Heterogeneous network epidemics: real-time growth, variance and extinction of infection
Abstract
Recent years have seen a large amount of interest in epidemics on networks as a way of representing the complex structure of contacts capable of spreading infections through the modern human population. The configuration model is a popular choice in theoretical studies since it combines the ability to specify the distribution of the number of contacts (degree) with analytical tractability. Here we consider the early real-time behaviour of the Markovian SIR epidemic model on a configuration model network using a multitype branching process. We find closed-form analytic expressions for the mean and variance of the number of infectious individuals as a function of time and the degree of the initially infected individual(s), and write down a system of differential equations for the probability of extinction by time t that are numerically fast compared to Monte Carlo simulation. We show that these quantities are all sensitive to the degree distribution—in particular we confirm that the mean prevalence of infection depends on the first two moments of the degree distribution and the variance in prevalence depends on the first three moments of the degree distribution. In contrast to most existing analytic approaches, the accuracy of these results does not depend on having a large number of infectious individuals, meaning that in the large population limit they would be asymptotically exact even for one initial infectious individual.
Keywords
SIR epidemic Configuration model Branching processMathematics Subject Classification
92D30 60J85 05C801 Introduction
1.1 Background
Models of infectious disease transmission have, from relatively modest beginnings (e.g. Bailey 1957), developed a rich domain of applicability covering the whole spectrum of human, animal and plant pathogens, and informing the study of questions from viral evolution, through epidemiology of infectious diseases, to public health policy (see Heesterbeek et al. 2015). Increasingly, networks have been seen as a way of modelling the complex, heterogeneous patterns of contacts between individuals (Danon et al. 2011).
In theoretical studies, the configuration model has been a popular choice due to the ability to specify the number of contacts each individual has that are capable of spreading disease, while allowing for analytic results to be obtained (e.g. Molloy and Reed 1995; Newman 2002). Ball and Neal (2008) used an effective degree approach (which we describe in Sect. 2.2 below—cf. Lindquist et al. 2010) to derive a system of ordinary differential equations that describes the deterministic limit of the epidemic model as the population size \(N \rightarrow \infty \). A much simpler (equivalent) system of only 4 ordinary differential equations was obtained by Volz (2008) and subsequently shown by Miller (2011), Miller et al. (2012) to be essentially one-dimensional (the 4 ODEs were also shown by House and Keeling (2010) to be a special case of the much higher dimensional pair approximation model of Eames and Keeling (2002), in which the degree structure is explicit). Fully rigorous proofs of convergence in probability of the scaled stochastic model to the deterministic limit are given by Decreusefond et al. (2012), Bohman and Picollelli (2012), Barbour and Reinert (2013) and Janson et al. (2014). These works are primarily concerned with the temporal behaviour of proportions of the population in different epidemiological compartments (susceptible, infectious and removed) over the main body of a large epidemic. Here, we are also concerned with temporal behaviour, but focus on numbers infected early in the epidemic, including the possibility of early stochastic extinction.
In a recent paper, Graham and House (2014) use a pairwise approximation in conjunction with the central limit theorem for density dependent population processes (Ethier and Kurtz 1986, Chap. 11) to obtain a closed-form approximation to the mean and variance of prevalence in the linearised model which approximates the early asymptotic exponential growth phase of a Markovian SIR epidemic on a configuration network. In particular, they find that, under these approximations, the variance in disease prevalence is determined by the first three moments of the network degree distribution. In this paper, we use the effective degree approach of Ball and Neal (2008) to approximate the early stages of the epidemic by a continuous-time, multitype Markovian branching process, which is then analysed in detail. For \(t\ge 0\), let Z(t) denote the total number of individuals alive in this branching process at time t, so Z(t) approximates disease prevalence in the epidemic model during its early asymptotic exponential growth phase. Explicit closed-form expressions are derived for the mean and variance of Z(t), the covariance of Z(t) and Z(s) to give the behaviour over time, and also for the probability of extinction \(\pi (t) = \mathbb {P}(Z(t)=0)\). As in Graham and House (2014), the mean and variance in disease prevalence depends on the degree distribution only through its first two and three moments, respectively.
The results in Graham and House (2014) assume implicitly that the initial number of infectives is sufficiently large for the density dependent population process central limit theorem to yield a good approximation. In contrast, our results assume any arbitrary, but specified, initial number of infectives. The asymptotic distribution of types in the branching process, when it does not go extinct, is also available in closed-form and enables us to obtain a Gaussian process approximation, with explicit mean and covariance function, for the prevalence in the early asymptotic exponential growth phase of an SIR epidemic, with few initial infectives, which takes off and becomes established. We show that this approximation can be applied together with the methods of Ross et al. (2006) to estimate epidemiological parameters from early prevalence data of a simulated epidemic provided the first three moments of the degree distribution are known.
