Data Curiosities
The same numbers, two opposite stories.
A growing set of small, interactive explainers for the ways data quietly fools us — each one a single idea you can drag a slider and watch happen, in plain English first, then grounded in a real study. Start anywhere.
A trend that flips when you zoom out
A pattern true in every group can reverse once you lump the groups together.
Exercise looks like it raises blood sugar — until you split by age and every band slopes the other way. With the real kidney-stone surgery case.
Explore →The hidden cause behind a fake link
Two things rise together only because a lurking third thing drives both.
Ice-cream sales "cause" drownings — until you remember it's hot out. Hold the weather fixed and the link melts away.
Explore →What's true of groups, but not of people
A pattern measured on whole groups need not hold for the individuals in them.
Robinson's 1950 puzzle: regions with more immigrants were more literate, while immigrants themselves were less so.
Explore →A trade-off conjured by the filter
Selecting a group can invent a negative link between things that are actually unrelated.
Why do talented actors seem less good-looking? They aren't — you only ever see the famous, and fame needs one or the other.
Explore →An effect hiding in plain sight
A real effect can stay invisible until you account for something working against it.
Practice looks useless — even harmful — for your test score, because keen learners attempt harder tasks. Control for difficulty and the payoff appears.
Explore →Two analysts, opposite conclusions
The same before-and-after data can show an effect, or none, depending on how you measure change.
Weigh students before and after a diet: compare the gains and it did nothing; adjust for the starting weight and it clearly worked.
Explore →When 'controlling for' backfires
Adjusting for the wrong variable can create bias rather than remove it.
Test prep looks unrelated to GPA — until you "control for" SAT, a collider, and conjure a link that isn't there.
Explore →An effect that changes size on adjustment
Some measures of effect shift when you account for a variable — even with no confounding at all.
An odds ratio of 2.0 in every patient group can read as 1.5 once you pool them — with nothing confounded.
Explore →