PHYSICAL REVIEW E 87, 062810 (2013)

Stochastic analysis of epidemics on adaptive time varying networks

Bhushan Kotnis* and Joy Kuri†
Indian Institute of Science, Department of Electronic Systems Engineering, Bangalore 560012, India
(Received 15 March 2013; published 19 June 2013)
Many studies investigating the effect of human social connectivity structures (networks) and human behavioral
adaptations on the spread of infectious diseases have assumed either a static connectivity structure or a network
which adapts itself in response to the epidemic (adaptive networks). However, human social connections are
inherently dynamic or time varying. Furthermore, the spread of many infectious diseases occur on a time scale
comparable to the time scale of the evolving network structure. Here we aim to quantify the effect of human
behavioral adaptations on the spread of asymptomatic infectious diseases on time varying networks. We perform
a full stochastic analysis using a continuous time Markov chain approach for calculating the outbreak probability,
mean epidemic duration, epidemic reemergence probability, etc. Additionally, we use mean-field theory for
calculating epidemic thresholds. Theoretical predictions are verified using extensive simulations. Our studies
have uncovered the existence of an “adaptive threshold,” i.e., when the ratio of susceptibility (or infectivity)
rate to recovery rate is below the threshold value, adaptive behavior can prevent the epidemic. However, if it is
above the threshold, no amount of behavioral adaptations can prevent the epidemic. Our analyses suggest that
the interaction patterns of the infected population play a major role in sustaining the epidemic. Our results have
implications on epidemic containment policies, as awareness campaigns and human behavioral responses can be
effective only if the interaction levels of the infected populace are kept in check.

DOI: 10.1103/PhysRevE.87.062810 PACS number(s): 89.75.Fb, 89.75.Hc

I. INTRODUCTION The other limit of time-scale separation occurs when the

epidemic process evolves at a time scale much larger than
Our modern societies are best described as complex systems
the underlying contact structure. This is equivalent to every
consisting of a large number of interacting components. Due
individual randomly contacting every other individual, leading
to this underlying complexity, understanding and containing
to homogeneous mixing among the population.
the spread of biological diseases in human societies becomes
However, many networks evolve at a time scale comparable
a challenge. In the last decade, mathematical modeling
to that of the diffusion process. Examples are the spread of
techniques [1] and agent-based simulations [2] have allowed
a Bluetooth-based virus on smart phones [11], information
researchers to gain a deeper understanding of such processes.
spreading during an emergency [12], and information diffusion
Traditional mathematical models assumed a homogeneous
in ad hoc networks [13]. In the context of biological epidemics,
population where individuals contacted others randomly, thus
the spread of various sexually transmitted Infections [14,15],
ignoring the connectivity structure of individuals [3]. Due
slow infections such as tuberculosis (TB) [16], and foot and
to the availability of high-quality data, researchers have
mouth disease among cattle [17] are transmitted at a time scale
unearthed detailed connectivity structures of various complex
comparable to the network evolution.
systems [4,5] and have represented them as networks. For ex-
Human behavioral response to the epidemic is another
ample, in social networks, individuals are described as nodes,
important phenomenon that affects epidemic dynamics. The
and relationships between them are represented by links.
awareness of an epidemic through news and other media can
The discovery of network structures led to a detailed study
change the underlying connectivity structure. This change can
of their effects on the epidemic spreading process. Mathemat-
be either due to self-initiated precautions, such as avoiding
ical analysis using mean-field theory has revealed valuable
contact, or due to government-mandated programs such as
insights, such as lack of an epidemic threshold on scale-
closing schools, airports, and other services. In recent years
free networks [6,7]. These studies assumed a connectivity
various studies have modeled this adaptive behavior; see
structure which evolved on a time scale much larger than
Ref. [18] for a recent review. In a few studies [19], the network
the time scale of the diffusion process. This assumption
is considered to be static, and susceptibility of individuals
allowed the system to be modeled as a static network, which
changes with number of infected neighbors. This approach
is reasonable for networks that evolve very slowly compared
is further refined by extending the standard susceptible
to the time scale of the diffusion process. Examples of
infected susceptible or susceptible infected recovered com-
this include the spread of computer viruses on the World
partmental model to include behavior-related compartments,
Wide Web network or traffic dynamics on transportation
such as a fear compartment [20–23]. Individuals shift to
networks. Such a time-scale separation assumption also holds
these compartments at a rate proportional to the infected
reasonably well in epidemic models for diseases which spread
populace. Parameters in these compartments differ from other
rapidly [8], e.g., SARS [9] or the spread of Mycoplasma
compartments, such as reduction of susceptibility in the fear
pneumoniae in psychiatric wards [10].
compartment.
Thus, the effect of adaptive behavior must be included in
*
bkotnis@dese.iisc.ernet.in epidemic models that do not exhibit time-scale separation. To
†
kuri@dese.iisc.ernet.in this end, many studies [24–26] have analyzed the effect of

