What the NHIS Autism Preprint Can and Can't Show

A month ago I wrote about Austin Bradford Hill and the nine viewpoints he proposed for judging whether an association is causal (Association or Causation).¹ He meant for them to be questions you asked before reaching a verdict. He also said none of them "can bring indisputable evidence for or against the cause-and-effect hypothesis and none can be required as a *sine qua non*."²

I took two causal claims and pitted them against Hill in that post. One was Paul Thacker's lab-leak piece in RealClearInvestigations which fell short on six of the nine viewpoints.³ It partly satisfied temporality and offered a defensible analogy, while plausibility was a no-go. Andrew Wakefield's 1998 *Lancet* paper had a dozen children but no comparison group, and its only apparent temporality came from the vaccination schedule, since MMR is given around the age autism is usually recognized in kids.⁴⁻⁵

Six authors, including Nicolas Hulscher and Peter McCullough of the McCullough Foundation, borrowed Hill's framework for their preprint and posted their results to Zenodo on October 5, 2026. They analyzed COVID-19 vaccination and parent-reported neurodevelopmental outcomes in National Health Interview Survey (NHIS) data. The preprint has not been peer-reviewed.⁶ They think that the findings "satisfy the strength, consistency, biological gradient, plausibility, and coherence considerations for causal inference," which covers five of Hill's nine viewpoints.

Hulscher's announcement posts had with them numbers such as "AUTISM: +32%" and drew hundreds of thousands of views on social media in the first two days, which, incidentally, was way more attention than the limitations section received.

The six authors pooled the 2022–2024 NHIS Sample Child files. NHIS is a nationally representative household survey run by the National Center for Health Statistics. It selects one child per household, and a parent or guardian answers the questions.

The primary analytic sample was 21,990 children aged 0–17 with a known COVID-19 vaccination status. Because the autism, ADHD, and learning-disability questions are asked only for ages 2–17, those analyses include 19,882 children, with 844 lifetime and 779 current autism cases. Exposure came from a parent's answer to whether the child had ever had at least one dose of a COVID-19 vaccine.

Survey-weighted logistic regressions were fit across five adjustment tiers. The most complete model adjusted for demographics, socioeconomic position, insurance, region, a past-year wellness visit, and a usual source of care. In that, vaccinated children had higher odds of:

• current autism spectrum disorder (ASD): aOR 1.32 (95% CI 1.08–1.60)
• lifetime ASD: 1.26 (1.04–1.52)
• ADHD: 1.33 (1.18–1.49)
• special-education or early-intervention plan: 1.19 (1.08–1.31)
• daily or weekly anxiety: 1.27 (1.15–1.42)
• mental-health medication: 1.57 (1.38–1.77)
• mental-health therapy: 1.60 (1.43–1.79)

They also report a dose pattern (current ASD 1.03, 1.25, and 1.40 for one, two, and three or more doses vs. none), larger estimates in some younger subgroups, influenza and HPV vaccination as negative-control exposures, and a propensity-score analysis.

According to its little conclusion section, the crude autism association was zip, with autism prevalence of 4.19% in vaccinated children and 4.28% in unvaccinated children (crude OR 0.98, 0.83–1.16). An association only popped up after adjustmenting for socioeconomic position. Vaccinated children more often came from higher-income households with more parental education and private insurance, and those households reported *less* autism. An adjusted estimate built like this may still be correct, because its accuracy depends on which covariates the model includes and omits. Anyone passing along the "+32%" figure might ask what it looked like before they did the adjustment.

Hill listed strength first, with examples such as scrotal cancer in chimney sweeps (about 200 times the rate in unexposed workers) and lung cancer in cigarette smokers ("nine to 10 times the rate in non-smokers"), and the cholera death rate in John Snow's data was 14 times higher among customers of the polluted water supply.² He set those against the smoking–coronary thrombosis association, "no more than twice," and noted that with weaker associations "it is surely much easier... to think of some features of life that may go hand-in-hand" with the exposure. Autism estimates in this preprint are about 1.2 to 1.3 after adjustment, with a crude estimate of 1.0, and its discussion section rates strength as only "modestly met." Hill also cautioned against dismissing a cause-and-effect hypothesis because an association looks slight. But an odds ratio this wishy-washy leaves bias and confounding a lot of room to produce the result.

