Confoundle · a reasoning trap

Simpson's paradox

One choice can win in every single group, yet lose the moment you lump all the groups together. It sounds impossible, but it's real. It happens when the groups aren't a fair comparison: one side quietly got the easy cases, the other got the hard ones. So the big combined number says one thing while the group-by-group numbers say the opposite, and it's the big number that fools you.

The rule

An overall trend can reverse once you account for a lurking variable that's split unevenly between the groups.

What it looks like

Treatment B cures more patients overall. Which would you pick?Two kidney-stone treatments, 350 patients each. On overall success rate, Treatment B comes out ahead. Same illness, same goal, one number to go on.
Treatment A actually wins, for both stone sizes.A and B weren't treating the same patients. A got mostly the hard cases (large stones), while B got mostly the easy ones. Everyone does worse on hard cases, so A's overall average sinks even though A wins in each group:

Why it works

The 'combined' score isn't a fresh measurement; it's the group scores blended together, and bigger groups count for more. When one side is packed with easy cases and the other with hard ones, that blend pulls their combined scores in opposite directions. So one option can lead in the easy group and in the hard group, yet still trail overall, because it handled most of the hard cases, and its blended score sits closer to that lower number. The cure is a fair split: give both sides the same mix of easy and hard cases (exactly what a randomised trial does), and the reversal can't happen.

Source

Julious SA, Mullee MA. Confounding and Simpson's paradox. BMJ. 1994;309(6967):1480-1481. Underlying clinical data: Charig CR, Webb DR, Payne SR, Wickham JEA. Br Med J (Clin Res Ed). 1986;292(6524):879-882.

See if it fools you

This page gives the answer away. The puzzle version shows you the same figures first and asks you to commit before the reveal.

Play this one