Bias and confounding
Bias, Confounding and Causation: Types of Bias and Bradford Hill Criteria
Selection and information bias, confounding, effect modification and the Bradford Hill criteria, with examples of how each is prevented or controlled. Explains why bigger samples reduce chance but not bias, and which biases belong to which study design.
Three explanations for any association
- Before accepting that an exposure causes a disease, rule out three alternatives: chance, bias and confounding. Only then judge causation.
| Source | Nature | How it is handled |
|---|---|---|
| Chance (random error) | Results scatter around the truth | Larger sample; judged by p values and confidence intervals |
| Bias (systematic error) | Results pushed consistently in one direction | Good design; a larger sample does not fix it |
| Confounding | A third factor mixes its effect with the exposure | Design (randomise, restrict, match) and analysis (stratify, adjust) |
- Random error affects precision (reliability); systematic error affects accuracy (validity).
Exam trap Increasing the sample size reduces random error only; a biased study simply gives a more precise wrong answer.
Selection bias
- Arises from how people get into or stay in a study, so the groups compared differ in ways other than the exposure.
| Type | What happens | Typical setting |
|---|---|---|
| Sampling bias | Study population not representative of the target | Non-random or convenience samples |
| Berkson (admission rate) bias | Hospital cases and controls differ in exposure because of admission patterns | Hospital-based case-control |
| Healthy worker effect | Employed people are healthier than the general population, hiding occupational harm | Occupational cohorts |
| Volunteer (self-selection) bias | Volunteers are healthier and more motivated | Surveys, screening uptake |
| Non-response bias | Responders differ from non-responders | Postal surveys |
| Attrition (loss to follow-up) | Drop-outs differ between groups | Cohorts and RCTs |
| Neyman (prevalence–incidence) bias | Using prevalent cases misses rapidly fatal or brief cases | Cross-sectional and case-control |
| Survivorship bias | Only survivors are studied | Retrospective studies |
- Prevented by random sampling, population-based controls, high response rates and full follow-up with intention-to-treat analysis.
Information (measurement) bias
- Arises from how data are collected, so exposure or outcome is measured wrongly or differently between groups.
| Type | What happens | Prevention |
|---|---|---|
| Recall bias | Cases remember exposures better than controls (mothers of malformed babies recalling drugs) | Records, structured questionnaires |
| Interviewer / observer bias | Knowing the group influences questioning or assessment | Blinding |
| Reporting (social desirability) bias | Participants under-report smoking or alcohol | Anonymous questionnaires |
| Hawthorne effect | People change behaviour because they are being watched | Control group, unobtrusive measures |
| Misclassification | Wrong category of exposure or disease; non-differential errors bias towards no effect | Validated tools |
| Performance bias | Groups treated differently apart from the intervention | Blinding, standard protocols |
| Publication bias | Positive studies published more often | Trial registration, funnel plot |
| Regression to the mean | Extreme first values move towards average on retesting | Control group, repeated baseline |
- Placebo effect (improvement from receiving any treatment) is different from the Hawthorne effect (change due to being observed); both are controlled by a comparison group.
Biases in screening
- Lead-time bias: screening finds disease earlier, so survival from diagnosis seems longer even if death occurs at the same time. Judge screening by mortality, not survival time.
- Length-time bias: slow-growing tumours spend longer in the detectable phase, so screening finds a surplus of less aggressive disease.
- Overdiagnosis: detecting disease that would never have caused harm (small prostate cancers, DCIS).
- Volunteer bias: people who attend screening are healthier than those who do not.
- Only an RCT with mortality as the outcome avoids all of these.
Confounding
- A confounder is linked to both the exposure and the outcome, and is not on the causal pathway between them.
- Example: studies show coffee drinkers have more lung cancer. Smoking is associated with coffee drinking and independently causes lung cancer, so smoking confounds the coffee–cancer link. Age confounds most comparisons between populations.
- A factor on the causal pathway (such as raised LDL between diet and heart disease) is a mediator, not a confounder, and should not be adjusted away.
- Control at design stage: randomisation (the only method that balances unknown confounders), restriction (study only non-smokers), matching (each case paired with a control of the same age and sex).
- Control at analysis stage: stratification (Mantel–Haenszel), multivariable regression, standardisation (direct or indirect).
- Effect modification (interaction): the exposure's effect genuinely differs across levels of another factor (asbestos multiplies lung cancer risk far more in smokers). It is a real finding to report by stratum, not a bias to remove.
- Reverse causation: the outcome causes the exposure (people with early illness stop drinking, making abstainers appear less healthy).
- Ecological fallacy: assuming a group-level association applies to individuals.
Remember A confounder is a third factor tied to both exposure and outcome but not between them.
Internal and external validity
- Internal validity: the result is true for the people studied, free of bias and confounding. It comes first.
- External validity (generalisability): the result applies to other populations and settings. Strict inclusion criteria raise internal validity but can reduce external validity.
Judging causation: Bradford Hill criteria
| Criterion | Question asked |
|---|---|
| Strength | Is the RR large? |
| Consistency | Is it found in different studies, places and people? |
| Specificity | Does one exposure lead to one disease? (weakest criterion) |
| Temporality | Did exposure precede disease? The only essential criterion |
| Biological gradient | More exposure, more disease (dose–response)? |
| Plausibility | Is there a biological mechanism? |
| Coherence | Does it fit what is known of the disease? |
| Experiment | Does removing the exposure reduce disease? |
| Analogy | Do similar exposures cause similar effects? |
- Smoking and lung cancer meet nearly every criterion: strong RR, consistent across countries, dose–response with pack-years, falling risk after stopping.
