COVID-19

[lpm]

f.ck.

I'll wait for our forum geniuses to assign into the proper date, but this is basically infections during the final week of lockdown. They should have fallen from 15,000, not gone up.

Looks like London is the culprit, plus South East, Midlands and Wales. The northern hotspots are still coming down like they should.

[sTeamTraen]

South Wales is on fire and the ICUs in some parts are apparently close to full. Welsh secondary schools are going online from next week.

[bagpuss]

This must surely be something to do with the way that the joinzoe app people calculate estimated cases, because otherwise something exceedingly weird happened last weekend in my area.

Screenshot_20201210-173911_COVID Symptom Study.jpg (275.63 KiB) Viewed 3502 times

[mediocrity511]

So now London is hit, there's an admission that maybe transmission does occur in secondary schools and mass testing available for their pupils and the surrounding South East. But when Hull was worried about essential services collapsing because so many kids were out with covid and self isolating, there was nothing.

Wales have put their secondary schools online for the last week of term.

[sTeamTraen]

This must surely be something to do with the way that the joinzoe app people calculate estimated cases, because otherwise something exceedingly weird happened last weekend in my area.

"581, down by 131" seems to make no sense (whether the spike is real or not). Unless it means "the number of cases on this one day today is 581 and a week ago on the same day it was 714", which would be stupid. If cases means "new cases" then they need to do a 7- or 14-day moving average, and if it means "people known to be actively spreading the lurgy" then they need to take into account the area under the line in some way (e.g., assuming a new case remains infectious for X days). Otherwise it's just factoid p.rn.

[Gfamily]

We had the same about a month ago

Screenshot_20201210-181958.jpg (92 KiB) Viewed 3479 times

[KAJ]

Last 8 weeks cases by specimen date. The dashed line is the quadratic fit without the weekday factor to better demonstrate trend. The empty circles are flagged at gov.uk as "incomplete" and are zero weighted in the regression - they are only likely to increase.

xy.png (15.47 KiB) Viewed 3455 times

The fit is really very good, R-sq > 90%, showing that the regression fits the recent data behaviour very well.

Code: Select all

Weighted Residuals:
     Min       1Q   Median       3Q      Max 
-0.13202 -0.05545  0.00000  0.04867  0.15552 

Coefficients:
               Estimate Std. Error t value Pr(>|t|)    
(Intercept)     9.53110    0.03047 312.757  < 2e-16 ***
poly(date, 2)1 -1.15045    0.09882 -11.642 9.93e-15 ***
poly(date, 2)2 -1.40361    0.10206 -13.752  < 2e-16 ***
dayMon          0.53560    0.04155  12.889 3.45e-16 ***
dayTue          0.43555    0.04153  10.489 2.65e-13 ***
dayWed          0.39649    0.04284   9.255 1.07e-11 ***
dayThu          0.31834    0.04279   7.439 3.45e-09 ***
dayFri          0.28716    0.04276   6.716 3.71e-08 ***
daySat          0.02593    0.04274   0.607    0.547    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.07994 on 42 degrees of freedom
Multiple R-squared:  0.9221,	Adjusted R-squared:  0.9072 
F-statistic: 62.13 on 8 and 42 DF,  p-value: < 2.2e-16

That all but the last (and most incomplete) of the most recent 7 points are above the regression line suggests to me that the behaviour has changed. This suggestion is stronger in the residuals plot.

res.png (8.98 KiB) Viewed 3455 times

I think I'm beginning to see a similar picture in hospital admissions, although the day of week effect is much less pronounced.

admits.png (15.08 KiB) Viewed 3455 times

Code: Select all

Residuals:
      Min        1Q    Median        3Q       Max 
-0.108057 -0.031836 -0.003901  0.030935  0.106776 

Coefficients:
               Estimate Std. Error t value Pr(>|t|)    
(Intercept)     7.20674    0.01896 380.168  < 2e-16 ***
poly(date, 2)1  1.06736    0.05402  19.760  < 2e-16 ***
poly(date, 2)2 -1.20579    0.05361 -22.493  < 2e-16 ***
dayMon          0.03660    0.02680   1.365  0.17862    
dayTue          0.05015    0.02682   1.870  0.06772 .  
dayWed          0.08073    0.02683   3.008  0.00421 ** 
dayThu          0.05632    0.02686   2.097  0.04145 *  
dayFri         -0.07174    0.02682  -2.675  0.01025 *  
daySat         -0.06252    0.02680  -2.332  0.02402 *  
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.0536 on 47 degrees of freedom
Multiple R-squared:  0.9551,	Adjusted R-squared:  0.9474 
F-statistic: 124.9 on 8 and 47 DF,  p-value: < 2.2e-16

