Assessing optimal time between doses in two-dose vaccination regimen in an ongoing epidemic of SARS-CoV-2

The SARS-CoV-2 pandemic is a major concern all over the world and, as vaccines20 became available at the end of 2020, optimal vaccination strategies were subjected to intense investigation. Considering their critical role in reducing disease burden, the increasing demand outpacing production, and that most currently approved vaccines follow a two-dose regimen, the cost-effectiveness of delaying the second dose to increment the coverage of the population receiving the first dose is often debated. Finding the best solution is complex due to the trade-25 off between vaccinating more people with lower level of protection and guaranteeing higher protection to a fewer number of individuals. Here we present a novel extended age-structured SEIR mathematical model that includes a two-dose vaccination schedule with a between-doses delay modelled through delay differential equations and linear optimization of vaccination rates. Simulations for each time window and30 for different types of vaccines and production rates were run to find the optimal time window between doses, that is, the one that minimizes the number of deaths. We found that the best strategy depends on an interplay between the vaccine production rate and the relative efficacy of the first dose. In the scenario of low first-dose efficacy, it is always better to apply the second dose as soon as possible, while for high first-dose efficacy,35 the optimal window depends on the production rate and also on second-dose efficacy provided by each type of vaccine. We also found that the rate of spread of the infection does not affect significantly the thresholds of the optimal window, but is an important factor in the absolute . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice. 2 number of total deaths. These conclusions point to the need to carefully take into account both vaccine characteristics and roll-out speed to optimize the outcome of vaccination strategies.40

– SARS-CoV-2; Compartmental models; Delay differential equations; Linear optimization models. I. I NTRODUCTION As the implementations of SARS-CoV-2 vaccination programs evolve in several countries around the world, governments must decide how to allocate doses among45 the population during the epidemic, usually prioritizing the most at-risk groups, such as healthcare workers, people with comorbidities, and older adults. In a context of limited vaccine supply, an optimized dose allocation strategy is critical to effectively immunize the population whilst achieving the best reduction in hospitalizations and deaths. On this basis, some countries considered partially protecting a greater percentage of the50 population by administering the first dose more widely at the expense of delaying second doses, and thus having less fully vaccinated individuals [16, 20, 28]. Vaccination programs can have different goals depending on the context and char- acteristics of both the disease and the available vaccines. The immunization program against Covid-19 has focused on reducing the burden of hospitalizations and deaths,55 which is justified given the risk of health care system collapse and the lack of vaccine supplies to quickly reach the high coverage required to substantially reduce infections. Another strong argument for that strategy is that many of the current vaccines are highly effective against hospitalization and death, but not in preventing new infections. For instance, AZD1222 (Oxford/AstraZeneca) has shown an efficacy of 86% (95%60 CI: 53-96%) against hospitalization [34], but 59.9% (95% CI: 35.8-75.0%) effective against new infections measured by nucleic acid amplification-positive tests [36]; for CoronaVac (Sinovac) this contrast is even starker with 83.7% (95% CI: 58.0-93.7%) efficacy against severe cases but only 50.7% (95% CI: 35.9-62.0%) for mild symptomatic cases as outcome [24]. Even BNT162b2 (Pfizer/BioNTech), which showed high efficacy65 against infections in the first trials [25] may provide a lower protection against new variants, as exemplified by the variant B.1.617.2 (also known as Delta) [6]. In this context, vaccination programs in many countries have prioritised the groups most at risk of developing severe disease, namely high exposure individuals (e.g. health care professionals), people with aggravating conditions, and older adults [29, 21].70 Mathematical models in epidemiology have been widely used to assess optimal vaccination strategies for a variety of communicable diseases [4, 11, 12, 13, 37]. In a previous study on influenza viruses, Matrajt et al. [19] explored how two different vaccination strategies (single dose or two dose) could be integrated into pandemic control plans of a limited vaccination supply. They conclude that the best strategy75 depends on the level of partial protection introduced by a single dose, but the study is limited to pre-pandemic vaccination scenario and does not capture vaccination roll-out during a pandemic. . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 3 Mass vaccination strategies have been proposed as the main approach to tackle the spread of SARS-CoV-2, complementing or replacing non-pharmaceutical interventions80 (NPIs). SIR-like models have pointed to the possibility of securely relaxing NPIs months after vaccination campaigns, depending on the rate of vaccination, as shown by Kraay et al. [17]. The same work also suggests that, given a low roll-out rate scenario, one- dose strategies allow safe relaxation of NPIs sooner if the first dose corresponds to more than 80% of the vaccination protection (vaccination parameters based on BNT162b2).85 Although one-dose vaccination allows earlier relaxation of NPIs, it requires a slower transition to pre-pandemic levels for more vaccines to be delivered. The one-dose scheme is modelled as an effectively weaker vaccine, with no second doses being considered in the long run. In this work we aim to find the optimal time-window between doses under different90 scenarios of relative efficacy of the first dose and infection rate in the population. We assess these scenarios for the three most used vaccine platforms, namely inactivated virus, viral vector of adenovirus, and mRNA-based vaccines. Given that most vaccination protocols requires a second dose, we introduce a mathematical model that rigorously de- scribes a two-dose vaccination scheme and optimal allocation of vaccines through delay95 differential equations. We explore the consequences of different strategies regarding the time-window between the two vaccine doses. The distribution of available vaccines for the first and second dose is optimized for each between-doses time interval, assuming a constant production rate. The partial protection offered by single-dose vaccination compared to double-dose vaccination is controlled by a single relative efficacy parameter.100 We find that the best strategy depends on the interplay between the vaccine production rate and the single-dose relative efficacy, with the effective reproduction number being an important factor for the reduction of deaths. II. M ETHODS To investigate the impact on total hospitalizations and mortality of delaying the second105 dose in a two-dose vaccination roll-out during an ongoing epidemic, we built a model that takes into account the severity of the disease, age classes (the main risk factor for severity and death), and vaccination status. Our model consists of an SEIR model extended to account for asymptomatic, se- vere/hospitalized, and deceased individuals, thus being named SEAIHRD, correspond-110 ing to susceptible, pre-symptomatic, asymptomatic, mildly symptomatic, severely symp- tomatic/hospitalized, recovered, and deceased classes respectively. Each epidemiological class comprises three age classes, namely children and teenagers (0-19), adults (20-59), and older adults (60+) classes. To account for vaccinated individuals with one or two doses, we have duplicated this set of classes for individuals who received one dose115 and two doses, which have different epidemiological parameters, as discussed below. The model structure is shown in Fig. 1 and the model equations are presented in the supplementary material (SM). . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 4 Fig. 1: Diagram representing the full model structure. Subscripts v and w indicate the first- and second-dose vaccinated classes, respectively. Black arrows indicate transitions between epidemiological stages, green arrows indicate vaccination. All classes pictured inside the gray box are infectious. Because epidemiological progressions happen at time-scales shorter than those related to vaccine effects, infectious classes are not vaccinated in the model. . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 5 T ABLE I: Parameters related to vaccine efficacy. The observed efficacies are obtained from trials or effectiveness studies. The parameters of protection are calculated through the method explained in the appendix A1. P arameter Description CoronaVac AZD1222 BNT162b2 Eβ ,w Observed efficacy against infectious contact given second dose 0.00a 0.599 [36] 0.90 [9] Eα,w Observed efficacy against clinical symptoms given second dose 0.50 [24] 0.813 [36] 0.94 [9] Eσ,w Observed efficacy against hospitalization given second dose 0.83 [24] 0.900b 0.87 [9] Eµ,w Observed efficacy against death given second dose 0.95b 0.950b 0.98b ϵβ ,w Protection against infectious contact with second dose 0.000 0.599 0.900 ϵα,w Protection against clinical symptoms given second dose [0.499, 0.498, 0.494] [0.533, 0.533, 0.537] [0.402, 0.415, 0.494] ϵσ,w Protection against hospitalization of infected individual given second dose 0.830 0.750 -0.300 ϵµ,w Protection against death of hospitalized cases given second dose 0.706 0.500 0.846 a Assumed. There is no data available yet. b Assumed. The value reported had no statistical significance. The demographic characteristics of the population were based on the State of S ˜ao Paulo, Brazil. The basic epidemiological parameters come from current literature and120 are detailed in the SM. The daily contact rates between age classes are based on the matrices projected by Prem et al. [26], from which we use the Brazilian all-locations average for simplicity, reducing the age compartments using the same approach as in [8] (Supplementary material, section 3). We assume a “leaky” vaccination effect, in which vaccinated individuals receives par-125 tial protection, in contrast to the “all-in” model where part of the vaccinated individuals receives full protection. Bubar et al. [7] has shown that the outcome of these models do not differ substantially and the leaky model is simpler to understand and to implement. The main parameters that describe the vaccination dynamics are the vaccine‘s protec- tive effect against (i) acquiring the disease, (ii) developing symptoms, (iii) developing130 severe symptoms (i.e., leading to hospitalization), and (iv) death. We present our analysis using sets of parameters that represent three types of vaccine developed for Sars-Cov-2; inactivated virus (CoronaVac, from Sinovac), adenovirus-based vaccine (AZD1222, from AstraZeneca-Oxford) and mRNA (BNT162b2, from Pfizer-BioNTech). The parameters for each vaccine are shown in Table I.135 As the required parameters may not be available in the studies for each vaccine, specially regarding the efficacy of a single dose, we simplify the problem by assuming that the efficacy for each outcome is proportional to the efficacy after the second dose by a fixed ratio, called relative efficacy of the first dose. This is a key parameter influencing the model outcome: if it is too small it is better to apply the second dose sooner, but140 the opposite should be expected if the efficacy with one dose is as large as with two doses. Thus, we vary this parameter over all reasonable values (0-100%). As the efficacy related to a specific outcome can be caused by a multiplicative effect of efficacy related to another outcome, (for example, lower number of deaths caused by a lower number of hospitalizations), we translate mathematically the observed efficacies145 that come from field studies to the parameters used in the mathematical model, as described in Appendix A1. The values of parameters used in the model are given in Table I. Notice that this mathematical translation makes the protection against clini- cal symptoms age-dependent and also makes the protection against hospitalization for BNT162b2 negative. The latter case occurs because the efficacy against hospitalization is150 . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 6 lower than the efficacy against infectious contact, but when the parameters of protection are multiplied as in the model’s equations (SM) the value of efficacy is recovered. Since we aim to study vaccination roll-out during an ongoing epidemic, we set the initial conditions for recovered and infectious populations using estimated values from hospitalization data. The procedure follows closely the one presented in Coutinho et al.155 [8], and is described in the SM. We use epidemiological and demographic data from the state of S ˜ao Paulo, Brazil, although we do not expect the results to change substantially from one region to another. We do not explicitly implement non-pharmaceutical interventions in our study, opting instead to model different scenarios by fixing the initial value of the effective reproduc-160 tion number Rt. Thus, we computed Rt using the method of Next Generation Matrix (for details, see Appendix B) and chose the value of probability of infection per contact (β) that yields the desired value of Rt at the beginning of the simulation. We vary this initial Rt between 0.9 and 1.4, accounting for potentially more transmissible variants [8, 10] or increased contact rates due to less rigid non-pharmaceutical interventions.165 The distribution of vaccines follows the priority of age groups: first older adults, then adults, and at last children and teenagers [35]. We assume that serological tests are not performed before vaccination, therefore vaccination does not depend on the individual’s previous serological status. Thus, susceptible and recovered individuals are vaccinated proportionally to the compartment’s population size. Severe and symptomatic170 individuals do not receive doses while infectious due to noticeable infection. We also exclude pre-symptomatic and asymptomatic from vaccination, since it would lead to unrealistic immediate benefits and there is a window between the date of application and the start of protective effects. Recovered individuals are assumed to be completely immune and, since we fix our simulation interval at 300 days, we consider that the175 effects of waning immunity can be negligible. Likewise, we ignore aging dynamics, since it would be relevant only at much longer time scales. Our model specifies that the time between the applications of the first and second doses (a) is fixed for all individuals, which leads to a delay differential equation formulation. In contrast, constant rate models, where individuals receive the second180 dose at a fixed rate with an exponential distribution of times, would imply that some of them immediately receive the second dose and lead to incorrect average times between doses when vaccination rates fluctuate over time. We also allow for a proportion θ of individuals that take the first but not the second dose, representing abandonment; this value is fixed at0.1 throughout our simulations (In the case of CoronaVac vaccine phase185 3 trial, the rate of abandonment was 0.16 [2]). To optimally allocate the vaccination, keeping the number of stocked vaccines as low as possible while ensuring that second doses are available, we solve a delayed optimiza- tion problem. We assume an initial stock of vaccines V0 and a constant production (or deployment) rate p and solve for the control variable that is the vaccination rate v(t),190 with an upper limit of vmax. The full mathematical description of the problem and its solution are described in Appendix C. . