key: cord-164718-f6rx4h3r
authors: Wellenius, Gregory A.; Espinosa, Swapnil Vispute Valeria; Fabrikant, Alex; Tsai, Thomas C.; Hennessy, Jonathan; Williams, Brian; Gadepalli, Krishna; Boulanger, Adam; Pearce, Adam; Kamath, Chaitanya; Schlosberg, Arran; Bendebury, Catherine; Stanton, Charlotte; Bavadekar, Shailesh; Pluntke, Christopher; Desfontaines, Damien; Jacobson, Benjamin; Armstrong, Zan; Gipson, Bryant; Wilson, Royce; Widdowson, Andrew; Chou, Katherine; Oplinger, Andrew; Shekel, Tomer; Jha, Ashish K.; Google, Evgeniy Gabrilovich; Inc.,; View, Mountain; CA,; Health, Department of Environmental; Health, Boston University School of Public; Boston,; MA,; Surgery, Department of; Brigham,; Hospital, Women's; School, Harvard Medical; Policy, Department of Health; Management,; Health, Harvard T. H. Chan School of Public; Institute, Harvard Global Health; Cambridge,
title: Impacts of State-Level Policies on Social Distancing in the United States Using Aggregated Mobility Data during the COVID-19 Pandemic
date: 2020-04-21
journal: nan
DOI: nan
sha: 
doc_id: 164718
cord_uid: f6rx4h3r

Social distancing has emerged as the primary mitigation strategy to combat the COVID-19 pandemic in the United States. However, large-scale evaluation of the public's response to social distancing campaigns has been lacking. We used anonymized and aggregated mobility data from Google Location History users to estimate the impact of social distancing recommendations on bulk mobility among users who have opted into this service. We found that state-of-emergency declarations resulted in approximately a 10% reduction in time spent away from places of residence. Implementation of one or more social distancing policies resulted in an additional 25% reduction in mobility the following week. Subsequent shelter-in-place mandates provided an additional 29% reduction. Our findings provide evidence that state-wide mandates are effective in promoting social distancing within this study group.

Social distancing has emerged as the primary strategy for slowing the spread of the COVID-19 pandemic. Social distancing is expected to reduce the frequency of close contact with others and thus minimize the transmission of the SARS-CoV-2 coronavirus. Prior experience with the 2009 H1N1 influenza and Ebola suggests that social distancing is effective in reducing disease transmission. (1, 2) In China, officials engaged in an unprecedented quarantine of Hubei province to contain COVID-19 transmission out of the initial epicenter city of Wuhan. (3) (4) (5) As the pandemic spread to new clusters of infection in the United States (U.S.) efforts at containment and then mitigation have been largely at the discretion of state and local governments, leading to a patchwork of directives to encourage social distancing. These policies have included state emergency declarations, work-from-home policies, school closures, closures of non-essential businesses and services, limits placed on large social gatherings, bans on in-restaurant service, and shelter-in-place orders. (6) To date, there is some limited data that these policies may be working. Several news outlets have shown substantial drops in mobility across the nation since early March, although the ties to specific policy interventions are less clear. Given the current reliance on social distancing policies to limit the spread of COVID-19, systematically quantifying the impact of these policies on mobility may be helpful to public health officials.

The availability of anonymized and aggregated mobility data represents a novel opportunity to quantify the effectiveness of individual social distancing interventions. (7) We used aggregated (anonymized with differential privacy) data from Google users across the U.S. who have opted-in to Location History (the Location History feature is off by default and requires explicit opt-in). (8, 9) Using these data, information on the dates of implementation of various social distancing policies (see Supplemental Material), and a regression discontinuity approach, we sought to: 1) quantify the average effect on mobility of declarations of states of emergency, social distancing policies, and shelter-in-place orders, and 2) gain insights into which policies are most effective in promoting social distancing. Our primary outcome of interest is the relative change in the time spent away from places of residence, and also considered relative changes in the number of visits to places of work and the number of visits to: 1) grocery stores and pharmacies, 2) retail, recreation, and eateries), 3) parks, and 4) transit stops.

