Statistics
Bias and Confounding in Research: Types, Examples and How to Control Them
When a study finds an association, there are four possible explanations: chance, bias, confounding, or a true effect
Resora Research Team · 14 Nov 2026 · 5 min read
When a study finds an association, there are four possible explanations: chance, bias, confounding, or a true effect. Chance is handled with sample size and statistics. Bias is a systematic error built into how participants were chosen (selection bias) or how information was collected (information bias), and it cannot be fixed after the study ends. Confounding occurs when a third factor linked to both exposure and outcome distorts the association, for example smoking in a study of coffee and heart disease. Confounding can be controlled at the design stage by randomisation, restriction or matching, and at the analysis stage by stratification or multivariable regression. Prevent bias and confounding by careful design, because analysis alone cannot rescue a flawed study.
Four explanations for any association
Before believing that exposure X causes outcome Y, rule out the alternatives in this order:
- Chance: assessed with p-values and confidence intervals.
- Bias: a systematic error in design or measurement.
- Confounding: a third factor mixing its effect with the exposure.
- Reverse causation: the outcome came first (common in cross-sectional studies).
Only after these are excluded is a causal explanation reasonable. Bias and confounding differ in one crucial way: bias is a flaw in the study and can usually not be corrected afterwards, while confounding can be controlled if the confounder was measured.
Selection bias
Selection bias arises when the people included in the study (or analysis) differ systematically from the target population, or when cases and controls are selected differently.
| Type | What happens | Example |
|---|---|---|
| Berkson's (admission) bias | Hospital patients differ from the community | A hospital-based case–control study shows a false link between two diseases because people with both are more likely to be admitted |
| Non-response / volunteer bias | Those who respond differ from those who do not | A diet survey answered mostly by health-conscious people |
| Healthy worker effect | Employed people are healthier than the general population | A factory cohort shows lower mortality than the community |
| Loss to follow-up | Dropouts differ from those who stay | Sicker patients leave a cohort study, making the treatment look better |
| Survivorship bias | Only survivors are studied | A study of long-term patients misses those who died early |
| Control selection bias | Controls are not comparable to cases | Using hospital patients with another illness as controls for a community exposure |
Information (measurement) bias
Information bias arises when exposure or outcome is measured differently or inaccurately between groups.
| Type | What happens | Example |
|---|---|---|
| Recall bias | Cases remember past exposures more thoroughly than controls | Mothers of children with a birth defect recall every medicine taken in pregnancy; mothers of healthy children forget |
| Interviewer bias | The interviewer probes cases more than controls | An interviewer who knows who is a case asks more detailed questions |
| Observer / detection bias | One group is examined more carefully | Smokers get more chest X-rays, so more disease is found in them |
| Misclassification | People are placed in the wrong exposure or outcome group | A faulty BP machine labels hypertensives as normal |
| Social desirability bias | Participants give acceptable answers | Under-reporting of alcohol or tobacco use |
| Lead-time bias | Screening finds disease earlier, so survival looks longer | Cancer detected 2 years earlier with no change in time of death |
When misclassification is equal in both groups (non-differential), it usually pulls the result toward no association. When it differs between groups (differential), the result can move in either direction.
Confounding
A confounder is a factor that is (1) associated with the exposure, (2) an independent risk factor for the outcome, and (3) not on the causal pathway between exposure and outcome.
Example. A study finds that coffee drinkers have more heart disease. But smokers are more likely to drink coffee (linked to exposure) and smoking causes heart disease (linked to outcome). Smoking may explain part or all of the apparent coffee effect. Compare coffee drinkers with non-drinkers within smokers and within non-smokers, and the association may shrink or vanish.
Common confounders in medical research are age, sex, socio-economic status, smoking, and BMI.
Controlling confounding at the design stage
| Method | How it works | Limit |
|---|---|---|
| Randomisation | Chance spreads known and unknown confounders equally | Only possible in trials |
| Restriction | Include only one category, e.g. non-smokers | Results apply only to that group |
| Matching | Choose controls with the same age and sex as cases | Cannot then study the matched factor; needs matched analysis |
Controlling confounding at the analysis stage
| Method | How it works |
|---|---|
| Stratification | Analyse the association separately in each level of the confounder |
| Mantel–Haenszel adjusted estimate | Combines the strata into one adjusted odds ratio or risk ratio |
| Multivariable regression | Logistic or linear regression adjusts for many confounders at once |
A useful check: if the crude and adjusted estimates differ by more than about 10%, confounding was present.
Confounding is not effect modification
In effect modification (interaction), the true effect of the exposure genuinely differs across groups, for example a drug that works in women but not in men. Unlike confounding, effect modification is a finding to report, not a nuisance to remove. Stratification reveals it: if the stratum-specific estimates differ meaningfully from each other, report them separately.
Residual and unmeasured confounding
Even after adjustment, factors you did not measure, or measured poorly, can still distort results. This is why observational studies can suggest but rarely prove causation, and why randomised trials sit higher in the evidence hierarchy.
Practical checklist
- Define exposure and outcome clearly, using objective measures where possible.
- Choose controls from the same source population as cases.
- Blind interviewers and assessors; use standard questionnaires.
- Achieve high response and follow-up rates; compare responders with non-responders.
- List likely confounders before data collection and measure them.
- Adjust in the analysis and report both crude and adjusted estimates.
Keywords
bias in research, types of bias, selection bias, information bias, recall bias, confounding, confounding variable, how to control confounding, effect modification, Berkson's bias, healthy worker effect, matching and randomisation, bias vs confounding, bias in epidemiology, thesis methodology
References
- Park K. Park's Textbook of Preventive and Social Medicine — chapter on epidemiology.
- Gordis L. Epidemiology. 5th ed. Elsevier; 2014.
- Hulley SB, Cummings SR, et al. Designing Clinical Research. 4th ed. Lippincott Williams & Wilkins; 2013.
- Delgado-Rodríguez M, Llorca J. Bias. Journal of Epidemiology and Community Health. 2004;58:635–641.
- Rothman KJ, Greenland S, Lash TL. Modern Epidemiology. 3rd ed. Lippincott Williams & Wilkins; 2008.
Related services
- Research Topic Selection & RefinementRefinement of a research topic based on feasibility, relevance, available literature and study setting. Final topic approval remains with the researcher/institution.
- Research Question, Objectives & HypothesisSupport to formulate a focused research question, primary/secondary objectives and testable hypothesis where applicable.
- Literature Search StrategySearch-term development, database search strategy and inclusion/exclusion framework for a literature search.
- Preliminary Literature ReviewStructured summary of relevant literature to help identify context and research gaps; scope and number of papers agreed in advance.
- Synopsis Preparation SupportStructured synopsis support based on the researcher's own study idea, institutional format and supplied references; includes one revision.
- Research Protocol Development SupportSupport with protocol structure, rationale, objectives, study design, methodology and analysis plan. Does not include fabricated content or guaranteed ethics approval.
- Protocol ReviewExpert review of an existing protocol for clarity, design, feasibility, outcomes, eligibility criteria and methodological gaps; written feedback.
- Study Design ConsultationConsultation to compare appropriate observational, experimental or other study designs for a defined research question.
- Sample Size CalculationSample size estimate with stated assumptions, formula/software approach and brief justification; final inputs must be confirmed by the researcher/supervisor.
- Power Analysis ConsultationPower and sample-size consultation based on the research question, outcome, assumptions and proposed analysis.