Jovrulor

J. Math. Biol. (1995) 33:661-675

mathematical
6ioio9y
(C) Springer-Verlag 1995

Susceptible-infected-removed epidemic models with

dynamic partnerships
M. Altmann
Division of Health Computer Sciences, University of Minnesota, Minneapolis,
MN 55455, USA
Electronic mail: [email protected]
Received 18 March 1994; received in revised form 22 August 1994

Abstract. The author extends the classical, stochastic, Susceptible-Infected-

Removed (SIR) epidemic model to allow for disease transmission through
a dynamic network of partnerships. A new method of analysis allows for
a fairly complete understanding of the dynamics of the system for small and
large time. The key insight is to analyze the model by tracking the configura-
tions of all possible dyads, rather than individuals. For large populations, the
initial dynamics are approximated by a branching process whose threshold
for growth determines the epidemic threshold, Ro, and whose growth rate, A,
determines the rate at which the number of cases increases. The fraction of the
population that is ever infected, Q, is shown to bear the same relationship to
Ro as in models without partnerships. Explicit formulas for these three
fundamental quantities are obtained for the simplest version of the model, in
which the population is treated as homogeneous, and all transitions are
Markov. The formulas allow a modeler to determine the error introduced by
the usual assumption of instantaneous contacts for any particular set of
biological and sociological parameters. The model and the formulas are then
generalized to allow for non-Markov partnership dynamics, non-uniform
contact rates within partnerships, and variable infectivity. The model and the
method of analysis could also be further generalized to allow for demographic
effects, recurrent susceptibility and heterogeneous populations, using the same
strategies that have been developed for models without partnerships.

Key words: Epidemiological models - S-I-R - Threshold - Sexual partners -

Stochastic processes

1 Introduction

One of the major limitations of most mathematical models of infectious

diseases is that they treat contacts as unrelated events, disregarding their
662 M. Altmann

organization into partnerships. However, for most diseases - be they trans-

mitted by skin contact, coughing, sexual intercourse, sharing drug injec-
tion equipment, or breast feeding - contacts occur within social ties
that last for many contacts. Only in rare circumstances, such as cholera
transmission through a contaminated water source, or HIV transmis-
sion through transfusion and anonymous sex, is there compelling bio-
logical evidence that these social ties, or "partnerships", are so weak that
they could be omitted from a model. Nonetheless, partnerships might
be omitted if the epidemiological impact were negligible. To measure
this impact, one must compare models that include partnerships and to
pure contact models.
Three types of models have incorporated partnerships, with limited
success. Models based on percolation theory [14] require a fixed structure
of partnerships arranged in a very regular fashion, typically a square lattice.
Analytic results are generally unavailable except in special cases. Secondly,
models that allow for pair formation [6, 7] can be analyzed with standard
methods, but the models require that partnerships form isolated units.
Because these models do not allow for several concurrent partnerships,
they are most applicable to serially monogamous sexual relationships.
Thirdly, several research groups have used computer simulations of discrete
event models [15, 11, 4, 13]. This type of model easily allows for multiple
partnerships of varying duration and arbitrary rules governing partner
selection and partnership dynamics, but is not very amenable to mathematical
analysis.
This paper develops a continuous time stochastic SIR model for a
homogeneous population that allows for concurrent partnerships. The
model is initially presented in its simplest form, with Markov transitions,
and is then generalized. At one extreme, the model reduces to the classical
SIR model with independent contacts. At the other extreme, it represents
the spread of disease through a fixed network of partnerships. There is no
restriction on the number of concurrent partners for an individual; on
the contrary, partnerships are independent stochastic processes. The model
is conceptually similar to the ones presented in [i1] and [161, but the
analysis is completely new. The key idea is that one should consider
the population of dyads, rather than individuals. This goes one step
beyond the approach taken by Dietz and colleagues in [-6, 7] (which strongly
resembles the methods used in modeling chemical kinetics). They track dyads
only when the individuals are actively partners, whereas the analysis given
below tracks the state of every dyad.
We begin with a review of the classical model without partnerships.
Thereafter, a Markov model for partnerships is developed, and its threshold
behavior is analyzed and compared with the traditional model. Finally, the
model is generalized, and the generalized model is used to compute the
fraction of the population ever infected and the initial growth rate in infectious
individuals.
S-I-R models with dynamic partnerships 663

