UKC

This is my fourth set of plots, and it's very much not the update I had hoped to have.

• Plot 1 and 2 - this week's ONS report is missing a value for the infection rate.  I can't figure out what the report means when it details the absence...  A shame but hopefully we'll get a new estimate next Friday.
• Plot 3 - The peaks in infections then deaths gives us a good estimate of the current IFR, which looks to be about 0.75% as given by the crossing of the various curves in mid-November; the x-axis on this plot is by date of infection, not date of death.  This is really quite high.  Some random sampling surveys suggest more infections than the ONS would would give a lower IFR.  It's concerning that the IFR is currently rising regardless of which lag is used to estimate it; perhaps this is a sign of increasing prevalence in more advanced ages.  I haven't dug in to the data that much.
• There's no plot 4 as there's no new ONS data
• Plot 5 - this has changed a bit.  Some good discussion with a couple of posters on the previous thread hi lighted just how sensitive projections are at the leading edge of the data when there's false drops and rises caused by the under-sampling on weekends and the spike as this drains on Monday's.  Any convolutional filtering (7-day moving average, my preferred SG filters) is still mislead by these, as the filters are really only valid for uncorrelated (random) noise.  So I've made a method detailed on the previous thread to "de-weekend" the data.  This dramatically reduces the variance in a way that embodies the source of the noise.  I've started to use this data in some other plots - it allows me to reduce the severity of filtering on cases data in them and improves the (always tenuous and twitchy) projections made from the leading edge of the data.  

Previous updates 

https://www.ukclimbing.com/forums/off_belay/friday_night_covid_plotting-728...
https://www.ukclimbing.com/forums/off_belay/friday_night_covid_plotting__2-...

• Cases (UK) look to be rising.  This is a twitchy leading edge of the data and the trend-line could look very different by next Friday.  Un-shown provisional data for the next few days makes me think it'll change for the worse
• Admissions to hospital (UK) also seem to be rising
• Characteristic times plot - there's a lot going on here...
  • I show markers for where there is a "halving time" - the underlying measure decreases every X days.
  • I add a solid line for a "doubling time" - where the underling measure doubles every X days.
  • When swapping between doubling and halving, the characteristic time of one goes up to infinity, then comes back down as the other sort.  You can see that cases did this, forming a "U" shape bottoming out with a halving time of about 18 days on the 17th of November.  Since then, the decay slowed and the halving time shot up, before returning as a doubling time, implying more growth.
  • Hospital admissions are developing a very similar "U" shape with a lag from detected cases of about a week.
  • It doesn't take a lot of imagination to see what's likely to happen with the deaths curve which already appears to be bottoming out.
  • Everything from ~ 21st November onwards in these plots is very provisional - they're measured by fitting an exponential to the data in a sliding window, and then SG filtering the result (see first thread for more details).  This means the measurements are somewhat delocalised in time and so sensitive to retrospective data releases (values can be revised over the last month for deaths, mainly up, sometimes down) and to future data releases.  

This is a bitterly disappointing update - it seems the steam ran out of the English lockdown half way through.  The ZOE random sampling survey doesn't show a rise in infections.  The latest ONS data is missing this measure and may have been a bit to old (it always lags by a week or so) to be relevant.  So the question is, is this rise in cases real, or is it better sampling of test/trace?  I think that it's quite real given the behaviour in admissions and deaths.

The progress of Covid is becoming very regional in an almost fractally complex way with differences across the 4 nations, across regions and down to very local differences.

I've done characteristic time plots for each nation.  Something in the Scottish data is tripping it up so that's missing for now.  For NI and Wales I've disabled the line drawing as that needs some more smarts to work when there's multiple separate regions of halving times.  I've also had to smooth the curves a lot more for these, as it's smaller number statistics so more noisy.

These plots are not great in terms of aesthetics.

The Wales data has the same motif of"staggered Us" for each curve in sequence as covid control measures start reducing cases then admissions then deaths, but then they start rising again.  I think everything after 11-23 in this plot is highly provisional and the apparent halving time of deaths is probably due to large reporting lag; but I've not looked much at data at the nation level (other than England) before so I don't have much of a feel for that.

Northern Ireland also has a "staggered Us" motif although theirs for cases is more of a "W", this fine details is smeared out for admissions, presumably by the statistical relationship between them.   

Plot 16

• The horizontal lines are per-UTLA cases/100k in England at a time where this had the closest correspondence to the Tier levels activated in early December.  The line colour gives the assigned tier level.
• The dot markers show where cases/100k where on the most recent day with reasonably complete data.  Major metropolitan areas are show by the dot colours.  UTLAs that are not in such an area and are above their Tiering level are annotated.

Plot 16 detail 1

• London UTLAs stacked that have risen above their level used for Tier assignment. This has bene de-weekended and SG filtered.

Plot 16 detail 2

• The demographic breakdown (note: raw cases, not cases/100k) for the UTLAs from Plot 16 detail 1.  This has been de-weekended but othewise has no filtering.  The high case rates are creeping up the age range each day are now rising in ages where the IFR exceeds 1%.

