key: cord-0915275-8f76vhyz authors: Davies, Nicholas G; Klepac, Petra; Liu, Yang; Prem, Kiesha; Jit, Mark; Eggo, Rosalind M title: Age-dependent effects in the transmission and control of COVID-19 epidemics date: 2020-03-27 journal: nan DOI: 10.1101/2020.03.24.20043018 sha: be06a4f1d96c62391496a8f06a3d00cdf0ae0f9c doc_id: 915275 cord_uid: 8f76vhyz The COVID-19 pandemic has shown a markedly low proportion of cases among children. Age disparities in observed cases could be explained by children having lower susceptibility to infection, lower propensity to show clinical symptoms, or both. We evaluate these possibilities by fitting an age-structured mathematical model to epidemic data from six countries. We estimate that clinical symptoms occur in 25% (95% CrI: 19-32%) of infections in 10-19-year-olds, rising to 76% (68-82%) in over-70s, and that susceptibility to infection in under-20s is approximately half that of older adults. Accordingly, we find that interventions aimed at children may have a relatively small impact on total cases, particularly if the transmissibility of subclinical infections is low. The age-specific clinical fraction and susceptibility we have estimated has implications for the expected global burden of COVID-19 because of demographic differences across settings: in younger populations, the expected clinical attack rate would be lower, although it is likely that comorbidities in low-income countries will affect disease severity. Without effective control measures, regions with older populations may see disproportionally more clinical cases, particularly in the later stages of the pandemic. / infections, as shown for influenza 22 . Subclinical cases are sometimes called "asymptomatic" but very mild symptoms may not be noticed, even though they occur. Distinguishing which of these hypotheses is most supported by available data has important implications for policies that aim to control transmission 23 We tested three hypotheses-(1) no age variation in susceptibility or severity, with the age distribution of cases driven by age-dependent contact patterns alone; (2) varying susceptibility to infection by age; and (3) varying clinical fraction by age-using an age-structured dynamic transmission model (see Methods) with heterogeneous contact rates between age groups drawn from social contact surveys in Shanghai 17 . We generated model variants for each hypothesis ( Fig 1a ) and fitted to three data sources from the early epidemic in Wuhan for each hypothesis ( Fig, 1b, 1c ). We included school closures, for which we decreased the school contacts of children. We also estimated the effect of the holiday period, and the travel and movement restrictions in Wuhan, on transmission ( Fig 1d ) . We found that under each hypothesis, the basic reproduction number R 0 was 3.2-3.6 initially, was inflated 1.2-1.5-fold during the pre Lunar New Year holiday period, and then fell by 80-95% during restrictions in Wuhan, which brought R 0 below 1 ( Fig 1e ) . All model variants fitted the daily incident number of confirmed cases equally well ( Fig 1f ) . However, hypothesis 1 did not reproduce the observed age distribution of cases, overestimating the number of cases in children and underestimating cases in older adults 4 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. . https://doi.org/10.1101/2020.03. 24.20043018 doi: medRxiv preprint / ( Fig 1g ) . Hypotheses 2 and 3 fitted the age distribution of cases, but each implied a very different fraction of subclinical infections by age, with much higher numbers of subclinical infections under hypothesis 2 (Fig 1h) . Comparison using Deviance Information Criterion 6 (DIC) showed that hypothesis 2 (DIC: 691) and 3 (DIC: 558) were preferred over hypothesis 1 (DIC: 880), and hypothesis 3 was better supported than hypothesis 2. Under hypothesis 2, where the severity was equal by age, 20% of both clinical and subclinical infections occurred in the 70-100 year old age group ( Fig 1h ) . Under hypothesis 3, 20% of clinical cases but less than 5% subclinical cases are in this group. Recent work has demonstrated an age-dependent severity in hospitalised confirmed cases 24, 25 , which suggests that a high rate of subclinical infection in older adults may not be realistic. Close follow-up of contacts of cases in Shenzhen, China, found that children were infected at the same rate as adults 16 , lending more weight to hypothesis 3. Additionally, evidence of increased severity by age 19 suggests that clinical signs are more likely in older adults, which further decreases the plausibility of hypothesis 2. . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. . https://doi.org/10.1101/2020.03. 