UK mortality and COVID statistics - sars2.net

Contents

• Martin Neil and Norman Fenton: March 2024 preprint about the cheap trick
• motherfunkr: Unvaccinated ASMR reached near vaccinated ASMR during the last month of data included in the ONS dataset
• Clare Craig: New baseline used by ONS to calculate excess deaths
• canceledmouse: Neonatal deaths in Scotland
• Clare Craig: May 2024 UKHSA FOI response for deaths in vaccinated people
• Monthly deaths in FOI response compared to ONS data
• Were dying people vaccinated on their deathbed?
• Mortality rate by week of vaccination up to the end of 2022
• Joel Smalley's comment about my plot
• Joel Smalley's plots with an incorrect baseline
• Clare Craig: Category 1 ambulance incidents in England
• TheRustler83: Percentage of COVID deaths in unvaccinated people in ONS dataset for mortality by vaccination status
• Jikkyleaks: ASMR in people with two doses in the ONS dataset
• Steve Kirsch: Diagnoses for cataracts in England
• Clare Craig: Email by ONS which is supposed to have vindicated Fenton and Neil
• Steve Kirsch: Lancet paper which said there was about 8% excess deaths in the UK in 2022 and 2023
• Camus: Mortality rate in ages 10 to 14 in an old version of the ONS dataset
• Clare Craig: Hospital admissions for multiple sclerosis
• Daily Telegraph article about a 40% increase in hospitalizations for strokes
• Clare Craig: ASMR within 3 weeks from vaccination for first and second doses
• Clare Craig: Searches for vaccine compensation at Google Trends
• Alex Berenson: Plot for percentage of COVID deaths out of all deaths in the first version of the ONS dataset
• Problems with Fenton and Neil's n-1 dose misclassification hypothesis
• Mahadva Setty: Mortality rate for two doses relative to time since vaccination in ONS data
• Clare Craig: Difference in unvaccinated population size for cases and deaths in a Scottish report
• Clare Craig: Maternal mortality rate by race in England
• David Dickson: Cumulative excess deaths in England and Wales with a 2010-2019 average baseline
• Clare Craig: Stillbirth rate in UK

Martin Neil and Norman Fenton: March 2024 preprint about the cheap trick

In March 2024, Martin Neil, Norman Fenton, and Scott McLachlan published a preprint titled "The extent and impact of vaccine status miscategorisation on covid-19 vaccine efficacy studies". [https://www.medrxiv.org/content/10.1101/2024.03.09.24304015v1.full.pdf] The abstract said: "Systematic review identified thirty-nine studies that suffered from one particular and serious form of bias called miscategorisation bias, whereby study participants who have been vaccinated are categorised as unvaccinated up to and until some arbitrarily defined time after vaccination occurred. Simulation demonstrates that this miscategorisation bias artificially boosts vaccine efficacy and infection rates even when a vaccine has zero or negative efficacy. Furthermore, simulation demonstrates that repeated boosters, given every few months, are needed to maintain this misleading impression of efficacy. Given this, any claims of Covid-19 vaccine efficacy based on these studies are likely to be a statistical illusion."

Note how they relied on a simulation to demonstrate that the cheap trick would increase vaccine efficacy.

The methodology they used to do the simulations was poorly described, because for example they didn't mention that they only applied the cheap trick to the numerator but not to the denominator, and they didn't explain which number of new vaccine doses they simulated on each day or week of the simulation. And they didn't even explain why unvaccinated people only had one line in their plots, even though there should've been three different lines that would've corresponded to the three lines for vaccinated people. And at first I even thought that the weeks in their simulation were weeks since vaccination, but actually they were weeks of the simulation instead. This shows the results of the simulation:

For example in scenario A, the vaccine was supposed to be a placebo which had no effect on the infection rate, and all people had a 1% constant infection rate regardless of whether they were vaccinated or not. So I thought that even if people would be classified as unvaccinated for the first 1 to 3 weeks after vaccination, both unvaccinated and vaccinated people should still have a constant infection rate of 1%. But I only understood how the model worked after Uncle John Returns posted this tweet: [https://x.com/UncleJo46902375/status/1770085722550133215]

I've simulated Fenton's simulation. I've only shown the 1 week unvaccinated line (to hit 1% in week 7). You can get almost any shape you like by adjusting the vaccinations per week. Curves smoothed by Excel.

I reproduced Uncle John's plot, but I didn't add the smoothing, and I also added lines for unvaccinated people in the scenarios where people were considered as unvaccinated for 2 or 3 weeks after vaccination:

library(ggplot2)

weeks=11
vax=cumsum(c(10,190,360,200,70,20,rep(0,5)))

lv=c("No lag","1 week","2 weeks","3 weeks")
lag=factor(rep(lv,each=weeks),lv)
lagged=embed(c(rep(0,3),vax),4)

xy=data.frame(x=1:weeks,y=c(1000-lagged)/c(1000-vax),lag,status="Unvaccinated")
xy=rbind(xy,data.frame(x=1:weeks,y=c(lagged)/vax,lag,status="Vaccinated"))

# xy=data.frame(x=1:weeks,y=1,lag,status="Unvaccinated")
# xy=rbind(xy,data.frame(x=1:weeks,y=1,lag,status="Vaccinated"))

xstart=1;xend=weeks;ystart=0;yend=5

pal=c("black",hcl(c(200,250,330)+15,100,50))

ggplot(xy,aes(x,y))+
geom_hline(yintercept=c(ystart,yend),linewidth=.3,lineend="square")+
geom_vline(xintercept=c(xstart,xend),linewidth=.3,lineend="square")+
geom_line(aes(color=lag,linetype=status),linewidth=.4)+
# annotate(geom="label",fill=alpha("white",.8),x=1.24,y=3,hjust=0,vjust=.5,label=paste0("All lines have a constant 1% infection rate if people who are classified as unvaccinated because of the cheap trick are also added to the unvaccinated population size and not only to unvaccinated cases.")|fw(30),size=2.3,label.r=unit(0,"lines"),label.padding=unit(.4,"lines"),label.size=.3,lineheight=1.1)+
# labs(title="Reproduction of Neil and Fenton's scenario A (corrected version where people who are classified as unvaccinated because of the cheap trick are added to both unvaccinated cases and unvaccinated population size)"|>stringr::str_wrap(52),x="Week of simulation",y="Infection rate")+
labs(title="Reproduction of Neil and Fenton's scenario A (original version where people who are classified as unvaccinated because of the cheap trick are only added to unvaccinated cases but not to unvaccinated population size)"|>stringr::str_wrap(52),x="Week of simulation",y="Infection rate")+
scale_x_continuous(limits=c(xstart,xend),breaks=seq(xstart,xend))+
scale_y_continuous(limits=c(ystart,yend),breaks=seq(ystart,yend),labels=\(x)paste0(x,"%"))+
coord_cartesian(clip="off",expand=F)+
scale_color_manual(values=pal)+
guides(linetype=guide_legend(title="Status"),color=guide_legend(title="Cheap trick lag"))+
theme(axis.text=element_text(size=7,color="black"),
  axis.ticks=element_blank(),
  axis.ticks.length=unit(0,"lines"),
  axis.title=element_text(size=8),
  legend.background=element_blank(),
  legend.box.background=element_rect(fill=alpha("white",.85),color="black",linewidth=.3),
  legend.box.just="center",
  legend.direction="vertical",
  legend.justification=c(1,1),
  legend.key=element_rect(fill=alpha("white",0)),
  legend.key.size=unit(.8,"lines"),
  legend.margin=margin(.3,.4,.3,.4,"lines"),
  legend.position=c(1,1),
  legend.spacing.x=unit(.15,"lines"),
  legend.spacing.y=unit(-.2,"lines"),
  legend.text=element_text(size=7,vjust=.5),
  legend.title=element_text(size=8,margin=margin(0,0,.8,0,"lines")),
  panel.background=element_rect(fill="white"),
  panel.grid.major=element_line(linewidth=.3,color="gray88"),
  plot.margin=margin(.5,.7,.3,.4,"lines"),
  plot.title=element_text(size=7.5,margin=margin(.2,0,.4,0,"lines")))
ggsave("1.png",width=3,height=3,dpi=450)

Later I found out that in 2023 Neil and Fenton had published a Substack post where they described their simulation procedure in more detail. And they even included a table similar to the table that Uncle John made, which made it clear that they only applied the cheap trick to the numerator but not to the denominator: [https://wherearethenumbers.substack.com/p/the-illusion-of-vaccine-efficacy]

They could've also included a similar table in their new preprint to make their methodology more clear.

People in Substack comments rarely address any specific details related to the topic of the post, and even if the author makes some obvious error, it's rare for people in the comments to mention it (or if they do then they just get banned like me). But this time several people in the comment section were wondering why the cheap trick was not also applied to the denominator in the models, and people were asking Neil and Fenton to cite any actual study which would've worked like their models so that the cheap trick was only applied to the numerator but not the denominator:

Neil and Fenton failed to cite any actual study which would've worked like their simulation, but they responded to the Substack comments by adding this note to their Substack post: "6 May 2023 Update: Quite a few people have asked why we are including those vaccinated within the last 21 days in the 'total vaccinated' denominator for the vaccinated infection rate if such people are classified as unvaccinated. The answer is that, while those infected within 21 days are classified as unvaccined in observational trials, those who are not infected are generally treated as vaccinated. In observational studies, which is what we are simulating here, there is no pre-defined vaccine and control group as there would be in a controlled trial. And, unlike a controlled trial, people are getting vaccinated at different times. Hence, everything is driven by looking at 'cases' i.e. those defined to have been infected in any given week; all those infected within 21 days of a vaccination are classified as unvaccinated. However, for the efficacy calculation at any time, for the 'denominators' they simply use the total number of people vaccinated so far (e.g. from NIMS) and for this, there is generally no '21 day delay'." However they didn't cite any source for their claim that there was generally no 21-day delay applied to the denominator.

