The Aluminium Low-Dose Problem

 · 26 min read
 · Nulla Verba
Table of contents

When aluminium hydroxide is injected into muscle, the body can respond in two ways. If the dose triggers enough local inflammation, immune cells wall it off in a granuloma at the injection site. If it doesn't, macrophages, immune cells that engulf foreign particles, carry them elsewhere, including the brain.1

This is not the same as dietary aluminium. We ingest aluminium daily, but the oral ionic form is absorbed at roughly 0.1-0.4%, filtered by the kidneys, and largely blocked by the blood-brain barrier.2 The particulate form in vaccines is different: it is 100% in tissue from the moment of injection, engulfed intact by macrophages, and transported via immune pathways that bypass the blood-brain barrier entirely.3

Three controlled experiments, in mice, sheep, and rats, have found that injected aluminium adjuvant reaches the brain and causes measurable effects: microglial activation and behavioural changes in mice, gene disruption in sheep, and statistically significant brain aluminium accumulation in rats.1 13 25 These effects did not follow a linear dose-response. They depended on whether the aluminium stayed at the injection site or escaped.

The question for vaccine safety is not whether more aluminium is worse than less. It is how much escapes the injection site. The regulatory model assumes aluminium is excreted within weeks.4 The surveillance systems that would need to catch a problem are not designed to detect chronic or delayed effects.5

In 2017, the first of those three experiments tested this directly. Three doses of aluminium hydroxide adjuvant were injected into mice. The lowest dose produced 50 times more brain aluminium than controls. The two higher doses produced less brain aluminium than controls.1 The explanation: at the lowest dose, no granulomas formed. The particles escaped.

This matters because every large human study of aluminium adjuvant safety, including the 1.2-million-child Danish cohort,6 used a linear model that assumes more is worse than less. A linear model measures gradients. It is poorly suited to detect a step from zero to any exposure.6 And if the relationship reverses at lower doses, opposing slopes can cancel, returning null even when the data contain a real signal.

The largest dataset ever assembled on this question contains a non-monotonic pattern in its own supplementary data: the lowest-dose group had more neurodevelopmental cases than the highest-dose group.7 The authors described this as results that "did not suggest nonlinear associations." A linear model can detect harm when the pattern is linear (as Daley 2023 found for asthma). It cannot detect a pattern it was not designed to see.

Background: Aluminium Is Not Linear

Aluminium interacts non-linearly with biological systems. This has been documented in cell cultures and animal studies since at least 1987, using soluble ionic aluminium (a different form from the particulate adjuvant in vaccines, but the same element):

Study System Finding
Lieberherr et al. 1987 Mouse osteoblasts (in vitro) Cell proliferation stimulated 30% at 10⁻⁶ M, inhibited 26% below controls at 10⁻⁵ M. Same biphasic pattern across six endpoints. Reversal at ~1 μM.8
Platt et al. 1995 Rat hippocampal neurons (in vitro) Baseline neuronal excitability: low concentration slightly increased, high concentration suppressed and abolished (in vitro only; not seen in vivo). LTP impairment was monotonic.9
Tsunoda & Sharma 1999 Mouse hypothalamus (oral Al, 4 doses) Lowest dose (5 mg/L) produced the largest decrease in dopamine. Higher doses (25, 125 mg/L) had less effect.10
Kim 2003 Rat frontal cortex (oral Al, 3 doses) Low dose increased nNOS expression 10%, high dose decreased it 17%. Explicitly described as "biphasic effect."11

These studies used oral or in vitro ionic aluminium, not injected particulate adjuvant. The granuloma-trapping mechanism does not apply to them. What they establish is that aluminium as an element has non-linear biological properties across multiple cell types and endpoints. The question was whether the same applies to the particulate form injected in vaccines, and if so, through what mechanism.

