Thinking with evidence · Lesson 1 of 3

What an association actually supports

An association is a fact about the data. Which explanation it supports is a separate argument that has to be made.

9 min read · Free — no account needed to read

When two things vary together, there are several possible worlds and the data alone usually cannot tell them apart.

A causes B. B causes A — reverse causation, which is far more common than people expect. Something else causes both — a confounder. The association is an artefact of how the sample was selected. Or it is chance, which is what significance testing addresses and does not eliminate.

A confounder is a third factor associated with both. Ice cream sales and drownings rise together; temperature drives both. The example is famous because it is obvious. Real confounders are not obvious, which is the entire difficulty.

The ways researchers address this are worth recognising when you read: randomisation, which breaks the link between the treatment and everything else; statistical control, which adjusts for confounders you thought of and measured, and does nothing about the others; natural experiments, which use something outside the system to create variation; and longitudinal designs, which at least establish what came first.

So when you meet an association, the productive question is not "is this causal?" but "what would have to be false for this to be causal, and did anyone check?"

See it happen

Employees who use the company learning platform get promoted more often.

The read that arrives on its own

Learning drives advancement — so promote usage of the platform.

The same situation, with this idea

You generate the alternatives before accepting the first. Reverse: people who have been told they are on a promotion track are given time and encouragement to use it. Confounder: ambitious employees both use the platform and seek promotion, and the underlying trait drives both. Selection: the platform is rolled out first to teams already marked as high-potential. Each alternative suggests a different check, and none of them is ruled out by the data you have.

After this, you can

generate three alternative explanations before accepting the first one.

The takeaway

An association is a fact about the data. Which explanation it supports is a separate argument that has to be made.

Easy to take the wrong way

That "correlation is not causation" means associations are worthless. They are how nearly every causal question begins — the phrase is a reminder to keep going, not to stop.

Try this

Take an association you believe is causal and write three alternative explanations. For each, name the check that would rule it out.