Our government, and its PR machine, the mainstream media, have got the covid death caldron nicely on the boil. “Double, double toil and trouble; Fire, burn; and, caldron, bubble… Eye of newt, and toe of frog, Wool of bat, and tongue of dog…Like a hell-broth boil and bubble.” At first glance, the recent daily covid death counts are alarming, but are the reported covid counts real? Is there a worrying excess in all deaths, or just a normal seasonal rise? The answers to these question are most readily found by looking at all cause mortality. Although all cause mortality is coarse grained — in truth it has no grain, it is just a single count of all deaths — it nonetheless has one cardinal virtue that comes about precisely because it doesn’t try to be too clever, it is robust. The diagnosis of death is rarely got wrong, and the UK has a reliable system of death registration. We know how many people have died overall, and when and where they died.
Dr No has already shown in recent posts that overall 2020 mortality was bad, but only as bad as was normal a decade or two ago, depending on the detail of how the analysis is done. What this means is that someone from around 2005 using a time machine to look into the future over the intervening years at 2020 would see nothing exceptional. 2020 was just another normal year. But 2021 is a new year, and already the witches are at work, fire, burn; and, caldron, bubble, frothing about the ever rising covid deaths: another all-time high, record death tolls, a new daily high. Let us see if we can use the latest figures to get an idea whether the covid death caldron really is boiling over, or is instead just simmering, at high normal seasonal levels.
The first thing to get out of the way is Z-scores. Promoted by EuroMOMO, the pan-European organisation that aims to “to detect and measure excess deaths related to seasonal influenza, pandemics and other public health threats”, and favoured by PHE, who deploy them on the front page of its weekly excess deaths report, the Z-score is a classic example of statistical obfuscation. EuroMOMO claim that its modified Z-score allows direct apple to apple comparisons, both between countries and between years in one country by setting a benchmark and so thresholds based on the standard deviation (see footnote for a refresher on standard deviations) of the mean number of deaths, where the Z-score is the observed number of deaths minus the mean (and so expected) number, divided by the standard deviation. The Z-score tells us how many standard deviations the observed number is above (or below) the expected number. EoroMOMO defines a Z-score above four as a ‘substantial increase’ in the number of deaths.
In theory, this just might be plausible, were it not that EuroMOMO cook the books by using a baseline based on, in their own words, a “de-trended and de-seasonalized series, after a 2/3 powers transformation according to the method described in Farrington et al. 1996″. What this wonderful example of obfuscation means, once it has been de-jargoned and de-obfuscated after a 2/3 powers transformation according to the method described by Dr No in 2021, is that the baseline, and so all the thresholds above the baseline, are based only on spring and autumn data, the two typically unexceptional seasons of the year for mortality. This makes as much sense as calculating an average annual temperature based only on spring and autumn months, and then declaring ‘by gum, that summer was hot’.
Another perhaps more serious flaw in the EuroMOMO algorithm is that the variability (the dispersion or scatter) about the mean number of deaths will depend on the population size, with smaller populations, and so a smaller number of observations, or deaths, more prone to wider variability, which in turn will feed through into bigger standard deviations, and so smaller Z-scores (recall, the Z-score is the difference between the observed number of deaths and the mean, expected number of deaths, divided by the standard deviation). The four UK nations demonstrate this flaw very clearly (Figure 1). The smaller the population, the smaller the Z-scores. For Northern Ireland, (population ~1.8m vs England’s ~56m), EuroMOMO has achieved what months of epic masking up and locking down have failed to achieve: all but total ablation of the pandemic.
Figure 1: EuroMOMO charts for the four UK nations, with data from the last five years, up to week 2 2021
That’s what happens when you weaponise statistics, and the simple solution is don’t do it. So let us turn back to raw numbers of deaths, uncomplicated by numerology. Here are the latest ONS weekly all cause deaths for England and Wales**, and it looks like the covid death caldron is boiling nicely. Week 1 2021 sure looks bad (Figure 2).
Figure 2: ONS weekly all cause deaths registered for England and Wales**, 2020 plus week 1 2021
Not quite as striking, perhaps, but nonetheless very visible, we can also see that the two weeks before the last week were also unusual, by having low counts, and there is a well known reason, the lag in registrations caused by Christmas and the New year. What happens if we apply a 2/3 weeks transformation, according to a method described by Dr No in a recent post, which takes 4,000 deaths from the final week on the chart, and puts then back in the penultimate week (+ 2,500 deaths) and the week before the penultimate week (+ 1,500 deaths). What we see is far less alarming, with deaths roughly in line with a year ago, in January 2020 (Figure 3).
Figure 3: ONS weekly all cause deaths registered for England and Wales**, 2020 plus week 1 2021, adjusted for reporting delays over Christmas and the New Year
Viewed with this adjustment, it seems the caldron has rather gone off the boil, and is simmering at something approaching normal, with a moderate season increase in weekly deaths. There is no winter covid spike visible. If there is no spike in all cause deaths, then there is no spike in deaths cause by covid, because there is no spike.
This leaves just one other chart to consider, the weekly covid deaths chart on the coronavirus dashboard. This is based on ONS data, plus equivalent (but not identical) Scotland and Northern Ireland data, and for the UK as a whole it is very clearly another OMG chart. To make the data comparable to Figures 2 and 3, Dr No has downloaded the data for England and Wales and summed and plotted them (Figure 4).
Figure 4: Weekly deaths attributed to covid–19, by date of registration, England and Wales (source: data downloaded from coronavirus dashboard for each country, then summed and plotted locally).
Surely this is proof that covid–19 is not just out of control, it is on the rampage. But is it? Consider that there have been no similarly striking rise in all cause mortality for in recent weeks, once the bank holidays lags have been corrected. How can this be? Covid deaths all but double, yet overall mortality remains pretty stable? The only sensible explanation is that, rightly or wrongly, the proportion of deaths attributed to covid–19 all but doubled over a week. And it just happens that this, and the period just before it, was the period in which testing rocketed up, all but doubling from around 300,000 tests a day to approaching 600,000 tests a day on some days.
Dr No has good confidence that when ONS reports its next set of weekly death registrations for England and Wales, the numbers will have moved back towards normal for this time of year. His grounds for this confidence are not based on models or numerology or even ouija boards, but on his favourite grounds for confidence, real observations in the real world. As it happens, Scotland has published more recent data, as has Northern Ireland, and both show the numbers of deaths registered returning to normal. There is no reason to suppose England and Wales will not do the same.
Footnote: Standard Deviations. Standard deviations are a measure of the scatter, or dispersion, of readings about the mean. Consider two groups of adults males: one made up of men all of similar heights, and another made up of men widely varying heights. The mean height will likely be similar, but in the first group, the standard deviation will be small, in the second it will be large, indicating wider scatter, or dispersion. Now suppose we want to identify some outliers in each group, the extremely tall men. No single measure of height will do. For instance, the outliers in the similar height group will very likely be within the normal range for the more varied height group. This is where the standard deviation can help. Given a Normal distribution (the bell shaped curve), it is a characteristic of standard deviations that fixed proportions of the population lie within so many standard deviations from the mean. The best known range is from minus to plus two standard deviations, which will include just over ninety five percent of the population. This means we can use a single measure, two standard deviations above the mean (though of course the actual number will be different, because the groups are different), to define the threshold above which a man is considered tall.
** Correction added 1145 26th Jan 2021: Dr No apologises, these references to England and Wales are incorrect, they (and Figures 2 and 3) refer to just England. Nonetheless, he believes the point made remains the same, and so will leave them as they are – and better an original post plus correction rather than the smoke and mirrors of original posts that mysteriously change.