betaFeedback

WorksAgricultural and Biological Sciences

A resident-traveller model for global aspects of vaccination

Abstract

A model for the spread of a vaccine-preventable infectious disease is set in a resident-traveller metapopulation context to investigate global aspects of vaccination. Of particular interest here is the effect of varying vaccine coverage in some patches on the spread of disease in the whole system. Analytically and numerically, the resident-traveller framework is compared with a classic metapopulation model where individuals become residents of the patch they move to. A numerical example inspired by the 2003–2004 poliomyelitis resurgence in western and central Africa is then considered, illustrating how heterogeneous vaccine coverage can sustain widespread regional transmission that classic metapopulation models predict to be much smaller.

Introduction

The increase and acceleration of travel has dramatically eased the global dissemination of infectious pathogens. This was made clear when SARS spread across the globe in 2003 [15, 50], with the propagation of pandemic H1N1 influenza in 2009 [37] and, most recently, with the global spread of SARS-CoV-2 and its variants [2]. However, aside from the rapid spread of emerging or reemerging infectious diseases, globalization of human movement poses other problems. One of them is that it is becoming harder to devise public health policies within the restricted span of individual countries or regions and, as a consequence, to consider policies that do not take into account the practices of other health jurisdictions. While public policies are typically formulated and conducted within a country or local public health jurisdiction, the movement of individuals between countries or jurisdictions implies that issues or deficiencies in the public health system of one jurisdiction can have consequences in others.

For instance, the 2009 influenza pandemic was under way when the yearly pilgrimage to Mecca took place. The gathering of pilgrims from many countries in these circumstances led to much elevated risk of importation of the disease to countries that were at that time almost unaffected by it [38], although this particular pandemic turned out to be generally benign enough that this did not cause major issues. The use of antivirals against influenza during the 2009 pandemic varied greatly from country to country [11]. While some countries were attempting to curtail a potential surge of resistance against antivirals, others were not, so that it is not clear what consequences could have been expected of the policies of the former had widespread resistance indeed emerged. Tuberculosis is another disease where the difference between health systems plays a role. In Canada, for instance, the four countries providing the most new permanent residents in recent years are China, India, Pakistan and the Philippines. These four countries have high tuberculosis burden and not surprisingly, the rate of tuberculosis among new migrants to Canada is much higher than it is for the Canadian-born population [43]. Similarly, in the United States, because latent tuberculosis prevalence and drug-resistance profiles among arriving migrants reflect conditions in their countries of origin [39], decision-analysis studies have shown that it would be cost-saving for the United States to invest directly in expanding tuberculosis control programmes in high-incidence source countries [53]. However, this type of recommendation is rarely taken into account by public health authorities.

Another aspect of public health that is subject to spatial and temporal effects is vaccination. The motivation for the present work rests in the fact that there is evidence that, in 2003–2004, the interruption of poliomyelitis vaccination in the northern Nigerian state of Kano [16, 25, 31] led over the course of several months to an increase of the number of cases of polio in neighbouring regions [47]. Vaccination then resumed [21], but problems persisted long after this event [35]. Thus, eradication efforts, when they are conducted, should attack problems at a global scale [40]. At a more local scale, there have also been several outbreaks of childhood and other vaccine-preventable diseases in countries where vaccine coverage is generally high [41]. This is a worrying trend: because of misinformation about vaccines [18] or religious convictions [27], the number of individuals who are not vaccinated is on the rise. This phenomenon happens at the local scale and can even be driven by varying physicians’ attitudes towards vaccination [58].

Vaccination poses interesting ethical issues [30] as it lies at the intersection of the public and private spheres. Debates concerning vaccination are likely to persist [10]. The reasons that lead to refusal of vaccination are diverse [13], but whatever the reasons, the increased prevalence of the Internet as a place for sharing often incorrect information [32, 33] and the persistence of anti-vaccination movements means that it is of considerable importance to understand the effect of varying vaccine coverage levels.

With these examples in mind, in order to illustrate the effect of local policies on global health, the effect of vaccination is investigated in a metapopulation model that describes travel of individuals between patches or regions. The purpose here is not to formulate full polio or measles vaccination models. Such models would indeed need to include other characteristics of the virus and vaccines that for simplicity are not considered here. Previous work has considered the effect of spatial aspects of vaccination for specific diseases. In particular, [40] focuses on disease elimination when infections are few, e.g., polio, and [46] uses data on rubella in South Africa to consider the outcome of heterogeneous vaccination.

The present work uses a model framework of travel between patches introduced in [51], used in [52] and others and studied analytically in [7, 8, 29]. This type of model is sometimes called a mover-stayer or resident-traveller model (the latter name being used here) because it keeps track of the place of residence of individuals. As in further work of the authors with this model [1, 9], the original model of [51] is extended to allow individuals to travel between remote patches. It is established that the behaviour of the resulting travel model, i.e., the model describing the dynamics of the total population in each subpopulation, is similar to that in [7, 8], namely that with more general travel, each patch population has an asymptotically stable positive equilibrium and the resident population in each patch is asymptotically constant. A model for vaccination in different patches is then formulated and mathematically and computationally analysed.

This work also aims to compare the resident-traveller formulation with classic metapopulation models. In classic models, individuals who travel immediately adopt the parameters and status of their destination. This simplification is problematic when modelling jurisdiction-based public health interventions such as vaccination, where individuals are subject to the policies of their residence rather than their temporary location. By contrasting the two frameworks, the aim is to show how applying local vaccination rates to visitors in classic models can distort predictions of disease persistence and control.

In the computational investigation of the models, the focus is on two public health questions:

  1. Can a public health authority that does vaccinate help non-vaccinating neighbouring authorities, or protect its own population?

  1. What happens to the global system when, like in the polio example motivating this work, a jurisdiction ceases to vaccinate?

This is complemented by a numerical thought experiment inspired by the 2003–2004 polio resurgence in western and central Africa.

The mobility model in the absence of disease

Formulation of the mobility model

The model tracks individuals as they move in a multidigraph comprising nn patches (regions) as vertices and links between the patches as (multi) arcs. For simplification purposes, it is assumed that an individual who is born in a given patch stays until death a resident of this patch. This individual may however spend an indefinite amount of time in other patches. An individual resident of one patch who is currently in another patch is a traveller. The number of residents of patch ii present in patch jj at time tt is denoted Nij(t)N_{ij}(t); they are referred to as subpopulation (i,j)(i,j).

The total number of residents of patch ii is given by Nir=∑k=1nNikN_i^r = \sum_{k=1}^{n} N_{ik} and the number of individuals who are physically present in patch ii at time tt, both residents and travellers, is Nip=∑k=1nNkiN_i^p = \sum_{k=1}^{n} N_{ki}. The rates of movement of individuals between patches are written with three ordered subscripts as

mresidence,destination,origin.m_{\mathrm{residence,destination,origin}}.

Thus, individuals from patch ii who are visiting patch jj travel to patch kk at a rate mikj≥0m_{ikj} \ge0. The travel structure defines a weighted multi-digraph, with patches as vertices and weighted arcs representing the links between these patches, the weight on the arcs indicating the rates of travel. Between any two patches (vertices) there are at most nn arcs in each direction.

Assume that death occurs at the per capita rate dij>0d_{ij} > 0 for residents of patch ii currently in patch jj. only in the residence patch, at the rate biib_{ii}. Dispersion is instantaneous, so death during travel is negligible.

Taking the previous assumptions into account and assuming that population numbers are sufficiently large for a deterministic formulation, ordinary differential equations can be derived for the dynamics of the population. Firstly, the evolution of the number of residents of patch ii present in patch ii is given by

ddtNii=bii−diiNii+∑k=1k≠inmiikNik−(∑k=1k≠inmiik)Nii.\frac{d}{dt}N_{ii}=b_{ii}-d_{ii}N_{ii}+\sum_{\substack{k=1\\k\ne i}}^{n}m_{iik}N_{ik}-\left(\sum_{\substack{k=1\\k\ne i}}^{n}m_{iik}\right)N_{ii}.

The evolution of the number of residents of patch ii who are present in patch j≠ij\ne i is governed by

ddtNij=−dijNij+∑k=1k≠jnmijkNik−(∑k=1k≠jnmikj)Nij.\frac{d}{dt}N_{ij}=-d_{ij}N_{ij}+\sum_{\substack{k=1\\k\ne j}}^{n}m_{ijk}N_{ik}-\left(\sum_{\substack{k=1\\k\ne j}}^{n}m_{ikj}\right)N_{ij}.

To make the notation more compact, denote

bij={bii>0if i=j0if i≠jb_{ij}= \begin{cases} b_{ii}>0 & \text{if }i=j\\ 0 & \text{if }i\ne j \end{cases}

the birth constant for subpopulation (i,j)(i,j) and

mijj=−∑k=1k≠jnmikjm_{ijj}=-\sum_{\substack{k=1\\k\ne j}}^{n}m_{ikj}

the rate of movement outbound from patch jj for residents of patch ii. The general equation for the demographic model is then given for i,j=1,…,ni,j=1,\ldots,n by

ddtNij=bij−dijNij+∑k=1nmijkNik.(1)\frac{d}{dt}N_{ij}=b_{ij}-d_{ij}N_{ij}+\sum_{k=1}^{n}m_{ijk}N_{ik}. \tag*{(1)}

System (1.1) subject to initial values Nij(0)≥0N_{ij}(0)\ge0 for all i,j=1,…,ni,j=1,\ldots,n and Nir(0)>0N_i^r(0)>0 for all $i=1,\ldots,n constitutes the linear nonhomogeneous mobility model in the absence of disease.

Behaviour of the total population

Lemma 1. With the given initial conditions, the resident population NirN_i^r in each patch and the total population are positive and bounded.

Proof. The evolution of the total resident population in patch i=1,…,ni=1,\ldots,n takes the form

ddtNir=ddt∑j=1nNij=bii−∑j=1ndijNij.(2)\frac{d}{dt}N_i^r=\frac{d}{dt}\sum_{j=1}^{n}N_{ij}=b_{ii}-\sum_{j=1}^{n}d_{ij}N_{ij}. \tag*{(2)}

where the travel terms have cancelled out. Indeed,

∑j=1n(∑k=1nmijkNik)=∑j=1n(∑k=1k≠jnmijkNik−∑k=1k≠jnmikjNij)=∑j=1n∑k=1k≠jnmijkNik−∑j=1n∑k=1k≠jnmikjNij=0.\begin{aligned} \sum_{j=1}^{n}\left(\sum_{k=1}^{n}m_{ijk}N_{ik}\right)=\sum_{j=1}^{n}\left(\sum_{\substack{k=1\\k\ne j}}^{n}m_{ijk}N_{ik}-\sum_{\substack{k=1\\k\ne j}}^{n}m_{ikj}N_{ij}\right) \\ &=\sum_{j=1}^{n}\sum_{\substack{k=1\\k\ne j}}^{n}m_{ijk}N_{ik}-\sum_{j=1}^{n}\sum_{\substack{k=1\\k\ne j}}^{n}m_{ikj}N_{ij} \\ &=0. \end{aligned}

Let d‾i=min⁡j{dij}\underline{d}_i=\min_j\{d_{ij}\} and d‾i=max⁡j{dij}\overline{d}_i=\max_j\{d_{ij}\}. Equation (1.2) gives, for all t≥0t\ge0,

0<Nir(0)e−d‾it+biid‾i(1−e−d‾it)≤Nir(t)≤Nir(0)e−d‾it+biid‾i(1−e−d‾it).\begin{aligned} 0<N_i^r(0)e^{-\overline{d}_it}+\frac{b_{ii}}{\overline{d}_i}\left(1-e^{-\overline{d}_it}\right)\le N_i^r(t) \\ &\le N_i^r(0)e^{-\underline{d}_it}+\frac{b_{ii}}{\underline{d}_i}\left(1-e^{-\underline{d}_it}\right). \end{aligned}

As Nir(0)>0N_i^r(0)>0, the result follows for the total resident population. Also, the total population in the system is given by ∑i=1nNir\sum_{i=1}^{n}N_i^r and is therefore also positive and bounded.

Equilibrium of the mobility model

To further analyze system (1.1), the following notation is introduced. Let a=(a1,…,an)∈Rn\boldsymbol{a}=(a_1,\ldots,a_n)\in\mathbb{R}^n be a vector. The notation a≥0\boldsymbol{a}\ge0 indicates that ai≥0a_i\ge0 for all i=1,…,ni=1,\ldots,n; a>0\boldsymbol{a}>0 means that a≥0\boldsymbol{a}\ge0 and there exists i∈{1,…,n}i\in\{1,\ldots,n\} such that ai>0a_i>0; finally, a≫0\boldsymbol{a}\gg0 indicates that a\boldsymbol{a} is entry-wise positive. The same notation is used for matrices. Here and later, it is convenient to use the following vector notation for state variables. For X=NX=N and, later, X∈{N,S,I,R,V}X\in\{N,S,I,R,V\}, denote

Xi=(Xi1,…,Xin)T(3)X_i=(X_{i1},\ldots,X_{in})^{T} \tag*{(3)}
X=(X1T,…,XnT)T=(X11,…,X1n,X21,…,Xnn)T.(4)\begin{aligned} X=(X_1^{T},\ldots,X_n^{T})^{T} \\ &=(X_{11},\ldots,X_{1n},X_{21},\ldots,X_{nn})^{T}. \tag*{(4)} \end{aligned}

Thus, NiN_i is the part of NN corresponding to the subpopulation residing in patch ii. Let M=⨁i=1nMi\mathcal{M}=\bigoplus_{i=1}^{n}\mathcal{M}_i be the block diagonal mobility matrix, with each block Mi\mathcal{M}_i describing the movement of residents of patch ii and given by

Mi=(−∑k=1k≠1nmik1mi12⋯mi1nmi21−∑k=1k≠2nmik2⋯mi2n⋮⋮⋱⋮min1min2⋯−∑k=1k≠nnmikn).(5)\mathcal{M}_i= \begin{pmatrix} -\sum_{\substack{k=1\\k\ne1}}^{n}m_{ik1} & m_{i12} & \cdots& m_{i1n}\\ m_{i21} & -\sum_{\substack{k=1\\k\ne2}}^{n}m_{ik2} & \cdots& m_{i2n}\\ \vdots& \vdots& \ddots& \vdots\\ m_{in1} & m_{in2} & \cdots& -\sum_{\substack{k=1\\k\ne n}}^{n}m_{ikn} \end{pmatrix}. \tag*{(5)}

By [3 Lemma 2], −Mi-\mathcal{M}_i is a singular M-matrix. From now on, assume that each matrix Mi\mathcal{M}_i is irreducible, i.e., the mobility digraph for residents of patch ii is strongly connected (if residents of patch ii only visit a subset of patches, Mi\mathcal{M}_i is restricted to that accessible subset, with Nij=0N_{ij}=0 for unvisited patches jj). It is not assumed that all Mi\mathcal{M}_i have the same zero/nonzero pattern, just that they all be irreducible on their respective accessible sets. Let

Di=diag⁡(di1,…,din)\mathcal{D}_i=\operatorname{diag}(d_{i1},\ldots,d_{in})

and

D=⨁i=1nDi\mathcal{D}=\bigoplus_{i=1}^{n}\mathcal{D}_i

be the diagonal matrices of death rates. Finally, let b=(b1T,…,bnT)T\boldsymbol{b}=(\boldsymbol{b}_1^{T},\ldots,\boldsymbol{b}_n^{T})^{T}, with the vector bi\boldsymbol{b}_i having iith entry biib_{ii} and zeros everywhere else. The vector b\boldsymbol{b} is the nonhomogeneous component of the system and system (1.1) can be written as

ddtN=(M−D)N+b.(6)\frac{d}{dt}N=(\mathcal{M}-\mathcal{D})N+\boldsymbol{b}. \tag*{(6)}

Because of the structure of the mobility model used, individuals resident in a given patch “radiate” out to all the patches to which they can travel, but never lose their connection to their patch of residence. As a consequence, the resident population of patch ii satisfies the equation

ddtNi=(Mi−Di)Ni+bi.(7)\frac{d}{dt}N_i=(\mathcal{M}_i-\mathcal{D}_i)N_i+\boldsymbol{b}_i. \tag*{(7)}

and thus can be considered independently for each i=1,…,ni=1,\ldots,n. Since Di>0\mathcal{D}_i>0, this model has a unique equilibrium as given by the following result.

