SSRN Id3569367 PDF
SSRN Id3569367 PDF
SSRN Id3569367 PDF
Abstract
This paper studies the SEIRD epidemic model for COVID-19. First, I show that the
model is poorly identified from the observed number of deaths and confirmed cases. There
are many sets of parameters that are observationally equivalent in the short run but lead to
markedly different long run forecasts. Next, I show that the basic reproduction number R0
can be identified from the data, conditional on the clinical parameters. I then estimate it for
the US and several other countries and regions, allowing for possible underreporting of the
number of cases. The resulting estimates of R0 are heterogeneous across countries: they are
2-3 times higher for Western countries than for Asian countries. I demonstrate that if one
fails to take underreporting into account and estimates R0 from the reported cases data, the
resulting estimate of R0 will be biased downward and the resulting forecasts will exaggerate
the number of deaths. Finally, I demonstrate that auxiliary information from random tests
can be used to calibrate the initial parameters of the model and reduce the range of possible
forecasts about the future number of deaths.
∗
I thank Andy Atkeson, Jeremy Fox, Oleg Itskhoki, David Slichter, Jim Stock, Ping Yan, and Tom Zohar
for their comments and suggestions. All remaining errors are mine.
†
Department of Economics, Binghamton University, 4400 Vestal Parkway East, PO Box 6000, Bingham-
ton, NY 13902-6000, USA. E-mail: [email protected].
The SIR (Susceptible, Infectious, Recovered) model and its variations is widely used in
epidemiology to model the spread of epidemics. Since the outbreak of COVID-19, it has seen
increased popularity among economists who are trying to asses the economic consequences
of the coronavirus and various mitigation policies, such as Atkeson (2020b,c), Berger et al.
(2020), Eichenbaum et al. (2020), Fernandez-Villaverde and Jones (2020), Piguillem and Shi
(2020), Toda (2020), and others. In this paper, I study identification and estimation of the
modification of the SIR model called SEIRD (Susceptible, Exposed, Infectious, Recovered,
and Dead) and present several findings.
First, I show that the SEIRD model has too many degrees of freedom and is poorly
identified from the deaths and confirmed cases data. Conditional on the values of clinical
parameters, i.e. parameters that reflect the clinical progression of the disease, the only model
parameter that is identified is the basic reproduction number R0 . While R0 governs the speed
of spread of the virus, the key driver of the long run number of deaths in the model is the
case fatality ratio (CFR), which is not identified separately from initial values. As a result,
when I estimate the model for different initial values, I obtain models that are observationally
equivalent in the short run but produce markedly different estimates of unobserved variables
and long run forecasts. For instance, the estimated number of people who had the virus in
the US on March 31 for different observationally identical models ranges from several million
to around 140 million, while the predicted death toll varies from around 30 thousand to over
a million.
Second, I estimate the basic reproduction number R0 , which is identified, and show that it
is heterogeneous across countries and regions. For the same values of clinical parameters, the
estimates of R0 for the US, UK, California, and New York state are about 2-3 times higher
than for Japan or Taiwan. Moreover, the estimates of R0 are highly sensitive to the values
of clinical parameters. There is no agreement in the medical literature on the length of the
incubation and infectious period for COVID-19, different values of these parameters result
In this paper I study a version of the SEIR model that includes dead among its compart-
ments. Similar models have been used in epidemiology by Chowell et al. (2007), Lin et al.
(2020), Wang et al. (2020), and others. More advanced versions of the model with more
compartments are considered in Chowell et al. (2003) and Chowell et al. (2006). I consider
a model with five groups of people: susceptible (S), exposed (E), infectious (I), recovered
(R), and dead (D). Susceptible are those who have not gotten the virus yet and can become
infected. Exposed are those who have gotten the virus but cannot transmit it to others yet.
This corresponds to the so called incubation period. Infectious are those who have the virus
and are contagious. Recovered are those who were sick in the past but have recovered from
the virus. Dead are those who have died because of the virus.
