Pandemic En Croûte – another slice

Something a little special for independence day: using generalised linear modelling to reveal just how flaky the pastry is (rev 1.0)

I hope subscribers are enjoying this dish on this most auspicious of days. The eagle-eyed will have noticed I started my modelling run at 16 Sep 2020 rather than at the onset of the pandemic back in Mar 2020. This is because I wanted to highlight the impact of lateral flow devices that became available during Sep 2020. The best way to generate the illusion of a pandemic is to ensure members of the public poke their nose as often as possible and to get them to tune in to sober news bulletins at breakfast: there’s something about a smart suit, blue tie and posh accent.

What I’ve done while the kettle boils is to run my model for the period 12 Mar 2020 – 15 Sep 2020 but omit the independent variable for rolling 7-day LFD tests. Here’s what popped out a few moments ago:

The Pearson bivariate correlation for this fetches up at a rather sexy  r = 0.984 (p<0.001, n=188), which means variation in the amount of testing and number of people tested using PCR accounts for  96.8% of the variation in daily cases that we observe. This leaves just 3.2% of the variance for anything the virus decided to do; an improvement on the 1.4% declared in the previous newsletter.

The plainest English I can speak at this point is to say our illustrious leaders have made a mountain out of a mole hill. They know it, you know it, they know that you know it, you know that they know that you know it and I can believe that it is not butter.

Cogitation & Coffee

Yesterday I sat thinking about the astonishing results I had baked whilst scoffing flapjack still warm from the oven. I asked myself several times how it was possible that a simple linear model based on tests and people tested alone could predict the daily variation in new cases so accurately that you couldn’t insert a credit card between observed and predicted time series.

I pondered on the inter-dependence between testing regime and disease prevalence in that we’d expect the authorities to ramp up testing when outbreaks were getting vicious and cool the regime down when things were calmer. This would generate a relationship for sure but because of the intrinsic inertia in such a colossal enterprise we wouldn’t ever see a perfect relationship. I got to thinking how well the authorities could predict matters in advance but that boils down to guessing what a (novel) virus is going to do next month. Nope, no joy there either!

I eventually settled on the notion that what we had was a dynamic situation so darn perfect that the tests themselves must somehow be generating the cases. If the tests were random in nature then my model would track daily new cases perfectly but with a great deal of noise (scatter). Non-random results were the only logical conclusion which got me thinking as to what the tests could be testing for. I realised it had to be something that was already endemic in the population that had both seasonal and non-seasonal components. It was then that my cheeky mind settled on the idea of a test that tests for fragments of stuff arising from old and not so novel viruses of all flavours, and the contents of those mysterious exosomes perhaps.

In a nutshell my models suggest that the PCR test appears to be a test that tests for itself, and it might be a jolly good idea to have another good long look at the primers used. Right now I’d put good money on them not being unique or novel.