While vaccines against severe acute respiratory syndrome coronavirus (SARS-CoV-2) are being administered, in many countries it may still take months until their supply can meet demand. The majority of available vaccines elicit strong immune responses when administered as prime-boost regimens. Since the immunological response to the first (“prime”) dose may provide already a substantial reduction in infectiousness and protection against severe disease, it may be more effective—under certain immunological and epidemiological conditions—to vaccinate as many people as possible with only one dose instead of administering a person a second (“booster”) dose. Such a vaccination campaign may help to more effectively slow down the spread of SARS-CoV-2 and reduce hospitalizations and fatalities. The conditions that make prime-first vaccination favorable over prime-boost campaigns, however, are not well understood. By combining epidemiological modeling, random-sampling techniques, and decision tree learning, we find that prime-first vaccination is robustly favored over prime-boost vaccination campaigns even for low single-dose efficacies. For epidemiological parameters that describe the spread of coronavirus disease 2019 (COVID-19), recent data on new variants included, we show that the difference between prime-boost and single-shot waning rates is the only discriminative threshold, falling in the narrow range of 0.01–0.02 day1 below which prime-first vaccination should be considered.

In the current stage of the coronavirus disease 2019 (COVID-19) pandemic, many countries are still facing limited vaccination supply and fear the increasingly wide-spread emergence of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) variants of concern such as the Delta and Gamma strains. For the majority of available vaccines, prime-boost regimes are considered immunologically efficient since they elicit a strong immune response. However, given the current vaccine supply and distribution constraints, it may be epidemiologically more effective to first vaccinate as many people as possible with only one (“prime”) dose, before administering a second (“boost”) dose. Such a strategic vaccination campaign may help to reach a sufficient level of population immunity more rapidly and effectively slow down the spread of SARS-CoV-2, thereby reducing fatalities and the risk of collapsing health care systems. Still, the conditions that make prime-first vaccination favorable over prime-boost regimens are not clear and subject to debate. To identify decisive conditions for strategic vaccination against SARS-CoV-2, we combine epidemiological modeling, random-sampling techniques, and decision tree learning. This approach allows us to conclude that prime-first vaccination campaigns are epidemiologically more effective than prime-boost campaigns even for high vaccination rates, a relatively low single-dose efficacy, and for a large degree of uncertainty in key epidemiological data.

After the initial identification of the novel severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) in Wuhan, China in December 2019, the virus quickly reached pandemic proportions and caused major public health and economic problems worldwide.1 The disease associated with SARS-CoV-2 infections was termed coronavirus disease 2019 (COVID-19). As of August 28, 2021, the number of confirmed COVID-19 cases exceeded 216 million, and more than 4.5 million COVID-19 deaths in at least 221 countries and territories were reported.2 Large differences between excess deaths and reported COVID-19 deaths across different countries suggest that the actual death toll associated with COVID-19 is even higher.3 

With the start and continuation of vaccination campaigns against SARS-CoV-2 in many countries,4 millions of people will receive partial and full immunization in the coming months. The mRNA vaccines BNT162b2 (BioNTech–Pfizer) and mRNA-1273 (Moderna) have been approved in several countries.5 When administered as a prime-boost regimen, these vaccines have a reported protective efficacy of 95%6 and 94.1%,7 respectively. An effectiveness evaluation of the BioNTech–Pfizer vaccine shows that it may offer about 50% protection against SARS-CoV-2 infections about 2–3 weeks after receiving the first dose.8 The adenovirus-based vaccine ChAdOx1 (Oxford–AstraZeneca) is being used in the UK, EU, and other countries with a reported single-dose efficacy between 62%–79%.9,10 Estimates of the vaccine effectiveness of prime-boost regimens against symptomatic disease are reported to be 88% and 80% for Alpha and Delta variants,11 respectively. Studies on the effectiveness of prime-boost vaccines against hospitalization12,13 suggest values of 92% and 94% for Alpha and Delta strains, respectively.

Taken together, the majority of currently available SARS-CoV-2 vaccines elicit strong immune responses against all studied variants when administered as prime-boost regimens. However, given the current distribution and production constraints, it may take months until the production of COVID-19 vaccines can meet the actual global demand. Similar to vaccination campaigns in previous disease outbreaks, it may, therefore, be a favorable alternative to administer a single vaccination dose to twice as many people. In 2016, a single-dose vaccination campaign against cholera was implemented in Zambia because of the insufficient number of vaccination doses that were available to complete a standard two-dose campaign.14 Other vaccines, such as the oral cholera vaccines that require two doses, are highly effective after a single dose, but their protection is short lived compared to that obtained with prime-boost vaccination.15,16

While there are countries that adopted a prime-boost protocol with an inter-dosing interval of about 1 month for COVID-19 vaccinations,17 some countries implemented vaccination protocols with longer inter-dosing intervals.18 For example, the Canadian Ministry of Health describes in Ref. 19 that they follow the National Advisory Committee on Immunization in recommending an interval of 4 months between the first and second dose to “increase the number of individuals benefiting from the first dose of vaccine.”

Despite clear advantages of such prime-first vaccination campaigns, including faster immunization of a larger number of people and lower vaccine-distribution infrastructure requirements and costs, any deviation from the immunologically favorable prime-boost protocol may negatively affect the level of vaccination-induced immunity. An analysis of blood samples from COVID-19 patients suggests that T cells play an important role in the long-term defense against SARS-CoV-220 since antibody concentrations were found to decay faster than those of T cells that respond to SARS-CoV-2 epitopes. Clinical trial results21 on the COVID-19 vaccine BNT162b1 show that the vaccination-induced CD4+ and CD8+ T-cell responses are significantly reduced if no boost shot was administered, indicating that booster doses are important for T-cell-mediated immunity against SARS-CoV-2. In the same study, antibody concentrations in patients who received prime-boost regimes were found to be about 5–20 times higher than those observed in patients who only received a single vaccination dose, highlighting the need for boosting. Similar observations were made for type-1 inactivated poliovirus vaccine (IPV), for which clinical trial results22 suggest that boost injections are needed to increase the level of neutralizing antibodies. However, for type-2 and -3 IPV, the first vaccination dose already elicits a neutralizing antibody response. In addition, single-dose vaccination may provide already a substantial degree of protection against infection, as confirmed in studies for BioNTech–Pfizer8 and Oxford–AstraZeneca23 vaccines. Still, the mechanisms underlying vaccination-induced humoral (antibody-mediated) and cell-mediated immune responses against SARS-CoV-2 are not well understood, and data on immunity waning are scarce.24,25

Here, we study population dynamics of SARS-CoV-2 that account for vaccine-induced protection levels, immunity waning, and other immunological factors to analyze under which epidemiological conditions prime-first vaccination is favorable over prime-boost vaccination. The advantages and disadvantages of these alternative vaccination protocols are subject to debate in many countries, including the US,26,27 the UK,28,29 and Germany,18 as they are fearing the spread of more contagious and more deadly SARS-CoV-2 mutants,30 such as the Delta and Gamma variants, and the risk of collapsing health care systems.31 

The current controversy around prime-first and prime-boost vaccination strategies raises two connected questions, which we address in this paper: How do shortages in vaccine supplies and uncertainties in epidemiological parameters alter the possible advantage of prime-first over prime-boost vaccination? And how do possible differences in vaccine efficacy and loss of vaccine-induced immunity affect the decision boundary separating prime-first and prime-boost vaccination regimes in a high-dimensional parameter space? By combining methods from epidemiological modeling, statistical mechanics,32,33 and decision tree learning, we explore position, extent, and sensitivity of the decision boundary and provide a characterization of discriminative criteria,34 sufficiently simple and immediately accessible to decision makers.

Different vaccination campaigns may lead to different proportions of infected, recovered, and deceased individuals at a given time. We study the differences between prime-first (Fig. 1) and prime-boost campaigns by accounting for a vaccination-induced reduction in transmissibility in a susceptible–exposed–infected–recovered–deceased (SEIRD)-based model35 (see  Appendices A– C and Fig. 1). To quantify the effect of prime-first and prime-boost vaccination protocols on the overall disease-induced fatality, we use two fatality measures. The first measure is based on fatality rates, and the second one is based on cumulative deaths. Specifically, let d1 and d2 be the maximum (daily) changes in the total number of deaths within the time horizon of about 10 months (T=300 days) for prime-first and prime-boost protocols, respectively. In addition, we use D1 and D2 to denote the corresponding cumulative numbers of disease-induced deaths. As a measure of the relative difference between d1 and d2, we use the relative fatality change (RFC-δ),

δ(d1,d2)=d2d1max(d1,d2).
(1)

As a cumulative measure, we study the relative change in the cumulative number of disease-induced deaths (RFC-Δ),

Δ(D1,D2)=D2D1max(D1,D2),
(2)

defined within the same time horizon as RFC-δ.

FIG. 1.

Schematic of vaccination campaigns and model. (a) and (b) Evolution of the number of individuals that received prime (n) and prime-boost (n) doses in a population of size N. (a) Prime-first () vaccination: Prime doses are administered with a maximal rate as long as unvaccinated susceptible individuals exist. Then, booster doses are administered to prime-vaccinated individuals. (b) Prime-boost () vaccination: Prime and booster doses are balanced equally. Boosting starts after td days. (c) In our model, susceptible individuals (S, S, S) become exposed (E, E, E) at rates β,β,β and transition to an infected state (I, I, I) at rates σ,σ1,σ2. Infectious individuals either recover (R) or die (D) at rates x(1fx)γx and xfxγx, respectively. Prime vaccination doses () are administered to susceptible individuals at rate ν1, and prime-vaccinated individuals receive a booster dose () at rate ν2. Transitions from S, S, and R to S (“waning immunity”) occur at rates η1, η2, and η3, respectively. Notation: xβxIx=βI+βI+βI.

FIG. 1.

Schematic of vaccination campaigns and model. (a) and (b) Evolution of the number of individuals that received prime (n) and prime-boost (n) doses in a population of size N. (a) Prime-first () vaccination: Prime doses are administered with a maximal rate as long as unvaccinated susceptible individuals exist. Then, booster doses are administered to prime-vaccinated individuals. (b) Prime-boost () vaccination: Prime and booster doses are balanced equally. Boosting starts after td days. (c) In our model, susceptible individuals (S, S, S) become exposed (E, E, E) at rates β,β,β and transition to an infected state (I, I, I) at rates σ,σ1,σ2. Infectious individuals either recover (R) or die (D) at rates x(1fx)γx and xfxγx, respectively. Prime vaccination doses () are administered to susceptible individuals at rate ν1, and prime-vaccinated individuals receive a booster dose () at rate ν2. Transitions from S, S, and R to S (“waning immunity”) occur at rates η1, η2, and η3, respectively. Notation: xβxIx=βI+βI+βI.

Close modal

A positive sign of RFC-Δ and RFC-δ, respectively, indicates that the cumulative number and the maximum daily number of disease-induced deaths is larger under a prime-boost vaccination strategy than under a prime-first vaccination campaign. A negative sign of RFC-Δ and RFC-δ indicates the opposite. In  Appendix D, we show that and how the two measures are correlated.