Even though YouTube comments are not generally posted by the sharpest kind of people, even people in Fenton's YouTube comments were saying that his simulation was fishy: [https://www.youtube.com/watch?v=Gkh6N-ZL3_k]

In the old Substack post where Fenton and Neil described their simulation procedure in more detail, even people in their comment section were able to tell that it was BS, so I wonder if that's why they didn't include a proper description of their simulation procedure in their new preprint.

In any case Neil and Fenton knew that their method of doing the simulations was controversial, so they could've mentioned in their preprint that there would've also been an alternative way to perform the simulations where the 1-to-3-week delay was also applied to the denominator, and that they were not sure which method was used in the real studies they listed in their paper.

Added later: Peter Hegarty went through the 39 studies listed in the cheap trick paper with a fine-toothed comb, and he wrote: "Moreover, their text and simulations suggest they were concerned with an even much more egregious form of the bias, whereby only Covid cases amongst the partially vaccinated contributed to the infection rate for the unvaccinated. In other words, partially vaccinated cases were counted in the numerator of the infection rate for the unvaccinated, but partially vaccinated individuals were counted in the denominator for the fully vaccinated. We can state categorically that we found no evidence in any of the 39 studies of this 'trick' being employed." [https://researchgate.net/publication/387220055_A_DETAILED_ANALYSIS_OF_CLAIMS_OF_MISCATEGORIZATION_BIAS_IN_STUDIES_OF_COVID-19_VACCINE_EFFECTIVENESS]

He wrote the following summary of his findings:

Here is a summary of our findings for the 39 papers on NFMs list:

• 2 of the items are the same, thus there are 38 different items.
• 1 item is a meta-analysis and 1 is not a VE study at all. We omit both from consideration.
• NFM claim that each of the remaining 36 items has Miscategorization bias. We found that only 10 items had such bias. Indeed, 8 items were RCTs, for which such bias doesn't even make sense.
• In addition, 1 item had a type of miscategorization bias not covered by NFMs definition, but it affected a tiny proportion of the cohort and was not significant.
• NFM claimed that only 2 items had Exclusion bias. In fact, 28 items had this bias in their primary analysis and it was by far the most common form of statistical bias. - Of the 28 items with Exclusion bias, 7 had secondary analyses which corrected for it more or less in full, while 3 more corrected for it in part. In all 10 cases, corrected VE remained significant.
• In a further 2 items with Exclusion bias and 1 with Miscategorization bias, it was clear that the bias affected too small a proportion of the cohort to be significant.
• 1 item had none of the biases defined by NFM at all.
• NFM claimed that 4 items had Unverified bias, whereas we found only 1 item definitely had that bias, and for 1 other item we were not sure.
• NFM claimed that 5 items had Uncontrolled bias, but we found no item with that bias.
• NFM claimed that 2 items had Undefined bias, but we found no item with that bias.
• Overall, we found that NFM correctly assigned categories of bias to only 6 out of 39 items (#3, 7, 28, 34, 36, 39).
• Of the 22 items with either Miscategorization or Exclusion bias which could not be corrected for from the data provided in the study, only 1 gave a clear suggestion of negative VE in the period soon after a first vaccination. But VE was already modest in this study for fully vaccinated individuals, and the vaccine involved was SpikoGen, a non- mRNA vaccine deployed outside of Western countries. Of the 29 studies mainly focused on mRNA vaccines, 1 had evidence for negative VE, even for fully vaccinated individuals, against Covid-19 infection but not against more severe outcomes. Otherwise, there was scant evidence that, in NFMs words, "any claims of vaccine efficacy are likely to be a statistical illusion" resulting from the biases they defined.

Added in February 2025: Neil and Fenton now wrote that when they tried to submit their cheap trick paper to a journal, it didn't pass peer review. [https://wherearethenumbers.substack.com/p/dirty-deeds-done-dirt-cheap] They published a list of comments written by one of their reviewers, who didn't seem to initially understand that the classification delay was only applied to the numerator in the simulations, because the reviewer wrote: "Can you explain why in Figure 1, Panel A, the infection rate would increase for the unvaccinated in week 2? The unvaccinated group is composed of vaccinated people and unvaccinated people because of the delay but both groups have been assigned an infection rate of 1% right? And why is the rate for the unvaccinated group consistent across time regardless of the amount of delay used?" The reviewer also wrote: "Do you have a citation for the statement that miscategorisation selection biases inevitably exaggerate vaccine efficacy? I would expect nondifferential exposure misclassification to move the effect estimate towards the null while differential exposure misclassification could bias the estimate in any direction. See Rothman, page 137." Like me, the reviewer also pointed out that the methodology of the simulations was not described in a detailed enough way: "I'd like more information on the simulation methods. For example, how are you calculating vaccine effectiveness? And how do you treat the infected population - do they recover, are they eligible for reinfection?" Then Neil and Fenton wrote that they added a supplementary file to their paper which described the methodology of their simulations more clearly, so now the same reviewer wrote:

Thank you for providing the supplementary materials - this clarifies your simulation methods and explains your results. Unfortunately, you have made a major error in your simulation design.

Using Scenario A, Case 1 as an example, in week 2, you calculate the infection rate among the vaccinated as Cases in fully vaccinated (>1 week) divided by Cumulative ever vaccinated (i.e., 100/210,000). The denominator of this calculation is incorrect since the 200,000 newly vaccinated individuals in week 2 should not count as vaccinated people - their outcomes are attributed to the unvaccinated population so they should also be part of the unvaccinated population, in accordance with both fundamental study design principles and as stated in the specific study design decision that is the topic of this paper. The correct calculation for the vaccinated rate in week 2 is Cases in fully vaccinated (>1 week) divided by Total fully vaccinated (>1 week) which is 100/10,000 resulting in a rate of 1%, which is consistent with your simulation specifications. This also corrects the calculations for the unvaccinated population - you currently do (new cases + cases in newly vaccinated)/Total unvaccinated when the correct calculation is (new cases + cases in newly vaccinated)/(Total unvaccinated + Newly vaccinated). This corrected calculation is (7900+2000)/(790000+200000) resulting in a rate of 1%, which is again consistent with your simulation specifications. Notably, with a rate of 1% in both your vaccinated and unvaccinated groups, this reproduces a vaccine effectiveness of no effect as intended by your simulation. This mistake is also seen in the other scenarios.

In making this mistake, your simulation diverges from the study design decision that you claim introduces bias - when researchers delay assigning individuals to the vaccinated group for 14 days after a vaccination, they assign both any outcomes that occur and these people's follow-up time to the unvaccinated group until 14 days have passed. Your simulations do not recreate this scenario.

I posted this comment to Neil and Fenton at Substack (but they blocked me and deleted my comment):

I now read the updated version of your paper.

1. Your abstract still says: "Simulation demonstrates that this miscategorisation bias artificially boosts vaccine efficacy and infection rates even when a vaccine has zero or negative efficacy." You could've specified that you referred to the form of miscategorization that is only applied to the numerator, but that you're not sure if the numerator-only form of miscategorization was employed in the papers you listed.

2. You also wrote: "This bias, which has been seen to take several different types, all of which help exaggerate vaccine efficacy, has recently become known colloquially as the 'cheap trick' (Fenton & Neil, 2023)."

However if vaccine manufacturers state that the vaccine is not yet designed to be fully effective immediately after vaccination or only after a single dose, then I don't think it exaggerates the stated efficacy of the vaccine if the efficacy is only measured starting from two or three weeks after the second dose.

And also I don't think all types of the bias would necessarily exaggerate efficacy like you claimed. In your first simulation where you assumed a constant infection rate and you assumed that vaccines have zero efficacy, if you also apply the classification delay to the numerator, then both unvaccinated and vaccinated people have a constant 1% reported infection rate, so the reported vaccine efficacy is zero which is the same as the actual efficacy in your simulation.

3. You also wrote: "We focus explicitly on miscategorisation biases, which in the Covid-19 studies cited have inevitably exaggerated vaccine efficacy." I would be more wary of using words like "inevitably". If the miscategorization bias is also applied to the denominator, then it does not inevitably exaggerate vaccine efficacy. If for example there was a study where there was a low incidence of COVID during the first half of the observation period and a high incidence of COVID during the second half, and where the risk ratio of vaccinated people relative to unvaccinated people would not be adjusted for the amount of observation time during different periods, then the misclassification might end up underestimating vaccine efficacy because it would reduce the amount of vaccinated observation time during the low-COVID period and it would increase the amount of unvaccinated observation time during the low-COVID period.

4. You wrote: "The later method (denominators unadjusted) is the convention, and our results are calculated using that method." But you should've provided a more detailed description of why you claim it is the convention.

5. You also wrote: "In practice, most studies do not report vaccine efficacy in the initial week(s) (when no cases are categorised as vaccinated) as this would show up as 100% efficacy." But you should've specified that your statement was dependent on the assumption that the cheap trick was only applied to the numerator.

motherfunkr: Unvaccinated ASMR reached near vaccinated ASMR during the last month of data included in the ONS dataset

The 9th and final version of the ONS dataset for deaths by vaccination status was published in August 2023. It included data up to May 2023, but the reason why June and July were excluded from the release was probably because they were still missing too many deaths because of a registration delay. In May it looked like the trend was that the unvaccinated ASMR was soon about to cross below the ever vaccinated ASMR: [https://x.com/mothrfunkr/status/1774625071140893040]

However I think it's because younger people are more likely to be unvaccinated than older people, and the registration delay for deaths is longer in younger people, so during the last months with data available there's a larger percentage of deaths missing in younger people, which introduces a bias where it reduces unvaccinated ASMR more than vaccinated ASMR.