1.2 Outline of the paper
The paper is organised as follows. The configuration network model and a Markov SIR epidemic on that network are described in Sect. 2.1. The effective degree construction of this epidemic is outlined in Sect. 2.2. Approximation of the early stages of this epidemic by a branching process is outlined in Sect. 2.3, where conditions are given for the mean, variance and covariance functions of the number of infectives in the epidemic process to converge to the corresponding quantities of the approximating branching process as the population size tend to infinity. The representation of the approximating branching process as a continuous-time, multitype Markov branching process is outlined in Sect. 2.3 and described more explicitly in Sect. 2.4. The mean, variance and covariance functions of the total number of individuals alive in the branching process are considered in Sects. 3, 4 and 5, respectively. Explicit closed-form expressions are obtained for each of these quantities and for their limits as time \(t \rightarrow \infty \). The arguments in Sects. 3, 4 and 5 assume that underlying degree distribution has a maximum degree. In Sect. 6, we show that these expressions continue to hold in the unbounded degree setting, subject to the degree distribution satisfying suitable moment conditions. The probability that the branching process is extinct at time t is studied in Sect. 7. Closed-form expressions for this probability, given the initial state of the branching process, are not available so asymptotic results as \(t \rightarrow 0\) and \(t \rightarrow \infty \) are considered.
The mean, variance and covariance functions derived in Sects. 3, 4 and 5 are unconditional, so they include realisations of the branching process which result in extinction. However, in the epidemic setting, we are often interested in analysing the behaviour of epidemics that take off and become established, which correspond to non-extinction of the branching process. In Sect. 8, we first derive the mean and variance of the total number of individuals alive in the branching process at time t, conditional upon the process having survived to time t; fully closed-form results are not available owing to the absence of a closed-form expression for the survival probability. We then consider realisations of the branching process which reach some specified size, K say, with time being set to zero the first time the total number of individuals alive is K. The results in Sect. 3 yield an explicit expression for the asymptotic distribution of types, given that the branching process does not go extinct, which, provided K is sufficiently large, enables the above branching process starting from K individuals to be approximated by a Gaussian process whose mean and covariance functions are determined explicitly. The theory is illustrated by numerical examples of both forward simulation and inference in Sect. 9 and some concluding comments are given in Sect. 10.
Notation used in multiple sections of this paper
Primary notation | Meaning | Equivalent notation |
---|---|---|
Network properties | ||
N | Size of the population | |
D | A random variable for an individual’s degree | \(D_1, D_2, \ldots \) |
\(p_k\) | Probability mass function for D evaluated at k | |
\(k_{\mathrm {max}}\) | The maximum degree | |
\(\mathscr {K}\) | The set of possible degrees | \(\{0,1,\ldots k_{\mathrm {max}}\}\) |
\({\tilde{D}}\) | A random variable for an individual’s size-biased degree | \({\tilde{D}}_1, {\tilde{D}}_2, \ldots \) |
\({\tilde{p}}_k\) | Probability mass function for \({\tilde{D}}\) evaluated at k | \(\mu _D^{-1}kp_k\) |
Vectors and matrices | ||
\( \mathbf {v} \) | A column vector whose kth entry is \(v_k\) | \((v_k)\) |
\( \mathbf {v} ^{\!\top }\) | A row vector (transpose of a column vector) | |
\( \varvec{M} \) | A matrix with (k, l)th entry \(M_{kl}\) or \(m_{kl}\) | \([M_{kl}]\), \([m_{kl}]\) |
\(| \varvec{M} |\) | Determinant of matrix \( \varvec{M} \) | |
\( \mathbf {1} \) | A column vector whose entries are all equal to 1 | |
\( \mathbf {n} \) | A column vector whose ith entry is i | |
\( \mathbf {n} _2\) | A column vector whose ith entry is \(i^2\) | |
\( \varvec{I} \) | The identity matrix | \([\delta _{k,l}]\) |
Probability | ||
\(\mathbb {P}(e)\) | Probability of event e | |
\(\mu _{f(X)}\) | Expected value of a function f of a random variable X | \(\mathbb {E}\left[ f(X)\right] \) |
\(M_X(\theta ) \) | Moment generating function for random variable X | \(\mathbb {E}\left[ \mathrm {exp}(\theta X)\right] \) |
\(\mathrm {var}(X) \) | Variance of random variable X | \(\mathbb {E}\left[ X^2\right] - \mathbb {E}\left[ X\right] ^2\) |
\(\mathrm {cov}(X,Y) \) | Covariance of random variables X and Y | \(\mathbb {E}\left[ X Y\right] - \mathbb {E}\left[ X\right] \mathbb {E}\left[ Y\right] \) |
Epidemic and branching process dynamics | ||