2 The pure contact model: a review

A number of authors (see [3] for a good presentation) have considered some
version of the following model for the spread of an infectious disease in a fixed
population. The population contains N individuals, each of whom is in one of
three states: susceptible to disease, infected and infectious, or removed (infec-
ted, but no longer infectious). The numbers of individuals in these three states
at time t is denote S(t), I(t), and R(t). Mortality, fertility, migration and other
demographic forces are neglected. Effective contacts arise between each pair of
individuals according to a Poisson process with rate c/N. Transmission
occurs whenever effective contact occurs between a susceptible and an infec-
tious individual. The susceptible individual then becomes infectious and
remains infectious for an amount of time that is exponentially distributed with
rate v. The contact processes and the infectious durations are all mutually
independent.
Although this is a description at the level of individuals, many authors
jump immediately to the population level and say that the rate at which new
cases arise is cS(t)I(t)/N. Other authors factor c into two terms; one is the
overall contact rate and the other is the probability that a contact is effective.
Not all contacts are effective because of the protective effect of condoms, the
probability that air borne droplets are inhaled, etc. From an epidemiological
point of view, only effective contacts matter; from a sociological point of view,
it may be helpful to measure the total contact rate and then determine the
fraction of these that are effective.
The analysis of this model has focused on three measures of the force of the
epidemic: the reproductive number, the fraction of the population ever infec-
ted, and the initial rate of growth in infectious individuals. In the limit when
N becomes large, the asymptotic values of these measures can be obtained by
considering either the stochastic linearization or the deterministic average. We
will say that a major epidemic occurs when the number of cases that ever
arises is of the same order of magnitude as the population - is O(N).
Otherwise, the number of cases is small compared to the population - is
o(N) - and we will say that a minor epidemic has occurred.
Let I(0) denote the initial number of infectious individuals. If I(0) is
kept fixed (rather than scaling with N), the number of infectious indi-
viduals initially grows like a continuous time branching process with each
infectious individual creating new infectious individuals at a rate c and
being removed at a rate v. To draw the parallel with the demographic
use of branching processes, c is the "birth" rate and v is the "death" rate.
The branching process is a good approximation while the total number of
cases is o(N). If c/v < 1, the probability of a major epidemic is
zero and the distribution of cases is the same as the distribution of the total
size of the branching process. If c/v > 1, the probability of a major epi-
demic is the probability that the branching process will grow to infinity
and the distribution of sizes of minor epidemics is again the same as for
the branching process. The quantity ely is denoted Ro, and is the average
664 M. Altmann

number of secondary cases arising from an isolated initial case. If individuals

are marked according to their "generation" in the contagious process (the
number of links in the path of transmission from the initial case t o the
individual), then the size of successive generations follows a discrete genera-
tion branching process, and R0 is the average number of new cases directly
produced by a member of any generation.
Regardless of the value of c/v, the average value of I(t) initially grows (or
shrinks) exponentially with rate

A = c - v = (Ro - 1)v .

The fraction of the population ever infected is given by R(oo). The

expected value of R( oo ), conditional on a major epidemic occurring, solves
the transcendental equation

x = 1 - e -R°x

This paper extends these results to a model in which contacts occur only
within partnerships.

3 The Markov model with partnerships

The Markov version of the model makes the following assumptions.