Similarly in the July 2022 release of the ONS dataset, the last month of data included was May 2022 when unvaccinated people had only about 6% higher ASMR than vaccinated people, and it looked like the trend was that unvaccinated ASMR was soon about to cross below vaccinated ASMR. But in the next release after more missing deaths were added, the unvaccinated ASMR in May 2022 became about 22% higher than vaccinated ASMR:

In the year 2021 in an ONS dataset titled "Impact of registration delays on mortality statistics", the proportion of deaths which had a registration delay of 3 months or longer was about 12% in ages 45-64 but only about 2% in ages 85 and over: [https://www.ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/articles/impactofregistrationdelaysonmortalitystatisticsinenglandandwales/2021]

Clare Craig: New baseline used by ONS to calculate excess deaths

In February 2024 the UK ONS announced that they had changed the methodology they used to calculate excess mortality. [https://www.ons.gov.uk/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/articles/estimatingexcessdeathsintheukmethodologychanges/february2024]

Previously they used a simple 5-year average where the year 2020 was excluded but the years 2021-2023 were not, so for example the baseline for the year 2023 was based on the years 2017-2019 and 2021-2022. However now they simply excluded a hardcoded set of months when COVID was listed as the underlying cause of death for at least 15% of all deaths, which were April and May 2020 and November 2020 to February 2021 (or weeks 14-22 of 2020 and week 45 of 2020 to week 8 of 2021 for weekly data).

A lot of people were saying that the new method inflated excess deaths because 2020 and later years were included in the fitting period of the baseline, but people didn't realize that the old method also included the years 2021-2023. The old 5-year average underestimated excess deaths because UK has an increasing trend in deaths because of the aging population, but the old method overestimated excess deaths relative to the prepandemic trend because 2021-2023 were included in the baseline fitting period. So the old method was inaccurate in two ways, which ended up partially canceling each other out starting from the year 2022 when 2021 started to be included in the baseline fitting period.

Clare Craig wrote: [https://www.hartgroup.org/too-many-deaths-are-to-be-expected/]

Figure 1: ONS expected deaths estimates showing smooth predictions of old methodology (black dotted line) compared to jumping estimates using modelling methodology (red line) and actual deaths (grey bars). Note y axis starts at 500,000

The consequence of the crazy estimate their new model spits out is that a total of 110,000 more deaths were "expected" from 2016-2019 compared to their old model. It also means that 2019 went from having an excess of 3k deaths to a deficit of 36k. If 36,000 people who were "expected" to die in 2019 did not have their deaths registered, then surely that paints the excess in registrations in 2020 in a different light.

However UK has an increasing trend in deaths per year because of the aging population, so in the years 2012 to 2019, the number of deaths was always above the average of the past five years. But a linear trend of the past five years was a better approximation of actual deaths, because the 5-year average was lagging 3 years behind the trend:

The ONS article about the new baseline methodology was accompanied by a spreadsheet which shows monthly and weekly deaths in the years 2005 to 2023 by age, region, and sex. In the code below I calculated the yearly sum of deaths in England and Wales based on Table 2 of the spreadsheet. When I used the average of the five previous years as the baseline for each year, there was a total of 68,840 excess deaths in 2016 to 2019. But when I used a linear trend of the five previous years as the baseline for each year, there's a total of -29,476 excess deaths in 2016 to 2019 instead. So difference is about 100,000 deaths, which is similar to the difference between the new and old ONS method that was mentioned by Craig:

download.file("https://www.ons.gov.uk/file?uri=/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/datasets/estimatingexcessdeathsintheukmethodologychanges/current/dataset20240220.xlsx","dataset20240220.xlsx")
t=as.data.frame(readxl::read_excel("dataset20240220.xlsx",skip=4,sheet=5))
t=t[t$Geography=="England and Wales, including non-residents",]
d=aggregate(list(dead=t$Death),list(year=t$Year),sum)

# # same as above
# d=data.frame(year=2005:2023)
# d$dead=c(512817,502405,503838,508906,491140,493062,484178,499145,506608,501243,529472,524866,533047,541568,530841,607922,586334,577377,581368)

d$linear=c(rep(NA,5),sapply(2010:2023,\(i)predict(lm(dead~year,d[d$year%in%(i-1:5),]),list(year=i))))
d$average=c(rep(NA,5),sapply(2010:2023,\(i)mean(d$dead[d$year%in%(i-1:5)])))

d$excess_linear=d$dead-d$linear
d$excess_average=d$dead-d$average

print.data.frame(round(na.omit(d)),row.names=F)
# year   dead linear average excess_linear excess_average
# 2010 493062 492765  503821           297         -10759
# 2011 484178 490455  499870         -6277         -15692
# 2012 499145 479676  496225         19469           2920
# 2013 506608 487341  495286         19267          11322
# 2014 501243 505932  494827         -4689           6416
# 2015 529472 508485  496847         20987          32625
# 2016 524866 531935  504129         -7069          20737
# 2017 533047 534559  512267         -1512          20780
# 2018 541568 541998  519047          -430          22521
# 2019 530841 551307  526039        -20466           4802
# 2020 607922 537791  531959         70131          75963
# 2021 586334 596821  547649        -10487          38685
# 2022 577377 611821  559942        -34444          17435
# 2023 581368 606942  568808        -25574          12560

round(colSums(d[d$year%in%2016:2019,-1]))
#    dead   linear_trend        average  excess_linear excess_average
# 2130322        2159798        2061482         -29476          68840

Craig also pointed out that the new ONS method produced a similar expected number of deaths for 2020, 2021, and 2022: [https://x.com/ClareCraigPath/status/1778336819270074865]

The new method uses a 5-year trend calculated with a quasi-Poisson regression. There was a big drop in the baseline between 2019 and 2020, because the baseline for the year 2020 was based on the years 2015-2019, and there was a low number of deaths in 2019 and a high number of deaths in 2015, so the slope of the trend was too low. The slope of a 2015-2019 linear trend is also too flat relative to the actual long-term trend. OWID also uses a 2015-2019 linear trend for all countries, so it overestimates excess deaths in countries that had a low number of deaths in 2019, like UK and Sweden.

In the plot below, I think my green baseline is more accurate than the new method used by the ONS. I first calculated a linear trend in crude mortality rate within each 5-year age group in 2010-2019. Then I multiplied the projected trend by the population size of the age group to get the expected number of deaths for the age group, and I added together the expected number of deaths for all age groups to get the total expected deaths for a year. I used a fairly long fitting period for the baseline, so anomalous years like 2019 have less impact on my baseline than in the new ONS method. But compared to my green baseline, the new ONS baseline actually seems far too low in the years 2021 and 2022:

library(readxl);library(data.table);library(ggplot2)

download.file("https://www.ons.gov.uk/file?uri=/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/datasets/estimatingexcessdeathsintheukmethodologychanges/current/dataset20240220.xlsx","dataset20240220.xlsx")

t=data.table(read_excel("dataset20240220.xlsx",skip=4,sheet=6))
t=t[Geography=="England and Wales, including non-residents"]
t[,pop:=`Population estimate`*lubridate::days_in_month(paste0(Year,"-",match(Month,month.name),"-1"))]
t[,age:=as.numeric(sub(" .*","",sub("Less.*",0,`Age group`)))]

a=t[,.(dead=sum(`Death registrations`),pop=sum(pop)),.(age,year=Year)]
years=unique(a$year)
a=merge(a,a[year%in%2010:2019,.(year=years,base=predict(lm(dead/pop~year),.(year=years))),age])[,base:=base*pop]
p=a[,.(dead=sum(dead),base=sum(base)),year]

t2=data.table(read_excel("dataset20240220.xlsx",skip=4,sheet=8))
t2=t2[Country=="England and Wales, including non-residents"&Sex=="Both sexes"&`Age group`=="All ages"]
p=merge(t2[,.(newons=sum(`Expected deaths`)),.(year=Year)],p,all=T)

p$oldons=c(rep(NA,5),sapply(2010:2023,\(i)mean(tail(p$dead[p$year<i&p$year!=2020],5))))

p$base[p$year<2010]=NA

p=p[,.(x=year,y=unlist(p[,-1]),z=rep(names(p)[-1],each=.N))]

p[,z:=factor(z,c("dead","newons","oldons","base"))]
color=c("black","#dd0000","gray50","#00aa00")
lab=c("Actual deaths","New ONS baseline (quasi-Poisson regression by age, sex, and location)","Old ONS baseline (average of past 5 years with 2020 excluded)","My baseline (derived from 2010-2019 trend in CMR by age group)")

xstart=2005;xend=2023;xbreak=xstart:xend
xlab=c(rbind("",xstart:xend),"")
ymin=min(p$y,na.rm=T);ymax=max(p$y,na.rm=T)
cand=c(sapply(c(1,2,5),\(year)year*10^c(-10:10)))
ystep=cand[which.min(abs(cand-(ymax-ymin)/5))]
ystart=ystep*floor(ymin/ystep)
yend=ystep*ceiling(ymax/ystep)
ybreak=seq(ystart,yend,ystep)

leg=data.frame(x=xstart+(xend-xstart)*.027-.5,y=ystart+(yend-ystart)*seq(.94,0,-1/15)[1:nlevels(p$z)],label=lab)