Hulscher and colleagues use consistency to describe an estimate that holds up across their own adjustment tiers and weighting schemes. Hill asked whether an association "has been repeatedly observed by different persons, in different places, circumstances and times," and he gave an example of why repetition within one design adds little. If a bias is built into how the data is collected, "no number of exact repetitions would remove or necessarily reveal that fallacy."² All five tiers analyze one parent-report survey and share its data-generating process. They show robustness to model choice within NHIS, and replication in independent data is still to be done. Rerunning one dataset five ways is a basically like rewatching the same *Severance* episode half a dozen times and calling it a season.

The dose pattern is offered as evidence of a biological gradient, and Hill did find a clear dose-response persuasive. In the crude data, weighted autism prevalence was flat across dose groups (4.28%, 4.27%, 4.22%, and 4.18% for zero, one, two, and three or more doses), and the authors explain that the gradient "emerges after adjustment because dose count rises with socioeconomic position." Dose count also tracks how often a family sees a clinician, and more healthcare contact would produce the same pattern as a biologic dose effect.

For plausibility and coherence, the authors draw on spike-protein persistence and blood–brain-barrier studies in cell culture, along with rat models of prenatal exposure. Hill wrote that plausibility is "a feature I am convinced we cannot demand," because "what is biologically plausible depends upon the biological knowledge of the day."² Coherence asks only that a causal reading "not seriously conflict" with what is known about the disease. Neither viewpoint can support an argument alone, and details in the preprint undercut both. One of the cited rat studies has been retracted, as the preprint notes, and human cohorts followed after in utero mRNA exposure found no neurodevelopmental differences. Their discussion calls coherence "partially met" and reports that the association appeared only in girls, which is the opposite of the male-predominant rat model. They write that this sex pattern "is more readily explained by detection than by biology," and elsewhere that "Plausibility motivates a test; it does not substitute for one."

Specificity gets the most candid treatment in the preprint. In the four-group model, compared with children who got neither vaccine, the lifetime ASD estimates were:

• influenza vaccine only: 1.11 (0.87–1.41)
• COVID-19 vaccine only: 1.12 (0.84–1.48)
• both: 1.42 (1.14–1.78)

A separate 2022 model of children aged 9–17 (n = 3,606) gave an HPV vaccination estimate of 1.17 (0.74–1.85). Influenza vaccination was as strongly or more strongly associated with ADHD, intellectual disability, developmental delay, asthma, ED visits, and hospitalization. A flu shot contains no spike protein, mRNA, or lipid nanoparticle. When unrelated vaccines show autism associations of about the same size, the likelier common factor is the family's contact with healthcare.

Hill phrased temporality as a question: "which is the cart and which the horse?"² He didn’t make any of his viewpoints mandatory. A cause has to come before its effect, though, so most epidemiologists come close to requiring temporality. This preprint does not come close to meeting it. NHIS "records lifetime diagnoses without dates," the authors write, and the design "cannot establish that vaccination preceded diagnosis." Most vaccinated children with autism received their first dose well after the usual age of diagnosis (median age at that dose 11; 84% were 6 or older). Adolescents in the sample were first vaccinated at 12 or older, and for those cohorts the preprint says the data establish *the reverse* order, so those estimates "cannot be causal." A reverse model in the same paper found that children with an existing autism diagnosis had higher adjusted odds of being vaccinated (1.26).

An experiment isn't possible with this design, and the paper concedes that point, and analogy isn't addressed. In the discussion, Hulscher and colleagues write:

"Specificity and temporality are not clearly satisfied by the present study, and the experimental criterion cannot be met by an observational design."

Their abstract (my favorite part) mentions the timing limitation and the similar influenza estimates before listing five considerations as satisfied, with specificity and temporality left off. The final section states, "These properties indicate a real effect." Reading that a few pages after the quote above gave me mild whiplash.

The table below sets each of the paper's claims beside my own read of the evidence. What do you think?

