medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020.

The copyright holder for this preprint

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

TESTING, TRACING AND SOCIAL DISTANCING: ASSESSING

OPTIONS FOR THE CONTROL OF COVID1 9

Lia Humphreya , Edward W. Thommesb , Roie Fieldsa , Laurent Coudevilleb

Naseem Hakimc , Ayman Chitd,e , Monica G. Cojocarua,∗

a Department of Mathematics and Statistics, University of Guelph, Guelph, Ontario,

Canada

b Vaccine Epidemiology and Modeling, Sanofi Pasteur, Toronto, Ontario, Canada

c Founder, covid-testing.org

d,e Sanofi Pasteur North-America, Leslie Dan School of Pharmacy, University of

Toronto, Canada

Abstract

In this work we present an analysis of non-pharmaceutical interventions implemented

around the world in the fight against COVID-19: Social distancing, shelter-in-place,

mask wearing, etc measures to protect the susceptible, together with, in various degrees,

testing & contact-tracing to identify, isolate and treat the infected. The majority of

countries have relied on the former, while ramping up their testing and tracing

capabilities. We consider the examples of South Korea, Italy, Canada and the United

States. By fitting a disease transmission model to daily case report data, we show that

in each of the four countries their combination of social-distancing and testing/tracing

to date have had a significant impact on the evolution of their pandemic curves. In this

work we estimate the average isolation rates of infected individuals needing to occur in

each country as a result of large-scale testing and contact tracing as a mean of lifting

social distancing measures, without a resurgence of COVID-19. We find that an average

isolation rate of an infected individual every 4.5 days (South Korea), 5.7 days (Canada)

and to 6 days (Italy) would be sufficient. We also find that a rate of under 3.5 days will

help in the United States, although it would not completely mitigate the second wave

the country is currently under.

Key Words: Pandemic modelling | Pandemic forecasting under policy | Testing

frequency policy modelling

July 26, 2020 1/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

YAll authors contributed to the paper equally, from inception, model

implementation to data fitting and forecast.

¶Edward W. Thommes, Laurent Coudeville and Ayman Chit are employees of Sanofi

Pasteur. Monica G. Cojocaru has received national research grants in the past in which

Sanofi Pasteur was a matching partner. The grants are for research completely

independent from this work.

*
E-mail: mcojocar@uoguelph.ca

1 Introduction 1

In late 2019, a novel betacoronavirus called SARS-CoV-2 emerged from a live animal 2

marketplace in Wuhan, Hubei Province, China, and has since inflicted a worldwide 3

pandemic of a disease now referred to as COVID-19. The disease is highly contagious, 4

with an estimated R0 between 2.2 and 4.6 [1, 2] although it is important to consider that 5

R0 is not strictly biologically determined but rather heavily influenced by host 6

behavioural and environmental factors [3]. The incubation period has been found to be 7

5.1-5.2 days, while 97.5% of patients display symptoms within 11.5 days [1, 4]. The 8

disease spreads primarily through the respiratory tract and respiratory secretions. 9

As of July 16 2020, there have been a total of 13.7 million confirmed cases of 10

COVID-19 worldwide, and over 580,000 deaths (WHO COVID-19 Situation Report 89, 11

https://www.who.int/emergencies/diseases/novel-coronavirus-2019/ 12

situation-reports). All outbreaks, apart from that of Hubei Province continue to be 13

active, with new cases reported daily, though a number of countries have clearly passed 14

a (first) peak. Countries have taken various degrees of social distancing measures: 15

lockdowns, shelter-in-place, banning gatherings, sport events, closing schools, mask 16

wearing etc. (e.g. [5–8]) in an effort to suppress the disease, or at least prevent it from 17

overwhelming a country’s critical care capacity. While proven effective to slow the 18

spread, these measures have had a large effect on daily lives and economies throughout 19

the world. Countries who managed to stave off their first wave, are now in the process 20

of implementing (in various degrees) relaxation measures, in an effort to restore, as 21

much and as safe as possible, their socio-economic landscapes. It is thus critical to 22

further examine and devise strategies which will allow more phasing out of social 23

July 26, 2020 2/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

distancing measures while preventing a resurgence of outbreaks [9], [10]. 24

In this work, we fit a disease transmission model to daily case reports in four 25

developed countries in order to assess the effectiveness of their COVID-19 26

countermeasures. We chose these due to the fact that they implemented 27

countermeasures in various degrees and combinations, with differing outcomes to date. 28

We begin with the examples of Italy, Canada and the United States, three countries 29

which have relied principally on social distancing via universal shelter-in-place measures, 30

while testing and contact tracing was implemented at an increased pace only after the 31

pandemic was more-or-less established in their populations. We compare their pandemic 32

evolution scenarios to that of South Korea, a country which has had and still has a tight 33

control on COVID-19 through a combination of early aggressive testing and contact 34

tracing, paired with social-distancing measures. We show that an analogous strategy 35

can still provide, from this moment onward, a feasible path to further relaxing 36

social-distancing in Italy, Canada and by extension other countries with similar 37

pandemic profiles. The case of the United States stands apart: while coordinated large 38

scale, frequent testing and contact tracing will help decelerate the current U.S. 39

pandemic trajectory, this must be accompanied by a tightening rather than relaxation 40

of social distancing measures if that country’s outbreak is to be brought under control. 41

