Causal Inference Dashboard

An interactive portfolio-grade causal inference dashboard built with Dash in Python. Applies a number of causal methods to a publicly available retail email marketing dataset (Hillstrom, 2008), allowing side-by-side comparison of estimates, uncertainty and assumptions.

---

Screenshot

---

Methods Covered

Tab	Method	Library
2	Matched-Control (Propensity Score Matching)	scikit-learn
3	Bayesian A/B Test	PyMC, ArviZ
4	Uplift Modelling / HTE (T-Learner, S-Learner)	scikit-uplift
5	Multi-Arm OLS w/ Interactions	statsmodels

---

Dataset

The MineThatData Email Analytics dataset (Hillstrom, 2008) captures a randomised marketing experiment across 64,000 US retail customers:

• Men's Email arm: 21,388 customers
• Women's Email arm: 21,307 customers
• Control Group (No Email comm): 21,305 customers
• Outcome: spend ($) in the 2 weeks after the campaign
• Covariates: recency, purchase history, gender catalogue, zip code, newbie status, channel

---

To Add

• Further feature engineering

How to Run Locally

1. Install dependencies

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2. Run the app

python app.py

Open your browser at http://localhost:8050.

First run: The app will pre-compute all causal models (~3–5 minutes for PyMC sampling, PSM bootstrap & uplift models) and cache results to .cache/results.pkl. Subsequent starts load instantly from cache.

3. Force recompute

After changing any estimation logic in causal_utils.py, flip the USE_CACHE flag at the top of that file to False and restart the app - the next run will rebuild from scratch and overwrite .cache/results.pkl. Set it back to True afterwards so subsequent starts (and the Plotly Cloud deployment) load instantly from the pickle.

---

What This Demonstrates

Methodological breadth

Selecting and applying the right method for the causal question: ATE via PSM and OLS, posterior distributions via Bayesian inference, and individual-level CATE via uplift modelling.

Uncertainty quantification

Every estimate is accompanied by an appropriate confidence interval: bootstrap CIs for PSM, HDI for Bayesian, and coefficient CIs for OLS.

Individual-level causal thinking

Uplift modelling moves beyond average effects to identify which customers respond most, enabling targeted campaign optimisation.

Assumption transparency

Every tab includes a collapsible "Methodology & Assumptions" section written for both technical and non-technical audiences.

---

Project Structure

.
├── app.py            # Dash app layout + callbacks
├── causal_utils.py   # All causal estimation logic (PSM, Bayesian, Uplift, OLS)
├── requirements.txt
├── README.md
└── .cache/
    └── results.pkl   # Auto-generated on first run