key: cord-1054684-32lau54s authors: Vollmer, M. A. C.; Mishra, S.; Unwin, H. J. T.; Gandy, A.; Mellan, T. A.; Bradley, V.; Zhu, H.; Coupland, H.; Hawryluk, I.; Hutchinson, M.; Ratmann, O.; Monod, M.; Walker, P.; Whittaker, C.; Cattarino, L.; Ciavarella, C.; Cilloni, L.; Ainslie, K.; Baguelin, M.; Bhatia, S.; Boonyasiri, A.; Brazeau, N.; Charles, G.; Cooper, L. V.; Cucunuba, Z.; Cuomo-Dannenburg, G.; Dighe, A.; Djaafara, B.; Eaton, J.; van Elsland, S. L.; FitzJohn, R.; Fraser, K.; Gaythorpe, K.; Green, W.; Hayes, S.; Imai, N.; Jeffrey, B.; Knock, E.; Laydon, D.; Lees, J.; Mangal, T.; Mousa, A.; Nedjati-Gilani, G.; Nouvellet, P. title: A sub-national analysis of the rate of transmission of COVID-19 in Italy date: 2020-05-09 journal: nan DOI: 10.1101/2020.05.05.20089359 sha: 9a1814a7a5232980ff0e7d60d844f19dd485427d doc_id: 1054684 cord_uid: 32lau54s Italy was the first European country to experience sustained local transmission of COVID-19. As of 1st May 2020, the Italian health authorities reported 28,238 deaths nationally. To control the epidemic, the Italian government implemented a suite of non-pharmaceutical interventions (NPIs), including school and university closures, social distancing and full lockdown involving banning of public gatherings and non essential movement. In this report, we model the effect of NPIs on transmission using data on average mobility. We estimate that the average reproduction number (a measure of transmission intensity) is currently below one for all Italian regions, and significantly so for the majority of the regions. Despite the large number of deaths, the proportion of population that has been infected by SARS-CoV-2 (the attack rate) is far from the herd immunity threshold in all Italian regions, with the highest attack rate observed in Lombardy (13.18% [10.66%-16.70%]). Italy is set to relax the currently implemented NPIs from 4th May 2020. Given the control achieved by NPIs, we consider three scenarios for the next 8 weeks: a scenario in which mobility remains the same as during the lockdown, a scenario in which mobility returns to pre-lockdown levels by 20%, and a scenario in which mobility returns to pre-lockdown levels by 40%. The scenarios explored assume that mobility is scaled evenly across all dimensions, that behaviour stays the same as before NPIs were implemented, that no pharmaceutical interventions are introduced, and it does not include transmission reduction from contact tracing, testing and the isolation of confirmed or suspected cases. New interventions, such as enhanced testing and contact tracing are going to be introduced and will likely contribute to reductions in transmission; therefore our estimates should be viewed as pessimistic projections. We find that, in the absence of additional interventions, even a 20% return to pre-lockdown mobility could lead to a resurgence in the number of deaths far greater than experienced in the current wave in several regions. Future increases in the number of deaths will lag behind the increase in transmission intensity and so a second wave will not be immediately apparent from just monitoring of the daily number of deaths. Our results suggest that SARS-CoV-2 transmission as well as mobility should be closely monitored in the next weeks and months. To compensate for the increase in mobility that will occur due to the relaxation of the currently implemented NPIs, adherence to the recommended social distancing measures alongside enhanced community surveillance including swab testing, contact tracing and the early isolation of infections are of paramount importance to reduce the risk of resurgence in transmission. 4th May 2020 Imperial College COVID-19 Response Team 2 we also overlay the timing of major NPIs (Appendix Table 6 .3). Due to very strong collinearity across mobility dimensions we only use residential, transit stations, and an average of the remaining four dimensions (i.e. retail and recreation, groceries and pharmacies, parks, and workplaces). The residential dimension is a proxy for household transmission and the transit dimension is a proxy for general travel within and between regions, including time spent at travel hubs. The average mobility is the mean of the other dimensions and is a proxy for general day-to-day activities. There is clear visual correspondence between the dates interventions were implemented and the observed reductions in mobility. This is demonstrated statistically by the large mean correlation of 0.81 obtained with a simple linear model regressing interventions (as piecewise constant) on the average mobility dimension. This suggests that mobility can act as a suitable proxy for the changes in behaviour induced by the implementation of the major NPIs. We do note however, that mobility does not capture all the heterogeneity in transmission, specifically missing factors such as case-based