\(\tau \) | Rate of transmission across a network link | |
\(\gamma \) | Rate of recovery from infection | |
\(\omega _k\) | Death rate for individual of type k | |
t | Real time | s |
\(\mathscr {B}\) | The limiting branching process | |
\(E_N\) | The epidemic process in a population of size N | |
K | A large value of infectious population size | |
\(Z^{(k)}_i(t)\) | Random number of individuals of type i in the branching process at time t given initial type k | |
\(\pi _k(t)\) | Probability that the branching process is extinct at time t given initial type k | \(1-q_k(t)\) |
2 Model and approximating branching process
2.1 Model
We consider the spread of an SIR epidemic on a network of N individuals, labelled \(1,2,\ldots ,N\), constructed using the configuration model as follows (see e.g. Newman 2002). Let D be a random variable which describes the degree of a typical individual and let \(p_k=\mathbb {P}(D=k)\)\((k=0,1,\ldots )\). Let \(D_1,D_2,\ldots ,D_N\) be independent realisations of D and, for \(i=1,2,\ldots , N\), attach \(D_i\) stubs (half-edges) to individual i. Pair up these stubs uniformly at random to form the edges in the network. If \(D_1+D_2+\cdots +D_N\) is odd, there will be a left-over stub, which is ignored; the resulting network may have other ‘defects’ such as self-loops and multiple edges between pairs of individuals but, provided that D has finite variance, such imperfections become sparse in the network as \(N \rightarrow \infty \) (see e.g. Durrett 2007, Theorem 3.1.2). An alternative to the degrees \(D_1,D_2,\ldots ,D_N\) being random is, for each \(N=1,2,\cdots \), to replace \( \mathbf {D} =(D_1,D_2,\ldots ,D_N)\) by \( \mathbf {D} ^{(N)}=\left( D_1^{(N)},D_2^{(N)},\ldots ,D_N^{(N)}\right) \), where the degree sequences \( \mathbf {D} ^{(N)}\)\((N=1,2,\ldots )\) are prescribed and satisfy \(p_k^{(N)}=N^{-1} \sum _{i=1}^N \delta _{k,D_i^{(N)}} \rightarrow p_k\) as \(N \rightarrow \infty \)\((k=0,1,\ldots )\), where the Kronecker delta \(\delta _{k,j}\) is 1 if \(k=j\) and 0 otherwise (see e.g. Molloy and Reed 1995).
The epidemic is defined as follows. Initially some individuals are infective and the remaining individuals are susceptible. Infective individuals have independent infectious periods, each having an exponential distribution with rate \(\gamma \) (and hence mean \(\gamma ^{-1}\)), after which they become recovered and play no further role in the epidemic. Throughout its infectious period, an infective contacts each of its susceptible neighbours in the network at the points of independent Poisson processes having rate \(\tau \), so the probability that a given infective contacts a given neighbour before the infective recovers is \(\tau /(\gamma +\tau )\). Any contacted susceptible immediately becomes infective and may transmit the infection to any of its neighbouring susceptibles; i.e. there is no latent period. All the infectious periods and Poisson processes governing transmission of infection are mutually independent. The epidemic ends as soon as there is no infective present in the network.
2.2 The effective degree model
Ball and Neal (2008) introduced an ‘effective-degree’ construction of the above epidemic, in which the network is constructed as the epidemic progresses. The process starts with some individuals infective and the remaining individuals susceptible, but with none of the stubs paired up. For \(i=1,2,\ldots ,N\), the effective degree of individual i is initially \(D_i\). Infected individuals transmit infection by pairing their stubs with stubs attached to susceptible individuals in the following fashion. An infected individual makes infectious contacts down its unpaired stubs independently at rate \(\tau \) and is removed at rate \(\gamma \). When an infective, individual i say, transmits infection down a stub that stub is paired with a stub (attached to individual j, say) chosen independently and uniformly at random from all the unpaired stubs, to form an edge. The effective degrees of individuals i and j are both reduced by 1. If \(i=j\) then the effective degree of individual i is reduced by 2 but this will not significantly affect the dynamics for large populations since the probability of it happening is \(O(N^{-1}\)). If individual j is susceptible then it becomes infective and can transmit infection down any of its unattached stubs. As before, the epidemic ends as soon as there is no infective present. The network is then typically only partially constructed but that does not matter if interest is focussed on properties of the epidemic. In the original formulation of Ball and Neal (2008), when an infective recovers its unpaired stubs, if any, were paired with stubs chosen uniformly at random without replacement from the set of unpaired stubs but that is unnecessary; the stubs from such an infective can simply be left in the set of unpaired stubs.