Current vaccination campaigns prioritize health care workers and vulnerable groups (e.g., elderly people with comorbidities) with a high risk of infection, leading to variations in vaccination rates. Further heterogeneity in model parameters may arise from infection rates that differ between age groups because of different degrees of susceptibility to infection36 and different mobility characteristics. Our model accounts for these variations in epidemiological parameters through a large degree of parameterization. Nine different infection rates describe contacts between (susceptible and infectious) unvaccinated, single-dose vaccinated, and prime-boost vaccinated individuals. This large degree of parameterization can effectively account for possible correlations between age group, transmissibility, and mobility. We, therefore, choose not to incorporate demographic compartmentalization in our model.33 In the following sections and in  Appendices E,  F,  G, and  H, we study the effects of age-stratification, natural immunity waning, and effects from parameter constraints in separate scenarios.

To provide mechanistic insight into the population-level differences between prime-first and prime-boost vaccination campaigns, we study how RFC-δ and RFC-Δ are impacted by epidemiological parameters and epidemic state. As a function of two parameters, green domains as shown in Fig. 2 indicate preference for prime-first campaigns, while red regions indicate that prime-boost is favored over prime-first vaccination. The parameter ranges follow the existing literature or are chosen sufficiently broad to cover uncertainties. Empirical data37 suggest an estimated range of the basic reproduction number R0[1,4] for the wild-type virus strain. Other SARS-CoV-2 variants such as the Delta strain may be outside this range. The following analysis and the data presented in Fig. 2 suggest that our main results also hold for larger values of R0.

FIG. 2.

Vaccination-campaign-preference diagrams. For combinations of basic reproduction number R0, waning-rate difference η1η2, initial disease prevalence I(0), maximum vaccination rate νmax, and relative efficacy for prime-first immunization (RE), β2/β1, we plot RFC-δ [Eq. (1)] and RFC-Δ [Eq. (2)]. Green-shaded regions indicate preference for prime-first (RFC-δ>0, RFC-Δ>0), and red-shaded regions indicate preference for prime-boost (RFC-δ<0, RFC-Δ<0). (a)–(l) Parameter domain as in Table I, assuming a moderate single-dose efficacy, i.e., β1=β/2, f=f/10. (k)–(m) Low single-dose efficacy domain: we set β1=0.9β, f=0.6×102=0.6f, and all remaining parameters as in Table I. For (m), we varied β1/β and β (hence R0) and set all other parameters according to Table I. The ratios β2/β1=β2/β1=β2/β1=1/5 are also as in Table I. Dashed lines: Decisive threshold η1η2=0.017day1 and Israel’s vaccination rate as of February 1, 20214 (νmax=0.013day1). Solid line in (c): decision boundary as a guide to the eye between y=η1η2 and x=I(0) as given by the following nonlinear relations: y=0.4x1/2+0.02 (RFC-δ) and y=0.06x1/4+0.017 (RFC-Δ).

FIG. 2.

Vaccination-campaign-preference diagrams. For combinations of basic reproduction number R0, waning-rate difference η1η2, initial disease prevalence I(0), maximum vaccination rate νmax, and relative efficacy for prime-first immunization (RE), β2/β1, we plot RFC-δ [Eq. (1)] and RFC-Δ [Eq. (2)]. Green-shaded regions indicate preference for prime-first (RFC-δ>0, RFC-Δ>0), and red-shaded regions indicate preference for prime-boost (RFC-δ<0, RFC-Δ<0). (a)–(l) Parameter domain as in Table I, assuming a moderate single-dose efficacy, i.e., β1=β/2, f=f/10. (k)–(m) Low single-dose efficacy domain: we set β1=0.9β, f=0.6×102=0.6f, and all remaining parameters as in Table I. For (m), we varied β1/β and β (hence R0) and set all other parameters according to Table I. The ratios β2/β1=β2/β1=β2/β1=1/5 are also as in Table I. Dashed lines: Decisive threshold η1η2=0.017day1 and Israel’s vaccination rate as of February 1, 20214 (νmax=0.013day1). Solid line in (c): decision boundary as a guide to the eye between y=η1η2 and x=I(0) as given by the following nonlinear relations: y=0.4x1/2+0.02 (RFC-δ) and y=0.06x1/4+0.017 (RFC-Δ).

Close modal
TABLE I.

Overview of model parameters for scenarios without natural immunity waning. The listed parameter values are used when the corresponding model parameters are held constant in the parameter-space plots shown in Fig. 2. As initial fractions of infected and susceptible individuals, we use I(0) = 10−2 and S(0) = 1 − I(0).

ParameterSymbolValueUnitsComments/references
Infection rates S β, β, β 3/14, β/10, β/20 day−1 β inferred from R0 = β/γ37  
Infection rates S β1, β1, β1 β/2, β/2, β/2 day−1 Estimate 
Infection rates S β2, β2, β2 β/10, β/10, β/10 day−1 Estimate 
Incubation rate E σ 1/5 day−1 Refs. 37 and 62  
Incubation rate E σ1 1/5 day−1 Estimate 
Incubation rate E σ2 1/5 day−1 Estimate 
Vaccination rate νmax 10−3 day−1 Ref. 4  
Waning rate (prime) η1 10−2 day−1 Estimate 
Waning rate (prime-boost) η2 3 × 10−3 day−1 Estimate 
Waning rate (recovered) η3 day−1 Estimate 
Resolution rate I γ 1/14 day−1 Refs. 57 and 63  
Resolution rate I γ 2γ day−1 Estimate 
Resolution rate I γ 4γ day−1 Estimate 
Fatality ratio I f 10−2 … Refs. 3 and 42  
Fatality ratio I f 10−3 … Estimated from reported efficacy5  
Fatality ratio I f 10−3 … Estimated from reported efficacy5  
Prime-boost delay td 21 day Ref. 5 and Fig. 4  
ParameterSymbolValueUnitsComments/references
Infection rates S β, β, β 3/14, β/10, β/20 day−1 β inferred from R0 = β/γ37  
Infection rates S β1, β1, β1 β/2, β/2, β/2 day−1 Estimate 
Infection rates S β2, β2, β2 β/10, β/10, β/10 day−1 Estimate 
Incubation rate E σ 1/5 day−1 Refs. 37 and 62  
Incubation rate E σ1 1/5 day−1 Estimate 
Incubation rate E σ2 1/5 day−1 Estimate 
Vaccination rate νmax 10−3 day−1 Ref. 4  
Waning rate (prime) η1 10−2 day−1 Estimate 
Waning rate (prime-boost) η2 3 × 10−3 day−1 Estimate 
Waning rate (recovered) η3 day−1 Estimate 
Resolution rate I γ 1/14 day−1 Refs. 57 and 63  
Resolution rate I γ 2γ day−1 Estimate 
Resolution rate I γ 4γ day−1 Estimate 
Fatality ratio I f 10−2 … Refs. 3 and 42  
Fatality ratio I f 10−3 … Estimated from reported efficacy5  
Fatality ratio I f 10−3 … Estimated from reported efficacy5  
Prime-boost delay td 21 day Ref. 5 and Fig. 4  

Differences in the waning rates η1 and η2 are not known at the present time, not even conclusive estimates,38 while clinical trials are still ongoing. The characteristic time scale of waning immunity for prime-boost vaccinated individuals associated with a waning rate η2=3×103day1 is about 1 year. Thus, we sample a broad parameter range, η1η2[104,101]day1 with η2=3×103day1, which includes waning time scales that were reported earlier for SARS-CoV.24 The corresponding sampling interval for η1 includes waning time scales between 10 days and 1 year. For sampling the initial disease prevalence I(0), we use the interval [104,101]. This interval includes up to 10% infected individuals but may lie outside prevalence estimates that were reported for some places with a very high prevalence such as Manaus, Brazil39 August 2020 and New York City, USA.40 For the range of the maximum vaccination rate νmax, we use [0,101]day1, which we inferred from current vaccination-campaign data.4 

We assume that the transmission rates β1 and β2 are proportional to the vaccine efficacies after single-dose and prime-boost vaccination, respectively. Thus, we identify the relative efficacy for single-dose immunization (RE) with the ratio β2/β1. Values close to one are favorable for prime-first campaigns, while a low RE disfavors prime-first.

In order to analyze the effect of RE on the effectiveness of prime-first and prime-boost vaccination campaigns, we sample β2/β1 from the interval [104,1]day1. We choose this rather broad range to account for the lack of reliable data, in particular, regarding new variants of SARS-CoV-2 and possible adverse effects in vaccine protection.41 Parameters that are held constant in our simulations are listed in Table I.

The vaccination-campaign-preference diagrams (Fig. 2) suggest that prime-first vaccination campaigns are associated with a smaller death toll compared to prime-boost campaigns for a wide range of R0, maximum vaccination rates, epidemic states, and relative efficacy ratios (green-shaded regions in Fig. 2).

As the main result of our study, we identify a two-parameter threshold combination that separates vaccination-campaign preferences (dashed black lines in Fig. 2). For a sufficiently small waning-rate difference η1η20.02day1 and a sufficiently low maximum vaccination rate νmax0.02day1, we observe that prime-first vaccination outperforms prime-boost vaccination in all projections where parameters are held constant as specified in Table I. In the projections involving η2η1, prime-boost preference is observed if immunity wanes significantly faster for prime-vaccinated individuals than for prime-boost vaccinated individuals.

All projections in Fig. 2, therefore, suggest that prime-boost vaccination should only be favored for νmax0.02day1, which largely exceeds SARS-CoV-2 immunization rates worldwide.4 

How a relatively low single-dose efficacy affects the preference for each campaign is shown in Figs. 2(k)2(m). In Figs. 2(k) and 2(l), we assume a transmission reduction of only 10% after \hbox{single-dose immunization, β1=0.9β, together with a 40% reduction in fatality,11,12f=0.6×102=0.6f, and all other parameters as in Table I.

The low single-dose efficacy domain is characterized by the occurrence of additional prime-boost preference regions in the parameter space [red-shaded regions in Figs. 2(k) and 2(l)]. However, even if the fatality rates of prime-first and prime-boost differ substantially, f/f0.8, only for low values of the relative prime-first efficacy RE=β2/β10.1, preference for prime-boost is observed [Fig. 2(k), shown range 0β2/β10.1]. Given this range and current data on SARS-CoV-2,8,11,12,23 the diagram suggests preference for prime-first vaccination campaigns. Regarding the waning-rate difference, η1η2, a low single-dose efficacy does not suggest a threshold lower than 0.017day1 for prime-first preference [Fig. 2(l)].

Figure 2(m) shows the dependence of RFC-δ and RFC-Δ on β1/β and R0. Values of β1/β1 indicate a very low single-dose efficacy, whereas β1/β0 indicates an unrealistically high efficacy. For large R0 and very low single-dose efficacies, β1/β0.8, prime-boost is preferred over prime-first [Fig. 2(m)]. These parameters are, however, unlikely to be characteristic of SARS-CoV-2,8,23 recent data on the Delta variant included.12 

Finally, the waning-rate threshold η1η2=0.017day1 robustly separates prime-first and prime-boost preference regions for varying natural immunity-waning rates and empirical vaccination time series data (see  Appendices I,  F, and  G). The waning-rate threshold below which preference for prime-first is observed depends only weakly on the initial infection prevalence: it slightly increases as I(0) decreases, η1η20.010 to 0.017day1 for I(0)=105102 (see  Appendix J).

