{"paper_id":"6c4adab4-4844-4301-aecf-999fa7e9904e","body_text":"An improved method to estimate the effective reproduction number of the COVID-19 \npandemic:  lessons from its application in Greece\nTheodore Lytras1,2, Vana Sypsa3, Demosthenes Panagiotakos4, Sotirios Tsiodras1,5\n1. National Public Health Organization, Athens, Greece\n2. Department of Medicine, School of Medicine, European University Cyprus, Nicosia, Cyprus\n3. Department  of  Hygiene,  Epidemiology  and  Medical  Statistics,  Medical  School,  National  and\nKapodistrian University of Athens, Athens, Greece\n4. School of Health Science and Education, Harokopio University, Athens, Greece\n5. 4th Department of Internal Medicine, Attikon University Hospital,  Medical School, National and\nKapodistrian University of Athens, Athens, Greece\nCorresponding author: \nTheodore Lytras, MD, PhD\nDepartment of Medicine, School of Medicine, European University Cyprus\n6 Diogenous str, 2404 Engomi, Nicosia, Cyprus\nEmail: t.lytras@euc.ac.cy\nWord count: 2,719 (full-text), 288 (abstract)\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \nNOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.\n\nAbstract\nIntroduction:  Monitoring the time-varying effective  reproduction number  Rt is  crucial  for assessing the\nevolution of the COVID-19 pandemic. We present an improved method to estimate Rt and its application to\nroutine surveillance data from Greece. \nMethods: Our method extends that of Cori et al (2013), adding Bayesian imputation of missing symptom\nonset  dates,  imputation of  infection times  using an external  estimate  of  the  incubation period,  and an\nadjustment for reporting delay. To facilitate its use, we provide an R software package named “bayEStim”.\nWe applied the method to COVID-19 surveillance data from Greece, and examined the resulting Rt estimates\nin relation to control measures applied, in order to assess their effectiveness. We also associated  Rt, as a\nmeasure of transmissibility, to population mobility as recorded in Google data and to ambient temperature.\nWe used a serial interval between 4 and 7.5 days, and a median incubation period of 5.1 days.\nResults: In Greece Rt fell rapidly as the first control measures were introduced, dropping below 1 at least a\nweek before a full lockdown came into effect. In mid-July Rt started increasing again, as increased mobility\nassociated with tourism activity was observed. Each 10% of increase in relative mobility increased  Rt by\n8.1% (95% CrI 6.1–10.2%), whereas each unit celsius of temperature increase decreased Rt by 4.6% (95%\nCrI 5.4–13.7%).\nConclusions: Mobility patterns significantly affect Rt. Most of the reduction in COVID-19 transmissibility in\nGreece occurred already before the lockdown, likely as a result of decreased population mobility. Lower\nviral transmissibility in summer does not appear sufficient to counterbalance the increased mobilit y due to\ntourism. Monitoring Rt is an essential component of COVID-19 surveillance, and it is crucial for correctly\nassessing the effect of control measures.\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nIntroduction\nThe coronavirus disease 2019 (COVID-19) pandemic originated in December 2019 in the city of Wuhan,\nChina, spreading across the globe and causing millions of cases and hundreds of thousands of deaths within a\nfew months [1,2]. As the world braces for likely further pandemic waves in late 2020, disease surveillance is\ncrucial in order to monitor the situation and guide public health action. An essential component of COVID-\n19 surveillance is estimation of the time-varying effective reproduction number Rt, defined as the average\nnumber of secondary cases at time  t produced by an infected individual over his/her infectious period .  Rt\nreflects the real-world transmissibility of a pathogen, and is affected by the mobility patterns, social mixing,\ncontrol measures, population immunity and other factors that are prevalent at each point in time.  Rt>1\nindicates exponential growth, Rt=1 indicates sustained transmission and Rt<1 suggests exponential decay of\nan epidemic.