Friday Night Covid Plotting #6

New Topic
This topic has been archived, and won't accept reply postings.
 wintertree 01 Jan 2021

The 4-day Christmas weekend has led to an exceptional low in sample taking pillar 2 testing, and there's a big spike for Tuesday 29th still developing in the provisional leading edge of the cases data.  Normally I try and sort out the spikes and lows in the data to get a decent analysis of the leading edge, but the extra delays mean the data I need isn't ready by this particular Friday, and it's beyond my normal de-weekending algorithm.  This makes it really hard to try and measure anything about cases in the last week without bias; I've picked the  windows I judged to be best and missed out the most compromised plots, but treat the right side of cases data plots as more provisional than normal this week.  There's still no ONS update on daily infection rates so that's some more plots still missing.

Plots 6, 7, 8

  • These stop a bit further back than normal due to the sampling lags discussed above.  They stop at just over 40k; the provisional data suggests cases are still rising after the end of this plot, but by how much it's hard to say accurately.  Lots.
  • There is now a clear rising signal in all 3 plots; the valley between the "wave 2" lockdown induced decrease  and "wave 3" rise brought on by the new variant is shallower in each successive plot due to the broad distributions in time between detection and admission, and admission and death, which blurs out the features.

Plot 9

  • All measures are now rising as indicated by positive exponential rate constants and doubling times.  
  • The exponential rate for cases has backed off from ~10 days to ~20 days.  This is still provisional and could change, but it's a promising sign - cases are still growing but they didn't stay on the frankly bonkers exponentials we saw around early December.  Later posts dig in to geographic and demographic breakdowns of this.
  • Hospitalisations are growing at a lower exponential rate than cases.  This means that the alarming rise in cases is not as disastrous as one might think - although I fear it's disastrous enough.  This is at least partly understandable through the demographics, covered in plots D1, D2 and D3 in a future post.

Plot 10

  • Not featured since thread #2 - measurements of Case Fatality Rate for different lag times between the detection of an infection as a case and the death.  
  • The NHS could produce an actual, demographically broken down version of this from longitudinal data.  Perhaps I've missed it and its out there publicly.
  • Not having proper data, I measure it for different lags.  
  • Which is the "best" one?  I lean towards around 17 days.
  • The curves cross each other when cases change from rising to falling and vice versa, discussed on a previous thread.  
  • All CFRs rose heavily in early December and are now dropping again.  This is puzzling to interpret without an independent estimate of actual infections numbers (vs detected cases).  Perhaps this is an indication that the fraction of infections detected as cases dropped for a while, when cases were rising fastest?  
    • I don't like Plot 10 and much prefer Plot 3 which was last seen on thread #4.  Until the ONS provide the data, we can't have that plot, so this is the next best thing.

Previous thread - https://www.ukhillwalking.com/forums/off_belay/friday_night_covid_plotting_5-7...

Post edited at 20:37

2
OP wintertree 01 Jan 2021
In reply to wintertree:

Plot 6 - nations

  • Wales looks to have turned a corner to decreasing cases.  From my understanding of supplemental figure S1 in the recent ICL report [1] on the new variant of the virus, it is unlikely that the recent exponential rise in Wales was driven by this new variant.
  • England - the  doubling time looks to be backing off from the peak of ~10 days; it's hard to say how much of this is real and how much is the unusually high sampling + reporting lag over Christmas.  We had some discussion on thread #5 along with some noddy plotting I did showing that fast doubling times tend not to last, which is reassuring. [2].  It's not good that the doubling times are still positive, but at this point every bit of hope is adding up to a way out of this.  
  • Scotland and NI - both look to be gearing up to go exponential.

Edit: Thanks "ablackett" for pointing out the axis label mistake - fixed!

[1] https://www.imperial.ac.uk/media/imperial-college/medicine/mrc-gida/2020-12... 

[2] https://www.ukhillwalking.com/forums/off_belay/friday_night_covid_plotting_5-7...

Post edited at 21:10

2
OP wintertree 01 Jan 2021
In reply to wintertree:

These looks similar to last weeks.  Key observations and changes:

  • The doubling times vary strongly with demographics.  They peak for 20-35 years of age, and there's a marked decrease for > 75 years of age which I take to be related more to shielding than mechanics of the virus.
    • The ICL report (linked above) suggests a slight bias towards being more transmissible at younger ages and less transmissible at older ones.
  • The doubling times in the 20-35 year bracket in London are still totally bonkers, but they've dropped from 4 days to 5 days.   This is a good start...
Post edited at 20:49

2
 ablackett 01 Jan 2021
In reply to wintertree:

The y-axis of plot 6e, 6n 6s and 9w should be cases, not deaths - right?

1
 Pedro50 01 Jan 2021
In reply to wintertree:

I truly admire your efforts and contributions but I really can't be bothered to read a single word. Perhaps we could climb at the Sunderland Wall when this is over. I've had covid and now I have covid information fatigue. Keep up the good work. 

1
 kwoods 01 Jan 2021
In reply to Pedro50:

To the contrary I read them and find them fascinating. Just as well you don't have to analyse the data because someone else has done it already.

3
OP wintertree 01 Jan 2021
In reply to wintertree:

  • If you look at the curve for the "red" regions on Plot 17, you can see that their total cases is rising roughly linearly with time - you can fit a straight edge to the curve holding it against your monitor.  So, it's not exponential - this is the increase in doubling times visible in plot 18.  They peak around Dec 7th and then get longer again.

Plot 18 - regional characteristic times

  • Doubling times for the “red” regions peaked around December 7th and have been getting longer since - this is from before London went to Tier 3 (16th) or tier 4 (19th). 

    • Key point: So, the growth rate reduced before control measures.

    • Several of us discussed this behaviour where the initially very rapid growth of an outbreak apparently self-moderates on thread #5 - shown by a very short doubling time then getting longer - often before control measures catch up.  We discussed various possible mechanisms but whatever it is, it’s very good that doubling times did not stay at ~ 5 days for the “red” regions.

    • The south west has now joined the “red” regions in terms of doubling times, although the absolute number of cases there is currently quite low.

    • The other “blue” regions all have cases increasing with their doubling times getting shorter.  I’m minded to think that a lot of the moderating effect seen that brings the doubling time down for the “red” regions is social as people take risk control and behavioural measures more seriously, and that people in the “blue” regions will be doing so as a result of the news stories on London and the delayed school terms driving home the seriousness of the current situation; so I don’t think we’ll see the doubling times for the “blue” regions going much above 10 days.  The growth in their doubling times slacked off at the same time it reduced for the “red” regions - which again makes me think behavioural factors in response to the news and messaging rather than related to the new variant, as it’s synchronous across all locations.

    • Tentatively this plot is a huge relief in that - combined with my speculative interpretations above - it suggests that the new variant hasn’t completely broken control measures as long as people stick with them really well.  The plots are not good news in that everything is still growing - but there are signs it can be brought under control.  

Breaking down below the level of regions, Plots 20a and 20b focus on the UTLAs within 2 areas that are furthest ahead with the new variant - perhaps a pandemic within a pandemic.  These cover UTLAs around the Thames Estuary and around London, and show the cases/100,000 people/day and their doubling times. 

  • Doubling times for all UTLAs are lengthening - the exponential growth is slowing.  This is a necessary step towards the growth becoming negative and the times becoming halving.
  • Some London UTLAs have pitched over in to decay, as have Medway and Southend-on-Sea.  
    • This has happened in the last 3-4 days of data (excluding the provisional window). It's not reflected in the doubling times yet, because that is measured over a window of ±7 days (or as much of that as exists), so it's a sort of average over a longer period.  This means it takes a bit longer for it to pick up on changes at the leading edge of the data.  It's a trade-off between immediacy and the quality of the measurement (i.e. too short a measurement window and it just tell us what the noise in the data is doing...)
  • This pitching over of cases in to decay happened within days of going in to Tier 4, but the infections must have happened before Tier 4 was activated (which happened with no notice), to have been detected when they are.
    • Tentatively, this is more good news - factors on the ground in these areas responded to and started to limit the growth of cases before the government response.  Some speculation on thread #5.

I try and avoid preaching or ranting on these threads, but this business where the doubling time appears to self-moderate down at the UTLA level is I think important. It tells me that people on the ground are responding once the infection gets to a "rampant" stage.  How are they responding?  By doing things that make them less likely to catch and transmit the virus; a suspicion is that part of it is reigning in a previously relaxed attitude to the rules and best practices around distancing etc.  It's becoming increasingly obvious from the news coverage over hospital occupancy levels that a very real and very serious nationwide crisis could break within 1-2 weeks over healthcare, so if ever there was a time for all of us to apply the precautionary principle and do what can to avoid transmitting the virus, now is that time.  

Those of us in the "blue" regions on plot 17 can respond before the infection gets rampant in these areas.  I think people already are given the way the +ve gradient of all curves on plot 18 becomes less positive around Dec 8th.  


1
OP wintertree 01 Jan 2021
In reply to Pedro50:

> I truly admire your efforts and contributions but I really can't be bothered to read a single word. Perhaps we could climb at the Sunderland Wall when this is over. I've had covid and now I have covid information fatigue. Keep up the good work. 

To be fair, when I read a scientific paper, 90% of the time I just look at the pictures...

I'm long (long) overdue a trip to Sunderland Wall and definitely need an excuse to get fit.   

1
OP wintertree 01 Jan 2021
In reply to wintertree:

Last plots, with a shifting focus given where we sadly now are.

Plot 20 - Data on admissions to English hospitals in the demographic bins provided by the dashboard.  

  • The astute reader will notice that they're rather coarse.  The last few days seem to be seeing sharp upticks in the number for all 3 adult age bins.  I think this is where the last 2-3 weeks of sharply rising cases start to really hit the NHS.

Plot 21 - An attempt to produce a higher demographic resolution version of plot 20. 

  • It's done by a rather dubious model fitting approach between demographic cases data and a demographic IFR and lag model re-binned in to the admissions data. 
  • The NHS have this as raw data but aren't releasing it in bins compatible with the demographic cases data. This is maddening, but more importantly it must surely be a hinderance to people doing science that is in the pipeline to SAGE. Something to be looked at I think once we are in happier times is sorting out reporting so that as consistent as set of geographic and demographic bins are used as possible across the myriad different reporting systems involved.
  • Take this plot with a pinch of salt, I would say it's a way to try and speculatively interpret the data from Plot 20. Suffice to say, it's not just the "extremely old" going in to hospital.  This echos what various medics are saying on Twitter and the news.
  • The x-axis is the date on which a case was detected, which will go on to be a hospitalisation at some point in the next few weeks.  It seems the older someone is, the less time that delay is.

