Modeling geographic vaccination strategies for COVID-19 in Norway

4.1. Data sources

The health data used in this study, covering the period from the 1st of January to the 31st of July 2021, were provided by the Norwegian emergency preparedness register for COVID-19 (Beredt C19) [36]. The database contains individually merged data collected from national health and other administrative registers. Data including positive tests, importations and deaths were obtained from the Norwegian Surveillance System for Communicable Diseases (MSIS) [37]. Additionally, data on hospitalizations and ICU admissions were collected from the Norwegian Intensive Care and Pandemic Registry (NIPaR) [38, 39]. Finally, the vaccination data were obtained from the Norwegian Immunization Registry (SYSVAK), a national electronic immunization registry that records vaccination status of all individuals [40, 41].

The data were pre-processed and aggregated by date, municipality, age, and risk status. The risk status, or medical risk group, refers to individuals with one or more defined diseases or conditions (e.g. organ transplantation, immunodeficiency, diabetes, etc.) that increase the risk of severe disease and death from COVID-19 [42]. Further details about the data used in this study can be found in S1 Appendix.

Ethics approval has been obtained for the use of data in this article, authorized through the Norwegian Health Preparedness Act, paragraphs 2-4. More information regarding this approval can be found at https://www.fhi.no/en/id/infectious-diseases/coronavirus/emergency-preparedness-register-for-covid-19.

4.2. The two geographic-age-structured models

Our study employs two stochastic mathematical models, an individual-based model (IBM) and a meta-population model (MPM), to conduct analyses of COVID-19 vaccine distribution scenarios in Norway. The models were developed to simulate the Norwegian population of approximately 5.4 million individuals, taking into account their demographics in terms of age and geographic location. Both models consider the actual age profile and geographic distribution of the population, and simulate the spread of SARS-CoV-2 from January to July 2021, when the majority of the first doses of vaccines were delivered to the population. The distribution of population size and density across the Norwegian municipalities is illustrated in Fig S9A and B in S1 Appendix. Moreover, we rely on unpublished social contact studies to determine the heterogeneity of contact patterns between individuals of different ages, while the contacts between and within regions are based on mobility data [43].

To capture the transmission dynamics of the virus, we employ a Susceptible-Exposed-Infectious-Removed (SEIR) transmission model that accounts for symptomatic individuals who may require hospitalization and have a risk of mortality due to severe infection. The models also incorporate interventions such as vaccination and restrictions of social contacts. The distribution of vaccines follows the priority order established by the Norwegian government, while the changes in non-pharmaceutical interventions (NPIs) and population contact behavior are captured by the transmission rates, as described in the model calibration section below.

4.3. SEIR transmission and hospitalization model

We employed an SEIR transmission model, which includes pre-symptomatic, symptomatic and asymptomatic infectious compartments, as illustrated in Fig S3 in S1 Appendix. In brief, individuals who have not been exposed to the virus or the vaccine are susceptible (S). Upon exposure, susceptible individuals (S) enter a latent, non-infectious state (E_a or E_s), before becoming infectious. Symptomatic individuals first enter a pre-symptomatic state (I_p) before developing symptoms (I_s), whereas those with asymptomatic infections (I_a) are assumed to recover (R). Individuals with a symptomatic infection (I_s) may recover (R), or develop severe illness requiring hospitalization (H). Hospitalization can result in discharge and recovery (R), or it can lead to ICU admission (U). ICU admissions (U) are followed by a second stay in the hospital ward (H), before discharge and recovery (R). Deaths (D) are assumed to occur from the symptomatic state or during hospitalization. The time spent in each compartment follows gamma distributions in the IBM and exponential distributions in the MPM, and the parameters are shown in Table S1 in S1 Appendix.

The probabilities of hospitalization and death depend on age and risk status, whereas the probabilities of ICU admissions depend solely on age. We consider that older people have a higher probability of hospitalization compared to younger people, and the risk increases exponentially with age, as reported by another study [49]. Similarly, people belonging to the risk group for severe illness have a higher probability of hospital admission and death than those who do not. For simplicity, we assume that the probability of ICU admission is independent of risk status. The probabilities of hospitalization, ICU admissions, and death are reported in Tables S3, S7, and S8 in S1 Appendix.