The number of people in different groups evolves over time as follows:
dS(t) S(t)
= −β I(t) (2.1)
dt N
dE(t) S(t)
=β I(t) − σE(t) (2.2)
dt N
dI(t)
= σE(t) − γI(t) (2.3)
dt
dR(t)
= (1 − α)γI(t) (2.4)
dt
dD(t)
= αγI(t) (2.5)
dt
dC(t)
= λγI(t) (2.6)
dt
N is the population size of a given country or region. I assume that it is fixed and does
not vary over time. I could model the dynamics of the population size to account for the fact
that some people die from the disease, but then I would also need to model births and deaths
due to other causes. In order to avoid these complications, I simply fix N , as is commonly
done in the literature. C(t) is the cumulative number of cases confirmed. It does not affect
the model dynamics but is used to match the model to the confirmed cases data. In my
In my estimation, I rely on the deaths and confirmed cases data for COVID-19. The
country level data is collected by the Center for Systems Science and Engineering at Johns
Hopkins University and is available online.1 The state level data for the US is collected by
the New York Times and is also available online.2 The population of different countries is
taken from World Population Prospects 2019 by United Nations.3 The population of different
states in the US is taken from the US Census Bureau.4
I use T = 70 observations in my sample, with the first observation being January 22,
2020. Around that time, cases of coronavirus were widely registered outside China, e.g. in
the USA (January 21),5 Germany (January 27),6 and the UK (January 31).7 However, as I
show below, the initial conditions and the epidemic start date are not identified separately
from the CFR and the fraction of cases observed. I discuss the identification challenges in
more detail below. The last observation in my data corresponds to March 31, 2020.
4 Identification
In this section, I study identification of the model parameters based on the deaths and
confirmed cases data. There are several earlier papers on identification of the parameters of
the SIR and related models, e.g. Marinov et al. (2014), Magal and Webb (2018), and Ducrot
et al. (2019), but they are not directly applicable in the current setting. In particular,
they do not study whether the parameters are identified based on the short run data only.
Atkeson (2020a), written concurrently and independently of this paper, attempts to answer
the question similar to mine in the context of the usual SIR model.
1
https://github.com/CSSEGISandData/COVID-19/tree/master/csse_covid_19_data/csse_covid_
19_time_series
2
https://github.com/nytimes/covid-19-data
3
https://population.un.org/wpp/Download/Standard/Population/
4
https://www.census.gov/data/tables/time-series/demo/popest/2010s-state-total.html.
5
https://www.cdc.gov/media/releases/2020/p0121-novel-coronavirus-travel-case.html
6
https://www.dw.com/en/germany-confirms-human-transmission-of-coronavirus/a-52169007
7
https://www.bbc.co.uk/news/health-51325192
In this section, I describe the estimation procedure that I use. It is similar to Method 2
in Chowell et al. (2007) but consists of several steps because I use both deaths and reported
cases data. As discussed above, I fix T0 = 1 and I0 = 0. I start from E0 = 1 and increase it
if needed as discussed below. For a given vector of parameters (α, R0 , λ) and each period t, I
compute the implied number of deaths D(t, α, R) and the implied number of reported cases
C(t, R, λ).8 Then I compute the residual sum of squares for the deaths series, given by
T
X
RSSD (α, R) = (D(t) − D(t, α, R))2 ,
t=1
and the residual sum of squares for the reported cases series, given by
T
X
RSSC (R, λ) = (C(t) − C(t, R, λ))2 ,
t=1
where D(t) and C(t) are the actual data. I then find the values (α̃D , R̃D ) that minimize
RSSD and the values (R̃C , λ̃C ) that minimize RSSC . I then estimate (α, R0 , λ) jointly by
minimizing
RSST (α, R, λ) = RSSD (α, R)/RSSD (α̃D , R̃D ) + RSSC (R, λ)/RSSC (R̃C , λ̃C )
I use the normalization by the preliminary values of the RSS for deaths and reported cases
so that both series contribute roughly equally to the final objective function. If I did not
normalize their contributions, then RSSC would dominate, because the number of reported
cases in the data is orders of magnitude larger than the number of deaths. I call the resulting
estimates (α̂, R̂, λ̂).
