How to Manufacture Safety

An Essay on the Formula Behind Vaccine Post-Marketing Studies

The Phillips 2020 paper is publicly available. Every number in this essay is taken from the paper itself. What the essay does is walk through what the paper does with them.

---

Adverse events following HPV vaccination: 11 years of surveillance in Australia, published in the journal Vaccine in July 2020, is one of the most cited pieces of post-marketing surveillance in the vaccine literature. It reports 4,551 adverse event reports across 9.4 million doses of quadrivalent HPV vaccine. Its concluding sentence reads: “affirms the safety profile of 4vHPV.”¹

The paper contains, in its own results, evidence that its own conclusion cannot be drawn from its own data.

During 2013 and 2014, Australian school nurses were told to record adverse events on the spot as they injected twelve- and thirteen-year-olds with quadrivalent HPV vaccine. Reporting rates for syncope in that age group quadrupled. When enhanced surveillance ended, rates returned to their prior level.¹

This is not about syncope. It is the difference between a reporting system operating at intensity and the same system operating passively. The paper contains this natural experiment in its results and applies its finding to nothing. The passive rate is used as the safety baseline. The multiplier, which is evidence that the reporting system misses most of what it should catch, is set aside.

This is not an oversight. It is a move.

The Phillips paper is one instance of a formula. Post-marketing vaccine safety studies are built by executing a repeatable sequence of decisions, each producing the appearance of safety without producing its measurement. Once the sequence is visible, it can be recognized in almost any major published safety analysis of the last two decades. The Phillips paper is typical of the genre. That is what makes it worth reading closely.

What follows walks through the formula in order. The Phillips paper is the running example; where the same move appears in other studies, they are named.

This work stays free because paid subscribers make it possible. They get the full book library, the Deep Dive Audio Library, and the Questions for Your Doctor, Before You Consent, and Package Insert series. No grants, no gatekeepers — your subscription is what keeps it that way.

Move 1: Choose an Instrument That Cannot See

Every post-marketing safety study begins with a data source. When the study is being built to affirm safety, the data source is a passive spontaneous reporting system.

Passive means the system does nothing to solicit reports. Providers, patients, and manufacturers may file if they wish. Almost no one does. A CDC-funded study run by the Harvard Pilgrim Health Care team from 2007 to 2010 across 715,000 patients built an automated reporting system that captured events from electronic medical records. The team found that fewer than one percent of adverse events reach VAERS, the American passive equivalent of Australia’s Adverse Events Management System.² The Harvard Pilgrim team offered to integrate their automated system with VAERS. The CDC stopped responding to their communications, and the project was terminated. The Australian AEMS operates on the same passive principle. It has never been shown to perform meaningfully better than VAERS.

Darrell Huff put the principle plainly in 1954: the result of a sampling study is no better than the sample it is based on.¹⁵ A passive reporting system is not a sample of harm. It is a sample of what got past the barriers to reporting.

An instrument that captures one percent of what it is designed to measure cannot be used to demonstrate safety. It can, however, be used to declare safety, because the low rate it produces is the rate that will be reported. The Phillips paper’s crude rate of 39.8 per 100,000 doses is the output of an instrument known to miss most of the events it should catch. The number is real. What it measures is not the rate of harm. It is the rate of harm that reached a system with almost no capacity to receive it.

Move 2: Never Activate It

The choice of a passive system is defensible only if active surveillance is impossible. It is not impossible. Australia demonstrated for two years, in the same paper, that active surveillance is straightforward: instruct the nurses at the injection site to record what happens. The overall reporting rate rose from 39.8 to 72.3 per 100,000 doses.¹ The rate of syncope in twelve- and thirteen-year-olds rose from 7.1 to 29.6.¹ These changes did not require new technology. They required an instruction.

The two-year enhanced period was framed as monitoring the introduction of boys to the program. When boys had been in the program for a year, enhanced surveillance was discontinued. The nine-year passive rate before and after enhanced surveillance is what the paper treats as the safety baseline. The paper does not offer, and the Australian regulator has not funded, any longer-term active surveillance covering the delayed-onset conditions that clinicians in Denmark, Japan, Ireland, and Colombia have documented in adolescent girls following injection.³ ⁴ ⁵

The Harvard Pilgrim system was buried in the same period. What both cases demonstrate is that the choice not to look actively is a choice. The passive rate is not the rate the world produces. It is the rate the surveillance system was structured to produce.

