Indomethacin RCT — placebo arm (indo_rct)

A real randomised, double-blind, placebo-controlled trial: 602 patients, rectal indomethacin versus placebo for the prevention of post-ERCP pancreatitis. 32 covariates per patient.

• Source: medicaldata R package, dataset indo_rct — https://cran.r-project.org/web/packages/medicaldata/
• Local copy: /mnt/c/dev/git-repos/asifin/data/indo_rct.csv (148 KB) — already in the asifin repo, together with a research plan, a source catalogue and prior analysis in findings.md
• Licence: MIT (the medicaldata package, Peter Higgins, plus an additional LICENSE file). Verified 2026-08-07.
• Original study: Elmunzer BJ et al., A randomized trial of rectal indomethacin to prevent post-ERCP pancreatitis. N Engl J Med 2012;366:1414–22.
• Attribution: cite the NEJM paper and the medicaldata package. The package exists explicitly for teaching, which is what this is.

Read the sensitivity section before using this. It is the last section and it is the reason this entry exists in the form it does.

Shape

Headline results, recomputed from the CSV 2026-08-07 (they match the published trial):

The drug works. That is not in dispute here and is not what this dataset is for in the jam.

Why it suits the jam

Because it is the jam’s thesis stated in clinical form. A placebo arm is a group of people who received a documented, inert intervention. Whatever varies between them cannot be the drug. So the placebo arm is a controlled laboratory for exactly the question the event is about: how much of an apparent effect is the world, and how much is measurement, context and who was counted?

Because the interesting variation is real and large, and the honest explanation is uninteresting while the dishonest one is fascinating. That asymmetry is the whole engine of the event.

Because it is small enough to check. 307 rows. Every subgroup can be counted by hand. There is nowhere to hide, which makes for a fast, brutal Defensio.

The traps

1. Same placebo, 2.3× the outcome — the site effect

Placebo arm only, by trial site:

Among patients who received no active drug, the complication rate differs by a factor of 2.3 between the two centres that hold 96% of the arm. Prior analysis in the asifin repo puts this at p = 0.002, with a multivariate odds ratio of 0.284 for Indiana versus Michigan after adjustment.

The as-if reading is irresistible: the institution is the treatment. Ritual, staff confidence, the room, the script — the context of care determines whether you get sick. The room will spend forty minutes on mechanism and enjoy every one of them.

The boring reading is case mix, referral patterns, procedural technique, and — most importantly — what each centre counts as post-ERCP pancreatitis. The endpoint is a consensus definition applied by clinicians. Two sites applying the same written criteria with slightly different thresholds will produce different rates from identical patients. That is a normaliser, in the jam’s exact sense: the rule that decides whether two things are “the same”. It is the fourth column of the table in CLAUDE.md, sitting in a hospital.

2. The site with one patient

Look at the bottom row again. Case Western contributed a single placebo patient, who did not develop pancreatitis. Rate: 0.0%.

That is a true, reproducible, correctly computed statistic. “One participating centre recorded a zero per cent complication rate in the placebo arm” is a sentence you could put on a poster today.

It is also worth noting how it behaves in practice: the prior analysis in /mnt/c/dev/git-repos/asifin/findings.md reports the site effect as “UM: 28.7%, IU: 12.6%, UK: 8.3%” — the n = 1 site is silently absent. That is not misconduct; dropping a singleton is the sensible thing to do. But it happened without being recorded as a decision, and it is exactly the class of step the platform is built to make visible. A team could reasonably build their whole entry on the difference between the analysis that mentions Case Western and the one that does not.

Use this in the method session. It is a live example, from a real analysis in this organisation’s own repo, of an entirely defensible exclusion that never got written down.

