Mathematical Biosciences 244 (2013) 125–134

Contents lists available at SciVerse ScienceDirect

Mathematical Biosciences
journal homepage: www.elsevier.com/locate/mbs

The scaling of contact rates with population density for the infectious
disease models q
Hao Hu ⇑, Karima Nigmatulina, Philip Eckhoff
Epidemiological Modeling (EMOD) Group, Intellectual Ventures Laboratory, 1555 132nd Ave. NE, Bellevue, WA 98005, USA

a r t i c l e i n f o a b s t r a c t

Article history: Contact rates and patterns among individuals in a geographic area drive transmission of directly-trans-
Received 31 August 2011 mitted pathogens, making it essential to understand and estimate contacts for simulation of disease
Received in revised form 28 April 2013 dynamics. Under the uniform mixing assumption, one of two mechanisms is typically used to describe
Accepted 30 April 2013
the relation between contact rate and population density: density-dependent or frequency-dependent.
Available online 9 May 2013
Based on existing evidence of population threshold and human mobility patterns, we formulated a spatial
contact model to describe the appropriate form of transmission with initial growth at low density and
Keywords:
saturation at higher density. We show that the two mechanisms are extreme cases that do not capture
Contact rate
Epidemic model
real population movement across all scales. Empirical data of human and wildlife diseases indicate that
Transmission scaling a nonlinear function may work better when looking at the full spectrum of densities. This estimation can
Population density be applied to large areas with population mixing in general activities. For crowds with unusually large
Crowd dynamics densities (e.g., transportation terminals, stadiums, or mass gatherings), the lack of organized social con-
tact structure deviates the physical contacts towards a special case of the spatial contact model – the
dynamics of kinetic gas molecule collision. In this case, an ideal gas model with van der Waals correction
fits well; existing movement observation data and the contact rate between individuals is estimated
using kinetic theory. A complete picture of contact rate scaling with population density may help clarify
the definition of transmission rates in heterogeneous, large-scale spatial systems.
Ó 2013 The Authors. Published by Elsevier Inc. All rights reserved.

1. Introduction pre-vaccination era, when a minimum population threshold called

Critical Community Size (CCS) affected disease persistence or
For directly transmitted diseases, contact patterns drive the fade-outs in the United Kingdom [18]. In turn, studies explored a
spread of infectious pathogens in both space and time. Recent number of scaling mechanisms in dynamic epidemic models
data-driven, detailed, large-scale spatial models and contact exper- to characterize the relation between transmission rate and
iments have improved our understanding of contact networks in population [19–24].
different social settings, as well as the corresponding spread At the other end of the density spectrum, if a location has extre-
patterns of infectious diseases [1–15]. However, the generalized mely high densities in a local area (for example in transportation
scaling rules of host contact patterns over population size or terminals, stadiums, or mass gatherings), it might pose high risks
density still need refinement. In epidemic models targeting a large that may facilitate disease spread, and recently a research agenda
heterogeneous area, it becomes important to set the appropriate has been proposed to explore the role of mathematical modeling
transmission rates to different local populations. in assessing health risks associated with mass gatherings [25].
The population effect on transmission, e.g., threshold of patho- However, in such settings, the contact rates between individuals
gen invasion and persistence [16], has frequently been examined remain unexplored and it is unclear whether a scaling mechanism
as a way to differentiate pathogen transmissibility in geographic relationship exists between transmission rates and population
locations. Stochastic effects [17] make the invasion threshold density. If a spatial epidemic model needs to include such locations
difficult to observe. However, the persistence threshold has and quantitatively assess the risk of disease outbreak, it might not
successfully been identified in measles endemics during the be appropriate to extrapolate the rules under general mixing to
crowds. In this case, people have very low mean free path and do
not have the ability to contact everyone else. Understanding the
q
This is an open-access article distributed under the terms of the Creative contact patterns in the extremely high density range is important,
Commons Attribution-NonCommercial-No Derivative Works License, which per-
and appropriate scaling of contacts and transmissibility needs to
mits non-commercial use, distribution, and reproduction in any medium, provided
the original author and source are credited.
be applied.
⇑ Corresponding author. Tel.: +1 425 691 3336. We aim to use simple physical models to explore the complete
E-mail address: [email protected] (H. Hu). spectrum of contact rate scaling with population density. A spatial

0025-5564/$ - see front matter Ó 2013 The Authors. Published by Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.mbs.2013.04.013
126 H. Hu et al. / Mathematical Biosciences 244 (2013) 125–134

model is used to construct a function which monotonically per-individual rate contact rate
increases with density, but saturates as density increases. We show 2
that the acquired density scaling function is comparable to a num- 0.002

rate
ber of physical models in different activity range limits, such as ki- 1
netic collision theory and homogeneous mixing. At the high 0.001
population density limit, when looking at the mixture of crowds,
0 0
it is possible to approximate crowd motion as random movements, 0 5×10
2
1×10
3 0 5×10
2
1×10
3

using a corrected model from kinetic collision theory. It is a special

example of the density scaling function to capture the change of 0.002 2
contact rates at the high density limit. A number of empirical

rate
examples are presented to illustrate the use of the density scaling 0.001 1
function: in low-density conditions, this scaling function is fitted to
a number of human and wildlife diseases: the mortality data of the 0 0
0 5×10
2
1×10
3 0 5×10
2
1×10
3

1918 influenza outbreak in the United States and England & Wales,
home range and transmission threshold density of red fox rabies, 0.002 2
contact rates in brushtail possums and voles for the transmission

rate
of bovine tuberculosis and cowpox. At the high density limit, the 0.001 1
contact rate fits well with observations from different mass gather-
ing events. 0 0
0 5×10
2
1×10
3 0 5×10
2
1×10
3

