COVID-19

[shpalman]

It made the national news that intensive care in the hospital in the centre of Como (the Valduce) is full, and the non-intensive covid beds are running out. (The place where I was teaching swing is right next to it; it's a small city, it still weirds me out a bit that there's space for a whole hospital in it.)

There's still a big out-of-town hospital though (the new St. Anna). There's also the old St. Anna just up the road from me which is where I went to get the 'flu vaccine last year and where they're doing drive-in covid tests but I'm not sure if there are any patients there anymore. There were suggestions to use is for isolation of light/asymptomatic covid cases, I'm not sure if they actually did anything.

lombardy.png (32.1 KiB) Viewed 3776 times

In the graph, total positives and isolated at home are on the right y-axis but hospitalized cases (further divided into intensive care and non-intensive care) are on the left y-axis.

[shpalman]

It's obvious in the first wave that lots of mild/asymptomatic cases were missed due to insufficient testing and there would have been a lot of people left at home who quickly went from "doing ok, don't need to go to hospital" to "too late, no point in going to hospital". Meanwhile in the first wave the ICU probably wasn't a great place to be since rapid prone intubation turns out to not be the best way to deal with SARS2 (as opposed to SARS1).

[shpalman]

Same thing but at a national level.

national.png (29.97 KiB) Viewed 3775 times

(which shows quite neatly that 1 in 20 positives are in hospital)

[shpalman]

The health commissioner of Calabria, Saverio Cotticelli, only just realized that it was his responsibility to organize his region's covid response plan.

he realized while he was being interviewed

[jimbob]

I've tried to extrapolate the exponential back to see if it coincides with the mid-September return to school, but it really kicked off at the end of October. (Whereas the bump you see before that is probably the end-of-August return from holidays which seems to have been kept under control during September.)

Based on the death or hospitalization rates we're probably picking up about 10 times more cases than during the first wave but yeah, hospitals are in trouble, and deaths are on course to be as bad as the first wave.

I consider it a big pile of f.cking b.llsh.t that we tried to go on with half-arsed restrictions until yesterday but then people here weren't really ready to accept a second preventative lockdown and are barely ready to accept it now.

Untitled.png

"day 102" is June the 1st, by the way. October the 1st is 224. Sorry about that but plotting dates is a pain.

I use the ECDC countrywide data, which makes it fairly easy

How would using the ECDC data instead of the one I get from Italy via a git help with the way LibreOffice charts deal with dates?

I use, excel which doesn't have a problem, but the ECDC has several different formats for the data, so one might be easier for Libre office to work it out.

Is the problem something to do with the location setting for your system? I'm surprised it has a problem

Anyway, here are my plots per 100k using that data: https://www.ecdc.europa.eu/en/publicati ... -worldwide

Italy 2020-11-07 141136.png (54.06 KiB) Viewed 3749 times

France 2020-11-07 141241.png (33.05 KiB) Viewed 3749 times

UK 2020-11-07 141402.png (31.7 KiB) Viewed 3749 times

[shpalman]

No, the problem is that by plotting every 28 days you get a different day every month (or two in one month every so often) and it takes a moment to realize that it's just a label every 28 days.

[KAJ]

I prefer cases by specimen date to cases by reporting date, otherwise there is a random shift in the date axis.

Screenshot 2020-11-07 at 20.56.42.png (21.48 KiB) Viewed 3682 times

I don't think the ECDC data includes that so I use the gov.uk data.

I've installed R and RStudio on my Chromebook (hat-tip to sTeamTraen), grabbed R code from gov.uk which imports the data to-date. and I've had a play. If anyone wants the code let me know - no warranties or support! I also have a URL which downloads the data as a CSV, same offer.

Using specimen date means omitting the last few days, whose results may not be available yet. I've used back to 56 days before that,can't really expect a simple model to fit over a longer period.

