Donate to Remove ads

Got a credit card? use our Credit Card & Finance Calculators

Thanks to gadjet,niord,88V8,Wasron,PeterGray, for Donating to support the site

Re: Coronavirus - Numbers and Statistics

The home for all non-political Coronavirus (Covid-19) discussions on The Lemon Fool
Forum rules
This is the home for all non-political Coronavirus (Covid-19) discussions on The Lemon Fool
servodude
Lemon Quarter
Posts: 4360
Joined: November 8th, 2016, 5:56 am
Has thanked: 1886 times
Been thanked: 1445 times

Re: Coronavirus - Numbers and Statistics

#472960

Postby servodude » January 14th, 2022, 3:32 am

9873210 wrote:
servodude wrote:
Julian wrote:Neither of those distortions happen with the CODC numbers but there is a delay in compiling that data due to the time it takes for the death certificates to be processed so I can see why the 28DA data is still published but I do worry that with the massive amount of testing and probably quite a few people testing themselves for the first time in the run up and over Christmas the distortions on the 28DA data might be in danger of making those data a very unreliable indicator now.


Indeed

it was useful for "dead reckoning" during the fog at the start of the pandemic when things were moving faster than the paperwork could keep up
- it has served its purpose and should be retired

the pun was accidental
-sd


The two numbers track very closely. Between Jan 1 2021 and Dec 15 2021 the r-square between the seven day averages of 28DA and the CODC is 0.99!!!! (and the death certificate numbers are 7% higher).

I think we can reasonably use 28DA as a timely proxy for CODC. In any case the alternatives that don't use it are worse. Either you make policy without data or you make policy based on data such as cases, which would result in a massive overreaction to Omicron.

!!!! I rarely use multiple explanation marks, but I have never seen an R^2 of .99 in real data. I was ready to work up a multi-variant model that included test numbers. But with only 1% left to explain it scarcely seems worth it.


I agree on the R^2 stuff
- it's the kind of thing that makes you think you've made a mistake (or I did anyway when I tried exponential fits to the rising hospital figures back in the mists of this thing)

The main issue I have with the 28DA as it is "going forward" is we know it's going to get less and less accurate
- particularly in the face of a variant with a proven degree of significant escape
- which leaves it open to some valid criticism (and then we're into politics :( )

Fixing that for me would just require the removal of the word "first" from the definition - but i think that might be fixing it in the wrong direction for most of its detractors (whereas i think likelihood of anyone dying in a given 28day period is probably quite well understood)

- sd

Julian
Lemon Quarter
Posts: 1215
Joined: November 4th, 2016, 9:58 am
Has thanked: 424 times
Been thanked: 524 times

Re: Coronavirus - Numbers and Statistics

#473044

Postby Julian » January 14th, 2022, 12:18 pm

servodude wrote:
9873210 wrote:
servodude wrote:
Julian wrote:Neither of those distortions happen with the CODC numbers but there is a delay in compiling that data due to the time it takes for the death certificates to be processed so I can see why the 28DA data is still published but I do worry that with the massive amount of testing and probably quite a few people testing themselves for the first time in the run up and over Christmas the distortions on the 28DA data might be in danger of making those data a very unreliable indicator now.


Indeed

it was useful for "dead reckoning" during the fog at the start of the pandemic when things were moving faster than the paperwork could keep up
- it has served its purpose and should be retired

the pun was accidental
-sd


The two numbers track very closely. Between Jan 1 2021 and Dec 15 2021 the r-square between the seven day averages of 28DA and the CODC is 0.99!!!! (and the death certificate numbers are 7% higher).

I think we can reasonably use 28DA as a timely proxy for CODC. In any case the alternatives that don't use it are worse. Either you make policy without data or you make policy based on data such as cases, which would result in a massive overreaction to Omicron.

!!!! I rarely use multiple explanation marks, but I have never seen an R^2 of .99 in real data. I was ready to work up a multi-variant model that included test numbers. But with only 1% left to explain it scarcely seems worth it.


I agree on the R^2 stuff
- it's the kind of thing that makes you think you've made a mistake (or I did anyway when I tried exponential fits to the rising hospital figures back in the mists of this thing)

The main issue I have with the 28DA as it is "going forward" is we know it's going to get less and less accurate
- particularly in the face of a variant with a proven degree of significant escape
- which leaves it open to some valid criticism (and then we're into politics :( )

Fixing that for me would just require the removal of the word "first" from the definition - but i think that might be fixing it in the wrong direction for most of its detractors (whereas i think likelihood of anyone dying in a given 28day period is probably quite well understood)

- sd

As I observed before there are 2 potential distortions in the 28DA data pulling in opposite directions and it is interesting (and gratifying) to see quite how well those seem to have cancelled out between 1-Jan-2021 and 15-Dec-2021. I have no issues with using 28DA as a timely proxy, even if it doesn't maintain the same level of accuracy as seen in 2021 timeliness in many cases is crucial enough to outweigh some level of error in the proxy data. My concern is the same as servodude's that going forward we might see increasing inaccuracy in how well the 28DA does act as a proxy for the CODC data. Time will tell and I assume that the sort of R^2 analysis you did is being done fairly regularly across varying time windows as part of the data analysis.

