key: cord-255575-640v00jg authors: Binny, R. N.; Baker, M. G.; Hendy, S. C.; James, A.; Lustig, A.; Plank, M. J.; Ridings, K. M.; Steyn, N. title: Early intervention is the key to success in COVID-19 control date: 2020-10-23 journal: nan DOI: 10.1101/2020.10.20.20216457 sha: doc_id: 255575 cord_uid: 640v00jg New Zealand responded to the COVID-19 pandemic with a combination of border restrictions and an Alert Level system that included strict stay-at-home orders. These interventions were successful in containing the outbreak and ultimately eliminating community transmission of COVID-19. The timing of interventions is crucial to their success. Delaying interventions for too long may both reduce their effectiveness and mean that they need to be maintained for a longer period of time. Here, we use a stochastic branching process model of COVID-19 transmission and control to simulate the epidemic trajectory in New Zealand and the effect of its interventions during its COVID-19 outbreak in March-April 2020. We use the model to calculate key outcomes, including the peak load on the contact tracing system, the total number of reported COVID-19 cases and deaths, and the probability of elimination within a specified time frame. We investigate the sensitivity of these outcomes to variations in the timing of the interventions. We find that a delay to the introduction of Alert Level 4 controls results in considerably worse outcomes. Changes in the timing of border measures have a smaller effect. We conclude that the rapid response in introducing stay-at-home orders was crucial in reducing the number of cases and deaths and increasing the probability of elimination. An outbreak of COVID-19, a novel zoonotic disease caused by the SARS-CoV-2 virus, was first detected in Wuhan, China in November 2019. The virus spread rapidly to other countries resulting in a pandemic being declared by the World Health Organisation in March 2020. Governmental policy responses to COVID-19 outbreaks have varied widely among countries, in terms of the nature and stringency of policy interventions, how quickly these interventions were implemented (Table 1 ) (Desvars-Larrive et al., 2020) and their effectiveness at reducing spread of the virus (Flaxman et al., 2020; Hsiang et al., 2020; Binny et al., 2020a) . While it is tempting to judge the success of interventions by comparison across jurisdictions, this assessment may be confounded by local context that may influence success, as well as by the fact that policy choices can be driven by the severity of initial outbreaks. Models of disease spread played an important role in the design and timing of interventions, but they can also be used post hoc, to evaluate the effectiveness of those interventions. For example, Flaxman et al. (2020) and Brauner et al. (2020) fitted models of disease dynamics to case count and death data in different countries to estimate the effect of specific non-pharmaceutical interventions on the transmission rate of COVID-19. In response to the escalating COVID-19 pandemic and the outbreak that was establishing in New Zealand in March 2020, a number of policy interventions were implemented to mitigate risk at the border and risk of community transmission. From 15 March 2020 (11.59pm), border restrictions were put in place requiring all international arrivals to 'self-isolate' (home quarantine) for 14 days. On 19 March 2020, the border was closed to everyone except returning citizens and residents. A system of four alert levels was introduced on 21 March with the Alert Level initially set at Level 2. On 23 March, it was announced that the Alert Level was increasing to Level 3 , and that the country would move to Alert Level 4 as of 11.59pm on 25 March, signalling that NZ was taking a decisive COVID-19 response that would become known as an elimination strategy (Baker et al., 2020a) . At the time Alert Level 4 came into effect, there had been 315 reported (confirmed and probable) cases. Alert Level 4 stayed in place until 27 April when restrictions were eased to Alert Level 3. On 13 May, after 16 days at Alert Level 3, daily new cases had dropped to 3 and there was a phased easing into Alert Level 2 ( Table 2) . The seven weeks spent under stringent Alert Level 3 or 4 restrictions, which included stay-at-home orders (see Appendix Table S2 for full list of measures) alongside systems for widespread testing, contact tracing and case isolation, were effective at reducing transmission (Reff = 1.8 prior to Alert Level 4; Reff = 0.35 during Alert Level 4; Binny et al., 2020b) . Daily numbers of new cases declined to between zero and one by mid May and the last case of COVID-19 associated with the March outbreak was reported on 22 May. On 8 June, it was estimated that New Zealand had very probably eliminated community transmission of COVID-19 after 17 consecutive days with no new reported cases . Between 22 May and 11 August, the only new cases detected were associated with international arrivals and during this period these arrivals were required to spend 14 days in government-managed isolation or quarantine facilities (Baker et al, 2020b) . On 9 August 2020, New Zealand reached a milestone of 100 days with no community transmission. For each hypothetical scenario, we simulate a model of COVID-19 spread and compare key outcomes, including the peak load on the contact tracing system, the cumulative numbers of cases