density density
2. Contact rate scaling functions
Fig. 1. Different scaling functions of transmission rates with population density: (A)
2.1. Density and frequency dependent mechanisms density-dependent. (B) frequency-dependent. (C) Mechanistic nonlinear scaling
function. For comparison, all plots have the same ranges. The area is assumed to be
2
In a homogeneously mixed, single location SIR model with 10 km . The reproductive number for frequency-dependent function is chosen as
R0 ¼ 2:0 (red) and R0 ¼ 1:5 (blue). For density-dependent function, we assume a
population N inside area A (S; I, and R representing the number 2
population threshold of 100 individuals=km (103 individuals), therefore the
of susceptible, infected, and recovered individuals respectively), constant per-individual rate is calculated using R0 ¼ 1 when density =
the per capita transmission rate b controls the rate of disease 2
100 individuals=km . For nonlinear scaling function, the saturation values are
spread. b is the product of infection probability p, and individual chosen the same as R0 for direct comparison. (For interpretation of the references to
colors in this figure caption, the reader is referred to the web version of this article.)
contact rate C, therefore b ¼ pC. Since the proportion of infec-
tious contacts is NI , the rate of increase for infectious individuals
2.2. The nonlinear transmission scaling function
is dI
dt
¼ bS NI ¼ pCSI
qA , given N ¼ qA. Homogeneous mixing assump-
tion results in a constant contact rate C within a local popula- Based on existing evidences, it is possible that neither
tion, and it depends on the population density: C  CðqÞ [23]. mechanism may adequately describe the correct scaling over the
C is typically determined by density- or frequency-dependent entire population density spectrum. Therefore, a nonlinear func-
mechanisms, both are widely used and are applicable to a range tion is usually preferred [16,29,30], such as the Holling Type 2 rela-
of infectious diseases [19,21]. Density-dependent transmission is tionship in ecological predator–prey models [31,32] or a power
also called ‘pseudo’ mass action [22]. Borrowing from the law of function with population [30,33]. These non-linear functions have
mass action in chemistry, it is analogous to the scaling between the expected features on both limits, and they continuously shift
reaction rates and reactant concentrations. With this approach, with host density. A number of mathematical forms have been
the contact rate scales in a linear fashion with population den- explored [19–21].
sity (or size); C ðqÞ ¼ c0 q, where c0 is a constant, and Here, we aim to use a spatial contact model with constant pop-
dI
dt
¼ pcA0 SI  cSI assuming c is a constant. The per-link contact ulation density and limited activity range to estimate contact rates.
rate is a constant c0 . Estimating the density-dependent contact Within a local population, individuals move around and contact
rates from collision frequency in the ideal gas model has the with each other randomly, and a spatial dispersal kernel controls
same result (See Section 2.3). Frequency-dependent transmis- the decay behavior of contact rate. Since the decay usually depends
sion, also called ‘true’ mass action [22] or proportionate mixing on both density and activity range, a dispersal kernel function uq ðrÞ
[26], assumes a constant contact rate in the local population in- belonging to exponential power distribution family is selected in
dependent of population size or density, C ¼ c1 , so order to calculate the contact rate within the activity range.
dI
dt
¼ pc1 S NI  c SIN. In this case, the per-link contact rate, cN1 , decreases We assume a land area with uniform population density q, and
with larger density. Fig. 1(A) and (B) show examples of transmission each individual has a constant fraction kð0 < k < 1Þ of contacts
scaling with population density under the above two mechanisms. among the calculated effective population. Physical contact of each
Choice of mechanism makes no difference when investigating a individual is confined to a maximum activity range rmax . Contact
single fixed size population as the parameters can be calibrated rate C within the reachable area is calculated by the spatial kernel
appropriately, but patterns of disease spread will differ when mod- uq ðrÞ, density q and maximum activity range rmax :
eling a large heterogeneous spatial area. The reproductive number Z Z r max
R0 for frequency-dependent transmission is fixed regardless of C ðqÞ ¼ uq ðrÞdrdh ¼ uq ðrÞ2prdr
0
population density, and therefore, does not have a population inva-
sion or persistence threshold for the given parameters because A number of disease transmission models include a distance-
R0 > 1. For density-dependent transmission, the actual reproduc- based transmission kernel to capture population movement and
tive number depends on the population size and an invasion the effect of spatial clustering. For instance, previous research shows
threshold. The mechanism, however, assumes indefinite linear examples of the functional decay based on daily travel distances, in
contact rate growth with the population size, and it is questionable this case a spatial kernel is included to calculate the risks of commu-
whether an individual has more contacts in a population of 106 nity infection [3,4]. Other models, such as that of foot-and-mouth
than 105 [24,27,28]. disease (FMD) in the United Kingdom in 2001, proves that a
 H. Hu et al. / Mathematical Biosciences 244 (2013) 125–134 127

distance-explicit exponential decay kernel captures the progression is easy to get the limits of both extreme conditions using the above
of FMD in a heterogeneous space [34]. Furthermore, studies of con- function: given q0 , the contact rate will approach density and fre-
2
tact rate heterogeneities within populations have demonstrated quency dependent limits at q ! 0 (C ðqÞ  pkr max q, and therefore,
that these can also alter transmission dynamics [35,36]. Using data C ðqÞ  q) and large q(C ðqÞ ! pkq0 ), respectively. Fig. 1(C) has an
from either dollar bill circulation or cell phone locations as proxies example describing the behavior of this nonlinear function form,
[37–39], population-level movement studies describe the quantita- and show the contact rate satisfies the two existing mechanisms
tive scaling of people’s movement patterns with distance. All of in different density limits.
these insights may help facilitate an appropriate form of spatial ker- A simulation of endemic disease die-outs illustrates the impact
nel uq ðrÞ to describe the pattern of transmission. of different choices of density scaling mechanisms, including the
We define an exponential power distribution with limited non-linear scaling function. This agent-based, stochastic, dis-
activity range to estimate how contact rates change with density crete-time simulation is performed with 100,000 individuals, but
2
and distance. In the function, contact rate decays with distance, with different densities between 10 and 200=km (area ranges
2
and the relationship between home range and population density from 500–10,000 km ) to represent a broad range of places from
tends to be inverse. Therefore, we start with the generalized decay sparsely-populated areas to crowded metropolitans. For each se-
form of spatial kernel with density and distance dependence, lected density, 10 random simulations are executed using a SEIRS
 a
c q
q
rb
model, and we apply the repetitive infected case importation rates
uq ðrÞ ¼ kqe . After integrating the spatial kernel to
0
(5 per 60 days) to all scenarios. We run the simulation for 730 days
calculate the contact rate, we have: and calculate the length of disease die-outs, i.e. the fraction of time
 a when no new case is reported. Latent and infectious period is set to
Z r max
c q
q
rb 3 days and to 4 days, respectively. People lose disease-specific
C ðqÞ ¼ kqe 0
2prdr immunity after 60 days on average, and at R0 ¼ 1:5, the disease
0
will remain endemic. Saturated value of R0 for the non-linear scal-
A numerical study is performed on different a and b. Fig. 2 ing function is set to 1.5 and we set the same threshold density
shows the effect of different a on the contact scaling function un- 2
(R0 ¼ 1 when qthreshold ¼ 32:96=km ) to both density-dependent
der different pre-selected b values. Under all conditions, the scaling and non-linear scaling relationships. In Fig. 3, it is apparent to
function approaches a constant value only when b ¼ 2a is satisfied. see the limits of both frequency- and density-dependent transmis-
Therefore, in order to have a saturated contact rate at high density, sion, while the non-linear scaling relationship scales nicely in a
the exponential powers on density and distance must satisfy: wide spectrum of densities. For frequency-dependent transmis-
 a
 q
q
r 2a
sion, very few die-outs are observed, because R0 remains constant.
uq ðrÞ ¼ kqe 0
Both density-dependent and non-linear scaling transmission are
able to show the effect of threshold density when R0 < 1, however,
a (a > 0) defines the scaling laws of different physical movement for density-dependent transmission, R0 will increase to unrealisti-
models. Choosing a ¼ 1 corresponds to a distribution of distance cally high values when density is large, in this case disease dies out
in Gaussian form; it represents the physical process of random by itself because all susceptible individuals are depleted quickly.
walk. For different a, the integration of the spatial kernel gives sim- In the above discussions, we assume that spatial kernel u de-
ilar saturation function forms as shown in Fig. 2. q pends on both density q and distance r, and r max is a constant. If u
 r2
When a ¼ 1, spatial kernel density becomes: uq ðrÞ ¼ kqe q0 . only depends on r, the same scaling functions and results can be
q0 is the characteristic density constant, which is determined by achieved by assuming a density dependent activity range rmax ðqÞ,
the type of disease. Integrating the contact rate function, we obtain: where the maximum home range decreases with greater density.
 q
r 2
CðqÞ ¼ pkq0 1  e max q0
2.3. Maximum range, width of spatial kernel and kinetic models
In the following discussions, we use the above form for kernel
density function as an example. A nonlinear function approximates Both dispersal kernel function and maximum range act together
density and frequency dependent mechanisms at low- and high- to create a mechanistic model of contact dynamics for both human
density ends in the general population mixing range, respectively: social networks and wildlife dynamics. In human social networks,
 q
daily life involves a certain number of human-to-human contacts
r 2
C ðqÞ ¼ pkq0 1  e max q0  c1 ð1  ec2 q Þ which saturates at high density. People have greater rates of con-
tacting individuals near them spatially, but still contact individuals
where c1 and c2 are constants. As density increases, this functional who live further away. In lower-density areas, people will travel
form results has initial growth, and it is saturated when q is high. It further on average per-contact, but there is a maximum range for

