2020 06 23 20111419v2 Full PDF

medRxiv preprint doi: https://doi.org/10.1101/2020.05.23.20111419.this version posted June 23, 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(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
It is made available under a CC-BY 4.0 International license .
Visualizing the invisible:

The effect of asymptomatic transmission on the outbreak dynamics of COVID-19
Mathias Peirlincka , Kevin Linkaa , Francisco Sahli Costabalb , Jay Bhattacharyac , Eran Bendavidc , John P.A.
Ioannidisc,d , Ellen Kuhla
a Departmentof Mechanical Engineering, Stanford University School of Engineering, Stanford, California, United States
b Departmentof Mechanical and Metallurgical Engineering and Institute for Biological and Medical Engineering, Schools of Engineering,
Biology and Medicine, Pontificia Universidad Catolica de Chile, Santiago, Chile
c Department of Medicine, Stanford University School of Medicine, Stanford, California, United States
d Department of Epidemiology and Population Health, Stanford University School of Medicine, Stanford, California, United States
Abstract
Understanding the outbreak dynamics of the COVID-19 pandemic has important implications for successful contain-
ment and mitigation strategies. Recent studies suggest that the population prevalence of SARS-CoV-2 antibodies, a
proxy for the number of asymptomatic cases, could be an order of magnitude larger than expected from the number
of reported symptomatic cases. Knowing the precise prevalence and contagiousness of asymptomatic transmission
is critical to estimate the overall dimension and pandemic potential of COVID-19. However, at this stage, the effect
of the asymptomatic population, its size, and its outbreak dynamics remain largely unknown. Here we use reported
symptomatic case data in conjunction with antibody seroprevalence studies, a mathematical epidemiology model, and
a Bayesian framework to infer the epidemiological characteristics of COVID-19. Our model computes, in real time,
the time-varying contact rate of the outbreak, and projects the temporal evolution and credible intervals of the effec-
tive reproduction number and the symptomatic, asymptomatic, and recovered populations. Our study quantifies the
sensitivity of the outbreak dynamics of COVID-19 to three parameters: the effective reproduction number, the ratio
between the symptomatic and asymptomatic populations, and the infectious periods of both groups. For nine distinct
locations, our model estimates the fraction of the population that has been infected and recovered by Jun 15, 2020 to
24.15% (95% CI: 20.48%-28.14%) for Heinsberg (NRW, Germany), 2.40% (95% CI: 2.09%-2.76%) for Ada County
(ID, USA), 46.19% (95% CI: 45.81%-46.60%) for New York City (NY, USA), 11.26% (95% CI: 7.21%-16.03%) for
Santa Clara County (CA, USA), 3.09% (95% CI: 2.27%-4.03%) for Denmark, 12.35% (95% CI: 10.03%-15.18%) for
Geneva Canton (Switzerland), 5.24% (95% CI: 4.84%-5.70%) for the Netherlands, 1.53% (95% CI: 0.76%-2.62%)
for Rio Grande do Sul (Brazil), and 5.32% (95% CI: 4.77%-5.93%) for Belgium. Our method traces the initial out-
break date in Santa Clara County back to January 20, 2020 (95% CI: December 29, 2019 - February 13, 2020). Our
results could significantly change our understanding and management of the COVID-19 pandemic: A large asymp-
tomatic population will make isolation, containment, and tracing of individual cases challenging. Instead, managing
community transmission through increasing population awareness, promoting physical distancing, and encouraging
behavioral changes could become more relevant.
Keywords: COVID-19; epidemiology; asymptomatic transmission; machine learning; uncertainty quantification;
Bayesian inference
1. Motivation
Since its outbreak in December 2019, the COVID-19 pandemic has rapidly swept across the globe and is now
affecting 188 countries with more than 5 million cases reported worldwide [9]. In the early stages of a pandemic,
doctors, researchers, and political decision makers mainly focus on symptomatic individuals that come for testing
and address those who require the most urgent medical attention [14]. In the more advanced stages, the interest
shifts towards mildly symptomatic and asymptomatic individuals who–by definition–are difficult to trace and likely
to retain normal social and travel patterns [34]. In this manuscript, we collectively use the term “asymptomatic” for
Preprint submitted to medRxiv June 23, 2020

medRxiv preprint doi: https://doi.org/10.1101/2020.05.23.20111419.this version posted June 23, 2020. The copyright holder for this preprint
individuals who have mild symptoms that are not directly associated with COVID-19 or display no symptoms at all.
Recent antibody seroprevalence studies suggests that the number of asymptomatic COVID-19 cases outnumbers the
symptomatic cases by an order of magnitude or more [3, 4, 8, 11, 13, 17, 23, 28, 30, 37, 46, 47, 52, 53, 56, 54, 55, 59,
60, 62, 64, 66, 68]. Estimating the prevalence and contagiousness of these asymptomatic cases is critical since it will
change our understanding of the overall dimension and the pandemic potential of COVID-19 [15]. Yet, at this stage,
the effect of the asymptomatic population, its size, and its outbreak dynamics remain largely unknown.
The first evidence of asymptomatic individuals in a family cluster of three was reported in late January, where
one individual was mildly symptomatic and two remained asymptomatic, with normal lymphocyte counts and chest
computer tomography images, but positive quantitative reverse transcription polymerase chain reaction tests [41]. As
of today, more than 50 studies have reported an asymptomatic population, twenty-three of them with a sample size of
at least 500 [27], with a median undercount of 20 across all studies, suggesting that only one in twenty COVID-19
cases is noticed and reported. These studies are based on polymerase chain reaction or antibody seroprevalence tests
in different subgroups of the population, at different locations, at different points in time [3, 6, 59]. To no surprise,
the reported undercount varies hugely, ranging from 3.5 and 5.0 in Luxembourg [56] and Germany [59] to 543 and
627 in Iran [53] and Japan [8] respectively. Most of these studies are currently only available on preprint servers,
but an increasing number is now passing peer review, including a study of 1402 individuals in Wuhan City with an
undercount of 22.1 [68], a study of 400 health care workers in London with an undercount of 35.0 [65], a community
spreading study of 131 patients with influenza-like symptoms in Los Angeles with an undercount of 100.0 [58], and a
seroprevalence study in Los Angeles county with an undercount of 43.5 [57]. The reported trend across all studies is
strikingly consistent: A much larger number of individuals displays antibody prevalence than we would expect from
the reported symptomatic case numbers. Knowing the exact dimension of the asymptomatic population is critical
for two reasons: first, to truly estimate the severity of the outbreak, e.g., hospitalization or mortality rates [15], and
second, to reliably predict the success of surveillance and control strategies, e.g., contact tracing or vaccination [18].
While there is a pressing need to better understand the prevalence of asymptomatic transmission, it is also be-
coming increasingly clear that it will likely take a long time until we can, with full confidence, deliver reliable mea-
surements of this asymptomatic group. In the meantime, mathematical modeling can provide valuable insight into
the tentative outbreak dynamics and outbreak control of COVID-19 for varying asymptomatic scenarios [34]. Many
classical epidemiology models base their predictions on compartment models in which individuals pass through dif-
ferent stages as they experience the disease [29]. A popular model to simulate the outbreak dynamics of COVID-19 is
the SEIR model [14], which is made up of four compartments for the susceptible, exposed, infectious, and recovered
populations [2]. Here, to explicitly account for the asymptomatic population, we introduce an SEIIR model, which
further divides the infectious population into symptomatic and asymptomatic groups. Similar models have recently
been used to study the general role of asymptomatic carriers in disease transmission [42] and to illustrate how asymp-
tomatic individuals have facilitated the rapid spread of COVID-19 throughout China [34], South Korea [63], and Italy
[19]. While it is tempting–and easily possible–to introduce many more sub-populations into the model, for example
a pre-symptomatic, hospitalized, or mortality group [45], here, we focus on the simplest possible model that allows
us to explore the role of the asymptomatic population throughout the COVID-19 pandemic. To systematically probe
different scenarios, we combine this deterministic SEIIR model with a dynamic effective reproduction number and
adopt machine learning and uncertainty quantification techniques to learn the reproduction number, in real time, and
quantify uncertainties in the symptomatic-to-asymptomatic ratio, and the initial exposed and infectious populations
[36]. We show that this not only allows us to visualize the dynamics and uncertainties of the dynamic contact rate, the
effective reproduction number, and the symptomatic, asymptomatic, and recovered populations, but also to estimate
the initial date of the outbreak.
2. Methods
2.1. Epidemiology modeling.