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 7 To assess which conditions is best to employ a longer (up to 12 weeks) or shorter (at least 3 weeks) interval between first and second doses, we simulate the model and compute the total number of deaths across all ages and vaccination status at the end195 of a period of 300 days, choosing as the optimal window the one that has the lowest number of deaths. We performed this comparison by varying the relative efficacy of the first dose and also the parameters related to the current rate of infection (R t) and vaccine production rate. III. R ESULTS200 The optimal vaccine roll-out with time between doses of 3 and 12 weeks is shown in Fig. 2. We can see that for reasonable parameters of production and vaccine application rate for S ˜ao Paulo State, the vaccine stock can be kept close to zero (i.e. all doses can be used immediately) after about 30 days after the beginning of dose application when using a time interval of 3 weeks. For the time interval of 12 weeks, the stock is kept205 close to zero during the first stage of the first dose vaccination, but doses need to be stored for the application of second-doses for a small time window. In both cases, we see a pattern of alternating stages of first and second dose vaccination, while optimally keeping the stock of vaccines close to zero. 3 Weeks 12 Weeks 0 100 200 300 0 100 200 300 0.0% 0.1% 0.2% 0.3% 0.4% 0.0% 0.5% 1.0% 1.5% 2.0% t (days) Vaccination rate (doses/pop−day) Vaccine doses stored (doses/pop) Fig. 2: Vaccination rate as a function of time for first (solid, black) and second doses (dashed, grey) – scale is given by the left y-axis. Number of stored vaccine doses as function of time (dot-dashed, blue), scale in the right y-axis. Taking into account 3 or 12 weeks (panels) as separation between doses. V0 = 2.83% of population in doses, ρ(t) = 0. 23% of population in doses/day, vmax = 0.45% of population in doses/day. By varying the initial values of Rt for a fixed production rate (Figure 3), we see that210 longer periods between the first and second doses led to a lower number of deaths when the relative efficacy of the first dose alone is higher than approximately 60%, whereas smaller time intervals are better if the relative efficacy is lower than approximately 50%, and intermediary periods are the optimal interval between doses if the relative efficacy is between 50% and 60%. It is also important to note that the value of initial Rt does215 not substantially affect the value of relative efficacy where transitions between optimal windows occur, but is an important factor in the absolute value of deaths. . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 8 CoronaVac AZD1222 BNT162b2 Rt: 0.9Rt: 1Rt: 1.1Rt: 1.2Rt: 1.3Rt: 1.4 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 10.0% 20.0% 30.0% 40.0% 50.0% 10.0% 20.0% 30.0% 40.0% 50.0% 10.0% 20.0% 30.0% 40.0% 50.0% 10.0% 20.0% 30.0% 40.0% 50.0% 10.0% 20.0% 30.0% 40.0% 50.0% 10.0% 20.0% 30.0% 40.0% 50.0% Relative Efficacy of First Dose Reduction in Total Number of Deaths Weeks 3 7 12 Fig. 3: Reduction in total number of deaths as function of the first dose relative efficacy, considering three time windows between doses: 3,7 and 12 weeks (colors); varying vaccine type (columns); and effective reproduction number at the start of simulation (rows). V0 = 2.83% of population in doses, ρ(t) = 0. 23% of population in doses/day,vmax = 0.45% of population in doses/day. Since our results show that the optimal time window is not strongly dependant on the value of Rt, we fixed this value at 1.1 and varied the relative efficacy and production . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 9 rate. Then, we compare the optimal window relative to the reduction in deaths for several220 time-window intervals (Figure 4). The result is computed by simulating a coarse grid of values and then linearly interpolating a finer mesh. For all vaccines, relative efficacy of the first dose below approximately 45% indicates that regardless of the vaccine production rate, the best strategy to reduce mortality is to complete the two-dose scheme 3 weeks after the initial dose. If the production rate is low (approximately below 0.2%225 of doses/population-day), increasing the relative efficacy above 45% rapidly shifts the optimal time-window to increasingly larger periods, and converges to the maximum interval of 12 weeks. However, on this same range for the relative efficacy of the first dose, increasing the production rates results in a non-linear transition to shorter time- windows. CoronaVac AZD1222 BNT162b2 0.1% 0.2% 0.3% 0.4% 0.1% 0.2% 0.3% 0.4% 0.1% 0.2% 0.3% 0.4% 0.00 0.25 0.50 0.75 1.00 Production Rate (doses/population−day) Relative Efficacy of First Dose Weeks 3 4 5 6 7 8 9 10 11 12 Fig. 4: Best window (color) for reduction of deaths as function of production rate (x axis) and relative efficacy of first dose (y axis). Each panel represents a different vaccine with the respective second dose parameters. The initial Rt is 1.1, V0 is set to zero, vaccination rate limited to approximately 0.45% of population in doses per day. 230 Although it is expected that the optimal time window should always increase with higher relative efficacy, and decrease with higher production rate, the latter hypothesis is sometimes violated in the simulations (Fig. 4). Despite that, the general pattern is largely unaltered. IV. D ISCUSSION235 Our model evaluates the optimal time windows between first and second doses for individuals in a context of limited vaccine supply in the ongoing COVID-19 pandemic. We found that the time window that best reduces the mortality depends on the interplay between two key parameters: vaccine production rate and the relative efficacy of the first dose. Within the simulated regimes with low single-dose efficacy, the resulting best240 strategy consistently relied on allocating the available doses to complete the two-dose scheme at the shortest time possible to provide the vaccine’s maximum protection and minimize the mortality of the population. However, when the first dose presents a higher level of relative efficacy and vaccine supply is restricted, the best strategy relied on prioritizing the allocation of doses to increase the proportion of the population receiving245 . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 10 the first dose, resulting in a longer interval between doses to effectively allocate the use of the limited vaccine supply. Further, this interval is shortened by increasing the vaccine supply, as it progressively alleviates the conflict on dose allocation. Given that the effective reproduction number has little effect when considering the optimal window between doses, this further ensures the applicability of our results to other locations.250 Another important parameter that varies across vaccines is the protection against infection, i.e. reduction in susceptibility, after the second dose, with values of 0, 0.6 and 0.9 for CoronaVac, AZD1222 and BNT162b2 respectively (See table I). This difference allows for larger delays with lower relative efficacy of the first dose, even with similar protection against deaths (0.95, 0.95, 0.98, in the same order) simply by avoiding new255 infections. Given the key role of the efficacy of the first dose in deciding the best strategy to reduce mortality, it is worth considering the best available knowledge for the vaccines evaluated in this study. Preliminary results of the effectiveness of the inactivated virus vaccine CoronaVac in Chile have shown that the first-dose efficacy of that vaccine is260 substantially lower before the application of the second dose, with effectiveness against symptomatic infection being 17.2% (95% CI: 15.8–18.6%) [14]. This result, together with previous studies concerning inactivated virus vaccines showing that the protection against infectious contact is likely to be low [1, 32], suggests that the safest strategy concerning second dose application of CoronaVac is to apply it in three weeks, the265 shortest recommended interval, to achieve the largest reduction in the number of deaths. In the case of the adenovirus-based AZD1222 vaccine, V oysey et al. [36] estimated 76.7% (95% CI: 47.0-89.8%) for the efficacy against symptoms 21 days after the first dose. Albeit this estimate provides reasonably high values, the small number of events

in great uncertainty on this estimate, reflected in its wide confidence intervals.270 Pritchard et al. [27] corroborates those results with effectiveness against having a positive RT-PCR test of 64% after 21 days (95% CI: 59–68%) in a study in the UK, but without differentiating between AZD1222 and BNT162b2. Using this value as a proxy of first dose efficacy compared to the efficacy against symptoms post second-dose, the relative efficacy would be around 80%, which, according to our results, would suggest that275 delaying the second dose can be an effective strategy, especially in lower production rates. The efficacy post second-dose using a window less than 6 weeks is substantially lower (55.1%, 95% CI: 33.0-69.9%) when compared to 12 weeks (81.3%, 95% CI: 60.3-91.2%) [36], providing another argument in support of delaying the second dose of AZD1222.280 For the mRNA vaccine BNT162b2, Pritchard et al. [27] have shown that its effective- ness is very high 21 days after the first dose (78%, 95% CI: 72-83%), results corrob- orated by a study conducted by the Joint Committee on Vaccination and Immunisation [16], from the UK. With these values, together with very high post second-dose efficacy against infectious contacts, our model results entail that postponing the application of285 the second dose is the best strategy. However, as this is the first known vaccine using the mRNA platform, it would be advisable to postpone the second dose only after studies . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 11 assuring the efficacy of such vaccine in these conditions, as pointed out by Robertson, Sewell, and Stewart [30]. Previous studies that have evaluated the best strategies for vaccination for COVID-19290 have focused on the interplay between vaccination and non-pharmaceutical interventions [7, 15, 23]. More recently, some agent-based models estimated the reduction in deaths and infections if postponing the application of the second dose: Moghadas et al. [22] found that delaying the second dose for BNT162b2 and mRNA-1273 (from Moderna) is usually the best strategy until it reaches peak values for longer intervals; Romero-Brufau295 et al. [31] found that the results are even better in the scenarios of low vaccine production rates. We confirm these initial results and extend the analysis to more combinations of first dose efficacy and production rates, and also consider more types of vaccine. Our model also accounts for optimal allocation of vaccination rates of first and second doses, an issue not addressed in those works.300 Using a differential equations formulation, Mak, Dai, and Tang [18] compared the strategies of “hold back” (storing half of the vaccines to guarantee the second dose) and “release” (applying as much as first doses possible, while minimizing the backlog of second dose vaccination) and found that first dose efficacy is essential when deciding optimal vaccine protection coverage if delaying the second dose. Their results agree305 with ours but it has the limitation of not guaranteeing that the mean period with single- dose protection is preserved in a varying vaccination rate, which is only possible using a more complex formulation, such as delay differential equations as used here. They also argue that, while the “release” strategy enables better allocation of vaccines, in a real-world situation the implementation of such strategy is too complicated to manage.310 While we agree that this strategy adds a complication factor, we also believe that stock optimization models such as the one developed in this study could be linked to vaccination databases and, as the vaccination roll-outs to tackle the SARS-CoV-2 pandemic were centralized by the national health systems, thus allowing more efficient allocation of vaccines.315 Throughout this study, we assumed that post second-dose efficacies do not depend on the time-window between doses, although evidence suggests that this is not the case for AZD1222 vaccine [36]. This simplifying assumption probably leads to a slight underestimation of the benefits of increasing the lag between doses for this vaccine. We have also assumed that first dose efficacy is constant over time, when in reality320 it builds up over approximately 2 weeks and wanes later on [27, 33]. This can be addressed using agent-based models [22], or age-of-infection models in a differential equations formulation, but that would greatly increase the complexity of the algorithm. This difficulty is also present when trying to compare model parameters to first dose efficacy in observational studies, as only the period after 14 days is usually considered,325 leaving a very short period in which events can occur. The impact of this on our

can go in either direction: for shorter time windows, the efficacy in the model should be lower than the observed one, since for most of the time – the first 14 days – protection is virtually non-existent; while for longer time windows the same . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 12 happens, that is, model parameters should be below measured efficacies due to waning330 effects of the vaccine. We currently do not have enough details of the profile of the immunological response over time to settle the issue definitively, yet we expect the pattern of the results to hold, given that we cover every possibility of relative efficacy and thus this can be compared to averaged values in time of the efficacy of the first dose.335 In summary, we developed a novel approach to assess the optimal conditions to delay the application of the second dose of three different vaccine platforms for COVID-19. Our approach consisted of using a SEIR-like model coupled with a delay differential equation vaccination model and optimization of vaccine stock. We found that the first dose efficacy relative to post second dose efficacy is an essential parameter when defin-340 ing the optimal window between doses, but the vaccine production rate in each country can be also decisive. Finally, we also found that those results have little dependence on the current epidemic situation at the start of the simulation. Further research in this theme would include non-constant production rate in the optimization scheme, together with starting such model in an ongoing vaccination roll-out, as well as considering345 lower efficacy in different age bins and other dose prioritization schemes.