The presented campaign-preference diagrams are two-dimensional projections of a 25-dimensional parameter space, with the majority of parameters kept arbitrarily fixed (Table I). Due to this clear limitation, we further examine if the prime-first vaccination scheme is supported by other mathematical methods.

Our previous results suggest a pronounced preference for prime-first vaccination for a wide range of key epidemiological parameters. To further substantiate this conclusion, we performed Monte Carlo sampling of the entire 25-dimensional parameter space (see  Appendix E). For the analyzed high-dimensional parameter space, our results support that preference for prime-boost occurs significantly less frequently than samples indicating an advantage of prime-first vaccination.

The relative frequencies of samples for which prime-boost vaccination outperforms prime-first vaccination, characterized by RFC-δ<0 and RFC-Δ<0, are estimated as 7.9% [standard error (SE): 0.2%] and 23.2% (SE: 0.4%); see orange bars in Fig. 3(a). For waning-rate differences η1η20.056day1 and vaccination rates νmax0.047day1, we find that the proportions of prime-boost-preference samples are 7.0% (SE: 0.2%) for RFC-δ<0 and 15.4% (SE: 0.3%) for RFC-Δ<0 [beige bars in Fig. 3(a)]. Further restricting the parameter space using the condition η1η2<0.017day1 (dashed black lines in Fig. 2) and reported vaccination rates νmax<0.013day14 lead to proportions of prime-boost-preference samples of 8.5% (SE: 0.2%) for RFC-δ<0 and 6.9% (SE: 0.2%) for RFC-Δ<0 [blue bars in Fig. 3(a)]. The corresponding distributions of the differences in the maximum change of fatalities, d2d1, and in the cumulative numbers of deaths, D2D1, are shown in Figs. 3(b) and 3(c). As indicated by the mean relative frequencies of prime-boost-preference samples in Fig. 3(a), the majority of samples satisfy RFC-δ>0 (i.e., d2d1>0) and RFC-Δ>0 (i.e., D2D1>0).

FIG. 3.

Monte Carlo sampling of entire and restricted high-dimensional parameter spaces. Unconditioned data (orange) are compared with two conditioned datasets: (i) data shown in beige: νmax0.047day1 and η1η20.056day1 and (ii) data shown in blue: νmax0.013day1 and η1η20.017day1. Thresholds (i) are inferred from our decision tree analysis (see  Appendix K). (a) Relative frequency of prime-boost-preference samples (RFC-δ<0, RFC-Δ<0) for the three datasets. Error bars are below 0.2% (not shown). (b) Probability density function (PDF) of the difference between death rates d2 (prime-boost) and d1 (prime-first). For the conditioned data, averages of d1d2 are about 3×105day1 (blue curve) and about 5×106day1 (beige curve), larger than the mean 4×106day1 of the unconditioned data. (c) PDF of the difference between the total number of deaths D2 (prime-boost) and D1 (prime-first). For the conditioned data, the means of D1D2 are about 2×103 (blue curve) and about 9×104 (beige curve), larger than the mean 7×104 of the unconditioned data. The presented data are based on 5×104 (blue and orange curves) and about 4.4×104 (beige curves) samples of the entire 25-dimensional parameter space. (d)–(f) Relative frequency of prime-boost-preference samples (RFC-δ<0, RFC-Δ<0) for additional datasets with the same constraints as used in (a). Error bars are below 0.5% (not shown). (d) We sampled natural immunity-waning rates η3 from the distribution U(0,0.1)day1. All remaining parameters are specified in Table II. (e) Fatality rates are fU(104,103), fU(104,f), and f=f. Natural immunity-waning rate η3 is sampled from U(0,0.1)day1. All remaining parameters are specified in Table II. (f) Fatality rates are fU(103,102), fU(103,f), and f=f. Natural immunity-waning rate η3 is sampled from U(0,0.1)day1. All remaining parameters are specified in Table II.

FIG. 3.

Monte Carlo sampling of entire and restricted high-dimensional parameter spaces. Unconditioned data (orange) are compared with two conditioned datasets: (i) data shown in beige: νmax0.047day1 and η1η20.056day1 and (ii) data shown in blue: νmax0.013day1 and η1η20.017day1. Thresholds (i) are inferred from our decision tree analysis (see  Appendix K). (a) Relative frequency of prime-boost-preference samples (RFC-δ<0, RFC-Δ<0) for the three datasets. Error bars are below 0.2% (not shown). (b) Probability density function (PDF) of the difference between death rates d2 (prime-boost) and d1 (prime-first). For the conditioned data, averages of d1d2 are about 3×105day1 (blue curve) and about 5×106day1 (beige curve), larger than the mean 4×106day1 of the unconditioned data. (c) PDF of the difference between the total number of deaths D2 (prime-boost) and D1 (prime-first). For the conditioned data, the means of D1D2 are about 2×103 (blue curve) and about 9×104 (beige curve), larger than the mean 7×104 of the unconditioned data. The presented data are based on 5×104 (blue and orange curves) and about 4.4×104 (beige curves) samples of the entire 25-dimensional parameter space. (d)–(f) Relative frequency of prime-boost-preference samples (RFC-δ<0, RFC-Δ<0) for additional datasets with the same constraints as used in (a). Error bars are below 0.5% (not shown). (d) We sampled natural immunity-waning rates η3 from the distribution U(0,0.1)day1. All remaining parameters are specified in Table II. (e) Fatality rates are fU(104,103), fU(104,f), and f=f. Natural immunity-waning rate η3 is sampled from U(0,0.1)day1. All remaining parameters are specified in Table II. (f) Fatality rates are fU(103,102), fU(103,f), and f=f. Natural immunity-waning rate η3 is sampled from U(0,0.1)day1. All remaining parameters are specified in Table II.

Close modal

Based on the results shown in Figs. 3(a)3(c), we conclude that constraining the studied parameter space by lowering νmax and η1η2 results in a substantially enhanced preference for prime-first in terms of reduced excess deaths, RFC-Δ. We also conclude that the preference of prime-first vaccination campaigns with respect to the fatality measure RFC-δ is almost independent of the chosen parameter restrictions, which is indicated by the observed narrow range between 7% and 9% in Fig. 3(a). This supports the robustness of our results.

The discriminative power of the thresholds associated with the parameters η1η2 and νmax is also supported by random-sampling results for different risk groups and in situations with natural immunity waning, as seen in Figs. 3(d)3(f). As detailed in  Appendix H, simulations were performed for two age groups42 with infection fatality ratios (IFRs) of less than 0.1% and between 0.1% and 1%, respectively. We observe slightly smaller proportions of prime-boost preference samples in the older age group than in the younger age group.

As another independent method for determining decisive conditions for strategic vaccination campaigns, we performed binary decision tree learning with repeated stratified cross-validation.43,44 This technique has proven useful to extract the most discriminative features in high-dimensional data. Our analysis suggests that νmax and η1η2 are the most discriminative parameters within the 25-dimensional parameter space (see  Appendix K). For the samples that we generated according to the distributions listed in Table I (orange lines and markers in Fig. 3) and the constraints νmax0.047day1 and η1η20.056day1, about 70% of vaccination preferences of simulated scenarios are correctly predicted (see  Appendix K for accuracy scores and details). Additionally, constraining the parameter space with the thresholds that we used in the previous paragraph (beige and blue lines and markers in Fig. 3) results in prime-first preference for 93% of the parameter-space volume. This suggests that for realistic vaccination rates, the vaccination-dose-dependent immunity-waning rate difference is the only highly discriminative factor.

Effective vaccination protocols are crucial to achieve a high immunization coverage, especially if vaccination supplies are limited. The ongoing debate on the most effective way of distributing prime-boost regimens against SARS-CoV-2 has been sparked by arguments suggesting that, from an epidemiological perspective, prime-first vaccination protocols may be more effective than immediate prime-boost administration if supply cannot meet demand.18,26,28,45,46 For many COVID-19 vaccines, prime-boost protocols are considered immunologically efficient due to their ability to elicit strong and long-term humoral and cellular immune responses.21 Immunologically efficient vaccination protocols, however, may be not epidemiologically favorable,47 in particular, for exponentially increasing infection numbers and vaccine doses shortages on times scales of months. By combining epidemiological modeling, methods from statistical mechanics, and decision tree learning, we have studied the effect of relevant immunological and epidemiological parameters (e.g., vaccine efficacy and immunity waning) on a possible advantage of prime-first over prime-boost vaccination. We have identified and analyzed decision boundaries separating the parameter regimes in which one or the other vaccination protocol is preferable. Our results suggest that prime-first campaigns are associated with a lower death toll compared to prime-boost vaccination campaigns even for relatively high vaccination rates and more surprisingly for low single-dose efficacies, which is in contrast to the existing literature.8,23,45,48

A related study45 compares prime-first and prime-boost vaccination campaigns against SARS-CoV-2, without accounting for immunity waning. This study reports that single-dose vaccination campaigns make optimal use of resources in the short term, given a sufficiently large single-dose efficacy that they identify as the main discriminative factor. In contrast, our study calls attention to immunity waning and vaccination rates as the highly discriminative factors, while we find that vaccine efficacies are less discriminative.

Previous works15,48 consistently emphasize that due to the complexity of underlying models and limitations from available data, a vaccine campaign recommendation can only be given once the precision in all key epidemiological parameters, such as field efficacy data and the level of protection obtained after a single dose, becomes sufficiently high. Laubenbacher et al.49 highlight the need for further data collection and model integration in infectious disease modeling, which are important steps to better estimate immunity-waning rates and vaccine effectiveness.8,15,48

Saad-Roy et al.50 focus on the long-term effects of waning and evolutionary immune response in a highly parameterized model. Certain scenarios they analyze suggest that prime-first campaigns may be favorable for some time scales but not for others, depending on a combination of parameters, waning rates included. In contrast, for the critical time scale of months, we provide a simple decision criterion by showing that the waning-rate difference is the main discriminative threshold.

Preference for prime-first vaccination is not unexpected. During the initial inter-dose interval, both vaccination strategies are identical since, regardless of the chosen strategy, booster doses are not yet administered. In the subsequent time interval, twice as many susceptible individuals can be immunized by following a prime-first protocol compared to prime-boost campaigns. This means that about 50% of individuals who could have been vaccinated will actually remain unvaccinated in a prime-boost campaign. Let us refer to this unvaccinated group as group A and denote by group B the group of individuals that receive two vaccination doses in a prime-boost scenario. One can assume that the infection rates of individuals in group A are larger than those of individuals in group B, who benefit from a more effective immune response. As a result, higher transmission in group A is the expected dominating differential adverse effect. As expected for effective prime-boost vaccines, one may assume that the prime-boost infection rate, β2, and the fatality rate, f, are significantly smaller than their counterparts in prime-first campaigns. Thus, the effective transmission rate for group A and B combined is dominated by group A’s rate but not critically dependent on β2, which intuitively explains why the relative efficacy ratio RE=β2/β1 is not a highly discriminative factor.