This data is for 5th December; it's the 11th today, the Tier revision is to be announced on the 16th and likely implemented on the 19th from news reports.   That means there are two weeks between this data shown and when the tier revisions might take effect - which in practice is just before they're temporarily relaxed for Christmas.  There isn't enough data on the rising phase yet to predict with any credibility how fast cases are rising in London, and all but the last 3 days of this data comes form when London was in lockdown rather than the new Tier 2 its is now in.  

Outstanding work Wintertree. Thank you.

Great stuff again. Thanks. 

Thankyou.....very impressive although I'm really not smart enough to follow it all!!!

Your contribution, reasoning, openness and objectivity is a credit to you and the value of this forum....

with the thousands of people that will have seen your posts and data on here over the last year,  there is no doubt that behavior will have been changed and as a result i guess its quite possible that there is someone alive today that wouldn't be if you'd just watched some t.v instead.  Good job

I'm finding the characteristic times plot is getting more and more muddled, so I've turned it on it's head.  

I now plot the exponential rate constant for each measure.  This is inversely proportional to the halving/doubling time.  The larger the magnitude of the rate constant, the more quickly the measure (cases, hospitalisations, deaths) is changing.  A positive rate constant means growth and a negative rate constant means decay.   There is a second Y-axis added on the right hand side which shows the corresponding halving or doubling times - how many days it takes for cases to halve or double with the behaviour of the underlying measure at that point in time.

This plot I think makes it much clearer how the steam ran out of the English lockdown half way through, and the asymmetry between growth and decay is quite stark - although not as bad as initiation might suggest as a lot of the growth happened at lower numbers of cases, and the recent decay happened at high numbers - so it had more absolute effect.

As always the last couple of weeks in this plot should be considered highly provisional.

This plot says to me that the old T2/T3 up north worked, and that to being with lockdown worked more, until it didn't.   Obvious contenders are worsening weather and behavioural response to the news over the Pfizer vaccine but I've not seen much discussion/evidence.  

I've been thinking about it and I think perhaps the lockdown was taken more seriously by people in areas of formerly high prevalence and so was more effective in those areas - I was seeing a lot of ambulances on the roads in late October and I'm only a couple of acquaintances away form people who died in this wave  Generally more behavioural levers were being pushed on people in these areas. Then what happens is falling cases in old T2/T3 regions make up the bulk of the national cases and drive them down, with small but rising in the south going on "under the radar(*)" until a crossing point is reached where the rising cases further south eclipse the fallen and falling number up north, and their rise drives this curve.  This interpretation seems compatible with the regional plots on the MRC nowcast [1].  I'd like to get regional level data in to my plotting pipeline and I think I've found an authoritative mapping for UTLAs to do this...

So, perhaps the most important step ahead is to get the people in regions of the south to take the new T2 rules seriously; these should be enough to make a difference given their similarity to the old T3 rules.  I can't see much of the south avoiding T3 this week, but I'm not going to try and predict the government's decisions. 

(*) as in, their effect is masked in national level plots.  There's been plenty of discussion of their rise on UKC for a few weeks now.

[1] https://www.mrc-bsu.cam.ac.uk/nowcasting-and-forecasting-11th-december-2020...

I suspect there are a few factors at play but the N / SE divide is the major one for the reasons you describe. A glance at the map on the official website shows that the SE and S Wales are the main problem areas at the moment at a local authority level (of course there will be mini hot spots in otherwise relatively low prevalence areas). 

This is much clearer (rate constant plot)

Edit: though I'd probably give days the linear scale and rate constant the log. The way it is it's easier to misinterpret.

Edit edit: not that it's straightforward to draw the right conclusions from either way. "The doubling time has halved faster... 🤯"

Much prefer the new plot, good work. Agree with your qualitative analysis. The problem is that we can't be sure of the root cause of the differences that have allowed cases to re-grow in the south east and therefore it's really hard to know how effective the move from tier 2 to 3 will be (for example, if there is indeed a lot of transmission in secondary schools, that would be unaffected... Although the Christmas holidays will obviously cut that route right down for a fortnight .) Whatever they do needs to move the dial significantly in terms of people's attitudes and perception of risk I think. 

​​​​Edit to speculate a bit, I do also wonder if the general attitude in London is affected by the fact that fewer people there in their under 40s live close to their parents. Ie many have moved there for work and close family aren't nearby. I think with a vaccine on the horizon, if people don't see their behavior as a threat to their nearest /dearest then more of them will be relaxed about restrictions.

Obviously people move to work all over the country but I suspect the proportion of people living in London without any older relatives close by is probably higher? 

The problem with that is twofold - the days values cross from halving to doubling via - and + infinity.  A log scale much better suits their dynamic range, and if using a linear scale they can’t meet at y=0 but have to meet via opposing infinites, which is where the chaotic look of the old plots comes from once there are both halving and doubling times in the plot.