- Julian

jfgw
Lemon Quarter
Posts: 2053
Joined: November 4th, 2016, 3:36 pm
Has thanked: 697 times
Been thanked: 837 times

Re: Coronavirus - Numbers and Statistics

#473166

Postby jfgw » January 14th, 2022, 7:34 pm

I think we need to look at the data again in another week. There could be an increasing number of non-Covid deaths within 28 days of a first positive test (or, for Wales, a first positive test within a six-week testing period).
Image
My graph. Contains public sector information licensed under the Open Government Licence v3.0. : https://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/


Julian F. G. W.

Julian
Lemon Quarter
Posts: 1215
Joined: November 4th, 2016, 9:58 am
Has thanked: 424 times
Been thanked: 524 times

Re: Coronavirus - Numbers and Statistics

#473167

Postby Julian » January 14th, 2022, 7:37 pm

BBC News just posted a news piece about concern that recent(*) 28 days after data might be becoming skewed - https://www.bbc.co.uk/news/health-60000391

- Julian

(*) emphasis on recent, not disputing the good R^2 correlation for data up to 15th Dec.

9873210
Lemon Slice
Posts: 566
Joined: December 9th, 2016, 6:44 am
Has thanked: 144 times
Been thanked: 160 times

Re: Coronavirus - Numbers and Statistics

#473198

Postby 9873210 » January 14th, 2022, 10:31 pm

jfgw wrote:I think we need to look at the data again in another week. There could be an increasing number of non-Covid deaths within 28 days of a first positive test (or, for Wales, a first positive test within a six-week testing period).
Image
My graph. Contains public sector information licensed under the Open Government Licence v3.0. : https://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/


Julian F. G. W.

What that graph shows is that there is a lag in the numbers from death certificates. That's why the grey area is grey.
The graphs of the most up-to-date numbers has been showing this pattern for over a year, yet as time passes it becomes clear that there is a reporting lag so that for the most recent dates death certificate data was incomplete.

I predict, with confidence, that in two weeks a similar graph will be available showing the two lines tracking closely up until Jan 1 then diverging. I also predict that people will still use this as evidence that "this time it's different".

scotia
Lemon Quarter
Posts: 3068
Joined: November 4th, 2016, 8:43 pm
Has thanked: 1676 times
Been thanked: 1468 times

Re: Coronavirus - Numbers and Statistics

#473287

Postby scotia » January 15th, 2022, 1:05 pm

9873210 wrote:I predict, with confidence, that in two weeks a similar graph will be available showing the two lines tracking closely up until Jan 1 then diverging. I also predict that people will still use this as evidence that "this time it's different".

Yes - when I first looked for correlations between admissions and deaths, I quickly discovered the considerable time lag before deaths by death date settled down to their final values - so were of little value when looking at recent trends, although they could be used for historical data. However I was pleasantly surprised by the reasonable correlations between admissions and deaths by publication date, which I have used ever since (over on the modelling discussions).

scotia
Lemon Quarter
Posts: 3068
Joined: November 4th, 2016, 8:43 pm
Has thanked: 1676 times
Been thanked: 1468 times

Re: Coronavirus - Numbers and Statistics

#473291

Postby scotia » January 15th, 2022, 1:31 pm

Julian wrote:BBC News just posted a news piece about concern that recent(*) 28 days after data might be becoming skewed - https://www.bbc.co.uk/news/health-60000391

- Julian

(*) emphasis on recent, not disputing the good R^2 correlation for data up to 15th Dec.


Thanks for the link - my thoughts on it are:-

They have chosen a bad period to analyse death rates - around the Christmas/New Year period the deaths by published date jump around - even when summed over 7 day periods. And deaths by death date, are at best, delayed by a significant time before settling down to a terminal value. I would like to reserve judgement, and wait for a period of reasonable stability, then take a careful look at any further computations that the author(s) provide.

9873210
Lemon Slice
Posts: 566
Joined: December 9th, 2016, 6:44 am
Has thanked: 144 times
Been thanked: 160 times

Re: Coronavirus - Numbers and Statistics

#473377

Postby 9873210 » January 15th, 2022, 7:47 pm

For those with OCD.

