I investigate a sample of 57,557 adult in-hospital deaths for the period 2017/w1 – 2021/w36 for an undisclosed NHS Trust with a view to estimating ‘internal’ excess death
1. We can define excess death for all cause deaths regardless of vaccination status using the 2017 - 2019 data but not for vaccinated cohorts for there's no baseline prior to vaccination! A slide of cumulative percentages of unvaccinated vs. dose 1 vs. dose 2 over time can be produced for interest (I may have done this already).
2. If vaccines were a particular solution to a particular problem then we'd expect to see a dramatic decline in excess death following the initial demise of vulnerable folk. If no other problems were extant within the catchment population then excess deaths should 'normalise'; that is they should bob about the zero axis. The fact that they do not indicates a major, and likely compounded problem that is ongoing. This is covered in parts 2 - 6.
Regarding #2 with respect to accumulated deaths, ‘if’ an initial surge of deaths were in an older population and many of those were brought forward in time from when they would otherwise have occurred, e.g. within a couple years, and ‘if’ there are no other abnormal events, ‘then’ should one expect not only a leveling off but also a decrease (in accumulated deaths)?
I take it you are referring to the alleged “dry tinder” effect that was floated at the start of the lockdown era.... can’t really see it in this graph!
Yes and no! The dry tinder effect of mild flu seasons stretches back before 2017 with just a hint in 2018 so it's not going to be that visible with this short series. Also, the technique that I'm using avoids this sort of problem because I'm not using prior 5-year means. In parts 2 - 6 you'll get to see the differences when I flip to using ONS registered deaths for England & Wales going back to 2010.
Dave - how about I establish a simulation so folk can get to see how different scenarios play out? I'll call it 'Understanding Excess Death' or something like that and it'll be 3 - 4 articles long.
1. Business as usual for a seasonal pattern; 2. Outbreak of nasty pathogen during peak season; 3. Outbreak of nasty pathogen during off-peak; 4. Ramping of underlying rate from a specific point. I haven't done any work on the timing of variants as yet but it's something to think about!
Understood- the message is that it may not matter in this data set.
Once a baseline model is established then I presume it appropriate to develop and event model in order to tease out whether there could be multiple driving events and how many (like a clustering problem).
In general, if a model of excess deaths for a first event properly accounts for deaths pulled forward (i.e. via a subsequent drop), and this drop is not seen in the observed data, then a second event may be present.
I believe one of the things Joel Smalley did was to model a first event with a Gompertz function, subtract its effects, and look for a second event. I don’t recall whether Gompertz function can include pull forward effects, or whether Joel added analysis to account for that if not - it’s been a long while since I read his analysis (though I see he has some more recent posts).
It depends on your method. If using prior 5-year means as a baseline and you are trying to model excess death then you need to approach it like Joel. If you are using ARIMA all this is done for you and ARIMA will churn out a table of outlying data points that provide evidence of event sequences. In place of a Gompertz function you can use the seasonal mean time series as an independent variable representing 'usual flu'. All this and more is covered in parts 2 - 6 as I take folk through five different methods so we may compare results.
How does/would ‘population change over time’ factor into the baseline model? (More deaths are expected in a larger population, all other things being equal.) How would population be defined for the cohort in this Trust?
The impact of population change (which will be negligible for the span of years studied) is embedded in the time series for observed counts (dependent variable). Parts 2 - 6 will hopefully illustrate this. The Trust will have a catchment area that is subject to contract but these are not hard and fast boundaries with some Trusts mopping up for others when it comes to specialised services.
Thats a good point - some people - including the frail elderly - very inconveniently choose to stay with relatives at Christmas and drop dead due to too much plum pudding - however my guess is that tis will sort of even out nationwide.
That's a really impressive analysis - now excited to see what the ARIMA approach can bring to the ONS data.
Joel Smalley has often cautioned on the use of the ONS data because it uses date of registration rather than date of death. Will you be able to use the data of death data series?
As an aside, I just wish I had known about the T4235H smoothing filter years ago - lots of smoothing with no phase shift.