and deaths, and the probability of elimination predicted by the model. In particular, we assess how important New Zealand's decision to move 'hard and early' was for the successful elimination of community transmission during the March-April outbreak. To this end, we compare scenarios with different timings until the start of Alert Level 4 to see how these choices could have affected the size of the outbreak. While Alert Level 4 was successful in achieving elimination, the benefits of elimination had to be weighed against the negative impacts of stringent stay-at-home measures, for example job losses, increased rates of domestic violence, disruption to education, and impacts on mental health. If careful border management could have avoided the need for a lockdown or reduced its intensity, this approach may have been preferable. For instance, Taiwan's early border closure, travel restrictions and 14-day quarantine for those entering the country have meant that, to date, Taiwan has avoided a mass lockdown (Summers et al., 2020) . We explore whether introducing border restrictions earlier in New Zealand might have been sufficient to eliminate or reduce transmission from international arrivals to the extent where stringent Alert Level 4 restrictions could have been avoided or less restrictive measures been sufficient. Compared to other countries, New Zealand was very quick to close its border to all except returning citizens and residents (Table 1) . We explore a scenario where border closure is delayed by 5 days to assess how much larger the outbreak might have been had New Zealand been slower to act. Finally, we consider a scenario with no AL4/3 restrictions to compare the size of outbreak that New Zealand could have experienced if border restrictions and closure had been the only control measures. In this study, we focus on the timing of interventions; we do not explicitly consider the duration of interventions, although this will be investigated in future work. Indeed, the likelihood of elimination was one of the factors taken into account in New Zealand government decision-making concerning the duration of Alert Levels (DPMC, 2020). We simulated a stochastic model of COVID-19 spread in New Zealand under alternative scenarios in which implementation of border restrictions, border closure and Alert Level 4 were either delayed or started earlier. In each scenario, we kept the duration of each Alert Level the same as actually occurred, i.e. 33 days at Alert Level 4 followed by 16 days at Alert Level 3. We explored the following scenarios (see Table 2 A full description of the model is provided in James et al. (2020a) . We obtained case data from ESR, containing arrival dates, symptom onset dates, isolation dates and reporting dates for all international cases arriving in New Zealand between February and June. Border restrictions, border closure and start of AL4 were all implemented at 11.59pm so we start simulating their effects on the day after their implementation date. For scenarios 0, 1, 2, 3 and 6, the model was seeded with the same number of international cases as were actually reported. In scenarios where border restrictions were implemented on the actual start date (15 March; Scenarios 0, 1, 2, 4 and 6), the self-isolation dates of international cases were set to the same isolation dates as were actually reported. In all scenarios, prior to 9 April the modelled effect of self-isolation is to reduce an individual's infectiousness to 65% of their infectiousness when not isolated. This reflects some risk of onward transmission for cases self-isolating at home. After 9 April, the model assumes that all international cases are placed in government-managed isolation and quarantine (MIQ) facilities and do not contribute to local transmission. We also simulated a Poisson-distributed random number of international subclinical cases in proportion to the number of international clinical cases (assuming 1/3 of all cases are subclinical), with arrival and symptom onset dates that were randomly sampled with replacement from the international case data. We assume that these international subclinical cases are not detected and therefore do not self-isolate, but those arriving after 9 April are placed in MIQ. To simulate border restrictions starting 5 days early (Scenarios 3 and 5a), international cases arriving between the earlier start date and the actual start date (11 -15 March, inclusive) were assumed to be . CC-BY-NC-ND 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 October 23, 2020. ; self-isolated on their date of arrival. To simulate a 5-day delay to border closure (Scenarios 4 and 5b), we delayed the arrival dates (and associated symptom onset, reporting and isolation dates) of international seed cases arriving after 19 March by 5 days. We then allowed for new international cases arriving over these 5 days (e.g. additional non-residents that may have chosen to travel had the border remained open for longer) by seeding an additional Poisson-distributed random number of international cases from 20 March to 24 March, with an average daily number of seeded cases equal to the actual average daily number of international cases arriving during the week prior to 19 March (33 international cases per day). These additional seeded cases were assumed to self-isolate on arrival and their delays from arrival-to-symptom onset and arrival-to-reporting