ggplot(p,aes(x=x,y=y))+
geom_hline(yintercept=c(ystart,yend),color="gray65",linewidth=.3)+
geom_vline(xintercept=c(xstart-.5,xend+.5),color="gray65",linewidth=.3)+
geom_line(aes(color=z),linewidth=.4)+
geom_point(data=subset(p,z=="dead"),aes(color=z),size=.5)+
geom_label(data=leg,aes(x=x,y=y,label=label),fill=alpha("white",.8),label.r=unit(0,"lines"),label.padding=unit(.04,"lines"),label.size=0,color=color,size=2.7,hjust=0)+
labs(title="Deaths in England and Wales by registration date",x=NULL,y=NULL)+
scale_x_continuous(limits=c(xstart-.5,xend+.5),breaks=seq(xstart-.5,xend+.5,.5),labels=xlab)+
scale_y_continuous(limits=c(ystart,yend),breaks=ybreak,labels=\(x)paste0(x/1e3,"k"))+
coord_cartesian(clip="off",expand=F)+
scale_color_manual(values=color)+
theme(axis.text=element_text(size=7.5,color="black"),
  axis.text.x=element_text(angle=90,vjust=.5,hjust=1),
  axis.ticks=element_line(linewidth=.3,color="gray65"),
  axis.ticks.x=element_line(color=alpha("gray65",c(1,0))),
  axis.ticks.length=unit(.2,"lines"),
  axis.title=element_text(size=8),
  legend.position="none",
  panel.background=element_rect(fill="white"),
  panel.grid=element_blank(),
  plot.background=element_rect(fill="white"),
  plot.margin=margin(.4,.6,.4,.4,"lines"),
  plot.subtitle=element_text(size=7.2),
  plot.title=element_text(size=8.8))
ggsave("0.png",width=4.5,height=3,dpi=400*4)

sub="\u00a0     Source: ons.gov.uk/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/\ndatasets/estimatingexcessdeathsintheukmethodologychanges, tables 2 and 4. Includes non-residents.
      The green baseline was calculated by first taking the linear trend in CMR by five-year age groups in 2010-2019, and then the projected trend was multiplied by the population size of each age group to get the expected deaths, and the expected deaths for each age group were added together to get the yearly total expected deaths.
      The gray baseline is a simple average of the past five years with 2020 excluded, so for example the baseline for 2023 consists of the years 2017-2019 and 2021-2022."

system(paste0("magick 0.png -resize 25% -trim 1.png;magick 1.png \\( -size `identify -format %w 1.png`x -font Arial -interline-spacing -3 -pointsize 42 caption:'",sub,"' -splice x30 \\) -append -bordercolor white -border 36 1.png"))

People were speculating that ONS adopted the new method of calculating the baseline in order to diminish the number of excess deaths caused by the vaccines. But my plot above shows that the new ONS baseline is likely far too low in the years 2021 and 2022, which would mean that it actually exaggerates excess deaths during the years when the most vaccines were given.

Craig also wrote: "However, you can't hide the truth forever. All that is needed is the calculation of an age-standardised mortality rate. This is a way of mapping mortality from a real population to one that has a standard age structure so different regions can be fairly compared with each other and across time. Using that methodology it is impossible to hide the issue of excess mortality." In the code below I calculated ASMR for England and Wales using the spreadsheet that accompanied the methodology article by the ONS. I used the 2020 population as the standard population, and I used the 2005-2019 second-degree polynomial trend in ASMR as the baseline, which gave me only about 0.12% excess ASMR in the years 2016-2019 (so it's closer to the new ONS baseline than to the old ONS baseline):

t=as.data.frame(readxl::read_excel("dataset20240220.xlsx",skip=4,sheet=5))
t=t[t$Geography=="England and Wales, including non-residents",]

dim=list(age=as.numeric(sub(" .*","",sub("Less.*",0,t$Age))),year=t$Year)
a=aggregate(list(dead=t$Death),dim,sum)
a$pop=c(tapply(t$Pop,dim,mean))*2

std=a$pop[a$year==2020]
d=aggregate(list(asmr=a$dead/a$pop*std[factor(a$age)]/sum(std)*1e5),a[,"year",drop=F],sum)

d$trend=predict(lm(asmr~year,subset(d,year%in%2010:2019)),d)
d$excesspct=(d$asmr/d$trend-1)*100

d$pop=tapply(a$pop,a$year,sum)
d$excessdead=(d$asmr-d$trend)/1e5*d$pop

print.data.frame(round(d),row.names=F)
# year asmr trend excesspct      pop excessdead
# 2005 1100   987        11 53569582      60692
# 2006 1062   982         8 53958732      43042
# 2007 1050   977         7 54389720      39671
# 2008 1048   972         8 54834344      41695
# 2009  994   967         3 55243469      14567
# 2010  978   963         2 55695362       8736
# 2011  941   958        -2 56162059      -9383
# 2012  951   953         0 56578624      -1114
# 2013  949   948         0 56995298        550
# 2014  921   943        -2 57442224     -13077
# 2015  959   939         2 57887460      11881
# 2016  936   934         0 58347631       1158
# 2017  936   929         1 58697682       3941
# 2018  937   924         1 59008790       7575
# 2019  902   919        -2 59293219     -10239
# 2020 1023   915        12 59445253      64287
# 2021  973   910         7 59704283      37939
# 2022  941   905         4 60243501      21604
# 2023  931   900         3 60856738      18938

with(d[d$year%in%2016:2019,],sum(asmr)/sum(trend)-1)*100
# 0.115167 (total excess ASMR percentage in 2016-2019)

sum(d$excessdead[d$year%in%2016:2019])
# 2435.476 (total excess deaths in 2016-2019 derived from ASMR)

png("1.png",800,500,res=150)
par(mar=c(2.3,2.3,2.1,.8),mgp=c(0,.6,0))
tit="ASMR in England and Wales"
leg=c("Actual","Baseline")
col=c("black","red")
plot(d$year,d$asmr,type="l",col=col[1],main=tit,xlab=NA,ylab=NA,lwd=1.5)
lines(d$year,d$trend,type="l",col=col[2],lwd=1.5)
legend("topright",legend=leg,col=col,lty=1,lwd=1.5)
dev.off()

This plot also shows that regardless of whether I used a 2010-2019 linear trend or 2005-2019 polynomial trend to calculate excess ASMR, my total excess ASMR percentage in 2021-2023 was lower than the total excess mortality percent in 2021-2023 using the new ONS baseline:

library(colorspace)

t=as.data.frame(readxl::read_excel("dataset20240220.xlsx",skip=4,sheet=5))
t=t[t$Geography=="England and Wales, including non-residents",]

dim=list(age=as.numeric(sub(" .*","",sub("Less.*",0,t$Age))),year=t$Year)
a=aggregate(list(dead=t$Death),dim,sum)
a$pop=c(tapply(t$Pop,dim,mean))*2

std=a$pop[a$year==2020]
asmr=tapply(a$dead/a$pop*std[factor(a$age)]/sum(std)*1e5,a$year,sum)
asmr=data.frame(year=unique(a$year),asmr)

trend=predict(lm(asmr~year,subset(asmr,year%in%2010:2019)),asmr)
trend2=predict(lm(asmr~poly(year,2),subset(asmr,year<2020)),asmr)

d=aggregate(a[,3:4],a[,2,drop=F],sum)

m=data.frame("ASMR using 2010-2019 linear trend"=(asmr$asmr/trend-1),check.names=F)
rownames(m)=2005:2023
m$"ASMR using 2005-2019 polynomial trend"=(asmr$asmr/trend2-1)

ave=c(rep(NA,5),sapply(2010:2023,\(i)mean(tail(d$dead[d$year<i&d$year!=2020],5))))
m$"Old ONS method (average of past 5 years without 2020)"=d$dead/ave-1

t4=as.data.frame(readxl::read_excel("dataset20240220.xlsx",skip=4,sheet=7))
t4=t4[t4$Country=="England and Wales, including non-residents"&t4$Sex=="Both sexes"&t4$Age=="All ages",]
new=tapply(t4$"Expected deaths",factor(t4$Year,2005:2023),sum)
m$"New ONS method (5-year quasi-Poisson regression)"=d$dead/new-1

ag=aggregate(list(dead=t$Death,pop=t$Pop),list(age=as.numeric(sub(" .*","",sub("Less.*",0,t$Age))),year=t$Year),sum)
ag$trend=c(t(sapply(split(ag,ag$age),\(x)lm(dead/pop~year,x[x$year%in%2010:2019,])|>predict(list(year=unique(ag$year))))))
cmrbased=tapply(ag$trend*ag$pop,ag$year,sum)
m$"Population size times 2010-2019 linear trend in CMR by age"=d$dead/cmrbased-1

owid=predict(lm(dead~year,subset(d,year%in%2015:2019)),d)
m$"Raw deaths using 2015-2019 linear regression (like OWID)"=d$dead/owid-1

m=t(m*100)[,-(1:5)]
maxcolor=max(abs(m),na.rm=T)
pal=hex(HSV(c(210,210,210,210,0,0,0,0,0),c(1,.8,.6,.3,0,.3,.6,.8,1),c(.3,.65,1,1,1,1,1,.65,.3)))

pheatmap::pheatmap(m,filename="0.png",display_numbers=T,number_format="%.1f",
  cluster_rows=F,cluster_cols=F,legend=F,cellwidth=20,cellheight=20,fontsize=9,fontsize_number=8,
  border_color=NA,na_col="gray90",
  number_color=ifelse(abs(m)>maxcolor*.6,"white","black"),
  breaks=seq(-maxcolor,maxcolor,,256),
  colorRampPalette(pal)(256))

system("f=0.png;w=`identify -format %w $f`;convert $f -gravity northwest \\( -splice x16 -size $[w-44]x -pointsize 38 caption:'Excess mortality percent in England and Wales (by date of registration, includes non-residents). Source: ONS dataset titled \"Estimating excess deaths in the UK, methodology changes\".
      ASMR was calculated using 5-year age groups so that the 2020 population of England and Wales was used as the standard population.
      On the row labeled \"Population size times 2010-2019 linear trend in CMR by age\", a linear trend was calculated for crude mortality rate within each 5-year age group in 2010-2019, and then the projected trend was multiplied by the yearly population sizes of each age group, and the expected deaths for all age groups were added together to get the yearly total expected deaths.' -extent $[w-44]x -gravity center \\) +swap -append -bordercolor white -border 6 +repage 1.png")

