Week 3: Causal Machine Learning

Core Concepts

• The difference between correlation and causation
• Understanding confounders
• Four types of treatment effects
• Why causal analysis matters for business decisions

Practical Tools

• SHAP for explaining predictions
• EconML for finding causes
• Real business applications

The Key Difference

Prediction tells you WHAT will happen

Prescription tells you HOW to change it

Business value comes from changing outcomes, not just predicting them

Key Lesson

Just because two things happen together doesn't mean one causes the other. There might be a third factor causing both.

What Do You Notice?

🟢 Young people (20s-30s): Exercise MORE (4-5 hrs) → Health Score HIGHER (80-85)

🟡 Middle age (40s-50s): Exercise LESS (2-3 hrs) → Health Score MEDIUM (68-72)

🔴 Older people (60s+): Exercise LEAST (0.5-1.5 hrs) → Health Score LOWER (55-62)

The Question: Does exercise cause better health? Or does AGE affect BOTH exercise habits AND health?

Definition

A confounder is a hidden influencer that affects both your action AND your result, making you think one causes the other when it doesn't.

Without Controlling for Age

Young people exercise more AND are healthier

Conclusion: "Exercise causes better health!" (Partially true, but age is hiding part of the story)

With Age Controlled

Compare 30-year-olds who exercise vs. don't exercise

Compare 60-year-olds who exercise vs. don't exercise

Now we see the TRUE effect of exercise at each age level

The Right Approach

Identify what makes customers engaged, then work on increasing engagement rather than just forcing review views.

What is a "Treatment"?

Any action, policy, or intervention you might take:

• Sending a discount email
• Offering free shipping
• Hiring more staff
• Launching a new product feature

What is a "Treatment Effect"?

The actual impact that action has on your outcome (sales, satisfaction, retention, etc.)

Scenario 3: Natural Experiment (Policy Changes)

External event creates treatment/control groups

• Example: Government policy applies to one region but not another
• Compare regions before and after policy change
• Useful when experiments are impossible or unethical

1. Average Treatment Effect (ATE)

Question: "Does it work overall?"

Measures: Impact across EVERYONE

Example: "Free shipping increases average order value by $12"

2. Heterogeneous Treatment Effect (HTE)

Question: "Does it work differently for different groups?"

Measures: How impact varies by group

Example: "Free shipping increases orders by 30% for students but decreases orders by 5% for seniors"

3. Local Average Treatment Effect (LATE)

Question: "Does it work for those actually affected?"

Measures: Impact only on those who changed behavior

Example: "For customers who used the free shipping offer, orders increased by $25"

4. Conditional Average Treatment Effect (CATE)

Question: "Does it work for people like YOU?"

Measures: Impact for specific individual characteristics

Example: "For 25-year-old female customers in Sydney who shop on weekends, free shipping increases orders by $18"

Business Value

Instead of one recommendation for everyone (ATE), personalize based on characteristics (CATE)

Result: Higher engagement, lower churn, more subscription renewals

The Historical Context

• 1970s Boston housing data
• Shows how correlation can mislead
• Demonstrates why causal analysis is essential for ethical AI

Why This Matters for Business Analytics

Predictive models can perpetuate bias and historical injustice

Causal models help us understand what we can actually change

Why This Matters

• Prediction tells you WHAT will happen
• Causation tells you HOW to change it
• Business value comes from changing outcomes, not just predicting them

Using Them Together

1. Use SHAP to understand your predictive model
2. Use EconML to find what you can actually change
3. Make data-driven decisions with confidence

Step 1: Linear Regression (15 minutes)

• Build intuition with simple model
• Understand basic relationships

Step 2: Predictive ML + SHAP (25 minutes)

• Train a Random Forest model
• Use SHAP to find strongest predictors
• Interpret correlation-based insights

Step 3: Causal ML + EconML (25 minutes)

• Apply Double Machine Learning
• Identify true causal factors
• Compare with predictive insights

Step 4: Business Recommendations (10 minutes)

• What actions would you take?
• What would you change first?
• How would you measure success?

What is SHAP?

SHAP (SHapley Additive exPlanations) explains ML predictions by showing how much each feature contributed to a specific prediction.

Based on game theory: distributes "credit" for a prediction fairly among all features.

1️⃣ Feature Importance Bar Chart

What it shows: Overall ranking of features by average absolute impact

Business use: "Which 5 factors matter most for customer churn?"

Contract Status     ████████████████ 0.42
Service Calls       ███████████      0.31
Monthly Charges     █████████        0.25
Account Age         ██████           0.18
Data Usage          ████             0.12

Higher bar = stronger correlation with prediction

How the Model Got There

🎯 The Ideal Workflow

1. Step 1: Build predictive ML model (LightGBM, XGBoost, etc.)
2. Step 2: Use SHAP to understand correlations and generate hypotheses
3. Step 3: Use EconML to test causal hypotheses and estimate treatment effects
4. Step 4: Make business decisions based on CAUSAL effects, not correlations
5. Step 5: Use CATE to personalize actions for different customer segments

🎓 Key Concepts Mastered Today

• Confounders bias treatment effect estimates if not controlled
• Treatment effects (ATE, HTE, CATE, LATE) answer different business questions
• SHAP shows correlation - which features predict outcomes
• EconML shows causation - which features cause outcomes
• Use causal methods (not SHAP alone) when making business decisions

💻 Today's Jupyter Notebook Exercise

You'll apply SHAP and EconML to the TeleConnect churn dataset and see:

• Which features are strong predictors (SHAP)
• Which features have true causal effects (EconML)
• Why correlation ≠ causation with real visualizations
• How to make actionable business recommendations

💡 Remember the Core Principle

Prediction tells you what will happen.

Causation tells you how to change it.

Real business value comes from taking the RIGHT actions, not just predicting outcomes.