Proposition 2. The mobility model (1.1) always has a unique asymptotically stable equilibrium N⋆≫0N^\star\gg0, in which the equilibrium value of the resident population of patch i=1,…,ni=1,\ldots,n is given by Ni⋆=(Di−Mi)−1biN_i^\star=(\mathcal{D}_i-\mathcal{M}_i)^{-1}\boldsymbol{b}_i.

Proof. From (1.6), the value of the equilibrium is given by Ni⋆=(Di−Mi)−1biN_i^\star=(\mathcal{D}_i-\mathcal{M}_i)^{-1}\boldsymbol{b}_i provided that the matrix Di−Mi\mathcal{D}_i-\mathcal{M}_i is invertible. Since Di>0\mathcal{D}_i>0 and Mi\mathcal{M}_i is irreducible, it follows from [3 Proposition 3(4)] that Di−Mi\mathcal{D}_i-\mathcal{M}_i is an irreducible nonsingular M-matrix and (Di−Mi)−1≫0(\mathcal{D}_i-\mathcal{M}_i)^{-1}\gg0. As a consequence, Ni⋆=(Di−Mi)−1biN_i^\star=(\mathcal{D}_i-\mathcal{M}_i)^{-1}\boldsymbol{b}_i is unique and Ni⋆≫0N_i^\star\gg0 since bi>0\boldsymbol{b}_i>0. Finally, since Di−Mi\mathcal{D}_i-\mathcal{M}_i is an irreducible nonsingular M-matrix, it has all its eigenvalues with positive real parts, implying that Mi−Di\mathcal{M}_i-\mathcal{D}_i has all its eigenvalues with negative real parts and Ni⋆N_i^\star is asymptotically stable. □

The SIRS model

Before considering the model with vaccination, properties of the underlying SIRS epidemic model are briefly summarised. To the best of the authors’ knowledge, properties of an SIRS model in a resident-traveller context have not been established.

Formulation of the model

The following assumptions are made.

  1. Each subpopulation is described by a compartmental SIRS model; the model structure in each subpopulation is as in Figure 1. Here, SS, II and RR are the numbers of individuals that are susceptible, infectious and recovered, respectively. The number of infectious residents in patch ii present in patch jj at time tt is denoted by Iij(t)I_{ij}(t), and similarly for the other variables. The average duration of infectiousness for an individual resident in patch ii currently in patch jj is 1/γij1/\gamma_{ij}, and upon recovery the average period of temporary immunity is 1/ψij1/\psi_{ij}.

Flow diagram for one subpopulation of the SIRS model

Figure 1. Flow diagram for one subpopulation of the SIRS model in a given population NijN_{ij}. There is birth into the SijS_{ij} compartment with i=ji=j, i.e., bij=0b_{ij}=0 if i≠ji\ne j.

  1. The disease transmission coefficient in patch jj from an infectious resident in patch kk to a susceptible resident in patch ii is denoted βkij≥0\beta_{kij}\geq0. It is assumed that ∑j=1nβkij>0\sum_{j=1}^{n}\beta_{kij}>0, i.e., the disease can be transmitted in some patch jj from residents of patch kk to individuals from patch ii visiting patch jj. Incidence is supposed to be of standard type.

  1. As in the underlying mobility model of Section 1, birth only occurs in the patch of residence and happens at the rate bijb_{ij}, with bij=0b_{ij}=0 for i≠ji\ne j. There is no vertical transmission of the disease, so all individuals are born susceptible.

  1. The rate of movement can depend on the epidemiological status of individuals. Using the same notation as in Section 1, a superscript is added to indicate the status of individuals, so movement is described by

mresidence,destination,originstatusm_{\mathrm{residence,destination,origin}}^{\mathrm{status}}

This defines a multi-digraph with at most 3n3n arcs in each direction between any two vertices. Unless otherwise indicated, it is assumed that the matrices MiX−Di\mathcal{M}_i^X-\mathcal{D}_i are irreducible, where MiX\mathcal{M}_i^X is given for X∈{S,I,R}X\in\{S,I,R\} by (1.4) with mijkm_{ijk} replaced by mijkXm_{ijk}^X.

  1. There is no disease induced mortality: diseases such as poliomyelitis, although debilitating, rarely cause death.

For i,j=1,…,ni,j=1,\ldots,n, the equations for this nn-patch model are

ddtSij=bij+ψijRij−(dij+∑k=1nβikjIkjNjp)Sij+∑k=1nmijkSSik(8)\begin{aligned} \frac{d}{dt}S_{ij} &= b_{ij}+\psi_{ij}R_{ij}-\left(d_{ij}+\frac{\sum_{k=1}^{n}\beta_{ikj}I_{kj}}{N_j^p}\right)S_{ij}+\sum_{k=1}^{n}m_{ijk}^{S}S_{ik} \tag*{(8)} \end{aligned}
ddtIij=∑k=1nβikjIkjNjpSij−(dij+γij)Iij+∑k=1nmijkIIik(9)\begin{aligned} \frac{d}{dt}I_{ij} &= \frac{\sum_{k=1}^{n}\beta_{ikj}I_{kj}}{N_j^p}S_{ij}-(d_{ij}+\gamma_{ij})I_{ij}+\sum_{k=1}^{n}m_{ijk}^{I}I_{ik} \tag*{(9)} \end{aligned}
ddtRij=γijIij−(dij+ψij)Rij+∑k=1nmijkRRik(10)\begin{aligned} \frac{d}{dt}R_{ij} &= \gamma_{ij}I_{ij}-(d_{ij}+\psi_{ij})R_{ij}+\sum_{k=1}^{n}m_{ijk}^{R}R_{ik} \tag*{(10)} \end{aligned}

System (2.1) is considered with initial conditions

∑j=1nSij(0)>0,Iij(0)≥0andRij(0)≥0\sum_{j=1}^{n}S_{ij}(0)>0,\qquad I_{ij}(0)\geq0\quad\text{and}\quad R_{ij}(0)\geq0

for all i,j=1,…,ni,j=1,\ldots,n and can easily be shown to have all solutions in the nonnegative orthant. It is also easily shown that the set

{(Sij,Iij,Rij)∈R3n2; ∀i,j=1,…,n, Iij=0}\{(S_{ij},I_{ij},R_{ij})\in\mathbb{R}^{3n^{2}};\ \forall i,j=1,\ldots,n,\ I_{ij}=0\}

is positively invariant under the flow of (2.1). From now on, it is further assumed that

∑i,j=1nIij(0)>0\sum_{i,j=1}^{n}I_{ij}(0)>0

to avoid trivial solutions.

It is convenient to use vector and matrix notation as in Section 1.3, so for X∈{N,S,I,R}X\in\{N,S,I,R\}, the state variable vectors are written using (1.3). Using the same ordering of first and second index, write Ψ=diag⁡(ψij)\Psi=\operatorname{diag}(\psi_{ij}) the matrix of rates of loss of immunity and Γ=diag⁡(γij)\Gamma=\operatorname{diag}(\gamma_{ij}) the matrix of rates of recovery. When used with only one index, variables refer to matrices relative to residents of a given patch. For instance, Γi=diag⁡(γi1,…,γin)\Gamma_i=\operatorname{diag}(\gamma_{i1},\ldots,\gamma_{in}), so Γ=⨁i=1nΓi\Gamma=\bigoplus_{i=1}^{n}\Gamma_i.

The disease-free equilibrium of the SIRS model

In order to continue the analysis, system (2.1) is investigated at the disease-free equilibrium (DFE), i.e., where Iij=0I_{ij}=0 for all i,j=1,…,ni,j=1,\ldots,n. In earlier studies [7, 8], the case of status-independent mobility was considered, in which case the population NijN_{ij} in subpopulation (i,j)(i,j) tends to Nij⋆N_{ij}^{\star} as given in Proposition 2. Because of status dependent travel, the situation is more complex here. First, at the DFE, there are no recovered individuals.

Lemma 3. Suppose that system (2.1) is at the DFE. Then lim⁡t→∞R(t)=0\lim_{t\to\infty}R(t)=0.

Proof. Substituting Iij=0I_{ij}=0 in (2.1c) gives

ddtRij=∑k=1nmijkRRik−(dij+ψij)Rij.(11)\frac{d}{dt}R_{ij}=\sum_{k=1}^{n}m_{ijk}^{R}R_{ik}-(d_{ij}+\psi_{ij})R_{ij}. \tag*{(11)}

or, in vector form,

ddtR=(MR−D−Ψ)R.\frac{d}{dt}R=(\mathcal{M}^{R}-D-\Psi)R.

The matrix −MR-\mathcal{M}^{R} is a singular M-matrix, so reasoning as in the proof of Proposition 2, D+Ψ−MRD+\Psi-\mathcal{M}^{R} is a nonsingular M-matrix. As a consequence, R=0R=0 is the unique equilibrium of R′=(MR−D−Ψ)RR'=(\mathcal{M}^{R}-D-\Psi)R, MR−D−Ψ\mathcal{M}^{R}-D-\Psi has all eigenvalues with negative real parts and therefore, R→0R\to0 as t→∞t\to\infty.

It follows that at the DFE,

0=bij−dijSij+∑k=1nmijkSSik.0=b_{ij}-d_{ij}S_{ij}+\sum_{k=1}^{n}m_{ijk}^{S}S_{ik}.

In vector notation, these equations take the form, for patch i=1,…,ni=1,\ldots,n,

0=(MiS−Di)Si+bi.(12)0=(\mathcal{M}_{i}^{S}-D_i)S_i+b_i. \tag*{(12)}

From this, the situation can be determined at the level of the patch of residence.

Lemma 4. Consider residents of patch i=1,…,ni=1,\ldots,n. If system (2.1) is at the DFE, then the DFE for residents of patch ii is unique and given by

Si⋆=(Di−MiS)−1bi≫0,Ii⋆=Ri⋆=0.(13)S_i^{\star}=(D_i-\mathcal{M}_i^{S})^{-1}b_i\gg0,\qquad I_i^{\star}=R_i^{\star}=0. \tag*{(13)}

Proof. Consider the DFE relation (2.3). The matrix Di−MiSD_i-\mathcal{M}_i^{S} is a nonsingular M-matrix and by the irreducibility assumption, has entry-wise positive inverse [3 Proposition 3(4)]. Solving (2.3) gives Si⋆S_i^{\star} as in (2.4); its entry-wise positivity follows from the inverse positivity of Di−MiSD_i-\mathcal{M}_i^{S}.

Note that if MS=M\mathcal{M}^{S}=\mathcal{M}, this is the equilibrium given by Proposition 2 for the demographic model; at the DFE (2.4), the total subpopulation sizes are given by Ni⋆=Si⋆N_i^{\star}=S_i^{\star}.

The basic reproduction number R0\mathcal{R}_0

The next generation operator approach [17] and its characterisation in the ordinary differential equations case given by [57] is used to determine the basic reproduction number R0\mathcal{R}_0 for the system. The reasoning closely follows that used in [8]. Consider only the infected variables and order them as in (1.3). Let the vectors F\mathcal{F} and −W-\mathcal{W} contain new infections and all other flows within infected compartments, respectively. Differentiating the vectors F\mathcal{F} and $\mathcal{W} with respect to the infected variables and evaluating at the disease-free equilibrium (2.4) gives the matrices

F=(F11⋯F1n⋮⋮Fn1⋯Fnn)(14)F=\begin{pmatrix} F_{11} & \cdots& F_{1n}\\ \vdots& & \vdots\\ F_{n1} & \cdots& F_{nn} \end{pmatrix} \tag*{(14)}

and

W=Γ+D−MI=⨁i=1n(Γi+Di−MiI).(15)W=\Gamma+D-\mathcal{M}^{I}=\bigoplus_{i=1}^{n}\left(\Gamma_i+D_i-\mathcal{M}^{I}_i\right). \tag*{(15)}

The blocks FijF_{ij} in matrix FF are the Jacobian matrices obtained by differentiating FiF_i with respect to IjI_j. They are diagonal and take the form Fij=diag⁡(fij1,…,fijn)F_{ij}=\operatorname{diag}(f_{ij1},\ldots,f_{ijn}), with

fijq=βijqSiq⋆Nqp=βijqNiq⋆Nqp⋆,(16)f_{ijq}=\beta_{ijq}\frac{S_{iq}^{\star}}{N_q^p}=\beta_{ijq}\frac{N_{iq}^{\star}}{N_q^{p\star}}, \tag*{(16)}

where q=1,…,nq=1,\ldots,n and Nqp⋆=∑k=1nNkq⋆N_q^{p\star}=\sum_{k=1}^{n}N_{kq}^{\star} is the total population of patch qq at the DFE. Hence FF is a nonnegative matrix. Also, WW is a nonsingular M-matrix and thus W−1W^{-1} is nonnegative. In fact,

W−1=⨁i=1n(Γi+Di−MiI)−1=:⨁i=1nWii−1(17)W^{-1}=\bigoplus_{i=1}^{n}\left(\Gamma_i+D_i-\mathcal{M}^{I}_i\right)^{-1}=:\bigoplus_{i=1}^{n}W_{ii}^{-1} \tag*{(17)}

and the Wii−1W_{ii}^{-1} are positive by the irreducibility assumption. As a consequence, FW−1FW^{-1} takes the form

FW−1=(F11W11−1⋯F1nWnn−1⋮⋮Fn1W11−1⋯FnnWnn−1),(18)FW^{-1}= \begin{pmatrix} F_{11}W_{11}^{-1} & \cdots& F_{1n}W_{nn}^{-1}\\ \vdots& & \vdots\\ F_{n1}W_{11}^{-1} & \cdots& F_{nn}W_{nn}^{-1} \end{pmatrix}, \tag*{(18)}

and [57 Theorem 2] gives the following result, in which ρ(⋅)\rho(\cdot) is the spectral radius, showing that the basic reproduction number R0R_0 acts as a threshold.

Theorem 5. Consider system (2.1) with FF and WW as defined above. Let

R0=ρ(FW−1).(19)\mathcal{R}_0=\rho(FW^{-1}). \tag*{(19)}

Then

  • if R0<1\mathcal{R}_0<1, the DFE is locally asymptotically stable;

  • if R0>1\mathcal{R}_0>1, the DFE is unstable, the disease is uniformly persistent and there exists at least one endemic equilibrium point.

Proof. The local asymptotic stability for R0<1\mathcal{R}_0<1 and in-stability for R0>1\mathcal{R}_0>1 follow from [57 Theorem 2]. It remains to show uniform persistence and existence of an endemic equilibrium when R0>1\mathcal{R}_0>1. Let Ω\Omega be the biologically feasible region defined by the non-negative orthant of the state space where Nir>0N_i^r>0 for all i=1,…,ni=1,\ldots,n. Because the total population in each patch is governed by a dissipative demographic equation, solutions are bounded, making Ω\Omega positively invariant and point dissipative. Let Ω0\Omega_0 be the subset of Ω\Omega where Iij>0I_{ij}>0 for at least one pair (i,j)(i,j) and define the disease-free subspace as ∂Ω0=Ω∖Ω0\partial\Omega_0=\Omega\setminus\Omega_0, where I=0I=0.

