Skip to content

Quickstart

Run a complete causal analysis in 5 minutes.

1. Get Data

# Download IHDP dataset (built-in)
causal-toolkit demo --dataset ihdp --out ./data

Or load your own CSV: 

import pandas as pd
df = pd.read_csv("your_data.csv")

2. Define Causal Model

from causal_toolkit import CausalModel
from causal_toolkit.wrappers import DoWhyWrapper
from causal_toolkit.core.base import EstimatorType, IdentificationStrategy

# Specify treatment, outcome, and confounders
model = CausalModel(
    data=df,
    treatment="treatment",      # Column name
    outcome="outcome",          # Column name
    common_causes=["age", "income", "education", "sex"]  # Confounders
)

3. Identify Causal Effect

dowhy = DoWhyWrapper(model)
estimand = dowhy.identify(IdentificationStrategy.BACKDOOR)
print(estimand)
# ATE: E[E[Y|T=t, X] - E[Y|T=t', X]]

4. Estimate Treatment Effect

# Choose estimator
estimate = dowhy.estimate(EstimatorType.DOUBLY_ROBUST)
print(f"ATE: {estimate.value:.4f} [{estimate.ci_lower:.4f}, {estimate.ci_upper:.4f}]")

Available estimators: - LINEAR_REGRESSION — Simple OLS adjustment - PROPENSITY_SCORE_MATCHING — Nearest-neighbor matching - PROPENSITY_SCORE_WEIGHTING — IPTW - DOUBLY_ROBUST — AIPW (recommended default) - CAUSAL_FOREST — Heterogeneous effects (EconML) - TWO_STAGE_LS — Instrumental variables

5. Refute (Sensitivity Checks)

refutations = dowhy.refute()
for r in refutations:
    print(r)

Methods tested: - Placebo treatment (random treatment assignment) - Random common cause (add noise covariate) - Data subset (remove random fraction)

6. Sensitivity Analysis

from causal_toolkit.analysis import SensitivityAnalyzer

analyzer = SensitivityAnalyzer(model)
analyzer.rosenbaum_bounds(estimate)      # Γ-sensitivity
analyzer.cinelli_hazlett(estimate)      # Robustness value
analyzer.e_value(estimate)              # E-value
print(analyzer.summarize())

Key outputs: - Rosenbaum Γ — How strong hidden confounding must be to flip conclusion - Robustness Value (RV) — Minimum confounding strength to change significance - E-value — Minimum risk ratio of unmeasured confounder

7. Config-Driven Pipeline (YAML)

# config.yaml
pipeline:
  identification:
    strategy: backdoor
    adjustment_set: ["age", "income", "education"]
  estimation:
    method: doubly_robust
    params:
      n_estimators: 100
  refutation:
    - placebo_treatment
    - random_common_cause
    - data_subset
  sensitivity:
    method: cinelli_hazlett
    benchmark_covariates: ["age", "income"]

causal-toolkit estimate --config config.yaml --data data.csv --out results/

What's Next?