1539-3755/2013/87(6)/062810(10) 062810-1 ©2013 American Physical Society

BHUSHAN KOTNIS AND JOY KURI
adaptive behavior on the disease and the contact structure
using adaptive networks. Healthy individuals break their
connections with infected individuals and reconnect to other
healthy individuals for avoiding contact with the infected.
Variations of this model include rewiring and connecting to
random individuals [27] or simply breaking contacts with the
infected [28]. Adaptive networks do not assume time-scale
separation; i.e., the infection spreads at a time scale comparable
to the evolution of the network.
However, this rewiring dynamic assumes that healthy
individuals can recognize infected individuals in order to
break contact with them. This is possible when the disease
is symptomatic. However, if the disease is asymptomatic or if
it has a large asymptomatic stage like TB or human papilloma
virus, then the diagnosis of the disease (in the asymptomatic
stage) can be carried out only by a medical practitioner
after performing a thorough medical examination. In such a
scenario healthy individuals may not be able to recognize an
infected individual. Thus, disease awareness can be generated
only through news, other media, and government advisories.
Furthermore, in such models, the underlying network structure
evolves only as a response to the epidemic.
The processes responsible for the time evolution of network
are not well understood [29]. Changes in network structure
can be due to either human behavioral response to the
epidemic or due to various other processes. We believe that it
is important to consider time varying networks, i.e., networks
which change their structure not only due to the epidemic
process but also due to other complex processes, as some
recent studies [30–33] have shown that dynamical processes
such as random walks behave differently on these networks.
These studies have prompted us to investigate the effect of
human behavioral adaptations on epidemic dynamics when
the disease is asymptomatic and does not exhibit time-scale
separation.
Our model incorporates the effects of
(1) an asymptomatic or a mildly symptomatic disease,
(2) a disease which does not exhibit time-scale separation,
(3) adaptive behavior in response to the epidemic, and
(4) a contact network which evolves independently of the
epidemic.
We use a susceptible-infected-susceptible (SIS) model to
study the epidemic dynamics and an activity-driven model [31]
to account for the changes in contact structure.
All the studies described above have used mean-field theory
to analyze their models. This analysis results in a system
of deterministic differential equations which describe the
epidemic process. Useful analytical insights can be obtained by
using mean-field analysis. However, a full stochastic analysis
can uncover subtle but important details such as probability
of an outbreak, probability distribution of number of infected
individuals, etc., which mean-field analysis cannot. In this
paper we carry out a detailed stochastic analysis of epidemic
dynamics taking into account the effect of the connectivity
structure adaptations caused by human behavioral response to
the epidemic.
We investigate the effect of adaptive human behavior on
epidemic outbreak probability, mean duration of the epidemic,
and outbreak size. We also study the effect of adaptive
behavior on the epidemic threshold and provide an explicit
062810-2
PHYSICAL REVIEW E 87, 062810 (2013)
analytical expression for the epidemic threshold. Additionally,
we calculate the probability of epidemic reemergence after
strong measures are implemented to curb the epidemic. And
finally, we quantify the accuracy of mean field theory in
predicting epidemic dynamics and compare it with predictions
based on stochastic analysis. The quasi-stationary probability
distribution of the number of infected individuals is also
computed.
Our investigations reveal that only under certain conditions
can adaptive human behavior arrest the epidemic. Contact
avoidance strategy pursued by healthy individuals is effective
only if the interaction patterns of the infected individuals
are within certain limits. We believe that our results can
inform policies drafted by governmental and nongovernmental
organizations to contain epidemics.
The paper is organized as follows. Section II introduces
a time varying network model in which each individual
interacts with others with a given activity rate. Healthy
individuals modify their activity rate as a response to the
epidemic. The epidemic is analyzed using continuous time
Markov chain (CTMC) theory in Sec. III. Various properties
of the epidemic model such as outbreak probability, mean
duration of the epidemic, etc., are investigated in Sec. IV.
Interpretations and implications of the results are presented
in Sec. V, and Sec. VI summarizes our conclusions and main
results.
II. MODEL
We use a simple activity-driven model, first proposed in
Ref. [31], to capture the interaction patterns of individuals. The
advantage of the activity-driven model is that it allows one to
model an individual’s interaction rate or activity, which enables
the model parameters to be estimated through measurements.
Each individual i is characterized by an activity level ai .
Activity level is the rate of interaction of the individual, i.e.,
number of the interactions initiated by the individual, in a
fixed time window of length t. Throughout the paper, it
is assumed that t is small. The model assumes that the
interaction generation process is memoryless. Individuals are
represented by nodes, and every interaction an individual
initiates is symbolized by a link, allowing the system to be
modeled as a network. In this study we do not consider
the effects of demography, and hence the total number of
individuals, N , is fixed. Network evolution is described by
the following process [31]:
(1) At each time step t, the network Gt contains N nodes
and zero links, i.e., edge set Et = φ.
(2) Each node becomes active with probability ai t. When
the node becomes active it picks m nodes uniformly at random
and establishes undirected links with them. Nonactive nodes
can receive links from active nodes.
(3) At time t + t, all the links in the network Gt are
removed; thus Et+t = φ and the process starts again from
step 1.
In our analysis we assume that all susceptible nodes have
the same activity level ah > 0, and all infected nodes have
activity level ainf > 0. The scenario where individuals can
have different activity levels is studied through simulations.
In general ainf  ah , as mild infection symptoms may result
STOCHASTIC ANALYSIS OF EPIDEMICS ON ADAPTIVE . . . PHYSICAL REVIEW E 87, 062810 (2013)

