Causal Inference

Causal inference is the process of determining cause-and-effect relationships from data, going beyond the observation that two things tend to occur together (correlation) to establishing that one actually produces the other (causation).

Correlation vs causation

This distinction is fundamental. Ice cream sales and drowning deaths both increase in summer — they are correlated. But ice cream does not cause drowning; both are caused by warm weather. Machine learning models excel at finding correlations but often cannot distinguish them from causation. A model might learn that umbrella purchases predict rain, but umbrellas do not cause rain.

Why AI needs causal inference

Standard machine learning models learn associations in data. This works well for prediction — if you want to predict which customers will churn, correlations are sufficient. But if you want to know what action will reduce churn, you need causation. A model might find that customers who contact support are more likely to churn, but that does not mean preventing them from contacting support would help.

Key approaches

• Randomised experiments (A/B tests): The gold standard. Randomly assign people to treatment and control groups and measure the difference.
• Instrumental variables: Use a third variable that affects the cause but not the outcome directly.
• Difference-in-differences: Compare changes over time between affected and unaffected groups.
• Causal graphs (DAGs): Map out the assumed causal structure and use it to identify which variables to control for.
• Do-calculus: A mathematical framework developed by Judea Pearl for reasoning about interventions.

Causal AI in practice

Causal inference is gaining importance as organisations move from AI that predicts to AI that recommends actions. Healthcare uses it to determine treatment effectiveness. Marketing uses it to measure campaign impact. Pricing teams use it to estimate demand elasticity.

Current limitations

Causal inference often requires assumptions about the data-generating process that are difficult to verify. No amount of observational data can definitively prove causation without some structural assumptions. The field is advancing rapidly but remains more challenging than standard prediction.