 a
q
c q rb
Fig. 2. Numerical Studies on different exponents of density and distance in the spatial kernel uq ðrÞ ¼ kqe 0
. b is pre-selected, and we show the effect on the saturation
of the scaling function when selecting different a. In all cases, the surface and contour plot on q and a is reported. Only when a ¼ 12 b, the function approaches a constant at
high density. The example uses rmax ¼ 5; q0 ¼ 10, and k ¼ 1. The results were calculated with different distance decay exponent: (A) b ¼ 0:8 , (B) b ¼ 1, (C) b ¼ 2.
128 H. Hu et al. / Mathematical Biosciences 244 (2013) 125–134

2.5 1

fraction of time for disease die-outs

A B
basic reproductive number
2 0.8

1.5 0.6

1 0.4
non-linear scaling function
0.5 frequency-dependent 0.2
density-dependent
0 0
0 50 100 150 200 0 50 100 150 200
density (people/square km) density (people/square km)

Fig. 3. Effect of different contact rate scaling functions in a SEIRS model. (A) Three scaling functions used in the simulation: density-dependent, frequency-dependent and
2
non-linear scaling function. Threshold density was set as qthreshold ¼ 32:96=km for both density-dependent and non-linear scaling functions, and saturated R0 ¼ 1:5 is used for
both frequency-dependent and non-linear scaling functions. (B) fraction of time with disease die-outs in the simulation. The plot shows results from 10 random stochastic
simulations for each density. The average waning time of immunity is 60 days, so that the disease will remain endemic at R0 ¼ 1:5. 5 infected cases are seeded every 60 days
and we run the simulation for 730 days. For other disease-specific parameters please refer to the main text.

effective routine contacts. In wildlife populations, higher densities 2.4. The contact rate scaling of crowds
decrease the effective area per animal, and the density kernel
shrinks in width. Animals are limited by mobility and thus at Local public sites with extremely high population density, such
low densities, the maximum effective range limits contact rates as train stations or large social, political, or religious mass gather-
and contact rates transition to a density-dependent regime. ings are regarded as high-risk for respiratory disease transmission
The activity range rmax represents the proximity of practical due to extended close contact between hosts [41]. The Islamic pil-
contact distances for each individual. It is also a parameter to grimage Hajj brings more than 2 million pilgrims a year to Mecca,
control the pattern of spread: whether it is well-mixed or Saudi Arabia, where densities reach about 7 people /m2 walking
wave-like, as these patterns have different rules of transmission around the Kaaba. This raises major concerns about the risk of
[40]. When r max ! 1; C ðqÞ ¼ pkq0 describes a well-mixed mod- respiratory diseases, like the 2009 H1N1 influenza outbreak [42].
el, every individual is able to contact everyone else. In this case, Enhanced disease surveillance and public health control measures
transmission is controlled by the overall contact rate, regardless have been performed at previous major events [43–46], but the
of population density. The interpretation of this limit is that peo- general epidemiology at mass gatherings remains poorly under-
ple have a constant contact rate regardless of population density, stood [47]. Mathematical modeling have potential applicability in
and travel as far as necessary to achieve that contact rate. When assessing effect of public health policies and assist policy planning
2
rmax ! 0; C ðqÞ ¼ pkrmax q, the spread pattern becomes wave-like, and strategic decision making[25]. Simulating these dynamics re-
which is controlled by the network connectivity with nearest quires better understanding of contact patterns and the appropri-
neighbors. Therefore, it scales directly with density, assuming ate adjustment of transmission rates.
that connectivity scales to all neighbors within this minimal The laminar movement of pedestrians can break down under
effective distance. In this case, this wave-like behavior is well high densities, resulting in self-collective behaviors, such as stop-
demonstrated by the ideal gas kinetic model. When population and-go waves and turbulence [48,49]. Existing studies use a variety
density is within the general population mixing range, the vol- of experiments and models, such as Newtonian physics [50,51] or
ume of individual is negligible compared to the mean free path, cognitive heuristics [48], to study the patterns of pedestrian flows
and we assumed the container was large enough so that each and crowd dynamics. In the disease modeling context, we are more
individual had no restrictions. Therefore, the average velocity v interested in calculating the contact rate of each individual. Sophis-
remains constant. Since the relative velocity of two particles is ticated mathematical expressions require large numbers of inputs
pffiffiffi for calibration [48]. Therefore, for simplicity we assume that peo-
2v , the average frequency of collision C, or the average number
pffiffiffi 2 ple are more likely to move and contact others randomly.
of contacts per unit of time can represented as: C ¼ 2pd qv ,
As people are only able to contact others nearby, the activity
where d is the collision diameter and the cross-sectional area
range r max , or the width of the spatial kernel, is close to 0, therefore
2
is pd . In this formulation, when density is low, the contact rate contact behavior leans towards a kinetic gas model, and uniform
C has linear dependence with density q. mixing assumption is no longer appropriate for modeling crowded
The kinetic model assumes a random movement with colli- areas. However, when q is extremely high, the assumption will be
sion between particles (or individuals), rather than population different from the ideal gas kinetic model discussed in Section 2.3.
movements with clustering or social organizations. This assump- Here, average speed of people is expected to be slowed down un-
tion does not apply when modeling human societies in a large der high densities, representing the impeded pedestrians in den-
spatial scale, since individuals form social groups and the daily sely packed crowds. Empirical measurements of crowd motion
trajectories are repetitive and highly predictable [39]. If physical follow Maxwell–Boltzmann’s theory [52,24]. We propose that the
contact behaviors were random, the kinetic model would be short, frequent pedestrian contact process is analogous to the col-
more applicable, for example, short and random contacts be- lision process using the crowd gaseous model. The relationship be-
tween people at public sites. However, corrections to the model tween average speed and population density is obtained in this
needs to be made under special conditions, e.g. when population case. Besides, the van der Waals correction may need to be consid-
are overcrowded in limited spaces. We explore these limits and ered, because the volumes of molecules and inter-molecule inter-
corrections in more details, and will discuss them in the next actions are not negligible any more. In the model, we consider
section. individuals as molecules in a restricted container with volume V.
 H. Hu et al. / Mathematical Biosciences 244 (2013) 125–134 129