The Experiment

In 2017, Crépeaux and colleagues at INSERM tested three doses of aluminium hydroxide adjuvant (200, 400, and 800 μg/kg) injected intramuscularly in mice, alongside a PBS control group. The dosage protocol was recommended by AFSSAPS, the French drug safety agency. The study was co-funded by ANSM (the French national drug agency) and CMSRI.1

The results at 180 days:

Outcome Control 200 μg/kg (low) 400 μg/kg (mid) 800 μg/kg (high)
Brain Al (median μg/g) 0.020 1.003 0.014 0.016
Microglial activation (brain inflammation) Baseline Significant increase No change No change
Locomotor activity Baseline Decreased No change No change
Grip strength Baseline Decreased No change No change
Injection-site granulomas N/A None (0/3) Present (3/3) Present (3/3)

The lowest dose produced 50 times more brain aluminium than controls. The two higher doses produced less brain aluminium than controls. Only the lowest dose caused behavioural changes. Only the lowest dose had no granulomas.

The explanation: at 200 μg/kg, the injected suspension contained only small particle agglomerates, in the size range that macrophages readily engulf and transport.1 At higher doses, larger agglomerates formed, triggering granuloma formation that walled off the aluminium at the injection site. The granulomas did their job. The particles that escaped did not.

The neurological effects tracked the brain aluminium: where brain Al was high (low dose), there were effects; where brain Al was low (mid and high dose), there were none. The key variable is not what aluminium does once in the brain. It is how much gets there.

The Sheep

The same pattern appeared independently in a different species, from a different lab.

At the University of Zaragoza, veterinary pathologist Lluís Luján investigated a neurological syndrome in Spanish sheep following mass bluetongue vaccination. His group designed a controlled experiment: 84 lambs, three groups, 19 subcutaneous injections over 15 months. Vaccine plus aluminium, aluminium alone, and PBS saline control.12

The granuloma findings mirrored Crépeaux: 100% of aluminium-exposed animals developed granulomas, zero in controls. The vaccine group had more severe granulomas than the aluminium-only group, because the vaccine antigens provoked stronger local inflammation.12

The brain findings went further. The aluminium-only group, the one with less local inflammation and fewer granulomas, showed five times more gene disruption in the brain than the vaccine-plus-aluminium group (63 genes with altered activity vs 12).13 Aluminium accumulated preferentially in spinal cord grey matter, with the highest individual values in the aluminium-only animals.14

Less inflammation at the injection site. More aluminium in the brain. The same direction as Crépeaux, in a different species, from the University of Zaragoza.15

The Human Data

No randomised trial has tested aluminium adjuvant against an inert placebo in children. The two trials that exist enrolled 84 adults, gave a single dose each, and measured acute reactions for days to weeks.20 Neither was designed to detect autoimmune or neurodevelopmental outcomes, which take months to years to manifest. The question of dose-response shape has been approached only through observational data, where cumulative aluminium is calculated from vaccine records, not controlled experimentally.

Within that limitation, one study formally tested whether the relationship is linear. In 2021, Glanz et al. examined 584,171 children in the US Vaccine Safety Datalink, testing three measures of the childhood immunisation schedule against type 1 diabetes. For schedule adherence and cumulative antigen exposure, the linear model held. For cumulative aluminium, it did not: a formal test rejected linearity at P=.007, forcing the authors to switch to categorical analysis.21 The non-linearity was specific to aluminium. Of three exposure variables tested in the same cohort with the same methods, only aluminium broke the linear model.

The Andersson 2025 Danish cohort used a linear model for its primary analysis and found no association between cumulative aluminium and chronic disease. It did not apply the formal linearity test that Glanz used.22 The categorical analysis in Supplement Figure 11 tells a different story. When children were grouped by exposure level, the lowest-dose group (>0-1.5 mg) had 11 more neurodevelopmental cases per 10,000 than the highest-dose group (>3-4.5 mg). The mid-dose group had 9.7 fewer. The lowest dose was the worst, not the highest.7

The Bernstein 2008 vaccine trial included arms with three aluminium concentrations (0.0875, 0.175, 0.350 mg), though antigen dose also varied across these arms.16 Four of four systemic reaction patterns (headache, redness, feverishness, nausea) peaked at the middle arm, not the highest. With n=30-59 per arm and confounded variables, this is suggestive at best.17

The formal test exists. It has been applied once, to one outcome, in one cohort. It rejected linearity. It has not been applied to the largest dataset on the question.