We are grateful for the collaborative work of the entire group of the Observat ´orio COVID-19 BR. In particular, we thank Monica de Bolle, Ver ˆonica Coelho and M ´arcia Castro for critical inputs. This work was supported by the Coordenac ¸˜ao de Aperfeic ¸oamento350 de Pessoal de N ´ıvel Superior, Brazil (Finance Code 001 to LSF and FMDM), Con- selho Nacional de Desenvolvimento Cient ´ıfico e Tecnol ´ogico, Brazil (grant number: 315854/2020-0 to MEB, 141698/2018-7 to RLPS, 312559/2020-8 to MASMV and 311832/2017-2 to RAK), Fundac ¸˜ao de Amparo `a Pesquisa do Estado de S ˜ao Paulo, Brazil (grant number: 2019/26310-2 and 2017/26770-8 to CF, 2018/26512-1 to OC,355 2018/24037-4 to SPL and contract number: 2016/01343-7 to RAK) and Swiss National Science Foundation (grant PCEFP3 181243 to VS). Declaration of Competing Interests The authors declare that they have no competing interests. Data Availability360 The algorithms used to produce the results presented here are available at https: //github.com/covid19br/VaxModel-paper. APPENDIX A. Parameterization of the model The parameters of the model are as follows: β is the probability of infection given365 infectious contact. α is the proportion of asymptomatic individuals. σ is the proportion . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 13 of severe/hospitalized cases amongst infected individuals. µ is the proportion of hospi- talized individuals that die. Such parameters are age-stratified unless otherwise stated and their values are given in the supplementary material. Parameters related to vaccine efficacy are given in table I.370 1) Efficacy parameters computation from observed efficacies: The vaccinated classes parameters are combined with vaccine efficacies as: βv = (1−ϵβ,v)β αv = 1− (1−ϵα,v)(1−α) βw = (1−ϵβ,w)β αw = 1− (1−ϵα,w)(1−α) (1) σv = (1−ϵσ,v)σ µv = (1−ϵµ,v)µ σw = (1−ϵσ,w)σ µw = (1−ϵµ,w)µ To avoid multiplicative effects in vaccine efficacies, we need to calculate the efficacy parameters from the reported values. Let us start with the risk of infection. In our model, this is given by β. Thus the observed efficacy against infection Eβ is given by:375 Eβ = 1− (1−ϵβ)β β =ϵβ (2) Therefore the protection against infection parameter is simply the observed efficacy. Note that we dropped the dose index as these expressions are valid for both first and second dose efficacies. The risk of individuals being hospitalized is given by βσ, therefore, the observed efficacy in reducing hospitalized cases Eσ is then given by: Eσ = 1− (1−ϵβ)β(1−ϵσ)σ βσ = 1− (1−ϵβ)(1−ϵσ) (3) In terms of known values, the protection against hospitalization is given by: ϵσ = 1− 1−Eσ 1−Eβ (4) Being µ the proportion of hospitalized individuals that die, we have that the risk of an individual being infected and die is given by βσµ, therefore the observed efficacy against death Eµ is given by: Eµ = 1− (1−ϵβ)β(1−ϵσ)σ(1−ϵµ)µ βσµ = 1− (1−ϵβ)(1−ϵσ)(1−ϵµ) (5) We then can obtain εµ in terms of known values: εµ = 1− 1−Eµ (1−ϵβ)(1−ϵσ) = 1− 1−Eµ (1−Eβ) (1−Eσ) (1−Eβ) = 1− 1−Eµ 1−Eσ (6) . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 14 In our model, symptomatic cases are given by severe (hospitalized) and mild cases, the risk of becoming a symptomatic individual is given by β[σ + (1−σ)(1−α)], then the observed efficacyEsymp is given by: Esymp = 1− (1−ϵβ)β{(1−εσ)σ + [1− (1−εσ)σ][(1−εα)(1−α)]} β[σ + (1−σ)(1−α)] (7) Thus (1−Esymp)[σ + (1−σ)(1−α)] 1−ϵβ = (1−ϵσ)σ + [1− (1−ϵσ)σ][(1−ϵα)(1−α)] (8) Then 1−ϵα = (1−Esymp) (1−ϵβ) [σ + (1−σ)(1−α)]− (1−ϵσ)σ [1− (1−ϵσ)σ](1−α) (9) Therefore, εα is given in terms of known variables as: εα = 1− (1−Esymp)[σ + (1−σ)(1−α)]− (1−Eσ)σ (1−Eβ) [ 1− (1−Eσ) (1−Eβ)σ ] (1−α) (10) Note that 1−Esymp does not multiply the whole expression. B. Effective reproduction number and initial conditions estimation380 Both initial conditions estimation and effective reproduction number calculations go through rewriting the model in a different notation. It is a system of equations for two different groups, infected (y ) and non-infected (z) populations, being y = (E, A, I, H)T , and z = (S, R, D)T . Note that none of the vaccinated classes are considered since at the initial condition no vaccine has been applied yet.385 We write the system ˙y = F (y, z)−G(y, z), (11) ˙z = J(y, z) , (12) where F are all entries of new Infected, coming from classes z, whilst G accounts for the transitions within infected classes and also recovery and death from the disease. J accounts for the exits of the susceptible population to exposed classes, and the entrance of recovered and deceased in their respective compartments. Consider a linearization390 around a fixed vector z = ˜ z, the equation for y becomes ˙y = ( ˆF− ˆG)y , (13) . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 15 where ˆF and ˆG are matrices that appear from linearizing functions F and G, re- spectively. Remembering that each of the compartments is divided into three age sub- compartments, that is S = (Syoung,Sadult,Selderly) and that the only entrance of new infected comes from the βSλ/N terms in the ˙E equations, we write395 ˆF = β N   ωˆb ˆb (1−ξ)ˆb (1−ξsev)ˆb 09,12   , (14) where ˆb = diag(S) ˆC , (15) being ˆC the contact matrices, from [26], ω the relative infectiousness of exposed individuals and ξ and ξsev the reductions in contacts of people that are symptomatic and hospitalized, respectively. Now, ˆG contains the terms of Exposed, E, developing the possible forms of disease considered in the model as the terms in its first 3 rows, while it’s main diagonal contains400 terms of recovery and death, writing ˆG =   γ−1 0 0 0 −α(1−σ)γ−1 ν−1 i 0 0 −(1−α)(1−σ)γ−1 0 ν−1 i 0 −σγ−1 0 0 ν−1 s  . (16) Now,F andG are important for both Rt calculation and initial conditions estimation. Note that for the linear problem, assuming, y(t) =∑ i aierit, ai constant vectors, yields ∑ i riaierit = ( ˆF− ˆG) ∑ i aierit. (17) Definingr∗ = maxri, a∗ the vector associated with the exponential coefficient r∗, and dividing the above equation by er∗t, we get405 ∑ i riaie(ri−r∗)t = ( ˆF− ˆG) ∑ i aie(ri−r∗)t , (18) so after some time t elapses, the tuple (a∗,r∗) dominate the dynamics, and we’re left with r∗a∗ = ( ˆF− ˆG)a∗ (19) The main eigenvector of ˆF− ˆG, a∗, gives a distribution of infected individuals among different classes. With hospitalizations per day data, we can fit a re-scaling factor for the eigenvector to match the term of hospital entrances (σγ −1E).410 Notably, the effective reproduction calculation can be performed with ˆF and ˆG as . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 16 Rt =ρ(FG−1) , (20) where ρ(FG−1) is the spectral radius of FG−1, which may be seen as the dominant eigenvalue of FG−1 in the simplest cases. The derivation of said result can be checked in multiple textbooks, see, for instance, chapter 6 in Allen et al. [3]. C. Optimizing vaccination roll-out415 We write a dynamical equation for the vaccine stock V (t) assuming a constant production rate p and a varying withdrawal rate is given by the vaccination rate, which can be chosen, that is, it is the control variable. We impose that a constant fraction θ′ = 1−θ of the people who take the first dose will receive the second one after a period a, so we must be careful that the total vaccination rate is the sum of both first420 and second dose vaccination rates, but the control variable v(t) is the vaccination rate of first doses only. We also assume that there’s an initial stock of vaccines V0. The equation for V (t) then is:    dV dt =p−v(t)−θ′v(t−a) V (0) = V0 , v (t) = 0∀t< 0 (21) We note already that we can solve this equation, obtaining V (t) =V0 +pt− ∫ t 0 v(t′)dt′−θ′ ∫ t−a 0 v(t′)dt′ (22) We define the optimization problem by stating the objective function to be minimized425 and the restrictions that the solution must obey. Since we want to use vaccine doses as quickly as possible, a reasonable goal is to minimize the stock of vaccines V (t). With that, we impose that the total vaccination rate is limited by a certain maximum value, and of course, it is positive; also, the vaccine stock V (t) is always positive. Finally, we must ensure that in the period after the simulation ends (t > T ) there will be enough430 doses left to apply the second doses on those who have already taken the first dose. These considerations lead to the following optimization problem: find min f J = ∫ T 0 V (t)dt subject to v(t)≥ 0 v(t) +θ′v(t−a)≤vmax V (t)≥ 0 V (T )≥θ′ ∫ T T−a v(t′)dt′−pa (23) . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 17 Solution: We can solve the problem defined by Eqs. (22, 23) using linear program- ming. This is feasible because the objective function and all constraints are linear functions of the control variable f and state variable V and, as Eq.(22) shows, V435 is linear on the control variable. We first discretize the time in n intervals of length ∆t = T n , each interval ending at ti, i = 1,...,n, and assume that the control function will be constant over each interval (that is, a step function), with values ⃗ x= (v(t1),v (t2),...,v (tn)). Eq.(22) then becomes V (ti) =V0 +pti− i∑ j=1 xj−θ′ i−ˆa∑ j=1 xj , (24) and we seek to minimize the objective function (given by Eq.(23), up to a constant) that is a linear function of ⃗ x, subject to the (linear) restrictions. The discrete version of the problem becomes: min ⃗ x J = min ⃗ x − n∑ j=1 [ j∑ i=1 xi +θ′ j−ˆa∑ i=1 xi ] subject to ⃗ x≥ 0 xi +θ′xi−ˆa≤vmax , for i = 1,...,n j∑ i=1 xi +θ′ j−ˆa∑ i=1 xi≤V0 +pti , for j = 1,...,n (1 +θ′) n∑ i=1 xi≤V0 +p(tn +a) , (25) where ˆa = a ∆t (chosen so that ˆa is integer), and, to simplify notation, xi is taken to be440 zero over values of i below 1. These conditions can readily be written in matrix form and solved using standard lin- ear programming algorithms. We implemented them in R using the package lpSolve [5] to solve the linear programming problem. REFERENCES445 [1] B. Afrough, S. Dowall, and R. Hewson. “Emerging viruses and current strategies for vaccine intervention”. In: Clinical & Experimental Immunology 196.2 (Apr. 2019), pp. 157–166. DOI: 10.1111/cei.13295. URL: https://doi.org/10.1111/cei. 13295. [2] Ag ˆencia Nacional de Vigil ˆancia Sanit ´aria. Avaliac ¸˜ao do Uso Emergencial da450 Vacina CoronaVac - Apresentac ¸˜ao. 2021. URL: https://www.gov.br/anvisa/pt- br/assuntos/noticias-anvisa/2021/confira-materiais-da-reuniao-extraordinaria-da- dicol/1-apresentacao-ggmed-coronavac.pdf (visited on 07/21/2021). [3] Linda JS Allen et al. Mathematical epidemiology. V ol. 1945. Springer, 2008. . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 18 [4] Julien Arino et al. “A model for influenza with vaccination and antiviral treat-455 ment”. In: Journal of Theoretical Biology 253.1 (July 2008), pp. 118–130. DOI: 10.1016/j.jtbi.2008.02.026. URL: https://doi.org/10.1016/j.jtbi.2008.02.026. [5] Michel Berkelaar et al. lpSolve: Interface to ’Lp solve’ v. 5.5 to Solve Lin- ear/Integer Programs. R package version 5.6.15. 2020. URL: https://CRAN.R- project.org/package=lpSolve.460 [6] Jamie Lopez Bernal et al. “Effectiveness of Covid-19 Vaccines against the B.1.617.2 (Delta) Variant”. In:New England Journal of Medicine (July 2021). DOI: 10.1056/ nejmoa2108891. URL: https://doi.org/10.1056/nejmoa2108891. [7] Kate M. Bubar et al. “Model-informed COVID-19 vaccine prioritization strategies by age and serostatus”. In: Science 371.6532 (Jan. 2021), pp. 916–921. DOI:465 10.1126/science.abe6959. URL: https://doi.org/10.1126/science.abe6959. [8] Renato Mendes Coutinho et al. “Model-based estimation of transmissibility and reinfection of SARS-CoV-2 P.1 variant”. In: MedRxiv (Mar. 2021). DOI: 10.1101/ 2021.03.03.21252706. URL: https://doi.org/10.1101/2021.03.03.21252706. [9] Noa Dagan et al. “BNT162b2 mRNA Covid-19 Vaccine in a Nationwide Mass470 Vaccination Setting”. In: New England Journal of Medicine 384.15 (Apr. 2021), pp. 1412–1423. DOI: 10.1056/nejmoa2101765. URL: https://doi.org/10.1056/ nejmoa2101765. [10] Nicholas G. Davies et al. “Estimated transmissibility and impact of SARS-CoV-2 lineage B.1.1.7 in England”. In: Science 372.6538 (Mar. 2021), eabg3055. DOI:475 10.1126/science.abg3055. URL: https://doi.org/10.1126/science.abg3055. [11] Matthias Ehrhardt, J ´an Ga ˇsper, and So ˇna Kilianov ´a. “SIR-based mathematical modeling of infectious diseases with vaccination and waning immunity”. In: Journal of Computational Science 37 (Oct. 2019), p. 101027. DOI: 10 . 1016 / j.jocs.2019.101027. URL: https://doi.org/10.1016/j.jocs.2019.101027.480 [12] Elamin H. Elbasha and Alison P. Galvani. “Vaccination against multiple HPV types”. In: Mathematical Biosciences 197.1 (Sept. 2005), pp. 88–117. DOI: 10. 1016/j.mbs.2005.05.004. URL: https://doi.org/10.1016/j.mbs.2005.05.004. [13] David Greenhalgh. “Vaccination campaigns for common childhood diseases”. In: Mathematical Biosciences 100.2 (July 1990), pp. 201–240. DOI: 10.1016/0025-485 5564(90)90040-6. URL: https://doi.org/10.1016/0025-5564(90)90040-6. [14] Alejandro Jara et al. “Effectiveness of an Inactivated SARS-CoV-2 Vaccine in Chile”. In: New England Journal of Medicine(July 2021). DOI: 10.1056/nejmoa2107715. URL: https://doi.org/10.1056/nejmoa2107715. [15] Peter C Jentsch, Madhur Anand, and Chris T Bauch. “Prioritising COVID-19490 vaccination in changing social and epidemiological landscapes: a mathematical modelling study”. In: The Lancet Infectious Diseases (Mar. 2021). DOI: 10.1016/ s1473-3099(21)00057-8. URL: https://doi.org/10.1016/s1473-3099(21)00057-8. [16] Joint Committee on Vaccination and Immunisation. Optimising the COVID-19 vaccination programme for maximum short-term impact. 2021. URL: https : / /495 www.gov.uk/government/publications/prioritising- the- first- covid- 19- vaccine- . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 19 dose - jcvi - statement / optimising - the - covid - 19 - vaccination - programme - for - maximum-short-term-impact (visited on 06/18/2021). [17] Alicia N.M. Kraay et al. “Modeling the use of SARS-CoV-2 vaccination to safely relax non-pharmaceutical interventions”. In: (Mar. 2021). DOI: 10.1101/2021.03.500 12.21253481. URL: https://doi.org/10.1101/2021.03.12.21253481. [18] Ho-Yin Mak, Tinglong Dai, and Christopher S. Tang. “Managing Two-Dose COVID-19 Vaccine Rollouts with Limited Supply”. In: SSRN Electronic Journal (2021). DOI: 10.2139/ssrn.3790836. URL: https://doi.org/10.2139/ssrn.3790836. [19] Laura Matrajt et al. “One versus two doses: What is the best use of vaccine in an505 influenza pandemic?” In: Epidemics 13 (Dec. 2015), pp. 17–27. DOI: 10.1016/j. epidem.2015.06.001. URL: https://doi.org/10.1016/j.epidem.2015.06.001. [20] Minist ´erio da Sa´ude. D ´ECIMO QUINTO INFORME T ´ECNICO - ORIENTAC ¸˜OES T ´ECNICAS RELATIVAS `A CONTINUIDADE DA CAMPANHA NACIONAL DE VACINAC ¸˜AO CONTRA A COVID-19. 