My training in health information management (HIM) and medical coding likes to look at the gap between what happened to a patient and what the record states. A coded record what was documented and what the coder could code based on whatever classification system is needed. Coders get the line "If it's not documented, it didn't happen" beaten into them as a warning about incomplete charts. What’s in there also depends on who came in for care and on who recorded the visit.

A household survey adds another step that gets between the patient and the dataset. The NHIS autism question in the 2023 Sample Child codebook reads:

"Has a doctor or other health professional ever told you that [child] had Autism, Asperger's disorder, pervasive developmental disorder, or autism spectrum disorder?"⁷

The interviewer may add that "health professionals can include school psychologists and school nurses." The special-education item asks whether the child has "ever had a special education or early intervention plan," such as an IEP or IFSP, and the exposure item asks whether the child "had at least one dose of a COVID-19 vaccination."

One parent answers all three, which are then not checked against an outside source such as a medical chart or an immunization registry. Each answer turns a messy clinical and social history (whether the child was evaluated, by whom, at what threshold, whether the family could get an appointment, and what the parent remembers) into a yes-or-no value for a logistic model.

Two mundane explanations fit these data. Shared method variance could produce the reported pattern, because errors in the exposure and in the outcome are correlated when one person reports both. A parent who keeps careful track of a child's health history, perhaps because the child has been through evaluations and IEP meetings, is more likely to recall a vaccine dose accurately and to report the child's diagnoses and services.

Healthcare-engagement confounding could produce it as well. Families already connected to diagnostic and special-education services see clinicians far more often, which gives them more opportunity to be offered a vaccine and to receive a diagnosis. Adjustment for "wellness visit in the past 12 months" and "usual source of care" captures only a small part of this contact history, and the authors' own discussion calls past-year utilization a "weak proxy" for "the years of clinical engagement that produce a developmental diagnosis."

The negative controls help separate the two. A specific biologic effect of COVID vaccination on neurodevelopment would leave the flu shot unassociated with autism. Under confounding by healthcare contact, any vaccine that marks an engaged family should show a similar modest association, and children with two vaccines should show a larger one. NHIS matches that prediction. The same mechanism predicts the largest excess in groups where autism tends to be underdiagnosed until someone looks harder, such as girls, and the girls-only association reported earlier fits too. The influenza pattern, the authors acknowledge, "may be a signature of confounding by healthcare engagement."

To rescue temporality, Hulscher and colleagues look to the youngest children, in whom vaccination could plausibly have preceded diagnosis. In children aged 5–7 with three or more doses, the aOR was 2.54 (1.33–4.84). That estimate produced the "154% higher odds" figure that circulated, and it comes from 18 exposed cases in a model with fewer covariates than the main analysis.

The cleanest test in the preprint uses a cohort built so that the first dose came at age 5 or younger and at least one year before the interview. In that group, any vaccination versus none gave 1.54 (0.91–2.63; 29 exposed cases), with a confidence interval that includes 1. Narrowing to three or more doses gave 2.46 (1.08–5.62), which rests on 9 exposed cases, and requiring a two-year lag instead of one gave a null 1.02 (0.43–2.40), also from 9 cases.

Changing one lag definition moves the estimate from 2.46 to 1.02. With numbers that unstable and case counts that small, temporality remains unestablished. According to the limitations section, the subgroup analyses were exploratory and uncorrected for multiple comparisons, with secondary analyses "developed sequentially rather than registered in advance."

Against confounding, the strongest defense offered is a propensity-score analysis. Overlap weighting balanced the measured covariates, and the autism association held (lifetime 1.23, 1.01–1.50; current 1.26, 1.03–1.56). The negative-control outcomes moved to the null, with ED visits at 0.95 and hospitalization at 1.02, and the preprint reads that contrast as evidence that confounding was removed. That reading has two problems.

• Weighting balances only the covariates in the propensity model, which left out diagnostic intensity, parental health attitudes, and years of developmental services.
• The ED and hospitalization estimates were already at or near the null under regression adjustment (ED 1.02; hospitalization 1.16, 0.93–1.44, not significant). As acute events outside the reporting pathway that produces a lifetime developmental diagnosis or an IEP, they make a poor test of the bias at issue.

Calling propensity-score methods "the most direct test," as the discussion does, overstates what a method limited to measured covariates can show.

