key: cord-0799969-bd27zwdo
authors: Rypdal, M.; Rypdal, V. G.; Jakobsen, P. K.; Ytterstad, E.; Lovsletten, O.; Klingenberg, C.; Rypdal, K.
title: Modelling Suggests Limited Change in the Reproduction Number from Reopening Norwegian Kindergartens and Schools During the COVID-19 Pandemic
date: 2020-08-15
journal: nan
DOI: 10.1101/2020.08.14.20174896
sha: 02ced679d2391105de5c71eab9808cefd75886cd
doc_id: 799969
cord_uid: bd27zwdo

Background: To suppress the COVID-19 outbreak, the Norwegian government closed all schools on March 13, 2020. The kindergartens reopened on April 20, and the schools on April 27 and May 11 of 2020. The effect of these measures is largely unknown since the role of children in the spread of the SARS-CoV-2 virus is still unclear. There are only a few studies of school closures as a separate intervention to other social distancing measures, and little research exists on the effect of school opening during a pandemic. Objective: This study aimed to model the effect of opening kindergartens and the schools in Norway in terms of a change in the reproduction number (R). A secondary objective was to assess if we can use the estimated R after school openings to infer the rates of transmission between children in schools. Methods: We used an individual-based model (IBM) to assess the reopening of kindergartens and schools in two Norwegian cities, Oslo, the Norwegian capital, with a population of approximately 680 000, and Tromsoe, which is the largest city in Northern Norway, with a population of approximately 75 000. The model uses demographic information and detailed data about the schools in both cities. We carried out an ensemble study to obtain robust results in spite of the considerable uncertainty that remains about the transmission of SARS-CoV-2. Results: We found that reopening of Norwegian kindergartens and schools are associated with a change in R of 0.10 (95%CI 0.04-0.16) and 0.14 (95%CI 0.01-0.25) in the two cities under investigation if the in-school transmission rates for the SARS-CoV-2 virus are equal to what Ferguson et al. have previously estimated for influenza pandemics [1]. Conclusion: We found only a limited effect of reopening schools on the reproduction number, and we expect the same to hold true in other countries where nonpharmaceutical interventions have suppressed the pandemic. Consequently, current R-estimates are insufficiently accurate for determining the transmission rates in schools. For countries that have not opened schools yet, planned interventions, such as the opening of selected schools, can be useful to infer general knowledge about children-to-children transmission of SARS-CoV-2.

In response to the COVID-19 pandemic, the governments of most countries have introduced nonpharmaceutical interventions (NPI) to limit the spread of the virus. The most identifiable NPIs are travel restrictions and social distancing measures, such as isolation for infected individuals and quarantines for their contacts (combined with testing and contact tracing), closure of physical workplaces, and school closures. The effect of these measures is currently a topic of intense research [2, 3] . In particular, the effect of school closures is mostly unknown [4] . By early April 2020, 188 countries had closed schools countrywide, and more than 90% of the world's learners were affected [5] .

On March 12, the Norwegian government announced a series of restrictive infection control measures, including school closure from March 13. However, due to a situation with few COVID-19 cases in Norway, in particular among children and the general disruptive effects of school closures, the government announced already mid-April, a gradual reopening starting with kindergartens from April 20, schools for children born later than 2013 from April 27, and schools for children born later than 2009 from May 11. The Norwegian Institute of Public Health (NIPH) and the Ministry of Education and Research have published guidelines for infection prevention and control in schools during the COVID-19 pandemic [6] .

Based on the initial outbreak in China, only 1% of confirmed cases were children aged 1-9 years [7] , and a similar proportion in this age group (1.4%) is reported in Norway [8] . Due to milder disease in children, we expect the proportion of COVID-19 cases ascertained to be significantly lower for children than adults, and several studies assume a uniform attack rate among age groups [9, 10] . A uniform attack rate would imply that testing in China only ascertained around 10% of cases in children because asymptomatic children are likely not tested. Similar proportions would be expected in Scandinavian countries, depending on testing regimes. If a large proportion of the infections in children are unrecorded, it is difficult to conclude on the effect of school closures based on the currently available data.

