Error Asymmetry: How FDA Decides Which Mistakes Matter More

Most statistical frameworks treat errors symmetrically. A false positive is bad. A false negative is bad. Control one, tolerate the other, and let the math do the rest.

Clinical reality is not that tidy.

Approving an ineffective therapy and withholding a potentially effective one are both errors, but they do not carry the same weight across diseases or regulatory contexts. This imbalance quietly drives FDA decisions. Once you see it, regulatory behavior that appears inconsistent becomes interpretable.

Aduhelm: when the asymmetry goes wrong

On June 7, 2021, FDA approved aducanumab for Alzheimer's disease under accelerated approval. Within days, three advisory committee members resigned in protest. One called it "probably the worst drug approval decision in recent U.S. history." Congress launched investigations. Biogen discontinued the drug in January 2024.

The evidence was thin. Biogen conducted two identical Phase III trials, EMERGE and ENGAGE, in early Alzheimer's disease. Both were halted for futility in March 2019. Months later, Biogen reported that EMERGE showed benefit after including additional data: a 0.39-point improvement on the Clinical Dementia Rating-Sum of Boxes, where clinically meaningful change is generally considered 1 to 2 points. ENGAGE remained negative. In November 2020, the advisory committee reviewed the conflicting results and voted 10 to 0 against approval.

FDA approved anyway. Not on demonstrated clinical benefit, but on amyloid plaque reduction, deemed "reasonably likely to predict benefit." This was an explicit error-asymmetry judgment: FDA concluded that withholding the first potential disease-modifying Alzheimer's therapy in 18 years posed greater harm than approving a drug that might not work.

The consequences were immediate and cascading. Clinically, no meaningful benefit emerged. ENGAGE was negative. EMERGE's effect was below clinical relevance. Forty-one percent of treated patients experienced brain swelling or microhemorrhages. Economically, Aduhelm launched at $56,000 per year. Medicare restricted coverage to clinical trial participants, an almost unprecedented move, and Part B premiums increased 14% in 2022, partly in anticipation of Aduhelm costs. Institutionally, Congressional investigations revealed unusually close FDA-Biogen collaboration during the review. The decision damaged FDA credibility and intensified scrutiny of the entire accelerated approval pathway.

Biogen ultimately discontinued Aduhelm and halted the confirmatory trial. The drug approved over unanimous advisory committee opposition disappeared without ever confirming benefit.

FDA's judgment, that a false negative was worse than a false positive, proved wrong. The false positive caused real harm. The hoped-for benefit never materialized.

When the asymmetry works: oncology

Error asymmetry does not always fail. Oncology shows when it works, and the contrast with Aduhelm reveals what actually determines whether the gamble pays off.

Consider tarlatamab for extensive-stage small cell lung cancer. FDA granted accelerated approval in 2024 based on response rate. In November 2025, confirmatory trials showed a survival benefit: 13.6 months with tarlatamab versus 8.3 months with chemotherapy. Patients gained access roughly a year before survival data matured. The surrogate predicted correctly. The error-asymmetry bet paid off.

Three differences separated this outcome from Aduhelm. First, surrogate strength. Tumor response has predicted survival across hundreds of oncology trials. Amyloid reduction had repeatedly failed to predict cognitive benefit, a record that should have given FDA pause. Second, effect size. Tarlatamab showed a response rate around 40%. Aduhelm showed a 0.39-point change on a scale where 1 to 2 points constitutes clinical relevance. Third, confirmatory feasibility. Lung cancer survival trials complete in 12 to 18 months. Alzheimer's trials take years, and in Aduhelm's case, the confirmatory trial was never completed.

When surrogates are validated, effects are substantial, and confirmation is prompt, error asymmetry enables earlier access without catastrophic false-positive risk. When those conditions are absent, it becomes a rationalization for lowering standards.

What determines FDA's tolerance

Several factors shape how much uncertainty FDA will accept, and they interact in ways that resist simple rules.

Disease severity and trajectory matter most. Fatal diseases with rapid progression receive more tolerance for uncertainty than chronic conditions where patients have time to wait for better evidence. Availability of alternatives amplifies this: zero options increases flexibility dramatically, which is why oncology and rare diseases have always been the natural home for accelerated approval.

Surrogate endpoint validation matters more than sponsors often acknowledge. Well-established surrogates like tumor response support traditional approval. Novel surrogates, like amyloid plaque reduction, require extraordinary confidence in the mechanistic link. Aduhelm is a case study in what happens when that confidence is misplaced.

Confirmatory feasibility matters because error asymmetry is only defensible if the uncertainty is temporary. An accelerated approval that can be verified within two years carries different risk than one where confirmation might take a decade. FDA now expects confirmatory trials to be underway at the time of approval, not merely planned, and the FDORA provisions give the agency explicit authority to initiate withdrawal if sponsors fail to conduct them with due diligence.

Aduhelm scored high on severity and unmet need but low on surrogate validation and confirmatory feasibility. That combination should have been a warning. The error-asymmetry argument was deployed where its preconditions were weakest.

The decision problem underneath

Bayesian methods can express the uncertainty precisely. A posterior probability answers: given the data, how plausible is benefit? But it does not answer the harder question: what decision rule should we adopt, given the consequences of being wrong?

For Aduhelm, a rough Bayesian reading of the conflicting trial results might put the posterior probability of meaningful clinical benefit somewhere around 30 to 50%, depending on how you weighted EMERGE against ENGAGE and what prior you started with. The precise number matters less than the framing. The question was not whether that estimate was correct. It was whether that level of confidence justified exposing millions of Alzheimer's patients to a drug with a 41% rate of brain swelling.

That is a decision problem, not an inference problem. The statistical framework can quantify what we know. It cannot tell us what to do with what we know. The values enter at the decision layer, and they are inescapable.

I've written about this tension in the context of Bayesian calibration and in the surrogate trap. The common thread is that the hardest part of trial design is not the math. It is the moral arithmetic underneath: which errors are tolerable, who pays for them, and whether the conditions that justify accepting more uncertainty are actually present or merely asserted.

What this means for your designs

If error asymmetry may apply to your indication, make the case explicitly. Quantify disease severity, progression rate, and the absence of alternatives. Model both error types: what happens if approval is wrong, and what happens if rejection is wrong. Do not assert the asymmetry. Demonstrate it.

Engage FDA early. RDEP, CID, and pre-IND meetings exist to align on how much uncertainty the agency will tolerate in a specific context. That conversation is harder to have after the trial reads out.

Design confirmatory trials concurrently. Accelerated approval now assumes verification is already underway. If your error-asymmetry argument depends on the uncertainty being temporary, the confirmatory trial is what makes that argument credible.

And look honestly at whether your case resembles tarlatamab or Aduhelm. Validated surrogate, large effect, feasible confirmation: the asymmetry argument holds. Novel surrogate, marginal effect, uncertain confirmation timeline: the asymmetry argument becomes a liability.

FDA decisions make sense once you stop asking whether a trial "worked" and start asking which mistake would have caused more harm. The answer is not always the one the sponsor wants to hear.

For analysis of how error asymmetry operates in rare disease approvals, see my recent post on single-arm trial design. For the specific case of surrogate failure, see The Surrogate Trap.

📬 For more essays on experimental design, regulatory evidence, and statistical decision-making across domains, subscribe to the Evidence in the Wild newsletter or browse the archive.