Distorção idade-série no Brasil: uma abordagem preditiva com modelo XGBOOST

Abstract

Age–grade distortion constitutes one of the main indicators of educational vulnerability. This study analyzes the risk profiles for age-grade distortion among children and adolescents aged 7 to 17 attending school in Brazil. Age-grade distortion is defined as a delay of two or more years between the student's chronological age and the appropriate grade. The research uses microdata from the 2024 IBGE Continuous PNAD, the XGBoost (eXtreme Gradient Boosting) machine learning model was employed, comparing its performance with the logistic regression model. Socioeconomic, demographic, territorial, household infrastructure, food insecurity, and social benefit variables were included in the model. The analysis revealed the predictive superiority of XGBoost over logistic regression, showing higher accuracy and sensitivity in identifying students at risk. Through SHAP (SHapley Additive exPlanations) values, age was identified as the main predictor, with risk accentuating from the age of 13. Factors such as labor market insertion, male gender, and low education level of the head of the household also proved to be critical determinants. The study demonstrates that machine learning algorithms outperform traditional methods in capturing complex patterns, providing a robust diagnosis to support school retention policies.