To predict and explain. How prediction improves our understanding of models: an application to the study of teacher bias

Out-of-sample prediction is not often applied within educational research, although it can complement existing methods in important ways. Prediction gives an intuitive measure of a model's (in)ability to structure an outcome of interest, and complements the aggregate statistics usually obtained from typical in-sample methods. In this paper, I illustrate the potential of prediction through the study of teacher bias in tracking in the Netherlands. I show how the use of prediction identifies misspecification in the simple interval-model often estimated in the field, and can be used to obtain insights when estimating less interpretable, albeit more appropriate models. Substantively, I find that girls are positively biased in tracking net of observed ability, while students of low parental education are negatively biased. Importantly, the latter effect may have been structurally under-estimated in prior work. I also identify the school level to be a more substantial source of bias than student-level demographics, lending further support to calls to study school-level heterogeneity in tracking. My findings further accentuate the risks involved in tracking and fall broadly in line with increasing calls to re-evaluate the Dutch tracking system.