14 Apr 20
@joe-shmo saidThis model underestimated the death number by very far.
I see what you mean about death being the better measure. I agree.
Thanks for the pointing out the Logistics Regression. I have an 89, didn't know it had the function. I used the Data Matrix App with the following data. This was the result: Starting at March 4 as Day 10 for the regression.
https://www.worldometers.info/coronavirus/country/us/
(Day, Deaths)
10 ...[text shortened]... its a nice fit for the data we currently have), but perhaps the measures we are taking are working?
That is the consequence of trying to a logistic model, when the inflection hasn't been reached.
14 Apr 20
@joe-shmo saidEven with a doubled max, the curve of Course failed miserbaly, since the inflection hadn't been reached.
I went ahead and did the full data set for US Deaths: ( I realized it was flawed as the origin of US Deaths was not until March 1. With the exceptions of ( 0,0 ) and ( 1,1 ). If I run the regression with those points I get a Singular Matrix. So I just took out rows until I got a result.
Data: ( same source as before: worldometer)
2 6
3 9
4 11
5 12
6 15
7 19
8 22 ...[text shortened]... 5
d = 10.519
The IP March 31st, 2300 Deaths.
Max hitting 4500 Deaths by Day 49 ( April 18th )
14 Apr 20
@eladar saidSince the exponential growth has been dampened this predicztion proved wrong also.
@DeepThought
This depends on the suppression strategy and whether social distancing can be effectively implemented. Suppose the measures the government is attempting to take in the UK are successful, then we're on the green curve and it should start going out of the exponential phase at some point in early-mid April, so in about 21 days
Seeing as this is day 24 yo ...[text shortened]... .
So you think we will have about 80 thousand dead by the time we stop seeing exponential growth.
14 Apr 20
@joe-shmo saidsame Problem with this one
@Eladar
I think I figured out the issue. The computation is crashing when it is working with numbers in the high hundreds/thousands ( its probably trying to do exponentials with those numbers and failing ). If I divide all the Y data by 10, it produces a curve. I just have to figure out how the actual curve relates to the adjusted curve. I suspect since I'm not adjustin ...[text shortened]... numbers are changing significantly. @Deepthought was obviously correct about very large error bars.
14 Apr 20
@joe-shmo saidThis was a hit.
Yeah, I've realized I have a scaling issue. I believe my logistical regression is failing to compute because of the sizeable data points ( My calculator has limited processing capabilities and its probably trying to exponentiate very large numbers ). I originally had thought that if I scaled the Y data by a factor, I could just multiply the resulting function by said facto ...[text shortened]... healthdata.org/
April 11 ( I believe your day 42 ) has between 17,642 - 26,602 Cumulative Deaths.
- Joined
- 10 Dec 06
- Moves
- 8528
14 Apr 20
6 edits
@ponderable saidPatting yourself on the back a little much? That's a nice way to present a near 100% Error. If you believed your model was going to reasonably hold until day 42, its also reasonable that it wouldn't be total garbage going just 1, 2 perhaps 3 days further in the future?
So time has fled.
The very trivial first shot I did here proved to overestiamte after day 34.
The overestimation for 21 days was About 50% which is okay for a very lean data fit of an exponential curve.
What is the percent error day 43, 44? What does it predict today. Day 45?
You show up at the end of our journey posting a string of comments on work that was revised and greatly improved upon weeks ago?
14 Apr 20
1 edit
@joe-shmo saidWell in fact I was just showing Eladar who seemed to have doubts about an exponential growth ( he pointed at three days with 50 death each and seemd to infer that the maximum had already been reached) that the data were compatible with exponential growth. It was quite clear from the beginning that exponential is only in the beginning of any growth function (see my post from 21st of March, page 3), since of course there are Limits to that growth. But the Limit can't be detected early on.
Patting yourself on the back a little much? That's a nice way to present a near 100% Error. If you believed your model was going to reasonably hold until day 42, its also reasonable that it wouldn't be total garbage going just 2 to 3 days further in the future?
What does your model predict by the end of today? Day 45?
Of Course it is quite clear that a good description of the Situation can only be made in retrospect. One owuld Need to compare death numbers for judging both cases:
* People who were in the dying process contacted additional the SARS-CV-2 (that would inflate numbers) and
* Hospitals with too much at Hands to care About giving numbers to authorities (reverse effect).
* If we do compare dying stats for the spring of 2020 nd say the average of 2014 to 2019 we can assume the Impact. (And that won't be the number of COVID-29 deaths alone).
- Joined
- 10 Dec 06
- Moves
- 8528
14 Apr 20
3 edits
@ponderable saidWell that's not clear to me, as you were attacking all my efforts in your random string of posts. You are commenting on work from 2 weeks ago. , the models had progressed. What is your % Error for Day 43, 44 ?
