Introduction To Compartmental Modeling For The Buddinginfectious Disease Modeler
Introduction To Compartmental Modeling For The Buddinginfectious Disease Modeler
Introduction To Compartmental Modeling For The Buddinginfectious Disease Modeler
EDUCATION ARTICLE
1. Introduction
Recent high profile infectious disease outbreaks around the globe – including SARS,
Ebola, and Zika – have reignited a general interest in understanding the spread of disease.
The utility of mathematical and computational approaches to simulate and numerically
analyze infections within populations has also benefited from major advances in com-
putational power over the past several decades. As a result, public health officials and
policy makers have increasingly relied on a quantitative understanding of infectious disease
dynamics.
Mathematical models of infectious disease dynamics have a rich history that dates back
more than 100 years. Mathematically simple formulations that describe the transition of
individuals in a population between ‘compartments’ that capture the infection status of
individuals leads to surprisingly significant insight. Their elegance and simplicity allow
the ease of expansion to more complexities through, for example, the addition of com-
partments. Expanding these models is often straight forward but the apparent simplicity
can mask subtle, but important, model structure and parameterization choices. Further-
more, while there are many wonderful texts focused on infectious disease modeling, there
exist several complexities that are rarely discussed in sufficient detail for a novice disease
modeler.
Here we highlight several of these potential pitfalls, describe how they arise and, when
possible, how to address them. We also point to further resources in the literature. We begin
with a brief introduction to disease modeling as a means to provide basic context as well
as introduce the notation that we will use throughout this paper. We often provide more
mathematical detail and discussion in sections in the Appendix.
Following our introduction, we focus on analysis of the compartmental epidemiolog-
ical model with an emphasis on computing the basic reproductive number, commonly
known as R0 . We focus on the following two difficulties: (i) deriving the force of infec-
tion and (ii) instances when R0 does not represent a precise disease extinction threshold.
Keep in mind, this is by no means meant to be a all-encompassing discussion of infec-
tious disease modeling but a resource to supplement other more comprehensive texts
(e.g. Anderson, May, & Anderson, 1992; Brauer, 2008; Brauer & Castillo-Chavez, 2001;
Brauer & Kribs, 2015; Brauer, vanden Driessche, & Wu, 2008; Keeling & Rohani, 2008;
Martcheva, 2015; Vynnycky & White, 2010).
some infections that are frequently modeled within the SIR framework include many com-
mon childhood diseases such as chicken pox and measles. An important common feature
among these pathogens is that infection is thought to confer life-long immunity so that,
from a modeling perspective, individuals move directly into (and remain in) the recovered
compartment upon recovery.
Now, to model the infection process we consider a system of ordinary differential equa-
tions describing the change in the population size of each compartment. Let λ be the force
of infection, or the per capita rate at which susceptible individuals acquire infection. It is
important to note that the force of infection is not a constant but rather a function of the
size of the infectious compartments; that is, λ = λ(I).
The definition of λ(I) can indeed be tricky and depends on assumptions about the
nature of interactions between individuals (see Section 4.1). However, complications in
the different ways to define the force of infection most often come into play when the total
population size is changing, which is not the case at the moment. Thus, we will ignore
this complexity for the time being. The force of infection is comprised of a transmission
rate β or, more precisely, the product of the contact rate and the probability of trans-
mission given contact, and an interaction term with infectious individuals. In order to
preserve the units of the equation, using our definition of β (which has units time−1 ),
we will think about the interaction with the proportion of infectious individuals: I/N.
Thus,
I
λ(I) = β .
N
The definition of λ(I) will change for each model, although it will always remain a function
of the infectious class(es). Here, we have written λ(I) as a function of the infectious class,
but frequently the dependence is not explicitly included. in model formulation.
For now, we ignore demography, i.e. natural birth and mortality, but relax this assump-
tion later. We can now write down a system of ordinary differential equations (ODEs) of
the SIR model:
dS dI dR
= −λ(I)S, = λ(I)S − γ I, = γ I. (1)
dt dt dt
A schematic diagram of this model is provided in Figure 1(A). Here, the nature of our
equations show us that the population is ‘closed’; that is, the total population size (defined
by N = S+I+R) is constant over time (dN/dt = 0). A nice property of this assumption
is that the system is entirely determined by two, rather than three, of the equations. For
example, we can write the recovered population as R = N−S−I so that only susceptible
and infected populations need to be determined explicitly.
A final comment before we move into the model analysis is that throughout this paper S,
I, and R always represent population densities (although we will refer to population density
and size interchangeably throughout this paper). However, some texts define S, I, and R as
proportions of the population in each class, respectively. Non-dimensionalizing the state
variables in this way for disease models is a common modeling choice, and the associated
derivation is demonstrated in the Appendix (see Appendix 2). Our choice to consistently
write the models in terms of densities, rather than proportions, will make it more straight
forward to interpret some common pitfalls in disease modeling.
198 J. C. BLACKWOOD AND L. M. CHILDS
Figure 1. (A) Schematic representation of the standard SIR model in the absence of demography, as in
Equation (1). S, I, and R represent the total number of susceptible, infected, and recovered individuals
in a population, respectively. λ(I) is the force of infection, and γ is the recovery rate. The dashed arrow
indicates the necessity of interaction with the I compartment for transition from S to I. (B) Same as above,
but this schematic includes demography in grey (as in Equation (2)).
immunity. Without such replacement, the number of infected individuals returns to zero
following an epidemic, as in the SIR model without demography (i.e. in Equation (1)). In
this section, we therefore introduce a small modification to the SIR model of Section 2.1:
demography, i.e. the inclusion of natural birth and death.
