Types of bias in epidemiology describe different ways a study or an evidence base can give a systematically misleading answer. For CPH review, begin with the mechanism: who enters or remains in the study, how variables are measured, or which findings become available.
Bias is not the same as random error. A larger sample can improve precision without repairing a flawed comparison. Confounding is another threat to interpretation; it requires attention to the causal relationships among variables, not just a label for a poorly designed study.
Selection bias: who enters and who remains
Selection bias can arise when recruitment, participation, retention or analysis eligibility creates a distorted exposure–outcome comparison. A sample’s lack of representativeness can limit generalizability, but it does not by itself establish that an association within the study is biased. Ask how selection is related to the exposure, outcome and their causes.
- Nonresponse or volunteer participation: a survey may misestimate population disease prevalence if participation is related to disease status.
- Loss to follow-up: estimates may be biased when the people whose outcomes are observed differ in relevant ways from those whose outcomes are missing.
- Hospital-based controls: in a case-control study, controls admitted for conditions related to the exposure may give an inappropriate exposure comparison. Simply using hospital controls does not automatically establish Berkson’s bias.
- Healthy-worker comparisons: comparing workers with the general population can be misleading because employment and continued employment are related to health.
Example: in a fictional cohort, exposed participants who develop early symptoms are especially likely to leave before their final outcome is recorded. An analysis of completers alone may undercount outcomes in the exposed group. The direction and size of the resulting bias depend on the selection mechanism.
Review the source population, recruitment procedures, eligibility rules and losses to follow-up. Recruitment and retention strategies, appropriate comparison groups and carefully justified weighting can help, but none guarantees that selection bias is absent.
Information bias: how variables are measured
Information bias arises from systematic measurement, recording or classification errors. It can affect exposure, outcome or covariate information.
- Recall bias: cases and controls may reconstruct past exposures differently.
- Interviewer or observer bias: knowledge of exposure, diagnosis or treatment can influence questioning or assessment.
- Misclassification: an exposure or outcome is assigned to the wrong category.
Differential misclassification means that classification errors differ according to another relevant variable, such as exposure errors that differ by outcome status. Nondifferential misclassification does not mean “random error” or guarantee bias toward the null under every set of conditions.
Example: in a fictional case-control study, interviewers ask cases detailed questions about a suspected exposure but ask controls only a short checklist. Even if participants were recruited appropriately, the exposure measurements are not comparable.
Use consistent definitions and collection procedures, train assessors, and assess the performance of measurement tools in the population of interest. Blinding relevant assessors can help when feasible. A tool validated elsewhere is not automatically valid in every language, setting or population.
Publication bias: which findings become available
Publication bias can occur when whether a study becomes available depends on its results. Selective outcome reporting is related but distinct: a study may be available while particular measured outcomes or analyses are withheld.
For example, if investigators complete ten fictional studies but publish only those with favorable findings, a review limited to the published studies may overstate the evidence. If all ten are available and a reviewer deliberately excludes unfavorable eligible studies, the reviewer has introduced a separate selection problem; that scenario alone does not establish that the studies were never published.
Prospective registration and prespecified protocols make missing studies and changed outcomes easier to identify. Complete reporting and searches of registries and other appropriate sources can reduce the problem. Registration does not force publication, and unpublished or gray-literature evidence still needs critical appraisal. See Cochrane’s guidance on missing evidence.
Avoid these common mistakes
- Random sampling and random assignment are different. Sampling concerns entry into a sample; assignment concerns allocation to study groups. Randomizing a volunteer sample does not make it representative of the population.
- Bias can arise after recruitment. Loss to follow-up and analysis restrictions can change the comparison.
- Do not infer the direction of bias from its name alone. Work through the mechanism and the data.
- Statistical significance does not establish validity. A precise estimate can still be systematically misleading.
Practice questions
Question 1: selection during follow-up
In a fictional cohort, exposed participants with early symptoms are more likely to leave the study before their final outcome is measured. Researchers analyze only participants who complete follow-up. Which concern is most directly illustrated?
- Selection bias from differential loss to follow-up
- Publication bias
- A guarantee of effect modification
- Random assignment to treatment
Answer and explanation
A) Selection bias from differential loss to follow-up. Restricting the analysis to completers can distort the exposure–outcome comparison when retention is related to exposure and developing illness. The scenario does not give enough information to quantify the bias.
Question 2: unequal questioning
Interviewers probe cases about past exposures much more thoroughly than controls. What problem should be considered?
Answer and explanation
Information bias. Exposure ascertainment differs between groups. Use comparable questioning procedures and consider blinding interviewers when feasible.
Question 3: missing studies
Investigators decide not to publish completed studies because the primary results are not statistically significant. What problem does this create for an evidence synthesis?
Answer and explanation
Publication bias. Availability depends on the results. Registration can help reviewers identify missing studies, but it cannot by itself make their results available.
Question 4: randomization and representativeness
A trial recruits volunteers from one clinic and randomly assigns them to two interventions. Does random assignment make the sample representative of everyone in the region?
Answer and explanation
No. Random assignment supports an intervention-group comparison within the recruited sample. Generalizability requires a separate assessment of recruitment, population differences and setting.
Sources and related reading
- CDC Field Epidemiology Manual: analyzing and interpreting data
- Cochrane Handbook: considering bias in studies and reviews
- Cochrane Handbook: bias due to missing evidence
- Methods to reduce bias
- Confounding and effect modification
Editorial update — October 11, 2026: consolidated the original and expanded bias guides, clarified selection versus generalizability and misclassification terminology, and revised practice explanations. Examples are hypothetical.
Humanities Moment
The featured image for this CPH Focus article is ‘A glimpse of the Sound,’ Conn by Louis Kinney Harlow. Louis Kinney Harlow was an American painter, etcher, and illustrator best known for his New England coastal scenes, fishermen, boats, farms, and landscapes rendered primarily in watercolor and etching media. He contributed widely distributed chromolithograph illustrations through the Prang Company and helped organize significant Boston-area exhibitions in the 1880s–1890s, making his work both commercially visible and important to regional art circulation. His works are held in institutions such as the Strong Museum and the Whyte Museum of the Canadian Rockies.
Interested in preparing more for the CPH exam? Check out our backlog of CPH Focus articles!