\nVarious methods to estimate Rt have been described in the literature [3–5]; we present an improve ment that\naddresses the inherent limitations of surveillance data, such as incompleteness and reporting delays, allowing\nfor almost real-time estimation of epidemic trends. We apply  our method in the case of the COVID-19\npandemic in Greece, and examine how transmissibility of the SARS-CoV-2 virus has varied over time and in\nrelation to control measures, population mobility and ambient temperature.\nMethods\nEstimation of the effective reproduction number Rt\nOur way of estimating Rt extends the methods by Cori et al [3], implemented in a Bayesian framework. In\nbrief, Rt is estimated as the ratio of new locally-acquired infections at time t, to the sum of already infected\nindividuals (local or imported) weighted by an infectivity function  ws that is approximated by the serial\ninterval distribution [3,6]. We parametrically model the serial interval with a Gamma distribution, with user-\nprovided mean and standard deviation (SD); a range can be specified for both parameters, from which values\nare randomly drawn, to include additional uncertainty in the serial interval estimates. As in the original\nmethod, in order to reduce noise in the  Rt estimates a rolling time window can be specified over which\ntransmission is assumed constant; this is often set at 7 days.\nAs dates of symptom onset are often partially missing in real-world surveillance data, we use the subset of\nobserved durations  between symptom onset and case ascertainment,  assumed to also follow a  Gamma\ndistribution,  to  perform  Bayesian  imputation  of  any  unknown  symptom  onset  dates.  This  includes\nasymptomatic cases, for whom we impute an onset date assuming similar duration between infection and\ncase detection as symptomatic cases. In similar fashion, we use a Gamma distribution for the incubation time\n(again with user-provided mean and SD) to impute infection times for all cases. Whereas using symptom\nonset dates results in exact but time-lagging estimates of Rt [3], using infection times produces time-accurate\nestimates, which are necessary in order to assess the effectiveness of control measures and other factors that\npotentially affect transmission. \n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nBayesian imputation of infection times incorporates the appropriate uncertainty in the underlying epidemic\ncurve of infections, reflecting the full range of curves that are compatible with the observed data. At the same\ntime this smoothens out both the epidemic curve and the resulting Rt estimates, making it harder to detect\nabrupt changes in transmissibility and their possible relation to control measures [7]. Although a degree of\nsmoothing is desirable, if this is deemed excessive, a shorter rolling time window can be used.\nDelays in case diagnosis and reporting (case ascertainment) will result in right-truncation, i.e. artificially low\ninfection counts for later days in the time series, biasing recent Rt estimates downwards. Several methods\nhave been proposed to adjust for this, which are often referred to as “nowcasting” [8,9]. Our approach was to\ndivide the later counts by the cumulative probability of ascertainment, as given by the distribution of the\nduration  between infection  and ascertainment.  This  simple  method  has  the  advantage  of  not  requiring\nadditional historical or other data, compared to more elaborate methods. As an alternative we used a “data\naugmentation” approach, whereby future cases were added in the dataset based on the reporting rate of the\nprevious week. A comparison of the two approaches is included in the online supplement (Supplementary\nFigure 1).