One approach, presented in the 13th report on the COVID-19 virus of the Imperial College COVID-19 Response Team [3] , is to estimate R as a function of time in different countries, and follow the evolution of this number as countries have put in place various interventions. The study from Imperial College analyzed 11 European countries and found that school closures have the effect of reducing the R by up to 50%, with a best estimate of 20-25%. The authors argued that there is considerable uncertainty regarding these numbers due to the short time between different interventions (schools closing, social distancing, banning of public events, self-isolation, and lockdown). However, these results are consistent with previous analyses comparing the transmission of infectious diseases on weekdays and weekends [11] . In a recent study of the COVID-19 outbreaks in Wuhan and Shanghai, Zhang et al. use contact surveys to set up contact matrices in an age-structured Susceptible-Infectious-Recovered (SIR) model, and found, in simulations, that school closures can reduce the peak incidence by 40-60%, but are alone, not sufficient to interrupt the transmission [12] . The results of Zhang et al. are consistent with evidence from the 1918 influenza pandemic, suggesting that school closures could have reduced the total number of infections by 15%, and the peak attack rates by 40% [13] [14] [15] [16] [17] .

Given the uncertainty of disease spread in schools, and the effect of keeping schools closed, we must evaluate decisions to reopen schools based on the state of the disease development in the actual country, region, and city. The decision to reopen schools in Norway followed an announcement on April 6 [18] , where the NIPH, presented an estimated R of 0.70 for Norway, with a 95% confidence interval (95%CI) of 0.45-1.

In this study we aimed to model the effects on after school opening in two Norwegian cities. We performed an ensemble of experiments in an individual-based model (IBM) that is similar to the one used by Ferguson et al. [1, 19, 20] and other complex transmission models [21] [22] [23] [24] [25] . There exist similar IBM studies that focus on the effects of NPIs on the spread of SARS-CoV-2 [26] . However, as far as we know, this is the first study that evaluates a controlled school reopening using this methodology, and no other study uses as detailed data as we have collected in our experiments. Moreover, we have used ensembles of model experiments to infer results that remain robust under the uncertainties about SARS-CoV-2 virus dynamics.

Oslo is the largest city and the epicenter of the COVID-19 outbreak in Norway. By May 22, the city had 3.73 confirmed cases per 1000 inhabitants, compared to 1.54 per 1000 in Norway as a whole. There were 2.18 cases per 1000 inhabitants by May 22 in the city of Tromsø.

The model used detailed demographic data for the two cities, statistics for Norwegian families, and detailed information about the network of city schools and kindergartens. We made a wide range of different assumptions about the transmission rates within and outside families and schools. The experiments carried out conformed with the Norwegian government's plan for school reopening and were evaluated by comparison with model simulations where the schools remain closed.

We built networks of families and schools that were consistent with demographic data and school data collected from Statistics Norway and the municipalities of the two cities [27, 28] . In the first step of the construction, we associated every infected individual with a family. Every child (under the age of 18 years of age) was associated with a kindergarten or a school. Consistent with demographic data, the number of families was 0.45 times the total population, 24% of families were couples without children, 18.5% of families comprised of two parents with children, and 7% of families were composed of one parent with children. For families with children, the average number of children was 1.75. The age of each child was drawn randomly according to the age distribution in each city. We assigned each family with children to one middle school, one elementary school, one high school, and one kindergarten. The middle schools and high schools were drawn randomly with probabilities proportional to the school sizes. The elementary schools were drawn randomly from the set of elementary schools associated with the middle school assigned to the family. Our model associated each elementary school with ten different kindergartens and assigned each family to a kindergarten associated with the family's elementary school. Depending on their age, we assigned children to their family's kindergartens, elementary school, middle school, or high school.

We modeled transmission in the networks stochastically. In each time step, we found the probability of infection for each individual by computing a force of infection for this individual. A weighted sum over three terms defined the force of infection. The first term was a weighted sum over infectious persons in each individual's family, the second over the infectious persons in each individual's school, and the third term over all infectious individuals in the city.

More precisely, in each time step → + Δ , a susceptible individual could be infected with probability ! = 1 − "# ! $% , where the force of infection 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 15, 2020. .