This is a bit of a specialist plot I think.  It’s very powerful for spotting and predicting trends between the various measures but it can also be misinterpreted.  I’ve tried to help with that with some big text labels.  Perhaps I should add “bad” and “good” as well.

An interesting observation on family ties and those acting as additional encouragement to follow the rules; I think the government must be relying on those ties to moderate the Christmas period.

One last big push on behaviour for the next 3 months could kick this to the curb. I worry that what’s actually going to happen is another round of people crying “Liberty”, “but the economy” and so one, once again to the net detriment of both.  There doesn’t seem to be much social media astroturfing going on or much protest agitation either in the lead up to the tiering announcement though.

I think the sort of MPs who would be doing that are probably preoccupied with Brexit (no) deals, at least until Monday...

To illustrate this, a stacked plot of case counts for England broken down by region.  Regions that have fallen over the last ten days are stacked below the centre-line, and rising regions are stacked above it.  We're just passed the point where  rising regions "drive" cases for England as defined by contributing over half the numbers.  For the most recent day here, rising regions contributed ~ 6000 cases/day, and falling regions ~5500 cases/day.  It certainly looks to a casual glance as if lockdown had much less effect on the regions with lower but rising prevalence. 

(This is de-weekended and SG filtered data).

Now I've got a region mapping in there, I can do some regional characteristic time plots to go with this...

Edit: These come from the UTLA demographics data and I'm not sure they sum to quite the same numbers as I get from the headline data so I'll have to do some cross checking when I next get a chance.

I’ve been monitoring daily cases in Greater Manchester, as that is where I live. Generally improving, 4 of the 10 boroughs below the national average 9 of the 10 boroughs with cases decreasing. Greater Manchester is of course tier 3.

I was aware parts of tier 2 London were getting, with Havering for example having over 400 cases per 100,000 (national average 134). However, I hadn’t realised until today that all 32 boroughs plus the City of London are showing rising case numbers, 25 of the boroughs are above the national average and 15 above 200 per 100,000 and 4 above 300 per 100,000.

Yup.  

Most of those infections you're looking at for London were caused when it was in lockdown.   There's only 3 days of non-provisional data from after it went to T2 - and very few of those cases will have been caused after lockdown ended.

Perhaps if the problem was non-compliance with the rules, it won't actually get that much worse (exponentially speaking) with the end of lockdown...

Thing is, the shops are open now. In London that means people are getting on public transport to get to the shops and that’s going to be a vector. Same with pubs and restaurants.  

A week ago it was just east London, parts of Kent and a bit of Essex that looked problematic. Now it's almost all* of London, almost all of Kent and most of Essex, as well as odd towns throughout the wider south east that have suffered outbreaks that are then spreading. Some bits of Kent and Essex are just as bad as London but the overall South East line on Wintertree's graph is only a small rise because it's brought down by areas which hadn't caught up with the rise (at least by 05/12.)**

You can see the virus gradually spreading throughout the south east quite clearly if you follow developments on the dashboard interactive map fairly closely. The dynamics of the spread would make an interesting study.

I wouldn't get too optimistic about Manchester to be honest. Although Wintertree's graph shows the North West region falling up to 05/12, you can see in the incomplete dashboard data that it has already flattened out (the incomplete most recent data always needs to be caveated, but it can only ever go up, not down, and now that testing timescales are a bit better the significant effect is usually only 2-3 days long rather than five.) If you look at Greater Manchester specifically the behaviour is similar to the wider region and in fact Manchester, Stockport and Bury have all suffered small rises in infection rate over the last week (they still look like tiny bumps and the trend is flat, but it's obvious things would be close to tipping upwards if there was a relaxation and this will only get more likely as infections spread slowly up from the south east if it isn't brought in check.)

* There is an area of London that seems to be holding out at low infection rates but is now circled by a sea of blue on the map. This is the western half of the centre and then a ribbon heading west-southwest from there. I think these are all very affluent areas and most people are probably able to isolate easily and work from home.

** A perfect example of this is Southampton, which I follow closely as my in-laws live there. Until very recently it was a lockdown success story with rates having dropped very effectively from over 200 to nearly 60, and the seven day average up until 05/12 still showing only a very small rise (hence compensating for areas like Kent in Wintertree's south east graph.) However, if you look at the very most recent data you can see, even though the data is incomplete, it has now clearly turned upwards more sharply again (unsurprisingly soon after similar in Portsmouth and East Hampshire.)

These are two plots of the exponential rate constant by region.  

I think the downtick on the right hand side of the plots (which would indicate things getting better) is a result of the residual weekend sampling effect in the data, and that in a few days everything will continue pointing up.

My interpretation of this is that

• Exponential (i.e. doubling times, not absolute case rates) rates were broadly similar in all regions at the start of lockdown
• Lockdown sees a decrease in rates everywhere (rate constant becoming more -ve meaning faster halving times), with faster decreases outside the south east.
• Then that downwards trend in rate constant reversed in all regions -  faster and earlier in the south east which has decisively returned to growth (positive rate constant, doubling time).  
• I think the the other regions all having increasing rate constants (with the last few days droop being the residual weekend sampling) so will be returning to growth soon enough.  