Craig wrote: "The ONS methodology is complex and opaque - for example, they have failed to publish any of the weightings they have used in their formula." [https://www.hartgroup.org/too-many-deaths-are-to-be-expected/] However the ONS published their R code at GitHub. [https://github.com/ONS-Health-modelling-hub/Excess_deaths/blob/main/ons_monthly_ed%2eR] Here's a simplified version of the code which just prints the expected number of deaths for England and Wales in December 2023:

library(data.table);library(readxl)

download.file("https://www.ons.gov.uk/file?uri=/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/datasets/estimatingexcessdeathsintheukmethodologychanges/current/dataset20240220.xlsx","dataset20240220.xlsx")

target=202312

d=setDT(read_excel("dataset20240220.xlsx",skip=4,sheet=6))
names(d)=c("year","month","weekdays","age","agecoarse","sex","geography","dead","pop")
d=d[geography=="England and Wales, including non-residents"]

d[,yearmonth:=100*year+match(month,month.name)]
d[,dateindex:=.GRP,.(year,month)]

sub=d[dateindex%in%(dateindex[yearmonth==target][1]-12:71)]
sub=sub[!yearmonth%in%c(202004,202005,202011,202012,202101,202102)]
lm=glm(dead~age+sex+dateindex+month+weekdays+agecoarse:sex+agecoarse:dateindex+agecoarse:month+offset(log(pop)),quasipoisson,sub)

sum(predict(lm,d[yearmonth==target],type="response"))
# 49728.14 (same as cell H7649 of table 4 in the spreadsheet)

Or this version always uses 2013-2019 as the fitting period instead which is probably more accurate:

d=setDT(read_excel("dataset20240220.xlsx",skip=4,sheet=6))
names(d)=c("year","month","weekdays","age","agecoarse","sex","geography","dead","pop")
d=d[geography=="England and Wales, including non-residents"]

d[,dateindex:=.GRP,.(year,month)]

sub=d[year%in%2013:2019]
lm=glm(dead~age+sex+dateindex+month+weekdays+agecoarse:sex+agecoarse:dateindex+agecoarse:month+offset(log(pop)),quasipoisson,sub)
d$expected=predict(lm,d,type="response")

At first I thought that the regression above was based on raw number of deaths and not mortality rates, but the methodology article says that when the population size is included as an offset variable, it's analogous to doing the regression on mortality rates: "When using a quasi-Poisson regression model, modelling the number of deaths as the dependent variable, and including the natural logarithm of population size as an offset term is analogous to modelling the mortality rate in each age-sex-geography stratum." [https://www.ons.gov.uk/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/articles/estimatingexcessdeathsintheukmethodologychanges/february2024]

ONS is not the only organization which has switched away from using a simple average baseline. OWID previously also used a 2015-2019 average to calculate excess deaths, but they later switched to a 2015-2019 linear trend which is usually more accurate. OWID's website says: "Previously we used a different expected deaths baseline: the average number of deaths over the years 2015-2019. We made this change because using the five-year average has an important limitation - it does not account for year-to-year trends in mortality and thus can misestimate excess mortality. The WMD projection, on the other hand, does not suffer from this limitation because it accounts for these year-to-year trends." [https://ourworldindata.org/excess-mortality-covid]

The Australian Bureau of Statistics previously used a 2015-2019 average to calculate excess deaths during COVID, but in 2023 they switched to a more sophisticated cyclical linear regression method which reduced the excess deaths during COVID, because Australia has an increasing trend in deaths per year similar to most OECD countries. [https://www.abs.gov.au/articles/measuring-australias-excess-mortality-during-covid-19-pandemic-until-august-2023]

TheRustler83 has been tweeting about how the ONS adopted the new method to calculate the baseline because they were trying to hide excess deaths caused by vaccines. I posted the plot for ASMR in England and Wales below to demonstrate to him that when I used a 2015-2019 average baseline, I got negative total excess ASMR in 2024, and I got much lower excess mortality in the post-vaccination era compared to a 2015-2019 linear trend. But I guess in the case of ASMR he would not be in favor of using an average baseline instead of a linear trend, because there was a decreasing trend in ASMR before COVID so the average baseline would reduce the number of excess deaths in the post-vaccine era:

OHID also uses a quasi-Poisson regression to calculate the baseline in their dataset for excess deaths in England: [https://fingertips.phe.org.uk/documents/EMMethodology.pdf]

Quasi-Poisson regression models were fitted on the logarithmic scale [reference 13]. Quasi-Poisson models were used because when counts of weekly deaths are independent of one another they theoretically follow a Poisson distribution. This has the characteristic property that as its mean (the expected number of deaths) increases, the variability of the observed count of deaths (its variance) rises in parallel such that the variance always equals the mean.

However, in the real world, the underlying risk of death varies between different population subgroups and as this cannot usually be modelled perfectly, observed counts of deaths are not completely independent. In consequence, the variance then increases faster than the mean and this is referred to as "overdispersion". Because Quasi-Poisson models allow the linear relationship between variance and mean to have a slope other than unity, they appropriately analyse rates of death when overdispersion exists.

The models contained the set of covariates outlined in the 'Data structure and covariates' section above. To allow for effects to vary between groups, interaction terms were added between age and sex, age and deprivation, age and time of year, age and ethnicity and ethnicity and deprivation. Population sizes in each subgroup are accounted for in the model as an offset.

The model generates expected death rates for each population subgroup for each week, which are then applied to relevant population estimates to estimate the expected numbers of deaths for each week in each subgroup.

canceledmouse: Neonatal deaths in Scotland

canceledmouse posted this plot of neonatal deaths in Scotland, and he suggested that the spikes in deaths in September 2021 and March 2022 might have been caused by the vaccines: [https://x.com/canceledmouse/status/1782970982367371664, https://scotland.shinyapps.io/phs-covid-wider-impact/]

However the data has so much noise that I don't know if the spikes in deaths can be blamed on the vaccines. His plot only showed data up to June 2023, so you weren't able to see that there was a third spike in neonatal deaths in July 2023. And also stillbirths and postneonatal deaths peaked during completely different months than neonatal deaths. Stillbirths peaked in July 2020 and April 2023 which is difficult to blame on the vaccines, and postneonatal deaths peaked in September 2020 and April 2023:

Eurostat only had yearly data for infant mortality, but when I selected countries that had a neonatal mortality rate available for each year out of 2010-2022, the average mortality rate actually decreased from about 24 deaths per 10,000 live births in 2020 to about 22 in 2021 (even though it would probably be better to calculate total neonatal deaths across all countries divided by total live births across all countries, so then small countries like Luxembourg with a lot of noise wouldn't be given that much weight, or I could've used a weighted average by population size instead of a regular average): [https://ec.europa.eu/eurostat/web/population-demography/demography-population-stock-balance/database]

The new version of the Scottish dashboard now shows quarterly data, but the rate of neonatal deaths was lower in July to September 2021 than in two other quarters: [https://scotland.shinyapps.io/phs-pregnancy-births-neonatal/]

It's hard to see any clear pattern in the yearly data either: [https://scotland.shinyapps.io/phs-pregnancy-births-neonatal/]

t=setDT(read_excel("c/stillbirths_and_infant_deaths_2025-06-12.xlsx",skip=6))

s=t[!Date%like%"2020-"&!`Sub-category`%like%"total|extended|infant"]
s=merge(s,s[`Sub-category`=="neonatal deaths",.(births=Denominator),Date])

p=s[,.(y=sum(Numerator)/sum(births)*1000),.(x=as.integer(substr(Date,1,4)),facet=`Sub-category`)]
p[,facet:=paste0(toupper(substr(facet,1,1)),sub("^.","",facet))]
p[,facet:=factor(facet,unique(facet))]
p=rbind(p,p[,.(y=sum(y),facet="Total"),x])

xstart=2016;xend=2025;xbreak=seq(xstart,xend,2)

ylim=p[,.(min=min(y),max=max(y)),facet]
rat=ylim[,max(max/min)]
ylim=ylim[,{x=(rat*min-max)/(1+rat);.(min=min-x,max=max+x,facet)}]

ggplot(p)+
facet_wrap(~facet,ncol=2,dir="v",scales="free")+
geom_rect(data=ylim,aes(ymin=min,ymax=max),xmin=xstart-.5,xmax=xend+.5,lineend="square",linejoin="mitre",fill=NA,color="gray75",linewidth=.4)+
geom_line(aes(x,y),linewidth=.6)+
geom_point(aes(x,y),stroke=0,size=1.5)+
labs(title="Stillbirths and infant deaths per thousand live births in Scotland",x=NULL,y=NULL)+
scale_x_continuous(limits=c(xstart-.5,xend+.5),breaks=xbreak)+
scale_y_continuous(breaks=pretty,labels=\(x)sprintf("%.1f",x))+
scale_alpha_manual(values=c(1,0))+
coord_cartesian(clip="off",expand=F)+
theme(axis.text=element_text(size=11,color="gray45"),
  axis.text.y=element_text(margin=margin(,1)),
  axis.ticks=element_line(linewidth=.4,color="gray75"),
  axis.ticks.length.x=unit(0,"pt"),
  axis.ticks.length.y=unit(4,"pt"),
  panel.background=element_blank(),
  panel.grid=element_blank(),
  panel.spacing=unit(2,"pt"),
  plot.title=element_text(size=11,hjust=.5,face="bold",margin=margin(1,,4)),
  strip.background=element_blank(),
  strip.text=element_text(margin=margin(,,2),size=11))
ggsave("1.png",width=5.1,height=3.3,dpi=300*4)

Clare Craig: May 2024 UKHSA FOI response for deaths in vaccinated people

Monthly deaths in FOI response compared to ONS data

Clare Craig posted this tweet: [https://x.com/ClareCraigPath/status/1796064400060625118]

Her blue line shows the 9th August 2023 edition of the ONS data and the green line shows the 7th July 2022 edition, so there's a gap of over a year between the editions. So therefore a lot of deaths that were missing because of a registration delay in the 7th edition have been added to the 9th edition:

The 3rd edition was similarly missing a lot of deaths that had been added by the 7th edition even though there was a shorter time gap between the editions. (I didn't include the first two editions in the plot above, because they showed weekly instead of monthly data.)