For mathematical simplicity, we make the following key biological assumptions: (i) the
birth rate and death rate are equal and given by μ, (ii) all individuals are capable of repro-
ducing and are equally subject to mortality, and (iii) all individuals are born susceptible
to infection. The first two assumptions maintain a constant population size and the third
assumption is made for simplicity (and is a valid assumption for many directly transmitted
pathogens). It is possible to model infectious disease dynamics without these assumptions.
For example, a discussion of models that include demography with variable population size
can be found in Brauer (2008), Hethcote (2000), and Ledder (2017). The simple SIR model
with demography is given by:
dS I dI I dR
= μN − β S − μS, = β S − (γ + μ)I, = γ I − μR (2)
dt N dt N dt
and a schematic representation of this model is provided in Figure 1(B). For this model,
we can find both a DFE and an endemic equilibrium, which are given by:
∗ ∗ ∗ γ +μ μ β ∗ ∗
(S , I , R ) = N, N − 1 ,N − S − I , (4)
β β γ +μ
respectively.
Finding equilibria is only the first step to understanding long-term behaviour in a sys-
tem. We must also determine which of the behaviours are typically realized. This involves
determining the stability of the equilibria and will show us whether we will approach the
equilibria or move away from it, assuming we have started nearby. In order to determine the
stability of each of these equilibria, we can apply a standard linear stability analysis (more
details can be found in any standard ODEs or linear algebra textbook). In brief, we lin-
earize our system around each equilibria using the Jacobian matrix evaluated at the chosen
equilibrium of interest. We then find the associated characteristic polynomial, whose roots
are the eigenvalues to the Jacobian matrix. For local stability of the equilibrium, we require
that all of the eigenvalues have negative real parts; otherwise, the equilibrium is locally
unstable. In some cases, the value of the eigenvalues can be calculated directly. However,
under more complicated scenarios an alternative method can be used to determine the
sign of the eigenvalues (e.g. with a symbolic programming language or analytically with
the Routh-Hurwitz criteria, see Brauer & Castillo-Chavez, 2001).
Using linear stability analysis, we can determine the local stability of each equilib-
rium. However, even for simple systems, this does not always give us a clear biological
interpretation on when disease will spread or die out globally. Furthermore, when using
more complicated models, there may be multiple stable equilibria present simultaneously.
In such cases, it is important to obtain a more global picture of the infection dynamics
using, for example, bifurcation analysis to supplement the linear stability analysis (as in
Section 4.2.1).
200 J. C. BLACKWOOD AND L. M. CHILDS
In the standard SIR model with demography, however, we can use a standard linear
stability analysis to identify conditions under which the DFE and endemic equilibrium are
stable as well as unstable. As it turns out, stability is closely related to an important quantity
in epidemiology: R0 , often called the basic reproductive number (or ratio). Therefore, we
introduce R0 before discussing the stability of this model.
dI I
= β S − (γ + μ)I.
dt N
We rearrange this inequality to a form which provides biological intuition, and now shifts
the threshold of interest to one (rather than zero when considering whether the size in the
compartment will grow or shrink). Thus,
β N−1
⇒ < 1.
γ +μ N
In other words, the threshold is determined by the product of the transmission rate β with
the quantity 1/(γ + μ) (which is related to the average time spent in the infectious class
and the average lifespan, see Appendix 1) and the initial fraction of susceptible individuals.
Combined this must be ‘small enough’ (i.e. < 1) so that the epidemic cannot take off. As
long as we are considering relatively large population sizes, we can look at the approximate
LETTERS IN BIOMATHEMATICS 201
β
<1
γ +μ
implies the disease will die out. A parallel analysis that begins with the question ‘under
what conditions will a disease persist?’ reveals that
β
>1
γ +μ
implies the disease will persist. This quantity, β/(γ + μ), is defined as the basic reproduc-
tive number for the SIR system. It is a threshold quantity where the ‘critical value’ is at
R0 = 1, and we are now better able to interpret our original definition: if a typical infec-
tious individual infects less than one person on average during the time they are infected
then a disease will die out. In contrast, if a typical infectious individual infects more than
one person on average during the infectious period then the disease will spread. Intuitively,
this makes sense: R0 is the product of the transmission rate and – roughly – the average
amount of time spent in the infectious class.
Notice that this definition includes several important assumptions: (i) R0 is the aver-
age number of onward infections produced by a single individual (and therefore excludes
individual variation), (ii) the average transmissibility among individuals is used in the
definition of R0 (again ignoring individual variation), and (iii) all individuals in the pop-
ulation are susceptible at time t = 0. In other more complicated settings, this definition
changes such as when one of these assumptions is broken by the model structure (see
Section 4.1.2 on structured populations). We note that the mathematical form of R0
will differ for each model and numerically will depend on model parameters. How-
ever, almost all methods of computing R0 require examination of when the DFE loses
stability.
As you may expect, the way we have derived the basic reproductive number here by
examining whether the infectious population grows or shrinks at time zero is only possible
when the model framework is simple. However, there are multiple ways to determine R0
in a more general setting, some of which we discuss later (see Section 3).
Jacobian matrix evaluated at the DFE have negative real part, to show:
Disease-free
equilibrium (DFE) is: Endemic
equilibrium is:
linearly stable if R0 < 1 unstable if R0 < 1
unstable if R0 > 1 linearly stable if R0 > 1
We emphasize that when using most standard methods for the computation of R0 , which
involve linearization around the DFE, we can really only say something definitively about
the local stability of the DFE. In other words, the derivation of R0 typically depends on
the stability of the DFE. As R0 increases past the critical threshold value of one, the DFE
is destabilized – thereby often, and indeed in the case of the SIR model with demography,
this leads to stability of the endemic equilibrium. We will introduce an example later where
this is not the case (see Section 4.2.1).