were randomly sampled with replacement from the corresponding delays in the actual international case data. We did not attempt to simulate scenarios with delayed border restrictions or earlier border closure because these would have required additional modelling assumptions about isolation dates of international arrivals and about the reduction in the volume of international arrivals resulting from border closure. Model predictions would have been highly sensitive to these assumptions and, without data available to validate them, this would introduce additional model uncertainty. For each scenario, we assessed the following key measures describing the dynamics of a COVID-19 outbreak: 1. The maximum contact tracing load, by calculating the maximum number of daily new reported cases and the date on which this occurred. 2. The number of daily new reported cases at the end of Alert Level 4. 3. Cumulative number of reported cases and the cumulative number of deaths at the end of the seven week period of Alert Level 3-4 restrictions. 4. Probability of elimination, P(elim), 5 weeks after the end of AL3. We performed 5000 realisations of the model and report the average value of each key measure as well as the interval range within which 90% of simulation results were contained (in square brackets throughout). Here, we define elimination as there being no active cases (we assume a case remains 'active' for 30 days after date of exposure) that could contribute to future community transmission. This definition excludes cases in MIQ, that is it excludes international arrivals after 9 April 2020. In the model, P(elim) was calculated as the proportion of all model realisations that resulted in elimination. Simulations were run using estimates of reproduction number Reff that provided the best fit to actual data : for the period prior to lockdown Reff = 1.8; during Alert Level 4 Reff = 0.35. Under the Alert Levels 3, 2 and 1, which followed the lockdown, the daily numbers of new cases were too low to obtain reliable estimates of the effective reproduction number Reff. Instead we simulated the model for assumed values of Reff = 0.95, 1.7 and 2.4, respectively. Reff = 2.4 is in line with estimates reported in Plank et al (2020b) for the pre-lockdown period of New Zealand's August-September outbreak, when AL1 restrictions were in place. AL2's Reff would likely be lower than AL1 but greater than one due to relatively high activity levels and contact rates as stay-at-home orders are lifted and public venues, businesses and schools re-open; Reff = 1.7 is in the range of estimated values for the prelockdown period of the March-April outbreak given in Plank et al (2020b) . The Reff for AL3 was chosen to be less than 1, however we tested the sensitivity of our results to using a value greater than one. For the scenario with no stringent Alert Level restrictions (Scenario 6), we simulated the model using Reff = 1.8 for the entire period (i.e. the same value as was used in all Scenarios for the period prior to AL4). . CC-BY-NC-ND 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 October 23, 2020. ; We investigated the sensitivity of our results to varying the length of delay for the start of AL4, to introducing border restrictions 10 days early (cf. 5 days early in Scenario 3), and to other choices of Reff under AL3. To check that the model could accurately replicate the outbreak, we first simulated our model with border restrictions, border closure and AL4 implemented on the dates they actually occurred. The predicted dynamics of daily new reported cases were a very good visual match to observed daily case data ( Fig. 1) and predicted key measures showed good agreement with the values that were actually observed ( Table 3 , bold text). After moving into AL4, the model prediction and the actual number of daily new reported cases both levelled off at 70-80 for around one week before case numbers started to decline (Fig. 1 ). In actual case data, the maximum of 84 new cases per day was observed at the start of this flat-topped peak, while our model predicted a similar maximum (80 [67, 99] new cases per day) occurring 6 days later. By the end of AL3, the model predicted similar cumulative totals to the 1502 cases and 21 deaths actually reported. Five weeks after AL3 restrictions were relaxed, elimination of community transmission of COVID-19 was achieved in 66% of model simulations, giving P(elim)=0.66 (Table 3 ). In the following scenarios with alternative timings of interventions, we use Scenario 0 as a baseline for comparing key measures. Under a scenario where AL4 was implemented 5 days earlier (only one day after border closure), the model predicts slightly lower values for most key measures than were actually observed: daily new cases peaked at a lower level of 69 [61, 79] cases around 26 March and at the end of AL4 had dropped to a similar level of 4 new cases per day as was actually observed (Fig. 1) . By the end of the 7 weeks of AL4/3 it predicts approximately 500 fewer cases in total and 10 fewer deaths (Table 3 ; Scenario 1 cf. Scenario 0). However, this estimate should be taken with caution because of the small numbers of daily cases and fine-scale variations involved: for instance, whether an outbreak occurred in an aged care facility or not. Five weeks after AL3, the probability of elimination was 63%, slightly lower than in Scenario 0. This