An association may arise in your dataset precisely because you selected on a common effect.
Berkson’s Paradox is a selection problem now mainly understood as collider bias. Two factors that are not negatively associated in the target population may become negatively associated within a selected dataset because both influence inclusion.
An association may arise in your dataset precisely because you selected on a common effect.
Berkson showed that hospital data can produce spurious associations between conditions when both conditions influence hospitalisation. Conditioning on admission statistically connects the causes.
Collider bias is an established mechanism in causal inference. Whether a variable is actually a collider depends on the causal structure; selection alone is not enough.
Conditioning on a common effect can induce, mask or reverse an association between its causes. The selection variable is part of the data-generating process.
AI transfer to investigate: AI trained on selected organisational traces may learn collider-created relationships as generalisable knowledge. AI carries the mechanism; it does not originate it.
Proposed test, not yet conducted: Generate target-population data without an association, select records through a collider and train models on that subset. Test on separate, representative data from the original target population and compare with selection-aware designs.
Provenance must also include the inclusion mechanism: why does this record exist in the dataset?
A collider is a variable affected by both factors under study. Selecting on this common effect can create an association; selection alone does not establish the correct causal model.
Berkson (1946) — pp. 47–53; full text unavailable; not used as passage evidence
Cole et al. (2010), Illustrating bias due to conditioning on a collider — abstract; pp. 417–420
The basic mechanism is discussed using the authors’ abstract of Cole et al. (2010). Berkson’s original derivation has not been assessed here from the full text. The AI application is a research hypothesis.
Help improve this case
We review your contribution before updating the text or sources.