Food Desert Assessment: An Analytical Framework for Comparing Utility of Metrics and Indices; Case Study of Key Factors, Concurrences, and Divergences

Background: Low accessibility to healthy and nutritious food has been hypothesized to increase health disparities. Low-accessibility areas, called food deserts, are particularly commonplace in lower-income neighborhoods. However, indices for modeling food desert intensity are subjectively defined, and there is little agreement in the literature on their validity or relative strength. Methods: We first propose an assessment framework for objectively defining and comparing the utility of food desert indices using machine learning models. We introduce the concept of food desert index utility score, based on which we can compare the strength of indices for describing the food environment. We then explore the effect of the geographic spatial resolution of models and the impact of neighborhood-level income on the utility of food desert indices. Results: We observe the following: 1) Including healthy food suppliers only within a half-a-mile distance (as opposed to 1, 10, or 20 miles) from residents and discarding other food source variables leads to models of food environment with the highest utility scores. 2) The widely-accepted distinction between rural and urban areas in modeling food deserts adds little value to the utility score of the resulting models in our study region (Metro Atlanta). 3) The highest utility food desert index identified in our study region used a combination of access to food suppliers and the share of residents with low income and low access to vehicles. This approach achieved a utility score of 93% for identifying food deserts.