The predictive value of COVID positivity

[Not_a_Number]

In my previous incarnation, I had a thread about how well COVID positivity correlates with death rates 2 weeks later .

That still appears to be true. Anyone been following it? 

[Bambam]

Could you explain again? 
Are you saying that the COVID positivity rate (say 7.0%) correlates with death rates 2 weeks later? So, if there were 100 active cases, you would say 7 would be the death total on the day exactly 2 weeks from today? Surely that can't be right - at least at a local level. The death rate still seems to be going down as well.  

[Not_a_Number]

Just now, Bambam said:

Could you explain again? 
Are you saying that the COVID positivity rate (say 7.0%) correlates with death rates 2 weeks later? So, if there were 100 active cases, you would say 7 would be the death total on the day exactly 2 weeks from today? Surely that can't be right - at least at a local level. The death rate still seems to be going down as well.  

I mean, it's not perfect like the example you said 🙂 . Past of the reason I follow this largely at the level of the whole US is that the data is a lot smoother at that level. 

But yes, the ratio between positivity today (in the US) and the death number 2 weeks letter (in the US) is not quite constant but it doesn't wiggle all that much. Right now, we're coming up from a positivity trough, which was around 4.5%. As the positivity goes up, I expect the deaths to go up with a 2-3 week lag. 

[Bambam]

So, give me solidish numbers (which help me understand) -
So, if the US positivity rate is 5.1% (7 day moving average from Johns Hopkins). 
So, what is the second half of the ratio? 

[Not_a_Number]

2 minutes ago, Bambam said:

So, give me solidish numbers (which help me understand) -
So, if the US positivity rate is 5.1% (7 day moving average from Johns Hopkins). 
So, what is the second half of the ratio? 

Well, right now the deaths per day are at a 7-day average of 717, and two weeks ago, the positivity rate was about 4.6%. So then the ratio is about 

717/4.6 = 155. 

It was closer to 140 last time I was checking it, which was a few months ago. But that's about right. It was mostly wavering between 100-150. 

[Bambam]

Still not 100% clear: 
Positivity rate for US was 4.6% two weeks ago.
Deaths for US are at a 7 day average of 717 currently. 

So you have {7 day average deaths (US)}/{Positivity rate 2 weeks ago (US)} = 155? What is the 155 number represent?

Not trying to be a dunce, just trying to figure this out!  

Edited October 15, 2020 by Bambam

[Not_a_Number]

Just now, Bambam said:

Still not 100% clear: 
Positivity rate for US was 4.6% two weeks ago.
Deaths for US are at a 7 day average of 717. 

So you have {7 day average deaths (US)}/{Positivity rate 2 weeks ago (US)} = 155? What is the 155 number represent?

Not trying to be a dunce, just trying to figure this out!  

It's just a number that shows the ratio, and the ratio isn't exactly FIXED, but it's close. So, assuming you subscribe to this theory, the positivity starting to creep up is bad news for deaths in 2 weeks. 

[Bambam]

I shall try to watch to see. Have you made any graphs illustrating this? 

[Not_a_Number]

14 minutes ago, Bambam said:

I shall try to watch to see. Have you made any graphs illustrating this? 

No, but if you look at the positivity graph on the JHU site and the death graph on Worldometer (or elsewhere, I'm sure -- they all look the same), you can see the shape similarity: 

https://coronavirus.jhu.edu/testing/individual-states/usa

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

See it? 

[Danae]

I’m still tracking MN numbers, and it’s still predictive.  I have graphs, I’ll see if I can figure out how to post them in the morning.

[SKL]

It depends on which populations are testing positive.  If it's the college students we keep hearing about, their COVID death rate is approximately zero.  Same for minors in face-to-face school.  So no, I don't think it's predictive, but I guess we'll see.

[Not_a_Number]

7 hours ago, Danae said:

I’m still tracking MN numbers, and it’s still predictive.  I have graphs, I’ll see if I can figure out how to post them in the morning.

Thank you!! Sorry the old threads died. NY is thankfully ONLY using positivity to identify clusters, and they’ve done sharp shutdowns in high positivity areas. 

I’m continuing to be pleased with leadership of NY state. I hope it’s enough.

[Not_a_Number]

3 hours ago, SKL said:

It depends on which populations are testing positive.  If it's the college students we keep hearing about, their COVID death rate is approximately zero.  Same for minors in face-to-face school.  So no, I don't think it's predictive, but I guess we'll see.

