## Abstract

When a vaccine for COVID-19 becomes available, limited initial supply will raise the question of how to prioritize the available doses and thus underscores the need for transparent, evidence-based strategies that relate knowledge of, and uncertainty in, disease transmission, risk, vaccine efficacy, and existing population immunity. Here, we employ a model-informed approach to vaccine prioritization that evaluates the impact of prioritization strategies on cumulative incidence and mortality and accounts for population factors such as age, contact structure, and seroprevalence, and vaccine factors including imperfect and age-varying efficacy. This framework can be used to evaluate and compare existing strategies, and it can also be used to derive an optimal prioritization strategy to minimize mortality or incidence. We find that a transmission-blocking vaccine should be prioritized to adults ages 20-49y to minimize cumulative incidence and to adults over 60y to minimize mortality. Direct vaccination of adults over 60y minimizes mortality for vaccines that do not block transmission. We also estimate the potential benefit of using individual-level serological tests to redirect doses to only seronegative individuals, improving the marginal impact of each dose. We argue that this serology-informed vaccination approach may improve the efficiency of vaccination efforts while partially addressing existing inequities in COVID-19 burden and impact.

## Introduction

SARS-CoV-2 has caused a public health and economic crisis worldwide. As of August 2020, there have been over 25 million cases and 820,000 deaths reported *(1)*. To combat this crisis, a variety of non-pharmaceutical interventions have been implemented, including shelter-in-place orders, limited travel, and remote schooling. While these efforts are essential to slowing transmission in the short-term, long-term solutions—such as vaccines that confer immunity to SARS-CoV-2 infection—are urgently needed. As the benefits of an effective vaccine for individuals and their communities may result in high and widespread demand, it is critical that decision-making on vaccine distribution is transparent, particularly in the initial phases when vaccine availability may be limited *(2)*.

Here, we employ a model-informed approach to quantify the impact of COVID-19 vaccine prioritization strategies on cumulative incidence and mortality. Our approach explicitly addresses variation in three areas that can influence the outcome of vaccine distribution decisions. First, we consider variation in the performance of the vaccine, including its overall efficacy, a hypothetical decrease in efficacy by age, and the vaccine’s ability to block transmission. Second, we consider variation in the dynamics and outcomes of infection by incorporating variation in both susceptibility to infection and the infection fatality rate by age. Third, we consider variation in the population, including the age distribution, age-stratified contacts, and the existing fraction of seropositive individuals, with uncertainty, by age.

There are two main approaches to vaccine prioritization: (1) directly vaccinate those at highest risk and (2) protect them indirectly by vaccinating those who do the most transmitting. For influenza, model-based investigations of the tradeoffs between these strategies have led to recommendations that children be vaccinated due to their critical role in transmission *(3, 4)*. In the context of COVID-19 vaccination, Matrajt et al. found that the tradeoff depends on vaccine efficacy, recommending direct vaccination of the elderly when the vaccine is ineffective but indirect protection otherwise *(5)*, while Gallagher et al. showed the importance of indirect effects more broadly *(6)*. The problem of vaccine prioritization also bears parallels to the more general problem of optimal resource allocation to reduce transmission, e.g. with masks *(7)*.

Our approach expands on previous literature in a number of ways. First, our model includes age-stratified seroprevalence data and hypothetical age-varying vaccine efficacy. Second, we demonstrate the generalizability of our model and recommendations by evaluating numerous geographic regions. Third, we consider how the use of serological testing prior to vaccination can re-prioritize doses to seronegative individuals, thereby increasing the effective number of available doses. Importantly, this serology-informed dose redirection has the potential to ameliorate existing inequities in the burden and impact of COVID-19. We identify an optimal vaccine allocation strategy and also evaluate five heuristic strategies which, while sub-optimal, may be easier to implement.

## Results

We evaluated the impact of vaccine prioritization strategies using an age-stratified SEIR model, because age has been shown to be an important predictor or correlate of susceptibility (*8–10*), seroprevalence (*8, 11*), severity (*12–14*), and mortality (*15, 16*). This model includes an age-dependent contact matrix, susceptibility to infection, and infection fatality rate (IFR), allowing us to estimate cumulative incidence and mortality via forward simulations. Cumulative incidence and mortality were then used as outcomes by which to compare the impacts of vaccine prioritization strategies (see Methods).

