The Loop

 · 15 min read
 · Nulla Verba
Table of contents

The previous post1 showed that post-market surveillance is structurally insensitive to chronic, diffuse, or delayed signals. The natural follow-up: has anyone studied the question directly, outside routine surveillance?

Two groups have. Their headline conclusions and their supplementary data do not tell the same story.

The Circularity

In 2019, researchers queried the European Medicines Agency directly, asking what specific safety studies supported the continued use of aluminium adjuvants in vaccines. The EMA replied that it held "no specific safety studies comparing aluminium alone with an inactive comparator." No new studies were needed, the agency explained, because there were "no new safety issues reported after decades of use."3

Map the logic:

  1. No safety signal appears in surveillance data
  2. Therefore no new safety studies are needed
  3. Therefore no study exists that could generate a new signal
  4. Therefore no safety signal appears in surveillance data

Each step is defensible in isolation. No individual decision-maker is acting irrationally. But "no evidence of harm"2 is the loop's output, not a finding. It is what happens when no one looks. A study that has never been conducted cannot generate a signal. The absence of a signal from systems not designed to detect it cannot substitute for the study that was never run.

The EMA also stated that AAHS, Merck's proprietary aluminium adjuvant used in Gardasil-9 and other vaccines, "was introduced without any prelicensure safety evaluation."4 A novel adjuvant entered population-wide use without its own safety data, and the absence of subsequent signals in systems not designed to detect adjuvant-specific signals was taken as confirmation that none was needed.

The Background Problem

To study whether aluminium adjuvant causes harm, you need to compare children who received it to children who did not. That comparison is harder than it sounds.

Every inactivated, toxoid-based, or adsorbed vaccine on the UK and Italian routine schedules contains aluminium. There is no aluminium-free alternative for any of them.5

Contains aluminium No aluminium
6-in-1 (DTaP/IPV/Hib/HepB) Rotavirus (live oral)
MenB (Bexsero) MMR / MMRV (live)
PCV (Prevenar 13) MenACWY (conjugate)
DTaP/IPV boosters Nasal flu (live)
HPV (Gardasil 9) Inactivated flu, pregnancy (no adjuvant)
Tdap (pregnancy) RSV, pregnancy (no adjuvant)

A child following the UK routine schedule receives aluminium-containing vaccines at 8 weeks (6-in-1 + MenB), 12 weeks (6-in-1 + MenB), and 16 weeks (6-in-1 + PCV), accumulating roughly 2.1 to 3.6 mg by 16 weeks depending on the products used. In Italy, the primary series runs at 3, 5, and 11 months with the same vaccine types.6 After the first appointment, there is no period where a vaccinated child is aluminium-free.

Within vaccinated children, you can compare "more aluminium" to "less aluminium." Comparing to unvaccinated children is the alternative, but they are a small fraction of the population and systematically different in healthcare-seeking behaviour.9 Neither comparison answers the binary question: aluminium or none.

The Studies That Exist

Two research groups broke through the loop and examined cumulative aluminium exposure directly.

In 2023, Daley et al. used the US Vaccine Safety Datalink to calculate cumulative aluminium from vaccine records for 326,991 children. They found a positive association with persistent asthma: an adjusted hazard ratio of 1.19 per 1 mg increase in aluminium exposure. The authors noted multiple limitations and urged cautious interpretation.10

In 2025, Andersson et al. used Danish registry data on 1,224,176 children and reported no association between cumulative aluminium exposure and any of 50 chronic conditions, using the same dose-response design: hazard ratios per 1-mg increase.7

Both studies compare children who received more aluminium to children who received less. Neither compares aluminium-exposed to aluminium-unexposed. The authors of the larger study conceded this directly: "It is fair to say that our study does not evaluate the hypothesis that any exposure to aluminum vaccine adjuvants, irrespective of the cumulative amount received, increases the risk of these outcomes."8

Both studies chose a dose-response design. A dose-response model measures whether more is worse than less, not whether any is worse than none. A null result then feeds back into the safety consensus. The loop continues: "no evidence of harm", not because harm was ruled out, but because the question was not asked in a form that could detect it.

The checklist from an earlier post applies here. Both studies measured dose-response gradients, not binary exposure. Both used vaccinated children with different aluminium totals as the comparator, not vaccinated against unvaccinated.