In the IBM we carried out a number of simulations with transmission coefficients ( ' , ) , and * ), each drawn randomly from uniform distributions on the interval 0.2-1. In the model experiments, schools remain closed for 14 days ( ) = 0), and then kindergartens opened. Grades 1-4 opened one week later. In the model, school openings are implemented by including the term ! ()) in the force of infection. The number of infectious individuals in the start of the simulations were taken to be 5000 and 300 in Oslo and Tromsø, respectively. This was roughly three times the number of confirmed cases in each city (at the time the simulations were set up) in order to account for unrecorded cases. Since our results concern changes in the reproductive number, and not the absolute number of cases, they are largely insensitive to the initial conditions. For the city of Oslo, we carried out 121 model experiments, and for the city of Tromsø we carried out 295. We varied model parameters to ensure that the results presented in this paper are not sensitive to uncertain assumptions.

We estimated R-values by fitting the number of infectious individuals to ( ) in simulations of a susceptible-exposed-infected-recovery (SEIR) model.

In the IBM we quantified the effect of school openings as the difference Δ between the estimated R-values in two simulations where schools were opened in one, and not in the other, but otherwise identical. Clearly, Δ depended on the in-school transmission rate ) . For IBMs of the type used here, the transmission rates estimated by Ferguson et al. [1] form reference values that have also been used in other modelling studies [29] . Ferguson et al. found the in-school transmission rate ) =0.94 per/day to match well with attack rates for children during the 1957-1958 influenza pandemic [30] . Based on this, we define a dimensionless relative in-school transmission rate = ) /(0.94 per day). The factor characterizes the in-school transmissibility of the SARS-CoV-2 compared to a pandemic influenza, i.e., = 1 corresponds to a pandemic influenza.

To estimate the basic reproduction number in Norway we used a one-population version of the SEIR model used by the NIPH [18] , and our estimation of R follows their approach (See Supplementary Material for details). The method was to assume fixed parameters (except ), and assume a stepwise time evolution of the reproduction number, with one constant value, (:) , before March 15, a second constant value, (') , between March 15 and April 20, and a third value, ()) , after April 20. The transitions were smoothed using the hyperbolic tangent function and two characteristic transition times that were estimated as a part of the procedure. Varying these five parameters, we minimized the square error (on a linear scale) between the time series of hospitalized patients with COVID-19 in Norway and the predicted number of hospitalized patients from the SEIR model. The NIPH has published the SEIR-parameters, the assumptions on the proportion of symptomatic individuals that require hospitalization, and assumptions on the length of hospitalizations [18] . The NIPH employs a metapopulation-version of the SEIR model, where each municipality in Norway is modelled separately and coupled via contact a matrix estimated from cellphone data provided by the telecommunication company Telenor, but we used the one-population version of the model. 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 15, 2020. . Uncertainty was estimated by repeated addition of normal distributed noise to the hospitalization data and re-estimation of parameters. We used a noise with a standard deviation twice as large as the sample standard deviation of the difference between the optimal fit and the hospitalization data. Data on the number of hospitalized patients with COVID-19 was downloaded from the The Norwegian Directorate of Health [6] on May 22.

As a supplement to this method we also applied a simple non-parametric method to the time series of confirmed infections in Norway, as well as in the cities of Oslo and Tromsø. The approach, which is described in the Supplementary Material, is similar to the method presented by Thomson et al. [31] . Figure 1 shows the development of the accumulated number of infections for selected sets of -parameters in our ensemble of simulations. The black curves represent simulations for a situation without school opening ( ) = 0), and red curves with school opening. The difference between the red and black curves, which represents the effect of school opening, is small for all examples except for the one shown in panel B. The reason for this is that R, after school opening, is less than unity for panels A, C, and D. For these examples, the epidemic decays even after the opening of school, and hence the opening has little consequence. For the example shown in panel B, however, the reproduction number R is less than unity without opening, and + Δ greater than unity with opening of schools. This is an example where school opening leads to a second wave of infection.