Well in fact I was just showing Eladar who seemed to have doubts about an exponential growth ( he pointed at three days with 50 death each and seemd to infer that the maximum had already been reached) that the data were compatible with exponential growth. It was quite clear from the beginning that exponential is only in the beginning of any growth function (see my post fro ...[text shortened]... e of 2014 to 2019 we can assume the Impact. (And that won't be the number of COVID-29 deaths alone).
14 Apr 20
@ponderable saidThe UK's Public Health England reports deaths of people in hospital who have tested positive for Covid-19. They do not report deaths that occur outside hospitals. They report all deaths, but what can happen is illustrated by this - someone is hospitalised on Monday and tested, they are moved to intensive care on Tuesday, die on Wednesday and their test results come back on Friday. They are included in Fridays figures and not Wednesdays. This means PHE have clear criteria for including people in the figures and don't go back attempting to correct the time series. That's fine as we know what their data means.
Well in fact I was just showing Eladar who seemed to have doubts about an exponential growth ( he pointed at three days with 50 death each and seemd to infer that the maximum had already been reached) that the data were compatible with exponential growth. It was quite clear from the beginning that exponential is only in the beginning of any growth function (see my post fro ...[text shortened]... e of 2014 to 2019 we can assume the Impact. (And that won't be the number of COVID-29 deaths alone).
If people are dying and contract Covid-19 then it arguable contributes to their deaths. I think that such cases are such a small minority we can ignore them.
The Office for National Statistics produce complete mortality figures based on death certificates for which provisional figures come out about a fortnight after the fact. These include all deaths where the cause of death was entered as Covid-19, whether in hospital or not. About 10,000 people died per week during March in 2019.
14 Apr 20
@joe-shmo saidWell the Point here was that your attempts at logistic models were made after my initia Posts and fell short far more than my original trivial exponential growth model, which I never claimed to be valid beyond day 42. In fact I wrote already in 21st of March that it was only for the beginning of the curve and that a reasonable calculation f a potential maximum was not feasible.
Well that's not clear to me, as you were attacking all my efforts in your random string of posts. You are commenting on work from 2 weeks ago. , the models had progressed. What is your % Error for Day 43, 44 ?
As others poinetd out in the thread one would need more than the sheer Progression of numbers to make a good model. And of Course the effect of measures is not so easy to include, as well as vulnerability.
Conclusion:
* It was shown that there has been an exponential growth of death as attributed to Covid-19, using numbers by woldometers has occured.
* The growth has (for the available data) ceased to be exponential. So measures have an effect (which is actually good News).
14 Apr 20
@deepthought saidThen the UK is faster than the Federal republic of Germany. Here our most recent numbers are from Nov. 2019 (https://www-genesis.destatis.de/genesis/[WORD TOO LONG]
The UK's Public Health England reports deaths of people in hospital who have tested positive for Covid-19. They do not report deaths that occur outside hospitals. They report all deaths, but what can happen is illustrated by this - someone is hospitalised on Monday and tested, they are moved to intensive care on Tuesday, die on Wednesday and their test results come back ...[text shortened]... ed as Covid-19, whether in hospital or not. About 10,000 people died per week during March in 2019.
In Germany we have About 80000 deeath per month, so a Variation of 3000 (actual number of corona death) would still be in the normal fluctuation...But if you have Weekly numbers for the UK it could already Show as significant.
- Joined
- 10 Dec 06
- Moves
- 8528
@deepthought saidWell, The Gaussian model in the US with the data between from day 15 to 41 is a very good fit for the data.
I implemented the IHME method for the UK and get vastly different answers depending on where I start the linear regression. I had the entire population of the UK dying if I included the earlier data. My conclusion from this is that it's difficult to get total deaths from curve fitting, using the logistic curve my best fit gives about 18,000 deaths. By fitting data by e ...[text shortened]... I can get any result you want depending on where I take the included data in the linear regression.
The Quadratic Regression has a R² = 0.987
Y = b*x² + c*x + d
The coefficients are:
m = -b = 0.006019626
a = c/(2m)= 44.32171905
d = ln(A) - m*a² = -4.15358265
A = 2146
Plotting g(x) = A*e^(-m*(x-a)² )
Is a very reasonable fit to the data at that point.
Now, IHME from the inflection point on is predicting an almost linear regression for g(x).
The Gaussian under predicts that model
And the Data is strongly under performing the Gaussian.
IMHE must have some substantial factors in consideration to justify this overestimation
April 12 they projected 1522 , Actual 1528
April 13 the Project 2150 ( There is a clear discontinuity in the projection ), Actual 1535
April 14 ( Today ) they project 1,953
I think there model is relying heavily on the effects of failed state of US health care that is just completely wrong at this point in time.
15 Apr 20
@eladar saidThis explains it. The state totals require testing, the worldometer numbers do not.
One problem with models is who determines the numbers. Case in point
The State of New York reports 6536 people have died of Coronavirus in New York.
Worldometer says 10834
New York numbers...https://www1.nyc.gov/site/doh/covid/covid-19-data.page
https://www.marketwatch.com/story/new-york-city-coronavirus-death-toll-jumps-as-city-includes-probable-deaths-2020-04-14