The structure of the paper is as follows: in Section 2 we introduce our model main 42

ideas, notation, and assumptions. We follow in Section 3 with the presentation of our 43

fitted infection curves for Italy, Canada, the U.S., and South Korea, wherein we infer 44

the net effect to date of social distancing, testing & contact tracing on the decrease of 45

the transmission rates in each of the 4 countries. We then evaluate scenarios for 46

countries to phase out social distancing while preventing a resurgence of COVID-19 47

outbreaks by assuming an increase in testing and contact tracing. We are able to derive 48

the frequency with which a tested (infected) individual and an exposed (traced) 49

individual need to be detected and isolated in order for each country to maintain an 50

effective reproduction number of 1 (that is to say, each country maintains a ”slow burn” 51

of their pandemic) while assuming a major relaxation of social-distancing rules. We 52

present a thorough discussion and conclusions in Section 5. Additional mathematical 53

background is included in Appendix A. 54

July 26, 2020 3/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

2 Materials and methods 55

2.1 The SEIRL model 56

The transmission of an infectious disease in a homogeneously mixed population is often 57

described by a Susceptible-Infectious-Recovered (SIR) model [11] or its variants, most 58

notably a Susceptible-Exposed-Infectious-Recovered (SEIR); see [12] for a recent review. 59

A SEIR model normalized to population size N is described by 4 differential equations 60

of the form: 61

ds
= −βsi (1)
dt
de
= βsi − σe
dt
di
= σe − γi
dt
dr
= γi
dt

S E I R
with s = N, e= N, i= N, r= N, s + e + i + r = 1 and where β is the rate of effective

contacts, 1/σ = Tlat is the mean latent period (which may differ from the incubation

period), and 1/γ = Tinf is the mean duration of infectiousness, with both times having

exponential distributions. We also have the auxiliary equation for the cumulative

number of cases,
dc
= σe
dt

The daily incidence of cases on day i is then 62

inci = ci − ci−1

In turn, a SIR model is similar to (1) but without the E compartment, thus only 3 63

differential equations subject to s + i + r = 1. 64

The spread of an infectious disease can be halted if its effective reproduction number 65

Ref f = R0 s, (2)

can be decreased below 1. The effective reproduction number of both the SIR and SEIR 66

July 26, 2020 4/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

compartmental disease transmission models is 67

β
Ref f = s. (3)
γ

In both the SIR and SEIR model, when s ≈ 1 and we are near the disease-free 68

equilibrium (1, 0, 0, 0), the early growth of both i and inc is exponential (see e.g. [12]): 69

i(t) = i0 eρt , inc(t) = inc0 eρt (4)

In the SIR model, the growth factor ρ is given by 70

ρSIR = β − γ = γ(R0 − 1), (5)

One then can express R0 in terms of ρ: 71

ρSIR + γ
R0SIR = (6)
γ

In a similar manner, in the SEIR model we have [12]: 72

p
−(σ + γ) + (σ − γ)2 + 4σβ
ρSEIR = (7)
2

and by solving for β from Equation 7 we get: 73

(ρSEIR + σ)(ρSEIR + γ)
R0SEIR = . (8)
σγ

In the limit as σ → ∞, the SEIR model reduces to the SIR model, and accordingly, as 74

can readily be shown by L’Hospital’s rule: 75

(ρSEIR + σ)(ρSEIR + γ) ρSIR + γ

lim = = (9)
σ→∞ σγ γ

For COVID-19, as for other pandemics (e.g. SARS, MERS, the 1918 Spanish flu), 76

we can assume the entire population to be initially susceptible. Therefore, in the early 77

stages of an outbreak, Ref f ≈ R0 . We will also assume that infection with COVID-19 78

confers subsequent immunity, which does not wane significantly over the time horizon 79

July 26, 2020 5/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

considered here. Thus, whether they die or recover, an infected person is considered 80

removed from the pool of susceptibles. In the absence of a vaccine or other control 81

measures, s(t) = 1 − c(t), where c(t) is the cumulative number of people infected at 82

time t. 83

From Equation 3, assuming β to be given, we see that Ref f can be decreased in two 84

ways: by decreasing s at a rate higher than that due to infection alone; or by increasing 85

γ. The former can be considered an abstraction of social distancing measures, since 86

these effectively take a part of the population “out of circulation” as far as disease 87

transmission is concerned. The latter can be achieved by identifying and isolating 88

infected individuals early, thus cutting short Tinf . 89

To explicitly depict the role of control measures, we adapt the SEIR model to a 90

pandemic setting by adding an isoLated (L) compartment. As before, we include the 91

auxiliary equation for C, the cumulative number of infected. The resulting SEIRL 92

model, is described by: 93

ds
= −βsi
dt
de
= βsi − σe − κ1 e
dt
di
= σe − (γ + κ)i
dt
dr
= γi
dt
dl
= κi + κ1 e
dt
dc
= σe
dt

where as in the standard SEIR model, β is the mean rate of effective contacts, 94

1/σ = Tlat is the mean latent period, and 1/γ = Tinf is the mean infectious period. 95

Finally, 1/κ1 = Tisol,lat and 1/κ = Tisol,inf are the mean times for the latent and 96

infectious, respectively, to be isolated as a consequence of either testing or contact 97

tracing. 98

Its effective reproduction number is then given by (see Appendix A - CHECK) 99

βσ
Ref f,SEIRL = s (10)
(σ + κ1 )(γ + κ)