24.20043018 doi: medRxiv preprint . (b) Susceptibility by age for the three hypotheses. Age-specific values were estimated for hypothesis 2 and all ages had equal susceptibility for hypothesis 1 and 3. Susceptibility is measured as the probability of infection on contact with an infectious person. (c) Clinical fraction ( y i ) by age for the 3 hypotheses. Age-specific values were estimated for hypothesis 3 and all ages were equal for hypothesis 1 and 2. (d) Fitted contact multipliers for holiday and restricted periods for each hypothesis showed an increase in non-school contacts beginning on January 12th (start of Lunar New Year) and a decrease in contacts following restrictions on January 23rd. (e) Estimated R 0 values under each hypothesis. The red barplot shows the inferred window of spillover of infection. (f) Incident reported cases (black), and modelled incidence of clinical cases for the three hypotheses as fitted to the cases reported by China Centers for Disease Control 1 with onset on or before February 1st, 2020. Line marks mean and shaded window is the 95% highest density interval (HDI). (g) Age distribution of cases by onset date as fitted to the age distributions reported by Li et al. 26 Data are shown in the hollow bars, and model predictions in filled bars, where the dot marks the mean posterior estimate. (h) Inferred distribution of subclinical cases by age under each hypothesis. Credible intervals on modelled values show the 95% HDIs; credible intervals on data for panels d-f show 95% HDIs for the proportion of cases in each age group. Since the initial outbreak in Wuhan, the virus has spread to other regions within China and internationally. Local epidemics have exhibited a less extreme, but still marked lack of reported cases among children. The expected proportion of children infected depends on 6 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. . https://doi.org/10.1101/2020.03. 24.20043018 doi: medRxiv preprint / mitigation measures in place in each region, and is expected to be lower in regions which have closed schools. Using the best fitting and most biologically plausible hypothesis, hypothesis 3 -age-varying clinical fraction -we estimated the age-specific clinical fraction for 32 settings across six countries by using the stationary distribution of the next generation matrix to reproduce the locally-reported age distribution of cases compiled from a variety of sources (Fig 2a) . We used setting-specific demographics, measured contact matrices where possible, and synthetic contact matrices otherwise 27 . The age-dependent clinical proportion was markedly lower in younger age groups in all regions (Fig 2b) , with 20% of infections in children under 10 resulting in clinical cases, rising to over 70% in adults over 70 in the consensus age distribution estimated across all regions. To determine whether this distribution was capable of reproducing epidemic dynamics, we fitted our dynamic model to the incidence of clinical cases in Beijing, Shanghai, South Korea and Italy ( Fig 2c ) . The consensus age-specific clinical fraction was largely capable of reproducing the age distribution of cases, although there are some outliers, for example the 20-30 age group in South Korea. This could be the result of clustered transmission within a church group in this country 4 . The predicted age distribution of cases for Italy is also less skewed towards older adults than reported cases show, suggesting potential differences in age-specific testing in Italy 28 . Locally-estimated age-varying clinical fraction captured these patterns more precisely (Fig. 2c is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint Fitted mean and 95% HDI for the age distribution in clinical fraction for all countries. (c) Fitted incidence of confirmed cases and resulting age distribution of cases using either the consensus (grey) or country-specific (colour) age-specific clinical fraction from b. School closures during epidemics 29, 30 and pandemics 31, 32 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. . https://doi.org/10.1101/2020.03. 24.20043018 doi: medRxiv preprint / and Bulawayo (Zimbabwe, low median age) (Fig 3b) , using measured contact matrices for each country. There were many more clinical cases for COVID-19 than influenza in all cities, with relatively more cases in children in the influenza-like scenario, and more cases in adults in simulated COVID-19 epidemics ( Fig 3c) . More clinical cases were seen in older adults in Milan compared with the other cities, and a markedly younger age distribution in clinical cases in Bulawayo. Using the same age-dependent clinical fraction drawn from high and upper-middle income countries in low and lower-middle income countries (LMIC) may underestimate clinical cases due to the presence of comorbidities. To fully explore the effect of school closure we simulated 3 months of school closures with varying infectiousness of subclinical cases, at either 0, 0.25, 0.5 or 0.75 times the infectiousness of clinical cases ( Fig 3d) . We found that school closures decreased peak incidence slightly for influenza-like infections, and delayed the peak substantially. For COVID-19 epidemics, the delay and decrease of the peak was smaller, and this was especially the case in Bulawayo, which has the highest proportion of children (Fig 3e) . Because children exhibit more subclinical cases for COVID-19, school closures were more effective at reducing transmission of COVID-19 when the subclinical infectiousness was assumed to be higher (Fig 3f) . . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. Age-specific attack rate for COVID-19 and influenza-like infections, assuming 50% subclinical infectiousness. (d) Daily incidence of clinical cases in exemplar cities for COVID-19 versus influenza-like infections. R 0 is fixed at 2.4. The rows show the impact of varying the infectiousness of subclinical infections to be 0%, 25%, 50%, or 75% as infectious as clinical cases while keeping R 0 fixed. (e) Change in peak timing and peak cases for the three cities, for either COVID-19 or pandemic influenza. (f) Change in median COVID-19 peak timing and peak cases for the three cities,depending on the infectiousness of subclinical infections. Strong age dependence in the fraction of COVID-19 infections that become clinical cases has implications for the projected global burden. Simulating COVID-19 outbreaks in 146 capital cities, we found that the total expected number of clinical cases in an unmitigated epidemic varied between countries depending on the median age of the population, which is a proxy for the age structure of the population. The total clinical attack rate was higher, and the peak height of the epidemic was greater, in older populations ( Fig 4a ) . By contrast, the number and peak of subclinical infections was lower in older populations ( Fig 4b ) . The mean estimated basic reproduction number, R 0 , was higher in cities with a lower median age ( Fig 4c ) , because of the greater proportion of children and the higher number of contacts made by children compared to adults. We applied the same age-dependent clinical fraction to all 10 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. . https://doi.org/10.1101/2020.03.24.20043018 doi: medRxiv preprint / countries, but the relationship between age and clinical symptoms may be different in different countries, perhaps because of the age distribution of comorbidities, or the presence of other possible comorbidities, such as HIV 33 . If the relationship between clinical fraction and age skews younger in low and lower-middle income countries, there would be higher clinical attack rates in these countries (Supplementary Section 4). The expected age distribution of cases shifted substantially over time, where in the early phase of the epidemic, the clinical case distribution tended to be skewed to younger ages, and the late phase showed more cases in older individuals ( Fig 4d ) . This impacts projections for likely healthcare burdens at different phases of the epidemic, particularly because older individuals tend to have higher healthcare utilisation on infection 1 ( Fig 4e ) . is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint We have shown a strong age dependence in the probability of displaying clinical symptoms for COVID-19, from around 20% in under 10s, to over 70% in older adults. Given evidence of a stark age gradient in severity 8 and mortality 21, 28 , and recent studies of close follow up of children at risk of infection 16 showing that infection was frequent in all age groups, the plausibility of age-specific severity is higher than age-specific susceptibility to infection. For a number of other pathogens, there is evidence that children (except for the very youngest) have lower rates of symptomatic disease 12 and mortality 26 . For these reasons, we find that age-specific clinical fraction is more supported than age-specific susceptibility to infection. Serological surveys will provide critical information on the true distribution of subclinical infections. The age-specific distribution of subclinical infection we have found is similar in shape (but larger in scale) to that generally assumed for pandemic influenza. However, for the 2009 influenza A/H1N1p pandemic, the age-specific susceptibility to infection was lower in older individuals compared to COVID-19. These differences have a large effect on how effective school closures may be in limiting transmission, delaying the peak of expected cases, and decreasing the total and peak number of cases. For COVID-19, school closures are likely to be much less effective than for influenza-like infections where children play a more substantial role in transmission. It is critical to determine how infectious subclinical infections are compared to clinical infections in order to properly assess predicted burdens both with and without interventions. It is biologically plausible that milder cases are less transmissible, for example, because of an absence of cough 28, 29 , but direct evidence is limited 35 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. . https://doi.org/10.1101/2020.03. 