[shpalman]

993 deaths reported in Italy today. The previous highest was 971, on 28 March. It seems to have all gone to sh.t in the last six weeks, far steeper than most comparable countries.

Have the "We want our freedumb" demonstrations of a few weeks ago died down yet?

The number for today is a bit lower, but still high: 814. Lower than last Friday though.

Today is also a Thursday, and today also has more deaths, 887, than any other since 993 last Thursday.

Numbers of cases and deaths registered would have been a bit low earlier this week though because the 7th is a holiday in Milan and the 8th is a holiday in the whole country (and people probably took Monday off anyway to make a "bridge"). The trend is still downwards and should drop below 600 per day next week.

In terms of new cases per 100,000 per week Italy and Lombardy are dropping sharply day by day, I currently get something like 200 for Italy and 165 for Lombardy. The most recent official number for England is 153 and (slowly) rising. Don't look at Wales.

[sTeamTraen]

Don't look at Wales.

Fark, some areas are approaching 0.1% per day. That's getting into Dakota territory.

[shpalman]

So now London is hit, there's an admission that maybe transmission does occur in secondary schools and mass testing available for their pupils and the surrounding South East. But when Hull was worried about essential services collapsing because so many kids were out with covid and self isolating, there was nothing.

Wales have put their secondary schools online for the last week of term.

London is losing its sh.t about having 200/100,000/week while Hull was at 770 and they didn't care.

But at the same time they're very relaxed about putting London in a higher tier even though cases are rising, because it might lead to mild inconvenience.

[shpalman]

WHO accused of conspiring with Italy to remove damning Covid report

The report said Italy’s pandemic plan had not been updated since 2006 and that, due to being unprepared, the initial response from hospitals was “improvised, chaotic and creative”. It took time for formal guidance to become available, the report added.

The document was allegedly removed at the request of Ranieri Guerra, the WHO’s assistant director general for strategic initiatives. Guerra was the director general for preventive health at the Italian health ministry between 2014 and late 2017, and was therefore responsible for updating the pandemic plan as per new guidelines laid out by the WHO and European Centre for Disease Prevention and Control (ECDC). Guerra is among the scientists on the Italian government’s Covid-19 taskforce.

[mediocrity511]

So now London is hit, there's an admission that maybe transmission does occur in secondary schools and mass testing available for their pupils and the surrounding South East. But when Hull was worried about essential services collapsing because so many kids were out with covid and self isolating, there was nothing.

Wales have put their secondary schools online for the last week of term.

London is losing its sh.t about having 200/100,000/week while Hull was at 770 and they didn't care.

But at the same time they're very relaxed about putting London in a higher tier even though cases are rising, because it might lead to mild inconvenience.

Laptop allocations for isolating children got slashed by 80% at the beginning of the half term holidays. DfE have just announced they are referring to the original allocations offered to schools. Just as London schools are the ones who will be ordering them. Meanwhile deprived children in Hull or Manchester missed out.

[shpalman]

993 deaths reported in Italy today. The previous highest was 971, on 28 March. It seems to have all gone to sh.t in the last six weeks, far steeper than most comparable countries.

Have the "We want our freedumb" demonstrations of a few weeks ago died down yet?

The number for today is a bit lower, but still high: 814. Lower than last Friday though.

Today is also a Thursday, and today also has more deaths, 887, than any other since 993 last Thursday.

And today, being the day after yesterday, has a high but lower number of deaths, just like last week.

Numbers of cases and deaths registered would have been a bit low earlier this week though because the 7th is a holiday in Milan and the 8th is a holiday in the whole country (and people probably took Monday off anyway to make a "bridge"). The trend is still downwards and should drop below 600 per day next week.

