Bonus Good News: The death-defying properties of Christmas
Yes you read that right. Fewer people died in the UK, in every age band, during a week of the greatest COVID surge recorded, than died in a week before COVID was recorded in the UK.
|Age||Deaths/week 3 Jan 2020||Deaths/week 1 Jan 2021|
We have just recorded the highest number of COVID cases, and recent news tells us that there are more COVID deaths in this wave than the first one.
So why are we seeing fewer total deaths than average? What’s going on with these numbers? What are we looking at? As data scientists, what should we look at?
The next few sections explore COVID-related datasets to try and answer these curious questions. If you’d rather skip all that and go straight to the answer – which includes the death-defying properties of Christmas – then start at the section “Death and Christmas”
[NB: Apparently links to data sources aren’t showing in some browsers, if you hover over text that talks about data you should see it change colour where there are links. All data here is sourced from responsible data
farmers producers, and only some of it was slightly harmed in the production of these graphs]
The Different COVID Dead
When there are several complicating conditions it can be difficult to isolate which is the cause of death. We sometimes get lazy and look then for more solid numbers, whether or not they are suitable.
For example COVID is a notifiable disease. Anyone who dies with it must be reported as such so that health authorities can track its extent in the same way as they try to track anthrax, another notifiable disease. This makes for a nice solid measure which is usually given as COVID deaths:
Public Health England is quite clear that these are deaths of people who tested positive for COVID within 28 days previously, not that these are deaths because of COVID. If you test positive for COVID on being taken into hospital, but die of cancer, or trauma, or old age, or a broken toenail, you will still be classed as a ‘COVID death’.
That means ‘COVID deaths’ includes the extent of COVID in the population as well as how dangerous it is. If more people in general have it, more people will have it when they die, and the number of ‘COVID deaths’ increases – no matter its lethality. We also test many more people now than we did in the first wave, and so of course we discover more COVID cases in the first place.
It therefore has next to no value by itself for assessing the impact of the disease.
Yet this is what most news feeds report when they talk about ‘COVID deaths’.
What else have we got? Perhaps we could compare the total deaths in a COVID year with previous years. This gives us ‘excess’ deaths and, all other things being equal*, tells us how many extra people have died of COVID rather than with it. [*Not all other things are equal, of course, as lockdown has its own effects, some of which are beneficial]
The ONS weekly deaths source above gives us total deaths, and splits them down by age band:
(I have aggregated the age bands somewhat to compare them with previous years).
We then subtract from this the ‘normal’ deaths of a year. In this case I have deliberately chosen the bad flu season of 2017-2018 so we can see how much worse (or similar) COVID is to the risks we are used to:
This shows the classic ‘river bed cross section’ of a bad flu year, with winter deaths nearly double those of summer. (Note the consistent odd dips at both Christmases; I’ve extended the year to include both).
Subtracting this dataset of weekly deaths from 2020’s dataset gives us the ‘excess deaths’; that is, the number of extra people who have died compared to a bad flu season 2017-2018:
The excess deaths in April are stark. These are almost certainly largely caused by COVID – but also by clearing infectious, elderly, frail patients from hospitals to care homes.
We can see that excess deaths are largely limited to over 45 year-olds; excess deaths in under 44 year olds may be statistically significant but are very low. The rare stories of individual young people killed by COVID are of course terrible, but overall the younger population is no more at risk than a normal year.
We can also clearly see that despite the recent news, the number of excess deaths in late 2020 is much smaller than April and has tailed off. Remember though that this is excess compared to a bad flu year; these are the extra deaths over flu. The odd spike before Christmas, and the dip after which inspired this post, I will cover later.
Not Quite Dead
Perhaps the reason deaths are dropping is because people are treated better? Capacity was supposedly increased during the first wave, treatments have improved, and more staff have been trained.
Unfortunately the NHS stopped reporting on Critical Care Bed Capacity in February 2020. The national data we do have is based on patients with COVID, not patients who are in hospital because of COVID. This means someone in an Intensive Care Unit with emphysema caused by, say, smoking will still be classed as a COVID patient if they test positive for COVID.
We can see that while excess deaths drop from over 2,000/week in late November to less than 1,500/week before Christmas, the admissions to ICUs also drops in a similar period. The fall in deaths therefore does not seem to be due to a rise in ICU admissions.
We can also see that admissions have the lower, broader peak that was intended in the ‘flatten the curve’ concept earlier in 2020, but sustained admissions along with long recovery times can still mean that total occupancy can rise. While I can’t find national statistics, the Welsh government publish total figures here which show this accumulation:
Looking at the capacity, an overcapacity of mechanical ventilator bed spaces were created in March in preparation for a surge in patients that didn’t materialise in general (although it did in hot-spots) so they were repurposed. Now that more patients require mechanical ventilation so the beds are being made available once again. We should be wary therefore of sensational news that most ventilator beds are occupied without understanding what beds can and cannot be repurposed; hospitals have very little interest in maintaining empty beds of any kind.
Who has it?
