Introduction to Causal Inference



Causal Inference

  • moving beyond simple connections
Introduction to Causal Inference

Watch: Correlation and Causation

Introduction to Causal Inference

The Correlation Trap

center

Introduction to Causal Inference

The Correlation Trap

  • The Observation: On days when ice cream sales spike, the number of drownings goes up
  • The Logical Error: Thinking that eating ice cream causes drowning
  • The Fix: Using causal inference to dig deeper and find hidden forces
Introduction to Causal Inference

Ladder of Causation

  • cause-and-effect thinking
    • visualized as
      • three-rung ladder
Introduction to Causal Inference

Rung 1: Association

  • Nature: Pure observation and finding patterns
  • Question: "How are these things related?"
  • Example: Noticing ice cream sales and drownings show up together
  • Utility: Great for prediction, but it doesn't tell you why
Introduction to Causal Inference

Rung 2: Intervention

  • Nature: Active participation through experiments
  • Question: "What happens if we act?"
  • The Ice Cream Test: If we banned ice cream, would drownings stop?
  • Discovery: Intervening breaks the correlation and reveals the hidden variable: hot weather
Introduction to Causal Inference

Rung 3: Counterfactuals

  • Nature: The highest form of causal thinking; imagining things that didn't happen
  • Question: "What if we had acted differently?"
  • Humanity: This level allows us to explore alternate versions of the past to understand the present
Introduction to Causal Inference

Case Study: WWII Bomber Planes

  • The Problem: Allied planes were being shot down; armor was needed but could only be placed in specific areas due to weight
  • Rung 1 Analysis: Engineers mapped bullet holes on returning planes and wanted to armor the wings and tail because that's where the hits were seen
Introduction to Causal Inference

Bomber bullet holes map

center

credit: By Martin Grandjean (vector), McGeddon (picture), US Air Force (hit plot concept) - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=102017718

Introduction to Causal Inference

Abraham Wald's Counterfactual

  • The Rung 3 Question: Wald asked, "Where are the missing bullet holes?"
  • Survivor Bias: The data only showed planes that made it back
  • The Genius Insight: Any plane hit in the engines or cockpit never returned to be counted
  • The Solution: Armor the places where you see no damage on survivors

center

Introduction to Causal Inference

Key Takeaways

  • Beyond the Data: Real wisdom requires looking at information the data isn't telling you
  • Discussion Questions:
    • What are the "missing bullet holes" in your own life or studies?
    • Whose stories are you not hearing at all?
Introduction to Causal Inference

References

Topic Source
Molak, A. (2023) Molak, A. (2023). Causal inference and discovery in Python: Unlock the secrets of modern causal machine learning with DoWhy, EconML, PyTorch and more. Packt Publishing.
Wikipedia contributors (2026) Wikipedia contributors. (2026, January 25). Abraham Wald. Wikipedia.
Vigen, T. (n.d.) Vigen, T. (n.d.). Spurious correlations.

⏱️ Slide Timing: 1 min - Welcome the group and frame the session: correlation is everywhere, causation is rare - Preview the ladder of causation as the throughline for the deck - Set expectation: one classic case study (WWII bombers) anchors the ideas

⏱️ Slide Timing: 2 min - Play this before or after the ice cream example — reinforces the same lesson from a different angle - Encourage the audience to note down their own "spurious correlation" examples while watching

⏱️ Slide Timing: 2 min - Let the chart sit for a moment before explaining — ask the audience what they notice first - The two lines track almost perfectly, which is exactly what makes the trap convincing ❓ Ask: "What do you think is really driving both of these numbers up?"

⏱️ Slide Timing: 3 min - Name the fallacy explicitly: this is a textbook confounding variable problem - Ice cream and drownings share a common cause rather than causing each other - Preview that the "hidden force" gets revealed once we start intervening, not just observing

⏱️ Slide Timing: 2 min - Framework from Judea Pearl, one of the founders of modern causal inference - Each rung requires a different kind of evidence — observation, action, or imagination - The rest of the deck walks up this ladder one rung at a time

⏱️ Slide Timing: 3 min - Most traditional statistics and ML models live entirely on this rung - A recommendation engine can predict what you'll click without knowing why you'll click it - Prediction is useful, but it breaks down the moment you try to change the underlying system

⏱️ Slide Timing: 3 min - This rung is where A/B tests and randomized controlled trials live - Banning ice cream is a thought experiment — nothing would happen to drownings, exposing the confounder - Hot weather drives both ice cream sales and swimming (and therefore drowning risk)

⏱️ Slide Timing: 3 min - Counterfactual reasoning is uniquely human — even advanced AI systems struggle here - Doctors use this rung constantly: "would the patient have recovered without the treatment?" - Bridges directly into the bomber case study, which is built entirely on counterfactual thinking

⏱️ Slide Timing: 3 min - Setup for one of the most famous applied statistics stories of the 20th century - Weight limits meant armor was a zero-sum trade-off — every extra pound on the wings came off somewhere else - The engineers' instinct was reasonable, but purely observational (rung 1) reasoning

⏱️ Slide Timing: 2 min - Red dots mark where returning planes were hit — dense on wings and fuselage, sparse on engines and cockpit - Point out that this dataset only contains survivors, which is the crux of the problem

⏱️ Slide Timing: 4 min - Wald was a statistician at Columbia's Statistical Research Group during WWII - The counterfactual question flips the entire analysis: absence of data is itself data - This insight directly saved lives by redirecting armor to the engines and cockpit ❓ Ask: "Where else might absence of evidence actually be evidence of something?"

⏱️ Slide Timing: 6 min - Tie back to the ladder: association told the engineers where the holes were, counterfactual thinking told Wald where they weren't - Survivorship bias shows up constantly — in business case studies, hiring, even startup advice - Open the floor for the discussion questions and let a few people share examples

⏱️ Slide Timing: 1 min - Molak's book is the best next step for anyone wanting to apply this hands-on with DoWhy or EconML - Tyler Vigen's site is a fun way to keep building intuition for spotting spurious correlations