2021. URL: https://www.gov.br/saude/pt-510 br/media/pdf/2021/maio/3/anexo-decimo-quinto-informe-tecnico.pdf (visited on 07/23/2021). [21] Minist ´erio da Sa ´ude. Plano Nacional de Operacionalizac ¸˜ao da Vacinac ¸˜ao contra Covid-19. 2021. URL: https://www.gov.br/saude/pt-br/coronavirus/publicacoes- tecnicas/guias-e-planos/plano-nacional-de-vacinacao-covid-19/view (visited on515 07/26/2021). [22] Seyed M. Moghadas et al. “Evaluation of COVID-19 vaccination strategies with a delayed second dose”. In: PLOS Biology 19.4 (Apr. 2021). Ed. by Andrew Fraser Read, e3001211. DOI: 10.1371/journal.pbio.3001211. URL: https://doi.org/10. 1371/journal.pbio.3001211.520 [23] Sam Moore et al. “Vaccination and non-pharmaceutical interventions for COVID- 19: a mathematical modelling study”. In: The Lancet Infectious Diseases 21.6 (June 2021), pp. 793–802. DOI: 10.1016/s1473-3099(21)00143-2. URL: https: //doi.org/10.1016/s1473-3099(21)00143-2. [24] Ricardo Palacios et al. “Efficacy and Safety of a COVID-19 Inactivated Vaccine in525 Healthcare Professionals in Brazil: The PROFISCOV Study”. In: SSRN Electronic Journal (2021). DOI: 10.2139/ssrn.3822780. URL: https://doi.org/10.2139/ssrn. 3822780. [25] Fernando P. Polack et al. “Safety and Efficacy of the BNT162b2 mRNA Covid- 19 Vaccine”. In:New England Journal of Medicine 383.27 (Dec. 2020), pp. 2603–530 2615. DOI: 10.1056/nejmoa2034577. URL: https://doi.org/10.1056/nejmoa2034577. [26] Kiesha Prem et al. “Projecting contact matrices in 177 geographical regions: An update and comparison with empirical data for the COVID-19 era”. In: PLOS Computational Biology 17.7 (July 2021). Ed. by Matthew (Matt) Ferrari, e1009098. DOI: 10.1371/journal.pcbi.1009098. URL: https://doi.org/10.1371/535 journal.pcbi.1009098. . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 20 [27] Emma Pritchard et al. “Impact of vaccination on new SARS-CoV-2 infections in the United Kingdom”. In: Nature Medicine (June 2021). DOI: 10.1038/s41591- 021-01410-w. URL: https://doi.org/10.1038/s41591-021-01410-w. [28] Public-Health-England. COVID-19 vaccination programme - Information for health-540 care practitioners. 2021. URL: https://assets.publishing.service.gov.uk/government/ uploads/system/uploads/attachment data/file/999527/COVID- 19 vaccination programme guidance for healthcare workers 6July2021 v3 . 9 . pdf (visited on 07/23/2021). [29] Robert Koch Institut. Epidemiologisches Bulletin 2/2021. 2021. URL: https : / /545 www.rki.de/DE/Content/Infekt/EpidBull/Archiv/2021/Ausgaben/02 21.html (visited on 07/26/2021). [30] John F R Robertson, Herb F Sewell, and Marcia Stewart. “Delayed second dose of the BNT162b2 vaccine: innovation or misguided conjecture?” In: The Lancet 397.10277 (Mar. 2021), pp. 879–880. DOI: 10.1016/s0140- 6736(21)00455- 4.550 URL: https://doi.org/10.1016/s0140-6736(21)00455-4. [31] Santiago Romero-Brufau et al. “Public health impact of delaying second dose of BNT162b2 or mRNA-1273 covid-19 vaccine: simulation agent based modeling study”. In: BMJ (May 2021), n1087. DOI: 10 . 1136 / bmj . n1087.URL: https : //doi.org/10.1136/bmj.n1087.555 [32] Rachel L Roper and Kristina E Rehm. “SARS vaccines: where are we?” In: Expert Review of Vaccines 8.7 (July 2009), pp. 887–898. DOI: 10.1586/erv.09.43. URL: https://doi.org/10.1586/erv.09.43. [33] Madhumita Shrotri et al. “Vaccine effectiveness of the first dose of ChAdOx1 nCoV-19 and BNT162b2 against SARS-CoV-2 infection in residents of long-560 term care facilities in England (VIV ALDI): a prospective cohort study”. In: The Lancet Infectious Diseases (June 2021). DOI: 10.1016/s1473-3099(21)00289-9. URL: https://doi.org/10.1016/s1473-3099(21)00289-9. [34] Julia Stowe et al. Effectiveness of COVID-19 vaccines against hospital admission with the Delta (B.1.617.2) variant. 2021. URL: https://media.tghn.org/articles/565 Effectiveness of COVID-19 vaccines against hospital admission with the Delta B. G6gnnqJ.pdf (visited on 07/26/2021). [35] Secretaria de Vigil ˆancia em Sa´ude. PLANO NACIONAL DE OPERACIONALIZAC ¸˜AO DA VACINAC ¸˜AO CONTRA A COVID-19. 2021. URL: https://www.gov.br/saude/ pt-br/media/pdf/2021/marco/23/plano-nacional-de-vacinacao-covid-19-de-2021570 (visited on 06/03/2021). [36] Merryn V oysey et al. “Single-dose administration and the influence of the timing of the booster dose on immunogenicity and efficacy of ChAdOx1 nCoV-19 (AZD1222) vaccine: a pooled analysis of four randomised trials”. In: The Lancet 397.10277 (Mar. 2021), pp. 881–891. DOI: 10.1016/s0140- 6736(21)00432- 3.575 URL: https://doi.org/10.1016/s0140-6736(21)00432-3. [37] Shoujun Zhao, Zhiyi Xu, and Ying Lu. “A mathematical model of hepatitis B virus transmission and its application for vaccination strategy in China”. In: . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint 21 International Journal of Epidemiology 29.4 (Aug. 2000), pp. 744–752. DOI: 10.1093/ije/29.4.744. URL: https://doi.org/10.1093/ije/29.4.744.580 . CC-BY-ND 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted July 31, 2021. ; https://doi.org/10.1101/2021.07.28.21261200doi: medRxiv preprint