July 26, 2020 6/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

It can be shown (see Appendix A) that the exponential growth rate (Equation 4) of 100

infected near the disease-free equilibrium is 101

p
−(σ + κ1 + γ + κ) ((σ + κ1 ) − (γ + κ))2 + 4σβs
ρSEIRL = + ,
2 2

and the rate of effective contacts is 102

(σ + κ1 + ρSEIRL )(γ + κ + ρSEIRL )

βSEIRL =
σ

We can also express the effective reproduction number in terms of ρSEIRL : 103

(σ + κ1 + ρSEIRL )(γ + κ + ρSEIRL )

R0SEIRL =
(γ + κ)σ

3 Results 104

3.1 Estimating R0 from early exponential growth 105

While growth is still exponential, we have from (Equation 4) that 106

log(inc(t)) ∝ ρt (11)

i.e. a log-linear plot of incidence versus time will have slope ρ. Indeed, early exponential 107

growth can be seen to be a near-universal feature in COVID-19 daily case count data 108

from around the world. Figures 1 and 2 plot log(inc) versus time for South Korea, Italy, 109

Canada and the U.S., using time series data of daily new cases compiled by the Johns 110

Hopkins University Center for Systems Science and Engineering (JHU CSSE) [13], 111

retrieved from 112

http://github.com/CSSEGISandData/COVID-19/tree/master/csse_covid_19_data 113

In all four countries the initial linear phase is clearly apparent, followed by a 114

transition to sub-exponential growth. This transition is sharpest for South Korea, where 115

growth switches abruptly to decay around 1 March. Regression fit results for ρ and the 116

corresponding doubling time, 117

ln(2)
Tdbl = (12)
ρ

July 26, 2020 7/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

Fig 1. Log-linear plots of daily COVID-19 incidence versus time for South Korea and
Italy. The initial linear phase corresponds to exponential growth, which subsequently
turns over into sub-exponential growth. The factor ρ and the corresponding doubling
time are estimated via a regression fit to the initial phase. R0 is calculated using Eq. 8
with σ = γ = (2.5d)−1 .

together with dates for the onset of major national-level protective measures, are given 118

in Figures 1 and 2 and Table 1. In all four cases, the transition to sub-exponential 119

growth occurred at or after the time that widespread protective measures were first 120

invoked. 121

Inferring R0 from ρ requires choosing values for the mean latent and infectious 122

periods. The sum of these is the mean serial interval: 123

Tser = Tlat + Tinf = σ −1 + γ −1 (13)

Estimates of the serial interval of COVID-19 range from 3.95 to 6.6 days [14–17]. We 124

adopt a value of Tser = 5 days. The latent period of the disease is not well constrained, 125

but it can be shown (Appendix A) that for a given value of Tser and ρ, the maximum 126

value of R0 is obtained when Tlat = Tinf = Tser /2. We assume this “worst-case” 127

scenario and let Tlat = Tinf = 2.5 days. 128

3.2 Quantifying the effectiveness of COVID-19 129

countermeasures thus far 130

We begin with the remark that all 4 countries have enacted social distancing via school 131

closures, nationwide shutdowns, shelter-in-place orders, mask wearing, etc., all in 132

July 26, 2020 8/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

Fig 2. As Fig 1, but for Canada and the U.S..

Table 1. Initial exponential growth rates and R0 values for South Korea,
Italy, Canada, and the U.S.
Country ρ initial growth R0
S. Korea 0.22(0.15,0.29) 2.38(1.88,2.94)
Italy 0.18(0.16,0.2) 2.1(1.98,2.22)
Canada 0.18(0.16,0.2) 2.11(1.97,2.26)
U.S. 0.3(0.27,0.33) 3.07(2.81,3.35)
counter-measures sub-exponential onset fraction of cases reported
S. Korea Feb 21 [18] March 1 0.84
Italy March 12 [19] March 22 0.11
Canada March 20 [20] April 3 0.26
U.S. March 13 [21] April 6 0.51
Estimates of initial exponential growth rate ρ are obtained from regression fits to the
early outbreak phases (Figs 1 and 2). Corresponding values of R0 assume
Tlat = Tinf = 2.5 days. Estimates of the fraction of cases reported are taken from [22].

July 26, 2020 9/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

various combinations. South Korea was the first to impose measures, followed by Italy, 133

where the measures were ordered and coordinated eventually country-wide. In Canada 134

most provinces enacted similar measures over the course of 1-2 weeks around March 20, 135

2020, while the United States has had the more heterogeneous spread of similar 136

measures, depending on specific states. 137

We interpret the transition to sub-exponential growth (Figs 1 and 2) as the first 138

signature of the effect of these measures in a given country, and use this as the starting 139

point to infer the net effect that these measures have had up to the end of June 2020. 140

