Flip Flop Flu (part 3)

With the NHS straining under impressive levels of seasonal respiratory infection I decided to take a closer look at influenza and pneumonia

In part 2 of this series I threw out some pertinent paragraphs that were extracted from ONS documents summarizing how death certificates are processed and how AI-laden software is used to automatically designate a single conforming cause of death (ACC). Multi-causal death is not permitted even though this represents a clinical reality; neither are certifying physicians trusted with what they have written. Each real world death, then, becomes an institutionally acceptable synthetic construct that fits in precisely with what WHO decree.

Let us remind ourselves of how the WHO want physicians to think when it came to COVID and death with another pertinent paragraph:

"With reference to section 4.2.3 of volume 2 of ICD-10, the purpose of mortality classification (coding) is to produce the most useful cause of death statistics possible. Thus, whether a sequence is listed as ‘rejected’ or ‘accepted’ may reflect interests of importance for public health rather than what is acceptable from a purely medical point of view. Therefore, always apply these instructions, whether they can be considered medically correct or not."

What are the chances that the majority certifying physicians across the UK ignored these recommendations (that would be bearing down on them from on high via lead clinicians and senior management) and went ahead as they saw fit, placing cause of death with viral/bacterial pneumonia and/or influenza and/or other non-COVID causes because that’s how it sure looked to them? Pretty much nil, I’d venture, but even if these physicians had gone ahead and made a bold stand the current incarnation of the MUSE ACC software would have simply overridden what they had written if what they had written didn’t conform to COVI-Law. All this is done in the best interests of public health, of course.

Flu Truth

We should thus expect a mangled version of flu truth for the years 2020 and 2021, that is mashed in with the coding-driven permanent flu dive of 2001 and temporary coding trial of 1984 – 1992, that are going to make a dog’s breakfast/pig’s ear of any historical analysis unless we get in there first with a few pit props (indicator variables). Then we’ve got some real world pandemics to deal with which, according to Wiki, stand at:

• Spanish flu, 1918 – 1919

• Asian flu, 1957 – 1958

• Hong Kong Flu, 1968 - 1970

• Russian Flu, 1977 – 1979

• Swine (H1N1/09) Flu, 2009 – 2010

Let us then collectively get that kettle on, pop some Darjeeling in the warmed pot and reach for the biscuit tin (maybe the one with the crunchy nut jobbies). Now that we’re settled in I shall crank the handle on the statisticalator™…

ARIMA Stage One (identification of outlying periods)

Autoregressive Integrated Moving Average (ARIMA) time series analysis is once again my spanner of choice in the assessment of things happening over time in a wibbly-wobbly manner prone to events and changes. It might be a good idea to see the time series under scrutiny again, so here it is:

We’ve clearly got to stuff an indicator variable down at 1918 – 1919 to absorb the deadly Spanish flu, and we better stuff an indicator variable to cover the coding trial of 1984 – 1992. The period 2001 onward needs another indicator; and we better conjure an indicator to cover the COVID years of 2020 – 2021. With these obvious features settled the rest of the time series is not so obvious, so we better get the data to tell us what it wants. We can do this (and cross-check our assumptions at the same time) by running a null model (no independent variables) and getting the ARIMA package to spew out a list of all outlying data points.

Believe it or not there are no less than 7 outlier types defined within the ARIMA module of my stats package and here they all are:

Being a wild and reckless sort I am going to run the null model with the options all flipped on to see what drops out of the other end. In the twinkling of an eye (about the same time span as a quick biscuit dunk) we come face to face with this rather interesting table for an ARIMA(2,1,0) model structure:

This tells us that 1918 was a seriously bad year with an estimated crude excess of +378.972 deaths per 100k population (p<0.001) in a series that is generally bobbing about at 100 – 150 deaths per 100k population. We observe that 1919 wasn’t quite as deadly, with an estimated crude excess of +91.853 deaths per 100k population (p<0.001), the methodological upshot of this being that separate indicators are going to be needed to identify 1918 and 1919.

The transient of -57.910 with a decay factor of 0.987 (a very slow decay) for 1984 (p<0.001) marks the introduction of the WHO rule 3 coding trial, with the level shift of +49.608 for 1993 bringing us back more or less into line as the trial ended. Permanent adoption of WHO rule 3 in 2001 is estimated at reducing crude mortality figures by 48.376 deaths per 100k.

Curious Stuff

So far all this makes perfect sense and ties in with our pit prop guesswork based upon eyeballing of the time series. However, we are left with two curious dates – 1915 and 1929 – marking elevated levels of mortality due to pneumonia/influenza with no acknowledged global pandemic behind them (at least according to Wiki).

Also curious is that we do not observe outliers for the Asian flu of 1957, Hong Kong flu of 1968, Russian flu of 1977 and Swine flu of 2009. Now that is a rather large surprise worthy of freshening the pot and reaching for the chocolate hob-nobs! Either these pandemics didn’t make a sufficient dent in the mortality stats of England & Wales or death certification wasn’t what it should have been: another case of flip flop flu?

ARIMA Stage Two (primary indicators)

What we can do now is devise indicator variables to mark 1918, 1919, 1984-1992 and 2001-2021 then run the model again with these firmly in place to see what transpires. In doing this we are trapping down four major sources of variance and this will enable us to squeeze a better fit from the method, thereby revealing a bit more detail. Here’s the revised model standing in naked glory:

All four primary indicators turn up as highly statistically significant as expected (p<0.001), and we have new estimates of their impact that pretty much mirror those popping out of stage one. With our base model incorporating coding changes and the Spanish flu now baked to a golden perfection we can have a look at what the revised outlier table offers:

There’s 1915 hanging in again as a possible pandemic year with an elevated estimated of +44.753 deaths per 100k population. This flu year is a new one for me so I did a bit of searching and came across an old 1916 paper entitled The “Grip” Epidemic Of The Winter Of 1915-1916 by Louis I. Dublin, a statistician - good man! Now this is totally wonderful for it confirms the value of ARIMA modelling in digging out the detail.