interventions and the effect of school and university closures. Figure 3 shows the average global effect sizes for the mobility dimensions used in our model. Due to collinearity, it is not statistically possible to identify which dimension has had the largest impact on R t . However, we do find that the transit dimension and the average mobility dimension are statistically significant, while the residential dimension is not (though the posterior mean is less than 0). We hypothesise that the residential covariate could increase R t due to household transmission between cohabitants. Transit Residential Average Other Mobility 0% (no effect on transmissibility) 25% 50% 75% 100% (ends transmissibility) Relative % reduction in R t Mobility Figure 3 : Mobility effect sizes: relative reduction in R t if the specified mobility was completely stopped. Despite Italy having the largest number of deaths attributable to COVID-19 in Europe, the estimated attack rate (percentage of the population that has been infected) is still relatively low across all regions ( Figure 6 shows the basic reproduction number (R 0 ) and effective reproduction number (R t ) as of 1st The primary mechanism driving dynamics in our model is R t , which is parameterised by mobility. Using our model, jointly fitted to all regions in Italy, we are able to simulate forwards 8 weeks with hypothetical scenarios where mobility increases. We do not differentiate what causes these increases in mobility but it stands to reason they would occur from a relaxation of NPIs and changes in behaviour. We also note that other mechanisms aside from mobility can increase R t and would yield in the same result. We choose three scenarios (a) constant mobility in which mobility remains at current lockdown levels for 8 weeks, (b) 20% return to pre-lockdown mobility and (c) 40% return to pre-lockdown mobility. Scenarios (b) and (c) are calculated using a weighted average between the current mobility and the nominal pre-lockdown level. Thus, for example, in scenario (b), 20% of the weight is on the nominal pre-lockdown level and 80% on the current mobility. Scenario (a) is equivalent to a 0% return to prelockdown mobility. Figures 7 and 8 show the estimated increases in R t due to a 40% return to pre-lockdown mobility. A 40% return represents a reasonably large change in mobility and for many regions shifts R t just above 1. . 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 May 9, 2020. . https://doi.org/10.1101/2020.05.05.20089359 doi: medRxiv preprint 4th May 2020 Imperial College COVID-19 Response Team 4 and 5 show the scenarios of 20% and 40% returns to pre-lockdown mobility. In the constant mobility scenario we predict a continued reduction in deaths, however in the 20% and 40% scenarios, while initially deaths may continue to decrease, there will eventually be a resurgent epidemic that, without accounting for additional interventions, may be larger in size than the first wave. Using our simulated scenarios we can calculate the deaths averted by keeping mobility at current levels. Table 2 shows the deaths averted under the 20% and 40% return to pre-lockdown mobility scenarios and no other intervention is put in place. Under the 20% scenario we estimate the total number of excess deaths to be between 3,000 and 5,000, and under the 40% scenario the total number of excess deaths would be between 10,000 and 23,000 (see Table 2 ). The deaths averted are largest in regions currently experiencing major epidemics; the reason they rebound to such a large extent is driven by a large number of ongoing infections. If more time is spent under current lockdown mobility levels before increases occur, the number of deaths averted is likely to be considerably lower in both scenarios. It should be noted that in our model we do not account for cross-region movement, which, given increased mobility, is likely to increase infections and subsequently deaths, in regions not experiencing major epidemics. DOI: https://doi.org/10.25561/78677 Page 9 of 35 . 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 May 9, 2020. is the 95% credible interval forecast scenario were mobility stays at lockdown levels, and red is the 95% credible interval forecast scenario where mobility returns by 20% to pre-lockdown levels. DOI: https://doi.org/10.25561/78677 Page 10 of 35 . 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 May 9, 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. (which was not certified by peer review) The copyright holder for this preprint this version posted May 9, 2020. 