Plot 22 - a Lissajous figure of the number of people in hospital in England in total and in ITU beds.

  • "Wave 1" is the first blue loop in the top right of the plot.
  • "Wave 2" is the red line ending in a small red loop just under where the blue loop starts.
    • Note how the hospital occupancy is as high has last time, but the ITU occupancy is less than half what it was.  That's how much better the health services have become at keeping people of ITU and by extension keeping more of them alive, and keeping more of the survivors more health.  This is massive and the work that went on to learn lessons fast is not getting much credit.  Other posters are better informed than me.
    • The stuff of nightmares though - imagine that the virus that first landed in the UK was the new, much more transmissible variant, and we hadn't had time to learn any of these lessons.  Given how much more transmissible it is, and how badly things went in March/April, this thought experiment should but the wind right up you about just how unprepared we were for something that was realistically possible.
  • We are now in "Wave 3" which is the red line coming back out from Wave 2.
Post edited at 21:49

1
OP wintertree 01 Jan 2021
In reply to ablackett:

> The y-axis of plot 6e, 6n 6s and 9w should be cases, not deaths - right?

Fixed, thanks!  Someone mentioned it last week, but the problem with doing this on a Friday night is that it goes on the TODO list, and then it's the weekend and there's better stuff to do. Friday is (excepting this week) the optimum day in the data for a view minimally biassed by sampling artefacts though.  Also, the village pub shut.  And lockdown etc.

1
 Misha 02 Jan 2021
In reply to wintertree:

Thanks for the graphs and analysis again. As you say, it's great that growth in cases appears to be going from exponential to linear but the issue is that even linear growth is pretty bad in the current situation... This also assumes that the testing system is still coping (there is capacity nationally, no idea about the local situation in the highest prevalence areas but haven't seen anything in the news about people not being able to get tested). With education settings closed, T4 or equivalent in most of the UK and people seeing on the news how bad things are getting in the NHS, we have a chance to avoid total disaster. The data next week and the week after will be key as it would reflect the impact of T4 and get away from the Xmas reporting distortion (though there's today's BH in the mix as well).

 bouldery bits 02 Jan 2021
In reply to wintertree:

This looks quite bad.

In reply to wintertree:

With the rate going linear in That London, could it be that they've had enough cases that they're so far up the logistic curve in some age groups that it's expected? Or is that still wishful thinking? 

 BusyLizzie 02 Jan 2021
In reply to wintertree:

Thank you for all this, you are amazing.

 minimike 02 Jan 2021
In reply to Longsufferingropeholder:

Overall population infections since March are estimated at 10%. Even with demographic variation it’s unlikely significant fractions of society are reaching the 70-80% necessary for population immunity to have an impact.

 bouldery bits 02 Jan 2021
In reply to BusyLizzie:

> Thank you for all this, you are amazing.

I agree!!! I mean, I'm too stupid to understand it all, like a labrador looking at a microwave. However, I find your interpretation and commentary far more informative and useful than that available through traditional media. 

 Bacon Butty 02 Jan 2021
In reply to Pedro50:

> I truly admire your efforts and contributions but I really can't be bothered to read a single word. Perhaps we could climb at the Sunderland Wall when this is over. I've had covid and now I have covid information fatigue. Keep up the good work. 


I totally agree with you. I just see graphs heading in a North East direction at the moment.

I know of individuals mixing around various households, one of whom is an NHS doctor, and they don't seem to care. Why do we bother?

In reply to minimike:

I'm not taking about reaching population immunity. You don't need anything like 70-80% to start looking less like a pure exponential. It should stop looking like exponential growth long before that, but I'm not sure what order of magnitude we're currently at nor what the coefficients should be. If it's >10% in That London, could this be the start of that effect? I don't know. 

As ever, Gilbert covers the basic case really well:

youtube.com/watch?v=TCkLSYxx21c&

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

Post edited at 09:12
 minimike 02 Jan 2021
In reply to Longsufferingropeholder:

‘Stop looking like an exponential’ is a strange term.. everyone uses it at the moment but it makes no mathematical sense. With a short enough time window anything can be fit as an exponential. Even a straight line is a special case of an exponential (k=0). So what do you mean? If you mean the rate constant falling from its absolute peak value, then yes, but that happens immediately after the first person is infected. As you point out is (very crudely) logistic, so a smooth function. Hard to define the point where it ‘starts’ to have an impact.

In reply to minimike:

I mean start looking like the second term becoming significant. If you zoom in on the bottom of the logistic curve, it very closely looks like an exponential. For small time it should. If you look at the middle bit it doesn't. There is a transition at some point where you can no longer get away with approximating only as an exponential. I'm asking whether we're there yet in certain areas. 

I'm not taking about rate constant changing (which there's no baked in reason for in a simple exponential), I'm taking about the time when the second term comes into play and "rate constant" isn't just one thing any more. Are we there yet? Dunno.

Post edited at 09:26
OP wintertree 02 Jan 2021
In reply to Longsufferingropeholder:

> Are we there yet? Dunno.

I'll do a cumulative cases plot by age for London later.  That'll give an idea of how close each age range is to the ~40% infected that would the about the half-way point of the logistic curve for the new variant.   Not forgetting that cases are << infections.  Complicated by the lack of recent ONS data to cross-calibrate these two.

My suspicion is that the sub-part of the 20-35 age bracket whose jobs (no choice) and behaviour (choice) puts them at higher risk of reception/transmission may well be getting to the half-way point of the logistic.  Not the whole age bracket, and by no means the whole population.  But enough to moderate the growth rate.

OP wintertree 02 Jan 2021
In reply to Misha:

> This also assumes that the testing system is still coping 

Very true.  I'm starting to look at the CFR and doubling time plots with some skepticism about exactly what happened around Dec 8th.  

I should do some plots for positivity; I think it will be rising which indicates that cases are becoming less reliable.  I hope that when the ONS sort out their issues (of which I can't make head nor tails from the language in their recent reports) that they can go back and provide estimate of infections/day for the missing period.

In reply to wintertree:

Yes, exactly this. Thanks, will look forward to it. Might even have a poke around the API later. It'd be interesting to try to fit to. Could possibly even use it to have a stab at how many unreported cases must have happened. You'd think someone would have already but I can't find that in the literature.

I'd be surprised if it's the half way point yet but will be keen to see if we're starting up the 'straight bit in the middle'.

 minimike 02 Jan 2021
In reply to wintertree:

I think there’s a fundamental issue here which I’m not communicating well because I’m cooking bacon with 2 small children..

In a totally simplistic one population, no intervention model of an epidemic you’re right. The logistic curve should represent the dynamics. In that case I would still argue there’s no ‘exponential phase’ per se as the logistic curve can be perfectly decomposed into an infinite basis of exponential with linearly variant rate coefficients. (Hence the rather blunt link to wiki above - apologies).

however we are looking at a complex scenario with multiple waves, variants, sub populations, interventions, behavioural factors, network complexity.. ..

so I think asking whether we are ‘there yet’ in terms of ‘exiting the exponential phase’ is at best ill defined. Even for a specific sub population.

wintertree’s plot 18 (and variants) illustrates this point perfectly. You can ALWAYS define an exponential as a local approximation to any curve and a sufficient number with various rate coefficients can, in linear combination, represent ANY curve of and complexity.

 Si dH 02 Jan 2021
In reply to Longsufferingropeholder:

> With the rate going linear in That London, could it be that they've had enough cases that they're so far up the logistic curve in some age groups that it's expected? Or is that still wishful thinking? 

Bit of fag packet maths coming up to test this out as I don't think we've had the discussion in light of recent events. (I don't know the answer before beginning.)

From the dashboard (link below), it's apparent that across London as an average, in the peak age groups, 25-30 and 30-35, up to nearly 8000 per 100000 people, 8%, have now tested positive (females - males slightly less). Obviously that will rise further yet, but it's correct up to the same data date that Wintertree is using above. The demographic average is just under 4300/100000, ie a bit over half the above.

Let's assume the same ratio holds, ie approximately double the number of cases in the two peak age groups compared to the average, at a more local level.

In Redbridge, which is one of the worst hit London UTLAs in recent weeks, again from the dashboard, the demographic average total number of cases per 100000 to date is about 6250. So in the peak age groups, lets assume it is up to 12500 in the 25-35 age range (the data download will enable an accurate calculation of this but I'm just working with the graphs and numbers presented on the dashboard website, which don't quantify the demographics well at UTLA level.)

So, from the testing data and the above assumption, about 12.5% of people in Redbridge aged 25-35 have tested positive. That's obviously way below any herd immunity threshold.

This is where some of the assumptions get a bit wild.

Let's assume half of cases are asymptomatic or untested. As far as I'm aware there is still no better estimate of this available. It might be possible to do something with the LFT data available for Liverpool now but the problem is the number of cases that LFT apparently misses.

So the above estimate of 12.5% increases to 25% in that age range in the worst hit UTLAs. Large uncertainty bands required due to asymptomatic unknowns.

Now, there was very little testing of 25-35 year olds in the first wave. Let's assume pessimistically that all of the above cases have occurred since June.

For the first wave, round 1 of the React-1 antibody study estimated London antibody prevalence in June of 13% (medrxiv link below). Let's assume (guess?) that the declines seen in Iater rounds of React-1 are unimportant and people retain immunity with T-cells. We'll just use this data to estimate the number of infections.

The national data from the same study estimated prevalence in those two age groups as only 15-20% higher than the demographic average for England as a whole. Let's assume that relationship held for London (the data isn't shown in the tables in the paper.) That increases the 13% figure above to approximately 15% (antibody prevalence in 25-35 yo Londoners in React-1 round 1.)

Between round 1 and round 2 of react (late June to early August approx) antibodies measured by React for these groups declined by 15-25%. Let's make a big assumption that the round 1 data is representative of the cases that occurred due to wave 1 before testing picked up (and hence before the cases occurred in this age group that are captured by the PCR testing figures above.) I think that's probably the best we can do in mixing these two methods, it doesn't feel unreasonable to me but it's definitely a big uncertainty.

We don't have data afaik about the geographic distribution within London in wave 1 so we'll have to assume it was broadly even.

So that means for Londoners aged 25-35 in the worst hit UTLAs we have estimated 15% were infected in wave 1 up to June and a further 25% of them since then - *40% in total*.

You could put big uncertainty bands on that, particularly because of the uncertainty about the by number of asymptomatic infections in this age group compared to symptomatic.