Wakefield's 1998 paper studied twelve children; *The Lancet* retracted it,⁵ and he was struck off the UK medical register.⁸ This preprint is far more transparent and documents its own weaknesses, yet both lean on a temporal coincidence near the age autism is usually recognized and claim more than their failed viewpoints allow.

A 2019 Danish registry cohort of 657,461 children found a fully adjusted autism hazard ratio of 0.93 (0.85–1.02) for MMR, with no increase in susceptible subgroups or after vaccination.⁹ A 2014 meta-analysis of five cohort studies (1,256,407 children) and five case-control studies (9,920 children) found no vaccine–autism association (cohort OR 0.99, 0.92–1.06).¹⁰ Neither covered COVID-19 vaccines, and the preprint's own negative controls point the same direction.

Hill ended his list of viewpoints with the question they are meant to serve:

"[I]s there any other way of explaining the set of facts before us, is there any other answer equally, or more, likely than cause and effect?"²

For this preprint, the other answer comes from its own data. A survey that cannot order exposure and diagnosis, combined with influenza and HPV associations of about the same size as the COVID-19 one, points to healthcare engagement and shared parent reporting. A crude association that turns positive only after modeling fits that reading at least as well as a vaccine effect. The authors raise healthcare-engagement confounding themselves, in their discussion of the negative controls. The causal conclusion can go out the nearest airlock.

If you missed September's post on Hill, start there.

References

1. Crocker AR. Association or causation. Notes from the Abstracts. September 11, 2026. Accessed October 11, 2026. https://notesfromtheabstracts.com/posts/association-or-causation
2. Hill AB. The environment and disease: association or causation? *Proc R Soc Med*. 1965;58(5):295-300. Reprinted in *J R Soc Med*. 2015;108(1):32-37. doi:10.1177/0141076814562718
3. Thacker PD. COVID cover-up: hiding star researcher Ralph Baric's ties to global pandemic. RealClearInvestigations. April 28, 2026. Accessed October 11, 2026. https://www.realclearinvestigations.com/articles/2026/04/28/covid\_cover-up\_campaign\_to\_hide\_star\_researcher\_ralph\_barics\_ties\_to\_global\_pandemic\_1179562.html
4. Wakefield AJ, Murch SH, Anthony A, et al. Ileal-lymphoid-nodular hyperplasia, non-specific colitis, and pervasive developmental disorder in children. *Lancet*. 1998;351(9103):637-641. Retracted in: *Lancet*. 2010;375(9713):445.
5. The Editors of The Lancet. Retraction—Ileal-lymphoid-nodular hyperplasia, non-specific colitis, and pervasive developmental disorder in children. *Lancet*. 2010;375(9713):445. doi:10.1016/S0140-6736(10)60175-4
6. Hulscher N, Mead MN, Taccetta C, Catanzaro JA, Honea RA, McCullough PA. Neurodevelopmental outcomes in COVID-19-vaccinated versus unvaccinated U.S. children: analysis of the National Health Interview Survey, 2022–2024. Preprint. Zenodo. Posted October 5, 2026. Accessed October 11, 2026. doi:10.5281/zenodo.23165256
7. National Center for Health Statistics. 2023 National Health Interview Survey (NHIS): Sample Child codebook. Centers for Disease Control and Prevention; 2024. Accessed October 11, 2026. https://ftp.cdc.gov/pub/Health\_Statistics/NCHS/Dataset\_Documentation/NHIS/2023/child-codebook.pdf
8. Kmietowicz Z. Wakefield is struck off for the "serious and wide-ranging findings against him." *BMJ*. 2010;340:c2803. doi:10.1136/bmj.c2803
9. Hviid A, Hansen JV, Frisch M, Melbye M. Measles, mumps, rubella vaccination and autism: a nationwide cohort study. *Ann Intern Med*. 2019;170(8):513-520. doi:10.7326/M18-2101
10. Taylor LE, Swerdfeger AL, Eslick GD. Vaccines are not associated with autism: an evidence-based meta-analysis of case-control and cohort studies. *Vaccine*. 2014;32(29):3623-3629. doi:10.1016/j.vaccine.2014.04.085