The nerdy amounts of pandemic detail thread

[elbows]

In regards to the hospital spread stuff, the SAGE minutes of 17th December, which would have looked at the aforementioned modelling group paper, said:

SAGE 73 minutes: Coronavirus (COVID-19) response, 17 December 2020

Record of the discussion that took place at SAGE's seventy-third meeting.

www.gov.uk

Data considered by SPI-M shows that nosocomial infection across England has steadily increased throughout October and November. This is supported by CO-CIN analysis. This will lead to onward transmission within the community and will worsen overall mortality. Measures to limit aerosol transmission should be considered as part of infection prevention and control. It will be important to understand the epidemiology of spread amongst healthcare workers.

ACTION: Cath Noakes to discuss with Hospital Onset COVID-19 infection working group potential approaches to identifying and limiting airborne transmission in hospitals and provide advice to the NHS through that group as required.

ACTION: Mark Wilcox and Angela McLean to identify what work is underway to identify and monitor transmission in healthcare workers and assess whether there are any research gaps which need to be addressed.

 

[elbows]

A HSJ article about hospital infections, contains some interesting per-trust detail.

Covid infections caught in hospital up by half in a week

Covid-19 infections likely to have been caught in hospital have increased by almost half in a week, with several trusts seeing significant rises, HSJ can reveal.

www.hsj.co.uk

 

[elbows]

That article was extremely useful to me because it enabled me to confirm that I was interpreting NHS England daily data properly.

Their graph was the same shape as mine, which means I can publish these without fear that I've got it all wrong.

Using data from Statistics » COVID-19 Hospital Activity and subtracting daily spreadsheet numbers from the third table from those in the second table.

These are smoothed out using 7 day everages. The first shows England as a whole.

 

[cupid_stunt]

A HSJ article about hospital infections, contains some interesting per-trust detail.

Covid infections caught in hospital up by half in a week

Covid-19 infections likely to have been caught in hospital have increased by almost half in a week, with several trusts seeing significant rises, HSJ can reveal.

www.hsj.co.uk

Sussex seems to be poorly performing.

East Sussex Health Care Trust, from 27 cases to 86 in the latest week (31.7 per cent of cases), and at Brighton and Sussex University Hospitals Trust, where they increased from 23 to 69 (35.9 per cent of cases).

Pleased to see my trust, Western Sussex Hospitals, not mentioned.

 

[elbows]

I guess that my basic question is: what's the advantage of using the by-date-of-report numbers instead of the by-date-of-death ones - when both are available? Does the former allow us to make a better guess at the current situation, and the direction of travel in the coming days?

Now that some time has passed, I can repeat my exercise with the two forms of data on top of each other. You can judge for yourself whether the blue 7 day average by reporting date was a somewhat useful guide as to what the deaths by actual date of death would be like once that data became more complete via the passage of time.

I dont think I have managed to align the two forms of data with exatly the same distance between them as I did the first time, but it shouldnt really matter.

The original graph when we first spoke about this:



The latest version:



So yeah I would say that at least at certain stages of waves, there are useful clues about trajectory and scale to be found in the deaths by reporting date that will then eventually emerge in the deaths by date of death data.

I will repeat this exercise again in future to see how useful the deaths by reported date were for gaining a proper sense of peak, plateau and drop.

But in future I will probably do the graph properly and not perform the shift in time of one of the forms of data. I will probably also add ONS death certificate deaths, an even laggier measure which seems to be tracking the other measures quite differently in the second wave compared to the first. Since ONS data comes out on Tuesdays and Tuesday is also a weekend catchup day for the dashboard figures, I will probably do the new graph tomorrow.

edit - made correction to the latest graph as I accidentally missed off a couple of days of reporting date averaged deaths.

 

[teuchter]

Now that some time has passed, I can repeat my exercise with the two forms of data on top of each other. You can judge for yourself whether the blue 7 day average by reporting date was a somewhat useful guide as to what the deaths by actual date of death would be like once that data became more complete via the passage of time.

