Estimating short-range traveling using daily commute volume.

Daily commute volumes count the total number of home to work trips originating from a given county within 24 hours. Cuebiq Inc. provides a list of obfuscated “personal areas” for each user. Observations geolocated from within these locations are deemed to be coming for either the home or the work location of the individual and are therefore up-leveled to preserve user privacy. That is, these coordinates are aggregated to the centroid of the Census block group level that each observation falls into. In order to quantify the changes in commuting behavior to and from work, we classify personal areas into the home or work location to be able to count commute flows. In particular, we consider the most commonly-visited personal area during nighttime hours (9:00pm—5:00am) as the home location of the user, while the most common non-home personal area visited during daytime hours (9:00am—5:00pm) is classified as the work location of the user. This method is imperfect (i.e., it may obfuscate users who exclusively work night shifts), but it is based on assumptions about the typical worker in the United States. Then, one commute is defined as a user visiting their “home” and “work” in a given day. Lastly, in this study we take as reference the definition and location of personal areas as identified in the period immediately prior to the lockdown measures. Therefore, our commute metric reflects changes with respect to the status quo existing prior to the COVID-19 pandemic.

Estimating mobility range using the radius of gyration.

The radius of gyration [31] characterizes the extent of a given user’s trajectory in a single day and its formal definition is: (1) where n is the user’s number of observations on that day, is the i^th observed position of the user, i = 1, 2, …, n, and is the center of mass of the trajectory. This measure gives us a standardized way to tell how far an individual is traveling from their average daily position (center of mass), most likely their home and work locations, in a given day. I.e. a larger radius of gyration corresponds to a trajectory with positions that are further away from the person’s center of mass. The radius of gyration provides us with a way to measure the “characteristic distance” [31] traveled daily by each individual and how their spatial range changed during the different phases of the pandemic. From an epidemiological standpoint, this measure is particularly relevant as reductions in people spatial ranges will reduce the rate at which an epidemic will diffuse between (possibly distant) locations. Lastly, in order to compute typical mobility within a given region, we sum the total daily radius of gyration that by users in that region.

Estimating long-range traveling using inter-CSA transit.

To study the changes in long-range traveling, we calculated the number of users who visited at least two separate U.S. Census Combined Statistical Areas (CSAs) in a single day. In particular, CSAs are defined so that each CSA groups together neighboring areas that share significant employment interchange and that show a substantial degree of economic or social connection between them, often measured by commuting patterns. In other words, we measure the volume of long-range trips between major metropolitan areas which allows us to capture—among other things—also variations in air traffic and long-range train and road trips.

Estimating social mixing using daily contacts.

For the purpose of contact tracing, the CDC defined a close contact as someone who was “within 6 feet of an infected person for at least 15 minutes” [32]. Using this guidance, we operationalize the definition of contact as two devices being within the same rectangular area of approximately 38m × 20m for at least 15 minutes and we define two measures quantifying contact mixing between individuals. Even though defining contacts in this way can be noisy or imprecise due to the spatial resolution considered, we show in the following that the measures introduced in this section positively correlate with key epidemiological indicators, e.g., new deaths.

The method works as follows. Each time the time-stamped spatial coordinates of a user are recorded, we associate the longitude-latitude coordinates to an 8-character geohash. A geohash is a short string of letters and digits that allows to encode coordinates into a hierarchical spatial data structure that tessellates the world surface into a grid. In our case, we consider geohashes at an 8-digits resolution which encode rectangular cells of dimensions that are approximately 38m × 19m at the equator [33]. We define two users to be co-located if they are observed in the same geohash for at least 15 consecutive minutes. For each user, we compute the number of unique users that a device is in contact with during a single day and the total dwell time of their daily contacts that fit the above criteria. We average these values across users in a given region to arrive at county, state, and nationwide daily average contact duration and average number of distinct contacts. Note, this method can record only contacts occurring outside of personal areas due to Cuebiq’s privacy protection procedures that obfuscate the exact coordinates for location events occurring within personal areas (e.g. users’ home and work locations).

Typical behavior.

