See discussions, stats, and author profiles for this publication at: [Link]

net/publication/339240777

Estimation of the final size of coronavirus epidemic by the logistic model

Method · February 2020

CITATION READS

1 19,679

1 author:

Milan Batista
University of Ljubljana
138 PUBLICATIONS   559 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Port of Koper Industrial Risk Assessment View project

Navigability Study; Tanker Berth ”JET” View project

All content following this page was uploaded by Milan Batista on 18 March 2020.

The user has requested enhancement of the downloaded file.

18.03.2020 19:46

Estimation of the final size of the coronavirus epidemic by the logistic

model (Update 4)

Milan Batista
University of Ljubljana, Slovenia
[Link]@[Link]
(Mar 2020)

Abstract
In the note, the logistic growth regression model is used for the estimation of the final
size and its peak time of the coronavirus epidemic in China, South Korea, and the rest
of the World.

1 Introduction

In the previous article [1], we try to estimate the final size of the epidemic for the whole
World using the logistic model and SIR model. The estimation was about 83000 cases.
Both models show that the outbreak is moderating; however, new data showed a linear
upward trend. It turns out that the epidemy in China was slowing but is begin to
spread elsewhere in the World.

In this note, we will give forecasting epidemic size for China, South Korea, and the rest
of the World and daily predictions using the logistic model. We assume that the model
is a reasonable description of the epidemic. Full daily reports for China, Iran, Italy,
Slovenia, South Korea and counties outside of China generated are available as linked
data at

[Link]
-19_epidemic_size

The MATLAB program fitVirus used for calculations is freely available from

[Link]

We note that logistic models give similar results as the SIR model (at least for the case
of China and South Korea). However, the logistic model is given by explicit formula
and is thus much simpler for regression analysis than the SIR model, where one must
1
18.03.2020 19:46

on each optimization step solve a system of ordinary differential equations. (One may,
however, use approximate solution and thus obtain four-parameter problem which can
be very sensitive to initial guess). Yet, the logistics model has its drawbacks as the
epidemic approaches its final stage: the actual number of cases may be slightly larger
than that predicted by the logistics model. If the actual number of cases begins to
exceed the predicted end-state systematically, then a second phase of the epidemic is
likely to occur, and the model will no longer be applicable.

2 Logistic grow model

In mathematical epidemiology, when one uses a phenomenological approach, the

epidemic dynamics can be described by the following variant of logistic growth model
[2-5]

dC  C
 rC 1   , (1)
dt  K 

where C is an accumulated number of cases, r  0 infection rate, and K  0 is the

final epidemic size. If C 0  C 0  0 is the initial number of cases then the solution of
(1) is

K
C  , (2)
1  A exp rt 

K C 0
where A  When t  1 , assuming K  C 0 , and therefore A  1 we have
C0
the natural growth

Ke rt C0 e rt
C  rt   C 0e rt (3)
e  A 1 C 0 K 1  e A
rt

When t   the number of cases follows the Weibull function

C  K 1  Ae rt    K 1  e
r t t0  
 
 

(4)

dC d 2C
The growth rate reaches its maximum when  0 . From this condition, we
dt dt 2
obtain that the growth rate peak occurs in time time

2
18.03.2020 19:46

ln A
tp  (5)
r

At this time the number of cases is

K
Cp  , (6)
2

and the growth rate is

dC 
   rK (7)
 dt  4
p

To answer the question about doubling time t , i.e., the time takes to double the
number of cases we solve C t  t   2C t  for t . Result is

ln 2 1  1 
t   ln   e rt   t (8)
r r  A 

The first term represents initial exponential growth, then t increases with t. When
ln A
t  tp  , i.e., when C  K 2 , then t   . For C  K 2 doubling time lost
r
its meaning.

Figure 1. Doubling time (data up to 14 mar 2020)

3
18.03.2020 19:46

Now, if C 1,C 2 , ,C n are the number of cases at times t1, t2 , , tn , then the final size
predictions of the epidemic based on these data are K 1, K 2 , , K n . When convergence
is achieved, then one may try to predict the final epidemic size by iterated Shanks
transformation [6]

K n 1K n 1  K n2
K  . (9)
K n 1  2K n  K n 1

There is no natural law or process behind this transformation; therefore, it must be

used with some care. In particular, the calculated limit is useless if K  C n , i.e., it is
below the current data.