Overall, we observed three waves of state-level responses to COVID-19: 1) a first wave occurring during the first two weeks of March with state of emergency declarations, 2) a second wave during the week of March 16 where a variety of specific social distancing orders were implemented, and 3) a third wave during the last two weeks of March consisting of orders for residents to shelter in place orders ( Figure 1 ). The first state of emergency related to COVID-19 was declared by Washington State on February 29, and most recently by Oklahoma and Maine on March 15. Many states subsequently ordered that schools close (led by Louisiana and Virginia on March 13, 2020), and/or placed limits on specific activities and businesses in order to promote social distancing. Within a week, 48 states and Washington DC had implemented at least 1 social distancing policy. In 78% of states, the first social distancing order imposed was the closure of schools. On March 16, Nevada enacted orders advising residents to shelter in place, followed by California on March 19, 2020. As of April 5, 80% of states had ordered residents to shelter in place.

We used the same data that was used to prepare the Community Mobility Reports published by Google (8, 9) to compute aggregated visit trends to a number of specific categories of locations (see Supplemental Material). Trends were aggregated to the county level (including Washington, DC and independent cities that are not otherwise included in county boundaries), and available daily from January 1 through March 29, 2020. Within each county we applied a regression discontinuity analysis to estimate the relative change in the amount of time spent away from places of residence of Location History users and the relative change in the number of visits to specific categories of locations in the week after versus before each intervention. The following results should be interpreted in light of several important limitations including that our data are limited to Google users who have opted in to Location History. On average across the country, a declaration of a state of emergency was associated with a 9.9% (95% confidence interval [CI]: -10.1%, -9.7%) decrease in time spent away from places of residence, 11.4% (95% CI: -11.8%, -11.0%) fewer visits to the workplace, 11.5% (-11.7%, -11.2%) fewer visits to retail outlets and recreational sites, and 9.3% fewer visits to transit stops (-9.6%, -9.0%) in the following week ( Figure 2A ). These changes in mobility are noteworthy given that emergency declarations did not necessarily specifically call for increased social distancing and suggests that government messaging, news coverage, and/or actions observed in other countries could have influenced people's activities. Visits to parks were also affected by emergency declarations, with a small 3.5% (95% CI: -4.4%, -2.6%) average reduction. The smaller impact of emergency declarations on visits to parks versus other venues is likely at least partly explained by the warmer weather and transition to spring during this period. On the other hand, emergency declarations coincided with a relative increase in visits to grocery stores and pharmacies of 8.2% (95% CI: 7.9%, 8.5%), consistent with news reports of individuals stocking up on dry goods, cleaning supplies, and medications at the end of February and early March in anticipation of impending social distancing orders. (10) We next examined the impact on mobility of the first social distancing policies implemented in each state. We found that on average across the country these orders resulted in additional reductions in mobility above and beyond the changes observed following emergency declarations ( Figure 2B ). Specifically, implementation of one or more social distancing policies resulted in a further 24.5% (95% CI: -24.7%, -24.3%) reduction in time spent away from places of residence, a further 33.0% (-33.3%, -32.8%) reduction in visits to retail and recreational outlets, and a further 27.9% (-28.3%, -27.5%) reduction in visits to work in the following week. The same pattern was evident for visits to parks, and for visits to grocery stores and pharmacies. The impacts of social distancing orders varied substantially between states ( Figure 3 ). For example, implementation of social distancing policies was associated with a 36% decrease in the time spent away from places of residence in New Jersey versus a 12% decrease in Louisiana, but we note that differences in mobility between states may be due to a number of factors beyond social distancing policies. The median of state-specific changes in time spent away from places of residence was a decrease of 19%. States that enacted multiple social distancing measures tended to experience greater reductions in mobility. The impact of social distancing orders also varied substantially across counties within each state regardless of the average impact across the state ( Supplemental Table 1 ). Results were consistent when considering changes in visits to work, visits to grocery stores and pharmacies, visits to retail, recreation, and eateries, visits to parks, and visits to transit stops ( Supplemental Figure 1 ).

