Does Shingles Vaccination Reduce Dementia Risk?

01Why birth dates offer a better comparison

Health awareness, access to care and early cognitive changes can affect both vaccination and dementia diagnosis. Matching recipients to non-recipients may leave these differences unmeasured.

Wales's 2013 programme created another comparison. People born just before 2 September 1933 remained ineligible for life, whereas those born just after had at least a year of eligibility. They were close to age 80 and differed little in age. The paper reported uptake increasing from about 0.01% to 47.2% around the cutoff. Local regression estimated the discontinuity, rather than directly comparing all vaccinated with all unvaccinated adults. s1

Two questions follow. The eligibility effect concerns the policy's overall influence on recorded outcomes. The receipt effect additionally divides that effect by the vaccination increase induced by eligibility. Under appropriate assumptions, this identifies the complier average causal effect (CACE, or LATE): the effect among people vaccinated because they became eligible. It is not an effect for all vaccinees, all older adults or each individual dose.

This interpretation requires other risks to change smoothly near the cutoff, no people whose eligibility makes them less likely to receive vaccination, and no pathway from eligibility to the outcome other than vaccination. Balance tests, placebo cutoffs and negative-control outcomes challenge these conditions. Failure to detect an imbalance does not prove every condition holds.

02Similar directions, different policy contrasts

Positive numbers below denote reductions in diagnosis probability; parentheses retain the published 95% confidence intervals. These local policy contrasts cannot be pooled into a universal vaccine effectiveness estimate.

Sources: Wales s1, Australia s4, Canada s5. Ontario's primary follow-up began in January 2017; its secondary analysis began with the programme in September 2016. Both ended in June 2022.

Ontario's analyses share data and some birth cohorts, so they are not independent replications. The three countries also share research-team members: they test the result in different data environments, rather than constitute three independent analytical-team reviews.

The Welsh main cohort contained 282,541 adults, but the principal dementia bandwidth included 56,098. Australia's dataset contained 101,219 patients, with 18,402 in the main bandwidth. The total database size does not describe the information nearest the cutoff. Australia's 482-week main bandwidth merits attention to functional-form assumptions; the authors also reported narrower-window, time-to-event and competing-risk sensitivity analyses. s1 s4

These studies improve on simple healthy-vaccinee comparisons and check prior conditions, other preventive services and non-policy cutoffs. Non-random healthcare samples, under-recording, follow-up selection and local-regression assumptions remain. The policy effects have meaningful quasi-experimental support, conditional on their particular designs.

03Why dividing once does not recover the receipt effect

Wales reported a receipt CACE reduction of 3.5 percentage points (0.6–7.1), or a 20.0% relative reduction (6.5–33.4). The relative denominator is the estimated untreated risk among compliers; the Methods describe its identification separately. s1

Dividing 1.3 by the displayed first stage of 47.2% gives approximately 2.75 percentage points. That is insufficient to declare the reported 3.5 wrong, or to infer the actual compliance rate from 1.3/3.5. Different outcomes and specifications use different bandwidths. The code also reports conventional point estimates with robust bias-corrected intervals, which need not be centred on those estimates. Without the corresponding first stage, joint covariance and patient data, this audit cannot reconstruct the original CACE or confidence interval. s1 s2

The Australian and Canadian authors did not report receipt effects suitable for direct use. Australia's authors explicitly warned that under-recorded vaccination could shrink the first stage and inflate a rescaled effect. Canada measured uptake in the CHSS survey and dementia in CPCSSN medical records. The authors considered those populations systematically different and declined to interpret their ratio as a receipt effect. Survey first stages were 27.4 percentage points (0.0–54.7) and 13.9 points (−5.8 to 33.6). The latter interval crosses zero, further undermining precise ratio interpretations. s4 s5

If vaccination-recording sensitivity were 60%, specificity were 100%, and recording worked identically across the cutoff, a recorded uptake jump of 16.4 points would imply a larger true jump. Rescaling a 1.8-point eligibility effect would then produce roughly 6.6 rather than 11.0 points. This is a conditional sensitivity calculation: 60% is not a measured sensitivity in this study. A direct effect of eligibility through healthcare contact could also be mistaken for vaccine action in a Wald ratio.