Models of causation
- Epidemiological triad: agent, host and environment interacting (best for infections).
- Web of causation: many interlinked factors for chronic disease such as coronary heart disease.
- Sufficient-component cause model: a disease occurs when a complete set of component causes is present; a necessary cause must be in every set (*M. tuberculosis* for TB), but is rarely sufficient alone.
How FCPS asks about this
- Hospital-based case-control study where admission patterns distort exposure: Berkson bias.
- Occupational cohort appearing healthier than the public: healthy worker effect.
- Mothers of malformed babies recalling drug use more thoroughly: recall bias; reduce with records.
- Participants changing behaviour because they are observed: Hawthorne effect.
- Screening making survival look longer without delaying death: lead-time bias; use mortality.
- Best way to control unknown confounders: randomisation; restriction and matching handle known ones.
- Only essential Bradford Hill criterion: temporality.
Test yourself on this topic
5 of the 16 questions in our bank on this topic. Choose an answer to see why every option is right or wrong.
A screening programme detects many slow-growing tumours that would rarely have caused symptoms, so screen-detected cancers seem to have better survival. What is this bias?
Answer: Length-time bias
Length-time bias: screening at intervals preferentially picks up slowly progressing disease, which has a long preclinical phase, while aggressive cancers appear between screens (interval cancers). Screen-detected cases therefore look less lethal. Its extreme form is overdiagnosis. Lead-time bias: diagnosis is made earlier, so survival from diagnosis seems longer even if death occurs at the same time. Both are avoided by comparing mortality in randomised screened and unscreened groups.
Screening biases
- Lead-time: diagnosed earlier, dies at the same time
- Length-time: catches slow tumours
- Overdiagnosis: finds harmless disease
- Fix: RCT with mortality end point
Remember Lead is about timing; length is about the type of tumour.
Why the other options are wrong
- Lead-time bias Earlier diagnosis without later death.
- Recall bias Case-control studies; patients remember differently.
- Berkson bias Hospital-based case-control studies.
- Hawthorne effect Behaviour changes because people know they are studied.
A study of hypertension recruits the first 200 patients attending a hospital outpatient clinic. What is the main weakness of this sample?
Answer: Selection bias (convenience)
Convenience sampling uses whoever is easiest to reach. Hospital attendees differ from the general population (more severe disease, better access, referral patterns—Berkson's bias in hospital case-control studies), so findings may not be generalisable (poor external validity). Other non-probability methods: purposive (chosen for specific characteristics, used in qualitative research), quota (filling set numbers by category) and snowball (participants recruit others; useful for hidden groups such as drug users or sex workers).
Non-probability samples
- Convenience: easiest to reach → selection bias
- Purposive: chosen for traits (qualitative)
- Quota: fixed numbers per category
- Snowball: hidden populations
Remember Convenient for you, unrepresentative for science.
Why the other options are wrong
- The sample is too large Size is not the problem.
- It is a stratified sample No strata were used.
- It always overestimates BP Direction of bias is unpredictable.
- It cannot be analysed It can be analysed, but generalisation is limited.
A screening programme finds cancers earlier, and survival from diagnosis appears longer, but the age at death is unchanged. What explains this?
Answer: Lead-time bias
Lead-time bias: earlier diagnosis adds time between detection and death without changing when death occurs, so survival looks longer. Length-time bias: screening preferentially detects slow-growing tumours. Proper evaluation uses mortality in randomised screening trials.
Screening biases
- Lead-time: earlier diagnosis, same death
- Length-time: slow tumours found more
- Overdiagnosis: harmless disease found
Remember Starting the clock earlier isn't living longer.
Why the other options are wrong
- Recall bias Case–control studies.
- Selection bias by randomisation Randomisation reduces selection bias.
- Hawthorne effect Behaviour change when observed.
- Confounding by smoking Not explained by confounding.
Which of Bradford Hill's criteria is considered essential for a causal relationship?
Answer: Temporality
Bradford Hill criteria for causation: strength, consistency, specificity, temporality, biological gradient (dose–response), plausibility, coherence, experiment and analogy. Only temporality is essential: the cause must come before the effect.
Bradford Hill criteria
- Essential: temporality
- Strong support: strength, dose–response, consistency
- Others: plausibility, coherence, experiment, analogy, specificity
Remember Cause must come first.
Why the other options are wrong
- Strength of association Supportive, not essential.
- Consistency Supportive.
- Biological gradient Supportive.
- Analogy Weakest criterion.
A study finds that coffee drinkers have more lung cancer. Coffee drinkers in the study also smoke more. What is smoking in this situation?
Answer: A confounder
A confounder is linked to both the exposure (coffee) and the outcome (lung cancer) and is not on the causal pathway, creating a false association. It can be controlled by randomisation, restriction, matching, stratification or multivariable analysis.
Confounding
- Linked to exposure and outcome, not on the pathway
- Control: randomise, restrict, match, stratify, adjust
- Randomisation controls unknown confounders too
Remember Smoking hides behind coffee.
Why the other options are wrong
- An effect modifier only An effect modifier changes the effect size across groups; here smoking explains the association.
- Recall bias Not related to memory.
- A random error Not chance.
- An outcome Smoking is not the outcome here.
Where this sits in the CPSP syllabus
- Paper 1: Epidemiology, Bias, confounding and causation