The predictive value of COVID positivity

[Ausmumof3]

12 minutes ago, Not_a_Number said:

No, population in millions. So for SD that'd be 0.9 or so, not what you did. And you leave the positivity number as how many percent it was. So then you get 

0.9/3*50 =15, 

which seems pretty plausible, I gotta say. 

Is that deaths per day?

[Not_a_Number]

10 minutes ago, Happymomof1 said:

Only 15 dead people?  That makes no sense.  I'm not understanding the math then at all.

Per day. 15 dead per day on average, in 2 weeks. 

[Not_a_Number]

1 minute ago, Ausmumof3 said:

Is that deaths per day?

Correct. The 7-day average thereof, since that's a little more stable. 

[Not_a_Number]

8 minutes ago, Ausmumof3 said:

And here is the cases.  Deaths lag is 4-6 weeks of more so this could be quite scary in weeks to come.  Obviously the earlier spike is probably underrepresented due to lack of testing.

Not generically. I have not seen countries for which this is true, whatever people say. 

[Happy2BaMom]

15 dead people a day is actually a lot for a less-populated state like SD.

[Not_a_Number]

2 minutes ago, Happy2BaMom said:

15 dead people a day is actually a lot for a less-populated state like SD.

Indeed. And it's at 14 per day now, so it's not a bad prediction. 

[Ausmumof3]

3 minutes ago, Not_a_Number said:

Not generically. I have not seen countries for which this is true, whatever people say. 

Hmm I wonder how that works?  US previous case peak was roughly July 24 and deaths about August 4 so 10 days to 2 weeks.  Maybe more people start seeking testing when people get hospitalised?  I’m just curious I mean we know it takes time to die so if cases and deaths peaks at the same time there must be some human behaviour factor driving that.

[Not_a_Number]

1 minute ago, Ausmumof3 said:

Hmm I wonder how that works?  US previous case peak was roughly July 24 and deaths about August 4 so 10 days to 2 weeks.  Maybe more people start seeking testing when people get hospitalised?  I’m just curious I mean we know it takes time to die so if cases and deaths peaks at the same time there must be some human behaviour factor driving that.

From what we've seen on this thread, 2 weeks is about right. I think @Danae ran correlations between positivity and deaths and that was the best she could do. I think it takes a while for people to get tested, and then for tests to be reported. So the case to death lag for almost every country I've monitored is 2-3 weeks. 

Edited November 10, 2020 by Not_a_Number

[Ausmumof3]

1 minute ago, Not_a_Number said:

From what we've seen on this thread, 2 weeks is about right. I think @Danae ran correlations between positivity and deaths and that was the best she could do. I think it takes a while for people to get tested, and then for tests to be reported. So the case to death lag for almost every country I've monitored is 2-3 weeks. 

Interestingly with Australia there is a month between case peak and deaths peak.  So maybe when there’s adequate testing and a solid push for people to seek testing and a population that seeks testing the data is more in line with that.

[Not_a_Number]

3 minutes ago, Ausmumof3 said:

Interestingly with Australia there is a month between case peak and deaths peak.  So maybe when there’s adequate testing and a solid push for people to seek testing and a population that seeks testing the data is more in line with that.

Not seeing that in the Worldometer graph. Seeing 2-3 weeks: 

https://www.worldometers.info/coronavirus/country/australia/

ETA: ignore the giant peaks, though, measure to the middle. Giant peaks are data someone is dumping late. 

Edited November 10, 2020 by Not_a_Number

[Ausmumof3]

2 minutes ago, Not_a_Number said:

Not seeing that in the Worldometer graph. Seeing 2-3 weeks: 

https://www.worldometers.info/coronavirus/country/australia/

ETA: ignore the giant peaks, though, measure to the middle. Giant peaks are data someone is dumping late. 

Using 7 day average I’m seeing August 5 and September 4.  We did get daily reports on deaths and cases aside from one data dump from aged care one day so I think the data is reasonably accurate.

[Not_a_Number]

2 minutes ago, Ausmumof3 said:

Using 7 day average I’m seeing August 5 and September 4.  We did get daily reports on deaths and cases aside from one data dump from aged care one day so I think the data is reasonably accurate.