counter-intuitive result is due to the presence of an international case in the data that had an arrival date prior to the start of AL4 (25 March) but a much later symptom onset date near the end of AL4. In Scenario 0, when international cases are seeded in the model, this individual's peak infectiousness occurs during AL4. However, in Scenario 1, the earlier start to AL4 means that the individual is instead most infectious during AL3. With a lower Reff in AL3, this individual infects more people, on average, in this scenario than in Scenario 0. Similarly, any simulated subclinical cases with the same arrival and symptom onset dates (drawn from the international case data) will also be most infectiousness during AL3. These subclinicals do not appear in the numbers of reported cases but will reduce the probability of elimination. If this international case outlier is excluded from the data, the model predicts a very similar probability of elimination in both scenarios. Delaying the move into Alert Level 4 would have led to a higher peak in daily new cases, and greater cumulative totals of cases and deaths. For a delay of 20 days (Scenario 2c), the outbreak would have reached a considerably higher maximum of close to 500 daily new cases (cf. 80 cases in Scenario 0; Table 3 , Fig. 1 and Fig. 2 ). This number would certainly have overwhelmed the contact tracing system, which was already pushed close to capacity in places by the 70-80 daily new cases in late March 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 October 23, 2020. ; https://doi.org/10.1101/2020.10.20.20216457 doi: medRxiv preprint (Verrall, 2020) . After a week in AL4, case numbers would start to decline and by the end of the 4 weeks in AL4 daily new cases would still have been as high as 34 [22, 49] (close to the actual number of domestic daily reported cases when New Zealand went into AL4 on 25 March). By the end of the 7 week period of stringent restrictions (i.e. end of AL3) the incidence would have dropped to approximately 4 new cases per day (Fig. 1) , but the cumulative total could have climbed to 11,534 [8854, 15048] reported cases and 200 [147, 266] deaths, substantially more than Scenario 0 and the 1,502 cases and 21 deaths actually reported on 13 May. Additionally, the probability of elimination 5 weeks after the end of AL3 was only 7%, much lower than Scenario 0. We next investigated a scenario where border restrictions were put in place 5 days earlier, but border closure and AL4 were started on their actual dates. Border restrictions would therefore have been in place for 9 days (cf. actual 4 days) before the border was closed. Our model predicted this would have had very little impact for the initial trajectory ( Fig. 1) or eventual size of the outbreak, with values for all key measures very similar to those in Scenario 0 ( Table 3 ). This finding suggests that key measures are more sensitive to varying the timing of AL4 than to the timing of border restrictions. In reality, out of the 563 international cases who arrived prior to the start of MIQ and could have contributed to local transmission, only 78 (14%) arrived before border restrictions were implemented on 15 March and were not required to self-isolate. Furthermore, out of these 78 cases, 52 arrived between 10 and 15 March and 19 of these were reported to have voluntarily self-isolated immediately on arrival (the model simulates these 19 cases as being self-isolated on arrival in all scenarios). Therefore, under this scenario, only an additional 33 international cases have their infectiousness reduced by early self-isolation requirements. This reduction is not sufficient to prevent an outbreak, nor does it reduce transmission to an extent where AL4/3 restrictions would not have been necessary to control the outbreak. Under a scenario where closure of the border (to all except returning residents and citizens) was delayed by 5 days (24 March; 9 days after border restrictions and 1 day before AL4), our model predicted slightly worse outcomes, on average, for key measures than were predicted in Scenario 0. However, due to the stochasticity of individual simulations, the range of key measures always had overlap with the Scenario 0 values and actual values, suggesting that a 5 day delay to border closure alone would not have made a significant difference. A delayed border closure did, however, have a greater impact for the probability of elimination 5 weeks after AL3 restrictions were relaxed, which was only 55%, compared to 66% chance in Scenario 0. This reduced probability of elimination is partly due to the additional international clinical cases (captured in the key measures of reported cases) and international subclinicals (not captured in reported cases) arriving prior to the delayed border closure. It is also likely affected by the international case outlier with the pre-MIQ arrival date and late onset date, discussed above. If border restrictions were implemented 5 days early and AL4 came into effect 5 days early (Scenario 5a), this would have led to outcomes very similar to those predicted in Scenario 1 (where only AL4 started early) ( Table 3 ). This again suggests that results are more sensitive to changes in timing for the start of AL4 than to an earlier start to border restrictions. In contrast, if border closure and the start of AL4 had both been delayed by 5 days (Scenario 5b), outcomes would have been