MODERATOR NOTE: coronavirus resources & updates: ⭐ [COVID] Computational Research HUB For Novel Coronavirus: Data, Code, Visualizations, Notebooks - Staff Picks - WOLFRAM COMMUNITY
– you have earned Featured Contributor Badge 
Your exceptional post has been selected for our editorial column Staff Picks http://wolfr.am/StaffPicks and Your Profile is now distinguished by a Featured Contributor Badge and is displayed on the Featured Contributor Board.
This post has been listed in the main resource-hub COVID-19 thread: https://wolfr.am/coronavirus in the section Computational Publications. Please feel free to add your own comment on that discussion pointing to this post ( [NB] Predicting Coronavirus Epidemic in United States ) so many more interested readers will become aware of your excellent work. Thank you for your effort!
Oh dear. This model seems to be using RATE information as its source - which leads to the usual problems with noisy inputs.
No, the fits are done with the cumulative case data, then the fit is used to predict what the daily rate changes would be according to the fit of the cumulative case data up to that point. Growth rate changes are what will be noticed in the news. The fits raise the interesting question that the early data may not be very good as the rates derived from the daily log differences of the reported cases do not fit any pattern. Here are the plots for New Jersey including today’s data.
Using the default fitting, which uses Norm to minimize the fit residuals, the early data are not adding much to the fit as shown by the last log plot. Either the dynamics are changing after the peak or the early data are not very good. The plots of the early daily rates show a chaotic pattern suggesting the problem may be with the early data. Using the default fitting method to the cumulative data limits the influence of the early data points, so hopefully the predictive value gets better with more data, but the data from South Korea looked a little peculiar at the end, but the late case increases were not huge even though they failed to fit the model.
I looked over the three plots above and I’m guessing that there was a logarithmic fit, judging by the log ordinate for the case plot. If you differentiated your logistic function, you SHOULD have seen the linear scaled case rate growing in the usual bell shaped curve for a maximum. Your "Daily Log case differences " did not show the desired shape, so you did not have the opportunity to refit a non linear regression to the post-peak case data - to show the departure of the logistic parameters at later dates to a higher asymptotic level.
Let me show you the Oklahoma deaths and death rate.
Below is the Wolfram Alpha plot of daily Death rate
Day 1 is March 23rd, peak occurs around April 7th
I notice that when I did the same for Cases and Case Rate, the daily fit for Cases needed daily revision after the day for the Peak case rate. When I discarded the pre-peak date data points and ran a new regression, the new best fit parameters showed a considerably higher plateau, suggestiing the bell-shape time series is not in fact symmetrical.
Sincerely
Brian Whatcott Altus.
Robert,
please take a look at the notebook below, which generously copied from yours, analysing both positives and number of deaths. Two new ideas are:
- Introduction of a seven-day moving average, which seem to reveal some interesting behavior, and
- The addition of a manipulate environment that allows one to easily see how the projections change with the availability of new data. I wonder if someone has any thoughts on the somewhat unexpected behavior of the data.
Dietmar,
I am not sure what to make of the dynamic the moving average introduces, but mid March most of the datasets show a jump in cases for one day, but that wouldn’t affect deaths. I think probably the cyclic phenomenon is introduced by using a linear moving average on data that is more or less growing exponentially. I think if you want to do smoothing that is consistent with the data, you should try an ExponentialMovingAverage, I have also looked at the parameters as the data accumulates. The L parameter in general keeps growing, while k decreases. My best answer is that the early data doesn’t have any particular shape to it, rather it is chaotic – you could fit most anything to it. There is probably good reason for this. Public health officials were playing catch up; the spread of the virus was much wider than was realized. So initially there was a mix of unrestrained viral spread and minimal quarantine. So we are seeing a situation that doesn’t fit any model. Eventually the quarantine effort becomes more effective, but it may not become perfect and behave as a leaky quarantine. In the end the South Korea data, showed a persistent growth not at all consistent with the logistic model. Notebook attached.
Another interesting way to look at models is that the logistic model is the same as multiplying L times any unimodal statistical distribution. I am also attaching a notebook showing what fitting with that idea shows. Basically if we want to predict when it will end, The tail behavior of the distribution is what is important, and it is the right tail that will be important, because it doesn’t really matter if timing screwed up the initial data accumulation. The ultimate conclusion I think is that you cannot make predictions unless you know in advance the shape of the entire curve. If the tail behavior is exponential, then the logistic distribution should eventually succeed, but it wasn’t in S. Korea.
One thing that probably is predictable is that once there is a clear decline in case rates, it should eventually end, and we may be able to detect the shape of the decline. Right now we are probably still too close to the peak in the U.S.
The attached notebooks are not annotated, but you clearly will be able to understand them.
