longmi

longmi creates multiple plausible versions of missing repeated outcomes, fits the same analysis to each version, and combines the estimates and uncertainty using Rubin's rules.

It is designed for one incomplete longitudinal outcome and makes the assumptions behind every result explicit — as part of the result object, not only in a paper.

Incomplete longitudinal data
            ↓
Fit an imputation model
            ↓
Generate M completed datasets
            ↓
Fit the same analysis M times
            ↓
Pool estimates and uncertainty
            ↓
Report the result and its validity conditions

Installation

pip install longmi               # core imputation and pooling
pip install "longmi[analysis]"   # plus the statsmodels GEE/GLM adapters

The quickstart below uses the analysis extra. From a checkout for development: pip install -e ".[dev]".

Quickstart

from longmi import LongitudinalData, pool_rubin
from longmi.analysis import StatsmodelsGEE
from longmi.impute import NegativeBinomialImputer

data = LongitudinalData(
    frame,                        # long format, NaN where the outcome is missing
    id_col="subject",
    time_col="period",
    outcome_col="seizures",
    predictor_cols=("treatment", "baseline_seizures", "age"),
    outcome_type="count",
    times=(1, 2, 3, 4),           # declared design grid
)

# treatment-by-wave terms keep the analysis's interaction represented
# in the imputation model (assumption A8)
imputer = NegativeBinomialImputer(time_interactions=("treatment",))

fit = imputer.fit(data)           # validate, estimate, diagnose once
completed = fit.impute(m=50, random_state=20260723)

analysis = StatsmodelsGEE(
    "seizures ~ treatment * period + baseline_seizures + age",
    groups="subject",
    family="poisson",
    cov_struct="exchangeable",
)

result = pool_rubin(completed.analyze(analysis), validity=completed.declaration)
print(result.summary())
print(result.validity_report())

The same fitted model reruns under a missing-not-at-random scenario:

from longmi import DeltaAdjustment
import numpy as np

shifted = fit.impute(m=50, random_state=20260723,
                     delta=DeltaAdjustment(np.log(0.8)))  # means x 0.8

What is implemented

• Data contract — LongitudinalData validates the participant-wave grid, preserves observed outcomes bit-for-bit through completion, and enforces count support (src/longmi/data.py).
• Imputers — JointGaussianImputer (continuous outcomes; exact conjugate data augmentation, wave-saturated mean, unstructured covariance) and NegativeBinomialImputer (longitudinal counts; NB random intercept, Gauss–Hermite ML with verified convergence, large-sample posterior-approximation parameter draws, adaptive-grid random-intercept draws, gamma–Poisson outcome draws). Both support delta-adjusted MNAR sensitivity analysis and report numerical diagnostics; a failed fit refuses to impute.
• Analyses — StatsmodelsGEE (marginal GEE, robust sandwich, verified to reproduce direct statsmodels fits exactly), StatsmodelsGLM, and CallableAnalysis for custom estimators.
• Pooling — pool_rubin, multivariate, bit-compatible with mice::pool.scalar (verified at 1e-12 against R), Barnard–Rubin degrees of freedom, fraction of missing information, and validity_report() distinguishing verified properties from declared assumptions.

Validation status

Deterministic components are cross-validated against R (mice); imputation backends have unit, invariant, and numerical-diagnostic tests plus a seeded simulation suite for bias and confidence-interval coverage (pytest -m slow tests/simulation). See docs/project-status.md for the canonical per-feature maturity table — the backends are not claimed statistically validated beyond what that table states.

Documentation

• docs/index.md — documentation home (build with mkdocs serve after pip install -e ".[docs]");
• Why impute the response? and MCAR, MAR, MNAR — concepts;
• Choosing an imputer, interpreting diagnostics, reporting an analysis — task guides;
• Mathematical foundations and validity conditions, assumptions A1–A8, GEE after imputation — theory;
• algorithm specifications — what the software actually computes, approximations included;
• worked examples — the epil count example and the external validation oracles.

Development

pip install -e ".[dev]"
pytest                       # fast suite
pytest -m slow tests/simulation -q   # simulation studies
Rscript validation/r/rubin_reference.R   # regenerate R pooling reference

longmi implements established missing-data theory (Rubin 1976, 1987; Barnard–Rubin 1999; Meng 1994; Wang–Robins 1998); it does not introduce a new proof of multiple imputation. See REFERENCES.md for sources and CITATION.cff to cite the software. MIT license.