The logistic model (2) contains three parameters: K, r, and A, which should be
determined by regression analysis. Because the model is nonlinear, some care should be
taken for initial guess. First of all, in the early stage, the logistic curve follows an
exponential growth curve (3) , so the estimation of K is practically impossible. With
enough data, the initial guess can be obtain in the following way. Expressing t from (2)
and use three equidistant data point yield the following system of three equations:

1  AC k 2m  1  AC k m  1  AC k 

tk  2m  ln  , tk  m  ln   ,…. tk  ln   (10)
r  K  C k 2m  r  K  C k m  r  K  C k 

This system has a solution [7]

C k m C k 2mC k m  2C k 2mC k  C k mC k 

K  . (11)
C k2m  C kC k 2m

1 C k C k m  C k 2m 
r ln . (12)
m C k 2m C k  C k m 

tk m

C k  C k m C k m  C k 2m  C k C k m  C k 2m  m

A   . (13)

C k2m  C kC k 2m  C k 2m C k  C k m 
 

The solution is acceptable when all the unknowns are positive.

Formulas (10),(11),(12) are used to calculate the initial approximation in the fitVirus03
program. For practical calculation, we take the first, the middle, and the last data
point. If this calculation fails, we consider regression analysis as questionable. Using
4
18.03.2020 19:46

the calculated initial guess, the parameters K, r, and A are then calculated by least-
square fit using the MATLAB functions lsqcurvefit and fitnlm.

Figure 2. Epidemic phases

Before we proceed, we for convenience, introduce the following epidemy phases (see Fig
2):

1. Phase 1 – exponential growth (lag phase, slow growth) : t  t p  2 r

2. Phase 2 – fast growth (positive growth phase, acceleration phase) to an epidemy

turning point: t p  2 r  t  t p

3. Phase 3 – fast growth to steady-state (negative growth phase, deceleration

phase): t p  t  t p  2 r

4. Phase 4- steady-state (transition phase, slow growth, asymptotic):

t p  2 r  t  2t p

5. Phase 5 – steady) ending phase (plateau stage): t  2t p

The duration of the fast-growing period is thus equal to   4 r . We note that the

names of the phases are not standard, and are arbitrarily chosen.

5
18.03.2020 19:46

3 Results

3.1 China ([Link] 2020)

On the base of available data, one can predict that the final size of coronavirus epidemy
in China using the logistic model will be approximately 81 000 ± 500 cases (Table 1)
and that the peak of the epidemic was on 8 Feb 2020 (Table 2). It seems that the
epidemic in China is in the ending stage (Fig 3, Fig 4).

The short-term forecasting is given in Table 3 where we see that the discrepancy of
actual and forecasted number of cases is within 2%. However, actual and predicted
daily new cases are scattered and vary between 13% to 300%. On 7 Mar 2020, the
actual number of cases was 80695, and the daily number of cases was 44. Prediction in
Table 3 is cumulative 80588 cases and 39 daily cases. The errors are 0.1% and 11%,
respectively.

Table 1. Estimated logistic model parameters for China (data up to [Link] 2020)

Estimate SE tStat pValue

________ _________ ______ __________

K 80772 489.04 165.16 2.0205e-72

r 0.22568 0.0062941 35.855 2.4094e-38
A 191.23 26.978 7.0882 3.5795e-09

Number of observations: 55, Error degrees of freedom: 52

Root Mean Squared Error: 1.9e+03
R-Squared: 0.997, Adjusted R-Squared 0.997
F-statistic vs. zero model: 1.61e+04, p-value = 4.02e-77

6
18.03.2020 19:46

Coronavirus epidemic in China

90 5000

80 4500

4000
70
Predicted
Actual 3500
60
Infection rate
3000
50
Infected (1000)

Infected/day
2500
40
2000

30
1500

20
1000

10 500

0 0
01/02/20 01/03/20
Date

Figure 3. Predicted evaluation of coronavirus epidemic in China

Estimated final size of epidemic for China

100

90

80

70

60
Cases (1000)

50

40
SIR mode

Logistic model
30

20

10

0
26 28 30 32 34 36 38 40 42 44 46

Day after [Link] 2020

Figure 4. Predicted final size of coronavirus epidemic in China (prediction from 2.

Mar 2020)

7
18.03.2020 19:46

Table 2. Results of daily logistic regression for China (data from [Link] 2020)