We model the epidemiology of COVID-19 using an SEIIR model with five compartments, the susceptible, ex-
posed, symptomatic infectious, asymptomatic infectious, and recovered populations. Figure 1 illustrates our SEIIR
2
Figure 1: SEIIR epidemiology model. The SEIIR model contains five compartments for the susceptible, exposed, symptomatic infectious,
asymptomatic infectious, and recovered populations. The transition rates between the compartments, β, α, and γ are inverses of the contact period
B = 1/β, the latent period A = 1/α, and the infectious period C = 1/γ. The symptomatic and asymptomatic groups have the same latent period A,
but they can have individual contact periods Bs = 1/βs and Ba = 1/βa and individual infectious periods Cs = 1/γs and Ca = 1/γa . The fractions
of the symptomatic and asymptomatic subgroups of the infectious population are νs and νa . We assume that the infection either goes through the
symptomatic or the asymptomatic path, but not both for one individual.
model, which is governed by a set of five ordinary differential equations,
Ṡ = − S [ βs Is + βa Ia ]
Ė = + S [ βs Is + βa Ia ] − αE
I˙s = + νs α E − γs I s (1)
I˙a = + νa α E − γa Ia
Ṙ = + γs Is + γa Ia ,
where the fractions of all five populations add up to one, S + E + Is + Ia + R = 1. We assume that both the symptomatic
group Is and the asymptomatic group Ia can generate new infections. We introduce these two groups by fractions νs
and νa of the subjects transferring from the exposed group E. We postulate that the two infectious groups Is and Ia
have the same latent period A = 1/α, but can have individual contact periods Bs = 1/βs and Ba = 1/βa to mimic their
different community spreading, and individual infectious periods Cs = 1/γs and Ca = 1/γa to mimic their different
likelihood of isolation. From the infectious fractions, we can derive the overall contact and infectious rates β and γ
from their individual symptomatic and asymptomatic counterparts, βs , βa , γs , and γa ,
β = νs βs + νa βa and γ = νs γs + νa γa . (2)
Similarly, we can express the overall contact and infectious periods B and C in terms of their symptomatic and
asymptomatic counterparts, Bs , Ba , Cs , and Ca ,
Ba Bs Ca Cs
B= and C = . (3)
νs Ba + νa Bs νs Ca + νa Cs
Naturally, the different dynamics for the symptomatic and asymptomatic groups also affect the basic reproduction
number R0 , the number of new infections caused by a single one individual in an otherwise uninfected, susceptible
population,
C Ca Cs νs Ba + νa Bs νs βs + νa βa β
R0 = = = = . (4)
B Ba Bs νs Ca + νa Cs νs γs + νa γa γ
For a large asymptomatic group νa → 1, the basic reproduction number approaches the ratio between the infectious
and contact periods of the asymptomatic population, R0 → Ca /Ba , which could be significantly larger than the basic
reproduction number for the symptomatic group, R0 = Cs /Bs , that we generally see reported in the literature. To
characterize the effect of changes in social behavior and other interventions that may affect contact, we assume that
the contact rate β(t) can vary as a function of time [36], but is the same for the symptomatic and asymptomatic groups,
C Ca Cs β(t) β(t)
β = βs = βa = β(t) such that R(t) = = = = . (5)
B(t) [νs Ca + νa Cs ] B(t) νs γs + νa γa γ
This introduces a time-varying effective reproduction number R(t), which is an important real time characteristic of
the current outbreak dynamics. For the special case when the dynamics of the symptomatic and asymptomatic groups
are similar, i.e., βs = βa = β and γs = γa = γ, we can translate the SEIIR model (1) into the classical SEIR model (6)
with four compartments, the susceptible, exposed, infectious, and recovered populations [35]. For this special case,
3
Figure 2: SEIR epidemiology model. The SEIR model contains four compartments for the susceptible, exposed, infectious, and recovered
populations. The transition rates between the compartments, β, α, and γ are inverses of the contact period B = 1/β, the latent period A = 1/α, and
the infectious period C = 1/γ. If the transition rates are similar for the symptomatic and asymptomatic groups, the SEIIR model simplifies to the
SEIR model with Is = νs I and Ia = νa I.
we can back-calculate the symptomatic and asymptomatic groups from equation (6.3) as Is = νs I and Ia = νa I. Figure
2 illustrates the SEIR model, which is governed by a set of four ordinary differential equations [25],
Ṡ = −β SI
Ė = +β SI − αE
(6)
I˙ = + αE − γI
Ṙ = + γI.
2.2. Seroprevalence studies.

We consider seroprevalence studies where the representation ratio, the ratio between the amount of seropreva-
lence samples and the representative population, is larger than 0.02%. Subsequently, we draw the daily number of
confirmed reported cases for Heinsberg, Ada County, New York City, Santa Clara County, Denmark, Geneva Canton,
Netherlands, Rio Grande do Sul, and Belgium from online data repositories [5, 10, 24, 39, 40, 61]. The symp-
tomatic fraction νs follows from the ratio between the reported confirmed cases for the studied region and the amount
of seroprevalence-estimated cases on the last day of the seroprevalence study. Table 1 summarizes the study date,
number of samples, representation ratio, population, and symptomatic fraction of each location.
Table 1: Seroprevalence studies with a representative population larger than 0.02%, location, study date, number of samples, representation
ratio, population, and symptomatic fraction νs .
Location Study date Samples Population Repr.ratio Symptomatic fraction νs Ref.
Heinsberg, NRW, Germany March 30 - April 6 919 41946 2.191% 20.00% (95%CI: 16.32% - 25.20%) [59]
Ada County, ID, USA Late April 4856 481587 1.008% 7.90% (95%CI: 6.40%-9.82%) [4]
New York City, NY, USA May 5 - Jun 5 28523 8398748 0.340% 5.76% (95%CI: 5.68%-5.84%) [46]
Santa Clara County, CA, USA April 2 - 3 3330 1781642 0.187% 1.77% (95%CI: 1.05% - 3.82%) [3]
Denmark April 6 - 17 9496 5824857 0.163% 6.95% (95%CI: 5.13%-13.12%) [13]
Geneva Canton, Switzerland April 20-27 576 504031 0.114% 10.34% (95%CI: 7.66%-16.44%) [60]
Netherlands April 1-15 7361 17282163 0.043% 17.31% (95%CI: 15.14%-19.95%) [55]
Rio Grande do Sul, Brasil April 25 - 27 4188 11377239 0.037% 8.10% (95%CI: 3.72%-22.07%) [54]
Belgium April 14 - May 13 2700 11492641 0.023% 10.21% (95%CI: 8.61%-12.18%) [52]
2.3. Uncertainty quantification.

Our SEIIR model uses the following set of parameters, ϑ = { A, Cs , Ca , νs , β(t), E0 , Is0 , Ia0 , σ }. To reduce the
set of unknowns, we fix the latency period A = 2.5 days and the symptomatic infectious period Cs = 6.5 days
[32, 33, 50]. Since the asymptomatic infectious period Ca is unreported, we study three cases with Ca : Cs = 0.5, 1.0,
and 2.0, resulting in infectious periods of Ca = 3.25, 6.5, and 13.0 days. We impose the symptomatic fraction of the
infectious group, νs to follow the distributions from Table 1. We estimate the remaining parameters including the time-
varying contact rate β(t), the initial exposed population E0 , and the initial symptomatic and asymptomatic infectious
populations Is0 and Ia0 using Bayesian inference. For the time-varying contact rate β(t), we set a log-Gaussian random
walk prior, which we construct with weakly informative priors with a drift µRW and a daily step width σRW . For the
initial symptomatic infectious population Is0 , we set a weakly informative log-normal prior distribution with a mean
equal to the amount of detected confirmed cases on day 0 and a standard deviation equal to one. We express the
initial asymptomatic infectious population as Ia0 = [ 1 − νs ]/νs Is0 , and approximate the initial exposed population as
E0 = [ A/Cs ] [ Is0 + Ia0 ] [12]. Table 2 summarizes the choice of our priors.
For each parameter set, we deterministically calculate the time series of each compartment using an explicit time
integrator. For each point in time, we calculate the detected cases, D(t) = Is (t) + Rs (t) with Ṙs (t) = γ s Is (t). By scaling
4
Table 2: Prior distributions for SEIIR model parameters.
Parameter Interpretation Distribution

νs symptomatic fraction Table 1
A latent period fixed; 2.5 days [33, 32, 45]
Cs symptomatic infectious period fixed; 6.5 days [33, 50]
Ca asymptomatic infectious period fixed; 3.25 or 6.5 or 13.0 days [33, 50]
log(β(t)) dynamic contact rate Gaussian random walk(µRW , σRW )
µRW drift normal(µ = 0.0, σ = 1.0)
σRW daily step width half-normal(σ = 0.02)
E0 initial exposed determinstic([ A/Cs ] [ Is0 + Ia0 ]) [12]
Is0 initial symptomatic log-normal(D̂(t = 0), 1.0)
Ia0 initial asymptomatic deterministic([ 1 − νs ]/νs Is0 )
σ likelihood width half-Cauchy(β = 1)
the confirmed reported cases with the total population N, we obtain the relative detected population D̂(t). The day
on which each location went into lockdown marks day 0 and the beginning of our simulation. From this day on,
we calculate the simulated detected population, D(t) = Is (t) + Rs (t) with Ṙs (t) = γs Is (t). We quantify the likelihood
of the parameter set and model outcome D(t) in correlation to the reported cases D̂(t) [51, 40, 24], using Student’s
t-distribution,
p(D̂(t) | D(t, ϑ)) ∼ student Tν=4 ( mean = D(t, ϑ); width = σ ). (7)
We choose this distribution because it resembles a Gaussian distribution and makes the Markov-Chain Monte Carlo
more robust with respect to outliers [7, 31]. Here, σ represents the width of the likelihood p(D̂(t) | ϑ) between the time-
varying reported and the modeled symptomatic populations. Using Bayes’ rule, we compute the posterior distribution
of the parameters [44, 48] to account for the prior knowledge on the parameters and the reported confirmed cases
themselves,
p(D̂(t) | D(t, ϑ)) p(ϑ)

p(ϑ | D̂(t)) = . (8)
p(D̂(t))
Since we cannot describe the posterior distribution over the model parameters ϑ analytically, we adopt approximate-
inference techniques to calibrate our model on the available data. We use the NO-U-Turn sampler (NUTS) [26], which
is a type of Hamiltonian Monte Carlo algorithm as implemented in PyMC3 [49]. We use four chains, and the first 4
times 500 samples serve to tune the sampler and are later discarded. We use the subsequent 4 times 1000 samples
as the posterior distribution for the parameters ϑ. From the converged posterior distribution, we sample multiple
combinations of parameters that describe the time evolution of reported cases. Using these posterior samples, we
quantify the uncertainty of each parameter based on the reported case data. As such, each parameter set provides a set
of values for the initial exposed population E0 , the initial symptomatic and asymptomatic populations Is0 and Ia0 , the
symptomatic fraction νs , and the time evolution of the contact rate β(t). From these values, we quantify the effective
reproductive number R(t) and the time evolution of the susceptible, exposed, symptomatic infectious, asymptomatic
infectious and recovered populations, S (t), E(t), Is (t), Ia (t), and R(t) and report their values with the associated 95%
credible interval.
2.4. Hierarchical asymptomatic infectious period Ca estimation.

To estimate the asymptomatic infectious period Ca , we combine the confirmed case data from all locations and set
up a hierarchical model for Ca . Here we postulate that each location-specific Caloc i is drawn from a hyperdistribution
Ca . We assume that the initial reproduction number, which is disease-specific, is the same for all locations [16] and is
constant during a two-week time window around each local lockdown date: 10 days prior till 3 days (∼ A, the latent
period) after each local lockdown date. The priors for these distributions are summarized in Table 3. We perform
5
Bayesian inference as explained above and subsequently use the hierarchical posterior Ca distribution to quantify the
time-varying asymptomatic infectious population size at each of the locations after the lockdowns.
Table 3: Prior distributions for SEIIR-based hierarchical Ca estimation.
Parameter Interpretation Distribution

νs symptomatic fraction Table 1
A latent period fixed; 2.5 days [33, 32, 45]
Cs symptomatic infectious period fixed; 6.5 days [33, 50]
Ca asymptomatic infectious period normal(µ = 6.5, σ = σCa )
σCa half-normal(σ = 2.0)
Caloc i normal(µ = Ca , σ = 2.0)
R0 initial reproduction number normal(µ = 3.87, σ = 0.44) [16]
E0 initial exposed determinstic([ A/Cs ] [ Is0 + Ia0 ]) [12]
Is0 initial symptomatic log-normal(D̂(t = 0), 1.0)
Ia0 initial asymptomatic deterministic([ 1 − νs ]/νs Is0 )
σ likelihood width half-Cauchy(β = 1)
2.5. Estimating the outbreak date.

