Do COVID Vaccines Work? (part 3)

In this article I utilise data from an unknown NHS Trust and use split-file multivariate logistic regression in the prediction of vaccination status prior to death

In part 2 of this series we eyeballed four slides that suggested the mRNA vaccines are not conferring benefit in terms of reducing in-hospital death from acute and chronic respiratory disease/COVID-19. Incidence rates for both vaccinated and unvaccinated cohorts over the period 2020/w50 – 2021/w37 pretty much mirrored each other and I suggested that world governments get their money back. I cautioned readers over the very small sample sizes for weekly deaths and suggested we note the general failure of the vaccines to show up to the party. I also exposed my careless disregard for confounding variables - especially age - but this shortcoming will be addressed in this issue.

Chopping the data into weeks, though illuminating from the point of view of disease progression across the nation, has weakened the analysis and so this morning I am going to mash everything together in one big bowl and season with just salt and pepper. We shall start with the simplest tabulation I can muster, noting that the period in question is 2021/w1 – 2021/w37, and that I have included in-hospital deaths for those aged 18 years and older. We should also note that these figures are not derived from death certificates or from the ONS, for they derive directly from the EPR of an unknown Trust. We cannot be certain how reliable this is until somebody audits it.

In terms of diagnostic grouping that which I shall call a ‘COVID death’ is not necessarily causal. All this means is that the ICD-10 codes for COVID-19 (U07.1; U07.2) were entered in one of ten data entry fields along with all other diagnoses made. Thus, a ‘respiratory death’ is not necessarily causal, and neither is an ‘acute respiratory COVID death’ necessarily indicative of causality, e.g. they may have died from liver failure. I mention the liver a lot in these instances because this organ becomes a major determinant if a patient is pumped with drug after drug. With all that in mind, let me furnish the first seriously simple and utterly misleading table:

We have 9,894 in-hospital deaths in total, some 1,786 of which (18.1%) possessed a COVID diagnosis. We also find that 5,055 (51.1%) were vaccinated prior to death at some point. When I say ‘vaccinated’ I mean with any dose whether this was just the first shot or the booster.

More interesting than these figures are the percentages within the inner table for we observe 29.7% of unvaccinated deaths with a COVID designation as opposed to 6.9% for the vaccinated cohort. If taken at face value this would suggest that the vaccines have been working. Whilst this is the naïve approach taken by the authorities we must note this rests entirely on the assumption that all things are equal in all other aspects. This is certainly not the case and we can easily list a number of confounding factors starting with the distinct possibility that unvaccinated patients may well have been tested more frequently. This isn’t a mere whim for several nurses have confided that this is exactly how their unit goes about matters.

There’s also the possibility of reluctance to consider a COVID diagnosis for vaccinated patients. This sort of ugly goes beyond the decision of the SHO/registrar for we are now talking about the Trust clinical coding team working under senior management instructions, and again I’ve had staff confide in unsavoury practices within their Trusts.

Then we’ve got to consider reliability of the PCR test (forget hello fresh, we’re talking hello flu) and whether a death merited a COVID designation; that is, whether a ‘COVID death’ involved an acute respiratory phase. Let’s have a look at this very issue:

In this table asymptomatic COVID means the EPR carries an ICD-10 coding pointing to a positive test result with no mention of any respiratory condition. Non-acute symptomatic COVID points to those positive-testing cases that also suffered chronic respiratory disease but didn’t enter an acute, and thus potentially life-threatening phase. The final category of acute symptomatic COVID is the real deal  - somebody in a bad way with a positive test result.

It would appear that unvaccinated folk didn’t fare as well as vaccinated counterparts no matter how we slice the cake but this assumption still rests on many confounding factors as stated. What we ought to do is account for as many confounding variables as possible to ensure we are comparing like-with-like in a fair and just manner

Auto Efficacy

Except we can’t ever compare like-with-like in a fair and just manner. This hastily crayoned total rubbish diagram explains why…

Since national COVID cases were high when national vaccine uptake was low, with COVID cases petering-out when vaccine uptake was high, then any attempt to analyse this dataset as a homogeneous block will always result in apparent vaccine benefit even if the vaccine was saline. Whoops-a-daisies!

If we consider this in relation to in-hospital deaths then somebody being admitted early in the pandemic with advanced heart failure was more likely to register as COVID positive at a time when their likelihood of vaccination was low. Somebody being admitted late in the pandemic with advanced heart failure was less likely to register as COVID positive at a time when their likelihood of vaccination was high. Combine these two cases in an analysis and you’ll get instant vaccine benefit. This is what I am calling ‘auto efficacy’, and the inherent bias within the data sample will equally favour tap water as therapy for COVID as big pharma’s best efforts.

