Machine learning and public policy: Early detection of physical violence against children

Abstract

Physical violence against children is a widespread and grossly underreported phenomenon with substantial short and long-term negative consequences. In Latin America and the Caribbean, 43% of children under the age of 15 experience corporal punishment at home, yet reporting rates are alarmingly low. This paper aims to demonstrate how household data can be considered for a future predictive analytics model in Argentina. Based on the 2019–20 MICS survey we apply machine learning techniques to predict physical violence against children (understood as physical discipline) at the household level in Argentina. The scope and potential benefits of using predictive models in this context are assessed, as well as the main limitations and risks. The results suggest that, by analyzing the situation of the 30% of households with the highest risk scores, 43 out of 100 households in which children experience physical violence could be identified at an early stage.