For each country, we fit the SEIR model solutions for daily incidence 141

{incmodel,1 , ..., incmodel,n } to daily case reports. The model output is multiplied by a 142

factor kf , where k is an estimate of the fraction of symptomatic cases reported, 143

obtained using delay-adjusted case fatality rates [22], and f is the fraction of cases 144

which are symptomatic, estimated to be f = 0.5, from a recent CDC report 1 . 145

We compute − log L, the normal negative log likelihood of the time series of 146

observed daily incidences, {incobs,1 , ..., incobs,n }, given the model output, as a function 147

of model parameters 148

x = (i0 , q1 , q2 , ..., qm )

where i0 is the initial number of infected and the qi are reduction factors on the rate of 149

disease transmission, varying over time, such that βi0 = qi β (see Table 1). R0 for each 150

country is fixed at the respective values obtained via regression above. Parameters are 151

drawn using unweighted (uniform) Latin hypercube sampling. The best-fit solution is 152

the one which minimizes L. 153

We present our best fits for Italy, Canada, the U.S. and South Korea in the next 154

figures, including: 155

• Linear and semi-log plots of daily incidence data of confirmed cases per country, 156

together with maximum-likelihood model fit (“with measures, confirmed”). 157

• Inferred true number of infected, taking into account under-reporting and 158

asymptomatic cases (“with measures, all”). Shown for comparison are the number 159

of confirmed cases (“no measures, confirmed”) and all cases (“no measures, all”) 160

1 COVID-19 Pandemic Planning Scenarios, https://www.cdc.gov/coronavirus/2019-ncov/hcp/

planning-scenarios-h.pdf

July 26, 2020 10/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

expected to have occurred in the absence of countermeasures. 161

• Cumulative incidence and inferred reduction in effective contact rates (i β) due to 162

social-distancing, mask wearing etc. 163

• These fits assume κ = κ1 = 0, in other words, the reduction in the effective 164

contact rates qi0 s are a measure of each country’s combination of measures to date, 165

including social-distancing, testing and tracing. 166

Fig 3. (A) Left two panels: Linear and semi-log plots of daily incidence data of
confirmed cases in Italy, together with maximum-likelihood model fit (“with measures,
confirmed”) Also shown is the inferred true number of infected, taking into account
under-reporting and asymptomatic cases (“with measures, all”). Shown for comparison
are the number of confirmed cases (“no measures, confirmed”) and all cases (“no
measures, all”) expected to have occurred in the absence of countermeasures. (B)
Right two panels: Cumulative incidence (top) and inferred reduction of effective
contacts, together with the corresponding effective reproduction numbers (vertical
numbers) (bottom).

In all four countries, interventions arrested the initial exponential rise in cases and 167

brought the effective reproduction number below. In Italy, South Korea and Canada, 168

daily case numbers have since been brought far below their peak values. 169

South Korea effected the strongest and most rapid reduction in transmission. South 170

Korea experienced a very similar early exponential growth in cases, and hence has a 171

similar inferred R0 , as the other three countries. However, its mitigation and control 172

measures stood out from the beginning in the fact that the country employed a rapid 173

scale-up of testing, concurrent with contact tracing and isolating of infected individuals. 174

There are also social distancing measures imposed, but notably no shelter-in-place. Last 175

but not least, mask wearing is a regular policy that the population adopts widely (bot 176

July 26, 2020 11/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

Fig 4. Canadian daily COVID-19 incidence (left panels). Canadian cumulative

incidence fit and inferred effective contact rates reductions

Fig 5. U.S. daily COVID-19 data and model fit for incidence and cumulative incidence;
see caption of Fig 3 for details

only for this pandemic, but also for flu for example). Members of the population also 177

participate in a surveillance of contacts in order to identify potential spread early. In 178

contrast, mask wearing was instituted much later in the other 3 countries and some 179

regions of the US are still struggling to effectively adopt it. 180

After some relaxation of measures in South Korea, together with a series of national 181

holidays from April 30 to May 5 (“Golden Week”) possibly playing a role, a resurgence 182

occurred. This appears to have since been stabilized, with (see Fig 1 above). Our fit is 183

presented below: 184

July 26, 2020 12/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

Fig 6. South Korea daily COVID-19 data and model fit for incidence and cumulative
incidence; see caption of Fig 3 for details

July 26, 2020 13/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

4 Large-scale frequent testing and contact tracing 185

as a way to relax social-distancing measures and 186

control the spread - the SEIRL model 187

4.1 Theoretical estimates 188

As presented in Section 2, for a given set of values of β, σ and γ, Equation 10 gives us a 189

closed form expression for Ref f,SEIRL as a function of κ and κ1 . This relationship is 190

1
depicted as a surface plot in Fig 7 for σ = γ = 2.5 and σ = γ = 15 . In the latter case, 191

Ref f,SEIRL is nearly twice as large as in the former. It is interesting to note, though, 192

that for both cases, the combinations of κ and κ1 that make Ref f,SEIRL = 1 (i.e. the 193

intersections of the respective surfaces with the R0 = 1 plane) are quite similar. This 194

can be understood as follows: As σ and γ become small compared to κ1 and κ, 195

respectively, it is the latter which increasingly dominate the rate at which 196

exposed/infected people are isolated. 197

Fig 7. Effects of isolation rates due to testing and contact tracing on the initial value
1
of R0S(Q)EIRL model. We computed σ = γ = 2.5 (upper most surface), σ = γ = 15
(middle surface) and the reference surface R0 = 1.