Figure 2. Schematic representation of the SEIR model. λ(I) is the force of infection, σ is the rate at which
individuals transition from the exposed class to the infectious class, γ is the recovery rate, and μ is the
natural mortality rate. The dashed arrow indicates the necessity of interaction with the I compartment
for transition from S to E.
dE
= λ(I)S − (σ + μ)E, (6)
dt
dI
= σ E − (γ + μ)I, (7)
dt
dR
= γ I − μR, (8)
dt
where N = S+E+I+R. In this model, the force of infection λ(I) is the same as in the SIR
model (λ(I) = βI/N). There are no new infections contributed through interactions with
the exposed individuals, as they are not infectious. If we try to find R0 using the same meth-
ods as we did for the SIR model, we run into trouble. More precisely, when examining the
behaviour of I at t = 0, in contrast to the SIR model, we cannot factor out an I. Fortunately,
the next generation matrix method allows us to generalize the process of computing R0 .
A nice feature of the next generation matrix method is that it only requires use
of the DFE which is often easy to compute. In the SEIR model, the DFE is given by
(S∗ , E∗ , I ∗ , R∗ ) = (N, 0, 0, 0). After determining the DFE, we must create a sub-model that
only considers the ‘disease’ compartments, a subset of the equations in the SEIR model. The
disease compartments are those that include individuals that are in any stage of infection
which, for the SEIR model, includes both the exposed and infectious individuals. There-
fore, this sub-model will only contain the E and I equations (i.e. Equations (6)–(7)), which
we write in the form
dx
= F (x) − V (x) (9)
dt
where x is a vector of the j disease compartments; in the SEIR model j = 2 because the
disease compartments are E and I. The right hand sides of Equations (6)–(7) are therefore
contained in the vectors F (x) and V (x).
Here, F (x) contains any terms that directly lead to new infections entering each com-
partment j. Notice that the second element of F is zero because no new infections enter the
I compartment, rather they transition from the E compartment into the I compartment.
V (x), on the other hand, can be further broken down as
V (x) = V − (x) − V + (x) (10)
where V − (x) and V + (x) contain all other outputs and inputs, respectively, from each dis-
ease class. This includes terms such as mortality or transition between classes. Although
204 J. C. BLACKWOOD AND L. M. CHILDS
Notice, this is exactly equivalent to our sub-model original equations (Equations (6)
and (7)).
Next, we must linearize around the DFE, which involves the Jacobian, i.e. the
matrix of partial derivatives, evaluated at the DFE. Recalling that the DFE is given by
(S∗ , E∗ , I ∗ , R∗ ) = (N, 0, 0, 0), the Jacobian matrix of the sub-model evaluated at the DFE
is given by
⎡ ∂E ∂E ⎤
∗ ∗ ∗ ∗ ⎢ ∂E ∂I ⎥
J(S , E , I , R ) = ⎣ ⎦ (12)
∂I ∂I
∂E ∂I N,0,0,0
S
−(σ + μ) β
= N (13)
σ −(γ + μ) N,0,0,0
−(σ + μ) β
= . (14)
σ −(γ + μ)
As a result, we can factor out the vector x = EI on the right hand side, leaving us with
⎡ ⎤
dE
⎢ dt ⎥ 0 β σ +μ 0 E
⎣ dI ⎦ = − . (15)
0 0 −σ γ +μ I
dt
Notice that in the F matrix the only non-zero term β arises from the derivative of the
first row of F with respect to I, evaluated at S = N. To simplify, we define matrices in
Equation (15) as F and V, respectively, and rewrite Equation (15) as
dx
= (F − V)x. (16)
dt
Recall from Section 2.2.2 that the ultimate goal is to determine when the number of infec-
tious individuals, I, is increasing or decreasing when the population is near the DFE,
LETTERS IN BIOMATHEMATICS 205
i.e. when a single infected individual is introduced into an otherwise susceptible popu-
lation. We can now investigate a similar question: when are the compartments containing
the infectious classes increasing or decreasing near the DFE? From linear stability anal-
ysis we know that stability of the DFE is equivalent to the real part of all eigenvalues of
(F − V) being less than zero. From Equation (16), we can extract what is known as the
next generation matrix, or
FV−1 .
Here, an analogous threshold (to the eigenvalues of the Jacobian) for when the DFE is stable
is given by the spectral radius of the next generation matrix denoted by
R0 = ρ FV−1 < 1.
where the spectral radius (ρ) is the eigenvalue with the largest magnitude. While we
omit the proof here, it can be found in other sources, e.g. van den Driessche and Wat-
mough (2002).
For the SEIR model, the next generation matrix is given by
⎡ ⎤
βσ β
FV−1 = ⎣ (σ + μ)(γ + μ) γ + μ ⎦ .
0 0
We find the eigenvalues of this matrix are βσ /(γ + μ)(σ + μ) and 0. Here, R0 (or the
spectral radius of FV−1 ) is given by
βσ
R0 = .
(γ + μ)(σ + μ)
For the SEIR model, notice that R0 is very similar to the basic reproductive number
in the SIR model with demography, (β/(γ + μ)), multiplied with one additional term,
(σ/(σ + μ)). This term leads to a smaller R0 for the SEIR model because some individu-
als are lost to natural mortality while in the E class and never actually reach the infectious
class.
