Berkson's Paradox, explained.
Berkson's paradox is a selection bias in which conditioning on a shared consequence of two variables can create an association absent from the original population.
Why it happens
If either strength is sufficient for selection, a selected person with a low value on one trait must often have a high value on the other. That constraint makes the selected sample look like a tradeoff even though the original traits were independent.
Berkson's paradox is a selection effect: conditioning on a common consequence of two variables can create an association between them even when they are independent in the original population.
Read the result
Toggle the shortlist and compare the dot pattern and correlation. No applicant's score changes. Only the rule for including applicants in the displayed sample changes.
A worked example
An either-skill shortlist
A pool contains every combination of writing and coding scores from one to ten equally often.
Admission requires writing at least eight OR coding at least eight, so low-low pairs disappear.
Among admitted applicants, low writing implies high coding and vice versa. The negative association is induced by selection.
OPTIONAL DEEPER DETAILGo deeper: inside the model
Inside this model
The 100 applicants form a 10-by-10 grid of independent writing and coding scores from 1 to 10. The shortlist admits an applicant if writing is at least 8 OR coding is at least 8. The scatterplot and Pearson correlation use only the currently shown rows.
Where this idea is useful
A practical use
If admission rewards either test skill or athletic skill, comparing admitted students can suggest a tradeoff that is absent from the full applicant pool.
A common misconception
“A negative shortlist correlation proves the skills conflict.”
The admission rule can create that pattern. Check the broader population and the selection process before asserting a causal tradeoff.
What this explanation leaves out
- The threshold and independent grid are invented. Real traits can already be related, and different selection rules can create different associations.
How does this differ from ordinary confounding?
Here the conditioning variable is influenced by the traits being compared. Conditioning on that common consequence can introduce a relationship rather than remove a preexisting common-cause influence.
Which rule determined who was allowed into the dataset?
Associated thinkers
Further reading
Explore the original research or the teaching reference behind this experiment.