The DFE is the unique equilibrium in ∂Ω0\partial\Omega_0 and is globally asymptotically stable within ∂Ω0\partial\Omega_0 by Lemmas 3 and 4. When R0>1\mathcal{R}_0>1, the spectral bound s(F−W)>0s(F-W)>0. It is now shown that the DFE is a weak repeller for Ω0\Omega_0. Assume by contradiction that there exists a solution starting in Ω0\Omega_0 such that lim sup⁡t→∞I(t)=0\limsup_{t\to\infty}I(t)=0. Then R(t)→0R(t)\to0 and S(t)→S⋆S(t)\to S^\star as t→∞t\to\infty, so that Siq(t)/Nqp(t)→Siq⋆/Nqp⋆S_{iq}(t)/N_q^p(t)\to S_{iq}^\star/N_q^{p\star}. Consequently, for any η>0\eta>0 there exists t1>0t_1>0 such that for all t>t1t>t_1, I′(t)≥(F−W−ηM)I(t)I'(t)\ge(F-W-\eta M)I(t) for some non-negative matrix MM. By the continuity of the spectral bound, η\eta can be chosen small enough such that s(F−W−ηM)>0s(F-W-\eta M)>0. This implies that I(t)I(t) must eventually grow exponentially, which contradicts the assumption that I(t)→0I(t)\to0. Therefore, the DFE repels solutions starting in Ω0\Omega_0.

Because the system is point dissipative, Ω\Omega is positively invariant and ∂Ω0\partial\Omega_0 contains only the isolated and acyclic invariant set {E0}\{E_0\} (the DFE), it follows from standard uniform persistence theory [56 Theorem 4.6] that the disease is uniformly persistent, which together with the convexity of Ω\Omega implies the existence of at least one endemic equilibrium point in Ω0\Omega_0.

Basic reproduction number in isolated patches

The basic reproduction number in isolated patches is useful in numerical considerations and is easily obtained from the previous reasoning. Suppose that the patches are isolated, i.e., MS=MI=MR=0\mathcal{M}^{S}=\mathcal{M}^{I}=\mathcal{M}^{R}=0. Then the model can be considered with no subpopulations, as in the absence of movement, the number of visitors to a patch is either identically zero or goes to zero. Using (2.4) with MiS=0\mathcal{M}^{S}_i=0, the DFE takes the form, for i=1,…,ni=1,\ldots,n,

(Sii⋆,Iii⋆,Rii⋆)=(biidii,0,0).(S_{ii}^{\star},I_{ii}^{\star},R_{ii}^{\star})=\left(\frac{b_{ii}}{d_{ii}},0,0\right).

Matrix FF reduces to the scalar

Fii=βiiiSii⋆Nii⋆=βiii,F_{ii}=\beta_{iii}\frac{S_{ii}^{\star}}{N_{ii}^{\star}}=\beta_{iii},

while Wii−1W_{ii}^{-1} is

Wii−1=1γii+dii,W_{ii}^{-1}=\frac{1}{\gamma_{ii}+d_{ii}},

so the basic reproduction number in patch ii in isolation is

R0i(2.1)=βiiiγii+dii.\mathcal{R}_{0i}^{(2.1)}=\frac{\beta_{iii}}{\gamma_{ii}+d_{ii}}.

Case of movement equal for all epidemiological states

Lemma 6. Suppose that in system (2.1), individuals in all disease states travel at the same rate, i.e., M:=MS=MI=MR\mathcal{M} := \mathcal{M}^{S} = \mathcal{M}^{I} = \mathcal{M}^{R}. Then the number of individuals NijN_{ij} in each subpopulation (i,j)(i,j) is convergent.

Proof. Denote mijk:=mijkS=mijkI=mijkRm_{ijk} := m_{ijk}^{S} = m_{ijk}^{I} = m_{ijk}^{R}. For a given subpopulation (i,j)(i,j),

ddtNij=ddt(Sij+Iij+Rij)=bij−dijNij+∑k=1nmijkNik.\frac{d}{dt}N_{ij}=\frac{d}{dt}(S_{ij}+I_{ij}+R_{ij})=b_{ij}-d_{ij}N_{ij}+\sum_{k=1}^{n}m_{ijk}N_{ik}.

This is equation (1.1) and thus Proposition 2 holds here and NijN_{ij} converges for all i,j=1,…,ni,j=1,\ldots,n. As previously, this limit is denoted Nij⋆N_{ij}^{\star}.

This result is similar to those obtained in the SIS model in [8] and the SEIRS model in [7], where movement is equal for all compartments. Because the NijN_{ij} converge, it is possible to use a comparison theorem for asymptotically autonomous limit systems to prove that the DFE is globally asymptotically stable; see Appendix § 11.

The model with vaccination

The model in the presence of a disease for which a vaccine is available is now considered.

Formulation of the model with vaccination

The following assumptions are made additionally to those in Section 2.

  1. Each subpopulation is described by a compartmental SIRS model with vaccination, i.e., an SVIRS model; the model structure in each subpopulation is as in Figure 2. Here, SS, II, RR, VV are the numbers of individuals that are susceptible, infectious, recovered, vaccinated, respectively. The number of vaccinated residents in patch ii present in patch jj at time tt is denoted by Vij(t)V_{ij}(t), and similarly for the other variables.

Flow diagram for one subpopulation of the SVIRS model with vaccination in a given population

Figure 2. Flow diagram for one subpopulation of the SVIRS model with vaccination in a given population NijN_{ij}. There is birth into the S and V compartments when i=ji=j, i.e., bij=0b_{ij}=0 if i≠ji\ne j.

  1. A proportion pijp_{ij} (with 0≤pij<10 \le p_{ij} < 1) of newborns in subpopulation (i,j)(i,j) are vaccinated at birth (since birth only occurs in the patch of residence, this only applies when i=ji=j), and susceptible individuals resident in patch ii are vaccinated at the rate ϕij\phi_{ij} when they are in patch jj. Vaccination reduces disease susceptibility by a factor πij\pi_{ij} with 0≤πij≤10 \le\pi_{ij} \le1, thus 1−πij1-\pi_{ij} is the vaccine efficacy in subpopulation (i,j)(i,j). Thus, in this population, πij=0\pi_{ij}=0 means the vaccine is completely effective at preventing infection, whereas πij=1\pi_{ij}=1 means that the vaccine is useless.

Once infected, vaccinated and unvaccinated individuals transmit the disease at the same rate.

Note that the vaccination model used here is simple. In particular, omitting vaccine waning avoids the standard vaccine-waning mechanism for backward bifurcation in an isolated patch [6] (and rules out backward bifurcation altogether when immunity is permanent, ψij=0\psi_{ij}=0). Although backward bifurcations are known to occur in metapopulation epidemic models (see, e.g., [5]), the point here would be less clear if a backward bifurcation were to occur.

For i,j=1,…,ni,j=1,\ldots,n, the equations for this nn-patch model are

ddtSij=(1−pij)bij+ψijRij+∑k=1nmijkSSik−(dij+ϕij+∑k=1nβikjIkjNjp)Sij(20)\frac{d}{dt}S_{ij}=(1-p_{ij})b_{ij}+\psi_{ij}R_{ij}+\sum_{k=1}^{n}m_{ijk}^{S}S_{ik} -\left(d_{ij}+\phi_{ij}+\frac{\sum_{k=1}^{n}\beta_{ikj}I_{kj}}{N_{j}^{p}}\right)S_{ij} \tag*{(20)}
ddtIij=(Sij+πijVij)∑k=1nβikjIkjNjp−(dij+γij)Iij+∑k=1nmijkIIik(21)\frac{d}{dt}I_{ij}=(S_{ij}+\pi_{ij}V_{ij})\frac{\sum_{k=1}^{n}\beta_{ikj}I_{kj}}{N_{j}^{p}}-(d_{ij}+\gamma_{ij})I_{ij} +\sum_{k=1}^{n}m_{ijk}^{I}I_{ik} \tag*{(21)}
ddtRij=γijIij−(dij+ψij)Rij+∑k=1nmijkRRik(22)\frac{d}{dt}R_{ij}=\gamma_{ij}I_{ij}-(d_{ij}+\psi_{ij})R_{ij}+\sum_{k=1}^{n}m_{ijk}^{R}R_{ik} \tag*{(22)}
ddtVij=pijbij+ϕijSij−(dij+πij∑k=1nβikjIkjNjp)Vij+∑k=1nmijkVVik.(23)\frac{d}{dt}V_{ij}=p_{ij}b_{ij}+\phi_{ij}S_{ij}-\left(d_{ij}+\pi_{ij}\frac{\sum_{k=1}^{n}\beta_{ikj}I_{kj}}{N_{j}^{p}}\right)V_{ij} +\sum_{k=1}^{n}m_{ijk}^{V}V_{ik}. \tag*{(23)}

As with the SIRS model (2.1), if there are initially no infectious individuals in the entire system, then this remains true for all times. As a consequence, system (3.1) is considered with initial conditions

∑j=1nSij(0)>0,∑i,j=1nIij(0)>0,Rij(0)≥0 and Vij(0)≥0\sum_{j=1}^{n} S_{ij}(0)>0,\qquad\sum_{i,j=1}^{n} I_{ij}(0)>0,\qquad R_{ij}(0)\ge0\ \text{and}\ V_{ij}(0)\ge0

for all i,j=1,…,ni,j=1,\ldots,n and can easily be shown to have all solutions in the nonnegative orthant.

As earlier, it is convenient to use vector and matrix notation, so for X∈{S,I,R,V}X\in\{S,I,R,V\}, (1.3) is used. Using the same ordering of first and second index, write Φ=diag⁡(ϕij)\Phi=\operatorname{diag}(\phi_{ij}) the matrix of vaccination rates, Γ=diag⁡(γij)\Gamma=\operatorname{diag}(\gamma_{ij}) the matrix of rates of recovery and Π=diag⁡(πij)\Pi=\operatorname{diag}(\pi_{ij}) the matrix of vaccine-induced reduction of transmissibility. When used with only one index, variables refer to matrices relative to residents of a given patch. For instance, Φi=diag⁡(ϕi1,…,ϕin)\Phi_i=\operatorname{diag}(\phi_{i1},\ldots,\phi_{in}), so Φ=⨁i=1nΦi\Phi=\bigoplus_{i=1}^{n}\Phi_i.

The disease-free equilibrium of the SVIRS model

In order to continue the analysis, system (3.1) is investigated at the disease-free equilibrium (DFE), i.e., where Iij=0I_{ij}=0 for all i,j=1,…,ni,j=1,\ldots,n. In earlier studies [7, 8], the case of status-independent mobility was considered, in which case the population NijN_{ij} in subpopulation (i,j)(i,j) tends to Nij⋆N_{ij}^{\star} as given in Proposition 2. Because of vaccination and status dependent travel, the situation is more complex here. First, similarly to the SIRS model of § 5, at the DFE there are no recovered individuals, since Lemma 3 holds true for (3.1) without any change to the proof. It follows that at the DFE,

0=(1−pij)bij−(dij+ϕij)Sij+∑k=1nmijkSSik0=(1-p_{ij})b_{ij}-(d_{ij}+\phi_{ij})S_{ij}+\sum_{k=1}^{n}m_{ijk}^{S}S_{ik}
0=pijbij+ϕijSij−dijVij+∑k=1nmijkVVik.0=p_{ij}b_{ij}+\phi_{ij}S_{ij}-d_{ij}V_{ij}+\sum_{k=1}^{n}m_{ijk}^{V}V_{ik}.

In vector notation, these equations take the form, for patch i=1,…,ni=1,\ldots,n,

(00)=(MiS−Φi−Di0ΦiMiV−Di)(SiVi)+((Π−Pi)biPibi)(24)\begin{pmatrix}0\\0\end{pmatrix} = \begin{pmatrix} \mathcal{M}_{i}^{S}-\Phi_i-D_i & 0\\ \Phi_i & \mathcal{M}_{i}^{V}-D_i \end{pmatrix} \begin{pmatrix}S_i\\V_i\end{pmatrix} + \begin{pmatrix}(\Pi-P_i)b_i\\P_i b_i \tag*{(24)}\end{pmatrix}

where Pi=diag⁡(pi1,…,pin)P_i=\operatorname{diag}(p_{i1},\ldots,p_{in}) is the diagonal matrix of vaccination at birth proportions for residents of patch ii. From this, the situation at the patch level can be determined.

Lemma 7. For system (3.1), the DFE of residents in patch i=1,…,ni=1,\ldots,n is unique and given by

Si⋆=(Φi+Di−MiS)−1(Π−Pi)bi≫0,Ii⋆=Ri⋆=0,S_i^{\star}=(\Phi_i+D_i-\mathcal{M}_i^{S})^{-1}(\Pi-P_i)b_i\gg0,\qquad I_i^{\star}=R_i^{\star}=0,
Vi⋆=(Di−MiV)−1(ΦiSi⋆+Pibi).(25)V_i^{\star}=(D_i-\mathcal{M}_i^{V})^{-1}(\Phi_iS_i^{\star}+P_i b_i). \tag*{(25)}

Proof. Since Φi+Di>0\Phi_i+D_i>0, Di>0D_i>0 and MiS,MiV\mathcal{M}_i^{S},\mathcal{M}_i^{V} irreducible, it follows from [3 Proposition 3(4)] that Φi+Di−MiS\Phi_i+D_i-\mathcal{M}_i^{S} and Di−MiVD_i-\mathcal{M}_i^{V} are nonsingular M-matrices with entry-wise positive inverses. Solving (3.2) at the equilibrium gives Si⋆S_i^{\star} and Vi⋆V_i^{\star} as in (3.3). Here, Si⋆S_i^{\star} is entry-wise positive (since pii<1p_{ii}<1) and Vi⋆V_i^{\star} is entry-wise positive provided that Φi\Phi_i has a positive entry or pii>0p_{ii}>0.

The vaccination reproduction number

The reasoning here closely follows that used in Section 2.3. Matrix FF has the same block structure as (2.5) with blocks FijF_{ij} taking the form Fij=diag⁡(fij1,…,fijn)F_{ij}=\operatorname{diag}(f_{ij1},\ldots,f_{ijn}), with

fijq=βijqSiq⋆+πiqViq⋆Nqp⋆,(26)f_{ijq}=\beta_{ijq}\frac{S_{iq}^{\star}+\pi_{iq}V_{iq}^{\star}}{N_q^{p\star}}, \tag*{(26)}

where Nqp⋆=∑ℓ=1n(Sℓq⋆+Vℓq⋆)N_q^{p\star}=\sum_{\ell=1}^{n}(S_{\ell q}^{\star}+V_{\ell q}^{\star}), q=1,…,nq=1,\ldots,n. The following is the pendant of Theorem 5.

Theorem 8. Consider system (3.1) with FF and WW defined by (2.5) and (2.6), the blocks in (2.5) having entries of the form (3.4). Let the system-wide vaccination reproduction number be

Rv=ρ(FW−1).(27)\mathcal{R}_{v}=\rho(FW^{-1}). \tag*{(27)}

Then

  • if Rv<1\mathcal{R}_{v}<1, the DFE is locally asymptotically stable;

  • if Rv>1\mathcal{R}_{v}>1, the DFE is unstable, the disease is uniformly persistent, and there exists at least one endemic equilibrium point.

Proof. The local stability of the DFE follows directly from [57 Theorem 2]. Furthermore, because the state space remains positively invariant, point dissipative, and the demographic processes bound the population, the proof of uniform persistence and the existence of an endemic equilibrium point follows the same arguments as those detailed in the proof of Theorem 5 via standard uniform persistence theory [56 Theorem 4.6].

As in § 5.4, it is useful to compute the vaccination number that holds when patches are isolated from each other, i.e., when there is no movement. This value is indeed used in numerical simulations to set some of the parameters. Suppose that patches are isolated; as in Section 2.4, the model can be considered with no subpopulations, as in the absence of movement, the number of visitors to a patch is either identically zero or goes to zero. Using (3.3) with MiS=MiV=0\mathcal{M}_i^{S}=\mathcal{M}_i^{V}=0, the DFE takes the form

(Sii⋆,Iii⋆,Rii⋆,Vii⋆)=((1−pii)biidii+ϕii,0,0,bii(piidii+ϕii)dii(dii+ϕii)).(S_{ii}^{\star},I_{ii}^{\star},R_{ii}^{\star},V_{ii}^{\star}) = \left( \frac{(1-p_{ii})b_{ii}}{d_{ii}+\phi_{ii}}, 0,0, \frac{b_{ii}(p_{ii}d_{ii}+\phi_{ii})}{d_{ii}(d_{ii}+\phi_{ii})} \right).

