A strong odds ratio or relative risk is a big clue, but numbers alone don't prove that an exposure causes a disease. Over more than a century, scientists have built increasingly powerful frameworks for testing causation.
1880s–1890Koch's Postulates
Written for single-agent infectious diseases, before anyone understood asymptomatic carriers or viruses that couldn't be cultured:
The agent must be found in every case of the disease, and not in healthy hosts.
The agent must be isolated from a host and grown in pure culture.
The cultured agent must cause the disease when introduced into a healthy, susceptible host.
The same agent must be re-isolated from that newly infected host.
Koch's postulates were the first formal guidelines for deciding whether something causes a disease. Their purpose: establish a causal relationship between a microbe and a disease. A classic test question: if a sick animal doesn't have the organism, or healthy animals do, the 1st postulate is violated.
Where Koch's rules break down: They can't explain carriers (who harbor an agent without getting sick, which breaks the 1st postulate), agents that can't be grown in a lab (viruses need living cells, so they can't be grown in pure culture: the 2nd postulate), or diseases (like heart disease or cancer) caused by many factors working together rather than one single agent.
To fix these gaps, modern epidemiology leans on Bradford Hill's criteria (1965): nine features that, together, make a cause-and-effect relationship more believable. No single criterion proves causation alone, and not every criterion needs to be met, with one exception: temporality is the only criterion considered essential. A cause has to come before its effect.
1
Strength
A large relative risk or odds ratio
2
Consistency
Repeated in different studies, places, times
3
Specificity
Exposure links to one specific disease
4
Temporality
The exposure clearly comes before the disease
5
Biological Gradient
More exposure = more disease (dose-response)
6
Plausibility
Makes sense given known biology
7
Coherence
Doesn't conflict with other known facts
8
Experiment
Removing the exposure lowers the risk
9
Analogy
A similar exposure causes a similar effect
A tenth check some tests add:consideration of alternative explanations. Before blaming an exposure, rule out chance, bias and confounding, the way a doctor tests for other diseases before settling on a diagnosis.
🍦 Real-World Example: Ice Cream and Drowning
Every summer, both ice cream sales and drowning deaths go up together. Does eating ice cream cause people to drown? Running this through just two of Hill's criteria shows why not. Plausibility fails, since there's no believable biological mechanism connecting sugar to swimming accidents. And Temporality can't be established either, since the two numbers rise together, with neither clearly coming before the other in a cause-then-effect sequence. What's really happening is a hidden third factor (hot weather) driving both numbers up at once. A hidden factor like that, linked to both the exposure and the disease, is called a confounder: in a study where coffee drinkers seem to get more lung cancer, smoking is the confounder, because coffee drinkers in that group also smoke more. This is the classic case behind every "correlation isn't causation" headline, and exactly the kind of trap Hill's criteria exist to catch.
Evans's Postulates (1976): Epidemiologist Alfred Evans later blended Koch's rigor with Hill's flexibility into a single updated checklist that works for infectious and chronic diseases alike, comparing disease frequency between exposed and unexposed groups, checking that exposure comes before illness, and confirming that removing the exposure lowers risk.
The Sufficient-Component Cause Model ("Causal Pies")
Rothman's causal pie model shows why real diseases rarely have just one cause. Each complete "pie" represents a combination of factors that, together, are sufficient to cause disease, and each pie can be assembled a different way:
Click each component cause below to add it as a slice. Once every slice is in place, the pie is "complete" and disease occurs: then try removing just one slice and watch what happens:
Pie incomplete: 0 of 4 component causes present. No disease yet.
Why this matters for prevention: You don't need to remove every slice of the pie to stop disease. Blocking just one component cause (like cooking food to a safe temperature) breaks the whole pie, even if the agent and host susceptibility are still present.
Necessary vs. sufficient: a whole pie is a sufficient cause: that complete set of factors is enough to produce disease. A component that shows up in every pie for a disease is a necessary cause: the disease can't happen without it, but it isn't always enough alone. The tuberculosis bacterium is necessary for TB, yet most people who carry it never get sick.
Causal Diagrams (DAGs) State/Nats
A directed acyclic graph (DAG) draws each factor as a box and each cause-and-effect link as an arrow pointing from cause to effect ("acyclic": no loops). It shows at a glance what to control for. Here the question is whether exercise lowers heart disease:
🔀
Confounder
A common cause of the exposure and the outcome (Age → Exercise and Age → Heart disease). It opens a false "back-door" path. Control for it.
➡️
Mediator
A step on the causal path (Exercise → Blood pressure → Heart disease). Controlling for it hides part of the real effect, so don't when you want the total effect.
🎯
Collider
A common effect: two arrows point into it (Exercise → Hospital visit ← Heart disease). A collider already blocks its path; controlling for it (say, studying only hospital patients) opens a false link. Don't control for it.
Reading paths: a causal path follows the arrows from exposure to outcome (Exercise → Blood pressure → Heart disease). If you can't get from one variable to another by following arrows forward, there's no causal path between them. To block a non-causal path, control for the confounders on it and leave colliders alone.
GRADE (Grading of Recommendations, Assessment, Development and Evaluation) rates how much to trust the evidence behind a health recommendation: high, moderate, low or very low. Randomized trials start high and observational studies start low; evidence moves down for bias, inconsistent results or imprecision, and up for a large effect or a dose-response. Example: guidelines on when to prescribe antibiotics use GRADE, so doctors don't overprescribe and add to antibiotic resistance.
✓ Check Yourself
✓ Complete
Q12According to the causal pie model, do you need to remove every component cause to stop a disease from occurring?
No. Removing just one component cause (one "slice") breaks that entire sufficient cause, even if the other component causes are still present.