Perhaps the drop in deaths is due to a drop in the extent of the disease?
Let me please emphasise yet again, that we should estimate infection rates instead of counting cases if we actually care about making sense of this. Nobody seems to bother with this, not even the UK public health authority, so here is their data on cases:
They do at least, at last, also provide the total tests. From these we can estimate the proportion of infected people in the sampled population by dividing the 7-day average of the number of positive results by the 7-day average of the number of tests (Using the 7 day average largely removes the effects of weekends and short hiccups).
Note that the sampled population might not represent the wider UK population. In April most of the tests were ‘pillar 1’ tests of people with symptoms or those who were dealing with COVID patients, so the proportions of positive results would be high (sometimes over 30%) compared to the overall UK population. Now the vast majority of tests are ‘pillar 2’ tests of the wider population so while there is still likely some skew this will give us something close to an upper limit:
Around 5% of the population had COVID in October, rising to an 8% peak in early November. Tier 4 lock-down rules were then introduced and by the end of November the proportion had dropped to around 5%. The rise in early December to 9% suggests lock-down rules were not sufficient or not obeyed, and the sharp rise after Christmas to 14% also suggests many people met many others for the celebration.
Note that these accumulate. People usually recover (or die) within two weeks, so roughly speaking this suggests that well over a third of the UK population have had COVID in the last quarter of 2020.
From the 21st November the proportion of those infected is steady or rising, yet the number of excess deaths two weeks later from 4th December drop (ignore the Christmas spike and trough, I’ll come back to that):
That suggests that COVID is becoming less dangerous, but not (yet) less dangerous than flu.
Death and Christmas
So what was that about the last week of 2020? Why have we seen this strong drop in weekly deaths, so strong that even in a pandemic, in lockdown, the deaths during that week dropped below the average? Why is there a strange spike and dip in the excess deaths graph in those last two weeks?
In fact every year we see a very strong reduction in weekly deaths around Christmas: the average weekly deaths drop from 12,000 per week to 8,000 for the Christmas week and then back up again:
That’s an extraordinary event. Thousands of people don’t die during any week that contains Christmas and Boxing Day. And while there are two other interesting dips (end of May and end of August) they are not nearly as large.
There could be a number of causes. It might be bureaucratic; perhaps paperwork is delayed until the week after. It might be that frail people make an effort to last past the sociableness of Christmas, or at least the good options in the Quality Street bucket. It might be the miracle of Christmas. It might be that when you’re with people you tend to be more, well, lively. It might be that you are much more likely to be rescued from that fatal fall, or the stroke, or the heart attack. It might simply be that sociableness helps us live longer; in which case we should consider how many of the excess deaths are due to the loneliness of lock-down.
Whatever the cause, this ‘dip’ is why we see the dramatic drop in deaths in the last week of 2020 compared to the first. The last week of 2020 in the ONS statistics ended 1st January 2021 and so included Boxing Day. The first week of 2020 however ended on 3rd January, so started on the 28th December 2019, and so did not include the crucial days.
To compensate we can total the two weeks that cover the Christmas 2020 period and compare them with the two from previous year:
|Age||27 Dec 2019||03 Jan 2020||Christmas 2019||25 Dec 2020||01 Jan 2021||Christmas 2020|
We can see 1,800 or so excess deaths over these two weeks in 2020 compared to 2019. That’s around 900 excess deaths per week which seems more sane.
Nevertheless this recent data strongly suggests that the disease is becoming less lethal.
Estimated infections rose from around 5% on 21st November to 8% by Christmas (Cases rose from 14,000 to nearly 40,000). 5% of the population is around 3 million people; by the end of December positive tests were around 14% which suggests around 9 million people in the UK had COVID.
Yet excess deaths from the 5-year average – not just from the bad flu year – dropped every week throughout December:
Even if we ‘correct for Christmas’ and multiply those last two weeks by 50% (the average difference between normal Christmas-week deaths and other winter-week deaths), excess deaths still drop.
The sudden drop in the proportion of infected from 1st January is also remarkable and unprecedented. This may be bureaucratic, ie delays in reporting, but incomplete days are shown in grey on the PHE page and are not included here. It suggests therefore either that lock-down is suddenly ‘working’ (even though the fall after the November Tier 4 lock-down is gradual not sudden, and the fast rise after Christmas shows that many people are not isolating) or that the disease is starting to reach the limits of infection.
Optimistically, and assuming that people recover (or die) within two weeks, about 10% of the population had COVID in October, another 10% in November, and another 20% in December. Then there are the people who must have got it in early 2020, and the small number of vulnerable people vaccinated. It might just be that most people have either been infected or are naturally immune, but my optimistic bias is probably showing through.
The Right Data
In summary, again, we need to use the right data not the easy data. Deaths with COVID are no use whatsoever if we don’t adjust for prevalence in the general population. Instead we should check the excess deaths – and remember that it include effects of lock-down as well as COVID.
And cases should never never ever ever be used without knowing how many tests have been run. Use and report the percentage of positive tests, not the number of cases.