(1) Removal from the infectious state occurs at a constant rate.
(2) Unpaired individuals begin a partnership at a constant rate, partner-
ships dissolve at a constant rate, and partnerships behave independently.
Therefore the intervals during which two individuals are unpaired and paired
are independent, exponentially distributed random variables.
(3) Partnerships always begin with a contact, which may or may not be
effective, and thereafter effective contacts occur at a constant rate. Therefore,
the contacts other than the initiating contact may be modeled as Poisson
processes.
(4) An effective contact between an infectious individual and a susceptible
individual results in transmission.
We will also need to keep track of all of the effective contacts that an
infectious individuals has, regardless of the disease status of the other indi-
vidual. These will be called potential transmissions.
Let the population size be denoted as N + 1. We use N + 1 rather than
N so that each individual has N potential partners.
Let p/N be the rate at which a partnership is formed and a be the rate at
which it dissolves, p is scaled by N to keep the total partner acquisition rate
roughly constant as the population size is changed.
Let fl denote the probability that a partnership begins with an effective
contact.
Let effective contacts within a partnership arise at a constant rate 7.
Let v be the removal rate.
S-I-R models with dynamic partnerships 665

Mathematical remarks: This description is sufficient to define a continu-

ous time jump stochastic process with finite state space. The partnerships are
an ergodic stochastic process. We will assume that it has equilibrated before
the disease is introduced. The number of partners for an individual at any
given time is binomially distributed with parameters N and p/(p + N~r). Thus
the mean is Np/(p + N~). As N becomes large the number of partnerships for
an individual converges in distribution to a Poisson random variable with
mean p/cr.

3.1 Secondary cases production

To understand the initial dynamics of the model, we look at what happens

when a few cases are introduced into a large population of susceptible
individuals. So long as these initial cases are distributed randomly in the
population, it suffices to consider what happens when a single initial case is
introduced. The independence of the partnership processes implies that the
secondary cases arising from a particular infectious individual are unlikely to
be partners of each other, and thus the number of infected individuals will
initially grow like a branching process, just as for the classical model. The
threshold for a major epidemic is when the average number of secondary cases
produced by an isolated initial case, R0, is greater than 1. The remainder of
this section is devoted to the calculation of Ro. In the next section we
investigate the temporal distribution of the secondary cases and show how
this can be used to establish the initial growth rate, A.
Let us denote the initial case by individual i. The number of secondary
cases is the sum over the N other individuals of the probability that i transmits
to the other individual. The individuals are identical, so it suffices to compute
the probability that i transmits to some individual j. The key to computing
this probability, and more generally to understanding the dynamics of the
epidemic, is to consider the configuration of the dyad (i,j). Since each indi-
vidual can be in one of three possible disease states and the individuals may or
may not be partners at a given time, there are 18 possible configurations for
the dyad. These configurations will be denoted by two letters, indicating the
disease status of the two individuals in the dyad. A centered dot will be used to
indicated that i a n d j are unpaired, a dash to indicate that they are paired. For
example, let S-S denote the configuration where both are susceptible and they
are partners, I • S denote the configuration where i is infected and j is
susceptible and they are not partners, etc. Each dyad will move among the
possible configurations according to a Markov process. To compute Ro we
need only consider the life history of a dyad from the first infection to the
removal of that case or the infection of the other member of the dyad. The
states and the corresponding transitions for this reduced Markov process are
shown in Fig. 1. The transition rates only account for the force of infection
due to the case in the dyad, ignoring the possibility that some other secondary
case that arises early in the infectious period will infect j. The probability of
666 M. Altmann

7
tip~N/
3'
Fig. 1. Transitions rates among the states for
a dyad of individuals, from first infection to
-~ transmission or removal. External sources of
infection are ignored. 1. S refers to the state
1 O" where the first individual is infectious, the
second is susceptible and they are not currently
V partners, R - S refers to the state where the
first individual is removed, the second is
I
susceptible and they are currently partners, etc.
_~ The dyad begins with one infected and one

F susceptible individual, either in state I. S or

state I - S

such external infection is negligible for large populations with independent

partnership processes, which is why the branching process a p p r o x i m a t i o n is
valid. This reduced d i a g r a m has three absorbing states: I - I, R . S, and R - S.
Because the partnership process is assumed to have equilibrated before the
disease was introduced, the dyad starts in I - S or I . S with probability

