Berkson’s paradox is what happens when a filter creates a fake anti-correlation.
The raw intuition is simple: once you only look at people who passed through a gate, strength on one dimension reduces how much strength was needed on the other.
Simple Picture
Imagine a school that admits students who are either very strong at math or very strong at writing.
Inside the admitted class, math and writing can look negatively correlated. The student who got in with extraordinary math ability no longer needed extraordinary writing ability, and vice versa.
That does not mean math skill causes low writing skill. It means the admission filter manufactured the pattern.
Why It Happens
The trap appears when you condition on a combined outcome:
- getting admitted
- getting hired
- showing up in a hospital sample
- becoming notable enough to be observed
Once selection depends on multiple routes to success, the selected group will often make those routes look like substitutes.
ELI5: if either key opens the door, then among the people inside, having one key makes it less necessary to have the other.
Correlation Becomes Anti-Correlation At The Extremes
This is why “correlation becomes anti-correlation at the extremes” is often a warning about filters rather than about the world itself.
At the top of a selected distribution, people may look like they trade off against each other even when the underlying traits are neutral or positively related in the full population.
Common Misread
The dimwit take is “these two traits are opposed.”
The midwit take is “the selected sample reveals the deepest truth because it contains the important cases.”
The better take is that selected samples are often the most misleading ones. The more aggressive the filter, the more likely you are seeing the logic of the gate instead of the logic of reality. A related trap is reference-point-bias, where the “filter” is simply your own position — you condition on where you stand and read everything else as deviation.
Main Payoff
Berkson’s paradox is a reminder to ask one question before believing an anti-correlation:
What process decided who ended up in the sample?
If the sample was created by a gate that either trait could satisfy, the anti-correlation may be mostly an artifact of selection.