Autism Environmental and Genetic Studies Must Test for Both Genetics and Environment - and for Effect Modification by Familial Liability

Adjusting for sibling effect can bury the signal in the interaction term, which, when unstudied, will not be found by parents, the HELP committee, or Senator Cassidy.

Main-Effects Only Models Bury the Point — Especially in Autism and Vaccine Research. Null conclusions are being manufactured by design.

Dear Reader: I have been railing against this specific means of burying environmental signals in autism studies since 2015. It’s now 2026. It is the most important problem second only to the outright fraud conducted by choosing co-predictors as covariates (collision bias). Please stick with me on this and read to the end. And please SHARE it with your Senators and Congressional Representatives. Secretary Kennedy has said that autism studies must study both genetic and environmental factors. They must also study if and how they interact.

For more than two decades, the autism research literature has repeated the same analytic maneuver with unwavering confidence: treat familial liability as a nuisance variable, adjust it away, and declare the remaining exposure effect authoritative. This pattern appears across studies of pharmaceuticals, air pollution, nutrition, infections, and chemical exposures. It has produced hundreds of pooled estimates and a steady stream of reassuring headlines.

This is not a minor technical preference. It is a categorical modeling error.

And it is intentional.

When investigators control for family history, genetic risk scores, or shared familial factors—but never test whether an exposure’s effect differs across levels of familial liability—they impose a biologically extreme assumption without naming it. They assume the exposure operates identically in families with minimal baseline neurodevelopmental risk and in families already carrying substantial inherited or developmental vulnerability.

That assumption is rarely justified, never innocuous, and often decisive.

It’s misleading, and it has to stop.

Calling results generated under that assumption “definitive” is not warranted.

Adjustment Is Not Interaction: A Distinction the Literature Keeps Violating

Most autism–environment studies estimate models of the following form:

Outcome = Exposure + Familial Liability | Presumed Covariates

where “|” means “given”, or in this case “after adjusting for.”

Translated plainly, this model structure asks: What is the average association between exposure and outcome, assuming the exposure effect is the same in all families?

What it does not ask is the question that actually matters:

Does the exposure have different effects in families with different baseline liability for autism in a way that is detectable by studying how genetics modifies environmental risk?

That question requires an explicit interaction term:

Outcome = Exposure + Familial Liability + (Exposure×Familial Liability) | KNOWN Covariates

Omitting the interaction term is not agnosticism. It is an assertion—made silently—that the interaction equals zero without looking at it. In other words, the literature routinely assumes effect homogeneity without testing it.

In a condition as heterogeneous and heritable as autism, this assumption borders on the absurd.

Familial Liability Is Not a Confounder to Be “Controlled Away”

Autism is not a binary disease waiting for a universal trigger. It is a spectrum emerging from layered developmental vulnerability. Familial liability captures far more than DNA sequence alone. In real datasets, it bundles:

• Polygenic risk and rare inherited variants

• Shared prenatal and early-life environments

• Epigenetic inheritance

• Maternal immune and metabolic physiology

• Family-level diagnostic practices and surveillance intensity

Treating this composite as a mere confounder misstates and mis-specifies misunderstands the causal structure of the problem. Familial liability is not simply upstream of exposure and outcome. It is the substrate on which environmental effects plausibly operate.

If an exposure increases autism risk only in families already carrying elevated liability, a main-effects-only model will average that signal into oblivion. A pooled null estimate in that setting does not mean “no effect.” It means effect modification was mathematically erased.

Omitting the interaction term is like running a drug trial, discovering it only works for people with a specific gene variant, and then averaging the results across everyone—treated and untreated, variant and non-variant—and declaring the drug is effective - or ineffective - for everyone.

It’s worth noting that studies might also “control” for environment as a covariate leaving only genetics behind. It works in reverse, depending on whichever narrative they feel is important to publish on at the moment.

---

The Sibling-Comparison Mirage

Sibling-comparison designs are frequently described as the gold standard of observational autism research. They are praised for “controlling shared genetics and environment” and are often treated as closing causal debates.

They do not.

Sibling comparisons control only for familial factors that are constant across siblings. They do not test whether the exposure effect differs across families with different baseline liability. They simply average the within-family exposure effect across the subset of families that happen to be exposure-discordant.

Critically, sibling designs are especially vulnerable to exposure misclassification. Over-the-counter drugs, self-reported symptoms, and retrospectively reconstructed prenatal exposures are measured with error. Within-family contrasts amplify that error, biasing estimates toward the null.

A null sibling-comparison result therefore answers a narrow question: Is there a detectable average within-family effect under noisy exposure measurement? It does not answer whether susceptible families experience harm.

---

Why Effect Modification via Interaction Must Be Tested (Also) on the Additive Scale

Most studies report interaction—if they report it at all—on multiplicative scales such as odds ratios. This is a mistake.

Public health relevance lives on the additive scale: absolute risk differences. A small relative effect in a high-liability group can translate into a large absolute burden, even if the population-average odds ratio hovers near one.

Adequate analyses must therefore report:

• Stratum-specific absolute risks

• Risk differences by familial liability level

• Additive interaction metrics (RERI, attributable proportion)

Without these, claims of “no meaningful risk” are ungrounded.

---

The Lancet Paper: Strong Methods, the Wrong Question, and Overconfident Conclusions

A recent Lancet systematic review and meta-analysis on acetaminophen use in pregnancy and autism/ADHD exemplifies the problem. It represents yet another example of what I called in 2015 “statistical shamwizardry” in my book, Cures vs. Profits. The authors restricted their analysis to sibling-comparison studies and pooled adjusted estimates across countries and cohorts, reporting odds ratios near unity.

