key: cord-1025754-dyqgc75k authors: Shearston, J. A.; Martinez, M. E.; Nunez, Y.; Hilpert, M. title: Social-distancing Fatigue: Evidence from Real-time Crowd-sourced Traffic Data date: 2021-03-08 journal: medRxiv : the preprint server for health sciences DOI: 10.1101/2021.03.04.21252917 sha: 9c2a514fd6678d0e97ac7d1a1032f65c4fb2c603 doc_id: 1025754 cord_uid: dyqgc75k Introduction: To mitigate the COVID-19 pandemic and prevent overwhelming the healthcare system, social-distancing policies such as school closure, stay-at-home orders, and indoor dining closure have been utilized worldwide. These policies function by reducing the rate of close contact within populations and results in decreased human mobility. Adherence to social distancing can substantially reduce disease spread. Thus, quantifying human mobility and social-distancing compliance, especially at high temporal resolution, can provide great insight into the impact of social distancing policies. Methods: We used the movement of individuals around New York City (NYC), measured via traffic levels, as a proxy for human mobility and the impact of social-distancing policies (i.e., work from home policies, school closure, indoor dining closure etc.). By data mining Google traffic in real-time, and applying image processing, we derived high resolution time series of traffic in NYC. We used time series decomposition and generalized additive models to quantify changes in rush hour/non-rush hour, and weekday/weekend traffic, pre-pandemic and following the roll-out of multiple social distancing interventions. Results: Mobility decreased sharply on March 14, 2020 following declaration of the pandemic. However, levels began rebounding by approximately April 13, almost 2 months before stay-at-home orders were lifted, indicating premature increase in mobility, which we term social-distancing fatigue. We also observed large impacts on diurnal traffic congestion, such that the pre-pandemic bi-modal weekday congestion representing morning and evening rush hour was dramatically altered. By September, traffic congestion rebounded to approximately 75% of pre-pandemic levels. Conclusion: Using crowd-sourced traffic congestion data, we described changes in mobility in Manhattan, NYC, during the COVID-19 pandemic. These data can be used to inform human mobility changes during the current pandemic, in planning for responses to future pandemics, and in understanding the potential impact of large-scale traffic interventions such as congestion pricing policies. The COVID-19 pandemic threatens almost every single country on Earth. 1,2 To curb the 62 pandemic and prevent overwhelming the healthcare system with COVID-19 cases, social-63 distancing policies such as school closures, non-essential business closures, curfews, and stay-at- 64 home orders have been put into place. Social-distancing policies, specifically quarantines, are 65 some of the oldest and most utilized strategies for epidemic control. 3 Social distancing policies 66 function by reducing physical contact and decreasing human mobility; adherence can 67 substantially reduce transmission 4 and case counts. 5 Thus, quantifying human mobility and 68 social-distancing compliance, especially at high temporal and spatial resolution, is critical for 69 controlling the epidemic. Several measures of social-distancing compliance have been used previously in the literature. Studies evaluating individual-level adherence to social-distancing policies frequently use self-73 report through online surveys. 6-8 While these methods are very useful for evaluating 74 determinants of adherence, they are not as useful for real-time epidemic monitoring because they 75 are time-and labor-intensive and only capture a snapshot of the population's behavior in a small 76 sample at a single timepoint. In contrast, highly resolved data such as mobile phone call records 77 or locations can be used to measure social-distancing 9 and to inform mobility components of 78 infectious disease transmission models. 10 These measures, especially if publicly available, can 79 contribute to epidemic monitoring. One such option is vehicular traffic condition smartphone 80 apps, which base their maps on cellphone movement data. Traffic congestion maps are publicly 81 available, updated in real-time, and can be used for public health research. For example, we have 82 previously shown that colors in Google traffic maps correspond to relative vehicle speed, and 83 used this information to infer traffic-related air pollution. 11 App-based traffic information offers a pathway for assessing social-distancing, as increased 86 traffic congestion is indicative of spending time outside the home that may result in opportunities 87 for human interaction and contact. In fact, traffic data has been used to evaluate mobility during 88 the pandemic in South Korea, finding that in some cities increases in traffic were correlated with 89 increases in cases, but that in others traffic and cases were negatively correlated. 