the new UK ONS data is out and it's worse than before

this was not a fix, it was further breakage

back in may, the UK statistical agency (ONS) stopped reporting all cause deaths numbers by vaccine status. in response to a number of serious criticisms, they went so far as to admit that their data was not, in fact, suited to purpose and was of too low a quality to be relied upon to assess vaccine efficacy our outcomes.

chief among this list of complaints was reliance upon an old census to back into the number of people who were unvaccinated by taking that out of date population count and subtracting from it the number of people jabbed.

this is going to severely undercount the unvaccinated because population has risen since then.

that, in turn, will mean that bad outcomes (like deaths) from this larger population will be ascribed to too small a group.

that in turn, will mean that the apparent rate per person or person year of bad outcomes will look too high.

and that will destroy the validity of the comparisons by greatly inflating base rate risk.

it is a massive and enduring problem with their methodology and one the proposed to fix by updating to the 2021 census. none other than ONS’s own sarah caul described this update and it’s reasons to me. (interestingly, one of them was to include info from a 4th booster and this data seems to appear nowhere in the new release)

in particular, i had a lot of questions about the number of unvaxxed as when you used a better number (like UK HSA) it massively inverted the signals and made it clear that vaccines were, on an age stratified basis, associated with much higher all cause death rates.

and this seemed a deeply suspicious time to cease reporting.

my expectation was that fixing this data would make the negative signal for all cause mortality from vaccines clear. this was going to be too big to hide.

well, i was half right. it was too big, but they seem to have gone to massive lengths to hide it. the new set of ONS data (available here) is flat out worse than before. they did not fix the denominator undercount issue. they double downed on it. this data appears highly manipulated and contains such improbable/impossible assumptions that i’m not sure what to call it other than deliberate misrepresentation.

this is not coming clean. it’s a whitewash.

let’s look:

we start here in table 1. i selected all cause deaths for december.

the core claim from this data is that vaccines are associated with lower all cause deaths. 944.9 deaths per 100k person years vs 1026.7 in never vaxxed. but this claim is rooted in some truly herculean data manipulation.

1. when we calculate the simple ratio of deaths per 100k person years, we get 206 for unvaxxed vs 1113.6 for ever vaxxed. vaxxed is over 5X higher. this does not mean it’s wrong to claim that vaxxed is better on an age standardized basis. the unvaxxed skew younger and thus have lower expected death rates. but it DOES mean that it’s age adjustment doing all the work here and the output will only be as good as that assumption/model.

2. worse, when we calculate the assumption about % ever vaxxed vs never vaxxed by comparing person years, we get to their core assumption about the size of the unvaxxed: 15.2% of the total population. and this is WILDLY low. there is no data anywhere else in the UK that looks remotely like this. and this assumption taints every other aspect of this analysis to the point of extreme inversion.

according to the UK HSA (health security agency) this number is absurdly low. there is no tranche of UK society under age 50 that is that heavily vaccinated. some have tried to claim that the UKHSA DATA cannot be compared to this ONS data because it is a subset, but i find this objection unconvincing for several reasons:

1. this is a large subset comprising the majority of UK population

2. the data is MUCH higher quality as it comes from counting actual medical records instead of inferring/assuming the count of the unvaccinated that constitutes the base rate risk used for all comparisons

3. the subset it includes disproportionately excludes the very young and more recent immigrants. this would seem likely to, if anything, undercount the unvaccinated relative to the full population

4. we need some sort of sanity check on these ONS claims and this seems like the most authoritative one. perhaps it is not exact, but i suspect it’s FAR higher quality and less error prone than the ONS estimation that looks increasingly rigged. it’s the one used to measure most disease issues in UK.

to (as much as possible) avoid the issues of the opaque ONS age adjustment and get a more granular look, i’m going to switch to looking at the age stratified data. unfortunately, nearly all of this data is (perhaps deliberately) presented in a bayesian mess where it is split up by vaccine status as though that is some sort of independent variable when it is not. everyone who has 3 doses at some point had 2 and 1 and passed through that risk strata. this creates a “left truncation” issue where only the healthiest (and least culled) make it to dose 3.

the only real exception is table 5 which has very little data but is still instructive to look at as we can see what % of deaths by age strata were in vaxxed vs unvaxxed. the problem with this data is that it gives death by status, but not a count of people or person years. this seems an odd omission given that it is present in many other parts of the data and they clearly have it. (and is part and parcel to the sort of “hide the ball” games endemic in this data.) but we can still calculate the % of deaths in each group. if we then use the UKHSA data to estimate group size, we can get some risk ratios.

these look quite bad for the vaccines. all cause mortality is higher in all groups. (i had to combine 80-9 and 90+ into one group because that is how UKHSA reports vaccination and there is no granularity above 80)

this is nothing like the data claim that ONS is seeking to pass off. it shows highly elevated risk of death in the vaxxed if we use more plausible population statistics. is this exact? no. but it’s likely good enough to give us an intuition that all is not well and what the trick here is.