in decreased activity. When individuals with mild symptoms at least one of those infected neighbors transmits the disease
visit a medical practitioner, they become aware of their infected in t}:
state, which may lead to a reduction in their interaction levels.
As ah  ainf , this scenario is different from the homogeneous m    k  
m I I m−k
mixing assumption, since the probability of a susceptible φ= 1− (1 − (1 − β)k )asus (I )t
k N N
contacting a susceptible is different from the probability of k=1
a susceptible contacting an infected node.  I  
I
We use the classical compartmental SIS epidemic spreading + (θ )k (1 − θ )I −k (1 − (1 − β)k )ainf t − o(t),
model. Individuals are assumed to belong to either the k=1
k
susceptible compartment or the infected compartment. A (1)
susceptible node can become infected if any of its infected
neighbors transmits the disease. Let β be the probability of where θ is the probability that the infected node establishes one
transmitting disease. This model allows for the recovery of −1
link with the susceptible node of interest: ∴ θ = ( N N
m − 1 )/( m ).
infected nodes; furthermore, we assume that the infection is
nonfatal. Recovered nodes become susceptible and can be The first term of (1) represents the probability of a suscep-
reinfected. Let μ be the rate of recovery for each node, i.e., tible node actively connecting to an infected node and the
each node recovers with probability μt in time t. probability that infection is transmitted. The second term
Since the disease is asymptomatic or mildly symptomatic, represents the chance of infected nodes actively connecting
individuals will gain information about the epidemic only to the susceptible node and passing on the infection. The third
through news, Internet, and government advisories and an- term is negligible as it represents the chance that the two
nouncements, but not through contact with infected individu- nodes will be active simultaneously. Since (t)2 ≈ 0, the
als. We assume that this information depends on the number of probability of two or more susceptible individuals contracting
infected individuals. Additionally, we assume that all healthy the infection in time t is negligible. Therefore the probability
individuals will show the same response to the epidemic [19]. that k susceptible nodes get infected out of a population
We use the exponential model in Ref. [19] to model the of S susceptible nodes is given by ( Sk )φ k (1 − φ)S−k . Since
behavioral response. Let I (t) and S(t) be the number of φ 2 is negligible (t 2 ≈ 0), the chance that more than one
infected and susceptible individuals at time t. We replace the susceptible node catches the infection out of S susceptible
activity level of healthy population ah with asus (I ) = ah e−δ·I (t) nodes is very small. Therefore, the probability that one
where δ is the risk perception factor, which models the susceptible node gets infected from S susceptible nodes is
perceived risk felt by healthy individuals. An exponential given by Sφ + o(t) = (N − I )φ + o(t).
function is used to model the nonlinearity of the response It can be easily shown that (see Appendix B for details)
as the awareness results from news and other media [19].
Individuals who have contracted the disease may receive P{I (t + t) = i + 1|I (t) = i} = (N − i)φ + o(t),
information about their condition through a routine medical
checkup. Such knowledge may drive them to reduce their P{I (t + t) = i − 1|I (t) = i} = iμt + o(t).
activity level (ainf  ah ); however, they will not perceive a risk
of being infected since they have knowledge of their condition. Thus, the CTMC becomes a birth death process with absorbing
state 0. The birth rate for state i given by
III. ANALYSIS
m    k  
m i i m−k
We use a continuous time Markov chain (CTMC) to analyze bi := (N − i) 1−
k=1
k N N
the epidemic process.
The network evolves on a time scale comparable to that × [1 − (1 − β)k ] · asus (i) + (N − i)
of the epidemic process. In our analysis we assume that the  i  
two time scales are the same [30,31]. The network formation i
× (θ )k (1 − θ )i−k [1 − (1 − β)k ] · ainf (2)
process happens every t time units, and since the recovery k=1
k
process should also have the same time scale, we assume
that in time t, each infected node recovers with probability and death rate
μt. For t → 0, it can be shown that the recovery process
becomes a Poisson process (see Appendix A for details). The di := μi. (3)
probability that two infected nodes will recover simultaneously
in time t = μ2 (t)2 is negligible since t is very small.
Therefore the probability that one infected node recovers out This allows us to write the rate matrix Q:
of a population of I infected nodes is given by I μt + o(t).
Since the epidemic and network evolution have the same qi,j = bi for j = i + 1, i = 1,2,3, . . . ,N − 1,
time scale, we assume that during time interval t, an infected qi,j = di for j = i − 1, i = 1,2,3, . . . ,N,
node can pass on the infection to another susceptible node with qi,j = −(bi + di ) for j = i, i = 1,2,3, . . . ,N − 1,
probability β.
P{a susceptible node gets infected in t} = φ := qN,N = −dN ,
P{susceptible node has at least one infected neighbor, and qi,j = 0, otherwise,

062810-3
BHUSHAN KOTNIS AND JOY KURI PHYSICAL REVIEW E 87, 062810 (2013)

⎛ ⎞
0 0 0 0 ··· 0 Absorption probability depends on the initial number of
⎜ d1 infected nodes. Let zi be the probability of absorption into state
⎜ −(b1 + d1 ) b1 0 ··· 0 ⎟⎟
⎜0 0 when initial number of infected node is i for i = 1,2,3, . . .
Q=⎜ d2 −(b2 + d2 ) b2 ··· 0 ⎟⎟. and z0 = 1:
⎜ ⎟
⎜ .. .. .. .. .. .. ⎟
⎝. . . . . . ⎠ zi = lim P{X(t) = 0|X(0) = i}.
t→∞
0 0 0 0 ··· −dN
Due to the birth death process we can write the following
(4)
relation:
bi di
IV. RESULTS zi = zi+1 + zi−1 , ∀ i = 1,2,3, . . . .
bi + di bi + di
In this section we investigate the effect of human behavioral This recurrence relation can be solved to obtain (see Ref. [35],
response on the epidemic dynamics by studying various Ch. 6)
properties of the model. We study outbreak probability [34] in ∞
 ∞
Sec. IV A, quasistationary distribution [34] in Sec. IV B, mean i=k ρi
zk =  if ρi < ∞
epidemic duration in Sec. IV C, and epidemic reemergence 1+ ∞ i=1 ρi i=1
probability in Sec. IV D. Calculation of the epidemic threshold
and final size of the outbreak is done using mean-field analysis ∞ d1 ·d2 ·d3 ···di
and zk = 1 if i=1 ρi = ∞, where ρi = b1 ·b2 ·b3 ···bi for
in Sec. IV E. Finally, the accuracy of mean-field theory in i = 1,2,3, . . . . The probability of outbreak for the finite state
predicting the epidemic is discussed in Sec. IV F using Monte C.T.M.C when initial infected individuals i, is approximately
Carlo simulations. equal to 1 − zi .
To study the effect of behavioral adaptation due to risk
perception factor δ, we numerically calculate this probability
A. Probability of an outbreak
for various values of δ in Figs. 1(a) and 1(b). Outbreak
Due to the existence of an absorption state in the SIS probability variation with δ in Fig. 1(a) is very different from
epidemic CTMC, the epidemic will eventually die out. This that in Fig. 1(b). For lower values of μβ ratio and lower activity
can happen in two ways: either the disease never spreads and levels ah and ainf , even a small quantity of risk perception
dies out immediately, or the disease spreads to a significant factor (δ) can completely stop the epidemic, whereas, for a
proportion of nodes and then dies out eventually. The state relatively higher μβ ratio and larger activity levels, a high risk
where a significant proportion of the population becomes perception factor is unable to stop the epidemic.
infected is termed an outbreak.
Let pi be the probability of an outbreak when the initial
number of infected individuals is i. pi can be estimated by B. Quasistationary probability distribution
replacing the finite state CTMC by a random walk with infinite The epidemic will eventually die out, however, in the
states and an absorbing state 0 [34]. Let X(t) be an infinite event of an outbreak, the epidemic appears to enter a stable
state non-negative random walk process, with absorbing state equilibrium before dying out. Thus, in the preextinction stage
0, and birth and death rates given by (2) and (3). For the random the probability distribution of the epidemic approaches a sta-
walk, at t → ∞ the random walker either will be in state 0 or tionary distribution. This distribution is called quasistationary
will be in a nonzero state: Probability of Absorption in state probability distribution [34]. Let qi (t) be the probability of the
0 = limt→∞ P{X(t) = 0}. event that number of infected nodes = i at time t, conditioned

