Fig. 5

We start with a model that best describes the food environment at the census tract level. For each tract, we take the geographic centroid to represent the tract. We pull all the food supplier (restaurants, supermarkets, etc.) information from Yelp for each representative point. For each of the retailers in the previous step, we query Google Maps for actual driving and walking duration. We then remove all retailers that need a commute longer than 20 min from our calculations. At the next step, we fuzzy-match the retailer names compiling a list of the most frequent names. We ask five evaluators through Amazon Mechanical Turks to answer a questionnaire identifying the retailer type and the availability of different kinds of food in each retailer. Using all the information gathered about each point, we create 40 additional features that describe the retail food environment and as viewed from that representative point. Finally, we combine these features with another 1,000 features that describe other aspects of the census tract. We then build a model that predicts whether each of the centroids belongs to a “food desert”. (C) Emory University, reproduced under the CC BY-SA license