Master's Thesis: Causal AI in Marketing Measurement

This repository contains the code and notebooks developed for my Master's thesis on Causal Artificial Intelligence (Causal AI) and its applications in marketing measurement, experimentation, and decision-making. The work applies causal inference tools like DoWhy, EconML, Causica, and GCM to simulated marketing data.

---

Contents

Causal Graph & Assumptions

• CausalGraphValidation.ipynb
  Purpose: Formalizes structural assumptions as Directed Acyclic Graphs (DAGs) and validates them using DoWhy and domain knowledge.
  Related Chapter: Chapter 6: Methodology and Results 1: Conventional Statistics vs Causal Inference

• Causal_assumptions_checks.ipynb
  Purpose: Implements checks for key assumptions such as ignorability, overlap, and consistency using graphical and empirical methods.
  Related Chapter: Chapter 6: Methodology and Results 1: Conventional Statistics vs Causal Inference

---

Estimation & Simulation

• CausalAI_DoWhy_EconML_masters_.ipynb
  Purpose: Estimates Average Treatment Effects (ATE), Conditional ATE (CATE), and Individual Treatment Effects (ITE) using DoWhy + EconML meta-learners on linear and nonlinear datasets.
  Related Chapter: Chapter 6: Methodology and Results 1: Conventional Statistics vs Causal Inference

• SiMMMulator_GCM_DoWhy_EconML.ipynb
  Purpose: Simulates marketing data using siMMMulator, then applies multiple causal inference pipelines (DoWhy, GCM, EconML) to estimate treatment effects.
  Related Chapter: Chapter 6: Methodology and Results 1: Conventional Statistics vs Causal Inference

• SiMMMulator.R
  Purpose: R script for generating synthetic marketing data (adstock, saturation, noise, etc.) with known causal structure using the siMMMulator package. The actual dataset generated is also included in the repo.
  Related Chapter: Chapter 6: Methodology and Results 1: Conventional Statistics vs Causal Inference

---

Root Cause & Causal Discovery

• DoWhy_GCM_RootCauseMasters_.ipynb
  Purpose: Explores root-cause attribution and anomaly explanation using DoWhy-GCM methods such as arrow strength, intrinsic contribution, and counterfactual change attribution.
  Related Chapter: Chapter 7: Methodology and Results 2: Graphical Causal Models

• CausicaMasters (1).ipynb
  Purpose: Trains Microsoft's Causica model to learn latent causal structure and infer treatment effects using probabilistic graphical models.
  Related Chapter: Chapter 7: Methodology and Results 2: Graphical Causal Models

---

Educational Examples & Classic Cases

• Simpson_Paradox_Berkeley_Admissions.ipynb
  Purpose: Analysis of Simpson's Paradox applied to the famous Berkeley admissions scandal, demonstrating how aggregate statistics can mislead without proper causal reasoning and the importance of controlling for confounders.
  Related Chapter: Chapter 3: Correlation and Causation / Chapter 4: Frameworks for Causal Inference

---

Repository Structure

├── README.md
├── notebooks/
│   ├── CausalGraphValidation.ipynb
│   ├── Causal_assumptions_checks.ipynb
│   ├── CausalAI_DoWhy_EconML_masters_.ipynb
│   ├── SiMMMulator_GCM_DoWhy_EconML.ipynb
│   ├── DoWhy_GCM_RootCauseMasters_.ipynb
│   ├── CausicaMasters (1).ipynb
│   └── Simpson_Paradox_Berkeley_Admissions.ipynb
├── data/
│   └── [Generated siMMMulator datasets]
└── scripts/
    └── SiMMMulator.R

Key Research Questions Addressed

1. RQ1: Assumption validation & diagnostics
2. RQ2: Structural discovery and causal graph learning
3. RQ3: Effect estimation (ATE/ITE) accuracy
4. RQ4: Intervention simulation ("what-if" futures)
5. RQ5: Counterfactual reconstruction ("alternative pasts")
6. RQ6: Root-cause analysis of performance shifts

Technologies Used

• Python: DoWhy, EconML, Causica, DoWhy-GCM
• R: siMMMulator package for synthetic data generation