causalsoap

Causal Feature Selection and Dimensionality Reduction using Residual-Based ATE Estimation
Author: Kazi Sakib Hasan

💡 What is causalsoap?

causalsoap is a Python library for causal inference-driven feature selection and dimensionality reduction technique.
It ranks features based on their Average Treatment Effect (ATE) on an outcome variable by applying the Frisch–Waugh–Lovell (FWL) theorem using residualization and double machine learning.

This method is particularly useful when:

• You want interpretable ranking of features by causal effect
• The dataset has confounders
• Traditional correlation-based selection is misleading

Link to preprint will be available soon.

📦 Installation

pip install causalsoap

🚀 Quickstart

import pandas as pd
import numpy as np
from causalsoap import CausalDRIFT

# Simulated data
df = pd.DataFrame({
    'X1': np.random.randn(100),
    'X2': np.random.rand(100),
    'X3': np.random.randn(100),
    'X4': np.random.choice([0, 1, 2], size=100),  # categorical numeric
    'Y': np.random.randn(100)
})

# Run model
X = df.drop(columns='Y')
y = df['Y']

model = CausalDRIFT()
model.fit(X, y, outcome_type='continuous', categorical_features=['X4'])

print(model.get_feature_ate())

⚙️ Parameters

fit(X, y, outcome_type, categorical_features=None) 

X : Feature matrix (all numeric) pd.DataFrame y : Target variable pd.Series outcome_type : Continuous or categorical str categorical_features: List of column names in X that are categorical but encoded numerically list[str]

## 📈 How it Works

For each feature:

1. Predict the outcome using confounders → compute residual (Ro)

2. Predict the feature (treatment) using confounders → residual (Rt)

3. Estimate ATE via linear regression: Ro ~ Rt