Game Theory of Social Distancing in Response to an

Epidemic

Timothy C. Reluga*

Department of Mathematics, Pennsylvania State University, State College, Pennsylvania, United States of America

Abstract

Social distancing practices are changes in behavior that prevent disease transmission by reducing contact rates between

susceptible individuals and infected individuals who may transmit the disease. Social distancing practices can reduce the

severity of an epidemic, but the benefits of social distancing depend on the extent to which it is used by individuals.

Individuals are sometimes reluctant to pay the costs inherent in social distancing, and this can limit its effectiveness as a

control measure. This paper formulates a differential-game to identify how individuals would best use social distancing and

related self-protective behaviors during an epidemic. The epidemic is described by a simple, well-mixed ordinary differential

equation model. We use the differential game to study potential value of social distancing as a mitigation measure by

calculating the equilibrium behaviors under a variety of cost-functions. Numerical methods are used to calculate the total

costs of an epidemic under equilibrium behaviors as a function of the time to mass vaccination, following epidemic

identification. The key parameters in the analysis are the basic reproduction number and the baseline efficiency of social

distancing. The results show that social distancing is most beneficial to individuals for basic reproduction numbers around 2.

In the absence of vaccination or other intervention measures, optimal social distancing never recovers more than 30% of

the cost of infection. We also show how the window of opportunity for vaccine development lengthens as the efficiency of

social distancing and detection improve.

Citation: Reluga TC (2010) Game Theory of Social Distancing in Response to an Epidemic. PLoS Comput Biol 6(5): e1000793. doi:10.1371/journal.pcbi.1000793

Editor: Carl T. Bergstrom, University of Washington, United States of America

Received December 3, 2009; Accepted April 23, 2010; Published May 27, 2010

Copyright: ß 2010 Timothy C. Reluga. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits

unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Funding: This research was funded in part by the National Science Foundation (DMS-0920822) and the Bill and Melinda Gates Foundation (Grant Number 49276).

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing Interests: The author has declared that no competing interests exist.

* E-mail: timothy@reluga.org

Introduction

Epidemics of infectious diseases are a continuing threat to the

health of human communities, and one brought to prominence in

the public mind by the 2009 pandemic of H1N1 influenza [1].

One of the key questions of public health epidemiology is how

individual and community actions can help mitigate and manage

the costs of an epidemic. The basic problem I wish to address here

is how rational social-distancing practices used by individuals

during an epidemic will vary depending on the efficiency of the

responses, and how these responses change the epidemic as a

whole.

Social distancing is an aspect of human behavior particularly

important to epidemiology because of its universality; everybody

can reduce their contact rates with other people by changing their

behaviors, and reduced human contact reduces the transmission of

many diseases. Theoretical work on social distancing has been

stimulated by studies of agent-based influenza simulations

indicating that small changes in behavior can have large effects

on transmission patterns during an epidemic [2]. Further research

on agent-based models has argued that social distancing can arrest

epidemics if started quickly and maintained for a relatively long

period [3]. Compartmental epidemic models have also been used

to study social distancing by including states that represent

individuals employing specific behaviors. For instance, Hyman

and Li [4] formulate and begin the analysis of flu disease

transmission in SIR models where some individuals decrease their

activity levels following infection. Reluga and Medlock [5] uses

this approach to show that while social distancing can resemble

immunization, it can generate hysteresis phenomena much more

readily than immunization.

Rather than treating behaviors as states, some models treat

behaviors as parameters determined by simple functions of the

available information. Reluga et al. [6] studies dynamics where

contact rates can depend on the perceived disease incidence.

Buonomo et al. [7] investigates the impact of information

dynamics on the stability of stationary solutions in epidemic

models. Chen [8] considers a similar system but allows individuals

to learn from a random sample of neighbors. Funk et al. [9]

considers the information dynamics associated with social

distancing in a network setting by prescribing a reduction in

contacts based on proximity to infection. Related work by Epstein

et al.[10] explicitly considers the spatial and information dynamics

associated in response to an ongoing epidemic.

Building on the ground-breaking work of Fine and Clarkson

[11], there has been substantial recent interest in the application of

game theory to epidemiology [12–17]. The games studied so far

have primarily considered steady-state problems, and have not

allowed for dynamic strategies. One notable exception to this is the

work of Francis [18], which determines the time-dependent game-

theoretical solution of a vaccination problem over the course of an

epidemic. In another, van Boven et al. [19] studies the optimal use

of anti-viral treatment by individuals when they take into account

the direct and indirect costs of treatment.

PLoS Computational Biology | www.ploscompbiol.org 1 May 2010 | Volume 6 | Issue 5 | e1000793

To study the best usage of social distancing, we apply

differential-game theory at a population-scale. Differential games

are games where strategies have a continuous time-dependence; at

each point in time, a player can choose a different action. For

instance, a pursuit-game between a target and a pursuer is a two-

player differential game where each player’s strategies consist of

choosing how to move at each successive time until the target is

caught by the pursuer or escapes. Geometrically, one might think

of differential games as games where strategies are represented by

curves instead of points. Two-player differential-game theory was

systematically developed by Isaacs [20] as an extension of optimal

control theory [21–23]. Here, we employ an extension of

differential game theory to population games of the form described

by Reluga and Galvani [24]. The analysis in this paper will be

limited to the simplest case of the Kermack–McKendrick SIR

model with strong mixing [25].