0.7 0.94
β = 0.025, μ = 0.2, a = 20, a = 4, m = 1, β = 0.6, μ = 0.5, a h = 100, a inf = 50, m =1,
h inf
0.6 N = 50, I(0) = 1. 0.92 N = 50, I(0) = 1.
Outbreak Probability
Outbreak Probability

0.5
0.9
0.4
0.88
0.3
0.86
0.2

0.84
0.1

0 0.82
0 0.01 0.02 0.03 0.04 0 2 4 6 8 10
Risk Perception Factor ( δ ) Risk Perception Factor ( δ )
(a) (b)

FIG. 1. (Color online) Outbreak probability vs risk perception.

062810-4
STOCHASTIC ANALYSIS OF EPIDEMICS ON ADAPTIVE . . . PHYSICAL REVIEW E 87, 062810 (2013)

0.16 C. Duration of the epidemic

β = 0.05, μ = 0.1, m = 1, Approximation
0.14 a h =30, a inf = 10, Actual The duration of the epidemic is the time until the process
N = 100, I(0) = 1. hits the zero absorption state [39]. The duration of the epidemic
Probability of Infection

0.12 depends on the initial number of infected nodes. Let Ti be the

duration of the epidemic when initial number of infected nodes
0.1
is i. Clearly T0 = 0.
0.08 By obtaining the mean and higher order moments of Ti , we
can study the effect of the adaptive behavior on the epidemic
0.06 duration. To obtain the mean and higher order moments, we
0.04 first calculate the probability generating function (PGF) of Ti .
The PGF of Ti is given by Gi (z) = E[z(Ti ) ]
0.02 Due to the birth death process, we can write the following
recurrence relation:
0
0 20 40 60 80 100
Number of Individuals E[z(Ti ) ] = bi tE[z(Ti+1 +t) ] + di tE[z(Ti−1 +t) ]
+ [1 − (bi + di )t]E[z(Ti +t) ].
FIG. 2. (Color online) Quasistationary probability distribution.
After simplification we get
on the event that epidemic has not died till time t. Formally, E[z(Ti ) ](1 − zt ) = t{bi E[zTi+1 ] + di E[zTi−1 ]
qi (t) = P{I (t) = i|I (s) > 0,t > s} for i = 1,2,3, . . . ,N. − (bi + di )E[zTi ]}
We use the approach in Ref. [34] to calculate the quasista-
tionary distribution. It is possible to show and mean duration T̄i = dGdzi (z) |z = 1.
The above equation then becomes
dq
= Q̃q + μq1 q, −T̄i (bi + di ) + bi T̄i+1 + di T̄i−1 = −1.
dt
This can be written in matrix form as B · T̄ = −1 where T̄ is
where Q̃ is the same as matrix Q with the first row and column the column vector with elements T̄i and 1 is a column vector
removed and q is the column vector of probabilities qi : of ones and
⎛ ⎞
−(b1 + d1 ) b1 0 0 ··· 0
Q̃T q∗ = −μq1∗ q∗ . ⎜ −(b2 + d2 ) b2 0 · · · 0 ⎟
⎜ d2 ⎟
B=⎜ ⎜ .. .. .. .. . .
⎟
.. ⎟.
This equation cannot be solved directly; however, it can ⎝ . . . . . . ⎠
be solved iteratively [36,37]. Approximating the quasista- 0 0 0 0 ··· −dN
tionary distribution is a well-studied problem, and several
approximations have been proposed [36–38]. We use a simple Matrix B is irreducibly diagonally dominant; hence the system
approximation proposed in Ref. [37] to calculate the quasista- of equations will have a unique solution [40].
tionary distribution. This approximation assumes d1 = 0; i.e., We calculate the expected duration numerically and study
we remove the absorbing state 0 from the CTMC. Figure 2 its variation with risk perception factor δ in Figs. 3(a) and 3(b).
shows a good fit between the approximation and the exact We find a large variation in mean duration with δ for a lower
β
calculation. μ
ratio and activity levels. In such a regime, adaptive behavior

6
10
β = 0.025, μ = 0.2, a h = 20, a inf = 4, m = 1, β = 0.06, μ = 0.5, a = 100, a = 50, m = 1,
h inf
Expected Duration (Time units)
Expected Duration (Time units)

N = 100, I(0) = 1. N = 100, I(0) = 1.