Why It Matters

A linear dose-response model assumes one shape: a straight line. It can detect whether more is worse than less. It cannot detect whether intermediate doses are worse than either extreme. It cannot detect a threshold.18

If local containment determines how much aluminium reaches the brain, then:

  • A study finding "no linear dose-response" has not ruled out harm. It has ruled out one shape of harm.
  • A study comparing 3 mg to 4 mg cannot detect effects that depend on whether granulomas formed.
  • The safety conclusion depends on the model, not on the data.

Every large human study cited for aluminium adjuvant safety used a linear model as its primary analysis.19 20 The one study that formally tested the assumption found it failed. No counter-evidence for non-linear dose response exists for injected particulate aluminium, because no one except the two groups above has tested it.

What This Does and Does Not Show

Each experiment tested a different variable: dose, inflammation level, antigen presence. All three changed where the aluminium ended up.

The effect sizes in the animal studies are large. Crépeaux's lowest dose produced 50 times more brain aluminium than controls. The sheep Al-only group had five times more gene disruption than the vaccine group. Granulomas were present in 100% of higher-dose animals and 0% of the lowest-dose mice. These are not borderline findings rescued by statistics.23 24

One independent study deserves close reading. Weisser et al. 2019, from Germany's Paul-Ehrlich-Institut, injected actual marketed vaccines intramuscularly in rats and measured brain aluminium at day 80.25 The authors concluded that brain increases were "chance findings." Their own data shows all three vaccine groups had statistically significant brain Al elevations; neither plain adjuvant group did. Plain Al hydroxide stayed at the injection site. Vaccines mobilised it.

The relationship to the sheep findings is not straightforward. In the sheep, Al alone (less inflammation) produced more brain gene disruption than Al plus vaccine, but behavioural changes were actually more pronounced in the vaccine group.26 In Weisser, plain Al went nowhere (total retention), while vaccines mobilised Al to the brain. The CNS patterns run in different directions across species and endpoints.

Three controlled experiments, testing three different variables (dose in mice, inflammation level in sheep, antigen presence in rats), each found that the immune context of the injection determined where aluminium ended up. None produced a simple more-is-worse pattern. The mechanism connecting them is not yet clear, but the assumption that every human study relies on, that more aluminium means more risk, has not been confirmed in any animal model that tested it.

The doses in Weisser are 7-18x above Crépeaux's range, so the granuloma-trapping model alone cannot explain the vaccine-group brain Al. But the study's own data, read past the authors' conclusions, raises questions the authors did not pursue.

What is missing is direct replication. Crépeaux's finding comes from one lab (INSERM Créteil), one mouse strain (CD1), one time point (180 days). No second lab has repeated the experiment in mice. The Zaragoza sheep work is independent confirmation through a different variable (inflammation level rather than dose), but uses a different species, route (subcutaneous vs intramuscular), and total dose.15

The doses differ from the human schedule. The sheep received roughly 80 mg total aluminium over 15 months,12 roughly 20 times what a UK infant receives by 16 weeks. Crépeaux's lowest mouse dose (200 μg/kg) was scaled to approximate two human vaccine doses using standard body-surface-area allometry (×12.3 factor from FDA 2005 guidance, which the authors state was recommended to them by AFSSAPS toxicologists), but mice received three injections over eight days rather than six over four months.1 The granuloma-trapping mechanism depends on local tissue response, which varies with route, volume, and timing.

The background evidence from in vitro and oral studies shows that aluminium interacts non-linearly with biological systems through multiple mechanisms. These studies used ionic aluminium, not particulate adjuvant, so the granuloma mechanism does not apply to them. No study using ionic oral aluminium can confirm or rule out the transport mechanism that Crépeaux proposes for injected particles.

The Model and Its Predictions

Crépeaux's model predicts that local containment determines systemic distribution. Less containment, whether from lower dose, smaller particle size, or weaker inflammation, means more aluminium reaches the brain. Two independent experiments tested this through different variables and both confirmed it: in mice through dose, in sheep through inflammation level.

The brain effects are not hypothetical. Crépeaux found microglial activation and behavioural changes. The sheep showed disruption of genes required for neural differentiation. These are functional neurological effects, not just tissue measurements.