I dont think I have managed to align the two forms of data with exatly the same distance between them as I did the first time, but it shouldnt really matter.

The original graph when we first spoke about this:

View attachment 249885

The latest version:

View attachment 249886

So yeah I would say that at least at certain stages of waves, there are useful clues about trajectory and scale to be found in the deaths by reporting date that will then eventually emerge in the deaths by date of death data.

I'd like to come back to this in, say a week or so (in the hope that both lines might have passed a peak by then) and take a view then.

 

[elbows]

I'd like to come back to this in, say a week or so (in the hope that both lines might have passed a peak by then) and take a view then.

Sure. I suppose the simple thing to watch for first is when the deaths by date of reporting give indications that the peak has been reached. The data is after all already hinting at it, but some more days have to be given to make sure its not just a brief blip, especially if we were looking at this data in isolation rather than already starting to have expectations set by infections and hospitalisations data.

And then contrast whatever sense of peak is established via that data on a particular day, with what can be seen in the deaths by date of death figures at that time. As we already discussed, they are the same deaths, but due to the way the deaths by date of death picture builds it probably takes longer for us to notice trends emerging. eg a slowing of the rate of increase will be exist in both data, but should stand out more clearly in one (by reporting date) sooner than the other. At the time anyway, for with hindsight the deaths by date of reporting will lag behind on the graphs, as my non-time-shifted graph will demonstrate clearly when I get round to sharing it tomorrow. But for nearer-realtime daily monitoring of trends, the 7 day rolling average of reported deaths simply shows up accelerating or decelerating trends in a pretty obvious way without needing to wait for a particular days actual deaths to become close to what the final number for that day will ultimately be.

 

[elbows]

I'm not that happy with my attempts to explain this using words by the way, I think the graphs have turned out pretty usefully but not so much the words. So here is another way I just came up with to try to explain.

Pretend we've only got the deaths by date of death data. We know there is a large chunk of the most recent deaths missing from this data. If we wanted to be able to spot changing trajectories, peaks etc in this data, we would have to come up with ways to compensate for that missing data. We'd probably look at what the typical reporting delays are and what sort of quantities are expected to come in later for today, yesterday, the day before that etc. And then we'd need to look at moments where these catchup figures are no longer growing larger all the time, or are starting to shrink.

That sounds like a lot of hassle. Now consider the extent to which rolling averages of deaths by reporting data can actually perform much the same task automatically without us having to think about it.

 

[elbows]

Or in other words, some sense of what that complicated compensation I described would actually look like can be gleaned from using the colour-coded graph I often share. And the colour-coding itself come from the reporting date for those deaths. So instead of looking at the distribution of a reporting days worth of cases over time, the shape of it and how its size changes over time, I could just assume the shape/distribution will remain fairly constant and then I can just use the whole total number instead as a guide. And that total number for a particular colour is that days reported total. So just use the reported totals.

 

[teuchter]

David Spiegelhalter pointed out on the radio this morning an issue with the deaths data published on the gov.uk dashboard: because it is deaths within 28 days of a positive covid test, it's affected by the much higher levels of testing going on now, compared to the first wave. Therefore, first wave deaths are under-counted.

This means that in fact, in terms of daily deaths, we are currently some way off the "real" first wave peak which he said was something like 1400, not around 1100 as shown on the dashboard.

 

[elbows]

Yes I will be covering that with a new version of my graph.

Not that death certificate deaths are epected to capture the whole picture either. Thats why excess deaths, excess winter deaths etc are also used in normal times to get a sense of the fuller burden from flu etc. These alternative means of measuring things are somewhat less necessary now that we have a mass testing system we get a fuller view via direct, positive-tested means, but still not a complete picture, especially when the 28-day cutoff is used. But I'll have more to say about that when I am ready to share the graph. I am waiting for todays further catch-up daily death numbers to come out first.

