Causal Inference: Loyalty Program Impact Analysis

Estimating the causal effect of a loyalty program on customer spending using DoWhy and propensity score matching.

Overview

This project demonstrates a rigorous causal inference workflow to answer: Does enrolling in a loyalty program cause customers to spend more?

Using simulated observational data, we construct a causal DAG, apply propensity score matching to control for confounders, estimate the Average Treatment Effect on the Treated (ATT), and validate with refutation tests.

Quick Start

# Install
pip install -e ".[dev]"

# Run the full pipeline (CLI)
python scripts/run_analysis.py

# Or with a fixed seed
python scripts/run_analysis.py --seed 42

# Run tests
pytest tests/ -v

# Launch the demo notebook
jupyter notebook notebooks/causal_inference_demo.ipynb

Or use the Makefile:

make install
make run
make test
make notebook

Methodology

Causal Framework

1. Causal Graph (DAG) — Encodes the assumed data-generating process: signup timing, pre-enrollment spending, and unobserved confounders influence both treatment assignment and the outcome.

2. Identification — The backdoor criterion is applied to the DAG to identify valid adjustment sets.

3. Estimation — Propensity score matching creates comparable treated/control groups, then estimates the ATT.

4. Refutation — Three robustness checks validate the estimate:

   • Placebo treatment — Replace treatment with a random variable; effect should vanish.
   • Random common cause — Add a random confounder; estimate should remain stable.
   • Data subset — Re-estimate on a 90% subset; estimate should hold.

Architecture

├── config/
│   └── default.yaml                 # Simulation & model parameters
├── src/
│   ├── data/
│   │   └── simulator.py             # Synthetic data generation
│   ├── models/
│   │   └── causal.py                # DoWhy pipeline wrapper
│   ├── visualization/
│   │   └── plots.py                 # DAG & treatment effect plots
│   └── utils.py                     # Config loading, helpers
├── notebooks/
│   └── causal_inference_demo.ipynb   # Interactive demo (imports from src/)
├── scripts/
│   └── run_analysis.py              # CLI entry point
├── assets/
│   └── causal_dag.png               # DAG visualization
├── tests/
│   └── test_simulator.py            # Data generation tests
├── pyproject.toml
├── Makefile
├── requirements.txt
└── LICENSE

Data Schema

The simulation generates panel data with:

Column	Description
user_id	Unique customer identifier
signup_month	Month of loyalty program enrollment (0 = never enrolled)
month	Observation month (1–12)
spend	Monthly spending (Poisson baseline with seasonal decay)
treatment	Whether the customer enrolled (boolean)

The cohort aggregation step computes per-user pre_spends and post_spends relative to a target signup month.

Configuration

All simulation and model parameters are defined in config/default.yaml:

simulation:
  num_users: 10000
  num_months: 12
  treatment_effect: 100
  signup_month: 3

model:
  estimation_method: iv.propensity_score_matching
  target_units: att

License

MIT — see LICENSE.