04What the above-100% criticism establishes

A 2026 critique by Kivimäki and colleagues combined eligibility and uptake differences and reported implied relative reductions of 18% in Wales, 177% in Australia, and 84% and 111% in Canada. It questioned those magnitudes and their compatibility with national dementia trends. This is material contrary evidence. We obtained and read the abstract, not the full calculations, and do not claim to reproduce its national trends, interval propagation or complete argument. s7

A genuine relative risk reduction cannot exceed 100%. However, absolute CACE divided by the untreated whole-population risk need not be relative CACE. Consider an explicitly chosen example:

• Eligibility induces 16.4% of people to vaccinate. Their untreated risk is set to 20%, and their treated risk to 9.02%.
• Everyone else's risk is set to 3.49%, unaffected by eligibility.
• Overall risks are then 6.20% and 4.40%, an eligibility difference of 1.80 percentage points.
• Absolute CACE is 10.98 points. Dividing by the compliers' 20% untreated risk gives a true relative reduction of 54.9%. Dividing by the population's 6.2% instead gives 177.0%.

Every probability is valid. This is a chosen one-sided-compliance counterexample; Australia had vaccination below the cutoff, and the example does not fit its full compliance-type structure. It shows that an impossibility claim based on the latter ratio requires additional denominator assumptions. It does not show that Australian compliers actually had 20% risk, that missing records explain all differences, or that the critique's empirical plausibility concerns are wrong.

Even fixing the eligibility difference, uptake difference and population baseline leaves multiple valid complier risks in this illustrative model with one-sided compliance, no direct policy path and identical recording processes. Actual risks require corresponding data, rather than choosing a value that makes a study look plausible.

A claimed absence of a national decline must also align coverage, local ages, effect timing, diagnostic practice and population changes. It can challenge a very large universal effect without directly reproducing a local birth-date contrast. Lancet correspondence and an author reply published in July 2026 were identified, but only their bibliographic records were available. Their existence establishes neither resolution of the dispute nor a point-by-point response to the September paper. s11

05Diagnosis, disease onset and competing death

The Welsh endpoint incorporates medical records and dementia on death certificates. Fewer diagnoses could reflect fewer cases, later disease or changed recognition. The authors investigated healthcare use, other conditions and encounters following shingles. These analyses weaken some diagnostic-channel explanations, but adjustment for post-intervention healthcare variables requires additional causal assumptions. An unchanged adjusted estimate does not exclude every diagnostic bias. s1

Death can prevent survival to a dementia diagnosis. In an illustrative constant-rate model, the probability of a first recorded dementia event before competing death is

P_D(T) = λ_D/(λ_D + λ_M) × [1 − exp(−(λ_D + λ_M)T)]

Here, λ_D is an effective diagnosis-event rate and λ_M a competing death rate. Choosing rates of 0.035 and 0.08 per year produces approximately 16.8% at seven years. Lowering only the death rate by 20% raises it to 17.7%. Lowering the effective diagnosis rate by 20% reduces it to 13.8%; the observed curve cannot itself determine whether that rate changed through disease prevention or delayed recognition. This model is not fitted to patients, estimates no vaccine efficacy and does not establish a constant-rate biological description of dementia.

A 2025 Cell study extended the same Welsh system to MCI and deaths among people with established dementia. Over nine years, reported eligibility reductions were 1.5 points (0.5–2.9) for MCI, 8.5 points (0.6–18.5) for dementia deaths, and 6.5 points (2.1–12.5) for all-cause deaths in the baseline dementia cohort. Receipt-CACE intervals were wide. The main mortality bandwidth contained only 3,418 patients, with a baseline ischemic-heart-disease imbalance addressed in sensitivity analyses. s8

Death certificates and all-cause mortality make a purely diagnostic-label explanation less sufficient, but this is not an independent Welsh replication. A death-probability difference among patients does not estimate extra life-years for the general population. Comparing its percentage points with shingles-diagnosis percentage points cannot establish a biological ceiling: populations, baseline risks, endpoints and windows differ.