Pl - s (0) = P
p + No"
and
No"
P~.s(0) = N-------~"
p +

If transmission occurs, the pair will move to I - I. If removal occurs before

transmission, the dyad will m o v e to either R . S or R - S, depending on whether
f a n d j are partners at the time of removal. To compute the probability of trans-
mission from i to j we simply need to calculate the probability of absorption into
I - I as opposed to R . S or R - S. This calculation can be done easily using
standard methods for M a r k o v processes. Let A b s denote the probability of
absorption into state I - I , starting from state I . S and let A~_ s denote the
probability of absorption into state I - I, starting from state I - S.
A b s and A I _ s satisfy
(p/N + v)Ai.s = flp/N + (1 -- fl)p/NAI-s
(v + o" + 7 ) A I - s = 7 + o"Ai.s •
Solving this system of equations gives

1 flp(v + o") + P7
AI.s = ~ D

flpo"/N + 7 (p/U + v)
AI-S =
D
S-I-R models with dynamic partnerships 667

where
D = ( p i n + v)(v + a + ~ , ) - (1 - fl)pa/g.
Then the probability of transmission from i to j is given by
NaAi.s + p A l - s
P= Na+p

aflp(v + a) + pTa + flp2a/N + 7p(p/N + v)

(Na + p)D
The expected number of new cases arising from a single initial case is given
by
Naflp(v + a) + NpTa + fipZa + ,/p(p/N + vN)
Ro = N v
(Na + p)D
For large N this converges to

lira Ro = P7 tip p~,(1 - fl) (1)

N-,~ a(v + a + ~) + - + (v + a + 7)v "
This formula may be obtained in another way by arguing that for large N,
the probability that more than one partnership episode between i a n d j occurs
during the infectious period is O(N-2) and can therefore be ignored. The first
term in (1) arises if i and j are paired at the start of the period. The second term
accounts for the probability that i and j become partners and transmission
occurs on the initial contact, and the third term accounts for the probability
that they become partners, transmission does not occur on the initial contact,
but effective contact occurs before either the partnership or the infectious
period ends.
The expression for Ro simplifies under several special limiting cases.

Case 1: Short partnerships

Consider the limit where a goes to infinity. Then (1) reduces to

lim lim Ro = - - ,
o'~ N ~ V

which is the standard formula for a model where the transmission occurs only
through isolated contacts.

Case 2: Long partnerships

We consider the limit where a goes to zero. To keep the average number of
current partnerships for an individual constant, we take p to zero at the same
rate. Replace p and a by ep and ~a. Then the limiting value of Ro is given by

P 7
lim lim Ro =
~ 0 N-~ o (v + 7) "
The limit for long partnerships is the same as percolation through a ran-
dom graph, which has already been studied [10, 14]. The graph is formed by
668 M. Altmann

including each possible partnership with probability p/(p + Na). Then the
probability that partnership results in transmission is simply the probability
that the effective contact that would result in transmission occurs before
removal, which is 7/(v + 7). The average number of secondary cases produced
by the initial case is therefore

Np 7
(p + Na) (v + 7)"

Even if v goes to zero, Ro stays bounded by p/a because the growth of the
epidemic is limited by the fan-out of the partnership graph.

Case 3: Rapidly fluctuating partnerships

Consider the limit where a and p go to infinity at the same rate. For the limit
to be finite we must take fl = 0. Replacing p and a by ep and ea and letting e go
to infinity gives

lim lim Ro = P__Z

Case 4: Transmission on a per partnership basis

There is a suggestion that for some diseases, such as HIV [16], the variability
in transmission is due to characteristics of the individuals. Thus, for pairs of
histo-compatible individuals, transmission occurs as soon as one member of
the partnership becomes infected. Non-histo-compatible pairs show no trans-
mission, regardless of the partnership duration. This corresponds to letting
7 become large (so that transmission occurs as soon as either is infected) and
interpreting p as the mean rate at which individuals acquire histo-compatible
partners.
lim lim Ro = p + P .
7~Qo N'--) ¢o (9" V

This is in agreement with the result that Watts and May obtained [16].
Our Markov model is essentially the same as their model when the latent
period is zero. In that case, they show that Ro is the sum of the average
number of partners an individual has at any one time (p/a in our notation, Po
in theirs) and the average number of new partners acquired during the
infectious period (p/v in our notation, Rb in theirs).
Remark: It is easy to check that the same results are obtained if the limits
on partnership dynamics and population size are interchanged.
Let us compare (1) with the estimate obtained by a pure contact model.
The average rate at which effective contacts are being made between any i
and j is

C= +
NIp + II. NIp + II.
S-I-R models with dynamic partnerships 669

and so the usual estimate of Ro would be

Ro = --Nc = N flpa + N~p

v vaN+vp vNa+vp"
For large N this converges to

lim [~o = tip + 7P.

