key: cord-0967721-w0i3oy97
authors: Garchitorena, A.; Gruson, H.; Cazelles, B.; ROCHE, B.
title: Quantifying the efficiency of non-pharmaceutical interventions against SARS-COV-2 transmission in Europe
date: 2020-08-18
journal: nan
DOI: 10.1101/2020.08.17.20174821
sha: 6743bfbc7e9f56649c1b56eb7745d42bc52dc47e
doc_id: 967721
cord_uid: w0i3oy97

Since the emergence of SARS-CoV-2, governments around the World have implemented a combination of public health responses based on non-pharmaceutical interventions (NPIs), with significant social and economic consequences. Though most European countries have overcome the first epidemic wave, it remains of high priority to quantify the efficiency of different NPIs to inform preparedness for an impending second wave. In this study, combining capture-recapture methods with Bayesian inference in an age-structured mathematical model, we use a unique European dataset compiled by the European Centre for Disease Control (ECDC) to quantify the efficiency of 24 NPIs and their combinations (referred to as public health responses, PHR) in reducing SARS-Cov-2 transmission rates in 32 European countries. Of 166 unique PHR tested, we found that median decrease in viral transmission was 74%, which is enough to suppress the epidemic. PHR efficiency was positively associated with the number of NPIs implemented. We found that bans on mass gatherings had the largest effect among NPIs, followed by school closures, teleworking, and stay home orders. Partial implementation of most NPIs resulted in lower than average response efficiency. This first large-scale estimation of NPI and PHR efficiency against SARS-COV-2 transmission in Europe suggests that a combination of NPIs targeting different population groups should be favored to control future epidemic waves.

Since its emergence in China in December 2019, the SARS-COV-2 pandemic has affected almost every country on Earth (1) . With more than 18 million cases reported and almost 800,000 deaths in the first 6 months of the pandemic (2) , this virus will leave a lasting imprint in human history. To control the epidemic, many governments and national public health authorities have implemented national public health responses (PHR) that combine several nonpharmaceutical interventions simultaneously (NPIs). To ultimately achieve either a suppression or a mitigation of the epidemic, NPIs aim to reduce transmission by (i) lowering contact rates in the general population or specific groups, minimizing contact with infectious individuals, and (ii) reduce the infectiousness of contacts (e.g. use of masks). The NPIs implemented by most countries involve the closure of schools and universities, banning gatherings of various sizes, issuing stay at home recommendations or orders, and the use of individual protective equipment (e.g., wearing masks), with different levels of law enforcement. The unprecedented breadth and scale of these responses, in addition to variations in combination, timing and level of implementation across European countries has created a context that could allow to quantify the efficiency of different NPIs in reducing SARS-COV-2 transmission rates (3) . Given the huge economic and social costs of these interventions, such information can be key to better control an impending second epidemic wave, allowing to prioritize the most efficient NPIs early on.

In this study we combine a capture/recapture method and Bayesian inference in an agestructured mathematical model, to understand the efficiency of NPIs implemented in Europe.

We apply this method to a dataset of country response measures to COVID-19(4) compiled by the European Centre for Diseases Control (ECDC) to estimate the efficiency of different NPIs and public health responses (i.e., a combination of NPIs). This dataset contains the precise timing for the different PHR implemented in each European country paired with epidemiological data on COVID-19 case incidence and mortality (5) . We find a median PHR All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted August 18, 2020. . efficiency of 74%, enough to suppress a country's epidemic. PHR efficiency is positively associated with the number of NPIs implemented. Among those NPIs implemented in Europe in the first half of 2020, we identify bans to mass gathering as the most efficient, followed by school closures, teleworking and stay at home orders. Partial implementation of most NPIs resulted in lower than average response efficiency.

Twenty-four different NPIs had been implemented from February 1 st to June 26 th , 2020 across the 32 European countries that reported data to the ECDC. Among them, bans on mass gatherings and school closures have been the most widely used, implemented by more than 80% of countries ( figure 1A ). These NPIs were implemented for a median time of two to three months. Across all the 166 PHRs analyzed, stay at home orders, bans on mass gatherings and school/university closures were typically implemented together (figure 1B). Despite large variability in terms of PHRs and sociocultural backgrounds across Europe, we were able to accurately quantify viral transmission rates over time in each country before and after implementation of PHRs, reproducing the observed national epidemiological dynamics ( figure   2 ). This allowed us to estimate the efficiency of the different country-level responses and isolate the effect of each NPI implemented in this period.

