Applied Causal Inference & Structural Causal Models: Beyond Passive Correlation

Machine learning predicts $P(Y \mid X)$. Strategic decisions require $P(Y \mid \text{do}(X))$. An applied framework across DAGs, do-calculus, Meta-Learners, and real-world counterfactual evaluation.

---

📌 Executive Summary

Most data science applications fail at decision-making because they optimize for passive statistical correlation. In observational data, intervening on a variable based solely on predictive correlation introduces expensive policy errors due to confounders, colliders, and selection bias.

This repository is an applied causal inference framework bridging theoretical econometrics with production Python code:

• Implements the foundational Neyman-Rubin Potential Outcomes and Pearl's Structural Causal Models (SCMs).
• Deconstructs graphical separation (d-separation) to identify and block spurious non-causal paths.
• Compares advanced Meta-Learners (S/T/X-Learners) to estimate Conditional Average Treatment Effects (CATE).
• Applies counterfactual structural modeling to a real-world biological system: evaluating canopy management interventions on Vitis vinifera L. cv. Sangiovese quality.

---

🔍 The Causal Inference Architecture

The repository is structured into modular, interactive research notebooks progressing from foundational theory to empirical counterfactual estimation:

Applied-Causal-Inference/notebooks/
├── 01_potential_outcomes_economic.ipynb   # Rubin Causal Model: ATE, ATT, SUTVA & Selection Bias
├── 02_inference_with_pgmpy.ipynb          # Bayesian Networks: Conditional Independence & Belief Propagation
├── 03_d_separation.ipynb                  # Graphical Criteria: Chains, Forks, Colliders & M-Bias
├── 04_causal_inference_with_dowhy.ipynb   # The 4-Step DoWhy Pipeline: Model, Identify, Estimate, Refute
├── 05_metalearners.ipynb                  # Heterogeneous Treatment Effects: S-Learner, T-Learner, X-Learner
├── 06_Structural_causal_models.ipynb      # Pearl's do-calculus: Abduction, Action, and Counterfactual Queries
├── 07_scm_matrix_form.ipynb              # Linear SCMs: Covariance Decomposition & Path Coefficients
└── sangiovese.ipynb                       # Empirical Case Study: Viticulture Counterfactual Intervention

---

📐 Core Methodological Modules

1. The Identification-Estimation Decoupling (04_causal_inference_with_dowhy.ipynb)

Following Judea Pearl and Microsoft Research's DoWhy protocol, causal queries are strictly separated into 4 distinct phases:

1. Model: Encode assumptions explicitly into a Directed Acyclic Graph (DAG).
2. Identify: Prove mathematical identifiability using Backdoor, Frontdoor, or Instrumental Variable criteria.
3. Estimate: Fit propensity score matching, regression discontinuity, or IPW estimators.
4. Refute: Stress-test the effect using placebo treatments, random unobserved confounders, and subset validation.

2. Heterogeneous Treatment Effects via Meta-Learners (05_metalearners.ipynb)

Moving beyond population-average ATE to person-specific CATE:

• S-Learner (Single Model): Models treatment as a feature; prone to regularizing away treatment effects in high dimensions.
• T-Learner (Two Models): Trains separate models for control and treatment groups; avoids treatment attenuation.
• X-Learner (Cross-Learner): Optimized for extreme treatment imbalance ($N_1 \ll N_0$), computing imputed counterfactual residuals to maximize sample efficiency.

3. Empirical Case Study: Viticulture Counterfactuals (sangiovese.ipynb)

• Domain: Precision viticulture and enological optimization (referencing Magrini, Di Blasi & Stefanini, Biometrical Letters).
• Objective: Quantify how agronomic canopy interventions (bud load, defoliation, cluster thinning) and harvest timing causally alter must sugar concentration, total acidity, and anthocyanin profiles.
• Mechanism: Disentangles seasonal climatic confounders from direct human agronomic interventions to guide optimal vineyard yield management.

---

🛠️ Stack & Dependencies

• Language: Python 3.10+
• Causal Engines: dowhy (causal graph identification & refutation), pgmpy (Bayesian networks), econml / scikit-learn (meta-learners)
• Data & Visualization: networkx, matplotlib, seaborn, pandas, numpy

Quick Start

# Clone repository
git clone https://github.com/FRA-0023/Applied-Causal-Inference.git
cd Applied-Causal-Inference

# Install requirements
pip install dowhy pgmpy networkx scikit-learn pandas matplotlib

---

Author: Francesco Colombini
GitHub Profile · LinkedIn