The matrix FF reduces to the scalar

Fii=βiiiSii⋆+πiiVii⋆Niip⋆,F_{ii}=\beta_{iii}\frac{S_{ii}^{\star}+\pi_{ii}V_{ii}^{\star}}{N_{ii}^{p\star}},

while Wii−1W^{-1}_{ii} is

Wii−1=1γii+diiW^{-1}_{ii} = \frac{1}{\gamma_{ii}+d_{ii}}

so the vaccination reproduction number in patch ii in isolation is

Rvi(3.1)=βiiγii+diiSii⋆+πiiVii⋆Nii⋆.(28)\mathcal{R}_{vi}^{(3.1)} = \frac{\beta_{ii}}{\gamma_{ii}+d_{ii}}\frac{S_{ii}^{\star}+\pi_{ii}V_{ii}^{\star}}{N_{ii}^{\star}} . \tag*{(28)}

As in the case of the SIRS model, when movement is the same for all epidemiological states, the following result holds.

Lemma 9. If individuals in all disease states travel at the same rate, i.e., M:=MS=MI=MR=MV\mathcal{M}:=\mathcal{M}^{S}=\mathcal{M}^{I}=\mathcal{M}^{R}=\mathcal{M}^{V}, then the number of individuals NijN_{ij} in each subpopulation (i,j)(i,j) is convergent.

Proof. Denote mijk:=mijkS=mijkI=mijkR=mijkVm_{ijk}:=m_{ijk}^{S}=m_{ijk}^{I}=m_{ijk}^{R}=m_{ijk}^{V}. For a given subpopulation (i,j)(i,j),

ddtNij=ddt(Sij+Iij+Rij+Vij)=bij−dijNij+∑k=1nmijkNik.\frac{d}{dt}N_{ij}=\frac{d}{dt}(S_{ij}+I_{ij}+R_{ij}+V_{ij})=b_{ij}-d_{ij}N_{ij}+\sum_{k=1}^{n}m_{ijk}N_{ik}.

This is equation (1.1) and thus Proposition 2 holds here and NijN_{ij} converges for all i,j=1,…,ni,j=1,\ldots,n. As previously, this limit is denoted Nij⋆N_{ij}^{\star}.

The high-mobility limit

Part of the objective of this paper is to consider the differences between the resident-traveller and classic metapopulation models. In order to do so, the classic SVIRS model in Appendix B is used. Basic properties of that model are given in the appendix, with, in particular, the expressions of the reproduction numbers R0(B.1)\mathcal{R}_{0}^{(\mathrm{B.1})} and Rv(B.1)\mathcal{R}_{v}^{(\mathrm{B.1})}. The relation between both model types is investigated computationally in Section 4.

However, before proceeding further, the distinction between the two modelling paradigms is formalised by considering the analytical limit of the vaccination reproduction number as the mobility rate m→∞m\to\infty. Let the movement matrix be parameterized as Mc=mM‾c\mathcal{M}_{c}=m\overline{\mathcal{M}}_{c}, where M‾c\overline{\mathcal{M}}_{c} defines the digraph topology.

In the classic metapopulation framework, the spectral radius of the next-generation matrix Rv(B.1)=ρ(FcWc−1)\mathcal{R}_{v}^{(\mathrm{B.1})}=\rho(F_{c}W_{c}^{-1}) depends on the transition matrix Wc=Γc+Dc−mM‾cW_{c}=\Gamma_{c}+\mathcal{D}_{c}-m\overline{\mathcal{M}}_{c}. Assuming that disease-independent mortality and recovery rates are uniform across patches (di=dd_{i}=d, γi=γ\gamma_{i}=\gamma), it follows that Γc+Dc=(d+γ)I\Gamma_{c}+\mathcal{D}_{c}=(d+\gamma)I. By factoring out mm, the inverse is written as

Wc−1=1m(d+γmI−M‾c)−1.W_{c}^{-1}=\frac{1}{m}\left(\frac{d+\gamma}{m}I-\overline{\mathcal{M}}_{c}\right)^{-1}.

Because M‾c\overline{\mathcal{M}}_{c} defines a strongly connected movement digraph (its columns sum to zero), it has a simple eigenvalue at 00 with right eigenvector uu (the stationary distribution, normalised so that 1Tu=1\mathbf{1}^{T}u=1) and left eigenvector 1T\mathbf{1}^{T}. Using singular perturbation theory for Markov chains (see, e.g., [60 Theorem 2.1] or [34 Chapter II, Equation (1.14)]), as m→∞m\to\infty the inverse Wc−1W_{c}^{-1} converges to a rank-1 multiple of a projection onto the null space of M‾c\overline{\mathcal{M}}_{c}, given by 1d+γu1T\frac{1}{d+\gamma}u\mathbf{1}^{T}.

Furthermore, as m→∞m\to\infty, the disease-free equilibrium states perfectly mix across the network. The proportion of susceptible and vaccinated individuals in every patch converges to the global network averages, denoted sˉ⋆\bar{s}^{\star} and vˉ⋆\bar{v}^{\star}. Therefore, the transmission matrix Fc(m)F_{c}(m) converges to a fully homogenized matrix F‾c=diag⁡(βi(sˉ⋆+πivˉ⋆))\overline{F}_{c}=\operatorname{diag}(\beta_{i}(\bar{s}^{\star}+\pi_{i}\bar{v}^{\star})).

Consequently, the next generation matrix FcWc−1F_{c}W_{c}^{-1} converges to a rank-1 matrix, and its spectral radius converges to the trace of this limiting matrix:

lim⁡m→∞Rv(B.1)=ρ(1d+γF‾cu1T)=tr⁡(1d+γF‾cu1T)=∑i=1nuiβid+γ(sˉ⋆+πivˉ⋆).(29)\lim_{m\to\infty}\mathcal{R}_{v}^{(\mathrm{B.1})} =\rho\left(\frac{1}{d+\gamma}\overline{F}_{c}u\mathbf{1}^{T}\right) =\operatorname{tr}\left(\frac{1}{d+\gamma}\overline{F}_{c}u\mathbf{1}^{T}\right) =\sum_{i=1}^{n}u_{i}\frac{\beta_{i}}{d+\gamma}(\bar{s}^{\star}+\pi_{i}\bar{v}^{\star}). \tag*{(29)}

Thus, infinite mobility in the classic model completely homogenizes the system. The system-wide reproduction number becomes a weighted arithmetic mean of the transmission potentials, driven entirely by the globally averaged vaccine coverage vˉ⋆\bar{v}^{\star}. If a single patch completely halts vaccination, its true isolated potential exceeds unity (Rvi(B.1)>1\mathcal{R}_{vi}^{(\mathrm{B.1})}>1). However, if the surrounding patches are highly vaccinated, the globally averaged susceptibility (sˉ⋆+πivˉ⋆)(\bar{s}^{\star}+\pi_{i}\bar{v}^{\star}) can still fall to a level where the trace drops below unity, erroneously allowing analytical eradication in the classic model.

Conversely, in the resident-traveller framework, the population does not assimilate into destination cohorts. Residents of an unvaccinated patch ii (Φi=0\Phi_{i}=0, Pi=0P_{i}=0) distribute across the digraph according to their specific stationary visiting distribution u(i)u^{(i)}, but remain completely unvaccinated (Si⋆=Ni⋆S_{i}^{\star}=N_{i}^{\star}, Vi⋆=0V_{i}^{\star}=0) regardless of how heavily other patches vaccinate their own residents. Because the next-generation matrix FW−1FW^{-1} tracks infections by residency, the principal submatrix FiiWii−1F_{ii}W_{ii}^{-1} corresponding to transmission exclusively among unvaccinated residents of patch ii (or, more generally, the principal submatrix corresponding to the set of all non-vaccinating patches) remains strictly positive. By the Perron-Frobenius theorem (see, e.g., [28 Theorem 8.1.18]), the spectral radius of the full system is bounded below by the spectral radius of any principal submatrix:

Rv≥ρ(FiiWii−1),(30)\mathcal{R}_{v}\geq\rho\left(F_{ii}W_{ii}^{-1}\right), \tag*{(30)}

where FiiWii−1F_{ii}W_{ii}^{-1} represents transmission among the susceptible residents of patch ii. Whenever this submatrix spectral radius (or the spectral radius of the subnetwork of unvaccinated patches) exceeds unity, global eradication is impossible regardless of the vaccination rate in the vaccinating patches, in contrast to the classic model where high vaccination in some patches can drive the system-wide reproduction number below unity. Note that, unlike the classic result, this lower bound is not a limit as m→∞m \to\infty: it holds for every movement rate. As mm increases, residents of patch ii spend a growing fraction of their time in other patches, where they make up a smaller share of the population present, so the entries of FiiF_{ii} decrease and ρ(FiiWii−1)\rho(F_{ii}W_{ii}^{-1}) can fall below unity (see § 7.3). The bound is therefore most informative at low to moderate mobility.

Investigation of specific vaccination situations

A few special cases are now considered assuming different types of vaccination patterns in the patches. These scenarios are investigated analytically or numerically, with a focus on comparing the outcomes predicted by the resident-traveller model versus the classic metapopulation model.

Setting of the numerical work

In the cases where numerical results are sought, three graph architectures are considered.

  1. Complete graph – In this situation, each vertex is connected to the n−1n - 1 other vertices.

  1. Lattice – In this case, patches are laid out on a lattice and residents of a given patch are only able to travel to patches immediately “north”, “south”, “east” and “west” of their patch of residence. A torus is considered, with the “top” connecting to the “bottom” and “right” connecting to the “left”. This is an approximation of a two-dimensional diffusion process with periodic boundary conditions.

  1. Ring graph – Patches form a ring, with access to exactly the two adjacent patches. This is an approximation of a one-dimensional diffusion process with periodic boundary conditions.

Note that only graph architectures that are regular, i.e., have an equal number of neighbours for all patches, are considered. This avoids artificial demographic shifts in the classic metapopulation model where patches with more neighbours would accumulate individuals when a uniform total leaving rate is divided equally among outgoing edges. The simple lattice with reflecting boundaries is therefore omitted from theoretical comparisons, though non-equal degrees will be considered in the real-world example.

Baseline values of parameters are generally chosen as in Table 1. In all simulations on the synthetic networks (n=49n = 49 patches), infectious individuals are assumed to travel at 10% of the rate of susceptible, recovered and vaccinated individuals (MI=0.1MS\mathcal{M}^{I} = 0.1\mathcal{M}^{S}, MR=MV=MS\mathcal{M}^{R} = \mathcal{M}^{V} = \mathcal{M}^{S}; similarly McI=0.1McS\mathcal{M}^{I}_{c} = 0.1\mathcal{M}^{S}_{c} in the classic model). Values of the vaccination rate ϕij\phi_{ij} and birth rate biib_{ii} are computed

ParameterInterpretationBaseline value
1/γij1/\gamma_{ij}Duration infection21 days
1/ψij1/\psi_{ij}Duration immune period1 year
ϕij\phi_{ij}Vaccination rate10−510^{-5}–10−210^{-2} day−1^{-1} (or computed)
1−πij1-\pi_{ij}Vaccine efficacy0.7 (70%)
cij∗c^{*}_{ij}Vaccine coverage at DFE0.8 (80%, reference value)
biib_{ii}Birth rate for residents of iiComputed (diNird_{i}N^{r}_{i})
1/dij1/d_{ij}Life expectancy70 years
NirN^{r}_{i}Resident population of patch ii10,000
mmTotal outbound movement rate10−310^{-3}–10−110^{-1} day−1^{-1}
R0(2.1)R^{(2.1)}_{0}Return rate to home patch0.09 day−1^{-1}
R0(2.1)R^{(2.1)}_{0}Isolated basic reproduction number1.2
Rvi(3.1)R^{(3.1)}_{v_{i}}Isolated vaccination reproduction number (at c∗=0.8c^{*}=0.8)0.528 (reference value)

Table 1. Parameters and values in the model and baseline values used in simulations. Computed values are deduced from values of other parameters. The coverage c∗c^{*} and the corresponding isolated Rvi(3.1)\mathcal{R}^{(3.1)}_{v_i} are reference values illustrating the effect of vaccination in an isolated patch; in the simulations on synthetic networks, the vaccination rate ϕ\phi is varied directly over the range indicated.

from assumptions about quantities they determine, as explained now.

Vaccine coverage Obtaining realistic parameter values for the vaccination rate is difficult. Indeed, the model does not include age classes, which are the points in life history with known vaccination action. Instead, a continuous vaccination rate ϕij\phi_{ij} is incorporated alongside a birth vaccination proportion pijp_{ij}. Vaccination applies to the entire susceptible population, such as may approximate the vaccination for measles following an outbreak. The vaccination rate is then set indirectly, using an observable consequence, namely, the vaccine coverage, i.e., the proportions cij=Vij/Nij∈[0,1]c_{ij} = V_{ij}/N_{ij} \in[0,1] of individuals in the population that are vaccinated. This is particularly relevant since data are usually given in terms of vaccine coverage. At the DFE (25), this can be written as

cij∗=Vij∗Sij∗+Vij∗(31)c^{*}_{ij} = \frac{V^{*}_{ij}}{S^{*}_{ij} + V^{*}_{ij}} \tag*{(31)}

but this estimation only accurately reflects the true vaccination coverage in a completely susceptible population, which does not hold when an endemic state emerges for Rv>1\mathcal{R}_{v} > 1. Also, even when Rv<1\mathcal{R}_{v} < 1, (3.3) involves ϕij\phi_{ij} through a nontrivial matrix inverse, so obtaining ϕ\phi using (4.1) involves a difficult inverse problem. Since, in the numerics, vaccination occurs only in the home patch and no vaccination at birth is assumed (Pi=0P_{i} = 0), the expression for (4.1) can be simplified using (3.3), giving that at the DFE, Vi∗=(Di−MiV)−1ϕiSi∗V^{*}_{i} = (\mathcal{D}_{i} - \mathcal{M}^{V}_{i})^{-1}\phi_{i}S^{*}_{i}. Using the diagonal structure of Φi\Phi_{i} and dividing numerator and denominator by Sii∗S^{*}_{ii} gives

cii∗=[(Di−MiV)−1]iiϕii1+[(Di−MiV)−1]iiϕii.c^{*}_{ii} = \frac{\left[(\mathcal{D}_{i} - \mathcal{M}^{V}_{i})^{-1}\right]_{ii}\phi_{ii}}{1 + \left[(\mathcal{D}_{i} - \mathcal{M}^{V}_{i})^{-1}\right]_{ii}\phi_{ii}}.

This implies that complete vaccine coverage cij∗=1c^{*}_{ij} = 1 can never be achieved, except in the limit as ϕii→∞\phi_{ii} \to\infty. More generally, whenever vaccination at birth is not complete (pij<1p_{ij}<1), all non-vaccinated newborns spend some time, however short, in the susceptible compartment.

Setting birth rates – The birth rate bb is chosen such that the demographic component plays a minimal role during simulations and each patch maintains a target equilibrium resident population Nir⋆N_i^{r\star}. In the resident-traveller model, births occur only in the home patch (bij=0b_{ij}=0 for i≠ji\ne j). With uniform per capita death rate dij=dd_{ij}=d, setting bii=dNir⋆b_{ii}=dN_i^{r\star} determines the unique demographic equilibrium Ni⋆=(Di−Mi)−1biN_i^\star=(D_i-M_i)^{-1}b_i, which satisfies ∑j=1nNij⋆=Nir⋆\sum_{j=1}^nN_{ij}^\star=N_i^{r\star}. In the absence of disease (I=R=0I=R=0) and with MS=MV=M\mathcal{M}^S=\mathcal{M}^V=\mathcal{M}, the DFE components satisfy Si⋆+Vi⋆=Ni⋆S_i^\star+V_i^\star=N_i^\star, so that starting with N(0)=N⋆N(0)=N^\star, the subpopulation sizes remain constant at the demographic level. When movement of infectious individuals differs from that of other states (MI≠MS\mathcal{M}^I\ne\mathcal{M}^S) so that Lemma 6 does not hold during an outbreak, there are only small transient variations of the spatial distribution Ni(t)N_i(t) while the total resident population Nir(t)=Nir⋆N_i^r(t)=N_i^{r\star} remains strictly constant.

