Human mobility and poverty as key factors in strategies against COVID-19

The unprecedented COVID-19 pandemic that swiped across the globe led many countries to apply 14 heavy nationwide restrictions and control measures. Analyzing aggregate and anonymized 15 mobility data from the cell-phone devices of >3 million users in Israel, we identified that poorer 16 regions exhibited lower and slower compliance with the restrictions. We integrated these mobility 17 patterns into age-, risk- and region-structured transmission model, and showed how we can explain 18 the spatiotemporal dynamics of 250 regions covering Israel. Model projections suggest that 19 applying localized and temporal interventions that focus on high-risk groups can substantially 20 reduce mortality, particularly in poorer regions, while enabling daily routine for a vast majority of 21 the population. These trends were consistent across vast ranges of epidemiological parameters, 22 possible seasonal forcing, and even when we assumed that vaccination would be commercially 23 available in 1-3 years. Our findings can help policymakers worldwide identify hotspots and apply 24 designated strategies against future COVID-19 outbreaks.


31
Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) was identified in Wuhan, China,32 in December 2019. It has since developed into a pandemic wave affecting over 200 countries, 33 causing over 5.4 million cases and claiming over 340 thousand lives, as of May 24, 2020 (1). The 34 rapid growth of the SARS-CoV-2 pandemic led to unprecedented control measures on a global 35 scale. As of May 2020, travel bans, restrictions on mobility of varying degrees, and nationwide 36 lockdowns have emerged sharply in over 200 countries (2). In Israel, since March 9, 2020, travelers 37 from any country are being denied entry unless they can prove their ability to remain under home 38 isolation for 14 days. From March 16 onward, daycare and schools were shut, and work was 39 limited to less than a third of the capacity. On March 26, inessential travel was limited to 100 40 meters away from home, and three lockdowns were applied in most regions in Israel to prevent 41 crowding due to holiday celebrations (3). 110 To explore the spatiotemporal effect of human mobility and poverty on transmission, we calculated 111 the number of new cases and the amount of travel between zones observed during three periods:  Supplementary materials). This lag includes the incubation period, the time from symptom onset 122 until a test is conducted, and the time until the test results arrive.

124
We integrated the daily mobility data into an age-, region-, and risk-stratified model for SARS- overestimation of disease transmission. We also found that a model that accounted for seasonal 132 forcing yielded a higher, but not significant (p value<0.35), likelihood than a model that did not 133 account for seasonal forcing (Table S5, Supplementary materials).

135
Focused lockdowns reduce mortality 136 As transmission varied considerably among regions, we projected the number of total deaths for 137 1-3 years under local and temporal lockdown strategies. Specifically, we simulated three strategies 138 triggered by a threshold of daily COVID-19 incidence in each of the 250 regions where we 139 considered a lockdown for 1) the entire population in the region, 2) daycare-and school-age 140 children (between 0-19 years of age (children), and 3) high-risk groups and individuals >65 years 141 of age (high-risk). To examine the efficiency of local strategies compared to nationwide strategies, 142 we also simulated a global strategy triggered by similar national daily incidence. When a lockdown 143 is applied, we consider the same compliance rate as that observed during previous lockdowns, 144 which is reflected in our data for each region by different values of the MI and travel between 145 zones.
We evaluated the efficiency of the lockdown strategies, defined as the number of deaths averted 148 per lockdown day (Figure 4). We found that the local strategy of targeting the high-risk group was 149 substantially more efficient than any other strategy. For example, assuming the proportion of 150 unreported cases is 85% and a lockdown threshold of 5/10,000 (cases/individuals), a strategy 151 targeting the high-risk group is 4.3-5.5 times more efficient than a global strategy ( Figure 4C and 152 D).

