Causality and correlation

Correlation means two things move together. Causation means intervening on one changes the other. UCLA’s DAG intro puts it sharply: “Are coffee drinkers more likely to have pancreatic cancer?” is not the same question as “If I stop drinking coffee, will my risk decrease?”

AI systems often exploit correlation. Business and policy decisions need causal care—shipping a model that predicts well can still fail when you *act*, because action changes the world the correlations came from.

Ask: if I *change* X, does Y change? Watch for confounders (hidden Z that drives both). Randomized experiments are a gold-standard way to break confounding; when you can’t randomize, you need explicit assumptions (graphs, potential outcomes) and identification strategies. Hernán & Robins’ free book *Causal Inference: What If* teaches potential-outcomes methods in increasing difficulty; Pearl’s *Causality* is the canonical structural/DAG tradition.

Ice cream sales and sunburns rise together—summer confounds them. Banning ice cream won’t fix burns. UCLA’s coffee–pancreatic-cancer story: smoking confounded coffee.

Pick a workplace metric correlated with “success.” Name one confounder and one intervention you’d test (even hypothetically).

⚠️ Myth: Big data removes the need for causal thinking.

✓ Reality: Bigger correlated piles can still mislead interventions (UCLA: associations alone don’t answer causal questions).

⚠️ Myth: Feature importance proves causation.

✓ Reality: Importance is about predictive role under a model class—not causal proof.

• Hernán & Robins *What If*: https://miguelhernan.org/whatifbook ↗
• Pearl *Causality* (Cambridge): https://www.cambridge.org/core/books/causality/B0046844FAE10CBF274D4ACBDAEB5F5B ↗
• UCLA Intro to DAGs: https://stats.oarc.ucla.edu/wp-content/uploads/2025/06/Intro-do-DAGs-UCLA-OARC.html ↗
• Pearl probabilistic calculus of actions (do-operator): https://arxiv.org/abs/1302.6835 ↗