Vaccination only in the patch of residence Although the model allows for vaccination in all subpopulations, it is supposed in numerical work that individuals are vaccinated only when present in their place of residence, i.e., ϕij=0\phi_{ij}=0 for i≠ji\ne j. Then the matrix Φ\Phi takes the form

Φ~=diag⁡(Φ~1,…,Φ~n),\widetilde{\Phi}=\operatorname{diag}(\widetilde{\Phi}_1,\ldots,\widetilde{\Phi}_n),

where Φ~i\widetilde{\Phi}_i is the matrix with the only nonzero entry (i,i)(i,i) equal to ϕii\phi_{ii}. Consider the iith diagonal block (Di−MiV)−1(D_i-\mathcal{M}_i^V)^{-1} in (D−MV)−1(D-\mathcal{M}^V)^{-1} and let [(Di−MiV)−1]ij[(D_i-\mathcal{M}_i^V)^{-1}]_{ij} be the (i,j)(i,j) entry in that matrix. Because of the structure of Φ~i\widetilde{\Phi}_i, the matrix (Di−MiV)−1Φ~i(D_i-\mathcal{M}_i^V)^{-1}\widetilde{\Phi}_i has only one nonzero column that consists of ϕii\phi_{ii} times the entries in the iith column of (Di−MiV)−1(D_i-\mathcal{M}_i^V)^{-1} and thus Vi⋆V_i^\star takes the form

Vi⋆=([(Di−MiV)−1]1i,…,[(Di−MiV)−1]ni)TϕiiSi⋆.(32)V_i^\star=\left([(D_i-\mathcal{M}_i^V)^{-1}]_{1i},\ldots,[(D_i-\mathcal{M}_i^V)^{-1}]_{ni}\right)^{T}\phi_{ii}S_i^\star. \tag*{(32)}

Note that this assumption is realistic from a public health perspective. Routine childhood immunisation schedules and publicly funded vaccination programmes are typically administered through local health systems and registries tied to an individual’s place of residence, rather than to temporary visitors.

Vaccination in only one patch

Suppose that vaccination takes place only in one patch. To consider implications on the model, assume without loss of generality that the patch vaccinating is patch 1. In this case, Φ=diag⁡(Φ1,0,…,0)\Phi=\operatorname{diag}(\Phi_1,0,\ldots,0) and Pi=0P_i=0 for i=2,…,ni=2,\ldots,n. Using this in (3.3), the DFE takes the form

S1⋆=(Φ1+D1−M1S)−1(I−P1)b1,I1⋆=R1⋆=0,S_1^\star=(\Phi_1+D_1-\mathcal{M}_1^S)^{-1}(I-P_1)b_1,\qquad I_1^\star=R_1^\star=0,
V1⋆=(D1−M1V)−1(Φ1S1⋆+P1b1),V_1^\star=(D_1-\mathcal{M}_1^V)^{-1}(\Phi_1S_1^\star+P_1b_1),

and, for i=2,…,ni=2,\ldots,n,

Si⋆=(Di−MiS)−1bi,Ii⋆=Ri⋆=Vi⋆=0.S_i^\star=(D_i-\mathcal{M}_i^S)^{-1}b_i,\qquad I_i^\star=R_i^\star=V_i^\star=0.

Thus, V⋆=(V1⋆T,0,…,0)TV^\star=(V_1^{\star T},0,\ldots,0)^T and diagonal entries in FijF_{ij} take the form (3.4) when i=1i=1 and (2.7) when i=2,…,ni=2,\ldots,n.

Numerically, the case investigated has patch 1 vaccinating (which can be thought of as the patch in position (1,1)(1,1) in a torus grid) and is such that, in isolation, Rv1(3.1)<1\mathcal{R}_{v_1}^{(3.1)}<1, while none of the others vaccinate and all have Rvi(3.1)>1\mathcal{R}_{v_i}^{(3.1)}>1 for i≠1i\ne1. This setting addresses the first question posed in the Introduction: can patch 1 contribute to disease eradication in the other patches, and can it protect its own population?

In the classic metapopulation framework (B.1), the vaccination matrix takes the form Φc=diag⁡(ϕ1,0,…,0)\Phi_c=\operatorname{diag}(\phi_1,0,\ldots,0). Using (B.2), the number of vaccinated individuals at the disease-free equilibrium is given by Vc⋆=(Dc−Mc)−1ΦcSc⋆V_c^\star=(D_c-\mathcal{M}_c)^{-1}\Phi_cS_c^\star. Since Dc−McD_c-\mathcal{M}_c is an irreducible, nonsingular M-matrix for a strongly connected digraph, its inverse is strictly positive. Consequently, the product (Dc−Mc)−1ΦcSc⋆(D_c-\mathcal{M}_c)^{-1}\Phi_cS_c^\star yields a vector where every single entry is strictly positive. This means that at the DFE, every patch’s resident population in the classic model becomes partially vaccinated because visitors to patch 1 are vaccinated and subsequently assimilated into other patches. This alters the basic reproduction number computation for all patches globally. In contrast, the resident-traveller model ensures Vi⋆=0V_i^\star=0 for all i≠1i\ne1, preserving the fully susceptible status of residents of non-vaccinating patches.

To visualize this divergence, a parameter sweep of the system-wide reproduction number Rv\mathcal{R}_v is performed across varying movement rates (mm) and vaccination rates (ϕ\phi) for the scenario where only one patch vaccinates. For the remaining 48 untreated patches in the 7×77\times7 grid, the isolated local reproduction number is configured to R0=1.2\mathcal{R}_0=1.2.

In the resident-traveller formulation, vaccination is strictly governed by residence. Thus, individuals from the 48 untreated patches are not subject to vaccination when travelling to the vaccinating patch. Because these untreated patches remain fully susceptible reservoirs, the overall Rv\mathcal{R}_v is analytically bounded below by the spectral radius of the unvaccinated subnetwork, remaining approximately constant at Rv≈1.20\mathcal{R}_v\approx1.20. Consequently, the answer to the first part of the question is negative in the resident-traveller framework: for the parameter range considered, it is impossible to eradicate the disease globally by vaccinating only a single patch.

This numerical observation can be formalized analytically. In the resident-traveller formulation, a lack of vaccination in a specific jurisdiction inherently places a lower bound on the system-wide reproduction number, which prevents eradication whenever this bound exceeds unity. Let R~0k=ρ(FkkWkk−1)\widetilde{\mathcal{R}}_0^k=\rho(F_{kk}W_{kk}^{-1}) be defined as the digraph-adjusted basic reproduction number for residents of patch kk in the absence of vaccination. This reflects transmission within the interconnected system, as opposed to the fully isolated R0k\mathcal{R}_0^k defined in Section 2.4.

Theorem 10. Assume vaccination only occurs in the patch of residence (ϕij=0\phi_{ij}=0 for i≠ji\ne j). Let patch kk be a non-vaccinating patch, meaning there is no ongoing vaccination (ϕkk=0\phi_{kk}=0) and no vaccination at birth (pkk=0p_{kk}=0). Then the system-wide vaccination reproduction number satisfies Rv≥R~0k\mathcal{R}_v\ge\widetilde{\mathcal{R}}_0^k.

Proof. Because patch kk does not vaccinate its residents, the kk-th block of the ongoing vaccination matrix is Φk=0\Phi_k=0. Furthermore, since there is no vaccination at birth for residents of patch kk, the birth vaccination proportion matrix satisfies Pkbk=0P_kb_k=0. It follows that any initially vaccinated residents of this patch eventually die, i.e., Vk→0V_k\to0 at the DFE.

From Lemma 7, the equilibrium vaccinated population for residents of patch kk at the disease-free equilibrium is given by Vk⋆=(Dk−MVk)−1(ΦkSk⋆+Pkbk)V_k^\star=(D_k-M_{V_k})^{-1}(\Phi_kS_k^\star+P_kb_k). Substituting Φk=0\Phi_k=0 and Pkbk=0P_kb_k=0 yields Vk⋆=0V_k^\star=0, meaning residents of patch kk remain fully susceptible (Sk⋆=Nk⋆S_k^\star=N_k^\star).

Consequently, the kk-th diagonal block FkkF_{kk} of the transmission matrix FF has entries

fkkq=βkkqSkq⋆+πkqVkq⋆Nq⋆=βkkqNkq⋆Nq⋆.f_{kkq}=\beta_{kkq}\frac{S_{kq}^\star+\pi_{kq}V_{kq}^\star}{N_{q}^\star}=\beta_{kkq}\frac{N_{kq}^\star}{N_{q}^\star}.

This matches the localized transmission block in the complete absence of vaccination.

The global next-generation matrix is FW−1FW^{-1}. Because W−1W^{-1} is block-diagonal with respect to the patch of residence, the non-negative matrix FW−1FW^{-1} contains the principal submatrix FkkWkk−1F_{kk}W_{kk}^{-1}, which isolates transmission exclusively among residents of patch kk.

By the Perron-Frobenius theorem and standard properties of non-negative matrices, the spectral radius of a matrix is bounded below by the spectral radius of any of its principal submatrices. Therefore,

Rv=ρ(FW−1)≥ρ(FkkWkk−1)=R~0k.\mathcal{R}_v=\rho(FW^{-1})\ge\rho(F_{kk}W_{kk}^{-1})=\widetilde{\mathcal{R}}_0^k.

Thus, if the local digraph-adjusted reproduction number of the untreated patch exceeds unity (R~0k>1\widetilde{\mathcal{R}}_0^k>1), global eradication (Rv<1\mathcal{R}_v<1) is impossible, regardless of how heavily other patches vaccinate. Regarding the second part of the question, vaccination in patch 1 only partially protects its own residents: while vaccinated residents of patch 1 have reduced susceptibility (π11<1\pi_{11}<1), which lowers their infection rate compared to unvaccinated patches, imperfect vaccine efficacy (π11>0\pi_{11}>0) and ongoing mixing with infectious visitors from the 48 non-vaccinating patches prevent patch 1 from completely shielding its population as long as the disease remains endemic in the network. In contrast, as illustrated in Figure 3, the classic model applies vaccination by current physical location. When individuals from the untreated patches travel to the single vaccinating patch, they are vaccinated at its local rate and subsequently return or move on as vaccinated individuals. As the movement rate mm increases, this export of vaccinated individuals reduces the proportion of susceptibles across all patches, driving the global Rv\mathcal{R}_v below the eradication threshold (Rv<1\mathcal{R}_v<1).

System-wide reproduction number heatmap

Figure 3. System-wide reproduction number Rv\mathcal{R}_v for the classic metapopulation model when only one patch vaccinates. The dashed contour line indicates the critical eradication threshold (Rv=1\mathcal{R}_v=1). The resident model (not shown) maintains a nearly constant Rv≈1.20\mathcal{R}_v\approx1.20 across the same parameter space because untreated patches remain fully susceptible reservoirs.

Figure 4 shows the effect of this difference by plotting the epidemic curve I(t)I(t) for identical parameters (m=0.01 day−1m=0.01\ \text{day}^{-1}, ϕ11=1/365 day−1\phi_{11}=1/365\ \text{day}^{-1}) where Rv(B.1)<1<Rv\mathcal{R}_v^{(\mathrm{B.1})}<1<\mathcal{R}_v, alongside the difference in cumulative incidence across the parameter space. In this regime, the classic model predicts only a small transient wave before the outbreak dies out. Conversely, the resident-traveller model resolves the large epidemic wave across the unprotected patches, resulting in substantially higher cumulative infections.

Figure 4. Epidemic curve time-series I(t)I(t) for identical parameters (m=0.01m = 0.01 day−1^{-1}, ϕ11=1/365\phi_{11} = 1/365 day−1^{-1}) where Rv(B.1)<1<Rv(3.1)R_v^{(\mathrm{B.1})} < 1 < R_v^{(3.1)} (a), and the corresponding difference in cumulative new infections between the resident-traveller and classic models over a one-year period (b). The outbreak originates in an untreated patch in the 7×77 \times7 grid.

Vaccination in all but one patch

Suppose that individuals are vaccinated in all patches except one. Without loss of generality, assume that the non-vaccinating patch is patch nn. Patch nn could be the core of a city or region where individuals refuse to be vaccinated. Patch nn does not vaccinate and has Rvn(3.1)>1\mathcal{R}_{vn}^{(3.1)}>1; all others vaccinate and have Rvi(3.1)<1\mathcal{R}_{vi}^{(3.1)}<1 for i≠ni\ne n. This scenario addresses the second question posed in the Introduction: what happens to the global system when a single jurisdiction ceases to vaccinate while surrounding locations continue to vaccinate?

For the classic metapopulation model, assuming patch nn does not vaccinate, the vaccination matrix takes the form Φc=diag⁡(ϕ1,…,ϕn−1,0)\Phi_c = \operatorname{diag}(\phi_1,\ldots,\phi_{n-1},0). Because individuals from patch nn travel to the other n−1n-1 vaccinating patches, they become subject to vaccination there. Mathematically, the nn-th component of the DFE vaccinated population Vc∗\mathbf{V}_c^* is given by the inner product of the nn-th row of (Dc−Mc)−1(\mathcal{D}_c-\mathcal{M}_c)^{-1} and the vector ΦcSc∗\Phi_c\mathbf{S}_c^*. Due to the strict positivity of (Dc−Mc)−1(\mathcal{D}_c-\mathcal{M}_c)^{-1}, this component is strictly positive. Consequently, the population of patch nn acquires vaccinated individuals through spatial mixing, which suppresses its effective susceptibility in the interconnected system. In the resident-traveller formulation, by contrast, residents of patch nn are not vaccinated when visiting other patches and remain a fully susceptible cohort.

Figure 5 illustrates the system-wide reproduction number RvR_v across a parameter space of movement rates and vaccination rates in the surrounding locations for both models. In the classic metapopulation model (left panel), even modest vaccination rates in the surrounding locations (ϕ≳1.5×10−5\phi\gtrsim1.5 \times10^{-5} day−1^{-1}) create a “safe” region (Rv<1R_v < 1) across all movement rates, suggesting that immunizing surrounding patches easily compensates for a core patch’s refusal to vaccinate. However, the resident-traveller formulation (right panel) answers the second question from the Introduction differently: for low to moderate movement rates (m≲0.02m \lesssim0.02 day−1^{-1}), where residents spend most of their time in their home patch so that R~0n>1\widetilde{R}_{0}^{n} > 1, RvR_v strictly exceeds unity regardless of how heavily the surrounding locations vaccinate. Only when the movement rate is sufficiently high (m>0.02m > 0.02 day−1^{-1}) that residents of the unvaccinated core patch spend a large fraction of their time diluted among the vaccinated surrounding locations (R~0n<1\widetilde{R}_{0}^{n} < 1) can high vaccination in those locations bring the system-wide RvR_v below unity.

(a) Classic metapopulation

(b) Resident-traveller

Figure 5. System-wide vaccination reproduction number Rv\mathcal{R}_{v} when a single core patch refuses to vaccinate (ϕn=0\phi_{n}=0) while all surrounding locations vaccinate at rate ϕ\phi in the classic model (a) and the resident-traveller model (b). The dashed white contour indicates the eradication threshold (Rv=1\mathcal{R}_{v}=1).

