# GLP-1 medicine reporting patterns in FDA FAERS, 2026 Q2

The complete April–June 2026 ASCII release contained **422,458 eligible unique cases** after retaining the latest available report per case. All four prospectively chosen medicine–reaction comparisons showed elevated reporting odds. These are associations within submitted reports; they do not establish that a medicine caused an event, estimate its clinical incidence, or show that one medicine is safer than another.

## Primary results

Exposure means a report mentions the named ingredient or a prespecified brand alias in any drug role. Reactions are the exact normalized MedDRA preferred-term strings **Pancreatitis** and **Cholelithiasis**. Each comparator is every other eligible report in the same release.

| Medicine | Exact reaction term | Reports with medicine and term (a) | Medicine without term (b) | Term without medicine (c) | Neither (d) | Reporting odds ratio (95% CI) | Holm-adjusted p |
|---|---|---:|---:|---:|---:|---|---:|
| Semaglutide | Pancreatitis | 160 | 18,141 | 847 | 403,310 | 4.20 (3.54–4.98) | 1.39 × 10⁻⁴⁴ |
| Semaglutide | Cholelithiasis | 130 | 18,171 | 628 | 403,529 | 4.60 (3.80–5.56) | 3.98 × 10⁻⁴⁰ |
| Liraglutide | Pancreatitis | 5 | 587 | 1,002 | 420,864 | 3.58 (1.48–8.65) | 0.0145 |
| Liraglutide | Cholelithiasis | 7 | 585 | 751 | 421,115 | 6.71 (3.17–14.19) | 0.000230 |

There were 18,301 semaglutide-mentioning reports and 592 liraglutide-mentioning reports; 93 mentioned both. The full background contained 1,007 reports with the exact term Pancreatitis and 758 with Cholelithiasis. Cases can contribute to both medicine comparisons and both reaction comparisons, but never more than once to a particular 2×2 table.

For example, the semaglutide–pancreatitis estimate compares 160/18,141 with 847/403,310: it is the ratio of the odds of the recorded term among medicine-mentioning versus other reports. It is **not** a 4.20-fold risk of pancreatitis among treated patients.

The intervals are approximate log-Wald 95% intervals. Tests are two-sided Fisher exact tests, with Holm adjustment across the four primary comparisons; intervals themselves are not multiplicity-adjusted. All four adjusted p-values are below 0.05, but statistical significance cannot remove reporting bias or confounding. No zero-cell correction was needed. Liraglutide has only five and seven co-reporting cases, respectively, giving imprecise estimates and limited reliability of the asymptotic intervals.

## Prespecified drug-role sensitivity

Restricting the exposure flag to a primary- or secondary-suspect drug role gave:

| Medicine | Reaction | ROR (95% CI) | Co-reporting cases |
|---|---|---|---:|
| Semaglutide | Pancreatitis | 4.58 (3.86–5.43) | 157 |
| Semaglutide | Cholelithiasis | 4.79 (3.95–5.82) | 123 |
| Liraglutide | Pancreatitis | 4.33 (1.79–10.48) | 5 |
| Liraglutide | Cholelithiasis | 8.13 (3.84–17.20) | 7 |

The denominator remained 422,458. Reports mentioning a medicine only in other roles moved to its comparator; they were not discarded. Suspect-role exposure totals were 16,506 for semaglutide and 490 for liraglutide. This is a descriptive robustness check on the same reports, not independent confirmation or causal attribution to a particular drug–event pair.

