Reply: Data Interpretation and Temporal Risk of Semaglutide-Associated NAION

By
Kent R. Heberer
Adam P. Bress
Jennifer S. Lee
October 1, 2026

Jama Ophthalmology

Overview

The authors address concerns about risk-table accounting, time-varying effects, and NAION outcome misclassification in their semaglutide analysis. They report that the risk-table counts reconcile, their analyses did not detect a meaningful violation of proportional hazards, and the NAION coding algorithm has an estimated positive predictive value of approximately 74.5%.

Background

Nonarteritic anterior ischemic optic neuropathy (NAION) is the outcome examined in a Brief Report comparing patients initiating semaglutide with those initiating sodium-glucose cotransporter-2 inhibitors (SGLT2i). A subsequent letter raised questions about the Figure 2 at-risk table and whether a single hazard ratio adequately represented risk over time. The reply also addresses the accuracy of administrative coding algorithms used to identify NAION in electronic health record data. The authors frame their findings as a rare but evident risk following semaglutide initiation, while noting the low absolute risk.

Data Highlights

Analysis or measure	Reported result
SGLT2i cohort accounting at day 182.5	87,135 at risk + 35 cumulative NAION events + 3,713 cumulative censored = 90,883 patients at day 0
Cause-specific treatment HR, year 1	1.78 (95% CI, 0.90–3.51)
Cause-specific treatment HR, after year 1	2.93 (95% CI, 1.75–4.91)
Between-period HR contrast	HR difference, 1.15; P = .25
Scaled Schoenfeld residual assessment	χ²₁ = 0.084; P = .77
ICD-10 algorithm positive predictive value for NAION	Approximately 74.5%

Key Findings

The at-risk count represents the number immediately before each time point, whereas cumulative event and censoring totals include occurrences through that time point, including day 0.
In the SGLT2i group, 510 patients were censored on day 0; the day 182.5 accounting reconciled to the full cohort of 90,883 patients.
Censoring included deaths, prespecified competing events, and end of follow-up. The authors reported no indication that censoring was informative regarding treatment and did not anticipate resulting hazard-ratio bias.
In the overlap-weighted cohort, cause-specific treatment HRs were 1.78 (95% CI, 0.90–3.51) during year 1 and 2.93 (95% CI, 1.75–4.91) thereafter; the between-period contrast was not statistically significant (HR difference, 1.15; P = .25).
Scaled Schoenfeld residuals in the unweighted cohort showed no detected temporal trend (χ²₁ = 0.084; P = .77); the authors interpreted the reported HR as an effect summary averaged across follow-up.
The cited ICD-10 algorithm has an approximately 74.5% positive predictive value for NAION. The authors reported no evidence of differential misclassification by treatment group and stated that nondifferential misclassification would tend to bias the HR toward the null.

Clinical Implications

The authors state that, despite the low absolute risk, consideration should be given to counseling patients about the rare but evident risk of NAION following semaglutide initiation. Their reply reports no detected appreciable violation of proportional hazards and no evidence that NAION misclassification varied by treatment group.

Conclusion

The reply reconciles the risk-table counts and presents two analyses that did not detect a meaningful temporal departure from proportional hazards. The authors maintain that the findings support counseling about a rare NAION risk after semaglutide initiation.