 

[teuchter]

If a death is "within 28 days of a positive test" as per the gov.uk dashboard, that doesn't actually tell us that the death is "because" of Covid - is that right? And there is a separate measure which depends on whether Covid is mentioned on the death certificate ... that's the measure that appears in the less regularly updated ONS numbers. Is that right?

 

[elbows]

'because' of Covid vs 'with Covid' is an area where I might fall out with people so I dont intent to indulge such attempts to undercount. Indeed I've gone out of my way to ensure that my graph will also include a version of the figures for England that include people who died within 60 days of a positive test rather than within 28 days.

ONS figures are indeed based on Covid being mentioned on the death certificate. This isnt necessarily a stable measure either because its a measure thats affected by the perceptions of those writing death certificates at that time. Assumptions are made. The nature of them can change over time, for example during a harsh wave there are different perceptions about causes of death than there will be during a pandemic lull. When there is mass testing, there will be less inclination to make assumptions instead. Some deaths are still missed, as first wave excess death figures compared to death certificate Covid deaths indicates, although thats one I'll have to cover another day since I dont have excess deaths on the forthcoming graph that I keep mentioning.

 

[elbows]

I was unable to update my graph in the way I wanted today due to a problem with the official dashboard currently not showing UK deaths by date of death. So here is a graph using yesterdays data instead.

The red and blue lines use the data we were discussing before in relation to whether one offered an advanced clue about how the other would eventually evolve when data caught up. So the standard daeths within 28 days of death, by date of death measure that I always use. And the rolling average of those by date reported instead of by date of death. Unlike my previous graph where I manipulated the timing in order to align the curves to see how well they ended up tracking each other, this time I have not messed with the timing.

In addition I have added a green line where I have subtracted the 'deaths within 28 days of a positive test' figures for England from the UK total, and then added the 'deaths within 60 days of a positive test' figures for England which are available on the dashboard. I would do it for all UK nations, but I dont have that data.

And I have added an orange line which shows Covid-19 death certificate deaths, published by the ONS weekly, but which also includes figures from NISRA and NRS covering Northern Ireland and Scotland.

We can see that in the first wave, as we knew at the time, lots of deaths were missed by a weak testing system etc, eg deaths recorded by this measure were mostly hospital deaths and many deaths at home, and in care homes, were missing, So the orange death certificate curve is way higher than the others during the first wave, and this period is where the big disparities between different totals we used to see came from.

Its a different story in this second wave. We can see that so far, the death certificate death figures have tended to track somewhere in between the deaths within 28 days and deaths within 60 days of a positive test figures. I do not know if that trend will continue all the way up to the peak, and we can see that the death certificate deaths are a more laggy measure that does not currently cover the most recent period of peak/plateau.

As mentioned earlier, death certificate deaths were likely not the full picture either, and total deaths from all causes/excess death figures were also of some use for checking whether there was still a part of the picture missing. At some point in the future I will try to cover this with a graph. But I'd say that this measure is not proving to be as useful a guide in the second wave as it was in the first, for reasons I've talked about before (less deaths happening from other causes than is normal). So I doubt it will end up offering strong clues about whats missing from the second wave picture, whereas it offered a quite compelling guide to some aspects of the first wave. This does add additional complications should I attempt to properly answer questions about how the second wave really compares to the first. At some point I will try, but now is not that time.



Also note it is the fact that unlike in the first wave the death certificates data isnt capturing more deaths for the record than the other types of death reporting is, when combined with the fact the death certificate numbers dont cover the most recent period at all yet, that means that unlike during the first wave, the following sort of BBC graphics no longer show a large difference between three different sorts of death measurement. Plus the stuff I already rambled on about in terms of excess death figures not capturing the picture either. At some point I will probably find a screenshot I took of this sort of BBC graphic much earlier on in the pandemic, to illustrate the difference, but I dont have that on hand at the moment, will add it some other time,

 

[elbows]

The following is preliminary because I stumbled on it by accident when messing around with some graphs looking at the timing differences between different sorts of data.