If any of α̃D , λ̃C , α̂, or λ̂ are at the upper bound of 1, then I increase E0 until the
8
Note that the fatality rate does not affect the number of reported cases in the model, while the observable
fraction of cases does not affect the number of deaths.
10
T
(log C(t) − log C(t, R, λ))2 1{D(t) > 25}
X
RSSC,logs (R, λ) =
t=1
When I compute the RSS for deaths in logs, I trim observations where the number of
deaths is below 2. For the RSS for cases in logs, I trim observations where the number of
cases is below 26. These choices are arbitrary, but they are motivated by the fact that the
log series, which should be linear in the model in the short run, typically exhibit breaks in
the data. During the early stages of the epidemic, they are almost flat, but then their slope
increases. By using trimming, I fit the model to the latter portion of the data, where thee
linearity of the log series in time appears to hold, but exclude the early observations before
the break.
I abstract away from the statistical properties of my estimators and consider them purely
as curve fitting algorithms. There are several reasons for this. First, the SEIRD model is
deterministic and does not contain any error terms. In order to conduct inference properly,
one would need to introduce the error term in the model and discuss its properties. Because
11
6 Results
This section presents the estimation results. Before I move on to my main results, I discuss
computational issues associated with estimation of the model. Based on the arguments from
the previous section, when I change the initial parameters, the estimate of R0 should remain
unchanged, while the estimates of α and λ should change proportionally to the changes initial
parameters. In practice, however, this is not always the case. Both α and λ are constrained
to lie between 0 and 1, and when these constraints are binding or close to binding, changes
in the initial values can have a substantial effect on the estimate of R0 .
Table 1 reports the estimates of R0 , α, and λ for different values of the initial parameters
when the RSS is computed in levels or in logs. When α and λ are reasonably far from the
upper and lower bound, as in Panel B, the estimate of R0 is fairly robust to changes in the
initial parameters. However, the left parts of Panels A and C demonstrate that when α is
close to its lower bound, the estimate of R0 can be very sensitive to the choice of initial
parameters. This issue seems to be more pronounced when the RSS is computed in levels
rather than logs. I have tried checking the sensitivity of R0 to the initial values by estimating
the model for several sets of initial values. Whenever possible, I pick the initial values so
that the constraints are not binding and the estimate of R0 is fairly stable.
The estimates of R0 for different values of clinical parameters σ and γ are presented in
Table 2. The upper panel of the table uses the RSS computed in levels, while the bottom
12
13
14
15
16
17
In this paper, I show that the SEIRD model for COVID-19 is poorly identified from the
short run data on deaths and reported cases. There can be many different models that are
indistinguishable in the short run but result in different estimates of unobserved variables and
markedly different long run forecasts. The estimated number of active COVID-19 cases in the
US based on observationally equivalent models can vary from several million to 140 million,
while the forecasted number of deaths from the epidemic ranges from tens of thousands to over
a million. Thus, this paper highlights that long run forecasts for COVID-19 heavily depend
on arbitrary choices made by the researcher. Available data cannot be used to determine
which forecast is correct because there are many models that are observationally equivalent
in the short run.
Next, I show that the basic reproduction number R0 is identified from the data, conditional
on the values of clinical parameters, and estimate it. My model takes into account possible
underreporting of cases. I show that there appears to be no single correct value of the basic
reproduction number R0 , as its estimates vary depending on clinical parameters. However,
for all values of clinical parameters the estimates of R0 are heterogeneous across countries:
they are 2-3 times higher in the US and Western countries than in Asian countries. I also
demonstrate that the estimates of R0 based on the confirmed cases data under the assumption
that all cases are reported are biased downward. The resulting models are inconsistent with
the observed data and dramatically overestimate the long run number of deaths.