While we provided a brief introduction to the next generation matrix in this section
using a particular example (the SEIR model), more formal mathematical treatments of this
can be found in other references, e.g. Diekmann and Heesterbeek (2000), Martcheva (2015)
and van den Driessche and Watmough (2002). We also note that analytically finding the
roots of the characteristic polynomial (and thereby finding R0 ) can be challenging; it is
common to use a symbolic mathematical programme (e.g. Mathematica) or to find it
numerically (e.g. using MATLAB). Furthermore, a great deal of tractability, both math-
ematically and biologically, may be gained by first non-dimensionalizing the model. Such
simplifications may reveal biologically relevant quantities that were obscured by relying
solely on the mathematical computation. We refer the reader to Ledder (2017) for a more
detailed description of non-dimensionalization of models.
instances in which the construction of a model is not amenable to this – or other – standard
methods for analytically finding R0 .
One common example in which R0 cannot be directly computed is when there is
‘importation’. Suppose a particular pathogen follows the standard SEIR structure with
demography as in Equations (5)–(8). Now, assume that there is some external source of
infection that does not depend on the number of infected individuals present in the popu-
lation. For example, an infected individual from another population may briefly encounter
a susceptible individual and transmit infection. If we assume the importation rate is given
by φ, then this modifies the force of infection so that
I
λ(I) = β + φ.
N
The S and E equations change to ensure that the total population size remains constant.
While R0 cannot be computed explicitly for a model containing infectious imports, some
modeling studies still attempt to take advantage of the information R0 provides in similar
standard models. For example, R0 can first be computed for the model in the absence of
imports (i.e. when φ = 0). This R0 then represents an ‘intrinsic’ basic reproductive num-
ber; that is, R0 in the absence of any spatial processes. The dynamical impact of importation
can then be explored numerically through simulations (e.g. Blackwood, Streicker, Altizer,
& Rohani, 2013). Other models with importation focus on other aspects of the model all
together, such as direct parameterizations that are independent of the precise R0 value itself
(e.g. Shrestha et al., 2015, 2013).
the force of infection λ(I) depended on the infectious proportion of the total population:
I/N.
There are two main ways to incorporate the infectious population into the force of infec-
tion λ(I). We generally assume that transmission is either ‘frequency dependent’ or ‘density
dependent’. In the former, which we have used throughout this paper, λ(I) is independent
of population size and depends on the fraction of individuals who are infected in a pop-
ulation so that λ(I) = βI/N. In the latter, λ increases proportionally with the number of
infected individuals I such that λ(I) = βI. Here, if the density of the infectious population
increases, i.e. there are more infectious individuals in the same area, the force of infection
also increases.
Frequency dependent transmission, also known as standard incidence, is typically used
for sexually transmitted pathogens in human population (Begon et al., 2002; Keeling
& Rohani, 2008). As the population size changes, the number of interactions between
susceptible and infectious individuals does not change. Hence, the relevancy for sexual
transmission, where the number of partners an individual has is likely independent of
the number of individuals near by, at least when the population is not very low. In con-
trast, density dependent transmission allows for reduced interaction between susceptible
and infectious individuals as the population size changes (Begon et al., 2002; Keeling
& Rohani, 2008). Consider a population that is declining due to disease-induced mortal-
ity from a disease with high mortality. As fewer individuals remain, it is more difficult for
individuals to come into contact, reducing the force of infection (Keeling & Rohani, 2008).
Density dependent transmission is typically used to model airborne and directly transmit-
ted infections such as measles or influenza. We remind the reader that these definitions
only differ when the total population size is variable.
Although the vast majority of the literature uses one of these two functions, a variety of
phenomenological non-linear formulations have also been introduced (Begon et al., 2002;
McCallum, Barlow, & Hone, 2001). When altering the transmission function it is essential
that careful consideration is given to the estimation of the transmission rate, β, as the units
of the term must be consistent within the equation. For example, a non-linear force of
infection such as the following
λ(I) = βI q
where β is the transmission rate, I is the number of infected individuals, and 0 < q < 1 is
a constant defining a power law relationship (Hochberg,
√ 1991; Liu, Levin, & Iwasa, 1986).
The exponent on I requires that β take the units ( individuals · time)−1 rather than time−1
q
Figure 3. Schematic representation of the age-structured SIR. SC , IC , and RC represent the total number
of susceptible, infected, and recovered children in a population, and SA , IA , and RA represent the same
classes but in the adult population. λC (IC , IA ) and λA (IC , IA ) are the forces of infection for the child and
adult classes, respectively, and γC and γA are their recovery rates. Dashed lines indicate transmission
between classes which contribute to the forces of infection, originating from infectious adults (red lines)
or infectious children (blue lines).
a continuous-age model with, for example, partial differential equations (Feng et al., 2016;
Martcheva, 2015), we examine an age-structured model with two discrete age classes so
that our population is sub-divided into children and adults. Such a model would be rel-
evant for infections that spread much faster among children than adults, whether due
to higher susceptibility or greater contact rates. We will use the subscript C to refer to
children populations and the subscript A to refer to adult populations. For example, the
number of susceptible children will be denoted as SC . We assume that the total number
of children is given by NC = SC + IC + RC and the total number of adults is given by
NA = SA + IA + RA . We also assume that children age into the adult compartment at a
rate f and that only adults reproduce and all births enter the SC class. Finally, only adults
are subject to natural mortality so that the total population size remains fixed. Now,
dSC
= μA NA − λC (IC , IA )SC − fSC ,
dt
dIC
= λC (IC , IA )SC − γC IC − fIC ,
dt
dRC
= γC IC − fRC ,
dt
dSA
= −λA (IC , IA )SA − μA SA + fSC ,
dt
dIA
= λA (IC , IA )SA − (γA + μA )IA + fIC ,
dt
dRA
= γA IA − μA RA + fRC .
dt
A schematic representation of this model is provided in Figure 3.