This leaves us with 1929 and 1951 as contenders for pandemics that haven’t been widely acknowledged, possibly because they were considered to be mere endemics. Again the NIH National Library of Medicine comes to the rescue with a 1930 paper by Selwyn D. Collins entitled, The Influenza Epidemic of 1928-1929 with Comparative Data for 1918-1919 . As if that wasn’t boon enough, the NLM also comes to the rescue with a 2006 paper by Cécile Viboud et al entitled, 1951 Influenza Epidemic, England and Wales, Canada, and the United States. Get in there!

So what’s with the two negative coefficients for 1930 and 1934? It is impossible to determine whether this represents a genuine improvement in the nation’s health during these specific years or whether a curve-ball was thrown by death certification and/or coding during a chaotic period. There’s also the possibility of survivorship bias and other effects.

ARIMA Stage Three (secondary indicators)

It would seem churlish to ignore 1915, 1929 and 1951 simply because they are classed as endemics, so let us run the model again with indicators for these three years thrown into the cooking pot:

With this refinement (and setting the two war years aside) we appear to have potential endemics detected for 1922, 1937 and 1940, so let’s go see what the NIH National Library of Medicine holds…

The good news is that I found a 1937 JAMA paper by Thomas Francis et al entitled Studies With Human Influenza Virus During The Influenza Epidemic Of 1936-1937, along with a paper in Nature, no less. Yaroo! As yet I’ve been unable to find confirmatory references for UK outbreaks in 1922 and 1940, so I shall leave these in the outlier pool rather than constructing indicators.

The Pudding

Some keen folk will no doubt ask me how well this revised model fits the data and the answer is probably best served in a warm dish with plenty of custard. Have a taste of this:

Those with beady eyes will note the disparity between observed and predicted mortality for the COVID years of 2020 and 2021.

If I run the model again, but limit the data window to 1901 – 2019, then the model predicts a rate of 48.8 deaths/100k for 2020 and 46.5 deaths/100k for 2021, which may be compared to observed rates of 33.7 deaths/100k for 2020 and 27.2 deaths/100k for 2021. In terms of factor reductions we’re talking x0.69 for 2020 and x0.58 for 2021; that is to say pneumonia and influenza mortality is down by 31% for 2020 and down by 42% for 2021 compared with what we may expect.

Whether this is a genuine result arising from viral dominance of SARS-COV-2 or a pointer to dodgy coding I cannot be certain, but what is weird is that even if we accept the viral dominance theory then why were severe COVID-19 cases not dying of pneumonia during 2020 and 2021? This is where WHO coding under rule 3 screws everything up, for pneumonia would not be permitted as the single (conforming) underlying cause of death if a patient tested positive for SARS-COV-2 at any point.

As it stands we cannot determine whether the outbreak of SARS-COV-2 induced more pneumonia than normal (with this fact hidden by cock-eyed rule 3 coding) or whether pneumonia didn’t feature in the progression of what thus was essentially a mild respiratory disease that was served with a monstrous helping of fear, panic and dystopian response sprinkled with glittery media-mediated propaganda.

Tea Break: On The Origin Of Species

I’ve only read the first chapter of Hope-Simpson’s seminal The Transmission of Epidemic Influenza (Springer, 1992) so I possess a rather vague grasp of his controversial hypothesis. I gather that he’s in favour of slowly evolving viral strains that have been an intrinsic part of human physiology since the beginning of all things, which then surface from time to time in a more aggressive form. If this is the case then the deadly 1918-19 pandemic presumably would have been the product of an earlier viral challenge, and that may well be the 1915-1916 outbreak whose signal has been hidden down in the UK data.

This sort of dangerous thinking finds me wondering about the supposedly novel SARS-COV-2 virus leading to a cluster of complex and puzzling symptoms we’re calling COVID-19. If we go all ‘Hope-Simpson’ on this we should be looking back a few years to identify the ancestral challenge, which plonks us down at Swine (H1N1/09) flu of 2009 or the mysterious Russian flu (H1N1) of 1977. Here’s what Wiki has to say about the latter:

The 1977 Russian flu was a relatively benign flu pandemic, mostly affecting population younger than the age of 26 or 25. It is estimated that 700,000 people died due to the pandemic worldwide. The cause was H1N1 virus strain, which was not seen after 1957 until its re-appearance in China and the Soviet Union in 1977. Genetic analysis and several unusual characteristics of the pandemic have prompted speculation that the virus was released to the public through a laboratory accident.

Is that intriguing or what? Those 25 – 26 year-olds would have been 68 – 70 years of age when COVID came to town so it kinda stacks up in a loose leaf tea sort of way, and especially when we consider older folk in spring 2020 likely succumbed to knee-jerk management rather than COVID itself. The last sentence from Wiki certainly raises an eyebrow given the current speculation (and paper trail) regarding a Wuhan manufactured product/bioweapon: has COVID been a natural variant that has been a long time coming?

Enough Excitement For One Day

I reckon that’s enough excitement for one day, so I’ll pull the shutters down and take a siesta. The phrase there are “lies, damned lies and statistics” was popularised by Mark Twain (who attributed it to Disraeli). I am hoping that my figure-foraging in this little series has fostered a little more faith for a most remarkable field of study that, in truth, can provide the most incredible insights. For this reason we must expect many booby traps and mines to have been planted by those keen on big careers and even bigger bank balances. All depends on where you steer your soul.

Kettle On!