From Equation 10, we obtain the relationship between κ and κ1 that makes 198

Ref f,SEIRL = 1: 199

sβσ sβσ
1= =⇒ k = −γ (14)
(σ + κ1 )(γ + κ) σ + κ1

Extracting the current values of s from our simulations of last section, we can now 200

July 26, 2020 14/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

compute average isolation rates due to testing and tracing that would ensure an 201

effective reproduction number of 1, while assuming that the effective contact rates β 202

will reverse to their values near the disease-free equilibrium, before any other social 203

distancing measures were employed. We further assume that the isolation rate due to 204

contact tracing ( κ1 ) is the same as the isolation rate due to testing (κ) and we used 205

Equation 14. We present our results compactly in Table 2 below. 206

Table 2. Model results for South Korea, Italy, Canada, and the U.S. for
Ref f = 1.
1 1
β s Ref f κ = κ1 ( , )
κ κ1
SK 0.961 0.999 1 0.22 (4.5, 4.5) days
Italy 0.841 0.937 1 0.161 (6.2,6.2) days
Can 0.841 0.982 1 0.175 (5.7,5.7) days
U.S. 1.225 0.96 1 0.288 (3.5, 3.5) days
1
In all cases, we chose γ = σ = 2.5 for a Tser = 5 days and where s are the estimated
values of susceptibles remaining in each country in mid July 2020.

While the U.S. and South Korea do not seem to have an effective reproduction 207

number under 1 at the moment, Canada and Italy have their effective reproductive 208

number estimated to Ref f = 0.83, as seen from Figures 1, 2. In their cases, we can redo 209

our estimates for the 2 countries and compute the isolation rates due to testing and 210

tracing so that they maintain their current value of Ref f = 0.83 (see Table 3): 211

Table 3. Model results for South Korea, Italy, Canada, and the U.S. for
Ref f = 0.83.
1 1
β s Ref f κ = κ1 ( , )
κ κ1
Italy 0.841 0.937 0.83 0.216 (4.62,4.62) days
Can 0.841 0.982 0.83 0.23 (4.33,4.33) days
1
In all cases, we chose γ = σ = 2.5 for a Tser = 5 days and where s are the estimated
values of susceptibles remaining in each country in mid July 2020. We used Eq (8).

4.2 Numerical results 212

We present next pandemic forecasts under different testing and contact tracing rates, in 213

the four countries under consideration. We show how the theoretical estimates arise in 214

the context of the simulated pandemic evolution in each of the 4 countries. 215

We depict first Canada and Italy, as they have with similar estimates, in several 216

respective testing and contact tracing scenarios: 217

July 26, 2020 15/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

Fig 8. Predicted daily cases in the Canada under different rates of isolation due to
testing, or testing plus contact-tracing, accompanied by a cessation of distancing
measures.

In Fig 9 below, We clearly see that for values of κ−1 = κ−1

1 shorter than the 218

thresholds given in Table 2, a second pandemic wave is averted in each respective 219

country. This threshold is the least stringent in Italy, where about 6% of the population 220

(accounting for asymptomatic and/or unreported cases) is inferred to have been infected 221

in the first wave, as opposed to Canada. 222

Next we present the simulations for South Korea and the United States: We see 223

that the isolation rates are the most stringent (lowest values in days) for the U.S., where 224

the value of R0 inferred from the initial exponential rise of cases is right now higher 225

than that of the other three countries (≈ 1.2). In the case of the United States, a 226

large-scale testing and tracing operation alone will not be able to curtail the current 227

epidemic curve around, thus strong social-distancing measures will be still be needed. 228

Current testing guidelines for social-distancing relaxation measures are established 229

by the WHO in such a way that countries can relax these if positivity rates for testing 230

are under 5% for 14 days in a row. Currently, South Korea is at 1.08%, Italy = 2.39%, 231

2
Canada = 4.34% and U.S. = 6.28% From publicly available data we have that daily 232

testing (in numbers per day/per 1 million) are now as: Italy = 0.09823, 233

Canada=0.085104, U.S.=0.1282 3 . We note that all these rates are lower than what we 234

2 https://coronavirus.jhu.edu/testing/international-comparison
3 https://www.statista.com/statistics/1104645/covid19-testing-rate-select-countries-worldwide/

July 26, 2020 16/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

Fig 9. Predicted daily cases in Italy under different rates of isolation due to testing, or
testing plus contact-tracing, accompanied by a cessation of distancing measures
accompanied by different with various isolation rates due to testing only, or a
combination of testing and contact tracing

require in our tables above, however this is consistent with the fact that at the moment 235

none of South Korea, Canada, Italy or U.S. have completely removed all 236

social-distancing measures. These values give us an idea of a possible current isolation 237

1 1
rate due to testing/tracing: 0.1282 = 7.8 days for the U.S. and of 0.085 = 11 days for 238

Canada. 239

5 Discussion 240

In keeping with other published findings for these and other countries, our results 241

suggest that the COVID-19 countermeasures taken in South Korea, Italy, Canada and 242

the United States have had a substantial impact on the course of the disease. Even 243

accounting for estimated under-reporting, the number of cases in these countries 244

appears thus far to have been suppressed by roughly one order of magnitude in Italy 245

and the U.S., two orders of magnitude in Canada, and three orders of magnitude in 246

South Korea. The development of effective vaccines and treatments is still critical to the 247

future control of this disease, however in the interim, non-pharmaceutical interventions 248

are the only recourse and can be effective. 249

Modeling studies and the United States case show that, barring a proportion of 250