24.20043018 doi: medRxiv preprint / overall burden in the groups at risk of severe disease, primarily older populations for COVID-19 21 , remains high. If those with subclinical infections are (much) less efficient at transmitting, then the overall contribution to the burden of clinical disease in the population may be proportionally lower. At the same time, lower relative infectiousness would reduce the impact of interventions targeting younger ages, such as school closure. By analysing epidemic dynamics before and after school closures, it may be possible to estimate the infectiousness of subclinical infections, however this will rely on granular data by age and time. A great deal of concern has been directed toward the expected burden of COVID-19 in low and middle income countries (LMIC), which have lower population median age than many high income countries. Our results show that these demographic differences, coupled with a lower symptomatic fraction in younger ages, can result in proportionally fewer clinical cases than would be expected in higher-income countries with flatter demographic pyramids. This should not be interpreted as few cases in LMIC, because the projected epidemics are still very large, resulting in high numbers infected. Moreover, the particular relationship found with age here is drawn from high income countries, primarily in East Asia, and may reflect not only age, but also the increasing frequency of comorbidities with age. This relationship, therefore, may differ in LMIC for two key reasons: first, the distribution of non-communicable comorbid conditions-which are already known to increase the risk of severe disease from COVID-19 21 may be differently distributed by age, often occurring in younger age groups 34 , along with other possible risk factors such as undernutrition 36 ; and second, communicable comorbidities such as HIV 33 , TB coinfection (which has been suggested to increase risk 37 ), and others 38 may alter the distribution of severe outcomes by age. Observed severity and burden in LMIC may also be higher due to a lack of health system capacity for intensive treatment of severe cases. . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. . https://doi.org/10.1101/2020.03.24.20043018 doi: medRxiv preprint / There are some limitations to the study. The true explanation for the age distribution could be a combination of age-specific susceptibility and clinical fraction, although some recent studies indicate children are infected at similar 24 , or slightly lower rates 39 than adults, and children are not commonly spared from other coronavirus infections 40, 41 . It is not possible to simultaneously estimate both effects from available data, so we were unable to validate a mixture model. While information drawn from the early stages of the epidemic are subject to uncertainty, age-specific information is drawn from several regions and countries, and clinical studies support the hypothesis presented here. We assumed that clinical cases are reported at a fixed fraction throughout the time period, although there may have been changes in reporting. We assumed that subclinical infections were less infectious than clinical infections but were not able to estimate how infectious subclinical infections were, instead testing the sensitivity of our findings to this parameter. We have used mixing matrices from the same country, but not the same location as the fitted data. We used contact matrices that combined physical and conversational contacts. We therefore implicitly assume that they are a good reflection of contact relevant for the transmission of SARS-CoV-2. If fomite, or faecal-oral routes of transmission are important in transmission, these contact matrices may not be representative of transmission risk. The role of age in transmission is critical to designing interventions aiming to decrease transmission in the population as a whole, and to projecting the expected global burden. Early evidence 24 , including presented here, suggests that there is age dependence in the is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint We use an age-structured deterministic compartmental model (Fig. 1a , main