N-+m V VO"

Therefore, the ratio of the Ro from the pure contact model to the true
Ro is

lim --R° = fl + '--?

, (v +or- ~,T']' (2)
N-~Ro fl + Z k v-+-a--+ T

This shows that the growth rate in infectious individuals is always less than
would have been predicted by a model with the same average contact rate in
which the contacts are made with independently selected individuals. Below,
we show that the final epidemic size is similarly reduced.

3.2 Initial growth

The last section examined R0, which measures the increase in the number of
cases in successive "generations" of infections. Epidemiologists are often
more concerned with the rate of growth as measured against real time. This
temporal growth rate can be calculated by tracing the configuration of
dyads (i,j). Roughly speaking, the persistence of partnerships leads us
to expect that transmissions will tend to occur earlier in the infectious
period because later some effective contacts may be "wasted" on partners
who are already infected. There are N dyads involving i, so the expected
number of secondary cases produced after i has been infected for s units
of time is N times the probability of transmission from i to j in s units of
time. Referring to Fig. 1, this probability is the probability of being in ! - I
by time s. By the usual methods for Markov processes, the probabilities of
being in the various configurations satisfy a system of differential equations.
In particular,

dPI~I (s) = -7"7

itiP
v PI.S(S) + YPl-s(S) •
Define #(s) by
#(s) = lim NPi_l(s),
N~oo

which is the expected number of new secondary cases produced in t amount of

time after infection, in a very large population.
670 M. Altmann

Solving the linear system of differential equations for the P's shows that
the
#(s)=I(1-fl)PT+flPll(1-e-~S)v+a

+ Ip , fl)p'/_](--]'1 e -('+~+~)s )
(1 ~2~--
7+a+v"
A little algebra confirms that
p~ ~p pr(1 - 9)
lira/~(s) = a(v + a + 7) + --v + (v + a + 7)v '
S--+ ¢X~

which agrees with (1).

The function/~(s) is also called the reproduction function and can be used
to construct the continuous time general branching process that approximates
the initial dynamics of the infective population [9, pp. 123-126]. Infection is
treated as the "birth" of an infectious individual and removal as "death".
According to [9, pp. 131-132], as long as R0 is finite (as must be the case for
any biologically reasonable model), the number of cases that have arisen up to
time t and the number of infective individuals at time t grow exponentially
with a rate A that is determined by the equation

f o e - Asd#(s) = 1 I (3)

From the formula for p(s), it follows that A satisfies

P~ + tip pT(1 - f l ) = 1. (4)

a(v+A+a+~) ~ +(v+A+a+~,)(v+A)
This shows that A is the amount by which the removal rate would have to be
increased to bring the reproductive number to 1 and thus prevent an epidemic.
This is an example of the more general principle that the exponential growth
rate of a demographic system is the harvesting rate that would bring the
population to zero growth. Subtracting (4) from (1) gives

Ro - 1 = ~ A p +
v+A\vJ\7+a/ 7+a+v+AkaJkT+a/
Because this is a quadratic equation for A, the expression for A in terms of R0
and the model parameters is much messier than the comparable formula for
a pure contact model, namely
A = (Ro - 1)v .