day date C K r A t_peak dC_peak date_peak

(cases) (cases) (1/day) (day) (cases/day)
16 [Link].20 11789 17884 0.452 460.109 13 2019 [Link].20
17 [Link].20 14378 21578 0.415 398.528 14 2238 [Link].20
18 [Link].20 17203 25879 0.383 353.422 15 2476 [Link].20
19 [Link].20 20438 31161 0.354 319.58 16 2753 [Link].20
20 [Link].20 24332 38706 0.324 294.507 17 3136 [Link].20
21 [Link].20 28028 44255 0.308 283.606 18 3406 [Link].20
22 [Link].20 31161 45828 0.303 279.609 18 3476 [Link].20
23 [Link].20 34546 47586 0.298 272.671 18 3545 [Link].20
24 [Link].20 37198 48066 0.296 269.981 18 3561 [Link].20
25 [Link].20 40171 49272 0.292 261.091 19 3594 [Link].20
26 [Link].20 42638 50275 0.288 252.341 19 3614 [Link].20
27 [Link].20 44438 50724 0.286 247.719 19 3621 [Link].20
28 [Link].20 59800 70500 0.226 154.092 22 3989 [Link].20
29 [Link].20 63850 92522 0.198 138.603 24 4585 [Link].20
30 [Link].20 66492 99501 0.192 137.601 25 4785 [Link].20
31 [Link].20 68500 96788 0.195 138.393 25 4712 [Link].20
32 [Link].20 70548 93100 0.199 141.105 24 4625 [Link].20
33 [Link].20 72436 90217 0.203 144.985 24 4569 [Link].20
34 [Link].20 74199 88248 0.206 149.059 24 4540 [Link].20
35 [Link].20 74576 86134 0.21 155.327 24 4519 [Link].20
36 [Link].20 75464 84630 0.213 161.403 23 4512 [Link].20
37 [Link].20 76288 83577 0.216 166.902 23 4513 [Link].20
38 [Link].20 76936 82810 0.218 171.816 23 4518 [Link].20
39 [Link].20 77150 82149 0.22 176.855 23 4525 [Link].20
40 [Link].20 77658 81677 0.222 181.054 23 4533 [Link].20
41 [Link].20 78064 81329 0.223 184.574 23 4541 [Link].20
42 [Link].20 78495 81087 0.224 187.292 23 4547 [Link].20
43 [Link].20 78824 80912 0.225 189.429 23 4552 [Link].20
44 [Link].20 79251 80805 0.226 190.85 23 4556 [Link].20
45 [Link].20 79824 80774 0.226 191.288 23 4557 [Link].20
46 [Link].20 80026 80755 0.226 191.566 23 4558 [Link].20
47 [Link].20 80151 80741 0.226 191.792 23 4558 [Link].20

8
18.03.2020 19:46

Table 3. Short-term forecasting for China

Daily Daily
Day Date Actual Predicted Error % Error %
actual predicted
44 [Link].20 79251 79814 0.945 427 232 45.667
45 [Link].20 79824 80000 0.407 573 186 67.539
46 [Link].20 80026 80149 0.302 202 149 26.238
47 [Link].20 80151 80268 0.265 125 119 4.8
48 [Link].20 - 80363 - - 95
49 [Link].20 - 80439 - - 76
50 [Link].20 - 80500 - - 61
51 [Link].20 - 80549 - - 49
52 [Link].20 - 80588 - - 39

3.2 South Korea ([Link] 2020)

On the base of available data, one can predict that the final size of coronavirus epidemy
in of South Korea using the logistic model will be approximately 8050 ±70 cases (Fig
5, Table 4) and that the peak of the epidemic was on 1 Mar 2020. The epidemic in
South Korea appears to be in the steady-state transition phase. These figures were
already predicted on 4. Mar 2020 (Table 5), i.e., the prediction was approximately 7500
to 8500 cases and that the peak will be around 2 Mar.

On 7 Mar 2020, the actual number of cases was 7134, and the daily number of cases
was 367. Prediction in Table 5 is cumulative 6572 cases and 259 daily cases. The errors
are 8 % and 30%, respectively.

Table 4. Estimated logistic model parameters for South Korea up to [Link] 2020

Estimate SE tStat pValue

________ ________ ______ __________

K 8046.5 71.784 112.09 7.8279e-32

r 0.36657 0.010259 35.733 5.581e-21
A 87.97 9.8747 8.9087 9.4612e-09

Number of observations: 25, Error degrees of freedom: 22

Root Mean Squared Error: 129
R-Squared: 0.998, Adjusted R-Squared 0.998
F-statistic vs. zero model: 1.26e+04, p-value = 1.03e-35

9
18.03.2020 19:46

Coronavirus epidemic in South Korea

9 800

8
700

7
Predicted 600

Actual
6
Infection rate
500

5
Infected (1000)

Infected/day
400

300
3

200
2

100
1

0 0
14/02/20 19/02/20 24/02/20 29/02/20 05/03/20 10/03/20 15/03/20 20/03/20 25/03/20

Date

Figure 5. Predicted evaluation of coronavirus epidemic in South Korea ([Link] 2020)

Daily estimated final size of epidemic for South Korea

16

14

12

10
Infected (1000)

0
01/03/20 08/03/20 15/03/20

Date

Figure 5a. The predicted final size of the epidemic in South Korea ([Link] 2020)