We next considered the impact on mobility of state-wide orders to shelter in place. Among the 7 states that had issued shelter in place orders on or before March 23, we found substantial reductions in time spent away from places of residence and in visits to all categories of locations ( Figure 2C ). Specifically, time spent away from places of residence was 29.0% (95% CI: -29.4%, -28.5%) lower in the week following implementation of shelter in place orders versus the prior week. Note that these changes are multiplicative over time as all states had already declared a state of emergency and implemented at least one social distancing policy. For comparison, we also show in Figure 2C the change in mobility during the same time frame (March 23-29 versus March 14-20, 2020) among those states that had not yet issued shelter-in-place orders by March 23. However, note that comparisons between states reflect the influence of a number of factors on mobility in addition to policy differences.

Given that most states enacted multiple policies to encourage social distancing over a short time period, it is not possible to estimate the independent effects of individual policies. However, in secondary analyses we sought to identify the combinations of social distancing orders that were associated with greater changes in mobility ( Figure These results should be interpreted in light of several important limitations. First, our source data are limited to smartphone users who have opted in to Google's consumer Location History feature. These data may not be representative of the population as whole, and furthermore their representativeness may vary by location. Additionally, these limited data are only viewed through the lens of differential privacy algorithms, specifically designed to protect user anonymity and obscure fine detail (8) . Moreover, comparisons across rather than within locations are only descriptive since these regions can differ in substantial ways besides the policy environment. Second, our analyses are focused on state-level policies, whereas individual metropolitan areas and counties within a state may have implemented specific social distancing policies prior to implementation of state-level policies (see Supplemental Figure 4 for detailed examples from King County, WA, Westchester County, NY, New York County, NY, and Santa Clara County, CA). Third, we did not assess the impact of social distancing policies or resulting changes in mobility on published counts of confirmed COVID-19 cases or deaths. Although such analyses are of critical importance, they are complicated by the heterogeneous availability of testing, delays in reporting, and the inherent latency before social distancing impacts are evident in clinical outcomes, and thus beyond the scope of this report.

In summary, using anonymized, aggregated, and differentially private data from Google users who opted in to Location History, we found that state-mandated social distancing orders were effective in decreasing time spent away from places of residence, as well as reducing visits to work, and visits to both grocery stores/pharmacies and retail/recreational locations. While the majority of states declared states of emergency by early March, the emergency declaration per se had only a modest effect on mobility. In contrast, implementation of one or more specific social distancing orders was associated with an almost 25% additional reduction in time spent away from places of residence and a 33% additional reduction in visits to retail and recreational locations. These effects were evident in every state and in virtually every county. Although we were unable to estimate the independent effects of different social distancing measures due to their close temporal proximity within each state, we did observe that those states that implemented multiple such measures experienced more pronounced declines in mobility. In addition, limits on bars and restaurants appeared to be the single most effective social distancing order.

We conclude that state-based orders intended to promote social distancing appear to be very effective in accomplishing the public health goals of encouraging individuals to stay at home in order to minimize the risk of COVID-19 transmission. Our findings not only illustrate the importance of specific social distancing orders, but also demonstrate the magnitude of change in mobility that we might expect from these policies. This information can help public health officials better calibrate and understand the extent to which social distancing can slow down the disease.

The anonymized and aggregated dataset analyzed herein was the same one that was used to create the publicly-available Google COVID-19 Community Mobility Reports (published at http://google.com/covid19/mobility on April 2, 2020). The data analyzed in this paper consisted of anonymized, aggregated, and differentially private counts of visits to places in different categories. The publicly available data reflects percentage ratios computed using these counts. The information on dates of policy interventions was aggregated from publicly available data as described in the Supplemental Material. 

Our overall approach was to use regression discontinuity using each county's recent past as its own control to assess the impact of state declarations of emergency and targeted social distancing policies on the relative changes in the average time spent away from places of residence, the number of visits to work, and the number of visits to: 1) grocery stores and pharmacies, 2) retail stores, recreational sites, and eateries, 3) transit stops, and 4) parks.

