CausalEstimate is a Python library designed for causal inference, providing a suite of methods to estimate treatment effects from observational data. It includes doubly robust techniques such as Targeted Maximum Likelihood Estimation (TMLE), alongside propensity score-based methods like inverse probability weighting (IPW) and matching. The library is built for flexibility and ease of use, integrating seamlessly with pandas and supporting bootstrap-based standard error estimation and multiple estimators in one pass.
- Causal inference methods: IPW, AIPW, TMLE, Matching, etc.
- Supports multiple effect types: ATE, ATT, Risk Ratio, etc.
- Bootstrap standard error estimation and confidence intervals
- Common-support filtering and matching (greedy, optimal)
- Plotting utilities for distribution checks (e.g., propensity score overlap)
pip install CausalEstimateOr for local development:
git clone https://github.com/kirilklein/CausalEstimate.git
cd CausalEstimate
pip install -e .You can import any estimator class (e.g., IPW, AIPW, TMLE) and call compute_effect(df) directly. Columns (treatment, outcome, propensity score) are passed to the estimator in its constructor.
import numpy as np
import pandas as pd
from CausalEstimate.estimators import IPW
# Simulate data
np.random.seed(42)
n = 1000
ps = np.random.uniform(0, 1, n) # true propensity for treatment
treatment = np.random.binomial(1, ps) # actual treatment assignment
outcome = 2 + 0.5 * treatment + np.random.normal(0, 1, n)
df = pd.DataFrame({
"ps": ps,
"treatment": treatment,
"outcome": outcome
})
# Create an IPW Estimator for ATE
ipw_estimator = IPW(
effect_type="ATE",
treatment_col="treatment",
outcome_col="outcome",
ps_col="ps",
# optionally stabilized=True if you want stabilized IP weights
)
results = ipw_estimator.compute_effect(df)
print("IPW estimated effect:", results)In this case, results is simply a dictionary with the effect estimate computed from a single sample run (n_bootstraps=1). When no bootstrapping is applied, the output includes the key "n_bootstraps": 0.
If you want to run multiple estimators (e.g., IPW, TMLE, AIPW) on the same dataset in one pass—optionally applying bootstrap or common-support filtering—you can use the MultiEstimator.
from CausalEstimate.estimators import IPW, AIPW, TMLE, MultiEstimator
ipw = IPW(
effect_type="ATE",
treatment_col="treatment",
outcome_col="outcome",
ps_col="ps"
)
aipw = AIPW(
effect_type="ATE",
treatment_col="treatment",
outcome_col="outcome",
ps_col="ps",
probas_t1_col="predicted_outcome_treated",
probas_t0_col="predicted_outcome_control"
)
tmle = TMLE(
effect_type="ATE",
treatment_col="treatment",
outcome_col="outcome",
ps_col="ps",
probas_col="predicted_outcome",
probas_t1_col="predicted_outcome_treated",
probas_t0_col="predicted_outcome_control"
)
multi_estimator = MultiEstimator([ipw, aipw, tmle])
# Apply bootstrap (n_bootstraps > 1 triggers bootstrapping), common support, etc.
results = multi_estimator.compute_effects(
df,
n_bootstraps=50, # If n_bootstraps > 1, bootstrapping is applied.
apply_common_support=True,
common_support_threshold=0.05,
return_bootstrap_samples=True # Optionally return raw bootstrap estimates.
)
print(results)Here, results is a dictionary with keys corresponding to each estimator's class name (e.g., "IPW", "AIPW", "TMLE"). For estimators that perform bootstrapping (i.e. when n_bootstraps > 1), the output dictionary includes:
"effect": The mean effect across bootstrap samples."std_err": The standard deviation of the bootstrap estimates."CI95_lower"and"CI95_upper": The 95% confidence interval computed using the percentile method."n_bootstraps": The number of bootstrap samples (e.g., 50).- Optionally, if
return_bootstrap_samples=True, a"bootstrap_samples"key with the raw bootstrap estimates (e.g., for the overall effect, treated, and untreated effects).
When no bootstrapping is performed (i.e. n_bootstraps is set to 1), "n_bootstraps" is set to 0 and the bootstrap summary keys (like "std_err", "CI95_lower", "CI95_upper") may not be present.
The library supports both optimal and greedy (a.k.a. eager) matching. For example:
import pandas as pd
import numpy as np
from CausalEstimate.matching import match_optimal, match_eager
df = pd.DataFrame({
"PID": [101, 102, 103, 202, 203, 204],
"treatment": [1, 1, 1, 0, 0, 0],
"ps": [0.30, 0.35, 0.90, 0.31, 0.34, 0.85],
})
# Optimal matching (with caliper=0.05, 1 control per treated)
matched_optimal = match_optimal(
df, n_controls=1, caliper=0.05,
treatment_col="treatment", ps_col="ps", pid_col="PID"
)
print("Optimal Matching Results:")
print(matched_optimal)
# Eager (greedy) matching
matched_eager = match_eager(
df, caliper=0.05,
treatment_col="treatment", ps_col="ps", pid_col="PID"
)
print("Eager Matching Results:")
print(matched_eager)Both functions return a DataFrame of matched pairs (or sets), typically with columns like [treated_pid, control_pid, distance].
CausalEstimate provides basic plotting utilities to visualize distributions of propensity scores or predicted outcome probabilities across treatment vs. control.
📌 Generated from this notebook
import matplotlib.pyplot as plt
from CausalEstimate.vis.plotting import plot_propensity_score_dist, plot_outcome_proba_dist
# Suppose df has columns "ps", "treatment", and "predicted_outcome"
fig, ax = plot_propensity_score_dist(df, ps_col="ps", treatment_col="treatment")
plt.show()
fig, ax = plot_outcome_proba_dist(df, outcome_proba_col="predicted_outcome", treatment_col="treatment")
plt.show()See CONTRIBUTING.md for details on setting up a dev environment, running tests, and contributing to this project.
CausalEstimate is licensed under the MIT License. See LICENSE for more details.
- GitHub: kirilklein
- Email: kikl@di.ku.dk
Please open issues or pull requests if you find any bugs or want to propose enhancements.
If you use CausalEstimate in your research, please cite it using the following BibTeX entry:
@software{causalestimate,
author = {Kiril Klein, ...},
title = {CausalEstimate: A Python Library for Causal Inference},
year = {2024},
url = {https://github.com/kirilklein/CausalEstimate},
version = {X.Y.Z},
note = {GitHub repository}
}