A straight exponential fit isn't too bad but there is clear:
a) day of week effect
b) structure in the residuals (the line isn't really straight)

Code: Select all

Call:
lm(formula = log(SpecCases) ~ date, data = aDF[5:60, ])

Coefficients:
              Estimate Std. Error t value Pr(>|t|)    
(Intercept) -7.332e+02  4.362e+01  -16.81   <2e-16 ***
date         4.005e-02  2.352e-03   17.02   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.2845 on 54 degrees of freedom
Multiple R-squared:  0.8429,	Adjusted R-squared:   0.84 
F-statistic: 289.8 on 1 and 54 DF,  p-value: < 2.2e-16

Screenshot 2020-11-07 at 21.30.43.png (52.65 KiB) Viewed 3682 times

Adding a weekday effect and quadratic in 'date' fixes those problems and gives an R^sq of 97.5%
Note Saturday and Sunday have substantially lower cases than the rest of the week, probably reflecting testing rates.

Code: Select all

Call:
lm(formula = log(SpecCases) ~ poly(date, 2) + weekdays(date), 
    data = aDF[5:60, ])

Coefficients:
                        Estimate Std. Error t value Pr(>|t|)    
(Intercept)              9.34618    0.04251 219.859  < 2e-16 ***
poly(date, 2)1           4.87366    0.12113  40.234  < 2e-16 ***
poly(date, 2)2          -1.54479    0.12021 -12.850  < 2e-16 ***
weekdays(date)Monday     0.10200    0.06018   1.695 0.096690 .  
weekdays(date)Saturday  -0.22587    0.06011  -3.758 0.000473 ***
weekdays(date)Sunday    -0.29432    0.06013  -4.894  1.2e-05 ***
weekdays(date)Thursday   0.08418    0.06011   1.400 0.167941    
weekdays(date)Tuesday    0.05308    0.06024   0.881 0.382703    
weekdays(date)Wednesday  0.11784    0.06014   1.959 0.056004 .  
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.1202 on 47 degrees of freedom
Multiple R-squared:  0.9756,	Adjusted R-squared:  0.9714 
F-statistic: 234.9 on 8 and 47 DF,  p-value: < 2.2e-16

Screenshot 2020-11-07 at 21.34.34.png (60.86 KiB) Viewed 3682 times

ANOVA shows both refinements matter

Code: Select all

Model 1: log(SpecCases) ~ date
Model 2: log(SpecCases) ~ date + weekdays(date)
Model 3: log(SpecCases) ~ poly(date, 2) + weekdays(date)
  Res.Df    RSS Df Sum of Sq       F    Pr(>F)    
1     54 4.3715                                   
2     48 3.0649  6    1.3066  15.072 1.581e-09 ***
3     47 0.6791  1    2.3859 165.130 < 2.2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

[jimbob]

No, the problem is that by plotting every 28 days you get a different day every month (or two in one month every so often) and it takes a moment to realize that it's just a label every 28 days.

Got it.

The step changes in the gradient in Italy are striking - at about the 26th September and 10th October for cases and deaths respectively.

It looks as though the UK is starting to slow down it's growth now.

[shpalman]

log.png (43.79 KiB) Viewed 3665 times

Nearly 40,000 new cases today. We're at 2634 patients in intensive care with a doubling time of about 12 days. Apparently the total capacity in Italy is about 10,500 so the end of the month will be terrible if the lockdown measures don't start to work soon.

The number of deaths per day - about 400 - is roughly half what we had at the peak at the end of March but the excess deaths point to there having been a peak double that slightly earlier.

[KAJ]

It looks as though the UK is starting to slow down it's growth now.

The rate of increase of cases is slowing (see my post above). Deaths (by date of death, not date published) are still rising exponentially.

Code: Select all

Residuals:
     Min       1Q   Median       3Q      Max 
-0.46648 -0.09575  0.02241  0.10271  0.50488 

Coefficients:
              Estimate Std. Error t value Pr(>|t|)    
(Intercept) -1.130e+03  2.529e+01  -44.69   <2e-16 ***
date         6.120e-02  1.364e-03   44.86   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.165 on 54 degrees of freedom
Multiple R-squared:  0.9739,	Adjusted R-squared:  0.9734 
F-statistic:  2012 on 1 and 54 DF,  p-value: < 2.2e-16

Screenshot 2020-11-07 at 23.13.28.png (30.25 KiB) Viewed 3651 times

[jimbob]

It looks as though the UK is starting to slow down it's growth now.