Data on state social distancing policies were obtained from official documents issued by state governors and health and education officials. Documents were linked from the Kaiser Family Foundation's State Data and Policy Actions Tracker (11) and supplemented with manual searches of state public health websites. Dates of policy enactment were cross-checked with the AEI COVID-19 Action Tracker (12) and the NYTimes Shelter in Place Tracker (13) . Policies tracked were categorized as follows: 1) state-declared state of emergency, 2) state-mandated school closures, 3) state-mandated closing of non-essential businesses and services, 4) state-mandated limits on large gatherings, 5) state-imposed bans on in-restaurant service, and 6) state-imposed mandatory quarantines. State-mandated closing of non-essential businesses included any order closing gyms, theaters, and other businesses even if it did not extend to all non-essential businesses. Limits on large gatherings referred to any ban on gatherings larger than a certain number of people, though that threshold varied between states. For states that issued additional orders reducing the size of permitted gatherings, the date of the first such order was taken. Bans on in-restaurant service excluded mandatory reductions in restaurant capacity and included only those orders that prohibited any restaurant activity except pick-up and delivery. These bans often also included bars and clubs. Mandatory quarantine referred to any stay-at-home or shelter-in-place order that prohibited non-essential travel away from the home for all residents. Shelter-in-place orders specifically for high-risk individuals were excluded. Orders that went into effect at any time after 12:00 pm were considered to begin on the following day.

We obtained aggregated and anonymized data from groups of Google users on mobile devices in all 50 states and Washington, DC who have opted in to having their Location History data stored. The anonymized dataset used for these analyses is the same as the one used to create the publicly-available Google COVID-19 Community Mobility Reports (published at http://google.com/covid19/mobility on April 2, 2020). The aggregation and anonymization process applied to the initial version of Google COVID-19 Community Mobility Reports has been previously described in detail. (8) The Community Mobility Reports leverage signals such as relative frequency, time and duration of visits to calculate metrics related to places of residence and places of work of Location History users as described elsewhere. (8) The anonymization process is designed to ensure that no personal data, including an individual's location, movement, or contacts, can be derived from the resulting metrics.

Data were aggregated to the county level (and Washington, DC) and available daily from January 1 through March 29, 2020. Within each county we applied a regression discontinuity analysis to estimate the relative change in time spent away from the place of residence (primary outcome) and the relative change in the number of visits to public locations (secondary outcomes) associated with: 1) declaration of a state of emergency, 2) ordering of one or more social distancing measures, and 3) orders for people to shelter-in-place.

We note that less populous counties are more likely to have days with missing data for visits to one or more categories of places (e.g., pharmacies) due to privacy filtering and other technical aspects. However, we believe that missing data has negligible effects on the state and national estimates provided because: 1) state level estimates are weighted by county population, and populous counties are extremely unlikely to have any metrics which fall below the limits of detection, and 2) the (unweighted) correlations between the relative changes in the combined metrics used in this paper (e.g., grocery stores and pharmacies considered together), compared to an unbiased, but lower-coverage alternative (e.g., grocery stores only) are very high (>0.95).

For each county we compared the value of each metric in the week after the date of implementation versus the 7-day period 9 to 2 days prior to the date of implementation. We included a two-day washout period prior to the implementation date given the public messaging that typically precedes implementation of these orders. Given data until March 29 were available, the effects of policies enacted on or before March 23 could be evaluated. The state level estimates reflect population-weighted aggregates of the county-specific estimates. The national estimates are a simple average of the state estimates.

Because each county and state is compared to its recent past these estimates are causally interpretable. However, comparisons across counties or states are only descriptive since locations can differ substantially in terms of the proportion of the population opted-in to Location History; the demographics of this group; the quality of the mobility data and of the Google Maps data about local establishments; and a number of other factors that may influence the observed changes in mobility beyond differences in the policy environment.

We performed a sensitivity analysis to assess the distribution of each outcome metric under the null hypothesis of the policies having no effect. We interpret the relative change across two periods of time (period 1: January 27-February 2, 2020; period 2: February 5-11, 2020), prior to the enactment of any state-level social distancing measures, as observations under the null hypothesis. During this time there were only 5 reported cases of COVID-19 in the U.S., which were isolated in Washington state. Although there was a small drop in time spent away from the place of residence across these pre-intervention periods, the much larger effects on mobility observed after each of the policies suggests that our observed effect during the exposure period is in fact related to the implementation of social distancing ( Supplement Figure 2 ). Figures 2 and 3 in the main text to gain a sense of how different these effects look when the first social-distancing measures were implemented. Under the null hypothesis of no policy-affect, the median observed change is slightly below -2%. However, during the first social-distancing measures, the median drop is nearly 10 times as large at -19%. For further context, all but two counties during the first social-distancing measures are below the null distribution median, and over 96% of counties are below the null distribution 2.5% percentile. A comparison between the changes in mobility during first social-distancing measures and changes in mobility under the null hypothesis for each metric is presented in Supplemental Tables 2.A and 2.B .

