What can we do to help local governments "flatten the curve"?We realized we could move quickly to apply the Real World Graph® data engine that we’ve developed over the past five years to analyze real-time data in a way that would accurately measure how effective various social distancing strategies perform in the real world. By focusing the geospatial human mobility insights our technology can surface, we created this pro bono Social Distancing Scoreboard as the first of many tools we are developing for a Unacast COVID-19 Toolkit — designed to provide high-quality insights to public agencies, healthcare organizations, local governments and businesses to enable them to learn and act in the best interest of at-risk populations and the general public.
Our Initial Approach
In developing a social distancing score that most valuable to organizations, especially those unfamiliar with human mobility data, we started with a generalized score that takes into account relevant underlying metrics — each of which explains one facet of social distancing behavior.
For our first iteration, we explored:
- People dwelling at home vs dwelling outside their homes (a proxy for how good an area is at "shelter-in-place" behavior)
- Changes in average time spent in and around home, aggregated over time (a proxy for how much time people spend at home versus other venues)
- Change in dispersion of activity clusters, or how many people no longer gathered in the same location at the same time (a proxy for change in number of encounters)
- Change in average distance traveled
Hitting a Moving Target
Speed and accuracy were the two most important requirements for the first iteration; we aimed to provide users with the most useful information possible in the quickest way possible.Our team quickly realized that there was a second underlying challenge: as the country's behavior changed, so did the underlying data. The models, which we had built and optimized towards a non-COVID-19 world, now need revisiting. For example, our previous model for detecting place of residents was geared towards high confidence that an assigned home is actually a device's home. However, with many people moving to be in another area (e.g. parents' houses, getaway cabins, etc.), the original home assignment model might no longer work well.
Additionally, due to the collection methods of app partners, our methodology for normalizing fluctuating supply levels also needs adjustment.
Identifying the Strongest Signals
After testing each metric, we found the change in average distance traveled worked best as a starting point:
- The metric correlates well with the number of confirmed cases: the more cases are confirmed, the greater the decrease in the average distance traveled on the county level.
- It requires no strong assumptions like an assumed home location.
- It works independently of supply side fluctuations — we get the most signals when people are moving so the metric is unaffected by changes in the ping frequency due to inactivity when people dwell at home.
The metric (while admittedly simple) captures how people adapt their everyday behavior in a few significant areas:
- Switching to home office strongly reduces the daily travel distance.
- Avoiding non-essential trips to entertainment places or spare-time facilities strongly reduces the travel distance.
- Canceling travels and vacations will strongly reduce travel distance.
Continuous Improvement & Refinement
Travel distance is one aspect, but of course people can travel far without meeting a soul or travel 50 feet and end up in a crowd — so we know that the real world picture can be quite complex. As noted above, changing behavior will trigger adjustments in our data strategy. That's why, post launch, we will be continuously working to improve our social distancing models.
For example, as of this writing we are in the process of understanding the best way to add layers that capture more of the complexity of social distancing: exploring how a change in the number of encounters for a given area, as well as a change in the number of locations visited, contribute to an area's social distancing score. Since many important social distancing metrics can be derived from an accurate understanding of residences, we are also dedicating resources to adjusting our home assignment algorithms so that they can better identify new common evening locations.As we better understand our data and how it describes the new world under COVID-19, we will continuously update our social distancing score as well as develop other datasets that can be used in the fight against COVID-19.