For very low single-dose efficacies or very high single-dose disease-induced fatality rates, the single-dose efficacy β1/β, the relative prime-first efficacy RE, and R0 may be discriminative, depending on the circumstances. Current data on SARS-CoV-2 vaccination campaigns,8,23 however, suggest that those parameter combinations are unlikely to occur. Furthermore, if immunity wanes substantially faster after the first dose than after the additional booster dose, prime-boost vaccination may become favorable over prime-first, depending on the value of R0. Unvaccinated and susceptible individuals should also receive both vaccination doses if a few percent of a jurisdiction’s total population can be vaccinated daily. However, even for the relatively large vaccination rate of about 1% per day, as realized in Israel,4 our analyses suggest that prime-first vaccination is still favorable over prime-boost campaigns.

A recent study in single-dose vaccinated SARS-CoV-2 patients infected with Beta or Delta variants showed neutralizing antibody concentrations well below the limit of detection.51 While antibody titers correlate with protection against severe disease,52 they are only a single component of the intricate immune response mechanisms and are not a necessary condition for effectiveness of vaccines against symptomatic disease or hospitalization. In fact, recent effectiveness estimates suggest that the first dose of BioNTech–Pfizer and Oxford–AstraZeneca is about 78% and 75% effective against hospitalization after an infection with the Alpha and Delta variants, respectively.12,13 These data strongly support our conclusions given the high correlation between hospitalization and fatality.

The effectiveness estimates for Alpha and Delta variants as reported in Refs. 12 and 13 are as low as 31% for single-dose immunization when compared to the values of about 80% that are reported for prime-boost vaccinated individuals. While efficacies below 50% may strongly suggest that the delay between the first and second dose should be kept limited, our study finds that prime-first vaccination should be considered if the primary health objective is minimizing hospitalizations and fatalities. This means that vaccination campaigns that deviate from the recommended immunization protocol are particularly relevant in countries where new variants, such as the Delta and Gamma strains, are highly prevalent.19,53,54

In summary, our results contrast the existing literature15,45,48,50 in the sense that not all key epidemiological data are required to be collected to identify most effective vaccination protocols. On the contrary, our analysis suggests that, even for a large degree of uncertainty in key epidemiological data, prime-first vaccination is robustly preferred over prime-boost vaccination—if the waning-rate difference between prime-first and prime-boost is sufficiently small. For epidemiological parameters that describe the spread of COVID-19, we found this threshold to be in the narrow range of 0.01–0.02 day1. Unfortunately, to date, there are no reliable data available on waning time scales.24,38 While some estimates suggest that no significant waning occurs for Oxford–AstraZeneca in the first 90 days after receiving the first dose,23 recent estimates from Israel may suggest substantial immunity waning regarding protection from infection after 6 months for BioNTech–Pfizer prime-boost regimens.55 For a characteristic prime-boost immunity-waning time scale of about 6 months, the waning immunity threshold condition suggests that prime-first protocols are more effective than prime-boost protocols if immunity waning of prime-vaccinated individuals is not expected to occur already within the first two months. Once vaccination-dependent waning rates can be estimated from data25 and adverse immunological effects can be assessed or excluded, our criterion may become highly valuable for decision makers in countries facing vaccine shortages.

Although clinical studies of the approved SARS-CoV-2 vaccines may suggest that these vaccines are safe and effective, only little is known about their possible long-term adverse effects.56 Clearly, for the comparison of different vaccination strategies, we assume that negative long-term effects are negligible. In addition, we do not consider harm measures covering non-hospitalized symptomatic cases. Adverse effects and different levels of protection may be incorporated in models that account for different subgroups.57 

To conclude, while current vaccine supply does not meet global demand, especially in low- and middle-income countries, and new variants of concern may reduce the effectiveness of currently available vaccines,58 it is desirable to provide decision makers with transparent tools that support them in assessing different vaccination protocols. This study may be of help to healthcare officials and decision makers since, in contrast to the existing literature, our combination of tools results in unexpectedly robust and highly decisive criteria. More generally, the presented framework establishes how epidemiologically efficient vaccine dosing strategies59,60 can be integrated into effective pandemic control plans.

L. B. acknowledges financial support from the SNF (No. P2EZP2_191888), the National Institutes of Health (NIH) (No. R01HL146552), and the Army Research Office (ARO) (No. W911NF-18-1-0345). Parts of the simulations were performed on the ETH Euler cluster.

The authors declare no competing interests.

The data that support the findings of this study are openly available in GitHub at https://github.com/lubo93/vaccination, Ref. 67.

We adapt the SEIRD model32,35 to account for immunity waning and a vaccination-induced reduction in transmissibility [Fig. 1]. The fractions of susceptible, exposed, infected, recovered, and deceased individuals at time t are denoted by S(t), E(t), I(t), R(t), and D(t) respectively. Moreover, we denote the fractions of prime and prime-boost vaccinated susceptible individuals by S(t) and S(t), respectively. With rate ν1, susceptible individuals receive their first vaccination dose, and with rate ν2, prime-vaccinated susceptible individuals receive their booster doses. The time dependence in the vaccination rates reflects temporal variations in the availability of vaccination doses, as explained below. The corresponding fractions of vaccinated exposed and infected individuals are denoted by E(t) and E(t) and I(t) and I(t). We use three constant rates η1, η2, η3 to model immunity waning (i.e., transitions from S, S, and R to S). Characteristic time scales of waning immunity,24 defined by the inverse of the corresponding rates, are much longer than those associated with entering and leaving exposed and infected compartments; therefore, we do not explicitly model waning immunity in these compartments. For long time horizons, additional birth and death processes may be employed to model birth and age-related death.

The resulting dynamics of the susceptible and exposed classes is described by the following rate equations:

dSdt=βSIβSIβSIν1+η1S+η2S+η3R,dSdt=ν1β1SIβ1SIβ1SIν2η1S,dSdt=ν2β2SIβ2SIβ2SIη2S,dEdt=βSI+βSI+βSIσE,dEdt=β1SI+β1SI+β1SIσ1E,dEdt=β2SI+β2SI+β2SIσ2E.
(A1)

The maximum proportion of susceptible individuals that can be prime and prime-boost vaccinated is S(t) and S(t), respectively. Based on vaccination data from Israel (Fig. 4), we assume linearly increasing immunization over time in our model and use the vaccination rates

ν1(μ1,μ2,S,S,td,t)=(μ1+μ2)H[S(t)]H[tdt]+μ1H[S(t)]H[ttd]
(A2)

and

ν2(μ1,μ2,S,S,td,t)=μ2H[S(t)]H[ttd]+μ1(1H[S(t)])H[S(t)]H[ttd],
(A3)

where μ1=νmax and μ2=0 for prime-first vaccination and μ1=μ2=νmax/2 for prime-boost vaccination. Here, H[x] denotes the Heaviside step function, which is zero for x<0 and one for x0. The function H[ttd] describes the delay td of about 2–3 weeks5 (Fig. 4) between prime and booster doses. Up to time td, susceptible individuals get vaccinated with rate μ1+μ2. If no susceptible individuals are left, prime-vaccinated individuals get vaccinated with rate μ1 too, leading to the term μ1(1H[S(t)])H[S(t)]H[ttd] in Eq. (A3). In our model, only susceptible individuals are vaccinated. This can be justified by the assumption that susceptible individuals outnumber those in other disease states.

FIG. 4.

Proportions of prime and prime-boost vaccinated individuals in Israel. The proportions of prime (orange) and prime-boost (green) vaccinated individuals in Israel. The first prime dose was administered on December 20, 2020. On January 10, 2021, the first individuals received booster doses. The prime-boost delay td is thus 21 days. The solid and dashed black lines are guides to the eye, representing a maximum daily vaccination rate of νmax=1.3%day1 and νmax/2, respectively. All data are taken from https://github.com/owid/covid-19-data/tree/master/public/data/vaccinations/country_data; accessed May 23, 2021.

FIG. 4.

Proportions of prime and prime-boost vaccinated individuals in Israel. The proportions of prime (orange) and prime-boost (green) vaccinated individuals in Israel. The first prime dose was administered on December 20, 2020. On January 10, 2021, the first individuals received booster doses. The prime-boost delay td is thus 21 days. The solid and dashed black lines are guides to the eye, representing a maximum daily vaccination rate of νmax=1.3%day1 and νmax/2, respectively. All data are taken from https://github.com/owid/covid-19-data/tree/master/public/data/vaccinations/country_data; accessed May 23, 2021.

Close modal
FIG. 5.

Selection of vaccination-campaign-preference diagrams for small incidence rates. For selected combinations of basic reproduction number R0, waning-rate difference η1η2, and maximum vaccination rate νmax, we plot RFC-δ [Eq. (1)] and RFC-Δ [Eq. (2)]. Green-shaded regions indicate preference for prime-first (RFC-δ>0, RFC-Δ>0) and red-shaded regions indicate preference for prime-boost (RFC-δ<0, RFC-Δ<0). In (a) and (b), we set I(0)=105; in (c) and (d), we set I(0)=107. The remaining parameters are as in Table I. Dashed lines are guides to the eye: Thresholds η1η2=0.01day1 and Israel’s vaccination rate as of February 1, 20214 (νmax=0.013day1).

FIG. 5.

Selection of vaccination-campaign-preference diagrams for small incidence rates. For selected combinations of basic reproduction number R0, waning-rate difference η1η2, and maximum vaccination rate νmax, we plot RFC-δ [Eq. (1)] and RFC-Δ [Eq. (2)]. Green-shaded regions indicate preference for prime-first (RFC-δ>0, RFC-Δ>0) and red-shaded regions indicate preference for prime-boost (RFC-δ<0, RFC-Δ<0). In (a) and (b), we set I(0)=105; in (c) and (d), we set I(0)=107. The remaining parameters are as in Table I. Dashed lines are guides to the eye: Thresholds η1η2=0.01day1 and Israel’s vaccination rate as of February 1, 20214 (νmax=0.013day1).

Close modal
FIG. 6.

Empirical vaccination data from Israel does not lower decision threshold, η1η2. For variations in η1η2 and R0, we plot RFC-δ [Eq. (1)] and RFC-Δ [Eq. (2)] for empirical vaccination data (Fig. 4) and two natural immunity-waning rates. In (a), we set η3=6×103day1 (waning within about 6 months); in (b), we set η3=3×103day1 (waning within about 12 months). Green-shaded regions indicate preference for prime-first (RFC-δ>0, RFC-Δ>0), and red-shaded regions indicate preference for prime-boost (RFC-δ<0, RFC-Δ<0). The simulation horizon is T=150 days. The remaining parameters are as in Table I. Dashed lines are guides to the eye of the original threshold η1η2=0.017day1. We observe preference for prime-first over prime-boost vaccination even for faster waning, η1η2>0.017day1, compared to other epidemiological scenarios with a constant and, on average, a larger vaccination rate.

FIG. 6.