Overall, the combination of NPIs implemented as part of PHRs across Europe had a high efficiency at reducing transmission rates, with a median reduction of 74%. With an estimated basic reproduction number R0 of about 3, standard epidemiological models of COVID-19 suggest that to decrease the effective reproductive ratio below 1, transmission rates need to be reduced by at least 66% (6) . According to this, nearly two thirds of the PHRs evaluated could be part of a "suppression strategy" (figure 3A), halting the epidemic before sustainable herd immunity is achieved(7). All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted August 18, 2020. . https://doi.org/10.1101/2020.08.17.20174821 doi: medRxiv preprint Efficiency was positively associated with the number of NPIs implemented as part of the PHR (figure 3A). To assess the effect of each NPI, we first estimated the added efficiency of each PHR when a particular NPI was present in univariate linear models, disaggregated by the number of NPIs implemented simultaneously (figure 3B). We found wide heterogeneity in efficiency between NPIs and for the same NPI at varying number of NPIs implemented simultaneously; only the presence of mandatory use of masks had consistent positive effects on PHR efficiency (figure 3B). In contrast, when we restricted the analysis to the most efficient PHRs, mass gatherings, teleworking and stay at home orders were the NPIs most consistently present across responses (figure 3C).

Importantly, these efficiencies were estimated for countries with very different sociocultural and economic backgrounds, NPIs were implemented for different lengths of time, and there was substantial overlap between them. To address this, we estimated the change in response efficiency over time when adding or removing a particular NPI, while controlling for GDP, number of NPIs, length of implementation, and country of implementation (figure 4). We show that adding bans on all mass gatherings to the response resulted in the largest increase in efficiency (22.5%); this increase in efficiency was slightly lower when restricting gatherings to 50 people (figure 4A). Stay at home orders, teleworking, and school closures all resulted in significant increases in efficiency over 15% regardless of the number of NPIs implemented, but partial implementation of these NPIs resulted in lower than average (or non-significant) response efficiency. Among the five groups of NPIs considered, all were associated with statistically significant increases in PHR efficiency, except for the use of masks (figure 4B).

Our empirical results on NPI implementation in 32 European countries show that most PHRs implemented between March and June 2020 had a sufficient efficiency to achieve suppression of the COVID-19 epidemic in the countries assessed. The most efficient PHRs typically All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted August 18, 2020. . https://doi.org/10.1101/2020.08.17.20174821 doi: medRxiv preprint targeted multiple societal activities and population groups, including bans of mass gatherings, stay home orders, teleworking, and some form of school closure. This is consistent with previous modeling studies suggesting that packages of NPIs, not single interventions, are necessary to achieve epidemic suppression (8) (9) (10) .

We find that stay at home orders (i.e. lockdowns) significantly increased the efficiency of PHRs. However, one third of responses with efficiencies higher than 66% did not involve stay at home orders but a combination of the other NPIs. In addition, adding stay at home orders to the PHR did not increase response efficiency more than other NPIs, such as bans to public gatherings, school closures, or teleworking. These results are in contrast with previous evidence indicating that the effect of lockdowns is substantially larger than other NPIs (3), and that only lockdown periods can reduce R0 below 1 in order to suppress the epidemic(10). This discrepancy can be explained by the fact that estimating empirically the relative efficiency of lockdowns compared to other combinations of NPIs requires large amounts of data, with a level of temporal and spatial heterogeneity in NPI implementation that has only recently been possible to achieve. Therefore, our study suggests that, although effective at reducing COVID-19 transmission, lockdowns may not always be necessary to achieve epidemic suppression.

There has recently been increasing consensus around recommendations for the universal use of masks by the general population, as a cost-effective solution to allow activities in public places while minimizing the risk of viral transmission between individuals (11) (12) (13) . Although our analysis did not find that adding the use of masks to existing PHRs led to a significant increase in efficiency, the results should be interpreted with caution. Few European countries had implemented this type of NPI at the time of our analyses, and implementation was recent compared to other NPIs ( figure 3 ). In addition, enforcement of the use of masks in Europe has been mild and, in contrast with many Asian countries, adoption of this individual protective measure by the population has been very variable. Contrary to the use of masks, we found that All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted August 18, 2020. This study had several limitations. First, we used an age-structured model where each age class has the same susceptibility to the virus, and we assume that detected and non-detected cases have the same transmissibility capacity because non-detected cases can be both asymptomatic and symptomatic cases. However, it is still unclear whether asymptomatic cases and children have similar or lower SARS-COV-2 transmissibility capacity compared to other population groups, which could affect our results. Second, although our mathematical model allowed us to accurately quantify PHR efficiency, responses were comprised of 24 different NPIs that frequently overlapped over space and time. We estimated the added efficiency provided by each NPI via statistical time series analyses controlling for relevant factors, but these should not be interpreted as NPI independent efficiencies because we could not control for the implementation of other NPIs in our multivariate framework. Third, we use a comprehensive database on NPI implementation compiled by the ECDC from country reports, but other interventions that could impact epidemic progression, such as travel restrictions (17, 18) , were not available in this dataset. This could have biased the estimates for the NPIs evaluated here.

In the case of a second wave, our study suggests that comprehensive PHRs that include bans to mass gatherings, teleworking, and school closures will be most effective at suppressing All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted August 18, 2020. . https://doi.org/10.1101/2020.08.17.20174821 doi: medRxiv preprint the epidemic. Stay at home orders substantially contribute to response efficiency but may not be always necessary to achieving epidemic suppression. Further evidence is needed to assess whether enforcing the use of masks significantly improves overall PHR efficiency.

