key: cord-275281-baxst5an authors: Dimke, C.; Lee, M. C.; Bayham, J. title: Working from a distance: Who can afford to stay home during COVID-19? Evidence from mobile device data date: 2020-07-26 journal: nan DOI: 10.1101/2020.07.20.20153577 sha: doc_id: 275281 cord_uid: baxst5an As local and state governments reopen parts of the economy while balancing public health through social distancing, it is important to understand the heterogeneity in how the population has reacted to the COVID-19 pandemic. We match census block group level Safegraph mobile device data with demographic data from the American Community Survey to identify trends amongst different subgroups of the population. We find evidence that people's ability to work from home is a determinant of time spent at home since the beginning of the pandemic. On April 15th, census block groups classified as being better able to work from home spent 3 more hours at home compared to those who were not. We see supporting trends amongst block groups with differences in income and educational attainment. The extent to which people reduce potentially infectious contacts is a function of economic conditions because people employed in essential positions may not have the flexibility to work from home. Baker et al. (2020) estimates that 18.4% of the US population works in occupations where they are exposed to 5 COVID-19 at least once per month. These jobs tend to pay lower wages and are disproportionately held by minority populations, or people with lower educational attainment (Mongey & Weinberg, 2020) . Additionally, households at lower percentiles of earnings experience larger drops in income during recessions (Heathcote et al., 2009) . The dependence on this income makes it difficult for 10 individuals to choose to stay home. Occupation is a key determinant of who can stay home during the pandemic response. Dingel & Neiman (2020) identify which occupations have high and low ability to work from home. Workers in high-personal-proximity occupations with low-work-from home ability are less likely to have college degrees and less 15 likely to be white. They are more likely to have below median income and more likely to work in small firms (Mongey & Weinberg, 2020) . These small firms are less likely to remain open after crises (Mongey & Weinberg, 2020) . While useful, this static analysis does not indicate who has and who has not been distancing during this pandemic. Here we propose a simple method 20 for parsing anonymized mobile device location data by sociodemographic characteristics using publicly available Census data. Our method yields up-to-date estimates of time spent at home across demographic groups, a classification unavailable using mobile device data alone. Our analysis extends the work of Jay et al. (2020) , who document heterogeneous mobility by income quintiles, 25 by evaluating education levels and occupations with the ability to work from home. 2 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 26, 2020. . https://doi.org/10.1101/2020.07.20.20153577 doi: medRxiv preprint Our objective is to estimate social distancing by socioeconomic and demographic characteristics. We merge Census Block Group (CBG) level mobile de-30 vice data from SafeGraph along with demographic data from the US Census Bureau's American Community Survey (ACS) (U.S. Census Bureau, 2016). Safe-Graph (www.safegraph.com) aggregates anonymized mobile device data that can be used to understand movement patterns during the COVID-19 epidemic. To enhance privacy, SafeGraph excludes census block group information if fewer 35 than five devices are observed on any day. The CBG level data is granular, but also preserves device anonymity. There are over 200,000 census block groups in the US with an average population of just over 1500 (U.S. Census Bureau, 2016). We classify each CBG based on the composition of the population along the following characteristics: education, 40 household income, and occupations with ability to work from home. Specifically, we identify CBGs with a majority of the population in one category. The education classification is based on two levels of education, those who have a Bachelor's degree or higher and those who do not. The household income classification is based on three income brackets: $0-50,000, $50,000-100,000, 45 and greater than $100,000. We assigned groups to having high ability to work from home or low ability based on the classification in (Dingel & Neiman, 2020) . Specifically, we multiply the fraction of workers in each occupation that are likely able to work from home (Dingel & Neiman, 2020) by the number of people employed in each occupation, and sum this product across all individuals in the 50 CBG. We classify each census block group as dominated by a category of interest if the fraction of the population in one of those categories exceeds 51%. Not all census block groups have a dominant population, so these estimates are based on a subset of the sample. We run the analysis for thresholds between 50% 55 and 80%. While the ordinal rank of time spent at home remains constant, the sample size falls as the thresholds are increased resulting in larger standard error 3 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 26, 2020. . https://doi.org/10.1101/2020.07.20.20153577 doi: medRxiv preprint estimates. Non-binary categories remain qualitatively stable between 40% and 70%. We implement our social distance parsing approach via regression of the 60 following form, where Y it is the vector of median time spent at home in CBG i on day, t, and X it is the matrix of categorical dummies in a given regression. This model does not include an intercept, so β t is interpretable as the group mean on day t. In contrast to the pre-post approach used in Bushman et al. (2020) , this 65 analysis allows us to see trends based on actual CBG data instead of hypothetical differences. We estimate β t for each day from January 1 to June 23, 2020. We can apply this methodology to any demographic variable available from the ACS. We omit three demographic variables from our report to succinctly describe results related to who can work from home. These variables are gender, 70 age, and race, and they influence our understanding despite their omission here. These omitted categories are correlated with education, income, and employment status. We choose to keep this analysis separate rather than including controls because of barriers to identification and a recognition that such an analysis is biased. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 26, 2020. . https://doi.org/10.1101/2020.07.20.20153577 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 26, 2020 . . https://doi.org/10.1101 We explore the heterogeneity of this response along education, income, and ability to work from home (Figure 1 ). We find that those with Bachelor's degrees or higher, household incomes greater than $100,000, and a greater ability to work from home spent significantly more time at home relative to the rest 100 of the population. On April 15th, the initial peak of the COVID-19 response, people in occupations with higher ability to work from home spent approximately 3 more hours at home than those in occupations without the ability to work from home. As states and municipalities have reopened, people in all sociodemographic groups are spending less time at home. 105 6 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 26, 2020. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 26, 2020. . https://doi.org/10. 1101 This simple approach to parsing distancing metrics has its limitations. It is possible that the population of device users in a CBG does not align with the category that it has been assigned in our analysis. Mobile device ownership is unequally distributed across society with younger people more likely to have smartphones and access to the internet. In 2018, 95% of individuals ages 18-34 110 in the U.S. had smartphones while only 67% of individuals older than 50 owned smartphones (Silver, 2019). Our classification method may be more robust to these effects since we focus on more homogeneous CBGs. However, more research is needed to understand the extent to which mobile device users are representative of the population at large. Our analysis has several implications for social policy during the COVID-19 response. Those who are able to work from home spend significantly more time at home relative to their less able counterparts. If this at home work is as productive as the work that they performed under pre-COVID-19 circumstances, 120 there is little to be gained by this portion of the population returning to work as usual. Employers can complement public health policy by working hard to accommodate at risk populations. We acknowledge support from Amazon Web Services Diagnostic Develop-125 ment Initiative. We thank SafeGraph (www.safegraph.com) for generously providing data. Baker, M. G., Peckham, T. K., & Seixas, N. S. (2020). Estimating the burden of united states workers exposed to infection or disease: A key factor in contain-130 ing risk of COVID-19 infection, . 15 , e0232452. URL: https://journals. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 26, 2020. . https://doi.org/10. 1101 Effectiveness and compliance to social distancing during COVID-19 How many jobs can be done at Unequal we stand: An empirical analysis of economic inequality in the united states Neighborhood income and physical distancing during the COVID-19 pandemic in the u Characteristics of workers in low work-from-home and high personal-proximity occupations L. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted July 26, 2020.