pharmaverse — CDISC SDTM/ADaM test data

A complete synthetic clinical trial in regulatory submission format: subject-level, adverse events, labs, vitals, exposure, PK — plus the raw-to-SDTM-to-ADaM pipeline that produced them.

• Examples site: https://pharmaverse.github.io/examples/
• Packages (CRAN, all Apache License 2.0 — verified 2026-08-12):
  • pharmaverseraw 0.1.1 — raw source datasets
  • pharmaversesdtm 1.5.0 — SDTM domains (64 datasets)
  • pharmaverseadam 1.3.0 — ADaM analysis datasets (30 datasets)
• Direct download without R’s package machinery: https://github.com/pharmaverse/pharmaverseadam/raw/main/data/<name>.rda
• Attribution: cite the package (pharmaverseadam, Apache-2.0) and note the CDISC pilot origin.

⚠ Not eligible as a jam build dataset

The data is synthetic. From the pharmaversesdtm documentation: some datasets are sourced from the CDISC pilot project, others were “constructed ad-hoc by the {admiral} team”, the pilot data is used as input to “create realistic synthetic data”, and the project states that no personal data is to be used.

That fails B4 directly — “Your source must be real, public, and licensed for this use. Synthetic or simulated data is out of scope.”

So no team may build a claim on this. It is in the folder for three other reasons, and the first one solves a real problem.

Why it is here anyway

1. It is the safe replacement for the placebo dataset

indo-rct-placebo is the highest-sensitivity item in this folder: real patients, a real complication, four named American hospitals, and a recommendation that it be organiser-presented with sites anonymised rather than handed to a team.

This dataset supports the same clinical-trial traps with none of that risk. Subgroup hunting, analysis-population choice, endpoint definitions, missing data, site effects — all present, all demonstrable, and nobody is harmed by a false conclusion about synthetic subject 01-701-1015.

If the organisers decide indo_rct is too sharp to hand out, this is what replaces it, and almost nothing is lost pedagogically.

2. It is the best available fixture for the rule checker

The jam’s rules/ machinery needs realistic lineage to validate against, and the current fixture is a toy. Here the pipeline itself is published: raw → SDTM → ADaM, with the transformation scripts, in a format regulators actually accept — so the lineage has real depth rather than three steps invented for a demonstration.

3. It is what the room’s professional half already uses

Anyone at Schmiede who works in pharma will recognise ADaM instantly. Showing that the traps in this folder are equally at home in a regulatory submission format — one designed end to end for traceability — is a much stronger claim than showing them in municipal tree data.

Shape

ADSL (subject level), the spine — 306 subjects, 55 columns:

Companion datasets:

curl -sL -o adsl.rda https://github.com/pharmaverse/pharmaverseadam/raw/main/data/adsl.rda
Rscript -e 'load("adsl.rda"); str(adsl)'

13 KB for adsl, 1.2 MB for adlb. Trivially laptop-sized.

The traps

1. ARM versus TRT01A — the analysis population, and it is not subtle

Computed 2026-08-12. 12 of 306 subjects randomised to Xanomeline High Dose actually received Low Dose. So the dataset ships two correct, standard, documented answers to “who was in the high-dose group”: 84 (as randomised) or 72 (as treated).

Proportion of subjects with at least one adverse event:

The apparent dose-response effect more than triples — from 2.3 to 7.6 percentage points — by selecting one documented column instead of another. Nothing is filtered, nothing is dropped, no row is touched.

Both choices are orthodox. Intention-to-treat is the regulatory default for efficacy; as-treated is standard for safety. A methods sentence saying “safety analyses were performed on the as-treated population” is correct, conventional, and does the entire job.

This is the same structure as totalCost versus ecMaxContribution in cordis-eu-research-funding: two correctly-labelled columns measuring different things, and one word in your sentence covering both. That it recurs in a regulatory submission standard is the point worth making in a Defensio.

2. SAFFL — a 17% exclusion in a single flag

52 of 306 subjects are screen failures with SAFFL = "N". Filtering to SAFFL == "Y" is correct, universal, and required — and it silently removes 17% of the enrolled population.

Overall proportion with at least one AE:

Screen failures contribute 0 of 52 AEs, because they never received treatment. So including them deflates every rate, and a team that quietly uses the enrolled count as its denominator has an honest-looking, fully reproducible, 15-point understatement of harm.

The ADaM standard exists precisely to make population choices explicit — and the flag is what makes the exclusion invisible, because everyone applies it without reporting the count.

3. Sites of size 1

17 sites, ranging from 51 subjects down to 1, with two sites under 5. Exactly the structure that made indo-rct-placebo dangerous — a “centre effect” bar chart where one bar rests on a single patient — available here without naming a real hospital.

Combine with no error bars (see articles/techniques/06-showing-and-framing.md) and the chart makes itself.

4. 47 lab parameters × 29 visits

adlb is a garden of forking paths with a controlled vocabulary. 47 parameters, 29 visits, times arm, times subgroup: tens of thousands of defensible comparisons, every one a real clinical measure with a plausible mechanism attached. Change from baseline, percent change, shift tables, last observation, worst observation — each a standard derivation, each a different answer.

1.3% of AVAL is missing, which is small enough that nobody checks and non-random enough to matter.

5. The traceability irony

ADaM’s entire design purpose is traceability: every analysis value carries AVAL, AVALC, PARAM, AVISIT, and derivation flags, so a reviewer can walk from a table back to the raw observation. It is the closest thing in industry to what the jam’s platform is building.

And every trap above survives it intact. Perfect traceability, standard variables, submission format — and the dose-response effect still triples on a column choice. That is the jam’s thesis in the one domain that has spent thirty years and a great deal of money trying to make it false.

If a single slide has to justify the whole event to a pharmaceutical audience, it is this one.

Sensitivity

None — and that is unusual for clinical data, so it is worth being explicit about why. The subjects are synthetic, no personal data is present by project policy, the licence is Apache 2.0, and no institution is identifiable. A false conclusion here cannot be repeated as medical misinformation about a real treatment, because there is no real treatment: Xanomeline dosing in this dataset describes nobody.

The only care needed is the reverse of the usual one — label it as synthetic, so nothing derived from it is ever mistaken for evidence about a real drug.

Notes

• pharmaversesdtm has 64 datasets including therapeutic-area variants; if a team wants oncology response criteria (rs_onco_recist, rs_onco_irecist, rs_onco_imwg, rs_onco_pcwg3) the choice of response criterion is a normaliser trap of the first order — the same scans, four standards, four answers.
• REGION1 and COUNTRY have one level each, so no geographic slicing.
• The examples site also covers TLG generation (rtables, tern), logging and eSub packaging — i.e. the reporting layer, which is where family 06 lives.
• ADaM specifications need a free CDISC account; the data does not.