\nThe overall model is run using Markov Chain Monte Carlo (MCMC) in JAGS [10]. To facilitate its use, we\nhave created an R software package named “bayEStim” ( https://github.com/thlytras/bayEStim) providing a\nuser-friendly interface, in similar fashion to the  “EpiEstim”  package for the original  method  [3]. At a\nminimum the package requires only a vector of symptom onset dates, an indicator of whether the case is\nlocal or imported, and a mean and SD for the serial interval. In addition, the dates of ascertainment and the\nmean and SD of the incubation period can be specified for optimal inference.\nData sources and analysis\nWe applied our method for estimating Rt on national-level COVID-19 surveillance data collected in Greece\nby the National Public Health Organization (NPHO) from 26 February (when the first case was identified) to\n3 August 2020. All cases were laboratory-confirmed with RT-PCR testing, and were considered imported if\nthey had arrived in Greece in the last 3 days before a swab was obtained. For the serial interval we specified\nranges of 4.0-7.5 days for the mean and 2.0-5.0 days for the SD, in order to cover the estimates reported in\nthe literature [11–13]. Similarly, the mean incubation period was specified as 5.1 days with an SD of 3 days\n[14]. A 7-day rolling time window was used for  Rt estimation. Sensitivity analyses were undertaken with\nshorter  time  windows,  as  well  as  with  the  subsets  of  hospitalized cases  and severe  cases  (defined  as\nhospitalized in intensive care or dead). Dates when major control measures were implemented (or relaxed)\nwere overlaid on the resulting time series of  Rt estimates, in order to assess their effects on COVID-19\ntransmissibility. \nIn addition, we sought to examine the association between population mobility, ambient temperature and Rt\nduring the study period. We downloaded the Google mobility data for Greece for the study period [15], and\naveraged the “retail and recreation”, “parks”, “transit stations” and “workplaces” categories into a single\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nrelative mobility measure (compared to the baseline period of 3 January to 6 February 2020). As a sensitivity\nanalysis we omitted “parks” from the mobility measure, as it comprises mostly outdoors activities. From the\nNational  Oceanic  and  Atmospheric  Administration  (NOAA)  website  [16],  we  downloaded  mean  daily\ntemperatures at 48 weather stations across Greece and used a population-weighted average as the overall\ncountrywide temperature of each day. Then, given the 7-day rolling window, we fitted a linear regression of\nthe logRt estimates on the 7-day moving averages of mobility and temperature. To incorporate the uncertainty\nof the dependent variable ( Rt), we ran the regressions separately on each Rt series obtained in each MCMC\niteration and combined the regression coefficients using Rubin’s rules [17]. Alternative specifications of this\nregression  were  explored  as  sensitivity  analyses.  All  analyses  were  undertaken  in  the  R  software\nenvironment, version 4.0.2 [18].\nResults\nUntil 3 August 2020 a total of 4,737 COVID-19 cases had been reported in Greece. Of those, 280 cases were\nfrom two large clusters neither linked to nor representative of the general population (a large cruiseferry\narriving in Greece from abroad, and a migrant hosting facility); these were excluded, leaving 4,459 cases for\nthe  Rt estimation. Table 1 presents the age distribution of these cases and their breakdown by location of\ntransmission and severity; approximately one quarter were imported, one third were hospitalized, and 8%\nwere severe cases, i.e. were hospitalized in intensive care or died. Imported cases were on average younger\nthan local cases, while hospitalized and severe cases were older than non-hospitalized (p<0.001, Table 1). A\ndate of symptom onset was reported for 2,510 cases (56.3%).\nThe epidemic curves based on reporting  date and estimated date of infection (according to the Bayesian\nimputation performed in our model) are shown on Figure 1. The lag between infection and reporting is\napparent, as is the smoothing introduced by the Bayesian imputation that diminishes the noise of day-to-day\nreporting. The vertical dashed lines represent the dates when the most important major control measures\nwere introduced or relaxed (see also Supplementary Table 1): on 11 March 2020 schools were closed, on 14\nMarch the food service industry and shopping centres were closed, and on 23 March all non-essential\nmovement was restricted (i.e. a lockdown was implemented). On 4 March the lockdown was lifted, with\nmost  business  restrictions  gradually relaxed until  the  end of  May.  