ONS provide the person years data on an age stratified basis in table 2 and then despite this stratification also age adjust it. unfortunately, they no longer provide “ever vaccinated” vs never and split it into the bad bayseian buckets that have become so familiar and unlike back in may are no longer providing enough data to even assess the 21 day period post each jab.

this seems like a serious omission as the period was showing massively elevated risk levels for many back in may.

taking their stratified and bayes boggled ASMR (age std mortality rate per 100k) data at face value (and ignoring this issue) we get this:

you can literally see how the risk is being pushed backwards and out of categories. 1 dose is extremely negative as it has nowhere to shift to. 2 doses have a strong negative signal, but suddenly a third makes it efficacious? (though really only in 40-79)

this seems biologically implausible. if a vaccine cannot teach you to resist a disease, more of it probably cannot either. this idea that you need to keep antibodies high by boosting all the time has never had any basis in fact. it would seem to imply that you cannot remember the response and replicate it even 6 months post inoculation. nothing about that has ever passed the smell test. more likely this is just statistical games. based on what alberta accidentally admitted, you can see just how bad that can be, but if you stop boosting and redefining “vaxxed” then the signal can emerge. this whole idea seems like “extend and pretend” using math games.

even if this signal were real, this also implies that anyone arriving at “boosted” has already likely run a gauntlet of higher risk rates and is a culled cohort with the weak removed. this would, by itself, make the claims about d3 working look problematic.

but there is another reason to discard this data and its claims as well: the baseline risk rate for the unvaxxed appears massively exaggerated and not only was this not fixed in the new version, it was accentuated.

using the “new census” somehow caused every age group apart from 80% to contain even fewer unvaxxed as a %.

and the margins were considerable and this effect therefore pronounced.

the meat of the curve from 40-79 that forms the core of “reported efficacy for boosters” saw the size of the unvaxxed cohort shrink an avg of 10.4% from may to dec 2022.

anyone who thinks that 1 in 10 people who made it to may 2022 unjabbed decided to go ahead and get the shot, please raise your hand.

i do not know a single person above college age in my entire social network who made that choice. does anyone?

this seems incredibly implausible and bears out literally nowhere in other UK data or in any sort of anecdotal report i have heard.

these numbers are fanciful and is has rendered this data a rig job to misattribute the all cause deaths to the unvaxxed by even further under counting them.

the effect from this is profound.

using the UKHSA data roughly doubles the risk ratios and causes them to pass through unity (a risk ratio of 1 indicated no effect)

those 50-9 who were being described as 37% less likely to die from all causes are now 25% MORE likely to die. 80+ leaps from a 4% benefit to an 88% increased likelihood of death.

how precise is this? it’s hard to say, but likely plenty good for this sort of gross analysis. the UKHSA data looks far more sound and likely errs toward undercounting the unvaxxed vs genpop and so i think it gives us a pretty good signpost, especially when it’s clear that ONS is playing silly buggers with the denominators and moving them to implausible levels to make trends they do not like disappear.

if ONS is correct and UKHSA wrong, then the group outside UKHSA must be vaccinated at incredibly high rates vs what is generally deemed the “representative group” for UK in health matters. i’m not even clear it’s mathematically possible but lack the data to do the analysis. (if someone knows where to get it, please let me know)

igor took a look at this and saw similar.

his conclusion that the unvaxxed were being undercounted by about 50% (which is to say that the group is roughly 2X the size claimed) very much foots with mine.

the uniformity of it is really quite striking

and so once more we land on “there is simply no way to trust a vaccine claim made with this ONS data.”

it shows every sign of being a massive rig job that was suppressed post may to find a way to rig it further before releasing it because the old level of rigging was insufficient to hide an increasingly bad signal.

the vaccination numbers are deeply implausible and look to be grossly undercounting the unvaccinated in order to basically double the base rate risk being ascribed to them. astonishingly, this is STILL not enough to make 1 or 2 jabs look efficacious and were we to apply the reduced base rate to them, it would get worse still.

this was not a data clarification or the adoption of better practice, it looks to have been a last ditch effort to manipulate the data into hiding a debacle.

we can argue about the UKHSA data and its comparability and just how accurate assumptions from it are, but it’s real medical record level data vs a calculated figure from ONS and nothing else seems to jibe with the ONS data, not even OWID which is notorious for overcounting vaccination.

to believe this shows efficacy for the boosted, you have to accept that:

1. vaccines that do not work as a double dosed “full course” suddenly become effective as a booster

2. the issue with bayesian rigging is small and time spending as single and double dosed and esp the clearly dangerous periods in the 21 days post jab is small enough not to flip the risk ratio. (it’s not, you can see it here)

3. that the ONS is not playing games with age adjustment (actually possible but not proven)

4. that the ONS is correctly calculating the unvaccinated %’s despite their data not agreeing with anyone else’s particularly those like UKHSA that use a far more reliable method of counting vs modeling

5. that the UKHSA database is not a reasonably good proxy here that is more likely to be conservative rather than overstated on counting the unvaxxed but somehow massively undercounts the jabbed and is overstating unvaxxed by ~2X

6. and that somehow either from end may to end dec, ~10% of key age groups who made it to june 2022 unvaxxed decided to go get jabbed or that the outdated census the ONS was using was actually overcounting the number of people in the UK.

sorry guys, but this past “a bridge too far” to accept and well into “all the bridges of venice past plausibility.”

nothing about this analysis or its foundational claims appears to hold together or even make sense. none of it foots with other data.

best i can tell, they are playing the same denominator games as CDC and layering bayesian datacrime on top for good measure.

i really had high hopes for this dataset as a way to get at the fabled “all cause mortality by vaxx status” data, but i must concede that this is just junk, unsuited for purpose, and likely manipulated to be contrary to it. gatopals™ martin neil and norman fenton appear to have been right all along about ONS. see footnotes 6-11 HERE)

this is bad data, possibly by design, it fails to pass any sort of sniff test, got more, not less stinky, and imposing even rudimentary adjustments to it to bring it into alignment with higher quality data strongly inverts the purported signal for booster efficacy, itself an implausible inversion of double dosed outcomes.

i’m just not seeing any particularly profitable way to work with it as is.

it’s just one more reason to stop trusting regulators and government agencies and demand open data.