What I’m telling you is that in the US, observationally, it’s been incredibly predictive. Much more predictive than total case numbers or even hospitalization numbers. You can argue that we could organize society so it’s no longer true, but the observation currently stands.

[Not_a_Number]

6 hours ago, kand said:

I think you explained it better in this thread for me than in the past one where I was still not 100% sure I was tracking you. So would that mean if in two weeks the positivity rate is 6%, you would expect two weeks after that the daily deaths would be somewhere around 930? Am I tracking correctly. 

Something like that. It’s not that perfect, but you do expect a relatively predictable increase when positivity goes up. 

[SKL]

3 minutes ago, Not_a_Number said:

What I’m telling you is that in the US, observationally, it’s been incredibly predictive. Much more predictive than total case numbers or even hospitalization numbers. You can argue that we could organize society so it’s no longer true, but the observation currently stands.

The results of school spread are pretty new though.

Then again, it also depends on who is getting tested.  If, over time, generally the only people getting tested are the people showing significant symptoms (and their close cohorts), meaning true positivity fluctuations are probably not being reflected, then the death rate will probably continue to correlate more or less with the reported positivity rate.

But to answer your OP question, no, I'm not following it.  To the extent I follow the numbers, I look at deaths, because the "case" numbers are clearly inaccurate and always have been.

[Not_a_Number]

1 minute ago, SKL said:

The results of school spread are pretty new though.

True. That'll be interesting to see. 

 

1 minute ago, SKL said:

Then again, it also depends on who is getting tested.  If, over time, generally the only people getting tested are the people showing significant symptoms (and their close cohorts), meaning true positivity fluctuations are probably not being reflected, then the death rate will probably continue to correlate more or less with the reported positivity rate.

True positivity fluctuations ARE being reflected if you test people with symptoms. The bigger fraction of people with symptoms are sick, the more the actual positivity rate is.  

I think the fraction of people who get sick who have symptoms is relatively stable, last I checked. Many people do have mild cases, but not so many stay entirely asymptomatic. 

 

1 minute ago, SKL said:

But to answer your OP question, no, I'm not following it.  To the extent I follow the numbers, I look at deaths, because the "case" numbers are clearly inaccurate and always have been.

Right. I agree. So I suggest that you follow the positivity, since so far, it's about 2 weeks ahead of the deaths. Of course, it could stop being predictive, but right now, it's a good guide to where we're going. 

[SKL]

14 minutes ago, Not_a_Number said:

 So I suggest that you follow the positivity, since so far, it's about 2 weeks ahead of the deaths. Of course, it could stop being predictive, but right now, it's a good guide to where we're going. 

No thanks.  The last thing I need is a tool to help me think about how many more people are going to die.

I periodically check the trends around my personal bubble, and they really don't seem to follow any of the reported predictors.  So we may have a situation where the overall numbers appear to follow a trend, but when broken down into smaller groups, we find a number of sub-trends that don't look at all like the overall trend.

For one thing, around my bubble, cases have been going up a lot while deaths have been going down.  Again, I think it's because of young people congregating after being apart all spring/summer.  There are dozens of universities near me.

[Not_a_Number]

1 minute ago, SKL said:

For one thing, around my bubble, cases have been going up a lot while deaths have been going down.  Again, I think it's because of young people congregating after being apart all spring/summer.  There are dozens of universities near me.

Yeah, it'd be interesting to see whether that stays contained. Universities are a different proposition than going home. 

I've been using the positivity to make local decisions to some extent. But partially, I'm just interested. 

[wintermom]

16 hours ago, Not_a_Number said:

I mean, it's not perfect like the example you said 🙂 . Past of the reason I follow this largely at the level of the whole US is that the data is a lot smoother at that level. 

But yes, the ratio between positivity today (in the US) and the death number 2 weeks letter (in the US) is not quite constant but it doesn't wiggle all that much. Right now, we're coming up from a positivity trough, which was around 4.5%. As the positivity goes up, I expect the deaths to go up with a 2-3 week lag. 

I'm curious about whether you (and others) are basing the predictions on a normal distribution. If so, what are you basing this on, other than the fact that it's the easiest. I'm not familiar with public health statistics, so there could be tried-and-true methods. I'm interested in hearing about them! Thanks!