July 26, 2020 17/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

Fig 10. Predicted daily cases in the South Korea under different rates of isolation due
to testing, or testing plus contact-tracing, accompanied by a cessation of distancing
measures.

asymptomatic cases so large that the majority of people have already been infected, a 251

second wave of disease is inevitable if distancing measures are halted or relaxed early. 252

As shown in the case of the US, it is vital not to rush into a relaxing of distancing 253

measures. As illustrated in Figures 9 to 10, if a change in control strategy causes Ref f 254

to exceed 1, how quickly a second wave builds depends on the number of cases at the 255

time the change has occurred. South Korea has a small number of cases, so (slightly) 256

exceeding Ref f = 1 would result in a gradual climb of cases. 257

In this work, we have attempted to quantify the level of testing which would be 258

needed to allow a country to make a near-complete return to a normal functioning of its 259

society. Among the countries considered here, we estimate that a frequency of isolating 260

individuals based on testing combined with contact-tracing raging from once every 6.2 261

days (Italy) to once every 3.5 days (U.S.) would work to keep the pandemic under a 262

“slow-burn” control. These estimates assume a test with sensitivity at or near 100% and 263

immediate isolation once a subject tests positive. Though reaching these targets would 264

necessitate an undeniably large logistic effort, home-test kits availability4 combined 265

5
with further advances in mobile device-based contact tracing can make these strategies 266

4 A recent discussion on home-test kits this can be found in “COVID-19: A cheap, simple way to control

the coronavirus” analysis https://www.nytimes.com/2020/07/03/opinion/coronavirus-tests.html

5 Canada started to test its electronic platform for volunteered participation in contact tracing in

July 2020 with Ontario being first province to test it on a larger scale: https://www.cp24.com/news/

July 26, 2020 18/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

Fig 11. Predicted daily cases in the U.S. under different rates of isolation due to
testing, or testing plus contact-tracing, accompanied by a cessation of distancing
measures.

possible. 267

Our work is subject to a number of limitations. We are simulating large, 268

heterogeneous, geographically widely-distributed populations with an unstratified 269

disease transmission model. Although we have taken the proportion of asymptomatic 270

COVID-19 cases to be 50%, informed by a recent CDC report, assuming asymptomatic 271

infection confers immunity, this would mean a smaller remaining pool of susceptibles 272

and thus a lower current effective reproduction number. Estimates of R0 from time 273

series data of cases depend, as always, on the assumed latent and infectious periods. As 274

we have demonstrated through (Fig 7), if these periods are longer than the isolation 275

time, then it is the latter which principally drives the disease dynamics. Thus, our 276

findings about threshold isolation times are relatively robust against the possibility of a 277

substantially longer COVID-19 serial interval. 278

6 Conclusion 279

Testing and tracing policy directions must be strongly dependant on public cooperation 280

and compliance. Populations will become anxious to resume more normal work, school 281

and social schedules, while compliance with measures will become harder to enforce. 282

covid-19-alert-app-starts-beta-testing-after-three-week-delay-1.5036434

July 26, 2020 19/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

The population must be relied on to comply with self-isolation if testing positive for the 283

virus, as well as self-isolation upon being exposed to an infected person. Last, 284

sustainable supply chains, accuracy and reliability of possible tests as well as privacy 285

issues around electronic contact tracing technology all present important, though not 286

insurmountable, hurdles countries must solve. Absent universal availability of effective 287

vaccines and treatments, the testing and tracing policies together with NPI measures are 288

much more desirable and should be the one to strive for in the immediate short-term. 289

References 290

1. Guan Wj, Ni Zy, Hu Y, Liang Wh, Ou Cq, He Jx, et al. Clinical characteristics 291

of coronavirus disease 2019 in China. New England journal of medicine. 292

2020;382(18):1708–1720. 293

2. Anastassopoulou C, Russo L, Tsakris A, Siettos C. Data-based analysis, 294

modelling and forecasting of the COVID-19 outbreak. PloS one. 295

2020;15(3):e0230405. 296

3. Neher RA, Dyrdak R, Druelle V, Hodcroft EB, Albert J. Potential impact of 297

seasonal forcing on a SARS-CoV-2 pandemic. Swiss medical weekly. 298

2020;150(1112). 299

4. Lauer SA, Grantz KH, Bi Q, Jones FK, Zheng Q, Meredith HR, et al. The 300

incubation period of coronavirus disease 2019 (COVID-19) from publicly reported 301

confirmed cases: estimation and application. Annals of internal medicine. 302

2020;172(9):577–582. 303

5. Tuite AR, Bogoch I, Sherbo R, Watts A, Fisman DN, Khan K. Estimation of 304

COVID-2019 burden and potential for international dissemination of infection 305

from Iran. medRxiv. 2020;. 306

6. Centor RM, Fisman DN. Annals On Call-Understanding the Spread of 307

COVID-19. Annals of Internal Medicine. 2020; p. OC1–OC1. 308

7. Tuite AR, Ng V, Rees E, Fisman D. Estimation of COVID-19 outbreak size in 309

Italy. The Lancet Infectious Diseases. 2020;20(5):537. 310

July 26, 2020 20/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

8. Ferguson N, Laydon D, Nedjati Gilani G, Imai N, Ainslie K, Baguelin M, et al. 311

Report 9: Impact of non-pharmaceutical interventions (NPIs) to reduce 312

COVID19 mortality and healthcare demand. 2020;. 313

9. Ferguson LDNGGea N M. Impact of non-pharmaceutical interventions (NPIs) to 314

reduce COVID-19 mortality and healthcare demand. 315

2020;doi:https://doi.org/10.25561/77482. 316

10. Kissler SM, Tedijanto C, Goldstein E, Grad YH, Lipsitch M. Projecting the 317

transmission dynamics of SARS-CoV-2 through the postpandemic period. 318

Science. 2020;doi:10.1126/science.abb5793. 319

11. Kermack WO, McKendrick AG. A contribution to the mathematical theory of 320

epidemics. Proceedings of the royal society of london Series A, Containing papers 321