In terms of new cases per 100,000 per week Italy and Lombardy are dropping sharply day by day, I currently get something like 200 for Italy and 165 for Lombardy. The most recent official number for England is 153 and (slowly) rising. Don't look at Wales.

Cases are decreasing with a halving time of a week.

Today's UK numbers haven't been posted yet but I expect them to be oh f.ck.

[shpalman]

Oh look today's new cases in the UK are higher than they were yesterday. KAJ, is it normal for Friday to be worse than Thursday?

[KAJ]

Oh look today's new cases in the UK are higher than they were yesterday. KAJ, is it normal for Friday to be worse than Thursday?

Up to date cases by specimen date (latest specimens for which we have results are dated yesterday).

xy.png (14.7 KiB) Viewed 3250 times

Results of regression:

Code: Select all

Call:
lm(formula = as.formula(aform), data = aDF, weights = wts)

Weighted Residuals:
     Min       1Q   Median       3Q      Max 
-0.17615 -0.04949  0.00000  0.06329  0.26345 

Coefficients:
               Estimate Std. Error t value Pr(>|t|)    
(Intercept)     9.51232    0.04247 223.971  < 2e-16 ***
poly(date, 2)1 -1.39415    0.13687 -10.186 6.46e-13 ***
poly(date, 2)2 -0.92649    0.14136  -6.554 6.33e-08 ***
dayMon          0.54239    0.05919   9.163 1.43e-11 ***
dayTue          0.43620    0.05922   7.365 4.38e-09 ***
dayWed          0.41806    0.05927   7.054 1.22e-08 ***
dayThu          0.35017    0.05933   5.902 5.50e-07 ***
dayFri          0.31733    0.05752   5.517 1.96e-06 ***
daySat          0.06539    0.05755   1.136    0.262    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.1107 on 42 degrees of freedom
Multiple R-squared:  0.8614,	Adjusted R-squared:  0.8351 
F-statistic: 32.64 on 8 and 42 DF,  p-value: 1.246e-15

The 'day' coefficients are the difference in natural log from Sunday. Their 'Pr and Signif codes are testing difference from Sunday. Friday is fitted a little lower than Thursday, but only about 0.03 Loge units which is substantially less than the standard error of the'day' coefficients.

But I guess your question refers to cases by publish date, rather than specimen date. That looks like this:

xy2.png (15.03 KiB) Viewed 3250 times

... and here are the regression results:

Code: Select all

Weighted Residuals:
     Min       1Q   Median       3Q      Max 
-0.36637 -0.07697  0.00000  0.05016  0.33031 

Coefficients:
               Estimate Std. Error t value Pr(>|t|)    
(Intercept)     9.82033    0.04937 198.921  < 2e-16 ***
poly(date, 2)1 -1.34394    0.17232  -7.799 1.07e-09 ***
poly(date, 2)2 -1.02582    0.17797  -5.764 8.68e-07 ***
dayMon         -0.09078    0.07246  -1.253    0.217    
dayTue         -0.09345    0.07244  -1.290    0.204    
dayWed          0.09810    0.07242   1.355    0.183    
dayThu          0.11748    0.07241   1.622    0.112    
dayFri          0.07868    0.07241   1.087    0.283    
daySat          0.05846    0.06971   0.839    0.406    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.1394 on 42 degrees of freedom
Multiple R-squared:  0.6702,	Adjusted R-squared:  0.6074 
F-statistic: 10.67 on 8 and 42 DF,  p-value: 4.983e-08

There is no substantial evidence of a weekday effect, and the R-sq has fallen from over 80% to about 60%. Introducing the largely unpredictable lag between specimens and result publication has muddied the waters.

[Bird on a Fire]

What's the best way to use these regressions to answer a question like "Is it normal for a Thursday to be higher than a Friday?"

I'm thinking some kind of post-hoc test comparing the day-of-week effects, while controlling for week number, would answer the question?

Or could we just regress Fridays and Thursdays for each week, and see whether the most recent week is within the confidence intervals.

Or we could use a Bayesian approach to ask how likely it is that a Friday would be higher than a Thursday, given the data we already have.

I'm not proposing anyone actually does it, but I can't immediately settle on a tight quantitative approach to shpalman's question.