For some reason the new FOI response has a higher number of deaths in January 2022 than December 2021, even though in the ONS dataset it's the reverse. I thought it might be if the FOI response was by registration date, but that doesn't seem to be the case:

library(ggplot2)

# download.file("https://www.ons.gov.uk/file?uri=/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/adhocs/1343dailydeathsbydateofoccurrence1stjune2014to31stmay2023bysingleyearofageengland/dailydeaths2014to2023england.xlsx","dailydeaths2014to2023england.xlsx")
# download.file("https://www.whatdotheyknow.com/request/deaths_in_nims_database/response/2653782/attach/4/Covid19VAccineDataForThoseWithADeathRecordv5.csv","Covid19VAccineDataForThoseWithADeathRecordv5.csv")
# download.file("https://www.ons.gov.uk/file?uri=/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/datasets/estimatingexcessdeathsintheukmethodologychanges/current/dataset20240220.xlsx","dataset20240220.xlsx")

t=read.csv("Covid19VAccineDataForThoseWithADeathRecordv5.csv")
t=t[t$Dose.Number==1,]
a=aggregate(list(y=t$Record.Count),list(x=sub("../(..)/(....)","\\2-\\1",t$Date)),sum)
a=a[a$x!="2023-08",]
a$z="May 2024 UKHSA FOI"

t0=read.csv("http://sars2.net/f/ons-table-1-2021-december.csv")|>subset(cause=="All causes"&status!="Unvaccinated")
t1=read.csv("http://sars2.net/f/ons-table-1-2022-july.csv")|>subset(cause=="All causes"&status=="Ever vaccinated")
t2=read.csv("http://sars2.net/f/ons-table-1-2023-august.csv")|>subset(cause=="All causes"&status=="Ever vaccinated")
t3=read.csv("http://sars2.net/f/ons-table-1-2023-february.csv")|>subset(cause=="All causes"&status=="Ever vaccinated")

d0=aggregate(list(y=t0$dead),list(x=sprintf("%d-%02d",t0$year,match(t0$month,month.name))),sum)|>cbind(z="ONS 3rd edition (December 2021)")
d1=data.frame(x=sprintf("%d-%02d",t1$year,match(t1$month,month.name)),y=t1$dead,z="ONS 7th edition (July 2022)")
d2=data.frame(x=sprintf("%d-%02d",t3$year,match(t3$month,month.name)),y=t3$dead,z="ONS 8th edition (February 2023)")
d3=data.frame(x=sprintf("%d-%02d",t2$year,match(t2$month,month.name)),y=t2$dead,z="ONS 9th edition (August 2023)")

reg=as.data.frame(readxl::read_excel("dataset20240220.xlsx",sheet=5,skip=4))
reg=reg[!grepl("Wales|Ireland|Scotland",reg$Geography),]
d4=aggregate(list(y=reg$Death),list(x=sprintf("%d-%02d",reg$Year,match(reg$Month,month.name))),sum)
d4=d4[d4$x>="2020-12",]
d4$z="All deaths by registration date"

occ=as.data.frame(readxl::read_excel("dailydeaths2014to2023england.xlsx",sheet=4,range="A6:CP3293"))
occ=aggregate(list(y=rowSums(occ[,-(1:3)])),list(x=sprintf("%s-%02d",occ$Year,occ$Month)),sum)
occ$z="All deaths by date of occurrence"
occ=occ[occ$x>="2020-12",]

xy=rbind(a,d3,d2,d1,d0,occ,d4)
xy$z=factor(xy$z,unique(xy$z))
xy$x=as.Date(paste0(xy$x,"-1"))

xstart=min(xy$x);xend=max(xy$x)
xbreak=sort(c(seq(xstart-15,xend+15,"month"),seq(xstart,xend,"month")))
xlab=c(rbind("",format(seq(xstart,xend,"month"),"%y %b")),"")
cand=c(sapply(c(1,2,5),\(year)year*10^c(-10:10)))
ymin=min(xy$y);ymax=max(xy$y)
ystep=cand[which.min(abs(cand-(ymax-ymin)/6))]
yend=ystep*ceiling(ymax/ystep)
ystart=0
ybreak=seq(ystart,yend,ystep)

color=c("black",hcl(c(310,270,240,180)+15,100,50),"gray35","gray70")

lab1=data.frame(x=xstart+.99*(xend-xstart),y=rev(seq(yend*.07,,yend/13,5)),label=levels(xy$z)[1:5])
lab2=data.frame(x=xstart+.99*(xend-xstart),y=seq(yend*.93,,-yend/13,2),label=levels(xy$z)[6:7])

kim=\(x)ifelse(x>=1e3,ifelse(x>=1e6,paste0(x/1e6,"M"),paste0(x/1e3,"k")),x)

ggplot(xy,aes(x=x,y=y))+
geom_vline(xintercept=seq(as.Date("2021-1-1")-16,xend,"3 month"),color="gray91",linewidth=.3)+
geom_vline(xintercept=seq(as.Date("2021-1-1")-16,xend,"year"),color="gray65",linewidth=.3)+
geom_hline(yintercept=c(ystart,yend),color="gray65",linewidth=.3,lineend="square")+
geom_vline(xintercept=c(xstart-15,xend+15),color="gray65",linewidth=.3,lineend="square")+
geom_line(aes(color=z),linewidth=.4)+
geom_label(data=lab1,aes(x=x,y=y,label=label),fill=alpha("white",.8),label.r=unit(0,"lines"),label.padding=unit(.04,"lines"),label.size=0,color=color[1:5],size=2.6,hjust=1)+
geom_label(data=lab2,aes(x=x,y=y,label=label),fill=alpha("white",.8),label.r=unit(0,"lines"),label.padding=unit(.04,"lines"),label.size=0,color=color[6:7],size=2.6,hjust=1)+
labs(title="Monthly deaths among vaccinated people in England: May 2024 UKHSA FOI response compared to different editions of the ONS dataset for mortality by vaccination status"|>stringr::str_wrap(63),subtitle="Sources: whatdotheyknow.com/request/deaths_in_nims_database, ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/
deaths/datasets/deathsbyvaccinationstatusengland, ons.gov.uk/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/
datasets/estimatingexcessdeathsintheukmethodologychanges, ons.gov.uk/peoplepopulationandcommunity/
birthsdeathsandmarriages/deaths/adhocs/
1343dailydeathsbydateofoccurrence1stjune2014to31stmay2023bysingleyearofageengland."|>stringr::str_wrap(90),x=NULL,y=NULL)+
coord_cartesian(clip="off",expand=F)+
scale_x_date(limits=c(xstart-15,xend+15),breaks=xbreak,labels=xlab)+
scale_y_continuous(limits=c(ystart,yend),breaks=ybreak,labels=kim)+
scale_color_manual(values=color)+
theme(axis.text=element_text(size=6.5,color="black"),
  axis.text.x=element_text(angle=90,vjust=.5,hjust=1),
  axis.ticks=element_line(linewidth=.3,color="gray65"),
  axis.ticks.x=element_line(color=alpha("gray65",c(1,0))),
  axis.ticks.length=unit(.15,"lines"),
  axis.title=element_text(size=8),
  legend.position="none",
  panel.background=element_rect(fill="white"),
  panel.grid=element_blank(),
  plot.background=element_rect(fill="white"),
  plot.margin=margin(.3,.3,.3,.3,"lines"),
  plot.subtitle=element_text(size=6.6,margin=margin(0,0,.6,0,"lines")),
  plot.title=element_text(size=8.2,margin=margin(.1,0,.5,0,"lines")))
ggsave("1.png",width=4,height=3.5,dpi=450)

For some reason in April 2021 the 9th edition has less deaths in ages 90+ than the 7th edition. So were some old people linked to the 2011 census but not the 2021 census? However the population size of ages 90+ is higher in the 9th edition than the 7th edition.

Were dying people vaccinated on their deathbed?

As an explanation why almost all deaths were in vaccinated people in late 2022, Craig suggested that dying people who were previously unvaccinated were vaccinated on their deathbed: [https://x.com/ClareCraigPath/status/1795770648721240528]

It has become almost impossible to die in this country without first being injected.

No longer is the priest by the bed side reading last rites, it's the vaccinator.

I will explain below.

The graph shows the percentage of adult deaths in England that were vaccinated https://whatdotheyknow.com/request/deaths_in_nims_database#incoming-2653782

out of total deaths in England. https://ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/adhocs/1343dailydeathsbydateofoccurrence1stjune2014to31stmay2023bysingleyearofageengland

From 2022, there were more covid deaths in the vaccinated with omicron - but I don't think that is what accounts for this.

This is all cause deaths - which had settled into a predictable pattern in summer 2021.

Even if the vaccinated died more, the unvaccinated should still be dying!

How can there be half the deaths in the unvaccinated in July 2022 compared to July 2021?

By Autumn it was only a quarter.

In summer 2021 the unvaccinated seem to have been left alone.

But in 2022, as well as the vaccinated dying more, it looks like those susceptible to death had been injected before dying.

Same as first graph with zoomed in y-axis showing percentage of deaths in vaccinated.