It is possible I have made some mistake so take these findings with a pinch of salt for now.

If I take the number of people with Covid-19 in mechanical ventilation beds in England, and then take the number of Covid-19 deaths within 28 days of a positive test, by date of death, for England, and then multiply the deaths figure by 3, I get an extremely strong correlation for much of the second wave. I do not know if this phenomenon will extend all the way to the ultimate peak level of death that will be shown by these figures once all the data for that period is in. I certainly would not make a strong claim that it will. But it should offer a pretty strong indication about timing of things and when to expect the level of death to decrease.

 

[elbows]

As for our earlier conveersation about whether the trajectory of rolling average daily reported deaths was useful for predicting what the deaths by date of death would eventually end up looking like for the period, nothing really happened to change all that I said previously about that. Certainly I think there are periods where it is more useful, eg when the trajectory is clearly changing, it does offer clues about momentum of things to come. I dont think a single version of my graph can really do the subject justice, since with the benefit of hindsight the actually deaths by date of death picture becomes more complete, and in this and other regards my graphs only capture the stagte of the picture at single moments in time.

It happens that on this occasion I had prepared an earlier version of the graph some days ago, but did not post it because at the time I completely ran out of energy to talk about it, so decided I might as well wait for more days data and an extr weeks ONS data. But I took a screenshot of it at that time so I may as well post it now so it can be compared to the earlier one. For reasons such as seeing how incomplete the death by date of death was then for a crucial period, and for seeing what stage the blue rolling average reported deaths had reached some days ago.

Anyway I would conclude that there were moments in time where the average of reporting offered clues, especially on the way up. It does offer clues about the rate of increase starting to slow, but it is also possible to misinterpret glitches in the curve that are actually caused by additional reporting delays, with a more genuine change in trajectory. So for example in the following version of the graph, it was not possible to tell whether a plateau was just starting to come into view, or whether it was just more stepping on the way up. So all I could have said with it at that point was that the average was not climbling quite as fast as it was in an earlier period. Whereas with the most recent version we can clearly see a plateau and are expecting to see continued signs of that plateau in the deaths by actual date of death stats.

I spent too long talking about this compared to how useful it actually is. Given the choice I would always rather use various hospital data as a guide.

 

[teuchter]

Interesting to look at all of these graphs - thanks.

 

[teuchter]

Thought I'd wait until today's results before trying to make any comment on the question of what the current "day reported" deaths average can tell us about where the actual deaths-by-date of death numbers are going.

I appreciate that this is not necessarily a hugely important question, but I'm just interested in the extent to which that number can give a false sense of confidence about what's actually happening.

I wanted to wait for today's results to see if they continued on the trend of being lower than those 7 days previously and happily they do. Therefore I feel we can say we are probably over the peak, or at least over "a" peak.

Your graphs suggest to me that the same is true of the second peak as the first peak - the "real" peak happened a few days earlier than the "by reporting date" average suggests. Obviously, that'll not be entirely clear for a few more days yet.

It was on the 19th January that I posted that I'd put my money on the peak already having passed, and that the peak itself would have been around the 14th of January. That was met with a fair bit of scepticism because the "by reporting date" average appeared at that point still to be on a fairly relentless climb.

Well, it's still a bit early really to say when the peak actually occurred but looking at the numbers as currently reported:



My guess at the 14th turning out to be the peak was a little optimistic but plausibly only 3 or 4 days early. And it looks like it might just have been true to say on the 19th that the peak had already been passed. At the moment the 16th has the highest number but this may change of course.

