Report to JCVI on vaccine numbers to prevent hospitalisation

[Arborbridge]

I can't see if this thread mentions it, but David Spiegelhalter was in a very good program yesterday around this issue on "Anti-Social" - I hope this is the link https://www.bbc.co.uk/sounds/play/w3ct2dmb
"How anti vax went viral".

However, my point is that this paper was mentioned, as were the statistics behind Andrew Bridgen's speech in the House of Commons which was bsed on statistics but, according to Professor Spiegelhalter, came to a completely wrong conclusion. BTW, he did agree with Bridgen that these policies should be tested and statistics used, but you need to understand them in the first place before drawing concludsions. Otherwise one ends up with "disinformation", "misinformation" or "malinformation". I learnt something there

Arb.

[Hallucigenia]

Caveats. This is a preprint ie. not peer reviewed yet.

This. It's amazing how many papers disintegrate on having someone else look at them.

It's also by what appear to be economists, when to be honest, the only stuff that seems worth reading on death rates seems to come from actuaries as they actually know what they're talking about.

For instance, this lot don't seem to be controlling for things like population growth and changing age structure when all the countries with >75% vax rates have growing populations and all but one of those with <60% vax rates have shrinking populations. If there's 5% more people in Ireland than in your control period, then you will expect 5% more deaths, and vice versa for Bulgaria. It gets more complicated as the demographics won't quite match, but that's why you need to age standardise. But on the face of it, population changes seem to account for most of the variation they see, certainly enough to leave you with just noise.

[9873210]

The graph shows estimated excess mortality for each of the months from January to September 2022. The points on this graph are displayed with substantial error bars, yet all of their centre points lie exactly on a straight line with no sign of any deviations.
Any comments from a statistician - or indeed from anyone who believe they can explain this pattern?

Error bar should be the result of independent inquiry about the reliability of data acquisition, not a function of the gathered data. If for example there is a possibility of a large systematic error (e.g. all the measurements are half the "true" value) you would get the pattern you describe.

For something as messy and difficult to repeat as a pandemic it will be very difficult to squeeze the error bars. For a well understood system that can be measured repeatedly the error bars may be dominated either by the actual phenomenon or something like instrument error that will also show up in the data. Neither "well understood" nor "measured repeatedly" applies here.

TL;DR: Error bars are supposed to be a reflection of our (the authors'/sciences'/humanity's) ignorance. That should be bounded below by the variability of the data, but unfortunately ignorance has no upper limit.

[scotia]

Error bar should be the result of independent inquiry about the reliability of data acquisition, not a function of the gathered data.

Not always - e.g. in nuclear physics scattering experiments, the error bars may be a direct function of the gathered data - due to the number of events following Poisson statistics. I can well remember, from many years ago, the months of data collection to get acceptably modest (Poisson statistical) error bars for publication. Other sources of possible error were less significant.

But I digress - in the Covid paper which we were discussing I could not understand in Figure 1A where the straight line comes from, with data points apparently sitting directly on it - yet with significant error bars. Servodude suggested that the straight line could have been some linear fit to data - but where were the data points used in the linear fit - which I would have expected to be scattered around the line? OK - I have had a look at the data source - "EuroStat Excess Mortality By Month". The version I down loaded was dated 17/2/23, and provided data from March 2022. The percentage excess numbers quoted for March to September are 6.6, 11.9, 7.9, 8.3, 17.0, 13.8, 10.1. I think these actual data points should have been displayed alongside any straight line fit - it provides a significantly different outlook.