14.56
10
4
10

14.53
10

2
10
14.5
10

0 14.47
10 10
0 0.5 1 1.5 2 0 5 10 15 20
Risk Perception Factor ( δ ) Risk Perception Factor ( δ )
(a) (b)

FIG. 3. (Color online) Mean duration of the epidemic vs risk perception.

062810-5
BHUSHAN KOTNIS AND JOY KURI PHYSICAL REVIEW E 87, 062810 (2013)

can cause considerable reduction of epidemic duration. How- 1

ever, in the scenario of high μβ ratio and activity levels, adaptive μ = 0.05
behavior causes very little variation in the epidemic duration. μ = 0.01
0.8 μ = 0.006

Reemergence Probability
D. Epidemic reemergence
0.6
In the event of an epidemic, the government can respond
by temporarily closing schools, churches, shopping centers,
and places where large numbers of people congregate. Such 0.4
measures can successfully stop the epidemic. However, the
disease can resurface and cause another epidemic if these
0.2
measures are not enforced for sufficient duration, i.e., till the
disease is eradicated from the community and thus does not
get a chance to reappear. Here we calculate the probability of 0
epidemic reemergence when such decrees have been lifted. 0 500 1000 1500
Duration of school closure (Time units)
We assume that the local administration enforces epidemic
containment policies if the epidemic hits a threshold value Ith . FIG. 4. (Color online) Probability of epidemic reemergence vs
Furthermore, we assume that such policies cause the activity duration D. Parameters: Ith = 50, β = 0.006, ah = 100, ainf =
level of both the healthy and the infected to drop to zero. Since 50, δ = 1.
the implementation of these policies is temporary, we assume
that life returns to normal after a time duration D, causing rate μ. This analysis allows one to calculate the appropri-
interaction levels to shoot up to their preepidemic values. ate duration of measures implemented to prevent epidemic
After the epidemic containment policy is enforced, birth reemergence.
rate bi = 0, and therefore the rate matrix Q of the CTMC I (t)
can be written as
E. Mean-field analysis
qi,j = di for j = i − 1, i = 1,2,3, . . . ,Ith Mean-field analysis is a very useful tool to understand
qi,j = −di for j = i, i = 1,2,3, . . . ,Ith , epidemic behavior. Here we aim to study the size of the
qi,j = 0, otherwise. outbreak using mean-field theory. A mean-field equation of
the epidemic process can be written from first principles or
Let t = 0 be the time when the epidemic containment policy can be derived from the Kolmogorov equations [34]. The
is applied, i.e., when the number of infected hits the threshold mean-field equation is a deterministic ordinary differential
Ith ; therefore pIth (0) = 1: equation describing the evolution of the mean number of
¯ The mean-field equation for an
infected individuals I (t).
P{Epidemic reemergence after duration D}
activity-driven model can be written as follows [31]:
= 1 − p0 (D)P{No epidemic outbreak
− p1 (D)P{No epidemic outbreak} · · · d I¯ I¯ I¯
= −μI¯ + βmah e−δI (N − I¯) + βmainf (N − I¯) .
¯

− pIth (D)P{No epidemic Outbreak}. dt N N

(5)
P{No epidemic outbreak} is calculated in Sec. IV A.
Substituting it in the above equation we get The first term represents the reduction in infections due to
recovery, the second term describes the rise in infections
P{Epidemic reemergence after duration D} due to susceptible individuals contacting infected individuals,
th
i=I while the final term represents the increase of infections when
≈1− pi (D)zi . infected individuals contact susceptible individuals.
i=0 Figure 5 compares the evolution of the epidemic predicted
Let p(t) be a column vector of probabilities pi (t). by mean-field analysis with the one predicted by calculating
From the Markov property, the mean of the CTMC. This is compared with data obtained
from Monte Carlo simulation of the epidemic.
p(D) = PT (D) · p(0),
where P(D) is the transition probability matrix. From the 1. Threshold
Kolmogorov equations we get SIS epidemic displays a phase transition behavior, where,
dQ above the μβ threshold, the infected population is nonzero,
= P(t),
dt while below the threshold it is zero. We calculate the threshold
whose solution is given by when δ = 0, i.e., in the absence of adaptations. Dividing (5)
by N we get
P(t) = eQ·t .
d ī
Thus P(D) = eQ·D . We calculate p(D) using numerical meth- = −μī + βmah (1 − ī)ī + βmainf (1 − ī)ī,
dt
ods. In Fig. 4 we plot the variation of epidemic reemergence
probability with duration D, for various values of recovery where ī = I¯/N .

062810-6
STOCHASTIC ANALYSIS OF EPIDEMICS ON ADAPTIVE . . . PHYSICAL REVIEW E 87, 062810 (2013)

80 Equilibrium point ī(t) = 0 is asymptotically stable if dV

dt
 0.
This translates to
70
Number of Infected Individuals

Mean Field Theory β 1

60 Simulation  −δN
.
C.T.M.C (mean)
μ mah e ī + mainf
50
At early time, ī ≈ ī0 . Thus if the above condition is true for
40 ī0 , then it holds for all ī > i0 . Notice that by putting δ = 0 we
30
can recover (7). Thus we get four cases:
(1) μβ  mah +ma
1
inf
20
(2) 1
mah +mainf
< μβ  1
mah e−δN ī0 +mainf
10 (3) 1
mah e−δN ī0 +mainf
< β
μ
 ma1inf
β
0 (4) μ
> ma1inf
0 10 20 30 40
Time units In the first case since R0  1, there will be no epidemic.
In the second case, if for a given ī0 , δ is sufficiently large
FIG. 5. (Color online) Monte Carlo simulation averaged over such that case 2 holds then epidemic will die down. In case 3,
5000 samples. Parameters: μβ = 0.5, m = 1, ah = 30, ainf = 10, δ = for given risk perception factor δ, there will be an epidemic
1, N = 100, I (0) = 1. outbreak; however, in this regime, there exists a δ́ > δ such that
β
μ
 ma e−δ́N1ī0 +ma , which can stop the epidemic. This shows
h inf

