative Methods for Education Research
A Data Mining Approach to Value-Added Models
Sharon Lohr
Arizona State University
Arturo Valdivia
Arizona State University
Many policymakers propose linking administrative decisions related to teachers to their effectiveness. Since teacher effectiveness cannot be measured directly, many researchers use value-added models to apportion changes in student achievement to the teachers and schools who have taught them. Several models, based on a linear mixed model framework, have been proposed and used for assessing value added by teachers: students' test scores are used as outcome variables and teachers' contribution is treated as random effect. The value-added score for a teacher is the empirical best linear unbiased predictor in the linear mixed model. However, the linear mixed model formulation has certain limitations, among them, its rigid structure. To address this, we use random forest. We introduce new variable importance measures and also use the existent measures to rank teacher effects. In addition, comparisons of traditional linear mixed model and random forest results are presented. We show that the random forest results may be more accurate when the linear model is misspecified. It is possible to use this approach as a complementary tool to linear models.