We first examined the impact of five straightforward vaccine prioritization strategies for a hypothetical infection- and transmission-blocking vaccine of varying efficacy. The strategies prioritized vaccines to children and teenagers, adults between ages 20 and 49, adults of 20 years or older, adults of 60 years or older, and all individuals irrespective of age (all ages; see Figure 1A). Although not exhaustive, these five strategies balance logistical simplicity while capturing a diverse set of possible outcomes from vaccine safety and efficacy trials.

We measured the reductions in cumulative incidence and mortality achieved by each strategy, varying the vaccine supply between 1% and 50% of the total population, and varying vaccine efficacy of 100%, 75%, and 50%, assuming that vaccines were administered before transmission begins in the population. We considered two ways to implement vaccine efficacy *υ*_{e} less than 100%: as an all-or-nothing vaccine, where the vaccine only protects *υ*_{e} percent of people, or as a leaky vaccine, where vaccinated people are *υ*_{e} percent protected against infection (see Methods).

Among the five strategies evaluated, and across all combinations of vaccine supply and vaccine efficacy, direct vaccination of adults 60 years and older consistently reduced mortality substantially more than alternative strategies (Figure 1B-D), followed by vaccination of adults 20+. In contrast, to minimize cumulative incidence, vaccination of adults 60+ performed the worst across all levels of vaccine efficacy. For an all-or-nothing vaccine, prioritizing adults 20-49y minimized cumulative incidence across all levels of vaccine supply and vaccine efficacy (Figure 1E,F,G). However, for a leaky vaccine with 50% efficacy, prioritizing children and teenagers minimized cumulative incidence (Supplementary Figure S5).

To determine whether our findings were robust across countries, we analyzed the ranking of prioritization strategies for demographics and modeled contact structures of Belgium (Figure 1) as well as the United States, India, Spain, Zimbabwe and Brazil. Across these countries, direct vaccination of adults 60+ minimized mortality irrespective of vaccine supply and for both types of vaccine efficacy at 100%, 75% and 50% (Supplementary Figs. S1, S2, S3, S4, and S5). Similarly, across countries, vaccination of adults 20-49y minimized infections for all levels of efficacy when assuming a perfectly effective vaccine or an all-or-nothing vaccine with 75% and 50% efficacy (Supplementary Figures S6, S7, S9). However, vaccinating children and teenagers was superior for a 50% effective leaky vaccine (Supplementary Fig. S10), with no consistent best strategy for a 75% effective leaky vaccine (Supplementary Fig. S8). The ranking of prioritization strategies thus depends on interactions between local context and the properties of the vaccine.

We also considered whether the rankings of prioritization strategies to minimize deaths would change, were a vaccine to block COVID-19 symptoms and mortality but not SARS-CoV-2 infection or transmission (Supplementary Text S2 and Supplementary Fig. S12) and found no meaningful changes in the rankings of strategies.

Vaccines may not be equally effective across age groups in preventing infection or transmission, a phenomenon known to affect influenza vaccines (*19–22*). To understand the impact of age-dependent vaccine efficacy, we incorporated a hypothetical linear decrease in efficacy of both infection and transmission prevention among those 60y and older (all-or-nothing: Figure 2A, leaky: Supplementary Figure S16). For all-or-nothing vaccines, the most effective strategies remained the prioritization of adults 20-49y to minimize cumulative incidence and the adults 60+ to minimize mortality.

Decreasing vaccine efficacy in older age groups diminished the benefits of any prioritization strategy that included this age group (Figure 2B). The reduction in mortality changed most when prioritizing vaccines to adults 60+, hinting that there might be a tipping point such that direct vaccination of the adults 60+ would no longer be the best approach to minimizing mortality. We investigated a variety of hypothetical vaccine efficacy curves to determine whether such a tipping point existed and if so, the new mortality-minimizing strategy. Across various specifications of total vaccine supply, baseline vaccine efficacy, and decreased efficacy among the adults 60+, the majority of scenarios had no tipping point and continued to favor direct vaccination of the adults 60+ (Table 1 and Supplementary Table S2). However, especially for leaky vaccines, in scenarios characterized by (i) high vaccine supply, (ii) high efficacy for those under 50, (iii) onset of decreased efficacy at a relatively young age, and (iv) very low efficacy among the oldest, vaccination of adults over 20 became the mortality-minimizing strategy. This analysis therefore strongly suggests that direct vaccination of the adults 60+ is likely to be the most effective way to minimize mortality.