[Not_a_Number]

4 minutes ago, wintermom said:

I'm curious about whether you (and others) are basing the predictions on a normal distribution. If so, what are you basing this on, other than the fact that it's the easiest. I'm not familiar with public health statistics, so there could be tried-and-true methods. I'm interested in hearing about them! Thanks!

Nope. I’m basically basing this on “looking at the graphs.” I’m not doing any statistical testing here: merely making qualitative observations. I think @Danaehas run correlations, but that’s as sophisticated as we’ve gotten.

[wintermom]

16 minutes ago, Not_a_Number said:

Nope. I’m basically basing this on “looking at the graphs.” I’m not doing any statistical testing here: merely making qualitative observations. I think @Danaehas run correlations, but that’s as sophisticated as we’ve gotten.

Thanks. So the message to underline is that this is not predictive, this is observational based on reported measures. 

You may want to change the thread heading. Or not, but it is misleading.

Edited October 15, 2020 by wintermom

[Not_a_Number]

36 minutes ago, wintermom said:

Thanks. So the message to underline is that this is not predictive, this is observational based on reported measures. 

You may want to change the thread heading. Or not, but it is misleading.

Huh? No, it seems to PREDICT deaths 2 weeks later. Obviously, I’m not concluding this in a very mathematical way, but I actually don’t think extrapolating from existing data is a bad way to go. 

[wintermom]

6 hours ago, Not_a_Number said:

Huh? No, it seems to PREDICT deaths 2 weeks later. Obviously, I’m not concluding this in a very mathematical way, but I actually don’t think extrapolating from existing data is a bad way to go. 

Of course it's not bad to attempt to make predictions from existing data. That is was statistics is all about. The key with this thread is that you have limited information about the data you are using. You don't know how accurate it is, how it was collected, what data was included and what data was excluded, the exact timing of the collection, etc. 

You are looking at some numbers and making assumptions and guesses. Your guesses could all be wrong. Or they might be right. But how can you know whether it's one, the other, or something completely different?

Edited October 15, 2020 by wintermom

[Not_a_Number]

Just now, wintermom said:

Of course it's not bad to attempt to make predictions from existing data. That is was statistics is all about. The key with this thread is that you have limited information about the data you are using. You don't know how accurate it is, how it was collected, what data was included and what data was excluded, the exact timing of the collection, etc. 

You are looking at some numbers and making assumption and guess. Your guesses could all be wrong. Or they might be right. But how can you know whether it's one, the other, or something completely different?

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. 

[wintermom]

1 minute ago, 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. 

You know now because you are looking at what happened - or at least the death count that is published. How are these numbers counted? I have no idea how you are attempting to determine what is going to happen. You said yourself that you do not know know if the data is normally distributed or how it's distributed. How are you going to determine the probability? Based on what statistic(s)? 

[Not_a_Number]

10 minutes ago, wintermom said:

You know now because you are looking at what happened - or at least the death count that is published. How are these numbers counted? I have no idea how you are attempting to determine what is going to happen. You said yourself that you do not know know if the data is normally distributed or how it's distributed. How are you going to determine the probability? Based on what statistic(s)? 

Do we ever know exactly how data is distributed in these situations?? You make assumptions. If you were going to make a statistical model that was going to run the tests, yes, you'd assume it was normally distributed, because that's the standard assumption for a lot of these tests, last I checked. But you can run your tests in all sorts of ways. One of the COVID predictive models (a fancy, statistical one!) was testing its data by comparing its performance to the "null model" of predicting the same number of deaths this week as last week. The "null model" turned out to be better than a lot of models out there 😉 . 

The positivity model is unambiguously better than a null model. So that's a test for you, if you like. If it were random, you'd expect it to do better than the null model half the time. I could calculate the probability for you that my model is doing better than the null model and I promise you the probability that my results are random are WAY under the "statistically significant" bound of 5%... 

[hopeallgoeswell]

I have been looking at my county's numbers and have found that the positivity rate in long term care facilities is a better predictor of the death rate, which isn't surprising/alarming considering most all-cause deaths occur in long-term care facilities and most patients in places like that don't have a very prolonged longevity.

https://www.ucsf.edu/news/2010/08/98172/social-support-key-nursing-home-length-stay-death

“One quarter of all deaths in the United States occur in nursing homes, and that figure is expected to rise to 40 percent by the year 2020,” says Smith.