The primary follow-up ran to age 5, but neurodevelopmental cases nearly tripled when extended to age 8 (6,386 to 19,466).11 The 15,237 children with zero aluminium-containing vaccines were in the dataset but never compared as a group. A sensitivity analysis excluding them found "similar results," but that analysis inherits the same limitation: it can detect gradients within exposure, not the effect of exposure itself.

A direct comparison of 15,237 unexposed children against 1.2 million exposed ones would have been the largest controlled comparison of aluminium exposure ever conducted. It was not reported. Holt et al. found that 55% of Danish children recorded as "unvaccinated" in the same registry infrastructure had actually been vaccinated according to medical records.14 The loop again: the one group that could break the circularity was set aside because the existing consensus said they were not comparable.

The main analysis used a linear model, and the authors reported no association. Their categorical analyses, they wrote, "did not suggest nonlinear associations."13 With 50 outcomes tested across multiple subgroups, some isolated findings are expected by chance.

But one is worth examining. When stratified by birth cohort, Asperger syndrome in children born 2007-2018 showed a statistically significant positive association (HR 1.67, 95% CI 1.01-2.77, 51 cases).12 The earlier cohort showed no association (HR 0.99, 124 cases), and the finding attenuated to null when follow-up was extended to age 8 (HR 1.02). This could be noise, a DSM-5 diagnostic artifact (Asperger was removed as a separate diagnosis in 2013), or a real signal too fragile to survive dilution. At the very least, it is hypothesis generation. What it cannot be is evidence that the question has been answered.

In the categorical analysis, neurodevelopmental outcomes also showed a non-monotonic pattern: the lowest-dose group had more cases than the highest-dose group, while the mid-dose group had fewer. A linear model would flatten this shape to null.

These are two misleading versions of "no evidence of harm": looked, but with a lens that sees only one shape of signal; measured, but at a follow-up too short for the outcomes that matter most.

The Neglected Field

In 2004, the WHO's Global Advisory Committee on Vaccine Safety characterised adjuvant safety research as "an important and neglected field." The committee called for improved post-licensure surveillance, clear case definitions for adjuvant-related adverse events, Phase IV clinical studies, and validated animal models, noting that the latter "do not exist, yet they will be required."15

Twenty years later: no standing surveillance system for adjuvant-specific outcomes has been created. No validated animal model exists. No Phase IV study comparing aluminium adjuvant to inert placebo has been conducted in children. The UK's CPRD, a primary care database covering 13 million patients and linkable to hospital episodes and death registrations, could theoretically isolate aluminium effects by exploiting schedule changes. No one has designed such a study.17

In 2025, the GACVS reaffirmed its safety conclusion without having fulfilled any of the four steps it said were needed to reach one.16 The loop closes: the committee reaffirmed safety using evidence generated within the constraints it identified as insufficient twenty years earlier. Both studies compare more to less within a near-universally exposed population. Neither was designed to answer the question the WHO committee identified in 2004.

What This Means

A comparative adjuvant safety review observed:

Even if a vaccine/adjuvant combination was thought to cause autoimmune disease, this would be very hard to prove, particularly if everyone in the population had received the vaccine.18

This is not a conspiracy. It is a structure. Each decision, no new studies because no new signals, excluding unvaccinated children because they are not comparable, using dose-response models that cannot detect threshold effects, is made for defensible reasons by reasonable people. Together, they produce a system that cannot generate the evidence it would need to detect the problem it cannot rule out.

The foundational studies opened a question with 4 rabbits and 84 adults.19 Surveillance was not built to close it. The largest study ever conducted on this question had 15,237 unexposed children in its dataset and did not compare them. The WHO called for the studies that could. They have not been done. Billions of doses were administered. The question, aluminium or none, was not asked.

If you spot an error in my reasoning, data, or sources, tell me. I'll correct it publicly.

  1. VAERS captures <1% to 76% of adverse events depending on severity and media visibility (Miller 2020). Active systems like the VSD use risk windows of days to weeks. No system monitors by shared ingredient. The structural blind spots are examined in What the Safety Net Can Catch. ↩

  2. "No evidence of harm" can mean four different things: we looked and found nothing, we looked but couldn't see, we didn't measure, or we measured at the wrong time. In practice these are often conflated. The distinction is examined in What "No Evidence of Harm" Actually Means. ↩