If naturally acquired antibodies correlate with protection from reinfection, seroprevalence will affect vaccine prioritization in two ways. First, depending on the magnitude and age distribution of seroprevalence at the time of vaccine distribution, the ranking of strategies could change. Second, distributing vaccines to seropositive individuals would reduce the marginal benefit of vaccination per dose.

To investigate the impact of vaccinating in mid-epidemic while potentially employing serology to target the vaccine to seronegative individuals, we included age-stratified seroprevalence estimates in our model by moving the data-specified proportion of seropositive individuals from susceptible to recovered status. We then simulated two approaches to vaccine distribution. First, vaccines were distributed according to the five prioritization strategies introduced above, regardless of any individual’s serostatus. Second, vaccines were distributed with a serological test, such that seropositive individuals would not be vaccinated, allowing their dose to be given to someone else. In a sub-population *i* with seroprevalence *θ _{i}*, reallocating vaccines in this manner leads to an effective gain of an additional

*θ*(1

_{i}/*− θ*) doses for each originally allocated dose.

_{i}We tested this serology-informed modeling approach by including age-stratified seroprevalence estimates, demographics, and age-contact structure from New York City in August 2020 (*23*) (Figure 3) in evaluations of the previous five prioritization strategies. The ranking of strategies to minimize mortality (Figure 3A) and incidence (Figure 3B) remained unchanged: prioritizing adults 60+ most reduced mortality and prioritizing adults 20-49y most reduced incidence, regardless of whether vaccination was limited to only seronegative individuals. These rankings were unchanged when we used age-stratified seroprevalence estimates, demographics, and age-contact structure from Belgium in July 2020 (Supplementary Figure S13) to test the consistency of results.

Preferentially vaccinating seronegative individuals yielded large additional reductions in cumulative incidence/mortality in locations with higher seroprevalence (New York City (*23*), August 2020 mean: 26.9%, Fig. 3) and modest reductions in locations with low seroprevalence (Belgium (*11*), July 2020 mean: 4.5%, Fig. S13). These results remained unchanged when statistical uncertainty, due to sample size and imperfect test sensitivity and specificity, were incorporated into the model (*24*) (Fig S13).

Rather than adhering to one of the five pre-specified strategies, we next asked whether an optimal prioritization existed which outperformed all others, given age-stratified data on serology, contact structure, and demographics. To do so, we use a nonlinear optimization approach which searched over all possible prioritization strategies, subject to a specified total vaccine supply.

To demonstrate the use of this optimization in practice, we computed optimal strategies using Belgian data, deriving incidence-minimizing strategies for vaccine supplies of 10%, 20% and 30% of the total population. For a 10% total supply, the optimal strategy immunizes only those aged 30-49 (Figure 4A; top), while a 20% supply leads to expanded vaccination of some aged 10-29. A 30% total supply leads to broader vaccination of those 10-49 with a small fraction of vaccines prioritized for those 50-59 (Figure 4A; bottom). All optimal strategies for minimizing incidence prioritize multiple age groups before completely vaccinating any given age group (Figure 4C). In contrast, to minimize mortality, vaccines with equal efficacy across age groups should be prioritized to the oldest first, starting from the very oldest in the population and moving down the age range (Supplementary Fig. S14). The optimal strategy suppressed the epidemic at a vaccine supply of only 42% (Figure 4B), outperforming, as expected, the five pre-specified strategies. We caution that this optimal allocation strategy depends on country- and time-specific variables included in the modeled scenario (in this case, *υ*_{e} = 100% and *R*_{0} = 2.6, and no age-dependence of efficacy), and of course on assumptions about age-specific susceptibility, transmissibility, etc.

## Discussion

This study demonstrated the use of an age-stratified modeling approach to evaluate and compare vaccine prioritization strategies for SARS-CoV-2. After accounting for country-specific age structure, age-contact structure, infection fatality rates, and seroprevalence, as well as the age-varying efficacy of a hypothetical vaccine, we found that across countries those 60y and older should be prioritized to minimize deaths. This recommendation is robust because of the dramatic differences in IFR by age. Our model identified a few scenarios in which prioritizing all adults over 20 would provide greater mortality benefits than prioritizing older adults. These scenarios were restricted to the conditions of a large enough vaccine supply, very low efficacy in older adults, very high efficacy in younger adults, and declines in efficacy starting at 49 or 59, for a leaky vaccine, and even more restrictive conditions for an all-or-nothing vaccine. Moreover, the advantage of prioritizing all adults vs. older adults was small even with these conditions. Thus, we conclude that for mortality reduction, prioritization of older adults is a robust strategy that will be optimal or close to optimal to minimize mortality for virtually all plausible vaccine characteristics.

