key: cord-0726947-9iaiiqqi authors: Scott, N.; Palmer, A.; Delport, D.; Abeysuriya, R.; Stuart, R.; Kerr, C. C.; Mistry, D.; Klein, D. J.; Sacks-Davis, R.; Heath, K.; Hainsworth, S.; Pedrana, A.; Stoove, M.; Wilson, D. P.; Hellard, M. title: Modelling the impact of reducing control measures on the COVID-19 pandemic in a low transmission setting date: 2020-06-12 journal: nan DOI: 10.1101/2020.06.11.20127027 sha: edcf9e16c1e2ae85d55fe951cf1c4a8d6ccbdac8 doc_id: 726947 cord_uid: 9iaiiqqi Aims: We assessed COVID-19 epidemic risks associated with relaxing a set of physical distancing restrictions in the state of Victoria, Australia - a setting with low community transmission - in line with a national framework that aims to balance sequential policy relaxations with longer-term public health and economic need. Methods: An agent-based model, Covasim, was calibrated to the local COVID-19 epidemiological and policy environment. Contact networks were modelled to capture transmission risks in households, schools and workplaces, and a variety of community spaces (e.g. public transport, parks, bars, cafes/restaurants) and activities (e.g. community or professional sports, large events). Policy changes that could prevent or reduce transmission in specific locations (e.g. opening/closing businesses) were modelled in the context of interventions that included testing, contact tracing (including via a smartphone app), and quarantine. Results: Policy changes leading to the gathering of large, unstructured groups with unknown individuals (e.g. bars opening, increased public transport use) posed the greatest risk, while policy changes leading to smaller, structured gatherings with known individuals (e.g. small social gatherings) posed least risk. In the model, epidemic impact following some policy changes took more than two months to occur. Model outcomes support continuation of working from home policies to reduce public transport use, and risk mitigation strategies in the context of social venues opening, such as >30% population-uptake of a contact-tracing app, physical distancing policies within venues reducing transmissibility by >40%, or patron identification records being kept to enable >60% contact tracing. Conclusions: In a low transmission setting, care should be taken to avoid lifting sequential COVID-19 policy restrictions within short time periods, as it could take more than two months to detect the consequences of any changes. These findings have implications for other settings with low community transmission where governments are beginning to lift restrictions. Following a rise in COVID-19 cases, in March 2020 the Australian government introduced mandatory quarantine periods for people returning from overseas, as well as a variety of physical distancing policies, including closing pubs, bars, entertainment venues, churches/places of worship, restricting restaurants and cafes to take-away only, and limiting public gatherings to two people [1] . Two months after these policies were introduced, available epidemic data indicate that they were successful in disrupting the spread of COVID-19, with fewer than 55 cases per day diagnosed between 12 April and 8 May, down from a peak of 469 diagnosed cases on 28 March [2, 3] . Acknowledging that maintaining these restrictions for an extended period would be socially and economically unfeasible, the federal government released a framework for a COVIDSAFE Australia [4] on 8 May. This framework outlined a three-step sequence of policy options across the education, retail, hospitality, sport and health sectors that would enable their reopening under sustainable conditions, and allowed states and territories to adopt different policies at different times based on their specific COVID-19 conditions. The federal and jurisdictional governments also implemented public health measures to mitigate risks associated with relaxing policies, including a scale-up of testing capacity and the release of the COVIDSafe smartphone app, which records community contacts via Bluetooth connection to enable contact tracing in the event that someone with the app is diagnosed. For countries entering COVID-19 response phases that involve relaxing restrictions, it is important to carefully consider the sequence and timing of relaxing policies so as not to compromise the overall effectiveness of the response. Each component that is relaxed will increase the risk of COVID-19 transmission, however it is currently unclear under which conditions the increased risk would be sufficient to allow an uncontrolled outbreak. Epidemic modelling is a crucial tool that can provide insight into the likely specific and combined impact of relaxing individual control measures, as well as the time required to monitor and observe the impact of relaxing these measures. Epidemic models can be broadly classified as either population-level or individual-level. Populationlevel models are simpler and faster to produce, but they can only provide population-average estimates of the impact of policy changes. This is because they divide a population into a small number of discrete risk categories and assume homogeneous mixing and transmission risks within each category. In contrast, agent-based models use a set of autonomous 'agents' to represent a population and offer a more complex method for simulating individual-level characteristics and human behaviour [5] . In reality, the risk of COVID-19 transmission is highly heterogeneous and driven by the contact networks of individuals, which are dependent on age, household structure and participation in different social and community activities. Moreover, the impact of interventions to slow the spread of . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint COVID-19, such as contact tracing and quarantine measures, are highly contact network dependent. Interventions that are specific to contact networks can only be captured in individual-level models. To our knowledge, no modelling is currently available for Australia that provides scenario analyses of the impact of "micro-policy" changes being proposed in the COVIDSAFE Australia framework. Population-level models [6] [7] [8] [9] have been useful in informing a "flatten the curve" narrative and supporting the initial roll-out of physical distancing policies. Agent-based models are increasingly being used to simulate the impact of social distancing measures on COVID-19 transmission (either influenza models that have been modified [10] [11] [12] [13] , including for Australia [11, 14] , or newly developed COVID-19 models [15] [16] [17] ); however these models are currently only considering the implementation of contact tracing, quarantine or social distancing policies rather than their release. In this study we used an agent-based model, Covasim [18] , to assess the risks associated with relaxing various physical distancing and lockdown policies in Victoria, Australia. In this low transmission settings, we estimate how effective interventions need to be to mitigate these risks. Providing modelled estimates for these policy changes will be crucial, not just for Australia but globally in other settings with low transmission, and as countries begin to reopen. We modelled the state of Victoria, Australia. Victoria is Australia's second most populous state with an estimated population of 6.63 million (~26% of the nation's total) [19] , and age structure as shown in Appendix A, Figure S1 . As of 17 May, Victoria had 1,561 confirmed COVID-19 cases, 835 (53%) of which were acquired through overseas travel, and 18 COVID-19-related deaths [20] . The Victorian epidemic has followed a similar trajectory to Australia as a whole; an increase in daily new diagnoses throughout March to a peak of 111 on 29 March, followed by a decline as various restrictions (outlined in Appendix B) were imposed. Throughout May there were two small "cluster" outbreaks detected at workplaces, where geographically isolated transmission occurred through known contacts leading to approximately 5-20 new diagnoses per day from 1 May to 17 May. . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint The Covasim model is described in detail on medrxiv [18] and has been applied to a number of high, middle and low-income settings, including a number of states in the USA and countries across Africa [21] . In brief, each person in the model is characterised by a set of demographic, disease and intervention status variables. Demographics variables include: age (one-year brackets); uniquely identified household contacts; uniquely identified school contacts (for people aged 5-18); uniquely identified work contacts (for people aged 18-65); and average number of daily community contacts in a collection of community networks and settings (described in the following sections). Disease variables include: infection status (susceptible, exposed, recovered or dead); viral load (time-varying); age-specific susceptibility (Table S2) ; and age-specific probabilities of being symptomatic, experiencing different disease severities (mild, severe, critical), and mortality (Table S2) . Person-level intervention status variables include: diagnostic status (untested, tested and waiting for results, tested and received results) and quarantine status (yes/no). Transmission is modelled to occur when a susceptible individual is in contact with an infectious individual through one of their contact networks. The per-day probability of transmission per contact with an infected person ("transmissibility") is calibrated to match the epidemic dynamics observed, and is weighted according to whether the infectious individual has symptoms, and the type/setting of the contact (e.g. household contacts are more likely to result in transmission than community contacts). The model allows people to be a part of multiple independent contact networks. Within each network, a "contact" is a link between two people indicating that transmission would be possible if one of them were infected. The model is designed so that each individual can be a part of an arbitrary number of contact networks used to approximate transmission dynamics associated with different activities or specific public spaces. For this analysis, we considered networks and settings most likely to be subject to a policy change in Australia, with contact networks explicitly modelled for: households; schools; workplaces; social networks; cafés and restaurants; pubs and bars; public transport; places of worship; professional sport; community sport; beaches; entertainment (cinemas, performing arts venues etc); national parks; public parks; large events (concerts, festivals, sports games etc.); child care; and aged care. Each contact network is defined by a set of properties: the percentage (and age range) of the population who are a part of it; the average number of contacts per day associated with these . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint activities; whether the contacts are known or random; the type of network structure (random or cluster -for example public transport is random while schools/workplaces are clustered); the risk of transmission relative to a household contact (scaled to account for frequency of some activities); the effectiveness of contact tracing that might occur; and the effectiveness of quarantine at reducing transmission (e.g. quarantine may be effective for workplace transmission, not effective for household transmission, and partially effective for community transmission due to imperfect adherence). Details of the contact networks are provided in Appendix D. The model population was initialized through the generation of households. Individual households were explicitly modelled based on the households size distribution for Australia [22] , with each person in the model assigned to a house. To assign people in the model an age, a single adult was selected for each household as an index, whose age was randomly sampled from a subset of the Victorian adult population (all adults 22 years and older and a percentage of 18-21 year olds -20%, 40%, 60%, and 80% of people aged 18, 19, 20 and 21, respectively) to ensure that at