CausaLens

A Practical Framework for Graph-Based Causal Analysis and Adjustment Set Discovery

---

📌 Overview

CausaLens is a causal inference project designed to move beyond correlation-based analysis by explicitly modeling causal structure, identifying confounders, and estimating interventional effects using principled causal reasoning.

The project focuses on:

• Representing causal assumptions as directed graphs
• Identifying valid adjustment sets using the backdoor criterion
• Preventing common causal errors such as conditioning on colliders or post-treatment variables
• Estimating causal effects under explicit assumptions

This repository is centered around an exploratory, notebook-based workflow that emphasizes interpretability, correctness, and causal validity.

---

🎯 Motivation

In many applied ML and data science problems (e.g., churn prediction, delivery delays, product quality analysis), models answer:

"What is correlated with the outcome?"

But decision-making requires a stronger question:

"What causes the outcome?"

Without causal reasoning:

• Confounders bias estimates
• Interventions fail in the real world
• Models break under distribution shifts

CausaLens provides a structured way to:

1. Explicitly encode causal assumptions
2. Diagnose confounding paths
3. Select correct controls
4. Estimate effects that correspond to real-world interventions

---

🧠 Core Causal Concepts Used

This project relies on foundational concepts from causal inference:

1. Directed Acyclic Graphs (DAGs)

• Represent causal relationships as directed edges (A → B means "A causes B")
• Nodes are variables; edges encode causal assumptions
• Used to visualize confounding paths and causal flows

2. Backdoor Criterion

• Identifies which variables to control for to block confounding
• A set of variables Z satisfies the backdoor criterion if:
  • No node in Z is a descendant of treatment
  • Z blocks all backdoor paths from treatment to outcome

3. Adjustment Sets

• The minimal set of variables needed to obtain unbiased causal estimates
• Derived from DAG structure using graphical criteria
• Controls for confounders while avoiding colliders and mediators

4. Confounders vs Colliders

• Confounder: Affects both treatment and outcome → must adjust for it
• Collider: Caused by both treatment and outcome → must NOT adjust for it
• Mediator: On the causal path (treatment → mediator → outcome) → adjusting blocks the effect

5. Temporal Ordering

• Causes must precede effects in time
• Post-treatment variables cannot be confounders
• Ensures causal direction is valid

---

🔬 Project Workflow

Step 1: Data Preprocessing

Goal: Transform raw e-commerce data into a causally valid analysis-ready dataset.

What We Do:

1. Import Multi-Table Data from Kaggle

   • Load four tables: orders, reviews, order_items, customers
   • Focus on order-level analysis (one row per order)

2. Parse Timestamps

   • Convert all date columns to consistent datetime format using pd.to_datetime()
   • Ensures temporal ordering can be validated

3. Create Intervention Variable (Treatment)

   • delayed_delivery_days = order_delivered_customer_date - order_estimated_delivery_date
   • delayed_delivery = binary indicator (1 if delay > 0 days, else 0)
   • This is the treatment whose effect we want to estimate

4. Merge Outcome Variable

   • Join reviews table to get review_score (1-5 stars)
   • This is the outcome we want to explain causally
   • Keep only one review per order (drop duplicates)

5. Derive Order-Level Features (Pre-Treatment Confounders)

   • order_price: Total cost of all items in order
   • freight_value: Shipping cost
   • num_items: Number of products ordered
   • These affect both likelihood of delay AND customer satisfaction

6. Derive Customer History Features (Pre-Treatment Confounders)

   • customer_tenure_days: Days since first purchase
   • prior_orders: Number of previous orders (using .cumcount() for temporal validity)
   • is_first_order: Binary flag for first-time customers
   • These capture customer loyalty and experience effects

7. Apply Validity Filters

   • Keep only delivered orders with valid timestamps and reviews
   • Remove incomplete or invalid observations

8. Save Preprocessed Dataset

   • Export clean CSV with treatment, outcome, and confounders
   • Ready for causal discovery and estimation

Output: olist_preprocessed_causal.csv (~96k orders, 12 columns)

---

Step 2: Causal Discovery & DAG Construction

Goal: Discover causal structure from data and identify which variables confound the treatment-outcome relationship.

What We Do:

1. Run Three Causal Discovery Algorithms

   • PC (Constraint-Based): Tests conditional independence using Fisher's Z test
   • GES (Score-Based): Optimizes Bayesian Information Criterion (BIC)
   • NOTEARS (Continuous Optimization): Uses gradient descent with acyclicity constraint
   • Each algorithm proposes edges based on different statistical principles

2. Sample Data for Efficiency

   • Use 10,000 random orders (from ~96k) to reduce computational cost
   • Causal structure learning doesn't require full data to detect relationships
   • Standardize variables (mean=0, std=1) for numerical stability

3. Build Consensus DAG (Conservative Approach)

   • Keep only edges agreed upon by ≥2 out of 3 algorithms
   • This reduces false positives and improves robustness
   • Pure data-driven structure discovery

4. Enforce Temporal Constraints

   • Remove edges that violate time ordering:
     • ❌ review_score → delayed_delivery (future cannot cause past)
     • ❌ delayed_delivery → order_price (treatment cannot cause pre-treatment)
   • Keep only valid temporal flows:
     • ✅ order_price → delayed_delivery (past → present)
     • ✅ delayed_delivery → review_score (present → future)
     • ✅ order_price → review_score (past → future)

