Jun 25, 2020
How to Select Food Distribution Sites during the COVID-19 Crisis
Food insecurity has nearly doubled during the COVID-19 pandemic — skyrocketing unmet financial need and a rapidly changing food supply chain are exacerbating food access disparities. This has left food banks and distribution programs struggling to keep up with demand. And even when food distribution is available, limited public transit service has limited access for those without a car, disproportionately affecting residents who are Black, Indigenous, and/or people of color.
Organizations distributing food are making difficult decisions every day about how to increase access to distribution sites for those with the most need. Every city faces unique challenges and questions when choosing distribution locations: how dense the population is, where people of different races and income levels live, and where the people particularly vulnerable to COVID-19-related job loss live.
Our team has been working on defining and measuring equitable transportation access by learning from transportation decisionmakers, planners, community leaders, and activists in four metropolitan regions facing very different challenges: Baltimore, Lansing, Nashville, and Seattle. We partnered with Conexión Américas, a nonprofit organization based in Nashville serving the Latino immigrant community, to help identify where food distribution would have the most impact for those in need.
Conexión Américas, whose work before COVID-19 primarily focused on providing social, economic, and civil opportunities to Latinos in Tennessee, reprioritized their programming to rapidly respond to program participants’ food insecurity by establishing a drive-through food distribution site. By June 4, they had already distributed 3,156 boxes of fresh produce. The need was visible and growing, but for many residents, public transit wasn’t a viable option because of limited schedules and challenges carrying a cumbersome food box on the bus. This meant many community members without access to a car could not reach the site. Conexión Américas needed to identify new, walkable food distribution sites that reached as many food insecure residents as possible.
Our team partnered with Conexión Américas to identify data sources that reflected the Nashville-specific factors affecting food insecurity. This analysis had to be done as soon as possible to accurately reflect the gaps in current food distribution locations and respond to increasing resident needs. In this blog post, we outline our efforts to create a community-informed analysis that balanced nuance and rigor with timely and useful recommendations.
Identifying food distributions sites
To identify the neighborhoods with the most need in Nashville, we worked with Conexión Américas to determine which characteristics were likely predictors of food distribution needs and to identify locations where they could get food to the most people in need at once. We landed on seven characteristics:
● areas without easy access to current food distribution sites (as measured by walking time to closest existing site)
● number of estimated job losses because of COVID-19 per square mile
● number of noncitizens per square mile (because they are often ineligible for other services and safety net programs)
● number of low-income households per square mile
● number of households without access to a car per square mile
● number of single female–headed households with kids per square mile
● number of households Conexión Américas has served in the past per square mile (we assume these households likely live near others also in need)
We normalized each of these metrics to create measures of geographic density, rather than totals, to serve the largest population in a small, walkable area. Our unit of analysis was the census tract because it is the smallest geographic unit for which reliable data are available for all our characteristics. Our analysis focuses on Davidson County, which contains the city of Nashville. For the rest of this post, we’ll use “Nashville” to refer to our Davidson County analysis.
Collecting the data and calculating measures in a timely fashion
This analysis would only be useful if it could be done quickly to help inform food distribution when it is needed most. So we worked together to split up tasks and draw from other work already in progress.
To determine which places did not have access to other food distribution sites, our partner at Conexión Américas helped us identify a list of the main existing food distribution sites in Nashville, including food banks, emergency COVID-19 food distribution sites, and summer food-delivery locations through Metropolitan Nashville Public Schools. One of our team members then used the Urban Institute’s geocoder to assign latitude and longitude to the site addresses, and we then used the open source OpenTripPlanner routing software to calculate walking distance from the population-weighted centroid of each block group in Nashville to the nearest distribution site. We then averaged across all of the block groups in each tract to produce tract-level estimates.
The local expertise of Conexión Américas enabled us to identify more than 100 food distribution sites very quickly. Even still, our analysis is undoubtedly missing informal or smaller food distribution efforts — such as mutual aid efforts and food pantries at places of worship — that may not be widely published. But given time and resource constraints, we focused on food distribution efforts identified by Conexión Américas with publicly available location data and consider it an effective proxy for access to food distribution overall.
We were able to run the analysis for walking time to current food distribution sites quickly by building off our work from a related concurrent project, in which we are building a pipeline to conduct OpenTripPlanner routing analyses at scale using PySpark, the Python API for the distributed computing platform, Apache Spark. For that project, we are mapping access to opportunities via walking, car, and public transit for Nashville, so the infrastructure was already in place to do this quickly. The big data system we’ve built using Spark is critical to conducting large volumes of routing analyses quickly. To calculate the walking distances for this analysis, we submitted a routing request to the OpenTripPlanner routers API endpoint for each set of our start and end points. For reference, Davidson County has 473 block groups, so calculating the walk time between all combinations of block groups in the county for this analysis required calculated 223,729 unique walking paths. For our other work that covers a broader geographic area, the number of routing requests quickly grows to hundreds of millions. Using Spark, we can run multiple instances of the OpenTripPlanner API and complete routing requests in parallel, dramatically increasing the speed of our analysis.
