Antecedent

Antecedent is an identification-first causal inference engine for Python (with a native Rust core). It takes an analysis from causal structure through estimation, diagnostics, interventions, and counterfactuals — without silently treating discovered graphs as ground truth.

Use it when you need a causal effect you can defend: identification is checked before anything is estimated, refuters run against the estimate by default, and uncertainty about the causal graph — a CPDAG, a PAG, a posterior over graphs — is carried through to the effect instead of being resolved by fiat.

Requires CPython 3.11–3.14. Install from PyPI:

pip install antecedent

Paste this block and run it. It simulates a confounded dataset, checks that the effect is identified, estimates it, and runs refuters against the estimate:

import numpy as np
from antecedent import AverageEffect, analyze

rng = np.random.default_rng(0)
n = 2000
season = rng.normal(size=n)  # confounder
price = 0.7 * season + rng.normal(size=n)  # treatment
sales = 1.5 * price + 2.0 * season + rng.normal(size=n)  # outcome, true effect = 1.5

result = analyze(
    data={"season": season, "price": price, "sales": sales},
    graph=[("season", "price"), ("season", "sales"), ("price", "sales")],
    query=AverageEffect(treatment="price", outcome="sales"),
)

print(
    result.identification
)  # NonparametricallyIdentified via backdoor.adjustment, adjusting for season
print(result.estimate)  # ate=1.48, analytic and bootstrap standard errors
print(result.validation)  # refuters ran and passed

The same analyze() call scales from this to temporal dose × horizon ResponseCurve surfaces, pulse and sustained contrasts, Bayesian graph-posterior mixtures that report unidentified structure mass, mediation, counterfactuals, and root-cause attribution — see the project README for worked examples and the documentation for the full API.

1.9.0

The 1.9.0 release notes cover the calibration of licensed intervals (a two-sided repeated-sampling coverage gate, with boundary records disclosed at runtime), ADMG interventional distributions (Rust Study API only), accepted-Dag counterfactuals, Frequentist DBN-posterior mediation, and staged attribution / transport / interference cells (Rust Study API only). Behaviour changes: rd.sharp defaults to the HC1 SE, NaN and null float cells are missing values, and the default bootstrap count is 199.

1.8.0

The 1.8.0 release notes cover the Bayesian remainder of the staged handle: functional path, distribution, and ADMG ATE, static mediation, counterfactuals, derivatives, and ConditionalEffect graph-posterior mixtures.

1.7.0

The 1.7.0 release notes cover class-preserving Bayesian analyses on incomplete temporal graphs. The 1.6.0 release notes cover per-horizon temporal identification, sequential overlays, and observation-adjusted temporal curves. The 1.5 Python walkthrough remains the guide for retargeting, CDFs, and joint interventions.

Development

For local development you need a Rust 1.85 toolchain. CI builds and smoke-tests wheels for the supported CPython range on Linux x86_64/aarch64 (manylinux), macOS arm64, and Windows x86_64 (default faer path; no system BLAS). Tagged releases publish wheels to PyPI and GitHub Release assets (see docs/development.md).

cd python
uv venv && source .venv/bin/activate
uv sync --group dev
maturin develop
pytest

Lint and types (local gate only — not part of wheel CI):

bash ../scripts/gate_python_lint.sh   # ruff check/format + mypy

Public API

Primary entry point is the OO facade:

import antecedent

g = antecedent.Dag.from_edges(["z", "t", "y"], [("z", "t"), ("z", "y"), ("t", "y")])
result = antecedent.analyze(
    data,  # dict[str, array] or pandas DataFrame
    graph=g,  # or an edge list
    query=antecedent.AverageEffect(treatment="t", outcome="y"),
    inference=antecedent.Frequentist(),  # or antecedent.Bayesian(...)
)
print(result.identification, result.estimate, result.validation)

# Identify without estimating:
id_only = antecedent.identify(
    graph=g,
    query=antecedent.AverageEffect(treatment="t", outcome="y"),
    identifier=antecedent.Identifier.BACKDOOR_ADJUSTMENT,
)

gcm = antecedent.model.fit_gcm(["z", "t", "y"], columns, list(g.edges()))
draws = gcm.sample_do({"t": 1.0}, n=200)

# Or discover then fit (never invents orientations; refuses incomplete PAG/FCI):
fitted, edges = antecedent.gcm.fit_gcm_discovered(
    data, discovery=antecedent.discovery.PC(alpha=0.05)
)

The root namespace is frozen at 52 names as of 1.9 (the 50-name 1.7 contract plus AnomalyAttribution and ChangeAttribution). Those two types exist for the query axis; analyze() refuses them — the licensed cells are Rust Study only. TransportQuery / InterferenceQuery stay stage modules (antecedent.transport, antecedent.interference). Everything else is reached through a stage module (antecedent.discovery, antecedent.priors, antecedent.errors, …).