In contrast, the ranking of infection-minimizing strategies depended on whether we considered an all-or-nothing vaccine or a leaky vaccine. Modeling all-or-nothing vaccines led to consistent recommendations to prioritize adults 20-49y across efficacy values and countries. However, for leaky vaccines, prioritization shifted toward children and teenagers for vaccine efficacy of 50%, in line with prior work (*5*). We also found that the transition point between top-ranked strategies varied by country demographics and contact patterns when considering a leaky vaccine. Because a vaccine is likely to have properties of both leaky and all-or-nothing models, empirical data on vaccine performance could help resolve this difference in model recommendations, recognizing that such data are difficult to obtain in practice (see, e.g. (*25,26*)).

While any pre-specified age-based prioritization strategy can be evaluated using this model, we showed how prioritization can be optimized to meet goals of either minimizing incidence or minimizing death most efficiently. Critically, this approach showed that an incidence-minimizing strategy prioritizes doses to those 10-59y, while a mortality-minimizing strategy prioritizes direct vaccination of the 60+ population. This finding is a consequence of the high infection fatality rates among the adults 60+ and an inability of lower incidence alone to provide indirect protection. However, a vaccine with declining efficacy among the elderly could change optimal strategies, highlighting the importance of age-stratified estimates of vaccine efficacy.

Due to the expedited testing and approvals process, it is not yet clear whether the first-generation of COVID-19 vaccines will be approved for the elderly or those under 18y (*27,28*). While our conclusions assumed that the vaccine would be approved for all age groups, both the evaluation and optimization approaches introduced here can be tailored to evaluate a subset of approaches restricted to those within the age groups for which a vaccine is licensed, using open-source tools such as those that accompany this manuscript. Furthermore, we considered two possible goals of vaccination: minimizing cumulative incidence or mortality. Our framework could also be adapted to include more morbidity, age-specific data in order to consider other goals of vaccination, such as minimizing hospitalizations or economic costs, or maximizing quality-adjusted life years.

We demonstrated that there would be value in pairing individual-level serological tests with vaccination, inclusive of the uncertainties in existing seroprevalence surveys (*24*), particularly in locations where seroprevalence rates are high. However, the marginal gain in effective vaccine supply, relative to no serological testing, must be weighed against the challenges of serological testing prior to vaccination. Delays in serological tests results would impair vaccine distribution.

Equitable vaccine distribution among social and demographic groups is a major concern for vaccine allocation strategies, and one our model did not explicitly address. Nevertheless, vaccinating seronegative individuals could help to ensure equitable allocation as long as natural infection confers protection at least as good as vaccination and vaccines are distributed to specific geographical areas for exclusively local use. For instance, in a hypothetical neighborhood with high seroprevalence, vaccinating only seronegatives would extend the local positive direct and indirect benefits of vaccination, but only if doses were reallocated locally. In this way, serology could allow locales which suffered larger burdens to extend their doses further. We caution that changing distribution in response to seroprevalence and rebalancing doses away from high seroprevalence neighborhoods could create a double blow of high initial burden followed by seroprevalence-based deprioritization during vaccination.

Our recommended strategies depend on the contact matrix, which is affected by social distancing. We used pre-pandemic contact matrices (*17*) with the goal of returning to normal life once a vaccine is available. The scenarios modeled here did not incorporate physical distancing or other non-pharmaceutical interventions, which we would expect to persist if vaccination coverage is incomplete. Our framework could easily be adapted to pandemic-era contact matrices like the CoMix survey (*29*). Doing so will indeed be critical, as mobility restrictions fluctuate.