worse than a delay in only one of these interventions ( 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 October 23, 2020. ; https://doi.org/10.1101/2020.10.20.20216457 doi: medRxiv preprint the end of the 7 week period in AL4/3, there would have been close to 1,050 more cases in total and nearly 20 more deaths than in Scenario 0 ( Table 3 ). The probability of elimination 5 weeks after AL3 would have also been reduced to 53%. Finally, we explored the impact of only having border restrictions and border closure in place, but without implementing AL4/3. Under this scenario, the international cases who arrived prior to 9 April and were either in self-isolation or were not isolated have a chance of seeding an outbreak which, without AL4/3 measures to reduce Reff below one, leads to community transmission and a large uncontrolled outbreak. New Zealand would have seen close to 1127 [841, 1492] new cases per day by 27 April, the date on which New Zealand moved from AL4 to AL3 in reality. By 13 May (the date on which New Zealand moved from AL3 to AL2), there could have been over 60,000 cumulative reported cases and over 1100 deaths. New cases would have continued to increase, reaching a peak of 47,592 [47, 240, 47, 962] daily new cases on 14 June (Table 3 ). By the end of the outbreak, around October 2020 on average, there could have been over 1.81 million reported cases in total and 31,905 [31, 606, 32, 204] deaths. No simulations resulted in elimination by 18 June (5 weeks after end of actual AL3), indicating a 0% chance of COVID-19 having been eliminated by this time, compared to the 66% chance on this date in Scenario 0. We assessed the effect that different lengths of delay (in days) until the start of AL4 (Fig. 2) had on key measures: maximum load on contact tracing system; cumulative total reported cases; total infected cases (including both clinical and subclinical); total deaths at end of AL3; and probability of elimination 5 weeks after the end of AL3. Measures of numbers of cases and deaths increased exponentially with increasing delay to AL4, emphasising the importance of acting quickly to reduce the risk of large outbreaks arising. Probability of elimination decreased linearly with increasing delays to AL4. Counterintuitively, earlier starts to AL4 slightly reduced the probability of elimination; again, this is caused by the international case outlier discussed previously. If the outlier is excluded from the international case data, the predicted probability of elimination is insensitive to AL4 starting 1 to 5 days early. Introducing border restrictions ten days earlier still results in an outbreak and gives very similar results to Scenario 3 (5 days early), with a maximum of 77 [65, 94] new daily cases, 1385 [1166, 1706] cumulative reported cases at the end of AL3, P(elim) = 0.68, and other measures the same as in Scenario 3. With border restrictions ten days earlier, the only difference compared to Scenario 3 is that an additional 16 cases who arrived between 5 and 10 March have their infectiousness reduced (a further 2 cases arriving in this 5-day period were voluntarily self-isolated on arrival in reality, so are simulated with self-isolation on arrival in all scenarios). This restriction has little impact on the overall contribution to local transmission by all 563 international cases who arrive prior to the start of MIQ. We also tested the sensitivity of all key measures to using different values of Reff under AL3 (Table S1 ; Reff = 1.1, 0.95 and 0.7). Different choices of AL3 Reff had very little effect on predicted cumulative totals of cases at the end of AL3 and no effect on total deaths at end of AL3. However, the predicted probability of elimination was sensitive to varying AL3 Reff ; for all scenarios, assuming a lower Reff = 0.7 (more effective AL3) gave a P(elim) that was approximately 0.14 higher than with Reff = 0.95, while a higher Reff = 1.1 (less effective AL3) reduced P(elim) by approximately 0.07. 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 October 23, 2020. ; Level 4 (i.e. end of Alert Level 3) , and the total number of deaths 7 weeks after the start of Alert Level 4. For each measure, except P(elim) , the mean value from 5000 simulations is reported alongside the interval range, in parentheses, in which 90% of simulations results are contained. 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 October 23, 2020. ; https://doi.org/10. 1101 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 October 23, 2020. ; https://doi.org/10.1101/2020.10.20.20216457 doi: medRxiv preprint New Zealand's decision to act quickly and to implement stringent restrictions to reduce SARS-CoV-2 transmission meant that, to date, New Zealand has experienced amongst the lowest mortality rates reported worldwide (Kontis et al., 2020) . On 8 June 2020, nearly 11 weeks after AL4 was initiated, New Zealand declared elimination of COVID-19. Over the course of the March-April outbreak, a total of 1504 cases and 22 deaths were reported before elimination was achieved. Our results suggest that the timing of Alert Level 4 is a much stronger driver of reductions in daily new cases than timings of border restrictions and closure. This finding makes sense because the effect of AL4 in the model is to greatly reduce Reff for all cases, domestic and international arrivals, to 0.35, while border restrictions reduce the delay until case