Also exposed:

• Typed graphs: Dag / Cpdag / Pag / Admg / TemporalDag (d_separated / latent_project on Dag; m_separated on Admg / Pag)
• Identifier / Estimator enums (wire ids) plus string kwargs
• identify(graph=…, query=AverageEffect(…)) — identify without estimating
• Queries: AverageEffect, MediationEffect, Counterfactual, PulseEffect, SustainedEffect, InterventionalDistribution, PathSpecificEffect, ConditionalEffect, TemporalMediationEffect, InterventionResponse, plus the response family (ResponseCurve, AverageDerivative, PointDerivative, Elasticity, SemiElasticity, DirectionalDerivative, ResponseJacobian), plus AnomalyAttribution / ChangeAttribution (root types; analyze() refuses). Temporal dose × horizon uses the same ResponseCurve / InterventionResponse types with keyword-only horizons, policy, and treatment_lag (see examples/python/temporal_response_curve.py).
• antecedent.discovery — PC, GES, LiNGAM, NOTEARS, FCI/RFCI, PCMCI family, Bayesian posteriors
• antecedent.validation.validate_pcmci_* — discovery stability (block bootstrap, FPR, grids, …)
• antecedent.model / antecedent.counterfactual — FittedGcm, sample_do, counterfactual_ite
• antecedent.population — PopulationRegistry / target_* for named predicates and custom-distribution IPW
• antecedent.gcm — fit_gcm_discovered / attribute_*_discovered discover-then-attribute composition
• antecedent.state.CausalState — incremental state with retained batches, events, suff-stats, particle filter
• refute="full"|"placebo"|False on static analyze (refute=True is a TypeError; leave it unset for the licensed default). Static and temporal ResponseCurve / InterventionResponse license validation none only; requesting a scalar refuter suite raises CausalUnsupportedError.
• RD: estimator="rd.sharp" with running_variable / cutoff / bandwidth
• Graph interchange on the classes: Dag.from_dot / .to_dot and the JSON / GML / NetworkX peers
• Design / state examples: examples/python/rank_designs.py, examples/python/causal_state_workflow.py, examples/python/temporal_response_curve.py, examples/python/staged_static_kinds.py (see ADR 0016 — no auto-rerun); catalog in examples/README.md

Build artifacts (_native.*.so) are gitignored; always maturin develop (or install a wheel) on a fresh checkout.

In 1.6, multi-step Sustained supports none/cheap/full validation in both inference modes, including DBN posterior mixtures. Full perturbations refit the sequential model; Bayesian checks retain mechanism PPC and composed-effect prior sensitivity. Temporal Soft multiplicative and truncated_shift compose across single, joint, and multi-step schedules; truncation applies to the propagated conditional mean, not a stochastic draw. Bayesian temporal observation curves and Sequence use an observed-data Gaussian SEM with latent-trajectory Gibbs sampling under an explicit ignorable trajectory coarsening assumption and distinct priors. See the observation contract for the conditioning assumptions; this is not Bayesian IPCW weighting.

In 1.3, antecedent.estimation.PreparedAnalysis also stages Frequentist DAG derivatives, static natural mediation, explicit-DAG unit counterfactuals, and the published observation-pair contract (selected AIPW, marginal KM, conditional Cox IPCW). Derivative cheap/full stay n/a; counterfactual sampling uncertainty is unavailable. See the 1.3 evidence ledger and the 1.2 ledger below for earlier Bayesian and sequential forms.

In 1.2, antecedent.estimation.PreparedAnalysis also supports the licensed Bayesian conditional, temporal-mediation and response forms, and multi-step SustainedEffect(..., window=(-2, -1)). Choose the validation suite when preparing these analyses; second-click refute() remains AverageEffect-only. export_artifact() retains the fitted posterior or response, while export_artifact(payload="query") retains the original query kind and variable IDs. Scalar posterior exports do not include the complete assumption or validation ledger; retain the analysis result alongside them. See the 1.2 evidence ledger for model restrictions and evidence.

Typed exceptions (CausalError and subclasses) mirror Rust CausalError categories. The native module antecedent._native remains available for advanced FFI use (including the flat AteAnalysisResult DTO; prefer nested AnalysisResult).