For each sample from the posterior distribution, we use the estimated initial exposed population E0 and the es-
timated initial asymptomatic infectious population Ia0 to estimate the date of the very first COVID-19 case in Santa
Clara County [51]. Specifically, for each parameter set, we create an SEIIR model and assume that the outbreak
begins with one single asymptomatic infectious individual. We fix the latency and symptomatic infectious periods
to A = 2.5 days and Cs = 6.5 days and impose the hierarchically estimated infectious period Ca . For each posterior
sample for the exposed, symptomatic infectious, and asymptomatic infectious population size on the lockdown date
(March 16, 2020), we use the Nelder-Mead optimization method [20] to find the most probable outbreak origin date.
Specifically, we solve the SEIIR model forward in time using an explicit time integration, starting from various start
dates before March 16, 2020, and iteratively minimize the difference between the computed exposed, symptomatic,
and asymptomatic infectious populations and the sample’s actual exposed, symptomatic, and asymptomatic infectious
populations. We concomitantly fit a static contact rate parameter β which is bounded between zero and the posterior
sample’s estimated contact rate on day 0, β(0). Repeating this process for each sample of the Bayesian inference
generates a distribution of possible origin dates. From this distribution, we compute the most probable origin date and
its uncertainty.
3. Results
3.1. Outbreak dynamics of COVID-19 in Santa Clara County.

Figure 3 illustrates the outbreak dynamics of COVID-19 in Santa Clara County. To explore the effects of Ca = on
the initial effective reproductive number, we started the simulations 10 days before the local lockdown date: March
6, 2020. The black dots highlight the reported cases D̂(t) from this day forward. Based on these data points, we learn
the posterior distributions of our SEIIR model parameters for fixed latent and symptomatic infectious periods A =
2.5 days and Cs = 6.5 days, and for three asymptomatic infectious periods, Ca = 3.25, 6.5, and 13.0 days, from left
to right. The gray and green-purple regions highlights the 95% credible intervals on the confirmed cases D(t), top
row, and the contact rate β(t), bottom row, based on the reported cases D̂(t), while taking into account uncertainties
on the fraction of the symptomatic infectious fraction νs , and the initial exposed and infectious populations E0 , Is0 ,
and Ia0 . The red, orange, and green histograms on the middle row display the inferred initial exposed and infectious
populations, E0 , Is0 , and Ia0 , for the three different asymptomatic infectious periods. The graphs confirm that our
dynamic SEIIR epidemiology model is capable of correctly capturing the gradual flattening and deflattening of the
6
shorter asymptomatic infectious period equally long asymptomatic infectious period longer asymptomatic infectious period
3000 Cs = 6.5 days Cs = 6.5 days Cs = 6.5 days
Ca = 3.25 days Ca = 6.5 days Ca = 13.0 days
2500
confirmed cases
2000
reported cases D(t)
1500 computed cases D(t)
1000
500
D(Jun 15)= 3230 D(Jun 15)= 3230 D(Jun 15)= 3230
0
Apr May Jun Apr May Jun Apr May Jun
500
0.0020% 0.047% 0.119% 0.0021% 0.047% 0.119% 0.0021% 0.048% 0.124%
400
posterior samples
initial populations
300 E0
Is0
200 Ia0
100
0
0.010% 0.100% 1.000% 0.010% 0.100% 1.000% 0.010% 0.100% 1.000%
0.9
(Mar 06): 0.688 (Mar 06): 0.529 (Mar 06): 0.463
0.8 (95%CI: 0.612-0.764) (95%CI: 0.464-0.593) (95%CI: 0.403-0.523)
0.7 (Jun 15): 0.393 (Jun 15): 0.210 (Jun 15): 0.109
(95%CI: 0.315-0.470) (95%CI: 0.163-0.256) (95%CI: 0.084-0.135)
0.6
0.5
(t)
0.4
0.3
0.2
0.1
0.0
Figure 3: Outbreak dynamics of COVID-19 in Santa Clara County. The simulation learns the time-varying contact rate β(t) for fixed latent
and symptomatic infectious periods A = 2.5 days and Cs = 6.5 days, and for three asymptomatic infectious periods Ca = 3.25 days, 6.5 days, and
13.0 days (from left to right). Computed and reported confirmed cases in Santa Clara County, D(t) = Is (t) + Rs (t) and D̂(t) (top), initial exposed
and infectious populations, E0 , Is0 , and Ia0 (middle), and dynamic contact rate, β(t) (bottom). The gray and green-blue regions highlight the 95%
credible intervals on the confirmed cases D(t) (top) and the contact rate β(t) (bottom) based on the reported cases D̂(t), while taking into account
uncertainties on the fraction of the symptomatic infectious population νs = Is /I, and the initial exposed and infectious populations E0 , Is0 , and Ia0 .
curve of confirmed cases in agreement with the decrease in new cases reported in Santa Clara County, top row. The
consistent downward trend of the contact rate β(t) after the lockdown date (March 16, 2020) quantifies the efficiency
of public health interventions. The different magnitudes in the contact rate highlight the effect of the three different
asymptomatic infectious periods Ca : For larger asymptomatic infectious periods Ca , from left to right, to explain the
same number of confirmed cases D(t) = Is (t) + Rs (t), the contact rate β(t) has to decrease. On March 6, 2020, the
mean contact rate β(t) was 0.688 (95% CI: 0.612 - 0.764) for an infectious period of Ca =3.25 days, 0.529 (95% CI:
0.464 - 0.593) for Ca = 6.5 days, and 0.463 (95% CI: 0.403 - 0.523) for Ca =13.0 days. By March 16, 2020, the day
Santa Clara County announced the first county-wide shelter-in-place order in the entire United States, these contact
rates β(t) were 0.491 (95% CI: 0.462 - 0.523) for an infectious period of Ca =3.25 days, 0.328 (95% CI: 0.305 - 0.352)
for Ca = 6.5 days, and 0.252 (95% CI: 0.234 - 0.271) for Ca =13.0 days.
3.2. Effect of asymptomatic transmission of COVID-19 in Santa Clara County.

Figure 4 visualizes the effect of asymptomatic transmission in Santa Clara County. The simulation learns the
time-varying contact rate β(t), and with it the time-varying effective reproduction number R(t), top row, for three
asymptomatic infectious periods Ca = 3.25 days, 6.5 days, and 13.0 days, from left to right. The effective reproduction
number R(t) follows a similar downward trend as the contact rate β(t). For larger asymptomatic infectious periods
Ca , from left to right, since R(t) = Cs β(t)/[ νs + νa Cs /Ca ], as Ca increases, Cs /Ca decreases, and R(t) increases.
Since R(t) represents the number of new infections from a single case, a decrease below R(t) < 1 implies that a
7
Figure 4: Effect of asymptomatic transmission of COVID-19 in Santa Clara County. The simulation learns the time-varying contact rate
β(t), and with it the time-varying effective reproduction number R(t), for fixed latent and symptomatic infectious periods A = 2.5 days and Cs =
6.5 days, and for three asymptomatic infectious periods Ca = 3.25 days, 6.5 days, and 13.0 days (from left to right). The downward trend of
the effective reproduction number R(t) reflects the efficiency of public health interventions (top row). The dashed vertical lines mark the critical
time period during which the effective reproductive reproduction number fluctuates around R(t) = 1. The simulation predicts the symptomatic
infectious, asymptomatic infectious, and recovered populations Is , Ia , and R (bottom row). The colored regions highlight the 95% credible interval
for uncertainties in the number of confirmed cases D, the fraction of the symptomatic infectious population νs = Is /I, the initial exposed population
E0 and the initial infectious populations Is0 and Ia0 .
single infectious individual infects less than one new individual, which indicates that the outbreak decays. The dashed
vertical lines indicate the date R(t) = 1 during which one infectious individual, either symptomatic or asymptomatic,
infects on average one other individual. For an asymptomatic infectious period of Ca =3.25 days, it took until March
28 before Santa Clara County managed to get R(t) below 1 for the first time after the outbreak. For Ca =6.5 days,
this only occurred by April 1 and for Ca =13.0 days, this occurred on April 8, 2020. This confirms our intuition that,
the larger the asymptomatic infectious period Ca , for example because asymptomatic individuals will not isolate as
strictly as symptomatic individuals, the higher the effective reproduction number R(t), and the more difficult it will be
to control R(t) by public health interventions. For each of the three cases, the symptomatic infectious, asymptomatic
infectious, and recovered population, are shown in the bottom row. For larger asymptomatic infectious periods Ca ,
from left to right, the total infectious population I increases and its maximum occurs later in time. Specifically, the
maximum infectious population since March 6, 2020 amounts to 0.70% (95% CI: 0.43%-0.97%) on March 28, 2020
for Ca = 3.25 days, 1.23% (95% CI: 0.72%-1.75%) on April 2, 2020 for Ca = 6.5 days, and 2.10% (95% CI: 1.25%-
2.94%) on April 7, 2020 for Ca = 13.0 days. For larger asymptomatic infectious periods Ca , from left to right, the
recovered population R decreases. Specifically, on June 15, 2020, the recovered population R amounts to 10.85%
8
(95% CI: 7.06%-16.09%) for an infectious period of Ca =3.25 days, 10.20% (95% CI: 6.07%-14.90%) for Ca =6.5
days, and 9.90% (95% CI: 6.07%-13.93%) for Ca =13.0 days. Similarly, and important when considering different
exit strategies, the total infectious population, I = Is + Ia , on June 15, 2020 is estimated to 0.39% (95% CI: 0.24%-
0.58%) for Ca =3.25 days, 0.68% (95% CI: 0.40%-1.01%) for Ca =6.5 days, and 1.25% (95% CI: 0.73%-1.77%) for
Ca =13.0 days.
3.3. Outbreak dynamics of COVID-19 worldwide.