Similar physical properties of gas/liquids can be used with differ- influenza pandemic in the United States, England, and Wales. The
ent meanings: temperature T reflects the energy mode or speed 1918 influenza pandemic is one of the more recent, well-docu-
of individuals, and pressure P represents the pressure of individu- mented global pandemics with detailed records on cities and
als going out of the system. towns. For United States, state-level cumulative mortality rates
Therefore, if the molecular weight is defined as a constant M, for influenza and pneumonia, population, and area information
the equation of state can be written as: are obtained from historical records [53] and 1920 Mortality Statis-
   tics [54]. Additional state reports on agriculture, industry, and min-
q2 a M ing come from the 1910 U.S. Census [55]. For England and Wales,
Pþ 2  b ¼ RT
M q large town and city-level influenza mortality data (population >
P is assumed to be constant (i.e., the overall momentum for 20,000) have been recorded by the UK government and made avail-
individuals going out of the system does not change). Using the able to the public [56,57]. Demographic information for population
qffiffiffiffiffiffi and area is from the 1911 National Census [58]. We use mortality
root mean square speed v ¼ 3RT M
and merging all constants, we
records during the second and third peaks, which are the two larg-
have the following relationship between v and q: est waves.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Invasion of an infectious disease pathogen into a local popula-
1 þ c 2 q þ c 3 q2 þ c 4 q3
v ¼ c1 tion is a two-stage process: initial importation from migration be-
q tween local populations, and subsequent increases in the number
qffiffiffi
of infected individuals due to contacts within the local population.
For ideal gas, using P ¼ qRT, the speed is v ¼ c1 q1 .
After the initial seeding, secondary spread of disease is driven
Comparing between the two models, at the numerator, the first
mainly by the force of infection among local contacts. For the
order of density is due to the size of molecules. The second and
1918 influenza pandemic, we investigate the local infection stage
third orders of density are introduced by the interaction potential
by looking at the cumulative influenza (or influenza and pneumo-
between people. In contrast to gas molecules, the inter-person
nia for the United States) mortality rate for each geographic loca-
potential is small enough to be ignored, even at the large density
tion. Mortality is used as a surrogate for the final size of the
limit. Therefore, the following equation describes the quantitative
outbreak, which scales with the average reproductive number.
relationship between people’s movement and population density:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi We do not calculate the exact reproductive number, but look for
1 þ c2 q scaling of transmissibility due to population effects. We assume
v ¼ c1 that the probability of infection, given contact with an infected
q
case, and the duration of the infectious period are the same for
Note that the density needs to be smaller than  c12 , the hard limit all individuals, and the contact rate is only related to population
when all molecules fill in the space. density. Under these assumptions, the reported mortality value di-
In densely-packed crowds, the average speed of people is rectly scales with contact rate.
slower and transmission potential is greatly reduced due to the Population density is calculated using the total area of the
decreasing frequency of contacts. Furthermore, the spread of dis- investigated region, therefore each region is expected to have uni-
ease is diffusive since people only contact nearby individuals and form population density with people living in the majority of areas.
there are no ‘‘small-world’’ links between contacts. Chemical reac- For this reason, influenza mortalities in U.S. states with more than
tions between molecules provide a good analogy to infectious dis- 5% of the population in mining occupations are reported separately
ease transmission. The transmission rate is similar to the reaction and not included in parameter fitting. In mining states, the popula-
rate, depending on collision frequency. tion is likely to form aggregated clusters rather than be spread uni-
According to results from kinetic theory, the contact rate can be formly across the state, making the actual population densities far
calculated as follows: from the estimated values. This effect is less obvious in non-mining
pffiffiffi 2 pffiffiffi 2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi states.
C¼ 2pd qv ¼ 2pd c1 qð1 þ c2 qÞ Fig. 4(A) shows the relationship between state-wide influenza
mortality rates and population densities. Generally, the low-density
states have lower reported mortality rates, while the rate is close to
3. Empirical examples
constant for higher density states. Our estimates from the spatial
kernel fit the overall trend better than the frequency-dependent
3.1. Human diseases
approach: for non-mining states, the best parameters are c1 ¼
6:55; c2 ¼ 0:04ðv2 ¼ 29:6Þ, compared with the frequency-depen-
To observe if real disease outbreaks have similar scaling with
dent fit, c ¼ 5:68 (v2 ¼ 39:3). Of note, the mining states have a
population density, we assess the transmissibility of the 1918

10 10 0.016
mortality rate per 100,000

mortality rate per 100,000

A B best fit C
8 8
total prevalence

0.014
6 6 0.012
4 best fit: non-linear scaling func. 4
best fit: frequency-dependent
0.01
2 non-mining state 2
mining state 0.008
0 0
0 100 200 300 400 500 600 0 5×10
4
1×10
5 0 5×10
4
1×10
5

density (people / square mile) density (people / square mile) density (people / square mile)

Fig. 4. The scaling between contact rate and density in general activities. (A) The relationship between population density and 1918 influenza and pneumonia mortality for
the United States, grouped by mining and non-mining states, with the best fit for frequency dependent and the nonlinear scaling function for non-mining states only.
Mortality and demographic data are obtained from [53]. (B) The relationship between population density and 1918 influenza mortality for cities and towns in England and
Wales. (C) Stochastic variation of total prevalence produced by the SEIR simulation, using the same size and density for cities and towns in England and Wales.
130 H. Hu et al. / Mathematical Biosciences 244 (2013) 125–134

higher mortality rate than non-mining states with comparable den- functions as shown above. For direct comparison, we use the same
sity, indicating the possible effects of overcrowding and highly het- population sizes and densities as the England and Wales cities data
erogeneous, scattered population clusters. in 1911. Ten stochastic simulations were performed for each pop-
For England and Wales, the mortality across cities and towns ulation size and density. In Fig. 4(C), the results are presented, and
appear to be frequency-dependent, as the best fit of the nonlinear they show a similar trend in incidence variance, which decrease
scaling function gave a form that was already saturated in the den- with population size.
sity range. This outcome is in accordance with findings from prior
studies of the England and Wales data [59], as well as those of large 3.2. Wildlife diseases
cities in the United States [60]. A prior report on the reproductive
number of measles in large England and Wales cities also suggests For small populations, scarcity of data makes it hard to compare
frequency-dependent transmission, which is independent of popu- outcomes against other human diseases. Therefore, we apply our
lation size [61]. Cities and towns in England and Wales have much scaling function to certain wildlife diseases where data is relatively
larger population densities when compared to states in the U.S, abundant.
and it might explain why the cumulative influenza mortality rate Rabies epizootic spreads steadily across Europe in the last cen-
is mostly independent of population density (Fig. 4(B)). The 1918 tury, mostly among red foxes (Vulpes vulpes). Various field studies
England and Wales influenza records lack original data for small have been carried out to investigate its spatial-temporal spread
towns and rural areas. The characteristic density constant q0 for [62]. There have been substantive efforts in mathematical
influenza and measles may be low: cities and towns with popula- modeling to study rabies virus transmission dynamics, in order
tions under 20,000 typically have densities of less than 500 people to provide insights on spatial invasion as well as effects of inter-
per square mile (the saturation region for the nonlinear scaling ventions, such as culling or vaccination. Based on early observa-
function obtained from the United States). Therefore, it is possible tions of the threshold density – below which the endemic fails to
that existing data only show the saturation of transmissibility in sustain, most models assume a density-dependent transmission
larger towns and do not speak to the dynamics in lower density rate.
rural areas. Better data would be required for validation of the non- From multiple empirical studies, a negative correlation exists
linear function in the England and Wales context. between fox density and home range size. Fig. 5(A) demonstrates
Large fluctuations at low population densities are likely due to the combined results of the two studies in [62,63]. This evidence
stochastic fluctuations and the small number statistics. To illus- suggests that contact rates among animals are largely - but not
trate this effect, we perform a agent-based, discrete, stochastic entirely - determined by population density [62]. When fox
SEIR simulation with R0 ¼ 2:1, using the fitted contact rate density is high, the contact model may be different from linear

red fox, rabies red fox, rabies

2.5 2
A data from Holmala et al. (2006) B
basic reproductive number

data from Trewhella et al. (1988)

home range radius (km)