Vaccinating the dying makes interpretation of any data on deaths almost impossible.

At first I thought that Craig's Twitter thread might have been tongue-in-cheek and she actually meant that there was some anomaly in the data she described and not that unvaccinated people who were soon about to die were actually being vaccinated. Someone else seems to have had the same impression, because they posted this reply to the thread: "Sorry, Clare, I find all that info very hard to take in, but are you saying that if you go to die in hospital, at any age, from any causes, they will vaccinate you? Or are you saying they are fiddling the figures?" However Craig indicated that she was being serious: "I think the former. Unless you die suddenly, I think there's huge pressure still to be injected." [https://x.com/ClareCraigPath/status/1795954376751632471]

However her theory doesn't hold water, because according to the new FOI response there were 269,247 people who died in the second half of 2022, but among them only 193 people had received the first vaccine dose on a week that started in the second half of 2022:

download.file("https://www.whatdotheyknow.com/request/deaths_in_nims_database/response/2653782/attach/4/Covid19VAccineDataForThoseWithADeathRecordv5.csv","Covid19VAccineDataForThoseWithADeathRecordv5.csv")
t=read.csv("Covid19VAccineDataForThoseWithADeathRecordv5.csv")
for(i in grep("Date",colnames(t)))t[,i]=as.Date(t[,i],"%d/%m/%Y")
d1=as.Date("2022-7-1");d2=as.Date("2022-12-31")
t=t[t$Dose.Number==1&t$Date.of.Death%in%d1:d2,]
sum(t$Record.Count) # 269247
sum(t$Record.Count[t$Dose.Date%in%d1:d2]) # 193

The reason why some of Craig's plots showed that nearly 100% of deaths were in vaccinated people in late 2022 was because she took the monthly number of deaths in vaccinated people from the new FOI response which was released in May 2024, but she took the number of deaths in the general population of England from an old ONS user request which was released in July 2023. The deaths in the ONS user request are by date of occurrence and not by registration date, so it has a fairly large number of deaths missing because of a registration delay, and the proportion of missing deaths is particularly high in late 2022 and in younger age groups. [https://www.whatdotheyknow.com/request/deaths_in_nims_database#incoming-2653782, https://www.ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/adhocs/1343dailydeathsbydateofoccurrence1stjune2014to31stmay2023bysingleyearofageengland] In the first plot below I took the number of deaths among the general English population from the same ONS user response that Craig used. But in the second plot I used a more recent dataset which was not missing as many deaths as the old ONS user reponse, so I got a much lower percentage of deaths in vaccinated people:

# download.file("https://www.ons.gov.uk/file?uri=/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/adhocs/1343dailydeathsbydateofoccurrence1stjune2014to31stmay2023bysingleyearofageengland/dailydeaths2014to2023england.xlsx","dailydeaths2014to2023england.xlsx")
# download.file("https://www.whatdotheyknow.com/request/deaths_in_nims_database/response/2653782/attach/4/Covid19VAccineDataForThoseWithADeathRecordv5.csv","Covid19VAccineDataForThoseWithADeathRecordv5.csv")
# download.file("https://www.ons.gov.uk/file?uri=/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/datasets/estimatingexcessdeathsintheukmethodologychanges/current/dataset20240220.xlsx","dataset20240220.xlsx")

t=read.csv("Covid19VAccineDataForThoseWithADeathRecordv5.csv")
t=t[t$Dose.Number==1,]
m=xtabs(t$Record~t$Age+sub("../(..)/(....)","\\2-\\1",t$Date))
ages=sort(as.numeric(sub("[-+].*","",unique(t$Age))))

death=as.data.frame(readxl::read_excel("dailydeaths2014to2023england.xlsx",sheet=4,range="A6:CP3293"))
month=sprintf("%s-%02d",death$Year,death$Month)
death=death[,-c(1:23)]
death=t(rowsum(t(death),cut(as.numeric(colnames(death)),c(ages,Inf),,T,F)))
death=t(rowsum(death,month))

d=as.data.frame(readxl::read_excel("dataset20240220.xlsx",sheet=5,skip=4))
d=d[!grepl("Wales|Ireland|Scotland",d$Geography),]
death2=xtabs(d$Death~pmin(80,as.numeric(sub(" .*","",sub("Less",0,d$Age))))+sprintf("%s-%02d",d$Year,match(d$Month,month.name)))
death2=death2[-(1:5),]

pal=colorRampPalette(colorspace::hex(colorspace::HSV(c(210,210,210,210,0,0,0,0,0),c(1,.8,.6,.3,0,.3,.6,.8,1),c(.3,.65,1,1,1,1,1,.65,.3))))(256)

m1=m
pick=intersect(colnames(m1),colnames(death));m1=m1[,pick];death=death[,pick]
m1=rbind(m1,Total=colSums(m1));death=rbind(death,Total=colSums(death))
m1=m1/death*100

m2=m
pick=head(intersect(colnames(m2),colnames(death2)),-1);m2=m2[,pick];death2=death2[,pick]
m2=rbind(m2,Total=colSums(m2));death2=rbind(death2,Total=colSums(death2))
m2=m2/death2*100

pheatmap::pheatmap(m1,filename="i1.png",display_numbers=round(m1),
  gaps_row=nrow(m1)-1,
  cluster_rows=F,cluster_cols=F,legend=F,cellwidth=17,cellheight=17,fontsize=9,fontsize_number=8,
  border_color=NA,na_col="gray90",number_color=ifelse(abs(m1-100)>50,"white","black"),
  breaks=seq(0,200,,256),pal)

pheatmap::pheatmap(m2,filename="i2.png",display_numbers=round(m2),
  gaps_row=nrow(m2)-1,
  cluster_rows=F,cluster_cols=F,legend=F,cellwidth=17,cellheight=17,fontsize=9,fontsize_number=8,
  border_color=NA,na_col="gray90",number_color=ifelse(abs(m2-100)>50,"white","black"),
  breaks=seq(0,200,,256),pal)

cap="      Deaths among vaccinated people in May 2024 FOI response as percentage of deaths among the general population of England. Sources: whatdotheyknow.com/request/deaths_in_nims_database#incoming-2653782, ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/adhocs/1343dailydeathsbydateofoccurrence1stjune2014to31stmay2023bysingleyearofageengland (by date of occurrence), ons.gov.uk/peoplepopulationandcommunity/healthandsocialcare/causesofdeath/datasets/estimatingexcessdeathsintheukmethodologychanges (by registration date)."
cap=paste0(cap,"
      The dataset for deaths by date of occurrence is missing many deaths in 2023 and 2022 because of a registration delay, and the proportion of missing deaths is particularly high in younger age groups in 2023. The dataset for deaths by registration date seems to be missing many deaths in December 2022, which might be if there was a delay registering deaths during the holidays.
      In the FOI response the age of each person is the age at the date of first vaccination, which might explain why for example the 70-74 age group has over 100% deaths in 2023 and 2022, because by 2023 it includes people who were up to 77 years old when they died.")

system("f=i1.png;mar=44;w=`identify -format %w $f`;convert -gravity northwest \\( -size $[w-mar]x -font Arial-Bold -pointsize 41 caption:'Deaths in general population of England from July 2023 ONS user response ID 1343 (by date of occurrence)' -extent $[w-mar]x -gravity north \\) $f -append l1.png")
system("f=i2.png;mar=44;w=`identify -format %w $f`;convert -gravity northwest \\( -size $[w-mar]x -font Arial-Bold -pointsize 41 caption:'Deaths in general population of England from spreadsheet about new baseline methodology (by registration date)' -extent $[w-mar]x -gravity north \\) $f -append l2.png")

system("convert -gravity northwest \\( -splice x18 l1.png \\) l2.png -append 0.png")

system(paste0("f=0.png;mar=40;w=`identify -format %w $f`;convert -gravity northwest \\( -size $[w-mar]x -font Arial -interline-spacing -5 -pointsize 41 caption:'",gsub("'","'\\\\''",cap),"' -extent $[w-mar]x -gravity north -splice x20 \\) $f -append 1.png"))

Mortality rate by week of vaccination up to the end of 2022

Clare Craig posted plots which showed that people vaccinated during the rollout peak subsequently had a low mortality rate up to the end of 2021, but people vaccinated either before or after the rollout peak had a higher mortality rate:

The same pattern is seen in older age groups.

1. Targeting of the dying in spring 2021, with a high proportion dead by the end of the year
2. Rollout to the healthy with a low proportion who then die
3. Targeting the dying again, perhaps pressuring those declined vaccinations if they then become ill or are admitted to hospital

[...]

Craig said that in during the third phase in the second half of 2021, "the dying" were targeted again because the mortality rate was higher than during the second phase. However I don't know if people who were soon about to die were even overrepresented among the people who were vaccinated in the third phase, because the plot below shows that the mortality rate of people vaccinated in the second half of 2021 is close to the baseline in most age groups (except it's below the baseline in youngest age groups and maybe a bit above it in the two oldest age groups, but on the other hand I didn't account for aging of people over time when I calculated the baseline, so the baseline is too low particularly in the two oldest age groups, because by the end of 2022 for example the age group 70-79 includes people who are up to 81 years old):

In the plot above ages 80+ are far above the baseline from February to April 2021, but it could be if people from the upper end of the age group were overrepresented among vaccinated people. One limitation of the FOI response is that it only has one broad age group for ages 80 and above.

Joel Smalley's comment about my plot

Joel Smalley wrote the following about a plot I posted on Twitter: [https://metatron.substack.com/p/the-ons-is-one-of-the-biggest-sources]

Well, apart from the fact that the deaths are in the wrong place, preventing any meaningful analysis, the missing deaths (the ones that are actually meaningful because they were unusual enough to warrant extra attention of the coroner), actually make a difference!