least one adult was in each household. The age of additional household members was then assigned according to Australian agespecific household contact estimates (from Prem et al. [23] , Figure S2 ), by drawing the age of the remaining members from a probability distribution based on the row corresponding to the age of the index member. The resulting age distribution of the model population, compared to the Victorian population, is provided in Figure S1 . School classrooms were explicitly modelled. Classroom sizes were drawn randomly from a Poisson distribution with mean 21, the Victorian average [24] . People in the model aged 5-18 years were assigned to classrooms with people of the same age. Each classroom had one randomly selected adult (>21 years) assigned to it as a teacher. The school contact network was then created as a collection of disjoint, completely connected clusters (i.e. classrooms). Similarly, a work contact network was created as a collection of disjoint, completely connected clusters of people aged 18-65 years. The size of each cluster was drawn randomly from a Poisson distribution with mean equal to the estimated average number of daily work contacts (Table S4) . . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Other clustered contact networks, such as places of worship, community sports, professional sports, child care and aged care were generated analogously (with transmissibility scaled to account for event frequency; Appendix D). Random contact networks (e.g. public transport) were generated by allocating each person a number of contacts drawn from a Poisson distribution with mean as per Appendix D. Unlike the clustered contact networks, the contacts in random contact networks were resampled at each time step in the model (representing days). Policy scenarios modelled were informed by the COVID-19 public health response in Victoria [1] and the COVIDSAFE Australia framework [4] , and included scenarios related to: the effectiveness of contact tracing; compliance with physical distancing; restricting access to hospitality and entertainment venues and other public spaces; restricting access to places of worship; restricting the size of social gathering; restricting community and professional sport; closing schools and childcare settings; closing non-essential workplaces, retail outlets and health care; and restricted travel across jurisdictional borders and domestic travel. Each policy change is linked to one or more networks, and can potentially influence the whole population. For example, if non-essential work begins, this would increase the size of the work network, as well as increasing transmissibility in public transport. See Appendix E for full list of modelled scenarios. Epidemiological data for the daily number of tests conducted, new diagnoses and new severe cases, critical cases and deaths was obtained from the Victorian Department of Health [20, 25] . Newly diagnosed cases were classified as "imported" to Victoria if their mode of acquisition was listed as travel overseas. Disease specific parameters, including duration of incubation, infectious and symptomatic periods, and age-specific risks associated with disease severity and outcomes, were based on global published estimates (Table S1 and Table S2 ). . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Parameters for contact networks and the effect of policy changes were obtained from a combination of the literature and a modified Delphi process (Appendix D). The modified Delphi process involved creation of a panel of 12 experts (a mixture of modellers, epidemiologists, qualitative researchers, social network researchers, infectious disease physicians and public health physicians), who participated in a video conference where they were introduced to the model and the interpretation of parameters. Panel members were then asked to make independent estimates of unknown parameters, which were collated and de-identified by the study team, and the median and range of each parameter was extracted. A follow-up video conference was held where the panel discussed the results and uncertainties and were provided an opportunity to revise any estimates. The distribution of responses for each parameter, as well as the final parameters used, are provided in Appendix D. A baseline scenario was run between 1 March and 30 April, including with the policy changes that had occurred over that period, and the parameter for the overall probability of transmission per contact was calibrated such that the model projections fit the data on number of diagnoses and deaths. Multiple scenarios were run, in which different policy restrictions were lifted in isolation starting from 15 May (the date of analysis): opening pubs/bars; allowing large events; opening cafes and restaurants; allowing community sports; allowing small social gatherings; opening entertainment venues (e.g. cinemas, performing arts); removing work from home directives (resulting in greater public transport use as well as more work interactions); and opening schools. The parameter and network configuration changes associated with relaxing each restriction are described in Appendix D. For each scenario, a number of new infections were introduced for modelling purposes (a theoretical five infections on 15 May) to restart the epidemic and test the robustness of the new policy configuration to outbreaks. . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint We estimated the threshold population-level coverage that a contact tracing smartphone app (i.e. COVIDSafe) would need to mitigate the risks of relaxing different policies. The threshold target was calculated to mitigate the risks associated with the policies of opening of pubs/bars and removing work from home directions, as these were the policies found to have the greatest risk (see results). Multiple scenarios were run where these policies were changed but with population-level coverage of the contact tracing app ranging from 0-50%. Policy options are available and being utilized by governments mitigate the risks associated with opening of cafés, restaurants, pubs and bars; for example, transmissibility in these settings could be reduced by implementing the "4 square metre rule" and similar limits on customer numbers, or restricting venues to outside service only. We estimate how effective these additional interventions would need to be to mitigate the risks associated with opening these venues. Opening pubs/bars was used as an example as it was found to pose the greatest risk, and multiple scenarios were run where transmissibility within pubs and bars was reduced by 0-50%. An additional policy option being used is for venues (pubs/bars/cafes/restaurants) to keep mandatory identification records of patrons, which would enable contact tracing following a diagnosed case (similar to a contact tracing smartphone app, but with much higher coverage over a smaller time and place). We estimate the threshold compliance with this policy required to mitigate the risks associated with opening these venues. Multiple scenarios were run where pubs/bars were opened but with the capacity to contact trace 40-80% of contacts following a within-venue transmission event. A reasonable model fit was obtained ( Figure 1 ) that included the initial increase in cases observed followed by the subsequent decline in cases following the introduction of specific policy changes. We estimate that by April 30, approximately 2000 people had been infected with COVID-19, of which . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint approximately 1600 (80%) had been diagnosed. The undiagnosed proportion primarily includes asymptomatic cases. To estimate which policy change is associated with the greatest or quickest increase in infections we assessed the impact of lifting individual policies separately, on May 15 th . The greatest risk of a rebound in cases comes from policy changes that facilitate random, once-off mixing in the community, or situations where individuals have a large number of contacts, particularly those that are unknown. This includes opening pubs and bars (without additional restrictions), removing work from home directives (which increases public transport and work interactions) or allowing large events (concerts, sporting crowds, protest marches). The least risk comes from policy changes that facilitate smaller numbers of contacts, or repeated contacts with the same people (e.g. small social gatherings) ( Figure 2 ). Importantly, for some policy changes the time before new infections begin to rapidly increase could be greater than two months ( Figure 2 , for example cafes/restaurants or entertainment venues opening). . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint To assess the potential impact of the contact tracing smartphone app on mitigating an increase in new infections, we modelled two highest risk policy changes -the re-opening of pubs and bars and removing working from home directives -with varying degrees of app coverage. Greater than 30% coverage was required before the app showed significant impact on mitigating population-level transmission risks (Figure 3 for pubs and bars being opened, and Figure S5 for working from home directives being removed). . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Opening pubs and bars (without additional restrictions) was found to be the policy that led to the greatest increase in new infections. However, the model suggests that if physical distancing policies within these settings could reduce transmissibility by more than 40% they could considerably mitigate the risks of them opening ( Figure 4 ). Alternatively, recording the identification of patrons attending pubs and bars to enable effective contact tracing would be an effective policy at a population-level if compliance was greater than 60% ( Figure S6 ). . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Using an agent-based model we have simulated the relaxation of a variety of policy restrictions in a low transmission setting in Australia. We found that policy changes leading to large, unstructured contact networks (e.g. pubs and bars opening, increased public transport use through removal of work from home directives, or large events) posed the greatest risk, while policy changes leading to smaller, structured contact networks with known individuals (e.g. small social gatherings) posed the least risk. Importantly, the model suggests that it could take more than two months to detect increases in new infections from a change in policy, and therefore care should be taken in introducing multiple policy changes within short time periods. These outcomes, and this modelling tool, have implications for other settings with low community transmission where governments are lifting restrictions following relatively successful early responses. Despite social and economic pressures to fast-track a return to normal conditions, our results suggest that restraint is needed, even in low transmission settings, because a resurgence in the epidemic . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint following some policy changes could take more than two months to establish and be detected. In the model, contact tracing is effective for known contacts (Table S7) ; however, transmission to unknown community contacts can still occur, with frequency dependent on the mixing allowed by the policy environment. Transmission to an unknown contact would only be detected through their eventual symptomatic testing, allowing opportunity for them to further transmit during their infectious/presymptomatic stage. As with the index case, some of these additional infections will be of known contacts and effectively traced and isolated, however it is the chains of transmission through unknown contacts that may represent a minority of new cases initially, but if allowed to continue provide an increasing cumulative risk for epidemic expansion. It is therefore essential that accessible and efficient testing sites are available to complement contract tracing programs and detect community transmission from unknown sources. Once community transmission is