To account for regional differences during the simulation period, we scaled the transmission rate parameter β by a scaling factor for each municipality. These scaling factors, or relative reproduction numbers, were assumed to be constant and calculated based on the proportion of cumulative number of reported cases and population size between February 2020 and May 2021 in each municipality, and are provided in Section S2.2 in S1 Appendix. A map showing the distribution of regional scaling factors is also included in Fig S9C in S1 Appendix.

We assume a seasonal effect on transmission across all municipalities, with a relative difference of 50% between the maximum (in winter) and minimum (in summer). The seasonality is calculated based on the mean daily temperature for Norway and illustrated in Fig S2E in S1 Appendix.

We consider cases imported from abroad and calculate the daily number of importations using monthly data that are aggregated by age and county, as shown in Fig S2F in S1 Appendix.

The simulations begin on the 1st of January 2021 concurrent to the vaccine roll-out. To initialize the number of individuals in each compartment, we rely on the results of a separate model used by the Norwegian Institute of Public Health (NIPH) for situational awareness and forecasting of the COVID-19 pandemic in weekly runs, enabling the monitoring of the pandemic in Norway since March 2020 [43, 50, 51]. The situational awareness model provides the total numbers of all age groups and risk factors per county. Within each county, we assume a uniform distribution across municipalities, age groups, and risk factors. This ensures that the total population count is divided among all subpopulations based on their respective population sizes.

4.4. Vaccination model

Based on observational studies during the pandemic, which suggest similar real-world effectiveness, we assume that both vaccines (Comirnaty and Spikevax) are equally effective in our models [52].

We vaccinate individuals who are not infected at the time of vaccination and assume that the vaccines are “leaky”, meaning that the vaccine effect of both the first and second doses is imperfect [16]. This implies that vaccinated individuals have a lower risk of infection upon exposure compared to unvaccinated individuals. The vaccine effectiveness (VE) of the two doses is assumed to be a two-step monotonically increasing function over time, as shown in Fig S2D in S1 Appendix. For simplicity, the MPM incorporates the leaky scheme by utilizing separate compartments. We assume that people receive their second dose 12 weeks after the first dose. We also assume that the time to reach full effect is 14 days after the second dose vaccination, which is consistent with another study [53]. Given the relatively short simulation period of seven months, we decided not to consider waning effects of vaccine-derived immunity in this study.

We characterize vaccine effectiveness by five types of protection, namely against (i) symptomatic infection, (ii) asymptomatic infection, (iii) hospitalization, (iv) death, and (v) transmissibility. The first two effects against infection represent a reduction in susceptibility to the virus, while the last effect against transmissibility is translated into a reduction in infectiousness for infected people. Vaccine effectiveness parameters are provided in Table S2 and detailed information with references on the assumptions of vaccine effectiveness can be found in S1 Appendix.

4.5. The actual Norwegian vaccination strategy

The national government established guidelines for prioritization and administration of the vaccine program, while the municipalities were responsible for vaccinating their residents [10].

During the initial roll-out at the turn of 2020/2021, vaccines were allocated to municipalities based on the number of people above 65 years old exclusively. Elderly people including residents in nursing homes and those with underlying medical conditions were given priority to protect those at greatest risk of severe illness and death from SARS-CoV-2 infection [54]. In practice, healthcare workers were also given priority. In our models, we positioned healthcare workers after the elderly population, and the distribution of vaccines to municipalities was based on the size of at-risk populations, with priority given in age-descending order (refer to Table S10 in S1 Appendix for more information):