of a mathematical and physical character. 1927;115(772):700–721. 322

12. Ma J. Estimating Epidemic Exponential Growth Rate And Basic Reproduction 323

Number. Infectious Disease Modelling. 2020;. 324

13. Dong E, Du H, Gardner L. An interactive web-based dashboard to track 325

COVID-19 in real time. The Lancet Infectious Diseases. 2020;. 326

14. Cereda D, Tirani M, Rovida F, Demicheli V, Ajelli M, Poletti P, et al. The early 327

phase of the COVID-19 outbreak in Lombardy. Italy arXiv. 2020;2003. 328

15. Ganyani T, Kremer C, Chen D, Torneri A, Faes C, Wallinga J, et al. Estimating 329

the generation interval for coronavirus disease (COVID-19) based on symptom 330

onset data, March 2020. Eurosurveillance. 2020;25(17):2000257. 331

16. Tindale L, Coombe M, Stockdale J, Garlock E, Lau W, Saraswat M, et al. 332

Transmission interval estimates suggest pre-symptomatic spread of COVID-19. 333

medRxiv. Preprint] doi. 2020;10(2020.03):03–20029983. 334

17. Bi Q, Wu Y, Mei S, Ye C, Zou X, Zhang Z, et al. Epidemiology and transmission 335

of COVID-19 in 391 cases and 1286 of their close contacts in Shenzhen, China: a 336

retrospective cohort study. The Lancet Infectious Diseases. 2020;. 337

July 26, 2020 21/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

18. McNeil DGJ. The virus can be stopped, but only with harsh measures, experts 338

say. The New York Times. 2020;. 339

19. D’Emilio F. Europe hopeful as Italy’s daily coronavirus death toll falls to lowest 340

in weeks. The Associated Press. 2020;. 341

20. News C. Coronavirus: Here’s what’s happening on March 22. CBC News. 2020;. 342

21. Alvarez P. Here’s what Trump’s coronavirus emergency declaration does. CNN. 343

2020;. 344

22. Using a delay-adjusted case fatality ratio to estimate under-reporting; 2020. 345

https://cmmid.github.io/topics/covid19/severity/global_cfr_ 346

estimates.html, retrieved 2020-07-18. 347

23. van den Driessche P. Reproduction numbers of infectious disease models. 348

Infectious Disease Modelling. 2017;2(3):288–303. 349

A Appendix 350

A.1 Local exponential growth around a disease free 351

equilibrium with s(0) ≤ 1 in an SEIRL model 352

We want to find a relation between the exponential growth of the infected compartment 353

in an SEIRL model (10) and the reproductive number R0 around a disease-free 354

equilibrium of the type (s(0) = s̃ ≤ 1, 0, 0, 0) which arises as a possibility in a first wave 355

(s̃ = 1) or a second wave of a pandemic such as COVID-19 (s̃ < 1). 356

In this case, we conduct a similar computation as in [12], but considering the 4 357

dimensional system of equations for s, e, i, l leads us to the Jacobian of the SEIRL: 358

 
−βi 0 −βs 0
 
 
 βi −(σ + κ1 ) βs 0 
J = .
 
 0 σ −(γ + κ) 0 
 
 
0 0 κ 0

July 26, 2020 22/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

If computed at the disease free equilibrium (s̃, 0, 0, 0) we further obtain: 359

 
0 0 −βs̃ 0
 
 
 0 −(σ + k1 ) βs̃ 0 
J = .
 
 0 σ −(γ + κ) 0 
 
 
0 0 κ + κ1 0

Again we note that the linearized equations for s and l are decoupled from the equations 360

of e and i, thus, to get information on the growth rate of the infected compartment, let 361

us try to solve the linearized reduced system in (e, i) based on the reduced Jacobian: 362

 
 −(σ + κ1 ) βs̃  Solve:
Jreduced =   =⇒ det(ρI2 − Jreduced ) = 0
σ −(γ + k)

Its characteristic equation is: 363

ρ2 − ρ((σ + κ1 ) + (γ + κ)) + (σ + κ1 )(γ + κ) − σβs̃ = 0

The eigenvalues of this matrix can be computed to be 364

−(σ + κ1 + γ + κ)
ρ1,2 = ±
2
p 365

(σ + κ1 + γ + κ)2 − 4((σ + κ1 )(γ + κ) − σβs̃)

