Two variables move together. Who's leading?
Three sliders for three hypotheses: a confounder lurks behind, X causes Y directly, or Y actually causes X. Click an example — the sliders snap to the documented reality. Pearson correlation and causal structure (DAG) react live.
When ice-cream sales rise, drownings rise too. Does ice kill people?
Both rise in summer. Heat drives swimmers to water AND to ice-cream stands. Control for temperature — the correlation collapses.
Ice cream ↔ Drownings
- 1. Strength
How large is the effect? Big effects are harder to explain via confounding.
- 2. Consistency
Do different studies, countries, methods replicate?
- 3. Specificity
One cause, one effect? (Weighted less today.)
- 4. Temporality
Cause must precede effect. The one strictly necessary criterion.
- 5. Dose-response
More exposure → more effect?
- 6. Plausibility
Can a mechanism explain it with current knowledge?
- 7. Coherence
Does the finding fit the broader epidemiological and biological picture?
- 8. Experiment
Confirmed by intervention (RCT, quitting, vaccination)?
- 9. Analogy
Are there analogous known cause-effect relationships?
Why the hormone-therapy story rewrote the textbook.
Major studies like the Nurses' Health Study found HRT users had 40–50 % lower heart risk. Hormones were considered cardioprotective. Millions of prescriptions followed.
The Women's Health Initiative randomised 16,608 women — and stopped the trial early: HRT raised heart risk by 29 %, strokes by 41 %, thromboses by over 200 %. The confounder: "healthy-user bias" — HRT users were systematically healthier.
Lesson: even high-quality observational studies can get the direction wrong. Only randomisation severs all known and unknown confounders — the gold standard.
How do you know a vaccine really works?
Medicine's largest field trial
1.8 million children in the US, Canada, Finland — double-blind, placebo-controlled. Efficacy 60–90 % against paralytic polio. Randomisation severs all confounders. The methodological proof that causality is measurable.
Francis Report 1955↗BNT162b2 Phase 3
43,448 participants randomised. 8 COVID cases in vaccine arm vs 162 in placebo arm from 7 days after dose 2. Efficacy 95.0 % (95 % CI 90.3–97.6). Consistent across age, sex, ethnicity, comorbidities.
Polack NEJM 2020↗Catching rare signals
Post-licensing: active surveillance (Swissmedic, VAERS, EMA). Myocarditis signal in young males was confirmed via Bradford-Hill (temporality, specificity, dose-response). Recommendations adjusted; net benefit remained positive.
WHO GACVS 2021↗Vaccine causality doesn't rest on one study but on Phase-3 RCT (internal validity) + real-world observation (external validity) + active surveillance (safety) + mechanistic research (immunology). All applicable Bradford-Hill criteria are satisfied.