Correlation vs causation in business decisions

Imagine a plant where batches made on the night shift fail inspection more often. The dashboard is clear, the correlation is strong, and the instinct is to intervene on the night shift.

Now add what the dashboard does not show. Night shifts run at a higher line speed, use a different mix of trained operators and handle a different product. Each of these could explain the defects. If the line speed is the real driver, retraining the night shift moves nothing.

A confounder is a variable that influences both the thing you are changing and the outcome you care about. It creates an association between them that does not reflect any direct effect.

Confounders are why “the customers who got the offer churned less” does not mean the offer worked. If the offer went to your most loyal customers, loyalty explains the result, and extending the offer to everyone would not reproduce it.

A few questions help separate a real lever from a coincidence.

The underlying issue is that the causal story usually exists only in someone's head. A causal graph writes it down: which variables exist, which influence which, and which are measured. Once it is on the page, a colleague can disagree with a specific arrow, which is a far better argument than disagreeing with a conclusion.

A written graph also lets you check something a spreadsheet cannot: whether the question you are asking can be answered from the structure at all. Sometimes the honest result is that an unmeasured common cause leaves the path open and no amount of data from this system will close it.

Before the next budget decision, write down the intervention, the outcome and everything you think influences both. If that list is empty, you have probably not looked hard enough. If it is long and some items are unmeasured, you have found the part of the decision most worth discussing.