154
We evaluated the effectiveness of each strategy in reducing mortality ( Figure 5). We found that a 155 strategy locally targeting the high-risk group yielded a lower number of deaths than a strategy 156 targeting children. For example, assuming the proportion of unreported cases is 85% and a 157 lockdown threshold of 5/10,000 (cases/individuals), a strategy targeting the high-risk group 158 resulted in 4,500-4,900 deaths while on targeting children resulted in 7,900-10,500 deaths after 159 one year ( Figure 5). In addition, for lockdown thresholds exceeded 5/10,000, which aligns with 160 the current practice in Israel, a strategy locally targeting the high-risk group either is projected to Our key findings suggest that COVID-19 infection does not spread uniformly in the population, 168 and thus, intervention strategies should be localized and temporal and should focus primarily on 169 protecting individuals at high risk. Such a strategy can reduce mortality while enabling daily 170 routine for a vast majority of the population. Furthermore, temporary lockdown strategies that 171 focus on the population at high risk were found to be most efficient and likely to result in 172 comparable mortalities to lockdown strategies of all individuals in a region.

174
Our work demonstrates that to understand the spatiotemporal dynamics of transmission, models 175 must account for mobility as well as behavioral aspects that are associated with sociodemographic 176 and socioeconomic factors. In particular, we found that SARS-Cov-2 is more likely to spread in 177 more impoverished regions and is affected by human mobility. The intensive interactions likely 178 led to higher transmission in developed countries than in developing countries. However, our 179 model suggested that people of low SES are at higher risk due to poorer compliance and larger 180 household size.

182
Our analyses indicate that localized lockdowns with incidence thresholds as low as five reported 183 cases in 10,000 individuals are essential to decrease mortality. This finding underscores the 184 importance of maintaining a high level of testing (17), particularly in regions with elevated risk of 185 transmission. However, with such a strategy, at least 2500 total years of lockdowns (equivalent to 186 a one-day lockdown of 912,500 individuals) are required to prevent a single death. Considering 6 that one day of lockdown is equivalent to a quality of life value that is ~0.85 times that in a routine 188 day (18), even local lockdowns should be prudently considered from a health economic 189 perspective. Thus, future modeling studies should also include localized and temporal massive 190 screening efforts, which result in more focused quarantines and isolations than massive lockdowns.

192
As in any modeling study, we made several simplifying assumptions. Our local lockdowns 193 correspond to regions with a population of ~36,000 people. A smaller lockdown may be more 194 efficient but could not be tested by our model. Additionally, with the growing evidence of a 195 disproportionate risk from COVID-19 to the elderly (10, 19), focused control measures are likely 196 to be conducted in retirement homes and facilities with populated communities at high risk, which 197 we did not explicitly account for in our model (20). Although the transmission dynamics are 198 unlikely to change with such focused interventions, the overall mortality is expected to be lower 199 than what we have found.

201
While there is a debate in the literature regarding the extent of infectiousness and transmissibility 202 in children (21), our results highlighted a not less important question: to whom do children 203 transmit? Our findings reveal that children are less likely to transmit to populations at risk, and 204 thus, a differential lockdown strategy that targets children is not the most efficient or effective in 205 reducing mortality.

207
In conclusion, we showed that using aggregated and anonymized human mobility data from 208 cellular phones under the General Data Protection Regulation (GDPR) guidelines is a powerful 209 tool to improve the understanding of transmission dynamics and to evaluate the effectiveness of 210 control measures. Our transmission model predicted that rather than nationwide lockdowns, 211 applying temporal and localized lockdowns that focus on groups at high risk can substantially 212 reduce mortality. Such focused measures will enable a vast majority of the population to maintain 213 a daily routine. Our findings can help policymakers worldwide identify hotspots and apply 214 designated strategies against the ongoing outbreak and future second waves.

218
Human mobility 219 Our data include mobility records based on cellular data of >3 million users from one of the largest 220 telecommunication companies in Israel. With the exception of children <10 years of age, the users 221 are well representative of Israel demographically, ethnically, and socioeconomically. In 222 accordance with the GDPR, the data include aggregated and anonymized information. The data 223 specifies movement patterns within and between 2,630 zones covering Israel, on an hourly basis, 224 from February 1, 2020, until May 16, 2020. To ensure privacy, if less than 50 individuals were 225 identified in the zone in a given hour, the number of reported individuals was set to zero.

227
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020. . https: //doi.org/10.1101//doi.org/10. /2020 We determined the location of individuals based on the triangulation of cell towers, which was 228 found to be accurate to 300 meters in most cases but varied by up to 1 km in less populated areas.

