Why does maternal smoking look protective among low-birth-weight infants?

## The observation In birth cohort data, babies born at low birth weight to mothers who smoked during pregnancy have *lower* infant mortality than low-birth-weight babies of non-smoking mothers. Taken at face value the finding says smoking protects small babies, which contradicts everything else known about smoking in pregnancy - including the fact that in the same data, over the whole cohort, smoking raises infant mortality. That contradiction is the interview question, and the answer is structural. ## The structure that produces it Write down what causes what: - Maternal smoking lowers birth weight. - Some unmeasured set of serious conditions - chromosomal abnormalities, severe placental insufficiency, undiagnosed infection - also lowers birth weight. - Those same serious conditions strongly raise infant mortality. - Smoking also raises mortality, through its own routes. Birth weight is therefore a **common effect** of smoking and of the unmeasured serious causes: a collider. Conditioning on it - which is exactly what 'among low-birth-weight infants' means - induces an association between smoking and those unmeasured causes that did not exist in the full cohort. The direction follows from explaining-away. A baby is small; something made it small. If the mother smoked, smoking is a sufficient explanation. If she did not, the smallness demands a different explanation, and the available explanations are the serious pathologies. So the non-smoking mothers' small babies are enriched for lethal conditions, and the smoking mothers' small babies are, comparatively, healthy babies who are merely small. The mortality comparison inside the stratum reflects that compositional difference, not any protective action of tobacco. Note what is *not* going on. The unmeasured serious conditions are not confounders of smoking and mortality in the whole cohort: there is no reason for mothers who smoke to have more chromosomal abnormalities. Unstratified, the analysis is fine. The bias is created entirely by the act of stratifying. ## The same shape elsewhere The **obesity paradox** - obesity appearing protective in studies restricted to patients who already have a disease - has repeatedly been argued to be at least partly the same structure. Obesity raises the chance of having the disease, and so do unmeasured risk factors that independently raise mortality. Restrict to patients with the disease, and the obese patients are enriched for 'got the disease because of weight', while the non-obese patients are enriched for 'got the disease because of something worse'. The generic name for this family is **index-event bias**: any analysis conditioned on units having already reached some state that the exposure helped cause. The recognisable warning sign in all of them is a reversal of a well-established direction that appears only inside a subgroup defined by a variable the exposure affects. ## What to do instead - **Estimate the effect in the full cohort.** Total infant mortality by smoking status is the answerable, decision-relevant question. Do not define the study population using a downstream consequence of the exposure. - **Treat 'we only looked at cases with X' as a modelling decision, not a neutral scoping choice.** If X is caused by the exposure, that filter is a conditioning step with statistical consequences. - **If a within-stratum estimand is genuinely wanted** - say, clinical decisions for babies who are already small - be explicit that the quantity is a different one, that it is not the causal effect of smoking, and that estimating it credibly needs assumptions about the unmeasured causes, ideally probed with a sensitivity analysis over how strong those causes could be. - **Do not fix it with more covariates inside the stratum.** Adjusting for the measured variables does not recover the induced association with the unmeasured ones. The remedy is to change the population, not the specification. ## How to say it in an interview The compact answer is three sentences: birth weight is a common effect of smoking and of unmeasured severe pathology; restricting to low birth weight makes those two explanations compete, so non-smoking mothers' small babies carry worse hidden causes; the apparent protection is an artefact of the restriction, and the whole-cohort analysis shows smoking increasing mortality as expected. Interviewers are checking that you reach for a structural explanation rather than proposing a measurement error or a new biological mechanism.