That's what I'm saying: focus on the actual shape of the graph in gray, ignore the single day data dump. That data dump is NOT from that day.  

Where's the middle? 

Edited November 10, 2020 by Not_a_Number

[Ausmumof3]

17 minutes ago, Not_a_Number said:

That's what I'm saying: focus on the actual shape of the graph in gray, ignore the single day data dump. That data dump is NOT from that day.  

Where's the middle? 

Hmm August 28 so 3 weeks?

[Not_a_Number]

Just now, Ausmumof3 said:

Hmm August 28 so 3 weeks?

Yeah, or I might actually say August 23 -- I'd guess the low numbers there are an artifact. But either way, you can see it's not 4-6 weeks. 

[Ausmumof3]

1 minute ago, Not_a_Number said:

Yeah, or I might actually say August 23 -- I'd guess the low numbers there are an artifact. But either way, you can see it's not 4-6 weeks. 

But I’m curious to know why?  We know that positive test to illness to death takes time.  Is it an artefact of lack of testing or something else?

[Not_a_Number]

Just now, Ausmumof3 said:

But I’m curious to know why?  We know that positive test to illness to death takes time.  Is it an artefact of lack of testing or something else?

I actually don't know how long death takes on average, do you? Recovery takes a while, but death might on average not take that long. (Or, on median. You don't really want to count outliers here.) 

[Ausmumof3]

14 minutes ago, Not_a_Number said:

I actually don't know how long death takes on average, do you? Recovery takes a while, but death might on average not take that long. (Or, on median. You don't really want to count outliers here.) 

21 days is what I’m seeing most but most of the data is from March April.  That may have changed  with better treatment protocols.

[Not_a_Number]

Just now, Ausmumof3 said:

21 days is what I’m seeing most but most of the data is from March April.  That may have changed  with better treatment protocols.

Matches Australia's data well, though 😉 . 

Most people don't die, so I don't know how much treatment is helping the people who will. The death rate is largely in very old and frail people. Many more people get very sick than die 😕 

[Ausmumof3]

For what it’s worth the death lag of three weeks quoted is from here 

so I have no idea where he’s getting that from but I didn’t make it up 🙂 just didn’t check it out.

 

[Not_a_Number]

11 minutes ago, Ausmumof3 said:

For what it’s worth the death lag of three weeks quoted is from here 

so I have no idea where he’s getting that from but I didn’t make it up 🙂 just didn’t check it out.

 

Yeah, he's kind of a publicity hound. Not a fan. 

ETA: he's probably right about 3-4 weeks, though, lol. I shouldn't let the antipathy prejudice me!! 

ETAA: actually, if he's saying that's what it looks like on the graphs, he's wrong. 

Edited November 10, 2020 by Not_a_Number

[Ausmumof3]

29 minutes ago, Not_a_Number said:

Yeah, he's kind of a publicity hound. Not a fan. 

ETA: he's probably right about 3-4 weeks, though, lol. I shouldn't let the antipathy prejudice me!! 

ETAA: actually, if he's saying that's what it looks like on the graphs, he's wrong. 

Lol yeah well he’s not the only one just the easiest one to find.  Have seen it from multiple sources.  But you’re right from looking at the peaks it’s more like three weeks with good testing and two weeks with less.  I’m also guessing the lead time is probably less accurate during the peaks because lack of testing means people get tested later in the infection.  

[Not_a_Number]

1 minute ago, Ausmumof3 said:

Lol yeah well he’s not the only one just the easiest one to find.  Have seen it from multiple sources.  But you’re right from looking at the peaks it’s more like three weeks with good testing and two weeks with less.  I’m also guessing the lead time is probably less accurate during the peaks because lack of testing means people get tested later in the infection.  

Yeah, that was my impression: lag was shorter for peaks than troughs.

Edited November 10, 2020 by Not_a_Number

[Halftime Hope]

On 10/15/2020 at 3:51 PM, Not_a_Number said:

How can you ever know? I've followed it long enough that it's a fairly robust pattern. If I did run statistical tests on it, it would be far too strong a correlation to be random. But I'm not going to bother, because I'm not publishing a paper -- I'm just sharing what I've observed. 