2
best fit from non-linear scaling function 1.5

1.5
1
1

0.5
0.5 threshold density: 0.19

0 0
0 1 2 3 4 5 0 0.5 1 1.5 2
density ( /square km) density (/square km)

brushtail possums, bovine tuberculosis voles, cowpox

0.25 80
C D
-1

0.2
contact rate males h

60
contact rate

0.15
40
0.1
male-male best fit nonlinear scaling func
20 power scaling function, Smith et al. (2009)
0.05 male-female best fit nonlinear scaling func
male-male best fit: non-linear scaling function
male-female
0 0
0 5 10 15 20 25 30 0 200 400 600 800 1000
-1 -1 population
males ha home range

Fig. 5. Density scaling function in wildlife diseases. (A) empirical data between fox home range and density relationship, and the best fit values using the spatial scaling
function. Home range radius (km) are calculated from home range (hectare or ) assuming the area is circular. (B) the scaling of basic reproductive number and the threshold
density, using best fit values in (A). (C) fitting the scaling function to mating contact rates among brushtail possums. (D) The comparison between the scaling function
C  N a ða ¼ 0:38Þ obtained for voles, for transmission of cowpox in [30]. The best fit for the non-linear scaling function in the same data range. The dashed lines show the
extrapolated curve when population is large.
 H. Hu et al. / Mathematical Biosciences 244 (2013) 125–134 131

density-dependent behavior and transmission rate does not in- Spatial scaling function is also fitted to contact rates between
crease linearly when density is large. brushtail possums when looking at bovine tuberculosis transmis-
The relationship between fox density and home range radius is sion[69]. The best fits (c0 ¼ 0:26; c1 ¼ 0:056 for male–male contact
in line with the spatial contact kernel. Therefore we fit the kernel and c0 ¼ 0:165; c1 ¼ 0:116 for male–female contact, see Fig. 5(C))
using reported field values of home range and density from both successfully replicates the non-linear convex-up relationship from
studies and derive the density scaling function. We assume home the obtained field data, which implies the contact rate during the
range distance for an individual is the radius where 50% of contacts breeding season does not decrease in proportion to reductions in
happen. The best estimates for the two constants: q0 and rmax are density [69]. The authors suggest this phenomenon is probably
acquired. Fig. 5(A) shows the best fit of the spatial contact kernel due to the increased male overlap of female ranges following the
(q0 ¼ 0:99 and r max ¼ 2:34), and Fig. 5(B) demonstrates the shape density reduction, which aligns well with the assumptions of the
of non-linear scaling function which saturates at R0 ¼ 1:5 . spatial kernel.
To test the validity of the obtained function, we calculate the Another example involves the transmissibility of cowpox in
threshold density of rabies in Europe. If the saturated R0 is assumed wildlife populations. In [30] the authors report a power function
as 1:5, Fig. 5(B) shows the scaling function and the corresponding C  N a ða ¼ 0:38Þ that represents the scaling between contact
2
threshold density, which is 0:19=km . This threshold density is rates and population. We compare this function with the best fits
consistent with estimated value of 0.2 from [64], and is close to using our density scaling function in Fig. 5(D) (c0 ¼ 38:43; c1 ¼
field observations from Europe [62]: 0.63 in central Europe [65], 0:018). Although both functions provide similar scaling results
as low as 0.25 [66], 0.3 in southern parts of western Siberia [67]. when the population is small, our non-linear scaling function gives
Compared with the linear relationship, because the non-linear potentially more realistic results for the saturation of infectivity
scaling function is convex-up, under the threshold density, R0 value when extrapolating to large populations.
obtained from the non-linear scaling function decays slower than
the one obtained from the density-dependent scaling function. 3.3. Contact rates of crowds
This confirms the finding that comparing with vaccination, culling
might not be effective in stopping the endemic transmission and Using existing pedestrian movement study, the relationship
achieve local eradication. In fact, field experiences from Europe between average speed of people and population density is fitted
suggest that the use of oral rabies vaccination is much more effec- into the state equation. Fig. 6(A) demonstrates the best fit with
tive and there are only very few examples where culling alone have the empirical passenger walking speed and crowd density data
either eliminated the disease or have prevented its spread to pre- obtained from a Japanese railway station [70], with
viously uninfected areas [68]. The spatial contact kernel explains c1 ¼ 1:30; c2 ¼ 0:137; v2 ¼ 0:005. As a comparison, we also re-
this nicely: as density decreases with culling, the width of each re- qffiffiffi
maining individual’s spatial contact kernel increases. This main- port the fitting results for the ideal gas model ¼ c1 q1 :
tains a consistent level of contacts until the resulting spatial c1 ¼ 1:07; v2 ¼ 0:205. The van der Waals state equation model fits
kernel width begins to increase beyond the mobility limiting r max . the data more closely. In addition, the maximum possible density

2 8
A B
van der Waals gas model
ideal gas model
Average Speed (m/s)

1.5 6
Contact Rate

1 4

0.5 2
x=7.29

0 0
0 2 4 6 8 0 2 4 6 8
density (people / square meter) density (people / square meter)

2 1.5
C van der Waals gas model (ρ<6) D van der Waals gas model
van der Waals gas model (all) ideal gas model
Average Speed (m/s)

Average Speed (m/s)

1.5 ideal gas model

1

0.5
0.5

0 0
0 2 4 6 8 10 0 1 2 3 4 5 6
density (people / square meter) density (people / square meter)

Fig. 6. Fitting the van der Waals state equation to population movement data. (A) Best fit between the pedestrian speed and population density, data were obtained from a
Japanese train station, reported in [70]. The ideal gas model fitting result is present as well. (B) The estimation of collision frequency based on population density, using the
fitted van der Waals state equation and value for close-contact distance. (C) and (D) Fitting the van der Waals gas model to two other empirical crowd movement studies: (C)
Hajj and (D) movements on a plane, respectively [49,71], and comparing the results to the best fits obtained for the ideal gas model.
132 H. Hu et al. / Mathematical Biosciences 244 (2013) 125–134