Notice that we distinguish between the force of infection to children λC (IC , IA ) and the
force of infection to adults λA (IC , IA ). In the former, the rate at which susceptible chil-
dren become infected will depend on the rate at which susceptible children contact both
infectious adults and children as well as the fraction of the adult and child populations
LETTERS IN BIOMATHEMATICS 209
that are infectious in their respective populations. Additionally, in contrast to the mod-
els presented earlier, this model contains multiple infectious classes. Therefore, the forces
of infection now depend on both infectious classes. We now assign two subscripts to the
transmission rate so that βij is the transmission rate from individuals in population j to
those in population i, where i and j can take on values of C or A. Notice that the proportion
of infectious individuals refers to the proportion of each class. Thus, the force of infection
for children and adults, respectively, are
IC IA
λC (IC , IA ) = βCC + βCA ,
NC NA
IC IA
λA (IC , IA ) = βAC + βAA .
NC NA
Figure 4. Schematic representation of the vector-borne disease model. SH and SV represent the total
number of susceptible humans and mosquitoes, respectively, and IH and IV represent the total number
of infected humans and mosquitoes, respectively. λH (IH , IV ) and λV (IH , IV ) are the forces of infection for
human and mosquito populations, respectively. Here, γ is the recovery rate for infections in humans and
this model assumes humans become susceptible to infection again after recovery. Dashed lines indicate
transmission between classes which contribute to the forces of infection: humans become infected only
through contact with an infected mosquito (red line), and mosquitoes only acquire infection through
biting a human (blue line). The only route for mosquitoes to leave the infectious class is by natural
mortality.
LETTERS IN BIOMATHEMATICS 211
Given how we defined the force of infection in previous sections, our first instinct may
be to define the forces of infection for human and vector populations as follows:
IV IH
λH (IH , IV ) = βHV , λV (IH , IV ) = βVH (17)
NV NH
such that the infection of humans is dependent upon the transmission rate βHV from vec-
tors to humans as well as fraction of the vector population that is infected; whereas the
infection of vectors is dependent upon the transmission rate βVH from humans to vec-
tors and the fraction of humans that are infected (NH = SH + IH is the total number of
humans). As in the previous models, the transmission rates βij include both the contact
rate and the probability of transmission given contact. Here, contacts that are capable of
generating infections occur through bites, bi , and we denote the probability of infection by
αij . In other words, we require that the transmission rate from j to i (βij ) to be equal to the
product of the contact rate bi and the infection probability αij so that
βij = bi αij .
Contact through bites has the symmetrical property such that every bite received by a
human is a bite taken by a vector, a concept referred to as ‘conservation of biting’. We
observed a similar idea in contacts between adults and children in Section 4.1.2. Conse-
quently, the number of bites depends on the population sizes of both the human and vector
populations. To demonstrate what this means, let us assume that bV is the biting rate of a
single mosquito and bH is the rate at which a single human gets bitten per unit time. Then
the total number of bites received by the human population and total number of bites taken
by the mosquitoes should equate as follows:
bH NH = bV NV . (18)
Solving for NV in the equation for conservation of biting (Equation (18)) and substi-
tuting in λH (IH , IV ) into our original force of infection equations (Equation (17)) leads
to
IV IH
λH (IH , IV ) = bV αHV , λV (IH , IV ) = bV αVH .
NH NH
As the same biting rate appears in both equations, typically the subscript is dropped. As
a note of caution, the canonical model of malaria (known as the Ross-Macdonald model
after Ronald Ross and George Macdonald who contributed to its formulation) often refers
to biting rate by the parameter a (Anderson et al., 1992; Macdonald, 1957; Ross, 1911).
Furthermore, although they use the notion of conservation of biting, the derivation is often
omitted. The appearance of the ratio of mosquitoes to humans, commonly denoted by m,
results from this assumption.
dS
= μN + θV + γ I − λ(I)S − (μ + φ)S,
dt
dV
= φS − λV (I)V − (μ + θ )V,
dt
dI
= λ(I)S + λV (I)V − (μ + γ )I,
dt
where N = S+V +I is the total population size, λ(I) = β(I/N) is the force of infection to
the susceptible class, λV (I) = σβ(I/N) is the force of infection to the susceptible class,
β is the transmission rate, μ is the natural birth/mortality rate, θ is the vaccine wan-
ing rate, σ is the ability of vaccinated individuals to produce new infections, γ is the
recovery rate, and φ is the vaccination rate. When the rate of waning of immunity from
LETTERS IN BIOMATHEMATICS 213
Figure 5. Schematic representation of the SIS model with an imperfect vaccine. Here, λ(I) = β(I/N)
and λV (I) = σβ(I/N) are the force of infection to the susceptible and vaccinated classes, respectively. β
is the transmission rate, θ is the rate of immunity loss, σ represents the ability of vaccinated individuals
to produce new infections, γ is the recovery rate, φ is the vaccination rate, and μ is the natural mortality
rate.
the vaccine exceeds that of recovery rate, it is possible to have a stable endemic equilib-
rium while R0 < 1. As found in Kribs-Zaleta and Velasco-Hernandez (2000), R0 (φ) =
(β/(μ + c))((μ + θ + σ φ)/(μ + θ + φ)). Varying R0 through the variation of the vac-
cination rate φ yields different dynamics dependent on the value of σ , i.e. the ability of
vaccinated individuals to produce new infections. Figure 6 shows an example of backward
Figure 6. Backward bifurcation in the presence of an imperfect vaccine. (A) The presence of a back-
ward bifurcation occurs with σ = 0.02. Solid line indicates a stable equilibrium while the dashed line
represents unstable. The grey shaded area indicates when two stable equilibria are simultaneously
present. (B) The backward bifurcation is not observed when when the vaccine is perfect σ = 0. Param-
eters are: μ = 0.01; θ = 0.01; γ = 0.5; β = 3.27. φ is varied between 0 and 100. Initial conditions
are: S(0) = 1 − η; I(0) = η; V(0) = 0 with η = 10−4 except near R0 = 1 where η = 10−10 is also
used to capture the stable DFE equilibrium. The model and parameters are motivated by Kribs-Zaleta
and Velasco-Hernandez (2000).