The image plot of the rate constants makes the bifurcation between the south east and the regions quite stark.

I struggle to think what we should expect to reach an equilibrium, if anything, at a given set of restrictions and environmental conditions. Do we think it's the exponential constant? That would have been my intuition but the behaviour of the pandemic has not exactly followed obvious patterns. Interesting to see what happens to these graphs in another week.

Edit to add, those latest two graphs are a really good example of how presenting the same data in two ways can lead to two completely different conclusions. Looking at the left hand graph (ignoring the far right hand side droop as a data artefact) everything is bad and everyone should look forward to lockdown 3, looking at the right hand graph most of the country is still falling and can look forward to dropping a Tier.  

Yes, I think the rate constant is the best place to look for a stable value indicating we're at an equilibrium - or R if you translate it to that.  Any such equilibrium is unstable I think however.

Unstable because compliance with rules decreases with time from the big messaging push around the change, and there's a lag of perhaps 1.5 to 2 typical doubling or halving time between a policy change being made, enacted and then working through the lags of infection, detection, hospitalisation, reporting, analysis, discussion and new policy creation.   If you look at it as a servo loop, as long as control is reactive it's always highly sub-optimal and unstable.  Looking at tends in the rate constant is another source of earlier information to beat some of the latency.

Instead, control should be predictive and there's enough insight and information to have a good go at that - but it would require a very different approach, much more like that in NZ, and for a whole host of reasons I think it would be exceptionally hard to pull the UK back to that.  Stable control would I think be better for the economy as well as for healthcare and infections/deaths.  Arguably that boat sailed in March, but I think we had another shot at it in September.

The next best thing is to work on reducing the latencies in the feedback loop.  One way of doing this is to have pre-determined action points for regions based on measurable, so that the policy forming time is removed from the control loop.  I've had high hopes this was happening a couple of times but then there never seems to be a coherent follow through.  

Bravo! This is a great graph, and I think a really important one which deserves a much bigger audience. The coherence between the English regions is striking, especially latterly. The change of slope in the middle of lockdown especially so: people relaxing as the news reports start to look good, or something else? If it were possible it would be really instructive to have the other countries of the UK on this graph too, given the different timings and nature of devolved Covid management measures.

I'm not sure what you mean by the red and green colours. What are 'rising' and 'falling' regions? If you mean regions with currently rising case rates (ie those above the mid line at the far right hand side of the graph), then you've either got the colours or the caption the wrong way round. Probably you mean something else and I just haven't read the text carefully enough, but either way, remember that some people struggle to distinguish red and green.

I'm asking a lot, but it would also be good to know which line was in which tier - might that be a better use of line colour in this graph? Have the lines change colour as they enter particular management regimes? Could get too fussy though. Stacked graphs grouping regions with the same management history would be another option. I think this could be a really powerful means of developing understanding of the relative impact of management measures, messaging, external events and, in the future, vaccination.

Thanks for the comments.

The criteria for rising/falling are given in an earlier post [1] althoughI I didn't  explicitly link it up to these ones.  It's the direction of change in cases over the last 10 days of the data.

The colours are backwards - thanks.  That's what you get for putting together a new plot during Jr's Saturday morning episode of The Gummi Bears (we're approaching the end of S5 which I missed as a kid, so it's riveting stuff.). I'll fix it for the next one.

Yes, I'm interested in doing this - particularly colouring UTLAs within the regional stacked plots in [1] by measure.  It could be fussy as you say, but not all plots have to be beautiful, some just have to be useful and not for a wide audience.  The main barrier to this is the amount of time involved in putting together a useful database of control measures per UTLA over time.  The Wikipedia article is good enough now that I can try this at some point...  It's a bit of a nightmare though as the measures are sometimes granular to a UTLA level not regional, and the classification of measure before the second lockdown is not so trivial, nor is mapping them on to post lockdown levels.   English half terms are another one to include - I had them some time ago on the doubling time plots but it got too confused.  The new plot might take them better.  It'd be interesting to know how regional the choice of the two weeks is - can I annotate a single week for each region?

I can do this with the other nations as well.

https://www.ukclimbing.com/forums/off_belay/friday_night_covid_plotting__3-...

I'm going to try and write up a coherent blog for LinkedIn for this Tuesday, my first attempt at "going viral" as the social media people call it.  We'll see what interest that nets.

Excellent news. If you think your blog post would benefit from a non-specialist proof read I'm sure there are plenty of people on here, myself included, who'd be more than happy to oblige.

News today sheds some light on what might be driving the different behaviour in the South East [1]:

"A new variant of coronavirus has been identified which may be associated with faster spread, MPs have been told. [...] We've currently identified over 1,000 cases with this variant predominantly in the South of England although cases have been identified in nearly 60 different local authority areas."