1. Elderly aged above 65 years (divided into three age categories)

2. Healthcare workers

3. Individuals aged 18-64 years with underlying medical conditions (divided into three age categories)

4. Adult population aged above 18 years without underlying medical conditions (divided into four age categories)

The geographic prioritization of COVID-19 vaccines in Norway was carried out in two steps. The initial step involved making a priority ranking of the 356 municipalities based on the cumulative hospitalization rates per resident population. On the 9th of March 2021, around the time most individuals above 85 years old had been vaccinated, the government decided to implement moderate geographic prioritization, whereby 20% additional vaccine doses were allocated to 10 municipalities, including the capital Oslo and four municipalities in the surrounding Viken county that had experienced consistently high infection levels over time [13]. These 10 municipalities, referred to as the Plus group, received vaccines at the expense of 330 municipalities, referred to as the Minus group, while the remaining 16 municipalities received the same number of vaccine doses and were referred to as the Neutral group. Later in March, the municipality distribution key was changed to population proportions of adults above 18 years old. The second phase began on the 19th of May 2021, around the time most individuals above 65 years old had been vaccinated, with a stronger geographic targeting of vaccines, where 60% extra doses were allocated to the prioritized regions [14]. This strategy involved adding 19 municipalities to the Plus group, primarily major cities and municipalities in Viken, while 319 and 13 municipalities were placed in the Minus and Neutral groups, respectively. The prioritization continued until all eligible adult individuals in the Plus group aged above 18 years had been offered vaccination.

To simulate the vaccination strategy used in Norway, we used vaccination data recorded by SYSVAK from January to July 2021. Real data from SYSVAK revealed that the prioritization plan was not consistently followed considering vaccine deliveries to each municipality and the vaccination of age groups within the municipalities, possibly due to vaccine hesitancy, logistical challenges, and unforeseen events. In our simulations, we proceeded by extracting the daily count per municipality, age, and risk status of individuals who received their first dose of the vaccine and assumed that they would all receive their second dose 12 weeks later. This assumption was supported by the high uptake rate of 90% among the adult population as of July 2021 (see Fig S10 in S1 Appendix), as well as the fact that 93% of those who received the first dose also received the second dose [36].

4.6. Model calibration

We calibrated our models to fit the national registry data between the 1st of January and the 31st of July 2021. The calibration process involved two steps, with the first step focusing on fitting the trend and age distributions of hospitalizations, and the second step focusing on fitting the age distributions of ICU admissions and deaths.

In March, the sharp increase in hospitalizations due to the third wave of COVID-19 infections and the arrival of the Alpha variant prompted the government to impose restrictions on social gatherings and travel. To account for changes in NPIs, we implemented two change points for the national transmission rate on the 28th of January and 11th of March 2021.

The calibration involved 12 free parameters including transmission rate parameters and susceptibility to infection parameters. The transmission rate parameters β controlled the time-varying behavior (contact rate) in three time periods, while the susceptibility to infection parameters described the relative probability of being infected upon exposure in each of the nine 10-year age group. Due to computational costs, the models were calibrated to fit the national daily hospitalization data by time and age, and we used a Latin hypercube sampling (LHS) to explore the parameter space. The goodness-of-fit was evaluated using a least squares method by the IBM, while the MPM used a Poisson likelihood.

The IBM was additionally calibrated to fit the proportions of the four transmission routes based on information provided by general practitioners. The data collected from general practitioners about the likely transmission revealed that 45.1%, 9.9%, 6.2%, and 38.6% of cases were infected within households, workplaces, schools, and other settings, respectively.

After calibrating the trend of hospitalizations and transmission, we estimated the probabilities of ICU admission given hospitalization and the probabilities of death given infection using age-stratified numbers of infections, hospitalizations, ICU admissions, and deaths over the entire simulation period. The fitted IBM and MPM to epidemiological data are shown in Fig 1. Both models estimate a significantly higher number (on the order of a factor two) of infections than reported cases, particularly in the younger age groups, indicating a high degree of under-ascertainment of cases [55]. Further details about the calibration of the two models can be found in S1 Appendix.

• Download figure
• Open in new tab

Fig 1. The calibrated individual-based model (IBM) and meta-population model (MPM) with data in the actual vaccination strategy.

The upper and lower panels show the IBM and MPM, respectively. A and C: The time series data of all ages. The lines show the model fits with their 50 and 95% prediction intervals represented by gray areas. The dots show the observed data. B and D: The age distribution of total counts. The gray bars show the counts of each simulation and the full bars with black borders show the mean of all simulations. The data are shown in dots with their exact numbers. During the whole period, there were 86289 confirmed cases, 2397 hospital admissions, 466 ICU admissions, and 266 deaths registered.