  3. Petersen SB, Gluud C. "Was amorphous aluminium hydroxyphosphate sulfate adequately evaluated before authorisation in Europe?" BMJ Evidence-Based Medicine 2021;26(6):285-289. doi:10.1136/bmjebm-2020-111419. PubMed 32763959. Based on direct queries to the EMA: ASK-50308 (October 2018) and ASK-53619 (January 2019). The EMA responded that no randomised clinical safety studies comparing aluminium adjuvant to inactive comparator existed, and that none were needed because of "decades of use." ↩

  4. Same source as above. The EMA stated that AAHS (Amorphous Aluminium Hydroxyphosphate Sulfate), used in Gardasil-9, Recombivax HB, and other Merck vaccines, "was introduced without any prelicensure safety evaluation." Petersen & Gluud 2021, citing EMA response ASK-53619. ↩

  5. Aluminium content verified from Summary of Product Characteristics (SmPCs) for each vaccine. All inactivated, toxoid-based, and adsorbed vaccines on the UK (NHS Green Book) and Italian (PNPV) routine schedules contain aluminium salts as adjuvant. The only aluminium-free vaccines are live vaccines (rotavirus, MMR, nasal flu), certain conjugate vaccines (MenACWY), and some unadjuvanted products (inactivated flu for pregnancy, RSV for pregnancy). ↩

  6. UK schedule: NHS Green Book, Chapter 11 (The UK Immunisation Schedule). From July 2025, the infant primary series is: 8 weeks (Infanrix Hexa + Bexsero), 12 weeks (Infanrix Hexa + Bexsero), 16 weeks (Infanrix Hexa + Prevenar 13). The previous schedule (pre-July 2025) had PCV at 12 weeks and MenB at 16 weeks; these were swapped but aluminium-containing vaccines remain at every primary visit. Italian schedule: Piano Nazionale di Prevenzione Vaccinale (PNPV), Ministero della Salute. Primary series at 3, 5, and 11 months. ↩

  7. Andersson NW, Bech Svalgaard I, Hoffmann SS, Hviid A. "Aluminum-Adsorbed Vaccines and Chronic Diseases in Childhood: A Nationwide Cohort Study." Ann Intern Med 2025;178(10):1369-1377. doi:10.7326/ANNALS-25-00997. Study population: 1,224,176 children born in Denmark 2001-2018. Of these, 15,237 (1.2%) had received zero aluminium-adsorbed vaccines by age 2. These children entered the main continuous model at 0 mg but were never compared as a discrete group. Categorical analyses used >0 mg groups only. A sensitivity analysis excluded them because of healthcare utilisation differences. No direct zero-vs-any comparison was reported. Primary neurodevelopmental results were in the protective direction: any neurodevelopmental disorder HR 0.93 (0.90-0.97), ASD 0.93 (0.89-0.97), ADHD 0.90 (0.84-0.96). Whether this reflects healthy vaccinee bias, a genuine protective effect, non-linear dose response, or something else cannot be determined from a dose-response design without an unexposed comparison group. A correction to the supplementary material was published two days after the original epub (doi:10.7326/ANNALS-25-03233) without specifying what changed. ↩

  8. Andersson NW, Bech Svalgaard I, Hoffmann SS, Hviid A. Response to journal comments on acpjournals.org, 2025. The authors stated: "It is fair to say that our study does not evaluate the hypothesis that any exposure to aluminum vaccine adjuvants, irrespective of the cumulative amount received, increases the risk of these outcomes." This concession was not in the published paper itself. Black SD and Salmon DA (Global Vaccine Data Network / Johns Hopkins) commented separately that "very little variability in terms of the exposure to alum" limits the study's ability to assess dose-related effects, and concluded the question "remains an open question with large public health implications." ↩

  9. Glanz JM, Newcomer SR, Narwaney KJ, et al. "A population-based cohort study of undervaccination in 8 managed care organizations across the United States." JAMA Pediatrics 2013;167(3):274-281. Children undervaccinated by parental choice had significantly lower rates of outpatient visits, emergency department visits, and hospitalisations than age-appropriately vaccinated children, consistent with lower healthcare utilisation and therefore lower ascertainment of diagnoses. See also Smith PJ, et al. "Children who have received no vaccines: who are they and where do they live?" Pediatrics 2004;114(1):187-195. ↩