Case in point, here, courtesy of henjin256, are the updates to the ONS bulletin on COVID vaccine mortality:

You see how many more deaths there really were in the vaccinated (black line resulting for Clare Craig's FOI request) compared to their original bulletin in Dec 2021 (green line) that was seized upon by the MSM to vilify all the dirty unvaccinated?

What a difference three years makes, eh? And we only know because of Clare's diligence. Left to their own devices, the ONS wouldn't have bothered with any further updates.

How much difference? Completely turns it on its head, that's what! You go from a spurious positive vaccine effectiveness to a negative one because, ultimately, there are thousands more deaths reported in the vaccinated (numerator) but the vaccinated population (denominator) hasn't changed [1].

[...]

1. In this case, the numerator is the proportion of deaths that are in the vaccinated and the denominator is the proportion of population that is vaccinated. Both numerator and denominator of deaths will increase as new mortality data is received but it is evidently disproportionately in the vaccinated. That alone is a story.

However the numerator in my plot was the absolute number of deaths in vaccinated people, and not the proportion of deaths. And the denominator did in fact change, because the black line includes deaths among the general resident population of England, but the colored lines which show data from the ONS dataset only include the subset of the English population that is linked in the ONS dataset.

The bulletin for the 9th release of the ONS dataset said: "The 2021 Census linked dataset is based on the population in Census 2021. This allows for analyses to be carried out that require a known living population with known characteristics. We linked deidentified Census 2021 records to NHS numbers using the personal demographics service to obtain NHS numbers for census identifiers. People with no NHS number or multiple entries are not included, and imputed individuals are not included. The individuals were then linked via NHS number to vaccination data from the National Immunisation Management Service (NIMS) and ONS death registrations. The population was restricted to people in England, alive on 1 April 2021 (51,786,812 people). This is 91.6% of the England population on Census Day 2021." [https://www.ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/bulletins/deathsinvolvingcovid19byvaccinationstatusengland/latest] And the bulletin for the 7th release of the ONS dataset said: "The PHDA is a linked dataset combining the 2011 Census, the General Practice Extraction Service (GPES) data for COVID-19 pandemic planning and research, and the Hospital Episode Statistics (HES). It combines demographic and socio-economic factors with pre-existing conditions based on clinical records. The PHDA covers England only and contains a subset of approximately 79% of the population of England aged 10 years." [https://www.ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/bulletins/deathsinvolvingcovid19byvaccinationstatusengland/deathsoccurringbetween1january2021and31may2022]

In the latest version of the ONS dataset for mortality by vaccination status, for example in May 2022 vaccinated people have 3,330,575 person-years, which corresponds to approximately 39,214,835 people (from 3330575/31*365). But in ZIP archive of data from the discontinued the UK Coronavirus Dashboard API, the average cumulative number of vaccinated people across all days of May 2022 is 44,876,003, which is about 14% higher than the number of vaccinated people in the ONS dataset: [https://ukhsa-dashboard.data.gov.uk/covid-19-archive-data-download]

$ wget https://archive.ukhsa-dashboard.data.gov.uk/coronavirus-dashboard/vaccinations.zip;unzip vaccinations.zip
$ grep England vaccinations/2022/cumPeopleVaccinatedFirstDoseByPublishDate_nation_2022.csv|grep 2022-05|awk -F, '/England/&&/2022-05/{a+=$NF}END{printf"%.0f\n",a/NR}'
44876003

Joel Smalley's plots with an incorrect baseline

Joel Smalley posted this plot where vaccinated people in ages 20-29 got high excess deaths relative to his baseline: [https://metatron.substack.com/p/debunking-the-lie-that-covid-vaccines]

Smalley is supposed to have calculated the expected deaths by multiplying the total deaths in England with the percentage of vaccinated people in England. However his percentage of vaccinated people was too low, because he accidentally calculated it by dividing the number of vaccinated people in England with the population estimate in the whole UK (and he also used ages 18-29 to calculate the percentage of vaccinated people in ages 20-29, even though it probably doesn't make much difference here): [https://ln5.sync.com/dl/321e94000/pn8agafm-g5mvaahq-yqrqkatg-8bht5qcs/view/doc/22875650440011]

His population size for ages 20-29 was about 42% bigger than the mid-2021 resident population estimate of ages 20-29 in England:

However his population size for ages 20-29 was only about 2% higher than the mid-2021 resident population estimate of ages 18-29 in the whole UK. [https://www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populationestimates/datasets/populationestimatesforukenglandandwalesscotlandandnorthernireland, file "Mid-2021 edition of this dataset", sheet "MYE2 - Persons"]

The number of people aged 18-29 in England with one or more vaccine dose on June 17th 2022 was 6,831,316 in a spreadsheet published by the ONS, and it 6,823,279 in Smalley's spreadsheet, so the figures were close to the same. [https://www.england.nhs.uk/statistics/wp-content/uploads/sites/2/2022/06/COVID-19-daily-announced-vaccinations-18-June-2022-1.xlsx] So Smalley seems to have divided the number of vaccinated people in England with the population size in the whole UK.

Another reason why Smalley's baseline was too low was that he took the deaths in England from his old ONS user request published in November 2022, where ages 20-29 were still missing a lot of deaths in 2021 due to a registration delay. [https://www.ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/adhocs/15018dailydeathsbydateofoccurrence1june2014to31july2022bysingleyearofageandsexengland] I don't know why he didn't use the updated version of the same user request that was published in July 2023: [https://ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/adhocs/1343dailydeathsbydateofoccurrence1stjune2014to31stmay2023bysingleyearofageengland]

dlf=\(x,y=sub(".*/","",x),...)for(i in 1:length(x))download.file(x[i],y[i],quiet=T,...)
dlf("https://www.whatdotheyknow.com/request/deaths_in_nims_database/response/2653782/attach/4/Covid19VAccineDataForThoseWithADeathRecordv5.csv")
dlf("https://www.ons.gov.uk/file?uri=/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/adhocs/1343dailydeathsbydateofoccurrence1stjune2014to31stmay2023bysingleyearofageengland/dailydeaths2014to2023england.xlsx")
dlf("https://www.ons.gov.uk/file?uri=/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/adhocs/15018dailydeathsbydateofoccurrence1june2014to31july2022bysingleyearofageandsexengland/dailydeathsnov22final.xlsx")

library(data.table);library(ggplot2)

t=fread("Covid19VAccineDataForThoseWithADeathRecordv5.csv")
t=t[`Dose Number`==1]

t[,date:=format(as.Date(`Date of Death Month Commencing`,"%d/%m/%Y"),"%Y-%m")]

p=t[`Age Group At First Dose`%like%"^2",.(y=sum(`Record Count`),z="Craig May 2024 UKHSA FOI (vaccinated only)"),.(x=date)]

dead=setDT(readxl::read_excel("dailydeaths2014to2023england.xlsx",sheet=4,range="A6:CP3293"))
dead$dead=rowSums(dead[,24:33])
p=rbind(p,dead[,.(y=sum(dead),z="Smalley July 2023 ONS user request"),.(x=sprintf("%d-%02d",Year,Month))])

dead0=setDT(readxl::read_excel("dailydeathsnov22final.xlsx",sheet=4,range="A5:CP2988"))
dead0$dead=rowSums(dead0[,24:33])
p=rbind(p,dead0[,.(y=sum(dead),z="Smalley November 2022 ONS user request (used for Smalley's baseline)"),.(x=sprintf("%d-%02d",Year,Month))])

p[,z:=factor(z,unique(z))]
p[,x:=as.Date(paste0(x,"-15"))]

xmin=as.Date("2020-1-1");xmax=as.Date("2024-1-1")
p=p[x>=xmin&x<=xmax]
ybreak=p[,pretty(y,6)];ymin=0;ymax=max(ybreak)

ggplot(p,aes(x,y))+
geom_vline(xintercept=seq(xmin,xmax,"year"),color="gray60",linewidth=.4)+
annotate("rect",xmin=xmin,xmax=xmax,ymin=ymin,ymax=ymax,linewidth=.4,fill=NA,color="black",lineend="square")+
geom_line(aes(color=z),linewidth=.6)+
geom_point(aes(color=z),stroke=0,size=1.4)+
labs(title="Monthly deaths in England in ages 20-29",x=NULL,y=NULL)+
scale_x_date(limits=c(xmin,xmax),breaks=seq(xmin,xmax,"6 month"),labels=c(rbind("",2020:2023),""))+
scale_y_continuous(limits=c(ymin,ymax),breaks=ybreak)+
scale_color_manual(values=c("black","#bbbbff","#5555ff","#ff8888"))+
coord_cartesian(clip="off",expand=F)+
theme(axis.text=element_text(size=11,color="black"),
  axis.text.x=element_text(margin=margin(3)),
  axis.ticks=element_line(linewidth=.4),
  axis.ticks.length.x=unit(0,"pt"),
  axis.ticks.length.y=unit(4,"pt"),
  legend.background=element_blank(),
  legend.box.spacing=unit(0,"pt"),
  legend.direction="vertical",
  legend.justification="left",
  legend.key.height=unit(14,"pt"),
  legend.key.width=unit(32,"pt"),
  legend.key=element_rect(fill="white"),
  legend.margin=margin(,,4),
  legend.position="top",
  legend.spacing.x=unit(1.5,"pt"),
  legend.spacing.y=unit(0,"pt"),
  legend.text=element_text(size=11),
  legend.title=element_blank(),
  panel.background=element_blank(),
  plot.margin=margin(5,6,5,6),
  plot.subtitle=element_text(size=10.6,margin=margin(,,5)),
  plot.title=element_text(size=11,face="bold",margin=margin(1,,4)))
ggsave("1.png",width=6,height=3.7,dpi=300*4)
system("magick 1.png -dither none -resize 25% -colors 256 1.png")