A Random Thought On The IHME Projections

[ProfessorGAC]

Ignoring completely the silliness in Georgia, I took a look at Worldometer.
The graph for cumulative deaths in the US is there. The y-axis is very broad, so it's in 10,000s.
Eyeball estimation shows 9,000 on April 1. 24 days later it's 53,000 and counting. 1,750 per day.
Now, we know that some of those early April days were under a thousand, now were 3 thousand, plus/minus for about a week.
So, as a thought experiment, optimistically, let's say this peak period only lasts 5 more days.
15,000 + 53,000 is 68,000. That's already higher than the IHME projection.
I think those folks need to rethink their methodology.

[BGBD]

1. The model

makes bad assumptions.

First it assumes that everyone is going to continue social distancing when we know some states aren't going to. It also assumes no deaths after June 1, which is insane to me.

Of course, the upper range of the IHME projection is 120k deaths.

[Igel]

4. It's hard to build in behavior that isn't predictable.

It's a model. It models what it does.

NE and ND are built in. Random events aren't.

And until we know what Georgia's laxing of the regs means, it's a question as to what the assumptions should be. All the assumptions for how social distancing and/or lockdowns would alter R0 and change the curve were just so much chum.

[BGBD]

11. Sure

but that is why you should build in more uncertainty. The model shouldn't show 0 deaths happening after June, it should show more uncertainty and not try to model anything after May. The range should widen.

[uponit7771]

5. +1, bad is charitable ... the assumptions are asinine at best.

[ProfessorGAC]

9. Brings Up Another Question

Wouldn't it be wiser & provide more accurate decision making information if they reported the middle of the range?
90+ thousand might get more serious attention than 65,000.
I'm suggesting that because observationally, I noticed people out & about, and open businesses got a little more diligent about every 10,000 deaths. Not standing around collecting data, but an interested observation.
When we were at 9,000 on April 1, I'd say one in 8 had their face covered & 1 in 20 wearing gloves.
About a week later we hit 20k and I'd estimate 25% had face cover and more gloves.
Over the last couple more weeks I have seeing more & more of both.
Went to grocery store today & all staff, at least 60% of customers were masked. All staff had gloves (well not one cashier) and probably 1 in 3 customers did.
If they want people to continue taking this seriously I'm not getting why they're reporting the low end.

[BGBD]

15. I think reporting out a specific number

is a bad idea. This would be better communicated as a range.

"Model says 60K people will die."

vs

"Model says between 50k and 120k people could die"

Most people don't have a great grasp on things like confidence intervals and probability. They hear a number and some math behind it and think that's what it is.

[SoonerPride]

2. The IMHE model is an unmitigated disaster

It isn’t used by epidemiologists and now you see why.

We will easily be at 100,000 or more by July

Probably 250,000 by Election Day

“But it’s just like the flu!”

The fuck it is. This is goddamned apocalyptic.

[Igel]

8. They've said why.

The IMHE actually hasn't been all that bad for the most part. It overshot reality, but they all did.

They don't think of it as a model. It doesn't start from first principles and then project. Theoreticians hate it just like theoretical physicists hate the Standard Model. Nothing predicts it. It has the parameters it has because that's what parameters were measured to be or need to be because of measurements.

It started from a real example--Wuhan--assumes that is to be the correct overall shape, uses assumptions to "personalize" that shape per region. It updates the "model" from time to time based on new actual results, and then has to figure out how where on that curve we are and what changes need to be made to the assumptions to account for the past data. Then, given that, project. It's a series of kluges. The theory is in the design. There is nothing simple about it, and screwing with the data like NY and the CDC did certainly didn't help its simplicity.

As for accurate, of course it's wrong. It's a model, right? But I've seen people say how wonderful their model was, accurate to the integer, although it was a fairly simple exponential function. The real curve may have all kinds of ups and downs, but apparently a really simple ae^bx curve does that. Who knew?

Or perhaps they were just being a bit overenthusiastic in their boasts of superiority because, in fact, their models were also wrong. And less useful because it's harder, with a simple model, to actually account for the data in any one state, much less the US. Remember, the US' projection is really the sum from i = 1 to a fairly large n of a lot of smaller projections, with the NYC metroplex being the godzilla in the bunch. The true horror that is the NYC metroplex is masked by having NYC diluted by the rest of the state and having the Jersey and CT parts sliced off by state lines. It, all on its own, accounts for half of all US deaths.

[qazplm135]

3. I'm going to guess 65K by 1 May

I will further guess, probably 90-95 by 1 June, assuming we drop down to around 1K a day averaged (more beginning of May, less towards the end of the month).

By 1 July I would guess probably 115K give or take 5K. With four months to go until election day, we will be hard-pressed to stay under 150K by the time people are voting. Maybe we get there, but it's definitely got a strong shot of being over 130K.

[uponit7771]

6. +1, I calculated very similar but that's if the country stays at the current isolation measures ...

... which it wont because the MAGA cluckers and suckers are going to sniff each others faces

[qazplm135]

13. I'm going to assume

enough gov are cautious enough to keep the daily death toll somewhere below 1500 a day once we hit June.

But yeah, that will definitely add up more than it should be.

[ProfessorGAC]

7. Aren't We Losing Almost 3,000 A Day?

65k on the 1st is only 2k per day.
You really think it will drop to an average of 2k? I'm surprised
Seems conservative.

[Igel]

12. Not even.

In the last week the range has been (Worldometer #s) from 1570 to 2683.

The numbers tend to cycle: Numbers drop Sunday and Monday, Tuesday things pick up, and Wednesday they catch up with all the death reports that weren't filed before. Part of that is when state agencies are open, part of it is Worldometer's restarting their days at 0:00 UTC when some state agencies haven't yet reported, so the numbers for a day slosh over to day + 1.

[qazplm135]

14. nope, our highest peak was mid-2000s and

the last couple of days have seen sub 2000 with today also on track for sub 2000.

I'm assuming around 2K the rest of this month.

I'm assuming then the first half of May drops to around 1500 or less, and the second half drops to around 1K or less.

After that I am assuming another drop to around 500 give or take for July and then August through October see it in the 100s each day.

That should get us to what I'd guess would be in the low 100s by election day.

This ignores a resurgence brought on by cooler weather or significantly lessened restrictions and people getting complacent (both certainly very possible).

I doubt it reaches 200K. But I wouldn't call it impossible either. I also doubt it's under 100K come election day. 40K over the next six months (starting 1 May and assuming we hit 60K+ on that day, which we will) is only 220 a day the rest of the way, and it's going to be July or August before we even reach that daily amount.

[BGBD]

16. but

the curve we see isn't a symmetrical hump. It's a quick incline and a much, much slower decline. The "peak" is going to be the days with the higher number, but isn't not the median of the epidemic. There are going to be more deaths after the peak than before it.

[qazplm135]

17. right which is why

I think we get to the low 100s by November. I didn't say anything else.

[Victor_c3]

10. I wouldn't be surprised to see 100k deaths by June first

We’d need to average less than 1,500 deaths a day to reach that milestone. Sadly, I don’t think that’ll be a problem. I don’t see any evidence that we’ve reached a plateau yet. If anything, I’ll gander that we’ll probably see more of an increase in the number of new deaths for the foreseeable future, especially with some states getting ready to loosen social distancing protocols.

At 1,500 deaths per day we’d average 45,000 new death in a month. That last day we had fewer than 1,500 new deaths reported was 6 April. We’re only seeing the beginning of things to come. It’s going to get worse before it gets better.