Skip to content
Scintria
ToolsGrant writerFunding
HomeLiteratureClinical trialsEvidence LabProtocol packData analysisLearnProjectsLibraryReferencesApplications
Start free
ScintriaStart free

Learn / Epidemiology / Bias, confounding and causation

From association to causation

13 minute read · 6-question quiz

Download this lesson as a colour Word or PDF study guide with the figures and answers. Included in paid plans; reading and quizzes stay free.Sign in

What you will learn

  • Apply the Bradford Hill considerations to judge causality

  • Explain the counterfactual idea and why randomised trials are the benchmark

  • Describe the sufficient-component cause model and why diseases have many causes

Epidemiology works with associations but aims at causes. A cause is something whose manipulation changes the outcome. The counterfactual definition says: the effect of an exposure on a person is the difference between what happened to them and what would have happened had they not been exposed. We can never observe both, so we compare groups that are alike except for the exposure. A randomised trial does this best; observational studies try to mimic it.

Before calling an association causal, rule out the alternatives

  1. Chance: use confidence intervals and tests; replication.
  2. Bias: selection and information bias.
  3. Confounding: measured and unmeasured.
  4. Reverse causation: did the disease cause the exposure? (Low weight might be a cause or a sign of early cancer.)

The Bradford Hill considerations (1965)

Sir Austin Bradford Hill proposed nine aspects to weigh. They are not a checklist and only temporality is essential.

Temporality

The cause must come before the effect. The only essential one.

Strength

Large relative risks are less likely to be explained by confounding (smoking and lung cancer, RR about 10 or more).

Dose-response

More exposure, more disease. A gradient strengthens the case.

Consistency

Similar findings in different populations, designs and times.

Plausibility and coherence

A mechanism that makes biological sense and agrees with other knowledge.

Specificity

One cause, one effect. Weak: many causes have several effects (smoking).

Experiment

Evidence from removing the exposure (trials, policy changes, natural experiments).

Analogy

Similar causes known for similar effects (thalidomide and other teratogens).

Biological gradient and reversibility

Disease falls when exposure stops (ex-smokers).

Smoking and lung cancer: how the case was built

Case-control studies in the early 1950s found smokers over-represented among lung cancer patients. Cohort studies (British Doctors Study, from 1951) showed smokers had about ten times the death rate from lung cancer, rising with number of cigarettes (dose-response), falling with cessation (reversibility). Results were consistent across countries and designs, a plausible mechanism (carcinogens in tobacco smoke) emerged, and temporality was clear. Alternative explanations, such as a genetic predisposition causing both smoking and cancer, could not account for the strength and gradient. No single study did it: the accumulated evidence converged.

Why diseases rarely have one cause: sufficient-component causes

Rothman's model pictures a sufficient cause as a complete 'pie' whose slices are component causes. Disease occurs when a pie is completed. A component present in every pie is necessary; most are neither necessary nor sufficient.

  • Tuberculosis: Mycobacterium tuberculosis is a necessary component (no bacterium, no TB) but not sufficient: crowding, HIV, malnutrition, diabetes and smoking complete different pies.
  • This explains why exposed people do not always get disease, unexposed people sometimes do, and why removing any one component of a pie prevents the cases that pie would have caused.
Pie 1HIV + crowdingPie 2Malnutrition +smokingBacterium
Two different sufficient causes of TB (each also needs the bacterium, the necessary cause in the overlap).

Strength of evidence

Judging causation is inference to the best explanation using many lines of evidence, not a single p-value. Randomised trials give the strongest evidence for interventions, and well-conducted cohort studies, consistent across settings, the best for harms that cannot be trialled ethically.

Common errors

'Correlation is not causation' is true, but so is 'no randomised trial, so no causation' as an excuse: smoking, asbestos and many other causes were established without trials. Equally, an observational study with strong adjustment can still be confounded by what was not measured. Say 'associated with' unless the design and the evidence support 'caused'.

Writing about causes

Match your verbs to your design: cross-sectional studies 'were associated with', cohort studies 'predicted', randomised trials 'reduced' or 'caused', and say what could weaken the conclusion.

Summary

  • A cause is something whose change would alter the outcome; the counterfactual idea underlies causal inference.

  • Rule out chance, bias, confounding and reverse causation before inferring cause.

  • Hill's considerations (temporality essential) guide a judgment, not a checklist.

  • Diseases usually result from several component causes completing a sufficient cause.

  • Causal wording should match the strength of the design and evidence.

Check your understanding

Question 1
1

Which Bradford Hill consideration is essential?

Question 2
2

A dose-response gradient between pack-years and lung cancer supports:

Question 3
3

In Rothman's model, a necessary cause is one that:

Question 4
4

The counterfactual for an exposed person is:

Question 5
5

An association that holds in different countries and designs shows:

Question 6
6

Which statement is best supported by a single cross-sectional survey?

Draft study guide. Spotted an error or something unclear? Tell us at hello@scintria.com. This is educational material, not clinical or regulatory advice.

Confounding, mediators, colliders and causal diagramsPrinciples of screening and its hidden biases