Finally, I demonstrate that auxiliary information from random tests for COVID-19 can
help calibrate the initial values of the model and reduce the range of possible forecasts that
are consistent with the observed data. Random, or nearly random, tests were conducted in
Iceland, and utilizing the information from these tests leads to a 5-fold reduction in the range
of the forecasted number of deaths.
I do not take a stand on which of the models used and forecasts made, if any, is correct.
The model I consider is fairly simplistic and does not take into account important factors
18
19
The table presents the estimates of α, R0 , and λ for USA for γ = 1/10, σ = 1/5, and different initial parameters.
20
The table presents the estimates of R0 for different countries and different values of clinical parameters σ
and γ. Panel A computes RSS in levels. Panel B computes RSS in logs.
21
The table presents the estimates of R0 for different countries and different values of clinical parameters σ and
γ when E(t) is used in equation (2.6) instead of I(t). Panel A computes RSS in levels. Panel B computes
RSS in logs.
22
The table presents the estimates of R0 for different countries and different values of clinical parameters σ
and γ. Panel A computes RSS in levels. Panel B computes RSS in logs. Only the first 60 observations are
used in estimation. Note: results for Taiwan become unreliable when RSS is computed in logs due to the
very small number of reported cases and deaths in the data and are thus omitted.
23
The upper panel shows the short run number of deaths and reported cases for three sets of parameters θ1 =
(0.01, 5, 0.2, 2, 2, 2), θ2 = (0.005, 5, 0.1, 4, 4, 2), and θ3 = (0.004, 5, 0.08, 2, 2, 10), where θ = (α, R, λ, E0 , I0 , T0 ).
The middle panel shows the long run forecasts from these models. The lower panel fixes the initial conditions
and shows the short run number of deaths and reported cases for (α, R, λ) = (0.01, 5, 0.2), (0.01, 3, 0.2), and
(0.05, 3, 0.8).
24
The upper panel shows the logarithms of the short run number of deaths and reported cases for three sets
of parameters θ1 = (0.01, 5, 0.2, 2, 2, 2), θ2 = (0.005, 5, 0.1, 4, 4, 2), and θ3 = (0.004, 5, 0.08, 2, 2, 10), where
θ = (α, R, λ, E0 , I0 , T0 ). The lower panel fixes the initial conditions and shows the logarithms of the short
run number of deaths and reported cases for (α, R, λ) = (0.01, 5, 0.2), (0.01, 3, 0.2), and (0.05, 3, 0.8).
25
The upper panel shows the short run number of deaths and reported cases for three sets of parameters θ1 =
(0.01, 5, 0.2, 2, 2, 2), θ2 = (0.005, 5, 0.1, 4, 4, 2), and θ3 = (0.004, 5, 0.08, 2, 2, 10), where θ = (α, R, λ, E0 , I0 , T0 ).
The middle panel shows the long run forecasts from these models. The lower panel fixes the initial conditions
and shows the short run number of deaths and reported cases for (α, R, λ) = (0.01, 5, 0.2), (0.01, 3, 0.2), and
(0.05, 3, 0.8).
26
The upper panel shows the fit of the actual deaths and reported cases by models with four different values
of clinical parameters σ and γ. The middle panel shows the fit of the logarithms of the actual deaths and
reported cases for the same four models. The lower panel shows the forecasts from the same four models.
27
The upper panel shows the fit of the actual deaths and reported cases by models with four different values
of clinical parameters σ and γ. The middle panel shows the fit of the logarithms of the actual deaths and
reported cases for the same four models. The lower panel shows the forecasts from the same four models.
28
The upper panel shows the fit of the actual deaths for the models that use I(t) and E(t) in equation (2.6).
The middle panel shows the fit of the reported cases for the same models. The bottom panel shows the
deaths forecasts. The left panel computes RSS in levels. The right panel computes RSS in logs.