Move 3: Search Where the Light Is

Having chosen an instrument that mostly cannot see, the study then narrows what it is looking for. Adverse events of special interest are pre-specified. The list determines what the analysis can find.

Phillips pre-specified seven: syncope, venous thromboembolism, anaphylaxis, conditions labeled autoimmune, postural orthostatic tachycardia syndrome, Guillain-Barré syndrome, complex regional pain syndrome, and primary ovarian insufficiency.¹ Each of these had already been the subject of at least one published epidemiological study reporting no signal. The list is the list of dismissed concerns.

The Chao 2012 VAERS analysis followed subjects for 180 days after each dose and tracked sixteen predetermined conditions.⁶ The Liu 2018 Ontario Grade 8 cohort study counted a girl as “exposed” only if her condition appeared between seven and sixty days after injection.⁶ The Verstraeten 2008 AS04 analysis pooled events into categories broad enough that specific signals disappeared into noise.⁷ In each case, the analytical window and the condition list were constructed such that most of what recipients report was either outside the window or outside the list.

Not on the Phillips list: menstrual irregularities short of amenorrhea. Chronic pain not meeting complex regional criteria. Cognitive symptoms. Sleep disturbance. Fertility outcomes across the cohort. Cancers of any type in follow-up. Fatigue syndromes. Any condition arising more than a few months after injection.

The conditions the paper searched for are conditions the reporting system might occasionally record. The conditions the reporting system almost never records are the conditions reported by the Japanese, Danish, Irish, and Colombian clusters of injured girls: pain, exhaustion, autonomic dysfunction, menstrual disruption, cognitive impairment, sleep collapse.³ ⁴ ⁵ The formula makes them invisible by definition.

Move 4: Set a Diagnostic Bar the Instrument Cannot Clear

Even for conditions on the pre-specified list, the study then requires diagnostic confirmation. The requirement sounds rigorous. Its effect is to eliminate cases.

The Phillips paper identifies thirteen reports of postural orthostatic tachycardia syndrome. For the seven coded specifically as POTS, it states: “there was insufficient information on symptoms, heart rate, blood pressure, investigations and/or duration of illness to establish a diagnosis of POTS according to published criteria” in any case.¹ It identifies twelve reports of primary ovarian insufficiency. Three are the Australian case series published by Little and Ward in 2014.⁸ Of the remaining nine, “none had sufficient information to confirm a diagnosis.”¹ It identifies thirteen reports of conditions labeled autoimmune. It notes: “There was no pattern regarding time of onset following vaccination.”¹

The diagnostic bar is set at a level a passive reporting system cannot reach. The reporting system does not require diagnostic workup. The regulator can request it and, as the paper notes, “despite multiple attempts, sufficient detail is not always obtained.”¹ The bar was set knowing the instrument could not clear it. Every case that fails to clear it disappears from the count.

Move 5: Exclude What Fails the Bar

The mechanism sits inside its own logic. The reporting system produces cases without full diagnostic workup. The study requires full diagnostic workup. Cases without workup are excluded. The remaining count is reported as reassurance.

Under this logic, the more thoroughly the reporting system fails to gather clinical detail, the safer the product appears. A perfectly executed passive system, one in which no clinician ever performed follow-up investigation, would produce zero confirmed cases of any adverse event of special interest. It would also produce the strongest possible safety declaration.

The same move appears far outside vaccine surveillance. Malcolm Kendrick documents a widely cited Swedish cervical cancer screening study published in the BMJ that excluded 390 women from its final analysis.¹⁶ Those 390 women contained 196 of the 372 total cervical cancer deaths in the study, or 53% of the deaths. The exclusion was not mentioned in the abstract, the introduction, or the discussion of the paper. The remaining data produced the desired conclusion.

Little and Ward’s three ovarian insufficiency cases survived only because they were already published in the peer-reviewed literature.⁸ Deirdre Little, working in New South Wales general practice, had documented that the “age-specific incidence of idiopathic POF in early to mid-adolescence is so rare as to be unknown.”⁸ The Australian regulator responded to her cases by stating the condition was “known to occur naturally in this age group.”⁹ The clinician who identified the condition and the regulator who dismissed it were describing different things. The regulator’s version prevailed.