text) stratified into 5-year age bands, with time approximated in discrete steps of 0.25 days. We assume that people are initially susceptible (S), and become exposed (E) after effective contact with an infectious person. After an incubation period, exposed individuals either develop a clinical or subclinical infection; an is the susceptibility to infection of an agei individual, is the number of agej u i c ij,t individuals contacted by an agei individual per day at time t , f is the relative infectiousness of a subclinical case, and is the effective probability that a random agej I I ) N ( P j + I C j + f S j / j 15 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint To calculate the basic reproductive number, R 0 , we define the next generation matrix as R 0 is the absolute value of the dominant eigenvalue of the next generation matrix. We use the local age distribution for each city or region being modelled, and synthetic or measured contact matrices for mixing between age groups ( is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint We used mixing matrices measured in Italy in 2005/2006 11 , adapted to the demographics of each region / city. This assumes that these contact patterns will still be representative of contact patterns in 2020. We used synthetic contact matrices, generated based on demographic information about the country 27 . We used synthetic contact matrices, generated based on demographic information about the country 27 . We used synthetic contact matrices based on demographic information about the country 27 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint We contrasted three possible hypotheses. In hypothesis 1, there were no age-related differences in susceptibility ( ) or symptomatic fraction ( ). In hypothesis 2, u i = u y i = y susceptibility varied by age ( ), but the proportion of exposed individuals who became (i) u i = u clinical cases did not vary by age ( ). In hypothesis 3, the clinical case probability varied y i = y by age ( ), but susceptibility did not ( ). Susceptibility and clinical fraction curves (i) y i = y u i = u were fitted using three control points for young, middle, and old age, interpolating between them with a half-cosine curve (see Methods for details). We assumed that the initial outbreak in Wuhan was seeded by introducing one exposed is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. reported by China CDC 1 . Because there was a large spike of incident cases reported on February 1 determined to have originated from the previous week, we amalgamated all cases from January 25 to February 1, including those in the large spike, into a single data point for the week. We assumed 10% of clinical cases were reported 19 . We used a Dirichlet distribution with a flat prior to obtain 95% HDIs for reported case data stratified by age group for display in figures. We used Markov-chain Monte Carlo to jointly fit each hypothesis to the two sets of empirical observations from the epidemic in Wuhan City, China. We used a negative binomial likelihood for incident cases and a Dirichlet-multinomial likelihood for the age distribution of cases, using the likelihood L = egBinom(C |size 00, ean ) Above, C k is the observed incidence on day k while c k is the model-predicted incidence for day k, for each of K days. A m is the observed age distribution for time period m (case counts for each age group) while a m is the model-predicted age distribution for the same period, and is the total number of cases over all age groups in time period m , measured for M time |a || | m periods. We set the precision of each distribution to 200 to capture additional uncertainty in data points that would not be captured with a Poisson or multinomial likelihood model. . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint Table 3 . Posterior estimates and 95% highest density intervals for parameters fitted to data from Wuhan, China. To infer age-specific susceptibility from reported case distributions, we assumed that reported cases follow the stationary distribution of cases reached in the early phase of an epidemic. Using our dynamic model would allow modelling any transient emphasis in the case distribution associated with the age of the individuals who seeded infection in a given region, but since the age of the true first cases is not generally known, we used the stationary distribution instead. Specifically, we used Bayesian inference to fit age-specific susceptibility to the reported case distribution by first generating the expected case distribution k i from (1) is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. . https://doi.org/10.1101/2020.03. 