The containment itself raises a further question. Granulomas are not inert vaults. They are living tissue maintained by active immune processes, documented at human injection sites up to 12 years after vaccination.27 If a granuloma breaks down, through trauma, infection, immune decline, or simply time, the aluminium it contains enters the same transport pathway. What was contained at 8 weeks could reach the brain at 8 years.

The hypothesis is specific. It makes testable predictions. Two controlled experiments in different species, testing different variables, produced results consistent with those predictions. The question has never been tested in humans with the right model.28

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

  1. Crépeaux G, Eidi H, David MO, et al. "Non-linear dose-response of aluminium hydroxide adjuvant particles: Selective low dose neurotoxicity." Toxicology 2017;375:48-57. doi:10.1016/j.tox.2016.11.018. PubMed 27908630. Critiqued by Hawkes & Benhamu 2017 on process grounds; authors responded (Crépeaux & Gherardi 2018, doi:10.1016/j.tox.2018.06.007). Dosage protocol recommended by AFSSAPS. Co-funded by ANSM and CMSRI; CMSRI had no role in design, analysis, or writing per the authors' disclosure. Co-authors Exley, Shaw, and Gherardi have been reported as members of CMSRI's Scientific Advisory Board, which was not separately disclosed as a conflict of interest in the paper. ↩↩↩↩↩↩

  2. Oral aluminium absorption varies widely by dose, form, and method of measurement. Greger 1993 (Annu Rev Nutr 13:43-63) reviews the range: from 0.006-0.036% (tissue accumulation in rats) to 0.78% (urinary excretion estimate at 5 mg/day dietary Al, from Ganrot) to ~1% (²⁶Al tracer, single subject, tiny dose). The ATSDR 2008 Toxicological Profile for Aluminum (atsdr.cdc.gov/toxprofiles/tp22.pdf) states gastrointestinal absorption is "generally in the range of 0.1-0.4% in humans" (Chapter 2) and uses 0.1% in deriving its oral Minimal Risk Level. A separate passage states "approximately 0.1-0.6%" including more bioavailable forms like aluminium citrate. The Mitkus 2011 FDA model uses 0.78% (the highest estimate in the range, from Ganrot via Greger), which is 7.8x the ATSDR's own value. The kidneys are the primary route of excretion for absorbed ionic aluminium. The blood-brain barrier limits but does not completely prevent ionic aluminium entry; some brain regions (pineal gland, area postrema, circumventricular organs) lack full BBB protection. See Four Rabbits for the full pharmacokinetic chain. ↩

  3. Gherardi RK, Eidi H, Crépeaux G, et al. "Biopersistence and brain translocation of aluminum adjuvants of vaccines." Front Neurol 2015;6:4. PMC4318414. Review documenting the "Trojan horse" mechanism: aluminium adjuvant particles are engulfed by macrophages at the injection site and transported via CCL2-dependent immune pathways to distant organs including the brain. This bypasses the blood-brain barrier because the aluminium travels inside cells, not as dissolved ions in blood. The mechanism was confirmed using fluorescent nanomaterials tracked to draining lymph nodes and brain (Khan et al. 2013). ↩

  4. The FDA's pharmacokinetic model for aluminium adjuvant safety (Mitkus 2011) traces to a 1997 study of 2 rabbits per adjuvant type, observed for 28 days, with lost bone samples. Of the four rabbits' brains, three were measured and all contained aluminium at 28 days; the fourth (AP rabbit #2, the animal with the highest blood aluminium) was lost during preparation (Flarend 1997, Table 2). When Weisser et al. 2019 injected actual vaccines into rats, they found statistically significant brain Al increases that the rabbit-derived model predicted should be undetectable.25 The model has not been tested in infants, who have immature blood-brain barriers and different immune responses from either rabbits or adult rats. See Four Rabbits. ↩

  5. Post-market surveillance systems (VAERS, VSD, Yellow Card) are optimised for acute, distinctive, temporally clustered events. Chronic, diffuse, or delayed signals, the kind the granuloma-trapping mechanism would produce, fall outside their detection range. See What the Safety Net Can Catch. ↩

