The Built Environment and Predicting Child Maltreatment: An Application of Random Forests to Risk Terrain Modeling

Abstract

An estimated one third of children in the United States will suffer from maltreatment. The use of spatial predictive analytics offers an opportunity to delineate places at elevated risk of child abuse. Risk terrain modeling is a spatial analytic framework for predicting instances of varied types of crime. This article compares a random forest negative binomial model in a risk terrain modeling framework to the question of predicting counts of substantiated child abuse in Portland, Oregon. The final model specification includes domestic incident data from the Portland Police Bureau, built environment data from the City of Portland, OpenStreetMap data, and a neighborhood deprivation index derived from American Community Survey data predicting counts of substantiated child maltreatment from the Oregon Department of Human Services administrative data. The random forest outperforms the negative binomial model, showing its superiority in a risk terrain modeling framework, though the relative lack of predictive importance of the built environment variables compared to the domestic incident neighborhood deprivation variables should encourage researchers to further investigate the role of the built environment in the problem of predicting child abuse and maltreatment.