In the Model section, we formulate an epidemiological-

economics model for an epidemic, accounting for the individual

and community costs of both social distancing practices and

infection. We then use differential game theory and numerical

methods to identify the equilibrium strategies over the course of an

epidemic. Numerical methods are used to investigate the finite-

time problem where vaccines become available after a fixed

interval from the start of the epidemic and the infinite-horizon

problem without vaccination. Fundamental results on the value

and timing of social distancing are obtained.

Model

In this article, social distancing refers to the adoption of

behaviors by individuals in a community that reduce those

individuals’ risk of becoming infected by limiting their contact with

other individuals or reducing the transmission risk during each

contact. Typically, social distancing incurs some costs in terms of

liberty, social capital, time, convenience, and money, so that

people are only likely to adopt these measures when there is a

specific incentive to do so. In addition to the personal

consequences, the aggregate effects of social distancing form an

economic externality, reducing the overall transmission of disease.

This externality needs to be accounted for in the determination

individuals optimal strategies, but, by definition, depends on the

choice of strategy.

To resolve this interdependence, we formulate our analysis as a

population game where the payoff to each individual is determined

by the individual’s behavioral strategy and the average behavioral

strategy used by the population as a whole. The model is related to

that previously studied by Chen [26]. We will use c

s

to represent

one specific individual’s strategy of daily investment in social

distancing. The population strategy

c

s

is the aggregate daily

investment in social distancing by the population. The overbar

notation is used to indicate that the aggregate investment

c

s

should

be thought of as an average investment aggregated over all

individuals in the population. In the limit of infinitely large

populations, c

s

and c

s

can be thought of as independent because

changes in one person’s behavior will have little affect on the

average behavior. Similarly, the epidemic’s dynamics depend on

the population strategy

c

s

but are independent of any one

individual’s behavior c

s

.

The effectiveness of social distancing is represented by a

function s(c

s

), which is the relative risk of infection given a daily

investment c

s

in social distancing practices. If there is no

investment, the relative risk s(0)~1. As the daily investment c

s

increases, the relative risk s(c

s

) decreases, but is bounded below by

0. We expect diminishing returns with increasing investment, so

we will also make the convenient assumption that s(c

s

) is convex.

Consider a Susceptible-Infected-Recovered (SIR) epidemic

model with susceptible (X), infected and infectious (Y), and

removed (Z) states. Suppose an epidemic starts with Y(0) cases in

a community of N~X(0)zY(0) total individuals (taking

Z(0)~0) and proceeds until time t

f

, at which point all the

individuals in the susceptible state are vaccinated. This epidemic is

fast relative to demographic processes and we do not distinguish

among the possible states of individuals leaving the infectious state,

so the population size N can be treated as constant. Between time

0 and time t

f

, the dynamics are described by

dX

dt

~{s(

c

s

)bYX, ð1aÞ

dY

dt

~s(

c

s

)bYX{cY , ð1bÞ

dZ

dt

~cY, ð1cÞ

where b is the transmission rate and c is the removal rate. This

SIR model assumes the population is homogeneous, strongly

mixed, and that the duration of infections is exponentially

distributed. At the start of the epidemic when there are few cases

of infection (Y (0)&0), the basic reproduction number R

0

~bN=c.

The total cost of the epidemic to the community, J, is the sum

of the direct costs plus the indirect costs of any economic

repercussions from the epidemic. To keep our analysis tightly

focused, we will only consider direct costs of the epidemic,

including the daily costs from infection, daily investments in social

distancing, and the costs of vaccination. Mathematically,

J~{

ð

t

f

0

(c

s

Xzc

I

Y )e

{ht

dt{c

V

X(t

f

)e

{ht

f

{c

I

Y (t

f

)e

{ht

f

1zh

ð2Þ

where c

I

is the daily cost of each infection, c

V

is the cost of

vaccination per person, and h is the discount rate. Note that while

the cost of infection c

I

is a constant, the investment in social

distancing

c

s

is a function of time. The last term in Eq. (2) is called

a salvage term and represents the cumulative costs associated with

individuals who are sick at the time the vaccine is made available

Author Summary

One of the easiest ways for people to lower their risk of

infection during an epidemic is for them to reduce their

rate of contact with infectious individuals. However, the

value of such actions depends on how the epidemic

progresses. Few analyses of behavior change to date have

accounted for how changes in behavior change the

epidemic wave. In this paper, I calculate the tradeoff

between daily social distancing behavior and reductions in

infection risk now and in the future. The subsequent

analysis shows that, for the parameters and functional

forms studied, social distancing is most useful for

moderately transmissible diseases. Social distancing is

particularly useful when it is inexpensive and can delay

the epidemic until a vaccine becomes widely available.

However, the benefits of social distancing are small for

highly transmissible diseases when no vaccine is available.