 

[elbows]

Well if you recall I wasnt skeptical about the timing difference between peak admissions and likely peak deaths that you were using to come up with January 14th. I was skeptical about the choice of January 14th because there was a higher hospital admission figure on a later day than the one you chose, which caused me to ask at the time why January 19th wasnt as good or better a guess. Not that it really matter at all which day exactly it was, since the key questions were whether we could deduce what would happen with deaths based on other sorts of data that we get regularly that is not as laggy. I know we have explored various detail on this in recent weeks and I dont think my conclusions have changed - the other data is a pretty reliable guide but with a few caveats, so it pays to accept the other indicators generally as strong clues, but leave a bit of vagueness in there so we dont end up quibbling over individual days.

Also I believe I was more interested in which day exactly was the peak of the first wave because there really was one single day that stood out distinctly above the others in the first wave. And because at the time, we were given far less daily data and the measure of death was very incomplete. Some of my charts showing how different data compares in terms of timing are not a fair reflection of the information we actually had access to at that stage in April. And we didnt know how well or how quickly lockdown would work back then. This time my attention is drawn more to the length of the peak, and at what pace the number of deaths will decline.

 

[teuchter]

Well if you recall I wasnt skeptical about the timing difference between peak admissions and likely peak deaths that you were using to come up with January 14th. I was skeptical about the choice of January 14th because there was a higher hospital admission figure on a later day than the one you chose, which caused me to ask at the time why January 19th wasnt as good or better a guess. Not that it really matter at all which day exactly it was, since the key questions were whether we could deduce what would happen with deaths based on other sorts of data that we get regularly that is not as laggy. I know we have explored various detail on this in recent weeks and I dont think my conclusions have changed - the other data is a pretty reliable guide but with a few caveats, so it pays to accept the other indicators generally as strong clues, but leave a bit of vagueness in there so we dont end up quibbling over individual days.

Also I believe I was more interested in which day exactly was the peak of the first wave because there really was one single day that stood out distinctly above the others in the first wave. And because at the time, we were given far less daily data and the measure of death was very incomplete. Some of my charts showing how different data compares in terms of timing are not a fair reflection of the information we actually had access to at that stage in April. And we didnt know how well or how quickly lockdown would work back then. This time my attention is drawn more to the length of the peak, and at what pace the number of deaths will decline.

The scepticism I refer to wasn't from you - it was I believe cupid_stunt who reckoned that deaths were obviously going to continue going upwards for another week or two following the 19th.

I agree with most of what you say here.

 

[cupid_stunt]

The scepticism I refer to wasn't from you - it was I believe cupid_stunt who reckoned that deaths were obviously going to continue going upwards for another week or two following the 19th.

I agree with most of what you say here.

My scepticism was about your 14th Jan claim, which you now admit was wrong, so you've now changed it to the 19th, as they are still adding number of deaths to that date & beyond, time will tell if you are wrong again.

 

[teuchter]

My scepticism was about your 14th Jan claim, which you now admit was wrong, so you've now changed it to the 19th, as they are still adding number of deaths to that date & beyond, time will tell if you are wrong again.

Nope, what I said, and what you said are on record here.

 

[elbows]

The scepticism I refer to wasn't from you - it was I believe cupid_stunt who reckoned that deaths were obviously going to continue going upwards for another week or two following the 19th.

I agree with most of what you say here.

Yeah I didnt think you were claiming the skepticism was from me, I was just using that comment to launch my points and reiterate my stance of that moment. I get worn out and make a meal of some aspects of discussing this stuff because of ambiguous potential meanings behind our straightforward language.

For example I would have sided with you in terms of there being indications in other data of the time that the number of daily deaths was not going to keep rocketing up at the pace seen previously. And that it was reasonable to assume that a plateau and peak were looming. But I suppose there are other possible interpretations of deaths keeping increasing. For example even when a sense of plateau was looming, I could not be sure that level that plateau and peak would reach. First I didnt know if the figures would breach 1100. Then I didnt know if they would breach 1200. And indeed right now I cannot say whether any of the highest days will end up going past 1300.

