Multilevel model versus recurrent neural network A case study to predict student success or failure revisited

Despite the success of machine learning in the past several years, the advantage of machine learning algorithms over simpler models when structured, tabular data are used has been debated. We outline a concrete example by revisiting a case study on predictive monitoring using educational data. In their work, the authors compared the performance of a simple regression model, a multilevel Bayesian regression approach, and a recurrent neural network to predict end-of-year grades. They found that the Bayesian multilevel model generally outperformed the significantly more complex machine learning method. Analyzing the data used in the original study, we revisit the models and showcase similarities and differences in variable selection and subgroup performances. We outline case characteristics that may contribute to the close performance of neural networks and multilevel models in the present case study. This article highlights the significance of tailoring the choice of model to the application and underlying data structure in process monitoring.