Mechanism versus Statistics

Mechanism versus statistics names the fundamental methodological and philosophical tension between explaining how a system works and describing what patterns it produces. The distinction is not merely a preference for depth over breadth; it is a disagreement about what constitutes scientific understanding.

Statistical approaches — regression, classification, prediction — seek to model the joint distribution of observable variables. They ask: given what we have seen, what is likely to happen next? Mechanistic approaches — causal modeling, simulation, structural analysis — seek to recover the generating process. They ask: what machine, running what rules, produced these patterns? Judea Pearl's ladder of causation formalizes this divide: association lives on the statistical floor; intervention and counterfactuals require mechanistic ascent.

The tension is visible across the sciences. In biology, population genetics describes statistical regularities in allele frequencies; molecular biology explains the mechanisms that produce them. In economics, macroeconomic aggregates are statistical summaries; microeconomic models are mechanistic stories. The causal intervention framework argues that statistical description, no matter how sophisticated, cannot answer mechanistic questions without additional structural assumptions.

The persistent elevation of statistical methods above mechanistic inquiry in fields from social science to machine learning is not a methodological choice. It is an ontological error — the mistake of treating the shadow as the thing that casts it.