From this analysis, we were able to produce this map of Davidson County:
Estimated job loss data come from the Urban Institute’s innovative job loss estimation tool, Where Low-Income Jobs Are Being Lost to COVID-19. This feature uses US Bureau of Labor Statistics estimates on employment change after the first wave of COVID-19 combined with American Community Survey five-year microdata to estimate local low-income net job change at the detailed industry level. These estimates are then rolled up to the two-digit North American Industry Classification System level and merged with Longitudinal Employer-Household Dynamics Origination-Destination Employment Statistics data to estimate job change by industry at the tract level. (For more information on how this data was calculated, see the technical appendix (PDF) for the feature.)
The data used in that feature are posted publicly on Github and the Urban Institute Data Catalog, which enabled us to quickly download the data and pull Nashville-specific information.
We were also able to quickly pull spatial demographic data at the census-tract level from the 2014–18 American Community Survey directly into R using the Tidycensus package. This package allowed us to compare our census data against different geographies, removed the need for any spatial joining or merging, and allowed us to quickly shift variable definitions if needed. Tidycensus saved our team hours of work downloading census data at different geographies and was crucial to keeping our shared files, workspace, and code clean and understandable for all team members.
We used the number of low-income residents (those earning less than 80 percent of the area median income) as our proxy for financial need. The Tidycensus package allowed us to quickly pull and sum all residents in median income brackets without downloading and transforming multiple spreadsheets.
Though this may sound smooth and straightforward, we did have to overcome some data challenges. For example, the data from Conexión Américas on how many households they served from 2019 through early June 2020 were at the zip code level, which doesn’t merge easily to the tract level (tracts are smaller than zip codes, and tract boundaries don’t match exactly with zip code boundaries). To make it even more challenging, census data are measured at Zip Code Tabulation Areas (or ZCTAs), not zip codes. ZCTAs are generalized representations of zip code areas constructed by the Census Bureau and will not always match zip codes. To translate from zip code to census tract, we ensured our zip code–to-ZCTA transformation correctly covered the Davidson County geography and our ZCTA-to-tract transformation, which was weighted by census tract population, added up to the correct number of Conexión Américas program participants. We also initially falsely assumed the zip codes of residents served by Conexión Américas would all be in Tennessee. After we realized many participants had out-of-county and out-of-state zip codes, we filtered these zip codes out and compared our pre- and postgeographic resident count information against accurate Davidson County participant counts.
Putting it all together into one index
Once we mapped out each of these measures, we combined them into a single index to identify the neighborhoods where Conexión Américas could distribute food to the most people in need. To make the index as easy to interpret as possible, we did not weight the measures, but rather created an evenly weighted average of Z-scores. For previous indices, we’ve gone back and forth on different kinds of weighted indices, and practitioners and policymakers always prefer the unweighted measures because they’re more intuitive. To make this calculation, we rescaled each variable to have a mean of 0 and a standard deviation of 1 so each of the measures are in comparable units. We then summed these Z-scores together and divided by the number of variables to get an average score to create our index.
Because some of our index scores were high outliers, likely because of their small geographies and the geographic normalization, we Winsorized the data at the 10 percent level at the top end (Winsorizing limits extreme values to reduce the effect of spurious outliers). We Winsorize by setting all data above the 90th percentile to the 90th percentile.
This set two of our index scores to the third-highest score level and addressed the outlier issue.
Finally, we mapped out each measure and the aggregate index using the urbnmapr and urbnthemes R packages. These packages allowed us to quickly create 10 clean, accurate, and publishable maps of every variable, the index, and preexisting distribution sites.
The map of our index below shows the neighborhoods located outside downtown Nashville and in southeast Davidson County had the greatest food distribution location need. Based on maps we created showing individual variables, we found that southeast Nashville, in particular, had census tracts with a denser population, larger concentrations of noncitizens, and disproportionate job loss because of COVID-19. Residents served by Conexión Américas were also more likely to be located in these neighborhoods than in other parts of the city. Though the travel time to the nearest current distribution was lower on average in these tracts, many residents did not have car access, and walking to current distributions locations could still take up to 30 to 45 minutes.
Conexión Américas is currently reviewing our individual variable and index maps, as well as a list of the 10 census tracts with greatest need. This analysis, coupled with their on-the-ground understanding of community need, will motivate their food distribution response strategy.
How other cities can do the same
Other organizations looking to identify food distribution sites can undertake a similar analysis by doing the following:
● determining which variables are most important to residents in their region related to food distribution
● normalizing each of these measures so they’re all in common units
● averaging them together into an index
● mapping them to identify ideal locations for food distribution
During a period of such great need, undertaking analyses quickly and using resources as efficiently as possible will help ensure as many people as possible who need resources can get them. We hope providing our methods and code examples here will help others do the same.