Game Theory of Social Distancing

PLoS Computational Biology | www.ploscompbiol.org 2 May 2010 | Volume 6 | Issue 5 | e1000793

(t

f

). The assumption that the entire remaining susceptible

population is vaccinated at time t

f

and that vaccination takes

effect instantly is, of course, unrealistic, but does provide an

approximation to the delayed release of a vaccine.

To simplify our studies, we will work with the dimensionless

version of the equations by taking:

S ~

b

c

X , I~

b

c

Y, R~

b

c

Z ,

^

tt~ct,

c~

c

s

c

I

, k~

c

V

c

I

,

^

hh~

h

c

,

^

ss(

c)~s(c

I

c):

ð3Þ

Under this choice of units, time will be measured in terms of

disease generations, social distancing costs will be measured

relative to the daily cost of infection, and population sizes will be

measured relative to the critical population size necessary to

sustain an epidemic.

Epidemics usually start with one or a few index cases, so we

focus on scenarios where I(0)&0. The dynamics can be described

in terms the shape of s(

^

cc), the discount rate

^

hh, and a single initial-

condition parameter

S (0)&

b

c

N: ð4Þ

From this, it follows that R

0

&S(0). Since epidemics are often

much faster than human demographic processes governing the

discount rate [27], we will also take

^

hh~0 in all calculations.

Henceforth, we will drop the hat-notation and work with the

dimensionless parameters. The dimensionless equations are

dS

dt

~{s(

c)IS, ð5aÞ

dI

dt

~s(

c)IS{I, ð5bÞ

J~{

ð

t

f

0

(cSzI)e

{ht

dt{

I(t

f

)e

{ht

f

1zh

{kS(t

f

)e

{ht

f

ð5cÞ

with the constraint that

c§0. Note that we drop the function

notation when necessary to simplify the presentation.

For our further analysis, we will assume

s(c)~

1

1zmc

, ð6Þ

with the maximum efficiency of social distancing s’(0)~{m. Eq.

(6) is nicely behaved for numerical solutions because of its

relatively fat tail.

The Social Distancing Game

We now formulate a differential game for individuals choosing

their best social distancing practices relative to the aggregate

behavior of the population as a whole. The following game-

theoretic analysis combines the ideas of Isaacs [20] and Reluga

and Galvani [24]. The premise of the game is that at each point in

the epidemic, people can choose to pay a cost associated with

social distancing in exchange for a reduction in their risk of

infection. The costs of an epidemic to the individual depend on the

course of the epidemic and the individual’s strategy of social

distancing. The probabilities p(t) that an individual is in the

susceptible, infected, or removed state at time t evolve according

to the Markov process

_

pp~Q(t; c)p ð7Þ

where c is the individual’s daily investment as a function of the

epidemic’s state-variables and the transition-rate matrix

Q(t; c)~

{s(c)I 00

s(c)I {10

010

2

6

4

3

7

5

: ð8Þ

Note that both c and I change over time. Along the lines discussed

above, c and

c represent different quantities in our analysis; c

represents one individual’s investment strategy and the population

strategy

c represents an aggregated average of all individual

investments. We also note that there are several different ways c

and

c can be parameterized. They may be parameterized in terms

of time, as c(t) and

c(t), or in implicit feedback form c(V

S

, I) and

c(V

S

, I), or in explicit feedback form c(S, I) and c(S, I). The form

used will be clear from the context.

Since the events in the individual’s life are stochastic, we can not

predict the exact time spent in any one state or the precise payoff

received at the end of the game. Instead, we calculate expected

present values of each state at each time, conditional on the

investment in social distancing. The expected present value is

average value one expects after accounting for the probabilities of

all future events, and discounting future costs relative to immediate

costs. The expected present values V of each state evolve

according to the adjoint equations

{

_

VV~ Q

T

{hI

Vzv ð9Þ

where v

T

(t; c)~{c(t), {1, 0½. The components V

S

(t; c,c),

V

I

(t; c,c), and V

R

(t; c,c) represent the expected present values of

being in the susceptible, infected, or removed state at time t when

using strategy c in a population using strategy

c. The expected

present values depend on the population strategy

c through the

infection prevalence I.

The adjoint equations governing the values of each state are

derived from Markov decision process theory. They are

{

dV

S

dt

~{hV

S

z(V

I

{V

S

)s(c)I{c, ð10aÞ

{

dV

I

dt

~{hV

I

zV

R

{V

I

{1, ð10bÞ

{

dV

R

dt

~{hV

R

, ð10cÞ

with the constraints that c(t)§0 for all time t. Solution of (10)b

and (10)c gives

V

I

(t

f

)~{1=(1zh), V

R

(t

f

)~0: ð11Þ

If it is impossible to make a vaccine, the equations must be solved

over an infinite horizon. Over an infinite horizon, V

S

(?)~{c=h,

Game Theory of Social Distancing

PLoS Computational Biology | www.ploscompbiol.org 3 May 2010 | Volume 6 | Issue 5 | e1000793

assuming c becomes constant. In the case of no discounting (h~0),

we still have V

S

(?)~0 provided c(S(t),0)~0 for sufficiently large

t. In the case where a perfect vaccine is universally available at

terminal time t

f

, the value of the susceptible and removed states

differs by the cost of vaccine k for t§t

f

. To avoid complications

with the choice of whether-or-not to vaccinate, we take k~0 so

V

S

(t

f

)~0. This is reasonable in scenarios where the cost of the

vaccine is covered by the government.

