pyrbmi

Python implementation of reference-based multiple imputation for regulatory clinical trials.

Direct feature-parity target: R's rbmi (openpharma) and the ICH E9(R1) estimands framework.

---

Motivation

Reference-based multiple imputation (MI) is a cornerstone of sensitivity analysis in Phase II–IV clinical trials, required under ICH E9(R1) estimands framework guidance adopted by FDA (2021) and EMA (2020). The canonical implementation is the R package rbmi (Gower-Page et al., JOSS 2022).

No validated Python equivalent exists. This package closes that gap: same algorithms, same defaults, tested output parity against R rbmi, targeting regulatory submission workflows.

---

Scope

Capability	R rbmi	pyrbmi	Status
Bayesian MI (MCMC)	✅	✅	v0.1.0
Approximate Bayesian MI	✅	✅	v0.2.0
Conditional Mean Imputation (CMI)	✅	✅	v0.3.0
MAR strategy	✅	✅	v0.1.0
Jump-to-Reference (J2R)	✅	✅	v0.2.0
Copy Reference (CR)	✅	✅	v0.2.0
Copy Increment from Reference (CIN)	✅	✅	v0.2.0
Last Mean Carried Forward (LMCF)	✅	✅	v0.2.0
Treatment Policy	✅	✅	v0.2.0
Rubin's rules pooling	✅	✅	v0.1.0
Jackknife resampling	✅	🔲	v0.3.0
Bootstrap resampling	✅	🔲	v0.3.0
Delta adjustments	✅	🔲	v0.5.0
Tipping point analysis	✅	🔲	v0.5.0
ICH E9(R1) estimands framework	✅	🔲	v0.4.0
MMRM base model	✅	🔲	v0.1.0
ADaM-compatible data structures	✅	🔲	v0.1.0
R parity validation suite	✅	🔲	v0.7.0
Regulatory output tables	—	🔲	v0.8.0

---

Installation

# Recommended: uv
uv add pyrbmi

# pip
pip install pyrbmi

# With Bayesian backend (PyMC)
uv add "pyrbmi[bayes]"

Requires: Python ≥ 3.11

---

Quick Start

from pyrbmi import RBMIDataset, Imputer, Strategy, pool

# 1. Load your ADaM-compatible longitudinal dataset
ds = RBMIDataset.from_dataframe(
    df,
    subject="USUBJID",
    treatment="TRT01A",
    visit="AVISIT",
    outcome="AVAL",
    baseline="BASE",
    reference_arm="Placebo",
)

# 2. Define imputation strategy (ICH E9R1 intercurrent event handling)
strategy = Strategy.jump_to_reference()

# 3. Run Bayesian multiple imputation
imputer = Imputer.bayesian(n_samples=200, n_chains=4)
imputed_datasets = imputer.fit_impute(ds, strategy=strategy)

# 4. Analyse each imputed dataset and pool via Rubin's rules
results = pool(imputed_datasets, estimand="difference_in_means")
print(results.summary())

---

Validation Against R rbmi

Numerical output parity is verified against R rbmi on reference datasets. See tests/test_vs_r/ for the full validation suite (requires rpy2).

uv run pytest tests/test_vs_r/ -v

---

Roadmap

See TODO.md for the full atomic task list organized by SemVer milestone.

• v0.1.0 — Bayesian MI + MAR + Rubin's rules + MMRM
• v0.2.0 — Reference-based strategies (J2R, CR, CIN, LMCF, TreatmentPolicy)
• v0.3.0 — Frequentist CMI + jackknife/bootstrap inference
• v0.4.0 — ICH E9(R1) estimands framework
• v0.5.0 — Sensitivity analysis (delta adjustments, tipping point)
• v0.6.0 — MMRM subsystem (full unstructured covariance support)
• v0.7.0 — R parity validation suite (rpy2-based)
• v0.8.0 — Regulatory output tables (RTF/HTML)
• v1.0.0 — JOSS paper submission, PyPI stable release

---

Reference

Gower-Page C, et al. (2022). rbmi: A R package for standard and reference based multiple imputation. Journal of Open Source Software, 7(78), 4251. https://doi.org/10.21105/joss.04251

ICH E9(R1) (2019). Statistical Principles for Clinical Trials: Addendum on Estimands and Sensitivity Analysis. Step 4, November 2019.

---

License

Apache-2.0 — see LICENSE.