Move 6: Invoke a Background Rate That Does Not Exist

The residual cases, the small number that survived pre-selection, diagnostic bar, and exclusion, must then be dismissed. The instrument for dismissal is the background rate.

Every safety study invokes it. Reported cases are declared consistent with the rate that would be expected in the general population. Very few safety studies measure it. The Phillips paper is explicit: “these conditions occur at a background rate in the population, irrespective of vaccination, although data on local and age-specific prevalence and incidence is often not available.”¹

Huff called this “the little figure that is not there”: the missing comparison whose absence goes unnoticed because absences do not attract the eye.¹⁵ The Phillips paper invokes a background rate without measuring one, and nothing on the page tells the reader a comparison is missing.

No age-matched unvaccinated cohort is examined. The rates of primary ovarian insufficiency, of new-onset conditions labeled autoimmune, and of orthostatic dysfunction in unvaccinated twelve- and thirteen-year-old Australian girls in 2007 are not established anywhere in the paper. Rates observed after injection are declared consistent with a background rate that is admitted, in the same paragraph, to be unknown.

The pattern repeats across the safety literature. The Ontario Grade 8 cohort study compared vaccinated and unvaccinated girls but counted a condition as vaccine-associated only if it arose between seven and sixty days after injection; a girl developing lupus 90 days after her third injection was, by definition, an “unvaccinated” case in the comparison.⁶ The Chao VAERS analysis compared reported rates to published incidence estimates in the general adult population, not to any measured baseline in the target age group.⁶ In each case, the comparator is a phantom, invoked but never measured, available to absorb whatever residual signal remains.

Move 7: Reframe the Residual Signals

Some signals will survive every prior step. The formula requires a reframing that keeps them from being read as harm.

Three reframes cover most cases.

The first is anxiety. Syncope is the most common acute reaction to injection. It occurs at rates high enough that no methodology can hide it. The Phillips paper handles this by adopting the World Health Organization category “immunisation anxiety-related reaction”. The category attributes the loss of consciousness not to the substance injected but to the psychological state of the child receiving it.¹ Adolescent girls fainting after injection becomes psychology. The higher rate in twelve- and thirteen-year-olds compared to fourteen- and fifteen-year-olds is attributed to greater anxiety in younger recipients. Lower body mass, less mature physiology, and the absence of any dose adjustment for a product delivered at the same volume regardless of the recipient’s size are not considered.

The second is media-induced illness. When clusters of injured girls have organized across countries (Japan, Denmark, Ireland, Colombia), the establishment explanation is media-induced psychogenic illness. The Phillips paper adopts this reading in a single sentence about Denmark: “clusters of non-specific symptoms attributed to POTS and CFS were reported in Denmark and increased following heightened media reporting in 2013 and 2015.”¹ The direction of causation is asserted without argument. Media coverage produced the symptoms, in this reading, rather than reporting them. The Danish Syncope Unit’s more than four hundred documented cases, the Uppsala Monitoring Centre’s signal analysis by Rebecca Chandler, and Denmark’s 2016 replacement of Gardasil with Cervarix following the collapse of public confidence go unmentioned.³

The third is coincidence. First-dose predominance is a strong pattern in the Phillips data. Of the 856 cases of syncope, dose number was documented in 811. All 811 followed dose one. Of the thirty cases of anaphylaxis, dose number was documented in 27. All 27 followed dose one. Of the thirteen cases labeled autoimmune, dose number was documented in nine. All nine followed dose one.¹ First-dose predominance is consistent with sensitization, the mechanism Charles Richet demonstrated in 1901 and received the Nobel Prize for describing in 1913: injection of foreign proteins produces a heightened response to subsequent exposure.¹⁰ It is also consistent with recipients declining subsequent doses after the first reaction, which is itself a signal of harm. The paper mentions neither reading. The numbers are reported and set aside.

The reframing family (anxiety, media-induced illness, coincidence) is not unique to vaccine surveillance. Kendrick documents parallel reframing across pharmaceutical categories: cancer screening benefits reported as saving 300 lives per thousand when the absolute figure is 0.3 per thousand; surrogate endpoints like cholesterol lowering treated as equivalent to preventing death; a “significant benefit” that dissolves the moment one asks how many people would need to take the drug for one to benefit.¹⁶ The labels differ. The function is consistent: relabeling the finding so the reader stops asking what was measured.