The dynamics are independent of R, so we need not consider

removed individuals further. Taking h~0 and V

I

~{1, we need

only study the reduced system

dS

dt

~{s(

c)IS, ð12aÞ

dI

dt

~s(

c)IS{I, ð12bÞ

{

dV

S

dt

~{ 1zV

S

ðÞs(c)I{c, ð12cÞ

with boundary conditions

S (0)~R

0

{I

0

, I(0)~I

0

, V

S

(t

f

)~0: ð12dÞ

The other conditions must be calculated from the solution of the

boundary-value problem and provide useful information. {V

S

(0)

will be the expected total cost of the epidemic to the individual.

The final size of the epidemic is given by S(0){S(t

f

).

Game Analysis

Solving a game refers to the problem of finding the best strategy

to play, given that all the other players are also trying to find a best

strategy for themselves. In some games, there is a single strategy

that minimizes a player’s costs no matter what their opponents do,

so that strategy can very reasonably be referred to as a solution. In

many games, no such strategy exists. Rather, the best strategy

depends on the actions of the other players. Any strategy played by

one player is potentially vulnerable to a lack of knowledge of the

strategies of the other players. In such games, it is most useful to

look for strategies that are equilibria, in the sense that every

player’s strategy is better than the alternatives, given knowledge of

their opponent’s strategies. A Nash equilibrium solution to a

population game like that described by System (12) is a strategy

that is a best response, even when everybody else is using the same

strategy. i.e. given V

S

(t; c,c), c

is a Nash equilibrium if for every

alternative strategy c, V

S

(0; c,c

)ƒV

S

(0; c

,c

). A Nash equilib-

rium strategy is a subgame perfect equilibrium if it is also a Nash

equilibrium at every state the system may pass through. I will not

address the problem of ruling out finite-time blowup of the

Hamilton–Jacobi equation and establishing existence and unique-

ness of subgame perfect equilibria. But numerical and analytical

analyses strongly support the conjecture that the stategies

calculated here are the unique global subgame perfect equilibria

to the social distancing game.

The equilibria of System (12) can be calculated using the

general methods of Isaacs [20]. The core idea is to implement a

greedy-algorithm; at every step in the game, find the investment

that maximizes the rate of increase in the individual’s expect value

V

S

. We represent strategies as functions in implicit feedback form.

c(V

S

,I) is the amount an individual invests per transmission

generation when the system is at state V

S

,I.Ifc

(V

S

,I) is a

subgame perfect equilibrium, then it satisfies the maximum

principle

c

(V

S

,I)~ argmax

c§0

{(1zV

S

)s(c)I{c

ð13Þ

when

c~c

everywhere. So long as s(c

) behaves well, in the sense

that it is differentiable, decreasing, and strictly convex, then c

is

uniquely defined by the relations

c

~0if{s’(0)I(1zV

S

)ƒ1,

{s’(c

)I(1zV

S

)~1 otherwise,

where s’~

ds

dc

:ð14Þ

Figure 1 shows the interface in V

S

|I phase space separating the

region where the equilibrium strategy will include no investment in

social distancing (c

~0,s(c

)~1) from the region where the

equilibrium strategy requires investment in social distancing

(c

w0,s(c

)v1).

Two cases are immediately interesting. The first is the infinite-

horizon problem – what is the equilibrium behavior when there is

never a vaccine and the epidemic continues on until its natural

end? The second is the finite-horizon problem – if a vaccine is

introduced at time t

f

generations after the start of the epidemic,

what is the optimal behavior while waiting for the vaccine? In both

of these cases, it is assumed that all players know if and when the

vaccine will be available.

The infinite-horizon and finite-horizon problems are distin-

guished by their boundary conditions. In the finite-horizon case,

we assume all susceptible individuals are vaccinated at final time

t

f

,soV

S

(t

f

)~0, V

I

~{1, S(0)~S

0

, I(0)~I

0

&

>

0 while V

S

(0)

and S(t

f

) are unknown. In the limit of the infinite-horizon case

(t

f

??), we solve the two-point boundary value problem with

terminal conditions V

S

(?)~0, I(?)~0, and initial conditions

S(0)~S

0

, I(0)~I

0

&

>

0 while V

S

(0) and S(?) are unknown. But

these conditions are insufficient to specify the infinite-horizon

problem. The plane I~0 is a set of stationary solutions to Eq. (12),

so we need a second order term to uniquely specify the terminal

condition when we are perturbed slightly away from this plane.