229
To prevent signal noise and identify stay points, we tracked only locations where users stayed for 230 at least 15 minutes within a distance threshold of 1.7 km. We defined users as residents of a zone 231 based on the location at which they had the highest number of signals on most nights during 232 February 2020.

234
To calculate the MI for each zone, we counted the daily number of individuals in each group that 235 showed a signal away from their home location. Conservatively, we counted only individuals who 236 were located more than 1.5 km away from home.

238
Next, we integrated data from the Central Bureau of Statistics (CBS) that specifies several 239 socioeconomic characteristics, including population size, household size, age distribution, 240 socioeconomic score, and dominant religion, for each zone. Each zone includes ~3,500 residents.

241
For each zone, we scaled the number of resident users of the telecommunication company to match 242 the actual number of residents in the zone, as reported by the Israeli CBS. The CBS specifies for 243 each zone a socioeconomic cluster from 1 to 10. Based on these clusters, we defined three SES 244 groups that were nearly equal in size: low (clusters 1-3), middle (clusters 4-7), and high (clusters 245 8-10). We aggregated the MI according to SES to test the mobility trends on a national level 246 ( Figure 1A). To evaluate the travel patterns based on an individual's SES ( Figure 1B and 1C), we 247 counted the mean daily number of travels between the 2,630 zones, including for those individuals 248 who stayed in their origin zone. Grouping by SES and scaling the daily number of travels to one 249 for each zone, we created an origin-destination travel probability matrix.

251
To analyze the relationship among poverty, mobility, and transmission ( Figure 2), we divided the 252 data into three periods: 13  May, corresponding to 1) 253 the early phase before restrictions started, 2) the time from restrictions until they were first lifted,

260
Transmission model 261 We developed a dynamic model for age-, risk-and region-stratified SARS-Cov-2 infection 262 progression and transmission in Israel. Our model is a modified susceptible exposed infected  (which was not certified by peer review) 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 6, 2020. . https://doi.org/10. 1101/2020 We distinguished high-risk and low-risk individuals in each age group based on the ACIP case 268 definition (23, 24). We also distinguished the 250 regions covering Israel in the model.

288
The rate at which individuals transmit depends on (i) contact mixing patterns between the infected 289 individual and his or her contact, (ii) age-specific susceptibility to infection, (iii) region-based 290 behavioral susceptibility, and (iv) potential seasonal forcing.

291
Age-specific contact rates were parameterized using data from an extensive survey of daily 292 contacts (35) and data from CBS regarding the household size in each region. In addition, we  We distinguished between in-home and out-of-home transmission. We evaluated the in-home 304 transmission is independent of age, and based on a previous retrospective studies, that suggested 305 All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

307
To account for behavioral susceptibility, we explicitly considered in our model a parameter 308 reflecting the order to maintain physical distancing, . The high regional variations in 309 susceptibility were parameterized based on fertility rates and socioeconomic characteristics.

310
Specifically, we computed for each region the relative change in mobility compared to routine.

311
Our analysis indicated that for regions of low SES, the change was lower, which was reflected in 312 our model by higher susceptibility (Supplementary materials). The use of regional fertility and 313 relative change in mobility allowed us to refrain from calibrating the model to an excessive number 314 of unknown parameters and avoid overfitting.  Due to the uncertainty regarding the proportion of unreported cases, we calibrated our model to 337 different scenarios. Specifically, underreporting is affected by testing policy and testing 338 All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020. . https://doi.org/10.1101/2020.06.04.20112417 doi: medRxiv preprint capabilities for each country, as well as individuals' tendency to seek care once clinical symptoms 339 appear. In addition, underreporting is affected by the severity of the infection, which is associated 340 with age (10). Thus, we chose different estimates for the proportion of underreporting, ranging 341 from 5.5-14 unreported cases for a single reported case. These estimates are based on observations 342 from screenings conducted in Denmark, Czechia, Netherlands; Santa Clara, California (10, 16, 40) 343 (Table S1, Supplementary materials). Due to the uncertainty related to positive predictive values 344 of serological screenings, we also tested a scenario of 2 unreported cases for a single reported case 345 to confirm the robustness of our findings.