4 The generalized model

So far, the model has assumed that partnerships have an exponential dura-
tion, and that effective contacts during a partnership occur according to
S-I-R models with dynamic partnerships 671

a Poisson process. Also, the infectious period was assumed to have an

exponential distribution and any effective contact that occurs during the
infectious period results in transmission if the other individual is susceptible.
These assumptions on the pattern of contacts and the infectivity can be
generalized using the mathematical machinery of point processes. The gener-
alized model has two ingredients.
First, the pattern of effective contacts is modeled by an arbitrary point
process on the real line [5]. Each dyad of individuals i and j will have an
associated, independent realization of the point process. For example, when
fl = 0 in our Markov model, the effective contacts form a Poisson process with
a random rate function that is one when a dyad is paired and zero when it is
not. This is called a doubly stochastic process, or Cox process.
In order to make the mathematics tractable, the only restriction that needs
to be imposed on this process is time invariance. This means that partnerships
are statistically equivalent, no matter when they start. Actually, once the point
process has been defined, the notion of partnerships starting and stopping
becomes secondary. As an example, in a model of sexual contacts, one might
use a process indicating that effective contacts are rare during an initial dating
phase, very common immediately after a commitment is made, and then
become less frequent over time, with complete cessation if the partnership
dissolves or perhaps during pregnancy.
Second, the course of infection is described by a random infectivity
function h(s), which gives the probability that an effective contact between
a susceptible individual and an infectious individual s units of time after
infection will result in transmission. For example, in the Markov model, h(t)
steps from 1 to 0 at a random time (the removal time) whose distribution is
exponentially distributed with rate v. This formulation is broad enough to
handle a wide variety of disease progression models for which infectivity
varies both within and between clinical stages. In this formulation, individuals
need not be explicitly removed; their infectivity simply becomes zero.
From the mathematical point of view, h thins the point process that
describe the effective contacts, to produce a point process that describes the
potential transmissions. Any particular h determines a function Oh(s), defined
as the probability that there is at least one point from the thinned process in
an interval (t, t + s). If an individual is infected at time t, then conditional on h,
the number of potential transmissions that have been made by time t + s is
binomially distributed with parameters N and Gh(S). As N becomes large, it is
reasonable to expect the contact processes to become sparser because in
a large population the probability that any two particular individuals have
contact in a finite time interval is small. Mathematically, the point processes
should scale with N in such a way that we can define a function Gh(S) by

Gh(s) = lim NGh(S).

N~zo

Therefore, conditional on h, the number of potential transmissions a case

makes by time t + s has a Poisson distribution with mean Gh(S). Let Eh denote
672 M. Altmann

the expected value of a random variable obtained by averaging over the

distribution ofh. The average probability that a case makes exactly k potential
transmissions can be written as

9k ~---lim Ehle-Gn(s)G~(S)']
s~® k! J"
The average number of potential transmissions by time t + s is
It (S) = Eh Gh (S),
and the probability of potential transmission from i to j can be written as

p = lira Eh[Gh(S)].

The average number of secondary cases produced by an initial case is

Ro = ~ kgk = lim #(s). (5)
8--.*00

4.1 Final size: minor epidemics

Recall that minor epidemics are those for which the number of cases ever
produced is small compared to the total population. In a minor epidemic, the
branching process may be used as an approximation during the entire epi-
demic. The 9k's describe the "offspring" distribution for the branching process
approximation to the initial spread of the disease. Let F denote the generating
function of the offspring distribution, and let H denote the generating function
of the distribution of the final size of the branching process. Standard results
for branching processes (see [8], for example) imply that the probability that
an major epidemic occurs is given by the smallest positive fixed point of the
offspring generating function, i.e.
q = F(q) = ~ gkq k.
k

The generating function for the conditional distribution of the size of the
minor outbreaks satisfies
qH(x) = F(xqH(x)) .

4.2 Final size: major epidemics

The branching process approximation can also be used to obtain the distribu-
tion of the fraction ever infected in major epidemics. Lefevre and Picard [12]
provide a formula for the generating function of the final size that can be
expressed in terms of the gk'S. Their machinery was developed for pure contact
models, but applies equally well to any model, including our generalized
S-I-R models with dynamic partnerships 673

model, for which the completed epidemic can be realized as follows. Each
individual chooses a random integer m~. These are independent and
P [ m = k] = Ok. Then each individual, i, plans potential transmission with
a random subset of size m~ drawn from the N other members of the popula-
tion. These selections are independent and all subsets have equal probability.
An individualj is marked as a case if there is a chain of potential transmissions
from an initial case to j. Although their formula for the generating function is
cumbersome, Lef~vre and Picard show that when N tends to infinity, the final
epidemic intensity converges to the limit of the deterministic model, and
therefore satisfies the equation