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 May 9, 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 May 9, 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 May 9, 2020. In this report we use a semi-mechanistic Bayesian hierarchical model fitted to sub-national death data for Italy. We parameterise the reproduction number, a fundamental measure of transmission intensity, as a function of an individual's mobility. We show that mobility, both visually and statistically, is associated with the onset and timing of major NPIs. Using our model, we estimate that the average reproduction numbers in all regions across Italy is currently below 1, suggesting that the major interventions implemented by the Italian government have controlled transmission and averted a major health catastrophe. We show that despite the large number of deaths attributable to COVID-19, the attack rates are far lower than required for herd immunity. Simulating 8 weeks into the future, we estimate that, if mobility remains the same, there will be a continued reduction in deaths and the epidemic will be suppressed. However, returns to pre-lockdown mobility of 20% or 40% from current levels may lead to a resurgence of the epidemic with more deaths than the current wave in the absence of additional interventions. Our modelling framework is novel in that we infer a latent function for infections, and, to the best of our knowledge for the first time, parameterise R t using mobility data. The use of mobility data as a proxy Given that interventions, such as extensive testing, contact tracing and social distancing are going to be implemented, our estimates can be viewed as being pessimistic. On the other hand, simulating 20% and 40% increase in mobility over the next 8 weeks is likely a conservative scenario. Our model uses the official deaths counts to estimate changes in transmission intensity. We did not use the reported number of confirmed COVID-19 cases because of potential biases arising from changes in the case definition and testing strategy adopted during the epidemic across the regions, which would be hard to correct for. As more information on new interventions are introduced we will include them in our modelling framework. Our results suggest that transmission, as well as mobility, need to be closely monitored in the future weeks and months. To date, it is hard to predict the extent to which new interventions will be able 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 May 9, 2020. Our model utilizes daily real-time death data provided by the Italian Civil Protection (publicly available at https://github.com/pcm-dpc/COVID-19) for the 20 Italian regions. For the Trentino Alto-Adige region, we report the results for the provinces of Trento and Bolzano separately, following the format of the death data provided by the Italian Civil Protection. For population counts, we use publicly available age-stratified counts from ISTAT ("Popolazione residente comunale per sesso anno di nascita e stato civile", from https://www.istat.it). Mobility data have been obtained from the Google Mobility Report (google.com/covid19/mobility/) which provides data on movement in Italy by region and highlights the percent change in visits to: • Grocery & pharmacy: Mobility trends for places like grocery markets, food warehouses, farmers markets, specialty food shops, drug stores, and pharmacies. • Parks: Mobility trends for places like local parks, national parks, public beaches, marinas, dog parks, plazas, and public gardens. • Transit stations: Mobility trends for places like public transport hubs such as subway, bus, and train stations. • Retail & recreation: Mobility trends for places like restaurants, cafes, shopping centers, theme parks, museums, libraries, and movie theaters. • Residential: Mobility trends for places of residence. • Workplaces: Mobility trends for places of work. The mobility data show the length of stay at different places compared to a baseline. It is therefore relative, i.e mobility of -50% means that, when compared to pre COVID-19, individuals are engaging in a given activity 50% less. We also catalogue data on the nature and type of major NPIs. We referred to government as well as official public health division webpages to identify the recommendations and laws being issued by the central government and local public health authorities. We collected the following: • School closure ordered: This intervention refers to nationwide extraordinary school closures which in most cases refer to both primary and secondary schools closing (for most regions this also includes the closure of other forms of higher education or the advice to teach remotely). The date DOI: https://doi.org/10.25561/78677 Page 19 of 35 . 