We never step in the same river twice. Heraclitus

Statistical versus causal explanations

We stated in the previous chapter that we look at causation as a primary concept and at correlation as a derived one. It is useful to explain in some detail what we mean. Doing so will also help the reader understand why we regard Bayesian nets as more than a tool for factorizing efficiently complex joint probabilities. We think that such a factorization view of Bayesian nets, powerful as it is, is too reductive, and misses their real potential. The ability afforded by Bayesian nets to ‘represent’ (conditional) independence in a transparent and intuitive way is only one of their strengths. The real power of Bayesian nets stems from their ability to describe causal links among variables in a parsimonious and flexible manner. See Pearl (1986, 2009) for a thorough discussion of these points. To use his terminology, casting our treatment in terms of causation will make our judgements about (conditional) (in)dependence ‘robust’; will make them well suited to represent and respond to changes in the external environment; will allow us to work with conceptual tools which are more ‘stable’ than probabilities; will permit extrapolation to situations or combination of events that have not occurred in history.