The rate of increase of cases is slowing (see my post above). Deaths (by date of death, not date published) are still rising exponentially.

Code: Select all

Residuals:
     Min       1Q   Median       3Q      Max 
-0.46648 -0.09575  0.02241  0.10271  0.50488 

Coefficients:
              Estimate Std. Error t value Pr(>|t|)    
(Intercept) -1.130e+03  2.529e+01  -44.69   <2e-16 ***
date         6.120e-02  1.364e-03   44.86   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.165 on 54 degrees of freedom
Multiple R-squared:  0.9739,	Adjusted R-squared:  0.9734 
F-statistic:  2012 on 1 and 54 DF,  p-value: < 2.2e-16

Screenshot 2020-11-07 at 23.13.28.png

Thanks - And yes I'd trust yours rather than mine, as it's on date reported although the 7-day average does smooth it very well

[AMS]

It looks as though the UK is starting to slow down it's growth now.

The rate of increase of cases is slowing (see my post above). Deaths (by date of death, not date published) are still rising exponentially.

That's as expected, isn't it? The lag in deaths v cases will mean the effect on the deaths curve will take longer to come through.

[KAJ]

It looks as though the UK is starting to slow down it's growth now.

The rate of increase of cases is slowing (see my post above). Deaths (by date of death, not date published) are still rising exponentially.

That's as expected, isn't it? The lag in deaths v cases will mean the effect on the deaths curve will take longer to come through.

Yes. Covid deaths are defined as within 28 days of the first positive test result. When I have time I'll compare death growth rates with case growth rate 28 days before. Now I have the tools (R and Rstudio, another hat-tip to sTeamTraen) and ready access to the data that's easy - but Mrs KAJ looks at me when I spend too long calculating

[Woodchopper]

Good thread on lack of UK govt transparency: https://twitter.com/edconwaysky/status/ ... 03520?s=21

[sTeamTraen]

For anyone (like me) who thought that the US might still be the place doing worst, especially with the recent numbers of over 100,000 cases per day, this might be sobering. 7-day moving averages of cases per 100K and deaths per million in the EU (27 countries) and the US (50 states etc etc). A variety of healthcare and lockdown policies in both cases, and also a lack of centralised leadership --- in the EU because they have no mandate, in the US because of, well, you know.

My guess is that the difference between the two is that several large US states that were hit hard in the first wave (e.g., New York) seem to be keeping a lid on new outbreaks this time around, whereas almost every EU country is going the wrong way. France reported 86,000 cases yesterday. Belgium's recent average, if scaled up to the US population, would be 450,000 per day. Shout out to Ireland, which has seen new cases halve in the two weeks since they went for a lockdown that was almost as strong as the one in the spring.

tl;dr: COVID makes idiots of us all, over and over again.

Untitled.png (21.27 KiB) Viewed 3586 times

(Apologies for the Excel charts, I still need to program the generation of those to go with the code that extracts the numbers from the ECDC data.)

[shpalman]

The health commissioner of Calabria, Saverio Cotticelli, only just realized that it was his responsibility to organize his region's covid response plan.

he realized while he was being interviewed

He's been replaced by Giuseppe Zuccatelli, who is currently in quarantine because he tested positive for covid last week, and during the first wave said that masks were useless, although he's insisting he didn't mean it and now says they're "indispensable".

[shpalman]

For anyone (like me) who thought that the US might still be the place doing worst, especially with the recent numbers of over 100,000 cases per day, this might be sobering. 7-day moving averages of cases per 100K and deaths per million in the EU (27 countries) and the US (50 states etc etc). A variety of healthcare and lockdown policies in both cases, and also a lack of centralised leadership --- in the EU because they have no mandate, in the US because of, well, you know.

My guess is that the difference between the two is that several large US states that were hit hard in the first wave (e.g., New York) seem to be keeping a lid on new outbreaks this time around, whereas almost every EU country is going the wrong way. France reported 86,000 cases yesterday. Belgium's recent average, if scaled up to the US population, would be 450,000 per day. Shout out to Ireland, which has seen new cases halve in the two weeks since they went for a lockdown that was almost as strong as the one in the spring.

tl;dr: COVID makes idiots of us all, over and over again.

