You’ve heard it a thousand times: correlation isn’t causation. But what does that actually mean for the decisions you make every day?
The problem with correlational thinking
When we see two things move together, our brains naturally assume one causes the other. Sales went up after we launched the new campaign? The campaign must have worked. Churn increased after we raised prices? Clearly the price increase drove customers away.
But these conclusions can be dangerously wrong.
Three reasons correlation misleads
Confounding variables. Maybe sales went up because of seasonality, not your campaign. Maybe churn increased because a competitor launched a better product the same week you raised prices.
Reverse causation. Perhaps you raised prices because you noticed increased demand, and the customers who churned were going to leave anyway.
Selection bias. The customers who saw your campaign might be systematically different from those who didn’t, making any comparison meaningless.
What causal inference offers
Causal inference provides a rigorous framework for answering “what would have happened if we had done X instead of Y?”, even when you can’t run a randomised experiment.
The key tools include:
- Directed acyclic graphs (DAGs) to make your assumptions explicit
- Identification strategies to find natural experiments in your data
- Sensitivity analysis to understand how robust your conclusions are
A practical example
Suppose you want to know whether your loyalty programme actually reduces churn. The naive approach, comparing churn rates between members and non-members, is meaningless. Members are different in ways that correlate with both their decision to join and their likelihood to stay.
A causal approach might:
- Map out the factors that influence both programme membership and churn
- Identify a source of variation in membership that’s independent of churn propensity
- Estimate the causal effect using that variation
- Test how the conclusion changes under different assumptions
The result isn’t just a number. It’s a defensible answer to a question that matters.