⇔
2

p
−(σ + κ1 + γ + κ) ± ((σ + κ1 ) − (γ + κ))2 + 4σβs̃
ρ1,2 = (15)
2

We first note that ((σ + κ1 ) − (γ + κ))2 + 4σβs̃ > 0, given all parameters are 366

positive. This implies that ρ1 6= ρ2 ∈ R and clearly ρ2 < 0. We check whether ρ1 > 0 by 367

looking at 368

p
((σ + κ1 ) − (γ + κ))2 + 4λβs̃ > σ + κ1 + γ + κ ⇔

(σ + κ1 ) − (γ + κ))2 + 4λβs̃ > ((σ + κ1 ) + (γ + κ))2 ⇔

July 26, 2020 23/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

−2(σ + κ1 )(γ + κ) + 4σβs̃ > 2(σ + κ1 )(γ + κ) =⇒

369

(σ + κ1 )(γ + κ)
σβs̃ > (σ + κ1 )(γ + κ) ⇔ βs̃ >
σβ

Therefore, as before, we have that 370

(σ + κ1 )(γ + κ)
ρ1 > 0 as long as s̃ > (16)
σβ

Inequality (16) simply shows that in order to not have an exponential growth from 371

our disease free equilibrium (in other words the infection dies out), we need to allow 372

that the initial fraction of susceptibles is lower than 373

(σ + κ1 )(γ + κ)
s̃ ≤
σβ

We note that β and γ are disease-dependent values on which we cannot exert control. 374

However, κ and κ1 are parameters on which we can exert an exogenous control 375

(specifically to increase them, thus raising the upper bound on fractions s̃ with no 376

exponential growth in infected) which will be outlined in detail in the next section. 377

Continuing as in [12], we express β as a function of s̃, σ, γ, κ, κ1 from (15) and we 378

get: 379

p
2ρ1 + (σ + κ1 ) + (γ + κ) = ((σ + κ1 ) − (γ + κ))2 + 4σβs̃ =⇒

4ρ21 + (σ + κ1 )2 + 2(σ + κ1 )(γ + κ) + (γ + κ)2 + 4ρ1 (σ + κ1 ) + 4ρ1 (γ + κ) =

380

2 2
(σ + κ1 ) − 2(σ + κ1 )(γ + κ) + (γ + κ) + 4σβs̃ ⇐⇒

4ρ21 + 4(σ + κ1 )(γ + κ) + 4ρ1 (σ + κ1 ) + 4ρ1 (γ + κ) = 4σβs̃ =⇒

381

ρ21 + (σ + κ1 )(γ + κ) + ρ1 (σ + κ1 ) + ρ1 (γ + κ)
= β ⇐⇒
σs̃

July 26, 2020 24/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

382
(σ + κ1 + ρ1 )(γ + κ + ρ1 )
=β
σs̃

Following [23], we can use the next generation matrix to deduce R0 as the dominant 383

eigenvalue of the next generation matrix: 384

 
βσs̃ βs̃
(σ+κ1 )(γ+κ) γ+κ βσs̃
F V −1 =   =⇒ R0 =
 
0 0 (σ + κ1 )(γ + κ)

Now using the expression of β in that of R0 we are able to express the effective 385

reproductive number as a function of the exponential growth and of γ, σ, κ, κ1 : 386

σs̃ (σ + κ1 + ρ1 )(γ + κ + ρ1 )
R0 = =⇒
(σ + κ1 )(γ + κ) σs̃
387
(σ + κ1 + ρ1 )(γ + κ + ρ1 )
R0 = and Ref f = s(t)R0 , ∀t > 0
(σ + κ1 )(γ + κ)

Similar to (8) we denote by R0S(Q)EIRL : 388

(σ + κ1 + ρ1 )(γ + κ + ρ1 )
R0SEIRL =
(γ + κ)σ

Clearly, if κ = κ1 = 0 then R0SEIRL reduces to R0SEIR in (8) of Section 2. 389

Let us now note that we have shown that the exponential growth factor (15), as well 390

as the R0SEIRL , are dependent on the rates κ and κ1 , that is to say, we denote by 391

p
−(σ + κ1 + γ + κ) + ((σ + κ1 ) − (γ + κ))2 + 4σβs̃
ρ(κ, κ1 ) = ,
2
392
(σ + κ1 + ρ1 )(γ + κ + ρ1 )
and by R0 (κ, κ1 ) =
(γ + κ)σ

A.2 The reproductive number as a function of Tlat 393

1
Let us express the reproductive number, in general, as a function of Tlat = σ and 394

1 1
Tser = Tinf + Tlat = = Tser − Tlat =⇒ γ = . 395
γ Tser − Tlat

July 26, 2020 25/26

 medRxiv preprint doi: https://doi.org/10.1101/2020.04.23.20077503.this version posted July 26, 2020. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.

From 8 we have that 396

1 1
(ρ + )(ρ + )
(ρ + γ)(ρ + σ) Tlat Tser − Tlat
R0 = =
γσ 1
Tlat (Tser − Tlat )
397
= (ρTlat + 1)(ρ(Tser − Tlat ) + 1).

Then 398

R0 = −(Tlat ρ − Tser ρ − 1)(Tlat ρ + 1) =⇒

399

dR
= −ρ(Tlat ρ + 1) − (Tlat ρ − Tser ρ − 1)ρ
dTlat

where we can solve for a Tlat value which maximizes R0 , namely 400

dR Tser
= 0 ⇐⇒ −2Tlat ρ + Tser ρ =⇒ Tlat = .
dTlat 2

July 26, 2020 26/26