  6. Andersson et al. 2025 is the largest study of cumulative aluminium exposure from vaccines: 1,224,176 Danish children, 50 chronic conditions tested. The study used a dose-response design (HR per 1-mg increase) and found no linear association. Its authors conceded the design "does not evaluate the hypothesis that any exposure... increases the risk." The study's structural limitations, its 15,237 unexposed children who were never compared as a group, and the regulatory loop that shaped its design are examined in The Loop. ↩↩

  7. Andersson NW, et al. Ann Intern Med 2025;178(10):1369-1377. Supplement Figure 11. Categorical risk differences comparing >0-1.5 mg and >1.5-3 mg groups against >3-4.5 mg reference. NDD: lowest-dose group +11.05 per 10,000 (CI -5.29, 27.40, not significant); mid-dose group -9.73 per 10,000 (CI -14.05, -5.41, significant). The pattern is non-monotonic: lowest dose worst, highest dose intermediate, mid-dose best. The authors described these results as showing "did not suggest nonlinear associations." The study and its supplementary findings are examined in The Loop. ↩↩

  8. Lieberherr M, Grosse B, Cournot-Witmer G, et al. "Aluminum action on mouse bone cell metabolism and response to PTH and 1,25(OH)2D3." Kidney International 1987;31:736-743. doi:10.1038/ki.1987.60. PubMed 3033386. Tested seven concentrations of ionic AlCl₃ (10⁻⁸ to 10⁻⁵ M) on osteoblast-like cells. DNA content increased 30% at 10⁻⁶ M and decreased 26% below controls at 10⁻⁵ M (Table 1). The same biphasic pattern appeared across six endpoints (alkaline phosphatase, acid phosphatase, ornithine decarboxylase, collagen synthesis, bone resorption). The reversal from stimulation to inhibition occurred consistently at ~10⁻⁶ M. This is ionic soluble aluminium, not particulate adjuvant; the granuloma mechanism does not apply, but it establishes that aluminium has non-linear biological properties at the cellular level. ↩

  9. Platt B, et al. "Aluminum Impairs Hippocampal Long-Term Potentiation in Rats In Vivo and In Vitro." Exp Neurol 1995;134(1):73-86. doi:10.1006/exnr.1995.1038. PubMed 7672040. Ionic Al applied to hippocampal neurons. In vitro (brain slices), baseline excitability was biphasic: 0.68 μg/ml slightly increased population spike amplitude, 2.7 μg/ml decreased it, 5.4 μg/ml abolished it. In vivo (intracerebroventricular injection), baseline PS amplitude was flat at ~100% across all three acute doses (Fig. 1A-D); the biphasic pattern was not observed. LTP impairment was monotonic in both preparations (more Al = more impairment). The authors describe the in vitro excitability finding as "biphasic." ↩

  10. Tsunoda M, Sharma RP. "Altered dopamine turnover in murine hypothalamus after low-dose continuous oral administration of aluminum." J Trace Elem Med Biol 1999;13(4):224-31. doi:10.1016/S0946-672X(99)80040-6. PubMed 10707345. Four doses of oral ionic Al ammonium sulfate (0, 5, 25, 125 mg/L) for 4 weeks. The lowest dose produced the largest decrease in hypothalamic dopamine, DOPAC, and HVA. Higher doses had progressively less effect. This is the Crépeaux-type pattern (lowest dose = strongest effect) but with oral ionic aluminium, so the granuloma mechanism does not apply. The non-linearity may arise at the gut absorption, BBB transport, or neurochemical response stage. ↩

  11. Kim K. "Perinatal exposure to aluminum alters neuronal nitric oxide synthase expression in the frontal cortex of rat offspring." Brain Res Bull 2003;61(4):437-41. doi:10.1016/s0361-9230(03)00159-x. PubMed 12909287. Three doses of oral ionic AlCl₃ (0, 5, 10 mM in drinking water, perinatal). Low dose increased nNOS expression 10%, high dose decreased it 17%. Explicitly described as "biphasic effect." This is a direction reversal (stimulation vs inhibition at different doses), a genuine dose-response non-linearity but a different pattern from Crépeaux's (where the lowest dose simply produced the largest effect in one direction). ↩

