Are we good at sheltering in place? And is it working? A deeper look at new data from Google

Recently, Google released Mobility reports based on Google maps data, that show changes in different types of activities such as shopping and park visits during shelter-in-place. This data provided an opportunity to understand how we are doing and if what we are doing is working.

Below, I describe a combined analysis of two data sources: Google’s mobility reports and COVID-19 deaths in different states. There are several goals for the analysis: understanding 1) to what degree the public is compliant with sheltering in place 2) where can we be more compliant 3) the relationship between compliance and the spread of the virus.

Some caveats to this analysis. 1) There is limited information on how mobility rates were computed and what exactly was included in each category for each state 2) coronavirus cases started growing at different time in different states and 3) shelter-in-place was announced on different dates. So get your grain of salt and hold on to it.

Are we compliant?

First, I took the Google data for US and chose the states that had more than 100 fatalities as of 4/4/2020 (CA, CT, FL, GA, IL, IN, LA, MA, MI, NJ, NY, PA, TX, WA). Google data includes 6 categories of mobility (this info is copied directly from their report):

• Retail & recreation – restaurants, cafes, shopping centers, theme parks, museums, libraries, and movie theaters
• Grocery & pharmacy – grocery markets, food warehouses, farmers markets, specialty food shops, drug stores, and pharmacies
• Parks – national parks, public beaches, marinas, dog parks, plazas, and public gardens
• Transit station – public transport, hubs such as subway, bus, and train stations
• Workplaces – places of work
• Residential – places of residence

Out of those categories we expect the first 5 to go down and the last one, Residential, to go up. Let’s see if that’s true.

The “box” plot above shows the summary stats for the top affected states combined (see above). The blue boxes show the 1st and 3rd quartiles and the whiskers show the min and max values. We see the greatest average reduction (>50%) in mobility in Retail & recreation and transit stations, not surprisingly so as many of these destinations and systems have closed to the public (e.g. the light rail by our house and Disneyland).

Grocery and pharmacy shows the lowest median reduction (-26%), which makes sense as these are considered essential services and many still shop in person for both food and medicine.

Workplace surprisingly only showed 39% reduction. I expected it to be higher given most of us are not working in essential services. There wasn’t a lot of variability in this category, so the lower reduction wasn’t caused by states that started sheltering-in-place later. This can be happening because 1) essential services account for a larger share of workers 2) there is a some type of bias in the maps data in terms of places of work.

The most interesting one was the Park category. As you can see in the plot, it has the biggest spread out of all categories. Three of the states Pennsylvania, Michigan and Indiana actually increased their park related mobility (by 7, 15 and 25 % accordingly)! Whereas in Massachusetts park mobility decreased by 56%. This might be due to the local regulations during sheltering in place. For example, in Pennsylvania parks are remaining open, while in Massachusetts many cities closed parks and playgrounds down.

Residential mobility increased as predicted, but by a surprisingly similar and small amount in all states (median 13%). It’s hard to tell why, not knowing exactly how these numbers are computed, but it’s fair to assume the amount of walking and running around the house someone can do doesn’t change much across the states.

The figure above shows average scores for all mobility categories for each state. As you can see, not all states are equally compliant. Massachusetts is heading the way with 53% mean reduction in mobility with New York and New Jersey following closely. Indiana, Georgia and Pennsylvania are trailing behind with less than 35% reduction. So there is still room for improvement in many states, especially in Park category. The next question is….

How does compliance affect COVID-19 spread?

To tackle this question, I took the data on deaths in each state and fitted exponential functions to this data. Why deaths? I consider deaths to be a more reliable gauge because they are not as dependant on testing as the total number of positive cases. After fitting exponentials, I took the estimated % of daily growth in deaths for each state. For example for California it is 21% daily growth at the moment.

Now what we can do, is correlate the daily growth rate with the change in mobility to see if there are any relationships. I correlated all of the six mobility categories and the mean values with daily death growth rates. Only Retail & recreation showed a significant correlation as shown below. All the other mobility types (and the mean mobility change) did not have a significant correlation with death growth rate.

The correlation coefficient was r=0.6 (the closer to 1 the more correlated are the data points) and was significant (p=0.02 both Pearson and Spearman). The interesting part is that the relationship was the opposite of what you’d expect. Higher compliance (more reduction in mobility) was correlated with higher daily growth in deaths.

Wait so what is going on? Does this mean that sheltering in place is not helping? I don’t think so. It looks like the effect here goes the other way. The mortality growth might be affecting public policy and compliance with sheltering in place. States that have higher growth rates, might be enforcing the rules more strictly and the public might be more compliant in these states as well. This is likely just the beginning, and the relationship will change as time passes. It might even out and eventually as we see compliance grow overtime in a given state, we should also see a reduction in mortality growth.