12 Here we 90 analyzed time series of traffic congestion that we derived from app-based maps. We quantify 91 changes in traffic in the Manhattan borough of New York City (NYC) during the COVID-19 92 pandemic (January 1 to December 31, 2020). Our objective was to describe how social 93 distancing interventions impacted population-level patterns of mobility throughout various stages 94 of the pandemic. From January 1 through December 31, 2020, we obtained 12 tiles from Google traffic maps to 100 view Manhattan's entire street network every three hours in real-time. We used image processing 101 methods, as described previously in Hilpert et al., 11 to identify the color-coded road segments. colors are proxies for vehicle speed. 11,13 In summary, 11 map pixels for each of the four colors 105 were counted and time series generated of the percent of the map area covered by each color. To investigate changes in traffic congestion during the pandemic, we first characterized all four 108 time series, but then restricted further analysis to the time series of red traffic congestion given 109 its clear pandemic-related signal (e.g., decrease in April) and simpler interpretability (increases 110 in red coverage indicate increases in congestion). We did this to avoid redundancy in analysis 111 since all color coverage time series showed correlated temporal patterns. There were substantial 112 changes in traffic congestion over the course of the pandemic and we captured this variation by 113 partitioning the time series into four distinct time periods, which we refer to as COVID periods. 114 We individually fit a generalized additive model (GAM) 14 to each COVID period. GAM models were constructed to predict the time series of percentage of total map area covered 117 by traffic congestion, ( ! ) using two discrete predictor variables: hour of day, ℎ ! , which can 118 assume values 0, 3, 6, 9, 12, 15, 18 and 21; and the binary variable ! indicating whether a 119 traffic map describes traffic on weekdays ( ",! ) or during weekends ( $,! ). We fitted the 120 following model to the measured ! , ",! , $,! and ℎ ! data: where % is the intercept (representing mean color coverage), and " and $ are cyclic cubic 125 regression splines with 8 degrees of freedom and zero mean. Degrees of freedom were selected 126 using the generalized cross-validation criterion. The fitted models were then used to predict 127 congestion levels hourly for a weekend and weekday, with 95% confidence intervals. Due to the 128 morning and evening rush hour, traffic congestion displayed a 24-hour periodicity. The range of 129 the diurnal fluctuation was measured as the deviation from the intercept. To investigate potential social distancing fatigue (increased mobility before relaxation of stay-at- missing observations (n = 54) were imputed using predictions from the GAM models. All analyses were conducted in R version 3.5.1 (R Foundation for Statistical Computing, Vienna, 139 Austria). Simon Wood's mgcv package was used to fit GAMs. 16 STL analysis was completed 140 using the `stl` function in the stats package. We found abrupt decreases in traffic congestion in Manhattan following school closures and the Table 1 . A map of the study area is shown in Figure 1 , with colors representing the crowd-149 sourced traffic data. There was a dramatic decrease in traffic congestion after the implementation 150 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) 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 March 8, 2021. where a more stable traffic pattern emerged. All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Morning and evening rush hour impose a periodic cycle to traffic patterns, seen as daily peaks 188 and troughs in traffic (Figure 2) . The pandemic not only caused changes in traffic trends, but 189 also drove changes in the period cycle. In addition to the daily periodicity, there was also a 190 weekly cycle where congestion was elevated during the five consecutive weekdays and was 191 dampened during the two consecutive weekend days. The weekly cycle was most apparent in the 192 Pre-COVID period and in the most recent period, COVID Period 3. Mean percent area with red traffic congestion changed dramatically throughout the time period 195 (Table 1) . Percent area with mean red traffic congestion (equivalent to the model intercept) was 196 highest during the pre-COVID time period, and then decreased abruptly during COVID Period 1 197 (from a mean of 0.99% to 0.41%) before steadily increasing for COVID Periods 2 and 3. By 198 COVID Period 3, the mean red traffic congestion area had rebounded to about 75% of the pre-199 pandemic average. All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The rebound in red congestion coverage began far in advance of the implementation of Phase 1 207 reopening (Figure 3, second panel) . In STL analysis of COVID Periods 1 and 2, congestion 208 appears to increase from approximately April 13 th onward, while Phase 1 reopening did not In addition to these changes in traffic trends, the pandemic substantially altered daily traffic 220 periodicity, such that during the height of the pandemic in NYC in the Spring, there was little 221 differentiation between weekday and weekend traffic patterns (Figure 4, tan lines) . During the 222 clear bimodal distribution with peaks around 9am and 5pm, and weekends a clear unimodal peak 224 around 5pm. However, during COVID Period 1, both weekday and weekend traffic peaks were 225 greatly dampened, and the bimodal weekday distribution shifted to nearly unimodal, becoming 226 very similar to the weekend distribution. During COVID Period 2 and 3 the daily traffic peaks 227 were greater than for Period 1, but still lower than pre-pandemic levels. Even as overall traffic 228 increased during these periods, the weekday distribution remained altered, such that the morning 229 peak was much smaller than the evening peak. A closer inspection of the GAM splines (Supplemental Figure 1) , including the deviation of the 239 fitted spline from the intercept, further highlight these differences in diurnal traffic congestion. In All rights reserved. No reuse allowed without permission. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. In this paper, we use changes in traffic as a measure of human mobility and an indicator for Out study has some notable limitations. First, we include a fairly limited geographic area (the 317 borough of Manhattan, NYC), and thus results of our study may not be generalizable to other 318 areas. However, studies in other cities also find dramatic decreases in traffic, on a similar scale 319 as that reported here. 18, 19, 21 Second, our data source does not allow for disaggregation of 320 passenger vehicles and trucks, which were likely differentially impacted by responses to the 321 pandemic (as trucking operations may be considered essential businesses). Studies in NYC 27 and 322 Somerville MA 20 have found decreases in both vehicles and trucks, though the decrease was 323 substantially less for trucks. Third, we do not disaggregate traffic changes by neighborhoods, and 324 thus do not report variation in congestion changes within Manhattan. We recommend that future 325 studies describe and evaluate these potential differences. Using highly temporally resolved, crowd-sourced traffic congestion data, we describe changes in 331 traffic congestion in Manhattan, NYC, during the COVID-19 pandemic. We report dramatic 332 declines in traffic congestion during the initial stages of the pandemic and implementation of 333 stay-at-home orders, followed by rebounds in congestion nearly two months before stay-at-home 334 orders were reversed, evidence of social-distancing fatigue. Additionally, we describe changes in 335 diurnal traffic congestion patterns for weekdays and weekends, by hour, for four time periods 336 during the pandemic. This data can be used to inform human mobility changes during the current 337 All rights reserved. No reuse allowed without permission. (which was not certified by peer review) 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 March 8, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 pandemic, in planning for responses to future pandemics, as well as in understanding the 338 potential impact of large-scale traffic interventions such as congestion pricing policies. Substantial undocumented infection facilitates the rapid dissemination 354 of novel coronavirus (SARS-CoV-2) Lessons from the history of quarantine, from plague to influenza A Quantifying the impact of physical distance measures 359 on the transmission of COVID-19 in the UK Effect of Timing of and Adherence to 361 Social Distancing Measures on COVID-19 Burden in the United States : A Simulation Modeling 362 Approach Experience with Social Distancing Early in the 364 COVID-19 Pandemic in the United States: Implications for Public Health Messaging. medRxiv Social Distancing Compliance under 367 COVID-19 Pandemic and Mental Health Impacts: A Population-Based Study Working memory capacity predicts individual differences in social-370 distancing compliance during the COVID-19 pandemic in the United States Social connections with COVID-19-affected areas 373 increase compliance with mobility restrictions On the Use of Human Mobility Proxies for Modeling 375 Epidemics A new approach for inferring traffic-related 377 air pollution: Use of radar-calibrated crowd-sourced traffic data The relationship between trends in COVID-19 prevalence and traffic 380 levels in South Korea No reuse allowed without permission. (which was not certified by peer review) 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 Wood SN. Fast stable restricted maximum likelihood and marginal likelihood estimation of 386 semiparametric generalized linear models Effect of Road Traffic on Air Pollution. Experimental Evidence 390 from COVID-19 Lockdown Implications for air quality management of 392 changes in air quality during lockdown in Auckland CoV-2 epidemic Reductions in traffic-related black carbon and 395 ultrafine particle number concentrations in an urban neighborhood during the COVID-19 396 pandemic Traffic Impacts of the COVID-19 Pandemic: 398 Statewide Analysis of Social Separation and Activity Restriction Effectiveness of non-pharmaceutical interventions on COVID-19 401 transmission in 190 countries from 23 Socioeconomic disparities in subway use and COVID-19 404 outcomes in New York City. medRxiv Impacts of COVID-19 on public transport ridership in Sweden: Analysis 406 of ticket validations, sales and passenger counts New York becomes first city in US to approve congestion pricing. The 412 Guardian The Effects of the COVID-19 Pandemic on Transportation Systems in 414