key: cord-0905219-qhffzj72 authors: Deforche, K.; Vercauteren, J.; Müller, V.; Vandamme, A.-M. title: Behavioral changes before lockdown, and decreased retail and recreation mobility during lockdown, contributed most to the successful control of the COVID-19 epidemic in 35 Western countries date: 2020-06-23 journal: nan DOI: 10.1101/2020.06.20.20136382 sha: eb2b08e0b51eeaba4e140bfba35a022b23f82c60 doc_id: 905219 cord_uid: qhffzj72 The COVID-19 pandemic has prompted a lockdown in many countries to control the exponential spread of the SARS-CoV-2 virus. This resulted in curbing the epidemic by reducing the time-varying basic reproduction number (Rt) to below one. Governments are looking for evidence to balance the demand of their citizens to ease some of the restriction, against the fear of a second peak in infections. More details on the specific circumstances that promote exponential spread (i.e. Rt>1) and the measures that contributed most to a reduction in Rt are needed. Here we show that in 33 of 35 Western countries (32 European, plus Israel, USA and Canada), Rt fell to around or below one during lockdown (March - May 2020). One third of the effect happened already on average 6 days before the lockdown, with lockdown itself causing another major drop in transmission. Country-wide compulsory usage of masks was implemented only in Slovakia 10 days into lockdown, and on its own reduced transmission by half. During lockdown, decreased mobility in retail and recreation was an independent predictor of lower Rt during lockdown, while changes in other types of mobility were not. These results are consistent with anecdotal evidence that large recreational gatherings are super-spreading events, and may even suggest that infections during day-to-day contact at work are not sufficient to spark exponential growth. Our data suggest measures that will contribute to avoiding a second peak include a tight control on circumstances that facilitate massive spread such as large gatherings especially indoors, physical distancing, and mask use. indication and quantification of the lockdown (Suppl 1). 1 From incidence data of deaths and diagnosed cases in 35 countries (ECDC data down-2 loaded on 3 June 2020, see Suppl 2), we estimated the change in transmission over time, while 3 differentiating between the effect of measures that preceded the lockdown, and the effect of the 4 lockdown itself. Since incidence data of deaths is considered more reliable than of diagnosed 5 cases (which is also highly dependent on testing policy and test availability), we estimated suit-6 able values for the dispersion parameters by estimating variance from the fits and then use a 7 slightly more cautious value (Suppl 3). For case data we then simply decreased the value to 8 give a higher weight to death data, favoring better fits of death data over case data. Only data 9 up to 60 days into the lockdown (up to about mid-May in European countries) were used to 10 avoid the confounding effect of the gradual lifting of restrictions, which started in many Eu-11 ropean countries in May. Given that time between infection and reported death was estimated 12 as 24 (95% IQR 18 -32) days, potential lifting of measures in the beginning of May for some 13 countries could not have confounded the estimates. 14 In order to obtain comparable estimates of R t , a simple SEIR compartment model was used, 15 with most epidemiological parameters kept constant. Parameters that model transmission rates 16 were allowed to change from an initial estimated value R t,0 during a transition period also 17 estimated from the data, to R t,1 until the day that mobility changes started, and then to R t,2 18 during the lockdown identified on the basis of mobility data, using a piecewise linear model 19 (Figure 1a and methods). The estimated models fitted well the reported incidence data for most 20 countries (see Suppl 4). Across all countries, the median of posterior estimates for R t,0 was 21 3.6 (95% IQR 2.5 -4.9) (see Figure 2a ). Before changes in mobility were observed (d 1 ), the 22 reproduction number was reduced to 2.3 (95% IQR 1.7 -3.0). During lockdown, transmission 23 was further reduced to 0.77 (95% IQR 0.58 -1.04). Only for Belarus and Moldova, the estimate 24 of R t,2 was higher than 1 (> 95% credible), while all other countries have the R t,2 value below 25 4 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. Figure 1 : Model for changes of the time-varying reproduction number R t as a piece-wise linear function. Dates d 1 and d 2 were estimated from mobility data. Date d 0 and values for R t,0 -R t,2 were estimated from incidence data on diagnosed cases and deaths. a. Model used for all countries; b. Model additionally used for Slovakia, which allowed an extra change during lockdown with dates d 0 , d 3 and d 4 , and values for R t,0 -R t,3 estimated from incidence data on diagnosed cases and deaths. 5 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 June 23, 2020. . https://doi.org/10.1101/2020.06.20.20136382 doi: medRxiv preprint or around 1. Estimates for R t,0 are expected to vary depending on setting, methodology and 1 assumptions on parameters (especially duration of infectious period and generation time) and 2 assumptions on how the number may vary over time (15). We find that our estimated values 3 for R t,0 tend to be higher, and estimated values for R t,1 lower, compared to estimates in other 4 studies for R 0 of around 2.7. Our estimates for R t,0 are similar to R 0 estimates obtained using 5 models that also consider interventions that preceded a full lockdown (14) . 