Figure 5 illustrates the hierarchical asymptomatic infectious period Ca estimation for nine locations that reported
COVID-19 antibody prevalence in a representative sample of the population: Heinsberg (NRW, Germany), Ada
County (ID, USA), New York City (NY, USA), Santa Clara County (CA, USA), Denmark, Geneva Canton (Switzer-
land), Netherlands, Rio Grande do Sul (Brazil) and Belgium. This estimate assumes an initial fixed reproduction
number R0 = 3.87 (95%CI: 3.01-4.66) during a two-week window before lockdown which results in an asymptomatic
infectious period of Ca = 5.76 (95%CI: 3.59-8.09) days. With this value, we simulate the outbreak dynamics of
COVID-19 in all nine locations. Figure 6 illustrates the learnt effective reproduction number R(t), and the symp-
tomatic and asymptomatic infectious populations Is and Ia , and the recovered population R for all nine locations.
Here, we assume fixed latent and infectious periods of A = 2.5 days, Cs = 6.5 days, and the hierarchical asymptomatic
infectious period Ca = 5.76 (95%CI: 3.59-8.09) days from Figure 5. For all nine locations, the calculated metrics
display the fine balance between the dynamics of the effective reproduction number and the control of the epidemic
outbreak. The downward trend of the effective reproductive number R(t) quantifies how fast each location managed to
control the spreading of COVID-19. The dashed vertical line indicates the first time each location managed to lower
the effective reproduction below R(t) = 1 after lockdown. This critical transition occurred on March 20 for Heinsberg,
April 4 for Ada County, April 28 for New York City, March 30 for Santa Clara County, April 8 for Denmark, April
2 for Geneva, April 14 for the Netherlands, June 8 for Rio Grande do Sul and April 11 for Belgium. Based on our
simulations, the maximum infectious population size amounted to 3.54% (95% CI: 2.96%-4.11%) on March 19 in
Heinsberg, 0.47% (95% CI: 0.41%-0.55%) on May 5 in Ada County, 6.11% (95% CI: 5.99%-6.21%) on April 13 in
New York City, 1.11% (95% CI: 0.69%-1.65%) on March 29 in Santa Clara County, 0.38% (95% CI: 0.27%-0.50%)
Heinsberg, NRW, Germany Ada County, ID, USA New York City, NY, USA
Ca = 4.06 Ca = 2.06 Ca = 2.32 Ca = 5.76 (95%CI: 3.59-8.05)
200 (95%CI: 2.32-6.59) 100 (95%CI: 1.15-3.22) 20000 (95%CI: 1.41-3.43) 0.35
confirmed cases
80 15000
150
60
100 4 6 2 3 10000 2 3
40 0.30
50 5000
computed 20
reported
0 0 0
Feb Mar Mar Mar Mar Mar
25 03 17 24 17 24 0.25
Santa Clara County, CA, USA Denmark Geneva Canton, Switzerland
500 2000
Ca = 8.71 Ca = 4.11 Ca = 8.75
(95%CI: 6.43-11.39) 1200 (95%CI: 1.77-7.83) 1750 (95%CI: 5.61-12.64)
400
confirmed cases
1000 1500
0.20
1250
Density
300 800
6 8 10 600 2 4 6 8 1000 7.5 10.0 12.5
200 750
400
500
100 0.15
200 250
0 0 0
Mar Mar Mar Mar Mar Mar
10 17 03 10 10 17
Netherlands Rio Grande do Sul, Brazil Belgium
8000 0.10
Ca = 10.36 100 Ca = 2.51 Ca = 7.50
7000 (95%CI: 7.90-13.13) (95%CI: 1.50-3.73) 4000 (95%CI: 5.55-9.77)
confirmed cases
6000 80
5000 3000
60
4000 0.05
8 10 12 2 3 4 2000 6 8 10
3000 40
2000 1000
20
1000
0 0 0 0.00
Mar Mar Mar Mar Mar Mar 2 4 6 8 10
17 24 10 17 10 17 Hyperposterior Ca [days]
Figure 5: Hierarchical asymptomatic infectious period Ca estimation. Assuming an initial fixed reproduction number R0 = 3.87 (95%CI:
3.01-4.66) [16], the simulation generates histograms of the asymptomatic infectious period Ca for each location based on the location-specific
symptomatic fraction νs . The black dots and grey regions represent the reported and simulated detected cases D̂(t) and D(t) respectively. The
hierarchical hyperdistribution for the asymptomatic infectious period results in Ca = 5.76 (95%CI: 3.59-8.09) days, right histogram.
9
Heinsberg, NRW, Germany Ada County, ID, USA New York City, NY, USA
3 R(Feb 28): 3.12 (95%CI: 2.73-3.49) 3 R(Mar 25): 4.07 (95%CI: 3.14-5.00) 3 R(Mar 22): 2.11 (95%CI: 1.81-2.41)
Mar 20
Apr 04
R(Jun 15): 0.69 (95%CI: 0.45-0.96) R(Jun 15): 2.13 (95%CI: 1.31-2.98) R(Jun 15): 1.20 (95%CI: 0.95-1.47)
2 2 2
Apr 28
R(t)
1 1 1
0 R(t) 1 0 R(t) 1 0 R(t) 1
32.0% I (Jun 15): 0.11% (95%CI: 0.06%-0.16%) 3.20% I (Jun 15): 0.12% (95%CI: 0.09%-0.15%) I (Jun 15): 0.52% (95%CI: 0.47%-0.58%)
R (Jun 15): 24.15% (95%CI: 20.48%-28.14%) R (Jun 15): 2.40% (95%CI: 2.09%-2.76%) R (Jun 15): 46.19% (95%CI: 45.81%-46.60%)
45%
24.0% 2.40%
population
30%
16.0% 1.60%
symptomatic infectious
8.0% asymptomatic infectious 0.80% 15%
recovered
0.0% 0.00% 0%
symptomatic infectious
0.60% 0.30%
population
0.03%
0.30% 0.02% 0.15%
0.00% 0.00% 0.00%
Mar Apr May Jun Apr May Jun Apr May Jun
Santa Clara County, CA, USA Denmark Geneva Canton, Switzerland
3 R(Mar 16): 1.24 (95%CI: 1.14-1.36) 3 R(Mar 13): 0.71 (95%CI: 0.59-0.84) 3 R(Mar 20): 0.92 (95%CI: 0.77-1.09)
Mar 30
Apr 08
Apr 02
2 2 2
R(t)
1 1 1
0 R(t) 1 0 R(t) 1 0 R(t) 1
I (Jun 15): 0.66% (95%CI: 0.38%-0.95%) I (Jun 15): 0.05% (95%CI: 0.04%-0.07%) I (Jun 15): 0.01% (95%CI: 0.00%-0.01%)
16.0% R (Jun 15): 11.26% (95%CI: 7.21%-16.03%) 4.00%
R (Jun 15): 3.09% (95%CI: 2.27%-4.03%) 16.0% R (Jun 15): 12.35% (95%CI: 10.03%-15.18%)
12.0% 3.00% 12.0%

population
8.0% 2.00% 8.0%
4.0% 1.00% 4.0%
0.0% 0.00% 0.0%

0.03%
0.02% 0.20%
population
0.02% 0.10%
0.01%
0.00% 0.00% 0.00%
Netherlands Rio Grande do Sul, Brazil Belgium
3 R(Mar 24): 1.42 (95%CI: 1.26-1.57) 3 R(Mar 20): 1.40 (95%CI: 1.03-1.81) 3 R(Mar 17): 2.04 (95%CI: 1.79-2.30)
Apr 14
Jun 08
Apr 11
2 2 2
R(t)
1 1 1
0 R(t) 1 0 R(t) 1 0 R(t) 1
I (Jun 15): 0.10% (95%CI: 0.09%-0.12%) I (Jun 15): 0.30% (95%CI: 0.14%-0.46%) I (Jun 15): 0.05% (95%CI: 0.04%-0.05%)
6.0% R (Jun 15): 5.24% (95%CI: 4.84%-5.70%) R (Jun 15): 1.53% (95%CI: 0.76%-2.62%) 6.0% R (Jun 15): 5.32% (95%CI: 4.77%-5.93%)
2.40%
4.5% 4.5%
population
1.60%
3.0% 3.0%
1.5% 0.80% 1.5%
0.0% 0.00% 0.0%

0.10% 0.02%
population
0.06%
0.05% 0.01% 0.03%
0.00% 0.00% 0.00%
Figure 6: Outbreak dynamics of COVID-19 worldwide. Dynamic effective reproduction number R(t) and symptomatic, asymptomatic, and
recovered populations at all nine locations. The simulation learns the time-varying contact rate β(t), and with it the time-varying effective repro-
duction number R(t), to predict the symptomatic infectious, asymptomatic infectious, and recovered populations Is , Ia , and R, for fixed latent and
infectious periods A = 2.5 days, Cs = 6.5 days, the hierarchical asymptomatic infectious period Ca = 5.76 (95%CI: 3.59-8.09) days from Figure
5. The dashed vertical lines mark the the first time each location managed to lower the effective reproduction below R(t) = 1 after lockdown. The
colored regions highlight the 95% credible interval for the effective reproductive number R(t) (top), the symptomatic and asymptomatic populations
Is and Ia , and the recovered population R (bottom plots), for uncertainties in the number of confirmed cases D, the fraction of the symptomatic
10infectious populations Is0 and Ia0 .
infectious population νs , the initial exposed population E0 , and the initial
on April 9 in Denmark, 2.10% (95% CI: 1.65%-2.69%) on April 2 in Geneva, 0.63% (95% CI: 0.57%-0.68%) on
April 15 in the Netherlands, 0.28% (95% CI: 0.14%-0.48%) on June 8 in Rio Grande do Sul, and 0.70% (95% CI:
0.62%-0.78%) on April 11 in Belgium. On Jun 15, 2020, the estimated recovered population reached 24.15% (95%
CI: 20.48%-28.14%) in Heinsberg, NRW, Germany 2.40% (95% CI: 2.09%-2.76%) in Ada County, ID, USA 46.19%
(95% CI: 45.81%-46.60%) in New York City, NY, USA 11.26% (95% CI: 7.21%-16.03%) in Santa Clara County, CA,
USA 3.09% (95% CI: 2.27%-4.03%) in Denmark 12.35% (95% CI: 10.03%-15.18%) in Geneva Canton, Switzerland
5.24% (95% CI: 4.84%-5.70%) in Netherlands 1.53% (95% CI: 0.76%-2.62%) in Rio Grande do Sul, Brazil 5.32%
(95% CI: 4.77%-5.93%) in Belgium.
3.4. Estimating the outbreak date.