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 May 9, 2020. . https://doi.org/10.1101/2020.05.05.20089359 doi: medRxiv preprint 4th May 2020 Imperial College COVID-19 Response Team of the school closure is taken to be the effective date when the schools started to be closed (if this was on a Monday, the date used was the one of the previous Saturdays as pupils and students effectively stayed at home from that date onwards). • Case-based measures: This intervention comprises strong recommendations or laws to the general public and primary care about self-isolation when showing COVID-19-like symptoms. These also include nationwide testing programs where individuals can be tested and subsequently selfisolated. Our definition is restricted to official advice to all individuals or to all primary care. These do not include containment phase interventions such as isolation if travelling back from an epidemic region such as China. • Public events banned: This refers to banning all public events of more than 100 participants such as sports events. • Social distancing encouraged: As one of the first interventions against the spread of the COVID-19 pandemic, the central government and many regions published advice on social distancing including the recommendation to work from home wherever possible and reduce the use of public transport and all other non-essential contacts. The dates used are those when social distancing has officially been recommended; the advice may include maintaining a recommended physical distance from others. • Lockdown decreed: There are several different scenarios that the media refers to as lockdown. As an overall definition, we consider regulations/legislations regarding strict face-to-face social interaction: including the banning of any non-essential public gatherings, closure of educational and public/cultural institutions, ordering people to stay home apart from essential tasks. We include special cases where these are not explicitly mentioned on government websites but are enforced by the police. The dates used are the effective dates when these legislations have been implemented. We note that lockdown encompasses other interventions previously implemented. The mobility data together with the intervention timings are shown in Figure 2 . In a previous report [4] , we introduced a new Bayesian framework for estimating the transmission intensity and attack rate (percentage of the population that has been infected) of COVID-19 from the reported number of deaths. This framework uses the time varying reproduction number R t to inform a latent function for infections, and then these infections, together with probabilistic lags, are calibrated DOI: https://doi.org/10.25561/78677 Page 20 of 35 . 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 May 9, 2020. Changes in testing strategies during the epidemic mean that the severity of confirmed cases as well as the reporting probabilities changed in time and may thus have introduced bias in the data. In this report, we adapt our original Bayesian semi-mechanistic model of the infection cycle to the 20 Italian regions. We infer plausible upper and lower bounds (Bayesian credible intervals) of the total populations infected (attack rates) and the reproduction number over time (R t ). In our framework we parameterise R t as a function of Google mobility data. We fit the model jointly to COVID-19 data from all regions to assess whether there is evidence that changes in mobility have so far been successful at reducing R t below 1. Our model is a partial pooling model, where the effect of mobility is shared but region-specific modifiers can capture differences and idiosyncrasies among the regions. We then simulate forwards using a simple assumption that mobility returns to 20% or 40% pre-lockdown levels of mobility from the latest lockdown levels and explore the impact of increased mobility on transmission intensity, infections and deaths. We note that future directions should focus on embedding mobility in realistic contact mechanisms to establish a closer relationship to transmission. Here, N (µ,σ) denotes a normal distribution with mean µ and standard deviation σ. We say that X follows a positive half normal distribution To mechanistically link our function for deaths to our latent function for infected cases, we use a previ- 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 May 9, 2020. . https://doi.org/10.1101/2020.05.05.20089359 doi: medRxiv preprint 4th May 2020 Imperial College COVID-19 Response Team the above, every region has a specific mean infection fatality ratio ifr m (see Table 1 ). To incorporate the uncertainty inherent in this estimate we allow the ifr m for every region to have additional noise around the mean. Specifically we assume ifr * m ∼ ifr m ·N (1,0.1). Using estimated epidemiological information from previous studies[16, 17], we assume the distribution of times from infection to death π (infection-to-death) to be π ∼ Gamma(5.1,0.86) + Gamma (17.8,0.45) . The expected number of deaths d t,m , on a given day t, for region, m, is given by the following discrete sum: where c τ,m is the number of new infections on day τ in region m and where π is discretized via π s = s+0.5 s−0.5 π(τ )dτ for s = 2,3,..., and π 1 = 1.5 0 π(τ )dτ , where π(τ ) is