A Brief Primer on SIR Models in Economics

Alexander Karaivanov
Simon Fraser University

May 15, 2020

SFU, May 2020

A. Karaivanov SIR Models

What is a SIR model?

compartmental epidemiological model

population of size N passes over time t through three stages (more can be
added)

{ Susceptible, St
{ Infectious, It
{ Resolved, Rt (recovered or dead)

N = St + It + Rt for all t

initial condition: some infected, I0 > 0; the rest N I0 susceptible

SFU, May 2020 1

A. Karaivanov SIR Models

SIR Equations

S to I
dSt It
= St
dt N

I to R
dIt It
= St It
dt N

R (absorbing state)
dRt
= It
dt

SFU, May 2020 2

A. Karaivanov SIR Models

Key equation

dIt It
= St It (*)
dt N

for a susceptible agent i 2 St, the probability (rate) of being infected is

It
Prob(S ! I) =
N

captures
the probability/rate of infection conditional on contact with I person
the contact rate
{ some authors write = 0 Mt

It
is the probability/rate of contact with I person (uniform mixing)
N

SFU, May 2020 3

A. Karaivanov SIR Models

Key parameters

\infectiousness",
{ tricky, since product of biology (the chance of passing the infection upon
meeting I person) and behavior/policy (the contact rate)

\removal rate",
{ (mostly) biological { how fast people recover and how many infected die
(fraction , approx. 0.3%{0.6% for COVID-19)
1
{ measured in time (1= is the expected time to recovery/death, e.g., in
days)

SFU, May 2020 4

A. Karaivanov SIR Models

Dynamics

re-write SIR equation (*) as:

dIt It
= St It =
dt N
St
= It ( 1) (1)
N

dIt
the current number of infected It will grow, > 0 if
dt

St
>1 (2)
N

and decrease otherwise

SFU, May 2020 5

A. Karaivanov SIR Models

Basic reproduction number R0

R0 = expected new infections per unit of time generated by the rst infected
person, when S0 ' N
note the confusing notation with Rt (resolved)

from (1) we see that

R0 =

hence the importance of R0 = = being larger or smaller than 1

{ if R0 < 1 the epidemic never takes o , even at S0 = N , see (2)

also, note that the early (St ' N ) infection growth rate is

dIt
=It '
dt

SFU, May 2020 6

A. Karaivanov SIR Models

E ective reproduction number

as the epidemic evolves in time (assuming R0 > 1) there are less susceptibles

so, the expected new infections per I person and unit of time,

St
Rt
N

decrease over time

Rt is called the e ective reproduction number

note that if interventions (e.g., lockdown, distancing, testing and

quarantine) or behavioral responses a ect or St; It they a ect Rt too

SFU, May 2020 7

A. Karaivanov SIR Models

Herd immunity

when Rt = 1 (in ow into state I equals out ow from I) it is said that \herd
immunity" is reached (unstable steady state)
{ herd immunity occurs when
dIt
=0
dt
{ that is, when the remaining fraction of susceptibles is
St 1
=
N R0

{ or, equivalently, when the fraction of recovered or dead (who were infected
at some point) is
Rt 1
=1
N R0

Example: if R0 = 2:5, need 60% of the population to have been infected

to reach herd immunity

SFU, May 2020 8

A. Karaivanov SIR Models

Overshoot

Z
60% may sound bad enough but the total number of infected, Itdt (and
dead) continues to grow after herd immunity is reached

this is known as overshoot

Why?

dIt
dt< 0 after herd immunity is reached means only that there are fewer daily
new infections (It decreases) but It > 0 still

important role for isolating known infectious persons

SFU, May 2020 9

A. Karaivanov SIR Models

Extensions

\Exposed" state { incorporate incubation time (around 5 days for COVID-19)

{ SEIR model

\Quarantine" state { e.g., if an individual tests positive and is isolated from

contacts
It It
{ reduces in St
N N

\symptomatic" vs. \asymptomatic" infectious states

SFU, May 2020 10

A. Karaivanov SIR Models

SIR model and social networks

the standard SIR model assumes population-level, uniform mixing

abstracts from locality and the fact that many infections occur via social
contacts (broadly de ned)