Figure 7 shows the estimated outbreak date of COVID-19 in Santa Clara County. For fixed latent and symptomatic
infectious periods A = 2.5 days and Cs = 6.5 days, and for the hierarchically estimated asymptomatic infectious
periods Ca = 5.76 (95%CI: 3.59-8.09) days, the graph highlights the estimated date of the first COVID-19 case in
the county. Based on the reported case data from March 16, 2020 onward, and taking into account uncertainty on
the fraction of the symptomatic infectious population νs , on the initial exposed population E0 , and on the initial
symptomatic and asymptomatic infectious populations Is0 and Ia0 , we systematically backtracked the date of the first
undetected infectious individual. Our results suggest that the first case of COVID-19 in Santa Clara County dates back
to January 20, 2020 (95% CI: December 29, 2019 - February 13, 2020).
samples
1.40% exposed
asymptomatic infectious
1.20%
1.00%
population
0.80%
0.60%
0.40%
0.20%
0.00%
Jan Feb Mar 0 500
0
samples
200 First COVID-19 case

Jan 20,2020 (95%CI:Dec 29-Feb 13)
Figure 7: Estimating the outbreak date of COVID-19 in Santa Clara County varying asymptomatic infectious periods Ca . Estimated date
of the first COVID-19 case in Santa Clara County for fixed latent and symptomatic infectious periods A = 2.5 days and Cs = 6.5 days, and for the
hierarchical asymptomatic infectious period Ca = 5.76 (95%CI: 3.59-8.09) days from Figure 5. The colored regions in the main plot highlight the
95% credible interval for the time evolution of the exposed and asymptomatic infectious populations E and Ia estimated based on the reported cases
D̂(t) from March 16, 2020 onward and taking into account uncertainties on the fraction of the symptomatic infectious population νs = Is /I, and the
exposed and asymptomatic infectious populations E0 and Ia0 on March 16, 2020 (right plot). The bottom plot histogram shows the distribution of
the most probable origin dates to January 20, 2020 (95% CI: December 29, 2019 - February 13, 2020).
4. Discussion
A key question in understanding the outbreak dynamics of COVID-19 is the dimension of the asymptomatic pop-
ulation and its role in disease transmission. Throughout the past three months, dozens of studies have been initiated
to quantify the fraction of the general population that displays antibody prevalence but did not report symptoms of
COVID-19. Here we assume that this subgroup of the population has been infected with the novel coronavirus,
11
but has remained asymptomatic, or only displayed mild symptoms that were not directly reported in the context of
COVID-19. We collectively map this subgroup into an asymptomatic population and additively decompose the total
infectious population, I = Is + Ia , into a symptomatic group Is and an asymptomatic group Ia . We parameterize this
decomposition in terms of a single scalar valued parameter, the symptomatic fraction νs . Within this paradigm, we
can conceptually distinguish two scenarios: the special case for which both subgroups display identical contact rates
β, latent periods A, and infectious periods C, and the general case for which these transition dynamics are different.
For comparable dynamics, the size of the asymptomatic population does not affect overall outbreak dy-
namics. For the special case in which both subgroups display identical contact rates β, latent rates α, and infectious
rates γ [67], our study shows that the overall outbreak dynamics can be represented by the classical SEIR model [25]
using equations (6). Importantly, however, since the reported case data only reflect the symptomatic infectious and
recovered groups Is and Rs , the true infectious and recovered populations I = Is /νs and R = Rs /νs could be about an
order of magnitude larger than the SEIR model predictions. From an individual’s perspective, a smaller symptomatic
group νs , or equivalently, a larger asymptomatic group νa = [1 − νs ], could have a personal effect on the likelihood
of being unknowingly exposed to the virus, especially for high-risk populations: A larger asymptomatic fraction νa
would translate into an increased risk of community transmisson and would complicate outbreak control [15]. From a
health care perspective, however, the special case with comparable transition dynamics would not pose a threat to the
health care system since the overall outbreak dynamics would remain unchanged, independent of the fraction νa of the
asymptomatic population: A larger asymptomatic fraction would simply imply that a larger fraction of the population
has already been exposed to the virus–without experiencing significant symptoms–and that the true hospitalization
and mortality rates would be much lower than the reported rates [27].
For different dynamics, the overall outbreak dynamics depend on both size and infectiousness of the
asymptomatic group. For the general case in which the transition rates for the symptomatic and asymptomatic
groups are different, the overall outbreak dynamics of COVID-19 become more unpredictable, since little is known
about the dynamics of the asymptomatic population [42]. To study the effects of different dynamics between the
symptomatic and asymptomatic groups, we decided to collectively represent a lower infectivity of the asymptomatic
population through a smaller infectious period Ca < Cs and a lack of early isolation of the asymptomatic population
through a larger infectious period Ca > Cs , while, for simplicity, keeping the latent period A and contact rate β similar
across both groups [33]. Our study shows that the overall reproduction number, R(t) = [Ca Cs ]/[νs Ca +νa Cs ] β(t), and
with it the outbreak dynamics, depend critically on the fractions of the symptomatic and asymptomatic populations
νs and νa and on the ratio of the two infectious periods Cs and Ca . To illustrate these effects, we report the results for
three different scenarios where the asymptomatic group is half as infectious, Ca = 0.5 Cs , equally infectious, Ca = Cs ,
and twice as infectious Ca = 2.0 Cs as the symptomatic group for Santa Clara County. The second case, the middle
column in Figures 3 and 4, corresponds to the special case with comparable dynamics and similar parameters. Our
learnt asymptomatic infectious periods of Ca = 5.76 days in Figure 5 suggest that Ca is smaller than the symptomatic
infectious period of Cs = 6.5 days and that the asymptomatic population is slightly less infectious as the symptomatic
population. This can be a combined effect of less viral shedding as opposed to the symptomatic individuals, whilst
concomitantly having more contacts because the asymptomatic individual does not realize he/she is spreading the
disease.
Dynamic contact rates are a metric for the efficiency of public health interventions. Classical SEIR epi-
demiology models with static parameters are well suited to model outbreak dynamics under unconstrained conditions
and predict how the susceptible, exposed, infectious, and recovered populations converge freely toward the endemic
equilibrium [25]. However, they cannot capture changes in disease dynamics and fail to converge towards a temporary
equilibrium before the entire population has become sufficiently immune to prevent further spreading [45]. To address
this limitation, we introduce a time-dependent contact rate β(t), which we learn dynamically from the reported case
data. Figures 3 and 6 demonstrate that our approach can successfully identify a dynamic contact rate that not only
decreases monotonically, but is also capable of reproducing local contact fluctuations. With this dynamic contact rate,
our model can capture the characteristic S-shaped COVID-19 case curve that plateaus before a large fraction of the
population has been affected by the disease, resembling a Gompertz function. Previous studies have inferred discrete
date points at which the contact rates vary [7] or used sliding windows over the amount of novel reported infections
[43] to motivate dynamic contact rates. As such, our framework provides a model-based method for statistical infer-
12
ence of virus transmissibility: It naturally learns the most probable contact rate from the changing time evolution of
new confirmed cases and concomitantly quantifies the uncertainty on that estimation.
The dynamics of the asymptomatic population affect the effective reproduction number. Our anal-
ysis in equations (4) and (5) and our simulations in Figure 4 illustrate how asymptomatic transmission affects
the effective reproduction number, and with it the outbreak dynamics of COVID-19. Our results show that, the
larger the infectious period Ca of the asymptomatic group, the larger the initial effective reproduction number,
R(t) = Cs β(t)/[ νs + νa Cs /Ca ], and the later the drop of R(t) below the critical value of one. A recent study an-
alyzed the dynamics of the asymptomatic population in three consecutive windows of two weeks during the early
outbreak in China [34]. The study found relatively constant latency and infectious periods A and C, similar to our
assumption, and a decrease in the contact rate β = 1.12, 0.52, 0.35 days−1 and in the effective reproduction number
R(t) = 2.38, 1.34, 0.98, which is consistent with our results. However, rather than assuming constant outbreak param-
eters within pre-defined time windows, our study learns the effective reproduction number dynamically, in real time,
from the available data. Figures 4 and 6 demonstrate that we can successfully learn the critical time window until R(t)
drops below 1, which, in Santa Clara County, took till March 28 for Ca =3.25 days, till April 1 for Ca = 6.5 days and
till April 8 for Ca = 13.0 days. Our findings are consistent with the observation that the basic reproduction number will
be over-estimated if the asymptomatic group has a shorter generation interval, and underestimated if it has a longer
generation interval than the symptomatic group [42]. Naturally, these differences are less pronounced under current
conditions where the effective reproduction number is low and the entire population has been sheltering in place. It
will be interesting to see if the effects of asymptomatic transmission become more visible as we gradually relax the
current constraints and allow all individuals to move around and interact with others more freely. Seasonality, effects
of different temperature and humidity, and other unknown factors may also influence the extent of transmission.
Estimates of the infectious asymptomatic population may vary, but general trends are similar. Through-
out the past months, an increasing number of researchers around the globe have started to characterize the size of
the asymptomatic population to better understand the outbreak dynamics of COVID-19 [27]. Two major challenges
drive the interest in these studies: estimating the severity of the outbreak, e.g., hospitalization and mortality rates
[15], and predicting the success of surveillance and control efforts, e.g., contact tracing or vaccination [18]. This is
especially challenging now–in almost complete lockdown–when the differences in transmission dynamics between
the symptomatic and asymptomatic populations are small and difficult to quantify. However, as Figure 4 suggests,
these transmission dynamics can have a significant effect on the size of the asymptomatic population: For infec-
tious periods of Ca = 0.5, 1.0, and 2.0 Cs , the maximum infectious population varies from 0.70% to 1.23% and
2.10%. Interestingly, not only the sum of the infectious and recovered populations, but also the uncertainty of their
prediction, remain relatively insensitive to variations in the infectious period. To explore whether this is a univer-
sal trend, we perform the same analysis for nine different locations at which COVID-19 antibody prevalence was
measured in a representative sample of the population, Heinsberg (NRW, Germany) [59], Ada County (ID, USA)
[4], New York City (NY, USA) [46], Santa Clara County (CA, USA) [3], Denmark [13], Geneva Canton (Switzer-
land) [60], Netherlands [55], Rio Grande do Sul (Brasil) [54], and Belgium [52]. The fraction of the symptomatic
population in these nine locations is νs = 20.00%, 7.90%, 5.76%, 1.77%, 6.95%, 10.34%, 17.31%, 8.10%, and
10.21% respectively, broadly representing the range of reported symptomatic versus estimated total cases worldwide
[3, 4, 8, 11, 13, 17, 23, 28, 30, 37, 46, 47, 52, 53, 56, 54, 55, 59, 60, 62, 64, 66, 68]. Of the nine locations we an-
alyzed here, Heinsberg tested IgG and IgA, Ada County tested IgG, New York City tested IgG, Santa Clara County
tested IgG and IgM, Denmark tested IgG and IgM, Geneva tested IgG, the Netherlands tested IgG, IgM and IgA,
Rio Grande do Sul tested IgG and IgM, and Belgium tested IgG. While we did include reported uncertainty on the
seroprevalence data, seroprevalence would likely have been higher if all locations had tested for all three antibodies.
Despite these differences, the effective reproduction numbers R(t) and the infectious and recovered populations Is , Ia ,
and R in Figure 6 display remarkably similar trends: In most locations, the effective reproduction number R(t) drops
rapidly to values below one within a window of about three weeks after the lockdown date. However, the maximum
infectious population, a value that is closely monitored by hospitals and health care systems, varies significantly rang-
ing from 0.28% and 0.38% in Rio Grande do Sul and Denmark respectively to 3.54% and 6.11% in Heinsberg and
New York City respectively. This is consistent with the reported ’superspreader’ events in these last two locations.
An effect that we do not explicitly address is that immune response not only results COVID-19 antibodies (humoral
response), but also from innate and cellular immunity [22]. While it is difficult to measure the effects of the unreported
13
asymptomatic group directly, and discriminate it precisely from innate and cellular immunity, mathematical models
can provide valuable insight into how this population modulates the outbreak dynamics and the potential of successful
outbreak control [34].
Simulations provide a window into the outbreak date. Santa Clara County was home to the first individual
who died with COVID-19 in the United States. Although this happened as early as February 6, the case remained
unnoticed until April 22 [1]. The unexpected new finding suggests that the new coronavirus was circulating in the
Bay Area as early as January. The estimated uncertainty on the exposed, symptomatic infectious, and asymptomatic
infectious populations of our model allows us to estimate the initial outbreak date dates back to January 20 (Figure 7).
This back-calculated early outbreak date is in line with our intuition that COVID-19 is often present in a population
long before the first official case is reported. Interestingly, our analysis comes to this conclusion purely based on a
local serology antibody study [3] and the number of reported cases after lockdown [51].
Limitations. Our approach naturally builds in and learns several levels of uncertainties. By design, this allows
us to estimate sensitivities and credible intervals for a number of important model parameters and discover impor-
tant features and trends. Nevertheless, it has a few limitations, some of them by design, some simply limited by
the current availability of data: First, our current SEIIR model assumes a similar contact rate β(t) for symptomatic
and asymptomatic individuals. While we can easily adjust this in the model by defining individual symptomatic and
asymptomatic rates βs (t) and βa (t), we currently do not have data on the temporal evolution of the hidden asymptomatic
infectious population Ia (t) and longitudinal large population antibody studies would be needed to appropriately cal-
ibrate βa (t). Second, the ratio between the symptomatic and asymptomatic populations νs : νa can vary over time,
especially, as we have shown, if both groups display notably different dynamics, in our model represented through Cs
