Causalmodels

0.4.0 · active · verified Fri Apr 17

Causalmodels is a Python library for defining, analyzing, and inferring causal relationships from data, drawing inspiration from Judea Pearl's do-calculus. It provides tools for building Bayesian causal models, performing matching, and conducting regression-based causal inference. The current version is 0.4.0, with an irregular release cadence.

Common errors

```
KeyError: "['NonExistentColumn']" or similar error indicating a missing column.
```
cause The column name provided for `treatment`, `outcome`, or `control_variables` does not exist in the input DataFrame.

fix Double-check the spelling of column names against `df.columns` to ensure they exactly match the DataFrame columns, respecting case sensitivity.
```
TypeError: unsupported operand type(s) for +: 'str' and 'int' or similar numerical calculation error.
```
cause One or more columns intended for numerical calculation (treatment, outcome, control) contain non-numeric data (e.g., strings, objects) that cannot be directly processed by the underlying statistical models.

fix Convert all relevant columns to appropriate numeric types (e.g., `float`, `int`) using `df['column'] = pd.to_numeric(df['column'], errors='coerce')` before passing the DataFrame to `causalmodels` methods.

Warnings

gotcha Causal inference methods in `causalmodels` (and generally) rely on strong assumptions (e.g., no unmeasured confounders, correct specification of the causal graph). Failing to meet these assumptions can lead to biased estimates.
Fix: Carefully define your causal graph, identify all potential confounders, and consider domain knowledge before applying methods. Validate assumptions where possible through sensitivity analysis or alternative approaches.
gotcha Input data to `causalmodels` methods must be clean and appropriately preprocessed. Missing values, incorrect data types, or inconsistent column names can lead to errors or silently biased results during estimation.
Fix: Ensure your DataFrame is clean, handles missing values (e.g., imputation or removal) appropriately, and all columns used in causal models are of the expected numerical type before passing the data.

Install

pip install causalmodels Install stable version

Imports

BayesianModel

from causalmodels.bayesian_model import BayesianModel

Matching

from causalmodels.matching import Matching

Regression

from causalmodels.regression import Regression

CausalDAG

from causalmodels.causal_dag import CausalDAG

CausalInference

from causalmodels.inference import CausalInference

Quickstart

This quickstart demonstrates how to use the `Regression` module to estimate the Average Treatment Effect (ATE) of a treatment variable 'X' on an outcome 'Y', while controlling for a confounder 'Z'. It simulates data reflecting a causal relationship and then applies the regression model.

import pandas as pd
import numpy as np
from causalmodels.regression import Regression

# Simulate some data with a known causal effect
np.random.seed(42)
n_samples = 1000

# Confounder Z affects both Treatment X and Outcome Y
Z = np.random.normal(0, 1, n_samples)
# Treatment X is affected by Z
X = 0.5 * Z + np.random.normal(0, 1, n_samples)
# Outcome Y is affected by X and Z
Y = 2.0 * X + 1.0 * Z + np.random.normal(0, 1, n_samples)

data = pd.DataFrame({'Z': Z, 'X': X, 'Y': Y})

# Initialize the Regression model
# X: treatment variable, Y: outcome variable, control_variables: confounders
model = Regression(data, treatment='X', outcome='Y', control_variables=['Z'])

# Estimate the Average Treatment Effect (ATE)
ate_estimate = model.estimate_ate()

print(f"Observed data with N={n_samples} samples.")
print(f"Estimated Average Treatment Effect (ATE) of X on Y, controlling for Z: {ate_estimate:.4f}")

view raw JSON →