Our study relies on both global and country-specific estimates of epidemiological parameters and therefore highlights key quantities that must be estimated as part of effective vaccine prioritization. In local contexts, these include age-structured seroprevalence and IFR, which vary by geographic region (*15, 16, 30*). Globally, key parameters include the degree to which antibodies protect against reinfection or severity of disease and relative infectiousness by age. From vaccine trials, we also need evidence of efficacy in groups vulnerable to severe outcomes, including the elderly. Additionally, it will be critical to measure whether the vaccine protects only against symptomatic disease, or whether it also blocks infection and transmission of SARS-CoV-2. If the vaccine does not block transmission or only protects against severe infection, the public health community will need to clearly communicate this to the public to discourage risky behavior that could increase negative externalities on the most vulnerable who cannot be vaccinated.

The role of children during this pandemic has been unclear. Under our assumptions about susceptibility by age, children are not the major drivers of transmission in communities. Thus, our results differ from the optimal distribution for influenza vaccines, which prioritize school-age children and adults age 30-39 (*4*). However, both the relative susceptibility and relative infectiousness of SARS-CoV-2 by age remain uncertain. While it is unlikely that susceptibility to infection is constant across age groups (*8*), we ran our model to test the sensitivity of this parameter. Under the scenario of constant susceptibility by age, vaccinating those under 20 has a greater impact on reducing cumulative cases than before, but the overall ranking of strategies remains the same (Supplementary Figures S17, S18). In our model, we assumed constant infectiousness (conditional on contact) by age, but a better understanding of this is paramount to accurately modeling disease dynamics.

Our study is subject to a number of limitations. First, our evaluation and optimization strategy considers variation in disease severity only by age. However, there are other factors that are have been shown to correlate with disease outcomes, such as treatment and healthcare access and comorbidities, which may correlate with factors like rural vs urban location, socioeconomic status, sex (*31, 32*), and race and ethnicity (*33*), that are not accounted for in this study. Inclusion of these factors in a model would be possible, but only with statistically sound measurements of both their stratified contact rates and disease outcomes. Even in the case of age stratification, contact surveys have typically not surveyed those 80 years and older, yet it is this population that suffers dramatically more severe COVID-19 disease and higher infection fatality rates. We extrapolated contact matrices to those older than 80, but direct measurements would be superior.

Our study also considers variation in disease risk only by age, via age-structured contact matrices, while many discussions around COVID-19 vaccine distribution have thus far focused on prioritizing healthcare or essential workers (*34, 35*). Contact rates, and thus transmission potential, vary greatly not only by occupation and age but also by living arrangement (e.g., congregate settings, dormitories), neighborhood and mobility (*36–39*), and whether the population has a coordinated and fundamentally effective policy to control the virus. With a better understanding of population structure during the pandemic, and risk factors of COVID-19, these limitations could be addressed. Meanwhile, the robust findings in favor of prioritizing those age groups with the highest IFR to minimize mortality could potentially be extended to prioritize those with comorbidities that predispose them to a high IFR, since the strategy of prioritizing the older age groups depends on direct rather than indirect protection.

Vaccine prioritization is not solely a question of science but a question of ethics as well. A hallmark of the COVID-19 pandemic, as with other global diseases, is one of inequalities and disparities. While our modeling efforts incorporate key parameters that focus on age and minimizing incidence and death within a simply structured population, other considerations are crucial, from equity in allocation between countries down to disparities in access to the benefits of vaccination within individual communities. Thus, the model’s simplistic representation of vulnerability (age) should be augmented over time with better information on the correlates of infection risk and severity. Fair vaccine prioritization should avoid further harming disadvantaged populations. We suggest that pairing serological testing with vaccination in the hardest hit populations is one possible equitable way to extend the benefits of vaccination in settings where vaccination might otherwise not be deemed cost-effective.

## Methods

We used a continuous-time, ODE-based compartmental model stratified by age. The model compartments (*S _{i}, E_{i}, I_{i}, R_{i}, V_{i}*), described below, represent mutually exclusive epidemiological states stratified by age group

*i, i*∈ {0-9, 10-19, …, 70-79, 80+}. We assume that people are initially susceptible, unless they have been effectively vaccinated or have infection-derived antibodies, assuming antibodies are protective against reinfection. Susceptible people (

*S*) transition to the exposed state (

*E*) after effective contact with an infectious individual. After a latent period, exposed individuals become infectious (

*I*). After an infectious period, individuals move to a recovered state (

*R*). We assume that recovered individuals are no longer infectious and are immune to reinfection. The duration of time spent in each compartment, in expectation, is specified in Table S1. Model equations were solved using

*lsoda*ODE solver from the package ‘deSolve’, R version 3.6.0 (

*40*).