Table 1 change significantly during the epidemic period. The density is cal-
Best fit values for several empirical crowd movement studies, using van der Waals culated using total area rather than actual living area, which may
and ideal gas state equations.
have biased the results. In the United States, we exclude several
qffiffiffiffiffiffiffiffiffiffiffi qffiffiffi
Study Van der Waals model, v ¼ c1 1þc2 q
Ideal gas, v ¼ c1 1
q
mining states likely to have highly heterogeneous distributions of
q
population with spatial clusters. This aligns with the observation
Train station [70] c1 ¼ 1:30; c2 ¼ 0:137ðv2 ¼ 0:005Þ c1 ¼ 1:07ðv2 ¼ 0:205)
that their dynamics more closely resemble those of more den-
Hajj [49] c1 ¼ 1:01; c2 ¼ 0:101ðv2 ¼ 1:005Þ c1 ¼ 0:714ðv2 ¼ 3:492)
sely-populated states, rather than similarly sparsely-populated
Hajj [49] (q < 6Þ c1 ¼ 1:22; c2 ¼ 0:133ðv2 ¼ 0:199Þ c1 ¼ 0:915ðv2 ¼ 0:727)
Plane [71] agricultural states. It is possible that other states with similar het-
c1 ¼ 0:99; c2 ¼ 0:184ðv2 ¼ 0:063Þ c1 ¼ 0:689ðv2 ¼ 0:337)
erogeneous clusters may not be fully excluded.
The nonlinear scaling function applies mainly to diseases with
for the van der Waals state equation ends at 7:3 =m2 , close to the direct transmission, where the density is one-dimensional. For vec-
reported maximum density value 7:4 =m2 in the literature [70]. tor-borne diseases, vector density and human vector ratio are
In comparison, the ideal gas model does not have density cut-offs. important, and these factors complicate the scaling relationships.
Data from other movement studies in a variety of settings re- For instance, historical records for the flea-borne plague (Black
veal a similar pattern. Results are reported in Table 1. Observed Death) show that mortality rates in Italy and France were highest
data from the Hajj has population densities up to 10 =m2 [49]. in sparsely-populated areas in the countryside rather than in cities
Since reaching this density limit seem to require collective human and towns. Even the rates in the largest cities, with more than
movement patterns as opposed to random movement, we perform 100,000 people, weren’t comparable [73]. The pathogen was trans-
a separate fit considering only data points from q < 6 =m2 . For all mitted by rat fleas to human beings. In the countryside, a rat col-
scenarios, the decreasing rates and cutoff values are similar, but ony cohabited with a household, while in urban areas several
differ slightly due to the ‘‘pressure’’ related to the specific environ- households were crowded together within the territory of a single
ment. The van der Waals ideal gas model is able to fit the empirical rat colony. Therefore, the rat/human ratio was likely to be lower in
data well. pffiffiffi 2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi urban areas than rural ones, causing higher mortality in the coun-
We calculate the contact rate C ¼ 2pd c1 qð1 þ c2 qÞ using tryside [73]. In rat-infested urban areas, natural selection might
results from the Japanese railway station study, assuming a colli- have also lowered the rat/human ratio by weeding out rats less
sion diameter d = 1 m, based on the WHO definition of a close con- able to defend themselves against the pathogen.
tact as interaction within 1 m [72]. Fig. 6(B) demonstrates the At the high density limit, the use of collision model is con-
density dependence of collision frequency. The contact rate keeps strained by its initial assumption that people move randomly.
increasing until it reaches maximum, and then decreases due to Therefore, the above scaling trend is closer to reality when people
the slower individual movement. It finally reaches zero when den- move in random directions, such as in railway stations. When peo-
sity is at its maximum. ple move collectively, as in the counterclockwise motion around
the Kaaba at the Hajj, the scaling trend will deviate from the behav-
4. Discussion ior of the estimated curve since the movement of individuals differs
from the random movement of gas molecules. Besides, although de-
Recent computational modeling studies use network models or tailed logs between population movement and density fit well into
social structures to represent contact heterogeneity when modeling the kinetic model, the contact duration, which is essential to deter-
the spread of infectious disease outbreaks, and empirical contact mine the transmission rate, remains unexplored. In mass gather-
experiments focus on revealing the connectivity and the heteroge- ings most contacts may occur in a very short time, therefore the
neous structure of the contact network. However, the scaling mech- overall transmissibility of respiratory disease through crowds
anisms are less focused. The heterogeneous contact structure seems may not be beyond the R0 ¼ 1 threshold, and it is reasonable to ar-
‘orthogonal’ to the transmission scaling mechanisms [20]. Using gue that transmission of influenza in crowds where frequent but
networks, an existing study was able to obtain a consistent scaling short contact occurs may not be as important as household and
relationship regardless of heterogeneous network connectivity workplace contacts, with low frequency but long durations. None-
[21]. Exploring population scaling on both dimensions is essential theless, the kinetic model provides a potential approach to model
for a full understanding of human mixing patterns. physical contacts under large density settings which could be used
Comparing with other contact rate functions, for example the in modeling transmission of respiratory diseases. More empirical
various forms in [19], the advantage of our scaling function in both observations and investigations on infectious diseases spread in
density ends is that it is constructed ab initio from physical con- mass gathering are needed to prove the efficiency of direct contact
tacts, it seems to follow the empirical data in both human and transmission in a variety of density settings.
wildlife diseases and it is possible to formulate and parameterize
a reasonable scaling relationship from independent field observa-
tions and get meaningful scaling results. For direct comparison of 5. Conclusion
real data with other models, different models have different free
parameters, and in many cases models are flexible enough that Substantial evidence indicates that when population density is
they will fit the data quite well. It is only easy to reject a model at the scale of general activities, the transmission mode is likely
when it makes prediction far from reality, for example the infinite to follow an initial sub-linear density-dependent pattern until
linear relationship of the density-dependent function. the saturation of transmission rate transitions to a frequency-
Our nonlinear scaling function has several limitations. We as- dependent pattern independent of population density. For high-
sume an exclusive scaling relationship between reproductive num- density clusters, such as crowds at mass gatherings, the random
ber and population density. Other factors, such as age, population movement and contact of individuals initially increases the fre-
immunity in previously isolated communities, and seasonal varia- quency of contacts. However, the contact rate decreases at extre-
tions may contribute to differences in transmissibility. The effect of mely high density, when it becomes difficult for people to move
important contact structures, such as household and living condi- and make contact with others. On both ends, the scaling effects
tions, is not considered either. For the transmissibility of the can be represented using simple functional forms from spatial con-
1918 influenza pandemic, the population data is acquired from tact models. Existing data, though scarce, support the proposed
the closest census records, and we assume those numbers do not models. We suggest that the conclusion of contact rate scaling with
 H. Hu et al. / Mathematical Biosciences 244 (2013) 125–134 133

