It reads like poetry to me, all these stat terms and notions, none of which I understand. But I can read your language and infer from that your absolute desire to cut out bias and arrive at an outcome that is as close to the “truth” as possible. And I can “read” your graphs 😉 thanks for all that unrelenting work❤️
Could you provide information on current "All cause mortality" for Canada? Stats Can nolonger seems to provide this. And the "Excess Mortality" is confusing for lay people like myself.
The coarse data like ACM is perhaps less scientific, but certainly enough for ordinary people when we know what the Historical numbers were.
Well obviously, when a public agency which is tasked with compiling and distributing information does not fulfill that role, then it is certain that the information would be harmful to the entity paying the salaries....
By inference only, I would guess that the "Official numbers"( if available), would show an All Cause Mortality, about 175% of the long term average. That is not yet threatening to our population levels (specially considering the government's increased focus on immigration?) However, when combined with the decrease in live births, a continuation of the trend will have large effects over a few years.
Good luck with this project Joel, and thank you for doing the work.
Brilliant work Joel. I've been immersed in this task while analyzing weekly New Zealand death data. I've used a simple linear fit, based on my spreadsheets linear regression function, only to comply with the eLife model of Karbinsky and Kobac. But this model is flawed. The fit has to take into account actuarial parameters like Demographics and improvements in life expectancy. Also the influence of bad flu seasons immediately preceding the pandemic. Ignoring these factors gives a flattering curve to many countries because the the expected death baseline is in inflated. Like you I was tempted to make a polynomial fit to the data.
Alas, polynomial is no better than linear fitting. I am not going to provide any insights derived from method #2. I am already on to method #3 which is Gompertz fitting. It is the most appropriate function and so far is generating very robust and reasonable results. Unfortunately it has to be solved numerically so it's taking a long time even with my 36 cores!
All statistical adjustments will need to comprehensively explain the following obs:
1. Almost perfect linear regression concordance between Australia and New Zealand between 2020 and currently, 2023, noting the 2020 excess deaths noted below the 2015 - 2019 mean (oecd.stat).
2. An inversion of the trend observed in Sweden when compared with Australia and New Zealand.
In the Netherlands we did a similar thought experiment (lineair model, link below). Your curve looks convincing and I think it would even be better if you start off by using just the summer periods (may-sep, maybe even jun-jul-aug). This will result in a flatter curve though and it will of course be placed far too low on the Y-scale. It should then be moved up in total, to match the total amount of real deaths. I think the curve will reflect the growth of the older part of the population and thus be a lot 'straighter' (less curvey, longer waves) than your curve.
Joel is like a dog with a bone! Good luck to the man. Very difficult to get any comprehensive data, given the amount of fraud/falsification/manipulation from 2020 onwards within the medical health industrial complex.
Of course it is as easy as the powers that should not be wish to make it. It is very, very easy to prove just how good or harmful the "vaccines" are, using two standardised control test groups of "vaccinated" and "unvaccinated". They can do this in large numbers and globally to prove the point. They won't and we all know why.
Mr. Smalley, you are not the first person to address the problem of understanding mortality rates. This is a recognized branch of the field of study known as "actuarial science" and there are professionals who work in this field called "actuaries". Their ability to do precise work is critical in industries like insurance and pensions. You should begin by learning from them and their hundreds of years of studies.
They don't approach this problem the way you are attempting to approach this problem and there is a very simple reason why: it isn't precise enough.
As far as I know, actuaries don't use polynomials and they don't use Gompertz curves. If they aren't using these techniques, then it is with good reason and you are wasting your time trying to invent a broken wheel.
Mathematical modelling can be properly applied only after you understand the dynamics of the system being modelled. You are attempting to do the opposite: guess what the model might be and then deduce what the dynamics must be. This won't work.
Joel, thanks for detailing the method. I'm following closely because want to use it to compare with other methods.
When you say periodic data I presume you mean the weekly deaths data? You are fitting some curve, originally linear, now suggesting 3rd order polynomial, to the weekly deaths data pre-pandemic. Then subtracting this fitted line to get an excess. This hovers about zero pre-pandemic and then shoots upwards. Then calculate the cumulative excess. I get confused in article when you say "In all cases, I fit a 3-order polynomial to the cumulative series to capture all potential dynamics" Is this fitting another curve, now to the cumulative excess data? Thanks for you patience. It's an important discussion. I like the idea that we have a fit where assume basically zero cumulative excess pre-pandemic. That fit has to reliably predict ahead. I know how once you go beyond linear things can shoot out of control.
