Understanding the impact of disease and vaccine mechanisms on the importance of optimal vaccine allocation^*

2.1. Transmission model

To model the effects of pre-epidemic vaccination, we extend the SEIR model for two population groups to include vaccination, asymptomatic and symptomatic disease progression, and death due to disease. We define our model from the following disease progression:

• Each population group is initially split into unvaccinated and vaccinated susceptibles,

• If infected, a susceptible person becomes exposed, where they are infected but not yet infectious,

• An exposed person will become infectious, developing either symptomatic or asymptomatic disease,

• A person with symptomatic disease can either recover from disease, or die due to infection,

• A person with asymptomatic disease will recover from disease.

We define an ODE model from these assumptions, as shown in the compartmental diagram in Figure 1.

• Download figure
• Open in new tab

Figure 1:

Compartmental diagram for our extended SEIR model. Initially people are either susceptible (U_i) or vaccinated (V_i) where i ∈ {1, 2} denotes population group. Both classes are able to become exposed (E_ji) and then infected (I_jki) where j ∈ {U, V} denotes whether they were initially unvaccinated or vaccinated and k ∈ {S, A} denotes if they experience symptomatic or asymptomatic disease. From here, symptomatic people or asymptomatic can recover from disease (R_jki) or symptomatic people can also die due to disease (D_ji). For each group, ω_i denotes the probability of developing symptomatic infection and σ_i denotes the probabiity of dying due to disease. We define vaccine effectiveness for each mechanism as 1 − α_m for m ∈ {1, 2, 3, 4}. We consider mechanisms: 1) reducing susceptibility, 2) reducing probability of developing symptomatic infection, 3) reducing probability of dying due to disease, and 4) reducing infectivity. In the force of infection λ_i, 1 − δ_A denotes the reduction in infectivity due to asymptomatic infection and LP_j denotes the number of people alive in Group j.

In our model, we define compartments X_jki, where

• X ∈ {U, V, E, I, R, D}: denotes compartment types; unvaccinated, vaccinated, exposed, infected, recovered and dead,

• j ∈ {U, V}: denotes unvaccinated or vaccinated disease progression,

• k ∈ {S, A}: denotes symptomatic or asymptomatic infection, and

• i ∈ {1, 2}: denotes population group (1: high risk of infection, 2: low risk of infection).

We incorporate the varying risk of infection between population groups by assuming people in Group 1 (high risk) are twice as susceptible to infection than people in Group 2 (low risk). From this assumption, we define a contact matrix C: where Group 1 is high risk and Group 2 is low risk, S denotes the susceptible groups (S_i and V_i) and I denotes the infected groups (I_SSi, I_SAi, I_{V Si} and I_{V Ai}).

To investigate the effects of a range of vaccine mechanisms, we consider four vaccine mechanisms [16]:

1. Reduction in susceptibility,

2. Reduction in probability of developing symptomatic infection,

3. Reduction in probability of dying due to disease, and

4. Reduction in infectivity.

We define the effectiveness of each vaccine as 1 − α_{m∈{1,2,3,4}} ∈ [0, 1], corresponds to the vaccine number as defined above. An 100% effective vaccine (1−α_m = 1) corresponds to a perfect vaccine, for example if α₁ = 0, vaccinated individuals are unable to be infected. Conversely, a 0% effective vaccine (1 − α_m = 0) corresponds to the vaccine having no effect. As we only consider pre-epidemic vaccination, vaccine allocation appears in the initial conditions for the vaccinated compartments for Groups 1 (high risk) and 2 (low risk).

In this model, we assume frequency dependent transmission i.e. the transmission rate depends on the proportion of infecteds in the population as the contact rate is independent of population size. This is preferred for modelling human populations over density dependent transmission as we determine our contacts by social constraints, which do not necessarily scale with population size [5]. We further assume progress on a fast enough timescale that we can ignore births and natural deaths and that the population is closed. Therefore, total population for each group P_i∈{1,2} is constant, but this population also includes people who have died due to disease. To implement frequency dependent transmission, we define a new population size LP_i∈{1,2}(t) which denotes the size of the living population for each group. This living population will change over time as people die due to disease. As people die and no longer have contact with living people (we assume), for the same number of infected people, the proportion of infected contacts will increase.

