Simpson's paradox
The next trap hides in the relationships between quantities themselves. Sometimes the trend visible in the data as a whole flips completely once you split it into groups. This is Simpson's paradox β it wrecks naive conclusions and regularly catches even seasoned analysts.
Before you is a cloud of points β say, employee tenure horizontally, salary vertically. Across all the points the line goes down: as if more tenure meant a lower salary. A strange conclusion.
The classic real example: a drug helped less overall, but better in every patient subgroup (mild and severe cases). The severe cases simply got it more often and dragged the overall result down.
That is why analysts almost reflexively slice data by key segments: device, region, age. The same chart "overall" and "by segment" can tell opposite stories.
The paradox has been caught in university admissions, hospital comparisons, ad effectiveness by channel. Everywhere the "average over everyone" misled until the data was split by group.
The lesson is simple: aggregates are convenient but dangerous. Always keep in mind which groups they are mixing.