Cheers! Yep, when Joel gets his hands on the latest FOI data dump I'll be able to run the same analysis over this. In parts 2 - 6 you'll see me try several methods for registered deaths. The date of death vs. date of registration issue can be massively problematic and it can be inconsequential, depending on how you are going about estimating excess death. I ran a study on this way back showing the difference - if I can find the article I'll drop a link. Yes indeed, a lack of phase shift is a god send!
Sure thing. I'm building a shopping list of extras that I can provide in part 7 (parts 2 - 6 already lined-up ready to roll). The alternative is to use substack notes for ad-hoc slides, which may be preferable.
Using Substack Notes for this is a bad idea, because you end up with slides that people remember seeing but cannot find when they search through your collected writing days, weeks, months even years later.
Understood but it's the only way to get specific slides to specific folk at the time of their request. I publish articles once every Monday and, as it stands, all article slots though to 27th May are now taken. If a request is likely to lead to a full article (e.g. David A's suggestion for looking at excess deaths for acute respiratory conditions) then that gets booked into my advance work programme. Jessica's request would not lead to a full article so I am left with the option of posting in notes immediately or politely declining.
I like the idea of posting a note immediately, and then every so often posting an article here: 'Odds and Ends: Stuff I posted to notes since the last time I made one of these articles' which just copies the stuff here. No longer article needed, because it's just for the purpose of having a searchable library.
I've heard about this but have not sat down to read what they've decided to do. With my series on excess death now running 7 parts (part 2 explores the original ONS method, and parts 3 - 6 explore alternatives) it makes sense to work through the new ONS method in part 8.
1. Would be nice to see the cumulative percents of 1 and 2-dosers overlayed.
2. Would you agree that leveling off (but not decreasing) is still incompatible with claims of vaccine efficacy?
3. Maybe to answer #2, cumulative excess of acute respiratory death can be looked at?
1. We can define excess death for all cause deaths regardless of vaccination status using the 2017 - 2019 data but not for vaccinated cohorts for there's no baseline prior to vaccination! A slide of cumulative percentages of unvaccinated vs. dose 1 vs. dose 2 over time can be produced for interest (I may have done this already).
2. If vaccines were a particular solution to a particular problem then we'd expect to see a dramatic decline in excess death following the initial demise of vulnerable folk. If no other problems were extant within the catchment population then excess deaths should 'normalise'; that is they should bob about the zero axis. The fact that they do not indicates a major, and likely compounded problem that is ongoing. This is covered in parts 2 - 6.
3. This is a stonkingly good idea!
I've made a note to produce slides for suggestions #1 and #3 but these will be published after part 6 in this series - I've been a busy bunny!
Regarding #2 with respect to accumulated deaths, ‘if’ an initial surge of deaths were in an older population and many of those were brought forward in time from when they would otherwise have occurred, e.g. within a couple years, and ‘if’ there are no other abnormal events, ‘then’ should one expect not only a leveling off but also a decrease (in accumulated deaths)?
Yes indeed, we should see a decrease until the situation 'normalises' wrt historic rates i.e. bobs about the zero axis.
I take it you are referring to the alleged “dry tinder” effect that was floated at the start of the lockdown era.... can’t really see it in this graph!
Yes and no! The dry tinder effect of mild flu seasons stretches back before 2017 with just a hint in 2018 so it's not going to be that visible with this short series. Also, the technique that I'm using avoids this sort of problem because I'm not using prior 5-year means. In parts 2 - 6 you'll get to see the differences when I flip to using ONS registered deaths for England & Wales going back to 2010.
Dave - how about I establish a simulation so folk can get to see how different scenarios play out? I'll call it 'Understanding Excess Death' or something like that and it'll be 3 - 4 articles long.
Sure - what scenarios were you thinking of? And is there any mileage in looking at prevalent Covid variants in this analysis?
1. Business as usual for a seasonal pattern; 2. Outbreak of nasty pathogen during peak season; 3. Outbreak of nasty pathogen during off-peak; 4. Ramping of underlying rate from a specific point. I haven't done any work on the timing of variants as yet but it's something to think about!
I may be though that term is new to me.
Understood- the message is that it may not matter in this data set.
Once a baseline model is established then I presume it appropriate to develop and event model in order to tease out whether there could be multiple driving events and how many (like a clustering problem).
In general, if a model of excess deaths for a first event properly accounts for deaths pulled forward (i.e. via a subsequent drop), and this drop is not seen in the observed data, then a second event may be present.