As for the data, it's not awesome, I agree. But then that's how it goes in a pandemic -- the data isn't awesome. This thread is just for people who are interested in what is the most predictive, and so far, it seems to be positivity and not other things. I've personally found that way of thinking about it helpful for my own decisions. 

But I don't understand how your positivity rate for any given date could possibly correlate with deaths for 2 weeks later.  If the majority of tests are done when people begin symptoms, the ones that die are not going to die two weeks later.  I believe that Dr Campbell said that for most people, deaths happened between 4-6 weeks after they tested positive?  

I must be missing something about how you are thinking about this.   

[Not_a_Number]

2 hours ago, Halftime Hope said:

But I don't understand how your positivity rate for any given date could possibly correlate with deaths for 2 weeks later.  If the majority of tests are done when people begin symptoms, the ones that die are not going to die two weeks later.  I believe that Dr Campbell said that for most people, deaths happened between 4-6 weeks after they tested positive?  

I must be missing something about how you are thinking about this.   

You are assuming that cases are reported immediately, which I doubt -- I'm sure there's a backlog. We also couldn't find reliable data about how long it takes people to die. 

Either way, the way I'm thinking about this is that I looked at a lot of countries, and every single country had a lag of 2-3 weeks from cases to deaths. For the US, that lag is not from case numbers but from positivity. You can see the graphs yourself -- they are somewhere on this thread. If you need a link, I'll link them up. 

[Laura Corin]

On 11/10/2020 at 12:06 AM, Happy2BaMom said:

I use John Hopkins. Here is what they show for SD - well over 50%.

And here is a summary of positivity rates by state, based on JH data.

Not looking to quibble over it, just backing up my stats above.

Wow.  Those positivity rates are high. Scotland is at 9 percent and we are only allowed out for essential purposes and aren't allowed to enter other homes.

[Not_a_Number]

3 hours ago, Laura Corin said:

Wow.  Those positivity rates are high. Scotland is at 9 percent and we are only allowed out for essential purposes and aren't allowed to enter other homes.

Yeah, some parts of the US aren’t taking this real seriously...

[Not_a_Number]

Positivity is above 9%. That’s double the bottom of the last trough. Expect to hit 2000 deaths per day on average in December, although I would guess not November. Current prediction for end of November is about 1500 deaths/day on average.

Edited November 13, 2020 by Not_a_Number

[Not_a_Number]

Positivity keeps going up. This is going to be bad, I think. 

Here's a map with the global positivity rates: 

https://ourworldindata.org/coronavirus-testing#the-positive-rate-a-crucial-metric-for-understanding-the-pandemic

Unsurprisingly, much of Europe is in deep trouble. 

[Not_a_Number]

@Danae, are you still keeping track? Has the scaling constant gone up? The one for the US overall definitely has, which makes a lot of sense.

@Corraleno, these correlations are why I tend to think we aren’t on track for 2000 deaths/day on average all that soon, although I DO assume we’ll hit that within December unless something changes.

[Guest]

My area seems to have stabilized at 10% BEFORE people start gathering for holidays and stores start extended hours for holiday shopping. I'm very worried about what the next two months will hold. 

[Not_a_Number]

3 minutes ago, dmmetler said:

My area seems to have stabilized at 10% BEFORE people start gathering for holidays and stores start extended hours for holiday shopping. I'm very worried about what the next two months will hold. 

Yeah, I’m seeing some hints of US-wide stabilization... but you’re right that the holidays are next week 😮. I didn’t even think of that.

[Corraleno]

On 11/9/2020 at 4:14 PM, Not_a_Number said:

Per day. 15 dead per day on average, in 2 weeks. 

10 days later (yesterday, the 19th) it's already at 20. ☹️

 

[Not_a_Number]

3 minutes ago, Corraleno said:

10 days later (yesterday, the 19th) it's already at 20. ☹️

 

Where are we talking about? I've lost track. 