4.7. Alternative strategies of geographic prioritization

We developed alternative geographic prioritization strategies for vaccine distribution based on the actual Norwegian vaccination strategy. We ensured the comparability of our models by keeping the calibrated parameters and setting the daily available doses identical across all strategies. As we distributed vaccine doses in pairs (first and second doses), all the numbers of doses refer to the first doses shown in Fig S11A in S1 Appendix.

We categorized the 356 municipalities into three groups: Plus, Minus, and Neutral, which corresponded to receiving extra doses, fewer doses, and the same number of doses, respectively. We simplified the geographic prioritization process into a single step and assumed that all municipalities followed the national guidelines for age and risk group prioritization. These strategies were defined by three key parameters:

• ∆p: The proportion of extra vaccine doses provided to the Plus group. The proportion received by each municipality is calculated relative to a national baseline strategy without geographic prioritization of eligible adult individuals aged above 18 years. We consider up to a maximum of 300% additional doses, corresponding to the municipalities in the Plus group receiving quadruple the number of doses relative to the adult population fraction.

• ∆n: The shift of municipality priority towards the Plus group (∆n > 0) or Minus group (∆n < 0). The selection of municipalities is based on the relative reproduction numbers of the regions. The shifts in either direction lead to a decreased population in the Neutral group, and the proportion of the population in the three groups changes accordingly. Geographic maps illustrating the distribution of various alternatives is presented in Fig 2, and more detailed descriptions of the shifts are provided in the following.

  

  • Download figure
  • Open in new tab

  Fig 2. The geographic distribution of municipality prioritization in alternative strategies.

  The 356 municipalities are classified into three groups, represented by the colors green (Minus), orange (Neutral), and blue (Plus). Each strategy comprises a different number of municipalities in each group. The baseline strategy (∆n = 0) represents the selection made in the real-world implementation. The shifts of municipality priority ∆n are indicated in brackets, and the number of municipalities in each group is shown accordingly in each panel title.

• ∆t: The timing of the start of the vaccination program measured in months. The geographic prioritization strategy is assumed to be implemented on the first day of the months, i.e., on the 1st of January, February, March, April, May, June and July 2021.

To derive alternative strategies, we employed the same grouping of municipalities as the original selection implemented in the second step of the actual strategy, with 24, 13, and 319 municipalities in the Plus, Neutral, and Minus groups, respectively. Given the large differences in sizes of municipalities, such as Oslo accounting for 12% of the Norwegian population, while other municipalities may contain only less than a thousand people, it is essential to consider the proportion of the population that is prioritized. In original selection (∆n = 0), the Plus, Neutral, and Minus municipalities represent approximately 50%, 20% and 30% of the Norwegian population, respectively.

The alternative strategies were selected in the following way:

1. We rank each municipality according to their relative reproduction numbers, as described in Section S2.2 and Fig S12 in S1 Appendix. Alternative target areas are chosen by shifting the priority of five municipalities at a time, considering the ranking and their priority group.

2. To increase the prioritized geographic region relative to the adopted strategy (∆n > 0), we add municipalities to the Plus group (originally 24 municipalities) in successive steps. The top five highest ranked municipalities in the Neutral and Minus groups are moved to the Plus group. This procedure is repeated until 40 additional municipalities are shifted to the Plus group, i.e., 29, 34, …, and 64 municipalities are in the Plus group.

3. To increase the non-prioritized geographic region relative to the adopted strategy (∆n < 0), we add municipalities to the Minus group (originally 319 municipalities) in successive steps. The bottom five lowest ranked municipalities in the Plus and Neutral groups are moved to the Minus group. This procedure is repeated until 30 additional municipalities are shifted to the Minus group, i.e., 324, 329, …, and 349 municipalities are in the Minus group.