This risk of a new wave of infections depends on the magnitude of R prior to school opening. As will be shown below, we found that if in-school transmission rates for the SARS-CoV-2 virus are similar to what is believed to be the case for influenza pandemics ( = 1), we estimated for Oslo Δ = 0.10 (95%CI 0.04-0.16) and for Tromsø, Δ = 0.14 (95%CI 0.01-0.25). This means that the most interesting range of R prior to school opening to explore is the interval 0.6-1.0. The result of such an exploration is shown in Figure 2 . The three panels show the estimated R before April 20 and after April 27. For each city, the figures are based on the whole ensemble of simulations but weighted by a probability density p(R) for the reproduction number prior to April 20, which is shown in the left part of each panel, and to the right, the R-values after school opening on April 27. In the three panels for each city, the gaussian weight distribution is centered around R=0.6, 0.8, and 1.0, respectively, with standard deviation 0.1. The most interesting case is probably R centered around 0.8 (panels B and E), because this is close to the most probable R estimated for the two cities (see Supplementary Figure 1 ). In panel B, the probability of having + Δ > 1 after school opening in Oslo is estimated to be 0.13. In Tromsø, this probability is somewhat higher, 0.28, as shown in panel E. Since the magnitude of R prior to school opening is only known with considerable uncertainty (Supplementary Figure 1) , these probabilities are also associated with great uncertainty, but their magnitudes still provide an indication of the risk for a second wave of infection associated with school openings.

The variation in Δ in Figure 2 was mostly determined by variation in the relative inschool transmission rate . We found the relationship between Δ and to be approximately linear over the ensembles of simulations, and for both cities. We estimated Δ = 0.10 for Oslo, and Δ = 0.14 for Tromsø (Figure 3 ). The standard errors were 0.004 for 0.10, and 0.005 for 0.14. It is on the basis of these linear relationships that we found Δ = 0.10 and Δ = 0.14, assuming that in-school transmission rates are similar to what is believed to be the case for influenza pandemics ( = 1). If in-school transmission rates are 50% of the estimate for influenza pandemics, the estimated changes in the reproduction number are Δ = 0.05 and Δ = 0.07 for the two cities, respectively.

Another objective of this study was to assess if we can use the estimated R after school openings to infer the rates of transmission between children in schools. The simulations show that we are not able to make such inference. The main reason for this is that the linear relationships between Δ and have small proportionality factors (0.10 and 0.14), implying that uncertainty of in-school transmission rates is amplified relative to uncertainty in Δ (Supplementary Figure 3) . In order to constrain in-school transmission rates from the change (or lack of change) in R in Norway, we would need to estimate Δ with significantly higher accuracy than we can obtain from currently available hospitalization data using SEIR models (Figure 4) , or from non-parametric estimation methods of R (Supplementary Figure 1 ).

For the COVID-19 pandemic, early evidence from the outbreak in China showed that children of all ages are susceptible, but that disease, in general, is milder in children than for adults [33] . There is currently little concrete data on how the attack rate varies with age, and how common asymptomatic infections are in different age groups. Recent data from Wuhan shows that the secondary attack rate in households was as high for children as adults [10] , but it is still unclear to what extent children pass on the virus, and consequently, the children-to-children transmission rate is unknown [34] [35] [36] . It can be argued that if children have milder symptoms than adults, they may also be less infectious, but this is unclear at this point [37] . Children with mild symptoms are less likely to be absent from school while they are infectious, and it may be more difficult for children to conform to strict hygiene measures and social distancing. Our findings suggest that it is challenging to infer these data through assessments of countrywide school openings, even if one makes the unrealistic assumption that the change in R after April 20 is not influenced by other factors than schools opening.

Our model results suggest that the controlled opening of schools in Norway will lead to a change in R of less than 0.25, and most likely in the range 0.10 to 0.14. This holds true despite the uncertainty about the role of children in the spread of the SARS-CoV-2 virus. Since Norway had strongly suppressed the COVID-19 outbreak by the middle of March, a change in R of the order of 0.10 to 0.14 will not have had a strong effect on the number of infections.

A strength of our study is that the estimates are robust under a wide range of assumptions about transmission rates and reflects changes in the contact networks associated with school openings. We model these network changes using detailed data about the actual 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 15, 2020. . school structure in the two Norwegian cities. This aspect of the modeling is carried out in greater detail in this work than in the previous modeling of influenza pandemics [1, 19, 20] .