06Recombinant vaccination and mechanisms

The US KPSC recombinant-vaccine study included 65,800 two-dose recipients and 263,200 matched unvaccinated adults. The dementia aHR was 0.49 (0.46–0.51), versus 0.73 (0.67–0.79) using Tdap recipients as an active comparator. The active comparator reduced the association without randomizing every selection process. Follow-up began six months after dose two among people who remained alive, enrolled and dementia-free. Approximately 63% of unvaccinated adults were later censored because they vaccinated. Near-null negative-control outcomes do not rule out dementia-specific confounding. GSK funded the study and employees participated. s10

A US vaccine-transition study compared different calendar-period cohorts predominantly receiving recombinant versus live vaccine, reporting a restricted-mean-time-lost ratio of 0.83. This is not a 17% reduction in everyone's dementia risk. The reported 164 days represent a time conversion for people subsequently diagnosed. Calendar period, patient characteristics and follow-up still require consideration. s12

Varicella-zoster virus persists latently; reactivation might contribute inflammatory and vascular injury. Preventing reactivation is therefore a testable pathway. Vaccination might also have pathogen-independent immune effects. Timing, dose associations and larger female effects in records do not independently distinguish these mechanisms. A 2025 observational study supports a reactivation–diagnosis association, but we read its abstract and did not directly measure central nervous system viral activity or validate mediation. s9 Nor does a significant female result alongside a non-significant male result substitute for an interaction test.

07Evidence that could change the answer

As checked on 7 October 2026, at least two registered recombinant-vaccine randomized trials directly include dementia outcomes. Both records report recruitment and no submitted results. Finnish trial NCT07502560 plans 33,609 adults aged at least 76, comparing vaccine with placebo. Danish DAN-ZOSTER, NCT07485283, plans 162,000 adults aged at least 65 in an open-label design, with dementia among its primary outcomes. s13 s14 Registered enrolment, dates and status may change; registration proves neither completion nor efficacy.

Randomization can better address healthy-vaccinee differences, but adherence, diagnostic blinding, competing mortality, multiple primary endpoints and follow-up still need assessment. Reliable positive or negative results would change the current judgment.

Current recombinant-vaccine indications and official prevention recommendations address shingles and its complications, not dementia prevention. The label lists common local and systemic reactions and a Guillain–Barré syndrome risk signal within 42 days of vaccination. Dementia observational studies cannot independently establish the overall benefit–safety balance. s15 s16 This article does not convert the new signal into a personal vaccination or treatment-replacement plan.

08Where the evidence reaches

Local eligibility policies for live vaccine have consistent quasi-experimental support for fewer recorded dementia diagnoses, with design conditions retained. Wales provides an explicit receipt CACE, but its 20% remains specific to its compliers, vaccine and window. Australia and Canada do not establish comparable receipt-relative efficacy from the available measurements. Recombinant vaccination has human observational signals; direct randomized results remain pending. Viral mechanisms, prevention versus delayed disease, healthspan and total lifespan do not follow automatically from these diagnosis records.

The reproducible work here covers published-result specification checks, the complier-denominator counterexample, conditional sensitivity models, and 144 education fits across 936 month-level census rows. For example, the all-sex uniform-kernel local-linear 24-month specification gives a no-qualification discontinuity of −0.24 percentage points, with a descriptive 95% interval of −1.94 to 1.45; the 96-month result differs. Census population, aggregation and survivor selection differ from the clinical cohort. This neither proves absence of confounding nor reproduces the authors' RDHonest inference or patient-level effects. s3

Reproduction package: code, inputs and complete outputs