2.3 Approximating multitype branching process
Suppose that the size N of the network is large and the initial number of infectives is small. Then during the early stages of an epidemic it is very likely that each time an infective individual transmits infection down a stub that stub is paired with a stub belonging to a susceptible individual. It follows that the early stages of such an epidemic can be approximated by a branching process in which each newly-infected individual has their “full” effective degree (i.e. their actual degree minus one for the stub that is paired with their infector). This approximation can be made fully rigorous by considering a sequence of epidemics, indexed by N, and using a coupling argument; see e.g. Ball and Neal (2008), which treats a more general model in which infective individuals also make contacts with individuals chosen uniformly at random from the population. Let \(E_N\) denote the epidemic on a network of N individuals and let \(\mathscr {B}\) denote the approximating branching process. Then following Ball and Neal (2008) (see “Appendix 1”) if \(\mu _D=\mathbb {E}\left[ D\right] \) is finite then the epidemics \(E_1,E_2\dots \) and the branching process \(\mathscr {B}\) can be constructed on a common probability space so that, with probability one, over any finite time interval [0, t] the process of infectives in \(E_N\) and the branching process \(\mathscr {B}\) coincide for all sufficiently large N. The same result holds for the model with prescribed degree sequences provided that \(p^{(N)}_k \rightarrow p_k\)\((k=0,1,\ldots )\) and \(\mu _D^{(N)}=\sum _{k=0}^\infty p^{(N)}_k k \rightarrow \mu _D\) as \(N \rightarrow \infty \), where \(\sum _{k=0}^{\infty } p_k=1\) and \(\mu _D < \infty \).
As indicated in “Appendix 1”, the branching process \(\mathscr {B}\) is not an almost sure upperbound for the process of infectives in \(E_N\), so unlike in Theorem 3.1 of Ball and Donnelly (1995) which considers homogeneously mixing epidemics, one cannot simply use the dominated convergence theorem to deduce convergence of moments of the number of infectives in the epidemic process to corresponding moments of the branching process as \(N \rightarrow \infty \). If there is a maximum degree \(k_{\mathrm {max}}\) (i.e. \(p_k=0\) for all \(k>k_{\mathrm {max}}\), or \(p_k^{(N)}=0\) for all \(k>k_{\mathrm {max}}\) and all N in the model with prescribed degrees) then, for all N, the process of infectives in \(E_N\) is bounded above by a branching process in which each newly-infected individual has the maximun effective degree \(k_{\mathrm {max}}-1\), so in that case the dominated convergence theorem can be used to prove convergence of moments. In “Appendix 1”, we consider the case when there is no maximum degree and use uniform integrability arguments to determine sufficient conditions for the mean and variance of the number of infectives at any given time \(t \ge 0\), and the covariance of the number of infectives at any given times \(t,s \ge 0\), in the epidemic \(E_N\) to converge to the corresponding mean, variance and covariance of the branching process \(\mathscr {B}\) as \(N \rightarrow \infty \). Specifically, we prove that (i) in the model with prescribed degrees these moments converge if, in addition to the conditions given above, there exists \(\delta >0\) such that \(\mu _{D^{3+\delta }}^{(N)}=\sum _{k=0}^\infty p^{(N)}_k k^{3+\delta } \rightarrow \mu _{D^{3+\delta }}=\sum _{k=0}^\infty p_k k^{3+\delta } \) as \(N \rightarrow \infty \), where \(\mu _{D^{3+\delta }} < \infty \); and (ii) in the model with random degrees they converge if the moment-generating function \(M_{D^2}(\theta )=\mathbb {E}\left[ \exp \left( \theta D^2\right) \right] \) of \(D^2\) is finite for some \(\theta _0>0\). Note that the latter condition implies that \(\mathbb {E}\left[ D^\alpha \right] < \infty \) for all \(\alpha \ge 0\).
In this context, we note that, for the model with prescribed degrees, the weakest conditions obtained on the moments of the degree distribution for convergence of the scaled stochastic epidemic on to its deterministic limit are given by Janson et al. (2014), who require uniform boundedness of the second moment of \( \mathbf {D} ^{(N)}\). However that paper, and the other related papers cited in the second paragraph of Sect. 1.1, (i) are concerned with the entire time course of the epidemic; (ii) assume that either the epidemic starts with a positive fraction of the population infected in the limit as \(N \rightarrow \infty \), or if that limiting fraction is zero then the convergence is for epidemics which take off and involves a random time translation describing when the epidemic becomes suitably established; and (iii) consider the evolution of the proportion of the population that is susceptible, infective or recovered. By contrast, this paper is concerned with epidemics initiated by few infectives and considers the number, rather than proportion, of infectives during the early phase of such an epidemic. Under the coupling mentioned above, in the limit as \(N \rightarrow \infty \), if an epidemic takes off then the duration of its early (exponentially growing) phase tends almost surely to infinity.
