Statistical Learning for Predicting School Dropout in Elementary Education: A Comparative Study

School dropout is a significant challenge for the education system. This phenomenon is present in different environments, modalities, and stages of education. In the Brazilian scenario, despite advances in some respects as a reduction of indexes, combating evasion is still one of the significant efforts. Identifying the factors that involve school dropout is supported by different decision support techniques such as Statistical Learning. Statistical learning consists of a method set for exploring and understanding data to establish an association between explanatory and response variables and develop an accurate model. We propose to examine the use of some regression methods commonly used in the Statistical Learning literature for estimating school dropout in the context of elementary school from the state of Pernambuco. The data involves educational indicators, and we defined phases in the study to understand, prepare, and model the data. For prediction, we apply models for estimating school dropout using kernel-based and linear regression methods. We measured the performance by the prediction error from the test data set using Mean Absolute Error and Root Mean Square Error. We considered Statistical tests to confirm the results. The findings show that kernel-based models are effective alternatives to provide greater precision in the estimation of school dropout in scope studied. The reason to explore more accurate predictive models is supporting intervening and targeting the most at-risk students of scholar dropout. The study provides knowledge about the applied scenario supporting policies to mitigate the problem.