4. For each of the 14 selected geographic regions in Step 2 (∆n > 0) and Step 3 (∆n < 0), as well as the baseline strategy (∆n = 0), we select a value for prioritization of vaccine doses ∆p ranging from 0% up to a maximum of 300%, and a date of implementation ∆t from January to July 2021. The maximum value of ∆p is determined by the population ratio between Minus and Plus, which is presented in Table S9 in S1 Appendix. The number of extra vaccine doses provided to the Plus group for each strategy combination of ∆n and ∆p is shown in Fig S13 in S1 Appendix.

5. As a baseline for comparison of the alternative strategies, we consider a national strategy without geographic prioritization (i.e. ∆p = 0).

To evaluate the strategies, we conducted 1000 simulations in both IBM and MPM. We used 100 stochastic seeds for each of 10 calibrated parameter sets. Each simulation in the IBM took approximately 3 minutes, while each simulation in the MPM took approximately 20 seconds. All simulations were run in parallel on large computer clusters. We compared each iterated simulation of the alternative strategies to the national baseline strategy without geographic prioritization. To assess the reduction in health outcomes, we calculated the relative risk reduction (RRR) using the following formula: where x represents the cumulative infections, hospitalizations, ICU admissions or deaths over the entire simulation period. The RRR values indicate the amount by which the alternative strategies reduce the health outcomes compared to the baseline strategy. An RRR = 0 means that the alternative is the same as the baseline, while a positive RRR = 50% means that the alternative strategy reduces the cumulative health outcome by 50% compared to the baseline strategy. The higher RRR indicates the more effective the strategy to reduce the burden, and an RRR = 1 represents a disease-free situation. In contrast, a negative RRR = −100% or −200% means doubling or tripling of the outcomes, respectively, compared to the baseline strategy. The following results shows the mean values and their 95% confidence intervals (95%CIs) using , where σ is standard deviation of RRR and N = 1000 is the number of simulations.

In addition, to evaluate the actual strategy as a special case, we generated another national strategy without geographic prioritization, using the real vaccination data. We adjusted the vaccine distribution starting from the implementation of the first step of geographic prioritization (i.e., 9th of March 2021) based on the adult population (aged 18 years or older) in each municipality, while keeping all the data prior to that day. We generated 100 realizations corresponding to the 100 stochastic seeds for each of 10 calibrated parameter sets. We chose to compare the actual strategy with this national strategy instead of the fully controlled baseline strategy above for several reasons. Firstly, healthcare workers were given a high priority in the actual strategy, while we vaccinated healthcare workers after the elderly in the baseline strategy. Secondly, the actual strategy in January and February prioritized based on the population of people aged 65 years or older, while we considered all adults throughout the entire period in the baseline. This resulted in prioritized municipalities receiving more doses than in the actual strategy, primarily due to more younger people living in Oslo and its surrounding areas. Thirdly, both the proposed geographic and age prioritization was not strictly followed, especially during the initial phase, in the actual strategy as we could control in our baseline simulations.

5.1. Evaluating the actual strategy

Table 1 presents the total mean number of infections, hospitalizations, ICU admissions, and deaths from the 1st of January to the 31st of July 2021 under the baseline strategy without geographic prioritization, along with the relative risk reduction (RRR) of the actual strategy and best strategies from both models. The RRR for infections, hospitalizations, ICU admissions, and deaths resulting from the actual strategy were nearly zero, implying that the actual strategy was similar to the baseline without geographic prioritization.

5.2. Alternative municipality priority

In Fig 3, we present the national RRR from two models for various combinations of proportions of extra doses (∆p) and shifts in municipality priority (∆n) for alternative strategies, given the starting time (∆t) of geographic prioritization on the 1st of January 2021. The leftmost column shows the baseline strategy without geographic prioritization (∆p = 0%) does not minimize any health outcome.

• Download figure
• Open in new tab

Fig 3. The mean relative risk reduction (RRR) of health outcomes for the alternative strategies modeled with (A) IBM and (B) MPM.