Also I previously hadn't appreciated the data you were using, ie following the same year of birth. Clever idea. Is there any reason the method would not be applicable to data such as deaths data in 5 year age bands, which is more commonly accessible?
Hi Andrew, it doesn't matter about the periodicity when I make the distinction between periodic and cumulative. But, otherwise, you are on the right track. There are two possible approaches - fit to the periodic data (daily, weekly, monthly, quarterly, annually, whatever) and then accumulate the results or fit to the cumulative data. The periodic data is not linear - it is, in fact, Gompertz over the whole distribution (i.e. 100 years). However, I assumed that over short intervals (5 years), a linear fit would be OK. It seemed so. But then we had the problem of convexity in the younger ages where I was identifying a massive excess problem. To test this, I considered fitting polynomials to the cumulative data so as to capture that "dynamic". Of course, a 3-order is better than a 2-order if there is more than one dynamic - this can be accelerating convexity (an exponential series), such as we expect in the younger ages, or it can be a reversal, such as we expect in the very old ages (eventually there is no-one left to die so the convexity eventually plateaus and then eventually decreases, i.e. turns concave. You get all these dynamics in a Gompertz. However, the problem with polynomials is that they are sometimes too good! Especially over shorty intervals, they can actually capture so much convexity and concavity, even if it is just noise, and when you forecast with it, you can get extreme values that are not natural or reasonable. So, that's where I am so far, strongly believing that the correct answer lies somewhere between - in the Gompertz. The great news is I can fit a Gompertz to each curve and I'm getting very nice results (between methods 1 and 2). It just takes a long time because each curve has to be derived numerically (not with a closed-form function).
It reads like poetry to me, all these stat terms and notions, none of which I understand. But I can read your language and infer from that your absolute desire to cut out bias and arrive at an outcome that is as close to the “truth” as possible. And I can “read” your graphs 😉 thanks for all that unrelenting work❤️
Mr. Smalley;
Could you provide information on current "All cause mortality" for Canada? Stats Can nolonger seems to provide this. And the "Excess Mortality" is confusing for lay people like myself.
The coarse data like ACM is perhaps less scientific, but certainly enough for ordinary people when we know what the Historical numbers were.
Thanks
Paul
They don't publish decent, raw data. It's a complete shambles.
Well obviously, when a public agency which is tasked with compiling and distributing information does not fulfill that role, then it is certain that the information would be harmful to the entity paying the salaries....
By inference only, I would guess that the "Official numbers"( if available), would show an All Cause Mortality, about 175% of the long term average. That is not yet threatening to our population levels (specially considering the government's increased focus on immigration?) However, when combined with the decrease in live births, a continuation of the trend will have large effects over a few years.
Good luck with this project Joel, and thank you for doing the work.
Paul
Brilliant work Joel. I've been immersed in this task while analyzing weekly New Zealand death data. I've used a simple linear fit, based on my spreadsheets linear regression function, only to comply with the eLife model of Karbinsky and Kobac. But this model is flawed. The fit has to take into account actuarial parameters like Demographics and improvements in life expectancy. Also the influence of bad flu seasons immediately preceding the pandemic. Ignoring these factors gives a flattering curve to many countries because the the expected death baseline is in inflated. Like you I was tempted to make a polynomial fit to the data.
Alas, polynomial is no better than linear fitting. I am not going to provide any insights derived from method #2. I am already on to method #3 which is Gompertz fitting. It is the most appropriate function and so far is generating very robust and reasonable results. Unfortunately it has to be solved numerically so it's taking a long time even with my 36 cores!
Looking forward to it.
All statistical adjustments will need to comprehensively explain the following obs:
1. Almost perfect linear regression concordance between Australia and New Zealand between 2020 and currently, 2023, noting the 2020 excess deaths noted below the 2015 - 2019 mean (oecd.stat).
