recoverlite 0.2.0

Pre-data recovery tests for planned study designs: simulate whether a planned design-analysis pair can recover its target estimand, with PASS/RISK/FAIL verdicts under versioned threshold profiles.

recoverlite (Python)

Pre-data recovery tests for planned study designs — the Python mirror of the R package.

A planned study can be unable to support its intended inferential claim even when the researcher's substantive assumptions are correct. recoverlite simulates a declared design–analysis pair over a crossed scenario grid — null and target effects, each under declared and pessimistically perturbed nuisance assumptions — and converts the diagnosands into a PASS / RISK / FAIL verdict under a pre-specified, versioned threshold profile.

Mirror contract

This package is protocol-identical to the R implementation: same declaration API, same scenario grid, same diagnosands (including the exact decomposition target bias = estimator bias + estimand drift), same threshold profiles (shared version string recoverlite-thresholds-0.2), same verdict rule, same report structure. Simulated numbers agree within Monte Carlo error, not byte-identically — R and Python cannot share RNG streams. The test suite enforces the contract by reproducing the R package's archived worked-example diagnosands within 4× combined MCSE.

	R package	Python mirror
Two-arm trial (baseline, additive measurement error, MAR/MCAR attrition, noncompliance)	✅	✅
Complete-case linear model	✅	✅
MI baseline-adjusted estimator (Rubin + Barnard-Rubin)	✅	✅
Cluster trial, LMM (Wald-z / Satterthwaite / Kenward–Roger)	✅ lme4 + lmerTest + pbkrtest	✅ internal exact REML fitter
Cluster-level t-test	✅	✅
Fragility curves (effect + nuisance)	✅	✅

The mixed-model machinery is self-contained (numpy/scipy only — no statsmodels): a closed-form REML fitter for the random-intercept model with Satterthwaite df (observed REML Hessian, the lmerTest algorithm) and Kenward–Roger adjusted covariance + df (expected REML information, the pbkrtest algorithm). Because these methods are deterministic given data, they are validated against lmerTest/pbkrtest to numerical precision (rel. ≤ 1e-4) on shared fixture datasets — balanced, unbalanced, near-boundary, and null — in tests/data/, a stronger contract than the MCSE-level simulation cross-checks.

Install

pip install recoverlite            # numpy + scipy only

Not yet on PyPI? Install from source: pip install git+https://github.com/heidihelena/recoverlite-py.

The workflow in one block

import recoverlite as rl

design = rl.declare_recovery(
    target=rl.target_estimand(
        estimand="ITT mean difference at 12 weeks",
        scale="latent-outcome standardized mean difference",
        sesoi=0.40,
    ),
    data_strategy=rl.two_arm_trial(n_per_arm=115),
    measurement=rl.measured_outcome(reliability=0.70),
    missingness=rl.attrition_model(rate=0.15, mechanism="differential"),
    answer_strategy=rl.planned_analysis(
        estimator="linear_model",
        formula="y_observed ~ treatment",
    ),
)

result = rl.recovery_test(design, sims=2000,
                          scenarios="confirmatory_grid", seed=1)
print(rl.verdict(result))   # PASS / RISK / FAIL (+ strict/lenient recompute)
rl.report(result)           # the standalone report always travels with it

Cluster designs use cluster_trial() with planned_analysis("lmm_random_intercept", "y_observed ~ treatment + (1 | cluster)") (Wald-z) or "cluster_mean_ttest". Fragility curves — effect_fragility(), nuisance_fragility() — are deliberately outside the verdict.

Citation

Andersen, H. H. (2026). Recovery before data: pre-data simulation diagnosis of planned study designs. Working paper; preprint forthcoming. https://github.com/heidihelena/recoverlite

A PASS is evidence about the instrument, not about the world.

License

Apache License 2.0.