What does this tell us? My interpretation would be that it's possible that pre existing immunity could start to slow the spread in this age group now but that it still has a long way to go before any sort of herd immunity point is reached, especially with the increased transmissibility of the new variant and the lower prevalence in older age groups. 

https://coronavirus.data.gov.uk/details/cases?areaType=region&areaName=...

https://www.medrxiv.org/content/10.1101/2020.10.26.20219725v1

Edit: that took me ages to write so I see I've crossed posts with several people.

Post edited at 09:54
In reply to minimike:

I don't think I've been thorough enough in the question to stand that level of scrutiny. Wintertree seems to be picking up what I'm putting down. I guess I could have asked 'are we at the point where is no longer reasonable to assume the susceptible population is big enough to be assumed large and so talk in single exponential terms?' Does that cover it? You know what I'm asking. Phrase it how you want. 

> You can ALWAYS define an exponential as a local approximation to any curve and a sufficient number with various rate coefficients can, in linear combination, represent ANY curve of and complexity.

Obviously you can represent a curve as a combination of arbitrary functions of any sort you choose. There comes a point where you can no longer get away with just the one. That's... Never mind. This is moot now.

In reply to Si dH:

Yes, this is the question I was getting at. BUT as a point of order I deliberately avoided suggesting "herd immunity" was near. Just that the single term exponential might be at the end of its useful life here. Look how far you don't need to be up the curve for that to happen. 

 minimike 02 Jan 2021
In reply to Longsufferingropeholder:

I kind of understand what you’re saying but my point is the logistic curve is not an exponential which ‘switches to linear’. You mentioned a ‘second term’.. it literally doesn’t have one! To analyse the point at which something changes mathematically you have to be able to define it precisely in mathematical terms. 

My issue isn’t with the concept of what you’re saying which I totally understand, it’s with how to mathematically define that behaviour so you can fit the data to estimate it. 

genuinely I’m not trying to be obtuse or unhelpful. I just don’t know how to define that ‘point on the curve’ you’re talking about. I could make a case for it being anywhere from zero to 50%.. that’s the problem.

In reply to minimike:

Second term in the form: Dy/dt=ay-by^2

And no, there's isn't strictly a point on the curve.

> To analyse the point at which something changes mathematically you have to be able to define it precisely in mathematical terms. 

​​​This is why mathematicians and physicists don't get on 😁. I mean the point where you go "yeah, doesn't look like e^x any more does it. Looks like a logistic curve to me".

 Si dH 02 Jan 2021
In reply to Longsufferingropeholder:

> Yes, this is the question I was getting at. BUT as a point of order I deliberately avoided suggesting "herd immunity" was near. Just that the single term exponential might be at the end of its useful life here. Look how far you don't need to be up the curve for that to happen. 

Yes, I only mentioned herd immunity because I wanted to be clear to other readers I was not suggesting it was becoming a viable "strategy". Quite a few people read Wintertree's threads.

> I guess I could have asked 'are we at the point where is no longer reasonable to assume the susceptible population is big enough to be assumed large and so talk in single exponential terms?' 

That has been the case for months albeit for a different reason. The different geographic nature of the pandemic is such that we are now in a second wave in London that will become a third wave in the North. There are lots of shades of grey in the middle. At no point since testing took off in June has a single exponential been a valid representation of what is happening in a given city or region, never mind across the country, but I think Wintertree's exponential rate constant graphs are still another good way of seeing how infection rate dynamics are changing.

Post edited at 10:12
In reply to Si dH:

Think the point of my first post has been lost to history here and now we're arguing about what it is we're arguing about.

Looking forward to more analysis from the OP

Post edited at 10:30
 Si dH 02 Jan 2021
In reply to Longsufferingropeholder:

No, sorry, I got your original post and I thought it was a good question given where we are now at. Hopefully my first answer was useful.

Didn't see the rest as an argument.

In reply to Si dH:

It was.

> Didn't see the rest as an argument.

Oh yes it was! 😁 😉 Debating, then.

Debate, discussion, argument... all synonymous and good things to have imo.

Post edited at 10:37
OP wintertree 02 Jan 2021
In reply to Longsufferingropeholder:

> Looking forward to more analysis from the OP

Si dH’s estimates are a much more grounded version of the quick estimates I’ve done; same ballpark answer.  I echo comments from upthread about the importance of choosing the right language for describing this.  

Not much more from me for some time - it’s a snow day, and a small spring in a steep field nearby has frozen into a sloping sheet of ice...

In reply to wintertree:  

> Not much more from me for some time - it’s a snow day, and a small spring in a steep field nearby has frozen into a sloping sheet of ice...

Not meant as a hurry up. I mean ongoing; these threads are great. 

 SDM 02 Jan 2021
In reply to wintertree:

Do you have any data on hospitalisations by age group in different regions or health authorities?

I've heard reports of large increases of hospitalisations among children in the worst affected areas (from people working on children's wards).

I don't expect the numbers would be big enough to show up on a national scale yet but they may show up if you have numbers just for London.

I'm hoping it is just a case of far more cases leading to a greater number of young people being hospitalised rather than a sign of greater severity in young people (which would presumably also mean greater severity in young adults too).

OP wintertree 02 Jan 2021
In reply to SDM:

> Do you have any data on hospitalisations by age group in different regions or health authorities?

Maybe!  The API documentation isn’t very complete on which keys can be retrieved on which areas.  There is a key for demographic admissions, and there are areas for NHS Regions and individual NHS trusts.  When I’ve got a laptop half hour I’ll put them together and see what comes back...

In reply to SDM:

Exactly what I was scouring this thread to try to find out.

I've heard some worrying anecdotes that covid wards are filling with children now which wasn't the case in the first wave. This article mentions it with no data though.

https://www.google.com/amp/s/www.mirror.co.uk/news/uk-news/new-covid-strain...

"Striking difference from last time - large family outbreaks with teenagers/young adults the focus. Multiple family members being admitted."

 Si dH 02 Jan 2021
In reply to cumbria mammoth and SDM:

Basic data here still shows very low totals but doesn't tell you anything about recent changes if they exist. If Wintertree doesn't find the data, you could monitor these graphs 'manually' for a couple of weeks and would quickly find the answer.

https://coronavirus.data.gov.uk/details/healthcare?areaType=nhsregion&a...

Post edited at 12:19
 mik82 02 Jan 2021
In reply to Si dH:

The annoying thing about that data is the age banding

  • 0-5
  • 6-17
  • 18-64
  • 65-84
  • 85+

The 18-64 age band spans a huge range and could do with being broken down further (I wonder whether that's due to a definition of "adult"  (<18 = child and >65 = elderly))

In reply to Si dH:

Thanks. I'll probably scour through that data tonight when I'll be on a laptop instead of this phone. 

 climbercool 02 Jan 2021
In reply to wintertree:

Someone on my facebook who posts every covid conspiracy going, posted how intensive care bed occupancy is down this year compared to historic averages, i thought id better check his data before i told him what an idiot he is and than i saw that he was actually right,

https://www.nuffieldtrust.org.uk/resource/hospital-bed-occupancy

https://data.spectator.co.uk/city/nhs

How is this possible with 800 covid deaths a day? 

OP wintertree 02 Jan 2021
In reply to climbercool:

> How is this possible with 800 covid deaths a day?

In part, almost nobody is getting flu it seems.   If you trawl through the combined influence and covid surveillance reports from PHE and compare to the flu ones from past years, the flu season is basically eliminated., increased vaccination and effects of covid transmission control measures I expect.

Also, look at the year-to-year variation in excess deaths as well as the mean / trend line, to get a better understanding.

Covid deaths are not at 800/day, that’s the effect of reporting lag over the seasonal period.  They may be soon enough however.

Post edited at 13:19
 mik82 02 Jan 2021
In reply to climbercool:

Look at the absolute numbers on the Spectator data tracker - 771 intensive care beds total in London in Dec 2019 - yet  1048 on the 27th December 2020

684 patients in intensive care on average Dec 2019, then 913 on 27/12/20

The apparent percentage occupancy has dropped as they're massively into "surge" capacity - expanding into operating theatres, wards etc and dropping from the usual 1:1 nursing ratios

If you look - all of the regions aside from the SW have increased absolute ITU bed numbers - so it will look like we've got loads of capacity when the reality is more desparate

You could tell them that there's 33% more in ITU in London than average last year, and they're effectively at 118% capacity, but with that kind of person they'll probably just stick their fingers in their ears and go "nananana"

Post edited at 13:34
OP wintertree 02 Jan 2021
In reply to SDM and cumbria mammoth:

The demographic admissions data is available at the "NHSRegion" level.  Plots below for London giving actuals and normalised to population levels by the ONS dataset "Population estimates for the UK, England and Wales, Scotland and Northern Ireland: mid-2019, using April 2020 local authority district codes".

I haven't scrutinised the population normalisation so no promises as to it's accuracy.


OP wintertree 02 Jan 2021
In reply to thread:

74,510 cases for the UK now on Dec 29th - and it's likely to grow more as it's still in the provisional window.

I think more than a bit of this is samples that were sat in the post system and other pent up demand from the long holiday weekend.  

I wonder if it'll top 80,000 cases for the day?

If this really Is using "date sample entered the lab" and not "Date written on the sample return envelope (as sanity checked by the system to be bounded to the appropriate time period)", there's a lesson here for how to improve data quality in future pandemics...

 mik82 02 Jan 2021
In reply to wintertree:

You can see why there's reports of hospitals in London seeing a younger cohort.

In the last week of November the under 65s made up 40% of admissions. It was 44% in week 2 and 51% in week 4 of December (using the London AdmissionsbyAge dataset off the dashboard)

 RobAJones 02 Jan 2021
In reply to mik82:

Is that all admissions or just covid admissions?

 mik82 02 Jan 2021
In reply to RobAJones:

Covid admissions - off the UK Gov dashboard. 

One admission every 2 minutes on the 28th Dec in London.

OP wintertree 02 Jan 2021
In reply to wintertree:

An update to my 20:49 Fri post.

I had a bit of a slow moment and didn't bring the final date for these plots as far forwards as I could have compared to last week... (The plots are correct and correctly labelled, they just could have been more current.). Doh.

Updated plots below - the doubling times in these have backed off more inline with Plot 9/9x.

The most interesting thing to me is the 10-15 age range in London; this did not go in to decrease at the start of the November lockdown unlike all other age ranges, and has gone in do decrease at the end of the plot, where nothing else has.  This seems quite a strong hint about secondary schools.

The fastest rise in ages 20-35 is a thing of the past, with age 60-65 now doubling slightly faster than other ages - every 7 days in London, 10 across the UK.  That's still really fast/bad. 

