Among low-birth-weight babies, those born to smokers have lower infant mortality than those born to non-smokers, although smoking raises mortality overall.
Yerushalmy reported the pattern in 1971, and it was read for decades as evidence that smoking somehow protected small babies. The resolution is selection. Restricting the comparison to low-birth-weight babies compares infants who are small for a known and relatively mild reason with infants who are small for unknown and often graver ones. Hernandez-Diaz, Schisterman and Hernan set it out in 2006 as the textbook case of collider stratification bias.
In practice
- A study of hospital patients finds smoking protective, because it compares only people already ill enough to be admitted.
- A firm finds that test scores predict nothing among its hires, because the low scorers who got through had something else going for them.
- Standing
- Settled rather than open: the pattern is an artefact of conditioning on a variable that both causes affect. Compare Berkson's paradox and the obesity paradox.
- Further reading
- Wikipedia
- Subject
- Science & Epistemology
- Essay
- Paradoxes in the data, and sayings about knowing