how to design virus containment policies? an agent-based

How to design virus containment policies? An agent-based analysis of economic and epidemic dynamics under the COVID-19 pandemic

How to design virus containment policies?An agent-based analysis of economic andepidemic dynamics under the COVID-19

pandemic

Alessandro Basurto Herbert Dawid Philipp Harting

Jasper Hepp Dirk Kohlweyer

Economic Theory and Computational Economics (ETACE), Bielefeld University

Behavioral Macro WS, Bamberg, July 8, 2021

1 / 28


Introduction

Motivation

I COVID-19 pandemic caused aglobal health and economiccrisis: > 180 Mio. infected,> 3.8 Mio. deaths, loss inGDP (≈ 5% in 2020 in Ger)

I Substantial variance acrosscountries wrt containmentmeasures

I Substantial variance acrosscountries wrt number ofinfected, casualties and alsowrt economic losses

I Occurrence of mutations,effect of vaccinations

2 / 28


Introduction

Research Questions

I What are the epidemic and economic implications of different typesof lock-down-policies?I Intensity of the lock-downI Degree of opening-up after lock-downI Time of (re-)start of the lock-down

I How do different policies compare with respect to ex-anteuncertainty about mortality and economic loss?

I How strongly are mortality and economic losses under differentpolicies affected by changes in the infectiousness of the virus ?

I How should the termination of containment measures be linked tothe number of vaccinations in the population?

3 / 28


Introduction

Our Approach

I Incorporate ’agentized version’ of standard model of virustransmission (SIRD) into an agent-based macroeconomic framework.

I Explicitly capture how interactions related to economic activity(production, consumption) act as virus transmission channels.

I Calibrate based on German micro and macro data.

I Systematically vary key parameters of the lock-down policy.

I Analyze effect on expected virus mortality / GDP loss and on thevariance of outcomes.

I Identify ’efficiency frontier’.

I Consider variation in virus infectiousness and vaccination dynamics.

4 / 28


Literature

Related LiteratureI Epidemic Models

I ODE: Kermack & McKendrick (1927), Giordano et al (2020),Britton et al. (2020), Kissler et al. (2020)

I Simulations: Epstein (2009), Ferguson et al. (2020), Lux (2020),LeBaron (2021)

I Eco-Pandemic ModelsI Optimization Problem + ODE: Alvarez et al. (2020), Miclo et al.

(2020), Acemoglu et al. (2020)I Dyn. Equil.: Eichenbaum et al. (2021a,b), Krueger et al. (2020),

Jones et al. (2020), Collard et al. (2020), ...I Stoch. Dynamics: Federico & Ferrari (2020)I Agent-based: Dawid et al. (2020), Delli Gatti & Reissl (2020),

Sharma et al. (2021), Mellacher (2020), ...

I Agent Based MacroeconomicsI Methodology: Testfatsion (2006), Foley & Farmer (2009), Dawid

and Delli Gatti (2018), Dosi et al (2019)I Eurace@UniBi: Dawid et al. (2014,2018,2019)

5 / 28


Model

The model: main ingredients economicsI Households (HH):

I Young: employed/unemployed, sector specific skillsI Old: retired, receive pension

I Firms in 4 economic sectors: manufacturing (M), service (S), food(F), public (P),I Labor is only production factor (no capital goods).I Heterogeneity of labor productivity across firms with sector specific

mean.I Price competition (with limited information and noise) in each sector.

I Behavioral rules of firms (pricing, quantity, labor demand) andhouseholds (saving/consumption, consumption choice) based onthose developed in the Eurace@Unibi model (empirical microfoundations).

I Wages homogeneous in each sector.

I Government adjusts tax rate to balance budget (apart fromlock-down periods).

6 / 28


Model

Model Overview

¢

Contacts at Workplace

Contacts at Mall

Firm: Production Planning

Goods Market

Labour Market

Firm: Production

Government

Job Search

Consume

Wages

Labour

Taxes

Pension / Unemployment

Benefit

Firm: Delivery

Hire / Fire

Sales

Sequence of Simulation Steps

Timing of events in Simulation (default policy)

Start: Virus End Lock-down VaccineSwitching between Lock-down and Opening-up

0 14 365 568 Day

Taxes

Social

Contacts

Households

7 / 28


Model

The model: main ingredients epidemics

I 4 potential health states for each HHI SusceptibleI InfectedI RecoveredI Dead

I At each meeting between an infectious and a susceptible HH, thesusceptible becomes infected with probability pinf .

I Newly infected HH is infectious for tinf days after a latency period(tlnt).

I Infected household recovers after trec days.

I Case-specific fatality rate depends on the age of the infected and thecurrent utilization of available intensive care unit beds.

8 / 28


Model

Transmission Channels

Young Agent

Firm

Firm

Mall

Young Agent

Young Agent

Old Agent

Old Agent

Mall

Mall

Contacts at Workplace Social Contacts Contacts during Consumption

9 / 28


Model

Policy Measures

i.) Individual prevention measures (IPM): pinf → (1− ξ)pinf .

ii.) Social distancing (SD): reduction in number of social contacts.

iii.) Working from home (WFH): sector specific fraction of employeeshave no contacts at workplace.

iv.) Reduction in consumption activity (CONS): weekly shoppingprobability of each HH reduced from 1 to ps.

v.) Economic support (ECON):I short-time work: fraction of workers not needed for production move

to short-time work rather being laid-off, financed by public transfers.I bailout: negative firm savings are balanced by public transfer.

I i.) - iii.) have no direct negative economic effects.