The limiting branching process may be described by a continuous-time multitype Markov branching process, with the type of an infective corresponding to its effective degree. Let \({\tilde{D}}\) denote the (size-biased) degree of a typical neighbour of a typical individual in the network and let \({\tilde{p}}_k=\mathbb {P}({\tilde{D}}=k)\)\((k=1,2,\ldots )\). Then \({\tilde{p}}_k=\mu _D^{-1} k p_k\), where \(\mu _D=\mathbb {E}\left[ D\right] \), since when a stub is paired it is k times as likely to be paired with a stub from a given individual having degree k than it is with a stub from a given individual having degree 1. Under the branching process approximation, the effective degree of a newly infected individual is distributed according to \({\tilde{D}}-1\), since one of that individual’s stubs is ‘used up’ when it is infected. Note for future reference that \(\mu _{{\tilde{D}}}=\mu _D^{-1}\mu _{D^2}\) and, more generally, \(\mu _{f({\tilde{D}})}=\mu _D^{-1} \mu _{Df(D)}\) for any real-valued function f. (For a random variable, X say, we use \(\mu _X\) to denote its expectation \(\mathbb {E}\left[ X\right] \). Thus, for example, \(\mu _{Df(D)}=\mathbb {E}\left[ Df(D)\right] \).)
2.4 Explicit form for the multitype branching process
For \(t \ge 0\), let \(\mathbf {Z}(t)=\left( Z_i(t)\right) \), where \(Z_i(t)\) denotes the number of individuals of type i alive at time t, and let \(Z(t)=Z_0(t)+Z_1(t)+\cdots +Z_{k_{\mathrm {max}}}(t)= \mathbf {1}^{\!\top }\mathbf {Z}(t) \) denote the total number individuals alive at time t. For \(k\in \mathscr {K}\), we use the notation \(\{\mathbf {Z}^{(k)}(t):t \ge 0\}\), where \(\mathbf {Z}^{(k)}(t)=(Z^{(k)}_i(t))\), to denote a process starting with a single individual, whose type is k, at time 0 (i.e. where \(Z_i^{(k)}(0) = \delta _{i,k}\), \(i \in \mathscr {K}\)). Further, \(Z^{(k)}(t)=\mathbf {1}^{\!\top }{\mathbf {Z}^{(k)}(t)}\) denotes the total number of individuals at time t in such a process.
3 Behaviour of means
For \(t\ge 0\), let \( \mathbf {m} (t)=(m^{(i)}(t))=( \varvec{M} (t)\mathbf {1})^{\!\top }\). Thus the kth element of \( \mathbf {m} (t)\) contains the mean total population size at time t given that the process starts with a single individual whose type is k. We derive a simple expression for \( \mathbf {m} (t)\). The following proposition is useful.
Proposition 1
Proof
4 Variance
The variance in infectious prevalence during the exponentially growing phase of an epidemic was considered in Graham and House (2014), but using the diffusion limit and an argument about the neighborhood around an infective. A branching process limit lets us be more explicit. For \(k\in \mathscr {K}\), let \( \mathbf {u} _k\) denote the length-\((k_{\mathrm {max}}+1)\) column vector whose element corresponding to type k is 1 and all other elements are 0, so \({ \mathbf {u} _k}=(\delta _{i,k})\). For \(t \ge 0\) and \(k \in \mathscr {K}\), let \(\sigma _{ij}^{(k)}(t)=\mathrm{cov}\left( Z^{(k)}_i(t),Z^{(k)}_j(t)\right) \)\((i,j \in \mathscr {K})\).
5 Covariance function
6 Unbounded degree distributions
The above results have all assumed that there is a maximum degree \(k_{\mathrm {max}}\). Suppose that is not the case, so the branching process has countably many types. For \(t\ge 0\), let \(\mathbf {Z}(t)=(Z_0(t), Z_1(t),\ldots )^{\!\top }\), where \(Z_i(t)\) denotes the number of individuals of type i alive at time t, and let \(Z(t)=\sum _{i=0}^\infty Z_i(t)\) denote the total number individuals alive at time t. (For ease of notation we drop explict reference to the type of the initial individual.) For \(k_{\mathrm {max}}=1,2,\ldots \), let \(\{\mathbf {Z}(t,k_{\mathrm {max}}):t \ge 0\}\) denote the branching process derived from \(\{\mathbf {Z}(t):t \ge 0\}\) by ignoring all individuals having type strictly greater than \(k_{\mathrm {max}}\) and any offspring of such individuals. For \(t \ge 0\), let \(Z(t, k_{\mathrm {max}})=\mathbf {1}^{\!\top }\mathbf {Z}(t,k_{\mathrm {max}})\) be the total number of individuals alive in \(\{\mathbf {Z}(t,k_{\mathrm {max}}):t \ge 0\}\) at time t. Now, for any \(t \ge 0\), \(Z(t, k_{\mathrm {max}})\) is monotonically increasing in \(k_{\mathrm {max}}\) and converges almost surely to Z(t) as \(k_{\mathrm {max}}\rightarrow \infty \). Thus, by the monotone convergence theorem, \(\mathbb {E}\left[ Z(t)\right] =\lim _{k_{\mathrm {max}}\rightarrow \infty }\mathbb {E}\left[ Z(t,k_{\mathrm {max}})\right] \).