The force of infection, *λ*_{i} for a susceptible individual in age group *i* is
where *β*_{i} is the probability of a successful transmission given contact with an infectious individual, *c _{ij}* is the number of age-j individuals that an age-i individual contacts per day, and

*I*/

_{j}*N*is the probability that a random age-

_{j}*j*individual is infectious. To calculate the basic reproductive number,

*R*

_{0}, we define the next-generation matrix as where

*D*

_{β}is a diagonal matrix with diagonal entries

*β*is the country-specific contact matrix, and

_{i}, C*D*

_{dI}is a diagonal matrix with diagonal entries

*d*, where

_{I}*d*

_{I}is the infectious period.

*R*

_{0}is the absolute value of the dominant eigenvalue of

*M*.

We scaled the estimated age-susceptibility *β*_{i} from Davies et al. (*9*) to obtain an *R*_{0} = 2.6 in Belgium, and used the same *β*_{i} for all countries (*29*). Thus, the realized *R*_{0} changes by country according to the contact matrix and demographics.

The *R*_{0} for SARS-CoV-2 has been estimated between 2.1 and 3.1, with a mean of 2.6 (*29*). We ran the Belgian scenario for this range of *R*_{0} to test the sensitivity of our results. The ranking of the strategies remained the same for all simulated values of *R*_{0} (Supplementary Fig. S11).

We used synthetic country-specific contact matrices from Prem et al. (*17*). These matrices include four types of contact: home, work, school, and other. We combined their 5-year age bins into 10-year age bins, and extrapolated an 80+ age group (See Supplementary Text S1). We used country age demographics from the UN World Population Prospects 2019 (*18*).

We ran simulations to completion of the outbreak, when the total number of infectious people in compartment *I* was less than 1. To compare the impact of different vaccination prioritization strategies, we calculated the cumulative number of infections and deaths. To estimate the total number of deaths, we multiplied the cumulative number of infections for each age group at the end of the simulation by an age-dependent IFR (*15*), and thus assumed deaths negligibly impacted transmission.

In the simplest version of the model, the vaccine is assumed to be transmission- and infection- blocking, and to work with 100% efficacy. To incorporate vaccinations, we initialized the model by dividing the total population of each age group between the susceptible compartment (*S*) and vaccinated compartment (*υ*), according to the vaccine prioritization strategy and number of vaccines available. The model was seeded with one infected person in each age group *i*, unless all age-*i* individuals were vaccinated.

We also implemented age-dependent vaccine efficacy, *υ*_{e}. A vaccine with less than perfect efficacy can either be a leaky vaccine, where vaccinated people are *υ*_{e} percent protected against infection, or an all-or-nothing vaccine, where the vaccine effectively protects *υ*_{e} percent of people but does not work in (1- *υ _{e}*) percent of people (See Supplementary Text S3). Since the vaccine efficacy is unknown at this time, we ran simulations with both types of vaccine efficacy. The type of vaccine efficacy does not change the ranking of the prioritization strategies (all-or-nothing vaccine: Results, leaky vaccine: Supplementary Fig. S16).

To incorporate age-dependent vaccine efficacy, we used a three parameter vaccine efficacy curve to define a baseline *υ _{e}*, hinge age and final

*υ*

_{e}(Supplementary Fig. S15). We assumed that the vaccine efficacy starts at a baseline

*υ*

_{e}for all ages younger than the hinge age, then decreases stepwise in equal increments for each decade of age to a final

*υ*

_{e}for the 80+ age group. We define the vaccine efficacy tipping point as the final

*υ*

_{e}where the best strategy switches from vaccinating the adults 60+ directly to minimize deaths. We estimated this point using the bisection method (

*41*) for various for baseline

*υ*

_{e}and hinge ages (Table 1 and S2).

To incorporate serology and compare areas with differing naturally-acquired immunity, we used data and seroprevalence estimates from New York City and Belgium (*11, 23*). We inferred the true seroprevalence distributions for these seroprevalence estimates when the sensitivity and specificity of the antibody test was available, according to (*24*). To assess how sensitive the model results are to the seroprevalence, we sampled from these distributions 100 times to get seroprevalence estimates for each age group, and then ran the simulation forward for each estimate. To set our initialization parameters, we moved the corresponding proportion of seropositive individuals into the recovered compartment, *R*.