"The average age of participants when they moved to a nursing home was about 83. The average length of stay before death was 13.7 months, while the median was five months. Fifty-three percent of nursing home residents in the study died within six months."

[Not_a_Number]

2 minutes ago, hopeallgoeswell said:

I have been looking at my county's numbers and have found that the positivity rate in long term care facilities is a better predictor of the death rate, which isn't surprising/alarming considering most all-cause deaths occur in long-term care facilities and most patients in places like that don't have a very prolonged longevity.

Not super surprising, no 🙂 . Do you have links to all the different graphs in your county? 

[hopeallgoeswell]

57 minutes ago, Not_a_Number said:

Not super surprising, no 🙂 . Do you have links to all the different graphs in your county? 

I'm not going to give my exact location (for reasons you probably understand).  Our death count remains the same until an outbreak happens in a care facility, and then they tick up.  On the bright side, our % positive rate for people over 65 was cut in half after the first few months.

[Not_a_Number]

15 minutes ago, hopeallgoeswell said:

I'm not going to give my exact location (for reasons you probably understand).  Our death count remains the same until an outbreak happens in a care facility, and then they tick up.  On the bright side, our % positive rate for people over 65 was cut in half after the first few months.

And your positivity doesn't change when there's an outbreak in a care facility? 

[wintermom]

4 hours ago, Not_a_Number said:

Do we ever know exactly how data is distributed in these situations?? You make assumptions. If you were going to make a statistical model that was going to run the tests, yes, you'd assume it was normally distributed, because that's the standard assumption for a lot of these tests, last I checked. But you can run your tests in all sorts of ways. One of the COVID predictive models (a fancy, statistical one!) was testing its data by comparing its performance to the "null model" of predicting the same number of deaths this week as last week. The "null model" turned out to be better than a lot of models out there 😉 . 

The positivity model is unambiguously better than a null model. So that's a test for you, if you like. If it were random, you'd expect it to do better than the null model half the time. I could calculate the probability for you that my model is doing better than the null model and I promise you the probability that my results are random are WAY under the "statistically significant" bound of 5%... 

Sure, if you don't have better information it's always easiest to assume normal distribution. But the more the distribution is skewed or otherwise not normal,  the less sure the predictions are. 

Remind me what exactly is testing as positive? Is it a positive confirmation that a person died of covid, or something else? I heard a medical expert mention a few weeks ago that looking at deaths due to covid was a lot more accurate than looking at living people testing positive for covid. The problem is that not every death was tested for covid. There could well be under-reporting because of this, but also because of many other factors. 

[Not_a_Number]

Just now, wintermom said:

Sure, if you don't have better information it's always easiest to assume normal distribution. But the more the distribution is skewed or otherwise not normal,  the less sure the predictions are. 

Which is where it's not a bad idea to do a comparison with something. Like the "null model" idea I described above for testing. 

 

Just now, wintermom said:

Remind me what exactly is testing as positive? Is it a positive confirmation that a person died of covid, or something else? I heard a medical expert mention a few weeks ago that looking at deaths due to covid was a lot more accurate than looking at living people testing positive for covid. The problem is that not every death was tested for covid. There could well be under-reporting because of this, but also because of many other factors. 

Testing positive just means a positive PCR test, I think. Maybe they have other tests now, actually -- the rapid tests. But anyway, it's the percentage of tests that have come back positive with a current infection. It's mostly on living people... 

The best data for deaths is always excess deaths, yes. From what Iv'e seen most places have actually had decent accuracy with the deaths they are counting as COVID... not all the excess deaths, but within a reasonable factor. 

[wintermom]

10 minutes ago, Not_a_Number said:

Which is where it's not a bad idea to do a comparison with something. Like the "null model" idea I described above for testing. 

 

Testing positive just means a positive PCR test, I think. Maybe they have other tests now, actually -- the rapid tests. But anyway, it's the percentage of tests that have come back positive with a current infection. It's mostly on living people... 

The best data for deaths is always excess deaths, yes. From what Iv'e seen most places have actually had decent accuracy with the deaths they are counting as COVID... not all the excess deaths, but within a reasonable factor. 

Good luck with that. 