Figure 1. Contour plots of relative risk surface for equilibrium

strategies. The relative risk is presented in feedback form with implicit

coordinates s(c

(V

S

,I)) (left) and transformed to explicit coordinates

s(c

(S,I )) (right) for the infinite-horizon problem with maximum

efficiency m~10. The greater the value of the susceptible state (V

S

),

the greater the instantaneous social distancing. We find that increasing

the number of susceptible individuals always decreases the investment

in social distancing, and the greatest investments in social distancing

occur when the smallest part of the population is susceptible. Note that

in the dimensionless model, the value of the infection state V

I

~{1.

doi:10.1371/journal.pcbi.1000793.g001

Game Theory of Social Distancing

PLoS Computational Biology | www.ploscompbiol.org 4 May 2010 | Volume 6 | Issue 5 | e1000793

Using Eq. (12), we can show solutions solve the second-order

terminal boundary condition

LV

S

LI

S ,I~0ðÞ~

{1

1{S

ð15Þ

for 0ƒSv1 as t

f

??.

Most of the equilibria we calculate are obtained numerically.

Some exceptions are the special cases where S(0)~0, I(0)w0.

Under these conditions, solutions can be obtained in closed-form.

First, I (t)~I(0)e

{t

. While mI(1zV

S

)v1, c

~0 and

V

S

(t)~½V

S

(0)z1e

I(0)(1{e

{t

)

{1: ð16Þ

When matched to the terminal boundary condition, we find that if

we write V

S

in feedback form as a function of I rather than t,

V

S

(I)~e

{I

{1 ð17Þ

is a solution so long as mIe

{I

v1 for all I. Inspecting the

inequality condition, we find that this holds as long as mve.

Results

A problem with solving Eq. (12) under Eq. (14) is that it requires

I(t) to be known from past time and V

S

(t) to be known from

future time. This is a common feature of boundary-value

problems, and is resolved by considering all terminal conditions

S(t

f

). Using standard numerical techniques, identifying an

equilibrium in the described boundary-value problem reduces to

scalar root finding for S (t

f

) to match the given S(0). The special

form of the population game allows the solution manifold to be

calculated directly by integrating backwards in time, rather than

requiring iterative approaches like those used for optimal-control

problems [23]. Code for these calculations is available from the

author on request.

Before presenting the results, it is helpful to develop some

intuition for the importance of the maximum efficiency m of

investments in social distancing. Given m for an arbitrary relative

risk function s, then in the best-case scenarios, where diminish-

ments on returns are weakest, one would have to invest atleast

1=m of the cost of infection per disease generation to totally isolate

themselves. The units here are derived from dimensional analysis.

This could be invested for no more than m generations, before

one’s expenses would exceed the cost of becoming infected. When

returns are diminishing, fewer than m generations of total isolation

are practical. Thus, the dimensionless efficiency m can be thought

of as an upper bound on the number of transmission generations

individuals can afford to isolate themselves before the costs of

social distancing outweigh the costs of infection.

For the infinite-horizon problem, an example equilibrium

strategy and the corresponding dynamics in the absence of social

distancing are shown in Figure 2. We can show that if social

distancing is highly inefficient (the maximum efficiency

mve&2:718), then social distancing is a waste of effort, no

matter how large R

0

. If social distancing is efficient, then there is a

threshold value of R

0

below which social distancing is still

impractical because the expected costs per day to individuals is too

small compared to the cost of social distancing, but above which

some degree of social distancing is always part of the equilibrium

strategic response to the epidemic (Figure 3).

The exact window over which social distancing is used depends

on the basic reproduction number, the initial and terminal

conditions, and the efficiency of distancing measures. The

feedback form of equilibrium strategies, transformed from (V

S

,I)

coordinates to the (S,I) coordinates of the phase-space is

represented with contour plots in Figure 1. Among equilibrium

strategies, social distancing is never used until part-way into the

epidemic, and ceases before the epidemic fully dies out.

The consequences of social distancing are shown in Figure 4.

The per-capita cost of an epidemic is larger for larger basic

reproduction numbers. The more efficient social distancing, the

more of the epidemic cost can be saved per person. However, the

net savings from social distancing reaches a maximum around

R

0

~2, and never saves more than 30% of the cost of the epidemic

per person. For larger R

0

’s, social distancing is less beneficial.

We can also calculate solutions of the finite-time horizon

problem where a vaccine becomes universally available at a fixed

time after the detection of disease (Figure 5). If mass vaccination

occurs soon enough, active social distancing occurs right up to the

date of vaccination. Using numerical calculations of equilibria over

finite-time horizons, we find that there is a limited window of

opportunity during which mass vaccine can significantly reduce

Figure 2. Epidemic solutions with equilibrium social distancing and without social distancing. Social distancing reduces the epidemic

peak and prolongs the epidemic, as we can see by comparing a time series with subgame-perfect social distancing (top left) and a time series with

the same initial condition but no social distancing (bottom left) (parameters R

0

~S(0)~4:46, m~20). In the phase plane (right), we see that both

epidemics track each other perfectly until S&3, when individuals begin to use social distancing to reduce transmission. Eventually, social distancing

leads to a smaller epidemic. The convexity change appearing at the bottom the phaseplane orbit with social distancing corresponds to the cessation

of social distancing.

doi:10.1371/journal.pcbi.1000793.g002

Game Theory of Social Distancing

PLoS Computational Biology | www.ploscompbiol.org 5 May 2010 | Volume 6 | Issue 5 | e1000793

the cost of the epidemic, and that social distancing lengthens this

window (Figure 6). The calculations show that increases in either

the amount of time before vaccine availability or the basic

reproduction number increase the costs of the epidemic. Smaller

initial numbers of infections allow longer windows of opportunity.