24.20043018 doi: medRxiv preprint / dates available; accordingly, we fit the proportion of confirmed cases with onset dates and the delay from onset to confirmation. We adjusted the size parameter of the negative binomial distribution used to model case incidence to 10 to reflect greater variability among fewer data points for these countries than for Wuhan. Beijing and Shanghai were fitted jointly, with separate dates of introduction but the same fitted susceptibility, large-scale restriction date and large-scale restriction magnitude. South Korea and Italy were each fitted separately; we fitted a large-scale restriction date and magnitude for both South Korea and Italy. For both the linelist fitting and validation, we assumed that schools were closed in China, but remained open in South Korea, Japan, Italy, Singapore, and Canada, as schools were open for the majority of the period covered by the data in the latter five countries. To determine the impact in other cities with different demographic profiles we used the inferred parameters from our linelist analysis to parameterise our transmission model for projections to other cities. We chose these to compare projections for a city with a high proportion of elderly individuals (Milan, Italy); a moderate-aged population (Birmingham, United Kingdom); and a city in a low-income country with a high proportion of young individuals (Bulawayo, Zimbabwe). For this analysis, we compared an outbreak of COVID-19, for which the burden and transmission is concentrated in relatively-older individuals, with an outbreak of pandemic influenza, for which the burden and transmission is concentrated in relatively-younger individuals. We assumed that immunity to influenza builds up over a person's lifetime, such that an individual's susceptibility to influenza infection plateaus at roughly age 35, and assumed that the severity of influenza infection is highest in the elderly and in children under 10 years old 7 . . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint 46 . We assumed that the epidemic was seeded by two infectious individuals in a random age group per week for 5 weeks to. We scaled the age-specific susceptibility u i by setting the "target" basic reproductive number R 0 = 2.4 as an illustrative example. We also performed a sensitivity analysis where we scaled u i to result in R 0 = 2 in Birmingham, and using the same setting for u i in all three cities, so that the actual R 0 changed depending upon contact matrices and demographics used to model each city. We projected the impact of school closure by setting the contact multiplier for school contacts school(t) to 0. Complete removal of school contacts may overestimate the impact of school closures because of alternative contacts children make when out of school 57 . This will however give the maximum impact of school closures in the model to demonstrate the differences. To project the impact of COVID-19 outbreaks in global cities, we used mixing matrices from Prem et al. 27 and demographic structures for 2020 from World Population Prospects 2019 to simulate a COVID-19 outbreak in 146 global capital cities for which synthetic matrices, demographic structures and total populations were available. For simplicity, we assumed that capital cities followed the demographic structure of their respective countries and took 24 . CC-BY 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. . https://doi.org/10.1101/2020.03. 24.20043018 doi: medRxiv preprint / the total population of each capital city from the R package maps . For each city, we scaled u i to result in an average R 0 = 2 in Birmingham, UK, and used the same setting for u i for all cities, so that the realised R 0 would change according to the contact matrices and demographics for each city. We simulated 20 outbreaks in each city, drawing the age-specific clinical fraction y i from the posterior of the estimated overall clinical fraction from our line list analysis (Fig. 2) , and analysed the time to the peak incidence of the epidemic, the peak clinical and subclinical incidence of infection, and the total number of clinical and subclinical infections. We took the first third and the last third of clinical cases in each city to compare the early and late stages of the epidemic. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. . https://doi.org/10.1101/2020.03. 