detected, interventions to suppress a second epidemic wave would be required, for example geographically targeted testing or re-introduction of restrictions, to reverse the increasing trends projected in our scenarios. These interventions have not been modelled in this study and further work is required to assess how outbreaks can be managed without needing a complete policy reversal. The greatest risks of a resurgence in cases were associated with policy changes that allowed individuals to have large contact networks (e.g. crowded public transport, crowded pubs/bars, sports events) that introduce once-off mixing between unknown individuals in the community. In particular, these findings support the Victorian government's decision to extend work from home directions for people who are able until at least July 2020, to minimise use of public transport [1] . Further modelling work could assess whether staggered work starting times (to limit the number of riders), increased ventilation and cleaning, or face mask recommendations could mitigate the risks associated with increased public transport use. Policies creating large networks with random mixing were also found to be high risk even when they only affected a small subset of the population. For example, even though a minority of the population attend sporting events regularly, the random mixing of large numbers of contacts at a sporting event creates connectivity between smaller, clustered networks of known contacts, such as workplaces and households. The lowest risks were associated with policy changes that led to smaller numbers of contacts for individuals, introduced organized contact network structure (e.g. known contacts), or introduced easily traceable contacts (e.g. family or small social gatherings). Under these network configurations, population-wide connectivity remains restricted, limiting the potential for wide-scale population spread. In addition, known contacts have a greater probability of being traced in a timely way when . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint transmission does occur. These findings suggest that a smaller number of high intensity contacts is favourable compared to a large number of low intensity random contacts. We found that a contact tracing smartphone app (i.e. COVIDSafe) would need greater than 30% effective population coverage to mitigate the risks associated with most policy relaxations. The effectiveness of the app relies on both the infected and susceptible person having the app and using it correctly (i.e. having their Bluetooth switched on and the app enabled), which means that if 30% of the population in a given time and place have downloaded the app, this would produce at most an additional 9% (30%*30%) of contacts able to be reliably traced. Importantly, the app also relies on a well-resourced, timely contact tracing and testing system to be in place to follow-up people identified as potentially exposed. As of end May, the COVIDSafe app had been downloaded by approximately 6 million Australians. This represents approximately 24% of the population, meaning that the app could trace at most an additional ~6% (24%*24%) of contacts. Therefore, while the app could be effective at high coverage, it is likely to have minimal impact for low-moderate coverage. Based on the current epidemiological situation, we estimated that to mitigate the risks of opening pubs and bars (the policy change found to pose the greatest risk), physical distancing strategies that can reduce COVID-19 transmissibility by at least 40% in these settings are required. The model cannot identify what interventions may be able to achieve this, but this provides a useful target for designing interventions that consist of a mix of hygiene measures, physical distancing and limits to patron numbers. An additional policy option being implemented is for pubs and bars to keep mandatory identification records of patrons, which would enable rapid contact tracing following a diagnosed case. The model identified that this could be an effective policy if it enabled greater that 60% of contacts to be traced if an infected person was found to have attended a venue ( Figure S5 ). Note that for mandatory identification to be as effective as the smartphone app, it needs to be more stringent, since the app has additional benefits by tracing multiple generations of transmissions rather than only those in the source setting. Schools have been a topic of considerable policy discussion in Australia and internationally. Data are extremely limited on the susceptibility and transmissibility of COVID-19 among children, as well as health outcomes associated with infection. In our projections, opening schools was not predicted to lead to major population-level epidemic rebound, however it is critical to understand that this conclusion is based on some input parameters for which there is limited evidence. First, people aged under 20 years were assumed to be less susceptible to infection than people over 20 years (people aged 0-9 or 10-19 have relative susceptibility of 0.34 or 0.67 respectively, Table S2 ). Second, the probability of people under 20 years being symptomatic was lower than for people over 20 years, with . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint asymptomatic cases having reduced transmissibility in the model. Third, the network structure for schools was clustered into classrooms with no links connecting different classrooms in a school, allowing more effective contact tracing following any outbreak. While this is the best available evidence at the time of writing, ongoing work is required to verify this preliminary data, and alternate inputs associated with COVID-19 dynamics and children would lead to different outcomes. The main limitations to this work are around model features, disease epidemiology parameters and contact network parameters. This model currently only attributes basic properties to individuals, specifically age, household structure and participation in different contact networks. Therefore, the model does not account for any other demographic and health characteristics such as socioeconomic status, comorbidities (e.g. non-communicable diseases) and risk factors (e.g. smoking) and so cannot account for differences in transmission risks, testing, quarantine adherence or disease outcomes for different population subgroups. Further work is required with the specific aims of assessing the impact of policy changes on different subsets of the community. The model also does not include a geospatial component and so cannot capture geographic clustering of infections or concentration of interventions, including differential tracing app uptake in urban versus rural settings or among people attending particular events or settings, or concentrated testing in response to a localised outbreak. This means that our projections may be overestimating outbreak sizes as geographic clustering may slow epidemic spread. Data reported on disease parameters such as duration of asymptomatic and infectious periods, as well as age-specific estimates of susceptibility, transmissibility and disease severity and are likely to be influenced by differences in surveillance systems in the countries they are being reported from. We have taken the best available data at the time, but this is likely to change as new information becomes available, and the model should be updated accordingly. Contact networks are the most important factor driving COVID-19 transmission yet limited studies are available that provide the parameters needed to model them. The modified Delphi process used has potential biases in the non-randomly selected panel, and the large variation in parameter estimates suggests a high degree of uncertainty in contact network parameters. Despite this uncertainty, we argue that it is still important to consider these contact networks and the impact of policy changes on them. For example, studies are not available to quantify the relative transmissibility among public . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint transport contacts compared to household contacts. However, omitting this parametrisation would implicitly either ignore public transport contacts, or assume that they are equal to household contacts. In this study we have instead assumed that they fall somewhere in between, but we do not know where and hence have used a panel to estimate. Similarly, if people are instructed to work from home, then the transmission risk on public transport would be expected to decrease. While the actual reduction is unclear, if this feature were not included then this would implicitly assume that there was no change. It is critical that these parameters are continually updated as new evidence becomes available, and that the model is used to compare multiple policy options rather than to directly estimate the effects of individual policies. In particular, conclusions should be drawn that relate to the network properties, rather than the policies; for example, statements about random contacts being worse than clustered contacts [26] , rather than statements about specific policies. In settings with low community transmission, care should be taken to avoid introducing multiple policy changes within short time periods, as it could take greater that two months to detect the consequences of any changes. When selecting which policies to relax, the greatest risks are associated with large gatherings of people who do not know each other. In particular, working from home directions should be maintained to minimise use of public transport. Opening pubs/bars was identified as the greatest risk in Victoria; however these risks could be mitigated if either greater than 30% population-level coverage of a contact tracing smartphone app was achieved, transmissibility within venues was reduced by more than 40% through physical distancing policies, or patron identification records were kept that enabled greater that 60% of contacts to be traced if an infected person was found to have attended a venue. . 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 June 12, 2020. . Left: within schools, students aged 5-18 were in classrooms with others of the same age, and one teacher per classroom (the asymmetry is due to all students having an adult teacher contact, but not all adults being teachers and having school contacts). The y-axis represents the age of the individual and the x-axis represents the age of their contacts. The colour represents average number of daily contacts. . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Figure S3 : Examples of age-mixing within workplaces and public spaces. Left: at workplaces, adults aged 18-65 could mix with adults of any other age. The extra intensity on the 25-35-year-old diagonal is due to the disproportionate population age distribution in Victoria. Right: in public spaces, all ages could mix together. The y-axis represents the age of the individual and the x-axis represents the age of their contacts. The colour represents average number of daily contacts. Figure S4 : Example contact network structures between in the model. Left: the workplace network was modelled as clusters with size drawn from a Poisson distribution, and was fixed throughout a simulation. Right: some the community transmission networks, such as public transport, were modelled such that each individual had a number of contacts that were randomly assigned, and were re-assigned each day. . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint . 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 June 12, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Assumption . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Prob[severe] 0.00004 0.00040 0.01100 0.03400 0.04300 0.08200 0.11800 0.16600 0.18400 Verity et al. [10] ;CDC [12] . [12, 13] CDC . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint The parameters in this appendix were obtained from the literature where available, or through a modified Delphi process where studies were not available (a Delphi process modified to be possible during the COVID-19 pandemic). A group of 12 experts (a mixture of modellers, epidemiologists, qualitative researchers and social network researchers) were invited to participate. A video conference was held where they were introduced to the model and the interpretation of parameters, and participants were asked to make independent estimates of unknown parameters following the conference. Estimates were then collated by the study team, and the median and range of each parameter was extracted. A follow-up video conference was held where the panel discussed the results, uncertainties and were offered an opportunity to update any parameters. Each contact network only applies to a subset of the model population; because not everyone participates in each activity, or attends each location, only a subset are able to be infected at these places or during these activities. The subset of the population that each network applies to is defined as a percentage of a given age range. . 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 June 12, 2020. . Public transport 15+ 11.5% 2016 census. 11.5% of people travelled to work by public transport [16] . National parks all 5.6% 1.38 million national park visitors in Australia in 2017 [17] , with an Australian population size of 24.6 million. "national park goers" are over counted due to multiple visits, however conversely this estimate does not include state parks. This would give ~5.6% (1.38 million / 24.6 million). Public parks all 60% Median estimate from panel: . 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 June 12, 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Each network can have a different structure, with people either being connected to their contacts randomly ("random") or people being grouped into disconnected clusters ("clustered", e.g. schools, where the network consists of disjoint classrooms, with students in each classroom connected to one another). The differences between a random and clustered network are illustrated in Figure S4 . Each person in the model has a specified number of contacts in each network layer. The epidemiological definition of a contact between two people is used, where a contact is defined as having a 15-minute face-to-face conversation, or spending one hour or more in a room together. For those who have a non-zero number of contacts in a particular network (i.e. they are inside the applicable age range and randomly-selected population fraction defined in Table S3 ), if the contact network is "random" type, then their number of contacts is drawn from a Poisson distribution with mean as per Table S4 . If the contact network is "clustered", then the size of each cluster is drawn from a Poisson distribution with mean as per Table S4 . Networks can also be time-varying or not. For example, contact networks for public spaces (e.g. public transport) are regenerated each day, to simulate once-off mixing, compared to work networks in which specific individuals remain connected to one another. . 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 June 12, 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 June 12, 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 June 12, 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint *Not size of large event but number of actual contacts during event . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Transmission of COVID-19 is likely to be highly variable depending on network. As well as an overall daily risk of transmission per contact (the calibration parameter for the model), the risk of transmission per contact per day is different for each network. Table S5 shows these estimated differences relative to the transmission risk per contact per day within households. Table S5 : Relative risk of transmission through a contact, compared to a household contact. No studies were available for these parameters, meaning that they were all are based on the median of the expert panel's estimates shown in Figure S9 below. . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Figure S9 : Expert panel estimates for the risk of transmission in each contact network, relative to household contacts. . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint People may not typically interact with the activities and public spaces corresponding to each network on a daily frequency; for example, community sport might be played once per week. The model currently does not include simulation of each activity with different frequencies, and so the impact of this was approximated by reducing the relative transmission risk in each contact network. The relative transmissibility (Table S5) was divided by the activity frequencies/365 to develop a proxy for per-day transmission risk. . 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 June 12, 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 June 12, 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 June 12, 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Quarantine and contact tracing People who are asked to self-isolate are likely to change their behaviour in ways that reduce their likelihood of transmission through different contact networks. For people in quarantine, their relative transmissibility in each contact network (Table S5) is reduced by the factors shown in Table S7 . For example, quarantine is modelled to have no impact on household transmission, to completely stop workplace and school transmission, and reduce (but not stop) other forms of community transmission due to imperfect adherence. When a person is diagnosed, there is a probability of tracing the people they are connected to in different contact networks, and an associated time to trace them. For example, we assume that household members would be notified on the day of diagnosis, while workplace contacts would have a 70% chance of being traced within 2 days. The effectiveness of quarantine, contact tracing probabilities and tracing time were estimated from the expert panel. . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint There were no studies available to estimate the impact of policy changes on each network. However, for many polices, the impact is based on turning on / off transmission within a particular network, and so the impact is derived from the network properties in Tables S3-S6. For some policies, there are logical impacts that extend beyond their specific network; for example, if non-essential work is cancelled, then the transmission risk on public transport would be expected to decrease. For these auxiliary effects, the actual impact size is unknown, and so has been estimated by the panel of experts. 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 June 12, 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint 23. Non-essential retail outlets closed 24. Non-essential work closed 25. Non-COVID-19 health services closed 26. Travel restrictions across state borders Any set of interventions can be run in combination, or staged according to policy change dates. . 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 June 12, 2020. . https://doi.org/10.1101/2020.06.11.20127027 doi: medRxiv preprint Victorian COVID-19 restrictions Australian Government Department of Health: Coronavirus (COVID-19) current situation and case numbers Early analysis of the Australian COVID-19 epidemic Australian Government: 3-Step Framework for a COVIDSafe Australia Agent-based models: Sage The effectiveness of full and partial travel bans against COVID-19 spread in Australia for travellers from China Delaying the COVID-19 epidemic in Australia: Evaluating the effectiveness of international travel bans Modelling the impact of COVID-19 in Australia to inform transmission reducing measures and health system preparedness Modelling the impact of COVID-19 upon intensive care services in New South Wales Interventions to mitigate early spread of SARS-CoV-2 in Singapore: a modelling study. The Lancet Infectious Diseases Modelling transmission and control of the COVID-19 pandemic in Australia. arXiv preprint arXiv Modeling layered non-pharmaceutical interventions against SARS-CoV-2 in the United States with Corvid Impact of non-pharmaceutical interventions (NPIs) to reduce COVID19 mortality and healthcare demand The effectiveness of social distancing in mitigating COVID-19 spread: a modelling analysis Modeling the impact of social distancing, testing, contact tracing and household quarantine on second-wave scen-arios of the COVID-19 epidemic Isolation and contact tracing can tip the scale to containment of COVID-19 in populations with social distancing Group CC-W: Effectiveness of isolation, testing, contact tracing and physical distancing on reducing transmission of SARS-CoV-2 in different settings Covasim: an agent-based model of COVID-19 dynamics and interventions M2YTY1MjI2NzdjIiwidCI6ImMwZTA2MDFmLTBmYWMtNDQ5Yy05Yzg4LWExMDRjNGViOW YyOCJ9. 21. Institute for Disease Modelling: Info hub Projecting social contact matrices in 152 countries using contact surveys and demographic data At any time point in a simulation, parameters can be varied to: • Change the number of imported infections (from other Australian jurisdictions or internationally) the overall probability of transmission per contact (e.g. due to general hand hygiene) • Scale the relative transmission risk for specific contact layers (e.g. a policy closing cafes and restaurants would set the transmission risk for the cafe Change the effectiveness of contact tracing for a particular contact network (e.g. the COVIDSafe app makes contact tracing possible for community transmission only if both the infected and susceptible person have the app) Policy changes are linked to one or more networks, and can potentially influence the whole population. For example, if non-essential work begins, this would increase the size of the work network, as well as increasing transmissibility in public transport The following are examples of policies that can be simulated: 1. Contact tracing (including the use of COVIDSafe app for different coverages) 2. Communication and enforcement of physical distancing (e.g. signs, advertisements, policing) 3. Cafes and restaurants take-away only 4. Cafes and restaurants implementing 4 square metre rule physical distancing rule 5. Pubs and bars closed 6. Pubs and bars implementing 4 square metre rule physical distancing rule 7 Outdoor settings restricted to <2 people 10. Outdoor settings restricted to <10 people 11. Outdoor settings restricted to <200 people 12. Indoor social catch ups with <10 people banned 13. Community sports 14. Professional sports (for players) 15. Child care closed 16. Schools closed 17 Projecting social contact matrices in 152 countries using contact surveys and demographic data The incubation period of coronavirus disease 2019 (COVID-19) from publicly reported confirmed cases: estimation and application The serial interval of COVID-19 from publicly reported confirmed cases Serial interval of novel coronavirus (COVID-19) infections Investigation of three clusters of COVID-19 in Singapore: implications for surveillance and response measures Temporal dynamics in viral shedding and transmissibility of COVID-19 Spread and dynamics of the COVID-19 epidemic in Italy: Effects of emergency containment measures Virological assessment of hospitalized patients with COVID-2019 Estimates of the severity of COVID-19 disease Changes in contact patterns shape the dynamics of the COVID-19 outbreak in China Prevention: COVID-19 Response Team. Severe outcomes among patients with coronavirus disease 2019 (COVID-19)-United States Impact of non-pharmaceutical interventions (NPIs) to reduce COVID19 mortality and healthcare demand Local Churches in Australia: Research Findings from NCLS Research The Sport Participation Research Project: Sport Participation RatesAggregation of 12 sports Census estimates on method of travel to work 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 Australian Government Director of National Parks: Anual Report 2016-17 Australian Institute of Health and Welfare: Accessed 15 State of Victoria Department of Education and Training: School classroom sizes Australian Government: 3-Step Framework for a COVIDSafe Australia