There are two ways to implement a vaccination strategy with seropositive individuals in the population: either vaccinate people regardless of serostatus, or pair the vaccine strategy with a serological test. That way, if someone is seropositive, they will not be vaccinated under the assumption that innate immunity is protective against reinfection. Without a paired serology test, the initial conditions for our model were
for an age group *i*, where *υ*_{i} is the number of people of age *i* vaccinated according to a given vaccination strategy, *P* (vaccinated_{i}) is the probability that an individual of age *i* is vaccinated according to a given vaccine allocation strategy, and *P* (seropositive_{i}) is the probability that a person of age *i* is seropositive. If an individual is both seropositive and vaccinated, they are put in the recovered compartment, *R*.

Alternatively, when vaccinating with paired serology tests, the initial conditions are

Finally, we inverted the problem and asked: given serology, a certain vaccine efficacy, contact structure, and demographics, what would be the optimal prioritization of vaccines? For this, we used a constrained optimization algorithm *cobyla* from the package ‘NLopt’. The constraints on our system were vaccinating a minimum of 0% and a maximum of 100% people in each age group, and fixing the total amount of vaccine supply. For the constrained optimization, we initially allocated vaccines uniformly across all age groups, with noise to make the initialization stochastic. Then, we ran the optimization algorithm 50 times for vaccine supply varying between 1% and 50% of the population and stored the allocation that performed best.

## Data Availability

Contact matrices, demographic data and model parameters are available via the cited references in the manuscript.

## Supplementary Materials

### S1 Extrapolating contact matrix to 80+

We made two adaptations to the contact matrices from Prem et al. (*17*) First, we combined their five year age bins into ten year bins to match our model’s age bins. Second, we extrapolated their final age bin (75+) to add an 80+ age bin. To extrapolate, we copied the contact rates from 70-79 y.o. to our new row and column for 80+, along the diagonal. Then we filled in the end of our new row and column with the 70-79 y.o. contact rates with 0-9 y.o., assuming interactions with 0-9 y.o. are similar for people 70+. Lastly, to account for increased housing in long term living facilities for 80+ y.o., we decreased their contacts for 0-60 y.o. by 10% and added it to the 70 and 80 y.o. contacts. Thus, 80+ year-olds have the same total number of contacts as 70-79 year-olds, but relatively fewer among 0-69-year-olds and proportionally more among 70+ year-olds.

### S2 Non-transmission blocking vaccine

We considered a vaccine that prevents severe manifestations of COVID-19 infection, including death, but does not block transmission of SARS-CoV-2. To implement, we used an SEIR model according to the schematic:
where the superscript V denotes vaccination. The force of infection depends on the vaccination state of the susceptible individual such that
and
where *α* scales the relative susceptibility to infection for susceptible individuals who were vaccinated and *ω* scales the infectiousness of vaccinated individuals who are infected with SARS-CoV-2. The impact of this non-transmission blocking model on minimizing incidence and mortality are shown in Supplementary Fig. S12 for *α* = *ω* = 1 and *α* = *ω* = 0.5. When *α* = *ω* = 1, there are no indirect effects of vaccination. The ranking of the strategies remains the same as a vaccine that blocks transmission of the virus: prioritize vaccines to adults 20-49y to minimize incidence and adults 60+ to minimize mortality.

### S3 Implementing vaccine efficacy: leaky vs all-or-nothing

A vaccine with less than perfect efficacy can either be a leaky vaccine, where vaccinated people are *υ _{e}* percent protected against infection, or an all-or-nothing vaccine, where the vaccine effectively protects

*υ*

_{e}percent of people but does not work in (1-

*υ*) percent of people. We considered both types of vaccine efficacy (leaky vaccine: Results, all-or-nothing vaccine: Supplementary Fig. S16). To implement, we used an SEIR model with initial conditions according to the following schematic:

_{e}### Supplementary Figures

### Supplementary Tables

## Acknowledgments

The authors wish to thank Sereina Herzog. The work was supported in part by the Interdisciplinary Quantitative Biology (IQ Biology) PhD program at the BioFrontiers Institute, University of Colorado Boulder. The work was supported in part by the Morris-Singer Fund for the Center for Communicable Disease Dynamics at the Harvard T.H. Chan School of Public Health. Reproduction code is open source and provided by the authors.