It depends a lot on what scale of map you're looking at. For example, in the first wave, half of Italy's cases were in Lombardy, and if Lombardy were a country it would have looked worse than Belgium (they have similar populations). So it seemed the second wave in Lombardy wasn't going too badly because people remembered, and in September that might have been true. But then Milan was relatively unaffected first time around because it was concentrated around Bergamo, and now Milan is having a much worse second wave (about a quarter of Lombardy's cases). (Lombardy's true death stats were about double the official covid deaths but Bergamo's true death stats were five times higher than the official covid deaths.)

Lombardy's numbers are always at https://www.lombardianotizie.online/cor ... lombardia/

The province of Como is trending up through 100 new cases per day per 100,000 on a doubling time of 8-9 days which is slightly above the Lombardy average (more like 90 new cases per day per 100,000). I don't have deaths data at a Province level but Lombardy as a whole is at about 10 deaths per day per million. Of course that will go up to follow the way that cases have been going up for the past month.

[jimbob]

For anyone (like me) who thought that the US might still be the place doing worst, especially with the recent numbers of over 100,000 cases per day, this might be sobering. 7-day moving averages of cases per 100K and deaths per million in the EU (27 countries) and the US (50 states etc etc). A variety of healthcare and lockdown policies in both cases, and also a lack of centralised leadership --- in the EU because they have no mandate, in the US because of, well, you know.

My guess is that the difference between the two is that several large US states that were hit hard in the first wave (e.g., New York) seem to be keeping a lid on new outbreaks this time around, whereas almost every EU country is going the wrong way. France reported 86,000 cases yesterday. Belgium's recent average, if scaled up to the US population, would be 450,000 per day. Shout out to Ireland, which has seen new cases halve in the two weeks since they went for a lockdown that was almost as strong as the one in the spring.

tl;dr: COVID makes idiots of us all, over and over again.


Untitled.png

(Apologies for the Excel charts, I still need to program the generation of those to go with the code that extracts the numbers from the ECDC data.)

I like Excel - it helps that I seem to be able to work out how Microsoft will approach their user interface (I'm often asked to see if I can work out how to do something at work in a package that I've not used). Handy hint - you can paste in different series into x-y plots in excel. After Excel 2003, you cannot use alt-ctrl-V but you can use "Alt then E then S" (you don't need to keep the alt key down to press the E but it works either way)

I think this is a pretty good way of presenting the last 20 years deaths data to show that it really was "of COVID" and not "with COVID"

The error bar is the number reported with COVID-19 mentioned on the death certificate

[shpalman]

FB_IMG_1603607834836.jpg (25.15 KiB) Viewed 3479 times

FB_IMG_1603607839312.jpg (51.35 KiB) Viewed 3479 times

Got these off Facebook a couple of weeks ago.

[lpm]

Holy f.ck, the US is getting bad unbelievably fast while we were looking the other way.

Doubled in a week in a place like Iowa - taking it to something like 3x worse than the UK. And nothing in place to stop it doubling again.

The most recent changes to the lockdown regime in Iowa are:

15 Sept - action on bars
16 Sept - decision on statewide mask order
2 Oct - new guidance on quarantine
2 Oct - further action on bars

However these actions might not be what you expect. To add a bit of detail to the above:

15 Sept - action on bars. Lifted the bar closure order in four counties.
16 Sept - decision on statewide mask order: governor refused calls to impose the order and blocked local officials from enforcing any local mask mandates.
2 Oct - new guidance on quarantine. It was relaxed in contradiction of CDC guidelines
2 Oct - further action on bars. Reopened in two more counties.

Of course, places like Iowa are now seeing voluntary restraint. Restaurants and bars, hairdressers, cinemas etc are all struggling badly even though allowed to be open. People are working from home so lunchtime diners and cafes suffer. College bars in Iowa, on the other hand, seem to be doing well - they were shut in September during the university flare up but only for a month.

It's truly mad. They aren't at the R=3 level due to voluntary restriction but they are a f.ck of a long way from R=1.

Can't they now forget about politics and mask culture wars?