7 Probability of extinction
We will now consider three regimes in which asymptotic methods can be used to bound the real-time behaviour of the probabilities of extinction. In particular, we will see that early real-time behaviour is bounded by the death rates \(\omega _k\), while late-time behaviour is bounded by the asymptotic real-time growth rate r provided \(r>-\gamma \).
7.1 Late behaviour of the subcritical case
7.2 Late behaviour of the supercritical case
7.3 Early behaviour and matched asymptotics
Matched asymptotics is a standard technique in mathematical biology for writing down approximations to non-linear models that match known asymptotic behaviour (Murray 2002, 2003). While numerical solution of the ODEs (7.2) is efficient (as we have noted above) we now obtain a crude approximation to the full system that takes a closed form in terms of elementary functions.
8 Fluctuations in the emerging phase of a major outbreak
Recent work by Constable and McKane (2014) considered the reduction of high-dimensional stochastic models to low-dimensional diffusions and this approach was shown to be asymptotically exact for some systems in the small-noise limit by Parsons and Rogers (2015). It is an open question whether the argument in Sect. 3.3 of Graham and House (2014) could be justified rigorously by a similar argument, however we note that a branching-process approach makes fewer assumptions than a low-dimensional diffusion limit and so will be more generally applicable.
Recall that the processes \(\{{\hat{Z}}_i(t):t \ge 0\}\)\((i=1,2,\ldots ,K)\) are mutually independent. It follows using the central limit theorem that, for sufficiently large K, the process \(\{Z(t):t \ge 0\}\), which approximates the prevalence of infection during the early growth of an epidemic, is approximately Gaussian with mean function given by (8.1) and covariance function given by (8.3).
9 Numerical examples
9.1 Forward simulations
We conducted a series of numerical experiments to provide specific examples of the general results presented here. \(M=10^4\) Monte Carlo simulations were performed on three different configuration model networks, each of size \(N=10^4\), and with the degree distributions shown in Fig. 2 Row (a). (Note that each Monte Carlo simulation consisted of first simulating a network and then simulating a single epidemic on it.) Two different scenarios were considered. In the first – most commonly considered in the literature when simulations are compared to analytic approaches – \(\mathrm {time}=0\) was defined as the first point when prevalence is at a given level, K. In our simulations we took \(K=100\), but in general K should take a value where the probability of extinction has become negligible, but the depletion of the susceptible population has not had a significant effect on the epidemic dynamics. In the second, each epidemic was started from one node, picked uniformly at random, so the probability of extinction played a major role. This scenario is less commonly considered when comparing real-time simulated epidemics to differential equation models because the latter are typically designed to hold when the epidemic is already established.
Since analytic results for the probabilities of extinction \(\pi _k(t)\) are not available, the branching process results required numerical integration of ordinary differential equations (in our case using Runge–Kutta methods). We stress that the computational effort required to do this is much less than that involved in performing Monte Carlo simulations, and has the benefit of not depending on N.
Figure 3 shows the results for the second approach in which there is one randomly chosen initial infective at time 0. Row (a) shows some sample trajectories. Row (b) shows the extinction probabilities, which are accurately captured in the branching process model until very close to the end of the epidemic when prevalence is low and extinction becomes likely again. Row (c) shows the mean, which does not start growing at a constant rate with the convergence rate accurately captured in the branching process model. Row (d) shows the variance and convergence onto its asymptotic value; in this case there is a single maximum just before the peak in prevalence.
Another important point is that while mean numbers infected are comparable between Figs. 2 and 3, the variability in the time for the epidemic to take off, as well as the contribution from extinct epidemics, makes the real-time variance in Fig. 3 orders of magnitude larger than in Fig. 2.
9.2 Statistical inference
In order to demonstrate the potential use of the real-time effective degree branching process model for statistical inference, we carried out a simulation study. Here we simulated one epidemic that took off on a configuration model network of size \(N=10^6\) with degree distribution \(D^{(3)}\) as in the right-hand column of Fig. 2 (\(d^{(3)}_1 = 1/8\), \(d^{(3)}_3 = 5/6\), \(d^{(3)}_9 = 1/24\)) and true rates \({\tau }_0 = 2\), \({\gamma }_0 = 1\). Letting I(t) be the prevalence of infection in the network model, we set time \(t=0\) when \(I(t)=100\) for the first time and make 40 evenly-spaced observations (with gap \(\delta t = 0.05\) between each) of I(t) up to \(t_{\mathrm {end}}=2\).
We then define an approximate likelihood based on the methods of Ross et al. (2006), in which a Gaussian process approximation based on known first and second moments is used, which will be more accurate for larger N, larger I(0) and smaller \(\delta t\). There should be, however, no a priori obstacle to fitting our model to data on smaller populations even with incomplete data, for example by using Markov chain Monte Carlo methods to perform multiple imputation as suggested by O’Neill and Roberts (1999).