I've revived my map plots for this - here the colour indicates not the cases/day or cases/100k/day but if cases are falling (R<1) or growing (R>1) and the trajectory of R (worsening, improving).  Geographically, the tip over into growth started in Kent and is spreading out from there.  That's certainly feels compatible with claims of a faster spreading variant of the virus with a geographical origin into south east. 

[1] https://www.bbc.co.uk/news/health-55308211

Heat maps of infection by age have gone up on the dashboard tonight down to LA level.

Whitty says they don't actually know if this variant is causing the spread, simply that it has more mutations than most variants seen and that it is more prevalent where rates are rising in the south east. They don't know yet whether this is cause or effect. 

I was looking at northern Kent in the last two weeks of lockdown and something odd appeared to be going on their as infection rates were still significantly increasing...I initially put it down to non compliance but it then rippled out, I know this could a effect of non compliance in the area as people move in and out of it but this mutation does make you think.... 

Wintertree, whoever you are, you really are a brick (said in my best Enid Blyton voice). Your efforts this year have been invaluable. Thanks. Poster of the year. 

Thanks.  

Armchair Expert and Self Appointed Expert in Absolutely ****ng Everything - qualifications awarded and proudly received via this forum.

And in reply to Si dH:

Indeed.  It's tempting to speculate, but it is just that - speculation.  Still, this presents itself at an opportune moment; I would be interested to see the time/geographic spread off the samples from the new strain visualised; if it started in Medway about a month ago, perhaps associated with a hospital outbreak, it could explain an awful lot.

I actually forced myself to listen to the press conference today.  By gods they let some inane questions through - what a waste of access and air time asking a political appointment to critics the government.  Ask hime something that can get a useful answer FFS.  It reminds me of that pillock asking Elon Musk what he was going to do about toilets for people on Earth a few years back.

My stats tell me live case numbers have been increasing since saturday on my daily calcs, and the 7 day count is also now showing an increase. Until then they'd been declining (but with a steadily slowing rate of decline), so the data looks correct.  We're where we were on the 21st October, just under 2 weeks before the November 5th lockdown and probably the number of cases at which the lockdown decision was taken.  We do, however, now have different tiering and levels of prevalence in different parts of the country now.  New cases as a percentage of the total are also increasing now.
 

It states in the article:

"The changes or mutations involve the spike protein of the virus - the part that helps it infect cells, and the target Covid vaccines are designed around."

Does anyone have any educated speculation about what this actually means? Is this saying that this mutated spike protein gives the virus more chance to gain a foothold in a new host?

The lack of movies in the UKC gallery has given me incentive to produce a better static visualisation.

This is measuring the exponential rate constant for each UTLA with the same methods as plot9a above.  Each UTLA is then assigned R<1 or R>1 based on the sign of that rate constant, for every day.  The detail plot shows when R was more or less than 1 for each UTLA, ranked vertically by when they tipped over in to growth (R>1).  This date is used to assign a colour code to the map regions (Map summary) and is plotted against distance from Medway (right).

The distance measurement isn't precise as it could be...

I think your dates might be a bit susceptible to bumps in data? Are there are some UTLAs for which r fell temporarily back below 1 after turning yellow on your detailed graph, but which you haven't shown?

Eg, you show Halton being r>1 since the end of November.  The dashboard data shows Halton's minimum 7 day rate at 112 per 100k on 3rd of December (ie, aligning your turning point), but it since went up and back down slightly and is now still only 114 per 100k, ie, basically the same as it was on 3rd December. (It does look like it is about to go up in the next couple of days though.)

Thanks for the detailed dig in to the plots - as ever!

Yes, I think anything where the measured change in R lands the last 4-5 days on the right hand side of the detail plot is quite subject to noise.  When measured using data from 1 days ago, the code I've got gives a +ve rate of change in cases for Halton over the last 5 days of data, today's download still has a -ve (but rising) rate of change.  

What I should have done is mark everything on the map plot after 12-03 as "provisional" and likewise on the "Detail" plot.

I'm always in two minds about how to handle the leading edge of the data - don't show it or shot it and qualify it as provisional.  Generally the day-on-day variation given by the changes in the noise gives an indication of the range of possibilities for where it's going...  It's right on the edge given that this variation is happening with the weekend fix and some polynomial filtering.  

I've put recent daily case numbers from the data I use below for Halton, with and without my de-weekending.  The case numbers in the most recent 10 days are genuinely all over the place, so any measurement of when R switches over is going to be very twitchy to new data points.  I use a polynomial filter to measure it.

What I really need to do is quantify the uncertainty on measurements of doubling times and rates of change and so on, and include this on the plots - but that's a lot more work; and I give myself half an hour a day for the plots...  Also how to accurately estimate the noise on the data and propagate that through to measurements is a job in itself...  For now I should stick to labelling certain points as "provisional"...

Edit: Looking at this, I think my weekend filter can be improved.

I think your graph charting distance from Medway to the date of change from R <1 to >1 is quite compelling. Perhaps all the more so when you consider all the human movements that are semi-independent of distance (ie motorways vs b-roads). I wonder if there are any economic/geographic data available on human movements between English regions that might provide a more pertinent measure than simple distance. Anyone know?