Empirical vaccination data from Israel does not lower decision threshold, η1η2. For variations in η1η2 and R0, we plot RFC-δ [Eq. (1)] and RFC-Δ [Eq. (2)] for empirical vaccination data (Fig. 4) and two natural immunity-waning rates. In (a), we set η3=6×103day1 (waning within about 6 months); in (b), we set η3=3×103day1 (waning within about 12 months). Green-shaded regions indicate preference for prime-first (RFC-δ>0, RFC-Δ>0), and red-shaded regions indicate preference for prime-boost (RFC-δ<0, RFC-Δ<0). The simulation horizon is T=150 days. The remaining parameters are as in Table I. Dashed lines are guides to the eye of the original threshold η1η2=0.017day1. We observe preference for prime-first over prime-boost vaccination even for faster waning, η1η2>0.017day1, compared to other epidemiological scenarios with a constant and, on average, a larger vaccination rate.

Close modal

Exposed individuals transition to an infected state at rates σ, σ1, and σ2. The evolution of the infected, recovered, and deceased compartments is described by

dIdt=σEγI,dIdt=σ1EγI,dIdt=σ2EγI,dRdt=γ(1f)I+γ(1f)I+γ(1f)Iη3R,dDdt=γfI+γfI+γfI.
(A4)

Only ten of equations (A1) and (A4) are independent since we employ the normalization condition S+S+S+E+E+E+I+I+I+R+D=1. Different transmissibilities β,β,β, β1,β1,β1, and β2,β2,β2 describe interactions between susceptible and infected individuals with different immunity levels.

For each infected compartment I, I, and I, we calculate the infection fatality ratios (IFRs)3 by dividing the associated cumulative number of deaths by the total number of infections in the unvaccinated, prime-vaccinated, and prime-boost-vaccinated compartments, respectively. The IFR of the unvaccinated pool of individuals is

IFR(t)=0tγfI(t)dtI(t)+0tγfI(t)dt+0tγ(1f)I(t)dt.
(A5)

For constant γ,f, we obtain

IFR(t)=γf0tI(t)dtI(t)+γ0tI(t)dt.
(A6)

As the number of infected individuals approaches zero for long time horizons [i.e., limtI(t)=0], the IFR satisfies limtIFR(t)=f. Similarly, limtIFR(t)=f and limtIFR(t)=f if γ,f and γ,f are time-independent.

Due to ergodicity breaking effects from multiplicative noise,61 deterministic models tend to overestimate infection and fatality. However, it is realistic to assume that the effects from noise are not discriminative as they do not differ for either vaccination campaign.

After all, the immunological intricacies of SARS-CoV-2 remain largely unknown, and there is no single commonly accepted epidemiological standard model. At the same time, we anticipate more reliable data on immunity waning and other immunological effects in the coming months. Our framework is transparent and flexible enough to change or augment the (already high) degree of parameterization, or compartmentalization, if warranted.

We calculate the basic reproduction number R0 of the epidemic model (A1) and (A4) using the next-generation matrix method.64 As a first step, we rewrite the rate equations (A1) and (A4) of the infected compartments in a matrix form,

x˙(t)=F(x,y)V(x,y),
(B1)

where x=(E,E,E,I,I,I), y=(S,S,S,R,D), F represents the vector of new infections, and V describes all remaining transitions. We thus find for the corresponding Jacobians of F and V at the disease-free equilibrium,

F=(000βSβSβS000β1Sβ1Sβ1S000β2Sβ2Sβ2S000000000000000000)

and

V=(σ000000σ1000000σ2000σ00γ000σ100γ000σ200γ).
(B2)

The basic reproduction number R0, the expected number of infections generated by an infectious individual in an otherwise completely susceptible population, is the spectral radius of the next-generation matrix FV1.64 Finding R0 for the general system (B2)

involves the analytically cumbersome task of finding roots of a cubic equation, which can be avoided by using numerical methods (e.g., the power method). For very effective vaccines, however, one may assume that the transmissibility of prime-boost vaccinated individuals is much lower than the transmissibility of unvaccinated individuals; that is, ββ, β1β, and β2β. In this approximation, we obtain

R0=βSγ+β1Sγ+β2γ2S2+2γγSS(2β1βββ1)+γ2β12S22γγ.
(B3)

For S(0)=1 and S(0)=0, the basic reproduction number is

R0=βγ.
(B4)

To solve Eqs. (A1) and (A4) numerically, we use the Dormand–Prince method65 with a maximum time step of 101 and simulate the evolution of different epidemics in the time interval [0,T], where T=300 days. For the simulation results that we show in Fig. 2, we set I(0)=102 and S(0)=1I(0). If model parameters are held constant in Fig. 2, we use the parameters that are listed in Table I. Transmissibilities of infection events that involve at least one vaccinated individual are smaller than or equal to the baseline transmissibility β as long as mobility and distancing characteristics of vaccinated individuals do not differ significantly from those who are unvaccinated. In our model, this means that βββ, ββ1β2, β1β1β1, and β2β2β2 (equality holds for very ineffective vaccines). As vaccination campaigns and vaccine effectiveness analyses are ongoing, we used estimates for β, β, β1, β1, β1, β2, β2, and β2 as reported in Table I. We also model the effect of small incidence rates and broader parameter ranges in a random-sampling analysis as reported in the following sections.

There are two more constraints that our model parameters have to satisfy to describe the impact of vaccination campaigns on disease transmission. First, the fatality ratio in the unvaccinated compartment is larger than the fatality ratios in the vaccinated compartments (i.e., fff). We assume that differences in f and f are negligible. Second, the waning rate in the prime-boost vaccinated compartment is smaller than the waning rate in the prime-vaccinated compartment (i.e., η2η1).

The fatality measures RFC-δ and RFC-Δ as defined in Eqs. (1) and (2) are correlated, [Figs. 8(a)8(c), R=0.94 (a), R=0.68 (b), and R=0.68 (c); corresponding p-values are smaller than machine precision]. The correlation observed for the threshold combination νmax0.013day1 and η1η20.017day1 confirms the discriminative power and robustness of our results regarding the choice of both fatality measures.

The parameter distributions that we use in our random sampling and decision tree analysis are summarized in Table II. We generate two datasets with N=50000 samples each and analyze the influence of different combinations of model parameters and initial conditions on RFC-δ(d1,d2) [Eq. (1)] and RFC-Δ(D1,D2) [Eq. (2)].

TABLE II.

Sampling distributions with and without natural immunity waning. The listed parameter values and distributions are used in our random-sampling analysis. As initial fraction of susceptible individuals, we use S(0)=1I(0)S(0)S(0). We set E(0) = 0, E(0)=0, E(0)=0, I(0)=0, I(0)=0, R(0) = 0, D(0) = 0. The minimum and maximum values of β are βmin = γ and βmax = 4γ, respectively. A uniform distribution with boundaries a and b is indicated by U(a,b).

ParameterSymbolValue/distributionUnits
Infection rates S β, β, β U(βmin,βmax), U(0,β), U(0,β) day−1 
Infection rates S β1, β1, β1 U(0,β), U(0,β1), U(0,β1) day−1 
Infection rates S β2, β2, β2 U(0,β1), U(0,β1), U(0,β1) day−1 
Incubation rate E σ U(0.2,0.5) day−1 
Incubation rate E σ1 U(0.2,0.5) day−1 
Incubation rate E σ2 U(0.2,0.5) day−1 
Vaccination rate νmax U(0,0.02) day−1 
Waning rate (prime) η1 U(0,0.1) day−1 
Waning rate (prime-boost) η2 U(0,η1) day−1 
Waning rate (recovered) η3 0 (scenarios without natural immunity waning) day−1 
  U(0,0.1) (scenarios with natural immunity waning)  
Resolution rate I γ 1/14 day−1 
Resolution rate I γ U(γ,2γ) day−1 
Resolution rate I γ U(γ,2γ) day−1 
Fatality ratio I f U(103,101) … 
Fatality ratio I f U(103,f) … 
Fatality ratio I f U(103,f) … 
Prime-boost delay td U(7,35) day 
Initially infected individuals I(0) U(104,3×101) … 
Initially prime-vaccinated individuals S(0) U(104,101) … 
Initially prime-boost vaccinated individuals S(0) U(104,101) … 
ParameterSymbolValue/distributionUnits
Infection rates S β, β, β U(βmin,βmax), U(0,β), U(0,β) day−1 
Infection rates S β1, β1, β1 U(0,β), U(0,β1), U(0,β1) day−1 
Infection rates S β2, β2, β2 U(0,β1), U(0,β1), U(0,β1) day−1 
Incubation rate E σ U(0.2,0.5) day−1 
Incubation rate E σ1 U(0.2,0.5) day−1 
Incubation rate E σ2 U(0.2,0.5) day−1 
Vaccination rate νmax U(0,0.02) day−1 
Waning rate (prime) η1 U(0,0.1) day−1 
Waning rate (prime-boost) η2 U(0,η1) day−1 
Waning rate (recovered) η3 0 (scenarios without natural immunity waning) day−1 
  U(0,0.1) (scenarios with natural immunity waning)  
Resolution rate I γ 1/14 day−1 
Resolution rate I γ U(γ,2γ) day−1 
Resolution rate I γ U(γ,2γ) day−1 
Fatality ratio I f U(103,101) … 
Fatality ratio I f U(103,f) … 
Fatality ratio I f U(103,f) … 
Prime-boost delay td U(7,35) day 
Initially infected individuals I(0) U(104,3×101) … 
Initially prime-vaccinated individuals S(0) U(104,101) … 
Initially prime-boost vaccinated individuals S(0) U(104,101) … 

For the critical time horizon of a few months that we consider in the main text, for scenarios (referred to as datasets) A and B, we assume robust natural (T cell) immunity in accordance with corresponding clinical data.66Figure 7 shows vaccination-campaign-preference diagrams for η1η2 vs R0 and for natural immunity-waning time scales of about 6 and 12 months. We observe that these variations in η3 do not change the campaign-preference threshold (dashed black line in Fig. 7).

FIG. 7.

Natural immunity waning does not lower decision threshold, η1η2. For variations in η1η2 and R0, we plot RFC-δ [Eq. (1)] and RFC-Δ [Eq. (2)] for two natural immunity-waning rates. In (a), we set η3=6×103day1 (waning within about 6 months); in (b), we set η3=3×103day1 (waning within about 12 months). Green-shaded regions indicate preference for prime-first (RFC-δ>0, RFC-Δ>0), and red-shaded regions indicate preference for prime-boost (RFC-δ<0, RFC-Δ<0). The remaining parameters are as in Table I. Dashed lines are guides to the eye of the original, unchanged threshold η1η2=0.017day1.

FIG. 7.

Natural immunity waning does not lower decision threshold, η1η2. For variations in η1η2 and R0, we plot RFC-δ [Eq. (1)] and RFC-Δ [Eq. (2)] for two natural immunity-waning rates. In (a), we set η3=6×103day1 (waning within about 6 months); in (b), we set η3=3×103day1 (waning within about 12 months). Green-shaded regions indicate preference for prime-first (RFC-δ>0, RFC-Δ>0), and red-shaded regions indicate preference for prime-boost (RFC-δ<0, RFC-Δ<0). The remaining parameters are as in Table I. Dashed lines are guides to the eye of the original, unchanged threshold η1η2=0.017day1.

Close modal
FIG. 8.