In this manuscript, we report the metrics as percent of typical activity. We select the period between January 16 and February 28, excluding holidays, as our baseline, therefore defining what constitutes, in the context of this work, typical behavior. For each measure, we divide the indicator daily value by its average value during the baseline period for the same day of the week (i.e., Mondays are compared to the average Monday during the baseline period). Therefore, values of 100% denote typical behavior. For example, the timeline of the percent of typical behavior for the commuting indicator at time t will be computed as where C_t denotes the number of commutes observed at time t, N_t denotes the number of active users at time t, denotes the number of commutes in the baseline reference period, and denotes the number of active users in the baseline reference period.

Commuting, working from home, and unemployment.

Commuting volume decreased dramatically in early March, and, in most states, did not increase in the same way that the other mobility/contact measures have. This is likely due to several reasons, from the historic waves of unemployment in the spring and early summer, to a dramatic increase in teleworking. Indeed, throughout the COVID-19 pandemic in the U.S., we see a strong negative correlation between the percentage of teleworkable jobs and commute volume in metropolitan statistical areas (MSAs). For public health officials planning for future pandemics, this relationship between telework and commute volume is especially insightful; many of the NPIs that have been introduced throughout the pandemic were designed to limit the amount of workplace infections, and as such, reductions in commute volume that are due to increased telework (as opposed to increased unemployment) could illustrate a preferable balance of economic and public health priorities.

During the COVID-19 pandemic, many employers eliminated in-person interactions, although many jobs in the United States cannot easily transition to remote work [28, 29, 38]. Furthermore, it is difficult to quantify the ubiquity of remote work across the United States and various surveys have been conducted trying to estimate this number [39]. In addition, typical commuting patterns were also impacted by an unprecedented surge in unemployment [40]. Therefore, we explore the relationships between commuting, unemployment, and teleworking by combining our commute metric with two additional data sources: the Bureau of Labor Statistics’s Local Area Unemployment Statistics (LAUS) dataset, which provides monthly estimates of county-level unemployment rates; and Dingel and Neiman’s and Dey et al.’s estimates of the proportion of jobs in an area that can be feasibly worked from home [28, 29]. The teleworkability estimates are at the Metropolitan Statistical Area (MSA) level, and the LAUS data at the county level. For this reason, in our analysis, we aggregate our measurements to the MSA level, excluding rural counties. Note that while local unemployment and commute volume vary month-to-month, the estimated proportion of teleworkable jobs is largely a static quantity.

In Fig 3, we present the relationship between commute volume, unemployment, and teleworkability in February, April, and July, with each month corresponding to a distinct phase of the pandemic. In February—the baseline period for which we use to define “typical” activity—there is no correlation between the percent of typical commutes in MSAs and the percent of teleworkable jobs, nor is there a correlation between commuting and unemployment at that point (Fig 3a and 3d). This is expected, as the United States had not yet experienced large disruptions resulting from the COVID-19 crisis. Then, in April, following the updated guidelines about physical distancing and the massive surges in unemployment, commuting volume dropped substantially across the United States (on average, about 40% of baseline levels, see Figs 1a and 3b). During this lockdown period, the percentage of jobs that can transition to telework show a -0.533 correlation with commute volumes (Fig 3b); we also observe a significant negative correlation between commute volume and unemployment rate during this period (Fig 3e).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Unemployment, teleworking and commuting patterns.

Grouping county-level employment data to the Metropolitan Statistical Area (MSA), we correlate commute volume with the percent of jobs that can readily transition to teleworking according to Dingel and Neiman [28] (top row) and unemployment rate (bottom row) over time. (a & d): February, during the baseline period; (b & e): April, during the peak lockdown; (c & f): July, after unemployment declined but commuting remained low—during the “new normal” phase).

https://doi.org/10.1371/journal.pdig.0000430.g003

Collective physical distancing in rural and urban areas.

We also observe different levels of collective physical distancing in different parts of the country, which reflects the heterogeneity in policy response, disease incidence, geography, and population structure across the US (see e.g. [41]). By grouping our collective physical distancing measures with the National Center for Health Statistics (NCHS) urban-rural county classification scheme [42], we can compare the responses of people living in urban versus rural settings. We observe that large central metro (code 1) areas showed the largest reductions in our collective physical distancing measures for contacts and mobility range (Fig 4a–4c); the more rural the county, the less reduction from typical behavior. However, users living in more urban counties also had lower baselines for these measures (Fig 4d–4f), which we show using a standardized index as opposed to the “percent of typical” values shown in Fig 4a–4c. As a point of comparison, by the beginning of April, the median reduction of the number of distinct contacts for users in large/medium metro counties approached a level similar to the baseline of a typical user in rural, micropolitan counties (Fig 4f).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Differences in county-level behavior based on rural-urban codes.