population density may provide a first-order approximation in- [18] M. Keeling, B. Grenfell, Disease extinction and community size: modeling the
persistence of measles, Science 275 (1997) 65.
spired by mechanisms of human social interaction and wildlife
[19] H. McCallum, N. Barlow, J. Hone, How should pathogen transmission be
ecology, when setting up transmission rates in large-scale spatial modelled?, Trends in Ecology & Evolution 16 (2001) 295
simulations. We hope that more experimental and theoretical so- [20] M. Begon, M. Bennett, R. Bowers, N. French, S. Hazel, J. Turner, A clarification of
cial contact studies will reveal the complete picture of the mixing transmission terms in host-microparasite models: numbers, densities and
areas, Epidemiology and Infection 129 (2002) 147.
patterns between individuals. [21] M. Ferrari, S. Perkins, L. Pomeroy, O. Bjornstad, Pathogens, social networks, and
the paradox of transmission scaling, Interdisciplinary Perspectives on
Infectious Diseases 2011 (2011) 1.
Acknowledgments [22] M. de Jong, O. Diekmann, H. Heesterbeek, How does transmission of infection
depend on population size, Epidemic Models: Their Structure and Relation to
The authors thank Daniel J. Klein and two anonymous reviewers Data (1995) 84.
[23] R. Anderson, R. May, Infectious Diseases of Humans: Dynamics and Control,
for their insightful discussions and suggestions. The 1918 influenza Oxford University Press, Oxford, 1991.
data was based on data UKDA 4350: Influenza deaths in England [24] C. Rhodes, R. Anderson, Contact rate calculation for a basic epidemic model,
and Wales, 22 June 1918–10 May 1919, deposited by N. Johnson Mathematical Biosciences 216 (2008) 56.
[25] J. Tam, M. Barbeschi, N. Shapovalova, S. Briand, Z. Memish, M. Kieny, Research
and distributed by the UK Data Archive. The data are Crown copy-
agenda for mass gatherings: a call to action, The Lancet Infectious Diseases 12
right. The original data creators, depositors or copyright holders, (2012) 231.
the funders of the Data Collections (if different), and the UK Data [26] K. Dietz, D. Schenzle, Proportionate mixing models for age-dependent
infection transmission, Journal of Mathematical Biology 22 (1985) 117.
Archive bear no responsibility for the further analysis or interpre-
[27] K. Dietz, Overall population patterns in the transmission cycle of infectious
tation in this publication. The authors thank Bill and Melinda Gates disease agents, Population Biology of Infectious Diseases (1982) 87.
for their active support of the EMOD Project and their sponsorship [28] D. Schenzle, K. Dietz, Critical population sizes for endemic virus transmission,
through the Global Good Fund. This work was performed in the Rumliche Persistenz und Diffusion von Krankheiten 83 (1987) 31.
[29] A. Fenton, J. Fairbairn, R. Norman, P. Hudson, Parasite transmission:
EMOD Program at Intellectual Ventures Laboratory, and useful dis- reconciling theory and reality, Journal of Animal Ecology 71 (2002) 893.
cussions with colleagues in this Program are likewise greatly [30] M.J. Smith, S. Telfer, E.R. Kallio, S. Burthe, A.R. Cook, X. Lambin, M. Begon, Host-
appreciated. pathogen time series data in wildlife support a transmission function between
density and frequency dependence, Proceedings of the National Academy of
Sciences 106 (2009) 7905.
References [31] C. Holling, The components of predation as revealed by a study of small-
mammal predation of the european pine sawfly, The Canadian Entomologist
91 (05) (1959) 293.
[1] I. Longini, A. Nizam, S. Xu, K. Ungchusak, W. Hanshaoworakul, D. Cummings,
[32] J. Antonovics, Y. Iwasa, M. Hassell, A generalized model of parasitoid, venereal,
M. Halloran, Containing pandemic influenza at the source, Science 309 (2005)
and vector-based transmission processes, The American Naturalist 145 (1995)
1083.
661.
[2] T. Germann, K. Kadau, I. Longini, C. Macken, Mitigation strategies for pandemic
[33] M. Hochberg, Non-linear transmission rates and the dynamics of infectious
influenza in the United States, Proceedings of the National Academy of
disease, Journal of Theoretical Biology 153 (1991) 301.
Sciences 103 (2006) 5935.
[34] M. Keeling, M. Woolhouse, D. Shaw, L. Matthews, M. Chase-Topping, D.
[3] N. Ferguson, D. Cummings, S. Cauchemez, C. Fraser, S. Riley, A. Meeyai, S.
Haydon, S. Cornell, J. Kappey, J. Wilesmith, B. Grenfell, Dynamics of the 2001
Iamsirithaworn, D. Burke, Strategies for containing an emerging influenza
UK foot and mouth epidemic: stochastic dispersal in a heterogeneous
pandemic in southeast Asia, Nature 437 (2005) 209.
landscape, Science 294 (2001) 813.
[4] N. Ferguson, D. Cummings, C. Fraser, J. Cajka, P. Cooley, D. Burke, Strategies for
[35] R. Larsson, K. Nigmatulina, Engineering responses to pandemics, in: W. Rouse,
mitigating an influenza pandemic, Nature 442 (2006) 448.
D. Cortese (Eds.), Engineering the System of Healthcare Delivery, 153, IOS
[5] S. Eubank, H. Guclu, V. Anil Kumar, M. Marathe, A. Srinivasan, Z. Toroczkai, N.
Press, Washington, DC, 2009, p. 311.
Wang, Modelling disease outbreaks in realistic urban social networks, Nature
[36] K. Nigmatulina, R. Larson, Living with influenza: impacts of government
429 (2004) 180.
imposed and voluntarily selected interventions, European Journal of
[6] S. Cauchemez, A. Bhattarai, T. Marchbanks, R. Fagan, S. Ostroff, N. Ferguson, D.
Operational Research 195 (2009) 613.
Swerdlow, S. Sodha, M. Moll, F. Angulo, Role of social networks in shaping
[37] D. Brockmann, L. Hufnagel, T. Geisel, The scaling laws of human travel, Nature
disease transmission during a community outbreak of 2009 H1N1 pandemic
439 (2006) 462.
influenza, Proceedings of the National Academy of Sciences 108 (2011) 2825.
[38] M. Gonzalez, C. Hidalgo, A. Barabasi, Understanding individual human
[7] S. Riley, Large-scale spatial-transmission models of infectious disease, Science
mobility patterns, Nature 453 (2008) 779.
316 (2007) 1298.
[39] C. Song, Z. Qu, N. Blumm, A. Barabasi, Limits of predictability in human
[8] R. May, Network structure and the biology of populations, Trends in Ecology &
mobility, Science 327 (2010) 1018.
Evolution 21 (2006) 394.
[40] D. Klein, J. Hespanha, U. Madhow, A reaction–diffusion model for epidemic
[9] D. Chao, M. Halloran, V. Obenchain, I. Longini Jr, Flute, a publicly available
routing in sparsely connected MANETs, in: Proceedings IEEE INFOCOM 2010,
stochastic influenza epidemic simulation model, PLoS Computational Biology 6
IEEE, 2010, p. 1.
(2010) e1000656.
[41] I. Abubakar, P. Gautret, G. Brunette, L. Blumberg, D. Johnson, G. Poumerol, Z.
[10] S. Brown, J. Tai, R. Bailey, P. Cooley, W. Wheaton, M. Potter, R. Voorhees, M.
Memish, M. Barbeschi, A. Khan, Global perspectives for prevention of
LeJeune, J. Grefenstette, D. Burke, S. McGlone, B. Lee, Would school closure for
infectious diseases associated with mass gatherings, The Lancet Infectious
the 2009 H1N1 influenza epidemic have been worth the cost? A
Diseases 12 (2012) 66.
computational simulation of Pennsylvania, BMC Public Health 11 (2011) 353.
[42] S. Ebrahim, Z. Memish, T. Uyeki, T. Khoja, N. Marano, S. McNabb, Pandemic
[11] B. Lee, S. Brown, P. Cooley, M. Potter, W. Wheaton, R. Voorhees, S. Stebbins, J.
H1N1 and the 2009 Hajj, Science 326 (2009) 938.
Grefenstette, S. Zimmer, R. Zimmerman, Simulating school closure strategies
[43] L. Jorm, S. Thackway, T. Churches, M. Hills, Watching the games: public health
to mitigate an influenza epidemic, Journal of Public Health Management and
surveillance for the Sydney 2000 Olympic Games, Journal of Epidemiology and
Practice: JPHMP 16 (2010) 252.
Community Health 57 (2003) 102.
[12] J. Mossong, N. Hens, M. Jit, P. Beutels, K. Auranen, R. Mikolajczyk, M. Massari, S.
[44] V. Demicheli, R. Raso, D. Tiberti, A. Barale, L. Ferrara, D. Lombardi,
Salmaso, G. Tomba, J. Wallinga, Social contacts and mixing patterns relevant to
Epidemiological Consultation Team, Surveillance system in place for the
the spread of infectious diseases, PLoS Medicine 5 (2008) e74.
2006 Winter Olympic Games, Torino, Italy, 2006, Euro Surveillance 11 (2006)
[13] D. Balcan, B. Goncalves, H. Hu, J. Ramasco, V. Colizza, A. Vespignani, Modeling
E060209.
the spatial spread of infectious diseases: the global epidemic and mobility
[45] J. Dvorak, A. Junge, K. Grimm, D. Kirkendall, Medical report from the 2006 FIFA
computational model, Journal of Computational Science 1 (2010) 132.
World Cup Germany, British Journal of Sports Medicine 41 (2007) 578.
[14] L. Isella, M. Romano, A. Barrat, C. Cattuto, V. Colizza, W. Van den Broeck, F.
[46] K. Schenkel, C. Williams, T. Eckmanns, G. Poggensee, J. Benzler, J. Josephsen, G.
Gesualdo, E. Pandolfi, L. Rav, C. Rizzo, Close encounters in a pediatric ward:
Krause, Enhanced surveillance of infectious diseases: the, 2006 FIFA World
measuring face-to-face proximity and mixing patterns with wearable sensors,
Cup experience, Germany, Euro Surveillance 11 (2006) 234.
PloS ONE 6 (2011) e17144.
[47] H. Rashid, E. Haworth, S. Shafi, Z. Memish, R. Booy, Pandemic influenza: mass
[15] J. Stehle, N. Voirin, A. Barrat, C. Cattuto, V. Colizza, L. Isella, C. Regis, J. Pinton, N.
gatherings and mass infection, Ambulatory Pediatrics 8 (2008) 526.
Khanafer, W. Van den Broeck, Simulation of an SEIR infectious disease model
[48] M. Moussaid, D. Helbing, G. Theraulaz, How simple rules determine pedestrian
on the dynamic contact network of conference attendees, BMC Medicine 9
behavior and crowd disasters, Proceedings of the National Academy of
(2011) 87.
Sciences 108 (2011) 6884.
[16] A. Deredec, F. Courchamp, Extinction thresholds in host-parasite dynamics,
[49] D. Helbing, A. Johansson, H. Al-Abideen, Dynamics of crowd disasters: an
Annales Zoologici Fennici 40 (2003) 115.
empirical study, Physical Review E 75 (2007) 046109.
[17] J. Lloyd-Smith, P. Cross, C. Briggs, M. Daugherty, W. Getz, J. Latto, M. Sanchez,
[50] D. Helbing, I. Farkas, T. Vicsek, Simulating dynamical features of escape panic,
A. Smith, A. Swei, Should we expect population thresholds for wildlife
Nature 407 (2000) 87.
disease?, Trends in Ecology & Evolution 20 (2005) 511
134 H. Hu et al. / Mathematical Biosciences 244 (2013) 125–134