10
18.03.2020 19:46

Table 5. Results of daily logistic regression for South Korea

day date C K r A t_peak dC_peak date_peak

(cases) (cases) (1/day) (day) (cases/day)
7 [Link].20 833 1105 0.824 47.332 4 227 [Link].20
8 [Link].20 977 1133 0.808 46.526 4 228 [Link].20
9 [Link].20 1261 1508 0.633 37.083 5 238 [Link].20
10 [Link].20 1766 3355 0.426 44.537 8 357 [Link].20
11 [Link].20 2337 7021 0.363 76.656 11 636 [Link].20
12 [Link].20 3150 17145 0.328 165.531 15 1405 [Link].20
13 [Link].20 3736 8639 0.359 95.534 12 775 [Link].20
14 [Link].20 4335 7214 0.378 88.434 11 682 [Link].20

Table 6. Short-term forecasting for South Korea

Daily Daily
day date actual predict Error % Error %
actual predict.
11 [Link].20 2337 2393 29.7 571 562 1.6
12 [Link].20 3150 3031 17.7 813 638 21.5
13 [Link].20 3736 3708 17.2 586 677 15.5
14 [Link].20 4335 4378 15.3 599 670 11.9
15 [Link].20 - 4997 - - 619
16 [Link].20 - 5532 - - 535
17 [Link].20 - 5971 - - 439
18 [Link].20 - 6313 - - 342
19 [Link].20 - 6572 - - 259
20 [Link].20 - 6761 - - 189

3.3. Rest of the World

The comparison of the predicted final sizes is shown in the graph in Figure 6. Based
on the data from 18. Mar 2020, a very rough estimate indicates that the number of
cases will be about 6000 00 (Fig 7); however, it is an early-stage epidemic, so this
estimate ise very questionable and will be changed with new data.

11
18.03.2020 19:46

4
Coronavirus epidemic in outside of China 10
600 3

Predicted

Actual

Infection rate
500 2.5

400 2
Infected (1000)

Infected/day
300 1.5

200 1

100 0.5

0 0
10/01/20 30/01/20 19/02/20 10/03/20 30/03/20 19/04/20 09/05/20 29/05/20 18/06/20 08/07/20

Date

Figure 6. Predicted evaluation of coronavirus epidemic outside of China (18 Mar

2020)

Daily estimated final size of epidemic for outside of China

1000

800

600
Infected (1000)

400

200

0
08/03/20 15/03/20

Date

Figure 7. Predicted final size of coronavirus epidemic outside of China ([Link] 2020)

4 Conclusion

On the base of available data, one can predict that the final size of coronavirus epidemy
in China will be around 81 000 cases. For South Korea, a prediction is about 8000

12
18.03.2020 19:46

cases. For the rest of the World, the forecasts are still very unreliable in is now
approximatly 380 000 cases.

We emphasize that the logistic model is a data-driven phenomenological model. Thus

its predictions are as good as useful data are and as good as it can mimics epidemy
dynamics. Because, as it said, the logistic model is phenomenological, it's in a way
mimic both epidemy spreading and its control i.e. prevention measures. When daily
predictions of epidemic size begin to converge, we can say that epidemy is in control.
Any deviation from the prediction curve may indicate that epidemy may be running
out of control (Fig 8). This happens when one looks at the data for China, were up to
say 25. Feb accumulated cases follow the logistic curve, and then they begin to deviate
from it. We now know that this was the beginning of the second stage of the epidemic,
which is now spreading all over the World. A similar linear trend can now be observed
for South Korea (Fig 5); we hope that this does not indicate the start of a second stage
of the epidemic in this country.

Figure 8. The uncontrolled and controlled phase of the epidemic (case of China)

References

[1] M. Batista, Estimation of the final size of the COVID-19 epidemic, medRxiv, (2020)
2020.2002.2016.20023606.

13
 18.03.2020 19:46

[2] X.S. Wang, J.H. Wu, Y. Yang, Richards model revisited: Validation by and
application to infection dynamics, J Theor Biol, 313 (2012) 12-19.
[3] B. Pell, Y. Kuang, C. Viboud, G. Chowell, Using phenomenological models for
forecasting the 2015 Ebola challenge, Epidemics, 22 (2018) 62-70.
[4] F. Brauer, Mathematical epidemiology: Past, present, and future, Infectious Disease
Modelling, 2 (2017) 113-127.
[5] S.L. Chowell G, Viboud C, Kuang Y., West Africa Approaching a Catastrophic
Phase or is the 2014 Ebola Epidemic Slowing Down? Different Models Yield Different
Answers for Liberia. , PLOS Currents Outbreaks., (2014).
[6] C.M. Bender, S.A. Orszag, Advanced mathematical methods for scientists and
engineers I asymptotic methods and perturbation theory, Springer, New York, 1999.
[7] P.F. Verhulst, Notice sur la loi que la population poursuit dans son accroissement,
Correspondance mathématique et physique, 10 (1838) 113-121.

14

View publication stats