Dec 8th stands out as a notable date even more than it did. Why the rapid drop in doubling times after this date across all ages? (or alternatively, why so high just before it?).   I'm all out of ideas for that, but it seems important.  Something with a feature that sharp and so correlated across the demographics stands out a an artefact - I don't think it's in my analysis and the only other obvious sort of artefacts are the testing system or test/trace.   

Post edited at 21:12

In reply to wintertree:

Drop in doubling time or drop in rate constant? I'm hearing one but seeing the other, unless I've got the wrong end of this stick.

Was Dec 8th around the time the new variant became dominant in the SE? Did schools start closing? When did Wales lock down? (remember the papers pointing towards a different, fast spreading strain in S Wales...)

OP wintertree 02 Jan 2021
In reply to Longsufferingropeholder:

> Drop in doubling time or drop in rate constant? I'm hearing one but seeing the other, unless I've got the wrong end of this stick.

No, that's me not being careful enough with words -i.e. backwards!.  After the 8th, the doubling time increases across the demographic board (the rate constant decreases, growth gets slower).

Re: Wales; I'm not sure the analysis in the ICL paper on the new variant supports it being behind the recent fast phase in Wales.  

Post edited at 21:50
In reply to wintertree:

> No, that's me not being careful enough with words -i.e. backwards!.  After the 8th, the doubling time increases across the demographic board (the rate constant decreases, growth gets slower).

I saw it the same way at first then twigged. Drop in your charts. Increase in doubling time.

> Re: Wales; I'm not sure the analysis in the ICL paper on the new variant supports it being behind the recent fast phase in Wales.  

Same. Just noted that there is a distinct emergence of the 501 change found there and then some hand waving and then blah blah faster transmission in Cardiff. It's not 'the' new variant. There's another, similar one that seems to have popped up there. I'll try to re-find the paper I read that in again. 

I'm just speculating out loud. Can't think what other significant events line up in time to have that effect then. 

Post edited at 22:00
In reply to wintertree:

For interest it was this, but it doesn't relate.

https://www.medrxiv.org/content/10.1101/2020.12.20.20248581v2.full

 Si dH 02 Jan 2021
In reply to wintertree:

Re 08/12, tbh I think you must have an artefact somewhere. It looks too systematic across the different data sets but having just looked at the government dashboard cases charts for the UK, England and regions, I cannot see the same behaviour anywhere.

Re: Wales, I think their rises were caused by straightforward relaxation of control measures. They came straight out of lockdown in to something approaching tier 1 across the whole country, including the areas where transmission had previously been very high and cases never dropped below around 300-400 from the firebreak. This must have caused a major shift in r. When they realised rates were going up they gradually reintroduced control measures, but too slowly.

This is easy to say in hindsight of course. I suspect they were struggling because the data is always a week or more out of date and the date on which each local authority transitioned out of decline in to a rise would have been hard to spot. However I'm sure they were also slow because they had committed themselves to a national approach and many less populated areas stayed at low rates for quite a while so an immediate re-lockdown would have been very unpopular.

Post edited at 22:31
OP wintertree 02 Jan 2021
In reply to Si dH:

> Re 08/12, tbh I think you must have an artefact somewhere. It looks too systematic across the different data sets but having just looked at the government dashboard cases charts for the UK, England and regions, I cannot see the same behaviour anywhere.

You might br right; it's a relatively small but demographically coordinated change in doubling times - this sort of thing is really hard to see "by eye" in the cases data.  The 8th looks to be when all 3 "red" regions maxed out their exponential growth rate, and the decrease looks coordinated across all 3 "red" regions using two independent sets of code, but it's very poor quality data given how irregular things get a couple of weeks afterwards.  It's one to keep an eye on.  

Thanks for the comments on Wales; the worse things get the more time I spend looking closer to home...  It's equal parts reassuring and maddening to see time and again that control measures work.  

OP wintertree 02 Jan 2021
In reply to Longsufferingropeholder:

> Yes, exactly this. Thanks, will look forward to it. Might even have a poke around the API later. 

Unless I'm missing something you can't get the demographic cases data from the API.  Actually, reviewing the docs, it's possible that "maleCases" and "femaleCases" may provide this; I get the demographic data from the "unstacked" auxiliary .csv download which is UTLA level, and from which I reconstruct regional level data.

Plot below.  This is using the same ONS population dataset as the hospital demographic stuff.  The 20-25 age bin has seen 9.5% of it's population within London test +ve since July.   With testing catching perhaps 40% of cases (*) this could be closer to 25% of that age bin who have been infected .  Assuming a similar number caught the virus in wave 1, that means around 50% of that particular age bin could have had the virus, which is enough to start moderating transmission to and by people of that age if immunity persists over these timescales.  This is based on a lot of egregious assumption making, and it doesn't apply so much to many of the other age bins.  

The higher age is interesting in that there's a dip centred around age 75.  One interpretation would be that the older people are, the more they're isolating themselves - for obvious reasons - but then they get to an age of high dependancy where isolation/shielding is rarely possible.

(*) it's hard to say without the ONS random estimate.  REACT would give another estimate but I've not run the numbers on that.  It doesn't always agree with ONS which puzzles me.  


In reply to wintertree and Si dH

Thanks for the pointers. Been having a nightmare with Excel and the .gov website tonight. 

I can't find anything in the .gov data for London that might suggest the new strain is hospitalising more children than the original strain. Could be that the effect doesn't exist, or it could be too early or not enough data, or who knows what?

0-5 year olds made up 1.18% of all hospital admissions to end Nov 2020 and make up a 7 day average of 1.4% as of 26/12/2020, but there's quite a bit of variation day by day. 

Edit - to add the disclaimer that I don't have the Excel skills or statistical knowledge that some on here have so I could easily be on completely the wrong track.

Post edited at 23:31
OP wintertree 02 Jan 2021
In reply to cumbria mammoth:

>  Could be that the effect doesn't exist, or it could be too early or not enough data, or who knows what?

We don't, that's for sure.  The NHS should be able to piece together the score from the RT-qPCR test(s) a patient has had to determine if they have an old variant or the new variant and their recent medical history, and they should be able to put it together in to coherent statistics on admission demographics, time from pillar 2 test to admissions, and lethality.  I don't think they publish anything like this routinely however, let alone with the new variant.

My sense is that the rise in cases in London was so very rapid, and hospitalisations have shot up across all demographics.  Critically, before they shot up, hospitalisations of children were so low that I doubt most hospitals in London were seeing them, but with the recent massive rise in infection, they're starting to see enough children admitted that they become a fixture on the wards of the bigger hospitals.  So, perhaps there's a "perception effect" at work as they cross a threshold of detection by medical workers.

> 0-5 year olds made up 1.18% of all hospital admissions to end Nov 2020 and make up a 7 day average of 1.4% as of 26/12/2020, but there's quite a bit of variation day by day. 

That doesn’t seem like a very significant difference given the very small numbers as work.  Either way, it’s much higher than many people would expect, with the perception this is an old person’s disease.

Post edited at 23:51
In reply to wintertree:

Maybe so. There were a lot of days in that spreadsheet where there were no admissions of 0-5 year olds but they coincided with days where there might have only been say 7 covid admissions in the whole of London so when there suddenly was one 0-5 year old it might be 14% of the total for that day so the data is hard to analyse.

 Misha 03 Jan 2021
In reply to wintertree:

Re 50% of the 20-25 age group having been infected. Bear in mind that most of those infections would have been mild or asymptomatic, so any immunity may have been pretty limited and/or short lived, in the same way that we don't tend to have much immunity against the common cold...

1
In reply to Misha:

We don't know that, and 'short lived' should still be worth a few months. Many of them will never have known about it, that's true. It's not on a scale that will change much at this stage anyway.

In reply to wintertree:

That's a pretty interesting plot. Thanks for pulling it together. Really is frustrating that the surveys are becoming so far between. But using wild-ass-guess logic it's not that hard to put gallic-shrug error bars on those estimates of the % of London millennials who might have some immunity, and it does look like even at the lower end it would begin to become significant about now.

Edit: not significant in any helpful way. Just enough to change the shape of the curves a tiny bit.

Post edited at 07:38
In reply to wintertree:

> The higher age is interesting in that there's a dip centred around age 75.  One interpretation would be that the older people are, the more they're isolating themselves - for obvious reasons - but then they get to an age of high dependancy where isolation/shielding is rarely possible.

Could also be convolved with access to tests? >80 more likely to be in a care home and get picked up by more thorough testing? 

In reply to wintertree:

.

Post edited at 08:05
 Si dH 03 Jan 2021
In reply to wintertree:

> > Re 08/12, tbh I think you must have an artefact somewhere. It looks too systematic across the different data sets but having just looked at the government dashboard cases charts for the UK, England and regions, I cannot see the same behaviour anywhere.

> You might br right; it's a relatively small but demographically coordinated change in doubling times - this sort of thing is really hard to see "by eye" in the cases data.  The 8th looks to be when all 3 "red" regions maxed out their exponential growth rate, and the decrease looks coordinated across all 3 "red" regions using two independent sets of code, but it's very poor quality data given how irregular things get a couple of weeks afterwards.  It's one to keep an eye on.  

Yep, I appreciate it's hard to see subtle changes in graphs of case rate. What convinced me to post the above was your graph 17 - the kink at 08/12 is really obvious in that, even for regions not affected by the stacking (although the stacking makes it more obvious.) This should be directly comparable to case rate charts on the dashboard but it isn't. For example the East of England graphs here:

https://coronavirus.data.gov.uk/details/cases?areaType=region&areaName=...

...shows a kink in the other direction in that time period - I think on about the 10th (I looked at this one purely because it is the positive unstacked region on your graph so is easier to compare by eye.) 

I don't like picking things up like this as I know how much time you must spend doing all this voluntarily, but I think this is definitely something that needs looking at to get confidence in your output because it seems to flow through to some of your other graphs, eg the exponential constant chart.

Might be worth sampling some of the individual data items inputted directly to your filters used in the graphs vs the raw dashboard data to see if you can identify any difference? Or perhaps the filters are doing something funny? Are they possibly being put off by the outlying low number on xmas day?

(Edit - I can't see the same effect in plot 6e? The doubling time looks like it is reducing but there is no kink in the cases data like that seen in plot 17)

Of course it might also be possible the dashboard is doing some further processing of the data in the download in order to present it in the charts. I'd hope not without further explanation though.

Post edited at 08:31
OP wintertree 03 Jan 2021
In reply to Si dH:

No, I’d like to understand it.  My best guess right now is that it’s the edge of the extended fitting window hitting a residual high or low in the data from a weekend sampling low or Monday spike.  