Have you factored in the 6+ hrs people wait at a testing centre, then they go home and may not decide to come back for testing? It's bonkers. How can you predict future deaths when the testing practices are  incredibly messy. You must have a lot of free time on your hands. 😉 

[Not_a_Number]

Just now, wintermom said:

Good luck with that. 

Have you factored in the 6+ hrs people wait at a testing centre, then they go home and may not decide to come back for testing? It's bonkers. How can you predict future deaths when the testing practices are  incredibly messy. You must have a lot of free time on your hands. 😉 

I haven't factored anything in. I'm just noticing an interesting pattern. I have a theory for why there's a correlation, but it's not something I'm going to bother testing -- just an idea. 

You should absolutely feel free to ignore this data if you think it's garbage. 🤷‍♀️

[Pen]

I’m still having trouble following this. 

Here are some screenshots for positivity rates and deaths for my state.  I wish I had taken and saved earlier ones too. 

To me it looks like positivity is going up (a bit of variation but basically up and up and up) 

but deaths aren’t following at all consistently 

How does it work?  The predictively? 

 

 

 

 

Edited October 16, 2020 by Pen
Removed screen shots

[Not_a_Number]

5 minutes ago, Pen said:

I’m still having trouble following this. 

Here are some screenshots for positivity rates and deaths for my state.

To me it looks like positivity is going up

but deaths aren’t following . 

How does it work? 

Honestly, with so few deaths, the data is incredibly noisy. I would look at the Oregon data and say that both positivity and deaths have been relatively stable, but again -- with so few deaths, the noise tends to be stronger than the signal. 

Edited October 16, 2020 by Not_a_Number

[hopeallgoeswell]

2 hours ago, Not_a_Number said:

And your positivity doesn't change when there's an outbreak in a care facility? 

No, positivity increase/decrease doesn't seem to correlate with when there is an outbreak in a care facility.  We had no care facility outbreaks from the start of reporting mid-April through the last week of May, with 928 cases and one death (not specified from where) from 4/14-5/26.  The first week of June we had our first two facility outbreaks, which added four deaths over that week and the next (not specified from where).  The second week of June through the first week of August we had no new facility outbreaks (case numbers more than doubled from 6/9 to 7/14 then started to decline).  Our death count went up by three that whole time (it's not specified if those were in a care facility or not) and our cases went up by 1702.  The last three weeks of August, as case numbers were steadily declining, 7 facilities reported outbreaks and our deaths went up by 7 (again not specified from where).  Oddly enough, our case numbers through September were at the lowest point (between 69 and 41 per week) since this all started, probably due to the wildfires keeping people at home, and we still had 2 outbreaks in the third week, and then none until 5 were reported this week.  The county only started to specify care facility case and death numbers the last week of August, and since then, all the deaths have been from there.

[Not_a_Number]

Just now, hopeallgoeswell said:

No, positivity increase/decrease doesn't seem to correlate with when there is an outbreak in a care facility.  We had no care facility outbreaks from the start of reporting mid-April through the last week of May, with 928 cases and one death (not specified from where) from 4/14-5/26.  The first week of June we had our first two facility outbreaks, which added four deaths over that week and the next (not specified from where).  The second week of June through the first week of August we had no new facility outbreaks (case numbers more than doubled from 6/9 to 7/14 then started to decline).  Our death count went up by three that whole time (it's not specified if those were in a care facility or not) and our cases went up by 1702.  The last three weeks of August, as case numbers were steadily declining, 7 facilities reported outbreaks and our deaths went up by 7 (again not specified from where).  Oddly enough, our case numbers through September were at the lowest point (between 69 and 41 per week) since this all started, probably due to the wildfires keeping people at home, and we still had 2 outbreaks in the third week, and then none until 5 were reported this week.  The county only started to specify care facility case and death numbers the last week of August, and since then, all the deaths have been from there.

But what does the positivity graph look like? I didn't say case numbers, I said percent positive. 

[Pen]

2 hours ago, Not_a_Number said:

Honestly, with so few deaths, the data is incredibly noisy. I would look at the Oregon data and say that both positivity and deaths have been relatively stable, but again -- with so few deaths, the noise tends to be stronger than the signal. 

 

So sounds like it only works if numbers get quite huge. ?

Maybe then there are enough people in positivity group to represent more vulnerable populations. 

Whereas our increased positives are tending to be amongst young people university age where few of any deaths are likely. 