1 -- ~ = e -RoQ

4.3. Initial growth

The computation of the initial growth rate for the general model is exactly as
for the Markov model, once p(s) has been determined. The continuous time
general branching process is constructed by assigning each individual a ran-
dom function h and then a birth process with cumultive birth rate Gh.
Averaging over h, the cumulative birth rate is # and again the initial growth
rate, A, is the solution of (3).

5 Discussion

Our assumption of independent partnerships should be contrasted with the

pair formation model. In its basic formulation, that model restricts an indi-
vidual to have at most one partner at a given time [6]. The model was
extended to allow for two concurrent partners [7], but only when the indi-
viduals have no partnerships outside the threesome. Pair formation models
therefore serve to define the other end of the spectrum, with the real world
probably lying somewhere in between, and so the two types of models can be
used to provide bounds on the behavior of real systems. Our partnership
model may also be used to validate complex simulation models and to
measure the relative importance of their additional degree of realism. If the
dynamic partnership model disagrees with a simulation model, that indicates
that either there is an error in the simulation software or that the simulation
model has significant features that cannot be captured by a partnership
model.
The reproductive number, the growth rate of the infectious population,
and the final epidemic size are the three most fundamental concepts in the
mathematical theory of infectious disease modeling. These quantities have
importance for the practicing epidemiologist because they determine the
minimal vaccination coverage to prevent disease spread, the rate at which it
will spread and the total impact of the epidemic. In many circumstances only
674 M. Altmann

one or two of these quantities can be measured, so it is important to be able to

make inference from one to another. We have shown how to compute the
three quantities for a general model that includes partnership formation and
dissolution. The results are pleasing because they show that the relationship
between the reproductive number and final size is the same as for a pure
contact model. A quadratic equation for the initial growth rate was derived,
though it is cumbersome.
In summary, the results show that if the overall contact rate is held
constant, then shorter partnerships lead to a larger Ro, a larger f2 and a
larger A. The relationship between Ro and t2 stays consistent with a pure
contact model, but A grows more slowly than suggested by a pure contact
model and the larger Ro. This behavior may be compared with the effect of
adding variability in contact rate to the basic pure contact model. As shown in
[2, Appendix E], if contacts are allocated according to proportional mixing,
increased variability leads to a larger Ro and a larger A, but a smaller g?. The
relationship between Ro and A remains as for the basic model.
It is common practice to calculate Ro from the formula Ro = c/v, where c,
the rate of effective contacts, is estimated by multiplying the average contact
rate (often obtained from behavioral studies) by the transmission probability
and v is estimated from tracking infected individuals. In the presence of
partnerships, this will always over-estimate Ro because the reproductive
number is diminished by the correlation of contacts occurring within partner-
ships. For the Markov model, the magnitude of this discrepancy can be
calculated from (2) once estimates are available for the basic biological
parameters. Suppose, however, that A is estimated directly from data. The
error should be less if Ro is computed using Ro = A/v + 1 because, for any
particular contact rate, the presence of partnerships will have reduced both R0
and A, keeping them more in line with each other than with the contact rate.
The generalized model handles many of the biological complications that
are necessary for practical models. The life history of partnerships can be
made more complex, affecting the timing of effective contacts. The disease
progression can be elaborated, with several stages of disease that differ in their
infectivity, or randomly and continually changing infectivity. If infectivity is
constant within each state and transitions are still Markov, Ro can be
computed from a larger state transition diagram for the dyad i and j; other-
wise, (5) can be used. It should be noted that the practical use of (3) and (5)
for these generalized models may require some numerical evaluation of
integrals.
Finally, structured populations could be added to the general framework.
If mixing is described by some mixing matrix and all subgroups become large
as N becomes large, then the initial dynamics of the stochastic process will
follow a multi-type branching process. Now p is replaced by Pro,, the probabil-
ity of transmission from an infected member of group m to a susceptible
member of group n. The p,,, then define the branching process and can be
expressed in terms of the partnership and infectivity processes just as for the
homogeneous population model.
S-I-R models with dynamic partnerships 675