> Are they possibly being put off by the outlying low number on xmas day?

Yes; that could ripple out quite a bit as a sudden discontinuity, and with a polynomial filter it could cause a high further left.  

Doing manual doubling time calculations agrees with my plots for random locations; I’m not going it for all of them though...  I think I’ll consider it part-real and part an artefact of likely the Christmas spike and one other weekend low I can see.  I’ll put up a plot without any filtering (other than the exponential fitting which is delocalised), and when the Christmas period is passed I’ll work out a patch to the data to redistribute the giant spike from the 29th.  I’ll try a plot without deweekeending as well although then they tend to be all over the place for obvious reasons...  It could be the interaction of this and some particularly bad sampling lows/spikes in the data.

I’ll also unstack the plot 17 curves and compare to the dashboard.

> Edit - I can't see the same effect in plot 6e? The doubling time looks like it is reducing but there is no kink in the cases data like that seen in plot 17)

17 is less smoothed than 6e and shows more residual weekly cycling.  It’s always a bit arbitrary how much to filter curves - too little and they’re gibberish to interpret, too much and important stuff gets missed.  I think plot 17 now shows residual weekly cycling from the irregular sampling; I also took the filtering way down on the doubling time heat maps after previous discussion so that could be why sampling effects show through more.

> Of course it might also be possible the dashboard is doing some further processing of the data in the download in order to present it in the charts.

I’ve occasionally sanity checked by reading off individual values and they match (only done at nation level).  They use a 7-day average which is less immediate at the leading edge but won’t have such problems with a weird event like Dec 29th.

Post edited at 08:45
OP wintertree 03 Jan 2021
In reply to Si dH:

There's no polynomial filtering in this plot anyway.  I've removed the de-weekending and taken the exponential fitting back to a ± 5 day window (11 days of data used to measure each point in the plot.), then added in de-weekending and then a ± 8 day window. 

I've split them so you can flip-book between them with the arrow keys when viewing on a computer (not mobile device).  

To my eye the de-weekending is removing a lot of vertical structure (systematic noise) from the plots and not adding anything weird.

In the data with a smaller fitting window for the exponential has two vertical bars of faster growth around early December.   With the larger fitting window the fine details are smoothed out and there's one bar landing between the two in the less filtered version.    

There is quite a large 1-day drop in the raw data around Dec 20th that isn't fully corrected by the de-weekending.  I *think* this is pulling exponential fits down as it enters their right hand side, which is a coordinated lowering across all age bands that kicks in in the right place to sharpen the right edge of the brown bands.  

So, the plots I think accurately represent the data, but GIGO (or better, perhaps NINIO for noise);  the only answers are:

  • Use more filtering, at the expense of useful, fine detail in the plots
  • Park the issue as being understood
  • Patch the raw data with additional re-distribution.  I feel that I have a good understanding of the effect around Dec 25th to 29th to justify a patch that one, but I don't feel that way about the weekend before.
  • Make an estimate of how much sampling error there is in the input data and annotate all the plots with regions at risk of being badly affected by that input sampling error (sampling error being massive under- or over- reporting by P1/P2 vs longer term trends for a single day). 

(I put the raw/de-weekended data comparisons into a more linear colour map than the normal pseudo-log on, as the later makes the specific differences much harder to grasp.  Images are ordered so you can flip-book from raw to small window).

Edit: Fixed problem with axis visibility on the plots.  I think I prefer the small window heat plots every though they're a lot noisier.  One possibility would be to increase the window size further left where the case numbers are lower and noise otherwise dominates.  

Post edited at 10:11

 Wicamoi 03 Jan 2021
In reply to wintertree:

> Dec 8th stands out as a notable date even more than it did. Why the rapid drop in doubling times after this date across all ages? (or alternatively, why so high just before it?).   I'm all out of ideas for that, but it seems important.  Something with a feature that sharp and so correlated across the demographics stands out a an artefact - I don't think it's in my analysis and the only other obvious sort of artefacts are the testing system or test/trace.   

I thought that way at first, but I've come to change my mind, aided by what I think may have been an error on Si dH's part. He thought there was a mismatch between your exponential rate constant stuff for Dec 8 (e.g. Plot 9) and the raw case date for the East of England. But all your ERC Dec 8 downturn is saying is that if we logged daily case data we would see the rising slope of the daily cases get a little less steep around Dec 8, which I think is pretty much exactly what we'd see if we logged the East of England plot he linked. I think Si dH's intuition may have momentarily got the better of him (but am happy to be corrected).

This gave me a small moment of epiphany. The great strength of plot 9 for those seeking to manage a pandemic is also perhaps a weakness when it comes to trying to understand one: it emphasises change. Humans aren't generally very intuitive when it comes to exponential rates - at least I'm not. The dramatic downturn in Plot 9 on Dec 8th encourages us to seek a dramatic explanation... but the truth is that, at our intuitive level of understanding, the change we're looking at is just a slight decrease in the rate of increase. This strikes me as exactly the scale of impact that changes in population behaviour caused by growing fear of rising cases (or more hopefully the impact of increasing levels of immunity in the population, or both) might be expected to generate.

But perhaps this intuition is also wrong. 

 Si dH 03 Jan 2021
In reply to Wicamoi:

Hi, no, you've misunderstood, I do get that plot 9 is looking at exponential constants and I understand how they work. That's why I'm mainly looking at plot 17 which is just case data and shows an obvious kink that is not replicated on the dashboard graphs. IME it's important to follow up on any obvious discrepancies like that in analyses, in case they lead to discovery of a wider issue.

I do agree though that graphs of exponential constants could be confusing or misleading for many people, who might not have seen them since they were at school (!) Same goes for graphs with log scales etc.

Regardless, wintertree I'm very impressed how much investigation you've done in the short time since I posted! Thanks

Post edited at 11:13
 Wicamoi 03 Jan 2021
In reply to Si dH:

Apologies, my mistake, and I certainly hadn't taken you for one who didn't understand. I see the obvious kink you mention in plot 17, and it agree it does look a bit artefactual. Anyway, my mistake about your non-mistake was helpful to me!

OP wintertree 03 Jan 2021
In reply to Si dH and others:

> ... plot 17 ...

Broken down by region, with raw, de-weekended and de-weekended and SG filtered traces (left) and exponential rate constants measured from those traces (right) over a ±7 day window from the x-axis date.  (This window becomes asymmetric and shorted for the last 7 days for obvious reasons around lack of data and so it's highly provisional there).  Y-axis offsets have been applied to the curves to aid in interpretation, as given in the legends.  For this reason, doubling times are not shown on the right hand side y-axis as being reciprocal of the exponential rate constant, they make no sense with offsets added...  The feint lines on the right hand side plot are y=0 for each respective axis.

The change in gradient in the South East looks to me to be the residual effect of the massive "weekend effect" in the samples data.   This happens to have occurred just before the exponential rates all start to slack off significantly - which I think is part real and part Christmas under-sampling.  The kink is in this data I think because of a different impulse response function in the filter compared to the 7-day moving average.  

Right now, I wouldn't like to say which curve is more accurate - the dashboard one without deweekending and with an asymetric 7-day moving average (all in the past 7 days from the x-axis point), or mine with de-weekending and a symmetric 21-day, 3rd order polynomial filter.  They're two different ways of interpreting very imperfect data.  If there was a "sane" noise model for the data I could tell you which is quantitatively better, but the noise (variation in cases from a trend line) doesn't follow a simple model - anything but - so it's basically impossible to adjudicate.  Mine correctly attribute effects to the centre data unlike the lagged moving average, and mine adapts to the asymmetric leading edge region in a way the moving average can't.  

Looking at the curves and scratching my head, I'm not concerned and think the combination of filtering and processing still represents an effective way of getting summary level plots out of some pretty cruddy data.  

There's this thing with the English language where if you say a word over and over again in isolation it starts to feel absurd in your head and to sound like gibberish.   (I just googled it, and this isn't just me being insane, it's called "Semantic satiation").  I think something similar happens when staring at data.  I often find myself going back to raw data and visualising it through a pipeline to pack those doubts away.  

Edit: Reading the wikipedia article, one interpretation would be that areas of ones brain end up hard-wired to grasp particular aspects of a plot, and over-use of these causes them to be reactively inhibited.  This kind of inhibition is really common in human vision and is one of the reasons the eye has constant, small amplitude high frequency oscillations.  Something I've wanted to do for years - and never managed to get all the parts and the free time in one place - is to build an eye tracking system that stabilises out these "Microsaccades" through a tip/tilt mirror as apparently then over a few seconds your field of vision dissolves away.

Post edited at 12:57

OP wintertree 03 Jan 2021
In reply to Si dH:

Edit: Comparing this to the dashboard data I see what you mean about my curve inflecting in the wrong direction though.  I think that's a result of the de-weekending which I think is doing what it's intended to do, and how the 7-day moving average responds to a very stark "weekend effect".  Others may disagree!

I think asking about very short term trends with the increasingly erratic nature of the data is of limited value right now.

This feature is going to change perhaps once the giant Dec 29th peak is finalised, and the cases can be back-apportioned to the 24th-28th, which then feeds back through the SG filter to a more accurate estimate levels to from the 14th onwards.

Post edited at 13:19

 Si dH 03 Jan 2021
In reply to wintertree:

Thanks. That all makes sense and seems like a plausible explanation. Hopefully once the bank holiday effects have all passed through and the testing lags revert to a more usual scenario it will be easier to see changes with more consistency between datasets again.

OP wintertree 03 Jan 2021
In reply to Si dH:

I’m not sure mine & 7-day moving average will ever converge for the second half of December though.  The 7-day moving average is looking very erratic on most recent data and isn’t up to the more extreme sampling discontinuities of the last few weeks.

They should come back in to more accord into January...

In reply to wintertree:

How are you treating the first and last few points when you SG filter? Just a 'for interest' question, not a nitpick. I vaguely remember scripting a SG filter years ago and that part being a headache that required a choice that had consequences. Apologies if it's already explained in the thread archeology.

OP wintertree 03 Jan 2021
In reply to Longsufferingropeholder:

It’s the scipy implementation which uses a convolution with the kernel for the centre section and polynomial fitting over the truncated windows near the edges.   So the final point in a 2n+1 order filter is done by a polynomial fit to that point and the proceeding n points.  This means the first and last n data points in the result are increasingly twitchy to the noise, hence the right side of the plots being couched as provisional.  