Further considerations on digraph topology

To illustrate how the difference between the two models depends on the movement network, the difference in the system-wide vaccination reproduction number ΔRv=Rv(3.1)−Rv(B.1)\Delta R_v = R_v^{(3.1)} - R_v^{(\mathrm{B.1})} is evaluated as a function of the movement rate mm at ϕ=0.01\phi= 0.01 day−1^{-1}. Figure 6 compares ΔRv\Delta R_v across the three regular digraph topologies: complete, lattice and ring.

(a) Single vaccinating patch

(b) Vaccination of surrounding locations (core refuses)

Figure 6. Difference ΔRv=Rv(3.1)−Rv(B.1)\Delta\mathcal{R}_{v}=\mathcal{R}_{v}^{(3.1)}-\mathcal{R}_{v}^{(B.1)} in the system-wide vaccination reproduction number as a function of movement rate across three digraph topologies (ϕ=0.01\phi=0.01 day−1^{-1}). (a) scenario where only a single patch vaccinates, (b) scenario where all but one patch vaccinate.

The left panel displays a scenario where only a single patch vaccinates (Section 4.2), while the right panel displays a scenario where a single core patch refuses to vaccinate (as in Section 4.3).

In the case where all but one patch vaccinate (Figure 6b), all three digraph types exhibit very similar ΔRv\Delta R_v curves across mm, showing that the influence of graph topology is minimal when 48 of 49 patches vaccinate and the main driver of ΔRv\Delta R_v is the movement rate.

On the other hand, for the single vaccinating patch scenario (Figure 6a), the density of the digraph has a strong influence on the difference between reproduction numbers. First, note that in all cases shown, Rv(3.1)>1R_v^{(3.1)} > 1, whereas Rv(B.1)R_v^{(\mathrm{B.1})} is smaller than 1 except for the dashed part of the curves at lower movement rates.

In complete graphs (representing direct connections between all patches), Rv(B.1)R_v^{(\mathrm{B.1})} drops well below Rv(3.1)R_v^{(3.1)} even at small movement rates, because visitors from every patch travel directly to the vaccinating patch. Conversely, ring graphs (representing nearest-neighbour movement) preserve spatial heterogeneity up to much higher movement rates before the classic model homogenises vaccine coverage across the network.

An example inspired by the 2003–2004 West and Central African polio crisis

The motivating example of the vaccination crisis that took place in 2003 and 2004 in northern Nigeria and led to the temporary suspension of oral polio vaccination is now considered. The intent here is to illustrate how such a study can be conducted; parametrisation of this particular problem is indeed very complicated and this is more a thought experiment than a proper epidemiological study. This example also serves to illustrate how different the behaviour is when classic and resident-traveller metapopulation models are used with more realistic graph architectures.

Movement data is scarce in general and even more so in the region under consideration. Except for air travel, most transportation in Africa is handled by informal transport or paratransit, i.e., operations run by micro-companies or individuals [26, 36]. Also, trade between countries in the area is extremely active but also very poorly documented: it is estimated that “informal cross border trade”, while escaping trade duties, amounts to upward of 40% of the GDP and 55% of employment in Africa [23, 42, 44]. For this reason, the volume of transportation happening between countries in the area affected by the 2003–2004 polio crisis is nigh impossible to estimate, although it can be assumed to be high. The following scenario thus uses assumptions on the network and vaccination uptake.

Figure 7 shows the percentage of children who had received the full three-doses OPV regimen in 2002 in Nigeria and countries at most two land borders away from Nigeria. Côte d’Ivoire and Ghana were added to the list although they are three borders away because of their close proximity to Nigeria. Specifically, Algeria, Benin, Burkina Faso, Cameroon, Central African Republic, Chad, Congo, Côte d’Ivoire, Equatorial Guinea, Gabon, Ghana, Libya, Mali, Niger, Nigeria, Sudan and Togo are considered. The yearly reported incidence from 2000 to 2005 for these countries is shown in Figure 8. For population and health data, United Nations, World Bank and World Health Organization sources are used.

Map of vaccination coverage and transportation networks in West and Central Africa

Figure 7. Percentage of children covered with the full three-doses regimen of oral polio vaccine in 2002 in countries at most two land borders away from Nigeria. Ghana and Côte d’Ivoire are also included because of their proximity to Nigeria. Because 2002–2003 data for Nigeria were missing in the series, its 2004 value (37%) is plotted here. Coverage values across these plotted countries range from 13% (Central African Republic) to 99% (Ghana). The red network shows the graph of shared land borders representing short-range ground movement, while the blue network represents long-range air travel connections. Note that the graph vertices are country centroids, not capitals as in the actual transport graph.

Reported cases of polio from 2000 to 2005 for the 17 countries in the region of interest

Figure 8. Reported cases of polio from 2000 to 2005 for the 17 countries in the region of interest. Nigeria numbers are on the right axis. Note that the data consists of discrete annual points, which have been joined by lines for visualization.

As indicated, transportation data is scarce for this region, so the movement matrix is constructed as a combination of inferred short and long range travel matrices, both of which are derived from basic principles and hypotheses rather than actual data. Short range ground travel is incorporated through the connection graph of shared land borders that has been derived, representing adjacent countries. Long range air travel is considered through the global air transportation network, using the wikipediaGATN package [4]. For the air network, a base weight AjiA_{ji} from country jj to country ii is calculated as the sum of the number of airlines operating on all individual airport-to-airport routes connecting the two countries. For the ground network, the base weight BjiB_{ji} is simply 1. if jj and ii share a border, and 0 otherwise. To compute the final connection rates for both networks, these base weights are multiplied by a gravity model that incorporates population sizes, distance and relative wealth. The standard gravity weight between countries jj and ii is taken as Njr⋆Nir⋆/Dji2N_j^{r\star}N_i^{r\star}/D_{ji}^{2}, where Nir⋆N_i^{r\star} denotes the population size of country ii and DjiD_{ji} is the great circle distance between their capital cities. To account for economic attractiveness, this gravity weight is multiplied by the destination’s share of the combined GDP per capita of the two countries. Consequently, travel from a poorer origin to a wealthier destination forms a larger proportion of the bidirectional gravity weight. The final directional rate of movement from jj to ii for a given transport mode with base weight WjiW_{ji} (AjiA_{ji} for air or BjiB_{ji} for ground) is therefore computed as

wjimode=WjiNjr⋆Nir⋆Dji2GiGj+Giw_{ji}^{\mathrm{mode}} = W_{ji}\frac{N_j^{r\star}N_i^{r\star}}{D_{ji}^{2}}\frac{G_i}{G_j+G_i}

where GG stands for GDP per capita, and mode ∈{air,ground}\in\{\text{air},\text{ground}\}. The combined per-day rate of movement wjiw_{ji} from jj to ii is then defined as a linear combination of the air and ground connection rates,

wji=αairwjiair+αgroundwjiground,w_{ji} = \alpha_{\mathrm{air}}w_{ji}^{\mathrm{air}}+\alpha_{\mathrm{ground}}w_{ji}^{\mathrm{ground}},

where αair\alpha_{\mathrm{air}} and αground\alpha_{\mathrm{ground}} are scaling parameters, taken as 2.26×10−122.26 \times10^{-12} and 9.01×10−129.01 \times10^{-12}, respectively. These values were chosen to scale the combined rate of movement wjiw_{ji} such that its maximum value does not exceed 0.05 day−1^{-1}. As a result, the non-zero values of wjiw_{ji} range from approximately 4.59×10−74.59 \times10^{-7} to 0.0438 day−1^{-1}.

In the resident-traveller model, residents of country ii leave their home patch for country jj at rate mjiij=wjim_{ji}^{ij}=w_{ji} and return to their home patch at rate mijii=1/14m_{ij}^{ii}=1/14 day−1^{-1} (corresponding to an average trip duration of 14 days), with infectious individuals travelling at 10% of the rate of other compartments (MI=0.1MSM^I=0.1M^S). To reflect socioeconomic and health-system heterogeneity across the 17 countries, the recovery period 1/γi1/\gamma_i is drawn uniformly in [18,24][18,24] days (consistent with the ∼20.4\sim20.4-day mean duration of faecal poliovirus excretion in non-immunes [20]), the isolated reproduction number R0i\mathcal{R}_{0i} is scaled linearly from 1.5 (highest GDP per capita) to 3.5 (lowest GDP per capita), and the relative susceptibility of vaccinated individuals πi\pi_i is scaled linearly from 0.1 (highest vaccine coverage) to 0.4 (lowest vaccine coverage), with ongoing vaccination rates set to ϕi=ci1−ci(ψi+di)\phi_i=\frac{c_i}{1-c_i}(\psi_i+d_i) (where cic_i is the country’s reported coverage, capped at 0.99) and initial conditions seeded with 10 infectious individuals in Nigeria. Note that while estimates of the basic reproduction number of wild poliovirus in a completely naive population are higher, typically ranging from 6 to 14 in developing-country settings (6 to 11 in medium-tier settings including northern Nigeria) [20, 19], our simulation initialises all non-vaccinated under-5 children as fully susceptible (R(0)=0R(0)=0) and does not explicitly model pre-existing natural or campaign-induced immunity from prior endemic circulation and supplementary immunisation activities. The range R0i∈[1.5,3.5]\mathcal{R}_{0i}\in[1.5,3.5] therefore acts as an effective reproduction number under which well-vaccinated countries sit near or below the local control threshold while under-vaccinated countries remain supercritical. Taking Ni⋆N_i^\star to be the under-5 population of country ii in the data, the home birth rate bii=diNi⋆b_{ii}=d_iN_i^\star (with bij=0b_{ij}=0 for j≠ij\ne i) and equilibrium resident distribution Nij⋆=(Di/λi)−1biN_{ij}^\star=(D_i/\lambda_i)^{-1}b_i satisfy ∑j=117Nij⋆=Nir⋆\sum_{j=1}^{17}N_{ij}^\star=N_i^{r\star}, so that (1.6) takes the form

ddtNi=(Mi−Di)Ni+bi=(Mi−Di)(Ni−Ni⋆),\frac{d}{dt}N_i=(\mathcal{M}_i-\mathcal{D}_i)N_i+b_i=(\mathcal{M}_i-\mathcal{D}_i)(N_i-N_i^\star),

and simulations starting with Ni(0)=Ni⋆N_i(0)=N_i^\star see no variation of the total resident population at the demographic level. For the classic metapopulation model, the movement matrix Mc\mathcal{M}_c and demographic steady state Nc⋆N_c^\star are obtained by lumping the resident-traveller model by current location at demographic equilibrium, so that the population present in each country (Nc,p⋆=Npr⋆=∑i=117Nip⋆N_{c,p}^\star=N_p^{r\star}=\sum_{i=1}^{17}N_{ip}^\star) and the daily flux of travellers between every pair of countries match those of the resident-traveller model (whereas applying the raw one-way gravity rates wjiw_{ji} without return trips would artificially redistribute populations toward wealthier destinations). The juxtaposition of the two model outputs illustrates how the classic model tends to homogenise vaccination coverage globally due to movement, while the resident-traveller model preserves the regional vulnerabilities associated with low vaccine coverage in specific countries. Unlike the generic simulations explored earlier, for the specific case of poliomyelitis, the rate of immunity loss ψij\psi_{ij} is set to zero in the numerics, reflecting the fact that immunity is life-long or long enough that immunity loss is not relevant over the time scales considered here.

Figure 9 and Figure 10 present the simulated output of both models using this derived data, specifically focusing on the under-5 demographic. Poliovirus infections are largely asymptomatic, with only roughly 0.5% (1 in 200) of infections resulting in irreversible paralysis [14, 59]. To ensure an apples-to-apples comparison between the two formulations, all quantities in both figures are tallied by location of presence or infection rather than by country of residence: Figure 9 plots 0.5% of the active infectious prevalence Ip(t)=∑i=117Iip(t)I_p(t)=\sum_{i=1}^{17}I_{ip}(t) physically present in country pp over the course of a year, and Figure 10 maps the accumulated paralytic cases (0.5% of cumulative new infections occurring in country pp, whether among local residents or visiting travellers) after one year, with both panels in each figure plotted on a common scale. (Notice that in the resident-traveller model, cumulative cases can also be tallied by country of residence; because Nigeria’s large under-5 cohort sends many unvaccinated travellers to neighbouring countries, 87,461 paralytic cases occur among Nigerian residents overall, of which 53,054 are acquired domestically within Nigeria and 34,407 are acquired while visiting neighbouring countries such as Cameroon, Niger and Benin.) While this thought experiment is not calibrated to reproduce the exact annual case counts of Figure 8, the resident-traveller formulation captures the qualitative pattern observed during the 2003–2004 resurgence: an earlier and larger epidemic wave in Nigeria accompanied by substantial transmission in neighbouring countries across the air and ground transportation networks (138,965 total paralytic cases region-wide after one year, compared with 87,367 in the classic model, where the assimilation and vaccination of travellers in higher-coverage countries slows and dampens the regional epidemic).

Two-panel line plots of infectious prevalence by current location for the classic metapopulation model and resident-traveller model

Figure 9. Infectious prevalence by current location, scaled by the paralysis-to-infection ratio (0.5% of Ipp(t)I_p^p(t)) over a 365-day period on a common vertical scale, indicating the timing and relative size of the outbreak in each country. Each curve represents a different country within the region (Nigeria in black). (a) Classic metapopulation model; (b) resident-traveller metapopulation.

Two-panel maps showing the geographic distribution of total accumulated paralytic cases for the classic metapopulation model and resident-traveller model

Figure 10. Geographic distribution of total accumulated paralytic cases after 1 year, tallied by country of infection and plotted on a shared size and colour scale. The classic model (a) predicts a smaller regional burden (87,367 paralytic cases overall, including 31,858 in Nigeria and 9,158 in Cameroon), whereas the resident-traveller model (b) predicts a substantially larger outbreak (138,965 paralytic cases overall, including 53,054 acquired in Nigeria and 26,708 in Cameroon).

Finally, two additional numerical sensitivity checks (not shown) are worth noting. First, when larger values of the basic reproduction number typical of completely naive populations (R0i∈[6,11]R_{0i} \in[6,11] or [6,14][6,14] [20, 19]) are used directly with R(0)=0R(0)=0 and imperfect vaccine protection (πi∈[0.1,0.4]\pi_i \in[0.1,0.4]), there is much less difference between the outcomes of the classic and resident-traveller models: because even high vaccine coverage leaves an effective susceptible fraction (1−ci)+πici≥πi(1-c_i)+\pi_i c_i \ge\pi_i, such large values of R0iR_{0i} place every country well above the local epidemic threshold (Rvi≫1R_{vi} \gg1) in both formulations, resulting in a rapid, synchronous epidemic wave across the entire region regardless of how traveller vaccination is bookkept. Second, applying a 50% global travel restriction, or a targeted 50% reduction in travel solely for individuals leaving the outbreak epicenter (Nigeria), yields virtually no reduction in the total accumulated paralytic cases after one year in either model. Because the outbreak is driven heavily by local transmission dynamics once introduced to a patch, moderate travel restrictions delay the epidemic peak rather than substantially altering the cumulative one-year burden.

Discussion

Comparing the resident-traveller framework with a classic Eulerian metapopulation model shows that tracking place of residence fundamentally alters predictions of vaccination impact across a mobility network. Because classic metapopulation models assimilate travellers into the destination population, they apply local vaccination rates to visitors and export newly vaccinated individuals across the digraph. In the high-mobility limit, this spatial mixing averages vaccine coverage globally, allowing high vaccination in one or more patches to drive the system-wide reproduction number Rv(B.1)R_v^{(B.1)} below unity even when other patches do not vaccinate at all. In contrast, when vaccination is restricted to residents in their home patch, the resident-traveller reproduction number Rv(3.1)R_v^{(3.1)} is bounded below by the digraph-adjusted basic reproduction number R~0k\widetilde{R}_0^k of any non-vaccinating patch kk. Consequently, whenever R~0k>1\widetilde{R}_0^k > 1, unvaccinated subpopulations act as persistent reservoirs that prevent global elimination regardless of vaccination effort elsewhere.