In any case the new strain certainly looks like a plausible explanation for the change in slope of exponential rate constant in the middle of lockdown (which you previously demonstrated, and which I had found perplexing).

A few days ago you mentioned that you were going to publish a blog post - care to provide UKC with a link?

It’s an attractive explanation for sure, but what I’ve shown is a long way from evidence.  It’s hints and breadcrumbs really.  

> I wonder if there are any economic/geographic data available on human movements between English regions that might provide a more pertinent measure than simple distance. Anyone know?

Google have this stuff from phone tracking and if they happened on my analysis and wanted to run a basic epidemic model using an epicentre from my plot and compare the timings, I’d be interested to see it...  The state will also have an approximation from APNR but the police are very protective of that data, presumably as if it was allowed to be used in ways risking a fuss over privacy it might get taken away from them.

The proof is likely in the sequencing data; the UK doesn’t seem very open in their approach to publishing that?  

I can share it by mail if anyone wants to PM me.

For what it's worth, infections rose higher in Swale (primarily Isle of Sheppey) earlier in October than happened in Medway. There was a parallel rise in Margate area as well at that time. Medway only overtook once they were already all in a bad place, so I don't think it was the source of the outbreak in Kent. I doubt this really changes the price of eggs though.

Medway has the convenient property of being in the middle of the “interesting” areas.  If there was an epicentre to a more virulent strain, I don’t have the resolution to find it.  I should really pick a “fair” location for my distance plot.  All suggestions welcomed.

Let me know if it bugs you, I enjoy digging but I don't want to be overly critical given you do this in your spare time!

I think it's pretty helpful for seeing trends at regional or national level. At LA you will run in to trouble because the weekend droop behaviour is nowhere near as consistent. Some LAs seem to just tend to have one day of low reporting each week or none at all. I guess it might depend how efficient local services are at delivering specimens to labs. It's also difficult to see for the noise when rates are low.

Quite the opposite thanks.  The intent is to both encourage and help other people to dig in to the data for themselves and not take the news at face value - although I know you were doing that anyway to more detail than me - and I feel more confident in the work when the methods or and/or results get challenged.   A lot of it happens during the half hour of downtime on an evening of “Wintertree University” (schooling Jr in 1980s cartoons) which to be fair isn’t the best work environment...

I have a suspicion the data would be regularised if they used the date from the sample not the date it arrived at the labs from its sojourn through the postal system.  I don’t actually know that this is what’s happening...

Agreed - it's correlation only of course, and not a very strong correlation at that, but it's very plausible (based on a thinking layman's understanding of how biology and human behaviour works at least), and there are no other very convincing candidates just now. But I wish you would incorporate the Welsh and Scottish data (even if T-in-E might resent Scotland being treated as an equivalent of an English region), so that a more considered examination of the role of management measures were possible.

Once again you show that you have already thought about all my suggestions long before I've made them - I am steadily moving toward the position that I have nothing to offer until I see you actually making a mistake!

Regards.

I had some thoughts on de-weekending, based on what I said on the other thread about standard deviation. It starts with your deconvolved data.

If the objective is to make a graph that looks "smooth" then you could define a fitness function which is the mean of the square of the error between the assumed daily cases and the (de-lagged) moving average.

I think you could then have 7 "de-weekending" filters, one for each day but they all work in the same way and are constrained. Each day is weighted so that filtered_day = w1*day(i-1)+w2*day(i-2)... for as far back as the history needs to go, eg you could say no result takes more than a week to report limits the filter length to 7, with 7 weights to determine.

You need 7 of these filters one for each day, each with its own set of weights, so with the previous assumption, the filter is a 7x7 matrix that has the constraint that the weights must all add up to 1 for any given day. So for example you could assume that Tuesday's figure is made of 30% of Monday's cases, 60% of Sunday's cases and 10% of Saturday. All the other entries for Tuesday have to be zero, and 70% of Monday is left to allocate to other days - so the 7x7 matrix is constrained and doesn't really have 49 unknowns.

Then subject to that constraint, and as a first estimate based on what you already know from your existing filter, have a stab at making the matrix. 

The tricky bit then would be to use the whole dataset of reported cases and the fitness function I made up above to feed this into a statistics optimisation package in Python, R, Octave, Mathematica, Prolog....whatever and see if it can find a matrix which minimises the fitness function. The matrix has lots of unknowns of course but there might be enough points in the data set for the optimiser to have a good stab if the starting point is credible.

You might be able to add some extra constraints or heuristics from your own experience to cut down the number of unknowns, or deliberately say that the numbers have to be 0, 25%, 50%, 75% or 100%. I don't know, maybe Friday is easy and all the results are Thursday's cases so a whole load of matrix goes to zero.

I don't know how well this approach would work, it might suffer from changing reporting times as the infrastructure has evolved over time so the weights are moving.