[Not_a_Number]

Here's @Danae's rough estimate — take the population of the region in millions and divide by three. Multiply that by the positivity rate today to get the death count two weeks from today.

Trying this for NY: 2 weeks ago, the positivity rate was something like 1.7%. This prediction gives 

1.7*20/3 = 1.7*6 = 10.2 

Definitely WAY too low. But then one does rather expect that for the states which are doing plenty of asymptomatic and precautionary testing. 

[Corraleno]

Just now, Not_a_Number said:

Where are we talking about? I've lost track. 

That was the prediction for South Dakota.

According to the chart that was posted upthread, US-wide positivity was only 7.6% 2 weeks ago (11/6). That gives a prediction for today of around 840. The average yesterday was already up to 1398, so that's way off.

[Not_a_Number]

19 minutes ago, Corraleno said:

That was the prediction for South Dakota.

According to the chart that was posted upthread, US-wide positivity was only 7.6% 2 weeks ago (11/6). That gives a prediction for today of around 840. The average yesterday was already up to 1398, so that's way off.

Yeah, that's not surprising -- I've been noticing the scaling factor ticking up. It's because we're testing many more asymptomatic people is my guess... the college and workplace testing. 

But that shifts pretty slowly, so the best way to predict what happens in 2 weeks is usually to just take the CURRENT scaling factor and use it. According to that, we'll be at about 

10/7.6*1398 = 1839

deaths in 2 weeks on average. So, not what I call good news, lol... but also less than you predict if you just use case numbers. 

We'll see if I'm right, of course. So far, it's tracked well, though. 

[Not_a_Number]

Of course, South Dakota is totally not testing asymptomatic people, lol! Their numbers are a total disaster. 

I don't know what to do with their data, frankly. It's very spiky. Frankly, given the quality of their data collection, being anywhere near in the ball park is good enough for me. 

Edited November 20, 2020 by Not_a_Number

[Terabith]

My city's positivity rate is 11.6%, a week before Thanksgiving.  

Sigh.

[Danae]

7 hours ago, Not_a_Number said:

@Danae, are you still keeping track? Has the scaling constant gone up? The one for the US overall definitely has, which makes a lot of sense.

@Corraleno, these correlations are why I tend to think we aren’t on track for 2000 deaths/day on average all that soon, although I DO assume we’ll hit that within December unless something changes.

It's gone way up. 

[Not_a_Number]

22 minutes ago, Danae said:

It's gone way up. 

How much? Theories why?

[Danae]

4 minutes ago, Not_a_Number said:

How much? Theories why?

I said a few weeks ago that the multiplier seems to go up when positivity is above 5.  Which makes sense with what epidemiologists have said that at 5% you know you're doing enough testing.  Above five means you aren't testing enough, which means a higher % of cases are being missed.

 

[Not_a_Number]

3 minutes ago, Danae said:

I said a few weeks ago that the multiplier seems to go up when positivity is above 5.  Which makes sense with what epidemiologists have said that at 5% you know you're doing enough testing.  Above five means you aren't testing enough, which means a higher % of cases are being missed.

 

Except that if we think of it as sampling, sampling MORE sick people ought to get the scaling factor down, not up... (Sampling asymptomatic people drives down positivity and all.)

I was keeping track of the factor calculated every week before, but sadly it got lost in the purge. It’s possible that the lag changes when there’s an outbreak, too? So we need to look fewer than 2 weeks back.

[Corraleno]

I think that when the numbers are this huge and cases cover the entire US, cases x CFR is not a bad predictor. And obviously the CFR changes in line with positivity, so case-based predictions include positivity in that sense. 

Predictions for the first two weeks of November based on cases three weeks earlier x a CFR of 1.7% were absolutely spot on, and for this week the prediction was low — 1380, when we hit 1398 yesterday and will be well into the 1400s by the end of today. So that suggests that the CFR is creeping up, closer to 1.8, which implies we're going to be well over 2000 in 2 weeks (12/4). I think the numbers are going to be all over the place next week, with what's essentially a 4-day holiday weekend , but I think the numbers the week after are going to be really grim. 😕

[Not_a_Number]

11 minutes ago, Corraleno said:

I think that when the numbers are this huge and cases cover the entire US, cases x CFR is not a bad predictor. And obviously the CFR changes in line with positivity, so case-based predictions include positivity in that sense. 