This is one of the many numerical tricks that are being played by those in authority who undoubtedly fully understand that they are playing mind games with a gullible and poorly-educated public.

Let us now consider the actual weekly COVID/vaccine deaths in my data sample to see how reality squares up to my theoretical, total rubbish diagram…

There you go! Analyse this lot as one big fat lump and you’ll get instant vaccine benefit regardless of whether the vaccine is doing anything useful. In my book if the vaccine is doing anything remotely useful then we ought to know about it and evaluate genuine benefit against genuine risk. It’s quite clear that the various authorities, NHS senior management, government ministers (with two notable exceptions), virologists, epidemiologists and other ‘ologists’, as well as the mainstream media and manufacturers don’t want us to do this, so let me get the coffee on and have a go at poking a stick at this…

Standing The Problem On Its Head

Setting aside the thorny problem outlined above, an obvious source of sample bias is that the vaccinated cohort will be older owing to the greater uptake among the elderly population. A less obvious source of bias stems from attitudes to vaccination in the wake of chronic disease. For example, several sick folk have confided that they don’t want any jabs thank-you-very-much because they’ve been poked around enough as it is.

Analysis seems impossible until we realise we can stand the problem on its head and try to predict the vaccination status of each individual prior to death instead of trying to predict what the vaccine does or doesn’t do. With vaccination status set as the binary dependent variable we can then process the EPR to identify key conditions of interest over the period 2021/w1 – 2021/w37 for those aged 18 years and over at death. Here’s my initial shopping list of basic ingredients:

Total diagnoses may require a bit of explanation. The data dump I was given contains ten fields in which ICD-10 codes may be recorded at death. Thus, a person may theoretically possess between 1 and 10 coded diagnostic fields, the exception being those who are declared dead on arrival. These folk get wheeled straight to the mortuary without much clinical coding on the hospital system. Later on, at post-mortem, the cause of death will be determined but this is registered on the death certificate and not the Trust EPR. This gives rise to those seemingly bizarre zero values of which there were just 37 DOA cases among 8,714 (0.42%).

Case Detection Rate may also require a bit of explanation. This is derived from the UK GOV coronavirus dashboard and is simply the rolling 7-day number of positive cases detected divided by the rolling 7-day number of viral tests undertaken expressed as a percentage. This provides a handle on disease progression within the population as opposed to test progression. We may note that the series (for England) runs from 0.10% to 16.20% with a grand mean of 2.50% over the period 2021/w1 – 2021/w37. Hopefully the remaining clinical factors are self-explanatory.

At this point we can either run a logistic regression for the entire dataset of 8,714 in-hospital deaths and run slap-bang into a wall of auto efficacy or we can wisely acknowledge the tricky issue and run a split-file analysis in an attempt to minimise this. If the term logistic regression sounds scary just think of it as a technique for predicting binary responses, in this case 0=Not vaccinated prior to death and 1=Vaccinated prior to death. We can then submit the independent variables listed above in a conditional selection procedure to determine which of these are statistically significant at the 95% level of confidence (i.e. can actually cut the mustard). Get that coffee on the stove and crumpets under the grill because we’re in for a turbulent ride…

Curiouser And Curiouser

Let us start by considering the classification table to give us confidence that the logistic regression model is doing something sensible:

There are our two periods of 2021/w1 – 2021/w18 and 2021/w19 – 2021/w37 divided into nice clean blocks. We can zip right over to the column headed ‘Percentage Correct’ to get a feel for how well the models are performing. In the case of weeks 1 – 18 we see that the model has correctly predicted 80.5% of vaccinated deaths (a.k.a. true positive rate) and 77.4% of unvaccinated deaths (a.k.a. true negative rate). That’s not bad going. In the case of weeks 19 – 37 the model is rather good at predicting vaccinated deaths (99.0% true positive rate) but is abysmal at trying to fathom the unvaccinated deaths (5.3% true negative rate) owing to the dwindling sample size.