214 J. C. BLACKWOOD AND L. M. CHILDS
bifurcation for this system when σ = 0.02 which disappears when the vaccine is perfect,
i.e. σ = 0.
Backwards bifurcations have arisen in the mathematical analysis of a variety of epi-
demiological models (e.g. Arino, McCluskey, & van den Driessche, 2003; Dushoff, Huang,
& Castillo-Chavez, 1998; Hadeler & van den Driessche, 1997; Martcheva & Thieme, 2003;
van den Driessche & Watmough, 2000). However, often the constructions necessary to pro-
duce these dynamics require biologically tenuous assumptions such as in the model above
where the rate of reinfection from the vaccinated state is greater than the rate of returning
to the susceptible state. Another example is a model for Tuberculosis in which a back-
ward bifurcation will occur when infection with re-exposure is more likely than infection
with primary exposure (Feng, Castillo-Chavez, & Capurro, 2000; Martcheva, 2015). To our
knowledge, however, there exists no real-world experimental evidence of the occurrence
of backwards bifurcation in infectious disease systems.
5. Conclusions
This paper provides an introduction to modeling infectious disease dynamics with a
focus on several potential pitfalls. We attempt to show when these issues are likely to
arise and how to proceed with models and analyses despite them. This paper provides
an introduction to modeling infectious disease dynamics with a focus on several poten-
tial pitfalls. We attempt to show when these issues are likely to arise and how to proceed
with models and analyses despite them. This paper is intended to serve as a supple-
ment for the many excellent texts on infectious disease dynamics (referenced throughout)
and it provides a concise reference that compiles a range of common issues in disease
modeling.
Acknowledgments
The authors thank two anonymous reviewers for their helpful feedback throughout the review
process. This work was, in part, made possible by the following inspiration:
R0 is so great
It provides so much info
Except when it fails
– ‘An R0 Haiku’
LETTERS IN BIOMATHEMATICS 215
Disclosure statement
No potential conflict of interest was reported by the authors.
Funding
LMC is supported by Simons Foundation Collaboration Grant 443528.
ORCID
Lauren M. Childs http://orcid.org/0000-0003-3904-3895
References
Anderson, R. M., May, R. M., & Anderson, B. (1992). Infectious diseases of humans: Dynamics and
control. Vol. 28. Oxford: Wiley Online Library.
Arino, J., McCluskey, C. C., & van den Driessche, P. (2003). Global results for an epidemic model
with vaccination that exhibits backward bifurcation. SIAM Journal on Applied Mathematics, 64(1),
260–276.
Begon, M., Bennett, M., Bowers, R. G., French, N. P., Hazel, S. M., & Turner, J. (2002). A clarification
of transmission terms in host-microparasite models: Numbers, densities and areas. Epidemiology
and Infection, 129(1), 147–153.
Blackwood, J. C., Streicker, D., Altizer, S. A., & Rohani, P. (2013). Resolving the role of immunity,
pathogenesis, and immigration for rabies persistence in vampire bats. Proceedings of the National
Academy of Sciences of the United States of America, 110, 20837–20842.
Brauer, F. (2008). Compartmental models in epidemiology. In F. Brauer, P. van den Driessche, & J.
Wu (Eds.), Mathematical Epidemiology (pp. 19–79). Berlin: Springer.
Brauer, F., & Castillo-Chavez, C. (2001). Mathematical models in population biology and epidemiol-
ogy. Vol. 40. New York: Springer.
Brauer, F., & Kribs, C. (2015). Dynamical systems for biological modeling: An introduction. Boca
Raton: CRC Press.
Brauer, F., vanden Driessche, P. & Wu, J. (Eds.). (2008). Mathematical epidemiology. Lecture Notes
in Mathematics. Vol. 1945. Berlin: Springer.
de Camino-Beck, T., Lewis, M. A., & van den Driessche, P. (2009). A graph-theoretic method for the
basic reproduction number in continuous time epidemiological models. Journal of Mathematical
Biology, 59(4), 503–516.
Diekmann, O., & Heesterbeek, J. A. P. (2000). Mathematical epidemiology of infectious diseases: model
building, analysis and interpretation. Vol. 5. West Sussex: John Wiley and Sons.
Diekmann, O., Heesterbeek, J. A. P., & Johan, A. J. (1990). On the definition and the computation of
the basic reproduction ratio R0 in models for infectious diseases in heterogeneous populations.
Journal of Mathematical Biology, 28(4), 365–382.
Diekmann, O., Heesterbeek, J. A. P., & Roberts, M. G. (2009). The construction of next-generation
matrices for compartmental epidemic models. Journal of the Royal Society Interface, 7, 873–882.
Dushoff, J., Huang, W., & Castillo-Chavez, C. (1998). Backwards bifurcations and catastrophe in
simple models of fatal diseases. Journal of Mathematical Biology, 36(3), 227–248.