This is as expected because the larger the initial portion of the

population infected, the shorter the time it takes the epidemic to

run its full course.

Discussion

Here, I have described the calculations necessary to identify the

equilibrium solution of the differential game for social distancing

behaviors during an epidemic. The benefits associated with the

equilibrium solution can be interpreted as the best outcome of a

simple social-distancing policy. We find that the benefits of social

distancing are constrained by fundamental properties of epidemic

dynamics and the efficiency with which distancing can be

accomplished. The efficiency results are most easily summarized

in terms of the maximum efficiency m, which is the percent

reduction in contact rate per percent of infection cost invested per

disease generation. As a rule-of-thumb, m is an upper bound on

the number of transmission generations individuals can isolate

before the costs of social distancing outweigh the costs of infection.

Social distancing is not practical if this efficiency is small compared

to the number of generations in the fastest epidemics (m v2:718).

While social distancing can yield large reductions in transmission

rate over short periods of time, optimal social-distancing strategies

yield only moderate reductions in the cost of the epidemic.

Our calculations have determined the equilibrium strategies

from the perspective of individuals. Alternatively, we could ask

what the optimal social distancing practices are from the

perspective of minimizing the total cost of the epidemic to the

community. Determination of the optimal community strategy

Figure 3. Social distancing threshold. This is the threshold that dictates whether or not equilibrium behavior involves some social distancing. It

depends on both the basic reproduction number R

0

and the maximum efficiency m, and is independent of the exact form of s. As rough rules of

thumb, if R

0

v1 or mv2:718, then equilibrium behavior involves no social distancing.

doi:10.1371/journal.pcbi.1000793.g003

Figure 4. Total costs and savings. Plots of the total per-capita cost of an epidemic {V

S

(0) (left) under equilibrium social distancing for the

infinite-horizon problem with several efficiencies m under Eq. (6), and the corresponding per-capita savings (right). Savings in expected cost

compared to universal abstention from social distancing are largest for moderate basic reproduction numbers, but are relatively small, even in the

limit of infinitely efficient social distancing. The m~? case corresponds to infection of the minimum number of people necessary to reduce the

reproduction ratio below 1.

doi:10.1371/journal.pcbi.1000793.g004

Game Theory of Social Distancing

PLoS Computational Biology | www.ploscompbiol.org 6 May 2010 | Volume 6 | Issue 5 | e1000793

leads to a nonlinear optimal control problem that can be studied

using standard procedures [23]. Yet, practical bounds on the

performance of the optimal community strategy can be obtained

without further calculation. The optimal community strategy will

cost less than the game-theoretic solution per capita, but must cost

more than 1{1=R

0

, as that is the minimum number of people

who must become sick to reduce the effective reproduction

number below the epidemic threshold. Preliminary calculations

indicate that optimal community strategies and game equilibrium

strategies converge as R

0

grows, and significant differences are

only observable for a narrow window of basic reproduction

numbers near 1.

The results presented require a number of caveats. I have, for

instance, only considered one particular form for the relative risk

function. Most of the analysis has been undertaken in the absence

of discounting (h~0), under the assumption that the epidemic will

be fast compared to planning horizons. Discounting would

diminish importance of long term risks compared to the instant

costs of social distancing, and thus should diminish the benefits of

social distancing. The benefits of social distancing will also be

diminished by incorporation of positive terminal costs of

vaccination (kw0). Realistically, mass vaccination cannot be

accomplished all-at-once, as we assume. It’s much more likely that

vaccination will be rolled out continuously as it becomes available.

This could be incorporated into our analysis, for instance, by

including a time-dependent forcing. Other approaches include

extending the model to incorporate vaccination results of Morton

and Wickwire [28], or to allow an open market for vaccine

purchase [18].

The simple epidemic model is particularly weak in its prediction

of the growths of epidemics because it assumes the population is

randomly mixed at all times. We know, however, that the contact

Figure 5. Solutions when vaccine becomes available after a fixed time. These are time series of an equilibrium solution for social distancing

when mass vaccination occurs 8:6 generations (left) and 6:5 generations (right) after the start of the epidemic. Investments in social distancing begin

well after the start of the epidemic but continue right up to the time of vaccination. Social distancing begins sooner when vaccine development is

faster. For these parameter values (m~20, S(0)~R

0

~3, I(0)~3|10

{6

), individuals save 50% of the cost of infection per capita (left) and 80% of the

cost of infection (right).

doi:10.1371/journal.pcbi.1000793.g005

Figure 6. Windows of Opportunity for Vaccination. Plots of how the net expected losses per individual ({V

S

(0)) depend on the delay

between the start of social-distancing practices and the date when mass-vaccination becomes universally available if individuals use a Nash

equilibrium strategy. The more efficient social distancing, the less individuals invest prior to vaccine introduction. The blue lines (m~0) do not use

social distancing, as the efficiency is below the threshold. The dotted lines represent the minimal asymptotic epidemic costs necessary to stop an

epidemic.