Each county in the United States is assigned a rural-urban code, ranging from 1 (large central metro) to 6 (highly rural, “non-core” counties). We average the percent of typical behavior per user (top row) and a standardized index (bottom row) across counties grouped by these six rural-urban code designations. The standardized index obscures raw values but preserves relative differences between groups; we do so by normalizing by the median value across all counties. (a & d): mobility range; (b & e): contact duration; (c & f): distinct contacts. Seven-day rolling averages are plotted in bold above raw values plotted as thin curves.

https://doi.org/10.1371/journal.pdig.0000430.g004

More-urban areas also began physical distancing behaviors earlier than more-rural areas; for example, large central and large fringe metro areas dipped to 80% of typical for their contact measures about five days earlier than small metropolitan areas, micropolitan areas, and non-core areas. During the first week of May, micropolitan and non-core counties showed mobility range that was around 75% of typical, while large central and fringe metro areas remained at around 50% of typical (Fig 4a). A similar rural-urban gap is seen in the percent of typical behavior for both measures of contacts (Fig 4b and 4c).

Collective physical distancing and the toll of COVID-19.

Lastly, we validate the use of the proximity/contact measures introduced in this manuscript as coarse-grained approximations for true person-to-person contacts. As such, we would expect to find a positive correlation between these measures and key epidemiological indicators, such as new deaths. More specifically, we would expect that a lagged correlation would best capture this relationship since it accounts for: a) the time from exposure to symptom onset (about 6 days); b) the median number of days from symptom onset to death (between 13 to 17 days depending on the age group considered); and c) the median number of days from death to reporting date (varying from 19 to 21 days depending on the age group) [43]. Because of these factors, the median delay we would expect in the correlation between our proxies for contacts and new deaths is in the range 38–44 days. Nationwide, we see that this is indeed the case.

In Fig 5, we plot the nationwide percents of typical average contact duration (a) and distinct contacts (b) against the daily number of new deaths per 100,000 people nationwide. The contact measures are correlated at a delay d ∈ [37, 44] days for each measure, which was selected by maximizing the correlation between contact patterns and new deaths. We do the same comparison for the number of new infections in the S4 Text (S11 Fig), highlighting the robustness of this correlation nationwide.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Collective physical distancing and new deaths.

Here we correlate daily contact measures nationwide with new reported deaths [30] between April 30 and November 5, 2020. The horizontal axes correspond to the percent of typical contact patterns, while the vertical axis corresponds to the (lagged) number of new deaths per 100,000. (a) Average contact duration (b) Distinct contacts. A lag of d days was selected for each state so as to maximize the correlation between new deaths and contact measures. Maximum correlation is observed at d ∈ [37, 44] (d = 44 is visualized) days that is consistent with CDC estimates [43] that account for disease dynamics and reporting delays. In each subplot, darker colors indicate later dates and marker size corresponds to an estimate of the median effective reproductive number (R_t) across all 50 states and District of Columbia (source: rt.live). These contact measures are also positively correlated with new reported cases (but at a shorter lag, see S11 Fig).

https://doi.org/10.1371/journal.pdig.0000430.g005

The color of the markers in Fig 5 corresponds to time; darker colors indicate later dates. Here, we see an important relationship between our contact measures and the course of the pandemic. Namely, as contact patterns increased in the early summer (lighter colored markers), new infections and new deaths followed; this, in turn, was followed by decreases in contact events, followed again by decreases in new infections by late August (darker colored markers). This is approximately the same time as when the curves in Fig 1b (the contact measures) started to level off, while mobility and inter-city transit continued to rise (Fig 1a). What this disconnect between mobility and contact patterns suggests is that our collective social behavior can reduce the rate of new infections and, as a result, new deaths. This finding is possibly trivial to epidemiologists and public health officials, but it is nonetheless important for our understanding of how our collective behavior impacts the trajectory of a pandemic, to validate our contacts measures as proxies for true person-to-person contacts, and it is also consistent with other findings throughout the literature on COVID-19 [11, 44–46]. The ability to measure these patterns in almost real time shows the potential benefits of using mobile device data in forecasting (or “nowcasting”) the trajectory of a virus, and moving forward, they present a baseline for our collective behavioral response to future pandemics.