29
The upper panel shows the fit of the actual deaths and reported cases by models with four different values
of clinical parameters σ and γ. The middle panel shows the fit of the logarithms of the actual deaths and
reported cases for the same four models. The lower panel shows the forecasts from the same four models.
30
The upper panel shows the fit of the actual deaths and reported cases by models with four different values
of clinical parameters σ and γ. The middle panel shows the fit of the logarithms of the actual deaths and
reported cases for the same four models. The lower panel shows the forecasts from the same four models.
31
The upper panel shows the fit of the actual deaths and reported cases by models with three different sets of
initial conditions. The middle panel shows the fit of the logarithms of the actual deaths and reported cases
by the same three models. The lower panel shows the forecasts from the same three models.
32
The upper panel shows the fit of the actual deaths and reported cases by models with three different sets of
initial conditions. The middle panel shows the fit of the logarithms of the actual deaths and reported cases
by the same three models. The lower panel shows the forecasts from the same three models.
33
The upper panel shows the fit of the actual deaths and reported cases by models with different estimates of
R0 and different initial conditions. The middle panel shows the fit of the logarithms of the actual deaths and
reported cases by the same three models. The lower panel shows the forecasts from the same three models.
34
The figure presents the results from the validation set exercise for the US. The vertical line marks the end of
the training period. Observations to the left of the line are used to estimate the model. Observations to the
right are used to assess the accuracy of forecasts. The red model allows for possible underreporting of the
number of confirmed cases. The blue model assumes that all cases are reported.
35
The upper panel shows the fit of the actual deaths and reported cases by models with different estimates of
R0 and different initial conditions. The middle panel shows the fit of the logarithms of the actual deaths and
reported cases by the same three models. The lower panel shows the forecasts from the same three models.
36
The figure presents the results from the validation set exercise for the US. The vertical line marks the end of
the training period. Observations to the left of the line are used to estimate the model. Observations to the
right are used to assess the accuracy of forecasts.
37
The figure presents the results from the validation set exercise for the US. The vertical line marks the end of
the training period. Observations to the left of the line are used to estimate the model. Observations to the
right are used to assess the accuracy of forecasts.
38
The figure presents the results from the validation set exercise for the US. The vertical line marks the end of
the training period. Observations to the left of the line are used to estimate the model. Observations to the
right are used to assess the accuracy of forecasts.
39
The figure presents the results from the validation set exercise for the US. The vertical line marks the end of
the training period. Observations to the left of the line are used to estimate the model. Observations to the
right are used to assess the accuracy of forecasts.
40
The figure presents the results for Iceland. The left panel does not use any additional information. The right
panel matches the number of active COVID-19 cases on March 21 to the one estimated based on testing a
random sample of population. The upper panel shows the deaths fit by models with three different initial
values E0 . The middle panel shows the reported cases fit by the same three models. The lower panel shows
the deaths forecasts from these models.
41
——— (2020b): “Lockdowns and GDP: Is there a tradeoff?” Working paper, UCLA.
——— (2020c): “What Will Be the Economic Impact of COVID-19 in the US? Rough Esti-
mates of Disease Scenarios,” Working Paper 26867, National Bureau of Economic Research.
Ducrot, A., P. Magal, T. Nguyen, and G. F. Webb (2019): “Identifying the number
of unreported cases in SIR epidemic models,” Mathematical Medicine and Biology.
42
Korolev, I. (2020): “What Does the Case Fatality Ratio Really Measure?” Working paper,
Binghamton University.
Magal, P. and G. Webb (2018): “The parameter identification problem for SIR epidemic
models: identifying unreported cases,” Journal of mathematical biology, 77, 1629–1648.
Piguillem, F. and L. Shi (2020): “The Optimal COVID-19 Quarantine and Testing Poli-
cies,” Tech. rep., Einaudi Institute for Economics and Finance (EIEF).
43