Before we look at the models themselves we need to cogitate on matters with that coffee and those crumpets. Whilst vaccines appear to be causing harm (e.g. accelerating cancer), it is also logically possible that people with cancer will decide to get vaccinated - correlation cuts both ways. Thus, in eyeballing the regression coefficients below we have to consider the possibility of bi-directional effects: we are in chicken and egg territory! Try this for size…

We now need to compare models across the two periods in order to get a handle on what is going on in chicken and egg world. If we consider age then we discover it is highly statistically significant in both periods (p<0.001). In the first period it takes up an odds ratio (OR) of 1.052 meaning that the likelihood of vaccination prior to death increases with age. This makes sense. In the second period age takes up an OR of 1.029, which is pretty similar. We may conclude that the age effect is consistent and meaningful in a real world sense.

Another consistent performer is COVID-19 diagnosis, with an OR of 0.429 for the first period (p<0.001) and an OR of 0.368 for the second period (p<0.001). This gives me cause for numerical concern and I shall explain why. These odds ratios of less than one indicate that vaccinated individuals were less likely to receive a COVID diagnosis prior to death. At the same time we have to entertain the alternative reality that those suffering COVID prior to death were less likely to get vaccinated, either by choice or because they were too sick to be messed with. This is the law of chicken and egg and is a right spanner in the works!

Do odds ratios of less than one in both periods mean that the vaccines were beneficial? Not necessarily, for the same result could easily arise through patient choice and case management, as stated. At this point we have to ask why the odds ratios for both periods were remarkably similar given very different disease and vaccine dynamics. Shouldn’t the benefit be much, much greater for the second period when second doses and boosters were applied? Sure, we observe a slight drop from OR=0.428 to OR=0.368, which points to possible vaccine benefit but the fact that we see somewhat similar odds ratios across two very different periods suggests to me that something murky might be going on. Things are getting curiouser and curiouser, and my guess is that COVID-19 coding is being rather dubiously handled. Either that or a Trust-level decision was made not to test seriously ill patients, especially if they have been vaccinated.

In Harm’s Way

Aside from age and COVID-19 diagnosis nothing else is constant apart from the… er… constant… and cancer. So here’s the thing, and a fine example of why statisticians often run split-file analysis to reveal that which is hidden. In the first period we observe an odds ratio of 0.757 for cancer diagnosis (p=0.001). Being a value less than one this means that individuals vaccinated prior to death during the first period were less likely to be suffering from any form of cancer. In the second period we observe an odds ratio of 1.272 (p=0.005). Being a value greater than one this means that individuals vaccinated prior to death were more likely to be suffering from cancer. So what changed?

What changed was folk dying in the second period were more likely to have received a second dose and booster. Admittedly, other relevant things may have changed (e.g. loss of diagnostic service provision during 2020 leading to a cancer time bomb) but dosing also changed. There are some who’ll grab the first explanation with both hands and ignore the latter fact. In response to such wilful arrogance I am going to take this as potential evidence of vaccine harm, and if we take both odds ratios we find a factor increase in risk of a cancer diagnosis of x1.68 between these two periods. In plain English this is close on doubling your risk of being diagnosed with cancer.

Chronic respiratory disease appears as a statistically significant predictor of vaccine status in the second period with an odds ratio greater than one (OR = 1.204, p=0.037). This indicates that those who were vaccinated prior to death were more likely to have also suffered from a chronic respiratory disease such as asthma or COPD (and vice versa). Since the clinically vulnerable were being vaccinated across the nation from the outset then why doesn’t this independent variable feature for week 1 – 18? Could it be that additional doses of the vaccine are causing respiratory issues?

Another oddity is why Total diagnoses appears with an odds ratio less than one during week 1 – 18 (OR=0.932, p=0.001). This indicates that those who were vaccinated prior to death during the first period also tended to present as less complex cases; a fact which goes against the grain of providing mRNA therapy to the elderly and vulnerable from the outset. We may think of this as ‘unusually clean death’, which makes me rather suspicious - are we again looking at evidence of iatrogenic death?

Case Detection Rate is an interesting predictor in that it only appears in the first period. This is a proxy measure for the spread of SARS-COV-2 amongst the population and we can see from the odds ratio that there is an inverse relationship between likelihood of vaccination prior to death and disease prevalence (OR=0.514, p<0.001), as we may readily observe in the time series slide. This is the sort of result that the authorities leap upon as evidence of vaccine efficacy, but if this is a genuine result and not an artefact arising from sample bias, then why don’t we see the same variable popping up as statistically significant for the second period? Occam would no doubt have something to say about this only find himself censored.

Tea Break!

There’s none better than Professors Fenton and Neil to hammer these issues home and I highly recommend their latest substack article on the matter. However, I think that’s enough crunching and mauling for today so I am going to pull up stumps and declare 2,691 words for 5 sections just before tea. Anybody care for a slice of Swiss roll?