Causation is a relationship in which a change in one factor brings about a change in another. Statistical evidence or predictive usefulness alone does not establish that relationship. section 2.6 section 9.6.2
Keep these inferences distinct:
- Bayesian inference updates beliefs about explanations; establishing causation typically also requires assumptions, carefully designed experiments, or supporting scientific knowledge. section 2.6
- Feature importance measures how much a model relies on a feature for prediction, not what would happen if that feature were changed through intervention. section 9.6.2
A predictive feature is not necessarily a cause
A feature may predict an outcome because it proxies for another factor, shares a common cause with the outcome, or occurs after the predicted process. A causally meaningful feature may also receive a low importance score. section 9.6.2