During the early time region, i.e., at t → 0, i¯2 ≈ 0: that adaptive behavior can prevent an epidemic. However, the
prevention strategy has its limits, as seen by case 4. If the
d ī condition in case 4 is satisfied, then even δ → ∞ is unable to
= −μī + βmah ī + βmainf ī. (6) stop the epidemic.
dt

Thus we obtain a linear ODE. Clearly, ī > 0 if and only if 2. Size of outbreak
μ < βmah + βmainf . Thus When the epidemic enters a steady state the rate of change
of infected individuals is zero. Thus the size of outbreak in
β steady state can be obtained by setting the right-hand side of
R0 = · (mah + mainf ), (7)
μ (5) to 0:

where R0 is the basic reproductive number. R0 > 1 results in a 1 −δ I¯ I¯ I¯
I=
¯ βmah e (N − I ) + βmainf (N − I )
¯ ¯ .
nonzero infected population, while if R0  1, then the primary μ N N
infection does not cause an epidemic outbreak.
Now we calculate the threshold condition for δ > 0. This is a fixed point equation whose solution will give the size
Consider the following Lyapunov function: V (t) = ī(t), where of outbreak in steady state. In Figs. 6(a) and 6(b) we plot the
ī(t) = I¯(t)/N. At t → 0,i¯2 ≈ 0. Therefore, size of outbreak versus the risk perception factor δ. Outbreak
size falls rapidly with increase in δ for the system parameters in
dV Fig. 6(a). However, for the system parameters in Fig. 6(b), δ has
= −μī + βmah īe−δN ī + βmainf ī, limited effect of outbreak size. The impact of adaptive behavior
dt

60 91
β = 0.25, μ = 0.2, a h = 20, a inf = 4, m = 1, β = 0.06, μ = 0.5, a h = 100, a inf = 50, m = 1,
90
Number of Infected Individuals
Number of Infected Individuals

50 N = 100, I(0) = 1. N = 100, I(0) = 1.

89
40
88

30 87

86
20
85
10
84

0 83
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Risk Perception Factor ( δ ) Risk Perception Factor ( δ )
(a) (b)

FIG. 6. (Color online) Epidemic outbreak size vs risk perception.

