Predicting Low Food Access: An Empirical Assessment of Machine Learning Algorithms and Socio-Demographic Predictors

Chronic disease is pervasive throughout the United States, with over 50% of adults having one or more preventable chronic illnesses relating to diet and/or physical activity. An understudied cause for poor nutrition is food insecurity. This study utilizes the 2019 Food Access Research Atlas (FARA) in order to assess what sociodemographic features are most correlated with food insecurity nationwide and train 5 predictive models for predicting whether a census tract has 'low access' to food within a geographic range defined by the USDA. The results showed that tree-based machine learning models (XGBoost, Random Forest) returned the highest raw accuracies of 76.30% and 75.89% respectively, and that the attention based tabular model TabNet did not outperform traditional approaches. An examination of feature importances also indicated that the level of White population, number of houses without registered vehicles, and urban/rural status of an area were the three most significant features for classifying census tracts as food insecure.