6 We estimated that the date d 0 of the first decline of transmission preceded observed mobility (there was no substantial decline in mobility during this period). The earliest of these dates 11 (for Italy, Ireland, Germany, France, and Belgium) were all estimated around 20 February (95% 12 IQR 12 -25 February), which might have been a result of awareness raised by the discovery of 13 the first European cluster around that time in Lombardy, Italy. Although the semantic meaning 14 given to d 0 (first date of decline of R 0 , before mobility changes), assumed that it would be esti-15 mated before d 1 , this order was not enforced. For Slovakia, the estimated model placed this date 16 around 1 April (95% IQR March 22 -April 7), later than the dates d 1 and d 2 that mark the mo-17 bility changes period (10 -17 March). At the same time, a suspiciously low value of 1.7 (95% 18 IQR 0.8 -2.6) for R t,0 was estimated. Both observations indicated that the assumption that 19 mobility changes leading to the lockdown were the final measure to control transmission, was 20 not applicable to data of Slovakia. We therefore re-estimated a model which allowed a further 21 reduction in transmission to a value R t,3 (assuming a prior distribution R t,3 − R t,2 = N (0, 1) 22 expressing no change with respect to R t,2 ), with a linear transition between co-estimated dates 23 d 3 and d 4 (see Figure 1b ). As prior distributions for d 0 , R t,0 -R t,2 , and infection-to-death dura-24 tion µ d , overall estimates obtained across all countries were used. This resulted in an additional 25 6 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 June 23, 2020. Table 1 : Univariate and multivariate associations of mobility changes during lockdown (per 10% mobility change) with the basic reproduction number R t,2 during lockdown. Mobility data related to transit stations and residential places were left out from the multivariate analysis since these variables were highly correlated with mobility data related to workplaces. This would imply that using face masks brought R t from 0.87 (95% IQR 0.56 -1.26) to 0.50 4 (95% IQR 0.30 -0.70), or thus a 43% (95% IQR 0 -64%) reduction in transmission. In further 5 statistical analyses, the parameters estimated for this latter model were used for Slovakia. To investigate how the reproduction number during lockdown R t,2 related to mobility changes, 7 Google Mobility report data was used. For all six mobility categories, mobility during lock-8 down was significantly different compared to baseline (Wilcoxon paired test p < 0.01, data 9 not shown). Lower R t,2 values during lockdown were significantly associated with a larger 10 mobility reduction related to retail, recreation, and workplaces, and a lower mobility reduction 11 related to residential (Figure 4) , Table 1 ). The association of mobility reduction related to retail 12 and recreation remained significant in a multivariate model which explained 40% of variance 13 (adjusted R 2 ) of the R t,2 value during lockdown (p < 0.002). This suggests that, in 35 West-14 ern countries, reductions of mobility related to retail and recreation during lockdown caused a 15 mean reduction of R t of 0.50 (95% CI 0.18 -0.81), or thus an average reduction of 22% (95% 16 7 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 June 23, 2020. . https://doi.org/10.1101/2020.06.20.20136382 doi: medRxiv preprint Posterior estimates of the initial basic reproduction numbers (R t,0 ), the reproduction number at start of lockdown (R t,1 ) and during lockdown (R t,2 ). b. Estimated median values (95% IQR) for d 0 (date of first reduction in transmission, presumably due to physical distancing, estimated from incidence data of deaths and diagnosed cases); and lockdown transition start and end dates d 1 and d 2 , estimated from mobility data (see also Suppl 1). 8 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 June 23, 2020. able. Alternatively, other types of changes that coincided in time with these mobility changes 6 could be responsible. By explaining only variation of R t,2 , instead of the reduction from R t,1 to R t,2 , the above 8 model side-stepped to a large extent the colinearity that exists between changes for different 9 mobility categories, which showed the same trend in all countries. To verify that the reduction 10 in R t as a result of the lockdown (comparing R t,1 before lockdown to R t,2 during lockdown) 11 was also associated with mobility changes, a multivariate model that directly predicted the 12 percentage change of R t due to lockdown using average mobility changes (comparing mobility 13 data from before d 1 to after d 2 ), was also estimated (data not shown). This model had a high 14 9 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 June 23, 2020. Figure 4 : Average percentage change in mobility, compared to baseline, for six location Google mobility categories. Colour reflects basic reproduction number R t during lockdown. Retail and Recreation: restaurants, cafes, shopping centers, theme parks, museums, libraries, and movie theaters; Grocery and Pharmacy: grocery markets, food warehouses, farmers markets, specialty food shops, drug stores, and pharmacies; Parks: local parks, national parks, public beaches, marinas, dog parks, plazas, and public gardens; Transit Stations: public transport hubs such as subway, bus, and train stations; Workplaces: places of work; Residential: places of residence. 