Move 8: Never Name What Was Injected

The formula’s structural silence is complete only when the product itself is not analyzed as a variable.

Gardasil contains an aluminum adjuvant. Merck’s product uses Amorphous Aluminum Hydroxyphosphate Sulfate, at a dose per three-dose series among the highest in the childhood injection schedule.¹¹ The clinical trials Merck submitted for the product’s licensure used AAHS itself as the primary control in most protocols, so that whatever damage the adjuvant caused would appear on both sides of the comparison and produce no signal. In Protocol 018, the one exception, the control group received a “carrier solution” containing polysorbate 80, sodium borate, yeast residuals, and L-histidine. This was everything in the injected product except the L1 protein particles and the aluminum.¹¹ The Australian Therapeutic Goods Administration itself confirmed to Deirdre Little in 2015 that the Merck package insert’s reference to a “saline” placebo was a misrepresentation.¹¹

Aluminum adjuvants biopersist. Work by Christopher Exley at Keele University, Romain Gherardi at the Université Paris-Est, and Christopher Shaw at the University of British Columbia has documented the movement of aluminum adjuvant particles from the injection site through the lymphatic system to distant tissues, including brain.¹² Mold and colleagues have found elevated brain aluminum in tissue from children diagnosed with autism.¹³

None of this research enters the surveillance literature. Exley’s findings have been met with institutional silence rather than replication; the Alzheimer’s Association’s website continues to state that “few believe that everyday sources of aluminum pose any threat.”¹² The research the safety studies would need to engage exists. It goes unengaged.

The Phillips paper does not mention aluminum. The word does not appear in the analysis. Adjuvants and sensitization are not discussed. The eleven-year surveillance analysis of an aluminum-adjuvanted product is conducted as though the aluminum were not there.

The formula requires this silence. Once the injected material is named as a variable, the framework required to see the data reappears. A study that named aluminum adjuvant, first-dose predominance, female clustering (every category of serious adverse event except syncope in the Phillips paper is overwhelmingly female), and Richet’s sensitization mechanism would be describing a signal. A study that names none of these describes coincidence.

The Ritual

The formula produces a document. The document is then packaged.

The paper appears in Vaccine, a journal whose editorial board and financial base are integrated with the industry it covers. It is peer-reviewed. Its institutional authorship lists the Therapeutic Goods Administration, the National Centre for Immunisation Research and Surveillance, the Victorian Cytology Service Foundation, and the University of Sydney. Conflicts are declared at the end. One of the authors, Julia Brotherton, discloses that she was an investigator on studies that received grants from Seqirus and Merck; the disclosure is not carried in the abstract or the interpretation section.¹ The final sentence of the abstract reads: “affirms the safety profile of 4vHPV.”¹

The paper then enters the citation chain. Future safety analyses will cite it. Regulators will cite it. Institutional review boards will cite it. Media will cite it. When a family asks whether their daughter’s post-injection condition might be related to the injection, they will be told that a comprehensive eleven-year Australian analysis affirmed safety. What that analysis actually contained will not enter the conversation.

The Formula Is Not New

The eight moves above are not novel to vaccine surveillance. They are the vaccine-surveillance application of a documented genre.

Darrell Huff cataloged the underlying techniques in 1954. How to Lie With Statistics remains in print seventy years later; Bill Gates has repeatedly named it among the books he most recommends, including telling everyone at TED to read it.¹⁷ The moves Huff described (biased samples, missing comparisons, selective inclusion, semantic substitution, the arithmetic that hides behind a chosen average) did not become obsolete. They were absorbed into the standard operating procedure of regulatory science.

Malcolm Kendrick, a British general practitioner writing in 2014, documented the pharmaceutical application. Doctoring Data records how relative risk framing conceals absolute risk, sometimes by a factor of a thousand; how surrogate endpoints like cholesterol and blood pressure are substituted for actual outcomes like death and stroke; how the odds that a commercially funded study will support its sponsor’s product are 5.3 times greater than the odds for a non-commercial study; how Richard Smith, former editor of the BMJ, cataloged the standard techniques for producing whichever result a sponsor requires.¹⁶

John Ioannidis’s 2005 paper “Why Most Published Research Findings Are False” is the most-downloaded paper in the modern medical literature.¹⁸ His analysis of highly cited findings in clinical medical research found that 31% had been contradicted or shown to overstate their effects by later studies. The population of research from which regulators, physicians, and journalists draw their consensus is a population known, from inside the discipline, to be substantially wrong.