The mean RRR of infections, hospitalizations, ICU admissions, and deaths is represented by color, ranging from red (negative RRR) to blue (positive RRR). The black empty circles and larger areas illustrate the strategies with the highest mean RRR and positive 95% confidence intervals for each outcome. To minimize infections, hospitalizations, or ICU admissions, the strategies with higher priority levels lead to higher mean RRR. However, for minimizing deaths, the highest level of prioritization has a negative impact, while medium levels of prioritization lead to the highest mean RRR. The y-axis shows the population fractions (%) of three groups (Minus, Neutral, and Plus) for each shift in municipality priority (∆n). The geographic distribution of municipality priority (∆n) can be found in Fig 2.

To minimize infections, a strong geographic prioritization to a small number of municipalities with high transmission levels is optimal. The best strategy involves reducing the priority by moving 25 municipalities to the Minus group (∆n = − 25) and giving 300% extra doses to the Plus group (∆p = 300%), resulting in the highest mean RRR (and their 95% CIs) of 16.4 (16.1 to 16.7)% and 16.0 (15.7 to 16.3)%, which is equivalent to saving 40,063 (39,190 to 40,936) and 31,563 (30,905 to 32,221) infections in the IBM and MPM, respectively. Similar strategies also minimize hospitalizations and ICU admissions. In the IBM, the optimal strategy involves slightly shifting the priority to ∆n = − 20 and ∆p = 250%, which produces the highest mean RRR (and their 95% CIs) of 19.4 (19.0 to 19.7)% and 19.2 (18.8 to 19.6)%, resulting in saving 524 (514 to 534) hospitalizations and 102 (99 to 104) ICU admissions, respectively. In the MPM, the optimal strategy is ∆n = − 25 and ∆p = 300%, with the mean RRR (and their 95% CIs)) of 15.4 (15.0 to 15.7)% and 15.6 (15.1 to 16.1)% for hospitalizations and ICU admissions, respectively, corresponding to saving 342 (334 to 351) hospitalizations and 67 (65 to 69) ICU admissions.

To minimize deaths, a moderate level of geographic prioritization to a larger number of municipalities is likely optimal, unlike the strategies that minimize infections, hospitalizations, or ICU admissions. In the IBM, increasing the priority by moving 10 municipalities to the Plus group (∆n = +10) and giving 100% extra doses (∆p = 100%) is the optimal strategy, which can avoid 20 (95%CI: 18 to 21) deaths with a mean RRR of 8.0 (95%CI: 7.5 to 8.6)%. In the MPM, the strategy that minimizes deaths is ∆n = − 15 and ∆p = 70%, with a mean RRR of 5.5 (95%CI: 4.9 to 6.2)%, corresponding to avoiding 15 (95%CI: 13 to 16) deaths. However, choosing the prioritization strategy that minimizes infections at ∆n = − 25 and ∆p = 300% results in a negative effect on deaths with mean RRR of -4.3 (95%CI: -5.0 to -3.6)% and -12.7 (95%CI: -13.4 to -12.0)%, leading to 10 (95%CI: 8 to 11) and 29 (95%CI: 28 to 31) extra lives lost compared to the baseline strategy in the IBM and MPM, respectively.

The gains in the Plus group are similar across all health outcomes, but the losses in the Minus group can be substantial under higher level of prioritization strategies. Fig S15 in S1 Appendix shows the trade-off between municipalities, i.e. the Plus and Minus groups. Similarly, Fig S16 in S1 Appendix shows the trade-off between age groups. When more doses are prioritized, the protection of the younger population increases, resulting in a higher average age of infections.

Taking the optimal strategy for minimizing infections in the IBM as an example, Fig 4 shows the trade-off between municipalities by plotting the RRR of each municipality on a map. The municipality-specific RRR of infections, hospitalizations, ICU admissions, and deaths are positive in Oslo and its surroundings because they receive more doses and hence prevent larger local outbreaks. Although the national benefit is larger than the losses in terms of infections, hospitalizations, and ICU admissions, it is not for deaths. Non-prioritized municipalities experience a significant negative impact under higher level of geographic prioritization, creating a large trade-off between age groups. Fig S14 in S1 Appendix shows the trade-off by age groups under different geographic prioritization strategies.

• Download figure
• Open in new tab

Fig 4. The geographic distribution of mean relative risk reduction (RRR) of the optimal scenario that minimizes infections.