10 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 June 23, 2020. Figure 5 : Summary of estimated contributions to the reduction of transmission in 35 Western countries. Initial basic reproduction numbers (R t,0 ), the reproduction number at start of lockdown (R t,1 ) and during lockdown (R t,2 ), and percentage reductions are shown as median values and 95% IQR. Estimates for individual countries (Italy, Spain and Slovakia) are shown as median posterior value and 95% cri. explanatory power (R 2 = 0.95) but as expected suffered from high colinearity. Nevertheless, 1 in agreement with the above finding, this model also attributed an estimated 31% (95% CI 11 -2 50) of the drop in R t during lockdown to a reduction in mobility related to retail and recreation 3 (p < 0.003). Our findings are summarized in Figure 5 and trends were found to be robust to different 5 assumptions of latent period duration (in the range of 2 to 4 days), and different assumptions of 6 generation time (in the range of 4.5 to 5.9 days), see Suppl 3. Although the estimated value for 7 R t,2 for the epidemic in Sweden was found to be relatively high (1.02, 95% cri 0.99 -1.06), it 8 was equally well explained based on mobility changes, suggesting that although not enforced, 9 in practice the country behaved the same as other Western countries in a lockdown. The fact that mobility changes linked to retail and recreation were significantly associated 11 with R t both in univariate and multivariate analyses, suggests that activities related to visiting 12 malls, bars, restaurants, or museums are linked to increased transmission and releasing those 13 mobility restrictions should be done with care since they may carry a high risk for reigniting 14 the epidemic. This is in line with mounting evidence of transmission being promoted by (loud) 15 speaking or singing (bars, choir (17)), and longer time spent in densely populated indoor loca-16 11 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 June 23, 2020. . I R σ β γ Figure 6 : Structure of a standard SEIR compartment model with four compartments: susceptible (S), exposed (E), infectious (I), and removed (recovered or deceased, R). Susceptible individuals become latently infected by infectious individuals, with transmission rate β. Latently infected become infectious at rate σ. Infectious people are removed at rate γ. tions with low air circulation (bars, restaurants, malls, events with mass gatherings (18)). In the 1 presence of general 'physical distancing' guidelines, which were active in all countries during 2 lockdown, mobility changes related to workplaces (which correlated with transit stations and 3 residential places), did not have a clear impact on R t , suggesting that easing of restrictions re-4 lated to work may be more easily manageable. The lack of association of these other mobility 5 parameters may suggest that rather than being associated with an individual, superspreading 6 is mainly associated with the circumstances of the interaction. This is adding weight to the 7 arguments that superspreading events, such as large indoor gatherings, need to be avoided. The following equations describe the dynamics of individuals in each of the four compartments 1 of a standard SEIR model (see Figure 6 ): The differential equations of the SEIR compartment models were numerically integrated 3 using deSolve (19). The duration of the latent period T lat = 1/σ was fixed to 3 days, and the The dates d 1 and d 2 which mark the transition period for mobility changes were estimated 7 from mobility data reports (22), by fitting a step function with a linear transition period through 8 the sum of mobility changes related to transit stations and workplaces (see Suppl 1). This was 9 motivated by the adoption of teleworking as a measure in most countries, even in those (like 10 Sweden) which had the lightest lockdown regimen. The value for β was modeled as a piece-11 wise linear function of time (Figure 1) . 12 We wanted a model that does not require assumptions on timing or number of introductions 13 per country. Instead, the models were seeded with an initial single exposed individual and it 14 was assumed that the estimated date d 0 , marking the first change in value of β, was linked to a (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 June 23, 2020. . Table 1 lists all parameters used in the model and during the estimation from data, with their values (either a constant, or a prior distribution for parameters that were estimated). 1 To estimate the model parameters, incidence data of diagnosed cases and deaths was used 2 (ECDC, https://opendata.ecdc.europa.eu/covid19/casedistribution/csv, 3 accessed on 6 June), within a Bayesian framework. The daily incidence of diagnosed cases and Univariate and multivariate linear models were used for quantifying the effect of the average 8 values of Google Mobility data (22) over the time period prior to lockdown to R t,1 and during 9 lockdown to R t,2 . Colinearity was assessed by calculating covariance-inflation factors (27). Statistical analysis was done in RStudio (28). All data files, R scripts and analysis steps are described in Suppl 2. (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 June 23, 2020. . Temporal dynamics in viral shedding and transmissibility of covid-12 19 Google COVID-19 community mobility reports Flexible Markov Chain Monte Carlo via Reparameter-17 ization (2020). R package version 1.0.1. 18 24. Li, Q. et al. Early transmission dynamics in wuhan, china, of novel coron-19 avirus-infected pneumonia A.M.V. contributed to the virological and epidemiological interpretation and revisions of the