3. Fifty-three subgroups, and one of them is always significant

Testing every single-covariate subgroup with n ≥ 20 in the placebo arm — 53 of them — against the 16.9% base rate:

A 4.1× spread between the extremes, from one file, with no interaction terms and no combinations. Cross two covariates and there are hundreds. A team that wants “patients who X are twice as likely to Y despite receiving no treatment” can have it in ten minutes, and every one of those subgroups is a real clinical category with a real mechanism story attached.

Note train — trainee involvement, 22.1% versus 12.6%. The operator changes the outcome is a genuinely interesting finding, a plausible mechanism, and also exactly what you would expect if teaching cases are systematically harder. Both stories fit. Nothing in the data separates them.

4. The 12-patient centre

Kentucky’s 8.3% rests on 1 event in 12 patients. Its 95% confidence interval covers essentially the whole plausible range. Put it on a bar chart next to Michigan’s 28.7% without error bars — a plotting-step parameter, recorded in the lineage, that nobody inspects — and you have a clean three-bar “institutional effect” chart in which one bar is noise and one bar is a rounding error.

Sensitivity — the strictest in this folder

This is real patient data about a real complication at four named American hospitals, and it is the one dataset here where I would not let a team proceed without a conversation first.

Three distinct risks, in increasing order of seriousness:

1. Medical misinformation. Anything implying that indomethacin does not work, or that placebo is an adequate substitute for treatment, is straightforwardly dangerous and must not be built. The trial’s actual result — the drug halves the complication rate — is not the material.
2. Named institutions. “University of Michigan gives you a 3× higher chance of pancreatitis” is a specific, attributable claim about a real hospital. This is worse than generic health misinformation because it has a defamation shape and a named target who is not in the room. Anonymise the sites to A/B/C/D in every published artefact. The finding survives anonymisation completely intact; nothing of value is lost.
3. Patients. They are de-identified and the data is openly published for teaching, so there is no re-identification concern. But the outcome is a painful complication that happened to 52 real people in the placebo arm, and posters should not be flippant about it.

What is safe, and it is the good part anyway

Frame every claim as about measurement and context, never about efficacy:

• ✅ “Where you are treated predicts your outcome more than who you are” — a claim about institutional variation and endpoint definition.
• ✅ “The definition of the endpoint is doing work the mechanism is getting credit for.”
• ✅ “A centre with one patient achieved a 0% complication rate.”
• ❌ Anything about whether the drug works.
• ❌ Anything a reader could act on medically.
• ❌ Anything naming a hospital.

Under F1 the honest assessment is that this dataset fails the “obviously a game” preference. Weather, trees and exoplanets are safe because absurdity is legible. Post-ERCP pancreatitis is not funny and an absurd conclusion about it does not read as absurd to someone outside the room.

Recommendation: use it in the jam-method session, presented by an organiser, with the sites anonymised — not as a team build dataset. It is the sharpest illustration in the folder of a normaliser deciding an outcome, and that lesson is worth having. Handing it to a team for three days of adversarial persuasion is a different proposition, and one for Michael, Erik and Walter to decide together rather than for this folder to authorise.

Related material in the asifin repo

/mnt/c/dev/git-repos/asifin/ is a separate workstream — placebo and as-if science research on open clinical data, doubling as a dogfooding case for the improve platform. Relevant files:

• placebo-studies-sources.md — catalogue of open datasets with placebo arms: PRO-ACT (ALS, 12 500+ patients), NIDA Data Share, NINDS archived trials, ImmPort, PhysioNet, ClinicalTrials.gov summary results. All larger and all subject to the same sensitivity analysis as above; none licence- checked for jam use.
• findings.md — the prior analysis this entry verifies and extends.
• research-plan.md, study-catalog.json, variables-placebo.csv.

If more clinical material is wanted for the jam, the ClinicalTrials.gov / AACT route is the safer one: summary-level data, no individual patients, no named sites, and the depression extract already described in findings.md (348 placebo-controlled studies, 2001–2024) has an obvious temporal-trend trap of its own — placebo response has been rising for two decades is a real, contested claim in the literature and would make an excellent hypothesis that harms nobody.