[See: D’Antonio F, Khalil A, et al. Prenatal paracetamol (acetaminophen) exposure and neurodevelopmental outcomes in childhood: systematic review and meta‑analysis. Lancet Obstet Gynecol Womens Health. 2026;published online January 16, 2026].

From this, they concluded that the “best available human evidence does not support a causal link.”

Media commentators escalated further, citing a doctor calling the findings “as definitive as we’re going to get.”

That claim is 100% false.

Matches and gasoline do not cause fire unless they interact.

The paper does not test whether acetaminophen exposure has different effects in families with high versus low neurodevelopmental liability. It does not present liability-stratified estimates. It does not model exposure×familial-liability interactions. It does not report additive-scale heterogeneity. It does not even acknowledge that its conclusions apply only to averaged effects.

Instead, it perpetuates the central myth of autism epidemiology: that averaging across liability strata produces generalizable causal knowledge.

It does not. It produces comforting but misleading nulls.

---

Why This Error Persists

The refusal to test effect modification by familial liability persists for three reasons.

First, interaction reveals the problem they do not want to exist.

Second, results complicate narratives. They fracture clean headlines and resist binary messaging.

Third, interaction analyses expose weak exposure measurement and underpowered designs.

Fourth, acknowledging susceptibility-dependent effects destabilizes regulatory and public-health reassurance frameworks.

None of these are scientific justifications. They are incentives.

---

What a Scientifically Adequate Autism Study Must Do

Any study claiming to inform autism causation or prevention must, at minimum:

• Explicitly define familial liability using biologically grounded measures

• Pre-specify hypotheses of effect modification by liability (state the model structures to be tested and design the study accordingly)

• Estimate exposure effects within liability strata

• Test interaction on both multiplicative and additive scales

• Report absolute risks and risk differences

• Quantify exposure misclassification and its impact on estimates

Journals, funding agencies, and peer reviewers should require these elements as conditions of publication and support. Their legitimacy hangs in the balance on this one issue.

---

Let’s use a cancer study as an example of how absurd their limited model designs are:

When a cancer risk study includes a familial-risk term—whether defined by family history, pedigree score, or a genetic liability index—but excludes an interaction term, it commits a decisive modeling error. This design choice creates an illusion of thoroughness while guaranteeing that any subgroup effect will be erased by construction.

In its simplest form (binary exposure E, binary familial risk F), a logistic model without interaction looks like this:

logit(P(cancer)) = β₀ + β₁·E + β₂·F

This formulation forces the exposure effect to remain constant across familial-risk strata. It encodes this claim: the exposure odds ratio equals ( (_1) ) in both low-risk and high-risk families.

But if reality looks like this:

logit(P(cancer)) = β₀ + β₁·E + β₂·F + β₃·(E × F)

Then ( (β1) ) reflects the exposure effect in low-risk families (F=0), while ( (β1 + β3) ) reflects the effect in high-risk families (F=1).

Omitting ( β3 ) does not “stay agnostic.” It enforces the assumption that ( β3 = 0 ).

This single modeling decision can generate a null “adjusted” exposure effect even when the exposure is harmful only in the high-risk stratum. The reported ( β1 ) becomes a weighted average of:

• The true effect in low-risk families (possibly null)

• The true effect in high-risk families (possibly large)

…and the weights are determined by:

1. The prevalence of familial risk

2. The exposure prevalence within each risk stratum

3. The scale on which heterogeneity matters (odds ratios vs risk differences)

In cancer research, this is especially dangerous: familial risk reflects biological susceptibility, which is exactly where environmental triggers are most likely to act.

A main-effects-only model answers a narrow technical question: “What is the average conditional association holding F constant?” But readers and policymakers hear a much broader (and false) claim: “There is no effect in anyone.”

When β3 is found to be highly significant, follow-up studies would be done to tease apart other shared environment effects from genetic effects.

---

The Trap is Deeper: Familial Risk Is Not Purely Genetic

In real datasets, “familial risk” almost never isolates inherited sequence variation. It is a composite marker—an entangled bundle of:

1. Inherited variants (e.g., pathogenic alleles or polygenic risk scores)

2. Shared environments (e.g., dietary patterns, household exposures, water/air contamination, occupational clustering)

3. Surveillance intensity (e.g., proactive screening, earlier workups, family medical culture)

Thus, “adjusting for familial risk” does two incompatible things:

• If familial risk is a confounder, adjustment blocks backdoor paths and helps isolate the exposure effect.

• If familial risk is a proxy for shared exposure, adjustment can over-control, scrubbing away the very signal of interest.

In this context, unmodeled interaction terms don’t just miss gene × environment interplay—they can also miss environment × environment compounding that tracks with family clustering.

This logic applies with even greater force to vaccine exposure studies, particularly where genetic susceptibility may modulate response.

If vaccine effects vary depending on immune-metabolic-genetic profiles—many of which cluster within families—then any vaccine safety study that adjusts for family history without modeling interaction commits the same error:

• It enforces effect homogeneity.

• It averages out liability-dependent signals.

• It risks declaring a false null that directly misguides risk communication.

Until this modeling failure is addressed, the autism, cancer, and vaccine literatures will continue to produce comfortable illusions—null results that are the direct mathematical consequence of the questions they refuse to ask.

Conclusion: Stop Declaring Victory Over Questions You Refuse to Ask

Autism research does not suffer from a lack of data. It suffers from a refusal to interrogate its own assumptions.

Adjusting for familial liability while refusing to test whether it modifies exposure effects is not conservative science. It is analytic malpractice dressed up as rigor.

The question is not whether environmental exposures increase autism risk on average. The question is whether they increase risk in susceptible families.

Until studies are designed to answer that question directly, proclamations of definitiveness—whether in The Lancet or on cable news—are not conclusions. They are evasions.