I suspect it's too complicated for its own good with insufficient data to make anything useful (too many unknows with insufficient training data). Might be an interesting experiment.

That’s the key - the filter kernel depends on the day of the week.  I started looking in to it; the next problem is that the scale of this effect seems to depend non linearly on the number of cases so I’d have to get a robust study of that effect, and your 7x7 matrix becomes at least 7x7x2 describing a 2nd order polynomial for each filter point in terms of cases.  I’m then trying to fit nearly 100 data points which is a *lot* of degrees of freedom for the data depending on how you look at it (days or weeks of data).

The second problem is one I didn’t really understand but Si dH’s message sheds some light - the residuals for different weeks are clearly well structured by not always similar to each other meaning that this 7x7(x2) matrix changes with time as well as case count.  If different UTLAs have different patterns to the weekend effect (different weekend postal services?), the details will change as the high case regions shift around the country - as they do.  I suppose I could try and build a model at UTLA level then re-generate national level data from that but I’m not looking for a PhD chapter’s worth of work - nor am I convinced enough it’d work to sink the effort...  I think SAGE could justify having a Bayesian person dive in to this and do something.

It’s frustrating because a lot of this is an “artificial” problem and the residual noise is very limiting - not just to “smooth curves” as you mention but to making accurate measurements from the most recent week of data.  The better the data, the sooner policy makers can spot and address a problem.  I think there’s an east fix to some of it which is to use the sample data written on postal samples not the date they enter the labs.... ??

I think this is the important bit. For those people using the data for real to drive policy, I would expect that they have a significantly more detailed information at their disposal than the reduced data on the govt website. As you say, postal data or fact like lab X working only 5 days a week due to maintenance or lab Y has limited capacity and when >10000 tests arrive, they ship the excess to another lab introducing another delay. This specific detail could be added to the original data and corrected better at source. 

The summarised / totalised values are full of special cases which blur underlying time varying patterns. I agree that brute forcing my crude linear model with loads of degrees of freedom to bodge over the non-linear patterns will not produce a robust solution.

I think it is laudable of the govt to publish as much raw data as they have in near realtime for everyone to look at as they see fit. This degree of openness is good, I just wish it were shared by the politicians!

Yes, this is fantastic.  It's a great resource for people interested in events, and for various professional organisations studying or communicating those events.  It's a really positive step towards an open data approach and I hope it continues with other events.

I'll try and work one up tomorrow.  Can you help out?  If you can put together some data on the control measures of the form (start date, measure, end date) I can use that to add annotations to the curves.  I'd also take that data on a level of English regions.  With it, I can try colouring the curves.

Happy to do my bit - I'll pm you later.  

The reasoning for the latest Tier allocations is just out.

https://www.gov.uk/government/speeches/review-of-local-restriction-tiers-17...

Might have known some of the information would be missing.

https://www.gov.uk/guidance/full-list-of-local-restriction-tiers-by-area

A good example of what looks like blatant contradiction to the stated cautious approach is Herefordshire, which is static on cases (tiny decrease) and next door to a welsh problem area but dropped to Tier 1 whilst surrounded by Tier 2. Another is East Suffolk...56% increase to just below national average. 

To be fair, l was looking the other day to see where if anywhere I thought  might be droppable to tier 1, and Herefordshire was top of the list (closely followed by Dorset if you exclude Poole and Bournemouth.) I agree about the Welsh border problem though. I'm sure the Government want to have some down as well as up to sell their system which is probably no bad thing to motivate people.

I'm slightly surprised but pleased they've decided to put several of the 'shires' into Tier 3 today. It would have been politically much easier to hold off longer yet but I think it's the right thing to do.

(Somewhere on a previous thread about a week or two ago I said that by this review date, without earlier action, we would be looking at more restrictions across most of the south east below a line from Portsmouth to Norwich via Northampton, excluding bits of Norfolk...it looks like I wasn't that far off.)

Wales looks really ugly with today's 11000 missing cases

Merthyr at over 1000 cases per 100k, that's the highest number I ever remember seeing for a local authority since the second wave began in early September.

The provisional leading edge of the data for England today is looking awful.

Crazy % increases all over the South East, loads of LAs >100%....highest I found was Braintree at 174%!

Yep, new data today is really bleak. Looks like we are back at the pre-lockdown-2 infection rate peak now.

London, South East and East of England all look to be on really steep upward curves but the rest of the England regions are now all rising too on, a regional average basis. I think it's just spreading out from London too fast.

Lockdown 3 before Christmas?

It is all very worrying.

I get what you say about Herefordshire if they are looking to drop some areas to Tier 1.  What I don't get is how the decisions add up given other factors and what the government say about caution. 

What are your odds that the Herefordshire area won't see a 30% weekly rise before the new year?

It's probably worth linking the Guardian map here as well, as people can click on the map and see the increasingly complex patchwork of growth alongside much fewer areas of decline and how it aligns with Tiers.

https://www.theguardian.com/world/2020/dec/17/covid-cases-and-deaths-today-...