346
To account for the age variation, we considered the detailed serological data from Santa Clara. We  Infectious Diseases, which suggested that a vaccine could be available by May 2021.

357
Each strategy considered includes a threshold for activation of a lockdown, and the groups 358 considered for lockdown were as follows: 1) the entire population in the region, 2) daycare-and 359 school-age children between 0-19 years of age (children), 3) high-risk groups and individuals >65 360 years of age (high-risk).

361
Thus, to model the lockdown strategies, we defined an indicator for each region as the weekly 362 number of new-reported cases per 10,000 people. Each week, we examined whether the indicator 363 exceeds a certain threshold for each region. If so, a lockdown was activated for the following week.

364
This process was continued for 1-3 years.

365
To project the number of individuals who will die under each strategy considered, we utilized 366 available detailed information from the Israeli Ministry of Health (Table S2, Supplementary 367 materials). Specifically, we calculated for each age and risk group the proportion of individuals 368 who died out of the reported cases. We multiplied these proportions with the daily model 369 projections of newly reported cases and summed this product to calculate the total projected 370 number of deaths. We also accounted for the uncertainty regarding the estimated probabilities. We 371 define the efficiency of a lockdown strategy as the total number of deaths averted per total 372 All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020. . https://doi.org/10.1101/2020.06.04.20112417 doi: medRxiv preprint lockdown days. The number of deaths averted is calculated as the projected number of deaths with 373 no lockdowns minus the number of deaths projected when the considered strategy is applied. 374 375 376 All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020.  (which was not certified by peer review) 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 6, 2020.      (which was not certified by peer review) 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 6, 2020.   (which was not certified by peer review) 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 6, 2020.  (which was not certified by peer review) 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 6, 2020.  (which was not certified by peer review) 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 6, 2020.   Competing interests: The authors declare that they have no competing interests.

600
Data and materials availability: The data that support the findings of this study are available 601 from the authors but restrictions apply to the availability of these data, which were used under 602 license for the current study and so are not publicly available. Data are however available from 603 the authors upon reasonable request.

605 606
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020.

615
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020.

626
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020.  (which was not certified by peer review) 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 6, 2020.

643
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

649 650
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020.  (which was not certified by peer review) 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 6, 2020. We developed a dynamic model for age-, risk-and regions-stratified SARS-Cov-2 infection 677 progression and transmission in Israel. Our model is a modified Susceptible-Exposed-Infected-  ≥70 years. (13, 41, 42). We distinguished between high-risk and low-risk individuals for each 683 age group based on the ACIP case definition (23,24). We also distinguish in the model between 684 250 regions covering Israel.

686
Multiple infections with SARS-Cov-2 is yet fully understood. A recent study indicated that there 687 is a protective immunity following infection in humans (31)

693
The mean incubation period of SARS-Cov-2 is 6.4 days (95% CI, 5.6 to 7.7 days) (25, 26), but 694 first evidence shows viral shedding occurs during a pre-symptomatic stage (27,28). Thus, we 695 considered an exposed period , and an early infectious period . Underreporting arises 696 from asymptomatic cases or mild cases of individuals that do not seek care (16,40,46,47). Thus, 697 following the early infectious phase, individuals in the model transition either to an infectious and 698 reported compartment , or to infectious and unreported compartment .

699
To enable in our model for a subset of the population to go for intervention (e.g., 30% of the 700 individuals from specific regions, age groups or risk-group to go under lockdown during a selected 701 time period), we also specifically distinguish between those who undergo and those who did not 702 undergo an intervention.

703
Accordingly, we stratified the population into six health-related compartments: (which was not certified by peer review) 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 6, 2020. consider a function of the initial spreaders with time , , ( ), that reflects the individuals exposed 726 to the virus the entered Israel from overseas between February 21 2020 -and March 9, 2020. Thus, 727 the transmission model is composed of the following system of difference equations: 728 All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020. . https://doi.org/10. 1101/2020 to the proportion of susceptible individuals in the house Age-specific susceptibility rate for 748 individuals out-of-home , was parameterized by calibrating our model with daily COVID-19 749 records (See Section 3. calibrated parameters).