10 Concluding comments
10.1 Summary of results
In this paper, we have provided explicit closed-form expressions for the real-time mean, variance and covariance function for disease prevalence during the early stages of the Markovian SIR model on a configuration model network, as well as deriving differential equations for the probabilities of extinction over time that are relatively numerically cheap to solve. These allow for a more explicit treatment of e.g. rates of convergence to asymptotic behaviour than has previously been possible.
10.2 Future directions
We believe that the methods of real-time, multitype branching processes could be more widely applied in infectious disease epidemiology, since they provide results concerning extinction and variance that are not available using deterministic differential equation models. For example, the effective-degree based methodology presented here may be extended to include degree correlation (e.g. in the sense of Newman 2002) by keeping track of the actual, as well as effective, degrees of individuals, though the type space becomes larger and explicit analytic results are unlikely to be available. We note that there is increasing interest in the eradication of infections (e.g. Klepac et al. 2013) and that arguably calculating extinction probabilities and variability in outbreak sizes is of equal or greater importance in this context than calculation of mean behaviour.
The explicit closed-form expressions derived have the potential to enhance statistical work on epidemic prevalence curves. In particular, many empirically observed epidemics of human pathogens exhibit more variability around the trend than simple models would predict (see Black et al. 2014, particularly Section 1, for a discussion of this), which can bias parameter estimation if an insufficiently variable model is used. Application of our methods to real data would be an interesting extension of our work.
The possibility of a more general non-Markovian stochastic epidemic being approximated by an appropriate real-time branching process is raised by the results of Barbour and Reinert (2013) and it would be interesting to investigate whether our analysis could be adapted to this scenario.
Finally, there is the question of low-dimensional PGF-based modelling of the whole network epidemic that incorporates stochasticity accurately. For example, the work of Miller (2014) considered accounting for early fluctuations and Graham and House (2014) considered a diffusion approximation once early fluctuations were negligible, but the results presented here as well as those of Barbour and Reinert (2013) suggest that a more unified low-dimensional stochastic approach that explicitly models early fluctuations may be possible.
Footnotes
Notes
Acknowledgements
We gratefully acknowledge support from the Isaac Newton Institute for Mathematical Sciences, Cambridge, where we held Visiting Fellowships under the Infectious Disease Dynamics programme and its follow-up meeting, during which this work was initiated. TH is supported by the Engineering and Physical Sciences Research Council (Grant number EP/N033701/1). We would like to thank Josh Ross for helpful comments on this manuscript. We would also like to thank the referees and associate editor for their constructive comments which have improved the presentation of the paper.
References
- Athreya KB, Ney PE (1972) Branching processes. Springer, BerlinCrossRefMATHGoogle Scholar
- Bailey NTJ (1957) The mathematical theory of epidemics. Griffin, LondonGoogle Scholar
- Ball F, Donnelly P (1995) Strong approximations for epidemic models. Stoch Process Appl 55(1):1–21MathSciNetCrossRefMATHGoogle Scholar
- Ball F, Neal P (2008) Network epidemic models with two levels of mixing. Math Biosci 212(1):69–87MathSciNetCrossRefMATHGoogle Scholar
- Barbour A, Reinert G (2013) Approximating the epidemic curve. Electron J Probab 18(54):1–30MathSciNetMATHGoogle Scholar
- Black AJ, House T, Keeling MJ, Ross JV (2014) The effect of clumped population structure on the variability of spreading dynamics. J Theor Biol 359:45–53MathSciNetCrossRefGoogle Scholar
- Bohman T, Picollelli M (2012) SIR epidemics on random graphs with a fixed degree sequence. Random Struct Algorithms 41(2):179–214MathSciNetCrossRefMATHGoogle Scholar
- Constable GWA, McKane AJ (2014) Fast-mode elimination in stochastic metapopulation models. Phys Rev E 89(3):032141CrossRefGoogle Scholar
- Daley DJ, Vere-Jones D (1988) An introduction to the theory of point processes. Probability and its applications. Springer, New YorkMATHGoogle Scholar
- Danon L, Ford AP, House T, Jewell CP, Keeling MJ, Roberts GO, Ross JV, Vernon MC (2011) Networks and the epidemiology of infectious disease. Interdiscip Perspect Infect Dis 2011:1–28CrossRefGoogle Scholar