062810-7
BHUSHAN KOTNIS AND JOY KURI
35
30
Difference in Percentage
25
20
15
10
5
0
−5
0 1 2 3
(β/μ)
FIG. 7. (Color online) Percentage difference in final size of the
outbreak between the one predicted by mean-field theory/CTMC and
the one obtained by Monte Carlo simulations. Monte Carlo simulation
averaged over 5000 samples. Parameters: m = 1, ah = 30, ainf =
10, δ = 1, N = 50, I (0) = 1.
on outbreak size depends on system parameters which are
governed by the disease and lifestyle patterns.
F. Accuracy of mean-field theory
We study the accuracy of mean-field theory in predicting
the final size of the epidemic by comparing it with the final size
of the epidemic obtained from Monte Carlo simulation. The
same is also done for the average infected individuals obtained
from the CTMC distribution. Results are illustrated in Fig. 7.
In general, final size of the epidemic obtained from mean-field
approximation is less accurate than the one calculated from the
CTMC. Accuracy of both mean-field theory increases with μβ .
This can be understood as follows: for a μβ ratio just above the
threshold, there is a significant chance that there will be no
epidemic outbreak. This is not captured by mean-field theory,
and hence it overestimates the epidemic size.
V. DISCUSSION
Our analysis suggests the existence of a threshold condition
which governs the effectiveness of human adaptive behavior
for stemming the epidemic. Since this threshold condition is
applicable only on populations with adaptive behavior hence
we term this as “adaptive threshold.” For given activity levels,
when ratio of infection probability and the recovery rate μβ are
below the adaptive threshold, exponential activity reduction
can successfully prevent an epidemic, while the lack of activity
reduction results in an epidemic outbreak. But, if μβ ratio
is above the threshold, then exponential activity reduction
strategy is unable to stop the epidemic. Thus, human behavioral
response to the epidemic can be effective in preventing the
epidemic only under conditions outlined in Sec. IV E1.
The existence of an adaptive threshold is supported by
the analysis of outbreak probability, mean duration of the
epidemic, and mean-field stability analysis. From Fig. 1(a) it
is clear that a small increase in risk perception factor δ results
PHYSICAL REVIEW E 87, 062810 (2013)
1
Mean Field Theory
C.T.M.C (mean)
0.8
δ=0
Outbreak Probability
δ=5
0.6 δ = 10
0.4
0.2
0
4 5 0 0.5 1 1.5 2
(β/μ)
β
FIG. 8. (Color online) Outbreak probability vs μ
. Parameters:
m = 1, ah = 20, ainf = 4, N = 100, I (0) = 1.
into a sharp drop in outbreak probability, while in Fig. 1(b)
even a large value of risk perception factor δ is unable to reduce
the outbreak probability. We see μβ = 0.269 in Fig. 1(a) while
β
= 4 in Fig. 1(b). This suggests the existence of a critical
μ
value of μβ above which behavior changes are unable to stop the
epidemic. In Fig. 8 we plot the outbreak probability versus μβ
for various risk perception factors δ. Above the threshold value
β
= 0.62 the outbreak probability becomes significant enough
μ
to cause an epidemic. Above this threshold, higher values of
δ have limited effect on outbreak probability. Analysis of
mean duration of epidemic [Figs. 3(a), 3(b)] and outbreak
size [Figs. 1(a) and 1(b)] show similar curves, supporting
the existence of an adaptive threshold. These results obtained
from numerical computation are confirmed through Lyapunov
stability analysis. Furthermore, the exact value of thresholds
are also calculated.
When the μβ ratio is above the adaptive threshold, an
increase in risk perception has a limited effect on the epidemic.
The primary reason for this behavior is the nonzero activity
level of the infected population. A nonzero activity level
enables infected individuals to connect with healthy populace,
thus sustaining the epidemic. Thus, if the activity level of
infected individuals ainf = 0, from (5) we get limδ→∞ I¯(t) =
0. Thus, if ainf = 0, then an increase in the risk perception
factor can quickly kill the epidemic.
We suspected that populations with heterogeneous activity
levels may also display an adaptive threshold behavior. In order
to test our intuitions we ran extensive simulations. The activity
levels of individuals were assumed to be distributed as a power
law, i.e., activity of node i = ai = ah P (x) where P (x) ∝ x −γ .
If a healthy node becomes infected, then we assume that
its activity level reduces by a factor of 4. Susceptible nodes
reduce their activity level exponentially as described in Sec. II.
In Fig. 9 through extensive simulations, we plot outbreak
size versus μβ for various risk perception factors δ. From the
figure, it is clear that up to β/μ = 1, adaptive behavior can
stop the epidemic; however, beyond this value an increase
of risk perception (from δ = 5 to δ = 10) cannot prevent the
062810-8
STOCHASTIC ANALYSIS OF EPIDEMICS ON ADAPTIVE . . .
100
δ=0
δ=5
80 δ = 10
Outbreak Size
60
40
20
0 measures are temporary, and eventually things must return to
0 1 2 3 4 5 6
(β/μ)
FIG. 9. (Color online) Outbreak size vs β/μ for a population with
heterogeneous activity rates. Results computed by averaging 5000
runs of Monte Carlo simulation. Parameters: m = 1, ah = 500, γ =
2.5, N = 100, I (0) = 1.
epidemic. This suggests an existence of an adaptive threshold
for populations with heterogeneous activity patterns.
Previous studies [19,21,23] have reported the existence of a
similar threshold condition. In these models, the susceptibility
β of individuals changes due to awareness of the epidemic.
Authors in Ref. [19] report the existence of a critical value
of risk perception factor beyond which the disease dies
out; thus they conclude that adequate precautions taken by
healthy population can stop the epidemic. Studies [21,23] also
report the existence of a similar threshold condition which is
dependent on fear or awareness parameters of the disease. If
we assume that individuals modify their interaction patterns
as opposed to reducing susceptibility, then as illustrated by
our study, this conclusion may not be true. Even if the healthy
population reduce their interaction patterns (exponentially),
a nonzero activity level of the infected population can still
sustain the epidemic. Thus, for an asymptomatic disease, it is
important to incentivize infected individuals for limiting their
activity level.
VI. CONCLUSION AND FUTURE WORK
In this paper we have investigated the effect of human
behavioral adaptations on SIS epidemic dynamics for a
population whose connection structure varies with time. We
considered a scenario where the disease is asymptomatic
