Our goal was to define regions that maximized mobility within clusters and minimized mobility between clusters. We also aimed to reflect some practical constraints when defining regions, such as the need for sufficient population size. We established the following criteria to reflect these goals:

1. Regions would not cross state lines.

2. Regions would not subdivide counties.

3. Each state would have at least two, but up to four regions.

4. Each region would have a population of at least 100,000.

5. All counties in a region would be geographically adjacent.

Population mobility data were obtained from SafeGraph, a company that aggregates anonymized location data from numerous applications to provide information about places people visit. SafeGraph reports the number of mobile device visits between all census block groups (CBG) by day in the U.S.* We used data from the period from January 01, 2021 to April 16, 2021—a time when mobility was impacted by COVID-19, but to a lesser extent than earlier phases of the pandemic.

We created mobility informed clusters for each state by taking the daily number of mobile device visits by CBG and aggregating this count up to the county level to determine a daily device count for every origin-destination county pair within each of the six Rocky Mountain West states. This count includes the number of origin county devices that were seen in the destination county, and the number of destination county devices that were seen in the origin county. The higher the device count, the more connected the two counties were deemed to be.

We wrote a clustering algorithm in R using capabilities from the igraph package to determine counties’ contact with each other. We utilized the usmap package to visualize how counties clustered with each other given certain constraints on cluster numbers, as well as the dendextend package to create dendrograms (which depict all the counties of a state and their connectedness to each other in a large tree-style diagram).

Cluster map using Utah as an example:

cluster map

We used this information to resolve regional configurations for each state that comprised at least two but no more than four regions, choosing the configuration with the maximum number of regions that conformed to our five criteria. Once this configuration was chosen, we created a dendrogram to go with it.

Utah corresponding dendrogram (below). Red dashed lines indicated the boundaries of the original mobility informed clusters determined by the algorithm. Dendrograms read like a family tree—the longer and more far back the branches between two counties go, the less connected they are.

dendrogram

To see the code and data used for this clustering analysis, click here to visit our GitHub page!

*To enhance privacy, SafeGraph excludes census block group information if fewer than two devices visited an establishment in a month from a given CBG.

We compared the preliminary clusters generated in Step 2 to existing public health planning regions for each state. If the mobility informed clusters did not align with these regions, we identified and assessed discordant counties, and investigated the dendrograms to determine whether to keep the counties in their mobility informed clusters or reallocate them to better match with public health planning regions.

For example, Utah is broken down into the following public health planning regions:

planning

The counties in our mobility informed clusters that did not agree with the public health planning regions were Piute, Sevier, and Wayne.

discordant

We investigated the dendrogram (shown in step 2) and determined that these three counties were not connected enough with the rest of their mobility informed cluster to require them to remain in the cluster. As a result, we reallocated these counties to better match with the public health planning regions.

utah final