Skip to content

Briefing · Research

· 3 MIN

Why mechanisms beat correlations

A correlation describes what moved together. A mechanism explains why, and therefore what happens when conditions change.

A correlation survives until conditions change.

Correlation is cheap and useful: it finds patterns faster than theory does. It is also silent about what produced the pattern, so it cannot say whether acting on it will work, or whether the relationship will still hold next year.

Correlation-based claims fail quietly. Nothing announces the moment the conditions that produced the pattern no longer apply.

That is not an argument against statistics. It is an argument about what a pattern can support: description, prioritisation, a hypothesis worth testing, not a claim about what to change.

The useful question is therefore not how strong a pattern is, but what would have to remain true for it to keep holding.

What a mechanism adds.

A mechanism is a claim about how something works: which conditions activate a capability, what interactions produce the effect, and what else must be true for the result to appear. It is a stronger claim than co-movement, and it is testable.

Mechanism reasoning states what evidence would count against it. That is precisely what makes it usable in a decision.

It is also what makes transfer possible. A claim about interaction can be tested in a new setting; a pattern fitted to one dataset usually has to be rebuilt from the beginning.

Mechanism claims are more honest about their own limits, too: they name the conditions under which the explanation holds, which is what a decision-maker needs to know before acting.

EVIDENCEDECISION USE

Where evidence tests the model.

The operating record is the test. If a mechanism is present, its effects should be observable in what happened. If it is absent, capability inferred from confidence alone will not survive the records, the decisions or the observed response to disruption.

A model that cannot be shown to be wrong cannot support a claim. Evidence graded before it is combined is what keeps the question open long enough to answer it.

Grading is what lets the model fail usefully: a weak grade is a prompt for more evidence rather than a verdict.

Why decisions need the mechanism, not the pattern.

A decision is an intervention: something is changed in the hope that the result changes. Only a mechanism says what to change and what should follow. A pattern says what happened to co-move, which is a weaker guide to action.

This is also where agents help without deciding. Agents gather and reconcile the evidence, run the counterfactuals and draft the case; people decide whether the mechanism is credible enough to act on.

Mechanism reasoning also changes what counts as a good answer. A mechanism-based case can be argued with; a correlation-based case can only be accepted or ignored.

The cost of thinking in mechanisms.

Mechanism work is slower. It needs domain knowledge, a record that can be traced, and some tolerance for uncertainty where the evidence is thin. It resists the tidy summary.

It pays back by travelling: a mechanism identified in one setting can be tested in another, whereas a correlation fitted to a single dataset rarely survives the move.

It also makes a portfolio of decisions intelligible: the same mechanism can be tested in several settings, and each test adds to one body of evidence rather than starting again.

The practical compromise is to use patterns to find candidates and mechanisms to keep them: fast where the cost of being wrong is low, deliberate where the decision is consequential.

In short

Correlations describe; mechanisms explain. An intelligence system that can explain will still be believed after the conditions change, which is the only test that matters.

All insights

Research

The research behind the briefing.

turns research, industry expertise and verified outcomes into methods the engine can run.

Explore Research