The same window truncation is used by my exponential rate measuring.  That’s done least squares fitting an exponential function to the data.  I’ve a nagging suspicion that as polynomial fitting can be done by a convolution, and with Taylor’s theorem, I could and should be able to make a kernel to measure the exponential rate.  I don’t have the beige jacket and beard to work it out though...

Post edited at 14:27
In reply to wintertree:

Thanks for that. All coming back to me now. I remember what I did. I had the luxury of enough record length where nothing happens to throw away the first and last (m-1)/2 points.

At this point I'll stop not-helping. 

OP wintertree 03 Jan 2021
In reply to Longsufferingropeholder:

> Thanks for that. All coming back to me now. I remember what I did. I had the luxury of enough record length where nothing happens to throw away the first and last (m-1)/2 points.

That works wonders for the opening end of this dataset...  

Getting the most accurate and most up to date information whilst still inside a pandemic is something that doesn’t seem to have had much academic or professional attention, and is something that would be important to a government looking to act as swiftly as possible to keep the effects of policy stable against a disease model with long lags from policy to critical effects.   Not exactly a motivating force for us right now then.

> At this point I'll stop not-helping. 

I’d rather people dug in to the plots and asked questions. Stops me going down some misdirected rabbit hole.  I should probably chuck the lot on GitHub and point people there.

Post edited at 14:53
In reply to wintertree:

> Getting the most accurate and most up to date information whilst still inside a pandemic is something that doesn’t seem to have had much academic or professional attention

Not joined the RAMP forums yet???!

OP wintertree 03 Jan 2021
In reply to Longsufferingropeholder:

> Not joined the RAMP forums yet???!

I’ve signed up but not ploughed through it yet then.  I’m not sure I can face a lot of modelling discussion...

In reply to wintertree:

It's not really modelling discussion. It's nothing like these threads. It's more about giving all the pre-prints the review they should get but aren't getting, and sorting the wheat from the chaff to pass on to those making the decisions.

 Si dH 03 Jan 2021
In reply to wintertree:

> 74,510 cases for the UK now on Dec 29th - and it's likely to grow more as it's still in the provisional window.

> I think more than a bit of this is samples that were sat in the post system and other pent up demand from the long holiday weekend.  

> I wonder if it'll top 80,000 cases for the day?

79,818...

Liverpool region all shot up today, the benefits of being in Tier 2 for too long - we'll be purple on the map again tomorrow without a doubt.

OP wintertree 03 Jan 2021
In reply to Si dH:

Yes, didn’t quite make 80k - almost certainly tomorrow.

I should have enough data for the 29th and 30th tomorrow to give an estimate evening out the xmas long weekend spike.   Probably closer to 65k with the spike sorted out, which is still not good.  

I’m finding it very hard to figure out trends in the most recent week of data right now other than “going up”.

 Misha 03 Jan 2021
In reply to Longsufferingropeholder:

I think it’s a reasonable assumption that mild or no symptoms mean little immunity. It’s what happens with the common cold and I understand that’s Covid’s closest relative. It would certainly be a dangerous assumption for someone to think they have immunity following a mild illness. I may be wrong of course. 

2
OP wintertree 03 Jan 2021
In reply to Misha:

I missed your last post.  It’s a good point, and the numbers are all very speculative.  The cumulative plot starts to put some bounds on them, but interpretation is still wide open.

 jkarran 03 Jan 2021
In reply to climbercool:

> Someone on my facebook who posts every covid conspiracy going, posted how intensive care bed occupancy is down this year compared to historic averages

> How is this possible with 800 covid deaths a day? 

A few possibilities spring to mind, none of which I'm sure of.

Intensive covid beds may not count in that total, they may have their own category. Wards are split covid and non-covid including ICUs. Covid control measures and mild weather will be keeping down non-covid ICU admissions. It'd take pretty careless or motivated reporting to miss that. 

Care rationing. Acutely Ill elderly patients may not be making it to ICU or even hospital if they'll derive little benefit, something which was unknown in wave one.

Best outcomes for most covid patients (even if they do count to the ICU total) may be being obtained with specialist (quite intensive) treatment not requiring a full ICU bed/staff, good early treatment may be preventing the serious complications like renal failure (and high care need) many suffered in wave one. Again covid infection control has benefits for non-covid winter admissions.

If it's expressed as a fraction of capacity and nationwide ICU capacity has increased.

It's all a big hoax

Jk

 David Alcock 04 Jan 2021
In reply to wintertree:

I haven't tilted a glass to you all for a couple of threads, so cheers for all the hard work. It seems probable I've caught it for the second time. I'm waiting on results. The first time was fever and short breath for only a couple of days (my bedfellow also had the smell thing) and I had six months of needle-lungs, to coin a phrase. This time - the cough - wow. It's exhausting. Also the lung thing is back with extra venom. Yeah I suppose this is a pity post, but I want to reiterate: no-one let your guard down. Don't let knowledge allow complacency. I've been so careful since March. It's most likely I've picked it up off one of my three teenage lads. Unavoidable, really. Well, stay safe, he wheezed! And keep the graphs coming - I'm only A level stats, but I can read them, and they're fascinating. 

Post edited at 05:35
 minimike 04 Jan 2021
In reply to thread:

Press conference at 8pm.. I’m guessing today’s figures look rather worse than the last few days in that case

 Michael Hood 04 Jan 2021
In reply to jkarran:

> If it's expressed as a fraction of capacity and nationwide ICU capacity has increased.

Almost certainly this or a variant thereof.

OP wintertree 04 Jan 2021
In reply to minimike:

> Press conference at 8pm.. I’m guessing today’s figures look rather worse than the last few days in that case

I'm not sure the cases data is going to be that much worse (famous last words?), but the reporting spike of > 80,000 on Dec 29th is a big, scary number if you're trying to make a point.

I suspect they've decided that the only way to preserve enough hospital capacity nationwide for locked-in hospitalisations from London and the South East is to lock everywhere down now.

Just throwing this in here, not because it's relevant to the discussion, but so that it's easy to find next time some knobhead starts the "it's no worse than flu" argument on another thread.

https://www.thelancet.com/journals/lanres/article/PIIS2213-2600(20)30527-0/...

OP wintertree 04 Jan 2021
In reply to minimike:

Yes, nothing that different about today's figures.

I've added a "shunt" to the data set with some manually chosen fraction of cases from the 29th re-assigned to the 24th, 25th and 28th (a Monday, which de-weekending then patches back to the weekend).

The aim here is to try and get to a better approximation of cases if there hadn't been a lot of sampling lags over the Christmas weekend.  This will then go on to reduce noise that would otherwise obscure trends in the doubling times plots.  Total cases are conserved in this process.

I would normally consider this sort of thing an outright no-no with data, but the cause and the effect are clear.  I'm happy to hear strong objections to use this going forwards!


 jkarran 04 Jan 2021
In reply to wintertree:

> Yes, nothing that different about today's figures.

Yes, looks like within a week or two we're going to need the north's residual capacity to handle patients from the south so we can't keep building demand here too. Nationwide lockdown is my bet, rules like March plus an exercise buddy (because it's dark). It may well be dressed up as another extension of tiers and a mere delay in school opening but I suspect the realty is two months minimum.

I wonder what the plan is for the hospitals, cancelling all non-essential operations, shuttling northern NHS staff south to staff the Nightingales or something more drastic, drafting vets, students and hospital trains?

> I would normally consider this sort of thing an outright no-no with data, but the cause and the effect are clear.  I'm happy to hear strong objections to use this going forwards!

Its clear what you've done and it won't much matter in a week (by which time you can probably slightly improve the estimate anyway). Might as well keep your filters working smoothly rather than have obviously anomalous data disturb them for weeks.

jk

Post edited at 16:52
1
In reply to wintertree:

Seems reasonable to me. Will know for sure with another few days' data, if you're still concerned enough then to look back over it.
I'm less convinced than ever that there's any merit in looking at the front end of these plots, as your many caveats always take care to point out. The last few days data are, for understandable reasons, all over the damn place.
And all the people who licked each others' plates on Christmas day are about to appear en masse on the charts, so, erm, good luck picking through that mess...

 balmybaldwin 04 Jan 2021
In reply to wintertree:

One thing I'm curious about is whether the test results lag has increased again (presumably as a result of bank holidays). It seems to me that none of the 58784 positives reported today (yesterday's processing presumably) are from samples later than the 30th December

 Toerag 04 Jan 2021
In reply to wintertree:

>

> I would normally consider this sort of thing an outright no-no with data, but the cause and the effect are clear.  I'm happy to hear strong objections to use this going forwards!

You've explained what you've done and why, it will be useful to see how the real figures pan out as time goes by, and this will allow you to make a better guess next holiday period. Will send you a PM that might help.

 Toerag 04 Jan 2021
In reply to wintertree:

> I suspect they've decided that the only way to preserve enough hospital capacity nationwide for locked-in hospitalisations from London and the South East is to lock everywhere down now.

Today's hospitalisations are pretty much based upon the 441k live cases / 41k new cases a week ago. Today there are 603k live case(36% increase) and 54k new cases (31% increase). A week ago the live case count was increasing by ~16k a day, it's now 23k a day. The rate has slowed (percentage wise) but the absolute increase is still climbing each day. There may not be enough capacity full-stop if the relatively high increase of prevalence in the elderly manifests itself everywhere. Or there may be enough capacity, but only if anything other than emergency treatment is stopped.  Anyone with a tobogganing injury in the next month is likely to have a relatively bad outcome. Patients in the SE were already being shipped to Plymouth hospitals a couple of days ago due to lack of capacity closer to home, that's a lot of ambulance motorway time!

 mik82 04 Jan 2021
In reply to Toerag:

There were 26,626 people in hospital with Covid in England this morning. London and East of England at record highs. Admissions per day in those areas also still going up and at record highs over weekend despite Tier 4.

 Michael Hood 04 Jan 2021
In reply to mik82:

How comes the main gov dashboard hasn't got admissions since 22/12, or total covid occupancy and mechanical ventilation since 28/12? - presumably later data is available though other channels, how comes they've not updated? Any thoughts?

OP wintertree 04 Jan 2021
In reply to Michael Hood:

> How comes the main gov dashboard hasn't got admissions since 22/12, or total covid occupancy and mechanical ventilation since 28/12? - presumably later data is available though other channels, how comes they've not updated? Any thoughts?

Not all nations are reporting fully over the festive period, and there’s lag.  If you click to view “by nation” some more up to data data is there for some nations.

More replies to thread coming later tonight.

 mik82 04 Jan 2021
In reply to Michael Hood:

Delays in data from the nations as above I think. Daily admissions data for England are available here:

https://www.england.nhs.uk/statistics/statistical-work-areas/covid-19-hospi...