And on 1 July international  airport\nconnections were restored with most countries, eff ectively allowing the Greek tourism sector to operate.\nAlthough the first peak in daily cases occurred at the time of the lockdown, infections appear to peak much\nearlier, specifically at the time of school closure. Similarly, a second peak on late August corresponds to an\nincrease of infections since early July, coincidental with an influx of imported cases (Figure 1).\nFigure  2  illustrates  the  Rt estimated  by  our  model,  overlaid  on  the  relative  mobility  and  mean  daily\ntemperature time series. From a high of 1.50 (95% CrI 1.14–2.08) just before major control measures were\nintroduced,  Rt fell rapidly below 1 already one week before the lockdown, reaching 0.68 (95% CrI 0.58–\n0.79) on 23 March and staying in this region until May. This rapid decline in Rt mirrors the sharp decline in\nmobility (more than 50%) recorded in the Google data, which was also sustained until May. A small bump in\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nlate May was followed by an Rt consistently below 1 until early July, when the border opening and influx of\nimported cases was subsequently followed by a rise in local COVID-19 cases and Rt (Figures 1 and 2).\nOur delay adjustment method performed relatively well, but consistently underestimated the Rt for the latest\n7-10 days in the time series (Figure 2 and Supplementary Figure 1); we therefore omitted the last week of\nestimates from our regression model (after 27 July). A significant association was found between Rt and both\nmobility and temperature; each 10% of increase in relative mobility increased  Rt by 8.1% (95% CrI 6.1–\n10.2%), whereas each unit celsius of increase in the 7-day temperature average decreased Rt by 4.6% (95%\nCrI 5.4–13.7%). The adjusted R-squared of the model was 54.3%, indicating a fairly good fit to the data, and\nhad the lowest AIC of all alternative models explored as sensitivity analyses (Supplementary Table 2).\nIn sensitivity analyses using hospitalized and severe cases, the  Rt time series was very similar to the one\nobtained  using  all  cases,  albeit  with  much  reduced  precision  given  the  lower  numbers  of  cases\n(Supplementary Figure 2). \nDiscussion\nMonitoring the transmissibility of the SARS-CoV-2 virus during the COVID-19 pandemic is crucial in order\nto assess the effectiveness of control measures and guide public health policy.  Rt is a useful and easily\nunderstandable metric for  this  purpose,  but its reliable estimation presents considerable challenges  [7],\nespecially given the inherent limitations of surveillance data [19]. Our method to estimate Rt does not require\nstructural assumptions other than the serial interval distribution, and introduces several improvements that\nmake it appropriate for use in a surveillance context. It works with minimal or incomplete data, it largely\nadjusts for reporting delays, and can produce time-accurate estimates using an external estimate of the\nincubation period; the latter is essential for correctly assessing the effectiveness of control measures, as any\nintervention to reduce infection rates will only be reflected in case reporting rates after a substantial period of\ntime. Our method is also very easy to use through our “bayEStim” package for the R software environment.\nOn the other hand, there are certain limitations with our approach. Imprecision can be high especially with\nlow case counts, as our model incorporates all sources of uncertainty in Rt, serial interval, incubation period,\nmissing symptom onset dates and delay distribution. Bayesian imputation of infection times also introduces\nsmoothing, which is both appropriate and to an extent desirable, but can also blur abrupt changes in Rt [7].