SFU, May 2020 11

A. Karaivanov SIR Models

Network SIR model (NSIR)

represent the population by graph G (nodes with social links among them)

probability of a susceptible person i becoming infected at time t depends on

i's social contacts
P
j2CG (i) 1xjt =I
Prob(S ! I) =
#CG(i)

where CG(i) is the set of person i0s contacts (other nodes in G)

#CG(i) is node i's degree (number of contacts)

SFU, May 2020 12

A. Karaivanov SIR Models

E ective reproduction number in the NSIR model

note that the graph structure a ects the infection dynamics

in the standard SIR model, the probability of S person i meeting an I person

It
is uniform, 8i (\representative agent model")
N

in the network model this probability is heterogeneous and depends on i's

social contacts P
j2CG (i) 1xjt =I
it (G)
#CG(i)

can model `superspreaders', clusters, local outbreaks, etc.

SFU, May 2020 13

A. Karaivanov SIR Models

Policy interventions

It
essentially, trying to make smaller the term St
N

main bene t: \ atten the curve" It

{ reduce its peak (ICU capacity constraint)
{ reduce the overshoot and total deaths

cost: can prolong the epidemic duration

SFU, May 2020 14

A. Karaivanov SIR Models

Policy interventions

simplest, reduced-form way

{ time-dependent , e.g., t = (1 lt) (Atkeson, 2020; Moll, 2020)
equivalent to reducing the e ective reproduction number Rt
{ model \distancing" or \lockdown" as period of lower t

It
alternative way: reduce the rate of meetings with infectious persons St
N
{ e.g., (1 )St meet with (1 )It where is fraction locked down
(Alvarez et al. 2020)
{ di erent meeting technology (Acemoglu et al., 2020)

SFU, May 2020 15

A. Karaivanov SIR Models

Testing and quarantine

introduce state P (tested positive) which evolves as

dPt
= It Pt
dt

where is the testing rate per unit of time

Note: the \new cases" reported in the news every day correspond to Pt and
not to It (infectious)

Quarantine or (self-)isolation:
{ assume P agents are removed from meetings
{ or meet on a restricted-contacts graph GQ

SFU, May 2020 16

A. Karaivanov SIR Models

Contact tracing

key role for the graph G

tracking, testing and isolating infected social contacts of positive (P) agents
X
Prob(I ! P) = + 1xjt=P
j2CG (i)

where is the contact tracing rate

SFU, May 2020 17

A. Karaivanov SIR Models

Examples and discussion

SFU, May 2020 18

 Source: Moll (2020)
Short and Tight Lockdown

24
A. Karaivanov SIR Models

One-o or intermittent distancing

SFU, May 2020 19

 Source: Karaivanov (2020)

A. Early short distancing D. Delayed short distancing

% population

% population
time (days) time (days)

B. Early medium long distancing E. Delayed medium long distancing

% population

% population
time (days) time (days)

C. Early long distancing F. Delayed long distancing

% population

% population

time (days) time (days)

Note: assumed initial infection rate 0.5%
 Source: Karaivanov (2020)

G. (30)-30-(30) distancing H. (60)-30-(60) distancing

% population

% population
time (days) time (days)

I. (30)-90-(30) distancing J. (60)-90-(60) distancing

% population

% population
time (days) time (days)

K. (30)-120-(30) distancing L. (60)-120-(60) distancing

% population

% population

time (days) time (days)

Note: assumed initial infection rate 0.5%
A. Karaivanov SIR Models

Lockdown exit

SFU, May 2020 21

 Source: Karaivanov (2020)

A. p=1, =70% lockdown

% population
infectious, It
lockdown period

B. p=0.5, =70% lockdown

% population

C. p=0, =30% lockdown

% population

D. p=0, =30% lockdown

% population

E. p=0, =70% lockdown

% population

F. p=0, =70% lockdown

% population

time (days)
 Source: Karaivanov (2020)

Infectious, It

70% lockdown policy

no testing and contact tracing
2% testing only
2% testing and 5% contact tracing
5% testing only
% population