2. An inversion of the trend observed in Sweden when compared with Australia and New Zealand.
https://drlatusdextro.substack.com/p/excess-deaths-nzzzz-ozzzz-and-sweden
Never mind my Gompertz function question. Still don't fully understand your method though. :D
In the Netherlands we did a similar thought experiment (lineair model, link below). Your curve looks convincing and I think it would even be better if you start off by using just the summer periods (may-sep, maybe even jun-jul-aug). This will result in a flatter curve though and it will of course be placed far too low on the Y-scale. It should then be moved up in total, to match the total amount of real deaths. I think the curve will reflect the growth of the older part of the population and thus be a lot 'straighter' (less curvey, longer waves) than your curve.
Our lineair approach, taking 10-year-trend instead of 5-year-trend. https://virusvaria.nl/en/werkelijke-oversterfte-2022-mogelijk-veel-lager-dan-gedacht/ (Google translation from Dutch)
Joel is like a dog with a bone! Good luck to the man. Very difficult to get any comprehensive data, given the amount of fraud/falsification/manipulation from 2020 onwards within the medical health industrial complex.
Of course it is as easy as the powers that should not be wish to make it. It is very, very easy to prove just how good or harmful the "vaccines" are, using two standardised control test groups of "vaccinated" and "unvaccinated". They can do this in large numbers and globally to prove the point. They won't and we all know why.
Mr. Smalley, you are not the first person to address the problem of understanding mortality rates. This is a recognized branch of the field of study known as "actuarial science" and there are professionals who work in this field called "actuaries". Their ability to do precise work is critical in industries like insurance and pensions. You should begin by learning from them and their hundreds of years of studies.
They don't approach this problem the way you are attempting to approach this problem and there is a very simple reason why: it isn't precise enough.
As far as I know, actuaries don't use polynomials and they don't use Gompertz curves. If they aren't using these techniques, then it is with good reason and you are wasting your time trying to invent a broken wheel.
Mathematical modelling can be properly applied only after you understand the dynamics of the system being modelled. You are attempting to do the opposite: guess what the model might be and then deduce what the dynamics must be. This won't work.
Joel, thanks for detailing the method. I'm following closely because want to use it to compare with other methods.
When you say periodic data I presume you mean the weekly deaths data? You are fitting some curve, originally linear, now suggesting 3rd order polynomial, to the weekly deaths data pre-pandemic. Then subtracting this fitted line to get an excess. This hovers about zero pre-pandemic and then shoots upwards. Then calculate the cumulative excess. I get confused in article when you say "In all cases, I fit a 3-order polynomial to the cumulative series to capture all potential dynamics" Is this fitting another curve, now to the cumulative excess data? Thanks for you patience. It's an important discussion. I like the idea that we have a fit where assume basically zero cumulative excess pre-pandemic. That fit has to reliably predict ahead. I know how once you go beyond linear things can shoot out of control.
Also I previously hadn't appreciated the data you were using, ie following the same year of birth. Clever idea. Is there any reason the method would not be applicable to data such as deaths data in 5 year age bands, which is more commonly accessible?
Hi Andrew, it doesn't matter about the periodicity when I make the distinction between periodic and cumulative. But, otherwise, you are on the right track. There are two possible approaches - fit to the periodic data (daily, weekly, monthly, quarterly, annually, whatever) and then accumulate the results or fit to the cumulative data. The periodic data is not linear - it is, in fact, Gompertz over the whole distribution (i.e. 100 years). However, I assumed that over short intervals (5 years), a linear fit would be OK. It seemed so. But then we had the problem of convexity in the younger ages where I was identifying a massive excess problem. To test this, I considered fitting polynomials to the cumulative data so as to capture that "dynamic". Of course, a 3-order is better than a 2-order if there is more than one dynamic - this can be accelerating convexity (an exponential series), such as we expect in the younger ages, or it can be a reversal, such as we expect in the very old ages (eventually there is no-one left to die so the convexity eventually plateaus and then eventually decreases, i.e. turns concave. You get all these dynamics in a Gompertz. However, the problem with polynomials is that they are sometimes too good! Especially over shorty intervals, they can actually capture so much convexity and concavity, even if it is just noise, and when you forecast with it, you can get extreme values that are not natural or reasonable. So, that's where I am so far, strongly believing that the correct answer lies somewhere between - in the Gompertz. The great news is I can fit a Gompertz to each curve and I'm getting very nice results (between methods 1 and 2). It just takes a long time because each curve has to be derived numerically (not with a closed-form function).
Thanks Joel. Looking forward to the next iteration.