\nThis should be kept in mind when interpreting the results. Substantial changes in testing or ascertainment\nrates over time can bias the results, as can the presence of few large clusters compared to overall cases;\ndown-weighting of cluster-related cases, for example according to their percentage positive compared to\noverall cases, may be a possible solution and has been implemented in “bayEStim”. Ideally our method\nshould be applied to hospitalized or severe case series, which can be less subject to underascertainment or\nnon-random testing, although in our case Rt estimates were not appreciably different.\nApplying our method of  Rt estimation to Greek COVID-19 surveillance data yielded several important\nresults, with clear implications for the past and future management of the pandemic. First, there was a clear\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nrelationship between population mobility and viral transmissibility, indicating that social distancing and\nstaying home can be very effective in bringing the Rt of COVID-19 below 1 [20]. Indeed both mobility and\nRt started falling rapidly just as the first control measures were introduced, indicating that the multiple\ncontrol measures introduced, combined with good public compliance, resulted in effective social distancing\nearlier than the full lockdown of 23 March . This was likely crucial, as the virus did not get enough time to\nspread widely among the population. Counterin tuitively, the lockdown appears to not have substantially\ncontributed to a further reduction in COVID-19 transmissibility, as Rt had already fallen significantly below\n1 at least a week earlier and remained stable thereafter. It is possible that without the lockdown the decrease\nin mobility and Rt might not have been sustained, but this purely hypothetical.\nAs social distancing measures were relaxed, Rt in Greece remained largely below 1 until early July. Then a\nlarge influx of tourists (with 1.3 million arrivals during all of July  [21], in a population of around 10.5\nmillion), with hundreds of imported COVID-19 cases (Figure 1), was followed by a steep increase in locally-\nacquired cases and  Rt climbing back to 1. This coincided with a continued increase in mobility beyond\nbaseline levels (Figure 2). It should be emphasized that the arrival of imported cases from abroad does not in\nitself  raise  Rt,  as  transmissibility  depends  entirely  on  local  conditions  of  social  mixing  and  mobility.\nHowever, summer in Greece is defined by people, Greek and foreign, going on vacation in large numbers\nand in proximity to each other, especially in tourist hotspots; moreover, an estimated 850,000 local jobs are\ndirectly or indirectly linked to the tourism industry  [22]. This situation creates the general conditions for\nsustained spread of the virus among the population, through increased mobility and crowding. It is the\ncombination of increased case importations and increased transmissibility that results in more COVID-19\ncases identified during the summer, a phenomenon which is expected to abate in September as the tourist\nseason winds down. In the meantime, is is important to push  Rt back below 1 as much as possible, with\ntargeted social distancing measures and widespread use of other protective measures like masks and hand\nhygiene.\nFinally, we found a relationship between ambient temperature and COVID-19 transmissibility; in fact the\n18.3 degrees celsius difference between the lowest and highest temperature in our data corresponds to a\n58.1% change in Rt (95% CrI 35.7–72.7%), which appears quite significant. This may be attributed to the\neffects of heat on the virus itself, but is likely more down to behavioural patterns in the population, with\nmore social interaction happening outdoors as the weather gets warmer and vice versa. This effect however\ndoes not ap pear sufficient to counterbalance the increased mobility associated with the summer season in\nGreece, which is unsurprising considering the lack of population immunity to SARS-CoV-2 [23]. Moreover,\nwith  the  higher  transmissibility  anticipated  during  the  winter,  a  substantially  higher  degree  of  social\ndistancing will likely be required to keep Rt below 1 and the pandemic under control.