This p a p e r has been c o n c e r n e d with the initial d y n a m i c s and the final size
of the epidemic. F o r the p u r e c o n t a c t model, the full d y n a m i c s of the system
for large p o p u l a t i o n s can be a p p r o x i m a t e d by a system of differential equa-
tions. A f o r t h c o m i n g p a p e r will use the a p p r o a c h i n t r o d u c e d a b o v e to deter-
mine the large p o p u l a t i o n limit for the m o d e l with p a r t n e r s h i p s [1].

Acknowledgements. This work was funded by NIH grant P41-RR01632. I would like to
thank K. Braswell, C. Bolan, D. Gruebele, J. M. Lundgren and D. Paschall-Zimbel at the
Simulation Resource for their ongoing logistical support. J. Altmann and E. Ann Stanley
provided helpful comments on several drafts. An anonymous reviewer made many
suggestions for this final version.

References

1. Altmann, M.: The deterministic limit of infectious disease models with dynamic part-
ners, To be submitted to J. Math. Biol.
2. Anderson, R. M. and May, R. M.: Infectious Diseases of Humans: dynamics and
control. New York: Oxford Univ Press 1992
3. Bartlett, M. S.: Stochastic Population Models. New York: Wiley 1960
4. Bolz, G. F.: Simulation on random graphs of the epidemic dynamics of sexually
transmitted diseases - a new model for the epidemiology of AIDS. In: Albeverio, S.,
Blanchard, P. and Testard., D. (eds.) Stochastics, Algebra and Analysis in Classical and
Quantum Dynamics. Dordrecht Boston: Kluwer Academic Publishers 1990
5. Cox, D. R. and Isham, V.: Point Processes. New York: Chapman & Hall, 1982
6. Dietz, K. and Hadeler, K. P.: Epidemiological models for sexually transmitted diseases,
J. Math. Biol., 26, 1-25 (1988)
7. Dietz, K. and Tudor, D.: Triangles in heterosexual HIV transmission. In: Jewell, N. P.,
Dietz, K. and Farewell, V. (eds.) AIDS Epidemiology: Methodological Issues,
pp. 143-155. Boston: Birkhiiuser 1992
8. Harris, T. E.: The Theory of Branching Processes, Englewood Cliffs: Springer, 1963
9. Jagers, P.: Branching Processes with Biological Applications. New York: Wiley 1975
10. Jaworski, J. and Smit, I. H.: On a random digraph, Ann. Discrete Math., 33, 111-127
(1987)
11. Kretzschmar, M., Reinking, D. P., Brouwers, H., van Zessen, G. and Jager, J. C.:
Network models: From paradigm to mathematical tool. In: Kaplan, E. and
Brandeau, M. (eds.) Modeling the AIDS Epidemic, pp. 561-583. New York: Raven
Press 1994
12. Lef6vre, C. and Picard, P.: The final size distribution of epidemics spread by infectives
behaving independently. In: Gabriel, J.-P., Lef6vre, C. and Picard, P. (eds.)
Stochastic Processes in Epidemic Theory. (Lecture Notes in Biomathematics, vol. 86,
pp. 155-169), New York: Springer-Verlag 1990
13. Leslie, W. D. and Brunham, R. C.: The dynamics of HIV spread: A computer simulation
model, Comput. Biomed. Res., 23, 380401 (1990)
14. Mollison, D.: Spatial contact models for ecological and epidemic spread, J. R. Stat. Soc.
B., 3, 283-313 (1977)
15. Peterson, D., Willard, K., Altmann, M., Gatewood, L. and Davidson, G.: Monte Carlo
simulation of HIV infection in an IV drug user community, J. Acquired Immune Def.
Syndr., 3, 1086-1095 (1990)
16. Watts, C. and May, R.: The influence of concurrent partnerships on the dynamics of
HIV/AIDS, Math. Biosci., 108, 89-104 (1992)