the density of π. The true number of infected individuals, c, is modelled using a discrete renewal process. We specify a generation distribution g with density g(τ ) as: g ∼ Gamma(6.5,0.62). Given the generation distribution, the number of infections c t,m on a given day t, and region, m, is given by the following discrete convolution function: where, similar to the probability of death function, the generation distribution is discretized by g s = We parametrise R t,m as a linear function of the relative change in time spent (from a baseline) across three (k = 3) Google mobility dimensions: residential, transit station and an average of retail and recreation, groceries and pharmacies, parks, and workplaces. The reason for taking an average was that these dimensions were extremely collinear. The effect of mobility on transmission is assumed to be multiplicative. R t,m is therefore a function of the mobility indicator I k,t,m in place at time t in region m: 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 May 9, 2020. where φ −1 is the inverse logit or sigmoid function. The impacts α k are shared between all M regions and β m,k allows for region specific effects. This model is therefore a partial pooling model. The prior distribution for the shared coefficients were chosen to be α k ∼ N (0,0.5), and the prior distribution for the pooled coefficients were chosen to be β m,k ∼ N (0,γ) with γ ∼ N + (0,0.5). The prior distribution for R 0,m [8] was chosen to be where κ is the same among all regions. We assume that seeding of new infections begins 30 days before the day after a region has cumulatively observed 10 deaths. From this date, we seed our model with 6 sequential days of an equal number of infections: c 1,m = · · · = c 6,m ∼ Exponential( 1 τ ), where τ ∼ Exponential(0.03). These seed infections are inferred in our Bayesian posterior distribution. We estimated parameters jointly for all regions in a single hierarchical model. Fitting was done in the probabilistic programming language Stan [2] using an adaptive Hamiltonian Monte Carlo (HMC) sampler. [9] Regione Emilia-Romagna. accessed on 10.04.2020. URL: https : / / www . regione . emiliaromagna.it/notizie/2020/febbraio/sospesa-attivita-scuole-nidi-e-manifestazionitutte-le-misure-adottate-contro-il-coronavirus. [10] Regione Emilia-Romagna. accessed on 10.04.2020. URL: https : / / www . regione . emiliaromagna . it / notizie / 2020 / febbraio / coronavirus -manifestazioni -pubblicheservizi-e-attivita-quelle-da-sospendere-e-quelle-che-possono-proseguire. Region specific % effect on Rt . 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 May 9, 2020. ribbon is the 95% credible interval forecast scenario were mobility stays the same, and blue is the 95% credible interval forecast scenario where mobility returns by 40% to pre-lockdown levels. DOI: https://doi.org/10.25561/78677 Page 31 of 35 . 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 May 9, 2020. . https://doi.org/10.1101/2020.05.05.20089359 doi: medRxiv preprint [5] Sicily Social distancing 08.03.2020 [5] Sicily Public events 08.03.2020 [5] Sicily Lockdown 10.03.2020 [6] Tuscany School closures 05.03.2020 [14] Tuscany Case-based measures 08.03.2020 [5] Tuscany Social distancing 08.03.2020 [5] Tuscany Public events 08.03.2020 [5] Tuscany Lockdown 10.03.2020 [6] Trento School closures 05.03.2020 [14] Trento Case-based measures 08.03.2020 [5] Trento Social distancing 08.03.2020 [5] Trento Public events 08.03.2020 [5] Trento Lockdown 10.03.2020 [6] Umbria School closures 05.03.2020 [14] Umbria Case-based measures 08.03.2020 [5] Umbria Social distancing 08.03.2020 [5] Umbria Public events 08.03.2020 [5] Umbria Lockdown 10.03.2020 [6] Veneto School closures 24.02.2020 [7] Veneto Case-based measures 08.03.2020 [5] Veneto Social distancing 08.03.2020 [5] DOI: https://doi.org/10.25561/78677 Page 34 of 35 . 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 May 9, 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. (which was not certified by peer review) The copyright holder for this preprint this version posted May 9, 2020. . https://doi.org/10.1101/2020.05.05.20089359 doi: medRxiv preprint Google COVID-19 Community Mobility Reports: Anonymization Process Description (version 1.0) Stan : A Probabilistic Programming Language The early phase of the {COVID}-19 outbreak in {Lombardy}, {Italy} Estimating the number of infections and the impact of non-pharmaceutical interventions on COVID-19 Basilicata Case-based measures 08 Bolzano Case-based measures 08 Liguria Case-based measures 08 Liguria Public events 25 Liguria Lockdown Lombardy Case-based measures 08 Lombardy Public events 23 Lombardy Lockdown We would like to thank Amazon AWS and Microsoft Azure for computational credits. We would like to thank the Stan Development team for their constant support. DOI: https://doi.org/10.25561/78677Page 24 of 35