The data in COVID-19 incidence in 32 European countries were retrieved from the ECDC epidemic intelligence database on worldwide data on COVID-19 (5) . The information about each NPI implemented over time as part of PHRs in each European country is available through the ECDC database on country response measures to COVID-19 (4). This database is regularly updated by the ECDC on a regular basis using authoritative sources for national authorities and international institutions (19) .

The mathematical model relies on contact data between the different age classes and has been derived for each country from Prem et al (20) . The age structure of each country has been retrieved from https://www.populationpyramid.net/. This data is released under a Creative Commons Attribution 3.0 (CC BY) license.

To model the evolution of the epidemic within each of the 32 European countries assessed, we built a model derived from a stochastic age-structured SEIR framework (21) . The population of each age class i is divided according to their infectious status: Susceptible (Si), Exposed (Ei, infected but not yet infectious), Detected infectious (Ii), Non-detected infectious (Ai) and recovered (Ri). The probability of case detection varies over time and has been estimated through a capture/recapture method ((22), see below). Model simulations were performed using a τ-leap algorithm (23) . Latency period (1/ε) and recovery period (1/σ) were assumed at 3 and 5 days respectively. Finally, 8 age classes have been considered (0-10, 11-20, 21-30, 31-40, 41-50, 61-60, and 70+ years).

All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted August 18, 2020. . https://doi.org/10.1101/2020.08.17.20174821 doi: medRxiv preprint

We estimated daily non-reported cases for each country using the daily new reported cases and daily new deaths and the capture/recapture methodology described by Böhning and collaborators(22) , where the number of daily new non-reported cases is:

where t is each time step (in days), ΔN(t) is the daily number of new cases over time and ΔD(t) is the daily number of new deaths.

We estimated transmission rates per country over time through Bayesian inference by running a particle Monte Carlo Markov Chain (pMCMC) with the epidemiological model described earlier. The chains had 10,000 iterations with a burn in of 90% and a thinning of 1/10. The likelihood was computed with a particle filter of 100 particles and a negative binomial function of parameter 0.5 on the total daily number of new cases expected by the model and observed in the data (i.e., number of cases notified to the national public health system and number of undetected cases estimated by the capture/recapture method). The ratio between transmission rates with and without implementation of the PHR evaluated provides an estimation of the efficiency of this response, therefore assuming that an instantaneous effect.

From the 166 PHR efficiencies estimated earlier, we first estimated in univariate linear regression analyses the effect on efficiency of having each NPI (presence/absence, dummy variables). Since the number of NPIs implemented was positively associated with response efficiency, we disaggregated these univariate analyses by the number of NPIs implemented as part of the response. To isolate the efficiency of each NPI given the substantial overlap of NPIs implemented together, we transformed the NPI presence/absence and response efficiency All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted August 18, 2020. . dataset to reflect changes over time. For this, we first estimated for each country the difference in response efficiency with the previous time period (either no PHR or previous response) and the difference in NPI implementation (0 = no change; 1 = NPI added to the response; -1 = NPI removed from the response). We then estimated the effect of adding each NPI on the change in response efficiency using a multivariate linear mixed model that controlled for country's GDP, number of interventions implemented, duration of implementation (fixed effects), and country of implementation (random intercept). Finally, we used the same model to estimate the effect of each of the five groups of NPIs implemented: teleworking, stay at home orders, bans on mass gatherings, use of masks and school closures.

All simulation and analyses were conducted using R 3.6 on the IRD itrop HPC (South Green Platform) at IRD Montpellier (https://bioinfo.ird.fr/). All code is available at https://github.com/Bisaloo/analysisECDC All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted August 18, 2020. . https://doi.org/10.1101/2020.08.17.20174821 doi: medRxiv preprint

Coronavirus disease 2019 (COVID-19) pandemic: increased transmission in the EU/EEA and the UK -seventh update. Impact of non-pharmaceutical interventions (NPIs) to reduce COVID-19 mortality and healthcare demand outbreak in China

Coronavirus disease 2019 (COVID-19) Situation Report

Modeling Infectious Diseases in humans and animals

Advice on the use of masks in the context of COVID-19: interim guidance-2

Using face masks in the community

Reopen website from the European Union

Optimal strategies for quarantine stopping in France. General expected patterns of strategies focusing on contact between age groups

Modeling Infectious Diseases

Results for each of the 24 NPIs. (B) Results for NPIs grouped together into 5 more general groups. A full description for each NPI is available in the section S1

HG and BR are supported by a grant from the "Agence Nationale de la Recherche" (ANR-

Difference in response efficiency when a particular NPI was present (univariate models). (C) Shows which NPIs were part of the most efficient responses (light blue). Both B and C are disaggregated by the number NPIs that were implemented simultaneously. A full description for each NPI is available in the section S1.

Results show the change in PHR efficiency over time when adding each NPI (mean effect and 95% confidence intervals). Each NPI was modelled separately using a multivariate linear mixed model that controls for country's GDP, number of interventions implemented, duration of implementation (fixed effects), and country of implementation (random intercept). (A)