Our main analyses estimate the incremental mobility changes after versus before each of three waves of policy orders ( Figure 2 ). However, it is also of interest to quantify the overall effect of social distancing by comparing mobility at the end versus the start of March. We define the overall effect as the change in mobility from the week before the term "social distancing" started increasing in Google Search (March 1-7th) to the last week available in the data (March 23-29th). We see a significant decrease in all metrics. The ordering of the magnitudes is comparable to those in Figure 2B . As expected the overall magnitudes of the drops are larger than any of the incremental effects in Figure  2 . Table 1 : ANOVA sum of squares decomposition of the variation in the relative change on average time spent away from places of residence across counties for the linear model with the relative pre-post difference in time spent away from the residence as the outcome and state as the only independent variable. The results show that approximately half of the variance in the outcome is explained by differences between states and the remaining variance is largely explained by differences across counties within states. Figure 2 .A Relative changes in national averages from January 27-February 2, 2020 to February 5-11, 2020 for all metrics of interest. The results are also shown in Supplemental Table 2 .A. This plot can serve as a reference point for Figure 2 in the sense that these periods occurred before any orders were issued and can be interpreted as observations under the null hypothesis of no effect of policy interventions .

Supplemental Figure 2 .B : Relative changes in county averages from January 27-February 2, 2020 to February 5-11, 2020 for time spent away from the residence. Counties are grouped by state to show the heterogeneity across counties and states before any orders were issued. The changes can be interpreted as observations under the null hypothesis of no policy effect because these periods occurred before any orders were issued

Population mobility reductions associated with travel restrictions during the Ebola epidemic in Sierra Leone: use of mobile phone data

Characterizing the epidemiology of the 2009 influenza A/H1N1 pandemic in Mexico

The effect of human mobility and control measures on the COVID-19 epidemic in China

An investigation of transmission control measures during the first 50 days of the COVID-19 epidemic in China

Effective containment explains sub-exponential growth in recent confirmed COVID-19 cases in China

Governmental Public Health Powers During the COVID-19 Pandemic: Stay-at-home Orders, Business Closures, and Travel Restrictions

Aggregated mobility data could help fight COVID-19

Google COVID-19 Community Mobility Reports: Anonymization Process Description

Panicked Shoppers Empty Shelves as Coronavirus Anxiety Rises

Kaiser Family Foundation, State Data and Policy Actions to Address Coronavirus

COVID-19 Action Tracker

See Which States and Cities Have Told Residents to Stay at Home

Supplemental Figure 1 : Effect of first social distancing order on visits to places of work (A) , visits to grocery stores and pharmacies (B) , visits to retail, recreation, and eateries (C) , visits to transit stops

(A) and overall average national effects of social distancing orders relative to period prior to social distancing awareness

We are grateful to Giorgia Abeltino, Matthew Abueg, Skip Addison, Elizabeth Adkison, Maha Afifi, Gerald Agapov, Ahmet Aktay, Putri Alam, Jan Antonaros, Neha Arora, Harry Askham, Boris Babenko, Arturo Bajuelos, Avi Bar, Davi Barbosa, Sean Barclay, Pinal Bavishi, Sherry Ben, Ashish Bora, Aleksey Boyko, Michael Bringle, Sander Bruens,

 Figure 4 : Timeline of change in average time spent away from places of residence in King County, Washington (i.e. Seattle area) (A), Westchester County, New York (B), New York County, New York (i.e., Manhattan) (C), and Santa Clara County, California (i.e. San Jose area). Colored boxes denote the declaration of a state of emergency, the implementation of the first county-level social distancing order, the implementation of the first state-level social distancing order, and county and/or state-level orders for residents to shelter in place. The height of each box corresponds to the change in average time Location History users spent away from places of residence in the week before (plus a 2-day washout period) versus the week after each policy date.