# Prospective protocol: FAERS 2026 Q2
Frozen 2026-10-08 before downloading or inspecting report/reaction counts.

## Question and scope
In the complete April–June 2026 FDA quarterly ASCII extract, are the exact reported terms Pancreatitis and Cholelithiasis disproportionately represented in reports mentioning semaglutide or liraglutide, relative to all other eligible reports? Four primary comparisons only. This is reporting disproportionality, not incidence, causation, or a ranking of medicine safety.

Source archive: https://fis.fda.gov/content/Exports/faers_ascii_2026q2.zip
Official index: https://fis.fda.gov/extensions/FPD-QDE-FAERS/FPD-QDE-FAERS.html
FDA limitations: https://www.fda.gov/drugs/fda-adverse-event-monitoring-system-aems/fda-adverse-event-monitoring-system-aems-public-dashboard

## Fixed exposure and reaction definitions
- Semaglutide aliases: SEMAGLUTIDE, OZEMPIC, RYBELSUS, WEGOVY.
- Liraglutide aliases: LIRAGLUTIDE, VICTOZA, SAXENDA.
- Normalize drug names and active-ingredient fields with Unicode NFKC, uppercase, punctuation-to-space, and whitespace collapse. Match whole normalized alias tokens/phrases in either DRUGNAME or PROD_AI. This includes any recorded product containing the named ingredient, including combinations; report combination matches explicitly. No fuzzy expansion or outcome-informed alias additions.
- Primary exposure: at least one matching drug row in any role. A report may mention both medicines and contributes independently to each medicine's comparison.
- Sensitivity: at least one matching row whose ROLE_COD is PS or SS. Same eligible universe; reports mentioning the medicine only in other roles move to the comparator. This is a descriptive role sensitivity, not an independent confirmatory analysis.
- Reactions: exact PT after trimming, case folding and whitespace normalization: PANCREATITIS; CHOLELITHIASIS. Do not add acute/chronic pancreatitis subterms, related diagnoses, or broad clinical groups. Each reaction appears at most once per report in analysis.
- Audit all matched source strings and nearby unmatched strings using target stems. Do not change definitions in response to outcomes.

## Case selection and complete background
Read the release README and deletion instructions. Apply the supplied deleted-case list, if any. From DEMO retain the latest available report per CASEID by FDA_DT, then numeric PRIMARYID. Inspect conflicting duplicate identifiers and report versions. Missing/invalid FDA dates sort before valid dates; use numeric PRIMARYID as the deterministic tie breaker and report such records.
Require usable numeric CASEID/PRIMARYID, at least one usable DRUGNAME or PROD_AI, and at least one nonempty reaction PT. Collapse duplicate drug/reaction membership. Do not restrict age, sex, geography, indication, reporter, seriousness, or target drugs in the background. The extract is not cumulative; deduplication is limited to versions present in this quarter and cannot identify all cross-case duplicates or later updates.
Record file-level counts, parse errors, deletion counts, case/report selection, unmatched join identifiers and unusable rows. Fail rather than silently skip malformed input rows. Investigate duplicate PRIMARYID records with conflicting information before proceeding.

## Statistics and verification
For each medicine/term: a=exposed/event, b=exposed/no event, c=unexposed/event, d=unexposed/no event. Every eligible report contributes once. ROR=ad/(bc); 95% log-Wald CI=exp(log(ROR) +/- 1.96 sqrt(1/a+1/b+1/c+1/d)).
Preserve raw counts and flag sparse cells (<5). For any zero cell, mark raw ROR as zero, infinity or undefined as appropriate and use 0.5 added to all cells for an explicitly labeled corrected estimate and interval. Two-sided Fisher exact tests on raw counts; Holm adjustment across the four primary comparisons. Sensitivity estimates and nominal tests are descriptive.
Check totals and margins, reconstruct 2x2 counts independently from report-ID sets joined to raw source rows, and verify ratio arithmetic with an independent calculation. No additional outcomes, subgroup mining, quarter substitution, or clinical-risk estimates.

## Execution, provenance and limits
Local native Notebook only, existing packages only, no plugins or remote services. No prior project scientific inputs. Total download ceiling 500,000,000 bytes including retries and supporting sources; extraction ceiling 3,000,000,000 bytes checked against the archive directory before extraction. Validate ZIP paths, CRCs and source SHA-256. Source provenance allows only URLs, retrieval times, byte counts, content type, ETag and Last-Modified; never save cookies or authorization values.
Deadline: 2026-10-08 16:12 Asia/Shanghai. Reserve time for validation/publication; if a complete valid quarter cannot be processed, report the limitation rather than analyze a partial quarter.
Publish original inputs, configuration, code, compact eligible-case matrix, source-name audit, selection/quality audit, result tables and report. Native environment capture and Replay availability are distinct from scientific completion. No .science export, application edits or installations. Operator may perform at most one native Replay after publication.

## Requirements checklist
1. One complete quarter and four fixed comparisons: archive identity, protocol and result-row checks.
2. Full eligible background: case uniqueness, selection reconciliation and equal 2x2 totals.
3. Correct membership and arithmetic: independent set-based counts and ratio calculation.
4. Bounded ordinary downloads/local analysis: byte ledger, archive inventory and Notebook Run evidence.
5. Inspectable rerun: source checksum, deterministic script/configuration, results, audits and version manifest.
6. Honest interpretation: no denominator-based incidence, causal claims or comparative safety ranking.
7. Traceable publication: input/output Artifact versions and exact producing Run IDs; unresolved native capture disclosed.