\nIn conclusion, we present an improved method and software for estimating Rt that is tailored to the realities\nof a disease surveillance context, and can be a valuable component of the surveillance activities during the\ncurrent  COVID-19  pandemic.  Applying  our  method  to  the  Greek  case  demonstrates  the  important\nconclusions that can be drawn, and that can guide the further management of the pandemic.\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nConflict of Interest statement\nST and VS serve on the Expert Advisory Group for COVID-19 of the Hellenic Ministry of Health, which is\nan unpaid position. TL and DP declare no competing interests.\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nBibliography\n1. Wu F, Zhao S, Yu B, Chen Y-M, Wang W, Song Z-G, et al. A new coronavirus associated with human \nrespiratory disease in China. Nature. 2020;579(7798):265–9. \n2. Rapid Risk Assessment: Coronavirus disease 2019 (COVID-19) in the EU/EEA and the UK – eleventh \nupdate: resurgence of cases [Internet]. European Centre for Disease Prevention and Control. 2020 [cited \n2020 Aug 12]. Available from: https://www.ecdc.europa.eu/en/publications-data/rapid-risk-assessment-\ncoronavirus-disease-2019-covid-19-eueea-and-uk-eleventh\n3. Cori A, Ferguson NM, Fraser C, Cauchemez S. A new framework and software to estimate time-varying \nreproduction numbers during epidemics. Am J Epidemiol. 2013 Nov 1;178(9):1505–12. \n4. Wallinga J, Lipsitch M. How generation intervals shape the relationship between growth rates and \nreproductive numbers. Proc Biol Sci. 2007 Feb 22;274(1609):599–604. \n5. Bettencourt LMA, Ribeiro RM. Real time bayesian estimation of the epidemic potential of emerging \ninfectious diseases. PLoS ONE. 2008 May 14;3(5):e2185. \n6. Thompson RN, Stockwin JE, van Gaalen RD, Polonsky JA, Kamvar ZN, Demarsh PA, et al. Improved \ninference of time-varying reproduction numbers during infectious disease outbreaks. Epidemics. \n2019;29:100356. \n7. Gostic KM, McGough L, Baskerville E, Abbott S, Joshi K, Tedijanto C, et al. Practical considerations \nfor measuring the effective reproductive number, Rt. medRxiv. 2020 Jun 20; \n8. Höhle M, an der Heiden M. Bayesian nowcasting during the STEC O104:H4 outbreak in Germany, \n2011. Biometrics. 2014 Dec;70(4):993–1002. \n9. McGough SF, Johansson MA, Lipsitch M, Menzies NA. Nowcasting by Bayesian Smoothing: A \nflexible, generalizable model for real-time epidemic tracking. PLoS Comput Biol. 2020;16(4):e1007735.\n10. Plummer M. JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling. In \nVienna; 2003. p. 125. \n11. Bi Q, Wu Y , Mei S, Ye C, Zou X, Zhang Z, et al. Epidemiology and transmission of COVID-19 in 391 \ncases and 1286 of their close contacts in Shenzhen, China: a retrospective cohort study. The Lancet \nInfectious Diseases [Internet]. 2020 Apr 27 [cited 2020 Apr 30];0(0). Available from: \nhttps://www.thelancet.com/journals/laninf/article/PIIS1473-3099(20)30287-5/abstract\n12. Du Z, Xu X, Wu Y , Wang L, Cowling BJ, Meyers LA. Serial Interval of COVID-19 among Publicly \nReported Confirmed Cases. Emerging Infect Dis. 2020 Mar 19;26(6). \n13. Liu T, Qi L, Yao M, Tian K, Lin M, Jiang H, et al. Serial Interval and Reproductive Number of COVID-\n19 Among 116 Infector-infectee Pairs — Jingzhou City, Hubei Province, China, 2020. CCDCW. 