As previously mentioned I was thwarted from doing some of my usual charts yesterday because the UK dashboard wasnt showing the UK figures by date of death properly, Wales was missing. Wales figures were still present on another part of the dashboard but when I tried to generate UK data by adding up the individual nations data it created some anomolies that meant I couldnt do my usual colour-coded graph with confidence. The dashboard data is fixed again as of todays data, but I have had to combine yesterdays and todays reported deaths into a single colour because I couldnt accurately snapshot things yesterday.


This is certainly the kind of moment where I am glad we have other sorts of data to guide our expectations for what will happen next. Because on the one hand by only looking at this graph we are starting to see falls that seem to go beyond the usual missing data for more recent days. But we are also perhaps seeing a slight increase in the delay in reporting of some deaths, eg the amount that has been added beyond 1200 for what may turn out to be the peak days this wave. So I'm glad we have stuff like the hospital admissions data and intensive care data that indicates we really should be looking for, and then eventually actually seeing for real, a drop in this graph that isnt just down to reporting delays.

 

[cupid_stunt]

Nope, what I said, and what you said are on record here.

Indeed, you put your money on the 14th...

I'm going to put my money on the peak in deaths having already occurred, somewhere around the 14th of January.

And, I questioned your lack of thinking...

Why on earth would you think that?
<snip>
Why do you think these increased hospital admissions will not result in higher deaths over the next week or two?

The 'next week or two' was clearly in reply to your 14th date, and numbers are still being added to that date and every day since*, your predications are totally baseless & useless.

*ETA - as illustrates in the graph just posted by elbows.

 

[teuchter]

elbows I don't know why the gov.uk dashboard doesn't do what you do, and colour-code by day of reporting, let's say for the past 7 days of reports or something like that. It strikes me that adding the colour coding allows a much more immediate understanding of what the graph represents, and allows more useful predictions about what's going to happen next, to be made.

 

[teuchter]

Indeed, you put your money on the 14th...

And, I questioned your lack of thinking...


The 'next week or two' was clearly in reply to your 14th date, and numbers are still being added to that date and every day since*, your predications are totally baseless & useless.

*ETA - as illustrates in the graph just posted by elbows.

I took "will not result in" (and the rest of that sentence) to be indicating future tense, and you were typing it on the 19th.

My money was not "on the 14th" it was "already having ocurred" by the 19th. And "somewhere around the 14th".

Anyway, we can come back to this in a week or two and see what the chart looks like then.

 

[elbows]

elbows I don't know why the gov.uk dashboard doesn't do what you do, and colour-code by day of reporting, let's say for the past 7 days of reports or something like that. It strikes me that adding the colour coding allows a much more immediate understanding of what the graph represents, and allows more useful predictions about what's going to happen next, to be made.

I dont know, maybe they never thought of the idea, maybe it opens a can of worms. It certainly requires more data storage structures to capture the 'by date of death' picture at different moments in time. ie If I dont download the right data each day from the site, then with the current system I lose the opportunity to get it later. Not that I should make that claim too strongly since I never properly investigated their API.

Plus we have seen in the past that if I try to cover too many days with colour-coding, I create an eyesore.

An animation of the graph evolving over time is another way to capture the same detail. Its not something I intend to attempt as I've already overdone it in terms of how much data I try to keep track of and graph, and I threw away a lot of my daily snaphots not too long ago. If it turns out that a 3rd party has been capturing all that data then I suppose I might revisit the idea, but what I already do does feel a bit morbid at times so I'll probably not.

 

[teuchter]

Have they only just added this graph to the gov.uk site - or have I just failed to notice it's there?

 

[elbows]

Yeah I think they only added that in recent days.

 

[2hats]

A new multiplexed RT-qPCR assay based around deletion in ORF1a Δ3675-3677, which improves on the ThermoFisher TaqPath spike gene target failure, and could be used as a proxy to rapidly detect all the current variants of concern (B.1.1.7, B.1.351, P.1). Also additionally to prioritise which samples should be sequenced.
DOI: 10.1101/2021.01.28.21250486.