  10. Daley MF, Reifler LM, Glanz JM, et al. "Association Between Aluminum Exposure From Vaccines Before Age 24 Months and Persistent Asthma at Age 24 to 59 Months." Academic Pediatrics 2023;23(1):37-46. PubMed 36180331. Study population: 326,991 children born 2008-2014 in the VSD. Adjusted hazard ratio for persistent asthma: 1.19 (95% CI 1.14-1.25) per 1 mg increase in cumulative aluminium exposure. The authors noted potential confounding by healthcare-seeking behaviour, limited breastfeeding data (~13% of cohort), and inability to capture dietary aluminium exposure. The effect was not robust across all subgroups. ↩

  11. Andersson 2025, main paper: 4,806 ASD and 1,580 ADHD diagnoses by age 5. Supplement Figure 12 (follow-up extended to age 8): 12,126 ASD and 12,091 ADHD diagnoses. Total neurodevelopmental outcomes: 6,386 at age 5, 19,466 at age 8. ↩

  12. Andersson 2025, Supplement Figure 4 (stratified by birth year). Asperger syndrome in children born 2007-2018: HR 1.67 (95% CI 1.01-2.77), 51 cases. This is the only individual neurodevelopmental outcome with a statistically significant positive association anywhere in the paper or supplement. Atopic dermatitis in the same cohort: HR 1.03 (95% CI 1.01-1.06), 11,255 cases. ↩

  13. Andersson 2025, Supplement Figure 11 (categorical risk differences, >0-1.5 mg and >1.5-3 mg vs >3-4.5 mg reference). Neurodevelopmental outcomes: >0-1.5 mg group had 11.05 more cases per 10,000 than the >3-4.5 mg group (95% CI -5.29, 27.40, not significant); >1.5-3 mg group had 9.73 fewer cases per 10,000 (95% CI -14.05, -5.41, significant). The lowest-dose group having more neurodevelopmental outcomes than the highest-dose group is consistent with a non-monotonic dose-response pattern. The authors described these categorical analyses as showing results that "did not suggest nonlinear associations." ↩

  14. Holt D, et al. "Danish MMR vaccination coverage is considerably higher than reported." Dan Med J 2017;64(2):A5345. PubMed 28157059. Of 246 children marked as unvaccinated in registry data, 55% had actually received the MMR vaccine according to medical records. The discrepancy was attributed to "administrative reimbursement errors." This study examined MMR specifically; the rate of misclassification for aluminium-adsorbed vaccines in the same registry has not been separately measured but uses the same data infrastructure. ↩

  15. WHO Global Advisory Committee on Vaccine Safety. "Adjuvants." Weekly Epidemiological Record 2004;79(41):369-376. WHO GACVS. The committee recommended: "(1) improved post-licensure surveillance of adjuvanted vaccines; (2) clear case definitions for potential adjuvant-related adverse events; (3) Phase IV clinical studies; (4) validated animal models." On models: "Validated animal models for adjuvant safety testing do not exist, yet they will be required." The 2004 discussion arose in the context of novel adjuvants being developed for HIV, malaria, and HPV vaccines. Whether the four recommendations apply only to novel adjuvants or to all adjuvants including aluminium is an interpretive question. None of the four has been fulfilled for aluminium either. ↩

  16. WHO Global Advisory Committee on Vaccine Safety, November 2025 meeting. The committee reviewed systematic reviews and reaffirmed that "high-quality scientific evidence indicates that vaccines, including those with aluminium, do not cause autism" and found "no association between aluminium-adjuvanted vaccines and chronic or systemic diseases." None of the 2004 committee's four specific recommendations (standing surveillance, case definitions, Phase IV trials, validated animal models) were reported as fulfilled. ↩

  17. The Clinical Practice Research Datalink (CPRD) contains anonymised primary care records for approximately 13 million UK patients, linkable to Hospital Episode Statistics (HES) and ONS death registrations. Schedule changes (e.g., the addition of MenB in 2015, changes in 6-in-1 products with different aluminium content) create natural experiments that could isolate adjuvant-specific effects in large populations. No such study has been published. ↩

  18. Petrovsky N. "Comparative Safety of Vaccine Adjuvants: A Summary of Current Evidence and Future Needs." Drug Safety 2015;38(11):1059-1074. doi:10.1007/s40264-015-0350-4. PMC4615573. The quoted passage appears in the context of discussing why adjuvant-related autoimmune disease would be nearly impossible to prove when population-wide exposure eliminates a clean control group. ↩

  19. The Mitkus 2011 FDA pharmacokinetic model and its empirical foundation (Flarend 1997, 2 rabbits per adjuvant type, 28 days) are examined in detail in Four Rabbits. ↩