Semaglutide and NAION: reading the reply

A reply letter published on 27 August 2026 in JAMA Ophthalmology defends the observational research linking semaglutide to nonarteritic anterior ischemic optic neuropathy (NAION). The authors, Heberer, Bress, and Lee^[1], are responding to a criticism letter by Yu CH that questioned whether the data behind the association adequately controls for confounding and detection bias.

To understand what the reply argues and why it matters, it helps to read it alongside the companion article covering Yu's original letter: Semaglutide and NAION: the data debate.

What the reply letter argues

Heberer, Bress, and Lee^[1] defend the active comparator design used in the original cohort work. Rather than comparing semaglutide users to non-users across the general population (which would import confounding by indication wholesale), the original studies compared semaglutide initiators to initiators of other medications prescribed for the same underlying condition, typically type-2 diabetes or obesity. The logic is that two patients starting different drugs for the same disease share a similar baseline disease burden, which reduces the chance that the higher NAION rate in semaglutide users reflects the disease rather than the drug.

The reply also points to consistency across independent datasets. When multiple studies using different data sources, different analytical approaches, and different geographic populations all find the same direction of association, the convergence is harder to dismiss as an artifact of one dataset's coding quirks or one team's analytic choices.

The core methodological disagreement

Yu's letter argued two things. First, that semaglutide patients tend to receive more intensive follow-up care than comparator patients, meaning any condition present in that group (including NAION) will be caught and coded at higher rates regardless of the drug's pharmacological effect. This is detection bias. Second, that NAION shares vascular risk factors with both diabetes and obesity^[2], so the patient population that gets semaglutide is already at above-average baseline risk for the condition.

Heberer, Bress, and Lee push back on both points. Their position is that the active comparator design substantially addresses the confounding problem because the comparator group shares the same clinical indication and the same prescriber attention. On detection bias, they argue that if the NAION signal were driven purely by more intensive ophthalmological follow-up, you would expect that pattern to affect multiple conditions in the same direction, not just NAION.

Both positions have merit. The disagreement is not about whether the studies were conducted carelessly; it is about what the inherent limits of retrospective claims data allow you to conclude.

The context: a regulatory label that already moved

The EMA's Pharmacovigilance Risk Assessment Committee reviewed the available evidence and concluded in June 2025 that NAION is a very rare side effect of semaglutide products^[3]. Very rare in EMA language means fewer than 1 in 10,000 patients. The product information for Ozempic, Rybelsus, and Wegovy was updated to list NAION in the adverse reaction table at that frequency. The Ozempic label was last revised in August 2026.

Neither Yu's letter nor the Heberer reply argues that the EMA acted wrongly. The label update reflects the precautionary standard regulators use when a plausible signal appears in multiple datasets: transparency with prescribers, not proof of causation. The methodological debate is about the strength of the underlying evidence, not about whether the label should revert.

Why the reply letter is worth reading separately

In academic correspondence, the reply from original authors is often where the clearest methodological statements appear. The original paper defends its design in the methods section; the reply is a direct response to the sharpest criticism and usually makes the key concessions (or refusals to concede) explicit.

In this case, Heberer, Bress, and Lee are not claiming that observational data settles the question. Both the critical letter and the reply agree that a prospective study with NAION as a pre-specified endpoint would provide better evidence than any retrospective cohort, however well designed. The Ziegler 2026 cohort found a hazard ratio of 2.33 (95% CI 1.54 to 3.54) for NAION in semaglutide initiators^[4], with an incidence rate much higher than general-population background rates, but in a population already enriched for diabetes and cardiovascular disease.

What this means for the broader research question

The exchange in JAMA Ophthalmology is part of the normal process by which a pharmacovigilance signal gets stress-tested in the literature. The signal was strong enough to prompt a regulatory label update. The methodology debate that follows is how researchers determine whether the signal reflects a true causal relationship or a detectable-but-confounded one.

For readers tracking semaglutide safety, the most honest summary is this: the association between semaglutide and NAION is on the label because regulators judged the signal plausible and the risk real enough to disclose. The original authors believe their methods support a real drug effect. A serious critic believes the methods leave important alternative explanations open. Both agree more prospective data would help.

What we do not yet know

No randomised trial has NAION as a pre-specified endpoint for semaglutide. No prospective ophthalmological monitoring cohort exists at scale for GLP-1 class drugs. The mechanism by which semaglutide could cause optic disc ischemia, whether through blood pressure changes, platelet effects, or microvascular pathways, has not been traced in controlled conditions. The absolute risk, even if the association is real, may be small enough that the drug's established cardiovascular and metabolic benefits outweigh it for most patients. That calculation requires better data than the field currently has.