Feng, Z., Castillo-Chavez, C., & Capurro, A. F. (2000). A model for tuberculosis with exogenous
reinfection. Theoretical Population Biology, 57(3), 235–247.
Feng, Z., Zheng, Y., Hernandez-Ceron, N., Zhao, H., Glasser, J. W., & Hill, A. N. (2016). Mathemat-
ical models of Ebola? Consequences of underlying assumptions. Mathematical Biosciences, 277,
89–107.
Hadeler, K. P., & van den Driessche, P. (1997). Backward bifurcation in epidemic control. Mathemat-
ical Biosciences, 146(1), 15–35.
Heffernan, J. M., Smith, R. J., & Wahl, L. M. (2005). Perspectives on the basic reproductive ratio.
Journal of the Royal Society Interface, 2(4), 281–293.
216 J. C. BLACKWOOD AND L. M. CHILDS
Hethcote, H. W. (2000). The mathematics of infectious diseases. SIAM Review, 42(4), 599–653.
Hochberg, M. E. (1991). Non-linear transmission rates and the dynamics of infectious disease.
Journal of Theoretical Biology, 153(3), 301–321.
Keeling, M. J., & Rohani, P. (2008). Modeling infectious diseases in humans and animals. Princeton:
Princeton University Press.
Kermack, W. O., & McKendrick, A. G. (1927). A contribution to the mathematical theory of
epidemics. Proceedings of the Royal Society of London A, 115, 700–721.
Kribs-Zaleta, C. M., & Velasco-Hernandez, J. X. (2000). A simple vaccination model with multiple
endemic states. Mathematical Biosciences, 164(2), 183–201.
Ledder, G. (2017). Scaling for Dynamical Systems in Biology. Bulletin of Mathematical Biology,
79(11), 2747–2772.
Li, J., Blakeley, D., & Smith, R. J. (2011). The failure of R0 . Computational and Mathematical Methods
in Medicine, 2011, 1–17.
Liu, W. M., Levin, S. A., & Iwasa, Y. (1986). Influence of nonlinear incidence rates upon the behavior
of sirs epidemiological models. Journal of Mathematical Biology, 23(2), 187–204.
Macdonald, G. (1957). The epidemiology and control of malaria. London: Oxford University Press.
Martcheva, M. (2015). Introduction to mathematical epidemiology. Vol. 61. New York: Springer.
Martcheva, M., & Thieme, H. R. (2003). Progression age enhanced backward bifurcation in an
epidemic model with super-infection. Journal of Mathematical Biology, 46(5), 385–424.
McCallum, H., Barlow, N., & Hone, J. (2001). How should pathogen transmission be modelled?.
Trends in Ecology and Evolution, 16(6), 295–300.
Ross, R. (1911). The prevention of malaria. London: John Murray.
Shrestha, S., Foxman, B., Berus, J., van Panhuis, W. G., Weinberger, D. M., Steiner, C., Viboud, C., &
Rohani, P. (2015). The role of influenza in the epidemiology of pneumonia. Scientific Reports, 5,
15314.
Shrestha, S., Foxman, B., Weinberger, D. M., Steiner, C., Viboud, C., & Rohani, P. (2013). Identifying
the interaction between influenza and pneumococcal pneumonia using incidence data. Science
Translational Medicine, 5, 191ra84.
Strogatz, S. H. (2014). Nonlinear dynamics and chaos: With applications to physics, biology, chemistry,
and engineering. Boca Raton: Westview press.
van den Driessche, P., & Watmough, J. (2000). A simple sis epidemic model with a backward
bifurcation. Journal of Mathematical Biology, 40(6), 525–540.
van den Driessche, P., & Watmough, J. (2002). Reproduction numbers and sub-threshold endemic
equilibria for compartmental models of disease transmission. Mathematical Biosciences, 180(1),
29–48.
Vynnycky, E., & White, R. (2010). An introduction to infectious disease modelling. Oxford: Oxford
University Press.
Wang, W., & Zhao, X-Q. (2008). Threshold dynamics for compartmental epidemic models in
periodic environments. Journal of Dynamics and Differential Equations, 20(3), 699–717.
Wesley, C. L., & Allen, L. J. S. (2009). The basic reproduction number in epidemic models with
periodic demographics. Journal of Biological Dynamics, 3(2–3), 116–129.
Appendices
Appendix 1. Relating constant rates and duration
The common usage of a constant rate of transition between compartments is frequently made for
mathematical simplicity. An underlying assumption for the choice of a constant rate is that average
duration spent within a compartment is exactly the reciprocal of the rate of exit from that compart-
ment. Consider a simplified equation for the infectious (I) compartment, such that only recovery
occurs and there are no new infections:
dI
= γ I. (A1)
dt
LETTERS IN BIOMATHEMATICS 217
The exit rate from the I compartment, γ , can be found by measuring the average time spent
infectious, d. Specifically, it turns out that γ = 1/d.
To understand the origin of this result, solve Equation (A.1) with an initial condition of I(0) = I0
to obtain the number of infectious individuals at time t
I(t) = I0 exp−γ t . (A2)
Dividing both sides by I0 allows us to rewrite this equation in terms of I(t)/I0 , i.e. the fraction of the
originally infectious individuals who are still infectious at time t. This is equivalent to the probability
of remaining infectious at time t.