Fatality measure correlations for entire and restricted high-dimensional parameter spaces. Correlation plots of RFC-δ and RFC-Δ. Unconditioned data (orange) are compared with two conditioned datasets: (i) data shown in beige: νmax0.047day1 and η1η20.056day1 and (ii) data shown in blue: νmax0.013day1 and η1η20.017day1. Thresholds (i) are inferred from our decision tree analysis. The presented data are based on 5×104 (blue and orange curves) and about 4.4×104 (beige curves) samples of the entire 25-dimensional parameter space. Parameters are as listed in Table II (no natural immunity waning).

FIG. 8.

Fatality measure correlations for entire and restricted high-dimensional parameter spaces. Correlation plots of RFC-δ and RFC-Δ. Unconditioned data (orange) are compared with two conditioned datasets: (i) data shown in beige: νmax0.047day1 and η1η20.056day1 and (ii) data shown in blue: νmax0.013day1 and η1η20.017day1. Thresholds (i) are inferred from our decision tree analysis. The presented data are based on 5×104 (blue and orange curves) and about 4.4×104 (beige curves) samples of the entire 25-dimensional parameter space. Parameters are as listed in Table II (no natural immunity waning).

Close modal

We further analyze the effect of natural immunity waning by sampling η3 from the uniform distribution U(0,0.1)day1. Figure 9 shows the corresponding distributions and correlation plots associated with the fatality measures RFC-δ and RFC-Δ.

FIG. 9.

Monte Carlo sampling of entire and restricted high-dimensional parameter spaces with natural immunity waning. Unconditioned data (orange) are compared with two conditioned datasets: (i) data shown in beige: νmax0.047day1 and η1η20.056day1 and (ii) data shown in blue: νmax0.013day1 and η1η20.017day1. Thresholds (i) are inferred from our decision tree analysis. We sampled natural immunity-waning rates η3 from the distribution U(0,0.1)day1. All remaining parameters are specified in Table II. (a) Relative frequency of prime-boost-preference samples (RFC-δ<0, RFC-Δ<0) for the three datasets [similar to Fig. 3(a) of the main text]. Error bars are below 0.5% (not shown). (b) Probability density function (PDF) of the difference between death rates d2 (prime-boost) and d1 (prime-first). For the conditioned data, averages of d1d2 are about 1.0×106day1 (blue curve) and about 8.3×107day1 (beige curve), larger than the mean 6.8×107day1 of the unconditioned data. (c) PDF of the difference between the total number of deaths D2 (prime-boost) and D1 (prime-first). For the conditioned data, the means of D1D2 are about 7.3×103 (blue curve) and about 2.0×103 (beige curve), larger than the mean 1.5×103 of the unconditioned data.

FIG. 9.

Monte Carlo sampling of entire and restricted high-dimensional parameter spaces with natural immunity waning. Unconditioned data (orange) are compared with two conditioned datasets: (i) data shown in beige: νmax0.047day1 and η1η20.056day1 and (ii) data shown in blue: νmax0.013day1 and η1η20.017day1. Thresholds (i) are inferred from our decision tree analysis. We sampled natural immunity-waning rates η3 from the distribution U(0,0.1)day1. All remaining parameters are specified in Table II. (a) Relative frequency of prime-boost-preference samples (RFC-δ<0, RFC-Δ<0) for the three datasets [similar to Fig. 3(a) of the main text]. Error bars are below 0.5% (not shown). (b) Probability density function (PDF) of the difference between death rates d2 (prime-boost) and d1 (prime-first). For the conditioned data, averages of d1d2 are about 1.0×106day1 (blue curve) and about 8.3×107day1 (beige curve), larger than the mean 6.8×107day1 of the unconditioned data. (c) PDF of the difference between the total number of deaths D2 (prime-boost) and D1 (prime-first). For the conditioned data, the means of D1D2 are about 7.3×103 (blue curve) and about 2.0×103 (beige curve), larger than the mean 1.5×103 of the unconditioned data.

Close modal

The relative frequencies of prime-boost-preference samples, characterized by RFC-δ<0 and RFC-Δ<0, are estimated as 10.4% (SE: 0.3%) and 41.2% (SE: 0.4%); see orange bars in Fig. 9(a). For waning-rate differences η1η20.056day1 and vaccination rates νmax0.047day1, the proportions of prime-boost-preference samples are 8.9% (SE: 0.3%) for RFC-δ<0 and 33.7% (SE: 0.4%) for RFC-Δ<0 [beige bars in Fig. 9(a)]. Using the condition η1η2<0.017day1 (dashed black lines in Fig. 2) and vaccination rates νmax<0.013day14 lead to proportions of prime-boost-preference samples of 6.6% (SE: 0.2%) for RFC-δ<0 and 9.3% (SE: 0.3%) for RFC-Δ<0 [blue bars in Fig. 9(a)].

As in the main text, we find that constraining the studied parameter space by lowering νmax and η1η2 yields a substantial increase in prime-first-preference samples that are associated with fewer total fatalities, as quantified by RFC-Δ. The proportions of prime-boost preference samples with RFC-δ<0 fall into the narrow rage between 7% and 10% and are less affected by the chosen parameter restrictions [Fig. 9(a)]. This again supports the robustness of our results. For randomly sampled parameters that account for different natural immunity-waning rates, prime-first vaccination is preferred regarding RFC-δ.

To study the effect age-related fatality rates, we perform a random-sampling analysis for two age groups. In the first group, we set fU(104,103), fU(104,f), f=f. Fatality rates f of less than 0.1% have been observed for individuals younger than 40 years.42 In the second group, we set fU(103,102), fU(103,f), f=f. Fatality rates of about 0.1%–1% have been reported for individuals with an age between 40 and 70 years.42 As in the previous section, we also account for natural immunity waning by setting η3U(0,0.1)day1. All remaining parameters are specified in Table II.

Figure 10 shows different distributions and correlation plots associated with the fatality measures RFC-δ and RFC-Δ for both age groups. The shown results are in agreement with those reported in the previous section and the main text. Prime-boost-preference samples (i.e., those samples with RFC-δ<0 and RFC-Δ<0) occur less frequently than prime-first-preference samples in both age groups. The conditions η1η2<0.017day1 (dashed black lines in Fig. 2) and νmax<0.013day14 again lead to significantly reduced proportions of prime-boost-preference samples (blue bars and curves in Fig. 10), supporting the validity of these decisive thresholds.

FIG. 10.

Monte Carlo sampling of entire and restricted high-dimensional parameter spaces with age-stratification and natural immunity waning. Unconditioned data (orange) are compared with two conditioned datasets: (i) data shown in beige: νmax0.047day1 and η1η20.056day1 and (ii) data shown in blue: νmax0.013day1 and η1η20.017day1. Thresholds (i) are inferred from our decision tree analysis. (a)–(c) Fatality rates are fU(104,103), fU(104,f), and f=f. (d)–(f) Fatality rates are fU(103,102), fU(103,f), and f=f. Natural immunity-waning rate η3 is sampled from U(0,0.1)day1. All remaining parameters are specified in Table II. (a) and (d) Same plots as in Fig. 3 of the main text. Relative frequency of prime-boost-preference samples (RFC-δ<0, RFC-Δ<0) for the three datasets. Error bars are below 0.5% (not shown). (b) and (e) Probability density function (PDF) of the difference between death rates d2 (prime-boost) and d1 (prime-first). For the conditioned data in (b), averages of d1d2 are about 1.2×108day1 (blue curve) and about 8.3×109day1 (beige curve), larger than the mean 6.1×109day1 of the unconditioned data. For the conditioned data in (e), averages of d1d2 are about 1.1×107day1 (blue curve) and about 8.7×108day1 (beige curve), larger than the mean 6.6×108day1 of the unconditioned data. (c) and (f) PDF of the difference between the total number of deaths D2 (prime-boost) and D1 (prime-first). For the conditioned data in (c), the means of D1D2 are about 9.1×105 (blue curve) and about 1.6×105 (beige curve), larger than the mean 3.2×105 of the unconditioned data. For the conditioned data in (f), the means of D1D2 are about 8.9×104 (blue curve) and about 1.7×104 (beige curve), larger than the mean 2.7×104 of the unconditioned data.

FIG. 10.

Monte Carlo sampling of entire and restricted high-dimensional parameter spaces with age-stratification and natural immunity waning. Unconditioned data (orange) are compared with two conditioned datasets: (i) data shown in beige: νmax0.047day1 and η1η20.056day1 and (ii) data shown in blue: νmax0.013day1 and η1η20.017day1. Thresholds (i) are inferred from our decision tree analysis. (a)–(c) Fatality rates are fU(104,103), fU(104,f), and f=f. (d)–(f) Fatality rates are fU(103,102), fU(103,f), and f=f. Natural immunity-waning rate η3 is sampled from U(0,0.1)day1. All remaining parameters are specified in Table II. (a) and (d) Same plots as in Fig. 3 of the main text. Relative frequency of prime-boost-preference samples (RFC-δ<0, RFC-Δ<0) for the three datasets. Error bars are below 0.5% (not shown). (b) and (e) Probability density function (PDF) of the difference between death rates d2 (prime-boost) and d1 (prime-first). For the conditioned data in (b), averages of d1d2 are about 1.2×108day1 (blue curve) and about 8.3×109day1 (beige curve), larger than the mean 6.1×109day1 of the unconditioned data. For the conditioned data in (e), averages of d1d2 are about 1.1×107day1 (blue curve) and about 8.7×108day1 (beige curve), larger than the mean 6.6×108day1 of the unconditioned data. (c) and (f) PDF of the difference between the total number of deaths D2 (prime-boost) and D1 (prime-first). For the conditioned data in (c), the means of D1D2 are about 9.1×105 (blue curve) and about 1.6×105 (beige curve), larger than the mean 3.2×105 of the unconditioned data. For the conditioned data in (f), the means of D1D2 are about 8.9×104 (blue curve) and about 1.7×104 (beige curve), larger than the mean 2.7×104 of the unconditioned data.

Close modal

We also study the influence of empirical vaccination data (Fig. 4) on the prime-first preference threshold η1η2=0.017day1. Figure 6 shows that preference for prime-first over prime-boost vaccination is given for faster immunity waning than indicated by the threshold η1η2=0.017day1, demonstrating the robustness of the original waning-rate difference threshold.

To study the effect of small incidence rates on the location and extent of prime-first and prime-boost preference regions, we set I(0)=105 and 107, which is three to five orders of magnitude smaller than the value I(0)=102 we used in Fig. 2, and show vaccination preference diagrams for η1η2 vs νmax and η1η2 vs R0 in Fig. 5. We observe that a threshold η1η2=0.01day1 separates prime-first and prime-boost preference regions in both diagrams. This value is smaller than the threshold of η1η2=0.017day1, which we used in Fig. 2.

A binary decision tree consists of a root condition and branches, where the left branch refers to the “yes”-branch while the right branch refers to the “no”-branch.

We employed binary decision tree learning with repeated stratified cross validation (k=5 folds, n=10 repeats). The algorithm RepeatedStratifiedKFold [available in the Python library scikit-learn (https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.RepeatedStratifiedKFold.html), accessed: 02-28-2021] optimizes for split purity using Gini as loss function (split criterion).

Gini impurity is a standard measure in tree learning that quantifies how often a randomly chosen sample from the training dataset would be incorrectly labeled if it was entirely randomly labeled given the distribution of (binary) labels in the subset. In our analysis, labels are “prime-first” and “prime-boost.”

