IHME model needs updating badly

[Loki Liesmith]

https://covid19.healthdata.org/united-states-of-america

Their median estimate is still 60k deaths between now and August.

If we are lucky we are just peaking now worth ~ 40k dead. Assuming a completely symmetric peak which is a best case scenario that’s at least 80k deaths by midsummer. But we know the plateau on case growth is a long one, 80k has got to be a lower bound.

On the upside their upper bound confidence envelope is really high so at least they’ve got the upside risk right.

On edit: their lower bound estimate is LOWER than the actual number of deaths right now. What does that even mean?

[sfstaxprep]

1. We May Reach That Before End Of April

I can see the original minimum prediction of 100K being hit in May.

[lapfog_1]

2. and all of that depends on staying locked down until the viral reservoir is near zero

and we have extensive testing to ensure that number stays there.

I don't see those conditions happening

[uponit7771]

3. K&R, That's if we stay at the current mitigation measures right? Once the MAGA Munchers let ...

... themselves out on the beaches & parks to breath each others faces things will change.

[Loki Liesmith]

4. We need a derogatory label for those fools. Something that makes them socially contagious.

Spreaders?

[uponit7771]

11. 👍, furtalizeres ?

[Amishman]

17. They'll wear it as a badge of honor, like 'Deplorables'

Shaming doesn't work very well anymore.

[Shermann]

5. Current social distancing assumed until infections minimized and containment implemented

Those are pretty bold assumptions.

https://www.democraticunderground.com/100213313222

[Loki Liesmith]

7. Model was looking ok last week.

They have not refit it to new data. Bad practice.

[Shermann]

10. We are still social distancing. That's the easy phase.

R{ .9 } or so. You can run a SIR simulation and get the same result.

The big unknowns are the R-naughts during Phase 1, 2, and 3.

[Igel]

14. The revised data are a problem.

The way some states reported it was as a single spike on one day--and not all states picked the same day and I'm not even sure all states have started that reporting.

No way to break down that spike over time until the states provide that information. Include it, and you have a real outlier that'll skew the curve in predictable but really undesirable ways. Leave out those spikes but use new data going forward, and you still have a step in your e-based curve, and that also deforms the curve. Leave out the add-on deaths (since the CDC wants them reported separately) and you get a clean curve that might be useful for predicting trends, but you leave out the out-of-hospital/untested deaths.

Retrofit to the bad data that's the new data, and you get a mess for your curve.

Bad practice is not to use the new data. Ever.

Less bad practice is to leave out the data as far as curve-fitting goes then, when there's enough data so that the offset for each date after the new reporting kicks is known you find a nice multiplier for your curve. Or just chuck the old curve and start from scratch, leaving out the old data entirely.

Best practice would have been the states' revising their reports for each date retroactively, which is what places like IMHE last I heard were trying to get the states to do. But that's not a best practice that IMHE or worldometer can implement.

Having a correct count was important more for politics than for forecasts. The final number will be different, in any event. Public health needs to have accurate predictions and for that we needed consistent data. As long as the number of out-of-hospital/untested victims stayed relatively proportional to those counted--an assumption, but it's hard to see why it wouldn't be fairly valid--the old data provided an adequate basis. Nobody expects IMHE's projection to be accurate to the day and within 5 dead, at least not anybody with any sense in their skulls.

[Squinch]

6. Ya think?

[Loki Liesmith]

8. Looked better last week

They even temporarily revised it upward to ~70k. Which was plausible at the time. Then inexplicably revised it downward again.

[Silent3]

9. For some strange reason, their model seems to assume or predict...

...a daily death rate from COVID-19 at or near zero by mid May. That makes no sense at all.

In states with lax safety measures, or no safety measures at all, I suspect those places will either be at or heading toward their peak death rates around mid May.

[Igel]

12. It means that their algorithm predicted something that didn't happen.

That's all.

Update the projection using current data, that problem goes away for the days up through the present.

[Loki Liesmith]

13. I get that, not really the point I was making

It was already wrong or close to it at the last update. The lower bound was effectively impossible when they made the prediction. That’s a model I’d revise the moment I saw the output.

[BGBD]

15. IHME has been very conservative

It is essentially a best-case scenario that makes a lot of assumptions that don't seem to be true right now.

I think this is a more accurate model.

It shows a median projection of close to 60k deaths by the end of April.


https://covid-19.tacc.utexas.edu/projections/

[Azathoth]

16. Look at their confidence intervals. They're predicting 65K beds today but...

with a confidence interval of like 15K to 220K. That's almost cartoonish from a useful forcasting standpoint. It's like the weatherman telling you that it will rain today unless it doesn't. You don't say, Nostradamus.