NYT case count for Iowa. It lets you select whichever state you want:
https://www.nytimes.com/interactive/202 ... cases.html

[Bird on a Fire]

Mixed messages here in Portugal too, as we enter a newly-declared State of Emergency.

On the one hand, we are told we have a civic duty to remain at home as much as possible. Remote working compulsory wherever possible. Not allowed to go anywhere between 11pm and 5am (apart from a few sensible exceptions).

On the other hand, bars and restaurants are allowed to remain open (though are limited to 6 people per table, unless they're from the same family group). I can't imagine any circumstances where it's necessary to go to a restaurant (as opposed to a takeaway), so I suspect the government is just trying to get out of having to compensate people, which presumably means they're also hoping enough people will ignore their civic duty to stop everywhere going bust.

Places are limited to 50% capacity and masks are still compulsory of course, plus there's early closing times and restrictions on alcohol. But we've had most of these restrictions for months and the cases keep ticking up and up

Interestingly, there's also provisions in the declaration that workplaces (I've seen it at hairdressers, for instance) are allowed to demand "non-invasive measures of body temperature" before allowing people inside, and can refuse anyone who's over 38°C. Some places - like educational facilities, accommodation blocks - are allowed to demand that people take tests. I've not heard of any testing shortages yet, and we're up to about 40,000 a day.

[KAJ]

The rate of increase of cases is slowing (see my post above). Deaths (by date of death, not date published) are still rising exponentially.

That's as expected, isn't it? The lag in deaths v cases will mean the effect on the deaths curve will take longer to come through.

Yes. Covid deaths are defined as within 28 days of the first positive test result. When I have time I'll compare death growth rates with case growth rate 28 days before. Now I have the tools (R and Rstudio, another hat-tip to sTeamTraen) and ready access to the data that's easy - but Mrs KAJ looks at me when I spend too long calculating

Well, I've had a play and you can colour me surprised .

I've used stats::ccf (cross-correlation) to find the lag which maximises correlation between SpecCases (cases by specimen date) and DateDeaths (deaths by date of death). I'd expected a lag of a couple of weeks or so, but the lag is zero!
Here's a plot of the two variables over the whole outbreak, the vertical grid lines are at 4/1 week intervals to ease comparison

Rplot.png (65.9 KiB) Viewed 3340 times

This plot shows quite clearly that the maximum correlation (+0.574) is at lag zero days.

Rplot01.png (14.17 KiB) Viewed 3340 times

I thought the lag might be "muddied" by using such a broad date range so I looked at just the last couple of months. Again the maximum correlation (+0.929) is at lag zero days. Here's a plot of that subset.

Rplot02.png (36.06 KiB) Viewed 3340 times

I really am surprised. I don't understand how there can not be a lag between positive cases and deaths. Any comments?

[shpalman]

When you have two things increasing exponentially with the same time constant, you'll have two parallel strain lines on your semilog plot, so it's ambiguous about whether you should shift one of them horizontally or vertically (or any combination) to match the other one.

If you made the assumption that deaths had to be 1% (for example) of cases, that fixes the vertical offset and so constrains the horizontal shift too.

[KAJ]

When you have two things increasing exponentially with the same time constant, you'll have two parallel strain lines on your semilog plot, so it's ambiguous about whether you should shift one of them horizontally or vertically (or any combination) to match the other one.

If you made the assumption that deaths had to be 1% (for example) of cases, that fixes the vertical offset and so constrains the horizontal shift too.

Good catch! I hadn't thought of that.

But changes in rate should show up, and that's what started me down this rabbit hole. There was no offset in the peaks in the first half (lag = 0, corr = 0.972) ...

Rplot.png (50.15 KiB) Viewed 3308 times

...but I'm a bit suspicious of the data then.

For the dip and rise in the second half,,,

Rplot02.png (57.4 KiB) Viewed 3308 times

,,,there is a max correlation (+0.621) at a lag of 14 days.

I'll take that as a typical case-to-death lag, although (as expected) it isn't very precisely determined,,,,

Rplot03.png (13.48 KiB) Viewed 3308 times

.

Thanks for your help

[geejaytee]

Does it change when you lag log(deaths) v log(cases)?