From the assumptions stated above, we derive the ODEs to describe our model. In addition to the standard SEIR model parameters and the vaccine effectiveness parameters α_m, ω_i denotes the proportion of exposed individuals that become symptomatically infected, σ_i denotes the proportion of symptomatic individuals who die due to disease, and 1 − δ_A denotes the reduction in infectivity for asymptomatic infecteds, all for a population group i. Furthermore, [C_ij] denotes the elements of the contact matrix as defined in Eq. (2.1), P_j denotes the size of Group j at the beginning of the epidemic, and LP_j(t) denotes the size of the living population of Group j. Our model is hence by the following equations for each population group:

Unvaccinated :

Vaccinated :

Force of infection

Adding Eqs. (2.2)–(2.15) we find: where we define P₁ and P₂ as the population sizes of Groups 1 and 2 at the beginning of the epidemic. For our analysis we assume P₁ = P₂. As the epidemic progresses and people die due to disease, we define the living population size LP_i∈{1,2}(t) as:

Given both Groups 1 and 2 start with I₀ infections each, and v₁ and v₂ vaccinations respectively Eqns (2.2) − (2.15) have the following initial conditions: for i ∈ {1, 2}, j ∈ {U, V} and k ∈ {S, A}. That is, we assume there are initially 2I₀ infected people with I₀ infected people in each population group. For each population group all initially infected people are unvaccinated with half symptomatic and the other half asymptomatic, which is why I_Uki(0) is equal to for k ∈ {S, A}. In our analysis we assume I₀ = 50. Complete modelling details are given in the supplementary material.

2.2. Optimising Vaccine Allocation

To find the optimal vaccine allocation strategy, we compute the number of vaccines allocated to each population group to minimise a given objective function. Given we consider a finite number of vaccines (v_max) and two population groups, the problem reduces to finding the number of vaccines to allocate to Group 1 (v_opt) to minimise the objective function: where y is the output from the model Eqs. (2.2)–(2.15) with initial conditions V₁(0) = v and V₂(0) = v_max − v.

The three objective functions that we consider are:

1. Total infections:

2. Total symptomatic infections:

3. Total deaths:

where denotes the size of compartment X at the end of the epidemic.

We aim to compare optimal and suboptimal strategies for discrete and continuous variables. For discrete variables, we will compare the outcomes of a variety of strategies generated by allocating all vaccines in varying proportions to each group. For continuous variables, we will compare the outcome of the optimal strategy with the outcomes of vaccinating each group equally (an uninformed strategy) and vaccinating poorly. We calculate the poor strategy by comparing the outcome of a variety of strategies and selecting the worst. For example, if the optimal strategy prioritises vaccinating for Group 1 (high risk) and a poor strategy was found to prioritise vaccinating for Group 2 (low risk), our three comparison strategies are:

1. 1. Optimal strategy: prioritise vaccinating Group 1,

2. Uninformed strategy: vaccinate Groups 1 and 2 equally,

3. Poor strategy: prioritise vaccinating Group 2.

The results presented here were generated using MATLAB 9.10. The model was simulated using ode45 and as the optimisation problem reduces to one dimension, the optimal strategy was computed using fmincon.

3.1. Varying disease characteristics

Here we present results that consider how the difference between optimal and suboptimal strategies is dependent on disease characteristics for various vaccine mechanisms and allocation objectives. We consider vaccination strategies for a 75% effective vaccine with 50% vaccine coverage. These values have been chosen arbitrarily to demonstrate the effects of varying disease characteristics. As we assume Group 1 and Group 2 are the same size, 50% vaccine coverage provides interesting results as it gives us the potential to vaccinate the entirety of one population group. Furthermore, we are interested in relatively effective vaccines, but still want to capture their imperfection and so we choose an effectiveness of 75%. Unless otherwise stated, we assume a total population of size of N = 10000, with 5000 people in each group (P₁ = P₂ = 5000).

The plots in Figure 2 show the effects of varying a parameter that affects transmission (R₀) and another that affects infection characteristics (σ) on the difference between optimal and suboptimal allocation strategies.

• Download figure
• Open in new tab

Figure 2:

Comparing the differences between strategies for varying vaccine mechanisms and allocation objectives as we vary R₀ and σ_i.

We consider the effect of varying R₀ on the difference in outcomes between strategies as shown in Figure 2a. Mathematically, there are two points of interest in this figure, the initial convexity of all the strategies and the point where the outcomes of the strategies meet. For low R₀, we see convex curves for each of the considered strategies due to the infection curves being dominated by the decay of initially infected people, rather than new infections. As R₀ increases, we recover the shape of a standard final size curve as our epidemic takes off (in spite of vaccination). Our second point of interest is where the outcomes of the three considered strategies meet, where we define R₀ = R₀^′ ≈ 2.5. For R₀ < R₀^′, the optimal strategy is to vaccinate Group 1 and the poor strategy is to vaccinate Group 2. For R₀ > R₀^′ we find that the optimal strategy is to vaccine Group 2 and the poor strategy is to vaccinate Group 1. Hence, as we approach R₀ = R₀^′, our optimal and poor strategies approach one another before swapping.

Qualitatively, we see a difference in outcome for all strategies as we vary R₀. While the difference in outcome for high and low R₀ depends on the strategy we choose, for all strategies total infections are monotonically increasing with R₀. The importance of choosing the optimal strategy also varies with R₀, with the largest differences between the optimal and poor strategies being when R₀ is between 1 and 2. While as R₀ increases above 5, the outcome for all strategies gets consistently worse and the difference between strategies decreases.

