Predicting access to healthful food retailers with machine learning.

Many U.S. households lack access to healthful food and rely on inexpensive, processed food with low nutritional value. Surveying access to healthful food is costly and finding the factors that affect access remains convoluted owing to the multidimensional nature of socioeconomic variables. We utilize machine learning with census tract data to predict the modified Retail Food Environment Index (mRFEI), which refers to the percentage of healthful food retailers in a tract and agnostically extract the features of no access-corresponding to a "food desert" and low access-corresponding to a "food swamp." Our model detects food deserts and food swamps with a prediction accuracy of 72% out of the sample. We find that food deserts and food swamps are intrinsically different and require separate policy attention. Food deserts are lightly populated rural tracts with low ethnic diversity, whereas swamps are predominantly small, densely populated, urban tracts, with more non-white residents who lack vehicle access. Overall access to healthful food retailers is mainly explained by population density, presence of black population, property value, and income. We also show that our model can be used to obtain sensible predictions of access to healthful food retailers for any U.S. census tract.