Predictive Modeling for Food Accessibility: Machine Learning Insights into Food Deserts and Swamps

Abstract

This thesis builds upon previous work that utilizes machine learning with census tract data to predict the percentage of healthful food retailers in a tract. Instead of using a modified Retail Food Environment Index (mRFEI) that refers to the percentage of healthful food retailers in a tract for classifications - this work focuses on taking a more granular approach based on data scraped from online food delivery websites. We process our data using the Nutrient Rich Food Index (NRF9.3) to calculate a Food Retailer Composite Score (FRCS) for each retailer, which corresponds to its relative cost and nutritional value. Using FRCS scores and the number of retailers in the vicinity of a tract, we label each tract as either a "Food Desert", "Food Swamp", or "Food Oasis" utilizing k-means clustering with k=3. After which we apply Random Forests to predict these labels using public demographic and economic data from the US Census Bureau for each tract. Our model detects food deserts, swamps, and oasis with a prediction accuracy of 79% - showing a reasonable improvement to previous work predicting at 72% accuracy. We find that College No Degree, Poverty Rate, Property Value, and Median income are the most important features for accurate predictions. This work highlights the need for further research on the links between food quality in a region and overall health outcomes in order to better identify locations at risk to target with policy suggestions.