The optimal scenario is the one of reducing the priority by moving 25 municipalities to Minus and giving 300% extra doses (∆n = − 25, ∆p = 300%). The geographic trade-off between municipalities is illustrated by color from blue to red representing mean values of municipality-specific RRR from positive to negative. The benefits in Oslo and its surroundings are much greater than the drawback in other municipalities to minimize (A) infections, (B) hospitalizations and (C) ICU admissions but not (D) deaths.

5.3. Alternative starting time

The starting time of changing from the national to geographic prioritization strategy is also a key factor in the strategy for prioritizing vaccines. We selected a subset of the combinations of ∆p and ∆n, and explored alternative implementation dates ∆t ranging from January to July 2021. Our results indicate that earlier implementation yields better control by reducing adverse health outcomes across most prioritization strategies.

Fig 5 shows that the RRR of infections, hospitalizations, and ICU admissions decreases by the starting time across all alternative prioritization strategies. However, the RRR of deaths decreases only when prioritizing in mild and moderate levels (approximately ∆p < 200%). For higher priority levels (approximately ∆p > 200%), initiating the geographic prioritization in February or March results in a higher RRR of deaths compared to initiating in January, as shown Fig S17 and Fig S18 in S1 Appendix. Nevertheless, the moderate level (∆p = 70% − 100%) with a starting time in January as the previous section shown, the RRR of deaths is the highest among all strategies. Additionally, we found that all the RRR approach zero if the starting time is after May 2021.

• Download figure
• Open in new tab

Fig 5. The mean relative risk reduction (RRR) of health outcomes for the alternative starting time modeled with (A) IBM and (B) MPM.

The mean RRR of infections, hospitalizations, ICU admissions, and deaths is represented by color, ranging from red (negative RRR) to blue (positive RRR). The panels of columns correspond to alternative start time, with the first column reflecting the results shown in Fig 3. The RRR of infections, hospitalizations, and ICU admissions decrease by starting time, considering all alternative strategies. However, the RRR of deaths decrease only when mild and moderate priority levels (approximately ∆p < 200%) are implemented. The black square areas indicate positive mean values and their 95% confidence intervals. The benefits are limited if the geographic prioritization started after May 2021.

5.4. Sensitivity analysis on the hospitalization probabilities

Doubling the probabilities of hospitalization does not significantly change the optimal strategies for minimizing infections, hospitalizations, ICU admissions, and deaths. For minimizing infections, hospitalizations, and ICU admissions, the strategies remain similar to those with the highest prioritization level. The moderate level of prioritization is still preferred for minimizing deaths. Similarly, implementing geographic prioritization after January leads to a decrease in RRR, which reaches nearly zero with no benefits if it starts after May 2021. For a more detailed comparison of the two assumptions on hospitalization probabilities, please refer to S1 Appendix.

5.5. Comparing the two models

Fig 6 presents a comparison between two models regarding the optimal strategy for minimizing infections. Apart from the initial growth being more rapid in the IBM, the primary discrepancy between the two models is the number of infections on the 28th of January and 11th of March 2021, when the transmission rates β undergo a change. Specifically, the infection curve from the MPM decreases instantly and more rapidly than that from the IBM. Moreover, the MPM correspondingly exhibits lower trends of hospitalizations and ICU admissions compared to the IBM. However, the MPM shows a higher trend of deaths than the IBM, attributed to the diverse age composition between the two models. More specifically, the IBM shows a greater number of infections among the younger population, whereas the MPM records more deaths in the oldest age group (aged 80 years or older). Furthermore, the complexity of the IBM, which includes heterogeneous contact structures in four transmission routes, contributes to additional variations. For a detailed comparison of the actual vaccination strategy for calibration and the counterfactual scenario without vaccination, please refer to Fig S20 and Section S5 in S1 Appendix, respectively.

• Download figure
• Open in new tab

Fig 6. The comparison of two models for the optimal strategy for minimizing infections.

A: The time series data of all ages. The lines show the model fits with their 95% prediction intervals represented by colored areas. B: The age distribution of total counts. The black lines show the median values, and the colored areas show the 95% prediction intervals.