[51] M. Moussaid, D. Helbing, S. Garnier, A. Johansson, M. Combe, G. Theraulaz, population densities and its implications for rabies control in Britain, Journal
Experimental study of the behavioural mechanisms underlying self- of Applied Ecology (1995) 693.
organization in human crowds, Proceedings of the Royal Society B: [64] G. Smith, D. Wilkinson, Modeling control of rabies outbreaks in red fox
Biological Sciences 276 (2009) 2755. populations to evaluate culling, vaccination, and vaccination combined with
[52] L. Henderson, The statistics of crowd fluids, Nature 229 (1971) 381. fertility control, Journal of Wildlife Diseases 39 (2003) 278.
[53] T. Garrett, Pandemic economics: the 1918 influenza and its modern-day [65] J. David, L. Andral, M. Artois, Computer simulation model of the epi-enzootic
implications, Federal Reserve Bank of St. Louis Review 90 (2008) 75. disease of vulpine rabies, Ecological Modelling 15 (1982) 107.
[54] U.S. Census Bureau, Mortality Statistics 1920, Twenty-First Annual Report, [66] B. Toma, L. Andral, Epidemiology of fox rabies, Advances in Virus Research 21
Government Printing Office, Washington, 1922. (1977) 1.
[55] U.S. Census Bureau, Census of population and housing, 1910. [67] Z. Tyul’ko, I. Kuzmin, Simulation of rabies epizootic process in fox populations
[56] HM Stationery Office, Supplement to the eighty first annual report of the at a limited carrying capacity of biotopes, Russian Journal of Ecology 33 (2002)
register general: report of the mortality from influenza in England and Wales 331.
during the epidemic 1918–1919, 1920. [68] W.H. Organization, WHO expert consultation on rabies: first report, World
[57] N. Johnson, 1918–1919 influenza pandemic mortality in England and Wales, Health, Organization, 2005.
2001. [69] D. Ramsey, N. Spencer, P. Caley, M. Efford, K. Hansen, M. Lam, D. Cooper, The
[58] Great Britain Historical GIS Project, A vision of Britain through time, 2011. effects of reducing population density on contact rates between brushtail
[59] G. Chowell, L. Bettencourt, N. Johnson, W. Alonso, C. Viboud, The 1918–1919 possums: implications for transmission of bovine tuberculosis, Journal of
influenza pandemic in England and Wales: spatial patterns in transmissibility Applied Ecology 39 (2002) 806.
and mortality impact, Proceedings of the Royal Society B: Biological Sciences [70] R. Smith, Density, velocity and flow relationships for closely packed crowds,
275 (2008) 501. Safety Science 18 (1995) 321.
[60] C. Mills, J. Robins, M. Lipsitch, Transmissibility of 1918 pandemic influenza, [71] U. Weidmann, Transporttechnik der Fuganger, Transporttechnische
Nature 432 (2004) 904. Eigenschaften des Fugangerverkehrs, Literaturauswertung, Schriftenreihe des
[61] O. Bjornstad, B. Finkenstadt, B. Grenfell, Dynamics of measles epidemics: IV, ETH Zurich, 1993.
estimating scaling of transmission rates using a time series SIR model, [72] WHO, Advice on the use of masks in the community setting in Influenza
Ecological Society of America 72 (2002) 169. A(H1N1) outbreaks, 2009.
[62] K. Holmala, K. Kauhala, Ecology of wildlife rabies in Europe, Mammal Review [73] O. Benedictow, The Black Death 1346–1353: the Complete History, Boydell
36 (2006) 17. Press, 2004.
[63] P. White, S. Harris, G. Smith, Fox contact behaviour and rabies spread: a model
for the estimation of contact probabilities between urban foxes at different