and Ca . Since this can have serious effects on the overall reproduction number R(t), and with it on required outbreak
control strategies, it seems critical to perform more tests and learn the dynamics of the fractions νs (t) and νa (t) of both
groups. Third, and this is not only true for our specific model, but for COVID-19 forecasts in general, all predictions
can be sensitive to the amount of testing in time. As such, they crucially rely on testing policies and testing capacities.
We expect to see a significant increase in the symptomatic-to-asymptomatic, or rather detected-to-undetected, ratio
as we move towards systematically testing larger fractions of the population and more and more people who have no
symptoms at all. The intensity of testing increases in most locations during our simulation period. For example, in
Santa Clara County, testing was extremely limited until early April, increased substantially in the first three weeks of
April, and even more after. Including limited testing and more undocumented cases during the early outbreak would
shift the case distribution towards earlier days, and predict an even earlier outbreak date. When longitudinal antibody
studies become available, additional methodologies can be developed to correct for this limitation. Fourth, while we
have included uncertainty in the seroprevalence data, the nine locations we analyzed here tested different types of
antibodies and had different sampling procedures. Seroprevalence could have been higher if all locations had tested
for the same three antibodies and data may differ depending on biases introduced by the sampling procedure. Finally,
our current model does not explicitly account for innate and cellular immunity. If the fraction of the population with
innate and cellular immunity is substantially high, we would anticipate a smaller susceptible population and a larger
and earlier protective immunity overall. These, and other limitations related to the availability of information, can be
easily addressed and embedded in our model and will naturally receive more clarification as studies and data become
available in the coming months.
5. Conclusions
The rapid and devastating development of the COVID-19 pandemic has raised many open questions about its
outbreak dynamics and unsuccessful outbreak control. From an outbreak management standpoint–in the absence
of effective vaccination and treatment–the two most successful strategies in controlling an infectious disease are
isolating infectious individuals and tracing and quarantining their contacts. Both critically rely on a rapid identification
of infections, typically through clinical symptoms. Recent antibody prevalence studies could explain why these
strategies have largely failed in containing the COVID-19 pandemic: Increasing evidence suggests that the number of
unreported asymptomatic cases could outnumber the reported symptomatic cases by an order of magnitude or more.
Mathematical modeling, in conjunction with reported symptomatic case data, antibody seroprevalence studies, and
machine learning allows us to infer, in real time, the epidemiology characteristics of COVID-19. We can now visualize
14
the invisible asymptomatic population, estimate its role in disease transmission, and quantify the confidence in these
predictions. A better understanding of asymptomatic transmission will help us evaluate strategies to manage the
impact of COVID-19 on both our economy and our health care system. A large asymptomatic population is associated
with a high risk of community spread and could require a conscious shift from containment to mitigation induced by
behavior changes. Our study suggests that, until vaccination and treatment become available, increasing population
awareness, encouraging increased hygiene, mandating the use of face masks, restricting travel, and promoting physical
distancing could be the most successful strategies to manage the impact of COVID-19 on both our economy and our
health care system.
Acknowledgements
This work was supported by a Stanford Bio-X IIP Seed Grant (M.P. and E.K.), by a DAAD Fellowship (K.L.),
and by the Stanford COVID-19 Seroprevalence Study Fund (J.B. and E.B.).
References
[1] E. Allday, M. Kawahara. First known U.S. coronavirus death occurred on Feb. 6 in Santa Clara County. San Francisco Chronicle, April 22,
2020. https://www.sfchronicle.com/health/article/First-known-U-S-coronavirus-death-occurred-on-15217316.php
assessed: Jun 15, 2020.
[2] J. L. Aron, I. B. Schwartz. Seasonality and period-doubling bifurcation in an epidemic model. Journal of Theoretical Biology 110 (1984)
665-679.
[3] E. Bendavic, B. Mulaney, N. Sood, S. Shah, E. Ling, R. Bromley-Dulfano, C. Lai, Z. Weissberg, R. Saavedra-Walker, J. Tedrow, D. Tversky,
A. Bogan, T. Kupiec, D. Eichner, R. Gupta, J.P.A. Ioannidis, J. Bhattacharya. COVID-19 antibody seroprevalence in Santa Clara County,
California. medRxiv doi:10.1101/2020.04.14.20062463.
[4] A. Bryan, G. Pepper, M.H. Wener, S.L. Fink, C. Morishima, A. Chaudhary, K.R. Jerome, P.C. Mathias, A.L. Greninger Performance Charac-
teristics of the Abbott Architect SARS-CoV-2 IgG Assay and Seroprevalence in Boise, Idaho J Clin Microbiol. 2020 doi:10.1128/JCM.00941-
20
[5] Painel de casos de doença pelo coronavı́rus 2019 (COVID-19) no Brasil pelo Ministerio da Saude www.covid.saude.gov.br/ assessed:
Jun 15, 2020.
[6] A.M. Cuomo. Amid ongoing COVID-19 pandemic, governor cuomo announces results of completed antibody testing study.
https://www.governor.ny.gov/news/amid-ongoing-covid-19-pandemic-governor-cuomo-announces-results-completed-
antibody-testing. assessed: Jun 15, 2020.
[7] J. Dehning, J. Zierenberg, F.P Spitzner, M. Wibral, J.P. Neto, M. Wilczek, V. Priesemann. Inferring COVID-19 spreading rates and potential
change points for case number forecasts arXiv (2020) preprint:2004.01105
[8] A. Doi, K. Iwata, H. Kuroda, T. Hasuike, A. Kanda, T. Nagao, H. Nishioka, K. Tomii, T. Morimoto, Y. Kihara. Seroprevalence of novel
coronavirus disease (COVID-19) in Kobe, Japan. medRxiv doi:10/1101/2020.04.26.20079822.
[9] E. Dong, L. Gardner. An interactive web-based dashboard to track COVID-19 in real time. Lancet Infectious Disease (2020)
doi:10.1016/S1473-3099(20)30120-1.
[10] European Centre for Disease Prevention and Control. 2020. Situation update worldwide. www.ecdc.europa.eu/en/geographical-dis-
tribution- 2019-ncov-cases assessed: Jun 15, 2020.
[11] M. Emmenegger, E. De Cecco, D. Lamparter, R.P.B. Jacquat, D. Ebner, M.M. Schneider, I. Condado Morales, D. Schneider, B. Dogancay,
J. Guo, A. Wiedmer, J. Domange, M. Imeri, R. Moos, C. Zografou, C. Trevisan, A. Gonzalez-Guerra, A. Carrella, I.L. Dubach, C.K. Xu,
G. Meisl, V. Kosmoliaptsis, T. Malinauskas, N. Burgess-Brown, R. Owens, J. Mongkolsapaya, S. Hatch, G.R. Screaton, K. Schubert, J.D.
Huck, F. Liu, F. Pojer, K. Lau, D. Hacker, E. Probst-Mueller, C. Cervia, J. Nilsson, O. Boyman, L. Saleh, K. Spanaus, A. von Eckardstein,
D.J. Schaer, N. Ban, C. Tsai, J. Marino, G.F.X. Schertler, J. Gottschalk, B.M. Frey, R. Reimann, S. Hornemann, A.M. Ring, T.P.J. Knowles,
I. Xenarios, D.I. Stuart, A. Aguzzi Population-wide evolution of SARS-CoV-2 immunity tracked by a ternary immunoassay. medRxiv 2020,
doi: 10.1101/2020.05.31.20118554
[12] R. Engbert, F.R. Drepper. Chance and chaos in population biology–Models of recurrent epidemics and food chain dynamics. Chaos, Solutions
& Fractals 4 (1994) 1147-1169.
[13] C. Erikstrup, C.E. Hother, O.B.V Pedersen, K. Molbak, R.L. Skov, D.K. Holm, S. Saekmose, A.C. Nilsson, P.T. Brooks, J.K. Boldsen,
C. Mikkelsen, M. Gybel-Brask, E. Sorensen, K.M. Dinh, S. Mikkelsen, B.K. Moller, T. Haunstrup,L. Harritshoj, B.A. Jensen, H. Hjal-
grim, S.T. Lillevang, H. Ullum Estimation of SARS-CoV-2 infection fatality rate by real-time antibody screening of blood donors medRxiv
doi:10.1101/2020.04.24.20075291.
[14] Y. Fang, Y. Nie, M. Penny. Transmission dynamics of the COVID-19 outbreak and effectiveness of government interventions. J. Med. Virol.
(2020) 1-15.
[15] A.S. Fauci, H.C. Lane, R.R. Redfield. Covid-19–Navigating the uncharted. New England Journal of Medicine 382 (2020) 1268-1269.
[16] S. Flaxman, S. Mishra, A. Gandy, H. Unwin, H. Coupland, T. Mellan, H. Zhu, T. Berah, J. Eaton, P. Perez Guzman, N. Schmit, L. Cilloni, K.
Ainslie, M. Baguelin, I. Blake, A. Boonyasiri, O. Boyd, L. Cattarino, C. Ciavarella, L. Cooper, Z. Cucunuba Perez, G. Cuomo-Dannenburg,
A. Dighe, A. Djaafara, I. Dorigatti, S. Van Elsland, R. Fitzjohn, H. Fu, K. Gaythorpe, L. Geidelberg, N. Grassly, W. Green, T. Hallett, A.
Hamlet, W. Hinsley, B. Jeffrey, D. Jorgensen, E. Knock, D. Laydon, G. Nedjati Gilani, P. Nouvellet, K. Parag, I. Siveroni, H. Thompson, R.
Verity, E. Volz, C. Walters, H. Wang, Y. Wang, O. Watson, P. Winskill, X. Xi, C. Whittaker, P. Walker, A. Ghani, C. Donnelly, S. Riley, L.
15
Okell, M. Vollmer, N. Ferguson, S. Bhatt Report 13: Estimating the number of infections and the impact of non-pharmaceutical interventions
on COVID-19 in 11 European countries Spiral Imperial College London 2020, doi: 10.25561/77731
[17] A.Fontanet, L. Tondeur, Y. Madec, R. Grant, C. Besombes, N. Jolly, S. Fernandes Pellerin M.N. Ungeheuer, I. Cailleau, L. Kuhmel, S.
Temmam, C. Huon, K.Y. Chen, B. Crescenzo, S. Munier, C. Demeret, L. Grzelak, I. Staropoli, T. Bruel, P. Gallian, S. Cauchemez, S.
van der Werf, O. Schwartz, M. Eloit, B. Hoen. Cluster of COVID-19 in northern France: A retrospective closed cohort study. medRxiv
10.1101/2020.04.20071134.
[18] C. Fraser, S. Riley, R.M. Anderson, N.M. Ferguson. Factors that make an infectious disease outbreak controllable. Proceedings of the National
Academy of Sciences 101 (2004) 6146-6151.
[19] M. Gatto, E. Bertuzzo, L. Mari, S. Miccoli, L. Carraro, R. Casagrandi, A. Rinaldo. Spread and dynamics of the COVID-19 epidemic in Italy:
Effects of emergency containment measures. Proceedings of the National Academy of Sciences; in press, doi:10.1073/pnas.2004978117.
[20] F. Gao, L. Han. Implementing the Nelder-Mead simplex algorithm with adaptive parameters. Computational Optimization and Applications
51(2012) 259-277.
[21] A. Gelman, ,J. Hill. Data Analysis using Regression and Multilevel/Hierarchical Models. Cambridge University Press, 2006.
[22] A. Grifoni, D. Weiskopf, S.I. Ramirez, J. Mateus, J.M. Dan, C. Rydyznski Moderbacher, S.A. Rawlings, A. Sutherland, L. Premku-
mar R.S. Jadi, D. Marrama, A.M. de Silva, A. Frazier, A. Carlin, J.A. Greenbaum, B. Peters. F. Krammer. D.M. Smith, S. Crotty, A.
Sette. Targets of T cell responses to SARS-CoV-2 coronavirus in humans with COVID-19 disease and unexposed individuals. Cell, doi:
10.1016/j.cell.2020.05.015.
[23] H. Hallal, F. Hartwig, B. Horta, G.D. Victora, M. Silveira, C. Struchiner, L.P. Vidaletti, N. Neumann, L.S. Pellanda, O.A. Dellagostin, M.N.
Burattini, A.M. Menezes, F.C. Barros, A.J. Barros, C.G. Victora Remarkable variability in SARS-CoV-2 antibodies across Brazilian 1 regions:
nationwide serological household survey in 27 states. medRxiv 2020, doi: 10.1101/2020.05.30.20117531
[24] Heinsberg 2020. Aktuelles aus dem Kreishaus. Coronavirus im Kreis Heinsberg. www.kreis-heinsberg.de/aktuelles. assessed: Jun
15, 2020.
[25] H. W. Hethcote. The mathematics of infectious diseases. SIAM Review 42 (2000) 599-653.
[26] M.D. Hoffman, A. Gelman. The No-U-Turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo. Journal of Machine
Learning Research (2014), 15(1), 1593-1623.
[27] J.P.A. Ioannidis. The infection fatality rate of COVID-19 inferred from seroprevalence data. medRxiv. doi:10.1101/2020.05.13.20101253.
[28] I. Jerkovic, T. Ljubic, Z. Basic, I. Kruzic, N. Kunac, J. Bezic, A. Vuko, A. Markotic, S. Andjelinovic SARS-CoV-2 antibody seroprevalence
in industry workers in Split-Dalmatia and Sibenik-Knin County, Croatia. medRxiv 2020, doi: 10.1101/2020.05.11.20095158
[29] W. O. Kermack, G. McKendrick. Contributions to the mathematical theory of epidemics, Part I. Proceedings of the Royal Society London
Series A 115 (1927) 700-721.
[30] V. Kraehling, M. Kern, S. Halwe, H. Mueller, C.P.C. Rohde, M. Savini, M. Schmidt, J. Wilhelm, S. Becker, S. Ciesek, R. Gottschalk
Epidemiological study to detect active SARS-CoV-2 infections and seropositive persons in a selected cohort of employees in the Frankfurt
am Main metropolitan area. medRxiv 2020, doi: 10.1101/2020.05.20.20107730
[31] K.L. Lange, R.J.A. Little, M.G. Taylor. Robust statistical modeling using the t distribution. Journal of the American Statistical Association
84 (1989) 881-896.
[32] S. A. Lauer, K. H. Grantz, Q. Bi, F. K. Jones, Q. Zheng, H. R. Meredith, A. S. Azman, N. G. Reich, J. Lessler. The incubation period
of coronavirus disease 2019 (COVID-19) from publicly reported confirmed cases: estimation and application. Annals of Internal Medicine
(2020) doi:10.7326/M20-0504.
[33] Q. Li, X. Guan, P. Wu, X. Wang, L. Zhou, Y. Tong, R. Ren, K.S.M. Leung, E.H.Y. Lau, J.Y. Wong, X. Xing, N. Xiang, Y. Wu, C. Li, Q.
Chen, D. Li, T. Liu, J. Zhao, M. Liu W. Tu, C. Chen, L. Jin, R. Yang, Q. Wang, S. Zhou, R. Wang, H. Liu, Y. Luo, Y. Liu, G. Shao, H. Li, Z.
Tao, Y. Yang, Z. Deng, B. Liu, Z. Ma, Y. Zhang, G. Shi, T.T.Y. Lam, J.T. Wu, G.F. Gao, B.J. Cowling, B. Yang, G.M. Leung, Z. Feng. Early
transmission dynamics in Wuhan, China, of novel coronavirus-infected pneumonia. New England Journal of Medicine 382 (2020) 1199-1207.
[34] R. Li, S. Pei, B. Chen, Y. Song, T. Zhang, W. Yang, J. Shaman. Substantial undocumented infection facilitates the rapid dissemination of
novel coronavirus (SARS-CoV2). Science 368 (2020) 489-493.
[35] K. Linka, M. Peirlinck, F. Sahli Costabal, E. Kuhl. Outbreak dynamics of COVID-19 in Europe and the effect of travel restrictions. Computer
Methods in Biomechanics and Biomedical Engineering (2020) in press; doi:10.1080/10255842.2020.1759560.
[36] K. Linka, M. Peirlinck, E. Kuhl. The reproduction number of COVID-19 and its correlation with public health interventions. medRxiv
doi:10.1101/2020.05.01.20088047.
[37] D. Ng, G. Goldgof, B. Shy, A. Levine, J. Balcerek, S.P. Bapat, J. Prostko, M. Rodgers, K. Coller, S. Pearce, S. Franz, L. Du, M. Stone,
S. Pillai, A. Sotomayor-Gonzalez, V. Servellita, C. Sanchez-San Martin, A. Granados, D.R. Glasner, L.M. Han, K. Truong, N. Akagi, D.N.
Nguyen, N. Neumann, D. Qazi, E. Hsu, W. Gu, Y.A. Santos, B. Custer, V. Green, P. Williamson, N.K. Hills, C.M. Lu, J.D. Whitman, S.
Stramer, C. Wang, K. Reyes, J. Hakim, K. Sujishi, F. Alazzeh, L. Pharm, C. Oon, S. Miller, T. Kurtz, J. Hackett, G. Simmons, M.P. Busch,
C.Y. Chiu SARS-CoV-2 seroprevalence and neutralizing activity in donor and patient blood from the San Francisco Bay Area. medRxiv 2020,
doi: 10.1101/2020.05.19.20107482
[38] N.B. Noll, I. Aksamentov, V. Druelle, A. Badenhorst, B. Ronzani, G. Jefferies, J. Albert, R. Neher. COVID-19 Scenarios: an interactive tool
to explore the spread and associated morbidity and mortality of SARS-CoV-2. medRxiv doi: 10.1101/2020.05.05.20091363
[39] New York City Health Department. NYC Coronavirus Disease 2019 (COVID-19) Data www.github.com/nychealth/coronavirus-data, as-
sessed: Jun 15, 2020.
[40] New York Times. An ongoing repository of data on coronavirus cases and deaths in the U.S. www.github.com/nytimes/covid-19-data, as-
sessed: Jun 15, 2020.
[41] X. Pan, D. Chen, Y. Xia, X, Wu, T. Li, X. Ou, L. Zhou, J. Liu. Asymptomatic cases in a family cluster with SARS-CoV-2 infection. Lancet
Infectious Diseases 20 (2020) 410-411.
[42] S.W. Park, D.M. Cornforth, J. Dushoff, J.W. Weitz. The time scale of asymptomatic transmission affects estimates of epidemic potential in
the COVID-19 outbreak. medRxiv 2020.03.09.20033514, doi 10.1101/2020.03.09.20033514.
[43] S.W. Park, K. Sun, C. Viboud, B.T. Grenfell, J. Dushoff. Potential roles of social distancing in mitigating the spread of coronavirus disease
16
2019 (COVID-19) in South Korea. medRxiv 2020.03.27.20045815, doi 10.1101/2020.03.27.20045815