 

[wintermom]

3 hours ago, Not_a_Number said:

I haven't factored anything in. I'm just noticing an interesting pattern. I have a theory for why there's a correlation, but it's not something I'm going to bother testing -- just an idea. 

You should absolutely feel free to ignore this data if you think it's garbage. 🤷‍♀️

Not garbage, but the error level in the data does seem overwhelmingly high at this point. 😉  Your power is probably pretty low, as is the likelihood of finding an effect if it exists. 

On a similar note, a stats geed acquaintance of mine was playing around with the graphs of covid deaths and mortality rates from the flu in Canada over similar months from Feb - July 2020. They were almost identical, apparently. Perhaps looking at flu rates could be a method of making predictions for future covid mortality trends. 

Edited October 16, 2020 by wintermom

[hopeallgoeswell]

2 hours ago, Not_a_Number said:

But what does the positivity graph look like? I didn't say case numbers, I said percent positive. 

Apologies.  I have been keeping an eye on case numbers.  I just looked specifically at positivity rates.  We had a 2% positivity each week until the middle of July, with two outbreaks the first week of June.  For the next six weeks it was 4.9%, 3.7%, 4.1%, 4.7%, 4.1%, and 4.2%.  During that time, 7 facilities reported outbreaks from weeks 4-6 of the high rates.  September was 2.8%, 2.5%, 2% (2 outbreaks that week), 2%, and 1.7%. October was 2% weekly with 5 outbreaks reported for this week.

Edited to add: Death counts are in my previous post through the end of August.  There were an additional 2 from the September outbreaks and 3 (so far) from the 5 outbreaks reported this week.

 

Edited October 16, 2020 by hopeallgoeswell

[Not_a_Number]

5 hours ago, hopeallgoeswell said:

Apologies.  I have been keeping an eye on case numbers.  I just looked specifically at positivity rates.  We had a 2% positivity each week until the middle of July, with two outbreaks the first week of June.  For the next six weeks it was 4.9%, 3.7%, 4.1%, 4.7%, 4.1%, and 4.2%.  During that time, 7 facilities reported outbreaks from weeks 4-6 of the high rates.  September was 2.8%, 2.5%, 2% (2 outbreaks that week), 2%, and 1.7%. October was 2% weekly with 5 outbreaks reported for this week.

Edited to add: Death counts are in my previous post through the end of August.  There were an additional 2 from the September outbreaks and 3 (so far) from the 5 outbreaks reported this week.

 

Oof, I have a hard time imagining the graph! These numbers are small enough that I’d again expect there to be very little signal in the noise. It works much better on a statewide level, and even better on the country level 🙂 .

[Not_a_Number]

5 hours ago, wintermom said:

Not garbage, but the error level in the data does seem overwhelmingly high at this point. 😉  Your power is probably pretty low, as is the likelihood of finding an effect if it exists. 

You can see the graphs yourself. It’s a strong correlation on a US-wide level. It’s also a fairly strong correlation at the state level for the bigger states. (For small states, it’s too noisy.)

As @Danaesaid in the previous version of this thread, lots of epidemiologists know this and this is the data they watch, because it’s ahead of the deaths, and case counts are very unreliable.

[Not_a_Number]

7 hours ago, Pen said:

 

So sounds like it only works if numbers get quite huge. ?

Maybe then there are enough people in positivity group to represent more vulnerable populations. 

Whereas our increased positives are tending to be amongst young people university age where few of any deaths are likely. 

 

I think most random things are like this! Law of Large Numbers and all that.

https://en.m.wikipedia.org/wiki/Law_of_large_numbers

So yeah, for small numbers of people, there’s a lot of heterogeneity about WHO is testing positive. But if you have enough people, it balances out.

Edited October 16, 2020 by Not_a_Number

[Pam in CT]

re data granular enough to have a sense of the impact of nursing home effects

14 hours ago, Not_a_Number said:

Not super surprising, no 🙂 . Do you have links to all the different graphs in your county? 

My mother lives in MA, which is astoundingly detailed and transparent with data, even more so than NY and CT. Here is their weekly data.  The community-level nursing home data is on pl 40. No graphs but the numbers are manageable enough that if you wanted to parse out an unusual weight for nursing homes it'd be a good sample.