isolation of international cases only (i.e. international cases have their infectiousness reduced earlier) and border closure reduces the daily numbers of international cases only. Out of the scenarios we considered, an earlier start to AL4 by 5 days resulted in the greatest reduction in numbers of cases and deaths, with approximately 500 fewer cases in total and 10 fewer deaths. However, in reality, the rapid escalation of the COVID-19 situation in mid-March may have made an earlier start to AL4 impractical and would have allowed less time to prepare for ongoing provision of essential services under AL4. Introducing border restrictions requiring 14-day self-isolation for international arrivals earlier than 15 March would have been unlikely to have much impact on the trajectory of New Zealand's March-April outbreak, unless such measures were started prior to the first case on 26 February and used methods that were particularly effective (notably full MIQ). The 563 international cases arriving between 15 March and 9 April were already required to self-isolate; had border restrictions been in place prior to the arrival of New Zealand's first case, this would have required self-isolation for, at most, an additional 56 international cases (22 cases who arrived prior to 15 March self-isolated voluntarily immediately on their arrival). In mid-March, there was a lower global prevalence of COVID-19 and between 2 and 12 cases arrived at the border each day in the week prior to 15 March. With a higher global prevalence and correspondingly higher numbers of international cases arriving per day, earlier implementation of border restrictions may have had a greater impact than our model predicted for this outbreak. Self-isolation is less stringent than MIQ and relies heavily on public compliance. Without additional safety nets, such as official monitoring and support for people who are self-isolating, there is a greater risk of the virus spreading into the community than in MIQ facilities. For example, risk of non-compliance may be higher for individuals who are concerned about loss of income (Bodas & Peleg, 2020) . Without Alert Level restrictions in place to require strong communitywide social distancing, any infected individuals who do not self-isolate effectively are more likely to spark an outbreak. Self-isolation restrictions for international arrivals can therefore reduce the frequency of cases leaking into the community but are unlikely to be sufficient to prevent an outbreak entirely, unless additional measures are also put in place. Delaying border closure by 5 days could have led to a slightly larger outbreak, but not as large as if AL4 had been delayed by 5 days. The full effect on local transmission potential of the additional international cases expected under a delayed border closure was partially dampened because international cases arriving after 9 April were still placed in MIQ and assumed not to contribute to community transmission. If the timing of this MIQ policy was also delayed, a larger outbreak may have occurred, but we did not model such a scenario here. If the start of AL4 had been delayed by 20 days, our results suggest New Zealand could have experienced over 11,500 reported cases and 200 deaths, reducing the chance of elimination to only 7%. As with other severe viral disease, the infection fatality risk for COVID-19 is greater for Māori and Pacific peoples (close to 50% higher for Māori than for non-Māori) Wilson et al., 2012; Verrall et al., 2010) . Therefore, in scenarios resulting in significantly higher numbers of COVID-19-related deaths (e.g. Scenario 2c), Māori and Pacific communities would likely have been disproportionately affected, however a population-structured model would be required to assess this consequence in detail. Delaying AL4 would have also increased the chance of a longer lockdown period being required to reduce daily new case numbers to low levels. With a 20 day delay to AL4, New Zealand could still have been experiencing close to 35 new reported cases per day at the end of AL4. While in reality, a 33-day period in AL4 was sufficient to reduce daily new cases to below 10, and the Government announced an easing to AL3, these higher case numbers predicted for a delayed start to AL4 may have motivated an extension to the lockdown to allow more time for cases to drop below a safe threshold. In terms of the key measures we considered, the counterfactual scenario with no AL4/3 restrictions (Scenario 6) had disastrous outcomes, including close to 2 million reported cases and tens of thousands of deaths. Our model uses a value of Reff =0.35 during AL4, which was estimated by Binny et al (2020b) by fitting the model to data, and is consistent with a later estimate of Reff from reconstructions of the epidemiological tree . This is a relatively low value of Reff compared to other countries who implemented interventions roughly equivalent to AL4 (Flaxman et al, 2020; Binny et al, 2020a) . A combination of a highly effective social distancing in AL4, fast contact tracing, effective case isolation, and the fact that the outbreak occurred at the end of the Southern hemisphere summer, likely contributed to this low Reff . For scenarios where the load on contact tracing exceeded system capacity (e.g. Scenario 2c with a maximum 500 daily new cases), this effect would have likely resulted in longer delays to isolation of