751
To account for behavioral susceptibility, we explicitly considered in our model a parameter 752 reflecting the order to maintain physical distancing, , as vast number of countries, including 753 Israel, adopted measures such as physical-distancing to control the susceptibility of SARS-Cov-2 754 (48). This parameter was calibrated to the epidemiological data of COVID-19 in Israel. Moreover, 755 the high regional variations in susceptibility were parameterized based on fertility rates and 756 socioeconomic characteristics relative to the national average, using the data from Central Bureau 757 of Statistics (CBS), . Specifically, we computed for each region the relative reduction in travels 758 >1.5 km compared to routine , , (See Section 2.2 Relative reduction in travels). Our analysis 759 indicated that for regions of low SES the change was lower, which was reflected by our model 760 with higher susceptibility.

762
Seasonal patterns have been observed in common circulating HCoVs, mostly causing infections 763 in humans between December and May in the Northern Hemisphere (37). The two human 764 coronaviruses 229 E and OC43 show distinct winter seasonality. In addition, many coronaviruses 765 in animals do exhibit a distinct seasonal pattern of incidence in their natural hosts (36). There is 766 growing evidence that SARS-CoV-2 is also seasonal, with the optimal setting for transmission in All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020.

786
Household contacts 787 We estimated the contact mixing at home for each region based on the average household size and 788 its age distribution from the Israeli Central Bureau of Statistics (CBS) (50, 51). We assume all 789 individuals in the same household will meet with each other daily regardless of the control 790 measures applied by the country (e.g. lockdowns). The CBS data suggest that low socioeconomic 791 status is characterized by larger and younger household size.

793
Work and leisure contact patterns 794 Age-specific contacts 795 We parametrized the age-specific contact rates using data from a survey of daily contacts collected 796 in eight European countries (35). This contact data includes contact rates for different locations: 797 works (or school for children <10), leisure. In addition, the data exhibits frequent mixing between 798 similar age-groups, moderate mixing between children and adults in their thirties (likely their 799 parents), and infrequent mixing between other groups. To generate the age-specific contact mixing 800 used in our model, we used the means of each age-group over the eight countries. To ensure the 801 matrices is symmetric and convert between age-groups used in the survey to those used in out 802 model, we adjusted the contact matrices according to the means for reciprocal age group pairing 803 (33).

805
Origin-destination from mobility data 806 807 Our data includes mobility records based on cellular data of >3 million users from one of the 808 largest telecommunication companies in Israel. The data specifies movement patterns within and 809 between 2,630 zones covering Israel, on an hourly basis, from February 1, 2020, and until May on the triangulation of cell towers, which was found accurate to 300 meters in most cases but 813 All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020. . https://doi.org/10.1101/2020 varied to 1 km in less populated areas. We defined users as residents of a zone based on location 814 in which they had the highest number of signals on most nights during February 2020.

816
We used this data to develop aggregated origin-destination (OD) matrices between and within 817 zones. To refrain from signal noises and identify stay points, we track only locations where users 818 stayed for at least 15 minutes within a distance threshold of 1.7 km. The OD matrices serve as a 819 proxy to the flow from each region to another.

821
Next, we integrated data from the Central Bureau of Statistics (CBS) that specifies for each zone 822 several socioeconomic characteristics, including population size, household size, age distribution,

835
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020. (which was not certified by peer review) 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 6, 2020. .

848
To integrate the age-specific contact matrices and the OD matrices we multiplied the number of 849 contacts for each age-group by the travel distribution for each region in the OD matrices. We 850 assumed that at work, children at the age of 0-9 years old, remains at their home region. We also 851 assumed that at leisure time children at the age of 0-9 years old movement patterns are like their 852 parents. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Relative reduction in travels 858
For each region, we computed the relative reduction in travels >1.5 km , , . This measure was 859 done scaling the daily proportion of travels more than 1.5 km out-of-home. To compute this minimal and maximal values and refrain from outliers, we averaged the three 862 minimal and three maximal values. This measure was found to be highly correlative with disease 863 growth factor ranging between 79.2-82.8% (p value<0.001) for a shift of 12-14 days ( Figure S3).

