Causality & the Bradford Hill Criteria
Bradford Hill's nine criteria are viewpoints — not a checklist — for judging whether an exposure–outcome association is causal. Board essentials: temporality is the only required criterion, dose–response and strength argue strongly for causation, and specificity is weakest — and chance, bias, confounding, and reverse causation must always be excluded first.
From association to causation
An observed association between an exposure and an outcome does not prove causation. Before invoking cause you must first exclude three impostors — chance (random error), bias (systematic error), and confounding (a lurking third variable) — plus reverse causation. In 1965, Sir Austin Bradford Hill, building on the smoking–lung cancer work he did with Richard Doll, proposed nine viewpoints for judging whether an association is causal.
Boards love one nuance: these are aids to judgment, not a rigid checklist. An association need not satisfy all nine, no single criterion is decisive, and only temporality is truly required. The criteria can strengthen — never prove — a causal claim.
- Temporality is the ONLY essential criterion — the exposure MUST precede the outcome.
- Strength: a larger effect size (high relative risk / odds ratio) is harder to explain away by confounding → more likely causal.
- Biological gradient (dose–response): more exposure → more disease; one of the strongest arguments for causation.
- Experiment (an RCT, or disease falling when the exposure is removed) is often the strongest single support for causation.
- Specificity is the WEAKEST criterion — most diseases are multifactorial, so one cause need not map to one effect.
- Criteria are not additive and not a checklist; failing one does not disprove causation.
- Association ≠ causation until chance, bias, confounding, and reverse causation are excluded first.
The nine Bradford Hill criteria
| Criterion | Meaning | Classic example |
|---|---|---|
| Temporality* | Cause precedes effect (only essential) | Smoking documented before cancer |
| Strength | Large RR/OR | RR of lung cancer ~10–20 in smokers |
| Biological gradient | Dose–response | More pack-years → more cancer |
| Consistency | Reproduced across studies/populations | Link seen in many countries |
| Plausibility | Credible biologic mechanism | Carcinogens damage bronchial DNA |
| Coherence | Fits known biology/natural history | Matches lab and pathology data |
| Experiment | Reducing exposure lowers risk | Quitting lowers cancer incidence |
| Specificity | One cause → one effect (weakest) | Often absent — smoking causes many diseases |
| Analogy | Resembles an established cause | Thalidomide → accept other teratogens |
Exclude these before claiming cause
| Explanation | What it is | Classic clue |
|---|---|---|
| Chance | Random error | Small n, wide confidence interval, p > 0.05 |
| Bias | Systematic error in design/measurement | Recall bias, selection bias |
| Confounding | Third variable tied to exposure and outcome | Coffee–cancer link confounded by smoking |
| Reverse causation | Outcome actually drives the 'exposure' | Illness causes inactivity, not vice versa |
| Causation | True cause–effect | Concluded only after excluding the above |
Vignette: A prospective cohort finds lung-cancer incidence rising stepwise with smoking intensity: 5/100,000 in nonsmokers, 50/100,000 in those smoking <1 pack/day, and 140/100,000 in >1 pack/day, with the same trend across increasing pack-years.
Q — Which Bradford Hill criterion is best demonstrated? Answer: Biological gradient (dose–response) — an increasing exposure produces an increasing effect; one of the strongest arguments for a causal relationship.
Reasoning: A clean monotonic dose–response makes uncontrolled confounding unlikely to be the sole explanation and supports causation — especially since the cohort design already satisfies temporality (exposure is measured before disease develops).
Vignette: A cross-sectional survey reports that adults with depression have lower serum vitamin D. The authors conclude that low vitamin D causes depression.
Q — Why is this flawed, and what is the best next step? Flaw: A cross-sectional study measures exposure and outcome at the same moment, so temporality cannot be established — depression could instead reduce sun exposure and dietary intake (reverse causation). Because temporality is the only essential criterion, causation cannot be claimed here.
Best next step: A prospective cohort — measure vitamin D in disease-free subjects and follow for incident depression — or ultimately an RCT of supplementation, to establish that the exposure precedes the outcome.
- Design dictates temporality: cohort and RCT can establish it; cross-sectional and ecological studies cannot.
- Ecological fallacy: group-level correlations (e.g., per-country fat intake vs heart disease) do not prove individual-level causation.
- Confounding vs effect modification: confounding distorts and must be controlled (randomization, restriction, matching, stratification, or multivariable adjustment); effect modification is a real difference across strata that should be reported, not removed.
- Strength cue: the higher the RR/OR, the less plausible that confounding alone explains the association.
- One-liner: Temporality is necessary but not sufficient — the criteria strengthen, never prove, a causal claim.
Practice Research Methods now
Board-style questions, spaced-repetition flashcards, and a Socratic AI tutor — free to start.