From a mathematical standpoint, while Theorems 5 and 8 establish the existence of at least one endemic equilibrium using uniform persistence theory when R0>1R_0>1 (respectively Rv>1R_v>1), the problem of uniqueness and global asymptotic stability of the endemic equilibrium was not resolved here. Furthermore, while Appendix § 11 proves the global asymptotic stability of the DFE for the baseline SIRS model under equal movement rates using a comparison theorem, establishing global stability for the DFE in the full SVIRS model when Rv(3.1)<1R_v^{(3.1)}<1 remains open, as temporary loss-of-immunity dynamics and breakthrough infections allow the effective susceptible pool Sij+πiVijS_{ij}+\pi_i V_{ij} to transiently exceed its DFE value. For resident-traveller models, this problem is notoriously complicated; the only work known to the authors addressing global stability in this context is [29], which proves a conjecture from [8] for the SIS case using the fact that the latter system is cooperative. From numerical investigations, it is suspected that in the present model, this type of behaviour also prevails, with a unique endemic equilibrium existing and attracting all solutions when Rv>1R_v>1.

The model used here is intentionally simple in its description of epidemiological and demographic processes so as to isolate the structural role of residence-based vaccination, and several limitations should be noted when interpreting the numerical examples. First, the model omits age structure, so demographic turnover in the under-5 poliomyelitis example is governed by the baseline mortality rate rather than explicit ageing out of the under-5 cohort, and routine childhood immunisation is approximated by a continuous rate ϕij\phi_{ij}. Second, mixing within each patch is assumed to be homogeneous, and the numerical simulations assume that infectious individuals travel at 10% of the rate of other compartments (MI=0.1MS\mathcal{M}^{I}=0.1\mathcal{M}^{S}). Third, the deterministic formulation does not capture stochastic fade-out at low prevalence or pathogen-specific features of oral poliovirus vaccines, such as vaccine-virus transmission or the emergence of circulating vaccine-derived polioviruses (cVDPV). At the same time, the general formulation (3.1) is flexible enough to accommodate non-resident vaccination policies through the off-diagonal rates ϕij>0\phi_{ij}>0 (i≠ji\ne j), which can represent transit vaccination posts at border crossings or vaccination requirements for incoming travellers, as were eventually deployed during polio eradication campaigns. It would also be interesting to undertake comparisons with residence-time (Lagrangian) models, which track individuals by home residence through a residence-time budgeting matrix rather than explicit movement fluxes (see, e.g., [12, 24]).

This added realism comes at a computational cost, as

Discussion

the number of equations rises quickly with network size: for nn patches, a classic SVIRS metapopulation model requires 4n4n equations (68 for the 17 countries in Section 4.5), whereas the resident-traveller formulation requires 4 equations per active residence-location pair, up to 4n24n^2 (1,156 for n=17n = 17) when all pairs of patches are connected, and requires specifying the zero/nonzero patterns of all movement matrices Mi\mathcal{M}_i.

As a final remark, this paper was a long time coming: the authors started working on it in 2004. For a variety of reasons, it was only finished recently. While an immense global public health effort has culminated in the African continent being officially certified free of wild poliovirus in August 2020 [22], poliomyelitis is not yet eradicated globally. Furthermore, vaccine coverage has decreased for other vaccine-preventable diseases, even in countries where vaccines are readily and widely accessible. Driven by rising vaccine hesitancy and disruptions from the COVID-19 pandemic, declining childhood immunization rates have recently fueled a severe resurgence of measles across Canada and the United States of America [45, 49], culminating in Canada officially losing its measles elimination status in November 2025 after more than 12 months of sustained transmission [48]. With localised, under-vaccinated pockets allowing imported cases to spark explosive outbreaks, the core spatial mechanics of jurisdiction-based vaccination explored in this paper are arguably more relevant today than they were in 2004.

Acknowledgements

The authors thank the Banff International Research Station (BIRS) for hosting the summer school and workshop during which this project was initiated. JA acknowledges partial support from an NSERC Discovery Grant.

Global Asymptotic Stability of the DFE

Proof for the SIRS model

The global asymptotic stability (GAS) of the disease-free equilibrium (DFE) for the SIRS model (2.1) when R0<1R_0 < 1 and movement rates are equal across all epidemiological states (MS=MI=MR=MM^S = M^I = M^R = M) is proved using the theory of asymptotically autonomous limit systems.

From Lemma 6, it is known that under the assumption of equal movement rates for all epidemiological states, the total population in each subpopulation converges to its equilibrium, lim⁡t→∞Nij(t)=Nij⋆\lim_{t \to\infty} N_{ij}(t) = N_{ij}^{\star}. Because Sij(t)≤Nij(t)S_{ij}(t) \leq N_{ij}(t) for all tt, it follows that for any η>0\eta> 0, there exists a time t1>0t_1 > 0 such that for all t>t1t > t_1,

Sij(t)≤Nij⋆+ηandNjp(t)≥Njp⋆−η,S_{ij}(t) \leq N_{ij}^{\star} + \eta\quad\text{and} \quad N_j^p(t) \geq N_j^{p\star} - \eta,

(recalling that Njp=∑kNkjN_j^p = \sum_k N_{kj} is the total population physically present in patch jj). The differential equations for the infectious compartments (2.1b) can therefore be bounded above for t>t1t > t_1 by the linear differential inequality:

ddtI≤(Fη−W)I,\frac{d}{dt} I \leq(F_{\eta} - W)I,

where WW is defined as in (2.6), and FηF_{\eta} is the matrix of new infections evaluated using the perturbed upper bounds Nij⋆+ηN_{ij}^{\star} + \eta and lower bounds Njp⋆−ηN_j^{p\star} - \eta.

Because FηF_{\eta} is a non-negative matrix and the off-diagonal entries of −W=MI−D−Γ-W = M^I - D - \Gamma consist of the non-negative movement rates from MIM^I, the matrix Fη−WF_{\eta} - W has non-negative off-diagonal entries (i.e., it is a Metzler matrix). This property implies that the bounding linear differential system is cooperative. Thus, standard comparison principles for monotone flows (such as Kamke’s comparison theorem) apply, despite the full nonlinear SIRS system not being cooperative [55].

When R0=ρ(FW−1)<1R_0 = \rho(FW^{-1}) < 1, it follows (see, e.g., [57 Theorem 2]) that the spectral bound satisfies s(F−W)<0s(F - W) < 0. By the continuity of eigenvalues with respect to matrix entries, η\eta can be chosen sufficiently small such that s(Fη−W)<0s(F_{\eta} - W) < 0.

By Kamke’s comparison theorem for cooperative linear ODEs, I(t)I(t) is bounded above by the solution to the linear system X′=(Fη−W)X\mathbf{X}' = (F_{\eta} - W)\mathbf{X}. Since the spectral bound of Fη−WF_{\eta} - W is strictly negative, X(t)→0\mathbf{X}(t) \to0 as t→∞t \to\infty, which implies lim⁡t→∞I(t)=0\lim_{t \to\infty} I(t) = 0.

Once I→0I \to0, the recovered equation (2.1c) reduces asymptotically to ddtR=(MR−D−Ψ)R\frac{d}{dt} R = (M^R - D - \Psi)R, which implies lim⁡t→∞R(t)=0\lim_{t \to\infty} R(t) = 0 by Lemma 3. Finally, with both I→0I \to0 and R→0R \to0, the susceptible population must converge to the limiting total population, lim⁡t→∞S(t)=N⋆\lim_{t \to\infty} S(t) = N^{\star}. Thus, the DFE of the SIRS model is globally asymptotically stable.

Adaptation to the SVIRS model

It is natural to ask whether this bounding argument extends to the SVIRS model (3.1) to prove global asymptotic stability of the DFE when Rv<1R_v < 1.

In the SIRS model, the DFE susceptible population is the entire population (Sij⋆=Nij⋆S_{ij}^{\star} = N_{ij}^{\star}), meaning the physical bound Sij(t)≤Nij(t)S_{ij}(t) \leq N_{ij}(t) provides the tight asymptotic bound Sij⋆+ηS_{ij}^{\star} + \eta. In the SVIRS model, the effective susceptible pool for subpopulation (i,j)(i,j) is Sij(t)+πijVij(t)S_{ij}(t) + \pi_{ij} V_{ij}(t). Because πij≤1\pi_{ij} \leq1, this pool is strictly bounded by the total living population:

Sij(t)+πijVij(t)≤Sij(t)+Vij(t)≤Nij(t).S_{ij}(t) + \pi_{ij} V_{ij}(t) \leq S_{ij}(t) + V_{ij}(t) \leq N_{ij}(t).

Consequently, for any arbitrarily small η>0\eta> 0 and sufficiently large tt, Sij(t)+πijVij(t)≤Nij⋆+ηS_{ij}(t) + \pi_{ij} V_{ij}(t) \leq N_{ij}^{\star} + \eta. Substituting this upper bound into the infectious subsystem yields the same bounding matrix inequality I′≤(Fη0−W)II' \leq(F_{\eta}^{0} - W)I, Fn0F_n^0 evaluates new infections assuming the entire population is fully susceptible. The spectral radius ρ(F0W−1)\rho(F^0W^{-1}) corresponds to the baseline, vaccine-free basic reproduction number R0(2.1)\mathcal{R}_0^{(2.1)}. Thus, if R0(2.1)<1\mathcal{R}_0^{(2.1)} < 1, Kamke’s theorem proves that the SVIRS disease-free equilibrium is globally asymptotically stable.

However, establishing global asymptotic stability specifically when Rv(3.1)<1<R0(2.1)\mathcal{R}_v^{(3.1)} < 1 < \mathcal{R}_0^{(2.1)} using this comparison method is not straightforward. The issue lies in bounding the effective transmission pool Sij(t)+πijVij(t)S_{ij}(t)+\pi_{ij}V_{ij}(t) tightly around its DFE value Sij⋆+πijVij⋆S_{ij}^{\star}+\pi_{ij}V_{ij}^{\star}. Unlike in the baseline SIRS model, where the natural excursion of individuals through the infectious and recovered classes (S→I→R→SS \to I \to R \to S) is strictly bounded by the demographic ceiling Nij(t)N_{ij}(t) with Sij⋆=Nij⋆S_{ij}^{\star}=N_{ij}^{\star}, temporary immunity (ψij>0\psi_{ij}>0) and breakthrough infections (V→I→R→SV \to I \to R \to S) in the SVIRS model can move individuals from the partially susceptible vaccinated class (πij<1\pi_{ij}<1) back into the fully susceptible class, allowing Sij(t)+πijVij(t)S_{ij}(t)+\pi_{ij}V_{ij}(t) to transiently exceed Sij⋆+πijVij⋆S_{ij}^{\star}+\pi_{ij}V_{ij}^{\star}. The standard comparison theorem thus fails. Establishing Global Asymptotic Stability (GAS) for the SVIRS resident-traveller model, either at the DFE or for the endemic equilibrium, remains a non-trivial open problem. Standard graph-theoretic Lyapunov methods [54] cannot be readily applied due to the distinct topologies of mobility and infection, alongside the non-cooperative nature of standard incidence mixing.

The classic metapopulation SVIRS model

Part of the justification for using a resident-traveller framework is that it allows consideration of problems such as those that arise when one location does not vaccinate while others do and residents maintain their vaccination status and behaviour even when travelling. In order to illustrate this claim, a classic metapopulation model analogous to (3.1) (with no vaccination at birth, matching Pi=0P_i=0 in the numerical comparisons) is now considered. In classic metapopulation models, individuals who travel from patch jj to patch ii immediately become, upon arrival in ii, members of patch ii. Consequently, they become subject to the vaccination policy and mixing patterns of their destination patch.

The classic metapopulation model takes the following form, where, for simplicity, it is assumed that movement is equal for all epidemiological states and mijm_{ij} denotes the rate of movement from patch jj to patch ii. Let Ni=Si+Ii+Ri+ViN_i=S_i+I_i+R_i+V_i be the total population in patch ii. The dynamics are governed by:

ddtSi=bi+ψiRi−(di+ϕi+βiIiNi)Si+∑j=1nmijSj(33)\frac{d}{dt}S_i=b_i+\psi_iR_i-\left(d_i+\phi_i+\beta_i\frac{I_i}{N_i}\right)S_i+\sum_{j=1}^{n}m_{ij}S_j \tag*{(33)}
ddtIi=(Si+πiVi)βiIiNi−(di+γi)Ii+∑j=1nmijIj(34)\frac{d}{dt}I_i=(S_i+\pi_iV_i)\beta_i\frac{I_i}{N_i}-(d_i+\gamma_i)I_i+\sum_{j=1}^{n}m_{ij}I_j \tag*{(34)}
ddtRi=γiIi−(di+ψi)Ri+∑j=1nmijRj(35)\frac{d}{dt}R_i=\gamma_iI_i-(d_i+\psi_i)R_i+\sum_{j=1}^{n}m_{ij}R_j \tag*{(35)}
ddtVi=ϕiSi−(di+πiβiIiNi)Vi+∑j=1nmijVj.(36)\frac{d}{dt}V_i=\phi_iS_i-\left(d_i+\pi_i\beta_i\frac{I_i}{N_i}\right)V_i+\sum_{j=1}^{n}m_{ij}V_j. \tag*{(36)}

A basic analysis follows steps from standard metapopulation models [1]. The vector form of the system involves vectors X=(X1,…,Xn)T\mathbf{X}=(X_1,\ldots,X_n)^T for X∈{N,S,V,I,R}X\in\{N,S,V,I,R\}, b=(b1,…,bn)T\mathbf{b}=(b_1,\ldots,b_n)^T and diagonal matrices Dc=diag⁡(d1,…,dn)\mathcal{D}_c=\operatorname{diag}(d_1,\ldots,d_n), Φc=diag⁡(ϕ1,…,ϕn)\Phi_c=\operatorname{diag}(\phi_1,\ldots,\phi_n), Γc=diag⁡(γ1,…,γn)\Gamma_c=\operatorname{diag}(\gamma_1,\ldots,\gamma_n) and Ψc=diag⁡(ψ1,…,ψn)\Psi_c=\operatorname{diag}(\psi_1,\ldots,\psi_n). Let Mc\mathcal{M}_c be the movement matrix with off-diagonal entries mijm_{ij} and diagonal entries mii=−∑k≠imkim_{ii}=-\sum_{k\ne i}m_{ki}. Because movement is equal for all epidemiological states, the total population satisfies

ddtN=b−DcN+McN,\frac{d}{dt}N=b-\mathcal{D}_cN+\mathcal{M}_cN,

so that lim⁡t→∞N(t)=Nc⋆:=(Dc−Mc)−1b\lim_{t\to\infty}N(t)=N_c^{\star}:=(\mathcal{D}_c-\mathcal{M}_c)^{-1}b. At the disease-free equilibrium (DFE), there are no infectious or recovered individuals (Ic⋆=Rc⋆=0I_c^{\star}=R_c^{\star}=0), and

0=b−(Dc+Φc−Mc)S0=b-(\mathcal{D}_c+\Phi_c-\mathcal{M}_c)S
0=ΦcS−(Dc−Mc)V0=\Phi_cS-(\mathcal{D}_c-\mathcal{M}_c)V

which yields the unique positive equilibrium

Sc⋆=(Φc+Dc−Mc)−1b,Vc⋆=(Dc−Mc)−1ΦcSc⋆.(37)S_c^{\star}=(\Phi_c+\mathcal{D}_c-\mathcal{M}_c)^{-1}b,\qquad V_c^{\star}=(\mathcal{D}_c-\mathcal{M}_c)^{-1}\Phi_cS_c^{\star}. \tag*{(37)}

The local asymptotic stability of the DFE is governed by the vaccination reproduction number Rv(B.1)=ρ(FcWc−1)\mathcal{R}_v^{(B.1)}=\rho(F_cW_c^{-1}), determined by the next-generation matrix method [57].

