Coronavirus Mobility Indicators

Much research has shown that measuring people’s mobility is a good way to predict trends in coronavirus infection spread. One great source of mobility measurements in the Google Mobility Report. This information for the entire country is based on Google Maps trip data. In its current form it reports on six different trip categories, based on destination type. In this research I gathered the Google Mobility Report data and analyzed it in a number of ways to try to correlate it with virus infection rates in Sonoma County. Later I will expand to include Marin county as a comparison.

Mobility

The mobility graph is based on Google Mobility Report data, which shows, for six different categories of trip, the daily percent change in mobility, measured by number of trips, showing departure in percent from a base period early in the year (January 2020) before the virus hit. The six trip categories are:

• Retail and Recreation (RR)
• Groceries and Pharmacy (QG)
• Parks and outdoors (PK)
• Transit Stations (TS)
• Workplaces (WO)
• Residential (HR)

Differences of Means from CA

To compare differences in the 6 mobility category means between the counties and California, and Pima county and Arizona, we use a t-test. On the left in each plot are the two categories which are usually higher in times of isolation and distancing. The other four are usually lower in times of isolation and distancing. Both Sonoma and Marin show much more park usage than California average. Sonoma is doing worse in the negative categories than California, whereas Marin is doing better. This shows the differences in behavior between Sonoma and Marin. In the HR category, Sonoma is doing slightly worse than California average, but nowhere near as well as Marin. Arizona is doing worse in all categories than California, and except for parks, Pima County and Arizona don’t differ much. Perhaps Arizona’s more wide-open geography accounts for some of the variation.

Sonoma and Marin Timeseries

• 

• 

It is useful to see a time series plot of the six mobility categories, with the means over the full time period data has been collected, and key dates shown.

• 

• 

• 

• 

Notice that in COVID-19 times in Sonoma County (green plots) mean RR is down 29% from normal, WO is down 28% from normal, and TS is down 35%. Grocery and pharmacy QG is down only just over 6%, and visits to HR and PK are actually up about 10% and 7% respectively, as one would expect when socially isolating and not able to do other things. People stay in their home and residence bubble, and go to parks (after the April and May ban). One can also clearly see how social isolation imposed much stricter controls than obtain now; the patterns clearly go way down, then move up (HR moves the reverse). The wildfires and smoke may also be imposing a late downward pressure on movements that are not isolating and distancing, and those would help with virus control. There are similar patterns in the Marin County (blue) plots. Note that we seem to be leveling out in October.

Observe that WO and HR series show a decided weekday-weekend pattern as we move closer to the current date. The HR pattern is further above base during the week, and nearer to the base on weekends; people behave more like normal on weekends but are staying home more than normal during the week. The WO pattern is the reverse. People are moving well below normal during the work days, and more like normal on weekends, when they would not be working anyway. A slight tendency to this pattern shows also in the TS graph, reflecting use of transit for work.

Mobility trends should give us an excellent way to estimate economic and social impact of the Coronavirus outbreak for the long term. For instance TS activity shows a 35% decline in Sonoma and a 40% decline in Marin, and they are holding level. From these facts, we should be able to estimate the shortfall in transit system revenues and trips, and perhaps even retail and recreation sales tax shortfalls.

Clusters derived from mobility

Our analysis found four clusters of data, which roughly align with four states of isolation and distancing during the virus: early, lockdown, transition, and current state. A bar graph of the six mobility categories of means, by cluster, clearly shows the differences between clusters we found. Each cluster is labeled with its 10th and 90th percentile dates, to eliminate any misclassifications, and the name I have given to the approximate time period: early (before social isolation rulemaking), lockdown (first restrictive set of social isolation rules), transition (not clear which cluster the point is in) and current (when further mobility was permitted).

• 

• 

• 

• 

The profiles of clusters are markedly different. An increase in HR (home and residence, red bar) traffic over the base percentages occurs after the early stage, most in the lockdown and successively less in the later stages. Parks and Outdoors (PK, yellowish bar) moved from above normal in early phase to drastically below normal in lockdown, slightly below normal in the pause, and much above normal in the current phase. The other four mobility categories were pretty normal in the early stage, drastically reduced in the lockdown stage, increasing slightly in the transition stage, and up more, though below normal, in the current phase. The mobility measures clearly show that we are at more risk in the current stage, and can expect reproduction rates for the virus.to be higher than they were in lockdown. Study shows that mobility increases are related, with about a 5-day lag, to R values. (McKinsey, 2020).