(I too am delighted that KAJ got R running on his Chromebook. That line for the day-of-week effect looks super funky, even if the Tufte in me thinks it should be a series of points )

[shpalman]

Yeah I'm specifically talking about whether it's more likely for a bunch of cases to finally get around to being reported on a certain day of the week rather than another one. Numbers of tests done each day is a different issue.

[KAJ]

What's the best way to use these regressions to answer a question like "Is it normal for a Thursday to be higher than a Friday?"

I'm thinking some kind of post-hoc test comparing the day-of-week effects, while controlling for week number, would answer the question?

Isn't the linear model I fitted just that? "comparing the day-of-week effects, while controlling for week number" where the control for week number is via a quadratic - which fits pretty well.

Or could we just regress Fridays and Thursdays for each week, and see whether the most recent week is within the confidence intervals.

Before I had R, and was using Google sheets, I calculated the (log) difference between each day and the lagging 7 day moving average. I averaged that difference for each weekday to estimate the weekday effect. I then subtracted the estimated weekday effect from the raw data to give a "corrected" count, which I regressed. I added the estimated weekday effect back to regression fits/predictions. I always regarded that as a poor mans substitute for the "proper" approach of simultaneously fitting models for time and for weekday. The choice of a quadratic is arbitrary and very much open to criticism. I justify it by goodness of fit (which has worsened in recent days) and that I am using the regression to describe the data, not to deduce mechanisms or make predictions.

Overall, I don't think there is much mileage in trying to fit cases by publish date. As I've said before, that implies trying to fit the variable and largely unpredictable delay between specimen date and reporting date.

(I too am delighted that KAJ got R running on his Chromebook. That line for the day-of-week effect looks super funky, even if the Tufte in me thinks it should be a series of points )

Yes, this data is inherently discrete - individual days with no intervening values - but in this case lines are much clearer than points would be, and I don't think anyone is going to interpolate.

[raven]

Wales have put their secondary schools online for the last week of term.

I think English schools are getting to close a measly day early for Christmas. Apparently so headteachers don't have to do contact tracing on Christmas day itself. Just right up to it. Because they don't need a break from the stress at all...

(My sister works in a primary. Staff were all hoping they'd shut a week early, because a) it would seem quite sensible given the circumstances and b)they suspect a fair few parents will take their kids out that last week anyway so that they can isolate before seeing Granny at Christmas. No such luck for the poor staff who will have to go in to the bitter end.)

[mediocrity511]

Wales have put their secondary schools online for the last week of term.

I think English schools are getting to close a measly day early for Christmas. Apparently so headteachers don't have to do contact tracing on Christmas day itself. Just right up to it. Because they don't need a break from the stress at all...

(My sister works in a primary. Staff were all hoping they'd shut a week early, because a) it would seem quite sensible given the circumstances and b)they suspect a fair few parents will take their kids out that last week anyway so that they can isolate before seeing Granny at Christmas. No such luck for the poor staff who will have to go in to the bitter end.)

What English schools have been offered is the ability to move an INSET day at short notice. So staff still have to be in and it means they need to move a day from elsewhere in the calendar.

I was tempted to take Miniocrity out to avoid being made to isolate over Christmas. But instead, we've negotiated an authorised absence for the first week back, to try and protect Mr Mediocrity from any rise caused by Christmas mixing.

Also schools that tried to move to online learning for the last week have been threatened with court action by the DfE.

[bagpuss]

Wales have put their secondary schools online for the last week of term.

I think English schools are getting to close a measly day early for Christmas. Apparently so headteachers don't have to do contact tracing on Christmas day itself. Just right up to it. Because they don't need a break from the stress at all...

(My sister works in a primary. Staff were all hoping they'd shut a week early, because a) it would seem quite sensible given the circumstances and b)they suspect a fair few parents will take their kids out that last week anyway so that they can isolate before seeing Granny at Christmas. No such luck for the poor staff who will have to go in to the bitter end.)

What English schools have been offered is the ability to move an INSET day at short notice. So staff still have to be in and it means they need to move a day from elsewhere in the calendar.

I was tempted to take Miniocrity out to avoid being made to isolate over Christmas. But instead, we've negotiated an authorised absence for the first week back, to try and protect Mr Mediocrity from any rise caused by Christmas mixing.

Also schools that tried to move to online learning for the last week have been threatened with court action by the DfE.