Predictions for the first two weeks of November based on cases three weeks earlier x a CFR of 1.7% were absolutely spot on, and for this week the prediction was low — 1380, when we hit 1398 yesterday and will be well into the 1400s by the end of today. So that suggests that the CFR is creeping up, closer to 1.8, which implies we're going to be well over 2000 in 2 weeks (12/4). I think the numbers are going to be all over the place next week, with what's essentially a 4-day holiday weekend , but I think the numbers the week after are going to be really grim. 😕

Do you want to see what my predictions would have been? I can test what I’d have done and see how well it’s performing. The problem is that the CFR is highly dependent on the fraction of caught cases.

[Roadrunner]

I wonder if the spike has anything to do with Halloween. We are in CA and my county has been in the purple (most restrictive) tier the entire time. We haven’t loosened any guidelines here locally and yet despite doing what we have been doing all along, we are still experiencing a spike. It’s very frustrating.

[Corraleno]

8 minutes ago, Not_a_Number said:

Do you want to see what my predictions would have been? I can test what I’d have done and see how well it’s performing. The problem is that the CFR is highly dependent on the fraction of caught cases.

Yes, that would be useful.

CFR of 1.7% x cases 3 weeks ago = 1380
CFR of 1.75% x cases 3 weeks ago = 1420 
CFR of 1.8% x cases 3 weeks ago = 1459

Edited November 20, 2020 by Corraleno

[Danae]

I've been playing around with numbers, and it looks like maybe what changes as the positivity goes higher is both the scaling factor and the lag.  Which is, unfortunately, much harder to model in a spreadsheet.  

 

Edit:  but is consistent with the idea that higher positivity means things are out of control and testing isn't keeping up.  

Edited November 20, 2020 by Danae

[Ausmumof3]

9 hours ago, Not_a_Number said:

@Danae, are you still keeping track? Has the scaling constant gone up? The one for the US overall definitely has, which makes a lot of sense.

@Corraleno, these correlations are why I tend to think we aren’t on track for 2000 deaths/day on average all that soon, although I DO assume we’ll hit that within December unless something changes.

Was 1800 per day two days ago and over 2000 yesterday.  Rolling average is at high 1300s which I assume is what you’re talking about.  Coming into the weekend you should have a couple of lower days potentially as several states don’t report deaths over the weekend.

[Corraleno]

2 hours ago, Not_a_Number said:

10/7.6*1398 = 1839

deaths in 2 weeks on average. So, not what I call good news, lol... but also less than you predict if you just use case numbers. 

The reason I'm skeptical that average deaths will only be 1839 in two weeks is that deaths have increased by more than 300 in just one week, while this prediction means that deaths will only increase by a little over 400 in the next two weeks. That implies that the death rate will be slowing down while cases are rapidly increasing, which would only be true if positivity were going way down, when it actually seems to be going up, which is why the CFR is creeping up.  

I don't think there is currently a lot of testing of asymptomatic people, I think that was more common when colleges first opened, but a lot of them are sending kids home for Thanksgiving and going virtual thereafter, so probably a lot less testing of students from this point forward. And a lot of people were saying in September that although the case rate then reflected lots of asymptomatic young people, those people were going to be spreading it to their families within a couple of months, and I think we're already seeing that — and about to see even more of it next week.

[Corraleno]

7 minutes ago, Ausmumof3 said:

Was 1800 per day two days ago and over 2000 yesterday.  Rolling average is at high 1300s which I assume is what you’re talking about.  Coming into the weekend you should have a couple of lower days potentially as several states don’t report deaths over the weekend.

The 7 day average includes weekends, though. Each day the new deaths for that day replace the deaths for that day in the previous week, so what matters for the average is the difference between, say, this Saturday and last Saturday, not between Saturday and Friday. So the average tends to increase a little over the weekend, and then catch up rapidly during the week. That's why I generally use data from Fridays.