## Input and quality checks

- One complete quarterly release only: 422,459 DEMO rows, 1,627,225 DRUG rows, and 1,394,751 REAC rows. All rows passed delimiter-width and encoding checks; none were silently skipped.
- All report identifiers were usable, and selected DRUG/REAC case identifiers agreed with DEMO. One case had two versions; the later FDA date and higher PRIMARYID retained version 7 instead of version 6.
- The supplied deletion file contained 4,217 distinct numeric identifiers, interpreted as CASEID. It had no overlap with the quarter's cases and changed no included records; it cannot remove unknown future deletions.
- Every selected case had a usable drug and reaction. Six individual drug rows lacked usable names/ingredients, but their reports had other usable drug rows. No age, sex, geography, indication, seriousness, or target-drug restriction was imposed.
- Two release records have FDA_DT 2025-12-30. They were retained because eligibility was defined by the complete quarterly release, not a new date filter. The quarter of release must not be interpreted as the quarter of symptom onset.
- Age was blank in 156,207 eligible cases and sex in 79,419. These fields were not used to restrict or adjust the comparisons.
- Fixed whole-token matching used generic ingredient names and prespecified brand names in DRUGNAME or PROD_AI. It includes combination products (including insulin degludec/liraglutide), compounded descriptions, and one ambiguous drug string, “repaglinide (post ozempic)”. The literal prespecified rule retained that match; the estimate is not restricted to verified current use or approved single-ingredient products. Full source-name counts are supplied. Broad near-name screening found only unrelated PLUVICTO strings among unmatched candidates; no aliases were added.
- Independent raw-CSV parsing, case selection and scalar token matching reproduced every case membership and all eight 2×2 tables. Decimal/SciPy odds-ratio checks, independent Holm arithmetic and zero-cell branch tests passed.
- The saved offline script was executed in a fresh namespace inside the native Notebook. Its results, complete case matrix, name audits, version/date audits and final analysis audit were byte-identical to the original outputs. This was a script reproducibility check, not a native Replay.

## What the patterns can and cannot support

These results identify disproportionately represented reporting combinations that can motivate clinical review and better-designed studies. FAERS has no denominator of exposed people or person-time, so event rates, excess absolute risk, and incidence cannot be calculated. Underreporting, publicity or stimulated reporting, different prescribing populations, indications, concurrent medicines, and the propensity to assign a suspect role can change the ratios without changing causal drug effects.

Reports may be incomplete, unverified or duplicated across distinct case identifiers. Within-quarter version deduplication cannot remove all such duplicates, reconcile all prior quarters, or incorporate later follow-ups. Reaction co-occurrence on a report does not prove that each listed medicine caused each listed event. Exact-term selection deliberately excludes related diagnoses such as “Pancreatitis acute”; it is not a comprehensive clinical case definition. Sparse counts, name ambiguity, combined products, and a single quarter further limit interpretation.

The four comparisons share reports and backgrounds, and each background can contain other GLP-1 medicines. This is not a head-to-head safety comparison. Neither an elevated ROR nor an absent disproportionality result alone would establish harm or safety. FDA explicitly cautions that submitted reports do not establish causation or permit incidence estimation.

## Reproducibility, resources and native status

The original ZIP is 63,223,614 bytes; its complete uncompressed inventory is 337,724,653 bytes. The cumulative download upper bound, including a conservative 6 MB allowance for the three planning-page reads, is 69,223,614 bytes. The 500 MB download and 3 GB extraction limits were respected. Tables were read as ZIP streams, without bulk extraction. ZIP CRC/path checks passed.

SHA-256: `90213500721ecaf976ee45d8ad04aa2dbb80861d93c87fef67ccaf515919bf2a`.

All calculations ran locally in the native Notebook using existing packages; no installation, remote computation, additional plugin, application edit, or .science export was performed. The reproduction bundle provides the frozen configuration, offline analysis.py, a portable notebook entry point, original calculation outputs, source documentation, software inventory, validation evidence and publication/Run references. Source metadata retains no cookie or authorization-header values.

Native environment capture was checked before substantive analysis and checked again after scientific publication. Artifact publication succeeded, but native discovery/lineage could not resolve the returned result version. Native environment capture therefore remains **unverified**; a recorded software inventory is not a substitute for native capture. No native Replay was performed, and no Replay command was exposed to the agent. The operator may attempt at most one native Replay after publication. These native-tool limitations are separate from the completed and independently checked scientific calculations.

## Sources

1. [FDA quarterly extract listing](https://fis.fda.gov/extensions/FPD-QDE-FAERS/FPD-QDE-FAERS.html), inspected 2026-10-08; lists April–June 2026 ASCII, posted 29-Jul-2026.
2. [Original FDA 2026 Q2 ASCII ZIP](https://fis.fda.gov/content/Exports/faers_ascii_2026q2.zip), downloaded 2026-10-08. Included Readme.pdf (revised February 2022), FAQs.pdf (revised October 2024), and ASCII/ASC_NTS.pdf define the release, identifiers and data fields.
3. [FDA AEMS public dashboard limitations](https://www.fda.gov/drugs/fda-adverse-event-monitoring-system-aems/fda-adverse-event-monitoring-system-aems-public-dashboard), inspected 2026-10-08. AEMS is the current FDA site name; the source ZIP retains its FAERS filename.