The complement of this quantity, which we define as F(t; γ ) = 1 − exp(−γ t), is then the prob-
ability of having left the infectious class by time t. Given that the initial time is assumed to be at
t = 0 (that is, there are no infections prior to the initial time), F(t; γ ) is exactly the cumulative dis-
tribution function for a probability distribution known as the exponential distribution where γ is
the rate parameter. Any standard statistical text will demonstrate that the mean of the exponential
distribution with parameter γ is exactly 1/γ , i.e. the average time an individual is infectious.
and R0 no longer applies in the same way. First, recall the SIR model without demography:
dS I dI I dR
= −β S, = β S − γ I, = γ I.
dt N dt N dt
In this model, there is only a DFE. In fact, there are infinitely many DFEs because as long as I = 0
then the value of S and R does not matter.
Suppose we try to find R0 using, for example, the next generation matrix. In this case, the F matrix
is simply a scalar, β(1/N)S and the V matrix is also a scalar given by γ (note that they are scalars
because there is only one infectious class, I). Evaluated at the DFE,
β
FV−1 =
γ
so that R0 = β/γ . The magnitude of R0 relative to one determines whether the disease initially grows
(R0 > 1) or declines (R0 < 1). In contrast to the SIR model with demography, you always return to
a disease free state.
Although R0 does not directly relate to the global asymptotic stability of the DFE, it does pro-
vide information on the shape of the epidemic curve. Namely, if R0 < 1 then I is always decreasing.
Consequently, there is no epidemic. If R0 > 1, then I increases until S is equal to γ N/β. At that
point, I decreases back towards zero. In this case, the dynamics follow a standard epidemic curve –
increasing cases followed by a peak and eventually extinction of the outbreak. Therefore, although
R0 does not directly impact the global stability of the DFE it does play a substantial role in the short
term epidemic dynamics.
Note that matrices will be indicated by bold font and vectors will have an arrow above them. We can
now simply write Equation (A.5) as
dn
= Jn
dt
where J is called the Jacobian matrix and here it is evaluated at the equilibrium. In the case of the
SIR model the Jacobian is
⎡ ⎤
I∗ S∗
⎢ −β − μ −β ⎥
⎣ N∗ N ⎦
I S∗
β β − (μ + γ )
N N
evaluated at the DFE (S∗ , I ∗ ) = (0, 0) makes the linearized system
⎡ dS ⎤
⎢ dt ⎥ −μ −β S
⎣ ⎦= 0 β − (μ + γ ) I
dI
dt
Note that the Jacobian can be extended to higher dimensions.
As our goal is to determine the long term behaviour of the perturbation from the equilibrium,
we need to understand how n will change. From calculus we know that in a one dimensional setting
dn(t)
= αn(t) =⇒ n(t) = n0 eα
dt
The dynamics of this equation can then be determined by the sign of α. If it is positive, the per-
turbation grows with time. In contrast if it is negative, the perturbation shrinks with time. Now, if
α represents the Jacobian evaluated around an equilibrium, these dynamics tell us about the stabil-
ity of the equilibrium, i.e. when α < 0 the equilibrium is stable and when α > 0 the equilibrium is
unstable.
220 J. C. BLACKWOOD AND L. M. CHILDS
A similar idea applies to higher order systems. Let’s return to our two dimensional system
dn
= Jn. (A6)
dt
It turns out it is possible to analyze equilibria in a similar manner and a solution of Equation (A.6)
exists and is of the form
n = eλt n0 (A7)
where n0 is a non-zero vector and λ is a constant. Differentiating Equation (A.7) leaves us with
dn
= λ eλt n0 .
dt
Substituting this into Equation (A.6) gives us that
λ eλt n0 = Jn
but from Equation (A.7), this means that
λ eλt n0 = J eλt .n0
Here, eλt is a non-zero constant so we can divide it out and we are left with
λn0 = Jn0 (A8)
Here, n0 is called an eigenvector and λ is an eigenvalue of the matrix J. Details and further expla-
nation on eigenvalues and eigenvectors can be found in any standard linear algebra or ordinary
differential equation textbook.
Keep in mind that our end goal is to determine how the dynamics change, which given
Equation (A.6) depend on λ. To find λ, we subtract λn0 from both sides of Equation (A.8), leaving
us with and we are left with
0 = Jn0 − λIn0
= (J − λI) n0
Notice that we cannot simply pull out n0 from the above equation because J and λ have different
dimensions. Therefore, we needed to introduce a 2x2 identity matrix I.
To have a non-zero solution of this equation the determinant of (J − λI) must equal zero. We
further you to a standard ordinary differential equation text to find the determinant.
Thus, a general solution the two-dimensional system
dn
= Jn
dt
is given by
n = c1 eλ1 t n 1 + c2 eλ2 t n 2
where n 1 and n 2 are the eigenvectors corresponding to the eigenvalues λ1 and λ2 , respectively, and
c1 and c2 are arbitrary constants.
In the case of the SIR model we find that the two eigenvalues of J are λ1 = −μ and λ2 = β −
(μ + γ ). As our parameters are all positive to ensure biological realism, λ1 < 0 always. However, λ2
may be positive or negative dependent upon the relative sizes of β and μ + γ . The DFE equilibrium
will be linearly stable when
β − (μ + γ ) < 0.
We can rearrange this quantity to find an equivalent comparison, i.e.
β
<1
(μ + γ )
Notice, this is in fact our definition of R0 . In fact, it can be shown that when R0 < 1 then our eigen-
values are negative and the DFE is linearly stable and when R0 > 1 then at least one eigenvalue is
positive and the DFE is unstable.
LETTERS IN BIOMATHEMATICS 221
We can now move on to determine what the values of λ mean in terms of linear stability of the
equilibria. Here, we will list the different cases; there are two main cases (each with three subcases
each):
To summarize, an equilibrium is linearly stable if the real part of all eigenvalues is less than zero,
i.e. Re(λi ) < 0 for all i. Otherwise we say the equilibrium is unstable.