Causal Impact of Promotions on Customer Retention

Research Question

Does receiving a promotional offer causally increase customer retention — after controlling for confounding variables?

A naive comparison overestimates the effect because higher-spending customers are both more likely to receive promotions and more likely to retain. This project applies causal inference methods to isolate the true treatment effect.

---

Methods

Method	Library	Purpose	Controls Confounding
Naive t-test	scipy	Biased baseline	❌
Propensity Score Matching	scikit-learn	ATT estimation	✅
Double ML (LinearDML)	EconML	ATE with 95% CI	✅
Uplift Modeling	CausalML	Individual-level effects	✅
Bootstrap (1,000 iter)	numpy	Validation	Partial
KMeans Clustering	scikit-learn	EDA segmentation	—
Random Forest Importance	scikit-learn	Feature selection	—

---

Dataset

Synthetic Dataset (used in this project)

Synthetically generated — 100,000 customer records, seed=42 (fully reproducible). Confounding is explicitly introduced: higher-spending customers are more likely to receive promotions, mimicking real-world selection bias.

Column	Description
age, region, segment	Customer demographics
tenure_months	How long they've been a customer
avg_monthly_spend	Average monthly spend (confounds treatment)
purchase_frequency	Purchases per month
last_purchase_days_ago	Recency
num_support_contacts	Support interactions
received_promotion	Treatment (binary, confounded)
retained	Outcome (binary)

True simulated ATE = 0.20 (20 percentage point retention lift from promotions)

Real-World Reference Datasets

The synthetic data structure is inspired by these publicly available datasets:

Dataset	Source	Link
UCI Online Retail Dataset	UCI ML Repository	https://archive.ics.uci.edu/dataset/352/online+retail
E-Commerce Behavior Data	Kaggle	https://www.kaggle.com/datasets/mkechinov/ecommerce-behavior-data-from-multi-category-store
Customer Retention / Store Data	Kaggle	https://www.kaggle.com/datasets/uttamp/store-data
E-Commerce Customer Churn	Kaggle	https://www.kaggle.com/datasets/ankitverma2010/ecommerce-customer-churn-analysis-and-prediction

---

Project Structure

causal-retention/
├── src/
│   ├── data_loader.py          # Synthetic data (pandas + PySpark)
│   ├── feature_engineering.py  # 18 features + KMeans + RF importance
│   ├── causal_model.py         # PSM + Double ML (EconML) + CausalML uplift
│   └── evaluation.py           # Bootstrap + t-test + results summary
├── tests/
│   └── test_pipeline.py        # 15 unit tests (pytest)
├── data/
│   ├── raw/                    # customer_data.csv (generated at runtime)
│   └── processed/              # features.csv, results_summary.csv
├── notebooks/                  # EDA plots (generated at runtime)
├── main.py                     # Full pipeline runner
├── requirements.txt
└── README.md

---

Quickstart

# 1. Clone
git clone https://github.com/ronithjuluri18/Causal-Impact-of-Promotions-on-Customer-Retention.git
cd Causal-Impact-of-Promotions-on-Customer-Retention

# 2. Install dependencies
pip install -r requirements.txt

# 3. Run the full pipeline
python main.py

# 4. View MLflow experiment dashboard
mlflow ui
# Open http://localhost:5000

# 5. Run unit tests
pytest tests/ -v

---

Key Results

Method	ATE Estimate	Confounding Controlled
Naive difference (biased)	~0.27	❌
Propensity Score Matching	~0.20	✅
Double ML (EconML)	~0.20	✅
Bootstrap (naive)	~0.27	❌

Causal methods recover the true simulated ATE of ~0.20 — a statistically significant 20 percentage point retention lift from promotions, after controlling for spend, segment, and tenure confounders. The naive estimate inflates this by ~35% due to selection bias.

---

Tech Stack

Python · PySpark · scikit-learn · EconML · CausalML · MLflow · pandas · NumPy · SciPy · statsmodels · pytest

---

Author

Ronith Pradyumna Juluri M.S. Data Science, Pace University (GPA: 3.89) linkedin.com/in/ronith-juluri