I Focus here is on iv.) !

10 / 28


Results

Reproducing German Dynamics: March - September 2020

Reported total infected Smoothed R0 value

Blue: (average) simulation data, Green: German data

Lockdown between days 14 - 63

11 / 28


Results


Casualties GDP loss



12 / 28


Results


Worker in short-time work Excess HH savings



13 / 28


Results

Lock-down Policy

I Social distancing (SD) and reduction in consumption activity(CONS) during the lock-down.

I Three key policy parameters:I αl: intensity of lock-down

I pl = (1, 1, 1)− αl(0.1, 0.33, 0)

I αo: restrictions during opening-up phaseI po = (1, 1, 1)− αo(0.1, 0.33, 0)

I βl: duration of openingI lock-down is re-started if incidence is above βl

I Default policy: (αl, αo, βl) = (1, 0, 50)

I Incidence threshold for stopping lock-down: βo = 5

14 / 28


Results

Default Simulation Settings

I Batches of 50 runs.

I Mutation of nmut = 5 infected HHs occurs at tmut = 162:I Two types of infected status: original or mutated virus

I Upon being infected a HH inherits the type from the infecting agent

I Latency, fatality rate, etc identical between original and mutatedvirus

I Individual infection probability of mutant 50% higher compared tooriginal virus: pmut

inf = 1.5 pinf

I After 18 months pinf → 0 and all restrictions are lifted (vaccinationdynamics introduced later)

I Consider total fraction of virus casualties and average weekly GDPreduction over time-horizon of 24 months

15 / 28


Results

Effects of Variations of the Key Policy Parameters

Red: αl, Blue: αo, Black: βl

16 / 28


Results

Dynamics

Total infected GDP (per capita)

Benchmark (A: (1, 0, 50)),

Low threshold (B: (1, 0, 5)),

Weak lock-down (C: (0.5, 0, 5)),

Cont. weak lock-down (D: (0.5, 0.5, 5))

17 / 28


Results

Differences wrt Expected Policy Effects and Uncertainty

A B C D

(αl, αo, βl) (1, 0, 50) (1, 0, 5) (0.5, 0, 5) (0.5, 0.5, 5)

benchmark policy low threshold weak lockdown cont. weak lockdown

GDP loss [%] 3.73 (0.76) 4.16 (0.91) 1.51 (0.33) 2.00 (0.12)

Mortality [%] 0.077 (0.016) 0.029 (0.008) 0.057 (0.025) 0.062 (0.025)

Duration in lockdown 421.4 (135,2) 495.2 (169.4) 572.74 (151.0) 529.9 (194.9)

Number of lockdowns 2.48 (0.61) 6.40 (2.11) 4.04 (1.35) 2.76 (1.17)

Pub. Acc. Deficit [% of GDP] 5.50 (1.28) 5.18 (1.56) 2.94 (0.75) 2.31 (0.46)

18 / 28


Results

Effects of a Variation of the Virus Infectiousness

I In particular at the beginning of the pandemic the virusinfectiousness is hard to estimate.

I Important to understand how robust qualitative findings are withrespect to variations in ’basic infection probability’ pinf .

I Consider two scenarios:

(i) No mutation: pinf stays at the level calibrated for spring 2020without any mutations

(ii) Higher infectiousness: pinf throughout the pandemic is constant25% higher than the calibrated value in (i)

19 / 28


Results

Policy Effects

No Mutation Higher Infectiousness

Red: αl, Blue: αo, Black: βl

20 / 28


Results

Robustness of Policy Effects wrt Virus Variations

Benchmark Policy A: (1, 0, 50)

Casualties GDP Loss

21 / 28


Results

Robustness of Policy Effects wrt Virus Variations

Cont. Weak Lockdown D: (0.5, 0.5, 5)

Casualties GDP Loss

22 / 28


Results

Effect of Vaccination Roll-out (in default scenario)

I Vaccination roll-out with finite speed (calibrated to German data).

I Old HHs vaccinated first, then young, random within each group.

I 95% effectiveness of the vaccine.

I Fraction qvacc = 0.75 of the population is willing to get vaccinated

Questions:

I Does the comparison between policy effects change qualitatively?

I At which percentage of vaccinated HHs should all containmentmeasures be stopped?

23 / 28


Results

When to Stop Containment Measures?

A: benchmark, B: low threshold, C: weak lock-down, D: cont. weak lock-down

24 / 28


Conclusions

Conclusions

I Agent-based macroeconomic model explicitly incorporatinginteractions giving rise to SARS-CoV-2 transmission channels.

I Calibrated model reproduces well German pandemic and economicdynamics after the outbreak of the COVID-19 crises.

I Main conclusions from the policy analysis:I Policies with long-lasting ‘weak lockdown’ tend to dominate policies

with switches between strong lock-down and full opening.I For given intensity of lockdown, a low incidence threshold for going

back into lockdown should be chosen (substantially lower thanincidence of 50).

I Choice of intensity of lock-down gives rise to trade-off betweenmortality and GDP loss.

I Ex-ante uncertainty wrt policy effects and robustness of effects wrtvirus variations varies considerably between policies.

I Containment policies should be stopped when between 25% to 40%of (vaccination-accepting) population have been vaccinated.

25 / 28


Conclusions

Thank you for your attention!

The code of the model is available on GitHub:https://github.com/ETACE/ace_covid19

26 / 28

https://github.com/ETACE/ace_covid19


Conclusions

Dynamics under Policies without direct economic effects

Reported total infected GDP

27 / 28

how to design virus containment policies? an agent-based

Documents