Theorem 11. The DFE of the classic metapopulation model is locally asymptotically stable if Rv(B.1)<1\mathcal{R}_v^{(B.1)}<1 and unstable if Rv(B.1)>1\mathcal{R}_v^{(B.1)}>1.

In this setup, the rate of new infections in patch ii is Fi=(Si+πiVi)βiIiNi\mathcal{F}_i=(S_i+\pi_iV_i)\beta_i\frac{I_i}{N_i}, and the other transitions are given by Vi=(di+γi)Ii−∑j=1nmijIj\mathcal{V}_i=(d_i+\gamma_i)I_i-\sum_{j=1}^{n}m_{ij}I_j. Linearizing at the DFE, the non-negative matrix FcF_c of new infections and the non-singular M-matrix WcW_c are

Fc=diag⁡(βiSci⋆+πiVci⋆Nci⋆)andWc=Γc+Dc−Mc.F_c=\operatorname{diag}\left(\beta_i\frac{S_{ci}^{\star}+\pi_iV_{ci}^{\star}}{N_{ci}^{\star}}\right) \qquad\text{and}\qquad W_c=\Gamma_c+\mathcal{D}_c-\mathcal{M}_c.

Similarly, the basic reproduction number R0(B.1)\mathcal{R}_0^{(B.1)} in the absence of vaccination (Φc=0\Phi_c=0, giving Vc⋆=0V_c^{\star}=0 and Sc⋆=Nc⋆S_c^{\star}=N_c^{\star}) is R0(B.1)=ρ(Fc0Wc−1)\mathcal{R}_0^{(B.1)}=\rho(F_c^0W_c^{-1}), where Fc0=diag⁡(βi)F_c^0=\operatorname{diag}(\beta_i).

References

  1. [1]J. Arino. Diseases in metapopulations. In Z. Ma, Y. Zhou, and J. Wu, editors, Modeling and Dynamics of Infectious Diseases, volume 11 of Series in Contemporary Applied Mathematics, pages 65–123. World Scientific, 2009.DOI
  2. [2]J. Arino. Describing, modelling and forecasting the spatial and temporal spread of COVID-19 – a short review. Fields Institute Communications, 85:25–51, 2022.DOI
  3. [3]J. Arino, N. Bajeux, and S. Kirkland. Number of source patches required for population persistence in a source–sink metapopulation. Bulletin of Mathematical Biology, 81(6):1916–1942, 2019.DOI
  4. [4]J. Arino and A.-S. Ciupeanu. wikipediaGATN. https://github.com/julien-arino/wikipediaGATN, 2026.
  5. [5]J. Arino, A. Ducrot, and P. Zongo. A metapopulation model for malaria with transmission-blocking partial immunity in hosts. Journal of Mathematical Biology, 64(3):423–448, 2012.DOI
  6. [6]J. Arino, C.C. McCluskey, and P. van den Driessche. Global results for an epidemic model with vaccination that exhibits backward bifurcation. SIAM J Appl Math, 64(1):260–276, 2003.DOI
  7. [7]J. Arino and P. van den Driessche. The basic reproduction number in a multi-city compartmental epidemic model. Lecture Notes in Control and Information Science, 294:135–142, 2003.DOI
  8. [8]J. Arino and P. van den Driessche. A multi-city epidemic model. Mathematical Population Studies, 10(3):175–193, 2003.DOI
  9. [9]J. Arino and P. van den Driessche. Metapopulations epidemic models. A survey. Fields Institute Communications, 48:1–12, 2006.
  10. [10]H. Bedford and D. Elliman. Concerns about immunisation. BMJ, 320(7229):240–243, Jan 2000.pubmed.ncbi.nlm.nih.gov/10642238
  11. [11]D. Berera and M. Zambon. Antivirals in the 2009 pandemic–lessons and implications for future strategies. Influenza Other Respiratory Viruses, 7 Suppl 3:72–79, Nov 2013.DOI
  12. [12]D. Bichara and C. Castillo-Chavez. Vector-borne diseases models with residence times–a Lagrangian perspective. Mathematical Biosciences, 281:128–138, 2016.DOI
  13. [13]S. Blume. Anti-vaccination movements and their interpretations. Soc Sci Med, 62(3):628–642, Feb 2006.pubmed.ncbi.nlm.nih.gov/16039769
  14. [14]Centers for Disease Control and Prevention. Polio disease and poliovirus. https://www.cdc.gov/polio/what-is-polio/index.html.
  15. [15]M.D. Christian, S.M. Poutanen, M.R. Loutfy, M.P. Muller, and D.E. Low. Severe acute respiratory syndrome. Clinical Infectious Diseases, 38:1420–1427, May 2004.pubmed.ncbi.nlm.nih.gov/15156481
  16. [16]C.J. Clements, P. Greenough, and D. Shull. How vaccine safety can become political–the example of polio in Nigeria. Curr Drug Saf, 1(1):117–119, Jan 2006.pubmed.ncbi.nlm.nih.gov/18690921
  17. [17]O. Diekmann and J.A.P. Heesterbeek. Mathematical Epidemiology of Infectious Diseases: Model Building, Analysis and Interpretation. Wiley, 2000.
  18. [18]J.S. Downs, W. Bruine de Bruin, and B. Fischhoff. Parents’ vaccination comprehension and decisions. Vaccine, 26(12):1595–1607, Mar 2008.DOI
  19. [19]Radboud J. Duintjer Tebbens, Mark A. Pallansch, Dominika A. Kalkowska, Steven G. F. Wassilak, Stephen L. Cochi, and Kimberly M. Thompson. Characterizing poliovirus transmission and evolution: Insights from modeling experiences with wild and vaccine-related polioviruses. Risk Analysis, 33(4):703–749, 2013.DOI
  20. [20]Paul E. M. Fine and Ilona A. M. Carneiro. Transmissibility and persistence of oral polio vaccine viruses: Implications for the global poliomyelitis eradication initiative. American Journal of Epidemiology, 150(10):1001–1021, 1999.pubmed.ncbi.nlm.nih.gov/10568615
  21. [21]F. Fleck. Nigerian state of Kano resumes polio vaccination. BMJ, 329(7457):70, Jul 2004.pubmed.ncbi.nlm.nih.gov/15242894
  22. [22]R. G. Fomban Leke, A. King, M. A. Pallansch, R. H. Tangermann, P. Mkanda, S. Chunsuttiwat, A. Jack, B. J. Kaboré, I. Kane, N. E. Khomo, R. J. Lopes-Feio, Z. M. Maiga, W. B. P. Matuja, N. E. Kironde, B. Kouassi, B. Sadrizadeh, and O. Tomori. Certifying the interruption of wild poliovirus transmission in the WHO African region on the turbulent journey to a polio-free world. The Lancet Global Health, 8(10):e1345–e1351, 2020.DOI
  23. [23]E. Gaarder, D. Luke, and L. Sommer. Towards an estimate of informal cross-border trade in Africa. Technical report, United Nations Economic Commission for Africa, 2021.
  24. [24]D. Gao and L. Cao. Vector-borne disease models with Lagrangian approach. Journal of Mathematical Biology, 88(2):22, 2024.DOI
  25. [25]I. Ghinai, C. Willott, I. Dadari, and H.J. Larson. Listening to the rumours: what the northern Nigeria polio vaccine boycott can tell us ten years on. Global Public Health, 8(10):1138–1150, 2013.DOI
  26. [26]A. Gil. Social and labour aspects of urban passenger transport in selected African cities. Technical report, International Labour Organisation, 1989.
  27. [27]J.D. Grabenstein. What the world’s religions teach, applied to vaccines and immune globulins. Vaccine, 31(16):2011–2023, Apr 2013.DOI
  28. [28]R. Horn and C. Johnson. Matrix Analysis. Cambridge University Press, 1990.
  29. [29]A. Iggidr, G. Sallet, and B. Tsanou. Global stability analysis of a metapopulation SIS epidemic model. Mathematical Population Studies, 19(3):115–129, 2012.DOI
  30. [30]D. Isaacs, H. Kilham, J. Leask, and B. Tobin. Ethical issues in immunisation. Vaccine, 27(5):615–618, Jan 2009.DOI
  31. [31]A.S. Jegede. What led to the Nigerian boycott of the polio vaccination campaign? PLoS Med, 4(3):e73, Mar 2007.pubmed.ncbi.nlm.nih.gov/17388657
  32. [32]A. Kata. A postmodern Pandora’s box: anti-vaccination misinformation on the Internet. Vaccine, 28(7):1709–1716, Feb 2010.
  33. [33]A. Kata. Anti-vaccine activists, Web 2.0, and the postmodern paradigm–an overview of tactics and tropes used online by the anti-vaccination movement. Vaccine, 30(25):3778–3789, May 2012.
  34. [34]T. Kato. Perturbation Theory for Linear Operators. Springer-Verlag, reprint of the 1980 second edition, 1995.
  35. [35]J.R. Kaufmann and H. Feldbaum. Diplomacy and the polio immunization boycott in Northern Nigeria. Health Affairs, 28(4):1091–1101, 2009.
  36. [36]T. Kerzhner. Is informal transport flexible? Journal of Transport and Land Use, 15(1):671–689, 2022.DOI
  37. [37]K. Khan, J. Arino, W. Hu, P. Raposo, J. Sears, F. Calderon, C. Heidebrecht, M. Macdonald, J. Liauw, A. Chan, and M. Gardam. Spread of a novel influenza A (H1N1) virus via global airline transportation. New England Journal of Medicine, 361(2):212–214, 2009.DOI
  38. [38]K. Khan, Z.A. Memish, A. Chabbra, J. Liauw, W. Hu, D.A. Janes, J. Sears, J. Arino, M. Macdonald, F. Calderon, P. Raposo, C. Heidebrecht, J. Wang, A. Chan, J. Brownstein, and M. Gardam. Global public health implications of a mass gathering in Mecca, Saudi Arabia during the midst of an influenza pandemic. Journal of Travel Medicine, 17(2):75–81, 2010.DOI
  39. [39]K. Khan, P. Muennig, M. Behta, and J. G. Zivin. Global drug-resistance patterns and the management of latent tuberculosis infection in immigrants to the United States. New England Journal of Medicine, 347(23):1850–1859, 2002.pubmed.ncbi.nlm.nih.gov/12466510
  40. [40]P. Klepac, C.J.E. Metcalf, A.R. McLean, and K. Hampson. Towards the endgame and beyond: complexities and challenges for the elimination of infectious diseases. Philos Trans R Soc Lond B Biol Sci, 368(1623):20120137, Aug 2013.DOI
  41. [41]M.J. Knol, A.T. Urbanus, E.M. Swart, L. Mollema, W.L. Ruijs, R.S. van Binnendijk, M.J. te Wierik, H.E. de Melker, A. Timen, and S.J. Hahn é. Large ongoing measles outbreak in a religious community in the Netherlands since May 2013. Eurosurveillance, 18(36):2–7, 2013.pubmed.ncbi.nlm.nih.gov/24079377
  42. [42]S. Koroma, J. Nimarkoh, N. You, V. Ogalo, and B. Owino. Formalization of informal trade in Africa: Trends, experiences and socio-economic impact. Technical report, Food and Agriculture Organization of the United Nations, 2017.
  43. [43]D. Langlois-Klassen, K.M. Wooldrage, J. Manfreda, K. Sutherland, E. Ellis, M. Phypers, B. Gushulak, and R. Long. Piecing the puzzle together: foreign-born tuberculosis in an immigrant-receiving country. Eur Respir J, 38(4):895–902, Oct 2011.
  44. [44]M.M. Larsen, R. Namatovu, and R. Narula. An actor-based perspective on informal cross-border trading. The Journal of Development Studies, 62(8):1275–1295, 2026.DOI
  45. [45]A.D. Mathis, K. Raines, N.B. Masters, T.D. Filardo, G. Kim, S.N. Crooke, B. Bankamp, P.A. Rota, and D.E. Sugerman. Measles — United States, January 1, 2020–March 28, 2024. MMWR Morbidity and Mortality Weekly Report, 73(14):295–300, 2024.DOI
  46. [46]C.J.E. Metcalf, C. Cohen, J. Lessler, J.M. McAnerney, G.M. Ntshoe, A. Puren, K. Klepac, A. Tatem, B.T. Grenfell, and O.N. Bjørnstad. Implications of spatially heterogeneous vaccination coverage for the risk of congenital rubella syndrome in South Africa. Journal of the Royal Society, Interface, 10(78):20120756, Jan 2013.DOI
  47. [47]M.A. Pallansch and H.S. Sandhu. The eradication of polio—progress and challenges. N Engl J Med, 355(24):2508–2511, Dec 2006.pubmed.ncbi.nlm.nih.gov/17167133
  48. [48]Public Health Agency of Canada. Statement from the Public Health Agency of Canada on Canada’s measles elimination status. https://www.canada.ca/en/public-health/news/2025/11/statement-from-the-public-health-agency-of-canada-on-canadas-measles-elimination-status.html, November 2025.
  49. [49]A. A. Rifai, D. C. Pittman Ratterree, J. D. Kong, and M. L. Ndeffo-Mbah. Epidemiological review on the resurgence of measles outbreaks in Canada during the post-elimination era: A scoping review. PLOS Global Public Health, 6(4):e0006295, 2026.DOI
  50. [50]S. Ruan, W. Wang, and S.A. Levin. The effect of global travel on the spread of SARS. Mathematical Biosciences and Engineering, 3(1):205–218, 2006.
  51. [51]L. Sattenspiel and K. Dietz. A structured epidemic model incorporating geographic mobility among regions. Mathematical Biosciences, 128:71–91, 1995.pubmed.ncbi.nlm.nih.gov/7606146
  52. [52]L. Sattenspiel and D.A. Herring. Simulating the effect of quarantine on the spread of the 1918-19 flu in central Canada. Bull. Math. Biol., 65:1–26, 2003.pubmed.ncbi.nlm.nih.gov/12597114
  53. [53]K. Schwartzman, O. Oxlade, R. G. Barr, F. Grimard, I. Acosta, J. Baez, E. Ferreira, R. E. Melgen, W. Morose, A. C. Salgado, V. Jacquet, S. Maloney, K. Laserson, A. P. Mendez, and D. Menzies. Domestic returns from investment in the control of tuberculosis in other countries. New England Journal of Medicine, 353(10):1008–1020, 2005.pubmed.ncbi.nlm.nih.gov/16148286
  54. [54]Z. Shuai and P. van den Driessche. Global stability of infectious disease models using Lyapunov functions. SIAM Journal on Applied Mathematics, 73(4):1513–1532, 2013.DOI
  55. [55]H.L. Smith. Monotone Dynamical Systems: An Introduction to the Theory of Competitive and Cooperative Systems, volume 41 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 1995.
  56. [56]H.R. Thieme. Persistence under relaxed point-dissipativity (with applications to an endemic model). SIAM J. Math. Anal., 24:407–435, 1993.DOI
  57. [57]P. van den Driessche and J. Watmough. Reproduction numbers and sub-threshold endemic equilibria for compartmental models of disease transmission. Mathematical Biosciences, 180:29–48, 2002.pubmed.ncbi.nlm.nih.gov/12387915
  58. [58]M. Weigel, K. Weitmann, C. Rautmann, J. Schmidt, R. Bruns, and W. Hoffmann. Impact of physicians’ attitude to vaccination on local vaccination coverage for pertussis and measles in Germany. Eur J Public Health, 24(6):1009–1016, 2014.DOI
  59. [59]World Health Organization. Poliomyelitis fact sheet. https://www.who.int/news-room/fact-sheets/detail/poliomyelitis.
  60. [60]G.G. Yin and Q. Zhang. Continuous-Time Markov Chains and Applications: A Singular Perturbation Approach. Springer, 1998.DOI

Paper details

vlnc:W_bwb49kta
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Agricultural and Biological Sciences

Not recorded for this paper

View statement
Looking for related papers…
Contents