There is uncertainty in the reference value for in-school transmission rates [1] , but our results show low sensitivity to the transmission rates. The model results for the city of Oslo show a change in the R less than 0.25, even for in-school transmission rates which is 20% higher than the influenza reference value. Moreover, there is evidence that a significant proportion of children continue to mix with other children after unplanned school closures [38] . If we had included this effect, it would reduce the estimated effect of school openings. On the other hand, there are also trickle-down effects of school openings, such as increased contacts between adults (teachers and parents) that we have not taken into account.

Based on our modeling results, we conclude that controlled reopening of schools in countries that have the first wave of the COVID-19 pandemic suppressed will only have a limited effect on the reproduction number. The benefits to children's health of opening schools are well-documented [34, 35] . April 20th April 20th

April 20th

April 27th

April 27th April 27th

April 27th

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 15, 2020. . The value r=1 corresponds to the transmission rate 0.94 per day, which is taken as a rough estimate for the in-school transmission rate during an influenza pandemic [1] . The black points are model simulations for randomly selected -parameters, and the contours show the conditional probability density (Δ | ) estimated using the method in [32] . 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 15, 2020. . https://doi.org/10.1101/2020.08.14.20174896 doi: medRxiv preprint 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 15, 2020. . https://doi.org/10.1101/2020.08.14.20174896 doi: medRxiv preprint

Strategies for containing an emerging influenza pandemic in Southeast Asia

How will country-based mitigation measures influence the course of the COVID-19 epidemic? The Lancet

Estimating the number of infections and the impact of nonpharmaceutical interventions on COVID-19 in 11 European countries

School closure and management practices during coronavirus outbreaks including COVID-19: a rapid systematic review. The Lancet Child & Adolescent Health

COVID-19 Educational Disruption and Response

The Norwegian Directorate of Health. Smittevernveiledere for barnehager og skoler

Characteristics of and Important Lessons From the Coronavirus Disease 2019 (COVID-19) Outbreak in China: Summary of a Report of 72 314 Cases From the Chinese Center for Disease Control and Prevention

Daily report and statistics about coronavirus and COVID-19

Estimates of the severity of coronavirus disease 2019: a model-based analysis. The Lancet Infectious Diseases

Epidemiology and transmission of COVID-19 in 391 cases and 1286 of their close contacts in Shenzhen, China: a retrospective cohort study. The Lancet Infectious Diseases

Estimating the impact of school closure on social mixing behaviour and the transmission of close contact infections in eight European countries

Changes in contact patterns shape the dynamics of the COVID-19 outbreak in China

The effect of public health measures on the 1918 influenza pandemic in U.S. cities

Quantifying social distancing arising from pandemic influenza

Closure of schools during an influenza pandemic. The Lancet. Infectious diseases

Public health interventions and epidemic intensity during the 1918 influenza pandemic

Nonpharmaceutical Interventions Implemented by US Cities During the 1918-1919 Influenza Pandemic

Covid-19-epidemien: Kunnskap, situasjon, prognose, risiko og respons i Norge etter uke 14

Strategies for mitigating an influenza pandemic

Impact of non-pharmaceutical interventions (NPIs) to reduce COVID19 mortality and healthcare demand

FluTE, a Publicly Available Stochastic Influenza Epidemic Simulation Model

Modelling disease outbreaks in realistic urban social networks

An influenza epidemic model with dynamic social networks of agents with individual behaviour

Containing Pandemic Influenza at the Source. Science

Multiscale mobility networks and the spatial spreading of infectious diseases

Interventions to mitigate early spread of SARS-CoV-2 in Singapore: a modelling study

Simulating school closure strategies to mitigate an influenza epidemic

Reviewing the History of Pandemic Influenza: Understanding Patterns of Emergence and Transmission. Pathogens

Improved inference of time-varying reproduction numbers during infectious disease outbreaks

Emergent constraint on equilibrium climate sensitivity from global temperature variability

Epidemiology of COVID-19 Among Children in China

School Reopening-The Pandemic Issue That Is Not Getting Its Due

Rapid risk assessment: Outbreak of novel coronavirus disease 2019 (COVID-19): increased transmission globally -fifth update

An analysis of SARS-CoV-2 viral load by patient age

The impact of unplanned school closure on children's social contact: rapid evidence review