  12. Asín J, et al. "Granulomas Following Subcutaneous Injection With Aluminum Adjuvant-Containing Products in Sheep." Vet Pathol 2019. doi:10.1177/0300985818809142. PubMed 30381018. 84 lambs, three groups (vaccine+Al, Al-only, PBS), 19 injections over 15 months. Published in a mainstream veterinary pathology journal, not retracted. ↩↩↩

  13. Asín J, et al. J Inorg Biochem 2020. doi:10.1016/j.jinorgbio.2019.110934. PubMed 31783216. Al-only group: 63 differentially expressed genes in brain vs control. Vaccine+Al group: 12. The lncRNA TUNA (required for neural differentiation) was downregulated only in the Al-only group. n=4 per group for transcriptome analysis. ↩↩↩

  14. de Miguel R, Asín J, et al. J Inorg Biochem 2020. doi:10.1016/j.jinorgbio.2019.110871. PubMed 31901536. Aluminium accumulated preferentially in spinal cord grey matter. Al-only group had the highest individual values. ↩

  15. Luján is a veterinary pathologist at the University of Zaragoza. No CMSRI funding documented; funded by Spanish Ministry of Education and University of Zaragoza. The transcriptome paper was originally published in Pharmacological Research (Nov 2018), withdrawn after anonymous complaint (Jan 2019), and republished in J Inorganic Biochemistry (Feb 2020) in a special issue edited by Christopher Exley. The granuloma paper (Vet Pathol) and spinal cord paper (J Inorg Biochem, separate from the special issue) have not been retracted. Luján collaborates within the broader Gherardi/Exley research network but the sheep work arose independently from a field veterinary problem (the 2007-2008 European bluetongue crisis). ↩↩

  16. Bernstein DI, et al. J Infect Dis 2008;197:667-75. doi:10.1086/527489. PubMed 18260764. H5N1 vaccine trial with three Al hydroxide concentrations. Non-monotonic patterns: headache 23%→35%→29%, redness 7%→16%→10%, feverishness 10%→13%→7%, nausea 7%→10%→3%. All peak at mid-dose. ↩

  17. With n=30 vs n=59 per arm, the minimum detectable difference at 80% power is ~19-25 percentage points depending on baseline rate. The observed mid-to-high-dose differences are 3-12pp, well below the detection threshold. The study was not powered to detect non-linear patterns. But the consistent direction (4/4 patterns peaking at mid-dose) has a sign-test probability of p=0.0625. ↩

  18. If your prior is "Al dose-response is probably linear because most things are," a null linear result is strong evidence for safety. If your prior is "Al dose-response may be non-linear because the granuloma mechanism predicts it and two animal studies confirmed it," the same null linear result is uninformative, because the model can't detect what the prior predicts. The study design determines which priors get updated. ↩

  19. Daley 2023 (n=326,991) and Andersson 2025 (n=1,224,176) both used Cox proportional hazards regression with cumulative aluminium as a continuous linear variable (HR per 1-mg increase). Daley tested the linearity assumption formally (it held for asthma, P=.506). Andersson did not apply the formal test. Glanz 2021 (n=584,171), from the same group as Daley, found the assumption fails for type 1 diabetes (P=.007). All three are examined in The Loop. ↩

  20. A 2022 systematic review identified 102 RCTs of aluminium adjuvants. Only two compared aluminium to an inert placebo, enrolling 84 adults total. Both found higher reactogenicity in the aluminium arm. Landrum 2017 measured reactions for 7 days before pooling the aluminium and saline arms; Basavaraj 2014 measured local reactions for 7 days and systemic for 21 days. Neither measured autoimmune, neurodevelopmental, or long-term outcomes. See 2 of 102. ↩↩

  21. Glanz JM, Clarke CL, Daley MF, et al. "The Childhood Vaccination Schedule and the Lack of Association With Type 1 Diabetes." Pediatrics 2021;148(6):e2021051910. doi:10.1542/peds.2021-051910. PubMed 34851413. CDC-funded (contract 200-2012-53582), CDC coauthor (DeStefano). N=584,171 in 8 VSD sites, 1,132 T1DM cases. Linearity tested via Kolmogorov-type supremum test: rejected for Al (P=.007), held for antigen and schedule adherence. The same group's 2023 asthma study applied the same test to the same variable and linearity held (P=.506), confirming the violation is outcome-specific. Cumulative Al range: 0-9.35 mg. The linearity violation and its implications for the Danish cohort are examined in The Loop. ↩