The Phillips paper is not exceptional. It is what a passive reporting analysis of an aluminum-adjuvanted vaccine looks like when the regulator responsible for the product’s licensure writes it, the industry’s own journal publishes it, and it is read as evidence of safety. The moves are visible in it because the moves are the same moves executed across statins, blood pressure medication, cholesterol drugs, and cancer screening. Recognizing them in one place is recognizing them in all.

How to Explain It to a Six-Year-Old

Imagine a big playground with hundreds of kids on it. There’s a new slide, and a grown-up wants to know whether the slide is safe.

She could stand next to the slide and watch what happens. That would take a long time. Instead, she puts a little wooden box near the school. She writes on it: “If a kid gets hurt on the slide, come put a note in this box.”

Some kids get hurt. Most kids don’t know about the box. Some kids know but are too shy to walk over. Some get told to sit back down. Some just want to keep playing. Most of the hurt kids never put a note in.

At the end of the year, the grown-up opens the box and counts three notes. She goes to a big meeting and tells everyone: “Only three kids got hurt on the slide this year. The slide is safe.”

But it wasn’t three kids. It was forty. Thirty-seven notes never made it into the box. The grown-up counted three because three was what her box could catch.

Here is the part where it becomes a trick.

For two weeks in the middle of the year, a different grown-up stood next to the slide and watched. She wrote down every hurt kid she saw. Her number was much bigger. She wrote it in a report.

The first grown-up read the report. She kept using her small box number anyway.

The next year, the box was still there. It was still catching almost none of the hurts. But the grown-ups all agreed the slide was safe, because that is what the box said.

Then when your friend fell off the slide and her mom asked, “Is this slide safe?”, the grown-ups pointed at their meeting and said, “Yes. We studied it for eleven years. Only a few kids ever get hurt. It is very safe.”

Recognition

The value of seeing the formula is that it becomes portable. The moves appear in every major HPV vaccine safety publication cited as evidence of a settled safety profile: Chao 2012, Verstraeten 2008, Arnheim-Dahlström 2013, Arana 2018, Frisch 2018, Liu 2018.⁶ ⁷ ¹⁴ Products, populations, and journals vary. The sequence does not.

The document exists. It says what it says. When the daughter of a family somewhere in Australia develops orthostatic dysfunction, autonomic collapse, chronic pain, or ovarian failure after her three-dose Gardasil series, and her family asks whether the injection may have been the cause, the Phillips paper will be produced as the answer. The paper cannot answer that question. It was not designed to.

Every safety study built the same way now reads the same way.

Truth Be Told: I’ve Accepted an Invitation to Speak on The Unvaccinated

On September 17th, I’ll be giving a one-hour presentation titled The Unvaccinated as part of a six-hour livestream called Truth Be Told. This is the first time I have accepted an invitation to an event, and I have been honoured with the opening act. The livestream begins at 12pm EST.

Vaccination is the subject closest to my heart, and this is another opportunity to spread the word. The format will preserve the pen name.

Jamie Andrews (Decentralized Science Projects) and Agent131711 (Dinosaurs) will also be presenting. Jamie’s Virology Control Studies work led to an interview here last year. Agent’s research shaped my essays on vitamin D and dinosaurs. Tickets are here. The code UNBEKOMING is $5 off and applies automatically at that link. Replay available afterwards. Hope you can make it.

---

References

1. Phillips A, Hickie M, Totterdell J, Brotherton J, Dey A, Hill R, Snelling T, Macartney K. Adverse events following HPV vaccination: 11 years of surveillance in Australia. Vaccine 2020;38:6038–6046. https://doi.org/10.1016/j.vaccine.2020.06.039

2. Lazarus R, Klompas M et al. Electronic Support for Public Health–Vaccine Adverse Event Reporting System (ESP:VAERS). Grant Final Report, Harvard Pilgrim Health Care, 2010. Inclusive dates 12/01/07–09/30/10. Available at https://digital.ahrq.gov/sites/default/files/docs/publication/r18hs017045-lazarus-final-report-2011.pdf

3. Holland M, Rosenberg KM, Iorio E. The HPV Vaccine on Trial: Seeking Justice for a Generation Betrayed. Skyhorse Publishing, 2018. Chapters on Denmark, Japan, Ireland, and Colombia; Brinth, Mehlsen, and Chandler documentation.