I believe one of the things Joel Smalley did was to model a first event with a Gompertz function, subtract its effects, and look for a second event. I don’t recall whether Gompertz function can include pull forward effects, or whether Joel added analysis to account for that if not - it’s been a long while since I read his analysis (though I see he has some more recent posts).
It depends on your method. If using prior 5-year means as a baseline and you are trying to model excess death then you need to approach it like Joel. If you are using ARIMA all this is done for you and ARIMA will churn out a table of outlying data points that provide evidence of event sequences. In place of a Gompertz function you can use the seasonal mean time series as an independent variable representing 'usual flu'. All this and more is covered in parts 2 - 6 as I take folk through five different methods so we may compare results.
Thank you for the context. I look forward them. ARIMA is quite the pressure cooker (appliance) to have on the countertop.
Graph of cum% by dose now posted as a note. The acute work will be meaty so I'll line this up for part 7...
How does/would ‘population change over time’ factor into the baseline model? (More deaths are expected in a larger population, all other things being equal.) How would population be defined for the cohort in this Trust?
The impact of population change (which will be negligible for the span of years studied) is embedded in the time series for observed counts (dependent variable). Parts 2 - 6 will hopefully illustrate this. The Trust will have a catchment area that is subject to contract but these are not hard and fast boundaries with some Trusts mopping up for others when it comes to specialised services.
Thats a good point - some people - including the frail elderly - very inconveniently choose to stay with relatives at Christmas and drop dead due to too much plum pudding - however my guess is that tis will sort of even out nationwide.
That's a really impressive analysis - now excited to see what the ARIMA approach can bring to the ONS data.
Joel Smalley has often cautioned on the use of the ONS data because it uses date of registration rather than date of death. Will you be able to use the data of death data series?
As an aside, I just wish I had known about the T4235H smoothing filter years ago - lots of smoothing with no phase shift.
Cheers! Yep, when Joel gets his hands on the latest FOI data dump I'll be able to run the same analysis over this. In parts 2 - 6 you'll see me try several methods for registered deaths. The date of death vs. date of registration issue can be massively problematic and it can be inconsequential, depending on how you are going about estimating excess death. I ran a study on this way back showing the difference - if I can find the article I'll drop a link. Yes indeed, a lack of phase shift is a god send!
Here's the first article in a series that compares date of registration data with date of death...
https://jdee.substack.com/p/counting-deaths-part-1
Is this all in-hospital deaths or a sample of those deaths? Thanks
All in-hospital deaths.
Thanks. Is there a chance you can post the graph of daily in-hospital deaths from 3/1/20-4/30/20 only?
Sure thing. I'm building a shopping list of extras that I can provide in part 7 (parts 2 - 6 already lined-up ready to roll). The alternative is to use substack notes for ad-hoc slides, which may be preferable.
Using Substack Notes for this is a bad idea, because you end up with slides that people remember seeing but cannot find when they search through your collected writing days, weeks, months even years later.
Understood but it's the only way to get specific slides to specific folk at the time of their request. I publish articles once every Monday and, as it stands, all article slots though to 27th May are now taken. If a request is likely to lead to a full article (e.g. David A's suggestion for looking at excess deaths for acute respiratory conditions) then that gets booked into my advance work programme. Jessica's request would not lead to a full article so I am left with the option of posting in notes immediately or politely declining.
I like the idea of posting a note immediately, and then every so often posting an article here: 'Odds and Ends: Stuff I posted to notes since the last time I made one of these articles' which just copies the stuff here. No longer article needed, because it's just for the purpose of having a searchable library.
In addition to the visual, I am looking for how many days of consecutive increase in any one stretch there was in late March/April
Graph posted as a note.
JD, I wonder if you have seen the February 2024 ONS rethink on excess deaths:
https://www.ons.gov.uk/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/articles/estimatingexcessdeathsintheukmethodologychanges/latest
How they love to impress us, the unwashed!
Peter Norman
I've heard about this but have not sat down to read what they've decided to do. With my series on excess death now running 7 parts (part 2 explores the original ONS method, and parts 3 - 6 explore alternatives) it makes sense to work through the new ONS method in part 8.