Figure 2b shows how differences in outcomes between strategies change as we vary the probability of dying due to infection for each population group (σ_i∈{1,2}). We consider three scenarios σ_i: σ₁ > σ₂, σ₁ = σ₂, and σ₁ < σ₂. The biggest difference in outcomes of strategies happens when σ₁ ≠ σ₂. When σ₁ = σ₂, the only factor influencing the difference in deaths between population groups is their relative susceptibility (as defined previously). Hence, while there is a difference in the number of people infected in each population group, vaccination will reduce the probability of dying due to disease to the same value for each group. However, when σ₁ ≠ σ₂, one group will be more likely to die due to disease. Therefore, vaccinating the group more likely to die due to disease will reduce more deaths than vaccinating the other group, widening the gap between the outcomes of various strategies. For parameters with only direct effects, such as σ_i, vaccination only affects vaccinated people. The difference between outcomes of allocation strategies is therefore dependent on how the parameters for each group compare to one another.

The two examples presented in Figure 2 demonstrate how direct and indirect affects of vaccination impact the differences in outcomes between optimal and suboptimal strategies. For parameters affecting transmission (e.g. R₀) the difference between outcomes of strategies is dependent on the value of the parameter itself. This is due to its affect on transmission for both population groups. However, for parameters affecting individual infection (e.g. σ_i) the difference between outcomes of strategies is dependent on the relative parameter values between each population group. The biggest differences between outcomes of strategies will occur with there is large disparity in the parameter between groups, and the smallest difference when both groups are equal.

3.2. Varying vaccine characteristics

We now consider an example population to demonstrate how vaccine effectiveness and coverage affect the differences in outcomes between allocation strategies. We define Group 1 to be more susceptible, more likely to develop symptomatic disease and more likely to die due to disease, and Group 2 to be less susceptible, less likely to develop symptomatic disease and less likely to die due to disease.

Figure 3 compares the differences between allocation strategies for various vaccine mechanisms and allocation objectives as we vary vaccine effectiveness and coverage.

• Download figure
• Open in new tab

Figure 3:

Comparing the differences between strategies for varying vaccine mechanisms and allocation objectives as we vary vaccine effectiveness and coverage.

Figures 3a and 3b show the total infections for vaccines that affect transmission as vaccine effectiveness and coverage are varied. For the considered parameters, the optimal strategy changes from prioritising vaccination for Group 2 to prioritising vaccination for Group 1 at 63% effectiveness (3a). For both vaccine effectiveness and coverage, as we saw in Figure 2a, as we increase our vaccine parameter we reduce our objective for all allocation strategies. As the vaccine tends towards becoming 100% effective, or there is high coverage (80%), the optimal strategy produces a much better result than other strategies. However, for lower effectiveness and coverage, increasing the either parameter decreases our objective for all strategies. This demonstrates that, in our model, securing better vaccines or more vaccines will work to decrease our objective, even under a suboptimal strategy.

Figures 3c and 3d show how total symptomatic infections and total deaths for vaccines that affect individual infection vary and we change vaccine effectiveness and coverage. In Figure 3d the difference between optimal and suboptimal strategies is much smaller for 20% and 80% coverage than 50% as there are fewer ways to allocate vaccines for 20% and 80% coverage than for 50% coverage. As vaccine effectiveness and coverage increase, there is a large reduction in outcome for optimal (and close to optimal) strategies, but a much smaller improvement for poor strategies. As vaccination only impacts infected people, vaccinating the less vulnerable, who are already unlikely to experience symptomatic or severe disease, does little to reduce an objective focused on these infection characteristics. Unlike a vaccine affecting transmission, where increasing quality or quantity of vaccines works to decrease the objective for all strategies, here there is only a sizeable reduction for the best strategies. This indicates that strategy remains important even when effective vaccines and high coverage is available. For our model, allocating poorly with an effective vaccine or high coverage could result in a much worse outcome than allocating well with an ineffective vaccine or lower vaccine coverage.

For our model and the parameters we chose, the difference between optimal and suboptimal strategies varies as we change vaccine effectiveness and coverage based on both vaccine mechanism and allocation objective. If a vaccine mechanism affects transmission, this will impact the total number of infections, and therefore also the total symptomatic infections and total deaths resulting from these infections. Hence for any of these objectives, increasing vaccine effectiveness or coverage will result in a better outcome for all strategies. However, if a vaccine mechanism only results in direct effects, this will only impact objectives associated with the mechanism. For example, a vaccine that reduces the probability of dying due to disease will have no impact on total infections or total symptomatic infections. When the vaccine mechanism does impact an objective, increasing vaccine effectiveness and coverage is only effective when an optimal, or close to optimal, strategy is employed.