cases and a higher Reff. We did not attempt to model this potential feedback effect and so our results for scenarios where contact tracing system capacity is exceeded may underestimate the size of the outbreak. We also did not attempt to directly model the burden of COVID-19 on the healthcare system (e.g. numbers of cases requiring hospitalisation or intensive care), or the effects of an overwhelmed healthcare system. Once numbers of daily new cases requiring hospitalisation or ICU admission exceed New Zealand's healthcare system capacity, this could result in increased fatality rates and considerably more deaths . These effects would have been most pronounced under the scenario with no AL4/3 restrictions. It is important to note that, while we report average values for outbreak dynamics, each individual realisation of the stochastic model can deviate (sometimes widely) from the average behaviour. When case numbers are small, as they were in New Zealand, the predicted dynamics are particularly sensitive to fine-scale variations. While Reff < 1 means that an outbreak will eventually die out, on average, it is still possible for a small number of cases to spark an outbreak in a particular stochastic realisation if interventions are relaxed too soon. Conversely, when case numbers are small, an outbreak can still die out by chance even when Reff > 1. It is therefore important to account for this stochasticity when weighing the effectiveness and risks of different intervention strategies, for example by considering the probability of elimination. On 18 June, five weeks after AL3 restrictions were relaxed, the probability that community transmission of COVID-19 had been eliminated in model simulations was estimated to be 66% in Scenario 0. This estimate is calculated by finding the proportion of stochastic realisations that resulted in elimination. In reality, as the outbreak died out and more days with zero new cases were observed, this provided additional information about which trajectory New Zealand was most likely experiencing. Each additional consecutive day with no new reported cases reduced the likelihood of . CC-BY-NC-ND 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 October 23, 2020. ; being on an upward trajectory. Making use of this information meant that the actual probability of elimination on 18 June was estimated to be 95% higher than in Scenario 0 . However, this estimate required up-to-date information about recent case numbers. The results reported in this paper compare average outcomes under different scenarios, which is appropriate for evaluating the effect of alternative actions and guiding future decision making. In general, for the other scenarios we explored, bringing in earlier interventions had very little impact on probability of elimination, while delaying border closure or AL4 reduced the chance of elimination. Our results are important for reflecting on the effectiveness of intervention timing in New Zealand's COVID-19 response, relative to alternative scenarios, to help guide future response strategies. Early intervention was critical to the successful control of New Zealand's March-April outbreak. For modelling future disease outbreaks, epidemiological parameters should be updated to reflect changes in national pandemic preparedness (e.g. improved policy and response plans) and behavioural changes influencing the dynamics of future outbreaks. For instance, the degree of compliance with alert level restrictions in future may differ dramatically from the March-April outbreak, resulting in different values of Reff. Further work is needed to explore the social dynamics affecting transmission and the effectiveness of interventions, for instance whether wearing masks in public spaces becomes more common, or whether more people will choose to work from home or avoid travel if a suspected new outbreak is reported. The key measures of outbreak dynamics assessed here should be considered alongside other measures of economic, social and health impacts (e.g. job losses, consumer spending, impacts for mental health, rates of domestic violence or disrupted education). Particular attention needs to be given to identifying vulnerable groups who may experience inequitable impacts so that future policies can be tailored to support these groups. At the end of AL3, health benefits (e.g. number of cases and deaths avoided) differed between scenarios. For cost-benefit analyses, age-dependent morbidity and mortality rates of COVID-19 (Kang & Jung, 2020) allow numbers of cases and deaths to be quantified in terms of disability-adjusted life years (DALYs) avoided (or quality-adjusted life years gained), which can (with obvious issues) be converted to monetary units to facilitate comparison with economic costs. Because the duration spent under AL4/AL3 was fixed at 7 weeks for Scenarios 0-5, the short-term economic costs of the different scenarios would have been similar, so we did not convert health benefits into DALYs here. After the end of AL3, benefits and costs would differ between scenarios depending on the value of Reff for AL1-2 and on whether or not elimination was achieved. Increased levels of activity and contact rates under AL1-2 mean that Reff is very likely to have been greater than one. In scenarios with lower probabilities of elimination, it is more likely that New Zealand would continue to experience new cases while under AL1-2 which, with Reff > 1, would likely lead to another outbreak and require a second lockdown (with its associated costs). Conversely, scenarios with higher probabilities of elimination mean there is a greater chance of the outbreak dying out entirely and less risk of a second lockdown being required. Future work could consider the costs and benefits of alternative scenarios where the duration of time spent in AL4 and AL3 is dictated by the need to achieve a certain threshold probability of elimination. . CC-BY-NC-ND 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 October 23, 2020. ; https://doi.org/10. 