Note Herefordshire is showing as 6% weekly increase on the map after todays updated figures, so the slight decrease earlier today has already gone. Next door, Forest of Dean is now up 147% on last week (to 188 cases per hundred thousand).

It has to be fairly likely I think. The evidence says tier 1 is useless and the proximity to high risk areas will mean infection spreads in.

Yes I like the guardian map too, although it can be a bit buggy and slow to be updated. Today there seem to be more bugs than usual (ie LAs that say unknown data.)

I have to confess I was surprised twice - firstly because I really didn't believe that the idea of a downgrade to tier 1 was anything more than a unicorn or a fairy tale for children (the current tier 1 areas all have some element of geography on their side to help maintain their low rates, plus cynically having somewhere still in tier 1 means there isn't a nationwide minimum at tier 2 but why else would you bother because it does naff all for infection control) - and then secondly if you were moving places down why rural Dorset didn't get a look in - no more connected to the rest of the world than Herefordshire and very low rates.

This is very true, but it doesn't convey the sense of negativity I've got right now.

Last time we reached lockdown, we did so with a very slow  increase (in day-on-day % or exponential terms) thanks to the old T2/T3 levels in the higher case rate areas up North.

This time we're approaching it with a much higher growth (in day-on-day % or exponential terms), closer to that last seen in October.  Monday the 14th may yet see the highest daily cases recorded to date.  This rise started during lockdown so I've no great faith that T2 escalating to T3 is going to halt it, and there aren't many places left now that haven't tipped over in to growth.   Although it's not been discussed anywhere I've seen, hospital occupancy by Covid patients is rising again, and is higher than the levels that prompted the November lockdown - with more aggressive growth in cases - and is higher than any time since April 22nd.  Then there seems to be a demographic shift going on towards older ages in London over the last 8 days or so of data.

The critical question to me is what exactly "it" is - just bleed through of cases, or something worse?   

I'm hoping that after tomorrow I don't have to leave the property for the next 14 days, put it that way.   Having work schools and many work places closing down for a couple of weeks should help.

Edit: on the positive side, today’s surveillance report continues to show an almost total absence of influenza hospitalisations.  

Yes, I didn't really expect it to happen at all either, it's just that those were the two places I'd have picked if I had to when I looked at the data yesterday or the day before.

Yeah cases and more importantly admissions and hospital numbers are going up and there isn’t much headroom to play with. A nearly 4 week long lockdown lite only succeeded in resetting casesto where they were a month pre lockdown and hospital numbers to where they were at the start of lockdown. Another 4 weeks in January? At the start of Lockdown 2, I fully expected that. Then I got a bit less pessimistic and thought 50/50. Now I’d say it’s near certain. Possibly as early as after the 5 day relaxation.

Northern Ireland going into full lockdown for 6 weeks from Boxing day....hospitals there are running at 104% capacity. 

If behaviour of the virus was fundamentally unchanged from 6 weeks ago just pre lockdown, then the latest measures should be enough to get a hold of it in London. All the evidence from the north and midlands is that this was enough to bring cases down. If the tide is stemmed in London then the surrounding areas (especially with similar measures) would follow. If it isn't, they will keep going up.

However the rate of increase and spread now does suggest something has changed, especially that seen in Kent while under tier 3. The three possibilities I see are climate, behaviours or this mutation that has been found. The temperature hasn't actually changed loads since the start of lockdown 2 (and in general, Kent is the warmest and sunniest bit of the country, right?) so I'd be surprised if it was that but I suppose it's possible. If that really is the case then I think we need another lockdown asap, because it will only get worse yet. There should be evidence for scientists to look at here from the US where there is lots of testing in a country with huge temperature variations between different states at any one time. Admittedly they are generally a drier climate.

If it's behaviours then what is needed is better messaging to turn the dial - hopefully the changes of the last week will make a big difference to people's mindset. Anecdotally from news reports and a few individuals I know living down south, it does seem that people there have all dropped straight back in to tier 1 mentality after leaving lockdown because they couldn't comprehend being in a worse state post lockdown than they were told they were in pre lockdown. (I had to tell my in-laws, who have been generally very cautious throughout and are not stupid, that tier 2 meant they weren't allowed to go for a coffee with friends - which they had just done!)  However I have no proper evidence for this, just heresay. The problem is that a change in behaviours now will take a couple of weeks to feed through. And this behavioural theory wouldn't properly explain some of the rises we are now starting to see further north, unless this is all down to crowds rushing to the shops for Christmas, which seems unlikely because in London and SE it started while they were still shut.

If it's a new mutation that is more infectious, and assuming it is just as severe a disease, then I think we need lockdown 3 as soon as possible to bring down it's spread across the country. Hopefully there are some frantic studies going on to test this out right now.

All good comments.  No idea on the Kentish climate, I haven't been south of the Thames in many years now...  I think there's something going on and very little information is out there to go off when it comes to picking which option it is.

The news reports seem to be leaning in that direction today, regardless of the root cause.

All of the above?