[44] M. Peirlinck, F. Sahli Costabal, K.L. Sack, J.S. Choy, G.S. Kassab, J.M. Guccione, M. De Beule, P. Segers, E. Kuhl. Using machine learning
to characterize heart failure across the scales. Biomechanics and Modeling in Mechanobiology 18 (2019) 1987-2001.
[45] M. Peirlinck, K. Linka, F. Sahli Costabal, E. Kuhl. Outbreak dynamics of COVID-19 in China and the United States. Biomechanics and
Modeling in Mechanobiology (2020) in press, doi:10.1101/2020.04.06.20055863.
[46] J. Reifer, N. Hayum, B. Heszkel, I. Klagsbald, V.A. Streva SARS-CoV-2 IgG Antibody Responses in New York City medRxiv
doi:10.1101/2020.2005.2023.20111427.
[47] E.S. Rosenberg, J.M. Tesoriero, E.M. Rosenthal, R. Chung, M.A. Barranco, L.M. Styer, M.M. Parker, S.J. Leung, J.M. Morne, D. Greene,
D.R. Holtgrave, D. Hoefer, J. Kumar, T. Udo, B. Hutton, H.A. Zucker Cumulative incidence and diagnosis of SARS-CoV-2 infection in New
York. medRxiv 2020, doi: 10.1101/2020.05.25.20113050
[48] F. Sahli Costabal, K. Matsuno, J. Yao, P. Perdikaris, E. Kuhl. Machine learning in drug development: Characterizing the effect of 30 drugs
on the QT interval using Gaussian process regression, sensitivity analysis, and uncertainty quantification. Computer Methods in Applied
Mechanics and Engineering 348 (2019) 313-333.
[49] J. Salvatier, T.V. Wiecki, C. Fonnesbeck. Probabilistic programming in Python using PyMC3. Peer Journal Computational Science 2 (2016)
e55.
[50] S. Sanche, Y.T. Lin, C. Xu, R. Romero-Severson, N. Hengartner, R. Ke. High contagiousness and rapid spread of severe acute respiratory
syndrome coronavirus 2. Emerging Infectious Disease (2020) doi:10.3201/eid2607.200282
[51] Santa Clara County COVID-19 Cases and Hospitalizations Dashboard. www.sccgov.org assessed: Jun 15, 2020.
[52] Sciensano: National Public Health Research Institute of Belgium Weekly Epidemiological Bulleting from May 29,2020
https://covid-19.sciensano.be assessed: Jun 15, 2020.
[53] M. Shakiba, S. Nazari, F. Mehrabian, S.M. Rezvani, Z. Ghasempour, A. Heidarzadeh. Seroprevalence of COVID-19 virus infection in Guilan
province, Iran. medRxiv doi:10.1101/2020.04.26.20079244.
[54] M. Silveira, A. Barros, B. Horta, L. Pellanda, G. Victora, O. Dellagostin, C. Struchiner, M. Burattini, A. Valim, E. Berlezi, J. Mesa, M.L.
Ikeda, M. Mesenburg, M. Mantesso, M. Dall’Agnol, R. Bittencourt, F. Hartwig, A.M. Menezes, F.C. Barros, P. Hallal, C.G. Victora Repeated
population-based surveys of antibodies against SARS-CoV-2 in Southern Brazil medRxiv doi: 10.1101/2020.05.01.20087205
[55] E. Slot, B.M. Hogema, C.B.E.M. Reusken, J.H. Reimerink, M. Molier, H.M. Karregat, J. Ijist, V.M.J. Novotny, R.A.W. van Lier, H.L.
Zaaijer. Herd immunity is not a realistic exit strategy during a COVID-19 outbreak. Research Square 2020. https://dx.doi.org/10.21203/rs.3.rs-
25862/v1.
[56] C.J. Snoeck, M. Vaillant, T. Abdelrahman, V.P. Satagopam, J.D. Turner, K. Beaumont, C.P.C. Gomes, J.V. Fritz, V.E. Schröder, A. Kaysen,
L. Pavelka, L. Stute, G. Ramos Meyers, L. Pauly, M. Hansen, C. Pauly, G.A. Aguayo, M. Perquin, A. Hanff, S. Ghosh, M. Gantenbein, L.
Huiart, M. Ollert, R. Krüger Prevalence of SARS-CoV-2 infection in the Luxembourgish population: the CON-VINCE study. medRxiv doi:
10.1101/2020.05.11.20092916
[57] N. Sood, P. Simon, P. Ebner, D. Eichner, J. Reynolds, E. Bendavid, J. Bhattacharya. Seroprevalence of SARS-CoV-2-specific an-
tibodies among adults in Los Angeles County, California, on April 10-11, 2020. Journal of the American Medical Association.
doi:10.1001/jama.2020.8279.
[58] B. Spellberg, M. Haddix, R. Lee, S. Butler-Wu, P. Holtom, H. Yee, P. Gounder. Community prevalence of SARS-CoV-2 among patients
with influenzalike illnesses presenting to a Los Angeles medical center in March 2020. Journal of the American Medical Association.
doi:10.1001/jama.2020.4958.
[59] H. Streeck, B. Schulte, B.M. Kümmerer, E. Richter, T. Höller, C. Fuhrmann, E. Bartok, R. Dolscheid, M. Berger, L. Wessendorf, M.
Eschbach-Bludau, A. Kellings, A. Schwaiger, M. Coenen, P. Hoffmann, B. Stoffel-Wagner, M.M. Nöthen, A.M. Eis-Hübinger, M. Exner2,
R.M. Schmithausen, M. Schmid, G. Hartmann. Infection fatality rate of SARS-CoV-2 infection in a German community with a super-
spreading event. medRxiv doi:10.1101/2020.05.04.20090076.
[60] S. Stringhini, A. Wisniak, G. Piumatti, A.S. Azman, S.A. Lauer, H. Baysson, D. De Ridder, D. Petrovic, S. Schrempft, K. Marcus, I. Arm-
Vernez, S. Yerly, O. Keiser, S. Hurst, K. Posfay-Barbe, D. Trono, D. Pittet, L. Getaz, F. Chappuis, I. Eckerle, N. Vuilleumier, B. Meyer, A.
Flahault, L. Kaiser, I. Guessous Repeated seroprevalence of anti-SARS-CoV-2 IgG antibodies in a population-based sample from Geneva,
Switzerland medRxiv doi:10.1101/2020.05.02.20088898
[61] Swiss Federal Office of Public Health FOPH COVID19 case numbers Cantons of Switzerland and Principality of Liechtenstein.
www.github.com/openZH/covid19/, assessed: Jun 15, 2020.
[62] M. Takita, T. Matsumura, K. Yamamoto, E. Yamashita, K. Hosoda, T. Hamaki, E. Kusumi, Regional Difference in Seroprevalence of SARS-
CoV-2 in Tokyo: Results from the community point-of-care antibody testing. medRxiv 2020, doi: 10.1101/2020.06.03.20121020
[63] B. Tang, F. Xia, N.L. Bragazzi, Z. McCarthy, X. Wang, S. He, X. Sun, S. Tang, Y. Xiao, J. Wu. Lessons drawn from China and South
Korea for managing COVID-19 epidemic: insights from a comparative modeling study. Bulletin of the World Health Organization. doi:
http://dx.doi.org/10.2471/BLT.20.257238.
[64] C. Thompson, N. Grayson, R.S. Paton, J. Lourenco, B.S. Penman, L. Lee. Neutralising antibodies to SARS coronavirus 2 in Scottish blood
donors – a pilot study of the value of serology to determine population exposure. medRxiv doi:10.1101/2020.04.13.20060467.
[65] T.A. Treibel, C. Manisty, M. Burton, A. McKnight, J. Lambourne, J.B. Augusto, X. Couto-Parada, T. Cutino-MOguel, M. Noursadeghi, J.C.
Moon. COVID-19: PCR screening of asympotmatic healthcare workers at London hospital. Lancet. doi:10.1016/S0140-6736(20)31100-4.
[66] L. Valenti, A. Bergna, S. Pelusi, F. Facciotti, A. Lai, M. Tarkowski, A. Berzuini, F. Caprioli, L. Santoro, G. Baselli, C. Della Ventura, E. Erba,
S. Bosari, M. Galli, G. Zehender, D. Prati SARS-CoV-2 seroprevalence trends in healthy blood donors during the COVID-19 Milan outbreak.
medRxiv 2020, doi: 10.1101/2020.05.11.20098442
[67] A. Viguerie, G. Lorenzo, F. Auricchio, D. Baroli, T.J.R. Hughes, A. Patton, A. Reali, T.E. Yankeelov, A. Veneziani. Simulating the spread of
COVID-19 via a spatially-resolved SEIRD model with heterogeneous diffusion. Oden Institute Report 20-09, Austin.
[68] X. Wu, B. Fu, L. Chen, Y. Feng. Serological tests facilitate identification of asymptomatic SARS-CoV-2 infection in Wuhan, China. Journal
of Medical Virology. doi:10.1002/jmv.25904.
17