doi:10.1371/journal.pcbi.1000793.g006

Game Theory of Social Distancing

PLoS Computational Biology | www.ploscompbiol.org 7 May 2010 | Volume 6 | Issue 5 | e1000793

patterns among individuals are highly structured, with regular

temporal, spatial, and social correlations. One consequence of

heterogeneous contact structure is that epidemics proceed more

slowly than the simple epidemic model naively predicts. Thus, the

simple epidemic model is often considered as a worst-case-

scenario, when compared with more complex network models

[29,30] and agent-based models [31–33]. In the context of social

distancing, it is not immediately clear how weaker mixing

hypotheses will affect our results. Weakened mixing will prolong

an epidemic, increasing the window over which social distancing is

needed. But under weakened mixing, individuals may be able to

use local information to refine their strategies in ways analogous to

the ideas of Funk et al. [9] and Perisic and Bauch [34]. In general,

the analysis of aggregate games with stochastic population

dynamics require a significant technical leaps, and are the subjects

of active research.

One of the fundamental assumptions in our analysis is that there

are no cost-neutral behavior changes that can reduce contact rates.

In fact, life-experience provides good evidence that many

conventional aspects of human behavior are conditional on

cultural norms, and that different cultures may adopt alternative

conventions. The introduction of a new infectious disease may

alter the motivational pressures so that behavioral norms that were

previously equivalent are no longer, and that one norm is now

preferred to the others. In such cases, there are likely to be

switching costs that retard the rapid adoption of the better

behaviors that conflict with cultural norms. The rate of behavior

change, then, would be limited by the rate of adoption of

compensatory changes in cultural norms that reduce the cost of

social distancing.

Another deep issue is that behavior changes have externalities

beyond influencing disease incidence, but we have not accounted

for these externalities. People’s daily activities contribute not just to

their own well-being but also to the maintenance of our economy

and infrastructure. Social distancing behaviors may have serious

negative consequences for economic productivity, which might

feed back into slowing the distribution of vaccines and increasing

daily cost-of-living expenses.

We can extend our analysis to include economic feedbacks by

incorporating capital dynamics explicitly. Individuals may accu-

mulate capital resources like food, water, fuel, and prophylactic

medicine prior to an epidemic, but these resources will gradually

be depleted and might be difficult to replace if social distancing

interferes with the economy flow of goods and services. Further

capital costs at the community and state scales may augment

epidemic valuations. These factors appear to have been instru-

mental in the recent US debate of school-closure policies. One

feature of a model with explicit capital dynamics is the possibility

of large economic shocks. This and related topics will be explored

in future work.

These calculations raise two important mathematical conjec-

tures which I have not attempted to address. The first is that the

social distancing game possesses a unique subgame-perfect Nash

equilibrium. There is reasonable numerical evidence of this in

cases where the relative risk function s is strictly convex, and

stronger unpublished arguments of this in cases of piecewise linear

s. I believe this will also be the case for non-convex but monotone

relative risks under some allowances of mixed-strategies. A second

conjecture, not yet addressed formally, is that increases in the

efficiency of social distancing always lead to greater use of social

distancing, all other factors being equal. This seems like common

sense, but the precise dependence of Figure 1 on the efficiency has

yet to be determined mathematically.

As with all game-theoretic models, human behavior is unlikely

to completely agree with our equilibria for many reasons,

including incomplete information about the epidemic and vaccine

and strong prior beliefs that impede rational responses. On the

other hand, our approach is applicable to a large set of related

models. We can analyze many more realistic representations of

pathogen life-cycles. For instance, arbitrary infection-period

distributions and infection rates can be approximated using a

linear chain of states or delay-equations [24]. Structured

populations with metapopulation-style mixing patterns may also

be analyzed. I hope to apply the methods to a wider variety of

community-environment interactions in the future.

Acknowledgments

The author thanks A. Bressan, A. Galvani, and E. Shim for helpful

discussion, and two anonymous referees for their valuable criticisms.

Author Contributions

Conceived and designed the experiments: TR. Performed the experiments:

TR. Analyzed the data: TR. Wrote the paper: TR.

References

1. Dawood F, Jain S, Finelli L, Shaw M, Lindstrom S, et al. (2009) Emergence of a

novel swine-origin influenza A (H1N1) virus in humans. New England Journal of

Medicine 360: 2605–15.

2. Del Valle S, Hethcote H, Hyman JM, Castillo-Chavez C (2005) Effects of

behavioural changes in a smallpox attack model. Mathematical Biosciences 195:

228–251.

3. Kelso J, Milne G, Kelly H (2009) Simulation suggests that rapid activa tion of

social distancing can arrest epidemic development due to a novel strain of

influenza. BMC Public Health 9: 117.

4. Hyman JM, Li J (2007) Infection-age structured epidemic models with behavior

change or treatment. Journal of Biological Dynamics 1: 109–131.

5. Reluga TC, Medlock J (2007) Resistance mechanisms matter in SIRS models.

Mathematical Biosciences and Engineering 4: 553–563.