5% testing and 5% contact tracing

5% testing and 10% contact tracing
10% testing and 20% contact tracing

Cumulative deaths
% population

time (days)
A. Karaivanov SIR Models

Behavioral responses

assume timely information on positive cases is available

responses (A, B, C) based on individual circumstances { e.g., a contact who

tested positive
{ permanent reduction in contacts (upper bound)
{ temporary reduction (while a contact is positive)

responses D, E based on aggregate data (x new cases in the past y days)

SFU, May 2020 22

 Source: Karaivanov (2020)

Behavioral responses A, B, C based on individual positive contacts

no behavioral response
behavioral response A
behavioral response B
Infectious I t, %
behavioral response C

Behavioral responses D based on aggregate new cases

2% testing, no behavioral response

2% testing, behavioral response D
Infectious I t, %

Behavioral responses E based on aggregate new cases

5% testing, no behavioral response

5% testing, behavioral response E
Infectious I t, %

time (days)
A. Karaivanov SIR Models

Impact on the economy?

requires (heroic) assumptions about

{ productivity in lockdown and/or when infectious (symptomatic vs.
asymptomatic)
{ value of life / years

current frontier working papers look at interventions and outcomes by

segments
{ by industry: e.g., retail vs. IT jobs are a ected very di erently
{ or, by population cohorts: elderly vs. work-age vs. school-age

SFU, May 2020 23

Economic activity index, Y t

no intervention
testing and quarantine
testing, quarantine and contact tracing
behavioral response E

time (days)
Economic activity index, Y t

no intervention
distancing D
distancing J
70% lockdown

time (days)
 Source: Moll (2020)
Some Unpleasant Lockdown Arithmetic
• If lockdowns only option, how long do effective ones need to last?
• Key: need to reach herd immunity. So: how long to reach that?
• Optimistically assume perm immunity, R0 ↓ to 2 (better hygiene...)
herd immunity threshold = 1 − 1/R0 = 50%
• Simple back of envelope calculation for U.S. Assumptions:
1. 10% have had disease ⇒ need additional 40% ≈ 100 million
2. lockdown suppresses Ret to 1, infections rolled over (sl 40-41)
(Ret = 1 close to current US estimate)

3. 200,000 new infections per day (current official count ≈ 30,000)

• ⇒ need some sort of lockdown / control for
100 million
= 500 days
200, 000
• Note: optimistic calculation assuming low R0 , permanent immunity
44
A. Karaivanov SIR Models

References
Acemoglu, D., V. Chernozhukov, I. Werning and M. Whinston (2020), \A Multi-Risk SIR Model with Optimally Targeted Lockdown"

Alvarez, F., D. Argente and F. Lippi (2020), \A Simple Planning Problem for COVID-19 Lockdown"

Atkeson, A. (2020), \What will be the Economic Impact of COVID-19 in the US? Rough Estimates of Disease Scenarios"

Azzimonti, M., A. Fogli, F. Perri and M. Ponder (2020), \Social Distance Policies in Network Cities"

Berger, D., K. Herkenho and S. Mongey (2020), \An SEIR Infectious Disease Model with Testing and Conditional Quarantine"

Chari, V.V., R. Kirpalani and C. Phelan (2020), \The Hammer and the Scalpel: On the Economics of Indiscriminate versus Targeted
Isolation Policies during Pandemics"

Farboodi, M., G. Jarosch and R. Shimer (2020), \Internal and External E ects of Social Distancing in a Pandemic"

Fernandez-Villaverde, J. and C. Jones (2020), \Estimating and Simulating a SIRD Model of COVID-19 for Many Countries, States, and
Cities"

Karaivanov, A. (2020), \A Social Network Model of COVID-19", http:[Link]~akaraivanKaraivanov [Link]

Kermack, W. and A. McKendrick (1927), \A contribution to the mathematical theory of epidemics", Proceedings of the Royal Society
A, 115(772), p.700-721

Kuchler, T., D. Russel and J. Stroebel (2020), \The geographic spread of COVID-19 correlates with structure of social networks as
measured by Facebook"

Moll, B. (2020), \Lockdowns in SIR Models", slides

SFU, May 2020 24