4. Ozawa K et al. Suspected Adverse Effects After Human Papillomavirus Vaccination: A Temporal Relationship Between Vaccine Administration and the Appearance of Symptoms in Japan. Drug Safety 2017;40(12):1219–1229.

5. Brinth LS, Pors K, Theibel AC, Mehlsen J. Orthostatic intolerance and postural tachycardia syndrome as suspected adverse effects of vaccination against human papilloma virus. Vaccine 2015;33(22):2602–2605.

6. Chao C et al. Surveillance of autoimmune conditions following routine use of quadrivalent human papillomavirus vaccine. Journal of Internal Medicine 2012;271(2):193–203. Liu EY et al. Quadrivalent human papillomavirus vaccination in girls and the risk of autoimmune disorders: the Ontario Grade 8 HPV Vaccine Cohort Study. CMAJ 2018;190(21):E648–E655.

7. Verstraeten T et al. Analysis of adverse events of potential autoimmune aetiology in a large integrated safety database of AS04 adjuvanted vaccines. Vaccine 2008;26:6630–6638.

8. Little DT, Ward HRG. Adolescent Premature Ovarian Insufficiency Following Human Papillomavirus Vaccination: A Case Series Seen in General Practice. Journal of Investigative Medicine High Impact Case Reports 2014;2(4):2324709614556129.

9. Therapeutic Goods Administration. Gardasil (quadrivalent human papillomavirus vaccine) update 2. Australian Government Department of Health, 2015.

10. Fraser H. The Peanut Allergy Epidemic: What’s Causing It and How to Stop It. Skyhorse Publishing, 2011. Chapter 4 on Richet’s 1901 experiments. Richet C, Anaphylaxis, Nobel Lecture, December 11, 1913.

11. Holland M et al. The HPV Vaccine on Trial, chapters on clinical trial design, Protocol 018, and AAHS as fauxcebo. See also Kennedy RF Jr., Vax-Unvax, Skyhorse Publishing, 2023.

12. Gherardi RK, Authier FJ. Macrophagic myofasciitis: characterization and pathophysiology. Lupus 2012;21:184–189. Khan Z et al. Slow CCL2-dependent translocation of biopersistent particles from muscle to brain. BMC Medicine 2013;11:99. Shaw CA et al. Aluminum-Induced Entropy in Biological Systems: Implications for Neurological Disease. Journal of Toxicology 2014;491316. On Exley’s institutional reception, see Maready F., Crooked: Man-Made Disease Explained, 2018.

13. Mold M, Umar D, King A, Exley C. Aluminium in brain tissue in autism. Journal of Trace Elements in Medicine and Biology 2018;46:76–82.

14. Arana JE et al. Post-licensure safety monitoring of quadrivalent human papillomavirus vaccine in the Vaccine Adverse Event Reporting System (VAERS), 2009–2015. Vaccine 2018;36:1781–1788. Arnheim-Dahlström L et al. Autoimmune, neurological, and venous thromboembolic adverse events after immunisation of adolescent girls with quadrivalent human papillomavirus vaccine in Denmark and Sweden: cohort study. BMJ 2013;347:f5906. Frisch M et al. Quadrivalent human papillomavirus vaccination in boys and risk of autoimmune diseases, neurological diseases and venous thromboembolism. International Journal of Epidemiology 2018;47:634–641.

15. Huff D. How to Lie With Statistics. W. W. Norton & Company, 1954.

16. Kendrick M. Doctoring Data: How to Sort Out Medical Advice from Medical Nonsense. Columbus Publishing, 2014. The Swedish cervical cancer screening study Kendrick analyzes is described in the chapter on relative vs absolute risk.

17. Bill Gates has recommended Huff’s How to Lie With Statistics on multiple public reading lists, including a widely reported endorsement to attendees of the TED conference.

18. Ioannidis JPA. Why Most Published Research Findings Are False. PLoS Medicine 2005;2(8):e124. Contradicted and Initially Stronger Effects in Highly Cited Clinical Research. JAMA 2005;294(2):218–228.

19. Miller NZ. Miller’s Review of Critical Vaccine Studies. New Atlantean Press, 2016.