Mapping our Food System
The USDA ERS publishes an atlas that maps the United States food system.
It is a fascinating resource that allows for deep dives into which parts of our country have access to food, and which parts of our country lack access to fresh food. These ‘food deserts’ are an oft-mentioned indicator of the health of our local food systems, and are defined as places with
“Limited access to supermarkets, supercenters, grocery stores, or other sources of healthy and affordable food may make it harder for some Americans to eat a healthy diet. There are many ways to measure food store access for individuals and for neighborhoods, and many ways to define which areas are food deserts — neighborhoods that lack healthy food sources.” — USDA
As impressive a resource as the interactive map on the USDA website is, I was curious if I could dig a little deeper into the underlying data and understand a bit more about the trends associated with US food deserts. Where were they? What characteristics did they share? What were the key drivers that predicted whether an area was a food desert?
The Process
All of the USDA data is publicly available on their website to download as raw spreadsheets. Survey data is available at the county (more general) and census tract (more granular) levels. I combined the data and added in geographic markers, so as to perform analyses and prepare the data for mapping.
Using a variety of statistical methods, I trained rudimentary machine learning models to predict whether an area was considered a food desert. If the model flagged one of the randomly sampled census tracts as a food desert, it collected all of associated characteristics available in the USDA. When the model was completed, it reported all of the most relevant characteristics that were used to best predict whether a given census tract was classified as a food desert. This methodology gave me insight into how the USDA classification was made, and delivered some interesting results about the underlying trends on food deserts in the US.
The Results
According to my most simple and intuitive model, the most significant factor that would determine whether a census tract is a food desert is if that area is Urban or Rural. I was surprised to see that Urban areas are the best predictor that the population lacks food access. This could be attributed to a number of factors, including not having enough stores or markets in densely populated areas, having higher populations in Urban areas, and having higher income levels in Urban areas overall and thus a higher poverty level (USDA defines poverty as income relative to the overall income level in a given area).
Other findings of note are that census tracts clustered in the Southwest US, California, and Colorado had the strongest effects on an area being considered a food desert, and census tracts clustered in New York, Florida, and Illinois had the strongest effects on an area not being considered a food desert.
Further Study
This analysis sets up a good foundation for transforming the raw USDA data into predictive models and maps of US food deserts. There is more work to be done in understanding the links between status as a food desert and community health statistics, and local food economies.