- Decreusefond L, Dhersin JS, Moyal P, Tran VC (2012) Large graph limit for an SIR process in random network with heterogeneous connectivity. Annal Appl Probab 22(2):541–575MathSciNetCrossRefMATHGoogle Scholar
- Dorman K, Sinsheimer J, Lange K (2004) In the garden of branching processes. SIAM Rev 46(2):202–229MathSciNetCrossRefMATHGoogle Scholar
- Durrett R (2007) Random graph dynamics. Cambridge University Press, CambridgeMATHGoogle Scholar
- Eames KTD, Keeling MJ (2002) Modeling dynamic and network heterogeneities in the spread of sexually transmitted diseases. PNAS 99(20):13330–13335CrossRefGoogle Scholar
- Ethier SN, Kurtz TG (1986) Markov processes: characterization and convergence. Wiley series in probability and mathematical statistics. Wiley, HobokenCrossRefGoogle Scholar
- Graham M, House T (2014) Dynamics of stochastic epidemics on heterogeneous networks. J Math Biol 68(7):1583–1605MathSciNetCrossRefMATHGoogle Scholar
- Grimmett GR, Stirzaker DR (2001) Probability and random processes, 3rd edn. Oxford University Press, OxfordMATHGoogle Scholar
- Heesterbeek H, Anderson RM, Andreasen V, Bansal S, De Angelis D, Dye C, Eames KTD, Edmunds WJ, Frost SDW, Funk S, Hollingsworth TD, House T, Isham V, Klepac P, Lessler J, Lloyd-Smith JO, Metcalf CJE, Mollison D, Pellis L, Pulliam JRC, Roberts MG, Viboud C, Isaac Newton Institute IID collaboration (2015) Modelling infectious disease dynamics in the complex landscape of global health. Science 347(6227):aaa4339Google Scholar
- Heinzmann D (2009) Extinction times in multitype Markov branching processes. J Appl Probab 46(1):296–307MathSciNetCrossRefMATHGoogle Scholar
- Holme P (2013) Extinction times of epidemic outbreaks in networks. PLoS ONE 8(12):e84429CrossRefGoogle Scholar
- House T, Keeling MJ (2010) Insights from unifying modern approximations to infections on networks. J R Soc Interface 8(54):67–73CrossRefGoogle Scholar
- Janson S, Luczak M, Windridge P (2014) Law of large numbers for the SIR epidemic on a random graph with given degrees. Random Struct Algorithms 45(4):726–763MathSciNetCrossRefMATHGoogle Scholar
- Klepac P, Metcalf CJE, McLean AR, Hampson K (2013) Towards the endgame and beyond: complexities and challenges for the elimination of infectious diseases. Philos Trans R Soc Lond B Biol Sci 368(1623): 20120137Google Scholar
- Lindquist J, Ma J, Driessche P, Willeboordse FH (2010) Effective degree network disease models. J Math Biol 62(2):143–164MathSciNetCrossRefMATHGoogle Scholar
- Miller JC (2011) A note on a paper by Erik Volz: SIR dynamics in random networks. J Math Biol 62(3):349–358MathSciNetCrossRefMATHGoogle Scholar
- Miller JC (2014) Epidemics on networks with large initial conditions or changing structure. PLoS ONE 9(7):e101421CrossRefGoogle Scholar
- Miller JC, Kiss IZ (2014) Epidemic spread in networks: existing methods and current challenges. Math Model Nat Phenom 9(2):4–42MathSciNetCrossRefMATHGoogle Scholar
- Miller JC, Slim AC, Volz EM (2012) Edge-based compartmental modelling for infectious disease spread. J R Soc Interface 9(70):890–906CrossRefGoogle Scholar
- Molloy M, Reed B (1995) A critical point for random graphs with a given degree sequence. Random Struct Algorithms 6:161–179MathSciNetCrossRefMATHGoogle Scholar
- Murray JD (1984) Asymptotic analysis, applied mathematical sciences, vol 48. Springer, New YorkCrossRefGoogle Scholar
- Murray JD (2002) Mathematical biology I, 3rd edn. Springer, BerlinGoogle Scholar
- Murray JD (2003) Mathematical biology II, 3rd edn. Springer, BerlinGoogle Scholar
- Newman MEJ (2002) Assortative mixing in networks. Phys Rev Lett 89(20):208701CrossRefGoogle Scholar
- Newman MEJ (2002) Spread of epidemic disease on networks. Phys Rev E 66(1):016128MathSciNetCrossRefGoogle Scholar
- O’Neill PD, Roberts GO (1999) Bayesian inference for partially observed stochastic epidemics. J R Stat Soc A 162:121–129CrossRefGoogle Scholar
- Parsons TL, Rogers T (2015) Dimension reduction via timescale separation in stochastic dynamical systems. [arXiv:1510.07031]
- Ross JV, Taimre T, Pollett PK (2006) On parameter estimation in population models. Theor Popul Biol 70(4):498–510CrossRefMATHGoogle Scholar
- Volz EM (2008) SIR dynamics in random networks with heterogeneous connectivity. J Math Biol 56(3):293–310MathSciNetCrossRefMATHGoogle Scholar
- Waugh WAO (1958) Conditioned Markov processes. Biometrika 45(1–2):241–249MathSciNetCrossRefMATHGoogle Scholar
- Windridge P (2015) The extinction time of a subcritical branching process related to the SIR epidemic on a random graph. J Appl Probab 52(4):1195–1201Google Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.