24.20043018 doi: medRxiv preprint The epidemiological characteristics of an outbreak of 2019 novel coronavirus diseases (COVID-19) in China. Chin Early epidemiological analysis of the coronavirus disease 2019 outbreak based on crowdsourced data: a population-level observational study The early phase of the COVID-19 outbreak in Transmission potential and severity of COVID-19 in South Korea Epidemiology, transmission dynamics and control of SARS: the 2002-2003 epidemic Epidemiology of 2009 Pandemic Influenza A (H1N1) in the United / States Age, influenza pandemics and disease dynamics Epidemiological Characteristics of 2143 Pediatric Patients With Coronavirus Disease in China Incidence, clinical characteristics and prognostic factor of patients with COVID-19: a systematic review and meta-analysis Clinical features of patients infected with 2019 novel coronavirus in Wuhan Social Contacts and Mixing Patterns Relevant to the Spread of Infectious Diseases Estimating the impact of school closure on influenza transmission from Sentinel data Measured Dynamic Social Contact Patterns Explain the Spread of H1N1v Influenza World Population Prospects -Population Division -United Nations Singapore COVID-19 Cases Increased Risk of Noninfluenza Respiratory Virus Infections Associated With Receipt of Inactivated Influenza Vaccine Age-related prevalence of common upper respiratory pathogens, based on the application of the FilmArray Respiratory panel in a tertiary hospital Rates of asymptomatic respiratory virus infection across age groups Clinical course and risk factors for mortality of adult inpatients with COVID-19 in Wuhan The Impact of Illness on Social Networks: Implications for Transmission and Control of Influenza Defining the Epidemiology of Covid-19 -Studies Needed Epidemiology and Transmission of COVID-19 in Shenzhen China: Analysis of 391 cases and 1,286 of their close contacts Epidemiological and clinical features of the 2019 novel coronavirus outbreak in China Early Transmission Dynamics in Wuhan, China, of Novel Coronavirus-Infected Pneumonia Projecting social contact matrices in 152 countries using Case-Fatality Rate and Characteristics of Patients Dying in Relation to COVID-19 in Italy Control of Severe Acute Respiratory Syndrome in Singapore Monitoring community responses to the SARS epidemic in Hong Kong: from day 10 to day 62 School closures during the 2009 influenza pandemic: national and local experiences Closure of schools during an influenza pandemic Severe Influenza-associated Respiratory Infection in High HIV Prevalence Setting Exhaled Mycobacterium tuberculosis output and detection of subclinical disease by face-mask sampling: prospective observational studies Determining the Provincial and National Burden of Influenza-Associated Severe Acute Respiratory Illness in South Africa Using a Rapid Assessment Methodology Active or latent tuberculosis increases susceptibility to COVID-19 and disease severity Potential Impact of Co-Infections and Co-Morbidities Prevalent in Africa on Influenza Severity and Frequency: A Systematic Review Age profile of susceptibility, mixing, and social distancing shape the dynamics of the novel coronavirus disease 2019 outbreak in China Coronavirus infections in hospitalized pediatric patients with acute respiratory tract disease Characterization of Human Coronavirus OC43 and Human Coronavirus NL63 Infections Among Hospitalized Children <5 Years of Age Incubation period of 2019 novel coronavirus (2019-nCoV) infections among travellers from Wuhan, China Early dynamics of transmission and control of COVID-19: a mathematical modelling study The Rate of Underascertainment of Novel Coronavirus (2019-nCoV) Infection: Estimation Using Japanese Passengers Data on Evacuation Flights Patterns of human social contact and contact with animals in Shanghai Social Contact Structures and Time Use Patterns in the Manicaland Province of Zimbabwe China Statistical Year Book Bayesian measures of model complexity and fit An overview of composite likelihood methods ET008/edit?usp=sharing Open access epidemiological data from the COVID-19 outbreak Population Statistics, Charts, Map and Location Age breakdown of the population of Birmingham -Office for National Statistics Population Statistics, Charts, Map and Location School's Out: Seasonal Variation in the Movement Patterns of School Children We acknowledge the following for funding: NGD: National Institutes of Health Research Author contributions RME conceived the study. NGD and RME designed the model with PK, and YL, KP and MJ providing input. NGD designed the software and inference framework and implemented the model. YL processed the data. NGD and RME wrote the first draft of the manuscript. All authors interpreted the results, contributed to writing, and approved the final version for submission. The data used for fitting are publicly available, but will also be made available with the code in the github repository for the project. Contact matrix data are available at zenodo 21, 22 . The authors have no competing interests. Supplementary Information is available for this paper. Correspondence and requests for materials should be addressed to Rosalind M Eggo or Nicholas G Davies at r.eggo@lshtm.ac.uk or nicholas.davies@lshtm.ac.uk