6. Reluga TC, Bauch CT, Galvani AP (2006) Evolving public perceptions

and stability in vaccine uptake. Mathematical Biosciences 204: 185–

198.

7. Buonomo B, d’Onofrio A, Lacitignola D (2008) Global stability of an sir

epidemic model with information dependent vaccination. Mathematical

Biosciences 216: 9–16.

8. Chen FH (2009) Modeling the effect of information quality on risk behavior

change and the transmission of infectious diseases. Mathematical Biosciences

217: 125–133.

9. Funk S, Gilad E, Watkins C, Jansen V (2009) The spread of awareness and its

impact on epidemic outbreaks. Proceedings of the National Academy of Sciences

106: 6872.

10. E pstein J, Parker J, Cummings D, Hammond R (2008) Coupled contagion

dynamics of fear and disease: mathematical and computational explorations.

PLoS ONE 3: e3955.

11. F ine PEM, Clarkson JA (1986) Individual versus public priorities in the

determination of optimal vaccination policies. American Journal of Epidemiol-

ogy 124: 1012–1020.

12. B auch CT, Galvani AP, Earn DJD (2003) Group interest versus self-interest in

smallpox vaccination polic y. Proceedings of the National Academy of Sciences

100: 10564–10567.

13. Chen FH (2006) A susceptible-infected epidemic model with voluntary

vaccinations. Journal of Mathematical Biology 53: 253–272.

14. R eluga TC, Medlock J, Poolman E, Galvani AP (2007) Optimal timing of

disease transmission in an age-structured population. Bulletin of Mathematical

Biology 69: 2711–2722.

15. T oxvaerd F (2009) The economics of infectious disease: Infection, treatment and

acquired immunity. in submission.

16. d’Onofrio A, Manfredi P (2009) Information-related changes in contact patterns

may trigger oscillations in the endemic prevalence of infectious diseases. Journal

of Theoretical Biology 256: 473–478.

17. R eluga TC (2009) An SIS game with two subpopulations. Journal of Biological

Dynamics 3: 515–531.

18. F rancis PJ (2004) Optimal tax/subsidy combinations for the flu season. Journal

of Economic Dynamics and Control 28: 2037–2054.

19. van Boven M, Klinkenberg D, Pen I, Weissing F, Heesterbeek H (2008) Self-

Interest versus Group-Interest in Antiviral Control. PLoS ONE 3: e1558.

Game Theory of Social Distancing

PLoS Computational Biology | www.ploscompbiol.org 8 May 2010 | Volume 6 | Issue 5 | e1000793

20. Isaacs R (1965) Differential Games: A mathematical theory with applications to

warfare and pursuit, control and optimizatio n. New York: John Wiley and Sons.

21. Arrow KJ, Kurz M (1970) Public Investment, the Rate of Return, and Optimal

Fiscal Policy. BaltimoreMD: Johns Hopkins Press.

22. Clark C (1976) Mathematical bioeconomics John Wiley & Sons New York.

23. Lenhart S, Workman J (2007) Optimal Control Applied to Biological Models

Chapman & Hall/CRC.

24. Reluga TC, Galvani AP (2010) A general approach for population games with

application to vaccination. in submission.

25. Kermack WO, McKendrick AG (1927) Contributions to the mathematical-

theory of epidemics. Proceedings of the Royal Society of London 115: 700–721.

26. Chen FH (2004) Rational behavioral response and the transmission of stds.

Theoretical Population Biology 66: 307–316.

27. Reluga TC, Medlock J, Galvani AP (2009) The discounted reproductive number

for epidemiology. Mathematical Biosciences and Engineering 6: 377–393.

28. Morton R, Wickwire KH (1974) On the optimal control of a deterministic

epidemic. Advances in Applied Probability 6: 622–635.

29. M eyers LA, Newman ME, Pourbohloul B (2006) Predicting epidemics on

directed contact networks. Journal of Theoretical Biology 240: 400–418.

30. M eyers LA (2007) Contact network epidemiology: Bond percolation applied to

infectious disease prediction and control. Bulletin of the American Mathematical

Society 44: 63–86.

31. Mniszewski S, Del Valle S, Stroud P, Riese J, Sydoriak S (2008) EpiSimS

simulation of a multi-component strategy for pandemic influenza. Proceedings of

the 2008 Spring simulation multiconference. pp 556–563.

32. L ongini IM, Nizam A, Xu S, Ungchusak K, Hanshaoworakul W, et al. (2005)

Containing pandemic influenza at the source. Science 309: 1083–1087.

33. F erguson N, Cummings D, Cauchemez S, Fraser C, Riley S, et al. (2005)

Strategies for containing an emerging influenza pandemic in Southeast Asia.

Nature 437: 209–214.

34. P erisic A, Bauch C (2009) Social Contact Network s and Disease Eradicability

under Voluntary Vaccination. PLoS Computational Biology 5: e1000280.

Game Theory of Social Distancing

PLoS Computational Biology | www.ploscompbiol.org 9 May 2010 | Volume 6 | Issue 5 | e1000793