AssessLite (Python)

Test what your analysis depends on.

Structural assumption assessment for causal analysis — the native Python interface to the AssessLite core specification v0.1. The R package (assesslite on r-universe) implements the same spec; audit records written by either conform to the same JSON Schema.

Every causal result borrows strength across units, places, or times, licensed by invariance assumptions usually left implicit inside exchangeability or i.i.d. sampling. AssessLite makes them explicit, attacks them, and records what survived — with three-way verdicts (stable / unstable / not resolvable at this n), because every stability gate is a bright line on a noisy estimate.

Install

pip install assesslite

Pure numpy + pandas; the Cox (Breslow partial likelihood) and GLM (IRLS) estimators are implemented in-package, so there is no heavy or version-fragile stats dependency.

Use

from assesslite import StructuralAudit

assessment = StructuralAudit(
    data=lung_registry,                 # pandas DataFrame
    outcome=("time", "status"),         # Cox; or "y" for a GLM
    exposure="adherence",
    covariates=["age", "sex", "stage"],
    cluster="hospital",
    time="diagnosis_year",
    subgroups=["stage"],
    unit="patient",
)

assessment.assume(
    "cluster_exchangeability",
    rationale="hospitals follow the same national guideline; assumed provisionally",
    licenses="one pooled effect across hospitals; transport to a hospital outside the sample",
)

assessment.test(
    ["unit_permutation", "cluster_holdout", "temporal_split", "subgroup_stability"]
).decide(abstain_if={"estimate_sign_changes": True})

print(assessment)
audit = assessment.export_audit("audit.json")   # durable audit record (also returned as a dict)
assessment.render_report("report.html")          # self-contained HTML report

Product, process, output

• AssessLite is the product.
• an assessment — a StructuralAudit — is the process: open it, declare its ledger, attack it, decide.
• an audit record is the durable output: export_audit() returns it as a dict and writes it as schema-conforming JSON.

Estimators

• ("time", "status") → Cox proportional hazards (Breslow ties), log hazard ratio.
• binary 0/1 outcome → logistic GLM, log odds ratio.
• continuous outcome → linear GLM, linear coefficient.

The Cox implementation reproduces R's coxph(ties="breslow") coefficients and standard errors exactly.

See examples/worked_example.py for a complete run on simulated multicentre data. Apache License 2.0.