1101 New Zealand's elimination strategy for the COVID-19 pandemic and what is required to make it work Successful elimination of covid-19 transmission in New Zealand Effect of Alert Level 4 on effective reproduction number: review of international COVID-19 cases Effective reproduction number for COVID-19 in Aotearoa New Zealand. medRxiv preprint Probability of elimination for COVID-19 in Aotearoa New Zealand Self-Isolation compliance in the COVID-19 era influenced by compensation: findings from a recent survey in Israel The effectiveness of eight nonpharmaceutical interventions against COVID-19 in 41 countries CAB-20-SUB-0270 Review of COVID-19 Alert Level 2 CCCSL: Complexity Science Hub Covid-19 Version 2.0 Estimating the effects of non-pharmaceutical interventions on COVID-19 in Europe The Effect of Large-Scale Anti-Contagion Policies on the Coronavirus (COVID-19) Pandemic. medRxiv preprint A structured model for COVID-19 spread: modelling age and healthcare inequities Successful contact tracing systems for COVID-19 rely on effective quarantine and isolation Model-free estimation of COVID-19 transmission dynamics from a complete outbreak Age-Related Morbidity and Mortality among Patients with COVID-19 Magnitude, demographics and dynamics of the effect of the first wave of the COVID-19 pandemic on all-cause mortality in 21 industrialized countries A stochastic model for COVID-19 spread and the effects of Alert Level 4 in Aotearoa New Zealand. medRxiv preprint Effective reproduction number and likelihood of cases outside Auckland Estimated inequities in COVID-19 infection fatality rates by ethnicity for Aotearoa New Zealand. medRxiv preprint Potential lessons from the Taiwan and New Zealand health responses to the COVID-19 pandemic. The Lancet Regional Health -Western Pacific Rapid Audit of Contact Tracing for Covid-19 in New Zealand. Ministry of Health, New Zealand. 0.95 80 Evaluated at end of outbreak, approx weeks after end of actual AL3) The authors acknowledge the support of StatsNZ, ESR, and the Ministry of Health in supplying data in support of this work. We are grateful to Samik Datta, Nigel French, Markus Luczak-Roesch, Melissa McLeod, Anja Mizdrak, Fraser Morgan and Matt Parry for comments on an earlier version of this manuscript. This work was funded by the Ministry of Business, Innovation and Employment and Te Pūnaha Matatini, New Zealand's Centre of Research Excellence in complex systems. Table S1 : Key measures from alternative scenarios of early or delayed implementation of policy interventions: the maximum number of daily new cases, date on which the peak occurs, the number of daily new cases at the end of the Alert Level 4 period, the cumulative number of cases 7 weeks after the start of Alert Level 4, and the total number of deaths 7 weeks after the start of Alert Level 4. For each measure, except P(elim) , the mean value from 5000 simulations is reported alongside the interval range, in parentheses, in which 90% of simulations results are contained. • Stay at home except for essential personal movement • Safe recreational activity allowed in local area.• Travel is severely limited.• All gatherings cancelled and public venues closed.• Businesses closed, except essential services and lifeline utilities.• Educational facilities closed.• Rationing of supplies and requisitioning of facilities possible.• Reprioritisation of healthcare services. • Stay at home except for essential personal movement -including going to work, school or for local recreation.• Work from home if possible.• Low risk local recreation activities allowed.• Inter-regional travel highly limited (e.g. allowed for essential workers, with limited exemptions for others).• Public venues closed (e.g. libraries, museums, cinemas, food courts, gyms, pools, playgrounds, markets).• Gatherings of up to 10 people allowed but only for wedding services, funerals and tangihanga, with physical distancing and public health measures maintained. • Two-metre physical distancing outside home. One metre in controlled environments, e.g. schools and workplaces.• People must stay within their immediate household bubble, but can expand this to reconnet with close family/whānau, or bring in caregivers, or support isolated people. This extended bubble should remain exclusive.• Businesses can open premises, but cannot physically interact with customers.• Schools (years 1 to 10) and Early Childhood Education centres can safely open but with limited capacity. Children should learn from home if possible. • Healthcare services use virtual, non-contact consultations where possible.• People at high risk of severe illness (older people or those with pre-existing medical conditions) are encouraged to stay at home and take extra precautions when leaving home. They may choose to work.