CausalScience

CausalScience provides structural causal model (SCM) simulation, graph-based identification and causal-effect estimation. This source prepares the 0.1.0a14 release candidate with 22 public exports, including estimate_capo and CSCM_capo Simpson support. Both source and installed imports use causalscience; runtime files are under src/causalscience/.

The GitHub repository and local checkout directory are named CausalScience. Historical snapshots and the CausalEngine Noncommercial License retain their original names.

Choose your task

A graph-free estimator executes your chosen identifying formula. It does not infer or verify a causal graph from the data.

Install

Python 3.11+ is required. To install or update the latest published release:

python -m pip install --upgrade causalscience
python -m causalscience.demo

Check python -c "import causalscience; print(causalscience.__version__)" before using version-specific features. Public a13 does not contain the new CAPO API or CSCM_capo Simpson extension. For the a14 candidate, install its reviewed wheel or this checkout in a project environment:

python -m pip install /path/to/causalscience-0.1.0a14-py3-none-any.whl
# Alternatively, from this source checkout:
python -m pip install .

Both routes use import causalscience. See installation for environment setup and editable development. Candidate preparation does not itself publish a release to PyPI.

One complete example

This binary synthetic example estimates the mean under action A=1. It is not an ATE; the adjustment assumption is supplied by the caller.

import numpy as np
import pandas as pd
from causalscience import estimate_bd

rng = np.random.default_rng(11)
c = rng.binomial(1, .5, 1200)
a = rng.binomial(1, .2 + .5*c)
y = rng.binomial(1, .1 + .3*c + .4*a)
data = pd.DataFrame(dict(C=c, A=a, Y=y))
fit = estimate_bd(data, treatment="A", outcome="Y", adjustment=["C"],
    action=1, schema={name: "binary" for name in data}, num_boost_round=30)
assert fit["estimation_status"] == "ACCEPT"
assert fit["graph_verification"] == "not_run"
assert fit["causal_interval"] is None
print(fit["estimate"])

Defaults use Unified Covariate Adjustment (UCA) full scores with XGBoost backward regressions and separate normalized minimum-L2 weights. Graph-based default fitting checks the Fixed-partition Multilinearity Criterion (FMC), then the Multilinearity Criterion (MC) if needed; graph-free fitting uses the supplied functional. Always inspect status and diagnostics. Successful fitting does not verify overlap, causal assumptions, population consistency or causal confidence-interval coverage.

CAPO estimator in the a14 candidate

estimate_capo learns one treatment regime's conditional expected outcome. Call it for actions 0 and 1 to obtain separate CAPO predictors. This API requires the a14 candidate or later and from causalscience import estimate_capo; it is not included in public 0.1.0a13. See the complete source example and target definitions. The existing public six-API tutorial retains its original scope.

Read more

• Document: workflows, assumptions, API contracts and implementation map.
• Six-API tutorial: four graph-free means plus paired ATE/CATE.
• Wiki: detailed modules and method routes.
• Indexed compiler reference: lower-level symbolic/finite tools.
• Whole-call benchmark and shared-fit component pilot: different timing scopes.
• Historical validation: dated version-specific checks.

Browser editions: Document, Wiki, Tutorial.

Repository layout

Source developers install the checkout with python -m pip install ., or python -m pip install --no-deps --no-build-isolation -e . after dependencies are installed. Source and installed packages use import causalscience and python -m causalscience.demo. Source checks use PYTHONPATH=src python -B -m unittest discover -s tests/checkout -v. Old causalengine.* and flat imports such as import SCM are not aliases; use causalscience.* imports. The compatibility notes cover external research scripts and frozen snapshots.

Boundaries and license

Default full-score execution supports declared discrete actions; continuous fixed-point actions and unsupported compiler/design cases are declined. CATE requires a supported baseline anchor and learning assumptions; it returns predictions, not automatic pointwise causal intervals. CSCM_capo has fast direct truth lookup but conditional centering can make mixed-X generation expensive. See limitations.

LICENSE is the CausalEngine Noncommercial License 1.0. LICENSING.md preserves prior MIT grants and separate rights. Engine provenance records current canonical hashes and earlier snapshots.