I'm not entirely sure that it was an "offer" exactly. The bagkitten's school initially said that they had decided not to take up this offer and school would be open on Friday as planned. But then, the next day, announced that the last day of term would be a homeschool day and apologised for the short notice but stuff announced late, yada yada. Now it may just be that the (excellent) head of the linked infant school persuaded the (slightly rubbish and out of his depth) head of juniors that this would be a good idea, but it read to me as if pressure had been applied from somewhere.

[KAJ]

What's the best way to use these regressions to answer a question like "Is it normal for a Thursday to be higher than a Friday?"

I'm thinking some kind of post-hoc test comparing the day-of-week effects, while controlling for week number, would answer the question?

Isn't the linear model I fitted just that? "comparing the day-of-week effects, while controlling for week number" where the control for week number is via a quadratic - which fits pretty well.

Or could we just regress Fridays and Thursdays for each week, and see whether the most recent week is within the confidence intervals.

Before I had R, and was using Google sheets, I calculated the (log) difference between each day and the lagging 7 day moving average. I averaged that difference for each weekday to estimate the weekday effect. I then subtracted the estimated weekday effect from the raw data to give a "corrected" count, which I regressed. I added the estimated weekday effect back to regression fits/predictions. I always regarded that as a poor mans substitute for the "proper" approach of simultaneously fitting models for time and for weekday. The choice of a quadratic is arbitrary and very much open to criticism. I justify it by goodness of fit (which has worsened in recent days) and that I am using the regression to describe the data, not to deduce mechanisms or make predictions.

Overall, I don't think there is much mileage in trying to fit cases by publish date. As I've said before, that implies trying to fit the variable and largely unpredictable delay between specimen date and reporting date.

(I too am delighted that KAJ got R running on his Chromebook. That line for the day-of-week effect looks super funky, even if the Tufte in me thinks it should be a series of points )

Yes, this data is inherently discrete - individual days with no intervening values - but in this case lines are much clearer than points would be, and I don't think anyone is going to interpolate.

This got me thinking and I thought:
We have a finite population so inferential statistics (significance tests, confidence intervals etc.) just aren't appropriate - we don't need to infer the properties of a population from a sample, we can directly measure those properties. Which leads me to this table, based on yesterdays cases by publication date data.

Code: Select all

		relative to yesterday		
Day		<	=	> 
Sunday		24	7	18
Monday		32	6	11
Tuesday		19	6	24
Wednesday	11	8	30
Thursday	16	6	27
Friday		18	5	27
Saturday	28	8	13

So Thursday was higher than the following Friday 18 times, equal 5 times, less than 27 times. I don't think it makes sense to deduce the properties of some underlying process, I would take a lot of convincing to accept that there was a constant underlying process.

[shpalman]

Ok I see what you mean that the day-of-the-week effect is much stronger in the cases-by-specimen-date data than the cases-by-date-reported data but cases are going up significantly.

[sTeamTraen]

Here are the percentage of cases (left) / deaths (right) announced per day of the week for the past 39 weeks (since which daily cases have not dropped below 300, except for 29 July when only 70 were reported as 70 because, I think, the figures for England were missing and were caught up the next day).

Code: Select all

Monday      12.29    8.38
Tuesday     13.36   19.15
Wednesday   15.27   18.15
Thursday    16.08   15.76
Friday      15.15   16.68
Saturday    14.23   14.13
Sunday      13.63    7.74

I don't know what the cutoff time is for reporting, so to what extent data reported at around 1600-1700 UK time on any given day includes events that took place that day versus the day before, but I imagine that the lower number of deaths reported on Sunday and Monday corresponds to fewer death certificates being written on Saturdays and Sundays.

Both of the figures for Saturdays are remarkably close to one-seventh.

[shpalman]

london-is-fucked.png (94.09 KiB) Viewed 3061 times

Scores of shoppers flocked to central London on the second-to-last weekend before Christmas.

Regent Street, one of the British capital’s busiest shopping streets, has been pedestrianised in the lead-up to Christmas to encourage more shopping and safe distances.

It comes as British health secretary, Matt Hancock, said it is “not inevitable” that London will be placed under tighter coronavirus restrictions ahead of next week’s review of the three-tier system in England.