SKorea.nb (426 KB)
StatidticalModels.nb (285 KB)
On your last remark, yes, that is the issue with trying to predict the future: It’s really hard to do in advance; it’s much easier to “predict” (=explain) the future once it’s not the future anymore, but that’s not as helpful in our current situation.
This is also the difficulty with those SIR/SEIR/etc. “first-principle” models. They are really pretty, and work alright once all the parameters are known with sufficient accuracy. Unfortunately it turns out that we typically have a good grasp of those parameters only after the epidemic is over.
Agreed. Most of the epidemiologic models are focused on the dynamics of the disease and recovery rates. Effective quarantine, however, can shut down an epidemic very rapidly, and the epidemiologic models also require transmission assumptions that are not known for a new disease.
I do suspect the S. Korea finale was unusual, and that we should be able to predict the rate of decline from the data pretty soon, then modelling with the tail behavior of statistical distributions might prove useful. If the decline is faster than exponential, then the gamma distribution offers the range of tail behaviors between exponential and normal distributions, if it is slower, then a Pareto distribution could model it. If there are recurrent new small outbreaks, which might be what happened in S. Korea, then no mathematical model will succeed.
I just updated the notebook I posted with the latest data.
This does not look good. Looks like the United States is turning the corner, the wrong way…
Robert, you asked about my data source. I use Worldometer.info. It provides data for a number of countries, and includes US by state - but no smaller divisions. I collect it daily. With a little effort, you can recapture prior data by clicking data points, as I recall
Brian W
Brian,
Just clicked on the link, they are now reporting 45318, deaths for 4/22. I don’t know how often they update, but it eventually may be showing up there as well. Are you analyzing with Mathematica?
We are getting to an interesting point in the epidemic, which should decide the fate of the logistic model. The logistic model requires an exponential decline, which is easiest to show in the log plot of the daily differences.
In the picture the orange dots are the daily differences of the raw data and the curve is the calculated daily difference from the logistic fit to the cumulative data as a continuous curve which is very close to the first derivative curve. If the points don’t follow something close to a log linear decline, I don’t think the logistic model holds, but with data anomalies it may be difficult to track. I am finding that the death data is lagging the case data by 7 to 8 days. In the end the death data may be more reliable, especially if people start reporting antibody tests as positive cases.
Bob
Day #48 of my data was rather like that: a step increase when NY State and several others decided to include non-hospital cases and fatalities in their totals. I excluded two prior days data - which is itself a kind of massaging of the input - an original sin of sorts.
Dietmar,
I believe you hit several nails on the head with your critique of logistical models. They necessarily provide the same time constant for the asymptotic approach as for the exponential increase. Moreover, when the daily growth rate becomes linear, the differential of the logistic function puts the middle of the straight line growth as the peak of the daily rate. I am pleased by an improved model function that David Bowman at Georgetown proposed. At the cost of a fourth, fitting parameter, the exponential growth time constant is uncoupled from the time constant of the asymptotic approach.
DEATHS = k1 / (1 + aexp(ab*(starting day number - present day number)))^(1/c)
It too suffers - this time from its virtue: the trailing exponential can easily follow a short term hollow - so that I find it advantageous to train the non-linear fitting with a fictitious value for a datum several days in the future, which best matches the recent early data points of the slowing daily numbers - so that the fit does not hunt unnecessarily to find the least squares fit of the linear data. Sadly the only data plot where this is presently useful is US Deaths, which appear to be turning back: other plots I maintain are currently best served by straight line models, which provide a direct value for daily rates.
Brian W
Dietmar,
I believe you hit several nails on the head with your critique of logistical models. They necessarily provide the same time constant for the asymptotic approach as for the exponential increase. Moreover, when the daily growth rate becomes linear, the differential of the logistic function puts the middle of the straight line growth as the peak of the daily rate. I am pleased by an improved model function that David Bowman at Georgetown proposed. At the cost of a fourth, fitting parameter, the exponential growth time constant is uncoupled from the time constant of the asymptotic approach.
DEATHS = k1 / (1 + aexp(ab*(starting day number - present day number)))^(1/c)
It too suffers - this time from its virtue: the trailing exponential can easily follow a short term hollow - so that I find it advantageous to train the non-linear fitting with a fictitious value for a datum several days in the future, which best matches the recent early data points of the slowing daily numbers - so that the fit does not hunt unnecessarily to find the least squares fit of the linear data. Sadly the only data plot where this is presently useful is US Deaths, which appear to be turning back: other plots I maintain are currently best served by straight line models, which provide a direct value for daily rates.
Brian W
About that step change in Deaths: I notice Worldometers updated US DEATHS about an hour ago - local time, actually 04:17 GMT tomorrow! Here’s what that one and only decaying plot looks like now:
Their usual update happens around noon. Looks like that double exponential may go back to its former linear growth…
Brian W