864
Thus, we incorporated for each region this measure in the model.  (which was not certified by peer review) 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 6, 2020. estimates of unreported ratios 1:5.5 (Scenario A), 1:9 (Scenario B), and 1:14 (Scenario C). It is not 892 clear how much reutilizing antibodies are sufficient to ensure protection, and thus it is possible 893 serological screenings serve as over estimation to determine exposure. Thus, to determine the 894 robustness of our findings, we also considered an extreme scenario of 1:2 (Scenario D). 895 We estimated the proportion of under reporting for each age-group by scaling the estimates from 896 Santa-Clara Study to the age reported cases in this region (52). This analysis suggested that 897 younger age-groups are more likely to be unreported. Conservatively, we assumed that all cases 898 among individuals at high-risk are reported. Using these estimates and based on the reported cases 899 in Israel between February 20 th -May 14 th ,2020, we obtained that overall proportion of unreported 900 cases is 85% for scenario A, 89% for scenario B, 93% for scenario C and 69% for scenario D. (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

901
The copyright holder for this preprint this version posted June 6, 2020. The probability of death for each age-and risk-group given a reported case was evaluated based on 907 the Israeli Ministry of Health case report data (Table S2).  (Table S1). We entered these initial spreaders, , , , ( ), to the exposed compartment.

919
Susceptibility at-home 920 We distinguish between in-home versus out-of-home transmission. Consistent with a previous 921 study (8). We specifically distinguish between the susceptibility of those settings. We estimated 922 the in-home susceptibility rate, ℎ , based on a previous study that showed a secondary attack 923 rate of 16.3% throughout the entire infectious period (11).

925
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020.
Fertility rate for each region k relative to the nation's mean.
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020.  The calibration was conducted on a 30 sub-district level rather than 250 regions to ensure there are 937 sufficient time-series data points in each location for each age group. The stratification is based on 938 the 16 formal districts, which we further stratified such that the sub districts will be homogenous 939 in terms of their SES and religious affiliation (Table S4). To calibrate the model to the incidence data,

940
we maximized the likelihood assuming a normal distribution of the error between model predictions and 941 incidence data. This was achieved by using the truncated Newton (TNC) algorithm. We calibrated the 942 model for 16 different scenarios of unreported cases and seasonal forcing. The final transmission model 943 included five parameters without constraints imposed from previous data: reduced susceptibility 944 due to physical distancing , and susceptibility rate based on age-groups j: 0-19, 20-39, 40-59, 945 and >60 (Table S5).

946
We used an F-test of equality of variances to compare between models 1) with vs. without 947 consideration of seasonal forcing, 2) with and without consideration of human mobility, 3) with 948 and without consideration of regional fertility. We denote that in all three comparisons, the number 949 of calibrated parameters is constant and equal to five. Our tests suggested that models that do not 950 include the mobility data (p.value<0.01), and the regional fertilities (p.value<0.01) were 951 significantly worse. We also found that models that accounted for seasonal forcing yielded higher, 952 but not significant (p value<0.35), likelihood than models that did not account for the seasonal 953 forcing.

Sub-district number Name Population Size
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
(which was not certified by peer review) 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 6, 2020. . https://doi.org/10.1101/2020.06.04.20112417 doi: medRxiv preprint 4. Further simulation results 963 We found that a global lockdown strategy had a larger temporal effect than local lockdowns and 964 had by greater oscillations ( Figure S4). We present here a model with seasonal forcing. Our model 965 projections suggested that global lockdowns were less efficient and effective compared to a 966 strategy that targets locally the elderly. However, due to high variability between the 250 regions 967 considered, some regions undergo multiple lockdowns, while others will not undergo lockdowns.

968
Local lockdowns that specifically target children decreases the local morbidity, but in the long run 969 increases mortality, while lockdowns of individuals at high-risk has a moderate impact on 970 transmission but decreases mortality.

971
These findings where robust across all settings considered (Table S3 and Table S5

980
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) 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 6, 2020.

986
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.