Stratified cross-validation is based on splitting the data into folds such that each fold has the same proportion of observations with a given categorical value. It is particularly useful for imbalanced datasets such as our datasets where there are more prime-first samples than prime-boost samples.

Following standard procedure, we split the dataset (randomly) into training and test datasets, 70% and 30%, respectively. Learning is performed using the training dataset, while the test dataset is cross-validated.

The training accuracy score (for a binary classification task) is defined as the relative number of correctly predicted labels; that is,

Accuracy=TP+TNTP+TN+FP+FN,
(K1)

where TP are true positives, TN are true negatives, FP are false positives, and FN are false negatives.

In our binary classification problem, positives are prime-first labeled samples and negatives are prime-boost labeled samples. Class “prime-boost” is defined by ΔD<0 (red-shaded regions in Fig. 2). Class “prime-first” is defined by ΔD>0 (green-shaded regions in Fig. 2).

An n-times repeated stratified cross-validation is based on the following iteration: (i) Shuffle the test dataset randomly; (ii) split the dataset into k folds; (iii) for each fold, take the fold as a test dataset and take the remaining folds as a training dataset; and (iv) fit the tree on the training dataset and evaluate it on the test dataset.

Accuracy and balanced accuracy are monitored as main cross-validation scores. We also monitored precision, F1-score based metrics, ROC AUC, and recall. Balanced accuracy is warranted for imbalanced datasets and defined as the arithmetic mean of sensitivity=TPTP+FN (true positive rate) and specificity=TNTN+FP (true negative rate).

Dataset A

Dataset A comprises 50 000 randomly sampled data points for parameter ranges and disease stages as described in Table II (without natural immunity waning). In Fig. 3, dataset A is called “unconditioned data” (displayed in orange).

Here, we analyze dataset A; see Fig. 11. Training performance is excellent and reaches 100% for large depths due to overfitting. Learning performance is satisfactory, as seen from similar behaviors for test accuracy and balanced accuracy, and reaches values of about 70% for depth=3. The resulting tree reveals two highly discriminative conditions for prime-boost preference, νmax0.047day1 and η1η20.056day1. This threshold combination is used in Fig. 3, referred to as the conditioned data, displayed in beige.

FIG. 11.

Decision tree analysis for dataset A. (a) Cross-validation: Training accuracy (green curve) reaches 100% for a large tree size. This indicates that fitting is satisfactory, while test training data are overfitted for large tree sizes, as expected. Accuracy (blue curve and shades) and balanced accuracy (orange curve and shades) show a fair and also similar performance for all depths. This indicates satisfactory class prediction. (b) Confusion-matrix: Number of instances for predicted and true (ground truth) labels for prime-first and prime-boost. Upper left: True positives. Upper right: False positives. Lower left: False negatives. Lower right: True negatives. Notation: Positives are prime-first and negatives prime-boost. (c) Decision tree: (Value) denotes the number of samples in class “prime-first” (left part) and class “prime-boost” (right part), respectively, for the given branch (brackets), while samples denote the total sum of samples at the given branch. Left branches satisfy the displayed condition (“yes” branch), and right descendants are “no” branches. Notation: diff\_eta=η1η2, nu\_max=νmax, and I0=I(0). Shown tree: depth=3.

FIG. 11.

Decision tree analysis for dataset A. (a) Cross-validation: Training accuracy (green curve) reaches 100% for a large tree size. This indicates that fitting is satisfactory, while test training data are overfitted for large tree sizes, as expected. Accuracy (blue curve and shades) and balanced accuracy (orange curve and shades) show a fair and also similar performance for all depths. This indicates satisfactory class prediction. (b) Confusion-matrix: Number of instances for predicted and true (ground truth) labels for prime-first and prime-boost. Upper left: True positives. Upper right: False positives. Lower left: False negatives. Lower right: True negatives. Notation: Positives are prime-first and negatives prime-boost. (c) Decision tree: (Value) denotes the number of samples in class “prime-first” (left part) and class “prime-boost” (right part), respectively, for the given branch (brackets), while samples denote the total sum of samples at the given branch. Left branches satisfy the displayed condition (“yes” branch), and right descendants are “no” branches. Notation: diff\_eta=η1η2, nu\_max=νmax, and I0=I(0). Shown tree: depth=3.

Close modal

Dataset B

Here, we study the conditioned data constrained by η1η20.017day1 and νmax<0.013day1, as analyzed in Fig. 3 (blue), here called dataset B. For this dataset, we uniformly sampled initial proportions of infected individuals, I(0), on a logarithmic scale from 107 to 3×101. This way of sampling allows us to study the robustness of decision boundary thresholds for a large range of initial disease prevalences.

Dataset B comprises 50 000 randomly sampled data points (without natural immunity waning) where samples simultaneously satisfy η1η20.017day1 and νmax<0.013day1. Class prime-boost is defined by ΔD<0 (red-shaded regions in Fig. 2). Class prime-first is defined by ΔD>0 (green-shaded regions in Fig. 2).

Results are presented in Fig. 12. The training accuracy curve (green curve) shows high accuracy levels from overfitting of the training set that reaches 100% for large tree depths. The accuracy curve (blue) shows the mean of the cross-validation of the accuracy for the test dataset.

FIG. 12.

Decision tree analysis for dataset B. (a) Cross-validation: Large spread between training accuracy (green curve) and balanced accuracy (orange) for all depths indicate poor learning performance regarding a possible subtree structure of prime-boost samples but excellent class prediction. High values of accuracy result from excellent prediction for prime-first samples (majority 13 724), while prediction of prime-boost is poor (1276 false negatives, 0 true negatives in a confusion matrix). (b) Confusion-matrix: Number of instances for predicted and true (ground truth) labels for prime-first and prime-boost. Upper left: True positives. Upper right: False positives. Lower left: False negatives. Lower right: True negatives. Notation: Positives are prime-first and negatives prime-boost. (c) Decision tree: (Value) denotes the number of samples in class “prime-first” (left part) and class “prime-boost” (right part), respectively, for the given branch (brackets), while samples denote the total sum of samples at the given branch. Left branches satisfy the displayed condition (“yes” branch), and right descendants are “no” branches. Notation: betap=β, diff\_eta=η1η2, beta\_2/beta\_1=β2/β1, I0=I(0). Shown depth=3. The resulting tree of depth=3 is essentially equivalent to always choosing prime-first, leading to an accuracy of 92%. With increasing depth, balanced accuracy (and recall) increase only slightly.

FIG. 12.

Decision tree analysis for dataset B. (a) Cross-validation: Large spread between training accuracy (green curve) and balanced accuracy (orange) for all depths indicate poor learning performance regarding a possible subtree structure of prime-boost samples but excellent class prediction. High values of accuracy result from excellent prediction for prime-first samples (majority 13 724), while prediction of prime-boost is poor (1276 false negatives, 0 true negatives in a confusion matrix). (b) Confusion-matrix: Number of instances for predicted and true (ground truth) labels for prime-first and prime-boost. Upper left: True positives. Upper right: False positives. Lower left: False negatives. Lower right: True negatives. Notation: Positives are prime-first and negatives prime-boost. (c) Decision tree: (Value) denotes the number of samples in class “prime-first” (left part) and class “prime-boost” (right part), respectively, for the given branch (brackets), while samples denote the total sum of samples at the given branch. Left branches satisfy the displayed condition (“yes” branch), and right descendants are “no” branches. Notation: betap=β, diff\_eta=η1η2, beta\_2/beta\_1=β2/β1, I0=I(0). Shown depth=3. The resulting tree of depth=3 is essentially equivalent to always choosing prime-first, leading to an accuracy of 92%. With increasing depth, balanced accuracy (and recall) increase only slightly.

Close modal

The results confirm that no conditions other than the constraints η1η20.017day1 and νmax<0.013day1 robustly characterize prime-first preference domains.

