Aggregate first, decide last
A pooled chart is a starting point. If segment membership affects both the input and the output, the pooled slope can reverse the real within-group relationship.
Free interactive learning tool
Simpson's Paradox Splitter
The same dataset can say "the metric is getting worse" in aggregate and "the metric is improving" inside every segment. Split the chart before you trust the headline.
Here's the pooled chart for an admissions trial: as investment rises, the overall success rate falls — the aggregate trend line slopes down. Now take those exact same points and split them into three segments (Enterprise, Mid-market, SMB). Which way does each segment slope?
🔒 Lock in your prediction above to unlock the split view.
Which conclusion survives?
Marketing says paid campaigns perform worse because the combined conversion rate is lower. Segmenting by device shows paid traffic beats organic on desktop and mobile.
Score 0 / 3
Pick the conclusion you would put in front of the team.
Ask what changed in the mix
Simpson's paradox usually appears when the composition moves: more hard cases, more low-intent traffic, more senior roles, or more customers in a different cohort.
Use segments responsibly
Do not slice forever until you get a story you like. Segment on plausible causes, inspect group sizes, and report both the aggregate and the meaningful subgroup view.
One line to carry out of here: a pooled trend is a hypothesis, not a verdict — when the segment mix moves together with the metric, the aggregate can point the opposite way from every group inside it, so split before you decide.