At least three different statistics and two distinctions have recently been mentioned in this thread.
The statistics are:
  • Deaths within 28 days of positive test by date reported
  • Deaths within 28 days of positive test by date of death
  • Daily deaths with COVID-19 on the death certificate by date of death

The distinctions being discussed
  • Delays in reporting and "weekend"effects. The reported date statistic for a date is available late on the same day but has a very strong weekend effect. The other two have noticable delays of about a week for the second and about four weeks for the death certificates. The last two are published before those periods but are usually revised upwards during that time frame.
  • The criteria for including the death as a covid related death. The first two are "automatic" while the last depends on the judgement of the person filling out the death certificate, and perhaps on the judgement of the person reading the certificate.

jfgw
Lemon Quarter
Posts: 2053
Joined: November 4th, 2016, 3:36 pm
Has thanked: 697 times
Been thanked: 837 times

Re: Coronavirus - Numbers and Statistics

#473392

Postby jfgw » January 15th, 2022, 8:50 pm

9873210 wrote:I predict, with confidence, that in two weeks a similar graph will be available showing the two lines tracking closely up until Jan 1 then diverging. I also predict that people will still use this as evidence that "this time it's different".

A difference here is a noticeable increase in deaths within 28 days of a first positive test. The question is whether the errors will continue to cancel out so precisely.


There are different criteria for being counted as dying within 28 days in Wales and the numbers tend to be significantly lower: https://coronavirus.data.gov.uk/metrics/doc/newDeaths28DaysByDeathDate.
Image
My graph. Contains public sector information licensed under the Open Government Licence v3.0. : https://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/


Julian F. G. W.

dealtn
Lemon Quarter
Posts: 4608
Joined: November 21st, 2016, 4:26 pm
Has thanked: 321 times
Been thanked: 1542 times

Re: Coronavirus - Numbers and Statistics

#473435

Postby dealtn » January 16th, 2022, 9:19 am

9873210 wrote:For those with OCD.

At least three different statistics and two distinctions have recently been mentioned in this thread.
The statistics are:
  • Deaths within 28 days of positive test by date reported
  • Deaths within 28 days of positive test by date of death
  • Daily deaths with COVID-19 on the death certificate by date of death

The distinctions being discussed
  • Delays in reporting and "weekend"effects. The reported date statistic for a date is available late on the same day but has a very strong weekend effect. The other two have noticable delays of about a week for the second and about four weeks for the death certificates. The last two are published before those periods but are usually revised upwards during that time frame.
  • The criteria for including the death as a covid related death. The first two are "automatic" while the last depends on the judgement of the person filling out the death certificate, and perhaps on the judgement of the person reading the certificate.


What's that got to do with OCD?

Julian
Lemon Quarter
Posts: 1215
Joined: November 4th, 2016, 9:58 am
Has thanked: 424 times
Been thanked: 524 times

Re: Coronavirus - Numbers and Statistics

#473484

Postby Julian » January 16th, 2022, 12:22 pm

scotia wrote:
Julian wrote:BBC News just posted a news piece about concern that recent(*) 28 days after data might be becoming skewed - https://www.bbc.co.uk/news/health-60000391

- Julian

(*) emphasis on recent, not disputing the good R^2 correlation for data up to 15th Dec.


Thanks for the link - my thoughts on it are:-

They have chosen a bad period to analyse death rates - around the Christmas/New Year period the deaths by published date jump around - even when summed over 7 day periods. And deaths by death date, are at best, delayed by a significant time before settling down to a terminal value. I would like to reserve judgement, and wait for a period of reasonable stability, then take a careful look at any further computations that the author(s) provide.

The conundrum here is that if my hunch is correct about high testing rates coupled with increased transmission due to socialising over Christmas dragging more non-Covid-19 deaths into the 28-days-after figures simply because of a coincidental positive test in the few weeks before dying those effects really become most pronounced over the holiday period and the few weeks after so if they are trying to determine if that effect is real they are forced to look at the data over that period. It is true however that the lumpiness of the data over that period and into early January is simply terrible. The fact of the matter is that we won't be able to do proper analysis for quite a few weeks yet (as you point out). Work worth doing though, in my opinion, since we do need to be constantly monitoring how accurate our more timely "canary in the cage" 28-days-after figures are.

- Julian

jfgw
Lemon Quarter
Posts: 2053
Joined: November 4th, 2016, 3:36 pm
Has thanked: 697 times
Been thanked: 837 times

Re: Coronavirus - Numbers and Statistics

#473563

Postby jfgw » January 16th, 2022, 5:14 pm

9873210 wrote:At least three different statistics and two distinctions have recently been mentioned in this thread.
The statistics are:
  • Deaths within 28 days of positive test by date reported
  • Deaths within 28 days of positive test by date of death
  • Daily deaths with COVID-19 on the death certificate by date of death

The first two can be elaborated upon thus:

Deaths within 28 days include,
- England: All deaths within 28 days of a first ever (since the dawn of time) positive test result.
- Northern Ireland: All deaths within 28 days of any positive test result.
- Scotland: All deaths within 28 days of a first ever (since the dawn of time) lab-confirmed positive test result.
- Wales: Deaths within Wales in hospitals and care homes where there is a lab-confirmed positive test result and where the clinician suspects that Covid was a causative factor.

(Based upon criteria cited on the UK government and individual countries' websites.)


Julian F. G. W.


Return to “Coronavirus Discussions”

Who is online

Users browsing this forum: No registered users and 2 guests