Clearly an increase in HR mobility increases social isolation. An increase in PK mobility does not really increase isolation, but does increase opportunities for social distancing. Grocery and Pharmacy (QG) and Retail and Recreation (RR) mobility yield less isolation, and decrease distancing over home, though the distancing can be more or less controlled by standards like six-feet distance and no indoor seating. Of course venues like bars and movie theaters offer little opportunity for distancing; both would be counted in the RR category. Workplace (WO) distancing may more or less be controlled depending on the employer. If you must go into work you are at the mercy of the job requirements. Some jobs such as meatpacking require close proximity; others (office work) have the chance to adjust proximity; work in open office carrels with a common ventilation system does not constitute much distancing; health care settings have mixed distancing possibilities. Transit Stations (TS) have limited ability to allow distancing; to do so will decrease service levels at peak times, thus making it difficult to support adequate social distancing. Cabs do not support distancing. Personal travel in your own car probably does till you must stop and interact with others. The number of people outside your ‘bubble’ of people known to be safe to you clearly constitutes a major source of risk, given asymptomatic and untested people in the general population.

Clusters through time

The following visualization gives a hint as to how the clusters match up with the time trends of the six trip categories.

• 

• 

• 

• 

Clusters match up quite well with time periods of social isolation practices. The numeric means and dashed lines on the charts show differences from the normal last January 2020 of current levels of travel in the six categories at present. This is consistent with observations in the time of the virus.

Mobility Category	Sonoma (current %)	Marin (current %)	Pima (current %)	Arizona (current %)
Restaurant and Recreation (RR)	-29	-32	-26	-22
Groceries and Pharmacy (QG)	-9	-10	-14	-11
Parks and outdoors (PK)	2	4	-34	-19
Transit Systems (TS)	-46	-32	-20	-24
Work (WO)	-29	-17	-15	-16
Home and Residence (HR)	11	8	6	5

Clustering can be visualized using two synthetic dimensions, created by principal components analysis; I call them isolation and distancing. The scales of the dimensions are in standard deviations of percent change since the base period. A positive score means that the percent change is above the mean of the dimension, and a negative score means that the percent change is below the mean for the dimension. Thus to the right of the graph is higher isolation, and to the top of the graph is more distancing.

In Sonoma County, Home and Residential (HR) travel is the major contributor to the positive isolation dimension scores. All other categories contribute negatively to isolation. Parks and outdoors (PK) contributes the most to distancing; Groceries and Pharmacy (QG) and Restaurants and Recreation (RR) make a slight positive contribution to distancing, while Travel Systems (TS) and Work (WO) make slight negative contributions to distancing.

In Marin County, Home and Residential (HR) travel is also the major contributor to the positive isolation dimension scores, but contributes negatively to distancing. All other categories contribute negatively to isolation. But Parks and outdoors (PK) makes a large negative contribution to distancing. Work (WO) contributes the most positively to distancing; Travel Systems (TS) also contributes positively. Groceries and Pharmacy (QG) and Restaurants and Recreation (RR) make a negative contribution to distancing.

In Pima County, Home and Residential (HR) travel is also the major contributor to positive isolation dimension scores, but contributes negatively to distancing. All other categories contribute negatively to isolation. But Parks and outdoors (PK) and Groceries (QG) make a large negative contribution to distancing. Work (WO) contributes the most positively to distancing; others are neutral for distancing.

In all of Arizona, Home and Residential (HR) travel is the major contributor to positive isolation dimension scores, and also contributes highly positively to distancing. All other categories contribute negatively to isolation. But Parks and outdoors (PK) and Groceries (QG) make a large positive contribution to distancing. Work (WO) contributes the most negatively to distancing; others are more or less neutral for distancing.

• 

• 

• 

• 

Mapping movements; KPIs for mobility

The isolation dimension for Sonoma county explains over 72% of the variation in mobility across all six categories; the distancing dimension explains another 11% or so. In Marin County, isolation explains 67% of the variation, and distancing explains 18%. In Pima County AZ, isolation explains 70% and distancing explains 16%, whereas in Arizona overall, isolation explains 68% and distancing 15%.

• 

• 

• 

• 

For Sonoma County, the early region therefore shows not much isolation and below normal distancing. The lockdown cluster shows high isolation without too much distancing. In the present and transition phases we are exhibiting relatively less isolation than lockdown, exactly what we observe in practice from the reports.

For Marin County, the early region shows not much isolation and just above normal distancing. The lockdown cluster shows high isolation above normal distancing. Marin was hit hard and people reacted to inIn the present and transition phases we are exhibiting relatively less isolation than lockdown, exactly what we observe in practice from the reports.

Similar observations can be made for Pima County and for Arizona at large.