Like CT and NY data, it also breaks down by age cohort. While it is unsurprisingly true that MOST of the deaths are 65+, it is absolutely not true that no one in younger cohorts are dying.  (And MA has been among the best-managed states in terms of medical sector resource response; it is not a function of overwhelmed hospital capacity or medical personnel at this point.)

 

 

FWIW (and to the surprise of many people outside New England, lol) MA has a GOP governor.  Mask policy, closure/phased opening/testing/data transparency are not intrinsically partisan.

[Not_a_Number]

5 minutes ago, Pam in CT said:

re data granular enough to have a sense of the impact of nursing home effects

My mother lives in MA, which is astoundingly detailed and transparent with data, even more so than NY and CT. Here is their weekly data.  The community-level nursing home data is on pl 40. No graphs but the numbers are manageable enough that if you wanted to parse out an unusual weight for nursing homes it'd be a good sample.

Thanks!! If I ever feel like doing more experimenting with this, I’ll keep their data in mind. MA is still pretty small, though, so the data is noisy even before I‘d try to break it up.

The reason I follow this on the level of the US (which is obviously not as personally useful!) is that there’s then enough data to see the patterns. I follow the death numbers because I think they are the most reliable indicator of infections 2-3 weeks before, so then the positivity correlation suggests that positivity at a specific time is the BEST predictor of current levels of infections in the population. 

Now, it’s possible the data doesn’t generalize locally in places that are contact tracing more thoroughly... my theory for why this works is that people largely self-select when to test based on symptoms. That stops being true if there’s plenty of contact tracing!!

 

5 minutes ago, Pam in CT said:

Like CT and NY data, it also breaks down by age cohort. While it is unsurprisingly true that MOST of the deaths are 65+, it is absolutely not true that no one in younger cohorts are dying.  (And MA has been among the best-managed states in terms of medical sector resource response; it is not a function of overwhelmed hospital capacity or medical personnel at this point.)

FWIW (and to the surprise of many people outside New England, lol) MA has a GOP governor.  Mask policy, closure/phased opening/testing/data transparency are not intrinsically partisan.

Yeah, Charlie Baker has been just fine.

[wintermom]

1 hour ago, Not_a_Number said:

You can see the graphs yourself. It’s a strong correlation on a US-wide level. It’s also a fairly strong correlation at the state level for the bigger states. (For small states, it’s too noisy.)

As @Danaesaid in the previous version of this thread, lots of epidemiologists know this and this is the data they watch, because it’s ahead of the deaths, and case counts are very unreliable.

What graphs are you referring to here? Do you have a link you could share. Sorry if you've already stated this somewhere in the thread. I've been battling crazy allergies and am struggling to stay on my feet.

Have you tried expanding your sampling to include data from outside of the US? With countries that have national health care and a history of excellent health databases,(e.g., Sweden) you could reduce some error. And I don't mean the World health o meter website. There are often local websites with much more detailed information. 

Edited October 16, 2020 by wintermom

[Carrie12345]

I don’t think the predictive statistics for the *nation are helpful at all to regions and individuals except maybe as a reminder to, you know, not let it get out of control.

My county hasn’t recorded a death in over 7 weeks, but more than 7% of COVID positive people overall have died. As we approach another increase in cases, my guess is that death rate will increase a bit if our facilities get overwhelmed, but continue to decline if the numbers stay low enough to reasonably manage treatment.  The ability to provide quality treatment (whether by numbers, PPE, experience, preparation, new scientific findings, etc.) seems to be a giant major factor.

Graph-wise, people have little reason to expect to die of COVID in my area.  If our system gets overwhelmed, they could have a lot more reason to expect to die of COVID in this area.

(ETA: My county’s population is under 200,000.)

 

Edited October 16, 2020 by Carrie12345

[Not_a_Number]

37 minutes ago, Carrie12345 said:

I don’t think the predictive statistics for the *nation are helpful at all to regions and individuals except maybe as a reminder to, you know, not let it get out of control.

They just show the pattern well. The pattern is that positivity is the best measure of infection rate in the population.

[Matryoshka]

2 hours ago, Not_a_Number said:

Thanks!! If I ever feel like doing more experimenting with this, I’ll keep their data in mind. MA is still pretty small, though, so the data is noisy even before I‘d try to break it up.

MA is small in area, but not really in population.  Of the 50 states, we're number 14 in population, with more than Tennessee, Missouri, and Indiana, among others.