Vaccines are not a panacea

The shingles vaccine has not cured dementia

Over the past month, four studies have been read as showing that the shingles vaccine prevents, or even reverses, dementia. A national newspaper has turned that reading into advice. The advice is to go and get the vaccine now. The evidence tells a different story.

Vinay Prasad has already gone through the four papers and laid the timings side by side, and I will not improve on that account.

● The Annals paper has a 5% absolute reduction in dementia by three years, with benefit appearing within a year of a nursing-home admission.

● The JAMA paper has it working by a hundred days.

● The Welsh study in Nature is a natural experiment, not a simple cohort, and reports a fifth fewer dementia diagnoses over seven years. It declines to publish a time-to-event plot at all.

● The American paper, in Nature Medicine, includes a hazard function whose curves separate almost immediately.

● And a meta-analysis concludes that it is not the shingles vaccine alone but every adult vaccine that appears to protect against dementia.

What a dementia diagnosis actually is

A dementia code in a health record is the end of a chain of soft decisions:

● someone has to notice a change,

● the patient has to present to the system,

● a threshold has to be applied,

● and a clinician has to enter the code.

For each of these steps there is wide variation in the timing between patients. The noticing depends on who is being watched closely and who is not. The same forgetful elderly person can be coded, or not coded, according to how hard anyone happens to be looking. Look harder and you might diagnose more.

Now ask who receives an elective shingles vaccine. It is, overwhelmingly, the people well enough to go and get one. The housebound, the frail, and above all those whose memory is already failing, the people nearest to a diagnosis, are the least likely to present for a non-urgent injection. So the unvaccinated group is enriched, at the very outset and before anyone is followed up, with exactly the people who will later be coded with dementia. This is the “healthy vaccinee effect”, and in this setting it is large enough to produce the whole of the apparent benefit on its own.

Why four studies agreeing is not reassurance

The natural objection is that four datasets agree, and agreement is supposed to be reassuring. But when the cohort studies make the same methodological error it is not reassuring. Every one of them carries the same confounding structure: in each, vaccination marks underlying health, healthcare contact and the simple capacity to turn up, and in each that marker is correlated with the outcome being measured. The errors are not independent results that happen to coincide. They are the same error made several times, because the same bias is built into each design. A biased estimate does not become unbiased by being repeated. These confounded studies in agreement tell you that the confounder is stable, not that the result is real.

Prasad puts this as the earth looking flat whichever way you turn, which is exactly right as an image of the confounded cohorts. The formal version is that replication adds evidence only when the replications are capable of failing in different ways. When they share their confounders, consistency is not corroboration, it is the same confounder doing the same job in each dataset in turn. To break the symmetry you do not add a fifth observational cohort, another flat horizon. You change the design to one whose errors are structured differently. That test already exists.

A regression-discontinuity study exploits the age cut-off for eligibility: people just below and just above the threshold are, for practical purposes, exchangeable, so the baseline-health gap that drives the observational studies is largely removed. The Welsh study is exactly this design, which is why it is the strongest evidence on the table; they claimed dementia was prevented in Wales and in Australia and Ontario. But would an independent team running a study with an identical design find the same thing? George Davey Smith and colleagues have now carried out such a study in England, across some 6.3 million people, roughly twenty times the Welsh sample; the work was presented in June 2026 at a conference and is not yet published. The design worked, in that eligibility produced the expected fall in shingles (although it is not clear what happened in the ‘grace period’ immediately after vaccination) but on dementia it found nothing. The fact that the finding was null in the far larger study should mean the smaller one has been overruled.

The speed matters

Speed is being sold as a measure of how powerful the vaccine is. It is better understood as evidence that the vaccine is doing nothing at all. Dementia is slow. The neuropathology accumulates over years, usually decades, and by the time anyone is coded with the disease the process is long established. Nothing done to a brain can produce a measurable difference in dementia incidence within a hundred days, because the disease does not move on that timescale. So when the curves separate almost immediately, the separation cannot be the vaccine acting. It can only be that the two groups already differed at the moment the vaccine was given.

It is a signature of selection. The rapidity that is being offered as the headline result is the clearest single indicator that what is on the page is the “healthy vaccinee” artefact and not a treatment effect.

From the literature to the newsstand

On 10 June the Daily Telegraph published an interview with John Todd, professor of precision medicine at Oxford, in which he told readers over fifty to get the shingles vaccine now. Jonathan Engler has lodged a complaint with the MHRA about that article.

Professor Todd is, on his own declaration, a paid consultant for GSK, which manufactures Shingrix, and the study he relies on is one he co-authored. When the paper was published, given the issues described above about selection bias the authors wrote that the dementia results “provide a rationale for conducting a randomized control trial aiming to confirm the findings”. Shingrix is not licensed for the prevention of dementia. The advice to buy it for that purpose rests on a paper whose findings, in its authors’ own words, have not been confirmed.

In the United Kingdom, the advertising of prescription-only medicines is governed by Part 14 of the Human Medicines Regulations 2012. Regulation 284 provides that a person may not publish an advertisement likely to lead to the use of a prescription-only medicine. There is proper latitude for journalism, and the MHRA’s guidance for journalists asks only that coverage be balanced, that it not exaggerate benefit, and that it not steer readers towards a named product. It does not require proof of a commercial arrangement before a breach is made out. Engler’s point, which is the right one, is that a piece does not stop being promotional by declining to call itself promotional, and that the absence of the label is itself the difficulty. Where a consultant to the manufacturer urges the purchase of that manufacturer’s unlicensed product, on the strength of his own unconfirmed research, with the consultancy unmentioned, the ordinary description of the result is advertising, whatever else it is also called.

Telegraph readers thinking of taking his advice to rush off and buy a private Shingrix vaccine, might want to take note that (a) Professor Todd has no medical qualification (his FRCP is honorary and his primary qualification is in biochemistry) and (b) the department he heads in Oxford was funded by GSK to the tune of £30 million.

What honest curiosity would do next

The editors of the observational papers could be asked to publish the all-cause mortality curves alongside the dementia curves: if the vaccinated are simply healthier people, their mortality will diverge in the same direction, and the confounding will be visible on the page. This is a control test. Every child at school is taught the importance of using a control group in scientific research but all too often they are left out by actual scientists. The randomised trial that the NIH is said to be considering might need to be run simply to put to bed the excitement around this issue that has been fuelled by media speculation.

The truth is that the studies that kept producing the same answer did so because they shared the same bias, while the largest independent test found nothing and a national newspaper has printed a paid consultant’s promotion of a drug when he will benefit from the sales. Perhaps neither of those stories will sell as many newspapers as the ones that do get printed or clicked on.