1.
World Health Organization
, “WHO Director-General’s opening remarks at the media briefing on COVID-19—11 March 2020,” see https://www.who.int/director-general/speeches/detail/who-director-general-s-opening-remarks-at-the-media-briefing-on-covid-19—11-march-2020; accessed April 18, 2020.
2.
See https://www.worldometers.info/coronavirus/ for “COVID-19 Statistics” (2020); accessed August 28, 2021.
3.
L.
Böttcher
,
M.
D’Orsogna
, and
T.
Chou
, “
Using excess deaths and testing statistics to determine estimates of COVID-19 mortalities
,”
Eur. J. Epidemiol.
36
,
545
558
(
2021
).
4.
See https://ourworldindata.org/covid-vaccinations for “Coronavirus (COVID-19) Vaccinations” (2020); accessed January 4, 2021.
5.
See https://www.nytimes.com/interactive/2020/science/coronavirus-vaccine-tracker.html for “Coronavirus Vaccine Tracker”; accessed August 28, 2021.
6.
F. P.
Polack
et al., “
Safety and efficacy of the BNT162b2 mRNA Covid-19 vaccine
,”
New Engl. J. Med.
383
,
2603
2615
(
2020
).
7.
Vaccines and Related Biological Products Advisory Committee Meeting, “FDA briefing document, Moderna COVID-19 vaccine,” see https://www.fda.gov/media/144434/download (2020); accessed January 13, 2021.
8.
G.
Chodcik
et al., “The effectiveness of the first dose of BNT162b2 vaccine in reducing SARS-CoV-2 infection 13–24 days after immunization: Real-world evidence,”
medRxiv
(2021).
9.
M.
Voysey
et al., “
Safety and efficacy of the ChAdOx1 nCoV-19 vaccine (AZD1222) against SARS-CoV-2: An interim analysis of four randomised controlled trials in Brazil, South Africa, and the UK
,”
Lancet
397
(
10269
),
99
111
(
2021
).
10.
K.
Kupferschmidt
, “AstraZeneca reports powerful COVID-19 protection in new vaccine trial,”
Science
(published online) (2021).
11.
J. L.
Bernal
N.
Andrews
,
C.
Gower
,
E.
Gallagher
,
R.
Simmons
,
S.
Thelwall
,
J.
Stowe
,
E.
Tessier
,
N.
Groves
,
G.
Dabrera
,
M.
Richard
,
C. N. J.
Campbell
,
G.
Amirthalingam
,
M.
Edmunds
,
M.
Zambon
,
K. E.
Brown
,
S.
Hopkins
,
M.
Chang
, and
M.
Ramsay
, “
Effectiveness of Covid-19 vaccines against the B.1.617.2 (Delta) variant
,”
N. Engl. J. Med.
385
(17),
585
594
(
2021
).
12.
See https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/997414/Variants_of_Concern_VOC_Technical_Briefing_16.pdf for “SARS-CoV-2 Variants of Concern and Variants under Investigation in England” (2021); accessed June 30, 2021.
13.
J.
Stowe
et al., “Effectiveness of COVID-19 vaccines against hospital admission with the Delta (B.1.617.2) variant,” Public Health Engl. (to be published).
14.
E.
Ferreras
et al., “
Single-dose cholera vaccine in response to an outbreak in Zambia
,”
New Engl. J. Med.
378
,
577
579
(
2018
).
15.
A. S.
Azman
et al., “
The impact of a one-dose versus two-dose oral cholera vaccine regimen in outbreak settings: A modeling study
,”
PLoS Med.
12
,
e1001867
(
2015
).
16.
F.
Qadri
et al., “
Efficacy of a single-dose regimen of inactivated whole-cell oral cholera vaccine: Results from 2 years of follow-up of a randomised trial
,”
Lancet Infect. Dis.
18
,
666
674
(
2018
).
17.
J. L.
Kriss
et al., “
COVID-19 vaccine second-dose completion and interval between first and second doses among vaccinated persons–United States, December 14, 2020-February 14, 2021
,”
Morb. Mortal. Wkly. Rep.
70
,
389
(
2021
).
18.
The Guardian, “Germany and Denmark consider delaying second Covid vaccine dose,” see https://www.theguardian.com/world/2021/jan/04/germany-denmark-second-dose-covid-vaccine (2021); accessed January 15, 2021.
19.
See https://www.health.gov.on.ca/en/pro/programs/publichealth/coronavirus/docs/vaccine/COVID-19_Pfizer_vaccine_admin.pdf for “Administration of Pfizer-BioNTech COVID-19 Vaccine” (2021); accessed June 30, 2021).
20.
A.
Bonifacius
et al., “
COVID-19 immune signatures reveal stable antiviral T cell function despite declining humoral responses
,”
Immunity
54
,
340
354
(
2021
).
21.
U.
Sahin
et al., “
COVID-19 vaccine BNT162b1 elicits human antibody and TH1 T cell responses
,”
Nature
586
,
594
599
(
2020
).
22.
S. Y.
Tzeng
et al., “
Stabilized single-injection inactivated polio vaccine elicits a strong neutralizing immune response
,”
Proc. Natl. Acad. Sci. U.S.A.
115
,
E5269
E5278
(
2018
).
23.
M.
Voysey
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
,”
Lancet
397
,
881
891
(
2021
).
24.
M.
Jeyanathan
et al., “
Immunological considerations for COVID-19 vaccine strategies
,”
Nat. Rev. Immunol.
20
,
615
632
(
2020
).
25.
F. J.
Ibarrondo
et al., “
Rapid decay of anti-SARS-CoV-2 antibodies in persons with mild Covid-19
,”
New Engl. J. Med.
383
,
1085
1087
(
2020
).
26.
The Washington Post, “Coronavirus cases among lawmakers who sheltered in lockdown show one vaccine dose may not immediately protect against infection,” see https://www.washingtonpost.com/health/2021/01/13/capitol-riot-lockdown-coronavirus-vaccine/ (2021); accessed January 15, 2021.
27.
F.
Collins
, “Is one vaccine dose enough after COVID-19 infection?,” see https://directorsblog.nih.gov/2021/02/23/is-one-dose-of-covid-19-vaccine-enough-after-covid-19-infection/ (2021); accessed March 22, 2021.
28.
G.
Iacobucci
and
E.
Mahase
, “Covid-19 vaccination: What’s the evidence for extending the dosing interval?,”
BMJ
372
, 18 (2021).
29.
Z.
Gorvett
, “How effective is a single vaccine dose against Covid-19?,” see https://www.bbc.com/future/article/20210114-covid-19-how-effective-is-a-single-vaccine-dose (2021); accessed March 22, 2021.
30.
N. G.
Davies
et al., “
Increased mortality in community-tested cases of SARS-CoV-2 lineage B.1.1.7
,”
Nature
593
,
270
274
(
2021
).
31.
CDC, “New COVID-19 variants,” see https://www.cdc.gov/coronavirus/2019-ncov/transmission/variant.html (2021); accessed January 15, 2021.
32.
A.
Arenas
et al., “
Modeling the spatiotemporal epidemic spreading of COVID-19 and the impact of mobility and social distancing interventions
,”
Phys. Rev. X
10
,
041055
(
2020
).
33.
A.
Aleta
et al., “
Modelling the impact of testing, contact tracing and household quarantine on second waves of COVID-19
,”
Nat. Hum. Behav.
4
,
964
971
(
2020
).
34.
S.
Pei
,
S.
Kandula
, and
J.
Shaman
, “
Differential effects of intervention timing on COVID-19 spread in the United States
,”
Sci. Adv.
6
,
eabd6370
(
2020
).
35.
M. J.
Keeling
and
P.
Rohani
,
Modeling Infectious Diseases in Humans and Animals
(
Princeton University Press
,
2011
).
36.
N. G.
Davies
et al., “
Age-dependent effects in the transmission and control of COVID-19 epidemics
,”
Nat. Med.
26
,
1205
1211
(
2020
).
37.
C.-C.
Lai
,
T.-P.
Shih
,
W.-C.
Ko
,
H.-J.
Tang
, and
P.-R.
Hsueh
, “
Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) and coronavirus disease-2019 (COVID-19): The epidemic and the challenges
,”
Int. J. Antimicrob. Agents
55
,
105924
(
2020
).
38.
C.
Aschwanden
, “
Five reasons why COVID herd immunity is probably impossible
,”
Nature
591
,
520
522
(
2021
).
39.
L. F.
Buss
et al., “COVID-19 herd immunity in the Brazilian Amazon,”
Science
371
, 288–292 (2020).
40.
See https://www.nytimes.com/2020/04/23/nyregion/coronavirus-antibodies-test-ny.html for “1 in 5 New Yorkers May Have Had Covid-19, Antibody Tests Suggest” (2020); accessed January 31, 2021.
41.
E.
Callaway
and
S.
Mallapaty
, “
Novavax offers first evidence that COVID vaccines protect people against variants
,”
Nature
590
,
17
(
2021
).
42.
H.
Salje
et al., “
Estimating the burden of SARS-CoV-2 in France
,”
Science
369
,
208
211
(
2020
).
43.
L.
Breiman
,
J.
Friedman
,
C. J.
Stone
, and
R. A.
Olshen
,
Classification and Regression Trees
(
CRC Press
,
1984
).
44.
T.
Hastie
,
R.
Tibshirani
, and
J.
Friedman
,
The Elements of Statistical Learning: Data Mining, Inference, and Prediction
(
Springer Science & Business Media
,
2009
).
45.
L.
Matrajit
,
J.
Eaton
,
T.
Leung
, et al., “
Optimizing vaccine allocation for COVID-19 vaccines shows the potential role of single-dose vaccination
,”
Nat Commun
12
,
3449
(
2021
).
46.
M.
Lipsitch
and
R.
Kahn
, “Interpreting vaccine efficacy trial results for infection and transmission,” medRxiv (2021); see https://www.medrxiv.org/content/early/2021/02/28/2021.02.25.21252415.
47.
Z.
Wang
,
G.
Röst
, and
S. M.
Moghadas
, “
Delay in booster schedule as a control parameter in vaccination dynamics
,”
J. Math. Biol.
79
,
2157
2182
(
2019
).
48.
L.
Matrajt
,
T.
Britton
,
M. E.
Halloran
, and
I. M.
Longini
, Jr., “
One versus two doses: What is the best use of vaccine in an influenza pandemic?
,”
Epidemics
13
,
17
27
(
2015
).
49.
R.
Laubenbacher
,
J. P.
Sluka
, and
J. A.
Glazier
, “
Using digital twins in viral infection
,”
Science
371
,
1105
1106
(
2021
).
50.
C. M.
Saad-Roy
et al., “
Epidemiological and evolutionary considerations of SARS-CoV-2 vaccine dosing regimes
,”
Science
372
,
363
370
(
2021
).
51.
E. C.
Wall
et al., “
Neutralising antibody activity against SARS-CoV-2 VOCs B.1.617.2 and B.1.351 by BNT162b2 vaccination
,”
Lancet
397
,
2331
2333
(
2021
).
52.
D. S.
Khoury
et al., “
Neutralizing antibody levels are highly predictive of immune protection from symptomatic SARS-CoV-2 infection
,”
Nat. Med.
27
,
1205
1211
(
2021
).
53.
E.
Mahase
, “
Covid-19: Vaccine brands can be mixed in ‘extremely rare occasions,’ says Public Health England
,”
BMJ
372
,
n12
(
2021
).
54.
X.
Liu
et al., “Safety and immunogenicity report from the Com-COV study—A single-blind randomised non-inferiority trial comparing heterologous and homologous prime-boost schedules with an adenoviral vectored and mRNA COVID-19 vaccine,” SSRN Scholarly Paper ID 3874014, Social Science Research Network, Rochester, NY, 2021; see https://papers.ssrn.com/abstract=3874014.
55.
See https://www.nytimes.com/2021/07/23/science/covid-vaccine-israel-pfizer.html for “Israeli Data Suggests Possible Waning in Effectiveness of Pfizer Vaccine” (2020); accessed August 30, 2021.
56.
J.
de Vrieze
, “
Pfizer’s vaccine raises allergy concerns
,”
Science
371
,
10
11
(
2021
).
57.
L.
Böttcher
,
M.
Xia
, and
T.
Chou
, “
Why case fatality ratios can be misleading: Individual-and population-based mortality estimates and factors influencing them
,”
Phys. Biol.
17
,
065003
(
2020
).
58.
W. F.
Garcia-Beltran
et al., “
Multiple SARS-CoV-2 variants escape neutralization by vaccine-induced humoral immunity
,”
Cell
184
,
2372
2383
(
2021
).
59.
S.
Riley
,
J. T.
Wu
, and
G. M.
Leung
, “
Optimizing the dose of pre-pandemic influenza vaccines to reduce the infection attack rate
,”
PLoS Med.
4
,
e218
(
2007
).
60.
R. M.
Anderson
,
C.
Vegvari
,
J.
Truscott
, and
B. S.
Collyer
, “
Challenges in creating herd immunity to SARS-CoV-2 infection by mass vaccination
,”
Lancet
396
,
1614
1616
(
2020
).
61.
F.
Stollmeier
and
J.
Nagler
, “
Unfair and anomalous evolutionary dynamics from fluctuating payoffs
,”
Phys. Rev. Lett.
120
,
058101
(
2018
).
62.
X.
He
et al., “
Temporal dynamics in viral shedding and transmissibility of COVID-19
,”
Nat. Med.
26
,
672
675
(
2020
).
63.
See https://www.cdc.gov/coronavirus/2019-ncov/hcp/duration-isolation.html for “Duration of Isolation and Precautions for Adults with COVID-19” (2020); accessed January 4, 2021.
64.
P.
Van den Driessche
and
J.
Watmough
, “
Reproduction numbers and sub-threshold endemic equilibria for compartmental models of disease transmission
,”
Math. Biosci.
180
,
29
48
(
2002
).
65.
J. R.
Dormand
and
P. J.
Prince
, “
A family of embedded Runge-Kutta formulae
,”
J. Comput. Appl. Math.
6
,
19
26
(
1980
).
66.
T.
Sekine
et al., “
Robust T cell immunity in convalescent individuals with asymptomatic or mild COVID-19
,”
Cell
183
,
158
168
(
2020
).
67.
L.
Böttcher
and
J.
Nagler
, “
Decisive Conditions for Strategic Vaccination against SARS-CoV-2
,”
GitHub
. https://github.com/lubo93/vaccination.