  22. Andersson 2025 studied 50 outcomes including T1DM using a linear model. They tested the proportional hazards assumption formally (Schoenfeld residuals) but checked linearity only through categorical secondary analysis, not the supremum test that Glanz and Daley applied to the same variable. Their T1DM result was null: HR 0.98 (0.88-1.10). Glanz 2021 showed the linear model fails for this exact outcome-variable pair (P=.007). Andersson cited Daley 2023 (where linearity held for asthma) but not Glanz 2021 (where it did not). See The Loop. ↩

  23. "Small n" is a valid concern when effect sizes are small. These are not. Crépeaux's brain Al: 1.003 vs 0.020 μg/g, a 50-fold difference (Kruskal-Wallis p=0.017, n=5/group). Microglial activation: 59% increase (p=0.033, n=3/group). Granulomas: 0/3 vs 3/3, perfect binary separation. The sheep transcriptome: 63 vs 12 differentially expressed genes. When effects are this large, the limitation is not statistical power but replication: one lab (mice) and one independent confirmation (sheep, different variable). A second mouse replication by an independent group would settle the question regardless of sample size. ↩

  24. Hawkes D, Benhamu J. "Questions about methodological and ethical quality of a vaccine adjuvant critical paper." Toxicology 2017;389:53-54. doi:10.1016/j.tox.2017.06.009. PubMed 28669868. Raised concerns about ethics approval clarity, methods clarity, and funding transparency. Did not challenge the data or findings. Crépeaux & Gherardi's published response (doi:10.1016/j.tox.2018.06.007) addressed each point: ethics approved by Darwin committee, dosage recommended by AFSSAPS, CMSRI had no design/analysis role. ↩

  25. Weisser K, Göen T, Oduro JD, et al. "Aluminium in plasma and tissues after intramuscular injection of adjuvanted human vaccines in rats." Arch Toxicol 2019;93(10):2787-2796. doi:10.1007/s00204-019-02561-z. PubMed 31522239. From Germany's Paul-Ehrlich-Institut. Single IM injection in male Wistar rats, six groups (three marketed vaccines, two plain adjuvants, saline), tissues at day 80. Doses 1,400-3,600 μg/kg, 7-18x above Crépeaux's range. ↩↩↩

  26. Asín J, et al. J Inorg Biochem 2020 (same cohort as 13). Behavioural testing (T-maze, Open Field, Novel Object, social observation) showed reduced affiliative behaviour, increased aggression, stereotypies, and elevated stress biomarkers in both Al groups vs PBS controls. Critically, these effects were MORE pronounced in the vaccine group than the Al-only group, which is the opposite direction from the transcriptome data (where Al-only showed 5x more gene disruption). This dissociation between behavioural and molecular endpoints in the same animals has not been explained. n=7 per group (single flock). ↩

  27. Gherardi RK, Coquet M, Chérin P, et al. "Macrophagic myofasciitis lesions assess long-term persistence of vaccine-derived aluminium hydroxide in muscle." Brain 2001;124(Pt 9):1821-31. PubMed 11222632. Aluminium-laden macrophage infiltrates documented at deltoid injection sites in patients biopsied up to 12 years after their last aluminium-containing vaccination. The lesion, termed macrophagic myofasciitis (MMF), demonstrates that injected aluminium hydroxide can persist in human tissue for over a decade. ↩

  28. Standard epidemiological tools for detecting non-monotonic dose-response include restricted cubic splines in Cox models (allowing the curve to bend), generalised additive models (GAMs), categorical exposure analysis with a true zero-exposure reference group, change-point or threshold models, and Bayesian model comparison with non-linear priors. None has been applied to cumulative aluminium adjuvant exposure data. The Andersson 2025 dataset, with 1.2 million children and categorical exposure groups already defined, could be re-analysed with these methods without collecting new data. ↩