5. Incorporate Domain Knowledge

   • Algorithms may miss important confounders due to sample size or weak signals
   • Manually add theoretically justified edges:
     • order_price → delayed_delivery (expensive orders → special handling → delays)
     • prior_orders → delayed_delivery (loyal customers → priority shipping)
     • is_first_order → delayed_delivery (first-timers → default slower tier)
   • Hybrid approach: data + domain expertise

6. Identify Adjustment Set (Backdoor Criterion)

   • Find variables that create backdoor paths from treatment to outcome
   • Backdoor path example:

     delayed_delivery ← order_price → review_score
     

     (Order price affects both delay likelihood AND review score → confounding)
   • Adjustment set: Variables we must control for to block these paths
   • In this project: [order_price, prior_orders, is_first_order, num_items]

7. Visualize Final DAG

   • Create graph with:
     • Red node: Treatment (delayed_delivery)
     • Teal node: Outcome (review_score)
     • Green nodes: Pre-treatment confounders
   • Save as high-resolution PNG for reporting

Output:

• causal_dag_edges.csv: List of directed edges in final DAG
• adjustment_set.txt: Variables to condition on in Step 3
• causal_dag_visual.png: DAG visualization

Why This Matters:

• Without this step, you'd either:
  • Control for nothing → biased by confounding
  • Control for everything → introduce collider bias or overadjustment
• The adjustment set tells you exactly which variables to include in regression

---

Step 3: Causal Effect Estimation

Goal: Estimate the unbiased causal effect of delivery delays on customer satisfaction.

What We Do:

1. Baseline: Naive Analysis

   • Simple comparison of delayed vs on-time orders without adjustments
   • Purpose: Demonstrates confounding bias when causal methods aren't used
   • Shows overestimated effect due to unmeasured confounding

2. Causal Estimation Methods We apply three complementary approaches (Regression Adjustment, Inverse Propensity Weighting, and Matching) to ensure robustness.

3. Triangulation

   • Compare estimates across all three methods
   • Agreement across methods → high confidence in causal estimate
   • Disagreement → signals potential assumption violations

Output:

• causal_effect_estimates.csv: Effect estimates from all methods
• Typical finding: Naive effect ≈ -1.73 stars, Causal estimates ≈ -1.74 stars (minimal confounding detected in this dataset)

Interpretation:

• "Delivery delays cause approximately a -1.7 star reduction in customer reviews, after accounting for order characteristics and customer history"
• This represents the interventional effect (expected impact if delays were eliminated)

---

Step 4: Heterogeneous Effects (Subgroup Analysis)

Goal: Determine if the causal effect varies across different customer segments.

What We Do:

1. Stratified Analysis by Customer Type

   • Split data into subgroups: first-time vs repeat customers
   • Estimate delay effect within each group separately
   • Compare: Do first-timers suffer more from delays than loyal customers?

2. Analysis by Order Value

   • Create categories: low, medium, high value orders (using quantiles)
   • Estimate effect within each value tier
   • Check if expensive orders are more sensitive to delays

3. Analysis by Customer Tenure

   • Categories: new, established, loyal customers
   • Test if relationship with company affects delay tolerance

4. Statistical Interaction Test

   • Regression with interaction term: delayed_delivery × is_first_order
   • Tests if effect modification is statistically significant
   • Coefficient = difference in delay effect between subgroups

Typical Finding:

• Interaction coefficient ≈ -0.15 (not significant, p > 0.05)
• Conclusion: Delay effects are uniform across customer types
• Business implication: No need for segment-specific interventions; universal on-time delivery improvement helps all customers equally

Why This Matters:

• Identifies high-value intervention targets (e.g., "prioritize first-timers")
• Or confirms effect is stable (one-size-fits-all strategy works)
• Prevents wasted resources on ineffective targeting

---

📊 Key Results

Main Causal Effect

• Treatment: Delivery delay (binary: on-time vs delayed)
• Outcome: Customer review score (1-5 stars)
• Estimated Effect: -1.73 stars (95% CI: [-1.76, -1.70])
• Interpretation: Delays cause customers to rate ~1.7 stars lower on average

Confounder Structure

Variables that affect BOTH delay likelihood AND review scores:

• order_price: Expensive orders get delayed more + rated differently
• prior_orders: Loyal customers get priority + are more forgiving
• is_first_order: First-timers get slower shipping + rate harshly
• num_items: More items → longer processing + affects satisfaction

Subgroup Findings

• No significant heterogeneity detected (interaction p > 0.05)
• Delay effects are consistent across:
  • First-time vs repeat customers
  • Low vs high value orders
  • New vs loyal customers
• Implication: Universal improvement strategy is optimal

---

🛠️ Technologies Used

Core Libraries

• Causal Discovery: causal-learn, gcastle (PC, GES, NOTEARS algorithms)
• Graph Analysis: networkx (DAG manipulation and visualization)
• Statistical Modeling: scikit-learn, statsmodels (regression, propensity scores)
• Data Processing: pandas, numpy

---