2020 \nJul 1;2(27):491–5. \n14. Lauer SA, Grantz KH, Bi Q, Jones FK, Zheng Q, Meredith HR, et al. The Incubation Period of \nCoronavirus Disease 2019 (COVID-19) From Publicly Reported Confirmed Cases: Estimation and \nApplication. Ann Intern Med. 2020 Mar 10; \n15. Google. COVID-19 Community Mobility Reports [Internet]. COVID-19 Community Mobility Reports. \n[cited 2020 Aug 1]. Available from: https://www.google.com/covid19/mobility?hl=en\n16. National Oceanic and Atmospheric Administration (NOAA). National Centers for Environmental \nInformation (NCEI) Climate Data Online Search [Internet]. [cited 2020 Aug 1]. Available from: \n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nhttps://www.ncdc.noaa.gov/cdo-web/search\n17. Rubin DB. Multiple imputation for nonresponse in surveys. New York: John Wiley & Sons; 1987. \n18. R Core Team. R: A Language and Environment for Statistical Computing [Internet]. Vienna, Austria: R \nFoundation for Statistical Computing; 2019. Available from: http://www.R-project.org/\n19. Giesecke J. Routine surveillance of infectious diseases. In: Modern infectious disease epidemiology. \nCRC Press; 2017. \n20. Sypsa V , Roussos S, Paraskevis D, Lytras T, Tsiodras S, Hatzakis A. Modelling the SARS-CoV-2 first \nepidemic wave in Greece: social contact patterns for impact assessment and an exit strategy from social \ndistancing measures. medRxiv. 2020 May 29;2020.05.27.20114017. \n21. Greek City Times. Number Of Imported Coronavirus Cases Remains Low, Says Greek Tourism Minister\n[Internet]. Greek City Times. [cited 2020 Aug 15]. Available from: \nhttps://greekcitytimes.com/2020/08/07/number-of-imported-coronavirus-cases-remains-low-says-greek-\ntourism-minister/\n22. ekathimerini.com. WTTC praises Greece’s 2019 tourism growth and response to Covid-19 [Internet]. \n[cited 2020 Aug 16]. Available from: \nhttp://www.ekathimerini.com/252473/article/ekathimerini/business/wttc-praises-greeces-2019-tourism-\ngrowth-and-response-to-covid-19\n23. Bogogiannidou Z, V ontas A, Dadouli K, Kyritsi MA, Soteriades S, Nikoulis DJ, et al. Repeated leftover \nserosurvey of SARS-CoV-2 IgG antibodies, Greece, March and April 2020. Euro Surveill. 2020;25(31). \n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nTable 1: Age distribution of laboratory-confirmed COVID-19 cases by location of transmission and \nseverity, Greece, 26 February to 3 August 2020\nAll cases Local Imported Hospitalized Severe\nMedian age (IQR) 46 (31–60) 48 (32–62) 41 (28–53) 60 (47–73) 70 (59-81)\n0-17 years 293 (6.6%) 251 (7.7%) 42 (3.6%) 37 (2.6%) 1 (0.3%)\n18-39 years 1385 (31.1%) 903 (27.5%) 482 (41.0%) 192 (13.7%) 6 (1.7%)\n40-64 years 1822 (40.9%) 1367 (41.7%) 455 (38.7%) 606 (43.1%) 123 (34.8%)\n65+ years 819 (18.4%) 701 (21.4%) 118 (10.0%) 569 (40.5%) 223 (63.2%)\nAge unknown 138 (3.1%) 58 (1.8%) 80 (6.8%) 2 (0.1%) 0 (0.0%)\nTotal 4457 (100.0%) 3280 (73.6%) 1177 (26.4%) 1406 (31.5%) 353 (7.9%)\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nFigure Legends\nFigure 1: Daily confirmed COVID-19 cases by reporting date and estimated day of infection, Greece,\nFebruary–August 2020\nFigure  2:  Daily  COVID-19  effective  reproduction  number  ( Rt),  relative  mobility  and  mean\ntemperature, Greece, February–August 2020\nFootnote for Figure 2: Vertical lines  represent the dates when the most important major control measures\nwere introduced or relaxed (see main text for details)\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nMar Apr May Jun Jul Aug\n0\n20\n40\n60\n80\n100\n120\n(a) COVID-19 cases by reporting date\nNumber of daily cases\nImported\nLocal\nMar Apr May Jun Jul Aug\n0\n20\n40\n60\n80\n100\n120\n(b) COVID-19 cases by estimated day of infection\nEstimated number of daily infections\nImported\nLocal 95% Credible Band\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint \n\nMar Apr May Jun Jul Aug\n0.0 0.5 1.0 1.5 2.0 2.5\n(a) COVID-19 Effective Reproduction Number (Rt)\nRt\n95% Credible Bands\nWith data augmentation\nNo delay adjustment\nMar Apr May Jun Jul Aug\n(b) Relative mobility, 7-day average\nMobility (%)\n0 50 100\nMar Apr May Jun Jul Aug\n(c) Mean daily temperature, 7-day average\nTemperature (℃)\n10 20 30\n . CC-BY 4.0 International licenseIt is made available under a \nperpetuity. \n is the author/funder, who has granted medRxiv a license to display the preprint in(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted September 22, 2020. ; https://doi.org/10.1101/2020.09.19.20198028doi: medRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}