 Michael Hood 04 Jan 2021
In reply to mik82 & wintertree: thanks

OP wintertree 04 Jan 2021
In reply to David Alcock:

I hope you start to pick up soon and can have - and enjoy - your curry.

>  And keep the graphs coming - I'm only A level stats, but I can read them, and they're fascinating. 

Thanks.  Useful graphs are a subject close to my heart, and visualisation has a long history with epidemiology - plotting data the right way sets people up to succeed in understanding the problems.

 Si dH 04 Jan 2021
In reply to jkarran:

> Yes, looks like within a week or two we're going to need the north's residual capacity to handle patients from the south so we can't keep building demand here too. 

Most of the fastest rising areas in the country are now in the North again (Cumbria, Liverpool region, some bits of Yorkshire and some bits of the NE.) This transition has taken place in the data over the last couple of days as the Christmas backlog was worked through. Absolute case numbers are still less than half of those in the SE but in most of these regions rates have more than doubled in the last week and we won't see the effects of even tier 3 in the data for Yorkshire, Cumbria or Liverpool for another week. In practice as we stand tonight I suspect the true rates in some areas here may already be matching those in most of the South East (800+ per 100k), we just can't see it yet. There ain't going to be much spare hospital capacity to share around.

(For balance there are still some bits of the South rising very fast too.)

Interesting announcement tonight. The effectiveness of the lockdown will depend heavily on whether people treat it like March (scared of everyone you pass on the pavement) or more like November. I'll be interested to read through the new law and guidance to see if there are additional limits he didn't mention like 1 hour exercise etc to really keep everyone at home again.

On vaccines, to do all the over 70s, health / social care workers and extremely vulnerable groups by mid Feb, does anyone know how many per week that is?

Edit, yes, guidance is up and exercise is limited to once per day in local area (your village/town/part of city.) Also emphasises the 2m rule again. If this doesn't turn things over we are truly stuffed.

Post edited at 20:41
 AJM 04 Jan 2021
In reply to Si dH:

The guidance is already out if you want a look.

A few things I noticed:

- back to exercise only. Outdoor socialising is no longer a thing that's permitted. Should be once a day and limited to local area.

- young children seem no longer excluded from counts on groups

- work in the home (nannies, tradespeople etc) can still continue

- so, I think, can church services

Post edited at 20:45
 Si dH 04 Jan 2021
In reply to AJM:

Sorry, edited in parallel. Glad I got my trip to Pex in yesterday!

In reply to Si dH:

> On vaccines, to do all the over 70s, health / social care workers and extremely vulnerable groups by mid Feb, does anyone know how many per week that is?

Numbers halfway down this. BBC had a more concise graphic but can't find it now.
https://www.telegraph.co.uk/news/2021/01/04/oxford-pfizer-vaccine-priority-...

 David Alcock 04 Jan 2021
In reply to wintertree:

Things are improving. NFB-1.

Normal for Boris = 7 days from everyone shrieking. 

1
 RobAJones 04 Jan 2021
In reply to Si dH:

>There ain't going to be much spare hospital capacity to share around.

Yep, doesn't sound like there is going to be any spare in Cumbria. Primary heads were being pressurised to open today after the director of public health Cumbria had  had been released this yesterday.  

Over recent days, North Cumbria Integrated Care NHS NHS Foundation Trust - which runs The Cumberland Infirmary in Carlisle and Whitehaven's West Cumberland Hospital - has faced a surge in admissions of seriously ill Covid-19 patients.

The situation is now so serious that senior managers have declared an Opel 4 Alert.

The move came as the trust issued a series of appeals to the public, urging people to stay away from A&E unless their condition is life-threatening.

The NHS definition of the conditions triggering an Opel 4 Alert are summarised as: "Pressure in the local health and social care system continues to escalate leaving organisations unable to deliver comprehensive care. There is increased potential for patient care and safety to be compromised."

Post edited at 20:55
OP wintertree 04 Jan 2021
In reply to wintertree:

> I'm not sure the cases data is going to be that much worse (famous last words?), but the reporting spike of > 80,000 on Dec 29th is a big, scary number if you're trying to make a point.

Just listening to the PMs announcement fro today. I deny all rumours that I am the PM's speech writer.

Post edited at 21:25
 RobAJones 04 Jan 2021
In reply to wintertree:

> Just listening to the PMs announcement fro today. II deny all rumours that I am the PMs speech writer.

From previous comment you have made, I assumed you lost all your influence when your best mate Cummings was sacked   

OP wintertree 04 Jan 2021
In reply to balmybaldwin:

> One thing I'm curious about is whether the test results lag has increased again (presumably as a result of bank holidays). It seems to me that none of the 58784 positives reported today (yesterday's processing presumably) are from samples later than the 30th December

The data needed to reconstruct lag in reporting isn't available from the API retrospectively.  I used to do a daily download of the files needed to reconstruct this but gave up as it was too time consuming and the data itself became a lot more important. 

I do have the files to show the delta released in today's update.  It looks like the lag kernel isn't to bad right now.


OP wintertree 04 Jan 2021
In reply to thread:

As mentioned at the start, the sampling delays over Christmas meant not all plots were as current as usual.  

That issue in the data is now cleared so here are some updates.  I was going to try out my patch to the data over Christmas to smooth out the obvious sampling artefact, but instead I'm going for a night walk in the snow.

The plot 16 updates below are the sort of thing peril sensitive sunglasses were invented for, but plot D2 has a glimmer of hope on the right hand side.  I've gone with London as it's a large population (so data not too noisy) and it's part of the leading edge of the new variant's pandemic within a pandemic.  There is blue there in the younger ages suggesting things are going in very much the right direction and that control measures - when taken seriously - still work against this variant.  That gives me hope for the new lockdown pulling things back from the brink,  although if it can do so in time is not something I will speculate on.   The data is pretty erratic with the sampling issues around Christmas, but this appearance of -ve exponential rates is not uniform across the demographic so I don't think it's a sampling artefact.


 jkarran 04 Jan 2021
In reply to wintertree:

The sketcher interpretation of that blue is testing rates in schoolkids may have dropped once schools closed and that institutional oversight was lost.

It's going to be a rough few months either way. 

Jk

OP wintertree 04 Jan 2021
In reply to jkarran:

> The sketcher interpretation of that blue is testing rates in schoolkids may have dropped once schools closed and that institutional oversight was lost.

Maybe.  10-15 remained positive during lockdown and went -ve over xmas, bucking the trend both times but in opposite directions.

> It's going to be a rough few months either way. 

Yup.

 balmybaldwin 04 Jan 2021
In reply to wintertree:

Ok, that looks more like it. Thanks for that. the Gov dashboards aren't showing any cases by specimen past 30/12 (at least when I last looked about 6).

 Misha 05 Jan 2021
In reply to wintertree:

> Just listening to the PMs announcement fro today. I deny all rumours that I am the PM's speech writer.

He seems to be getting better at these speeches, though I think he could have emphasised the gravity of the situation a bit more. 

 Misha 05 Jan 2021
In reply to the thread:

Thoughts on how long the lockdown will last realistically? I reckon at least until the end of February (may be schools back after half term but not convinced), followed by a gradual release into T3. With a fair wind, most areas would by T3 from Easter but it may well take longer.

Depends how far they want to drive down new cases as I reckon with the new variant they aren’t going to drop much even in lockdown (and certainly not far enough for T&T to cope). They might just take the view that the economy needs to open up once the most vulnerable groups have been vaccinated. That would ensure cases stay high and hospitalisations relatively high... I can’t imagine what this would be like  without the prospect of vaccines. 

 SDM 05 Jan 2021
In reply to Misha:

>They might just take the view that the economy needs to open up once the most vulnerable groups have been vaccinated. That would ensure cases stay high and hospitalisations relatively high...

That would be my guess.

It looks as though cases will begin to drop during lockdown as long as schools stay shut. I expect there will be a lot of pressure to open up once most of the vulnerable have been vaccinated, especially if cases are on their way down.

 Si dH 05 Jan 2021
In reply to wintertree:

Thanks for that latest info, good spot on the demographic chart. Hopefully the older age groups will follow the younger age groups down. A thought about this - do you have data from the download that allows cases picked up in hospitals to be distingushed from those picked up in the community? It would be significant if there is now a lot of hospital transmission that continues in older groups like it did last Spring well in to the lockdown.

The user friendly website reports tests by pillar but not cases.

Post edited at 07:06
In reply to Misha:

Early to mid-March. Halving time is very unlikely to be as fast as it was in April.
Conversely the pressure to take the brakes of once we've vaccinated granny and let it rip through you & me will be large, so the point where enough is deemed to be enough will probably be higher.

Post edited at 07:26
 jkarran 05 Jan 2021
In reply to Misha:

> Thoughts on how long the lockdown will last realistically? I reckon at least until the end of February (may be schools back after half term but not convinced), followed by a gradual release into T3. With a fair wind, most areas would by T3 from Easter but it may well take longer.

I think in part it hinges on the vaccination strategy and the efficacy of winter lockdown. If social controls in the absence of widespread vaccination in the older population do bring hospitalisation rates under control (we can only hope!) we have the option to re-prioritise opening schools and reinvigorating the economy using the vaccines in risk of transmission groups (as opposed to risk of hospitalisation). Realistically, weary and mislead as we are I doubt lockdown without significant compulsion is going to be more than marginally effective, holding us at best at or very near the brink of healthcare collapse which limits our options well into summer. I suspect we'll stick with 'oldest and carers first' to keep that bed pressure and HSC staff sickness coming down. I reckon we might start seeing some T3/4 trialed again in March but with the pent up demand for social contact, economic activity and assuming the new strain really is as transmissible as it seems most likely we won't see tier reductions sticking with significant progression into T2 and T1 until summer when the vaccine starts to make significant inroads into the working population and the weather really helps.  I'm assuming to get to something that feels much more normal: we deliver ~2M/doses per week from Feb, that we need ~30M people at least partially vaccinated and that one dose/person will be the norm and reasonably effective for the first couple of quarters. May/June ish, maybe. Then we're racing the weather and flu again to get the vaccination program finished and to start facing down any problems from less responsive variants. 2022 should be better to put an optimistic finish on this.

> Depends how far they want to drive down new cases as I reckon with the new variant they aren’t going to drop much even in lockdown (and certainly not far enough for T&T to cope). They might just take the view that the economy needs to open up once the most vulnerable groups have been vaccinated. That would ensure cases stay high and hospitalisations relatively high... I can’t imagine what this would be like  without the prospect of vaccines. 

