skcausal

skcausal is a machine learning library for estimating average causal responses from observational data. It uses one consistent fit(X, t, y) contract across categorical, continuous, and multi-column treatments, while letting you plug in familiar scikit-learn models for nuisance regression, density estimation, and final smoothing.

It is built around average potential outcomes and dose-response functions rather than per-unit counterfactual prediction.

Documentation | Continuous treatments | Categorical treatments | Multidimensional treatments

What The Package Includes

• Average causal response estimators for direct, GPS/IPW-style, and doubly robust workflows
• One public interface for categorical, continuous, and multidimensional treatment tables
• Treatment density estimators and density pipelines for nuisance modeling
• Synthetic and semi-synthetic datasets with known truth for benchmarking
• Discovery utilities for listing datasets, density estimators, and causal estimators
• Optional plotting helpers for comparing estimated curves against observed data or ground truth

Installation

Install the base package:

pip install skcausal

Optional extras:

pip install "skcausal[plotting]"
pip install "skcausal[torch]"
pip install "skcausal[optuna]"
pip install "skcausal[skpro]"

Core Contract

Average-response estimators follow the same public workflow:

estimator.fit(X, t, y)
curve = estimator.predict(requested_treatments)

Where:

• X contains pre-treatment covariates
• t contains the observed treatment table
• y contains the observed outcome table
• requested_treatments contains the intervention levels or grid you want to evaluate

For categorical treatments, requested_treatments is usually a short table of valid levels. For continuous treatments, it is usually a dense treatment grid. Predictions return one average response per requested row, averaging over the covariate sample stored during fit.

Quick Start

The example below fits a continuous-treatment direct regression estimator on a synthetic dataset with known ground truth.

import numpy as np
from sklearn.ensemble import RandomForestRegressor

from skcausal.causal_estimators import DirectRegressor
from skcausal.datasets import SyntheticDataset2
from skcausal.utils.treatment_grid import make_cartesian_treatment_grid

dataset = SyntheticDataset2(n=2000, n_features=6, random_state=0)
X, t, y = dataset.load()

estimator = DirectRegressor(
	outcome_regressor=RandomForestRegressor(
		n_estimators=200,
		min_samples_leaf=5,
		random_state=0,
	)
)

estimator.fit(X, t, y)

grid = make_cartesian_treatment_grid(t, n_continuous_points=25)
estimated_curve = estimator.predict(grid)
truth_curve = np.asarray(dataset.predict(X, grid)).reshape(-1)

This pattern generalizes across treatment types:

• Fit on aligned X, t, and y
• Create a treatment table that represents the interventions you want to query
• Call predict(...) on that treatment table
• On synthetic datasets, compare against dataset.predict(X, requested_treatments) when ground truth is available

Choosing A Starting Estimator

DirectNoCovariates is useful as a sanity-check baseline, but it does not adjust for confounding.

Discover Available Components

You can inspect the registered public objects programmatically:

from skcausal.utils.lookup import (
	all_causal_average_response_estimators,
	all_datasets,
	all_density_estimators,
)

all_causal_average_response_estimators(
	as_dataframe=True,
	filter_tags={"capability:t_type": "continuous"},
	return_tags=["capability:t_type", "capability:multidimensional_treatment"],
)

This is useful when you want to discover which estimators support a particular treatment type or whether an estimator handles multidimensional treatments.

Main Public Modules

• skcausal.causal_estimators: average-response estimators and estimator pipelines
• skcausal.density: treatment density estimators and density pipelines
• skcausal.datasets: synthetic and semi-synthetic benchmark datasets
• skcausal.plotting: optional plotting helpers for response curves and comparisons
• skcausal.utils.lookup: object discovery helpers built on top of skbase.lookup

Documentation

• Getting started
• Estimator contract
• Continuous-treatment workflow
• Categorical-treatment workflow
• Multidimensional-treatment workflow
• Dataset catalog
• Benchmarking notes

Scope

skcausal is designed around average-response estimation from observational data:

• Average potential outcomes for binary or categorical treatments
• Average dose-response functions for continuous treatments
• Average response surfaces for multidimensional treatments

It does not present predict(X_new) as the main API. Instead, prediction is framed around requested intervention tables, and estimators average over the covariate sample seen at fit time.