or mildly symptomatic, making selective contact avoidance
impossible. Thus, human behavioral response to the disease
results from awareness campaigns and advisories provided
by government and the media. We modeled the epidemic
as a CTMC, which enabled us to perform a full stochastic
analysis. Our investigations on the effect of risk perception
[1] O. Diekmann and J. Heesterbeek, Mathematical Epidemiology
of Infectious Diseases: Model Building, Analysis and Interpre-
tation, Vol. 5 (Wiley, New York, 2000).
062810-9
PHYSICAL REVIEW E 87, 062810 (2013)
on outbreak probability and mean duration of the epidemic
suggest an existence of an “adaptive threshold.” If the
ratio of spreading probability to recovery rate is below the
threshold, then human behavioral response can prevent an
epidemic; however, if it is above the threshold, then human
behavioral response is unable to suppress the epidemic. This
result has implications on epidemic containment policies.
Epidemic awareness campaigns can be effective only if the
interaction patterns of the infected population are limited.
Thus, incentives must be provided to the diseased population
to keep their interaction levels to minimum possible values.
Additionally, administrative authorities may consider closing
schools, airports, supermarkets, etc., to limit interactions. Such
normalcy. This may trigger another epidemic outbreak as all
diseased individuals may not have recovered. This motivated
us to calculate the probability of epidemic reemergence when
the said policy is applied for a fixed duration.
Our study does not consider the role of culture and ethnicity
in human behavior. We assumed that all healthy individuals
will exhibit the same response to the epidemic, which may
not be correct in situations where culture and ethnicity shape
the response to an epidemic. This work can be extended to
include the effect of varied response to the epidemic. Another
extension would be to capture the effect of seasonal variations
in activity patterns resulting from school vacations, holidays,
etc. We plan to explore these avenues in the near future.
APPENDIX A
P{Infected node recovers in t} = μt. Let A be the event
that an infected node does not recover in duration T . P{A} =
(1 − μt)T /t . Since t → 0, this quantity becomes e−μT .
Now, if the same node recovers and again becomes infected,
then the event that it does not recover in time T is independent
of event A and is given by e−μT . Thus, the interrecovery
times have exponential distribution and are independent, and
therefore the recovery process is a Poisson process.
APPENDIX B
P{I (t + t) = i + 1|I (t) = i} = P{One infection and no
recovery in time t}, since the chance of more than one
infection and recovery is negligible.
Now, P{one infection and a recovery in t} = P{one in-
fection in t}P{one recovery in t} = (Sφ + o(t))(μt +
o(t)) = φμt + o(t), which is negligible as t 2 ≈ 0.
Therefore, P{one infection and a recovery in t} ≈ P{one
infection in t}.
Hence, P{I (t + t) = i + 1|I (t) = i} = Sφ + o(t) =
(N − i)φ + o(t).
Using similar arguments it can be shown that P{I (t +
t) = i − 1|I (t) = i} = iμt + o(t).
[2] S. Eubank, H. Guclu, V. S. Anil Kumar, M. V. Marathe,
A. Srinivasan, Z. Toroczkai, and N. Wang, Nature (London)
429, 180 (2004).
BHUSHAN KOTNIS AND JOY KURI
[3] W. O. Kermack and A. G. McKendrick, Proc. R. Soc. London
A 115, 700 (1927).
[4] A. Vázquez, R. Pastor-Satorras, and A. Vespignani, Phys. Rev.
E 65, 066130 (2002).
[5] M. E. J. Newman, Phys. Rev. Lett. 89, 208701
(2002).
[6] R. Pastor-Satorras and A. Vespignani, Phys. Rev. Lett. 86, 3200
(2001).
[7] M. Boguñá and R. Pastor-Satorras, Phys. Rev. E 66, 047104
(2002).
[8] S. Bansal, J. Read, B. Pourbohloul, and L. A. Meyers, J. Biol.
Dyn. 4, 478 (2010).
[9] L. A. Meyers, B. Pourbohloul, M. Newman, D. M. Skowronski,
and R. C. Brunham, J. Theor. Biol. 232, 71 (2005).
[10] L. A. Meyers, M. Newman, M. Martin, and S. Schrag, Emerg.
Infect. Dis. 9, 204 (2003).
[11] P. Wang, M. C. Gonzalez, C. A. Hidalgo, and A.-L. Barabasi,
Science 324, 1071 (2009).
[12] J. P. Bagrow, D. Wang, and A.-L. Barabasi, PLoS ONE 6, e17680
(2011).
[13] A. Panisson, A. Barrat, C. Cattuto, W. V. den Broeck, G. Ruffo,
and R. Schifanella, Ad Hoc Netw. 10, 1532 (2012).
[14] J. Moody, Social Forces 81, 25 (2002).
[15] C. T. Butts, Science 325, 414 (2009).
[16] B. Song, C. Castillo-Chavez, and J. P. Aparicio, Math. Biosci.
180, 187 (2002).
[17] R. R. Kao, D. M. Green, J. Johnson, and I. Z. Kiss, J. R. Soc.
Interface 4, 907 (2007).
[18] S. Funk, M. Salathé, and V. A. Jansen, J. R. Soc. Interface 7,
1247 (2010).
[19] F. Bagnoli, P. Liò, and L. Sguanci, Phys. Rev. E 76, 061904
(2007).
[20] S. D. Valle, H. Hethcote, J. Hyman, and C. Castillo-Chavez,
Math. Biosci. 195, 228 (2005).
[21] N. Perra, D. Balcan, B. Gonçalves, and A. Vespignani, PloS
ONE 6, e23084 (2011).
062810-10
PHYSICAL REVIEW E 87, 062810 (2013)
[22] J. M. Epstein, J. Parker, D. Cummings, and R. A. Hammond,
PLoS ONE 3, e3955 (2008).
[23] F. D. Sahneh, F. N. Chowdhury, and C. M. Scoglio, Sci. Rep. 2,
632 (2012).
[24] T. Gross, C. D’Lima, and B. Blasius, Phys. Rev. Lett. 96, 208701
(2006).
[25] T. Gross and B. Blasius, J. R. Soc. Interface 5, 259 (2008).
[26] V. Marceau, P.-A. Noël, L. Hébert-Dufresne, A. Allard, and
L. Dubé, Phys. Rev. E 82, 036116 (2010).
[27] S. Risau-Gusman and D. H. Zanette, J. Theor. Biol. 257, 52
(2009).
[28] C. Lagorio, M. Dickison, F. Vazquez, L. Braunstein, P. Macri,
M. Migueles, S. Havlin, and H. Stanley, Phys. Rev. E 83, 026102
(2011).
[29] P. Holme and J. Saramaki, Phys. Rep. 519, 97 (2012).
[30] N. Perra, A. Baronchelli, D. Mocanu, B. Gonçalves, R. Pastor-
Satorras, and A. Vespignani, Phys. Rev. Lett. 109, 238701
(2012).
[31] N. Perra, B. Gonçalves, R. Pastor-Satorras, and A. Vespignani,
Sci. Rep. 2, 469 (2012).
[32] M. Starnini, A. Baronchelli, A. Barrat, and R. Pastor-Satorras,
Phys. Rev. E 85, 056115 (2012).
[33] L. E. Rocha, F. Liljeros, and P. Holme, PLoS Comput. Biol. 7,
e1001109 (2011).
[34] L. J. S. Allen, in Mathematical Epidemiology, Lecture notes
in Mathematics, edited by F. Brauer, P. Driessche, and J. Wu,
Vol. 1945 (Springer, Berlin, Heidelberg, 2008), pp. 81–130.
[35] M. Pinsky and K. Samuel, An Introduction to Stochastic
Modeling, 4th ed. (Academic Press, New York, 2011).
[36] I. Nasell, Adv. Appl. Prob. 28, 895 (1996).
[37] I. Nassel, Math. Biosci. 156, 21 (1999).
[38] D. Clancy and S. Mendy, Methodol. Comput. Appl. Prob. 13,
603 (2011).
[39] E. G. Leigh, Jr., J. Theor. Biol. 90, 213 (1981).
[40] F. Sauvage, M. Langlais, N. G. Yoccoz, and D. Pontier, J. Anim.
Ecol. 72, 1 (2003).