2020 06 23 20111419v2 Full PDF

Uploaded by

Copyright:

Available Formats

2020 06 23 20111419v2 Full PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2020 06 23 20111419v2 Full PDF

Uploaded by

Copyright:

Available Formats

medRxiv preprint doi: https://doi.org/10.1101/2020.05.23.20111419.this version posted June 23, 2020.

The copyright holder for this preprint

Visualizing the invisible:

Preprint submitted to medRxiv June 23, 2020

2.1. Epidemiology modeling.

model, which is governed by a set of five ordinary differential equations,

2.2. Seroprevalence studies.

2.3. Uncertainty quantification.

Table 2: Prior distributions for SEIIR model parameters.

Parameter Interpretation Distribution

p(D̂(t) | D(t, ϑ)) p(ϑ)

2.4. Hierarchical asymptomatic infectious period Ca estimation.

Table 3: Prior distributions for SEIIR-based hierarchical Ca estimation.

Parameter Interpretation Distribution

2.5. Estimating the outbreak date.

3.1. Outbreak dynamics of COVID-19 in Santa Clara County.

3.2. Effect of asymptomatic transmission of COVID-19 in Santa Clara County.

3.3. Outbreak dynamics of COVID-19 worldwide.

12.0% 3.00% 12.0%

8.0% 2.00% 8.0%

4.0% 1.00% 4.0%

0.0% 0.00% 0.0%

0.0% 0.00% 0.0%

3.4. Estimating the outbreak date.

200 First COVID-19 case

2019 (COVID-19) in South Korea. medRxiv 2020.03.27.20045815, doi 10.1101/2020.03.27.20045815

You might also like