To avoid healthcare overload from the mid-vaccine 'let 'er rip' we probably need to have vaccinated down into the high 40s. If we're going to do that we will need to be very patient or have a frank national conversation about care rationing and the sacrifices we the working population will be required to make. That hasn't happened yet and I doubt it will happen. I don't think the government has the stomach for it but we may get there by accident as it also lacks the spine and gumption to resist its backbenchers and donors once the vaccine (then widely deployed in older citizens) dramatically improves the CFR.

jk

Post edited at 11:06
 jkarran 05 Jan 2021
In reply to Misha:

> I can’t imagine what this would be like  without the prospect of vaccines. 

We'd probably be capping the mass graves of 1-2 million, fearing it'll come back bad next winter.

jk

OP wintertree 05 Jan 2021

In reply to geode:

I can’t follow that I’m afraid.  Unless “people with tiny errorbars” is a metaphor.

 Blunderbuss 05 Jan 2021
In reply to wintertree:

I think he means the number of people infected with the bars being the 95% confidence interval...

 minimike 05 Jan 2021
In reply to wintertree:

Now now, let’s not have an error bar waving contest..

OP wintertree 05 Jan 2021
In reply to Blunderbuss:

> I think he means the number of people infected with the bars being the 95% confidence interval...

If he's referring to the Dec 30th slides, it's number of cases vs time.  I'm surprised it has errorbars at all as this number is an actual, but the slide gives it as a per-population rate and I suppose there are errorbars on the population estimates which propagate in to the plot.

OP wintertree 05 Jan 2021
In reply to thread:

Updated plots on hospital occupancy - total (x) and ITU (y).  The second plot is a measurement of the fraction of people that are in ITU over the tail end of the "second wave" and the start of the third.

25% more cases are going in to ITU now compared to a couple of months ago by this simple linear fit.  I haven't calculated uncertainties on this so I can't say how significant the difference is.  


OP wintertree 05 Jan 2021
In reply to wintertree:

Shunting the same fractions of cases about as with [1] to try and even out the Christmas sampling lull/spike to get a more accurate view of exponential rates.

Plots for England with raw data, with the xmas shunt, and with the shunt and the de-weekending.  To my eye, the amount of clear artefacts (caused by the spikes in the raw data) in the exponential rates plot decreases with each step.

The exponential rate looks horrific before applying these steps, as does the non-provisional cases data on the government dashboard.  This attempt at sorting out the reporting spikes reigns that in a lot, but the exponential rates are still high - and applied to a high number of cases.  I think perhaps that a residual low-then-spike is still in the data and causing an exaggerated rate at the far right. 

The most right hand 5 days of the rate constant plot are always very provisional as there's less data available to measure the growth rates from (fewer or no points to their right).  We now go in to lagged reporting on the presumed New Year's Day sampling low and the subsequent weekend sampling low, so "good" updates aren't going to happen again until Friday.

Nail biting stuff, and not in a good way.

https://www.ukhillwalking.com/forums/off_belay/friday_night_covid_plotting_6-7...

Post edited at 21:51

OP wintertree 05 Jan 2021
In reply to wintertree:

Plots for London and several other UTLAs likely leading with the new variant.

Again, I think exponential rates may not be quite as bad as it looks, due to the sample lag over xmas.  

The flip to decaying cases in some younger demographics I noted up thread did not last for long.  

Post edited at 21:57

OP wintertree 05 Jan 2021

In reply to geode:

I'm still not following you.  I'm afraid you're not explaining yourself very well.

What exactly is your problem with that graph?  

OP wintertree 05 Jan 2021

In reply to geode:

> the errors for a start..

You're really not explaining yourself very well.

What is your problem with the graph?  Described in sufficient detail for someone living outside your head to understand the problem.

 Si dH 05 Jan 2021
In reply to wintertree:

> Plots for London and several other UTLAs likely leading with the new variant.

> Again, I think exponential rates may not be quite as bad as it looks, due to the sample lag over xmas.  

> The flip to decaying cases in some younger demographics I noted up thread did not last for long.  

Are those raw or shunted?

OP wintertree 05 Jan 2021
In reply to Si dH:

> Are those raw or shunted?

They're both shunted and de-weekened (any such processing now goes in the title of the left sub-plot).

I don't think this is fully fixing the flaws in the underlying data - I think there's net fewer samples been taken over the Christmas period leading to a low then high implying a faster rise of cases than is real.  I don't feel that it's appropriate to alter the total number of cases however, only to shunt them around.

So I think it's over-estimating growth on the RHS, but we really don't know and won't till more data is out - but that data is going to have the new year's sampling low in it...

Still, I think this is the best I can do for a current look.  

OP wintertree 06 Jan 2021
In reply to wintertree:

An update to plots 17 and 18 with data up to Dec 30th now included, using the shunting described above to mitigate much (but maybe not all) of the apparent massive growth caused by the low-then-spike over the 4-day Christmas weekend.

It looks to me like the "blue" regions are all headed towards the very fast doubling times initially seen in the "red" regions.  This always seems to self-moderate shortly after peaking, irrespective of changes to control measures (not just in recent events).  My best interpretation - which isn't very good - is that this is a combination of a faster growth in the fraction of the population most likely to catch and transmit the virus, and of people responding to contacts and contacts-of-contacts being hospitalised - the local grapevine informing part of their behaviour over how seriously they take this.

As ever the last 7 days or so in these plots are provisional and will change as more data comes out.  

They are not happy plots.  If the "blue" regions are all going to follow the same kind of peak-then-decay (but remain positive) trajectory in doubling times as the "red" regions, the number of cases, live infections and locked-in hospitalisations is going to go up by a lot.

We now go in to the part of the week where lagged data being reported covers a weekend sampling low, so no more updates to these until Friday.

This is looking bleak, and it's important we all do what we can to moderate the spread.  There are reasons in my view to be have some positive hopes - they are without exception not reasons to let your guard down.

  1. London is close to zero growth in this data, and London is ahead of the "blue" regions with the variant.   So, it looks like it can be brought under control after this "fast burn" phase but control measures weaker than we now have.
  2. The ratcheting up of tier levels in late December should start to show through in this data soon - the blue curves may show moderated growth sooner or faster than the red ones, preventing some of the growth in cases.  
  3. Restrictions are even tighter now again over those in 2 above.

Stay safe one and all.  


 Toerag 06 Jan 2021
In reply to wintertree:

> It looks to me like the "blue" regions are all headed towards the very fast doubling times initially seen in the "red" regions.  This always seems to self-moderate shortly after peaking, irrespective of changes to control measures (not just in recent events).  My best interpretation - which isn't very good - is that this is a combination of a faster growth in the fraction of the population most likely to catch and transmit the virus, and of people responding to contacts and contacts-of-contacts being hospitalised - the local grapevine informing part of their behaviour over how seriously they take this.

I've just had a thought - the initial fast doubling happens everytime there's a relaxation in restrictions. That would imply that either there's always a reservoir or super-susceptible people in the community which hasn't been exhausted yet, or it's pure behaviour being slack and moderating quickly.

Post edited at 13:54
OP wintertree 06 Jan 2021
In reply to Toerag:

> That would imply that either there's always a reservoir or super-susceptible people in the community which hasn't been exhausted yet, or it's pure behaviour being slack and moderating quickly.

Something like that I think.  Well beyond my ability, profession or time to dig any deeper than what this superficial look at the doubling times shows.  In terms of phenomenology it seems quite repeatable going way back in the data.  Below is a quick plot I did for plotting #5 showing exponential growth rate (y) vs absolute cases (x).  They are not independent variables, and exponential growth rates are likely to be significantly higher when cases are low than when they are high.

Post edited at 14:10

 Si dH 06 Jan 2021
In reply to wintertree:

> The ratcheting up of tier levels in late December should start to show through in this data soon - the blue curves may show moderated growth sooner or faster than the red ones, preventing some of the growth in cases.  

I hope you are right about this. I think if the restrictions do not have major effect then the quick increase may be prolonged at a national and even regional level. In the last week the rates have increased hugely in some areas of the midlands and north, but other parts of the relevant regiona are still behind and haven't begun the steep rise yet. So the gradient seen at regional level does not reflect the peak gradient seen at utla or even city region level, where many doubling times are under a week. This is best understood with a browse of the interactive dashboard map at utla level.

The areas that haven't yet gone in to steep rise surely will do soon unless the restrictions are very effective, so we should expect these increases to compensate for any flattening of the growth in areas that have increased steeply over the last week. Tier 4 was only introduced a week ago across the North so there is at least another week of growth to come in 7-day average case rates before can take effect. Conclusion - I think the steep average rise across these regions will be maintained a good while longer yet, even if individual city regions or UTLAs follow the flattening behaviour you have described. It's not good for total case numbers.

Post edited at 16:38
OP wintertree 06 Jan 2021
In reply to Si dH:

> I hope you are right about this

Indeed.  I think this Friday & next Wednesday are the main points where the data will be likely to shape that hope.

The effect you describe of regional and sub-regional fragmentation puts me in mind of fractals - the demographic and geographic distribution of cases has been almost fractal from the start, and the obvious attention is often where numbers are high, not where exponential growth rates are high.  

 upordown 06 Jan 2021
In reply to wintertree:

> London is close to zero growth in this data, and London is ahead of the "blue" regions with the variant.   So, it looks like it can be brought under control after this "fast burn" phase but control measures weaker than we now have.

Hi wintertree, could you please clarify the second sentence here. It's the 'but control measures weaker' bit I don't understand. Apologies if I'm just being dim. And thank you for doing this and keeping us updated. I don't understand a lot of it but I find your summaries really helpful.

OP wintertree 06 Jan 2021
In reply to upordown:

What I mean is that the exponential growth rate (a proxy for R, if you like) dropped from 12-07 or so until the end of this data at around 12-29 or 12-30.

This happened before the latest significant change in control measures - everywhere is now T4, schools are partially closed, T4 rules are strengthened.

This may not be obvious as I missed a word or better yet several - “this apparent decrease happened under weaker lockdown conditions than we now have”.

I see what happens with London and the other “red” regions as key to understanding where the others are going.  From pure phenomenology it feels like all regions are destined to have a brief peak of high doubling rates, control measures or not.    But then what happens?

Edit: the other two red regions look to be rising.  But it only takes one dropping to prove that it is *possible.*.  Which is an olive branch I’ll take right now.

> don't understand a lot of it

Ask questions if you like.  

Post edited at 17:55
 upordown 06 Jan 2021
In reply to wintertree:

Thanks


New Topic
This topic has been archived, and won't accept reply postings.
Loading Notifications...